diff options
| author | Juan Linietsky <reduzio@gmail.com> | 2019-06-07 13:07:57 -0300 |
|---|---|---|
| committer | Juan Linietsky <reduzio@gmail.com> | 2020-02-11 11:53:26 +0100 |
| commit | fd188ddd5176390efb22ed041d8aedad25d81dee (patch) | |
| tree | 18766edace54f7e3e5ed5c1efe1223de163cad86 /thirdparty/glslang/SPIRV | |
| parent | 3e3f8a47616327d7faeb17f558bb81a943385e82 (diff) | |
Initial work on Vulkan:
-Added VulkanContext
-Added an X11 implementation
-Added a rendering device abstraction
-added a Vulkan rendering device abstraction
-Engine does not work, only shows Godot logo (run it from bin/)
Diffstat (limited to 'thirdparty/glslang/SPIRV')
25 files changed, 22427 insertions, 0 deletions
diff --git a/thirdparty/glslang/SPIRV/GLSL.ext.AMD.h b/thirdparty/glslang/SPIRV/GLSL.ext.AMD.h new file mode 100644 index 0000000000..009d2f1cf0 --- /dev/null +++ b/thirdparty/glslang/SPIRV/GLSL.ext.AMD.h @@ -0,0 +1,108 @@ +/* +** Copyright (c) 2014-2016 The Khronos Group Inc. +** +** Permission is hereby granted, free of charge, to any person obtaining a copy +** of this software and/or associated documentation files (the "Materials"), +** to deal in the Materials without restriction, including without limitation +** the rights to use, copy, modify, merge, publish, distribute, sublicense, +** and/or sell copies of the Materials, and to permit persons to whom the +** Materials are furnished to do so, subject to the following conditions: +** +** The above copyright notice and this permission notice shall be included in +** all copies or substantial portions of the Materials. +** +** MODIFICATIONS TO THIS FILE MAY MEAN IT NO LONGER ACCURATELY REFLECTS KHRONOS +** STANDARDS. THE UNMODIFIED, NORMATIVE VERSIONS OF KHRONOS SPECIFICATIONS AND +** HEADER INFORMATION ARE LOCATED AT https://www.khronos.org/registry/ +** +** THE MATERIALS ARE PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS +** OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +** FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL +** THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +** LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +** FROM,OUT OF OR IN CONNECTION WITH THE MATERIALS OR THE USE OR OTHER DEALINGS +** IN THE MATERIALS. +*/ + +#ifndef GLSLextAMD_H +#define GLSLextAMD_H + +static const int GLSLextAMDVersion = 100; +static const int GLSLextAMDRevision = 7; + +// SPV_AMD_shader_ballot +static const char* const E_SPV_AMD_shader_ballot = "SPV_AMD_shader_ballot"; + +enum ShaderBallotAMD { + ShaderBallotBadAMD = 0, // Don't use + + SwizzleInvocationsAMD = 1, + SwizzleInvocationsMaskedAMD = 2, + WriteInvocationAMD = 3, + MbcntAMD = 4, + + ShaderBallotCountAMD +}; + +// SPV_AMD_shader_trinary_minmax +static const char* const E_SPV_AMD_shader_trinary_minmax = "SPV_AMD_shader_trinary_minmax"; + +enum ShaderTrinaryMinMaxAMD { + ShaderTrinaryMinMaxBadAMD = 0, // Don't use + + FMin3AMD = 1, + UMin3AMD = 2, + SMin3AMD = 3, + FMax3AMD = 4, + UMax3AMD = 5, + SMax3AMD = 6, + FMid3AMD = 7, + UMid3AMD = 8, + SMid3AMD = 9, + + ShaderTrinaryMinMaxCountAMD +}; + +// SPV_AMD_shader_explicit_vertex_parameter +static const char* const E_SPV_AMD_shader_explicit_vertex_parameter = "SPV_AMD_shader_explicit_vertex_parameter"; + +enum ShaderExplicitVertexParameterAMD { + ShaderExplicitVertexParameterBadAMD = 0, // Don't use + + InterpolateAtVertexAMD = 1, + + ShaderExplicitVertexParameterCountAMD +}; + +// SPV_AMD_gcn_shader +static const char* const E_SPV_AMD_gcn_shader = "SPV_AMD_gcn_shader"; + +enum GcnShaderAMD { + GcnShaderBadAMD = 0, // Don't use + + CubeFaceIndexAMD = 1, + CubeFaceCoordAMD = 2, + TimeAMD = 3, + + GcnShaderCountAMD +}; + +// SPV_AMD_gpu_shader_half_float +static const char* const E_SPV_AMD_gpu_shader_half_float = "SPV_AMD_gpu_shader_half_float"; + +// SPV_AMD_texture_gather_bias_lod +static const char* const E_SPV_AMD_texture_gather_bias_lod = "SPV_AMD_texture_gather_bias_lod"; + +// SPV_AMD_gpu_shader_int16 +static const char* const E_SPV_AMD_gpu_shader_int16 = "SPV_AMD_gpu_shader_int16"; + +// SPV_AMD_shader_image_load_store_lod +static const char* const E_SPV_AMD_shader_image_load_store_lod = "SPV_AMD_shader_image_load_store_lod"; + +// SPV_AMD_shader_fragment_mask +static const char* const E_SPV_AMD_shader_fragment_mask = "SPV_AMD_shader_fragment_mask"; + +// SPV_AMD_gpu_shader_half_float_fetch +static const char* const E_SPV_AMD_gpu_shader_half_float_fetch = "SPV_AMD_gpu_shader_half_float_fetch"; + +#endif // #ifndef GLSLextAMD_H diff --git a/thirdparty/glslang/SPIRV/GLSL.ext.EXT.h b/thirdparty/glslang/SPIRV/GLSL.ext.EXT.h new file mode 100644 index 0000000000..e29c055b9a --- /dev/null +++ b/thirdparty/glslang/SPIRV/GLSL.ext.EXT.h @@ -0,0 +1,38 @@ +/* +** Copyright (c) 2014-2016 The Khronos Group Inc. +** +** Permission is hereby granted, free of charge, to any person obtaining a copy +** of this software and/or associated documentation files (the "Materials"), +** to deal in the Materials without restriction, including without limitation +** the rights to use, copy, modify, merge, publish, distribute, sublicense, +** and/or sell copies of the Materials, and to permit persons to whom the +** Materials are furnished to do so, subject to the following conditions: +** +** The above copyright notice and this permission notice shall be included in +** all copies or substantial portions of the Materials. +** +** MODIFICATIONS TO THIS FILE MAY MEAN IT NO LONGER ACCURATELY REFLECTS KHRONOS +** STANDARDS. THE UNMODIFIED, NORMATIVE VERSIONS OF KHRONOS SPECIFICATIONS AND +** HEADER INFORMATION ARE LOCATED AT https://www.khronos.org/registry/ +** +** THE MATERIALS ARE PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS +** OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +** FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL +** THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +** LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +** FROM,OUT OF OR IN CONNECTION WITH THE MATERIALS OR THE USE OR OTHER DEALINGS +** IN THE MATERIALS. +*/ + +#ifndef GLSLextEXT_H +#define GLSLextEXT_H + +static const int GLSLextEXTVersion = 100; +static const int GLSLextEXTRevision = 2; + +static const char* const E_SPV_EXT_shader_stencil_export = "SPV_EXT_shader_stencil_export"; +static const char* const E_SPV_EXT_shader_viewport_index_layer = "SPV_EXT_shader_viewport_index_layer"; +static const char* const E_SPV_EXT_fragment_fully_covered = "SPV_EXT_fragment_fully_covered"; +static const char* const E_SPV_EXT_fragment_invocation_density = "SPV_EXT_fragment_invocation_density"; + +#endif // #ifndef GLSLextEXT_H diff --git a/thirdparty/glslang/SPIRV/GLSL.ext.KHR.h b/thirdparty/glslang/SPIRV/GLSL.ext.KHR.h new file mode 100644 index 0000000000..333442bb3e --- /dev/null +++ b/thirdparty/glslang/SPIRV/GLSL.ext.KHR.h @@ -0,0 +1,45 @@ +/* +** Copyright (c) 2014-2016 The Khronos Group Inc. +** +** Permission is hereby granted, free of charge, to any person obtaining a copy +** of this software and/or associated documentation files (the "Materials"), +** to deal in the Materials without restriction, including without limitation +** the rights to use, copy, modify, merge, publish, distribute, sublicense, +** and/or sell copies of the Materials, and to permit persons to whom the +** Materials are furnished to do so, subject to the following conditions: +** +** The above copyright notice and this permission notice shall be included in +** all copies or substantial portions of the Materials. +** +** MODIFICATIONS TO THIS FILE MAY MEAN IT NO LONGER ACCURATELY REFLECTS KHRONOS +** STANDARDS. THE UNMODIFIED, NORMATIVE VERSIONS OF KHRONOS SPECIFICATIONS AND +** HEADER INFORMATION ARE LOCATED AT https://www.khronos.org/registry/ +** +** THE MATERIALS ARE PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS +** OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +** FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL +** THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +** LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +** FROM,OUT OF OR IN CONNECTION WITH THE MATERIALS OR THE USE OR OTHER DEALINGS +** IN THE MATERIALS. +*/ + +#ifndef GLSLextKHR_H +#define GLSLextKHR_H + +static const int GLSLextKHRVersion = 100; +static const int GLSLextKHRRevision = 2; + +static const char* const E_SPV_KHR_shader_ballot = "SPV_KHR_shader_ballot"; +static const char* const E_SPV_KHR_subgroup_vote = "SPV_KHR_subgroup_vote"; +static const char* const E_SPV_KHR_device_group = "SPV_KHR_device_group"; +static const char* const E_SPV_KHR_multiview = "SPV_KHR_multiview"; +static const char* const E_SPV_KHR_shader_draw_parameters = "SPV_KHR_shader_draw_parameters"; +static const char* const E_SPV_KHR_16bit_storage = "SPV_KHR_16bit_storage"; +static const char* const E_SPV_KHR_8bit_storage = "SPV_KHR_8bit_storage"; +static const char* const E_SPV_KHR_storage_buffer_storage_class = "SPV_KHR_storage_buffer_storage_class"; +static const char* const E_SPV_KHR_post_depth_coverage = "SPV_KHR_post_depth_coverage"; +static const char* const E_SPV_KHR_vulkan_memory_model = "SPV_KHR_vulkan_memory_model"; +static const char* const E_SPV_EXT_physical_storage_buffer = "SPV_EXT_physical_storage_buffer"; + +#endif // #ifndef GLSLextKHR_H diff --git a/thirdparty/glslang/SPIRV/GLSL.ext.NV.h b/thirdparty/glslang/SPIRV/GLSL.ext.NV.h new file mode 100644 index 0000000000..ede2c570eb --- /dev/null +++ b/thirdparty/glslang/SPIRV/GLSL.ext.NV.h @@ -0,0 +1,78 @@ +/* +** Copyright (c) 2014-2017 The Khronos Group Inc. +** +** Permission is hereby granted, free of charge, to any person obtaining a copy +** of this software and/or associated documentation files (the "Materials"), +** to deal in the Materials without restriction, including without limitation +** the rights to use, copy, modify, merge, publish, distribute, sublicense, +** and/or sell copies of the Materials, and to permit persons to whom the +** Materials are furnished to do so, subject to the following conditions: +** +** The above copyright notice and this permission notice shall be included in +** all copies or substantial portions of the Materials. +** +** MODIFICATIONS TO THIS FILE MAY MEAN IT NO LONGER ACCURATELY REFLECTS KHRONOS +** STANDARDS. THE UNMODIFIED, NORMATIVE VERSIONS OF KHRONOS SPECIFICATIONS AND +** HEADER INFORMATION ARE LOCATED AT https://www.khronos.org/registry/ +** +** THE MATERIALS ARE PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS +** OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +** FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL +** THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +** LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +** FROM,OUT OF OR IN CONNECTION WITH THE MATERIALS OR THE USE OR OTHER DEALINGS +** IN THE MATERIALS. +*/ + +#ifndef GLSLextNV_H +#define GLSLextNV_H + +enum BuiltIn; +enum Decoration; +enum Op; +enum Capability; + +static const int GLSLextNVVersion = 100; +static const int GLSLextNVRevision = 11; + +//SPV_NV_sample_mask_override_coverage +const char* const E_SPV_NV_sample_mask_override_coverage = "SPV_NV_sample_mask_override_coverage"; + +//SPV_NV_geometry_shader_passthrough +const char* const E_SPV_NV_geometry_shader_passthrough = "SPV_NV_geometry_shader_passthrough"; + +//SPV_NV_viewport_array2 +const char* const E_SPV_NV_viewport_array2 = "SPV_NV_viewport_array2"; +const char* const E_ARB_shader_viewport_layer_array = "SPV_ARB_shader_viewport_layer_array"; + +//SPV_NV_stereo_view_rendering +const char* const E_SPV_NV_stereo_view_rendering = "SPV_NV_stereo_view_rendering"; + +//SPV_NVX_multiview_per_view_attributes +const char* const E_SPV_NVX_multiview_per_view_attributes = "SPV_NVX_multiview_per_view_attributes"; + +//SPV_NV_shader_subgroup_partitioned +const char* const E_SPV_NV_shader_subgroup_partitioned = "SPV_NV_shader_subgroup_partitioned"; + +//SPV_NV_fragment_shader_barycentric +const char* const E_SPV_NV_fragment_shader_barycentric = "SPV_NV_fragment_shader_barycentric"; + +//SPV_NV_compute_shader_derivatives +const char* const E_SPV_NV_compute_shader_derivatives = "SPV_NV_compute_shader_derivatives"; + +//SPV_NV_shader_image_footprint +const char* const E_SPV_NV_shader_image_footprint = "SPV_NV_shader_image_footprint"; + +//SPV_NV_mesh_shader +const char* const E_SPV_NV_mesh_shader = "SPV_NV_mesh_shader"; + +//SPV_NV_raytracing +const char* const E_SPV_NV_ray_tracing = "SPV_NV_ray_tracing"; + +//SPV_NV_shading_rate +const char* const E_SPV_NV_shading_rate = "SPV_NV_shading_rate"; + +//SPV_NV_cooperative_matrix +const char* const E_SPV_NV_cooperative_matrix = "SPV_NV_cooperative_matrix"; + +#endif // #ifndef GLSLextNV_H diff --git a/thirdparty/glslang/SPIRV/GLSL.std.450.h b/thirdparty/glslang/SPIRV/GLSL.std.450.h new file mode 100644 index 0000000000..df31092bec --- /dev/null +++ b/thirdparty/glslang/SPIRV/GLSL.std.450.h @@ -0,0 +1,131 @@ +/* +** Copyright (c) 2014-2016 The Khronos Group Inc. +** +** Permission is hereby granted, free of charge, to any person obtaining a copy +** of this software and/or associated documentation files (the "Materials"), +** to deal in the Materials without restriction, including without limitation +** the rights to use, copy, modify, merge, publish, distribute, sublicense, +** and/or sell copies of the Materials, and to permit persons to whom the +** Materials are furnished to do so, subject to the following conditions: +** +** The above copyright notice and this permission notice shall be included in +** all copies or substantial portions of the Materials. +** +** MODIFICATIONS TO THIS FILE MAY MEAN IT NO LONGER ACCURATELY REFLECTS KHRONOS +** STANDARDS. THE UNMODIFIED, NORMATIVE VERSIONS OF KHRONOS SPECIFICATIONS AND +** HEADER INFORMATION ARE LOCATED AT https://www.khronos.org/registry/ +** +** THE MATERIALS ARE PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS +** OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +** FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL +** THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +** LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +** FROM,OUT OF OR IN CONNECTION WITH THE MATERIALS OR THE USE OR OTHER DEALINGS +** IN THE MATERIALS. +*/ + +#ifndef GLSLstd450_H +#define GLSLstd450_H + +static const int GLSLstd450Version = 100; +static const int GLSLstd450Revision = 1; + +enum GLSLstd450 { + GLSLstd450Bad = 0, // Don't use + + GLSLstd450Round = 1, + GLSLstd450RoundEven = 2, + GLSLstd450Trunc = 3, + GLSLstd450FAbs = 4, + GLSLstd450SAbs = 5, + GLSLstd450FSign = 6, + GLSLstd450SSign = 7, + GLSLstd450Floor = 8, + GLSLstd450Ceil = 9, + GLSLstd450Fract = 10, + + GLSLstd450Radians = 11, + GLSLstd450Degrees = 12, + GLSLstd450Sin = 13, + GLSLstd450Cos = 14, + GLSLstd450Tan = 15, + GLSLstd450Asin = 16, + GLSLstd450Acos = 17, + GLSLstd450Atan = 18, + GLSLstd450Sinh = 19, + GLSLstd450Cosh = 20, + GLSLstd450Tanh = 21, + GLSLstd450Asinh = 22, + GLSLstd450Acosh = 23, + GLSLstd450Atanh = 24, + GLSLstd450Atan2 = 25, + + GLSLstd450Pow = 26, + GLSLstd450Exp = 27, + GLSLstd450Log = 28, + GLSLstd450Exp2 = 29, + GLSLstd450Log2 = 30, + GLSLstd450Sqrt = 31, + GLSLstd450InverseSqrt = 32, + + GLSLstd450Determinant = 33, + GLSLstd450MatrixInverse = 34, + + GLSLstd450Modf = 35, // second operand needs an OpVariable to write to + GLSLstd450ModfStruct = 36, // no OpVariable operand + GLSLstd450FMin = 37, + GLSLstd450UMin = 38, + GLSLstd450SMin = 39, + GLSLstd450FMax = 40, + GLSLstd450UMax = 41, + GLSLstd450SMax = 42, + GLSLstd450FClamp = 43, + GLSLstd450UClamp = 44, + GLSLstd450SClamp = 45, + GLSLstd450FMix = 46, + GLSLstd450IMix = 47, // Reserved + GLSLstd450Step = 48, + GLSLstd450SmoothStep = 49, + + GLSLstd450Fma = 50, + GLSLstd450Frexp = 51, // second operand needs an OpVariable to write to + GLSLstd450FrexpStruct = 52, // no OpVariable operand + GLSLstd450Ldexp = 53, + + GLSLstd450PackSnorm4x8 = 54, + GLSLstd450PackUnorm4x8 = 55, + GLSLstd450PackSnorm2x16 = 56, + GLSLstd450PackUnorm2x16 = 57, + GLSLstd450PackHalf2x16 = 58, + GLSLstd450PackDouble2x32 = 59, + GLSLstd450UnpackSnorm2x16 = 60, + GLSLstd450UnpackUnorm2x16 = 61, + GLSLstd450UnpackHalf2x16 = 62, + GLSLstd450UnpackSnorm4x8 = 63, + GLSLstd450UnpackUnorm4x8 = 64, + GLSLstd450UnpackDouble2x32 = 65, + + GLSLstd450Length = 66, + GLSLstd450Distance = 67, + GLSLstd450Cross = 68, + GLSLstd450Normalize = 69, + GLSLstd450FaceForward = 70, + GLSLstd450Reflect = 71, + GLSLstd450Refract = 72, + + GLSLstd450FindILsb = 73, + GLSLstd450FindSMsb = 74, + GLSLstd450FindUMsb = 75, + + GLSLstd450InterpolateAtCentroid = 76, + GLSLstd450InterpolateAtSample = 77, + GLSLstd450InterpolateAtOffset = 78, + + GLSLstd450NMin = 79, + GLSLstd450NMax = 80, + GLSLstd450NClamp = 81, + + GLSLstd450Count +}; + +#endif // #ifndef GLSLstd450_H diff --git a/thirdparty/glslang/SPIRV/GlslangToSpv.cpp b/thirdparty/glslang/SPIRV/GlslangToSpv.cpp new file mode 100644 index 0000000000..4ef6cd7fc1 --- /dev/null +++ b/thirdparty/glslang/SPIRV/GlslangToSpv.cpp @@ -0,0 +1,8066 @@ +// +// Copyright (C) 2014-2016 LunarG, Inc. +// Copyright (C) 2015-2018 Google, Inc. +// Copyright (C) 2017 ARM Limited. +// +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions +// are met: +// +// Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// +// Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following +// disclaimer in the documentation and/or other materials provided +// with the distribution. +// +// Neither the name of 3Dlabs Inc. Ltd. nor the names of its +// contributors may be used to endorse or promote products derived +// from this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS +// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE +// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, +// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, +// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; +// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN +// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +// POSSIBILITY OF SUCH DAMAGE. + +// +// Visit the nodes in the glslang intermediate tree representation to +// translate them to SPIR-V. +// + +#include "spirv.hpp" +#include "GlslangToSpv.h" +#include "SpvBuilder.h" +namespace spv { + #include "GLSL.std.450.h" + #include "GLSL.ext.KHR.h" + #include "GLSL.ext.EXT.h" +#ifdef AMD_EXTENSIONS + #include "GLSL.ext.AMD.h" +#endif + #include "GLSL.ext.NV.h" +} + +// Glslang includes +#include "../glslang/MachineIndependent/localintermediate.h" +#include "../glslang/MachineIndependent/SymbolTable.h" +#include "../glslang/Include/Common.h" +#include "../glslang/Include/revision.h" + +#include <fstream> +#include <iomanip> +#include <list> +#include <map> +#include <stack> +#include <string> +#include <vector> + +namespace { + +namespace { +class SpecConstantOpModeGuard { +public: + SpecConstantOpModeGuard(spv::Builder* builder) + : builder_(builder) { + previous_flag_ = builder->isInSpecConstCodeGenMode(); + } + ~SpecConstantOpModeGuard() { + previous_flag_ ? builder_->setToSpecConstCodeGenMode() + : builder_->setToNormalCodeGenMode(); + } + void turnOnSpecConstantOpMode() { + builder_->setToSpecConstCodeGenMode(); + } + +private: + spv::Builder* builder_; + bool previous_flag_; +}; + +struct OpDecorations { + spv::Decoration precision; + spv::Decoration noContraction; + spv::Decoration nonUniform; +}; + +} // namespace + +// +// The main holder of information for translating glslang to SPIR-V. +// +// Derives from the AST walking base class. +// +class TGlslangToSpvTraverser : public glslang::TIntermTraverser { +public: + TGlslangToSpvTraverser(unsigned int spvVersion, const glslang::TIntermediate*, spv::SpvBuildLogger* logger, + glslang::SpvOptions& options); + virtual ~TGlslangToSpvTraverser() { } + + bool visitAggregate(glslang::TVisit, glslang::TIntermAggregate*); + bool visitBinary(glslang::TVisit, glslang::TIntermBinary*); + void visitConstantUnion(glslang::TIntermConstantUnion*); + bool visitSelection(glslang::TVisit, glslang::TIntermSelection*); + bool visitSwitch(glslang::TVisit, glslang::TIntermSwitch*); + void visitSymbol(glslang::TIntermSymbol* symbol); + bool visitUnary(glslang::TVisit, glslang::TIntermUnary*); + bool visitLoop(glslang::TVisit, glslang::TIntermLoop*); + bool visitBranch(glslang::TVisit visit, glslang::TIntermBranch*); + + void finishSpv(); + void dumpSpv(std::vector<unsigned int>& out); + +protected: + TGlslangToSpvTraverser(TGlslangToSpvTraverser&); + TGlslangToSpvTraverser& operator=(TGlslangToSpvTraverser&); + + spv::Decoration TranslateInterpolationDecoration(const glslang::TQualifier& qualifier); + spv::Decoration TranslateAuxiliaryStorageDecoration(const glslang::TQualifier& qualifier); + spv::Decoration TranslateNonUniformDecoration(const glslang::TQualifier& qualifier); + spv::Builder::AccessChain::CoherentFlags TranslateCoherent(const glslang::TType& type); + spv::MemoryAccessMask TranslateMemoryAccess(const spv::Builder::AccessChain::CoherentFlags &coherentFlags); + spv::ImageOperandsMask TranslateImageOperands(const spv::Builder::AccessChain::CoherentFlags &coherentFlags); + spv::Scope TranslateMemoryScope(const spv::Builder::AccessChain::CoherentFlags &coherentFlags); + spv::BuiltIn TranslateBuiltInDecoration(glslang::TBuiltInVariable, bool memberDeclaration); + spv::ImageFormat TranslateImageFormat(const glslang::TType& type); + spv::SelectionControlMask TranslateSelectionControl(const glslang::TIntermSelection&) const; + spv::SelectionControlMask TranslateSwitchControl(const glslang::TIntermSwitch&) const; + spv::LoopControlMask TranslateLoopControl(const glslang::TIntermLoop&, std::vector<unsigned int>& operands) const; + spv::StorageClass TranslateStorageClass(const glslang::TType&); + void addIndirectionIndexCapabilities(const glslang::TType& baseType, const glslang::TType& indexType); + spv::Id createSpvVariable(const glslang::TIntermSymbol*); + spv::Id getSampledType(const glslang::TSampler&); + spv::Id getInvertedSwizzleType(const glslang::TIntermTyped&); + spv::Id createInvertedSwizzle(spv::Decoration precision, const glslang::TIntermTyped&, spv::Id parentResult); + void convertSwizzle(const glslang::TIntermAggregate&, std::vector<unsigned>& swizzle); + spv::Id convertGlslangToSpvType(const glslang::TType& type, bool forwardReferenceOnly = false); + spv::Id convertGlslangToSpvType(const glslang::TType& type, glslang::TLayoutPacking, const glslang::TQualifier&, + bool lastBufferBlockMember, bool forwardReferenceOnly = false); + bool filterMember(const glslang::TType& member); + spv::Id convertGlslangStructToSpvType(const glslang::TType&, const glslang::TTypeList* glslangStruct, + glslang::TLayoutPacking, const glslang::TQualifier&); + void decorateStructType(const glslang::TType&, const glslang::TTypeList* glslangStruct, glslang::TLayoutPacking, + const glslang::TQualifier&, spv::Id); + spv::Id makeArraySizeId(const glslang::TArraySizes&, int dim); + spv::Id accessChainLoad(const glslang::TType& type); + void accessChainStore(const glslang::TType& type, spv::Id rvalue); + void multiTypeStore(const glslang::TType&, spv::Id rValue); + glslang::TLayoutPacking getExplicitLayout(const glslang::TType& type) const; + int getArrayStride(const glslang::TType& arrayType, glslang::TLayoutPacking, glslang::TLayoutMatrix); + int getMatrixStride(const glslang::TType& matrixType, glslang::TLayoutPacking, glslang::TLayoutMatrix); + void updateMemberOffset(const glslang::TType& structType, const glslang::TType& memberType, int& currentOffset, + int& nextOffset, glslang::TLayoutPacking, glslang::TLayoutMatrix); + void declareUseOfStructMember(const glslang::TTypeList& members, int glslangMember); + + bool isShaderEntryPoint(const glslang::TIntermAggregate* node); + bool writableParam(glslang::TStorageQualifier) const; + bool originalParam(glslang::TStorageQualifier, const glslang::TType&, bool implicitThisParam); + void makeFunctions(const glslang::TIntermSequence&); + void makeGlobalInitializers(const glslang::TIntermSequence&); + void visitFunctions(const glslang::TIntermSequence&); + void handleFunctionEntry(const glslang::TIntermAggregate* node); + void translateArguments(const glslang::TIntermAggregate& node, std::vector<spv::Id>& arguments); + void translateArguments(glslang::TIntermUnary& node, std::vector<spv::Id>& arguments); + spv::Id createImageTextureFunctionCall(glslang::TIntermOperator* node); + spv::Id handleUserFunctionCall(const glslang::TIntermAggregate*); + + spv::Id createBinaryOperation(glslang::TOperator op, OpDecorations&, spv::Id typeId, spv::Id left, spv::Id right, + glslang::TBasicType typeProxy, bool reduceComparison = true); + spv::Id createBinaryMatrixOperation(spv::Op, OpDecorations&, spv::Id typeId, spv::Id left, spv::Id right); + spv::Id createUnaryOperation(glslang::TOperator op, OpDecorations&, spv::Id typeId, spv::Id operand, + glslang::TBasicType typeProxy); + spv::Id createUnaryMatrixOperation(spv::Op op, OpDecorations&, spv::Id typeId, spv::Id operand, + glslang::TBasicType typeProxy); + spv::Id createConversion(glslang::TOperator op, OpDecorations&, spv::Id destTypeId, spv::Id operand, + glslang::TBasicType typeProxy); + spv::Id createIntWidthConversion(glslang::TOperator op, spv::Id operand, int vectorSize); + spv::Id makeSmearedConstant(spv::Id constant, int vectorSize); + spv::Id createAtomicOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy); + spv::Id createInvocationsOperation(glslang::TOperator op, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy); + spv::Id CreateInvocationsVectorOperation(spv::Op op, spv::GroupOperation groupOperation, spv::Id typeId, std::vector<spv::Id>& operands); + spv::Id createSubgroupOperation(glslang::TOperator op, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy); + spv::Id createMiscOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy); + spv::Id createNoArgOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId); + spv::Id getSymbolId(const glslang::TIntermSymbol* node); +#ifdef NV_EXTENSIONS + void addMeshNVDecoration(spv::Id id, int member, const glslang::TQualifier & qualifier); +#endif + spv::Id createSpvConstant(const glslang::TIntermTyped&); + spv::Id createSpvConstantFromConstUnionArray(const glslang::TType& type, const glslang::TConstUnionArray&, int& nextConst, bool specConstant); + bool isTrivialLeaf(const glslang::TIntermTyped* node); + bool isTrivial(const glslang::TIntermTyped* node); + spv::Id createShortCircuit(glslang::TOperator, glslang::TIntermTyped& left, glslang::TIntermTyped& right); +#ifdef AMD_EXTENSIONS + spv::Id getExtBuiltins(const char* name); +#endif + void addPre13Extension(const char* ext) + { + if (builder.getSpvVersion() < glslang::EShTargetSpv_1_3) + builder.addExtension(ext); + } + + glslang::SpvOptions& options; + spv::Function* shaderEntry; + spv::Function* currentFunction; + spv::Instruction* entryPoint; + int sequenceDepth; + + spv::SpvBuildLogger* logger; + + // There is a 1:1 mapping between a spv builder and a module; this is thread safe + spv::Builder builder; + bool inEntryPoint; + bool entryPointTerminated; + bool linkageOnly; // true when visiting the set of objects in the AST present only for establishing interface, whether or not they were statically used + std::set<spv::Id> iOSet; // all input/output variables from either static use or declaration of interface + const glslang::TIntermediate* glslangIntermediate; + spv::Id stdBuiltins; + std::unordered_map<const char*, spv::Id> extBuiltinMap; + + std::unordered_map<int, spv::Id> symbolValues; + std::unordered_set<int> rValueParameters; // set of formal function parameters passed as rValues, rather than a pointer + std::unordered_map<std::string, spv::Function*> functionMap; + std::unordered_map<const glslang::TTypeList*, spv::Id> structMap[glslang::ElpCount][glslang::ElmCount]; + // for mapping glslang block indices to spv indices (e.g., due to hidden members): + std::unordered_map<const glslang::TTypeList*, std::vector<int> > memberRemapper; + std::stack<bool> breakForLoop; // false means break for switch + std::unordered_map<std::string, const glslang::TIntermSymbol*> counterOriginator; + // Map pointee types for EbtReference to their forward pointers + std::map<const glslang::TType *, spv::Id> forwardPointers; +}; + +// +// Helper functions for translating glslang representations to SPIR-V enumerants. +// + +// Translate glslang profile to SPIR-V source language. +spv::SourceLanguage TranslateSourceLanguage(glslang::EShSource source, EProfile profile) +{ + switch (source) { + case glslang::EShSourceGlsl: + switch (profile) { + case ENoProfile: + case ECoreProfile: + case ECompatibilityProfile: + return spv::SourceLanguageGLSL; + case EEsProfile: + return spv::SourceLanguageESSL; + default: + return spv::SourceLanguageUnknown; + } + case glslang::EShSourceHlsl: + return spv::SourceLanguageHLSL; + default: + return spv::SourceLanguageUnknown; + } +} + +// Translate glslang language (stage) to SPIR-V execution model. +spv::ExecutionModel TranslateExecutionModel(EShLanguage stage) +{ + switch (stage) { + case EShLangVertex: return spv::ExecutionModelVertex; + case EShLangTessControl: return spv::ExecutionModelTessellationControl; + case EShLangTessEvaluation: return spv::ExecutionModelTessellationEvaluation; + case EShLangGeometry: return spv::ExecutionModelGeometry; + case EShLangFragment: return spv::ExecutionModelFragment; + case EShLangCompute: return spv::ExecutionModelGLCompute; +#ifdef NV_EXTENSIONS + case EShLangRayGenNV: return spv::ExecutionModelRayGenerationNV; + case EShLangIntersectNV: return spv::ExecutionModelIntersectionNV; + case EShLangAnyHitNV: return spv::ExecutionModelAnyHitNV; + case EShLangClosestHitNV: return spv::ExecutionModelClosestHitNV; + case EShLangMissNV: return spv::ExecutionModelMissNV; + case EShLangCallableNV: return spv::ExecutionModelCallableNV; + case EShLangTaskNV: return spv::ExecutionModelTaskNV; + case EShLangMeshNV: return spv::ExecutionModelMeshNV; +#endif + default: + assert(0); + return spv::ExecutionModelFragment; + } +} + +// Translate glslang sampler type to SPIR-V dimensionality. +spv::Dim TranslateDimensionality(const glslang::TSampler& sampler) +{ + switch (sampler.dim) { + case glslang::Esd1D: return spv::Dim1D; + case glslang::Esd2D: return spv::Dim2D; + case glslang::Esd3D: return spv::Dim3D; + case glslang::EsdCube: return spv::DimCube; + case glslang::EsdRect: return spv::DimRect; + case glslang::EsdBuffer: return spv::DimBuffer; + case glslang::EsdSubpass: return spv::DimSubpassData; + default: + assert(0); + return spv::Dim2D; + } +} + +// Translate glslang precision to SPIR-V precision decorations. +spv::Decoration TranslatePrecisionDecoration(glslang::TPrecisionQualifier glslangPrecision) +{ + switch (glslangPrecision) { + case glslang::EpqLow: return spv::DecorationRelaxedPrecision; + case glslang::EpqMedium: return spv::DecorationRelaxedPrecision; + default: + return spv::NoPrecision; + } +} + +// Translate glslang type to SPIR-V precision decorations. +spv::Decoration TranslatePrecisionDecoration(const glslang::TType& type) +{ + return TranslatePrecisionDecoration(type.getQualifier().precision); +} + +// Translate glslang type to SPIR-V block decorations. +spv::Decoration TranslateBlockDecoration(const glslang::TType& type, bool useStorageBuffer) +{ + if (type.getBasicType() == glslang::EbtBlock) { + switch (type.getQualifier().storage) { + case glslang::EvqUniform: return spv::DecorationBlock; + case glslang::EvqBuffer: return useStorageBuffer ? spv::DecorationBlock : spv::DecorationBufferBlock; + case glslang::EvqVaryingIn: return spv::DecorationBlock; + case glslang::EvqVaryingOut: return spv::DecorationBlock; +#ifdef NV_EXTENSIONS + case glslang::EvqPayloadNV: return spv::DecorationBlock; + case glslang::EvqPayloadInNV: return spv::DecorationBlock; + case glslang::EvqHitAttrNV: return spv::DecorationBlock; + case glslang::EvqCallableDataNV: return spv::DecorationBlock; + case glslang::EvqCallableDataInNV: return spv::DecorationBlock; +#endif + default: + assert(0); + break; + } + } + + return spv::DecorationMax; +} + +// Translate glslang type to SPIR-V memory decorations. +void TranslateMemoryDecoration(const glslang::TQualifier& qualifier, std::vector<spv::Decoration>& memory, bool useVulkanMemoryModel) +{ + if (!useVulkanMemoryModel) { + if (qualifier.coherent) + memory.push_back(spv::DecorationCoherent); + if (qualifier.volatil) { + memory.push_back(spv::DecorationVolatile); + memory.push_back(spv::DecorationCoherent); + } + } + if (qualifier.restrict) + memory.push_back(spv::DecorationRestrict); + if (qualifier.readonly) + memory.push_back(spv::DecorationNonWritable); + if (qualifier.writeonly) + memory.push_back(spv::DecorationNonReadable); +} + +// Translate glslang type to SPIR-V layout decorations. +spv::Decoration TranslateLayoutDecoration(const glslang::TType& type, glslang::TLayoutMatrix matrixLayout) +{ + if (type.isMatrix()) { + switch (matrixLayout) { + case glslang::ElmRowMajor: + return spv::DecorationRowMajor; + case glslang::ElmColumnMajor: + return spv::DecorationColMajor; + default: + // opaque layouts don't need a majorness + return spv::DecorationMax; + } + } else { + switch (type.getBasicType()) { + default: + return spv::DecorationMax; + break; + case glslang::EbtBlock: + switch (type.getQualifier().storage) { + case glslang::EvqUniform: + case glslang::EvqBuffer: + switch (type.getQualifier().layoutPacking) { + case glslang::ElpShared: return spv::DecorationGLSLShared; + case glslang::ElpPacked: return spv::DecorationGLSLPacked; + default: + return spv::DecorationMax; + } + case glslang::EvqVaryingIn: + case glslang::EvqVaryingOut: + if (type.getQualifier().isTaskMemory()) { + switch (type.getQualifier().layoutPacking) { + case glslang::ElpShared: return spv::DecorationGLSLShared; + case glslang::ElpPacked: return spv::DecorationGLSLPacked; + default: break; + } + } else { + assert(type.getQualifier().layoutPacking == glslang::ElpNone); + } + return spv::DecorationMax; +#ifdef NV_EXTENSIONS + case glslang::EvqPayloadNV: + case glslang::EvqPayloadInNV: + case glslang::EvqHitAttrNV: + case glslang::EvqCallableDataNV: + case glslang::EvqCallableDataInNV: + return spv::DecorationMax; +#endif + default: + assert(0); + return spv::DecorationMax; + } + } + } +} + +// Translate glslang type to SPIR-V interpolation decorations. +// Returns spv::DecorationMax when no decoration +// should be applied. +spv::Decoration TGlslangToSpvTraverser::TranslateInterpolationDecoration(const glslang::TQualifier& qualifier) +{ + if (qualifier.smooth) + // Smooth decoration doesn't exist in SPIR-V 1.0 + return spv::DecorationMax; + else if (qualifier.nopersp) + return spv::DecorationNoPerspective; + else if (qualifier.flat) + return spv::DecorationFlat; +#ifdef AMD_EXTENSIONS + else if (qualifier.explicitInterp) { + builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter); + return spv::DecorationExplicitInterpAMD; + } +#endif + else + return spv::DecorationMax; +} + +// Translate glslang type to SPIR-V auxiliary storage decorations. +// Returns spv::DecorationMax when no decoration +// should be applied. +spv::Decoration TGlslangToSpvTraverser::TranslateAuxiliaryStorageDecoration(const glslang::TQualifier& qualifier) +{ + if (qualifier.patch) + return spv::DecorationPatch; + else if (qualifier.centroid) + return spv::DecorationCentroid; + else if (qualifier.sample) { + builder.addCapability(spv::CapabilitySampleRateShading); + return spv::DecorationSample; + } else + return spv::DecorationMax; +} + +// If glslang type is invariant, return SPIR-V invariant decoration. +spv::Decoration TranslateInvariantDecoration(const glslang::TQualifier& qualifier) +{ + if (qualifier.invariant) + return spv::DecorationInvariant; + else + return spv::DecorationMax; +} + +// If glslang type is noContraction, return SPIR-V NoContraction decoration. +spv::Decoration TranslateNoContractionDecoration(const glslang::TQualifier& qualifier) +{ + if (qualifier.noContraction) + return spv::DecorationNoContraction; + else + return spv::DecorationMax; +} + +// If glslang type is nonUniform, return SPIR-V NonUniform decoration. +spv::Decoration TGlslangToSpvTraverser::TranslateNonUniformDecoration(const glslang::TQualifier& qualifier) +{ + if (qualifier.isNonUniform()) { + builder.addExtension("SPV_EXT_descriptor_indexing"); + builder.addCapability(spv::CapabilityShaderNonUniformEXT); + return spv::DecorationNonUniformEXT; + } else + return spv::DecorationMax; +} + +spv::MemoryAccessMask TGlslangToSpvTraverser::TranslateMemoryAccess(const spv::Builder::AccessChain::CoherentFlags &coherentFlags) +{ + if (!glslangIntermediate->usingVulkanMemoryModel() || coherentFlags.isImage) { + return spv::MemoryAccessMaskNone; + } + spv::MemoryAccessMask mask = spv::MemoryAccessMaskNone; + if (coherentFlags.volatil || + coherentFlags.coherent || + coherentFlags.devicecoherent || + coherentFlags.queuefamilycoherent || + coherentFlags.workgroupcoherent || + coherentFlags.subgroupcoherent) { + mask = mask | spv::MemoryAccessMakePointerAvailableKHRMask | + spv::MemoryAccessMakePointerVisibleKHRMask; + } + if (coherentFlags.nonprivate) { + mask = mask | spv::MemoryAccessNonPrivatePointerKHRMask; + } + if (coherentFlags.volatil) { + mask = mask | spv::MemoryAccessVolatileMask; + } + if (mask != spv::MemoryAccessMaskNone) { + builder.addCapability(spv::CapabilityVulkanMemoryModelKHR); + } + return mask; +} + +spv::ImageOperandsMask TGlslangToSpvTraverser::TranslateImageOperands(const spv::Builder::AccessChain::CoherentFlags &coherentFlags) +{ + if (!glslangIntermediate->usingVulkanMemoryModel()) { + return spv::ImageOperandsMaskNone; + } + spv::ImageOperandsMask mask = spv::ImageOperandsMaskNone; + if (coherentFlags.volatil || + coherentFlags.coherent || + coherentFlags.devicecoherent || + coherentFlags.queuefamilycoherent || + coherentFlags.workgroupcoherent || + coherentFlags.subgroupcoherent) { + mask = mask | spv::ImageOperandsMakeTexelAvailableKHRMask | + spv::ImageOperandsMakeTexelVisibleKHRMask; + } + if (coherentFlags.nonprivate) { + mask = mask | spv::ImageOperandsNonPrivateTexelKHRMask; + } + if (coherentFlags.volatil) { + mask = mask | spv::ImageOperandsVolatileTexelKHRMask; + } + if (mask != spv::ImageOperandsMaskNone) { + builder.addCapability(spv::CapabilityVulkanMemoryModelKHR); + } + return mask; +} + +spv::Builder::AccessChain::CoherentFlags TGlslangToSpvTraverser::TranslateCoherent(const glslang::TType& type) +{ + spv::Builder::AccessChain::CoherentFlags flags; + flags.coherent = type.getQualifier().coherent; + flags.devicecoherent = type.getQualifier().devicecoherent; + flags.queuefamilycoherent = type.getQualifier().queuefamilycoherent; + // shared variables are implicitly workgroupcoherent in GLSL. + flags.workgroupcoherent = type.getQualifier().workgroupcoherent || + type.getQualifier().storage == glslang::EvqShared; + flags.subgroupcoherent = type.getQualifier().subgroupcoherent; + flags.volatil = type.getQualifier().volatil; + // *coherent variables are implicitly nonprivate in GLSL + flags.nonprivate = type.getQualifier().nonprivate || + flags.subgroupcoherent || + flags.workgroupcoherent || + flags.queuefamilycoherent || + flags.devicecoherent || + flags.coherent || + flags.volatil; + flags.isImage = type.getBasicType() == glslang::EbtSampler; + return flags; +} + +spv::Scope TGlslangToSpvTraverser::TranslateMemoryScope(const spv::Builder::AccessChain::CoherentFlags &coherentFlags) +{ + spv::Scope scope; + if (coherentFlags.volatil || coherentFlags.coherent) { + // coherent defaults to Device scope in the old model, QueueFamilyKHR scope in the new model + scope = glslangIntermediate->usingVulkanMemoryModel() ? spv::ScopeQueueFamilyKHR : spv::ScopeDevice; + } else if (coherentFlags.devicecoherent) { + scope = spv::ScopeDevice; + } else if (coherentFlags.queuefamilycoherent) { + scope = spv::ScopeQueueFamilyKHR; + } else if (coherentFlags.workgroupcoherent) { + scope = spv::ScopeWorkgroup; + } else if (coherentFlags.subgroupcoherent) { + scope = spv::ScopeSubgroup; + } else { + scope = spv::ScopeMax; + } + if (glslangIntermediate->usingVulkanMemoryModel() && scope == spv::ScopeDevice) { + builder.addCapability(spv::CapabilityVulkanMemoryModelDeviceScopeKHR); + } + return scope; +} + +// Translate a glslang built-in variable to a SPIR-V built in decoration. Also generate +// associated capabilities when required. For some built-in variables, a capability +// is generated only when using the variable in an executable instruction, but not when +// just declaring a struct member variable with it. This is true for PointSize, +// ClipDistance, and CullDistance. +spv::BuiltIn TGlslangToSpvTraverser::TranslateBuiltInDecoration(glslang::TBuiltInVariable builtIn, bool memberDeclaration) +{ + switch (builtIn) { + case glslang::EbvPointSize: + // Defer adding the capability until the built-in is actually used. + if (! memberDeclaration) { + switch (glslangIntermediate->getStage()) { + case EShLangGeometry: + builder.addCapability(spv::CapabilityGeometryPointSize); + break; + case EShLangTessControl: + case EShLangTessEvaluation: + builder.addCapability(spv::CapabilityTessellationPointSize); + break; + default: + break; + } + } + return spv::BuiltInPointSize; + + // These *Distance capabilities logically belong here, but if the member is declared and + // then never used, consumers of SPIR-V prefer the capability not be declared. + // They are now generated when used, rather than here when declared. + // Potentially, the specification should be more clear what the minimum + // use needed is to trigger the capability. + // + case glslang::EbvClipDistance: + if (!memberDeclaration) + builder.addCapability(spv::CapabilityClipDistance); + return spv::BuiltInClipDistance; + + case glslang::EbvCullDistance: + if (!memberDeclaration) + builder.addCapability(spv::CapabilityCullDistance); + return spv::BuiltInCullDistance; + + case glslang::EbvViewportIndex: + builder.addCapability(spv::CapabilityMultiViewport); + if (glslangIntermediate->getStage() == EShLangVertex || + glslangIntermediate->getStage() == EShLangTessControl || + glslangIntermediate->getStage() == EShLangTessEvaluation) { + + builder.addExtension(spv::E_SPV_EXT_shader_viewport_index_layer); + builder.addCapability(spv::CapabilityShaderViewportIndexLayerEXT); + } + return spv::BuiltInViewportIndex; + + case glslang::EbvSampleId: + builder.addCapability(spv::CapabilitySampleRateShading); + return spv::BuiltInSampleId; + + case glslang::EbvSamplePosition: + builder.addCapability(spv::CapabilitySampleRateShading); + return spv::BuiltInSamplePosition; + + case glslang::EbvSampleMask: + return spv::BuiltInSampleMask; + + case glslang::EbvLayer: +#ifdef NV_EXTENSIONS + if (glslangIntermediate->getStage() == EShLangMeshNV) { + return spv::BuiltInLayer; + } +#endif + builder.addCapability(spv::CapabilityGeometry); + if (glslangIntermediate->getStage() == EShLangVertex || + glslangIntermediate->getStage() == EShLangTessControl || + glslangIntermediate->getStage() == EShLangTessEvaluation) { + + builder.addExtension(spv::E_SPV_EXT_shader_viewport_index_layer); + builder.addCapability(spv::CapabilityShaderViewportIndexLayerEXT); + } + return spv::BuiltInLayer; + + case glslang::EbvPosition: return spv::BuiltInPosition; + case glslang::EbvVertexId: return spv::BuiltInVertexId; + case glslang::EbvInstanceId: return spv::BuiltInInstanceId; + case glslang::EbvVertexIndex: return spv::BuiltInVertexIndex; + case glslang::EbvInstanceIndex: return spv::BuiltInInstanceIndex; + + case glslang::EbvBaseVertex: + addPre13Extension(spv::E_SPV_KHR_shader_draw_parameters); + builder.addCapability(spv::CapabilityDrawParameters); + return spv::BuiltInBaseVertex; + + case glslang::EbvBaseInstance: + addPre13Extension(spv::E_SPV_KHR_shader_draw_parameters); + builder.addCapability(spv::CapabilityDrawParameters); + return spv::BuiltInBaseInstance; + + case glslang::EbvDrawId: + addPre13Extension(spv::E_SPV_KHR_shader_draw_parameters); + builder.addCapability(spv::CapabilityDrawParameters); + return spv::BuiltInDrawIndex; + + case glslang::EbvPrimitiveId: + if (glslangIntermediate->getStage() == EShLangFragment) + builder.addCapability(spv::CapabilityGeometry); + return spv::BuiltInPrimitiveId; + + case glslang::EbvFragStencilRef: + builder.addExtension(spv::E_SPV_EXT_shader_stencil_export); + builder.addCapability(spv::CapabilityStencilExportEXT); + return spv::BuiltInFragStencilRefEXT; + + case glslang::EbvInvocationId: return spv::BuiltInInvocationId; + case glslang::EbvTessLevelInner: return spv::BuiltInTessLevelInner; + case glslang::EbvTessLevelOuter: return spv::BuiltInTessLevelOuter; + case glslang::EbvTessCoord: return spv::BuiltInTessCoord; + case glslang::EbvPatchVertices: return spv::BuiltInPatchVertices; + case glslang::EbvFragCoord: return spv::BuiltInFragCoord; + case glslang::EbvPointCoord: return spv::BuiltInPointCoord; + case glslang::EbvFace: return spv::BuiltInFrontFacing; + case glslang::EbvFragDepth: return spv::BuiltInFragDepth; + case glslang::EbvHelperInvocation: return spv::BuiltInHelperInvocation; + case glslang::EbvNumWorkGroups: return spv::BuiltInNumWorkgroups; + case glslang::EbvWorkGroupSize: return spv::BuiltInWorkgroupSize; + case glslang::EbvWorkGroupId: return spv::BuiltInWorkgroupId; + case glslang::EbvLocalInvocationId: return spv::BuiltInLocalInvocationId; + case glslang::EbvLocalInvocationIndex: return spv::BuiltInLocalInvocationIndex; + case glslang::EbvGlobalInvocationId: return spv::BuiltInGlobalInvocationId; + + case glslang::EbvSubGroupSize: + builder.addExtension(spv::E_SPV_KHR_shader_ballot); + builder.addCapability(spv::CapabilitySubgroupBallotKHR); + return spv::BuiltInSubgroupSize; + + case glslang::EbvSubGroupInvocation: + builder.addExtension(spv::E_SPV_KHR_shader_ballot); + builder.addCapability(spv::CapabilitySubgroupBallotKHR); + return spv::BuiltInSubgroupLocalInvocationId; + + case glslang::EbvSubGroupEqMask: + builder.addExtension(spv::E_SPV_KHR_shader_ballot); + builder.addCapability(spv::CapabilitySubgroupBallotKHR); + return spv::BuiltInSubgroupEqMaskKHR; + + case glslang::EbvSubGroupGeMask: + builder.addExtension(spv::E_SPV_KHR_shader_ballot); + builder.addCapability(spv::CapabilitySubgroupBallotKHR); + return spv::BuiltInSubgroupGeMaskKHR; + + case glslang::EbvSubGroupGtMask: + builder.addExtension(spv::E_SPV_KHR_shader_ballot); + builder.addCapability(spv::CapabilitySubgroupBallotKHR); + return spv::BuiltInSubgroupGtMaskKHR; + + case glslang::EbvSubGroupLeMask: + builder.addExtension(spv::E_SPV_KHR_shader_ballot); + builder.addCapability(spv::CapabilitySubgroupBallotKHR); + return spv::BuiltInSubgroupLeMaskKHR; + + case glslang::EbvSubGroupLtMask: + builder.addExtension(spv::E_SPV_KHR_shader_ballot); + builder.addCapability(spv::CapabilitySubgroupBallotKHR); + return spv::BuiltInSubgroupLtMaskKHR; + + case glslang::EbvNumSubgroups: + builder.addCapability(spv::CapabilityGroupNonUniform); + return spv::BuiltInNumSubgroups; + + case glslang::EbvSubgroupID: + builder.addCapability(spv::CapabilityGroupNonUniform); + return spv::BuiltInSubgroupId; + + case glslang::EbvSubgroupSize2: + builder.addCapability(spv::CapabilityGroupNonUniform); + return spv::BuiltInSubgroupSize; + + case glslang::EbvSubgroupInvocation2: + builder.addCapability(spv::CapabilityGroupNonUniform); + return spv::BuiltInSubgroupLocalInvocationId; + + case glslang::EbvSubgroupEqMask2: + builder.addCapability(spv::CapabilityGroupNonUniform); + builder.addCapability(spv::CapabilityGroupNonUniformBallot); + return spv::BuiltInSubgroupEqMask; + + case glslang::EbvSubgroupGeMask2: + builder.addCapability(spv::CapabilityGroupNonUniform); + builder.addCapability(spv::CapabilityGroupNonUniformBallot); + return spv::BuiltInSubgroupGeMask; + + case glslang::EbvSubgroupGtMask2: + builder.addCapability(spv::CapabilityGroupNonUniform); + builder.addCapability(spv::CapabilityGroupNonUniformBallot); + return spv::BuiltInSubgroupGtMask; + + case glslang::EbvSubgroupLeMask2: + builder.addCapability(spv::CapabilityGroupNonUniform); + builder.addCapability(spv::CapabilityGroupNonUniformBallot); + return spv::BuiltInSubgroupLeMask; + + case glslang::EbvSubgroupLtMask2: + builder.addCapability(spv::CapabilityGroupNonUniform); + builder.addCapability(spv::CapabilityGroupNonUniformBallot); + return spv::BuiltInSubgroupLtMask; +#ifdef AMD_EXTENSIONS + case glslang::EbvBaryCoordNoPersp: + builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter); + return spv::BuiltInBaryCoordNoPerspAMD; + + case glslang::EbvBaryCoordNoPerspCentroid: + builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter); + return spv::BuiltInBaryCoordNoPerspCentroidAMD; + + case glslang::EbvBaryCoordNoPerspSample: + builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter); + return spv::BuiltInBaryCoordNoPerspSampleAMD; + + case glslang::EbvBaryCoordSmooth: + builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter); + return spv::BuiltInBaryCoordSmoothAMD; + + case glslang::EbvBaryCoordSmoothCentroid: + builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter); + return spv::BuiltInBaryCoordSmoothCentroidAMD; + + case glslang::EbvBaryCoordSmoothSample: + builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter); + return spv::BuiltInBaryCoordSmoothSampleAMD; + + case glslang::EbvBaryCoordPullModel: + builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter); + return spv::BuiltInBaryCoordPullModelAMD; +#endif + + case glslang::EbvDeviceIndex: + addPre13Extension(spv::E_SPV_KHR_device_group); + builder.addCapability(spv::CapabilityDeviceGroup); + return spv::BuiltInDeviceIndex; + + case glslang::EbvViewIndex: + addPre13Extension(spv::E_SPV_KHR_multiview); + builder.addCapability(spv::CapabilityMultiView); + return spv::BuiltInViewIndex; + + case glslang::EbvFragSizeEXT: + builder.addExtension(spv::E_SPV_EXT_fragment_invocation_density); + builder.addCapability(spv::CapabilityFragmentDensityEXT); + return spv::BuiltInFragSizeEXT; + + case glslang::EbvFragInvocationCountEXT: + builder.addExtension(spv::E_SPV_EXT_fragment_invocation_density); + builder.addCapability(spv::CapabilityFragmentDensityEXT); + return spv::BuiltInFragInvocationCountEXT; + +#ifdef NV_EXTENSIONS + case glslang::EbvViewportMaskNV: + if (!memberDeclaration) { + builder.addExtension(spv::E_SPV_NV_viewport_array2); + builder.addCapability(spv::CapabilityShaderViewportMaskNV); + } + return spv::BuiltInViewportMaskNV; + case glslang::EbvSecondaryPositionNV: + if (!memberDeclaration) { + builder.addExtension(spv::E_SPV_NV_stereo_view_rendering); + builder.addCapability(spv::CapabilityShaderStereoViewNV); + } + return spv::BuiltInSecondaryPositionNV; + case glslang::EbvSecondaryViewportMaskNV: + if (!memberDeclaration) { + builder.addExtension(spv::E_SPV_NV_stereo_view_rendering); + builder.addCapability(spv::CapabilityShaderStereoViewNV); + } + return spv::BuiltInSecondaryViewportMaskNV; + case glslang::EbvPositionPerViewNV: + if (!memberDeclaration) { + builder.addExtension(spv::E_SPV_NVX_multiview_per_view_attributes); + builder.addCapability(spv::CapabilityPerViewAttributesNV); + } + return spv::BuiltInPositionPerViewNV; + case glslang::EbvViewportMaskPerViewNV: + if (!memberDeclaration) { + builder.addExtension(spv::E_SPV_NVX_multiview_per_view_attributes); + builder.addCapability(spv::CapabilityPerViewAttributesNV); + } + return spv::BuiltInViewportMaskPerViewNV; + case glslang::EbvFragFullyCoveredNV: + builder.addExtension(spv::E_SPV_EXT_fragment_fully_covered); + builder.addCapability(spv::CapabilityFragmentFullyCoveredEXT); + return spv::BuiltInFullyCoveredEXT; + case glslang::EbvFragmentSizeNV: + builder.addExtension(spv::E_SPV_NV_shading_rate); + builder.addCapability(spv::CapabilityShadingRateNV); + return spv::BuiltInFragmentSizeNV; + case glslang::EbvInvocationsPerPixelNV: + builder.addExtension(spv::E_SPV_NV_shading_rate); + builder.addCapability(spv::CapabilityShadingRateNV); + return spv::BuiltInInvocationsPerPixelNV; + + // raytracing + case glslang::EbvLaunchIdNV: + return spv::BuiltInLaunchIdNV; + case glslang::EbvLaunchSizeNV: + return spv::BuiltInLaunchSizeNV; + case glslang::EbvWorldRayOriginNV: + return spv::BuiltInWorldRayOriginNV; + case glslang::EbvWorldRayDirectionNV: + return spv::BuiltInWorldRayDirectionNV; + case glslang::EbvObjectRayOriginNV: + return spv::BuiltInObjectRayOriginNV; + case glslang::EbvObjectRayDirectionNV: + return spv::BuiltInObjectRayDirectionNV; + case glslang::EbvRayTminNV: + return spv::BuiltInRayTminNV; + case glslang::EbvRayTmaxNV: + return spv::BuiltInRayTmaxNV; + case glslang::EbvInstanceCustomIndexNV: + return spv::BuiltInInstanceCustomIndexNV; + case glslang::EbvHitTNV: + return spv::BuiltInHitTNV; + case glslang::EbvHitKindNV: + return spv::BuiltInHitKindNV; + case glslang::EbvObjectToWorldNV: + return spv::BuiltInObjectToWorldNV; + case glslang::EbvWorldToObjectNV: + return spv::BuiltInWorldToObjectNV; + case glslang::EbvIncomingRayFlagsNV: + return spv::BuiltInIncomingRayFlagsNV; + case glslang::EbvBaryCoordNV: + builder.addExtension(spv::E_SPV_NV_fragment_shader_barycentric); + builder.addCapability(spv::CapabilityFragmentBarycentricNV); + return spv::BuiltInBaryCoordNV; + case glslang::EbvBaryCoordNoPerspNV: + builder.addExtension(spv::E_SPV_NV_fragment_shader_barycentric); + builder.addCapability(spv::CapabilityFragmentBarycentricNV); + return spv::BuiltInBaryCoordNoPerspNV; + case glslang::EbvTaskCountNV: + return spv::BuiltInTaskCountNV; + case glslang::EbvPrimitiveCountNV: + return spv::BuiltInPrimitiveCountNV; + case glslang::EbvPrimitiveIndicesNV: + return spv::BuiltInPrimitiveIndicesNV; + case glslang::EbvClipDistancePerViewNV: + return spv::BuiltInClipDistancePerViewNV; + case glslang::EbvCullDistancePerViewNV: + return spv::BuiltInCullDistancePerViewNV; + case glslang::EbvLayerPerViewNV: + return spv::BuiltInLayerPerViewNV; + case glslang::EbvMeshViewCountNV: + return spv::BuiltInMeshViewCountNV; + case glslang::EbvMeshViewIndicesNV: + return spv::BuiltInMeshViewIndicesNV; +#endif + default: + return spv::BuiltInMax; + } +} + +// Translate glslang image layout format to SPIR-V image format. +spv::ImageFormat TGlslangToSpvTraverser::TranslateImageFormat(const glslang::TType& type) +{ + assert(type.getBasicType() == glslang::EbtSampler); + + // Check for capabilities + switch (type.getQualifier().layoutFormat) { + case glslang::ElfRg32f: + case glslang::ElfRg16f: + case glslang::ElfR11fG11fB10f: + case glslang::ElfR16f: + case glslang::ElfRgba16: + case glslang::ElfRgb10A2: + case glslang::ElfRg16: + case glslang::ElfRg8: + case glslang::ElfR16: + case glslang::ElfR8: + case glslang::ElfRgba16Snorm: + case glslang::ElfRg16Snorm: + case glslang::ElfRg8Snorm: + case glslang::ElfR16Snorm: + case glslang::ElfR8Snorm: + + case glslang::ElfRg32i: + case glslang::ElfRg16i: + case glslang::ElfRg8i: + case glslang::ElfR16i: + case glslang::ElfR8i: + + case glslang::ElfRgb10a2ui: + case glslang::ElfRg32ui: + case glslang::ElfRg16ui: + case glslang::ElfRg8ui: + case glslang::ElfR16ui: + case glslang::ElfR8ui: + builder.addCapability(spv::CapabilityStorageImageExtendedFormats); + break; + + default: + break; + } + + // do the translation + switch (type.getQualifier().layoutFormat) { + case glslang::ElfNone: return spv::ImageFormatUnknown; + case glslang::ElfRgba32f: return spv::ImageFormatRgba32f; + case glslang::ElfRgba16f: return spv::ImageFormatRgba16f; + case glslang::ElfR32f: return spv::ImageFormatR32f; + case glslang::ElfRgba8: return spv::ImageFormatRgba8; + case glslang::ElfRgba8Snorm: return spv::ImageFormatRgba8Snorm; + case glslang::ElfRg32f: return spv::ImageFormatRg32f; + case glslang::ElfRg16f: return spv::ImageFormatRg16f; + case glslang::ElfR11fG11fB10f: return spv::ImageFormatR11fG11fB10f; + case glslang::ElfR16f: return spv::ImageFormatR16f; + case glslang::ElfRgba16: return spv::ImageFormatRgba16; + case glslang::ElfRgb10A2: return spv::ImageFormatRgb10A2; + case glslang::ElfRg16: return spv::ImageFormatRg16; + case glslang::ElfRg8: return spv::ImageFormatRg8; + case glslang::ElfR16: return spv::ImageFormatR16; + case glslang::ElfR8: return spv::ImageFormatR8; + case glslang::ElfRgba16Snorm: return spv::ImageFormatRgba16Snorm; + case glslang::ElfRg16Snorm: return spv::ImageFormatRg16Snorm; + case glslang::ElfRg8Snorm: return spv::ImageFormatRg8Snorm; + case glslang::ElfR16Snorm: return spv::ImageFormatR16Snorm; + case glslang::ElfR8Snorm: return spv::ImageFormatR8Snorm; + case glslang::ElfRgba32i: return spv::ImageFormatRgba32i; + case glslang::ElfRgba16i: return spv::ImageFormatRgba16i; + case glslang::ElfRgba8i: return spv::ImageFormatRgba8i; + case glslang::ElfR32i: return spv::ImageFormatR32i; + case glslang::ElfRg32i: return spv::ImageFormatRg32i; + case glslang::ElfRg16i: return spv::ImageFormatRg16i; + case glslang::ElfRg8i: return spv::ImageFormatRg8i; + case glslang::ElfR16i: return spv::ImageFormatR16i; + case glslang::ElfR8i: return spv::ImageFormatR8i; + case glslang::ElfRgba32ui: return spv::ImageFormatRgba32ui; + case glslang::ElfRgba16ui: return spv::ImageFormatRgba16ui; + case glslang::ElfRgba8ui: return spv::ImageFormatRgba8ui; + case glslang::ElfR32ui: return spv::ImageFormatR32ui; + case glslang::ElfRg32ui: return spv::ImageFormatRg32ui; + case glslang::ElfRg16ui: return spv::ImageFormatRg16ui; + case glslang::ElfRgb10a2ui: return spv::ImageFormatRgb10a2ui; + case glslang::ElfRg8ui: return spv::ImageFormatRg8ui; + case glslang::ElfR16ui: return spv::ImageFormatR16ui; + case glslang::ElfR8ui: return spv::ImageFormatR8ui; + default: return spv::ImageFormatMax; + } +} + +spv::SelectionControlMask TGlslangToSpvTraverser::TranslateSelectionControl(const glslang::TIntermSelection& selectionNode) const +{ + if (selectionNode.getFlatten()) + return spv::SelectionControlFlattenMask; + if (selectionNode.getDontFlatten()) + return spv::SelectionControlDontFlattenMask; + return spv::SelectionControlMaskNone; +} + +spv::SelectionControlMask TGlslangToSpvTraverser::TranslateSwitchControl(const glslang::TIntermSwitch& switchNode) const +{ + if (switchNode.getFlatten()) + return spv::SelectionControlFlattenMask; + if (switchNode.getDontFlatten()) + return spv::SelectionControlDontFlattenMask; + return spv::SelectionControlMaskNone; +} + +// return a non-0 dependency if the dependency argument must be set +spv::LoopControlMask TGlslangToSpvTraverser::TranslateLoopControl(const glslang::TIntermLoop& loopNode, + std::vector<unsigned int>& operands) const +{ + spv::LoopControlMask control = spv::LoopControlMaskNone; + + if (loopNode.getDontUnroll()) + control = control | spv::LoopControlDontUnrollMask; + if (loopNode.getUnroll()) + control = control | spv::LoopControlUnrollMask; + if (unsigned(loopNode.getLoopDependency()) == glslang::TIntermLoop::dependencyInfinite) + control = control | spv::LoopControlDependencyInfiniteMask; + else if (loopNode.getLoopDependency() > 0) { + control = control | spv::LoopControlDependencyLengthMask; + operands.push_back((unsigned int)loopNode.getLoopDependency()); + } + if (glslangIntermediate->getSpv().spv >= glslang::EShTargetSpv_1_4) { + if (loopNode.getMinIterations() > 0) { + control = control | spv::LoopControlMinIterationsMask; + operands.push_back(loopNode.getMinIterations()); + } + if (loopNode.getMaxIterations() < glslang::TIntermLoop::iterationsInfinite) { + control = control | spv::LoopControlMaxIterationsMask; + operands.push_back(loopNode.getMaxIterations()); + } + if (loopNode.getIterationMultiple() > 1) { + control = control | spv::LoopControlIterationMultipleMask; + operands.push_back(loopNode.getIterationMultiple()); + } + if (loopNode.getPeelCount() > 0) { + control = control | spv::LoopControlPeelCountMask; + operands.push_back(loopNode.getPeelCount()); + } + if (loopNode.getPartialCount() > 0) { + control = control | spv::LoopControlPartialCountMask; + operands.push_back(loopNode.getPartialCount()); + } + } + + return control; +} + +// Translate glslang type to SPIR-V storage class. +spv::StorageClass TGlslangToSpvTraverser::TranslateStorageClass(const glslang::TType& type) +{ + if (type.getQualifier().isPipeInput()) + return spv::StorageClassInput; + if (type.getQualifier().isPipeOutput()) + return spv::StorageClassOutput; + + if (glslangIntermediate->getSource() != glslang::EShSourceHlsl || + type.getQualifier().storage == glslang::EvqUniform) { + if (type.getBasicType() == glslang::EbtAtomicUint) + return spv::StorageClassAtomicCounter; + if (type.containsOpaque()) + return spv::StorageClassUniformConstant; + } + +#ifdef NV_EXTENSIONS + if (type.getQualifier().isUniformOrBuffer() && + type.getQualifier().layoutShaderRecordNV) { + return spv::StorageClassShaderRecordBufferNV; + } +#endif + + if (glslangIntermediate->usingStorageBuffer() && type.getQualifier().storage == glslang::EvqBuffer) { + addPre13Extension(spv::E_SPV_KHR_storage_buffer_storage_class); + return spv::StorageClassStorageBuffer; + } + + if (type.getQualifier().isUniformOrBuffer()) { + if (type.getQualifier().layoutPushConstant) + return spv::StorageClassPushConstant; + if (type.getBasicType() == glslang::EbtBlock) + return spv::StorageClassUniform; + return spv::StorageClassUniformConstant; + } + + switch (type.getQualifier().storage) { + case glslang::EvqShared: return spv::StorageClassWorkgroup; + case glslang::EvqGlobal: return spv::StorageClassPrivate; + case glslang::EvqConstReadOnly: return spv::StorageClassFunction; + case glslang::EvqTemporary: return spv::StorageClassFunction; +#ifdef NV_EXTENSIONS + case glslang::EvqPayloadNV: return spv::StorageClassRayPayloadNV; + case glslang::EvqPayloadInNV: return spv::StorageClassIncomingRayPayloadNV; + case glslang::EvqHitAttrNV: return spv::StorageClassHitAttributeNV; + case glslang::EvqCallableDataNV: return spv::StorageClassCallableDataNV; + case glslang::EvqCallableDataInNV: return spv::StorageClassIncomingCallableDataNV; +#endif + default: + assert(0); + break; + } + + return spv::StorageClassFunction; +} + +// Add capabilities pertaining to how an array is indexed. +void TGlslangToSpvTraverser::addIndirectionIndexCapabilities(const glslang::TType& baseType, + const glslang::TType& indexType) +{ + if (indexType.getQualifier().isNonUniform()) { + // deal with an asserted non-uniform index + // SPV_EXT_descriptor_indexing already added in TranslateNonUniformDecoration + if (baseType.getBasicType() == glslang::EbtSampler) { + if (baseType.getQualifier().hasAttachment()) + builder.addCapability(spv::CapabilityInputAttachmentArrayNonUniformIndexingEXT); + else if (baseType.isImage() && baseType.getSampler().dim == glslang::EsdBuffer) + builder.addCapability(spv::CapabilityStorageTexelBufferArrayNonUniformIndexingEXT); + else if (baseType.isTexture() && baseType.getSampler().dim == glslang::EsdBuffer) + builder.addCapability(spv::CapabilityUniformTexelBufferArrayNonUniformIndexingEXT); + else if (baseType.isImage()) + builder.addCapability(spv::CapabilityStorageImageArrayNonUniformIndexingEXT); + else if (baseType.isTexture()) + builder.addCapability(spv::CapabilitySampledImageArrayNonUniformIndexingEXT); + } else if (baseType.getBasicType() == glslang::EbtBlock) { + if (baseType.getQualifier().storage == glslang::EvqBuffer) + builder.addCapability(spv::CapabilityStorageBufferArrayNonUniformIndexingEXT); + else if (baseType.getQualifier().storage == glslang::EvqUniform) + builder.addCapability(spv::CapabilityUniformBufferArrayNonUniformIndexingEXT); + } + } else { + // assume a dynamically uniform index + if (baseType.getBasicType() == glslang::EbtSampler) { + if (baseType.getQualifier().hasAttachment()) { + builder.addExtension("SPV_EXT_descriptor_indexing"); + builder.addCapability(spv::CapabilityInputAttachmentArrayDynamicIndexingEXT); + } else if (baseType.isImage() && baseType.getSampler().dim == glslang::EsdBuffer) { + builder.addExtension("SPV_EXT_descriptor_indexing"); + builder.addCapability(spv::CapabilityStorageTexelBufferArrayDynamicIndexingEXT); + } else if (baseType.isTexture() && baseType.getSampler().dim == glslang::EsdBuffer) { + builder.addExtension("SPV_EXT_descriptor_indexing"); + builder.addCapability(spv::CapabilityUniformTexelBufferArrayDynamicIndexingEXT); + } + } + } +} + +// Return whether or not the given type is something that should be tied to a +// descriptor set. +bool IsDescriptorResource(const glslang::TType& type) +{ + // uniform and buffer blocks are included, unless it is a push_constant + if (type.getBasicType() == glslang::EbtBlock) + return type.getQualifier().isUniformOrBuffer() && +#ifdef NV_EXTENSIONS + ! type.getQualifier().layoutShaderRecordNV && +#endif + ! type.getQualifier().layoutPushConstant; + + // non block... + // basically samplerXXX/subpass/sampler/texture are all included + // if they are the global-scope-class, not the function parameter + // (or local, if they ever exist) class. + if (type.getBasicType() == glslang::EbtSampler) + return type.getQualifier().isUniformOrBuffer(); + + // None of the above. + return false; +} + +void InheritQualifiers(glslang::TQualifier& child, const glslang::TQualifier& parent) +{ + if (child.layoutMatrix == glslang::ElmNone) + child.layoutMatrix = parent.layoutMatrix; + + if (parent.invariant) + child.invariant = true; + if (parent.nopersp) + child.nopersp = true; +#ifdef AMD_EXTENSIONS + if (parent.explicitInterp) + child.explicitInterp = true; +#endif + if (parent.flat) + child.flat = true; + if (parent.centroid) + child.centroid = true; + if (parent.patch) + child.patch = true; + if (parent.sample) + child.sample = true; + if (parent.coherent) + child.coherent = true; + if (parent.devicecoherent) + child.devicecoherent = true; + if (parent.queuefamilycoherent) + child.queuefamilycoherent = true; + if (parent.workgroupcoherent) + child.workgroupcoherent = true; + if (parent.subgroupcoherent) + child.subgroupcoherent = true; + if (parent.nonprivate) + child.nonprivate = true; + if (parent.volatil) + child.volatil = true; + if (parent.restrict) + child.restrict = true; + if (parent.readonly) + child.readonly = true; + if (parent.writeonly) + child.writeonly = true; +#ifdef NV_EXTENSIONS + if (parent.perPrimitiveNV) + child.perPrimitiveNV = true; + if (parent.perViewNV) + child.perViewNV = true; + if (parent.perTaskNV) + child.perTaskNV = true; +#endif +} + +bool HasNonLayoutQualifiers(const glslang::TType& type, const glslang::TQualifier& qualifier) +{ + // This should list qualifiers that simultaneous satisfy: + // - struct members might inherit from a struct declaration + // (note that non-block structs don't explicitly inherit, + // only implicitly, meaning no decoration involved) + // - affect decorations on the struct members + // (note smooth does not, and expecting something like volatile + // to effect the whole object) + // - are not part of the offset/st430/etc or row/column-major layout + return qualifier.invariant || (qualifier.hasLocation() && type.getBasicType() == glslang::EbtBlock); +} + +// +// Implement the TGlslangToSpvTraverser class. +// + +TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion, const glslang::TIntermediate* glslangIntermediate, + spv::SpvBuildLogger* buildLogger, glslang::SpvOptions& options) + : TIntermTraverser(true, false, true), + options(options), + shaderEntry(nullptr), currentFunction(nullptr), + sequenceDepth(0), logger(buildLogger), + builder(spvVersion, (glslang::GetKhronosToolId() << 16) | glslang::GetSpirvGeneratorVersion(), logger), + inEntryPoint(false), entryPointTerminated(false), linkageOnly(false), + glslangIntermediate(glslangIntermediate) +{ + spv::ExecutionModel executionModel = TranslateExecutionModel(glslangIntermediate->getStage()); + + builder.clearAccessChain(); + builder.setSource(TranslateSourceLanguage(glslangIntermediate->getSource(), glslangIntermediate->getProfile()), + glslangIntermediate->getVersion()); + + if (options.generateDebugInfo) { + builder.setEmitOpLines(); + builder.setSourceFile(glslangIntermediate->getSourceFile()); + + // Set the source shader's text. If for SPV version 1.0, include + // a preamble in comments stating the OpModuleProcessed instructions. + // Otherwise, emit those as actual instructions. + std::string text; + const std::vector<std::string>& processes = glslangIntermediate->getProcesses(); + for (int p = 0; p < (int)processes.size(); ++p) { + if (glslangIntermediate->getSpv().spv < glslang::EShTargetSpv_1_1) { + text.append("// OpModuleProcessed "); + text.append(processes[p]); + text.append("\n"); + } else + builder.addModuleProcessed(processes[p]); + } + if (glslangIntermediate->getSpv().spv < glslang::EShTargetSpv_1_1 && (int)processes.size() > 0) + text.append("#line 1\n"); + text.append(glslangIntermediate->getSourceText()); + builder.setSourceText(text); + // Pass name and text for all included files + const std::map<std::string, std::string>& include_txt = glslangIntermediate->getIncludeText(); + for (auto iItr = include_txt.begin(); iItr != include_txt.end(); ++iItr) + builder.addInclude(iItr->first, iItr->second); + } + stdBuiltins = builder.import("GLSL.std.450"); + + spv::AddressingModel addressingModel = spv::AddressingModelLogical; + spv::MemoryModel memoryModel = spv::MemoryModelGLSL450; + + if (glslangIntermediate->usingPhysicalStorageBuffer()) { + addressingModel = spv::AddressingModelPhysicalStorageBuffer64EXT; + builder.addExtension(spv::E_SPV_EXT_physical_storage_buffer); + builder.addCapability(spv::CapabilityPhysicalStorageBufferAddressesEXT); + }; + if (glslangIntermediate->usingVulkanMemoryModel()) { + memoryModel = spv::MemoryModelVulkanKHR; + builder.addCapability(spv::CapabilityVulkanMemoryModelKHR); + builder.addExtension(spv::E_SPV_KHR_vulkan_memory_model); + } + builder.setMemoryModel(addressingModel, memoryModel); + + if (glslangIntermediate->usingVariablePointers()) { + builder.addCapability(spv::CapabilityVariablePointers); + } + + shaderEntry = builder.makeEntryPoint(glslangIntermediate->getEntryPointName().c_str()); + entryPoint = builder.addEntryPoint(executionModel, shaderEntry, glslangIntermediate->getEntryPointName().c_str()); + + // Add the source extensions + const auto& sourceExtensions = glslangIntermediate->getRequestedExtensions(); + for (auto it = sourceExtensions.begin(); it != sourceExtensions.end(); ++it) + builder.addSourceExtension(it->c_str()); + + // Add the top-level modes for this shader. + + if (glslangIntermediate->getXfbMode()) { + builder.addCapability(spv::CapabilityTransformFeedback); + builder.addExecutionMode(shaderEntry, spv::ExecutionModeXfb); + } + + unsigned int mode; + switch (glslangIntermediate->getStage()) { + case EShLangVertex: + builder.addCapability(spv::CapabilityShader); + break; + + case EShLangTessEvaluation: + case EShLangTessControl: + builder.addCapability(spv::CapabilityTessellation); + + glslang::TLayoutGeometry primitive; + + if (glslangIntermediate->getStage() == EShLangTessControl) { + builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputVertices, glslangIntermediate->getVertices()); + primitive = glslangIntermediate->getOutputPrimitive(); + } else { + primitive = glslangIntermediate->getInputPrimitive(); + } + + switch (primitive) { + case glslang::ElgTriangles: mode = spv::ExecutionModeTriangles; break; + case glslang::ElgQuads: mode = spv::ExecutionModeQuads; break; + case glslang::ElgIsolines: mode = spv::ExecutionModeIsolines; break; + default: mode = spv::ExecutionModeMax; break; + } + if (mode != spv::ExecutionModeMax) + builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode); + + switch (glslangIntermediate->getVertexSpacing()) { + case glslang::EvsEqual: mode = spv::ExecutionModeSpacingEqual; break; + case glslang::EvsFractionalEven: mode = spv::ExecutionModeSpacingFractionalEven; break; + case glslang::EvsFractionalOdd: mode = spv::ExecutionModeSpacingFractionalOdd; break; + default: mode = spv::ExecutionModeMax; break; + } + if (mode != spv::ExecutionModeMax) + builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode); + + switch (glslangIntermediate->getVertexOrder()) { + case glslang::EvoCw: mode = spv::ExecutionModeVertexOrderCw; break; + case glslang::EvoCcw: mode = spv::ExecutionModeVertexOrderCcw; break; + default: mode = spv::ExecutionModeMax; break; + } + if (mode != spv::ExecutionModeMax) + builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode); + + if (glslangIntermediate->getPointMode()) + builder.addExecutionMode(shaderEntry, spv::ExecutionModePointMode); + break; + + case EShLangGeometry: + builder.addCapability(spv::CapabilityGeometry); + switch (glslangIntermediate->getInputPrimitive()) { + case glslang::ElgPoints: mode = spv::ExecutionModeInputPoints; break; + case glslang::ElgLines: mode = spv::ExecutionModeInputLines; break; + case glslang::ElgLinesAdjacency: mode = spv::ExecutionModeInputLinesAdjacency; break; + case glslang::ElgTriangles: mode = spv::ExecutionModeTriangles; break; + case glslang::ElgTrianglesAdjacency: mode = spv::ExecutionModeInputTrianglesAdjacency; break; + default: mode = spv::ExecutionModeMax; break; + } + if (mode != spv::ExecutionModeMax) + builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode); + + builder.addExecutionMode(shaderEntry, spv::ExecutionModeInvocations, glslangIntermediate->getInvocations()); + + switch (glslangIntermediate->getOutputPrimitive()) { + case glslang::ElgPoints: mode = spv::ExecutionModeOutputPoints; break; + case glslang::ElgLineStrip: mode = spv::ExecutionModeOutputLineStrip; break; + case glslang::ElgTriangleStrip: mode = spv::ExecutionModeOutputTriangleStrip; break; + default: mode = spv::ExecutionModeMax; break; + } + if (mode != spv::ExecutionModeMax) + builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode); + builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputVertices, glslangIntermediate->getVertices()); + break; + + case EShLangFragment: + builder.addCapability(spv::CapabilityShader); + if (glslangIntermediate->getPixelCenterInteger()) + builder.addExecutionMode(shaderEntry, spv::ExecutionModePixelCenterInteger); + + if (glslangIntermediate->getOriginUpperLeft()) + builder.addExecutionMode(shaderEntry, spv::ExecutionModeOriginUpperLeft); + else + builder.addExecutionMode(shaderEntry, spv::ExecutionModeOriginLowerLeft); + + if (glslangIntermediate->getEarlyFragmentTests()) + builder.addExecutionMode(shaderEntry, spv::ExecutionModeEarlyFragmentTests); + + if (glslangIntermediate->getPostDepthCoverage()) { + builder.addCapability(spv::CapabilitySampleMaskPostDepthCoverage); + builder.addExecutionMode(shaderEntry, spv::ExecutionModePostDepthCoverage); + builder.addExtension(spv::E_SPV_KHR_post_depth_coverage); + } + + switch(glslangIntermediate->getDepth()) { + case glslang::EldGreater: mode = spv::ExecutionModeDepthGreater; break; + case glslang::EldLess: mode = spv::ExecutionModeDepthLess; break; + default: mode = spv::ExecutionModeMax; break; + } + if (mode != spv::ExecutionModeMax) + builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode); + + if (glslangIntermediate->getDepth() != glslang::EldUnchanged && glslangIntermediate->isDepthReplacing()) + builder.addExecutionMode(shaderEntry, spv::ExecutionModeDepthReplacing); + break; + + case EShLangCompute: + builder.addCapability(spv::CapabilityShader); + builder.addExecutionMode(shaderEntry, spv::ExecutionModeLocalSize, glslangIntermediate->getLocalSize(0), + glslangIntermediate->getLocalSize(1), + glslangIntermediate->getLocalSize(2)); +#ifdef NV_EXTENSIONS + if (glslangIntermediate->getLayoutDerivativeModeNone() == glslang::LayoutDerivativeGroupQuads) { + builder.addCapability(spv::CapabilityComputeDerivativeGroupQuadsNV); + builder.addExecutionMode(shaderEntry, spv::ExecutionModeDerivativeGroupQuadsNV); + builder.addExtension(spv::E_SPV_NV_compute_shader_derivatives); + } else if (glslangIntermediate->getLayoutDerivativeModeNone() == glslang::LayoutDerivativeGroupLinear) { + builder.addCapability(spv::CapabilityComputeDerivativeGroupLinearNV); + builder.addExecutionMode(shaderEntry, spv::ExecutionModeDerivativeGroupLinearNV); + builder.addExtension(spv::E_SPV_NV_compute_shader_derivatives); + } +#endif + break; + +#ifdef NV_EXTENSIONS + case EShLangRayGenNV: + case EShLangIntersectNV: + case EShLangAnyHitNV: + case EShLangClosestHitNV: + case EShLangMissNV: + case EShLangCallableNV: + builder.addCapability(spv::CapabilityRayTracingNV); + builder.addExtension("SPV_NV_ray_tracing"); + break; + case EShLangTaskNV: + case EShLangMeshNV: + builder.addCapability(spv::CapabilityMeshShadingNV); + builder.addExtension(spv::E_SPV_NV_mesh_shader); + builder.addExecutionMode(shaderEntry, spv::ExecutionModeLocalSize, glslangIntermediate->getLocalSize(0), + glslangIntermediate->getLocalSize(1), + glslangIntermediate->getLocalSize(2)); + if (glslangIntermediate->getStage() == EShLangMeshNV) { + builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputVertices, glslangIntermediate->getVertices()); + builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputPrimitivesNV, glslangIntermediate->getPrimitives()); + + switch (glslangIntermediate->getOutputPrimitive()) { + case glslang::ElgPoints: mode = spv::ExecutionModeOutputPoints; break; + case glslang::ElgLines: mode = spv::ExecutionModeOutputLinesNV; break; + case glslang::ElgTriangles: mode = spv::ExecutionModeOutputTrianglesNV; break; + default: mode = spv::ExecutionModeMax; break; + } + if (mode != spv::ExecutionModeMax) + builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode); + } + break; +#endif + + default: + break; + } +} + +// Finish creating SPV, after the traversal is complete. +void TGlslangToSpvTraverser::finishSpv() +{ + // Finish the entry point function + if (! entryPointTerminated) { + builder.setBuildPoint(shaderEntry->getLastBlock()); + builder.leaveFunction(); + } + + // finish off the entry-point SPV instruction by adding the Input/Output <id> + for (auto it = iOSet.cbegin(); it != iOSet.cend(); ++it) + entryPoint->addIdOperand(*it); + + // Add capabilities, extensions, remove unneeded decorations, etc., + // based on the resulting SPIR-V. + builder.postProcess(); +} + +// Write the SPV into 'out'. +void TGlslangToSpvTraverser::dumpSpv(std::vector<unsigned int>& out) +{ + builder.dump(out); +} + +// +// Implement the traversal functions. +// +// Return true from interior nodes to have the external traversal +// continue on to children. Return false if children were +// already processed. +// + +// +// Symbols can turn into +// - uniform/input reads +// - output writes +// - complex lvalue base setups: foo.bar[3].... , where we see foo and start up an access chain +// - something simple that degenerates into the last bullet +// +void TGlslangToSpvTraverser::visitSymbol(glslang::TIntermSymbol* symbol) +{ + SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder); + if (symbol->getType().getQualifier().isSpecConstant()) + spec_constant_op_mode_setter.turnOnSpecConstantOpMode(); + + // getSymbolId() will set up all the IO decorations on the first call. + // Formal function parameters were mapped during makeFunctions(). + spv::Id id = getSymbolId(symbol); + + // Include all "static use" and "linkage only" interface variables on the OpEntryPoint instruction + if (builder.isPointer(id)) { + // Consider adding to the OpEntryPoint interface list. + // Only looking at structures if they have at least one member. + if (!symbol->getType().isStruct() || symbol->getType().getStruct()->size() > 0) { + spv::StorageClass sc = builder.getStorageClass(id); + // Before SPIR-V 1.4, we only want to include Input and Output. + // Starting with SPIR-V 1.4, we want all globals. + if ((glslangIntermediate->getSpv().spv >= glslang::EShTargetSpv_1_4 && sc != spv::StorageClassFunction) || + (sc == spv::StorageClassInput || sc == spv::StorageClassOutput)) { + iOSet.insert(id); + } + } + } + + // Only process non-linkage-only nodes for generating actual static uses + if (! linkageOnly || symbol->getQualifier().isSpecConstant()) { + // Prepare to generate code for the access + + // L-value chains will be computed left to right. We're on the symbol now, + // which is the left-most part of the access chain, so now is "clear" time, + // followed by setting the base. + builder.clearAccessChain(); + + // For now, we consider all user variables as being in memory, so they are pointers, + // except for + // A) R-Value arguments to a function, which are an intermediate object. + // See comments in handleUserFunctionCall(). + // B) Specialization constants (normal constants don't even come in as a variable), + // These are also pure R-values. + glslang::TQualifier qualifier = symbol->getQualifier(); + if (qualifier.isSpecConstant() || rValueParameters.find(symbol->getId()) != rValueParameters.end()) + builder.setAccessChainRValue(id); + else + builder.setAccessChainLValue(id); + } + + // Process linkage-only nodes for any special additional interface work. + if (linkageOnly) { + if (glslangIntermediate->getHlslFunctionality1()) { + // Map implicit counter buffers to their originating buffers, which should have been + // seen by now, given earlier pruning of unused counters, and preservation of order + // of declaration. + if (symbol->getType().getQualifier().isUniformOrBuffer()) { + if (!glslangIntermediate->hasCounterBufferName(symbol->getName())) { + // Save possible originating buffers for counter buffers, keyed by + // making the potential counter-buffer name. + std::string keyName = symbol->getName().c_str(); + keyName = glslangIntermediate->addCounterBufferName(keyName); + counterOriginator[keyName] = symbol; + } else { + // Handle a counter buffer, by finding the saved originating buffer. + std::string keyName = symbol->getName().c_str(); + auto it = counterOriginator.find(keyName); + if (it != counterOriginator.end()) { + id = getSymbolId(it->second); + if (id != spv::NoResult) { + spv::Id counterId = getSymbolId(symbol); + if (counterId != spv::NoResult) { + builder.addExtension("SPV_GOOGLE_hlsl_functionality1"); + builder.addDecorationId(id, spv::DecorationHlslCounterBufferGOOGLE, counterId); + } + } + } + } + } + } + } +} + +bool TGlslangToSpvTraverser::visitBinary(glslang::TVisit /* visit */, glslang::TIntermBinary* node) +{ + builder.setLine(node->getLoc().line, node->getLoc().getFilename()); + + SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder); + if (node->getType().getQualifier().isSpecConstant()) + spec_constant_op_mode_setter.turnOnSpecConstantOpMode(); + + // First, handle special cases + switch (node->getOp()) { + case glslang::EOpAssign: + case glslang::EOpAddAssign: + case glslang::EOpSubAssign: + case glslang::EOpMulAssign: + case glslang::EOpVectorTimesMatrixAssign: + case glslang::EOpVectorTimesScalarAssign: + case glslang::EOpMatrixTimesScalarAssign: + case glslang::EOpMatrixTimesMatrixAssign: + case glslang::EOpDivAssign: + case glslang::EOpModAssign: + case glslang::EOpAndAssign: + case glslang::EOpInclusiveOrAssign: + case glslang::EOpExclusiveOrAssign: + case glslang::EOpLeftShiftAssign: + case glslang::EOpRightShiftAssign: + // A bin-op assign "a += b" means the same thing as "a = a + b" + // where a is evaluated before b. For a simple assignment, GLSL + // says to evaluate the left before the right. So, always, left + // node then right node. + { + // get the left l-value, save it away + builder.clearAccessChain(); + node->getLeft()->traverse(this); + spv::Builder::AccessChain lValue = builder.getAccessChain(); + + // evaluate the right + builder.clearAccessChain(); + node->getRight()->traverse(this); + spv::Id rValue = accessChainLoad(node->getRight()->getType()); + + if (node->getOp() != glslang::EOpAssign) { + // the left is also an r-value + builder.setAccessChain(lValue); + spv::Id leftRValue = accessChainLoad(node->getLeft()->getType()); + + // do the operation + OpDecorations decorations = { TranslatePrecisionDecoration(node->getOperationPrecision()), + TranslateNoContractionDecoration(node->getType().getQualifier()), + TranslateNonUniformDecoration(node->getType().getQualifier()) }; + rValue = createBinaryOperation(node->getOp(), decorations, + convertGlslangToSpvType(node->getType()), leftRValue, rValue, + node->getType().getBasicType()); + + // these all need their counterparts in createBinaryOperation() + assert(rValue != spv::NoResult); + } + + // store the result + builder.setAccessChain(lValue); + multiTypeStore(node->getLeft()->getType(), rValue); + + // assignments are expressions having an rValue after they are evaluated... + builder.clearAccessChain(); + builder.setAccessChainRValue(rValue); + } + return false; + case glslang::EOpIndexDirect: + case glslang::EOpIndexDirectStruct: + { + // Structure, array, matrix, or vector indirection with statically known index. + // Get the left part of the access chain. + node->getLeft()->traverse(this); + + // Add the next element in the chain + + const int glslangIndex = node->getRight()->getAsConstantUnion()->getConstArray()[0].getIConst(); + if (! node->getLeft()->getType().isArray() && + node->getLeft()->getType().isVector() && + node->getOp() == glslang::EOpIndexDirect) { + // This is essentially a hard-coded vector swizzle of size 1, + // so short circuit the access-chain stuff with a swizzle. + std::vector<unsigned> swizzle; + swizzle.push_back(glslangIndex); + int dummySize; + builder.accessChainPushSwizzle(swizzle, convertGlslangToSpvType(node->getLeft()->getType()), + TranslateCoherent(node->getLeft()->getType()), + glslangIntermediate->getBaseAlignmentScalar(node->getLeft()->getType(), dummySize)); + } else { + + // Load through a block reference is performed with a dot operator that + // is mapped to EOpIndexDirectStruct. When we get to the actual reference, + // do a load and reset the access chain. + if (node->getLeft()->getBasicType() == glslang::EbtReference && + !node->getLeft()->getType().isArray() && + node->getOp() == glslang::EOpIndexDirectStruct) + { + spv::Id left = accessChainLoad(node->getLeft()->getType()); + builder.clearAccessChain(); + builder.setAccessChainLValue(left); + } + + int spvIndex = glslangIndex; + if (node->getLeft()->getBasicType() == glslang::EbtBlock && + node->getOp() == glslang::EOpIndexDirectStruct) + { + // This may be, e.g., an anonymous block-member selection, which generally need + // index remapping due to hidden members in anonymous blocks. + std::vector<int>& remapper = memberRemapper[node->getLeft()->getType().getStruct()]; + assert(remapper.size() > 0); + spvIndex = remapper[glslangIndex]; + } + + // normal case for indexing array or structure or block + builder.accessChainPush(builder.makeIntConstant(spvIndex), TranslateCoherent(node->getLeft()->getType()), node->getLeft()->getType().getBufferReferenceAlignment()); + + // Add capabilities here for accessing PointSize and clip/cull distance. + // We have deferred generation of associated capabilities until now. + if (node->getLeft()->getType().isStruct() && ! node->getLeft()->getType().isArray()) + declareUseOfStructMember(*(node->getLeft()->getType().getStruct()), glslangIndex); + } + } + return false; + case glslang::EOpIndexIndirect: + { + // Array, matrix, or vector indirection with variable index. + // Will use native SPIR-V access-chain for and array indirection; + // matrices are arrays of vectors, so will also work for a matrix. + // Will use the access chain's 'component' for variable index into a vector. + + // This adapter is building access chains left to right. + // Set up the access chain to the left. + node->getLeft()->traverse(this); + + // save it so that computing the right side doesn't trash it + spv::Builder::AccessChain partial = builder.getAccessChain(); + + // compute the next index in the chain + builder.clearAccessChain(); + node->getRight()->traverse(this); + spv::Id index = accessChainLoad(node->getRight()->getType()); + + addIndirectionIndexCapabilities(node->getLeft()->getType(), node->getRight()->getType()); + + // restore the saved access chain + builder.setAccessChain(partial); + + if (! node->getLeft()->getType().isArray() && node->getLeft()->getType().isVector()) { + int dummySize; + builder.accessChainPushComponent(index, convertGlslangToSpvType(node->getLeft()->getType()), + TranslateCoherent(node->getLeft()->getType()), + glslangIntermediate->getBaseAlignmentScalar(node->getLeft()->getType(), dummySize)); + } else + builder.accessChainPush(index, TranslateCoherent(node->getLeft()->getType()), node->getLeft()->getType().getBufferReferenceAlignment()); + } + return false; + case glslang::EOpVectorSwizzle: + { + node->getLeft()->traverse(this); + std::vector<unsigned> swizzle; + convertSwizzle(*node->getRight()->getAsAggregate(), swizzle); + int dummySize; + builder.accessChainPushSwizzle(swizzle, convertGlslangToSpvType(node->getLeft()->getType()), + TranslateCoherent(node->getLeft()->getType()), + glslangIntermediate->getBaseAlignmentScalar(node->getLeft()->getType(), dummySize)); + } + return false; + case glslang::EOpMatrixSwizzle: + logger->missingFunctionality("matrix swizzle"); + return true; + case glslang::EOpLogicalOr: + case glslang::EOpLogicalAnd: + { + + // These may require short circuiting, but can sometimes be done as straight + // binary operations. The right operand must be short circuited if it has + // side effects, and should probably be if it is complex. + if (isTrivial(node->getRight()->getAsTyped())) + break; // handle below as a normal binary operation + // otherwise, we need to do dynamic short circuiting on the right operand + spv::Id result = createShortCircuit(node->getOp(), *node->getLeft()->getAsTyped(), *node->getRight()->getAsTyped()); + builder.clearAccessChain(); + builder.setAccessChainRValue(result); + } + return false; + default: + break; + } + + // Assume generic binary op... + + // get right operand + builder.clearAccessChain(); + node->getLeft()->traverse(this); + spv::Id left = accessChainLoad(node->getLeft()->getType()); + + // get left operand + builder.clearAccessChain(); + node->getRight()->traverse(this); + spv::Id right = accessChainLoad(node->getRight()->getType()); + + // get result + OpDecorations decorations = { TranslatePrecisionDecoration(node->getOperationPrecision()), + TranslateNoContractionDecoration(node->getType().getQualifier()), + TranslateNonUniformDecoration(node->getType().getQualifier()) }; + spv::Id result = createBinaryOperation(node->getOp(), decorations, + convertGlslangToSpvType(node->getType()), left, right, + node->getLeft()->getType().getBasicType()); + + builder.clearAccessChain(); + if (! result) { + logger->missingFunctionality("unknown glslang binary operation"); + return true; // pick up a child as the place-holder result + } else { + builder.setAccessChainRValue(result); + return false; + } +} + +bool TGlslangToSpvTraverser::visitUnary(glslang::TVisit /* visit */, glslang::TIntermUnary* node) +{ + builder.setLine(node->getLoc().line, node->getLoc().getFilename()); + + SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder); + if (node->getType().getQualifier().isSpecConstant()) + spec_constant_op_mode_setter.turnOnSpecConstantOpMode(); + + spv::Id result = spv::NoResult; + + // try texturing first + result = createImageTextureFunctionCall(node); + if (result != spv::NoResult) { + builder.clearAccessChain(); + builder.setAccessChainRValue(result); + + return false; // done with this node + } + + // Non-texturing. + + if (node->getOp() == glslang::EOpArrayLength) { + // Quite special; won't want to evaluate the operand. + + // Currently, the front-end does not allow .length() on an array until it is sized, + // except for the last block membeor of an SSBO. + // TODO: If this changes, link-time sized arrays might show up here, and need their + // size extracted. + + // Normal .length() would have been constant folded by the front-end. + // So, this has to be block.lastMember.length(). + // SPV wants "block" and member number as the operands, go get them. + + spv::Id length; + if (node->getOperand()->getType().isCoopMat()) { + spec_constant_op_mode_setter.turnOnSpecConstantOpMode(); + + spv::Id typeId = convertGlslangToSpvType(node->getOperand()->getType()); + assert(builder.isCooperativeMatrixType(typeId)); + + length = builder.createCooperativeMatrixLength(typeId); + } else { + glslang::TIntermTyped* block = node->getOperand()->getAsBinaryNode()->getLeft(); + block->traverse(this); + unsigned int member = node->getOperand()->getAsBinaryNode()->getRight()->getAsConstantUnion()->getConstArray()[0].getUConst(); + length = builder.createArrayLength(builder.accessChainGetLValue(), member); + } + + // GLSL semantics say the result of .length() is an int, while SPIR-V says + // signedness must be 0. So, convert from SPIR-V unsigned back to GLSL's + // AST expectation of a signed result. + if (glslangIntermediate->getSource() == glslang::EShSourceGlsl) { + if (builder.isInSpecConstCodeGenMode()) { + length = builder.createBinOp(spv::OpIAdd, builder.makeIntType(32), length, builder.makeIntConstant(0)); + } else { + length = builder.createUnaryOp(spv::OpBitcast, builder.makeIntType(32), length); + } + } + + builder.clearAccessChain(); + builder.setAccessChainRValue(length); + + return false; + } + + // Start by evaluating the operand + + // Does it need a swizzle inversion? If so, evaluation is inverted; + // operate first on the swizzle base, then apply the swizzle. + spv::Id invertedType = spv::NoType; + auto resultType = [&invertedType, &node, this](){ return invertedType != spv::NoType ? invertedType : convertGlslangToSpvType(node->getType()); }; + if (node->getOp() == glslang::EOpInterpolateAtCentroid) + invertedType = getInvertedSwizzleType(*node->getOperand()); + + builder.clearAccessChain(); + if (invertedType != spv::NoType) + node->getOperand()->getAsBinaryNode()->getLeft()->traverse(this); + else + node->getOperand()->traverse(this); + + spv::Id operand = spv::NoResult; + + if (node->getOp() == glslang::EOpAtomicCounterIncrement || + node->getOp() == glslang::EOpAtomicCounterDecrement || + node->getOp() == glslang::EOpAtomicCounter || + node->getOp() == glslang::EOpInterpolateAtCentroid) + operand = builder.accessChainGetLValue(); // Special case l-value operands + else + operand = accessChainLoad(node->getOperand()->getType()); + + OpDecorations decorations = { TranslatePrecisionDecoration(node->getOperationPrecision()), + TranslateNoContractionDecoration(node->getType().getQualifier()), + TranslateNonUniformDecoration(node->getType().getQualifier()) }; + + // it could be a conversion + if (! result) + result = createConversion(node->getOp(), decorations, resultType(), operand, node->getOperand()->getBasicType()); + + // if not, then possibly an operation + if (! result) + result = createUnaryOperation(node->getOp(), decorations, resultType(), operand, node->getOperand()->getBasicType()); + + if (result) { + if (invertedType) { + result = createInvertedSwizzle(decorations.precision, *node->getOperand(), result); + builder.addDecoration(result, decorations.nonUniform); + } + + builder.clearAccessChain(); + builder.setAccessChainRValue(result); + + return false; // done with this node + } + + // it must be a special case, check... + switch (node->getOp()) { + case glslang::EOpPostIncrement: + case glslang::EOpPostDecrement: + case glslang::EOpPreIncrement: + case glslang::EOpPreDecrement: + { + // we need the integer value "1" or the floating point "1.0" to add/subtract + spv::Id one = 0; + if (node->getBasicType() == glslang::EbtFloat) + one = builder.makeFloatConstant(1.0F); + else if (node->getBasicType() == glslang::EbtDouble) + one = builder.makeDoubleConstant(1.0); + else if (node->getBasicType() == glslang::EbtFloat16) + one = builder.makeFloat16Constant(1.0F); + else if (node->getBasicType() == glslang::EbtInt8 || node->getBasicType() == glslang::EbtUint8) + one = builder.makeInt8Constant(1); + else if (node->getBasicType() == glslang::EbtInt16 || node->getBasicType() == glslang::EbtUint16) + one = builder.makeInt16Constant(1); + else if (node->getBasicType() == glslang::EbtInt64 || node->getBasicType() == glslang::EbtUint64) + one = builder.makeInt64Constant(1); + else + one = builder.makeIntConstant(1); + glslang::TOperator op; + if (node->getOp() == glslang::EOpPreIncrement || + node->getOp() == glslang::EOpPostIncrement) + op = glslang::EOpAdd; + else + op = glslang::EOpSub; + + spv::Id result = createBinaryOperation(op, decorations, + convertGlslangToSpvType(node->getType()), operand, one, + node->getType().getBasicType()); + assert(result != spv::NoResult); + + // The result of operation is always stored, but conditionally the + // consumed result. The consumed result is always an r-value. + builder.accessChainStore(result); + builder.clearAccessChain(); + if (node->getOp() == glslang::EOpPreIncrement || + node->getOp() == glslang::EOpPreDecrement) + builder.setAccessChainRValue(result); + else + builder.setAccessChainRValue(operand); + } + + return false; + + case glslang::EOpEmitStreamVertex: + builder.createNoResultOp(spv::OpEmitStreamVertex, operand); + return false; + case glslang::EOpEndStreamPrimitive: + builder.createNoResultOp(spv::OpEndStreamPrimitive, operand); + return false; + + default: + logger->missingFunctionality("unknown glslang unary"); + return true; // pick up operand as placeholder result + } +} + +bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TIntermAggregate* node) +{ + SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder); + if (node->getType().getQualifier().isSpecConstant()) + spec_constant_op_mode_setter.turnOnSpecConstantOpMode(); + + spv::Id result = spv::NoResult; + spv::Id invertedType = spv::NoType; // to use to override the natural type of the node + auto resultType = [&invertedType, &node, this](){ return invertedType != spv::NoType ? invertedType : convertGlslangToSpvType(node->getType()); }; + + // try texturing + result = createImageTextureFunctionCall(node); + if (result != spv::NoResult) { + builder.clearAccessChain(); + builder.setAccessChainRValue(result); + + return false; + } else if (node->getOp() == glslang::EOpImageStore || +#ifdef AMD_EXTENSIONS + node->getOp() == glslang::EOpImageStoreLod || +#endif + node->getOp() == glslang::EOpImageAtomicStore) { + // "imageStore" is a special case, which has no result + return false; + } + + glslang::TOperator binOp = glslang::EOpNull; + bool reduceComparison = true; + bool isMatrix = false; + bool noReturnValue = false; + bool atomic = false; + + assert(node->getOp()); + + spv::Decoration precision = TranslatePrecisionDecoration(node->getOperationPrecision()); + + switch (node->getOp()) { + case glslang::EOpSequence: + { + if (preVisit) + ++sequenceDepth; + else + --sequenceDepth; + + if (sequenceDepth == 1) { + // If this is the parent node of all the functions, we want to see them + // early, so all call points have actual SPIR-V functions to reference. + // In all cases, still let the traverser visit the children for us. + makeFunctions(node->getAsAggregate()->getSequence()); + + // Also, we want all globals initializers to go into the beginning of the entry point, before + // anything else gets there, so visit out of order, doing them all now. + makeGlobalInitializers(node->getAsAggregate()->getSequence()); + + // Initializers are done, don't want to visit again, but functions and link objects need to be processed, + // so do them manually. + visitFunctions(node->getAsAggregate()->getSequence()); + + return false; + } + + return true; + } + case glslang::EOpLinkerObjects: + { + if (visit == glslang::EvPreVisit) + linkageOnly = true; + else + linkageOnly = false; + + return true; + } + case glslang::EOpComma: + { + // processing from left to right naturally leaves the right-most + // lying around in the access chain + glslang::TIntermSequence& glslangOperands = node->getSequence(); + for (int i = 0; i < (int)glslangOperands.size(); ++i) + glslangOperands[i]->traverse(this); + + return false; + } + case glslang::EOpFunction: + if (visit == glslang::EvPreVisit) { + if (isShaderEntryPoint(node)) { + inEntryPoint = true; + builder.setBuildPoint(shaderEntry->getLastBlock()); + currentFunction = shaderEntry; + } else { + handleFunctionEntry(node); + } + } else { + if (inEntryPoint) + entryPointTerminated = true; + builder.leaveFunction(); + inEntryPoint = false; + } + + return true; + case glslang::EOpParameters: + // Parameters will have been consumed by EOpFunction processing, but not + // the body, so we still visited the function node's children, making this + // child redundant. + return false; + case glslang::EOpFunctionCall: + { + builder.setLine(node->getLoc().line, node->getLoc().getFilename()); + if (node->isUserDefined()) + result = handleUserFunctionCall(node); + // assert(result); // this can happen for bad shaders because the call graph completeness checking is not yet done + if (result) { + builder.clearAccessChain(); + builder.setAccessChainRValue(result); + } else + logger->missingFunctionality("missing user function; linker needs to catch that"); + + return false; + } + case glslang::EOpConstructMat2x2: + case glslang::EOpConstructMat2x3: + case glslang::EOpConstructMat2x4: + case glslang::EOpConstructMat3x2: + case glslang::EOpConstructMat3x3: + case glslang::EOpConstructMat3x4: + case glslang::EOpConstructMat4x2: + case glslang::EOpConstructMat4x3: + case glslang::EOpConstructMat4x4: + case glslang::EOpConstructDMat2x2: + case glslang::EOpConstructDMat2x3: + case glslang::EOpConstructDMat2x4: + case glslang::EOpConstructDMat3x2: + case glslang::EOpConstructDMat3x3: + case glslang::EOpConstructDMat3x4: + case glslang::EOpConstructDMat4x2: + case glslang::EOpConstructDMat4x3: + case glslang::EOpConstructDMat4x4: + case glslang::EOpConstructIMat2x2: + case glslang::EOpConstructIMat2x3: + case glslang::EOpConstructIMat2x4: + case glslang::EOpConstructIMat3x2: + case glslang::EOpConstructIMat3x3: + case glslang::EOpConstructIMat3x4: + case glslang::EOpConstructIMat4x2: + case glslang::EOpConstructIMat4x3: + case glslang::EOpConstructIMat4x4: + case glslang::EOpConstructUMat2x2: + case glslang::EOpConstructUMat2x3: + case glslang::EOpConstructUMat2x4: + case glslang::EOpConstructUMat3x2: + case glslang::EOpConstructUMat3x3: + case glslang::EOpConstructUMat3x4: + case glslang::EOpConstructUMat4x2: + case glslang::EOpConstructUMat4x3: + case glslang::EOpConstructUMat4x4: + case glslang::EOpConstructBMat2x2: + case glslang::EOpConstructBMat2x3: + case glslang::EOpConstructBMat2x4: + case glslang::EOpConstructBMat3x2: + case glslang::EOpConstructBMat3x3: + case glslang::EOpConstructBMat3x4: + case glslang::EOpConstructBMat4x2: + case glslang::EOpConstructBMat4x3: + case glslang::EOpConstructBMat4x4: + case glslang::EOpConstructF16Mat2x2: + case glslang::EOpConstructF16Mat2x3: + case glslang::EOpConstructF16Mat2x4: + case glslang::EOpConstructF16Mat3x2: + case glslang::EOpConstructF16Mat3x3: + case glslang::EOpConstructF16Mat3x4: + case glslang::EOpConstructF16Mat4x2: + case glslang::EOpConstructF16Mat4x3: + case glslang::EOpConstructF16Mat4x4: + isMatrix = true; + // fall through + case glslang::EOpConstructFloat: + case glslang::EOpConstructVec2: + case glslang::EOpConstructVec3: + case glslang::EOpConstructVec4: + case glslang::EOpConstructDouble: + case glslang::EOpConstructDVec2: + case glslang::EOpConstructDVec3: + case glslang::EOpConstructDVec4: + case glslang::EOpConstructFloat16: + case glslang::EOpConstructF16Vec2: + case glslang::EOpConstructF16Vec3: + case glslang::EOpConstructF16Vec4: + case glslang::EOpConstructBool: + case glslang::EOpConstructBVec2: + case glslang::EOpConstructBVec3: + case glslang::EOpConstructBVec4: + case glslang::EOpConstructInt8: + case glslang::EOpConstructI8Vec2: + case glslang::EOpConstructI8Vec3: + case glslang::EOpConstructI8Vec4: + case glslang::EOpConstructUint8: + case glslang::EOpConstructU8Vec2: + case glslang::EOpConstructU8Vec3: + case glslang::EOpConstructU8Vec4: + case glslang::EOpConstructInt16: + case glslang::EOpConstructI16Vec2: + case glslang::EOpConstructI16Vec3: + case glslang::EOpConstructI16Vec4: + case glslang::EOpConstructUint16: + case glslang::EOpConstructU16Vec2: + case glslang::EOpConstructU16Vec3: + case glslang::EOpConstructU16Vec4: + case glslang::EOpConstructInt: + case glslang::EOpConstructIVec2: + case glslang::EOpConstructIVec3: + case glslang::EOpConstructIVec4: + case glslang::EOpConstructUint: + case glslang::EOpConstructUVec2: + case glslang::EOpConstructUVec3: + case glslang::EOpConstructUVec4: + case glslang::EOpConstructInt64: + case glslang::EOpConstructI64Vec2: + case glslang::EOpConstructI64Vec3: + case glslang::EOpConstructI64Vec4: + case glslang::EOpConstructUint64: + case glslang::EOpConstructU64Vec2: + case glslang::EOpConstructU64Vec3: + case glslang::EOpConstructU64Vec4: + case glslang::EOpConstructStruct: + case glslang::EOpConstructTextureSampler: + case glslang::EOpConstructReference: + case glslang::EOpConstructCooperativeMatrix: + { + builder.setLine(node->getLoc().line, node->getLoc().getFilename()); + std::vector<spv::Id> arguments; + translateArguments(*node, arguments); + spv::Id constructed; + if (node->getOp() == glslang::EOpConstructTextureSampler) + constructed = builder.createOp(spv::OpSampledImage, resultType(), arguments); + else if (node->getOp() == glslang::EOpConstructStruct || + node->getOp() == glslang::EOpConstructCooperativeMatrix || + node->getType().isArray()) { + std::vector<spv::Id> constituents; + for (int c = 0; c < (int)arguments.size(); ++c) + constituents.push_back(arguments[c]); + constructed = builder.createCompositeConstruct(resultType(), constituents); + } else if (isMatrix) + constructed = builder.createMatrixConstructor(precision, arguments, resultType()); + else + constructed = builder.createConstructor(precision, arguments, resultType()); + + builder.clearAccessChain(); + builder.setAccessChainRValue(constructed); + + return false; + } + + // These six are component-wise compares with component-wise results. + // Forward on to createBinaryOperation(), requesting a vector result. + case glslang::EOpLessThan: + case glslang::EOpGreaterThan: + case glslang::EOpLessThanEqual: + case glslang::EOpGreaterThanEqual: + case glslang::EOpVectorEqual: + case glslang::EOpVectorNotEqual: + { + // Map the operation to a binary + binOp = node->getOp(); + reduceComparison = false; + switch (node->getOp()) { + case glslang::EOpVectorEqual: binOp = glslang::EOpVectorEqual; break; + case glslang::EOpVectorNotEqual: binOp = glslang::EOpVectorNotEqual; break; + default: binOp = node->getOp(); break; + } + + break; + } + case glslang::EOpMul: + // component-wise matrix multiply + binOp = glslang::EOpMul; + break; + case glslang::EOpOuterProduct: + // two vectors multiplied to make a matrix + binOp = glslang::EOpOuterProduct; + break; + case glslang::EOpDot: + { + // for scalar dot product, use multiply + glslang::TIntermSequence& glslangOperands = node->getSequence(); + if (glslangOperands[0]->getAsTyped()->getVectorSize() == 1) + binOp = glslang::EOpMul; + break; + } + case glslang::EOpMod: + // when an aggregate, this is the floating-point mod built-in function, + // which can be emitted by the one in createBinaryOperation() + binOp = glslang::EOpMod; + break; + case glslang::EOpEmitVertex: + case glslang::EOpEndPrimitive: + case glslang::EOpBarrier: + case glslang::EOpMemoryBarrier: + case glslang::EOpMemoryBarrierAtomicCounter: + case glslang::EOpMemoryBarrierBuffer: + case glslang::EOpMemoryBarrierImage: + case glslang::EOpMemoryBarrierShared: + case glslang::EOpGroupMemoryBarrier: + case glslang::EOpDeviceMemoryBarrier: + case glslang::EOpAllMemoryBarrierWithGroupSync: + case glslang::EOpDeviceMemoryBarrierWithGroupSync: + case glslang::EOpWorkgroupMemoryBarrier: + case glslang::EOpWorkgroupMemoryBarrierWithGroupSync: + case glslang::EOpSubgroupBarrier: + case glslang::EOpSubgroupMemoryBarrier: + case glslang::EOpSubgroupMemoryBarrierBuffer: + case glslang::EOpSubgroupMemoryBarrierImage: + case glslang::EOpSubgroupMemoryBarrierShared: + noReturnValue = true; + // These all have 0 operands and will naturally finish up in the code below for 0 operands + break; + + case glslang::EOpAtomicStore: + noReturnValue = true; + // fallthrough + case glslang::EOpAtomicLoad: + case glslang::EOpAtomicAdd: + case glslang::EOpAtomicMin: + case glslang::EOpAtomicMax: + case glslang::EOpAtomicAnd: + case glslang::EOpAtomicOr: + case glslang::EOpAtomicXor: + case glslang::EOpAtomicExchange: + case glslang::EOpAtomicCompSwap: + atomic = true; + break; + + case glslang::EOpAtomicCounterAdd: + case glslang::EOpAtomicCounterSubtract: + case glslang::EOpAtomicCounterMin: + case glslang::EOpAtomicCounterMax: + case glslang::EOpAtomicCounterAnd: + case glslang::EOpAtomicCounterOr: + case glslang::EOpAtomicCounterXor: + case glslang::EOpAtomicCounterExchange: + case glslang::EOpAtomicCounterCompSwap: + builder.addExtension("SPV_KHR_shader_atomic_counter_ops"); + builder.addCapability(spv::CapabilityAtomicStorageOps); + atomic = true; + break; + +#ifdef NV_EXTENSIONS + case glslang::EOpIgnoreIntersectionNV: + case glslang::EOpTerminateRayNV: + case glslang::EOpTraceNV: + case glslang::EOpExecuteCallableNV: + case glslang::EOpWritePackedPrimitiveIndices4x8NV: + noReturnValue = true; + break; +#endif + case glslang::EOpCooperativeMatrixLoad: + case glslang::EOpCooperativeMatrixStore: + noReturnValue = true; + break; + + default: + break; + } + + // + // See if it maps to a regular operation. + // + if (binOp != glslang::EOpNull) { + glslang::TIntermTyped* left = node->getSequence()[0]->getAsTyped(); + glslang::TIntermTyped* right = node->getSequence()[1]->getAsTyped(); + assert(left && right); + + builder.clearAccessChain(); + left->traverse(this); + spv::Id leftId = accessChainLoad(left->getType()); + + builder.clearAccessChain(); + right->traverse(this); + spv::Id rightId = accessChainLoad(right->getType()); + + builder.setLine(node->getLoc().line, node->getLoc().getFilename()); + OpDecorations decorations = { precision, + TranslateNoContractionDecoration(node->getType().getQualifier()), + TranslateNonUniformDecoration(node->getType().getQualifier()) }; + result = createBinaryOperation(binOp, decorations, + resultType(), leftId, rightId, + left->getType().getBasicType(), reduceComparison); + + // code above should only make binOp that exists in createBinaryOperation + assert(result != spv::NoResult); + builder.clearAccessChain(); + builder.setAccessChainRValue(result); + + return false; + } + + // + // Create the list of operands. + // + glslang::TIntermSequence& glslangOperands = node->getSequence(); + std::vector<spv::Id> operands; + std::vector<spv::IdImmediate> memoryAccessOperands; + for (int arg = 0; arg < (int)glslangOperands.size(); ++arg) { + // special case l-value operands; there are just a few + bool lvalue = false; + switch (node->getOp()) { + case glslang::EOpFrexp: + case glslang::EOpModf: + if (arg == 1) + lvalue = true; + break; + case glslang::EOpInterpolateAtSample: + case glslang::EOpInterpolateAtOffset: +#ifdef AMD_EXTENSIONS + case glslang::EOpInterpolateAtVertex: +#endif + if (arg == 0) { + lvalue = true; + + // Does it need a swizzle inversion? If so, evaluation is inverted; + // operate first on the swizzle base, then apply the swizzle. + if (glslangOperands[0]->getAsOperator() && + glslangOperands[0]->getAsOperator()->getOp() == glslang::EOpVectorSwizzle) + invertedType = convertGlslangToSpvType(glslangOperands[0]->getAsBinaryNode()->getLeft()->getType()); + } + break; + case glslang::EOpAtomicAdd: + case glslang::EOpAtomicMin: + case glslang::EOpAtomicMax: + case glslang::EOpAtomicAnd: + case glslang::EOpAtomicOr: + case glslang::EOpAtomicXor: + case glslang::EOpAtomicExchange: + case glslang::EOpAtomicCompSwap: + case glslang::EOpAtomicLoad: + case glslang::EOpAtomicStore: + case glslang::EOpAtomicCounterAdd: + case glslang::EOpAtomicCounterSubtract: + case glslang::EOpAtomicCounterMin: + case glslang::EOpAtomicCounterMax: + case glslang::EOpAtomicCounterAnd: + case glslang::EOpAtomicCounterOr: + case glslang::EOpAtomicCounterXor: + case glslang::EOpAtomicCounterExchange: + case glslang::EOpAtomicCounterCompSwap: + if (arg == 0) + lvalue = true; + break; + case glslang::EOpAddCarry: + case glslang::EOpSubBorrow: + if (arg == 2) + lvalue = true; + break; + case glslang::EOpUMulExtended: + case glslang::EOpIMulExtended: + if (arg >= 2) + lvalue = true; + break; + case glslang::EOpCooperativeMatrixLoad: + if (arg == 0 || arg == 1) + lvalue = true; + break; + case glslang::EOpCooperativeMatrixStore: + if (arg == 1) + lvalue = true; + break; + default: + break; + } + builder.clearAccessChain(); + if (invertedType != spv::NoType && arg == 0) + glslangOperands[0]->getAsBinaryNode()->getLeft()->traverse(this); + else + glslangOperands[arg]->traverse(this); + + if (node->getOp() == glslang::EOpCooperativeMatrixLoad || + node->getOp() == glslang::EOpCooperativeMatrixStore) { + + if (arg == 1) { + // fold "element" parameter into the access chain + spv::Builder::AccessChain save = builder.getAccessChain(); + builder.clearAccessChain(); + glslangOperands[2]->traverse(this); + + spv::Id elementId = accessChainLoad(glslangOperands[2]->getAsTyped()->getType()); + + builder.setAccessChain(save); + + // Point to the first element of the array. + builder.accessChainPush(elementId, TranslateCoherent(glslangOperands[arg]->getAsTyped()->getType()), + glslangOperands[arg]->getAsTyped()->getType().getBufferReferenceAlignment()); + + spv::Builder::AccessChain::CoherentFlags coherentFlags = builder.getAccessChain().coherentFlags; + unsigned int alignment = builder.getAccessChain().alignment; + + int memoryAccess = TranslateMemoryAccess(coherentFlags); + if (node->getOp() == glslang::EOpCooperativeMatrixLoad) + memoryAccess &= ~spv::MemoryAccessMakePointerAvailableKHRMask; + if (node->getOp() == glslang::EOpCooperativeMatrixStore) + memoryAccess &= ~spv::MemoryAccessMakePointerVisibleKHRMask; + if (builder.getStorageClass(builder.getAccessChain().base) == spv::StorageClassPhysicalStorageBufferEXT) { + memoryAccess = (spv::MemoryAccessMask)(memoryAccess | spv::MemoryAccessAlignedMask); + } + + memoryAccessOperands.push_back(spv::IdImmediate(false, memoryAccess)); + + if (memoryAccess & spv::MemoryAccessAlignedMask) { + memoryAccessOperands.push_back(spv::IdImmediate(false, alignment)); + } + + if (memoryAccess & (spv::MemoryAccessMakePointerAvailableKHRMask | spv::MemoryAccessMakePointerVisibleKHRMask)) { + memoryAccessOperands.push_back(spv::IdImmediate(true, builder.makeUintConstant(TranslateMemoryScope(coherentFlags)))); + } + } else if (arg == 2) { + continue; + } + } + + if (lvalue) + operands.push_back(builder.accessChainGetLValue()); + else { + builder.setLine(node->getLoc().line, node->getLoc().getFilename()); + operands.push_back(accessChainLoad(glslangOperands[arg]->getAsTyped()->getType())); + } + } + + builder.setLine(node->getLoc().line, node->getLoc().getFilename()); + if (node->getOp() == glslang::EOpCooperativeMatrixLoad) { + std::vector<spv::IdImmediate> idImmOps; + + idImmOps.push_back(spv::IdImmediate(true, operands[1])); // buf + idImmOps.push_back(spv::IdImmediate(true, operands[2])); // stride + idImmOps.push_back(spv::IdImmediate(true, operands[3])); // colMajor + idImmOps.insert(idImmOps.end(), memoryAccessOperands.begin(), memoryAccessOperands.end()); + // get the pointee type + spv::Id typeId = builder.getContainedTypeId(builder.getTypeId(operands[0])); + assert(builder.isCooperativeMatrixType(typeId)); + // do the op + spv::Id result = builder.createOp(spv::OpCooperativeMatrixLoadNV, typeId, idImmOps); + // store the result to the pointer (out param 'm') + builder.createStore(result, operands[0]); + result = 0; + } else if (node->getOp() == glslang::EOpCooperativeMatrixStore) { + std::vector<spv::IdImmediate> idImmOps; + + idImmOps.push_back(spv::IdImmediate(true, operands[1])); // buf + idImmOps.push_back(spv::IdImmediate(true, operands[0])); // object + idImmOps.push_back(spv::IdImmediate(true, operands[2])); // stride + idImmOps.push_back(spv::IdImmediate(true, operands[3])); // colMajor + idImmOps.insert(idImmOps.end(), memoryAccessOperands.begin(), memoryAccessOperands.end()); + + builder.createNoResultOp(spv::OpCooperativeMatrixStoreNV, idImmOps); + result = 0; + } else if (atomic) { + // Handle all atomics + result = createAtomicOperation(node->getOp(), precision, resultType(), operands, node->getBasicType()); + } else { + // Pass through to generic operations. + switch (glslangOperands.size()) { + case 0: + result = createNoArgOperation(node->getOp(), precision, resultType()); + break; + case 1: + { + OpDecorations decorations = { precision, + TranslateNoContractionDecoration(node->getType().getQualifier()), + TranslateNonUniformDecoration(node->getType().getQualifier()) }; + result = createUnaryOperation( + node->getOp(), decorations, + resultType(), operands.front(), + glslangOperands[0]->getAsTyped()->getBasicType()); + } + break; + default: + result = createMiscOperation(node->getOp(), precision, resultType(), operands, node->getBasicType()); + break; + } + if (invertedType) + result = createInvertedSwizzle(precision, *glslangOperands[0]->getAsBinaryNode(), result); + } + + if (noReturnValue) + return false; + + if (! result) { + logger->missingFunctionality("unknown glslang aggregate"); + return true; // pick up a child as a placeholder operand + } else { + builder.clearAccessChain(); + builder.setAccessChainRValue(result); + return false; + } +} + +// This path handles both if-then-else and ?: +// The if-then-else has a node type of void, while +// ?: has either a void or a non-void node type +// +// Leaving the result, when not void: +// GLSL only has r-values as the result of a :?, but +// if we have an l-value, that can be more efficient if it will +// become the base of a complex r-value expression, because the +// next layer copies r-values into memory to use the access-chain mechanism +bool TGlslangToSpvTraverser::visitSelection(glslang::TVisit /* visit */, glslang::TIntermSelection* node) +{ + // see if OpSelect can handle it + const auto isOpSelectable = [&]() { + if (node->getBasicType() == glslang::EbtVoid) + return false; + // OpSelect can do all other types starting with SPV 1.4 + if (glslangIntermediate->getSpv().spv < glslang::EShTargetSpv_1_4) { + // pre-1.4, only scalars and vectors can be handled + if ((!node->getType().isScalar() && !node->getType().isVector())) + return false; + } + return true; + }; + + // See if it simple and safe, or required, to execute both sides. + // Crucially, side effects must be either semantically required or avoided, + // and there are performance trade-offs. + // Return true if required or a good idea (and safe) to execute both sides, + // false otherwise. + const auto bothSidesPolicy = [&]() -> bool { + // do we have both sides? + if (node->getTrueBlock() == nullptr || + node->getFalseBlock() == nullptr) + return false; + + // required? (unless we write additional code to look for side effects + // and make performance trade-offs if none are present) + if (!node->getShortCircuit()) + return true; + + // if not required to execute both, decide based on performance/practicality... + + if (!isOpSelectable()) + return false; + + assert(node->getType() == node->getTrueBlock() ->getAsTyped()->getType() && + node->getType() == node->getFalseBlock()->getAsTyped()->getType()); + + // return true if a single operand to ? : is okay for OpSelect + const auto operandOkay = [](glslang::TIntermTyped* node) { + return node->getAsSymbolNode() || node->getType().getQualifier().isConstant(); + }; + + return operandOkay(node->getTrueBlock() ->getAsTyped()) && + operandOkay(node->getFalseBlock()->getAsTyped()); + }; + + spv::Id result = spv::NoResult; // upcoming result selecting between trueValue and falseValue + // emit the condition before doing anything with selection + node->getCondition()->traverse(this); + spv::Id condition = accessChainLoad(node->getCondition()->getType()); + + // Find a way of executing both sides and selecting the right result. + const auto executeBothSides = [&]() -> void { + // execute both sides + node->getTrueBlock()->traverse(this); + spv::Id trueValue = accessChainLoad(node->getTrueBlock()->getAsTyped()->getType()); + node->getFalseBlock()->traverse(this); + spv::Id falseValue = accessChainLoad(node->getTrueBlock()->getAsTyped()->getType()); + + builder.setLine(node->getLoc().line, node->getLoc().getFilename()); + + // done if void + if (node->getBasicType() == glslang::EbtVoid) + return; + + // emit code to select between trueValue and falseValue + + // see if OpSelect can handle it + if (isOpSelectable()) { + // Emit OpSelect for this selection. + + // smear condition to vector, if necessary (AST is always scalar) + // Before 1.4, smear like for mix(), starting with 1.4, keep it scalar + if (glslangIntermediate->getSpv().spv < glslang::EShTargetSpv_1_4 && builder.isVector(trueValue)) { + condition = builder.smearScalar(spv::NoPrecision, condition, + builder.makeVectorType(builder.makeBoolType(), + builder.getNumComponents(trueValue))); + } + + // OpSelect + result = builder.createTriOp(spv::OpSelect, + convertGlslangToSpvType(node->getType()), condition, + trueValue, falseValue); + + builder.clearAccessChain(); + builder.setAccessChainRValue(result); + } else { + // We need control flow to select the result. + // TODO: Once SPIR-V OpSelect allows arbitrary types, eliminate this path. + result = builder.createVariable(spv::StorageClassFunction, convertGlslangToSpvType(node->getType())); + + // Selection control: + const spv::SelectionControlMask control = TranslateSelectionControl(*node); + + // make an "if" based on the value created by the condition + spv::Builder::If ifBuilder(condition, control, builder); + + // emit the "then" statement + builder.createStore(trueValue, result); + ifBuilder.makeBeginElse(); + // emit the "else" statement + builder.createStore(falseValue, result); + + // finish off the control flow + ifBuilder.makeEndIf(); + + builder.clearAccessChain(); + builder.setAccessChainLValue(result); + } + }; + + // Execute the one side needed, as per the condition + const auto executeOneSide = [&]() { + // Always emit control flow. + if (node->getBasicType() != glslang::EbtVoid) + result = builder.createVariable(spv::StorageClassFunction, convertGlslangToSpvType(node->getType())); + + // Selection control: + const spv::SelectionControlMask control = TranslateSelectionControl(*node); + + // make an "if" based on the value created by the condition + spv::Builder::If ifBuilder(condition, control, builder); + + // emit the "then" statement + if (node->getTrueBlock() != nullptr) { + node->getTrueBlock()->traverse(this); + if (result != spv::NoResult) + builder.createStore(accessChainLoad(node->getTrueBlock()->getAsTyped()->getType()), result); + } + + if (node->getFalseBlock() != nullptr) { + ifBuilder.makeBeginElse(); + // emit the "else" statement + node->getFalseBlock()->traverse(this); + if (result != spv::NoResult) + builder.createStore(accessChainLoad(node->getFalseBlock()->getAsTyped()->getType()), result); + } + + // finish off the control flow + ifBuilder.makeEndIf(); + + if (result != spv::NoResult) { + builder.clearAccessChain(); + builder.setAccessChainLValue(result); + } + }; + + // Try for OpSelect (or a requirement to execute both sides) + if (bothSidesPolicy()) { + SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder); + if (node->getType().getQualifier().isSpecConstant()) + spec_constant_op_mode_setter.turnOnSpecConstantOpMode(); + executeBothSides(); + } else + executeOneSide(); + + return false; +} + +bool TGlslangToSpvTraverser::visitSwitch(glslang::TVisit /* visit */, glslang::TIntermSwitch* node) +{ + // emit and get the condition before doing anything with switch + node->getCondition()->traverse(this); + spv::Id selector = accessChainLoad(node->getCondition()->getAsTyped()->getType()); + + // Selection control: + const spv::SelectionControlMask control = TranslateSwitchControl(*node); + + // browse the children to sort out code segments + int defaultSegment = -1; + std::vector<TIntermNode*> codeSegments; + glslang::TIntermSequence& sequence = node->getBody()->getSequence(); + std::vector<int> caseValues; + std::vector<int> valueIndexToSegment(sequence.size()); // note: probably not all are used, it is an overestimate + for (glslang::TIntermSequence::iterator c = sequence.begin(); c != sequence.end(); ++c) { + TIntermNode* child = *c; + if (child->getAsBranchNode() && child->getAsBranchNode()->getFlowOp() == glslang::EOpDefault) + defaultSegment = (int)codeSegments.size(); + else if (child->getAsBranchNode() && child->getAsBranchNode()->getFlowOp() == glslang::EOpCase) { + valueIndexToSegment[caseValues.size()] = (int)codeSegments.size(); + caseValues.push_back(child->getAsBranchNode()->getExpression()->getAsConstantUnion()->getConstArray()[0].getIConst()); + } else + codeSegments.push_back(child); + } + + // handle the case where the last code segment is missing, due to no code + // statements between the last case and the end of the switch statement + if ((caseValues.size() && (int)codeSegments.size() == valueIndexToSegment[caseValues.size() - 1]) || + (int)codeSegments.size() == defaultSegment) + codeSegments.push_back(nullptr); + + // make the switch statement + std::vector<spv::Block*> segmentBlocks; // returned, as the blocks allocated in the call + builder.makeSwitch(selector, control, (int)codeSegments.size(), caseValues, valueIndexToSegment, defaultSegment, segmentBlocks); + + // emit all the code in the segments + breakForLoop.push(false); + for (unsigned int s = 0; s < codeSegments.size(); ++s) { + builder.nextSwitchSegment(segmentBlocks, s); + if (codeSegments[s]) + codeSegments[s]->traverse(this); + else + builder.addSwitchBreak(); + } + breakForLoop.pop(); + + builder.endSwitch(segmentBlocks); + + return false; +} + +void TGlslangToSpvTraverser::visitConstantUnion(glslang::TIntermConstantUnion* node) +{ + int nextConst = 0; + spv::Id constant = createSpvConstantFromConstUnionArray(node->getType(), node->getConstArray(), nextConst, false); + + builder.clearAccessChain(); + builder.setAccessChainRValue(constant); +} + +bool TGlslangToSpvTraverser::visitLoop(glslang::TVisit /* visit */, glslang::TIntermLoop* node) +{ + auto blocks = builder.makeNewLoop(); + builder.createBranch(&blocks.head); + + // Loop control: + std::vector<unsigned int> operands; + const spv::LoopControlMask control = TranslateLoopControl(*node, operands); + + // Spec requires back edges to target header blocks, and every header block + // must dominate its merge block. Make a header block first to ensure these + // conditions are met. By definition, it will contain OpLoopMerge, followed + // by a block-ending branch. But we don't want to put any other body/test + // instructions in it, since the body/test may have arbitrary instructions, + // including merges of its own. + builder.setLine(node->getLoc().line, node->getLoc().getFilename()); + builder.setBuildPoint(&blocks.head); + builder.createLoopMerge(&blocks.merge, &blocks.continue_target, control, operands); + if (node->testFirst() && node->getTest()) { + spv::Block& test = builder.makeNewBlock(); + builder.createBranch(&test); + + builder.setBuildPoint(&test); + node->getTest()->traverse(this); + spv::Id condition = accessChainLoad(node->getTest()->getType()); + builder.createConditionalBranch(condition, &blocks.body, &blocks.merge); + + builder.setBuildPoint(&blocks.body); + breakForLoop.push(true); + if (node->getBody()) + node->getBody()->traverse(this); + builder.createBranch(&blocks.continue_target); + breakForLoop.pop(); + + builder.setBuildPoint(&blocks.continue_target); + if (node->getTerminal()) + node->getTerminal()->traverse(this); + builder.createBranch(&blocks.head); + } else { + builder.setLine(node->getLoc().line, node->getLoc().getFilename()); + builder.createBranch(&blocks.body); + + breakForLoop.push(true); + builder.setBuildPoint(&blocks.body); + if (node->getBody()) + node->getBody()->traverse(this); + builder.createBranch(&blocks.continue_target); + breakForLoop.pop(); + + builder.setBuildPoint(&blocks.continue_target); + if (node->getTerminal()) + node->getTerminal()->traverse(this); + if (node->getTest()) { + node->getTest()->traverse(this); + spv::Id condition = + accessChainLoad(node->getTest()->getType()); + builder.createConditionalBranch(condition, &blocks.head, &blocks.merge); + } else { + // TODO: unless there was a break/return/discard instruction + // somewhere in the body, this is an infinite loop, so we should + // issue a warning. + builder.createBranch(&blocks.head); + } + } + builder.setBuildPoint(&blocks.merge); + builder.closeLoop(); + return false; +} + +bool TGlslangToSpvTraverser::visitBranch(glslang::TVisit /* visit */, glslang::TIntermBranch* node) +{ + if (node->getExpression()) + node->getExpression()->traverse(this); + + builder.setLine(node->getLoc().line, node->getLoc().getFilename()); + + switch (node->getFlowOp()) { + case glslang::EOpKill: + builder.makeDiscard(); + break; + case glslang::EOpBreak: + if (breakForLoop.top()) + builder.createLoopExit(); + else + builder.addSwitchBreak(); + break; + case glslang::EOpContinue: + builder.createLoopContinue(); + break; + case glslang::EOpReturn: + if (node->getExpression()) { + const glslang::TType& glslangReturnType = node->getExpression()->getType(); + spv::Id returnId = accessChainLoad(glslangReturnType); + if (builder.getTypeId(returnId) != currentFunction->getReturnType()) { + builder.clearAccessChain(); + spv::Id copyId = builder.createVariable(spv::StorageClassFunction, currentFunction->getReturnType()); + builder.setAccessChainLValue(copyId); + multiTypeStore(glslangReturnType, returnId); + returnId = builder.createLoad(copyId); + } + builder.makeReturn(false, returnId); + } else + builder.makeReturn(false); + + builder.clearAccessChain(); + break; + + default: + assert(0); + break; + } + + return false; +} + +spv::Id TGlslangToSpvTraverser::createSpvVariable(const glslang::TIntermSymbol* node) +{ + // First, steer off constants, which are not SPIR-V variables, but + // can still have a mapping to a SPIR-V Id. + // This includes specialization constants. + if (node->getQualifier().isConstant()) { + spv::Id result = createSpvConstant(*node); + if (result != spv::NoResult) + return result; + } + + // Now, handle actual variables + spv::StorageClass storageClass = TranslateStorageClass(node->getType()); + spv::Id spvType = convertGlslangToSpvType(node->getType()); + + const bool contains16BitType = node->getType().containsBasicType(glslang::EbtFloat16) || + node->getType().containsBasicType(glslang::EbtInt16) || + node->getType().containsBasicType(glslang::EbtUint16); + if (contains16BitType) { + switch (storageClass) { + case spv::StorageClassInput: + case spv::StorageClassOutput: + addPre13Extension(spv::E_SPV_KHR_16bit_storage); + builder.addCapability(spv::CapabilityStorageInputOutput16); + break; + case spv::StorageClassPushConstant: + addPre13Extension(spv::E_SPV_KHR_16bit_storage); + builder.addCapability(spv::CapabilityStoragePushConstant16); + break; + case spv::StorageClassUniform: + addPre13Extension(spv::E_SPV_KHR_16bit_storage); + if (node->getType().getQualifier().storage == glslang::EvqBuffer) + builder.addCapability(spv::CapabilityStorageUniformBufferBlock16); + else + builder.addCapability(spv::CapabilityStorageUniform16); + break; + case spv::StorageClassStorageBuffer: + case spv::StorageClassPhysicalStorageBufferEXT: + addPre13Extension(spv::E_SPV_KHR_16bit_storage); + builder.addCapability(spv::CapabilityStorageUniformBufferBlock16); + break; + default: + break; + } + } + + const bool contains8BitType = node->getType().containsBasicType(glslang::EbtInt8) || + node->getType().containsBasicType(glslang::EbtUint8); + if (contains8BitType) { + if (storageClass == spv::StorageClassPushConstant) { + builder.addExtension(spv::E_SPV_KHR_8bit_storage); + builder.addCapability(spv::CapabilityStoragePushConstant8); + } else if (storageClass == spv::StorageClassUniform) { + builder.addExtension(spv::E_SPV_KHR_8bit_storage); + builder.addCapability(spv::CapabilityUniformAndStorageBuffer8BitAccess); + } else if (storageClass == spv::StorageClassStorageBuffer) { + builder.addExtension(spv::E_SPV_KHR_8bit_storage); + builder.addCapability(spv::CapabilityStorageBuffer8BitAccess); + } + } + + const char* name = node->getName().c_str(); + if (glslang::IsAnonymous(name)) + name = ""; + + return builder.createVariable(storageClass, spvType, name); +} + +// Return type Id of the sampled type. +spv::Id TGlslangToSpvTraverser::getSampledType(const glslang::TSampler& sampler) +{ + switch (sampler.type) { + case glslang::EbtFloat: return builder.makeFloatType(32); +#ifdef AMD_EXTENSIONS + case glslang::EbtFloat16: + builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float_fetch); + builder.addCapability(spv::CapabilityFloat16ImageAMD); + return builder.makeFloatType(16); +#endif + case glslang::EbtInt: return builder.makeIntType(32); + case glslang::EbtUint: return builder.makeUintType(32); + default: + assert(0); + return builder.makeFloatType(32); + } +} + +// If node is a swizzle operation, return the type that should be used if +// the swizzle base is first consumed by another operation, before the swizzle +// is applied. +spv::Id TGlslangToSpvTraverser::getInvertedSwizzleType(const glslang::TIntermTyped& node) +{ + if (node.getAsOperator() && + node.getAsOperator()->getOp() == glslang::EOpVectorSwizzle) + return convertGlslangToSpvType(node.getAsBinaryNode()->getLeft()->getType()); + else + return spv::NoType; +} + +// When inverting a swizzle with a parent op, this function +// will apply the swizzle operation to a completed parent operation. +spv::Id TGlslangToSpvTraverser::createInvertedSwizzle(spv::Decoration precision, const glslang::TIntermTyped& node, spv::Id parentResult) +{ + std::vector<unsigned> swizzle; + convertSwizzle(*node.getAsBinaryNode()->getRight()->getAsAggregate(), swizzle); + return builder.createRvalueSwizzle(precision, convertGlslangToSpvType(node.getType()), parentResult, swizzle); +} + +// Convert a glslang AST swizzle node to a swizzle vector for building SPIR-V. +void TGlslangToSpvTraverser::convertSwizzle(const glslang::TIntermAggregate& node, std::vector<unsigned>& swizzle) +{ + const glslang::TIntermSequence& swizzleSequence = node.getSequence(); + for (int i = 0; i < (int)swizzleSequence.size(); ++i) + swizzle.push_back(swizzleSequence[i]->getAsConstantUnion()->getConstArray()[0].getIConst()); +} + +// Convert from a glslang type to an SPV type, by calling into a +// recursive version of this function. This establishes the inherited +// layout state rooted from the top-level type. +spv::Id TGlslangToSpvTraverser::convertGlslangToSpvType(const glslang::TType& type, bool forwardReferenceOnly) +{ + return convertGlslangToSpvType(type, getExplicitLayout(type), type.getQualifier(), false, forwardReferenceOnly); +} + +// Do full recursive conversion of an arbitrary glslang type to a SPIR-V Id. +// explicitLayout can be kept the same throughout the hierarchical recursive walk. +// Mutually recursive with convertGlslangStructToSpvType(). +spv::Id TGlslangToSpvTraverser::convertGlslangToSpvType(const glslang::TType& type, + glslang::TLayoutPacking explicitLayout, const glslang::TQualifier& qualifier, + bool lastBufferBlockMember, bool forwardReferenceOnly) +{ + spv::Id spvType = spv::NoResult; + + switch (type.getBasicType()) { + case glslang::EbtVoid: + spvType = builder.makeVoidType(); + assert (! type.isArray()); + break; + case glslang::EbtFloat: + spvType = builder.makeFloatType(32); + break; + case glslang::EbtDouble: + spvType = builder.makeFloatType(64); + break; + case glslang::EbtFloat16: + spvType = builder.makeFloatType(16); + break; + case glslang::EbtBool: + // "transparent" bool doesn't exist in SPIR-V. The GLSL convention is + // a 32-bit int where non-0 means true. + if (explicitLayout != glslang::ElpNone) + spvType = builder.makeUintType(32); + else + spvType = builder.makeBoolType(); + break; + case glslang::EbtInt8: + spvType = builder.makeIntType(8); + break; + case glslang::EbtUint8: + spvType = builder.makeUintType(8); + break; + case glslang::EbtInt16: + spvType = builder.makeIntType(16); + break; + case glslang::EbtUint16: + spvType = builder.makeUintType(16); + break; + case glslang::EbtInt: + spvType = builder.makeIntType(32); + break; + case glslang::EbtUint: + spvType = builder.makeUintType(32); + break; + case glslang::EbtInt64: + spvType = builder.makeIntType(64); + break; + case glslang::EbtUint64: + spvType = builder.makeUintType(64); + break; + case glslang::EbtAtomicUint: + builder.addCapability(spv::CapabilityAtomicStorage); + spvType = builder.makeUintType(32); + break; +#ifdef NV_EXTENSIONS + case glslang::EbtAccStructNV: + spvType = builder.makeAccelerationStructureNVType(); + break; +#endif + case glslang::EbtSampler: + { + const glslang::TSampler& sampler = type.getSampler(); + if (sampler.sampler) { + // pure sampler + spvType = builder.makeSamplerType(); + } else { + // an image is present, make its type + spvType = builder.makeImageType(getSampledType(sampler), TranslateDimensionality(sampler), sampler.shadow, sampler.arrayed, sampler.ms, + sampler.image ? 2 : 1, TranslateImageFormat(type)); + if (sampler.combined) { + // already has both image and sampler, make the combined type + spvType = builder.makeSampledImageType(spvType); + } + } + } + break; + case glslang::EbtStruct: + case glslang::EbtBlock: + { + // If we've seen this struct type, return it + const glslang::TTypeList* glslangMembers = type.getStruct(); + + // Try to share structs for different layouts, but not yet for other + // kinds of qualification (primarily not yet including interpolant qualification). + if (! HasNonLayoutQualifiers(type, qualifier)) + spvType = structMap[explicitLayout][qualifier.layoutMatrix][glslangMembers]; + if (spvType != spv::NoResult) + break; + + // else, we haven't seen it... + if (type.getBasicType() == glslang::EbtBlock) + memberRemapper[glslangMembers].resize(glslangMembers->size()); + spvType = convertGlslangStructToSpvType(type, glslangMembers, explicitLayout, qualifier); + } + break; + case glslang::EbtReference: + { + // Make the forward pointer, then recurse to convert the structure type, then + // patch up the forward pointer with a real pointer type. + if (forwardPointers.find(type.getReferentType()) == forwardPointers.end()) { + spv::Id forwardId = builder.makeForwardPointer(spv::StorageClassPhysicalStorageBufferEXT); + forwardPointers[type.getReferentType()] = forwardId; + } + spvType = forwardPointers[type.getReferentType()]; + if (!forwardReferenceOnly) { + spv::Id referentType = convertGlslangToSpvType(*type.getReferentType()); + builder.makePointerFromForwardPointer(spv::StorageClassPhysicalStorageBufferEXT, + forwardPointers[type.getReferentType()], + referentType); + } + } + break; + default: + assert(0); + break; + } + + if (type.isMatrix()) + spvType = builder.makeMatrixType(spvType, type.getMatrixCols(), type.getMatrixRows()); + else { + // If this variable has a vector element count greater than 1, create a SPIR-V vector + if (type.getVectorSize() > 1) + spvType = builder.makeVectorType(spvType, type.getVectorSize()); + } + + if (type.isCoopMat()) { + builder.addCapability(spv::CapabilityCooperativeMatrixNV); + builder.addExtension(spv::E_SPV_NV_cooperative_matrix); + if (type.getBasicType() == glslang::EbtFloat16) + builder.addCapability(spv::CapabilityFloat16); + + spv::Id scope = makeArraySizeId(*type.getTypeParameters(), 1); + spv::Id rows = makeArraySizeId(*type.getTypeParameters(), 2); + spv::Id cols = makeArraySizeId(*type.getTypeParameters(), 3); + + spvType = builder.makeCooperativeMatrixType(spvType, scope, rows, cols); + } + + if (type.isArray()) { + int stride = 0; // keep this 0 unless doing an explicit layout; 0 will mean no decoration, no stride + + // Do all but the outer dimension + if (type.getArraySizes()->getNumDims() > 1) { + // We need to decorate array strides for types needing explicit layout, except blocks. + if (explicitLayout != glslang::ElpNone && type.getBasicType() != glslang::EbtBlock) { + // Use a dummy glslang type for querying internal strides of + // arrays of arrays, but using just a one-dimensional array. + glslang::TType simpleArrayType(type, 0); // deference type of the array + while (simpleArrayType.getArraySizes()->getNumDims() > 1) + simpleArrayType.getArraySizes()->dereference(); + + // Will compute the higher-order strides here, rather than making a whole + // pile of types and doing repetitive recursion on their contents. + stride = getArrayStride(simpleArrayType, explicitLayout, qualifier.layoutMatrix); + } + + // make the arrays + for (int dim = type.getArraySizes()->getNumDims() - 1; dim > 0; --dim) { + spvType = builder.makeArrayType(spvType, makeArraySizeId(*type.getArraySizes(), dim), stride); + if (stride > 0) + builder.addDecoration(spvType, spv::DecorationArrayStride, stride); + stride *= type.getArraySizes()->getDimSize(dim); + } + } else { + // single-dimensional array, and don't yet have stride + + // We need to decorate array strides for types needing explicit layout, except blocks. + if (explicitLayout != glslang::ElpNone && type.getBasicType() != glslang::EbtBlock) + stride = getArrayStride(type, explicitLayout, qualifier.layoutMatrix); + } + + // Do the outer dimension, which might not be known for a runtime-sized array. + // (Unsized arrays that survive through linking will be runtime-sized arrays) + if (type.isSizedArray()) + spvType = builder.makeArrayType(spvType, makeArraySizeId(*type.getArraySizes(), 0), stride); + else { + if (!lastBufferBlockMember) { + builder.addExtension("SPV_EXT_descriptor_indexing"); + builder.addCapability(spv::CapabilityRuntimeDescriptorArrayEXT); + } + spvType = builder.makeRuntimeArray(spvType); + } + if (stride > 0) + builder.addDecoration(spvType, spv::DecorationArrayStride, stride); + } + + return spvType; +} + +// TODO: this functionality should exist at a higher level, in creating the AST +// +// Identify interface members that don't have their required extension turned on. +// +bool TGlslangToSpvTraverser::filterMember(const glslang::TType& member) +{ +#ifdef NV_EXTENSIONS + auto& extensions = glslangIntermediate->getRequestedExtensions(); + + if (member.getFieldName() == "gl_SecondaryViewportMaskNV" && + extensions.find("GL_NV_stereo_view_rendering") == extensions.end()) + return true; + if (member.getFieldName() == "gl_SecondaryPositionNV" && + extensions.find("GL_NV_stereo_view_rendering") == extensions.end()) + return true; + + if (glslangIntermediate->getStage() != EShLangMeshNV) { + if (member.getFieldName() == "gl_ViewportMask" && + extensions.find("GL_NV_viewport_array2") == extensions.end()) + return true; + if (member.getFieldName() == "gl_PositionPerViewNV" && + extensions.find("GL_NVX_multiview_per_view_attributes") == extensions.end()) + return true; + if (member.getFieldName() == "gl_ViewportMaskPerViewNV" && + extensions.find("GL_NVX_multiview_per_view_attributes") == extensions.end()) + return true; + } +#endif + + return false; +}; + +// Do full recursive conversion of a glslang structure (or block) type to a SPIR-V Id. +// explicitLayout can be kept the same throughout the hierarchical recursive walk. +// Mutually recursive with convertGlslangToSpvType(). +spv::Id TGlslangToSpvTraverser::convertGlslangStructToSpvType(const glslang::TType& type, + const glslang::TTypeList* glslangMembers, + glslang::TLayoutPacking explicitLayout, + const glslang::TQualifier& qualifier) +{ + // Create a vector of struct types for SPIR-V to consume + std::vector<spv::Id> spvMembers; + int memberDelta = 0; // how much the member's index changes from glslang to SPIR-V, normally 0, except sometimes for blocks + std::vector<std::pair<glslang::TType*, glslang::TQualifier> > deferredForwardPointers; + for (int i = 0; i < (int)glslangMembers->size(); i++) { + glslang::TType& glslangMember = *(*glslangMembers)[i].type; + if (glslangMember.hiddenMember()) { + ++memberDelta; + if (type.getBasicType() == glslang::EbtBlock) + memberRemapper[glslangMembers][i] = -1; + } else { + if (type.getBasicType() == glslang::EbtBlock) { + memberRemapper[glslangMembers][i] = i - memberDelta; + if (filterMember(glslangMember)) + continue; + } + // modify just this child's view of the qualifier + glslang::TQualifier memberQualifier = glslangMember.getQualifier(); + InheritQualifiers(memberQualifier, qualifier); + + // manually inherit location + if (! memberQualifier.hasLocation() && qualifier.hasLocation()) + memberQualifier.layoutLocation = qualifier.layoutLocation; + + // recurse + bool lastBufferBlockMember = qualifier.storage == glslang::EvqBuffer && + i == (int)glslangMembers->size() - 1; + + // Make forward pointers for any pointer members, and create a list of members to + // convert to spirv types after creating the struct. + if (glslangMember.getBasicType() == glslang::EbtReference) { + if (forwardPointers.find(glslangMember.getReferentType()) == forwardPointers.end()) { + deferredForwardPointers.push_back(std::make_pair(&glslangMember, memberQualifier)); + } + spvMembers.push_back( + convertGlslangToSpvType(glslangMember, explicitLayout, memberQualifier, lastBufferBlockMember, true)); + } else { + spvMembers.push_back( + convertGlslangToSpvType(glslangMember, explicitLayout, memberQualifier, lastBufferBlockMember, false)); + } + } + } + + // Make the SPIR-V type + spv::Id spvType = builder.makeStructType(spvMembers, type.getTypeName().c_str()); + if (! HasNonLayoutQualifiers(type, qualifier)) + structMap[explicitLayout][qualifier.layoutMatrix][glslangMembers] = spvType; + + // Decorate it + decorateStructType(type, glslangMembers, explicitLayout, qualifier, spvType); + + for (int i = 0; i < (int)deferredForwardPointers.size(); ++i) { + auto it = deferredForwardPointers[i]; + convertGlslangToSpvType(*it.first, explicitLayout, it.second, false); + } + + return spvType; +} + +void TGlslangToSpvTraverser::decorateStructType(const glslang::TType& type, + const glslang::TTypeList* glslangMembers, + glslang::TLayoutPacking explicitLayout, + const glslang::TQualifier& qualifier, + spv::Id spvType) +{ + // Name and decorate the non-hidden members + int offset = -1; + int locationOffset = 0; // for use within the members of this struct + for (int i = 0; i < (int)glslangMembers->size(); i++) { + glslang::TType& glslangMember = *(*glslangMembers)[i].type; + int member = i; + if (type.getBasicType() == glslang::EbtBlock) { + member = memberRemapper[glslangMembers][i]; + if (filterMember(glslangMember)) + continue; + } + + // modify just this child's view of the qualifier + glslang::TQualifier memberQualifier = glslangMember.getQualifier(); + InheritQualifiers(memberQualifier, qualifier); + + // using -1 above to indicate a hidden member + if (member < 0) + continue; + + builder.addMemberName(spvType, member, glslangMember.getFieldName().c_str()); + builder.addMemberDecoration(spvType, member, + TranslateLayoutDecoration(glslangMember, memberQualifier.layoutMatrix)); + builder.addMemberDecoration(spvType, member, TranslatePrecisionDecoration(glslangMember)); + // Add interpolation and auxiliary storage decorations only to + // top-level members of Input and Output storage classes + if (type.getQualifier().storage == glslang::EvqVaryingIn || + type.getQualifier().storage == glslang::EvqVaryingOut) { + if (type.getBasicType() == glslang::EbtBlock || + glslangIntermediate->getSource() == glslang::EShSourceHlsl) { + builder.addMemberDecoration(spvType, member, TranslateInterpolationDecoration(memberQualifier)); + builder.addMemberDecoration(spvType, member, TranslateAuxiliaryStorageDecoration(memberQualifier)); +#ifdef NV_EXTENSIONS + addMeshNVDecoration(spvType, member, memberQualifier); +#endif + } + } + builder.addMemberDecoration(spvType, member, TranslateInvariantDecoration(memberQualifier)); + + if (type.getBasicType() == glslang::EbtBlock && + qualifier.storage == glslang::EvqBuffer) { + // Add memory decorations only to top-level members of shader storage block + std::vector<spv::Decoration> memory; + TranslateMemoryDecoration(memberQualifier, memory, glslangIntermediate->usingVulkanMemoryModel()); + for (unsigned int i = 0; i < memory.size(); ++i) + builder.addMemberDecoration(spvType, member, memory[i]); + } + + // Location assignment was already completed correctly by the front end, + // just track whether a member needs to be decorated. + // Ignore member locations if the container is an array, as that's + // ill-specified and decisions have been made to not allow this. + if (! type.isArray() && memberQualifier.hasLocation()) + builder.addMemberDecoration(spvType, member, spv::DecorationLocation, memberQualifier.layoutLocation); + + if (qualifier.hasLocation()) // track for upcoming inheritance + locationOffset += glslangIntermediate->computeTypeLocationSize( + glslangMember, glslangIntermediate->getStage()); + + // component, XFB, others + if (glslangMember.getQualifier().hasComponent()) + builder.addMemberDecoration(spvType, member, spv::DecorationComponent, + glslangMember.getQualifier().layoutComponent); + if (glslangMember.getQualifier().hasXfbOffset()) + builder.addMemberDecoration(spvType, member, spv::DecorationOffset, + glslangMember.getQualifier().layoutXfbOffset); + else if (explicitLayout != glslang::ElpNone) { + // figure out what to do with offset, which is accumulating + int nextOffset; + updateMemberOffset(type, glslangMember, offset, nextOffset, explicitLayout, memberQualifier.layoutMatrix); + if (offset >= 0) + builder.addMemberDecoration(spvType, member, spv::DecorationOffset, offset); + offset = nextOffset; + } + + if (glslangMember.isMatrix() && explicitLayout != glslang::ElpNone) + builder.addMemberDecoration(spvType, member, spv::DecorationMatrixStride, + getMatrixStride(glslangMember, explicitLayout, memberQualifier.layoutMatrix)); + + // built-in variable decorations + spv::BuiltIn builtIn = TranslateBuiltInDecoration(glslangMember.getQualifier().builtIn, true); + if (builtIn != spv::BuiltInMax) + builder.addMemberDecoration(spvType, member, spv::DecorationBuiltIn, (int)builtIn); + + // nonuniform + builder.addMemberDecoration(spvType, member, TranslateNonUniformDecoration(glslangMember.getQualifier())); + + if (glslangIntermediate->getHlslFunctionality1() && memberQualifier.semanticName != nullptr) { + builder.addExtension("SPV_GOOGLE_hlsl_functionality1"); + builder.addMemberDecoration(spvType, member, (spv::Decoration)spv::DecorationHlslSemanticGOOGLE, + memberQualifier.semanticName); + } + +#ifdef NV_EXTENSIONS + if (builtIn == spv::BuiltInLayer) { + // SPV_NV_viewport_array2 extension + if (glslangMember.getQualifier().layoutViewportRelative){ + builder.addMemberDecoration(spvType, member, (spv::Decoration)spv::DecorationViewportRelativeNV); + builder.addCapability(spv::CapabilityShaderViewportMaskNV); + builder.addExtension(spv::E_SPV_NV_viewport_array2); + } + if (glslangMember.getQualifier().layoutSecondaryViewportRelativeOffset != -2048){ + builder.addMemberDecoration(spvType, member, + (spv::Decoration)spv::DecorationSecondaryViewportRelativeNV, + glslangMember.getQualifier().layoutSecondaryViewportRelativeOffset); + builder.addCapability(spv::CapabilityShaderStereoViewNV); + builder.addExtension(spv::E_SPV_NV_stereo_view_rendering); + } + } + if (glslangMember.getQualifier().layoutPassthrough) { + builder.addMemberDecoration(spvType, member, (spv::Decoration)spv::DecorationPassthroughNV); + builder.addCapability(spv::CapabilityGeometryShaderPassthroughNV); + builder.addExtension(spv::E_SPV_NV_geometry_shader_passthrough); + } +#endif + } + + // Decorate the structure + builder.addDecoration(spvType, TranslateLayoutDecoration(type, qualifier.layoutMatrix)); + builder.addDecoration(spvType, TranslateBlockDecoration(type, glslangIntermediate->usingStorageBuffer())); +} + +// Turn the expression forming the array size into an id. +// This is not quite trivial, because of specialization constants. +// Sometimes, a raw constant is turned into an Id, and sometimes +// a specialization constant expression is. +spv::Id TGlslangToSpvTraverser::makeArraySizeId(const glslang::TArraySizes& arraySizes, int dim) +{ + // First, see if this is sized with a node, meaning a specialization constant: + glslang::TIntermTyped* specNode = arraySizes.getDimNode(dim); + if (specNode != nullptr) { + builder.clearAccessChain(); + specNode->traverse(this); + return accessChainLoad(specNode->getAsTyped()->getType()); + } + + // Otherwise, need a compile-time (front end) size, get it: + int size = arraySizes.getDimSize(dim); + assert(size > 0); + return builder.makeUintConstant(size); +} + +// Wrap the builder's accessChainLoad to: +// - localize handling of RelaxedPrecision +// - use the SPIR-V inferred type instead of another conversion of the glslang type +// (avoids unnecessary work and possible type punning for structures) +// - do conversion of concrete to abstract type +spv::Id TGlslangToSpvTraverser::accessChainLoad(const glslang::TType& type) +{ + spv::Id nominalTypeId = builder.accessChainGetInferredType(); + + spv::Builder::AccessChain::CoherentFlags coherentFlags = builder.getAccessChain().coherentFlags; + coherentFlags |= TranslateCoherent(type); + + unsigned int alignment = builder.getAccessChain().alignment; + alignment |= type.getBufferReferenceAlignment(); + + spv::Id loadedId = builder.accessChainLoad(TranslatePrecisionDecoration(type), + TranslateNonUniformDecoration(type.getQualifier()), + nominalTypeId, + spv::MemoryAccessMask(TranslateMemoryAccess(coherentFlags) & ~spv::MemoryAccessMakePointerAvailableKHRMask), + TranslateMemoryScope(coherentFlags), + alignment); + + // Need to convert to abstract types when necessary + if (type.getBasicType() == glslang::EbtBool) { + if (builder.isScalarType(nominalTypeId)) { + // Conversion for bool + spv::Id boolType = builder.makeBoolType(); + if (nominalTypeId != boolType) + loadedId = builder.createBinOp(spv::OpINotEqual, boolType, loadedId, builder.makeUintConstant(0)); + } else if (builder.isVectorType(nominalTypeId)) { + // Conversion for bvec + int vecSize = builder.getNumTypeComponents(nominalTypeId); + spv::Id bvecType = builder.makeVectorType(builder.makeBoolType(), vecSize); + if (nominalTypeId != bvecType) + loadedId = builder.createBinOp(spv::OpINotEqual, bvecType, loadedId, makeSmearedConstant(builder.makeUintConstant(0), vecSize)); + } + } + + return loadedId; +} + +// Wrap the builder's accessChainStore to: +// - do conversion of concrete to abstract type +// +// Implicitly uses the existing builder.accessChain as the storage target. +void TGlslangToSpvTraverser::accessChainStore(const glslang::TType& type, spv::Id rvalue) +{ + // Need to convert to abstract types when necessary + if (type.getBasicType() == glslang::EbtBool) { + spv::Id nominalTypeId = builder.accessChainGetInferredType(); + + if (builder.isScalarType(nominalTypeId)) { + // Conversion for bool + spv::Id boolType = builder.makeBoolType(); + if (nominalTypeId != boolType) { + // keep these outside arguments, for determinant order-of-evaluation + spv::Id one = builder.makeUintConstant(1); + spv::Id zero = builder.makeUintConstant(0); + rvalue = builder.createTriOp(spv::OpSelect, nominalTypeId, rvalue, one, zero); + } else if (builder.getTypeId(rvalue) != boolType) + rvalue = builder.createBinOp(spv::OpINotEqual, boolType, rvalue, builder.makeUintConstant(0)); + } else if (builder.isVectorType(nominalTypeId)) { + // Conversion for bvec + int vecSize = builder.getNumTypeComponents(nominalTypeId); + spv::Id bvecType = builder.makeVectorType(builder.makeBoolType(), vecSize); + if (nominalTypeId != bvecType) { + // keep these outside arguments, for determinant order-of-evaluation + spv::Id one = makeSmearedConstant(builder.makeUintConstant(1), vecSize); + spv::Id zero = makeSmearedConstant(builder.makeUintConstant(0), vecSize); + rvalue = builder.createTriOp(spv::OpSelect, nominalTypeId, rvalue, one, zero); + } else if (builder.getTypeId(rvalue) != bvecType) + rvalue = builder.createBinOp(spv::OpINotEqual, bvecType, rvalue, + makeSmearedConstant(builder.makeUintConstant(0), vecSize)); + } + } + + spv::Builder::AccessChain::CoherentFlags coherentFlags = builder.getAccessChain().coherentFlags; + coherentFlags |= TranslateCoherent(type); + + unsigned int alignment = builder.getAccessChain().alignment; + alignment |= type.getBufferReferenceAlignment(); + + builder.accessChainStore(rvalue, + spv::MemoryAccessMask(TranslateMemoryAccess(coherentFlags) & ~spv::MemoryAccessMakePointerVisibleKHRMask), + TranslateMemoryScope(coherentFlags), alignment); +} + +// For storing when types match at the glslang level, but not might match at the +// SPIR-V level. +// +// This especially happens when a single glslang type expands to multiple +// SPIR-V types, like a struct that is used in a member-undecorated way as well +// as in a member-decorated way. +// +// NOTE: This function can handle any store request; if it's not special it +// simplifies to a simple OpStore. +// +// Implicitly uses the existing builder.accessChain as the storage target. +void TGlslangToSpvTraverser::multiTypeStore(const glslang::TType& type, spv::Id rValue) +{ + // we only do the complex path here if it's an aggregate + if (! type.isStruct() && ! type.isArray()) { + accessChainStore(type, rValue); + return; + } + + // and, it has to be a case of type aliasing + spv::Id rType = builder.getTypeId(rValue); + spv::Id lValue = builder.accessChainGetLValue(); + spv::Id lType = builder.getContainedTypeId(builder.getTypeId(lValue)); + if (lType == rType) { + accessChainStore(type, rValue); + return; + } + + // Recursively (as needed) copy an aggregate type to a different aggregate type, + // where the two types were the same type in GLSL. This requires member + // by member copy, recursively. + + // SPIR-V 1.4 added an instruction to do help do this. + if (glslangIntermediate->getSpv().spv >= glslang::EShTargetSpv_1_4) { + // However, bool in uniform space is changed to int, so + // OpCopyLogical does not work for that. + // TODO: It would be more robust to do a full recursive verification of the types satisfying SPIR-V rules. + bool rBool = builder.containsType(builder.getTypeId(rValue), spv::OpTypeBool, 0); + bool lBool = builder.containsType(lType, spv::OpTypeBool, 0); + if (lBool == rBool) { + spv::Id logicalCopy = builder.createUnaryOp(spv::OpCopyLogical, lType, rValue); + accessChainStore(type, logicalCopy); + return; + } + } + + // If an array, copy element by element. + if (type.isArray()) { + glslang::TType glslangElementType(type, 0); + spv::Id elementRType = builder.getContainedTypeId(rType); + for (int index = 0; index < type.getOuterArraySize(); ++index) { + // get the source member + spv::Id elementRValue = builder.createCompositeExtract(rValue, elementRType, index); + + // set up the target storage + builder.clearAccessChain(); + builder.setAccessChainLValue(lValue); + builder.accessChainPush(builder.makeIntConstant(index), TranslateCoherent(type), type.getBufferReferenceAlignment()); + + // store the member + multiTypeStore(glslangElementType, elementRValue); + } + } else { + assert(type.isStruct()); + + // loop over structure members + const glslang::TTypeList& members = *type.getStruct(); + for (int m = 0; m < (int)members.size(); ++m) { + const glslang::TType& glslangMemberType = *members[m].type; + + // get the source member + spv::Id memberRType = builder.getContainedTypeId(rType, m); + spv::Id memberRValue = builder.createCompositeExtract(rValue, memberRType, m); + + // set up the target storage + builder.clearAccessChain(); + builder.setAccessChainLValue(lValue); + builder.accessChainPush(builder.makeIntConstant(m), TranslateCoherent(type), type.getBufferReferenceAlignment()); + + // store the member + multiTypeStore(glslangMemberType, memberRValue); + } + } +} + +// Decide whether or not this type should be +// decorated with offsets and strides, and if so +// whether std140 or std430 rules should be applied. +glslang::TLayoutPacking TGlslangToSpvTraverser::getExplicitLayout(const glslang::TType& type) const +{ + // has to be a block + if (type.getBasicType() != glslang::EbtBlock) + return glslang::ElpNone; + + // has to be a uniform or buffer block or task in/out blocks + if (type.getQualifier().storage != glslang::EvqUniform && + type.getQualifier().storage != glslang::EvqBuffer && + !type.getQualifier().isTaskMemory()) + return glslang::ElpNone; + + // return the layout to use + switch (type.getQualifier().layoutPacking) { + case glslang::ElpStd140: + case glslang::ElpStd430: + case glslang::ElpScalar: + return type.getQualifier().layoutPacking; + default: + return glslang::ElpNone; + } +} + +// Given an array type, returns the integer stride required for that array +int TGlslangToSpvTraverser::getArrayStride(const glslang::TType& arrayType, glslang::TLayoutPacking explicitLayout, glslang::TLayoutMatrix matrixLayout) +{ + int size; + int stride; + glslangIntermediate->getMemberAlignment(arrayType, size, stride, explicitLayout, matrixLayout == glslang::ElmRowMajor); + + return stride; +} + +// Given a matrix type, or array (of array) of matrixes type, returns the integer stride required for that matrix +// when used as a member of an interface block +int TGlslangToSpvTraverser::getMatrixStride(const glslang::TType& matrixType, glslang::TLayoutPacking explicitLayout, glslang::TLayoutMatrix matrixLayout) +{ + glslang::TType elementType; + elementType.shallowCopy(matrixType); + elementType.clearArraySizes(); + + int size; + int stride; + glslangIntermediate->getMemberAlignment(elementType, size, stride, explicitLayout, matrixLayout == glslang::ElmRowMajor); + + return stride; +} + +// Given a member type of a struct, realign the current offset for it, and compute +// the next (not yet aligned) offset for the next member, which will get aligned +// on the next call. +// 'currentOffset' should be passed in already initialized, ready to modify, and reflecting +// the migration of data from nextOffset -> currentOffset. It should be -1 on the first call. +// -1 means a non-forced member offset (no decoration needed). +void TGlslangToSpvTraverser::updateMemberOffset(const glslang::TType& structType, const glslang::TType& memberType, int& currentOffset, int& nextOffset, + glslang::TLayoutPacking explicitLayout, glslang::TLayoutMatrix matrixLayout) +{ + // this will get a positive value when deemed necessary + nextOffset = -1; + + // override anything in currentOffset with user-set offset + if (memberType.getQualifier().hasOffset()) + currentOffset = memberType.getQualifier().layoutOffset; + + // It could be that current linker usage in glslang updated all the layoutOffset, + // in which case the following code does not matter. But, that's not quite right + // once cross-compilation unit GLSL validation is done, as the original user + // settings are needed in layoutOffset, and then the following will come into play. + + if (explicitLayout == glslang::ElpNone) { + if (! memberType.getQualifier().hasOffset()) + currentOffset = -1; + + return; + } + + // Getting this far means we need explicit offsets + if (currentOffset < 0) + currentOffset = 0; + + // Now, currentOffset is valid (either 0, or from a previous nextOffset), + // but possibly not yet correctly aligned. + + int memberSize; + int dummyStride; + int memberAlignment = glslangIntermediate->getMemberAlignment(memberType, memberSize, dummyStride, explicitLayout, matrixLayout == glslang::ElmRowMajor); + + // Adjust alignment for HLSL rules + // TODO: make this consistent in early phases of code: + // adjusting this late means inconsistencies with earlier code, which for reflection is an issue + // Until reflection is brought in sync with these adjustments, don't apply to $Global, + // which is the most likely to rely on reflection, and least likely to rely implicit layouts + if (glslangIntermediate->usingHlslOffsets() && + ! memberType.isArray() && memberType.isVector() && structType.getTypeName().compare("$Global") != 0) { + int dummySize; + int componentAlignment = glslangIntermediate->getBaseAlignmentScalar(memberType, dummySize); + if (componentAlignment <= 4) + memberAlignment = componentAlignment; + } + + // Bump up to member alignment + glslang::RoundToPow2(currentOffset, memberAlignment); + + // Bump up to vec4 if there is a bad straddle + if (explicitLayout != glslang::ElpScalar && glslangIntermediate->improperStraddle(memberType, memberSize, currentOffset)) + glslang::RoundToPow2(currentOffset, 16); + + nextOffset = currentOffset + memberSize; +} + +void TGlslangToSpvTraverser::declareUseOfStructMember(const glslang::TTypeList& members, int glslangMember) +{ + const glslang::TBuiltInVariable glslangBuiltIn = members[glslangMember].type->getQualifier().builtIn; + switch (glslangBuiltIn) + { + case glslang::EbvClipDistance: + case glslang::EbvCullDistance: + case glslang::EbvPointSize: +#ifdef NV_EXTENSIONS + case glslang::EbvViewportMaskNV: + case glslang::EbvSecondaryPositionNV: + case glslang::EbvSecondaryViewportMaskNV: + case glslang::EbvPositionPerViewNV: + case glslang::EbvViewportMaskPerViewNV: + case glslang::EbvTaskCountNV: + case glslang::EbvPrimitiveCountNV: + case glslang::EbvPrimitiveIndicesNV: + case glslang::EbvClipDistancePerViewNV: + case glslang::EbvCullDistancePerViewNV: + case glslang::EbvLayerPerViewNV: + case glslang::EbvMeshViewCountNV: + case glslang::EbvMeshViewIndicesNV: +#endif + // Generate the associated capability. Delegate to TranslateBuiltInDecoration. + // Alternately, we could just call this for any glslang built-in, since the + // capability already guards against duplicates. + TranslateBuiltInDecoration(glslangBuiltIn, false); + break; + default: + // Capabilities were already generated when the struct was declared. + break; + } +} + +bool TGlslangToSpvTraverser::isShaderEntryPoint(const glslang::TIntermAggregate* node) +{ + return node->getName().compare(glslangIntermediate->getEntryPointMangledName().c_str()) == 0; +} + +// Does parameter need a place to keep writes, separate from the original? +// Assumes called after originalParam(), which filters out block/buffer/opaque-based +// qualifiers such that we should have only in/out/inout/constreadonly here. +bool TGlslangToSpvTraverser::writableParam(glslang::TStorageQualifier qualifier) const +{ + assert(qualifier == glslang::EvqIn || + qualifier == glslang::EvqOut || + qualifier == glslang::EvqInOut || + qualifier == glslang::EvqConstReadOnly); + return qualifier != glslang::EvqConstReadOnly; +} + +// Is parameter pass-by-original? +bool TGlslangToSpvTraverser::originalParam(glslang::TStorageQualifier qualifier, const glslang::TType& paramType, + bool implicitThisParam) +{ + if (implicitThisParam) // implicit this + return true; + if (glslangIntermediate->getSource() == glslang::EShSourceHlsl) + return paramType.getBasicType() == glslang::EbtBlock; + return paramType.containsOpaque() || // sampler, etc. + (paramType.getBasicType() == glslang::EbtBlock && qualifier == glslang::EvqBuffer); // SSBO +} + +// Make all the functions, skeletally, without actually visiting their bodies. +void TGlslangToSpvTraverser::makeFunctions(const glslang::TIntermSequence& glslFunctions) +{ + const auto getParamDecorations = [&](std::vector<spv::Decoration>& decorations, const glslang::TType& type, bool useVulkanMemoryModel) { + spv::Decoration paramPrecision = TranslatePrecisionDecoration(type); + if (paramPrecision != spv::NoPrecision) + decorations.push_back(paramPrecision); + TranslateMemoryDecoration(type.getQualifier(), decorations, useVulkanMemoryModel); + if (type.getBasicType() == glslang::EbtReference) { + // Original and non-writable params pass the pointer directly and + // use restrict/aliased, others are stored to a pointer in Function + // memory and use RestrictPointer/AliasedPointer. + if (originalParam(type.getQualifier().storage, type, false) || + !writableParam(type.getQualifier().storage)) { + decorations.push_back(type.getQualifier().restrict ? spv::DecorationRestrict : spv::DecorationAliased); + } else { + decorations.push_back(type.getQualifier().restrict ? spv::DecorationRestrictPointerEXT : spv::DecorationAliasedPointerEXT); + } + } + }; + + for (int f = 0; f < (int)glslFunctions.size(); ++f) { + glslang::TIntermAggregate* glslFunction = glslFunctions[f]->getAsAggregate(); + if (! glslFunction || glslFunction->getOp() != glslang::EOpFunction || isShaderEntryPoint(glslFunction)) + continue; + + // We're on a user function. Set up the basic interface for the function now, + // so that it's available to call. Translating the body will happen later. + // + // Typically (except for a "const in" parameter), an address will be passed to the + // function. What it is an address of varies: + // + // - "in" parameters not marked as "const" can be written to without modifying the calling + // argument so that write needs to be to a copy, hence the address of a copy works. + // + // - "const in" parameters can just be the r-value, as no writes need occur. + // + // - "out" and "inout" arguments can't be done as pointers to the calling argument, because + // GLSL has copy-in/copy-out semantics. They can be handled though with a pointer to a copy. + + std::vector<spv::Id> paramTypes; + std::vector<std::vector<spv::Decoration>> paramDecorations; // list of decorations per parameter + glslang::TIntermSequence& parameters = glslFunction->getSequence()[0]->getAsAggregate()->getSequence(); + + bool implicitThis = (int)parameters.size() > 0 && parameters[0]->getAsSymbolNode()->getName() == + glslangIntermediate->implicitThisName; + + paramDecorations.resize(parameters.size()); + for (int p = 0; p < (int)parameters.size(); ++p) { + const glslang::TType& paramType = parameters[p]->getAsTyped()->getType(); + spv::Id typeId = convertGlslangToSpvType(paramType); + if (originalParam(paramType.getQualifier().storage, paramType, implicitThis && p == 0)) + typeId = builder.makePointer(TranslateStorageClass(paramType), typeId); + else if (writableParam(paramType.getQualifier().storage)) + typeId = builder.makePointer(spv::StorageClassFunction, typeId); + else + rValueParameters.insert(parameters[p]->getAsSymbolNode()->getId()); + getParamDecorations(paramDecorations[p], paramType, glslangIntermediate->usingVulkanMemoryModel()); + paramTypes.push_back(typeId); + } + + spv::Block* functionBlock; + spv::Function *function = builder.makeFunctionEntry(TranslatePrecisionDecoration(glslFunction->getType()), + convertGlslangToSpvType(glslFunction->getType()), + glslFunction->getName().c_str(), paramTypes, + paramDecorations, &functionBlock); + if (implicitThis) + function->setImplicitThis(); + + // Track function to emit/call later + functionMap[glslFunction->getName().c_str()] = function; + + // Set the parameter id's + for (int p = 0; p < (int)parameters.size(); ++p) { + symbolValues[parameters[p]->getAsSymbolNode()->getId()] = function->getParamId(p); + // give a name too + builder.addName(function->getParamId(p), parameters[p]->getAsSymbolNode()->getName().c_str()); + } + } +} + +// Process all the initializers, while skipping the functions and link objects +void TGlslangToSpvTraverser::makeGlobalInitializers(const glslang::TIntermSequence& initializers) +{ + builder.setBuildPoint(shaderEntry->getLastBlock()); + for (int i = 0; i < (int)initializers.size(); ++i) { + glslang::TIntermAggregate* initializer = initializers[i]->getAsAggregate(); + if (initializer && initializer->getOp() != glslang::EOpFunction && initializer->getOp() != glslang::EOpLinkerObjects) { + + // We're on a top-level node that's not a function. Treat as an initializer, whose + // code goes into the beginning of the entry point. + initializer->traverse(this); + } + } +} + +// Process all the functions, while skipping initializers. +void TGlslangToSpvTraverser::visitFunctions(const glslang::TIntermSequence& glslFunctions) +{ + for (int f = 0; f < (int)glslFunctions.size(); ++f) { + glslang::TIntermAggregate* node = glslFunctions[f]->getAsAggregate(); + if (node && (node->getOp() == glslang::EOpFunction || node->getOp() == glslang::EOpLinkerObjects)) + node->traverse(this); + } +} + +void TGlslangToSpvTraverser::handleFunctionEntry(const glslang::TIntermAggregate* node) +{ + // SPIR-V functions should already be in the functionMap from the prepass + // that called makeFunctions(). + currentFunction = functionMap[node->getName().c_str()]; + spv::Block* functionBlock = currentFunction->getEntryBlock(); + builder.setBuildPoint(functionBlock); +} + +void TGlslangToSpvTraverser::translateArguments(const glslang::TIntermAggregate& node, std::vector<spv::Id>& arguments) +{ + const glslang::TIntermSequence& glslangArguments = node.getSequence(); + + glslang::TSampler sampler = {}; + bool cubeCompare = false; +#ifdef AMD_EXTENSIONS + bool f16ShadowCompare = false; +#endif + if (node.isTexture() || node.isImage()) { + sampler = glslangArguments[0]->getAsTyped()->getType().getSampler(); + cubeCompare = sampler.dim == glslang::EsdCube && sampler.arrayed && sampler.shadow; +#ifdef AMD_EXTENSIONS + f16ShadowCompare = sampler.shadow && glslangArguments[1]->getAsTyped()->getType().getBasicType() == glslang::EbtFloat16; +#endif + } + + for (int i = 0; i < (int)glslangArguments.size(); ++i) { + builder.clearAccessChain(); + glslangArguments[i]->traverse(this); + + // Special case l-value operands + bool lvalue = false; + switch (node.getOp()) { + case glslang::EOpImageAtomicAdd: + case glslang::EOpImageAtomicMin: + case glslang::EOpImageAtomicMax: + case glslang::EOpImageAtomicAnd: + case glslang::EOpImageAtomicOr: + case glslang::EOpImageAtomicXor: + case glslang::EOpImageAtomicExchange: + case glslang::EOpImageAtomicCompSwap: + case glslang::EOpImageAtomicLoad: + case glslang::EOpImageAtomicStore: + if (i == 0) + lvalue = true; + break; + case glslang::EOpSparseImageLoad: + if ((sampler.ms && i == 3) || (! sampler.ms && i == 2)) + lvalue = true; + break; +#ifdef AMD_EXTENSIONS + case glslang::EOpSparseTexture: + if (((cubeCompare || f16ShadowCompare) && i == 3) || (! (cubeCompare || f16ShadowCompare) && i == 2)) + lvalue = true; + break; + case glslang::EOpSparseTextureClamp: + if (((cubeCompare || f16ShadowCompare) && i == 4) || (! (cubeCompare || f16ShadowCompare) && i == 3)) + lvalue = true; + break; + case glslang::EOpSparseTextureLod: + case glslang::EOpSparseTextureOffset: + if ((f16ShadowCompare && i == 4) || (! f16ShadowCompare && i == 3)) + lvalue = true; + break; +#else + case glslang::EOpSparseTexture: + if ((cubeCompare && i == 3) || (! cubeCompare && i == 2)) + lvalue = true; + break; + case glslang::EOpSparseTextureClamp: + if ((cubeCompare && i == 4) || (! cubeCompare && i == 3)) + lvalue = true; + break; + case glslang::EOpSparseTextureLod: + case glslang::EOpSparseTextureOffset: + if (i == 3) + lvalue = true; + break; +#endif + case glslang::EOpSparseTextureFetch: + if ((sampler.dim != glslang::EsdRect && i == 3) || (sampler.dim == glslang::EsdRect && i == 2)) + lvalue = true; + break; + case glslang::EOpSparseTextureFetchOffset: + if ((sampler.dim != glslang::EsdRect && i == 4) || (sampler.dim == glslang::EsdRect && i == 3)) + lvalue = true; + break; +#ifdef AMD_EXTENSIONS + case glslang::EOpSparseTextureLodOffset: + case glslang::EOpSparseTextureGrad: + case glslang::EOpSparseTextureOffsetClamp: + if ((f16ShadowCompare && i == 5) || (! f16ShadowCompare && i == 4)) + lvalue = true; + break; + case glslang::EOpSparseTextureGradOffset: + case glslang::EOpSparseTextureGradClamp: + if ((f16ShadowCompare && i == 6) || (! f16ShadowCompare && i == 5)) + lvalue = true; + break; + case glslang::EOpSparseTextureGradOffsetClamp: + if ((f16ShadowCompare && i == 7) || (! f16ShadowCompare && i == 6)) + lvalue = true; + break; +#else + case glslang::EOpSparseTextureLodOffset: + case glslang::EOpSparseTextureGrad: + case glslang::EOpSparseTextureOffsetClamp: + if (i == 4) + lvalue = true; + break; + case glslang::EOpSparseTextureGradOffset: + case glslang::EOpSparseTextureGradClamp: + if (i == 5) + lvalue = true; + break; + case glslang::EOpSparseTextureGradOffsetClamp: + if (i == 6) + lvalue = true; + break; +#endif + case glslang::EOpSparseTextureGather: + if ((sampler.shadow && i == 3) || (! sampler.shadow && i == 2)) + lvalue = true; + break; + case glslang::EOpSparseTextureGatherOffset: + case glslang::EOpSparseTextureGatherOffsets: + if ((sampler.shadow && i == 4) || (! sampler.shadow && i == 3)) + lvalue = true; + break; +#ifdef AMD_EXTENSIONS + case glslang::EOpSparseTextureGatherLod: + if (i == 3) + lvalue = true; + break; + case glslang::EOpSparseTextureGatherLodOffset: + case glslang::EOpSparseTextureGatherLodOffsets: + if (i == 4) + lvalue = true; + break; + case glslang::EOpSparseImageLoadLod: + if (i == 3) + lvalue = true; + break; +#endif +#ifdef NV_EXTENSIONS + case glslang::EOpImageSampleFootprintNV: + if (i == 4) + lvalue = true; + break; + case glslang::EOpImageSampleFootprintClampNV: + case glslang::EOpImageSampleFootprintLodNV: + if (i == 5) + lvalue = true; + break; + case glslang::EOpImageSampleFootprintGradNV: + if (i == 6) + lvalue = true; + break; + case glslang::EOpImageSampleFootprintGradClampNV: + if (i == 7) + lvalue = true; + break; +#endif + default: + break; + } + + if (lvalue) + arguments.push_back(builder.accessChainGetLValue()); + else + arguments.push_back(accessChainLoad(glslangArguments[i]->getAsTyped()->getType())); + } +} + +void TGlslangToSpvTraverser::translateArguments(glslang::TIntermUnary& node, std::vector<spv::Id>& arguments) +{ + builder.clearAccessChain(); + node.getOperand()->traverse(this); + arguments.push_back(accessChainLoad(node.getOperand()->getType())); +} + +spv::Id TGlslangToSpvTraverser::createImageTextureFunctionCall(glslang::TIntermOperator* node) +{ + if (! node->isImage() && ! node->isTexture()) + return spv::NoResult; + + builder.setLine(node->getLoc().line, node->getLoc().getFilename()); + + // Process a GLSL texturing op (will be SPV image) + + const glslang::TType &imageType = node->getAsAggregate() + ? node->getAsAggregate()->getSequence()[0]->getAsTyped()->getType() + : node->getAsUnaryNode()->getOperand()->getAsTyped()->getType(); + const glslang::TSampler sampler = imageType.getSampler(); +#ifdef AMD_EXTENSIONS + bool f16ShadowCompare = (sampler.shadow && node->getAsAggregate()) + ? node->getAsAggregate()->getSequence()[1]->getAsTyped()->getType().getBasicType() == glslang::EbtFloat16 + : false; +#endif + + const auto signExtensionMask = [&]() { + if (builder.getSpvVersion() >= spv::Spv_1_4) { + if (sampler.type == glslang::EbtUint) + return spv::ImageOperandsZeroExtendMask; + else if (sampler.type == glslang::EbtInt) + return spv::ImageOperandsSignExtendMask; + } + return spv::ImageOperandsMaskNone; + }; + + std::vector<spv::Id> arguments; + if (node->getAsAggregate()) + translateArguments(*node->getAsAggregate(), arguments); + else + translateArguments(*node->getAsUnaryNode(), arguments); + spv::Decoration precision = TranslatePrecisionDecoration(node->getOperationPrecision()); + + spv::Builder::TextureParameters params = { }; + params.sampler = arguments[0]; + + glslang::TCrackedTextureOp cracked; + node->crackTexture(sampler, cracked); + + const bool isUnsignedResult = node->getType().getBasicType() == glslang::EbtUint; + + // Check for queries + if (cracked.query) { + // OpImageQueryLod works on a sampled image, for other queries the image has to be extracted first + if (node->getOp() != glslang::EOpTextureQueryLod && builder.isSampledImage(params.sampler)) + params.sampler = builder.createUnaryOp(spv::OpImage, builder.getImageType(params.sampler), params.sampler); + + switch (node->getOp()) { + case glslang::EOpImageQuerySize: + case glslang::EOpTextureQuerySize: + if (arguments.size() > 1) { + params.lod = arguments[1]; + return builder.createTextureQueryCall(spv::OpImageQuerySizeLod, params, isUnsignedResult); + } else + return builder.createTextureQueryCall(spv::OpImageQuerySize, params, isUnsignedResult); + case glslang::EOpImageQuerySamples: + case glslang::EOpTextureQuerySamples: + return builder.createTextureQueryCall(spv::OpImageQuerySamples, params, isUnsignedResult); + case glslang::EOpTextureQueryLod: + params.coords = arguments[1]; + return builder.createTextureQueryCall(spv::OpImageQueryLod, params, isUnsignedResult); + case glslang::EOpTextureQueryLevels: + return builder.createTextureQueryCall(spv::OpImageQueryLevels, params, isUnsignedResult); + case glslang::EOpSparseTexelsResident: + return builder.createUnaryOp(spv::OpImageSparseTexelsResident, builder.makeBoolType(), arguments[0]); + default: + assert(0); + break; + } + } + + int components = node->getType().getVectorSize(); + + if (node->getOp() == glslang::EOpTextureFetch) { + // These must produce 4 components, per SPIR-V spec. We'll add a conversion constructor if needed. + // This will only happen through the HLSL path for operator[], so we do not have to handle e.g. + // the EOpTexture/Proj/Lod/etc family. It would be harmless to do so, but would need more logic + // here around e.g. which ones return scalars or other types. + components = 4; + } + + glslang::TType returnType(node->getType().getBasicType(), glslang::EvqTemporary, components); + + auto resultType = [&returnType,this]{ return convertGlslangToSpvType(returnType); }; + + // Check for image functions other than queries + if (node->isImage()) { + std::vector<spv::IdImmediate> operands; + auto opIt = arguments.begin(); + spv::IdImmediate image = { true, *(opIt++) }; + operands.push_back(image); + + // Handle subpass operations + // TODO: GLSL should change to have the "MS" only on the type rather than the + // built-in function. + if (cracked.subpass) { + // add on the (0,0) coordinate + spv::Id zero = builder.makeIntConstant(0); + std::vector<spv::Id> comps; + comps.push_back(zero); + comps.push_back(zero); + spv::IdImmediate coord = { true, + builder.makeCompositeConstant(builder.makeVectorType(builder.makeIntType(32), 2), comps) }; + operands.push_back(coord); + spv::IdImmediate imageOperands = { false, spv::ImageOperandsMaskNone }; + imageOperands.word = imageOperands.word | signExtensionMask(); + if (sampler.ms) { + imageOperands.word = imageOperands.word | spv::ImageOperandsSampleMask; + } + if (imageOperands.word != spv::ImageOperandsMaskNone) { + operands.push_back(imageOperands); + if (sampler.ms) { + spv::IdImmediate imageOperand = { true, *(opIt++) }; + operands.push_back(imageOperand); + } + } + spv::Id result = builder.createOp(spv::OpImageRead, resultType(), operands); + builder.setPrecision(result, precision); + return result; + } + + spv::IdImmediate coord = { true, *(opIt++) }; + operands.push_back(coord); +#ifdef AMD_EXTENSIONS + if (node->getOp() == glslang::EOpImageLoad || node->getOp() == glslang::EOpImageLoadLod) { +#else + if (node->getOp() == glslang::EOpImageLoad) { +#endif + spv::ImageOperandsMask mask = spv::ImageOperandsMaskNone; + if (sampler.ms) { + mask = mask | spv::ImageOperandsSampleMask; + } +#ifdef AMD_EXTENSIONS + if (cracked.lod) { + builder.addExtension(spv::E_SPV_AMD_shader_image_load_store_lod); + builder.addCapability(spv::CapabilityImageReadWriteLodAMD); + mask = mask | spv::ImageOperandsLodMask; + } +#endif + mask = mask | TranslateImageOperands(TranslateCoherent(imageType)); + mask = (spv::ImageOperandsMask)(mask & ~spv::ImageOperandsMakeTexelAvailableKHRMask); + mask = mask | signExtensionMask(); + if (mask != spv::ImageOperandsMaskNone) { + spv::IdImmediate imageOperands = { false, (unsigned int)mask }; + operands.push_back(imageOperands); + } + if (mask & spv::ImageOperandsSampleMask) { + spv::IdImmediate imageOperand = { true, *opIt++ }; + operands.push_back(imageOperand); + } +#ifdef AMD_EXTENSIONS + if (mask & spv::ImageOperandsLodMask) { + spv::IdImmediate imageOperand = { true, *opIt++ }; + operands.push_back(imageOperand); + } +#endif + if (mask & spv::ImageOperandsMakeTexelVisibleKHRMask) { + spv::IdImmediate imageOperand = { true, + builder.makeUintConstant(TranslateMemoryScope(TranslateCoherent(imageType))) }; + operands.push_back(imageOperand); + } + + if (builder.getImageTypeFormat(builder.getImageType(operands.front().word)) == spv::ImageFormatUnknown) + builder.addCapability(spv::CapabilityStorageImageReadWithoutFormat); + + std::vector<spv::Id> result(1, builder.createOp(spv::OpImageRead, resultType(), operands)); + builder.setPrecision(result[0], precision); + + // If needed, add a conversion constructor to the proper size. + if (components != node->getType().getVectorSize()) + result[0] = builder.createConstructor(precision, result, convertGlslangToSpvType(node->getType())); + + return result[0]; +#ifdef AMD_EXTENSIONS + } else if (node->getOp() == glslang::EOpImageStore || node->getOp() == glslang::EOpImageStoreLod) { +#else + } else if (node->getOp() == glslang::EOpImageStore) { +#endif + + // Push the texel value before the operands +#ifdef AMD_EXTENSIONS + if (sampler.ms || cracked.lod) { +#else + if (sampler.ms) { +#endif + spv::IdImmediate texel = { true, *(opIt + 1) }; + operands.push_back(texel); + } else { + spv::IdImmediate texel = { true, *opIt }; + operands.push_back(texel); + } + + spv::ImageOperandsMask mask = spv::ImageOperandsMaskNone; + if (sampler.ms) { + mask = mask | spv::ImageOperandsSampleMask; + } +#ifdef AMD_EXTENSIONS + if (cracked.lod) { + builder.addExtension(spv::E_SPV_AMD_shader_image_load_store_lod); + builder.addCapability(spv::CapabilityImageReadWriteLodAMD); + mask = mask | spv::ImageOperandsLodMask; + } +#endif + mask = mask | TranslateImageOperands(TranslateCoherent(imageType)); + mask = (spv::ImageOperandsMask)(mask & ~spv::ImageOperandsMakeTexelVisibleKHRMask); + mask = mask | signExtensionMask(); + if (mask != spv::ImageOperandsMaskNone) { + spv::IdImmediate imageOperands = { false, (unsigned int)mask }; + operands.push_back(imageOperands); + } + if (mask & spv::ImageOperandsSampleMask) { + spv::IdImmediate imageOperand = { true, *opIt++ }; + operands.push_back(imageOperand); + } +#ifdef AMD_EXTENSIONS + if (mask & spv::ImageOperandsLodMask) { + spv::IdImmediate imageOperand = { true, *opIt++ }; + operands.push_back(imageOperand); + } +#endif + if (mask & spv::ImageOperandsMakeTexelAvailableKHRMask) { + spv::IdImmediate imageOperand = { true, + builder.makeUintConstant(TranslateMemoryScope(TranslateCoherent(imageType))) }; + operands.push_back(imageOperand); + } + + builder.createNoResultOp(spv::OpImageWrite, operands); + if (builder.getImageTypeFormat(builder.getImageType(operands.front().word)) == spv::ImageFormatUnknown) + builder.addCapability(spv::CapabilityStorageImageWriteWithoutFormat); + return spv::NoResult; +#ifdef AMD_EXTENSIONS + } else if (node->getOp() == glslang::EOpSparseImageLoad || + node->getOp() == glslang::EOpSparseImageLoadLod) { +#else + } else if (node->getOp() == glslang::EOpSparseImageLoad) { +#endif + builder.addCapability(spv::CapabilitySparseResidency); + if (builder.getImageTypeFormat(builder.getImageType(operands.front().word)) == spv::ImageFormatUnknown) + builder.addCapability(spv::CapabilityStorageImageReadWithoutFormat); + + spv::ImageOperandsMask mask = spv::ImageOperandsMaskNone; + if (sampler.ms) { + mask = mask | spv::ImageOperandsSampleMask; + } +#ifdef AMD_EXTENSIONS + if (cracked.lod) { + builder.addExtension(spv::E_SPV_AMD_shader_image_load_store_lod); + builder.addCapability(spv::CapabilityImageReadWriteLodAMD); + + mask = mask | spv::ImageOperandsLodMask; + } +#endif + mask = mask | TranslateImageOperands(TranslateCoherent(imageType)); + mask = (spv::ImageOperandsMask)(mask & ~spv::ImageOperandsMakeTexelAvailableKHRMask); + mask = mask | signExtensionMask(); + if (mask != spv::ImageOperandsMaskNone) { + spv::IdImmediate imageOperands = { false, (unsigned int)mask }; + operands.push_back(imageOperands); + } + if (mask & spv::ImageOperandsSampleMask) { + spv::IdImmediate imageOperand = { true, *opIt++ }; + operands.push_back(imageOperand); + } +#ifdef AMD_EXTENSIONS + if (mask & spv::ImageOperandsLodMask) { + spv::IdImmediate imageOperand = { true, *opIt++ }; + operands.push_back(imageOperand); + } +#endif + if (mask & spv::ImageOperandsMakeTexelVisibleKHRMask) { + spv::IdImmediate imageOperand = { true, builder.makeUintConstant(TranslateMemoryScope(TranslateCoherent(imageType))) }; + operands.push_back(imageOperand); + } + + // Create the return type that was a special structure + spv::Id texelOut = *opIt; + spv::Id typeId0 = resultType(); + spv::Id typeId1 = builder.getDerefTypeId(texelOut); + spv::Id resultTypeId = builder.makeStructResultType(typeId0, typeId1); + + spv::Id resultId = builder.createOp(spv::OpImageSparseRead, resultTypeId, operands); + + // Decode the return type + builder.createStore(builder.createCompositeExtract(resultId, typeId1, 1), texelOut); + return builder.createCompositeExtract(resultId, typeId0, 0); + } else { + // Process image atomic operations + + // GLSL "IMAGE_PARAMS" will involve in constructing an image texel pointer and this pointer, + // as the first source operand, is required by SPIR-V atomic operations. + // For non-MS, the sample value should be 0 + spv::IdImmediate sample = { true, sampler.ms ? *(opIt++) : builder.makeUintConstant(0) }; + operands.push_back(sample); + + spv::Id resultTypeId; + // imageAtomicStore has a void return type so base the pointer type on + // the type of the value operand. + if (node->getOp() == glslang::EOpImageAtomicStore) { + resultTypeId = builder.makePointer(spv::StorageClassImage, builder.getTypeId(operands[2].word)); + } else { + resultTypeId = builder.makePointer(spv::StorageClassImage, resultType()); + } + spv::Id pointer = builder.createOp(spv::OpImageTexelPointer, resultTypeId, operands); + + std::vector<spv::Id> operands; + operands.push_back(pointer); + for (; opIt != arguments.end(); ++opIt) + operands.push_back(*opIt); + + return createAtomicOperation(node->getOp(), precision, resultType(), operands, node->getBasicType()); + } + } + +#ifdef AMD_EXTENSIONS + // Check for fragment mask functions other than queries + if (cracked.fragMask) { + assert(sampler.ms); + + auto opIt = arguments.begin(); + std::vector<spv::Id> operands; + + // Extract the image if necessary + if (builder.isSampledImage(params.sampler)) + params.sampler = builder.createUnaryOp(spv::OpImage, builder.getImageType(params.sampler), params.sampler); + + operands.push_back(params.sampler); + ++opIt; + + if (sampler.isSubpass()) { + // add on the (0,0) coordinate + spv::Id zero = builder.makeIntConstant(0); + std::vector<spv::Id> comps; + comps.push_back(zero); + comps.push_back(zero); + operands.push_back(builder.makeCompositeConstant(builder.makeVectorType(builder.makeIntType(32), 2), comps)); + } + + for (; opIt != arguments.end(); ++opIt) + operands.push_back(*opIt); + + spv::Op fragMaskOp = spv::OpNop; + if (node->getOp() == glslang::EOpFragmentMaskFetch) + fragMaskOp = spv::OpFragmentMaskFetchAMD; + else if (node->getOp() == glslang::EOpFragmentFetch) + fragMaskOp = spv::OpFragmentFetchAMD; + + builder.addExtension(spv::E_SPV_AMD_shader_fragment_mask); + builder.addCapability(spv::CapabilityFragmentMaskAMD); + return builder.createOp(fragMaskOp, resultType(), operands); + } +#endif + + // Check for texture functions other than queries + bool sparse = node->isSparseTexture(); +#ifdef NV_EXTENSIONS + bool imageFootprint = node->isImageFootprint(); +#endif + + bool cubeCompare = sampler.dim == glslang::EsdCube && sampler.arrayed && sampler.shadow; + + // check for bias argument + bool bias = false; +#ifdef AMD_EXTENSIONS + if (! cracked.lod && ! cracked.grad && ! cracked.fetch && ! cubeCompare) { +#else + if (! cracked.lod && ! cracked.gather && ! cracked.grad && ! cracked.fetch && ! cubeCompare) { +#endif + int nonBiasArgCount = 2; +#ifdef AMD_EXTENSIONS + if (cracked.gather) + ++nonBiasArgCount; // comp argument should be present when bias argument is present + + if (f16ShadowCompare) + ++nonBiasArgCount; +#endif + if (cracked.offset) + ++nonBiasArgCount; +#ifdef AMD_EXTENSIONS + else if (cracked.offsets) + ++nonBiasArgCount; +#endif + if (cracked.grad) + nonBiasArgCount += 2; + if (cracked.lodClamp) + ++nonBiasArgCount; + if (sparse) + ++nonBiasArgCount; +#ifdef NV_EXTENSIONS + if (imageFootprint) + //Following three extra arguments + // int granularity, bool coarse, out gl_TextureFootprint2DNV footprint + nonBiasArgCount += 3; +#endif + if ((int)arguments.size() > nonBiasArgCount) + bias = true; + } + + // See if the sampler param should really be just the SPV image part + if (cracked.fetch) { + // a fetch needs to have the image extracted first + if (builder.isSampledImage(params.sampler)) + params.sampler = builder.createUnaryOp(spv::OpImage, builder.getImageType(params.sampler), params.sampler); + } + +#ifdef AMD_EXTENSIONS + if (cracked.gather) { + const auto& sourceExtensions = glslangIntermediate->getRequestedExtensions(); + if (bias || cracked.lod || + sourceExtensions.find(glslang::E_GL_AMD_texture_gather_bias_lod) != sourceExtensions.end()) { + builder.addExtension(spv::E_SPV_AMD_texture_gather_bias_lod); + builder.addCapability(spv::CapabilityImageGatherBiasLodAMD); + } + } +#endif + + // set the rest of the arguments + + params.coords = arguments[1]; + int extraArgs = 0; + bool noImplicitLod = false; + + // sort out where Dref is coming from +#ifdef AMD_EXTENSIONS + if (cubeCompare || f16ShadowCompare) { +#else + if (cubeCompare) { +#endif + params.Dref = arguments[2]; + ++extraArgs; + } else if (sampler.shadow && cracked.gather) { + params.Dref = arguments[2]; + ++extraArgs; + } else if (sampler.shadow) { + std::vector<spv::Id> indexes; + int dRefComp; + if (cracked.proj) + dRefComp = 2; // "The resulting 3rd component of P in the shadow forms is used as Dref" + else + dRefComp = builder.getNumComponents(params.coords) - 1; + indexes.push_back(dRefComp); + params.Dref = builder.createCompositeExtract(params.coords, builder.getScalarTypeId(builder.getTypeId(params.coords)), indexes); + } + + // lod + if (cracked.lod) { + params.lod = arguments[2 + extraArgs]; + ++extraArgs; + } else if (glslangIntermediate->getStage() != EShLangFragment +#ifdef NV_EXTENSIONS + // NV_compute_shader_derivatives layout qualifiers allow for implicit LODs + && !(glslangIntermediate->getStage() == EShLangCompute && + (glslangIntermediate->getLayoutDerivativeModeNone() != glslang::LayoutDerivativeNone)) +#endif + ) { + // we need to invent the default lod for an explicit lod instruction for a non-fragment stage + noImplicitLod = true; + } + + // multisample + if (sampler.ms) { + params.sample = arguments[2 + extraArgs]; // For MS, "sample" should be specified + ++extraArgs; + } + + // gradient + if (cracked.grad) { + params.gradX = arguments[2 + extraArgs]; + params.gradY = arguments[3 + extraArgs]; + extraArgs += 2; + } + + // offset and offsets + if (cracked.offset) { + params.offset = arguments[2 + extraArgs]; + ++extraArgs; + } else if (cracked.offsets) { + params.offsets = arguments[2 + extraArgs]; + ++extraArgs; + } + + // lod clamp + if (cracked.lodClamp) { + params.lodClamp = arguments[2 + extraArgs]; + ++extraArgs; + } + // sparse + if (sparse) { + params.texelOut = arguments[2 + extraArgs]; + ++extraArgs; + } + + // gather component + if (cracked.gather && ! sampler.shadow) { + // default component is 0, if missing, otherwise an argument + if (2 + extraArgs < (int)arguments.size()) { + params.component = arguments[2 + extraArgs]; + ++extraArgs; + } else + params.component = builder.makeIntConstant(0); + } +#ifdef NV_EXTENSIONS + spv::Id resultStruct = spv::NoResult; + if (imageFootprint) { + //Following three extra arguments + // int granularity, bool coarse, out gl_TextureFootprint2DNV footprint + params.granularity = arguments[2 + extraArgs]; + params.coarse = arguments[3 + extraArgs]; + resultStruct = arguments[4 + extraArgs]; + extraArgs += 3; + } +#endif + // bias + if (bias) { + params.bias = arguments[2 + extraArgs]; + ++extraArgs; + } + +#ifdef NV_EXTENSIONS + if (imageFootprint) { + builder.addExtension(spv::E_SPV_NV_shader_image_footprint); + builder.addCapability(spv::CapabilityImageFootprintNV); + + + //resultStructType(OpenGL type) contains 5 elements: + //struct gl_TextureFootprint2DNV { + // uvec2 anchor; + // uvec2 offset; + // uvec2 mask; + // uint lod; + // uint granularity; + //}; + //or + //struct gl_TextureFootprint3DNV { + // uvec3 anchor; + // uvec3 offset; + // uvec2 mask; + // uint lod; + // uint granularity; + //}; + spv::Id resultStructType = builder.getContainedTypeId(builder.getTypeId(resultStruct)); + assert(builder.isStructType(resultStructType)); + + //resType (SPIR-V type) contains 6 elements: + //Member 0 must be a Boolean type scalar(LOD), + //Member 1 must be a vector of integer type, whose Signedness operand is 0(anchor), + //Member 2 must be a vector of integer type, whose Signedness operand is 0(offset), + //Member 3 must be a vector of integer type, whose Signedness operand is 0(mask), + //Member 4 must be a scalar of integer type, whose Signedness operand is 0(lod), + //Member 5 must be a scalar of integer type, whose Signedness operand is 0(granularity). + std::vector<spv::Id> members; + members.push_back(resultType()); + for (int i = 0; i < 5; i++) { + members.push_back(builder.getContainedTypeId(resultStructType, i)); + } + spv::Id resType = builder.makeStructType(members, "ResType"); + + //call ImageFootprintNV + spv::Id res = builder.createTextureCall(precision, resType, sparse, cracked.fetch, cracked.proj, + cracked.gather, noImplicitLod, params, signExtensionMask()); + + //copy resType (SPIR-V type) to resultStructType(OpenGL type) + for (int i = 0; i < 5; i++) { + builder.clearAccessChain(); + builder.setAccessChainLValue(resultStruct); + + //Accessing to a struct we created, no coherent flag is set + spv::Builder::AccessChain::CoherentFlags flags; + flags.clear(); + + builder.accessChainPush(builder.makeIntConstant(i), flags, 0); + builder.accessChainStore(builder.createCompositeExtract(res, builder.getContainedTypeId(resType, i+1), i+1)); + } + return builder.createCompositeExtract(res, resultType(), 0); + } +#endif + + // projective component (might not to move) + // GLSL: "The texture coordinates consumed from P, not including the last component of P, + // are divided by the last component of P." + // SPIR-V: "... (u [, v] [, w], q)... It may be a vector larger than needed, but all + // unused components will appear after all used components." + if (cracked.proj) { + int projSourceComp = builder.getNumComponents(params.coords) - 1; + int projTargetComp; + switch (sampler.dim) { + case glslang::Esd1D: projTargetComp = 1; break; + case glslang::Esd2D: projTargetComp = 2; break; + case glslang::EsdRect: projTargetComp = 2; break; + default: projTargetComp = projSourceComp; break; + } + // copy the projective coordinate if we have to + if (projTargetComp != projSourceComp) { + spv::Id projComp = builder.createCompositeExtract(params.coords, + builder.getScalarTypeId(builder.getTypeId(params.coords)), + projSourceComp); + params.coords = builder.createCompositeInsert(projComp, params.coords, + builder.getTypeId(params.coords), projTargetComp); + } + } + + // nonprivate + if (imageType.getQualifier().nonprivate) { + params.nonprivate = true; + } + + // volatile + if (imageType.getQualifier().volatil) { + params.volatil = true; + } + + std::vector<spv::Id> result( 1, + builder.createTextureCall(precision, resultType(), sparse, cracked.fetch, cracked.proj, cracked.gather, + noImplicitLod, params, signExtensionMask()) + ); + + if (components != node->getType().getVectorSize()) + result[0] = builder.createConstructor(precision, result, convertGlslangToSpvType(node->getType())); + + return result[0]; +} + +spv::Id TGlslangToSpvTraverser::handleUserFunctionCall(const glslang::TIntermAggregate* node) +{ + // Grab the function's pointer from the previously created function + spv::Function* function = functionMap[node->getName().c_str()]; + if (! function) + return 0; + + const glslang::TIntermSequence& glslangArgs = node->getSequence(); + const glslang::TQualifierList& qualifiers = node->getQualifierList(); + + // See comments in makeFunctions() for details about the semantics for parameter passing. + // + // These imply we need a four step process: + // 1. Evaluate the arguments + // 2. Allocate and make copies of in, out, and inout arguments + // 3. Make the call + // 4. Copy back the results + + // 1. Evaluate the arguments and their types + std::vector<spv::Builder::AccessChain> lValues; + std::vector<spv::Id> rValues; + std::vector<const glslang::TType*> argTypes; + for (int a = 0; a < (int)glslangArgs.size(); ++a) { + argTypes.push_back(&glslangArgs[a]->getAsTyped()->getType()); + // build l-value + builder.clearAccessChain(); + glslangArgs[a]->traverse(this); + // keep outputs and pass-by-originals as l-values, evaluate others as r-values + if (originalParam(qualifiers[a], *argTypes[a], function->hasImplicitThis() && a == 0) || + writableParam(qualifiers[a])) { + // save l-value + lValues.push_back(builder.getAccessChain()); + } else { + // process r-value + rValues.push_back(accessChainLoad(*argTypes.back())); + } + } + + // 2. Allocate space for anything needing a copy, and if it's "in" or "inout" + // copy the original into that space. + // + // Also, build up the list of actual arguments to pass in for the call + int lValueCount = 0; + int rValueCount = 0; + std::vector<spv::Id> spvArgs; + for (int a = 0; a < (int)glslangArgs.size(); ++a) { + spv::Id arg; + if (originalParam(qualifiers[a], *argTypes[a], function->hasImplicitThis() && a == 0)) { + builder.setAccessChain(lValues[lValueCount]); + arg = builder.accessChainGetLValue(); + ++lValueCount; + } else if (writableParam(qualifiers[a])) { + // need space to hold the copy + arg = builder.createVariable(spv::StorageClassFunction, builder.getContainedTypeId(function->getParamType(a)), "param"); + if (qualifiers[a] == glslang::EvqIn || qualifiers[a] == glslang::EvqInOut) { + // need to copy the input into output space + builder.setAccessChain(lValues[lValueCount]); + spv::Id copy = accessChainLoad(*argTypes[a]); + builder.clearAccessChain(); + builder.setAccessChainLValue(arg); + multiTypeStore(*argTypes[a], copy); + } + ++lValueCount; + } else { + // process r-value, which involves a copy for a type mismatch + if (function->getParamType(a) != convertGlslangToSpvType(*argTypes[a])) { + spv::Id argCopy = builder.createVariable(spv::StorageClassFunction, function->getParamType(a), "arg"); + builder.clearAccessChain(); + builder.setAccessChainLValue(argCopy); + multiTypeStore(*argTypes[a], rValues[rValueCount]); + arg = builder.createLoad(argCopy); + } else + arg = rValues[rValueCount]; + ++rValueCount; + } + spvArgs.push_back(arg); + } + + // 3. Make the call. + spv::Id result = builder.createFunctionCall(function, spvArgs); + builder.setPrecision(result, TranslatePrecisionDecoration(node->getType())); + + // 4. Copy back out an "out" arguments. + lValueCount = 0; + for (int a = 0; a < (int)glslangArgs.size(); ++a) { + if (originalParam(qualifiers[a], *argTypes[a], function->hasImplicitThis() && a == 0)) + ++lValueCount; + else if (writableParam(qualifiers[a])) { + if (qualifiers[a] == glslang::EvqOut || qualifiers[a] == glslang::EvqInOut) { + spv::Id copy = builder.createLoad(spvArgs[a]); + builder.setAccessChain(lValues[lValueCount]); + multiTypeStore(*argTypes[a], copy); + } + ++lValueCount; + } + } + + return result; +} + +// Translate AST operation to SPV operation, already having SPV-based operands/types. +spv::Id TGlslangToSpvTraverser::createBinaryOperation(glslang::TOperator op, OpDecorations& decorations, + spv::Id typeId, spv::Id left, spv::Id right, + glslang::TBasicType typeProxy, bool reduceComparison) +{ + bool isUnsigned = isTypeUnsignedInt(typeProxy); + bool isFloat = isTypeFloat(typeProxy); + bool isBool = typeProxy == glslang::EbtBool; + + spv::Op binOp = spv::OpNop; + bool needMatchingVectors = true; // for non-matrix ops, would a scalar need to smear to match a vector? + bool comparison = false; + + switch (op) { + case glslang::EOpAdd: + case glslang::EOpAddAssign: + if (isFloat) + binOp = spv::OpFAdd; + else + binOp = spv::OpIAdd; + break; + case glslang::EOpSub: + case glslang::EOpSubAssign: + if (isFloat) + binOp = spv::OpFSub; + else + binOp = spv::OpISub; + break; + case glslang::EOpMul: + case glslang::EOpMulAssign: + if (isFloat) + binOp = spv::OpFMul; + else + binOp = spv::OpIMul; + break; + case glslang::EOpVectorTimesScalar: + case glslang::EOpVectorTimesScalarAssign: + if (isFloat && (builder.isVector(left) || builder.isVector(right))) { + if (builder.isVector(right)) + std::swap(left, right); + assert(builder.isScalar(right)); + needMatchingVectors = false; + binOp = spv::OpVectorTimesScalar; + } else if (isFloat) + binOp = spv::OpFMul; + else + binOp = spv::OpIMul; + break; + case glslang::EOpVectorTimesMatrix: + case glslang::EOpVectorTimesMatrixAssign: + binOp = spv::OpVectorTimesMatrix; + break; + case glslang::EOpMatrixTimesVector: + binOp = spv::OpMatrixTimesVector; + break; + case glslang::EOpMatrixTimesScalar: + case glslang::EOpMatrixTimesScalarAssign: + binOp = spv::OpMatrixTimesScalar; + break; + case glslang::EOpMatrixTimesMatrix: + case glslang::EOpMatrixTimesMatrixAssign: + binOp = spv::OpMatrixTimesMatrix; + break; + case glslang::EOpOuterProduct: + binOp = spv::OpOuterProduct; + needMatchingVectors = false; + break; + + case glslang::EOpDiv: + case glslang::EOpDivAssign: + if (isFloat) + binOp = spv::OpFDiv; + else if (isUnsigned) + binOp = spv::OpUDiv; + else + binOp = spv::OpSDiv; + break; + case glslang::EOpMod: + case glslang::EOpModAssign: + if (isFloat) + binOp = spv::OpFMod; + else if (isUnsigned) + binOp = spv::OpUMod; + else + binOp = spv::OpSMod; + break; + case glslang::EOpRightShift: + case glslang::EOpRightShiftAssign: + if (isUnsigned) + binOp = spv::OpShiftRightLogical; + else + binOp = spv::OpShiftRightArithmetic; + break; + case glslang::EOpLeftShift: + case glslang::EOpLeftShiftAssign: + binOp = spv::OpShiftLeftLogical; + break; + case glslang::EOpAnd: + case glslang::EOpAndAssign: + binOp = spv::OpBitwiseAnd; + break; + case glslang::EOpLogicalAnd: + needMatchingVectors = false; + binOp = spv::OpLogicalAnd; + break; + case glslang::EOpInclusiveOr: + case glslang::EOpInclusiveOrAssign: + binOp = spv::OpBitwiseOr; + break; + case glslang::EOpLogicalOr: + needMatchingVectors = false; + binOp = spv::OpLogicalOr; + break; + case glslang::EOpExclusiveOr: + case glslang::EOpExclusiveOrAssign: + binOp = spv::OpBitwiseXor; + break; + case glslang::EOpLogicalXor: + needMatchingVectors = false; + binOp = spv::OpLogicalNotEqual; + break; + + case glslang::EOpLessThan: + case glslang::EOpGreaterThan: + case glslang::EOpLessThanEqual: + case glslang::EOpGreaterThanEqual: + case glslang::EOpEqual: + case glslang::EOpNotEqual: + case glslang::EOpVectorEqual: + case glslang::EOpVectorNotEqual: + comparison = true; + break; + default: + break; + } + + // handle mapped binary operations (should be non-comparison) + if (binOp != spv::OpNop) { + assert(comparison == false); + if (builder.isMatrix(left) || builder.isMatrix(right) || + builder.isCooperativeMatrix(left) || builder.isCooperativeMatrix(right)) + return createBinaryMatrixOperation(binOp, decorations, typeId, left, right); + + // No matrix involved; make both operands be the same number of components, if needed + if (needMatchingVectors) + builder.promoteScalar(decorations.precision, left, right); + + spv::Id result = builder.createBinOp(binOp, typeId, left, right); + builder.addDecoration(result, decorations.noContraction); + builder.addDecoration(result, decorations.nonUniform); + return builder.setPrecision(result, decorations.precision); + } + + if (! comparison) + return 0; + + // Handle comparison instructions + + if (reduceComparison && (op == glslang::EOpEqual || op == glslang::EOpNotEqual) + && (builder.isVector(left) || builder.isMatrix(left) || builder.isAggregate(left))) { + spv::Id result = builder.createCompositeCompare(decorations.precision, left, right, op == glslang::EOpEqual); + builder.addDecoration(result, decorations.nonUniform); + return result; + } + + switch (op) { + case glslang::EOpLessThan: + if (isFloat) + binOp = spv::OpFOrdLessThan; + else if (isUnsigned) + binOp = spv::OpULessThan; + else + binOp = spv::OpSLessThan; + break; + case glslang::EOpGreaterThan: + if (isFloat) + binOp = spv::OpFOrdGreaterThan; + else if (isUnsigned) + binOp = spv::OpUGreaterThan; + else + binOp = spv::OpSGreaterThan; + break; + case glslang::EOpLessThanEqual: + if (isFloat) + binOp = spv::OpFOrdLessThanEqual; + else if (isUnsigned) + binOp = spv::OpULessThanEqual; + else + binOp = spv::OpSLessThanEqual; + break; + case glslang::EOpGreaterThanEqual: + if (isFloat) + binOp = spv::OpFOrdGreaterThanEqual; + else if (isUnsigned) + binOp = spv::OpUGreaterThanEqual; + else + binOp = spv::OpSGreaterThanEqual; + break; + case glslang::EOpEqual: + case glslang::EOpVectorEqual: + if (isFloat) + binOp = spv::OpFOrdEqual; + else if (isBool) + binOp = spv::OpLogicalEqual; + else + binOp = spv::OpIEqual; + break; + case glslang::EOpNotEqual: + case glslang::EOpVectorNotEqual: + if (isFloat) + binOp = spv::OpFOrdNotEqual; + else if (isBool) + binOp = spv::OpLogicalNotEqual; + else + binOp = spv::OpINotEqual; + break; + default: + break; + } + + if (binOp != spv::OpNop) { + spv::Id result = builder.createBinOp(binOp, typeId, left, right); + builder.addDecoration(result, decorations.noContraction); + builder.addDecoration(result, decorations.nonUniform); + return builder.setPrecision(result, decorations.precision); + } + + return 0; +} + +// +// Translate AST matrix operation to SPV operation, already having SPV-based operands/types. +// These can be any of: +// +// matrix * scalar +// scalar * matrix +// matrix * matrix linear algebraic +// matrix * vector +// vector * matrix +// matrix * matrix componentwise +// matrix op matrix op in {+, -, /} +// matrix op scalar op in {+, -, /} +// scalar op matrix op in {+, -, /} +// +spv::Id TGlslangToSpvTraverser::createBinaryMatrixOperation(spv::Op op, OpDecorations& decorations, spv::Id typeId, + spv::Id left, spv::Id right) +{ + bool firstClass = true; + + // First, handle first-class matrix operations (* and matrix/scalar) + switch (op) { + case spv::OpFDiv: + if (builder.isMatrix(left) && builder.isScalar(right)) { + // turn matrix / scalar into a multiply... + spv::Id resultType = builder.getTypeId(right); + right = builder.createBinOp(spv::OpFDiv, resultType, builder.makeFpConstant(resultType, 1.0), right); + op = spv::OpMatrixTimesScalar; + } else + firstClass = false; + break; + case spv::OpMatrixTimesScalar: + if (builder.isMatrix(right) || builder.isCooperativeMatrix(right)) + std::swap(left, right); + assert(builder.isScalar(right)); + break; + case spv::OpVectorTimesMatrix: + assert(builder.isVector(left)); + assert(builder.isMatrix(right)); + break; + case spv::OpMatrixTimesVector: + assert(builder.isMatrix(left)); + assert(builder.isVector(right)); + break; + case spv::OpMatrixTimesMatrix: + assert(builder.isMatrix(left)); + assert(builder.isMatrix(right)); + break; + default: + firstClass = false; + break; + } + + if (builder.isCooperativeMatrix(left) || builder.isCooperativeMatrix(right)) + firstClass = true; + + if (firstClass) { + spv::Id result = builder.createBinOp(op, typeId, left, right); + builder.addDecoration(result, decorations.noContraction); + builder.addDecoration(result, decorations.nonUniform); + return builder.setPrecision(result, decorations.precision); + } + + // Handle component-wise +, -, *, %, and / for all combinations of type. + // The result type of all of them is the same type as the (a) matrix operand. + // The algorithm is to: + // - break the matrix(es) into vectors + // - smear any scalar to a vector + // - do vector operations + // - make a matrix out the vector results + switch (op) { + case spv::OpFAdd: + case spv::OpFSub: + case spv::OpFDiv: + case spv::OpFMod: + case spv::OpFMul: + { + // one time set up... + bool leftMat = builder.isMatrix(left); + bool rightMat = builder.isMatrix(right); + unsigned int numCols = leftMat ? builder.getNumColumns(left) : builder.getNumColumns(right); + int numRows = leftMat ? builder.getNumRows(left) : builder.getNumRows(right); + spv::Id scalarType = builder.getScalarTypeId(typeId); + spv::Id vecType = builder.makeVectorType(scalarType, numRows); + std::vector<spv::Id> results; + spv::Id smearVec = spv::NoResult; + if (builder.isScalar(left)) + smearVec = builder.smearScalar(decorations.precision, left, vecType); + else if (builder.isScalar(right)) + smearVec = builder.smearScalar(decorations.precision, right, vecType); + + // do each vector op + for (unsigned int c = 0; c < numCols; ++c) { + std::vector<unsigned int> indexes; + indexes.push_back(c); + spv::Id leftVec = leftMat ? builder.createCompositeExtract( left, vecType, indexes) : smearVec; + spv::Id rightVec = rightMat ? builder.createCompositeExtract(right, vecType, indexes) : smearVec; + spv::Id result = builder.createBinOp(op, vecType, leftVec, rightVec); + builder.addDecoration(result, decorations.noContraction); + builder.addDecoration(result, decorations.nonUniform); + results.push_back(builder.setPrecision(result, decorations.precision)); + } + + // put the pieces together + spv::Id result = builder.setPrecision(builder.createCompositeConstruct(typeId, results), decorations.precision); + builder.addDecoration(result, decorations.nonUniform); + return result; + } + default: + assert(0); + return spv::NoResult; + } +} + +spv::Id TGlslangToSpvTraverser::createUnaryOperation(glslang::TOperator op, OpDecorations& decorations, spv::Id typeId, + spv::Id operand, glslang::TBasicType typeProxy) +{ + spv::Op unaryOp = spv::OpNop; + int extBuiltins = -1; + int libCall = -1; + bool isUnsigned = isTypeUnsignedInt(typeProxy); + bool isFloat = isTypeFloat(typeProxy); + + switch (op) { + case glslang::EOpNegative: + if (isFloat) { + unaryOp = spv::OpFNegate; + if (builder.isMatrixType(typeId)) + return createUnaryMatrixOperation(unaryOp, decorations, typeId, operand, typeProxy); + } else + unaryOp = spv::OpSNegate; + break; + + case glslang::EOpLogicalNot: + case glslang::EOpVectorLogicalNot: + unaryOp = spv::OpLogicalNot; + break; + case glslang::EOpBitwiseNot: + unaryOp = spv::OpNot; + break; + + case glslang::EOpDeterminant: + libCall = spv::GLSLstd450Determinant; + break; + case glslang::EOpMatrixInverse: + libCall = spv::GLSLstd450MatrixInverse; + break; + case glslang::EOpTranspose: + unaryOp = spv::OpTranspose; + break; + + case glslang::EOpRadians: + libCall = spv::GLSLstd450Radians; + break; + case glslang::EOpDegrees: + libCall = spv::GLSLstd450Degrees; + break; + case glslang::EOpSin: + libCall = spv::GLSLstd450Sin; + break; + case glslang::EOpCos: + libCall = spv::GLSLstd450Cos; + break; + case glslang::EOpTan: + libCall = spv::GLSLstd450Tan; + break; + case glslang::EOpAcos: + libCall = spv::GLSLstd450Acos; + break; + case glslang::EOpAsin: + libCall = spv::GLSLstd450Asin; + break; + case glslang::EOpAtan: + libCall = spv::GLSLstd450Atan; + break; + + case glslang::EOpAcosh: + libCall = spv::GLSLstd450Acosh; + break; + case glslang::EOpAsinh: + libCall = spv::GLSLstd450Asinh; + break; + case glslang::EOpAtanh: + libCall = spv::GLSLstd450Atanh; + break; + case glslang::EOpTanh: + libCall = spv::GLSLstd450Tanh; + break; + case glslang::EOpCosh: + libCall = spv::GLSLstd450Cosh; + break; + case glslang::EOpSinh: + libCall = spv::GLSLstd450Sinh; + break; + + case glslang::EOpLength: + libCall = spv::GLSLstd450Length; + break; + case glslang::EOpNormalize: + libCall = spv::GLSLstd450Normalize; + break; + + case glslang::EOpExp: + libCall = spv::GLSLstd450Exp; + break; + case glslang::EOpLog: + libCall = spv::GLSLstd450Log; + break; + case glslang::EOpExp2: + libCall = spv::GLSLstd450Exp2; + break; + case glslang::EOpLog2: + libCall = spv::GLSLstd450Log2; + break; + case glslang::EOpSqrt: + libCall = spv::GLSLstd450Sqrt; + break; + case glslang::EOpInverseSqrt: + libCall = spv::GLSLstd450InverseSqrt; + break; + + case glslang::EOpFloor: + libCall = spv::GLSLstd450Floor; + break; + case glslang::EOpTrunc: + libCall = spv::GLSLstd450Trunc; + break; + case glslang::EOpRound: + libCall = spv::GLSLstd450Round; + break; + case glslang::EOpRoundEven: + libCall = spv::GLSLstd450RoundEven; + break; + case glslang::EOpCeil: + libCall = spv::GLSLstd450Ceil; + break; + case glslang::EOpFract: + libCall = spv::GLSLstd450Fract; + break; + + case glslang::EOpIsNan: + unaryOp = spv::OpIsNan; + break; + case glslang::EOpIsInf: + unaryOp = spv::OpIsInf; + break; + case glslang::EOpIsFinite: + unaryOp = spv::OpIsFinite; + break; + + case glslang::EOpFloatBitsToInt: + case glslang::EOpFloatBitsToUint: + case glslang::EOpIntBitsToFloat: + case glslang::EOpUintBitsToFloat: + case glslang::EOpDoubleBitsToInt64: + case glslang::EOpDoubleBitsToUint64: + case glslang::EOpInt64BitsToDouble: + case glslang::EOpUint64BitsToDouble: + case glslang::EOpFloat16BitsToInt16: + case glslang::EOpFloat16BitsToUint16: + case glslang::EOpInt16BitsToFloat16: + case glslang::EOpUint16BitsToFloat16: + unaryOp = spv::OpBitcast; + break; + + case glslang::EOpPackSnorm2x16: + libCall = spv::GLSLstd450PackSnorm2x16; + break; + case glslang::EOpUnpackSnorm2x16: + libCall = spv::GLSLstd450UnpackSnorm2x16; + break; + case glslang::EOpPackUnorm2x16: + libCall = spv::GLSLstd450PackUnorm2x16; + break; + case glslang::EOpUnpackUnorm2x16: + libCall = spv::GLSLstd450UnpackUnorm2x16; + break; + case glslang::EOpPackHalf2x16: + libCall = spv::GLSLstd450PackHalf2x16; + break; + case glslang::EOpUnpackHalf2x16: + libCall = spv::GLSLstd450UnpackHalf2x16; + break; + case glslang::EOpPackSnorm4x8: + libCall = spv::GLSLstd450PackSnorm4x8; + break; + case glslang::EOpUnpackSnorm4x8: + libCall = spv::GLSLstd450UnpackSnorm4x8; + break; + case glslang::EOpPackUnorm4x8: + libCall = spv::GLSLstd450PackUnorm4x8; + break; + case glslang::EOpUnpackUnorm4x8: + libCall = spv::GLSLstd450UnpackUnorm4x8; + break; + case glslang::EOpPackDouble2x32: + libCall = spv::GLSLstd450PackDouble2x32; + break; + case glslang::EOpUnpackDouble2x32: + libCall = spv::GLSLstd450UnpackDouble2x32; + break; + + case glslang::EOpPackInt2x32: + case glslang::EOpUnpackInt2x32: + case glslang::EOpPackUint2x32: + case glslang::EOpUnpackUint2x32: + case glslang::EOpPack16: + case glslang::EOpPack32: + case glslang::EOpPack64: + case glslang::EOpUnpack32: + case glslang::EOpUnpack16: + case glslang::EOpUnpack8: + case glslang::EOpPackInt2x16: + case glslang::EOpUnpackInt2x16: + case glslang::EOpPackUint2x16: + case glslang::EOpUnpackUint2x16: + case glslang::EOpPackInt4x16: + case glslang::EOpUnpackInt4x16: + case glslang::EOpPackUint4x16: + case glslang::EOpUnpackUint4x16: + case glslang::EOpPackFloat2x16: + case glslang::EOpUnpackFloat2x16: + unaryOp = spv::OpBitcast; + break; + + case glslang::EOpDPdx: + unaryOp = spv::OpDPdx; + break; + case glslang::EOpDPdy: + unaryOp = spv::OpDPdy; + break; + case glslang::EOpFwidth: + unaryOp = spv::OpFwidth; + break; + case glslang::EOpDPdxFine: + unaryOp = spv::OpDPdxFine; + break; + case glslang::EOpDPdyFine: + unaryOp = spv::OpDPdyFine; + break; + case glslang::EOpFwidthFine: + unaryOp = spv::OpFwidthFine; + break; + case glslang::EOpDPdxCoarse: + unaryOp = spv::OpDPdxCoarse; + break; + case glslang::EOpDPdyCoarse: + unaryOp = spv::OpDPdyCoarse; + break; + case glslang::EOpFwidthCoarse: + unaryOp = spv::OpFwidthCoarse; + break; + case glslang::EOpInterpolateAtCentroid: +#ifdef AMD_EXTENSIONS + if (typeProxy == glslang::EbtFloat16) + builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float); +#endif + libCall = spv::GLSLstd450InterpolateAtCentroid; + break; + case glslang::EOpAny: + unaryOp = spv::OpAny; + break; + case glslang::EOpAll: + unaryOp = spv::OpAll; + break; + + case glslang::EOpAbs: + if (isFloat) + libCall = spv::GLSLstd450FAbs; + else + libCall = spv::GLSLstd450SAbs; + break; + case glslang::EOpSign: + if (isFloat) + libCall = spv::GLSLstd450FSign; + else + libCall = spv::GLSLstd450SSign; + break; + + case glslang::EOpAtomicCounterIncrement: + case glslang::EOpAtomicCounterDecrement: + case glslang::EOpAtomicCounter: + { + // Handle all of the atomics in one place, in createAtomicOperation() + std::vector<spv::Id> operands; + operands.push_back(operand); + return createAtomicOperation(op, decorations.precision, typeId, operands, typeProxy); + } + + case glslang::EOpBitFieldReverse: + unaryOp = spv::OpBitReverse; + break; + case glslang::EOpBitCount: + unaryOp = spv::OpBitCount; + break; + case glslang::EOpFindLSB: + libCall = spv::GLSLstd450FindILsb; + break; + case glslang::EOpFindMSB: + if (isUnsigned) + libCall = spv::GLSLstd450FindUMsb; + else + libCall = spv::GLSLstd450FindSMsb; + break; + + case glslang::EOpBallot: + case glslang::EOpReadFirstInvocation: + case glslang::EOpAnyInvocation: + case glslang::EOpAllInvocations: + case glslang::EOpAllInvocationsEqual: +#ifdef AMD_EXTENSIONS + case glslang::EOpMinInvocations: + case glslang::EOpMaxInvocations: + case glslang::EOpAddInvocations: + case glslang::EOpMinInvocationsNonUniform: + case glslang::EOpMaxInvocationsNonUniform: + case glslang::EOpAddInvocationsNonUniform: + case glslang::EOpMinInvocationsInclusiveScan: + case glslang::EOpMaxInvocationsInclusiveScan: + case glslang::EOpAddInvocationsInclusiveScan: + case glslang::EOpMinInvocationsInclusiveScanNonUniform: + case glslang::EOpMaxInvocationsInclusiveScanNonUniform: + case glslang::EOpAddInvocationsInclusiveScanNonUniform: + case glslang::EOpMinInvocationsExclusiveScan: + case glslang::EOpMaxInvocationsExclusiveScan: + case glslang::EOpAddInvocationsExclusiveScan: + case glslang::EOpMinInvocationsExclusiveScanNonUniform: + case glslang::EOpMaxInvocationsExclusiveScanNonUniform: + case glslang::EOpAddInvocationsExclusiveScanNonUniform: +#endif + { + std::vector<spv::Id> operands; + operands.push_back(operand); + return createInvocationsOperation(op, typeId, operands, typeProxy); + } + case glslang::EOpSubgroupAll: + case glslang::EOpSubgroupAny: + case glslang::EOpSubgroupAllEqual: + case glslang::EOpSubgroupBroadcastFirst: + case glslang::EOpSubgroupBallot: + case glslang::EOpSubgroupInverseBallot: + case glslang::EOpSubgroupBallotBitCount: + case glslang::EOpSubgroupBallotInclusiveBitCount: + case glslang::EOpSubgroupBallotExclusiveBitCount: + case glslang::EOpSubgroupBallotFindLSB: + case glslang::EOpSubgroupBallotFindMSB: + case glslang::EOpSubgroupAdd: + case glslang::EOpSubgroupMul: + case glslang::EOpSubgroupMin: + case glslang::EOpSubgroupMax: + case glslang::EOpSubgroupAnd: + case glslang::EOpSubgroupOr: + case glslang::EOpSubgroupXor: + case glslang::EOpSubgroupInclusiveAdd: + case glslang::EOpSubgroupInclusiveMul: + case glslang::EOpSubgroupInclusiveMin: + case glslang::EOpSubgroupInclusiveMax: + case glslang::EOpSubgroupInclusiveAnd: + case glslang::EOpSubgroupInclusiveOr: + case glslang::EOpSubgroupInclusiveXor: + case glslang::EOpSubgroupExclusiveAdd: + case glslang::EOpSubgroupExclusiveMul: + case glslang::EOpSubgroupExclusiveMin: + case glslang::EOpSubgroupExclusiveMax: + case glslang::EOpSubgroupExclusiveAnd: + case glslang::EOpSubgroupExclusiveOr: + case glslang::EOpSubgroupExclusiveXor: + case glslang::EOpSubgroupQuadSwapHorizontal: + case glslang::EOpSubgroupQuadSwapVertical: + case glslang::EOpSubgroupQuadSwapDiagonal: { + std::vector<spv::Id> operands; + operands.push_back(operand); + return createSubgroupOperation(op, typeId, operands, typeProxy); + } +#ifdef AMD_EXTENSIONS + case glslang::EOpMbcnt: + extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_ballot); + libCall = spv::MbcntAMD; + break; + + case glslang::EOpCubeFaceIndex: + extBuiltins = getExtBuiltins(spv::E_SPV_AMD_gcn_shader); + libCall = spv::CubeFaceIndexAMD; + break; + + case glslang::EOpCubeFaceCoord: + extBuiltins = getExtBuiltins(spv::E_SPV_AMD_gcn_shader); + libCall = spv::CubeFaceCoordAMD; + break; +#endif +#ifdef NV_EXTENSIONS + case glslang::EOpSubgroupPartition: + unaryOp = spv::OpGroupNonUniformPartitionNV; + break; +#endif + case glslang::EOpConstructReference: + unaryOp = spv::OpBitcast; + break; + + case glslang::EOpCopyObject: + unaryOp = spv::OpCopyObject; + break; + + default: + return 0; + } + + spv::Id id; + if (libCall >= 0) { + std::vector<spv::Id> args; + args.push_back(operand); + id = builder.createBuiltinCall(typeId, extBuiltins >= 0 ? extBuiltins : stdBuiltins, libCall, args); + } else { + id = builder.createUnaryOp(unaryOp, typeId, operand); + } + + builder.addDecoration(id, decorations.noContraction); + builder.addDecoration(id, decorations.nonUniform); + return builder.setPrecision(id, decorations.precision); +} + +// Create a unary operation on a matrix +spv::Id TGlslangToSpvTraverser::createUnaryMatrixOperation(spv::Op op, OpDecorations& decorations, spv::Id typeId, + spv::Id operand, glslang::TBasicType /* typeProxy */) +{ + // Handle unary operations vector by vector. + // The result type is the same type as the original type. + // The algorithm is to: + // - break the matrix into vectors + // - apply the operation to each vector + // - make a matrix out the vector results + + // get the types sorted out + int numCols = builder.getNumColumns(operand); + int numRows = builder.getNumRows(operand); + spv::Id srcVecType = builder.makeVectorType(builder.getScalarTypeId(builder.getTypeId(operand)), numRows); + spv::Id destVecType = builder.makeVectorType(builder.getScalarTypeId(typeId), numRows); + std::vector<spv::Id> results; + + // do each vector op + for (int c = 0; c < numCols; ++c) { + std::vector<unsigned int> indexes; + indexes.push_back(c); + spv::Id srcVec = builder.createCompositeExtract(operand, srcVecType, indexes); + spv::Id destVec = builder.createUnaryOp(op, destVecType, srcVec); + builder.addDecoration(destVec, decorations.noContraction); + builder.addDecoration(destVec, decorations.nonUniform); + results.push_back(builder.setPrecision(destVec, decorations.precision)); + } + + // put the pieces together + spv::Id result = builder.setPrecision(builder.createCompositeConstruct(typeId, results), decorations.precision); + builder.addDecoration(result, decorations.nonUniform); + return result; +} + +// For converting integers where both the bitwidth and the signedness could +// change, but only do the width change here. The caller is still responsible +// for the signedness conversion. +spv::Id TGlslangToSpvTraverser::createIntWidthConversion(glslang::TOperator op, spv::Id operand, int vectorSize) +{ + // Get the result type width, based on the type to convert to. + int width = 32; + switch(op) { + case glslang::EOpConvInt16ToUint8: + case glslang::EOpConvIntToUint8: + case glslang::EOpConvInt64ToUint8: + case glslang::EOpConvUint16ToInt8: + case glslang::EOpConvUintToInt8: + case glslang::EOpConvUint64ToInt8: + width = 8; + break; + case glslang::EOpConvInt8ToUint16: + case glslang::EOpConvIntToUint16: + case glslang::EOpConvInt64ToUint16: + case glslang::EOpConvUint8ToInt16: + case glslang::EOpConvUintToInt16: + case glslang::EOpConvUint64ToInt16: + width = 16; + break; + case glslang::EOpConvInt8ToUint: + case glslang::EOpConvInt16ToUint: + case glslang::EOpConvInt64ToUint: + case glslang::EOpConvUint8ToInt: + case glslang::EOpConvUint16ToInt: + case glslang::EOpConvUint64ToInt: + width = 32; + break; + case glslang::EOpConvInt8ToUint64: + case glslang::EOpConvInt16ToUint64: + case glslang::EOpConvIntToUint64: + case glslang::EOpConvUint8ToInt64: + case glslang::EOpConvUint16ToInt64: + case glslang::EOpConvUintToInt64: + width = 64; + break; + + default: + assert(false && "Default missing"); + break; + } + + // Get the conversion operation and result type, + // based on the target width, but the source type. + spv::Id type = spv::NoType; + spv::Op convOp = spv::OpNop; + switch(op) { + case glslang::EOpConvInt8ToUint16: + case glslang::EOpConvInt8ToUint: + case glslang::EOpConvInt8ToUint64: + case glslang::EOpConvInt16ToUint8: + case glslang::EOpConvInt16ToUint: + case glslang::EOpConvInt16ToUint64: + case glslang::EOpConvIntToUint8: + case glslang::EOpConvIntToUint16: + case glslang::EOpConvIntToUint64: + case glslang::EOpConvInt64ToUint8: + case glslang::EOpConvInt64ToUint16: + case glslang::EOpConvInt64ToUint: + convOp = spv::OpSConvert; + type = builder.makeIntType(width); + break; + default: + convOp = spv::OpUConvert; + type = builder.makeUintType(width); + break; + } + + if (vectorSize > 0) + type = builder.makeVectorType(type, vectorSize); + + return builder.createUnaryOp(convOp, type, operand); +} + +spv::Id TGlslangToSpvTraverser::createConversion(glslang::TOperator op, OpDecorations& decorations, spv::Id destType, + spv::Id operand, glslang::TBasicType typeProxy) +{ + spv::Op convOp = spv::OpNop; + spv::Id zero = 0; + spv::Id one = 0; + + int vectorSize = builder.isVectorType(destType) ? builder.getNumTypeComponents(destType) : 0; + + switch (op) { + case glslang::EOpConvInt8ToBool: + case glslang::EOpConvUint8ToBool: + zero = builder.makeUint8Constant(0); + zero = makeSmearedConstant(zero, vectorSize); + return builder.createBinOp(spv::OpINotEqual, destType, operand, zero); + case glslang::EOpConvInt16ToBool: + case glslang::EOpConvUint16ToBool: + zero = builder.makeUint16Constant(0); + zero = makeSmearedConstant(zero, vectorSize); + return builder.createBinOp(spv::OpINotEqual, destType, operand, zero); + case glslang::EOpConvIntToBool: + case glslang::EOpConvUintToBool: + zero = builder.makeUintConstant(0); + zero = makeSmearedConstant(zero, vectorSize); + return builder.createBinOp(spv::OpINotEqual, destType, operand, zero); + case glslang::EOpConvInt64ToBool: + case glslang::EOpConvUint64ToBool: + zero = builder.makeUint64Constant(0); + zero = makeSmearedConstant(zero, vectorSize); + return builder.createBinOp(spv::OpINotEqual, destType, operand, zero); + + case glslang::EOpConvFloatToBool: + zero = builder.makeFloatConstant(0.0F); + zero = makeSmearedConstant(zero, vectorSize); + return builder.createBinOp(spv::OpFOrdNotEqual, destType, operand, zero); + + case glslang::EOpConvDoubleToBool: + zero = builder.makeDoubleConstant(0.0); + zero = makeSmearedConstant(zero, vectorSize); + return builder.createBinOp(spv::OpFOrdNotEqual, destType, operand, zero); + + case glslang::EOpConvFloat16ToBool: + zero = builder.makeFloat16Constant(0.0F); + zero = makeSmearedConstant(zero, vectorSize); + return builder.createBinOp(spv::OpFOrdNotEqual, destType, operand, zero); + + case glslang::EOpConvBoolToFloat: + convOp = spv::OpSelect; + zero = builder.makeFloatConstant(0.0F); + one = builder.makeFloatConstant(1.0F); + break; + + case glslang::EOpConvBoolToDouble: + convOp = spv::OpSelect; + zero = builder.makeDoubleConstant(0.0); + one = builder.makeDoubleConstant(1.0); + break; + + case glslang::EOpConvBoolToFloat16: + convOp = spv::OpSelect; + zero = builder.makeFloat16Constant(0.0F); + one = builder.makeFloat16Constant(1.0F); + break; + + case glslang::EOpConvBoolToInt8: + zero = builder.makeInt8Constant(0); + one = builder.makeInt8Constant(1); + convOp = spv::OpSelect; + break; + + case glslang::EOpConvBoolToUint8: + zero = builder.makeUint8Constant(0); + one = builder.makeUint8Constant(1); + convOp = spv::OpSelect; + break; + + case glslang::EOpConvBoolToInt16: + zero = builder.makeInt16Constant(0); + one = builder.makeInt16Constant(1); + convOp = spv::OpSelect; + break; + + case glslang::EOpConvBoolToUint16: + zero = builder.makeUint16Constant(0); + one = builder.makeUint16Constant(1); + convOp = spv::OpSelect; + break; + + case glslang::EOpConvBoolToInt: + case glslang::EOpConvBoolToInt64: + if (op == glslang::EOpConvBoolToInt64) + zero = builder.makeInt64Constant(0); + else + zero = builder.makeIntConstant(0); + + if (op == glslang::EOpConvBoolToInt64) + one = builder.makeInt64Constant(1); + else + one = builder.makeIntConstant(1); + + convOp = spv::OpSelect; + break; + + case glslang::EOpConvBoolToUint: + case glslang::EOpConvBoolToUint64: + if (op == glslang::EOpConvBoolToUint64) + zero = builder.makeUint64Constant(0); + else + zero = builder.makeUintConstant(0); + + if (op == glslang::EOpConvBoolToUint64) + one = builder.makeUint64Constant(1); + else + one = builder.makeUintConstant(1); + + convOp = spv::OpSelect; + break; + + case glslang::EOpConvInt8ToFloat16: + case glslang::EOpConvInt8ToFloat: + case glslang::EOpConvInt8ToDouble: + case glslang::EOpConvInt16ToFloat16: + case glslang::EOpConvInt16ToFloat: + case glslang::EOpConvInt16ToDouble: + case glslang::EOpConvIntToFloat16: + case glslang::EOpConvIntToFloat: + case glslang::EOpConvIntToDouble: + case glslang::EOpConvInt64ToFloat: + case glslang::EOpConvInt64ToDouble: + case glslang::EOpConvInt64ToFloat16: + convOp = spv::OpConvertSToF; + break; + + case glslang::EOpConvUint8ToFloat16: + case glslang::EOpConvUint8ToFloat: + case glslang::EOpConvUint8ToDouble: + case glslang::EOpConvUint16ToFloat16: + case glslang::EOpConvUint16ToFloat: + case glslang::EOpConvUint16ToDouble: + case glslang::EOpConvUintToFloat16: + case glslang::EOpConvUintToFloat: + case glslang::EOpConvUintToDouble: + case glslang::EOpConvUint64ToFloat: + case glslang::EOpConvUint64ToDouble: + case glslang::EOpConvUint64ToFloat16: + convOp = spv::OpConvertUToF; + break; + + case glslang::EOpConvDoubleToFloat: + case glslang::EOpConvFloatToDouble: + case glslang::EOpConvDoubleToFloat16: + case glslang::EOpConvFloat16ToDouble: + case glslang::EOpConvFloatToFloat16: + case glslang::EOpConvFloat16ToFloat: + convOp = spv::OpFConvert; + if (builder.isMatrixType(destType)) + return createUnaryMatrixOperation(convOp, decorations, destType, operand, typeProxy); + break; + + case glslang::EOpConvFloat16ToInt8: + case glslang::EOpConvFloatToInt8: + case glslang::EOpConvDoubleToInt8: + case glslang::EOpConvFloat16ToInt16: + case glslang::EOpConvFloatToInt16: + case glslang::EOpConvDoubleToInt16: + case glslang::EOpConvFloat16ToInt: + case glslang::EOpConvFloatToInt: + case glslang::EOpConvDoubleToInt: + case glslang::EOpConvFloat16ToInt64: + case glslang::EOpConvFloatToInt64: + case glslang::EOpConvDoubleToInt64: + convOp = spv::OpConvertFToS; + break; + + case glslang::EOpConvUint8ToInt8: + case glslang::EOpConvInt8ToUint8: + case glslang::EOpConvUint16ToInt16: + case glslang::EOpConvInt16ToUint16: + case glslang::EOpConvUintToInt: + case glslang::EOpConvIntToUint: + case glslang::EOpConvUint64ToInt64: + case glslang::EOpConvInt64ToUint64: + if (builder.isInSpecConstCodeGenMode()) { + // Build zero scalar or vector for OpIAdd. + if(op == glslang::EOpConvUint8ToInt8 || op == glslang::EOpConvInt8ToUint8) { + zero = builder.makeUint8Constant(0); + } else if (op == glslang::EOpConvUint16ToInt16 || op == glslang::EOpConvInt16ToUint16) { + zero = builder.makeUint16Constant(0); + } else if (op == glslang::EOpConvUint64ToInt64 || op == glslang::EOpConvInt64ToUint64) { + zero = builder.makeUint64Constant(0); + } else { + zero = builder.makeUintConstant(0); + } + zero = makeSmearedConstant(zero, vectorSize); + // Use OpIAdd, instead of OpBitcast to do the conversion when + // generating for OpSpecConstantOp instruction. + return builder.createBinOp(spv::OpIAdd, destType, operand, zero); + } + // For normal run-time conversion instruction, use OpBitcast. + convOp = spv::OpBitcast; + break; + + case glslang::EOpConvFloat16ToUint8: + case glslang::EOpConvFloatToUint8: + case glslang::EOpConvDoubleToUint8: + case glslang::EOpConvFloat16ToUint16: + case glslang::EOpConvFloatToUint16: + case glslang::EOpConvDoubleToUint16: + case glslang::EOpConvFloat16ToUint: + case glslang::EOpConvFloatToUint: + case glslang::EOpConvDoubleToUint: + case glslang::EOpConvFloatToUint64: + case glslang::EOpConvDoubleToUint64: + case glslang::EOpConvFloat16ToUint64: + convOp = spv::OpConvertFToU; + break; + + case glslang::EOpConvInt8ToInt16: + case glslang::EOpConvInt8ToInt: + case glslang::EOpConvInt8ToInt64: + case glslang::EOpConvInt16ToInt8: + case glslang::EOpConvInt16ToInt: + case glslang::EOpConvInt16ToInt64: + case glslang::EOpConvIntToInt8: + case glslang::EOpConvIntToInt16: + case glslang::EOpConvIntToInt64: + case glslang::EOpConvInt64ToInt8: + case glslang::EOpConvInt64ToInt16: + case glslang::EOpConvInt64ToInt: + convOp = spv::OpSConvert; + break; + + case glslang::EOpConvUint8ToUint16: + case glslang::EOpConvUint8ToUint: + case glslang::EOpConvUint8ToUint64: + case glslang::EOpConvUint16ToUint8: + case glslang::EOpConvUint16ToUint: + case glslang::EOpConvUint16ToUint64: + case glslang::EOpConvUintToUint8: + case glslang::EOpConvUintToUint16: + case glslang::EOpConvUintToUint64: + case glslang::EOpConvUint64ToUint8: + case glslang::EOpConvUint64ToUint16: + case glslang::EOpConvUint64ToUint: + convOp = spv::OpUConvert; + break; + + case glslang::EOpConvInt8ToUint16: + case glslang::EOpConvInt8ToUint: + case glslang::EOpConvInt8ToUint64: + case glslang::EOpConvInt16ToUint8: + case glslang::EOpConvInt16ToUint: + case glslang::EOpConvInt16ToUint64: + case glslang::EOpConvIntToUint8: + case glslang::EOpConvIntToUint16: + case glslang::EOpConvIntToUint64: + case glslang::EOpConvInt64ToUint8: + case glslang::EOpConvInt64ToUint16: + case glslang::EOpConvInt64ToUint: + case glslang::EOpConvUint8ToInt16: + case glslang::EOpConvUint8ToInt: + case glslang::EOpConvUint8ToInt64: + case glslang::EOpConvUint16ToInt8: + case glslang::EOpConvUint16ToInt: + case glslang::EOpConvUint16ToInt64: + case glslang::EOpConvUintToInt8: + case glslang::EOpConvUintToInt16: + case glslang::EOpConvUintToInt64: + case glslang::EOpConvUint64ToInt8: + case glslang::EOpConvUint64ToInt16: + case glslang::EOpConvUint64ToInt: + // OpSConvert/OpUConvert + OpBitCast + operand = createIntWidthConversion(op, operand, vectorSize); + + if (builder.isInSpecConstCodeGenMode()) { + // Build zero scalar or vector for OpIAdd. + switch(op) { + case glslang::EOpConvInt16ToUint8: + case glslang::EOpConvIntToUint8: + case glslang::EOpConvInt64ToUint8: + case glslang::EOpConvUint16ToInt8: + case glslang::EOpConvUintToInt8: + case glslang::EOpConvUint64ToInt8: + zero = builder.makeUint8Constant(0); + break; + case glslang::EOpConvInt8ToUint16: + case glslang::EOpConvIntToUint16: + case glslang::EOpConvInt64ToUint16: + case glslang::EOpConvUint8ToInt16: + case glslang::EOpConvUintToInt16: + case glslang::EOpConvUint64ToInt16: + zero = builder.makeUint16Constant(0); + break; + case glslang::EOpConvInt8ToUint: + case glslang::EOpConvInt16ToUint: + case glslang::EOpConvInt64ToUint: + case glslang::EOpConvUint8ToInt: + case glslang::EOpConvUint16ToInt: + case glslang::EOpConvUint64ToInt: + zero = builder.makeUintConstant(0); + break; + case glslang::EOpConvInt8ToUint64: + case glslang::EOpConvInt16ToUint64: + case glslang::EOpConvIntToUint64: + case glslang::EOpConvUint8ToInt64: + case glslang::EOpConvUint16ToInt64: + case glslang::EOpConvUintToInt64: + zero = builder.makeUint64Constant(0); + break; + default: + assert(false && "Default missing"); + break; + } + zero = makeSmearedConstant(zero, vectorSize); + // Use OpIAdd, instead of OpBitcast to do the conversion when + // generating for OpSpecConstantOp instruction. + return builder.createBinOp(spv::OpIAdd, destType, operand, zero); + } + // For normal run-time conversion instruction, use OpBitcast. + convOp = spv::OpBitcast; + break; + case glslang::EOpConvUint64ToPtr: + convOp = spv::OpConvertUToPtr; + break; + case glslang::EOpConvPtrToUint64: + convOp = spv::OpConvertPtrToU; + break; + default: + break; + } + + spv::Id result = 0; + if (convOp == spv::OpNop) + return result; + + if (convOp == spv::OpSelect) { + zero = makeSmearedConstant(zero, vectorSize); + one = makeSmearedConstant(one, vectorSize); + result = builder.createTriOp(convOp, destType, operand, one, zero); + } else + result = builder.createUnaryOp(convOp, destType, operand); + + result = builder.setPrecision(result, decorations.precision); + builder.addDecoration(result, decorations.nonUniform); + return result; +} + +spv::Id TGlslangToSpvTraverser::makeSmearedConstant(spv::Id constant, int vectorSize) +{ + if (vectorSize == 0) + return constant; + + spv::Id vectorTypeId = builder.makeVectorType(builder.getTypeId(constant), vectorSize); + std::vector<spv::Id> components; + for (int c = 0; c < vectorSize; ++c) + components.push_back(constant); + return builder.makeCompositeConstant(vectorTypeId, components); +} + +// For glslang ops that map to SPV atomic opCodes +spv::Id TGlslangToSpvTraverser::createAtomicOperation(glslang::TOperator op, spv::Decoration /*precision*/, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy) +{ + spv::Op opCode = spv::OpNop; + + switch (op) { + case glslang::EOpAtomicAdd: + case glslang::EOpImageAtomicAdd: + case glslang::EOpAtomicCounterAdd: + opCode = spv::OpAtomicIAdd; + break; + case glslang::EOpAtomicCounterSubtract: + opCode = spv::OpAtomicISub; + break; + case glslang::EOpAtomicMin: + case glslang::EOpImageAtomicMin: + case glslang::EOpAtomicCounterMin: + opCode = (typeProxy == glslang::EbtUint || typeProxy == glslang::EbtUint64) ? spv::OpAtomicUMin : spv::OpAtomicSMin; + break; + case glslang::EOpAtomicMax: + case glslang::EOpImageAtomicMax: + case glslang::EOpAtomicCounterMax: + opCode = (typeProxy == glslang::EbtUint || typeProxy == glslang::EbtUint64) ? spv::OpAtomicUMax : spv::OpAtomicSMax; + break; + case glslang::EOpAtomicAnd: + case glslang::EOpImageAtomicAnd: + case glslang::EOpAtomicCounterAnd: + opCode = spv::OpAtomicAnd; + break; + case glslang::EOpAtomicOr: + case glslang::EOpImageAtomicOr: + case glslang::EOpAtomicCounterOr: + opCode = spv::OpAtomicOr; + break; + case glslang::EOpAtomicXor: + case glslang::EOpImageAtomicXor: + case glslang::EOpAtomicCounterXor: + opCode = spv::OpAtomicXor; + break; + case glslang::EOpAtomicExchange: + case glslang::EOpImageAtomicExchange: + case glslang::EOpAtomicCounterExchange: + opCode = spv::OpAtomicExchange; + break; + case glslang::EOpAtomicCompSwap: + case glslang::EOpImageAtomicCompSwap: + case glslang::EOpAtomicCounterCompSwap: + opCode = spv::OpAtomicCompareExchange; + break; + case glslang::EOpAtomicCounterIncrement: + opCode = spv::OpAtomicIIncrement; + break; + case glslang::EOpAtomicCounterDecrement: + opCode = spv::OpAtomicIDecrement; + break; + case glslang::EOpAtomicCounter: + case glslang::EOpImageAtomicLoad: + case glslang::EOpAtomicLoad: + opCode = spv::OpAtomicLoad; + break; + case glslang::EOpAtomicStore: + case glslang::EOpImageAtomicStore: + opCode = spv::OpAtomicStore; + break; + default: + assert(0); + break; + } + + if (typeProxy == glslang::EbtInt64 || typeProxy == glslang::EbtUint64) + builder.addCapability(spv::CapabilityInt64Atomics); + + // Sort out the operands + // - mapping from glslang -> SPV + // - there are extra SPV operands that are optional in glslang + // - compare-exchange swaps the value and comparator + // - compare-exchange has an extra memory semantics + // - EOpAtomicCounterDecrement needs a post decrement + spv::Id pointerId = 0, compareId = 0, valueId = 0; + // scope defaults to Device in the old model, QueueFamilyKHR in the new model + spv::Id scopeId; + if (glslangIntermediate->usingVulkanMemoryModel()) { + scopeId = builder.makeUintConstant(spv::ScopeQueueFamilyKHR); + } else { + scopeId = builder.makeUintConstant(spv::ScopeDevice); + } + // semantics default to relaxed + spv::Id semanticsId = builder.makeUintConstant(spv::MemorySemanticsMaskNone); + spv::Id semanticsId2 = semanticsId; + + pointerId = operands[0]; + if (opCode == spv::OpAtomicIIncrement || opCode == spv::OpAtomicIDecrement) { + // no additional operands + } else if (opCode == spv::OpAtomicCompareExchange) { + compareId = operands[1]; + valueId = operands[2]; + if (operands.size() > 3) { + scopeId = operands[3]; + semanticsId = builder.makeUintConstant(builder.getConstantScalar(operands[4]) | builder.getConstantScalar(operands[5])); + semanticsId2 = builder.makeUintConstant(builder.getConstantScalar(operands[6]) | builder.getConstantScalar(operands[7])); + } + } else if (opCode == spv::OpAtomicLoad) { + if (operands.size() > 1) { + scopeId = operands[1]; + semanticsId = builder.makeUintConstant(builder.getConstantScalar(operands[2]) | builder.getConstantScalar(operands[3])); + } + } else { + // atomic store or RMW + valueId = operands[1]; + if (operands.size() > 2) { + scopeId = operands[2]; + semanticsId = builder.makeUintConstant(builder.getConstantScalar(operands[3]) | builder.getConstantScalar(operands[4])); + } + } + + // Check for capabilities + unsigned semanticsImmediate = builder.getConstantScalar(semanticsId) | builder.getConstantScalar(semanticsId2); + if (semanticsImmediate & (spv::MemorySemanticsMakeAvailableKHRMask | spv::MemorySemanticsMakeVisibleKHRMask | spv::MemorySemanticsOutputMemoryKHRMask)) { + builder.addCapability(spv::CapabilityVulkanMemoryModelKHR); + } + + if (glslangIntermediate->usingVulkanMemoryModel() && builder.getConstantScalar(scopeId) == spv::ScopeDevice) { + builder.addCapability(spv::CapabilityVulkanMemoryModelDeviceScopeKHR); + } + + std::vector<spv::Id> spvAtomicOperands; // hold the spv operands + spvAtomicOperands.push_back(pointerId); + spvAtomicOperands.push_back(scopeId); + spvAtomicOperands.push_back(semanticsId); + if (opCode == spv::OpAtomicCompareExchange) { + spvAtomicOperands.push_back(semanticsId2); + spvAtomicOperands.push_back(valueId); + spvAtomicOperands.push_back(compareId); + } else if (opCode != spv::OpAtomicLoad && opCode != spv::OpAtomicIIncrement && opCode != spv::OpAtomicIDecrement) { + spvAtomicOperands.push_back(valueId); + } + + if (opCode == spv::OpAtomicStore) { + builder.createNoResultOp(opCode, spvAtomicOperands); + return 0; + } else { + spv::Id resultId = builder.createOp(opCode, typeId, spvAtomicOperands); + + // GLSL and HLSL atomic-counter decrement return post-decrement value, + // while SPIR-V returns pre-decrement value. Translate between these semantics. + if (op == glslang::EOpAtomicCounterDecrement) + resultId = builder.createBinOp(spv::OpISub, typeId, resultId, builder.makeIntConstant(1)); + + return resultId; + } +} + +// Create group invocation operations. +spv::Id TGlslangToSpvTraverser::createInvocationsOperation(glslang::TOperator op, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy) +{ +#ifdef AMD_EXTENSIONS + bool isUnsigned = isTypeUnsignedInt(typeProxy); + bool isFloat = isTypeFloat(typeProxy); +#endif + + spv::Op opCode = spv::OpNop; + std::vector<spv::IdImmediate> spvGroupOperands; + spv::GroupOperation groupOperation = spv::GroupOperationMax; + + if (op == glslang::EOpBallot || op == glslang::EOpReadFirstInvocation || + op == glslang::EOpReadInvocation) { + builder.addExtension(spv::E_SPV_KHR_shader_ballot); + builder.addCapability(spv::CapabilitySubgroupBallotKHR); + } else if (op == glslang::EOpAnyInvocation || + op == glslang::EOpAllInvocations || + op == glslang::EOpAllInvocationsEqual) { + builder.addExtension(spv::E_SPV_KHR_subgroup_vote); + builder.addCapability(spv::CapabilitySubgroupVoteKHR); + } else { + builder.addCapability(spv::CapabilityGroups); +#ifdef AMD_EXTENSIONS + if (op == glslang::EOpMinInvocationsNonUniform || + op == glslang::EOpMaxInvocationsNonUniform || + op == glslang::EOpAddInvocationsNonUniform || + op == glslang::EOpMinInvocationsInclusiveScanNonUniform || + op == glslang::EOpMaxInvocationsInclusiveScanNonUniform || + op == glslang::EOpAddInvocationsInclusiveScanNonUniform || + op == glslang::EOpMinInvocationsExclusiveScanNonUniform || + op == glslang::EOpMaxInvocationsExclusiveScanNonUniform || + op == glslang::EOpAddInvocationsExclusiveScanNonUniform) + builder.addExtension(spv::E_SPV_AMD_shader_ballot); +#endif + +#ifdef AMD_EXTENSIONS + switch (op) { + case glslang::EOpMinInvocations: + case glslang::EOpMaxInvocations: + case glslang::EOpAddInvocations: + case glslang::EOpMinInvocationsNonUniform: + case glslang::EOpMaxInvocationsNonUniform: + case glslang::EOpAddInvocationsNonUniform: + groupOperation = spv::GroupOperationReduce; + break; + case glslang::EOpMinInvocationsInclusiveScan: + case glslang::EOpMaxInvocationsInclusiveScan: + case glslang::EOpAddInvocationsInclusiveScan: + case glslang::EOpMinInvocationsInclusiveScanNonUniform: + case glslang::EOpMaxInvocationsInclusiveScanNonUniform: + case glslang::EOpAddInvocationsInclusiveScanNonUniform: + groupOperation = spv::GroupOperationInclusiveScan; + break; + case glslang::EOpMinInvocationsExclusiveScan: + case glslang::EOpMaxInvocationsExclusiveScan: + case glslang::EOpAddInvocationsExclusiveScan: + case glslang::EOpMinInvocationsExclusiveScanNonUniform: + case glslang::EOpMaxInvocationsExclusiveScanNonUniform: + case glslang::EOpAddInvocationsExclusiveScanNonUniform: + groupOperation = spv::GroupOperationExclusiveScan; + break; + default: + break; + } + spv::IdImmediate scope = { true, builder.makeUintConstant(spv::ScopeSubgroup) }; + spvGroupOperands.push_back(scope); + if (groupOperation != spv::GroupOperationMax) { + spv::IdImmediate groupOp = { false, (unsigned)groupOperation }; + spvGroupOperands.push_back(groupOp); + } +#endif + } + + for (auto opIt = operands.begin(); opIt != operands.end(); ++opIt) { + spv::IdImmediate op = { true, *opIt }; + spvGroupOperands.push_back(op); + } + + switch (op) { + case glslang::EOpAnyInvocation: + opCode = spv::OpSubgroupAnyKHR; + break; + case glslang::EOpAllInvocations: + opCode = spv::OpSubgroupAllKHR; + break; + case glslang::EOpAllInvocationsEqual: + opCode = spv::OpSubgroupAllEqualKHR; + break; + case glslang::EOpReadInvocation: + opCode = spv::OpSubgroupReadInvocationKHR; + if (builder.isVectorType(typeId)) + return CreateInvocationsVectorOperation(opCode, groupOperation, typeId, operands); + break; + case glslang::EOpReadFirstInvocation: + opCode = spv::OpSubgroupFirstInvocationKHR; + break; + case glslang::EOpBallot: + { + // NOTE: According to the spec, the result type of "OpSubgroupBallotKHR" must be a 4 component vector of 32 + // bit integer types. The GLSL built-in function "ballotARB()" assumes the maximum number of invocations in + // a subgroup is 64. Thus, we have to convert uvec4.xy to uint64_t as follow: + // + // result = Bitcast(SubgroupBallotKHR(Predicate).xy) + // + spv::Id uintType = builder.makeUintType(32); + spv::Id uvec4Type = builder.makeVectorType(uintType, 4); + spv::Id result = builder.createOp(spv::OpSubgroupBallotKHR, uvec4Type, spvGroupOperands); + + std::vector<spv::Id> components; + components.push_back(builder.createCompositeExtract(result, uintType, 0)); + components.push_back(builder.createCompositeExtract(result, uintType, 1)); + + spv::Id uvec2Type = builder.makeVectorType(uintType, 2); + return builder.createUnaryOp(spv::OpBitcast, typeId, + builder.createCompositeConstruct(uvec2Type, components)); + } + +#ifdef AMD_EXTENSIONS + case glslang::EOpMinInvocations: + case glslang::EOpMaxInvocations: + case glslang::EOpAddInvocations: + case glslang::EOpMinInvocationsInclusiveScan: + case glslang::EOpMaxInvocationsInclusiveScan: + case glslang::EOpAddInvocationsInclusiveScan: + case glslang::EOpMinInvocationsExclusiveScan: + case glslang::EOpMaxInvocationsExclusiveScan: + case glslang::EOpAddInvocationsExclusiveScan: + if (op == glslang::EOpMinInvocations || + op == glslang::EOpMinInvocationsInclusiveScan || + op == glslang::EOpMinInvocationsExclusiveScan) { + if (isFloat) + opCode = spv::OpGroupFMin; + else { + if (isUnsigned) + opCode = spv::OpGroupUMin; + else + opCode = spv::OpGroupSMin; + } + } else if (op == glslang::EOpMaxInvocations || + op == glslang::EOpMaxInvocationsInclusiveScan || + op == glslang::EOpMaxInvocationsExclusiveScan) { + if (isFloat) + opCode = spv::OpGroupFMax; + else { + if (isUnsigned) + opCode = spv::OpGroupUMax; + else + opCode = spv::OpGroupSMax; + } + } else { + if (isFloat) + opCode = spv::OpGroupFAdd; + else + opCode = spv::OpGroupIAdd; + } + + if (builder.isVectorType(typeId)) + return CreateInvocationsVectorOperation(opCode, groupOperation, typeId, operands); + + break; + case glslang::EOpMinInvocationsNonUniform: + case glslang::EOpMaxInvocationsNonUniform: + case glslang::EOpAddInvocationsNonUniform: + case glslang::EOpMinInvocationsInclusiveScanNonUniform: + case glslang::EOpMaxInvocationsInclusiveScanNonUniform: + case glslang::EOpAddInvocationsInclusiveScanNonUniform: + case glslang::EOpMinInvocationsExclusiveScanNonUniform: + case glslang::EOpMaxInvocationsExclusiveScanNonUniform: + case glslang::EOpAddInvocationsExclusiveScanNonUniform: + if (op == glslang::EOpMinInvocationsNonUniform || + op == glslang::EOpMinInvocationsInclusiveScanNonUniform || + op == glslang::EOpMinInvocationsExclusiveScanNonUniform) { + if (isFloat) + opCode = spv::OpGroupFMinNonUniformAMD; + else { + if (isUnsigned) + opCode = spv::OpGroupUMinNonUniformAMD; + else + opCode = spv::OpGroupSMinNonUniformAMD; + } + } + else if (op == glslang::EOpMaxInvocationsNonUniform || + op == glslang::EOpMaxInvocationsInclusiveScanNonUniform || + op == glslang::EOpMaxInvocationsExclusiveScanNonUniform) { + if (isFloat) + opCode = spv::OpGroupFMaxNonUniformAMD; + else { + if (isUnsigned) + opCode = spv::OpGroupUMaxNonUniformAMD; + else + opCode = spv::OpGroupSMaxNonUniformAMD; + } + } + else { + if (isFloat) + opCode = spv::OpGroupFAddNonUniformAMD; + else + opCode = spv::OpGroupIAddNonUniformAMD; + } + + if (builder.isVectorType(typeId)) + return CreateInvocationsVectorOperation(opCode, groupOperation, typeId, operands); + + break; +#endif + default: + logger->missingFunctionality("invocation operation"); + return spv::NoResult; + } + + assert(opCode != spv::OpNop); + return builder.createOp(opCode, typeId, spvGroupOperands); +} + +// Create group invocation operations on a vector +spv::Id TGlslangToSpvTraverser::CreateInvocationsVectorOperation(spv::Op op, spv::GroupOperation groupOperation, + spv::Id typeId, std::vector<spv::Id>& operands) +{ +#ifdef AMD_EXTENSIONS + assert(op == spv::OpGroupFMin || op == spv::OpGroupUMin || op == spv::OpGroupSMin || + op == spv::OpGroupFMax || op == spv::OpGroupUMax || op == spv::OpGroupSMax || + op == spv::OpGroupFAdd || op == spv::OpGroupIAdd || op == spv::OpGroupBroadcast || + op == spv::OpSubgroupReadInvocationKHR || + op == spv::OpGroupFMinNonUniformAMD || op == spv::OpGroupUMinNonUniformAMD || op == spv::OpGroupSMinNonUniformAMD || + op == spv::OpGroupFMaxNonUniformAMD || op == spv::OpGroupUMaxNonUniformAMD || op == spv::OpGroupSMaxNonUniformAMD || + op == spv::OpGroupFAddNonUniformAMD || op == spv::OpGroupIAddNonUniformAMD); +#else + assert(op == spv::OpGroupFMin || op == spv::OpGroupUMin || op == spv::OpGroupSMin || + op == spv::OpGroupFMax || op == spv::OpGroupUMax || op == spv::OpGroupSMax || + op == spv::OpGroupFAdd || op == spv::OpGroupIAdd || op == spv::OpGroupBroadcast || + op == spv::OpSubgroupReadInvocationKHR); +#endif + + // Handle group invocation operations scalar by scalar. + // The result type is the same type as the original type. + // The algorithm is to: + // - break the vector into scalars + // - apply the operation to each scalar + // - make a vector out the scalar results + + // get the types sorted out + int numComponents = builder.getNumComponents(operands[0]); + spv::Id scalarType = builder.getScalarTypeId(builder.getTypeId(operands[0])); + std::vector<spv::Id> results; + + // do each scalar op + for (int comp = 0; comp < numComponents; ++comp) { + std::vector<unsigned int> indexes; + indexes.push_back(comp); + spv::IdImmediate scalar = { true, builder.createCompositeExtract(operands[0], scalarType, indexes) }; + std::vector<spv::IdImmediate> spvGroupOperands; + if (op == spv::OpSubgroupReadInvocationKHR) { + spvGroupOperands.push_back(scalar); + spv::IdImmediate operand = { true, operands[1] }; + spvGroupOperands.push_back(operand); + } else if (op == spv::OpGroupBroadcast) { + spv::IdImmediate scope = { true, builder.makeUintConstant(spv::ScopeSubgroup) }; + spvGroupOperands.push_back(scope); + spvGroupOperands.push_back(scalar); + spv::IdImmediate operand = { true, operands[1] }; + spvGroupOperands.push_back(operand); + } else { + spv::IdImmediate scope = { true, builder.makeUintConstant(spv::ScopeSubgroup) }; + spvGroupOperands.push_back(scope); + spv::IdImmediate groupOp = { false, (unsigned)groupOperation }; + spvGroupOperands.push_back(groupOp); + spvGroupOperands.push_back(scalar); + } + + results.push_back(builder.createOp(op, scalarType, spvGroupOperands)); + } + + // put the pieces together + return builder.createCompositeConstruct(typeId, results); +} + +// Create subgroup invocation operations. +spv::Id TGlslangToSpvTraverser::createSubgroupOperation(glslang::TOperator op, spv::Id typeId, + std::vector<spv::Id>& operands, glslang::TBasicType typeProxy) +{ + // Add the required capabilities. + switch (op) { + case glslang::EOpSubgroupElect: + builder.addCapability(spv::CapabilityGroupNonUniform); + break; + case glslang::EOpSubgroupAll: + case glslang::EOpSubgroupAny: + case glslang::EOpSubgroupAllEqual: + builder.addCapability(spv::CapabilityGroupNonUniform); + builder.addCapability(spv::CapabilityGroupNonUniformVote); + break; + case glslang::EOpSubgroupBroadcast: + case glslang::EOpSubgroupBroadcastFirst: + case glslang::EOpSubgroupBallot: + case glslang::EOpSubgroupInverseBallot: + case glslang::EOpSubgroupBallotBitExtract: + case glslang::EOpSubgroupBallotBitCount: + case glslang::EOpSubgroupBallotInclusiveBitCount: + case glslang::EOpSubgroupBallotExclusiveBitCount: + case glslang::EOpSubgroupBallotFindLSB: + case glslang::EOpSubgroupBallotFindMSB: + builder.addCapability(spv::CapabilityGroupNonUniform); + builder.addCapability(spv::CapabilityGroupNonUniformBallot); + break; + case glslang::EOpSubgroupShuffle: + case glslang::EOpSubgroupShuffleXor: + builder.addCapability(spv::CapabilityGroupNonUniform); + builder.addCapability(spv::CapabilityGroupNonUniformShuffle); + break; + case glslang::EOpSubgroupShuffleUp: + case glslang::EOpSubgroupShuffleDown: + builder.addCapability(spv::CapabilityGroupNonUniform); + builder.addCapability(spv::CapabilityGroupNonUniformShuffleRelative); + break; + case glslang::EOpSubgroupAdd: + case glslang::EOpSubgroupMul: + case glslang::EOpSubgroupMin: + case glslang::EOpSubgroupMax: + case glslang::EOpSubgroupAnd: + case glslang::EOpSubgroupOr: + case glslang::EOpSubgroupXor: + case glslang::EOpSubgroupInclusiveAdd: + case glslang::EOpSubgroupInclusiveMul: + case glslang::EOpSubgroupInclusiveMin: + case glslang::EOpSubgroupInclusiveMax: + case glslang::EOpSubgroupInclusiveAnd: + case glslang::EOpSubgroupInclusiveOr: + case glslang::EOpSubgroupInclusiveXor: + case glslang::EOpSubgroupExclusiveAdd: + case glslang::EOpSubgroupExclusiveMul: + case glslang::EOpSubgroupExclusiveMin: + case glslang::EOpSubgroupExclusiveMax: + case glslang::EOpSubgroupExclusiveAnd: + case glslang::EOpSubgroupExclusiveOr: + case glslang::EOpSubgroupExclusiveXor: + builder.addCapability(spv::CapabilityGroupNonUniform); + builder.addCapability(spv::CapabilityGroupNonUniformArithmetic); + break; + case glslang::EOpSubgroupClusteredAdd: + case glslang::EOpSubgroupClusteredMul: + case glslang::EOpSubgroupClusteredMin: + case glslang::EOpSubgroupClusteredMax: + case glslang::EOpSubgroupClusteredAnd: + case glslang::EOpSubgroupClusteredOr: + case glslang::EOpSubgroupClusteredXor: + builder.addCapability(spv::CapabilityGroupNonUniform); + builder.addCapability(spv::CapabilityGroupNonUniformClustered); + break; + case glslang::EOpSubgroupQuadBroadcast: + case glslang::EOpSubgroupQuadSwapHorizontal: + case glslang::EOpSubgroupQuadSwapVertical: + case glslang::EOpSubgroupQuadSwapDiagonal: + builder.addCapability(spv::CapabilityGroupNonUniform); + builder.addCapability(spv::CapabilityGroupNonUniformQuad); + break; +#ifdef NV_EXTENSIONS + case glslang::EOpSubgroupPartitionedAdd: + case glslang::EOpSubgroupPartitionedMul: + case glslang::EOpSubgroupPartitionedMin: + case glslang::EOpSubgroupPartitionedMax: + case glslang::EOpSubgroupPartitionedAnd: + case glslang::EOpSubgroupPartitionedOr: + case glslang::EOpSubgroupPartitionedXor: + case glslang::EOpSubgroupPartitionedInclusiveAdd: + case glslang::EOpSubgroupPartitionedInclusiveMul: + case glslang::EOpSubgroupPartitionedInclusiveMin: + case glslang::EOpSubgroupPartitionedInclusiveMax: + case glslang::EOpSubgroupPartitionedInclusiveAnd: + case glslang::EOpSubgroupPartitionedInclusiveOr: + case glslang::EOpSubgroupPartitionedInclusiveXor: + case glslang::EOpSubgroupPartitionedExclusiveAdd: + case glslang::EOpSubgroupPartitionedExclusiveMul: + case glslang::EOpSubgroupPartitionedExclusiveMin: + case glslang::EOpSubgroupPartitionedExclusiveMax: + case glslang::EOpSubgroupPartitionedExclusiveAnd: + case glslang::EOpSubgroupPartitionedExclusiveOr: + case glslang::EOpSubgroupPartitionedExclusiveXor: + builder.addExtension(spv::E_SPV_NV_shader_subgroup_partitioned); + builder.addCapability(spv::CapabilityGroupNonUniformPartitionedNV); + break; +#endif + default: assert(0 && "Unhandled subgroup operation!"); + } + + const bool isUnsigned = typeProxy == glslang::EbtUint || typeProxy == glslang::EbtUint64; + const bool isFloat = typeProxy == glslang::EbtFloat || typeProxy == glslang::EbtDouble; + const bool isBool = typeProxy == glslang::EbtBool; + + spv::Op opCode = spv::OpNop; + + // Figure out which opcode to use. + switch (op) { + case glslang::EOpSubgroupElect: opCode = spv::OpGroupNonUniformElect; break; + case glslang::EOpSubgroupAll: opCode = spv::OpGroupNonUniformAll; break; + case glslang::EOpSubgroupAny: opCode = spv::OpGroupNonUniformAny; break; + case glslang::EOpSubgroupAllEqual: opCode = spv::OpGroupNonUniformAllEqual; break; + case glslang::EOpSubgroupBroadcast: opCode = spv::OpGroupNonUniformBroadcast; break; + case glslang::EOpSubgroupBroadcastFirst: opCode = spv::OpGroupNonUniformBroadcastFirst; break; + case glslang::EOpSubgroupBallot: opCode = spv::OpGroupNonUniformBallot; break; + case glslang::EOpSubgroupInverseBallot: opCode = spv::OpGroupNonUniformInverseBallot; break; + case glslang::EOpSubgroupBallotBitExtract: opCode = spv::OpGroupNonUniformBallotBitExtract; break; + case glslang::EOpSubgroupBallotBitCount: + case glslang::EOpSubgroupBallotInclusiveBitCount: + case glslang::EOpSubgroupBallotExclusiveBitCount: opCode = spv::OpGroupNonUniformBallotBitCount; break; + case glslang::EOpSubgroupBallotFindLSB: opCode = spv::OpGroupNonUniformBallotFindLSB; break; + case glslang::EOpSubgroupBallotFindMSB: opCode = spv::OpGroupNonUniformBallotFindMSB; break; + case glslang::EOpSubgroupShuffle: opCode = spv::OpGroupNonUniformShuffle; break; + case glslang::EOpSubgroupShuffleXor: opCode = spv::OpGroupNonUniformShuffleXor; break; + case glslang::EOpSubgroupShuffleUp: opCode = spv::OpGroupNonUniformShuffleUp; break; + case glslang::EOpSubgroupShuffleDown: opCode = spv::OpGroupNonUniformShuffleDown; break; + case glslang::EOpSubgroupAdd: + case glslang::EOpSubgroupInclusiveAdd: + case glslang::EOpSubgroupExclusiveAdd: + case glslang::EOpSubgroupClusteredAdd: +#ifdef NV_EXTENSIONS + case glslang::EOpSubgroupPartitionedAdd: + case glslang::EOpSubgroupPartitionedInclusiveAdd: + case glslang::EOpSubgroupPartitionedExclusiveAdd: +#endif + if (isFloat) { + opCode = spv::OpGroupNonUniformFAdd; + } else { + opCode = spv::OpGroupNonUniformIAdd; + } + break; + case glslang::EOpSubgroupMul: + case glslang::EOpSubgroupInclusiveMul: + case glslang::EOpSubgroupExclusiveMul: + case glslang::EOpSubgroupClusteredMul: +#ifdef NV_EXTENSIONS + case glslang::EOpSubgroupPartitionedMul: + case glslang::EOpSubgroupPartitionedInclusiveMul: + case glslang::EOpSubgroupPartitionedExclusiveMul: +#endif + if (isFloat) { + opCode = spv::OpGroupNonUniformFMul; + } else { + opCode = spv::OpGroupNonUniformIMul; + } + break; + case glslang::EOpSubgroupMin: + case glslang::EOpSubgroupInclusiveMin: + case glslang::EOpSubgroupExclusiveMin: + case glslang::EOpSubgroupClusteredMin: +#ifdef NV_EXTENSIONS + case glslang::EOpSubgroupPartitionedMin: + case glslang::EOpSubgroupPartitionedInclusiveMin: + case glslang::EOpSubgroupPartitionedExclusiveMin: +#endif + if (isFloat) { + opCode = spv::OpGroupNonUniformFMin; + } else if (isUnsigned) { + opCode = spv::OpGroupNonUniformUMin; + } else { + opCode = spv::OpGroupNonUniformSMin; + } + break; + case glslang::EOpSubgroupMax: + case glslang::EOpSubgroupInclusiveMax: + case glslang::EOpSubgroupExclusiveMax: + case glslang::EOpSubgroupClusteredMax: +#ifdef NV_EXTENSIONS + case glslang::EOpSubgroupPartitionedMax: + case glslang::EOpSubgroupPartitionedInclusiveMax: + case glslang::EOpSubgroupPartitionedExclusiveMax: +#endif + if (isFloat) { + opCode = spv::OpGroupNonUniformFMax; + } else if (isUnsigned) { + opCode = spv::OpGroupNonUniformUMax; + } else { + opCode = spv::OpGroupNonUniformSMax; + } + break; + case glslang::EOpSubgroupAnd: + case glslang::EOpSubgroupInclusiveAnd: + case glslang::EOpSubgroupExclusiveAnd: + case glslang::EOpSubgroupClusteredAnd: +#ifdef NV_EXTENSIONS + case glslang::EOpSubgroupPartitionedAnd: + case glslang::EOpSubgroupPartitionedInclusiveAnd: + case glslang::EOpSubgroupPartitionedExclusiveAnd: +#endif + if (isBool) { + opCode = spv::OpGroupNonUniformLogicalAnd; + } else { + opCode = spv::OpGroupNonUniformBitwiseAnd; + } + break; + case glslang::EOpSubgroupOr: + case glslang::EOpSubgroupInclusiveOr: + case glslang::EOpSubgroupExclusiveOr: + case glslang::EOpSubgroupClusteredOr: +#ifdef NV_EXTENSIONS + case glslang::EOpSubgroupPartitionedOr: + case glslang::EOpSubgroupPartitionedInclusiveOr: + case glslang::EOpSubgroupPartitionedExclusiveOr: +#endif + if (isBool) { + opCode = spv::OpGroupNonUniformLogicalOr; + } else { + opCode = spv::OpGroupNonUniformBitwiseOr; + } + break; + case glslang::EOpSubgroupXor: + case glslang::EOpSubgroupInclusiveXor: + case glslang::EOpSubgroupExclusiveXor: + case glslang::EOpSubgroupClusteredXor: +#ifdef NV_EXTENSIONS + case glslang::EOpSubgroupPartitionedXor: + case glslang::EOpSubgroupPartitionedInclusiveXor: + case glslang::EOpSubgroupPartitionedExclusiveXor: +#endif + if (isBool) { + opCode = spv::OpGroupNonUniformLogicalXor; + } else { + opCode = spv::OpGroupNonUniformBitwiseXor; + } + break; + case glslang::EOpSubgroupQuadBroadcast: opCode = spv::OpGroupNonUniformQuadBroadcast; break; + case glslang::EOpSubgroupQuadSwapHorizontal: + case glslang::EOpSubgroupQuadSwapVertical: + case glslang::EOpSubgroupQuadSwapDiagonal: opCode = spv::OpGroupNonUniformQuadSwap; break; + default: assert(0 && "Unhandled subgroup operation!"); + } + + // get the right Group Operation + spv::GroupOperation groupOperation = spv::GroupOperationMax; + switch (op) { + default: + break; + case glslang::EOpSubgroupBallotBitCount: + case glslang::EOpSubgroupAdd: + case glslang::EOpSubgroupMul: + case glslang::EOpSubgroupMin: + case glslang::EOpSubgroupMax: + case glslang::EOpSubgroupAnd: + case glslang::EOpSubgroupOr: + case glslang::EOpSubgroupXor: + groupOperation = spv::GroupOperationReduce; + break; + case glslang::EOpSubgroupBallotInclusiveBitCount: + case glslang::EOpSubgroupInclusiveAdd: + case glslang::EOpSubgroupInclusiveMul: + case glslang::EOpSubgroupInclusiveMin: + case glslang::EOpSubgroupInclusiveMax: + case glslang::EOpSubgroupInclusiveAnd: + case glslang::EOpSubgroupInclusiveOr: + case glslang::EOpSubgroupInclusiveXor: + groupOperation = spv::GroupOperationInclusiveScan; + break; + case glslang::EOpSubgroupBallotExclusiveBitCount: + case glslang::EOpSubgroupExclusiveAdd: + case glslang::EOpSubgroupExclusiveMul: + case glslang::EOpSubgroupExclusiveMin: + case glslang::EOpSubgroupExclusiveMax: + case glslang::EOpSubgroupExclusiveAnd: + case glslang::EOpSubgroupExclusiveOr: + case glslang::EOpSubgroupExclusiveXor: + groupOperation = spv::GroupOperationExclusiveScan; + break; + case glslang::EOpSubgroupClusteredAdd: + case glslang::EOpSubgroupClusteredMul: + case glslang::EOpSubgroupClusteredMin: + case glslang::EOpSubgroupClusteredMax: + case glslang::EOpSubgroupClusteredAnd: + case glslang::EOpSubgroupClusteredOr: + case glslang::EOpSubgroupClusteredXor: + groupOperation = spv::GroupOperationClusteredReduce; + break; +#ifdef NV_EXTENSIONS + case glslang::EOpSubgroupPartitionedAdd: + case glslang::EOpSubgroupPartitionedMul: + case glslang::EOpSubgroupPartitionedMin: + case glslang::EOpSubgroupPartitionedMax: + case glslang::EOpSubgroupPartitionedAnd: + case glslang::EOpSubgroupPartitionedOr: + case glslang::EOpSubgroupPartitionedXor: + groupOperation = spv::GroupOperationPartitionedReduceNV; + break; + case glslang::EOpSubgroupPartitionedInclusiveAdd: + case glslang::EOpSubgroupPartitionedInclusiveMul: + case glslang::EOpSubgroupPartitionedInclusiveMin: + case glslang::EOpSubgroupPartitionedInclusiveMax: + case glslang::EOpSubgroupPartitionedInclusiveAnd: + case glslang::EOpSubgroupPartitionedInclusiveOr: + case glslang::EOpSubgroupPartitionedInclusiveXor: + groupOperation = spv::GroupOperationPartitionedInclusiveScanNV; + break; + case glslang::EOpSubgroupPartitionedExclusiveAdd: + case glslang::EOpSubgroupPartitionedExclusiveMul: + case glslang::EOpSubgroupPartitionedExclusiveMin: + case glslang::EOpSubgroupPartitionedExclusiveMax: + case glslang::EOpSubgroupPartitionedExclusiveAnd: + case glslang::EOpSubgroupPartitionedExclusiveOr: + case glslang::EOpSubgroupPartitionedExclusiveXor: + groupOperation = spv::GroupOperationPartitionedExclusiveScanNV; + break; +#endif + } + + // build the instruction + std::vector<spv::IdImmediate> spvGroupOperands; + + // Every operation begins with the Execution Scope operand. + spv::IdImmediate executionScope = { true, builder.makeUintConstant(spv::ScopeSubgroup) }; + spvGroupOperands.push_back(executionScope); + + // Next, for all operations that use a Group Operation, push that as an operand. + if (groupOperation != spv::GroupOperationMax) { + spv::IdImmediate groupOperand = { false, (unsigned)groupOperation }; + spvGroupOperands.push_back(groupOperand); + } + + // Push back the operands next. + for (auto opIt = operands.cbegin(); opIt != operands.cend(); ++opIt) { + spv::IdImmediate operand = { true, *opIt }; + spvGroupOperands.push_back(operand); + } + + // Some opcodes have additional operands. + spv::Id directionId = spv::NoResult; + switch (op) { + default: break; + case glslang::EOpSubgroupQuadSwapHorizontal: directionId = builder.makeUintConstant(0); break; + case glslang::EOpSubgroupQuadSwapVertical: directionId = builder.makeUintConstant(1); break; + case glslang::EOpSubgroupQuadSwapDiagonal: directionId = builder.makeUintConstant(2); break; + } + if (directionId != spv::NoResult) { + spv::IdImmediate direction = { true, directionId }; + spvGroupOperands.push_back(direction); + } + + return builder.createOp(opCode, typeId, spvGroupOperands); +} + +spv::Id TGlslangToSpvTraverser::createMiscOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy) +{ + bool isUnsigned = isTypeUnsignedInt(typeProxy); + bool isFloat = isTypeFloat(typeProxy); + + spv::Op opCode = spv::OpNop; + int extBuiltins = -1; + int libCall = -1; + size_t consumedOperands = operands.size(); + spv::Id typeId0 = 0; + if (consumedOperands > 0) + typeId0 = builder.getTypeId(operands[0]); + spv::Id typeId1 = 0; + if (consumedOperands > 1) + typeId1 = builder.getTypeId(operands[1]); + spv::Id frexpIntType = 0; + + switch (op) { + case glslang::EOpMin: + if (isFloat) + libCall = spv::GLSLstd450FMin; + else if (isUnsigned) + libCall = spv::GLSLstd450UMin; + else + libCall = spv::GLSLstd450SMin; + builder.promoteScalar(precision, operands.front(), operands.back()); + break; + case glslang::EOpModf: + libCall = spv::GLSLstd450Modf; + break; + case glslang::EOpMax: + if (isFloat) + libCall = spv::GLSLstd450FMax; + else if (isUnsigned) + libCall = spv::GLSLstd450UMax; + else + libCall = spv::GLSLstd450SMax; + builder.promoteScalar(precision, operands.front(), operands.back()); + break; + case glslang::EOpPow: + libCall = spv::GLSLstd450Pow; + break; + case glslang::EOpDot: + opCode = spv::OpDot; + break; + case glslang::EOpAtan: + libCall = spv::GLSLstd450Atan2; + break; + + case glslang::EOpClamp: + if (isFloat) + libCall = spv::GLSLstd450FClamp; + else if (isUnsigned) + libCall = spv::GLSLstd450UClamp; + else + libCall = spv::GLSLstd450SClamp; + builder.promoteScalar(precision, operands.front(), operands[1]); + builder.promoteScalar(precision, operands.front(), operands[2]); + break; + case glslang::EOpMix: + if (! builder.isBoolType(builder.getScalarTypeId(builder.getTypeId(operands.back())))) { + assert(isFloat); + libCall = spv::GLSLstd450FMix; + } else { + opCode = spv::OpSelect; + std::swap(operands.front(), operands.back()); + } + builder.promoteScalar(precision, operands.front(), operands.back()); + break; + case glslang::EOpStep: + libCall = spv::GLSLstd450Step; + builder.promoteScalar(precision, operands.front(), operands.back()); + break; + case glslang::EOpSmoothStep: + libCall = spv::GLSLstd450SmoothStep; + builder.promoteScalar(precision, operands[0], operands[2]); + builder.promoteScalar(precision, operands[1], operands[2]); + break; + + case glslang::EOpDistance: + libCall = spv::GLSLstd450Distance; + break; + case glslang::EOpCross: + libCall = spv::GLSLstd450Cross; + break; + case glslang::EOpFaceForward: + libCall = spv::GLSLstd450FaceForward; + break; + case glslang::EOpReflect: + libCall = spv::GLSLstd450Reflect; + break; + case glslang::EOpRefract: + libCall = spv::GLSLstd450Refract; + break; + case glslang::EOpInterpolateAtSample: +#ifdef AMD_EXTENSIONS + if (typeProxy == glslang::EbtFloat16) + builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float); +#endif + libCall = spv::GLSLstd450InterpolateAtSample; + break; + case glslang::EOpInterpolateAtOffset: +#ifdef AMD_EXTENSIONS + if (typeProxy == glslang::EbtFloat16) + builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float); +#endif + libCall = spv::GLSLstd450InterpolateAtOffset; + break; + case glslang::EOpAddCarry: + opCode = spv::OpIAddCarry; + typeId = builder.makeStructResultType(typeId0, typeId0); + consumedOperands = 2; + break; + case glslang::EOpSubBorrow: + opCode = spv::OpISubBorrow; + typeId = builder.makeStructResultType(typeId0, typeId0); + consumedOperands = 2; + break; + case glslang::EOpUMulExtended: + opCode = spv::OpUMulExtended; + typeId = builder.makeStructResultType(typeId0, typeId0); + consumedOperands = 2; + break; + case glslang::EOpIMulExtended: + opCode = spv::OpSMulExtended; + typeId = builder.makeStructResultType(typeId0, typeId0); + consumedOperands = 2; + break; + case glslang::EOpBitfieldExtract: + if (isUnsigned) + opCode = spv::OpBitFieldUExtract; + else + opCode = spv::OpBitFieldSExtract; + break; + case glslang::EOpBitfieldInsert: + opCode = spv::OpBitFieldInsert; + break; + + case glslang::EOpFma: + libCall = spv::GLSLstd450Fma; + break; + case glslang::EOpFrexp: + { + libCall = spv::GLSLstd450FrexpStruct; + assert(builder.isPointerType(typeId1)); + typeId1 = builder.getContainedTypeId(typeId1); + int width = builder.getScalarTypeWidth(typeId1); +#ifdef AMD_EXTENSIONS + if (width == 16) + // Using 16-bit exp operand, enable extension SPV_AMD_gpu_shader_int16 + builder.addExtension(spv::E_SPV_AMD_gpu_shader_int16); +#endif + if (builder.getNumComponents(operands[0]) == 1) + frexpIntType = builder.makeIntegerType(width, true); + else + frexpIntType = builder.makeVectorType(builder.makeIntegerType(width, true), builder.getNumComponents(operands[0])); + typeId = builder.makeStructResultType(typeId0, frexpIntType); + consumedOperands = 1; + } + break; + case glslang::EOpLdexp: + libCall = spv::GLSLstd450Ldexp; + break; + + case glslang::EOpReadInvocation: + return createInvocationsOperation(op, typeId, operands, typeProxy); + + case glslang::EOpSubgroupBroadcast: + case glslang::EOpSubgroupBallotBitExtract: + case glslang::EOpSubgroupShuffle: + case glslang::EOpSubgroupShuffleXor: + case glslang::EOpSubgroupShuffleUp: + case glslang::EOpSubgroupShuffleDown: + case glslang::EOpSubgroupClusteredAdd: + case glslang::EOpSubgroupClusteredMul: + case glslang::EOpSubgroupClusteredMin: + case glslang::EOpSubgroupClusteredMax: + case glslang::EOpSubgroupClusteredAnd: + case glslang::EOpSubgroupClusteredOr: + case glslang::EOpSubgroupClusteredXor: + case glslang::EOpSubgroupQuadBroadcast: +#ifdef NV_EXTENSIONS + case glslang::EOpSubgroupPartitionedAdd: + case glslang::EOpSubgroupPartitionedMul: + case glslang::EOpSubgroupPartitionedMin: + case glslang::EOpSubgroupPartitionedMax: + case glslang::EOpSubgroupPartitionedAnd: + case glslang::EOpSubgroupPartitionedOr: + case glslang::EOpSubgroupPartitionedXor: + case glslang::EOpSubgroupPartitionedInclusiveAdd: + case glslang::EOpSubgroupPartitionedInclusiveMul: + case glslang::EOpSubgroupPartitionedInclusiveMin: + case glslang::EOpSubgroupPartitionedInclusiveMax: + case glslang::EOpSubgroupPartitionedInclusiveAnd: + case glslang::EOpSubgroupPartitionedInclusiveOr: + case glslang::EOpSubgroupPartitionedInclusiveXor: + case glslang::EOpSubgroupPartitionedExclusiveAdd: + case glslang::EOpSubgroupPartitionedExclusiveMul: + case glslang::EOpSubgroupPartitionedExclusiveMin: + case glslang::EOpSubgroupPartitionedExclusiveMax: + case glslang::EOpSubgroupPartitionedExclusiveAnd: + case glslang::EOpSubgroupPartitionedExclusiveOr: + case glslang::EOpSubgroupPartitionedExclusiveXor: +#endif + return createSubgroupOperation(op, typeId, operands, typeProxy); + +#ifdef AMD_EXTENSIONS + case glslang::EOpSwizzleInvocations: + extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_ballot); + libCall = spv::SwizzleInvocationsAMD; + break; + case glslang::EOpSwizzleInvocationsMasked: + extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_ballot); + libCall = spv::SwizzleInvocationsMaskedAMD; + break; + case glslang::EOpWriteInvocation: + extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_ballot); + libCall = spv::WriteInvocationAMD; + break; + + case glslang::EOpMin3: + extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_trinary_minmax); + if (isFloat) + libCall = spv::FMin3AMD; + else { + if (isUnsigned) + libCall = spv::UMin3AMD; + else + libCall = spv::SMin3AMD; + } + break; + case glslang::EOpMax3: + extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_trinary_minmax); + if (isFloat) + libCall = spv::FMax3AMD; + else { + if (isUnsigned) + libCall = spv::UMax3AMD; + else + libCall = spv::SMax3AMD; + } + break; + case glslang::EOpMid3: + extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_trinary_minmax); + if (isFloat) + libCall = spv::FMid3AMD; + else { + if (isUnsigned) + libCall = spv::UMid3AMD; + else + libCall = spv::SMid3AMD; + } + break; + + case glslang::EOpInterpolateAtVertex: + if (typeProxy == glslang::EbtFloat16) + builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float); + extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_explicit_vertex_parameter); + libCall = spv::InterpolateAtVertexAMD; + break; +#endif + case glslang::EOpBarrier: + { + // This is for the extended controlBarrier function, with four operands. + // The unextended barrier() goes through createNoArgOperation. + assert(operands.size() == 4); + unsigned int executionScope = builder.getConstantScalar(operands[0]); + unsigned int memoryScope = builder.getConstantScalar(operands[1]); + unsigned int semantics = builder.getConstantScalar(operands[2]) | builder.getConstantScalar(operands[3]); + builder.createControlBarrier((spv::Scope)executionScope, (spv::Scope)memoryScope, (spv::MemorySemanticsMask)semantics); + if (semantics & (spv::MemorySemanticsMakeAvailableKHRMask | spv::MemorySemanticsMakeVisibleKHRMask | spv::MemorySemanticsOutputMemoryKHRMask)) { + builder.addCapability(spv::CapabilityVulkanMemoryModelKHR); + } + if (glslangIntermediate->usingVulkanMemoryModel() && (executionScope == spv::ScopeDevice || memoryScope == spv::ScopeDevice)) { + builder.addCapability(spv::CapabilityVulkanMemoryModelDeviceScopeKHR); + } + return 0; + } + break; + case glslang::EOpMemoryBarrier: + { + // This is for the extended memoryBarrier function, with three operands. + // The unextended memoryBarrier() goes through createNoArgOperation. + assert(operands.size() == 3); + unsigned int memoryScope = builder.getConstantScalar(operands[0]); + unsigned int semantics = builder.getConstantScalar(operands[1]) | builder.getConstantScalar(operands[2]); + builder.createMemoryBarrier((spv::Scope)memoryScope, (spv::MemorySemanticsMask)semantics); + if (semantics & (spv::MemorySemanticsMakeAvailableKHRMask | spv::MemorySemanticsMakeVisibleKHRMask | spv::MemorySemanticsOutputMemoryKHRMask)) { + builder.addCapability(spv::CapabilityVulkanMemoryModelKHR); + } + if (glslangIntermediate->usingVulkanMemoryModel() && memoryScope == spv::ScopeDevice) { + builder.addCapability(spv::CapabilityVulkanMemoryModelDeviceScopeKHR); + } + return 0; + } + break; + +#ifdef NV_EXTENSIONS + case glslang::EOpReportIntersectionNV: + { + typeId = builder.makeBoolType(); + opCode = spv::OpReportIntersectionNV; + } + break; + case glslang::EOpTraceNV: + { + builder.createNoResultOp(spv::OpTraceNV, operands); + return 0; + } + break; + case glslang::EOpExecuteCallableNV: + { + builder.createNoResultOp(spv::OpExecuteCallableNV, operands); + return 0; + } + break; + case glslang::EOpWritePackedPrimitiveIndices4x8NV: + builder.createNoResultOp(spv::OpWritePackedPrimitiveIndices4x8NV, operands); + return 0; +#endif + case glslang::EOpCooperativeMatrixMulAdd: + opCode = spv::OpCooperativeMatrixMulAddNV; + break; + + default: + return 0; + } + + spv::Id id = 0; + if (libCall >= 0) { + // Use an extended instruction from the standard library. + // Construct the call arguments, without modifying the original operands vector. + // We might need the remaining arguments, e.g. in the EOpFrexp case. + std::vector<spv::Id> callArguments(operands.begin(), operands.begin() + consumedOperands); + id = builder.createBuiltinCall(typeId, extBuiltins >= 0 ? extBuiltins : stdBuiltins, libCall, callArguments); + } else if (opCode == spv::OpDot && !isFloat) { + // int dot(int, int) + // NOTE: never called for scalar/vector1, this is turned into simple mul before this can be reached + const int componentCount = builder.getNumComponents(operands[0]); + spv::Id mulOp = builder.createBinOp(spv::OpIMul, builder.getTypeId(operands[0]), operands[0], operands[1]); + builder.setPrecision(mulOp, precision); + id = builder.createCompositeExtract(mulOp, typeId, 0); + for (int i = 1; i < componentCount; ++i) { + builder.setPrecision(id, precision); + id = builder.createBinOp(spv::OpIAdd, typeId, id, builder.createCompositeExtract(operands[0], typeId, i)); + } + } else { + switch (consumedOperands) { + case 0: + // should all be handled by visitAggregate and createNoArgOperation + assert(0); + return 0; + case 1: + // should all be handled by createUnaryOperation + assert(0); + return 0; + case 2: + id = builder.createBinOp(opCode, typeId, operands[0], operands[1]); + break; + default: + // anything 3 or over doesn't have l-value operands, so all should be consumed + assert(consumedOperands == operands.size()); + id = builder.createOp(opCode, typeId, operands); + break; + } + } + + // Decode the return types that were structures + switch (op) { + case glslang::EOpAddCarry: + case glslang::EOpSubBorrow: + builder.createStore(builder.createCompositeExtract(id, typeId0, 1), operands[2]); + id = builder.createCompositeExtract(id, typeId0, 0); + break; + case glslang::EOpUMulExtended: + case glslang::EOpIMulExtended: + builder.createStore(builder.createCompositeExtract(id, typeId0, 0), operands[3]); + builder.createStore(builder.createCompositeExtract(id, typeId0, 1), operands[2]); + break; + case glslang::EOpFrexp: + { + assert(operands.size() == 2); + if (builder.isFloatType(builder.getScalarTypeId(typeId1))) { + // "exp" is floating-point type (from HLSL intrinsic) + spv::Id member1 = builder.createCompositeExtract(id, frexpIntType, 1); + member1 = builder.createUnaryOp(spv::OpConvertSToF, typeId1, member1); + builder.createStore(member1, operands[1]); + } else + // "exp" is integer type (from GLSL built-in function) + builder.createStore(builder.createCompositeExtract(id, frexpIntType, 1), operands[1]); + id = builder.createCompositeExtract(id, typeId0, 0); + } + break; + default: + break; + } + + return builder.setPrecision(id, precision); +} + +// Intrinsics with no arguments (or no return value, and no precision). +spv::Id TGlslangToSpvTraverser::createNoArgOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId) +{ + // GLSL memory barriers use queuefamily scope in new model, device scope in old model + spv::Scope memoryBarrierScope = glslangIntermediate->usingVulkanMemoryModel() ? spv::ScopeQueueFamilyKHR : spv::ScopeDevice; + + switch (op) { + case glslang::EOpEmitVertex: + builder.createNoResultOp(spv::OpEmitVertex); + return 0; + case glslang::EOpEndPrimitive: + builder.createNoResultOp(spv::OpEndPrimitive); + return 0; + case glslang::EOpBarrier: + if (glslangIntermediate->getStage() == EShLangTessControl) { + if (glslangIntermediate->usingVulkanMemoryModel()) { + builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeWorkgroup, + spv::MemorySemanticsOutputMemoryKHRMask | + spv::MemorySemanticsAcquireReleaseMask); + builder.addCapability(spv::CapabilityVulkanMemoryModelKHR); + } else { + builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeInvocation, spv::MemorySemanticsMaskNone); + } + } else { + builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeWorkgroup, + spv::MemorySemanticsWorkgroupMemoryMask | + spv::MemorySemanticsAcquireReleaseMask); + } + return 0; + case glslang::EOpMemoryBarrier: + builder.createMemoryBarrier(memoryBarrierScope, spv::MemorySemanticsAllMemory | + spv::MemorySemanticsAcquireReleaseMask); + return 0; + case glslang::EOpMemoryBarrierAtomicCounter: + builder.createMemoryBarrier(memoryBarrierScope, spv::MemorySemanticsAtomicCounterMemoryMask | + spv::MemorySemanticsAcquireReleaseMask); + return 0; + case glslang::EOpMemoryBarrierBuffer: + builder.createMemoryBarrier(memoryBarrierScope, spv::MemorySemanticsUniformMemoryMask | + spv::MemorySemanticsAcquireReleaseMask); + return 0; + case glslang::EOpMemoryBarrierImage: + builder.createMemoryBarrier(memoryBarrierScope, spv::MemorySemanticsImageMemoryMask | + spv::MemorySemanticsAcquireReleaseMask); + return 0; + case glslang::EOpMemoryBarrierShared: + builder.createMemoryBarrier(memoryBarrierScope, spv::MemorySemanticsWorkgroupMemoryMask | + spv::MemorySemanticsAcquireReleaseMask); + return 0; + case glslang::EOpGroupMemoryBarrier: + builder.createMemoryBarrier(spv::ScopeWorkgroup, spv::MemorySemanticsAllMemory | + spv::MemorySemanticsAcquireReleaseMask); + return 0; + case glslang::EOpAllMemoryBarrierWithGroupSync: + builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeDevice, + spv::MemorySemanticsAllMemory | + spv::MemorySemanticsAcquireReleaseMask); + return 0; + case glslang::EOpDeviceMemoryBarrier: + builder.createMemoryBarrier(spv::ScopeDevice, spv::MemorySemanticsUniformMemoryMask | + spv::MemorySemanticsImageMemoryMask | + spv::MemorySemanticsAcquireReleaseMask); + return 0; + case glslang::EOpDeviceMemoryBarrierWithGroupSync: + builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeDevice, spv::MemorySemanticsUniformMemoryMask | + spv::MemorySemanticsImageMemoryMask | + spv::MemorySemanticsAcquireReleaseMask); + return 0; + case glslang::EOpWorkgroupMemoryBarrier: + builder.createMemoryBarrier(spv::ScopeWorkgroup, spv::MemorySemanticsWorkgroupMemoryMask | + spv::MemorySemanticsAcquireReleaseMask); + return 0; + case glslang::EOpWorkgroupMemoryBarrierWithGroupSync: + builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeWorkgroup, + spv::MemorySemanticsWorkgroupMemoryMask | + spv::MemorySemanticsAcquireReleaseMask); + return 0; + case glslang::EOpSubgroupBarrier: + builder.createControlBarrier(spv::ScopeSubgroup, spv::ScopeSubgroup, spv::MemorySemanticsAllMemory | + spv::MemorySemanticsAcquireReleaseMask); + return spv::NoResult; + case glslang::EOpSubgroupMemoryBarrier: + builder.createMemoryBarrier(spv::ScopeSubgroup, spv::MemorySemanticsAllMemory | + spv::MemorySemanticsAcquireReleaseMask); + return spv::NoResult; + case glslang::EOpSubgroupMemoryBarrierBuffer: + builder.createMemoryBarrier(spv::ScopeSubgroup, spv::MemorySemanticsUniformMemoryMask | + spv::MemorySemanticsAcquireReleaseMask); + return spv::NoResult; + case glslang::EOpSubgroupMemoryBarrierImage: + builder.createMemoryBarrier(spv::ScopeSubgroup, spv::MemorySemanticsImageMemoryMask | + spv::MemorySemanticsAcquireReleaseMask); + return spv::NoResult; + case glslang::EOpSubgroupMemoryBarrierShared: + builder.createMemoryBarrier(spv::ScopeSubgroup, spv::MemorySemanticsWorkgroupMemoryMask | + spv::MemorySemanticsAcquireReleaseMask); + return spv::NoResult; + case glslang::EOpSubgroupElect: { + std::vector<spv::Id> operands; + return createSubgroupOperation(op, typeId, operands, glslang::EbtVoid); + } +#ifdef AMD_EXTENSIONS + case glslang::EOpTime: + { + std::vector<spv::Id> args; // Dummy arguments + spv::Id id = builder.createBuiltinCall(typeId, getExtBuiltins(spv::E_SPV_AMD_gcn_shader), spv::TimeAMD, args); + return builder.setPrecision(id, precision); + } +#endif +#ifdef NV_EXTENSIONS + case glslang::EOpIgnoreIntersectionNV: + builder.createNoResultOp(spv::OpIgnoreIntersectionNV); + return 0; + case glslang::EOpTerminateRayNV: + builder.createNoResultOp(spv::OpTerminateRayNV); + return 0; +#endif + default: + logger->missingFunctionality("unknown operation with no arguments"); + return 0; + } +} + +spv::Id TGlslangToSpvTraverser::getSymbolId(const glslang::TIntermSymbol* symbol) +{ + auto iter = symbolValues.find(symbol->getId()); + spv::Id id; + if (symbolValues.end() != iter) { + id = iter->second; + return id; + } + + // it was not found, create it + id = createSpvVariable(symbol); + symbolValues[symbol->getId()] = id; + + if (symbol->getBasicType() != glslang::EbtBlock) { + builder.addDecoration(id, TranslatePrecisionDecoration(symbol->getType())); + builder.addDecoration(id, TranslateInterpolationDecoration(symbol->getType().getQualifier())); + builder.addDecoration(id, TranslateAuxiliaryStorageDecoration(symbol->getType().getQualifier())); +#ifdef NV_EXTENSIONS + addMeshNVDecoration(id, /*member*/ -1, symbol->getType().getQualifier()); +#endif + if (symbol->getType().getQualifier().hasSpecConstantId()) + builder.addDecoration(id, spv::DecorationSpecId, symbol->getType().getQualifier().layoutSpecConstantId); + if (symbol->getQualifier().hasIndex()) + builder.addDecoration(id, spv::DecorationIndex, symbol->getQualifier().layoutIndex); + if (symbol->getQualifier().hasComponent()) + builder.addDecoration(id, spv::DecorationComponent, symbol->getQualifier().layoutComponent); + // atomic counters use this: + if (symbol->getQualifier().hasOffset()) + builder.addDecoration(id, spv::DecorationOffset, symbol->getQualifier().layoutOffset); + } + + if (symbol->getQualifier().hasLocation()) + builder.addDecoration(id, spv::DecorationLocation, symbol->getQualifier().layoutLocation); + builder.addDecoration(id, TranslateInvariantDecoration(symbol->getType().getQualifier())); + if (symbol->getQualifier().hasStream() && glslangIntermediate->isMultiStream()) { + builder.addCapability(spv::CapabilityGeometryStreams); + builder.addDecoration(id, spv::DecorationStream, symbol->getQualifier().layoutStream); + } + if (symbol->getQualifier().hasSet()) + builder.addDecoration(id, spv::DecorationDescriptorSet, symbol->getQualifier().layoutSet); + else if (IsDescriptorResource(symbol->getType())) { + // default to 0 + builder.addDecoration(id, spv::DecorationDescriptorSet, 0); + } + if (symbol->getQualifier().hasBinding()) + builder.addDecoration(id, spv::DecorationBinding, symbol->getQualifier().layoutBinding); + else if (IsDescriptorResource(symbol->getType())) { + // default to 0 + builder.addDecoration(id, spv::DecorationBinding, 0); + } + if (symbol->getQualifier().hasAttachment()) + builder.addDecoration(id, spv::DecorationInputAttachmentIndex, symbol->getQualifier().layoutAttachment); + if (glslangIntermediate->getXfbMode()) { + builder.addCapability(spv::CapabilityTransformFeedback); + if (symbol->getQualifier().hasXfbBuffer()) { + builder.addDecoration(id, spv::DecorationXfbBuffer, symbol->getQualifier().layoutXfbBuffer); + unsigned stride = glslangIntermediate->getXfbStride(symbol->getQualifier().layoutXfbBuffer); + if (stride != glslang::TQualifier::layoutXfbStrideEnd) + builder.addDecoration(id, spv::DecorationXfbStride, stride); + } + if (symbol->getQualifier().hasXfbOffset()) + builder.addDecoration(id, spv::DecorationOffset, symbol->getQualifier().layoutXfbOffset); + } + + if (symbol->getType().isImage()) { + std::vector<spv::Decoration> memory; + TranslateMemoryDecoration(symbol->getType().getQualifier(), memory, glslangIntermediate->usingVulkanMemoryModel()); + for (unsigned int i = 0; i < memory.size(); ++i) + builder.addDecoration(id, memory[i]); + } + + // built-in variable decorations + spv::BuiltIn builtIn = TranslateBuiltInDecoration(symbol->getQualifier().builtIn, false); + if (builtIn != spv::BuiltInMax) + builder.addDecoration(id, spv::DecorationBuiltIn, (int)builtIn); + + // nonuniform + builder.addDecoration(id, TranslateNonUniformDecoration(symbol->getType().getQualifier())); + +#ifdef NV_EXTENSIONS + if (builtIn == spv::BuiltInSampleMask) { + spv::Decoration decoration; + // GL_NV_sample_mask_override_coverage extension + if (glslangIntermediate->getLayoutOverrideCoverage()) + decoration = (spv::Decoration)spv::DecorationOverrideCoverageNV; + else + decoration = (spv::Decoration)spv::DecorationMax; + builder.addDecoration(id, decoration); + if (decoration != spv::DecorationMax) { + builder.addExtension(spv::E_SPV_NV_sample_mask_override_coverage); + } + } + else if (builtIn == spv::BuiltInLayer) { + // SPV_NV_viewport_array2 extension + if (symbol->getQualifier().layoutViewportRelative) { + builder.addDecoration(id, (spv::Decoration)spv::DecorationViewportRelativeNV); + builder.addCapability(spv::CapabilityShaderViewportMaskNV); + builder.addExtension(spv::E_SPV_NV_viewport_array2); + } + if (symbol->getQualifier().layoutSecondaryViewportRelativeOffset != -2048) { + builder.addDecoration(id, (spv::Decoration)spv::DecorationSecondaryViewportRelativeNV, + symbol->getQualifier().layoutSecondaryViewportRelativeOffset); + builder.addCapability(spv::CapabilityShaderStereoViewNV); + builder.addExtension(spv::E_SPV_NV_stereo_view_rendering); + } + } + + if (symbol->getQualifier().layoutPassthrough) { + builder.addDecoration(id, spv::DecorationPassthroughNV); + builder.addCapability(spv::CapabilityGeometryShaderPassthroughNV); + builder.addExtension(spv::E_SPV_NV_geometry_shader_passthrough); + } + if (symbol->getQualifier().pervertexNV) { + builder.addDecoration(id, spv::DecorationPerVertexNV); + builder.addCapability(spv::CapabilityFragmentBarycentricNV); + builder.addExtension(spv::E_SPV_NV_fragment_shader_barycentric); + } +#endif + + if (glslangIntermediate->getHlslFunctionality1() && symbol->getType().getQualifier().semanticName != nullptr) { + builder.addExtension("SPV_GOOGLE_hlsl_functionality1"); + builder.addDecoration(id, (spv::Decoration)spv::DecorationHlslSemanticGOOGLE, + symbol->getType().getQualifier().semanticName); + } + + if (symbol->getBasicType() == glslang::EbtReference) { + builder.addDecoration(id, symbol->getType().getQualifier().restrict ? spv::DecorationRestrictPointerEXT : spv::DecorationAliasedPointerEXT); + } + + return id; +} + +#ifdef NV_EXTENSIONS +// add per-primitive, per-view. per-task decorations to a struct member (member >= 0) or an object +void TGlslangToSpvTraverser::addMeshNVDecoration(spv::Id id, int member, const glslang::TQualifier& qualifier) +{ + if (member >= 0) { + if (qualifier.perPrimitiveNV) { + // Need to add capability/extension for fragment shader. + // Mesh shader already adds this by default. + if (glslangIntermediate->getStage() == EShLangFragment) { + builder.addCapability(spv::CapabilityMeshShadingNV); + builder.addExtension(spv::E_SPV_NV_mesh_shader); + } + builder.addMemberDecoration(id, (unsigned)member, spv::DecorationPerPrimitiveNV); + } + if (qualifier.perViewNV) + builder.addMemberDecoration(id, (unsigned)member, spv::DecorationPerViewNV); + if (qualifier.perTaskNV) + builder.addMemberDecoration(id, (unsigned)member, spv::DecorationPerTaskNV); + } else { + if (qualifier.perPrimitiveNV) { + // Need to add capability/extension for fragment shader. + // Mesh shader already adds this by default. + if (glslangIntermediate->getStage() == EShLangFragment) { + builder.addCapability(spv::CapabilityMeshShadingNV); + builder.addExtension(spv::E_SPV_NV_mesh_shader); + } + builder.addDecoration(id, spv::DecorationPerPrimitiveNV); + } + if (qualifier.perViewNV) + builder.addDecoration(id, spv::DecorationPerViewNV); + if (qualifier.perTaskNV) + builder.addDecoration(id, spv::DecorationPerTaskNV); + } +} +#endif + +// Make a full tree of instructions to build a SPIR-V specialization constant, +// or regular constant if possible. +// +// TBD: this is not yet done, nor verified to be the best design, it does do the leaf symbols though +// +// Recursively walk the nodes. The nodes form a tree whose leaves are +// regular constants, which themselves are trees that createSpvConstant() +// recursively walks. So, this function walks the "top" of the tree: +// - emit specialization constant-building instructions for specConstant +// - when running into a non-spec-constant, switch to createSpvConstant() +spv::Id TGlslangToSpvTraverser::createSpvConstant(const glslang::TIntermTyped& node) +{ + assert(node.getQualifier().isConstant()); + + // Handle front-end constants first (non-specialization constants). + if (! node.getQualifier().specConstant) { + // hand off to the non-spec-constant path + assert(node.getAsConstantUnion() != nullptr || node.getAsSymbolNode() != nullptr); + int nextConst = 0; + return createSpvConstantFromConstUnionArray(node.getType(), node.getAsConstantUnion() ? node.getAsConstantUnion()->getConstArray() : node.getAsSymbolNode()->getConstArray(), + nextConst, false); + } + + // We now know we have a specialization constant to build + + // gl_WorkGroupSize is a special case until the front-end handles hierarchical specialization constants, + // even then, it's specialization ids are handled by special case syntax in GLSL: layout(local_size_x = ... + if (node.getType().getQualifier().builtIn == glslang::EbvWorkGroupSize) { + std::vector<spv::Id> dimConstId; + for (int dim = 0; dim < 3; ++dim) { + bool specConst = (glslangIntermediate->getLocalSizeSpecId(dim) != glslang::TQualifier::layoutNotSet); + dimConstId.push_back(builder.makeUintConstant(glslangIntermediate->getLocalSize(dim), specConst)); + if (specConst) { + builder.addDecoration(dimConstId.back(), spv::DecorationSpecId, + glslangIntermediate->getLocalSizeSpecId(dim)); + } + } + return builder.makeCompositeConstant(builder.makeVectorType(builder.makeUintType(32), 3), dimConstId, true); + } + + // An AST node labelled as specialization constant should be a symbol node. + // Its initializer should either be a sub tree with constant nodes, or a constant union array. + if (auto* sn = node.getAsSymbolNode()) { + spv::Id result; + if (auto* sub_tree = sn->getConstSubtree()) { + // Traverse the constant constructor sub tree like generating normal run-time instructions. + // During the AST traversal, if the node is marked as 'specConstant', SpecConstantOpModeGuard + // will set the builder into spec constant op instruction generating mode. + sub_tree->traverse(this); + result = accessChainLoad(sub_tree->getType()); + } else if (auto* const_union_array = &sn->getConstArray()) { + int nextConst = 0; + result = createSpvConstantFromConstUnionArray(sn->getType(), *const_union_array, nextConst, true); + } else { + logger->missingFunctionality("Invalid initializer for spec onstant."); + return spv::NoResult; + } + builder.addName(result, sn->getName().c_str()); + return result; + } + + // Neither a front-end constant node, nor a specialization constant node with constant union array or + // constant sub tree as initializer. + logger->missingFunctionality("Neither a front-end constant nor a spec constant."); + return spv::NoResult; +} + +// Use 'consts' as the flattened glslang source of scalar constants to recursively +// build the aggregate SPIR-V constant. +// +// If there are not enough elements present in 'consts', 0 will be substituted; +// an empty 'consts' can be used to create a fully zeroed SPIR-V constant. +// +spv::Id TGlslangToSpvTraverser::createSpvConstantFromConstUnionArray(const glslang::TType& glslangType, const glslang::TConstUnionArray& consts, int& nextConst, bool specConstant) +{ + // vector of constants for SPIR-V + std::vector<spv::Id> spvConsts; + + // Type is used for struct and array constants + spv::Id typeId = convertGlslangToSpvType(glslangType); + + if (glslangType.isArray()) { + glslang::TType elementType(glslangType, 0); + for (int i = 0; i < glslangType.getOuterArraySize(); ++i) + spvConsts.push_back(createSpvConstantFromConstUnionArray(elementType, consts, nextConst, false)); + } else if (glslangType.isMatrix()) { + glslang::TType vectorType(glslangType, 0); + for (int col = 0; col < glslangType.getMatrixCols(); ++col) + spvConsts.push_back(createSpvConstantFromConstUnionArray(vectorType, consts, nextConst, false)); + } else if (glslangType.isCoopMat()) { + glslang::TType componentType(glslangType.getBasicType()); + spvConsts.push_back(createSpvConstantFromConstUnionArray(componentType, consts, nextConst, false)); + } else if (glslangType.isStruct()) { + glslang::TVector<glslang::TTypeLoc>::const_iterator iter; + for (iter = glslangType.getStruct()->begin(); iter != glslangType.getStruct()->end(); ++iter) + spvConsts.push_back(createSpvConstantFromConstUnionArray(*iter->type, consts, nextConst, false)); + } else if (glslangType.getVectorSize() > 1) { + for (unsigned int i = 0; i < (unsigned int)glslangType.getVectorSize(); ++i) { + bool zero = nextConst >= consts.size(); + switch (glslangType.getBasicType()) { + case glslang::EbtInt8: + spvConsts.push_back(builder.makeInt8Constant(zero ? 0 : consts[nextConst].getI8Const())); + break; + case glslang::EbtUint8: + spvConsts.push_back(builder.makeUint8Constant(zero ? 0 : consts[nextConst].getU8Const())); + break; + case glslang::EbtInt16: + spvConsts.push_back(builder.makeInt16Constant(zero ? 0 : consts[nextConst].getI16Const())); + break; + case glslang::EbtUint16: + spvConsts.push_back(builder.makeUint16Constant(zero ? 0 : consts[nextConst].getU16Const())); + break; + case glslang::EbtInt: + spvConsts.push_back(builder.makeIntConstant(zero ? 0 : consts[nextConst].getIConst())); + break; + case glslang::EbtUint: + spvConsts.push_back(builder.makeUintConstant(zero ? 0 : consts[nextConst].getUConst())); + break; + case glslang::EbtInt64: + spvConsts.push_back(builder.makeInt64Constant(zero ? 0 : consts[nextConst].getI64Const())); + break; + case glslang::EbtUint64: + spvConsts.push_back(builder.makeUint64Constant(zero ? 0 : consts[nextConst].getU64Const())); + break; + case glslang::EbtFloat: + spvConsts.push_back(builder.makeFloatConstant(zero ? 0.0F : (float)consts[nextConst].getDConst())); + break; + case glslang::EbtDouble: + spvConsts.push_back(builder.makeDoubleConstant(zero ? 0.0 : consts[nextConst].getDConst())); + break; + case glslang::EbtFloat16: + spvConsts.push_back(builder.makeFloat16Constant(zero ? 0.0F : (float)consts[nextConst].getDConst())); + break; + case glslang::EbtBool: + spvConsts.push_back(builder.makeBoolConstant(zero ? false : consts[nextConst].getBConst())); + break; + default: + assert(0); + break; + } + ++nextConst; + } + } else { + // we have a non-aggregate (scalar) constant + bool zero = nextConst >= consts.size(); + spv::Id scalar = 0; + switch (glslangType.getBasicType()) { + case glslang::EbtInt8: + scalar = builder.makeInt8Constant(zero ? 0 : consts[nextConst].getI8Const(), specConstant); + break; + case glslang::EbtUint8: + scalar = builder.makeUint8Constant(zero ? 0 : consts[nextConst].getU8Const(), specConstant); + break; + case glslang::EbtInt16: + scalar = builder.makeInt16Constant(zero ? 0 : consts[nextConst].getI16Const(), specConstant); + break; + case glslang::EbtUint16: + scalar = builder.makeUint16Constant(zero ? 0 : consts[nextConst].getU16Const(), specConstant); + break; + case glslang::EbtInt: + scalar = builder.makeIntConstant(zero ? 0 : consts[nextConst].getIConst(), specConstant); + break; + case glslang::EbtUint: + scalar = builder.makeUintConstant(zero ? 0 : consts[nextConst].getUConst(), specConstant); + break; + case glslang::EbtInt64: + scalar = builder.makeInt64Constant(zero ? 0 : consts[nextConst].getI64Const(), specConstant); + break; + case glslang::EbtUint64: + scalar = builder.makeUint64Constant(zero ? 0 : consts[nextConst].getU64Const(), specConstant); + break; + case glslang::EbtFloat: + scalar = builder.makeFloatConstant(zero ? 0.0F : (float)consts[nextConst].getDConst(), specConstant); + break; + case glslang::EbtDouble: + scalar = builder.makeDoubleConstant(zero ? 0.0 : consts[nextConst].getDConst(), specConstant); + break; + case glslang::EbtFloat16: + scalar = builder.makeFloat16Constant(zero ? 0.0F : (float)consts[nextConst].getDConst(), specConstant); + break; + case glslang::EbtBool: + scalar = builder.makeBoolConstant(zero ? false : consts[nextConst].getBConst(), specConstant); + break; + case glslang::EbtReference: + scalar = builder.makeUint64Constant(zero ? 0 : consts[nextConst].getU64Const(), specConstant); + scalar = builder.createUnaryOp(spv::OpBitcast, typeId, scalar); + break; + default: + assert(0); + break; + } + ++nextConst; + return scalar; + } + + return builder.makeCompositeConstant(typeId, spvConsts); +} + +// Return true if the node is a constant or symbol whose reading has no +// non-trivial observable cost or effect. +bool TGlslangToSpvTraverser::isTrivialLeaf(const glslang::TIntermTyped* node) +{ + // don't know what this is + if (node == nullptr) + return false; + + // a constant is safe + if (node->getAsConstantUnion() != nullptr) + return true; + + // not a symbol means non-trivial + if (node->getAsSymbolNode() == nullptr) + return false; + + // a symbol, depends on what's being read + switch (node->getType().getQualifier().storage) { + case glslang::EvqTemporary: + case glslang::EvqGlobal: + case glslang::EvqIn: + case glslang::EvqInOut: + case glslang::EvqConst: + case glslang::EvqConstReadOnly: + case glslang::EvqUniform: + return true; + default: + return false; + } +} + +// A node is trivial if it is a single operation with no side effects. +// HLSL (and/or vectors) are always trivial, as it does not short circuit. +// Otherwise, error on the side of saying non-trivial. +// Return true if trivial. +bool TGlslangToSpvTraverser::isTrivial(const glslang::TIntermTyped* node) +{ + if (node == nullptr) + return false; + + // count non scalars as trivial, as well as anything coming from HLSL + if (! node->getType().isScalarOrVec1() || glslangIntermediate->getSource() == glslang::EShSourceHlsl) + return true; + + // symbols and constants are trivial + if (isTrivialLeaf(node)) + return true; + + // otherwise, it needs to be a simple operation or one or two leaf nodes + + // not a simple operation + const glslang::TIntermBinary* binaryNode = node->getAsBinaryNode(); + const glslang::TIntermUnary* unaryNode = node->getAsUnaryNode(); + if (binaryNode == nullptr && unaryNode == nullptr) + return false; + + // not on leaf nodes + if (binaryNode && (! isTrivialLeaf(binaryNode->getLeft()) || ! isTrivialLeaf(binaryNode->getRight()))) + return false; + + if (unaryNode && ! isTrivialLeaf(unaryNode->getOperand())) { + return false; + } + + switch (node->getAsOperator()->getOp()) { + case glslang::EOpLogicalNot: + case glslang::EOpConvIntToBool: + case glslang::EOpConvUintToBool: + case glslang::EOpConvFloatToBool: + case glslang::EOpConvDoubleToBool: + case glslang::EOpEqual: + case glslang::EOpNotEqual: + case glslang::EOpLessThan: + case glslang::EOpGreaterThan: + case glslang::EOpLessThanEqual: + case glslang::EOpGreaterThanEqual: + case glslang::EOpIndexDirect: + case glslang::EOpIndexDirectStruct: + case glslang::EOpLogicalXor: + case glslang::EOpAny: + case glslang::EOpAll: + return true; + default: + return false; + } +} + +// Emit short-circuiting code, where 'right' is never evaluated unless +// the left side is true (for &&) or false (for ||). +spv::Id TGlslangToSpvTraverser::createShortCircuit(glslang::TOperator op, glslang::TIntermTyped& left, glslang::TIntermTyped& right) +{ + spv::Id boolTypeId = builder.makeBoolType(); + + // emit left operand + builder.clearAccessChain(); + left.traverse(this); + spv::Id leftId = accessChainLoad(left.getType()); + + // Operands to accumulate OpPhi operands + std::vector<spv::Id> phiOperands; + // accumulate left operand's phi information + phiOperands.push_back(leftId); + phiOperands.push_back(builder.getBuildPoint()->getId()); + + // Make the two kinds of operation symmetric with a "!" + // || => emit "if (! left) result = right" + // && => emit "if ( left) result = right" + // + // TODO: this runtime "not" for || could be avoided by adding functionality + // to 'builder' to have an "else" without an "then" + if (op == glslang::EOpLogicalOr) + leftId = builder.createUnaryOp(spv::OpLogicalNot, boolTypeId, leftId); + + // make an "if" based on the left value + spv::Builder::If ifBuilder(leftId, spv::SelectionControlMaskNone, builder); + + // emit right operand as the "then" part of the "if" + builder.clearAccessChain(); + right.traverse(this); + spv::Id rightId = accessChainLoad(right.getType()); + + // accumulate left operand's phi information + phiOperands.push_back(rightId); + phiOperands.push_back(builder.getBuildPoint()->getId()); + + // finish the "if" + ifBuilder.makeEndIf(); + + // phi together the two results + return builder.createOp(spv::OpPhi, boolTypeId, phiOperands); +} + +#ifdef AMD_EXTENSIONS +// Return type Id of the imported set of extended instructions corresponds to the name. +// Import this set if it has not been imported yet. +spv::Id TGlslangToSpvTraverser::getExtBuiltins(const char* name) +{ + if (extBuiltinMap.find(name) != extBuiltinMap.end()) + return extBuiltinMap[name]; + else { + builder.addExtension(name); + spv::Id extBuiltins = builder.import(name); + extBuiltinMap[name] = extBuiltins; + return extBuiltins; + } +} +#endif + +}; // end anonymous namespace + +namespace glslang { + +void GetSpirvVersion(std::string& version) +{ + const int bufSize = 100; + char buf[bufSize]; + snprintf(buf, bufSize, "0x%08x, Revision %d", spv::Version, spv::Revision); + version = buf; +} + +// For low-order part of the generator's magic number. Bump up +// when there is a change in the style (e.g., if SSA form changes, +// or a different instruction sequence to do something gets used). +int GetSpirvGeneratorVersion() +{ + // return 1; // start + // return 2; // EOpAtomicCounterDecrement gets a post decrement, to map between GLSL -> SPIR-V + // return 3; // change/correct barrier-instruction operands, to match memory model group decisions + // return 4; // some deeper access chains: for dynamic vector component, and local Boolean component + // return 5; // make OpArrayLength result type be an int with signedness of 0 + // return 6; // revert version 5 change, which makes a different (new) kind of incorrect code, + // versions 4 and 6 each generate OpArrayLength as it has long been done + return 7; // GLSL volatile keyword maps to both SPIR-V decorations Volatile and Coherent +} + +// Write SPIR-V out to a binary file +void OutputSpvBin(const std::vector<unsigned int>& spirv, const char* baseName) +{ + std::ofstream out; + out.open(baseName, std::ios::binary | std::ios::out); + if (out.fail()) + printf("ERROR: Failed to open file: %s\n", baseName); + for (int i = 0; i < (int)spirv.size(); ++i) { + unsigned int word = spirv[i]; + out.write((const char*)&word, 4); + } + out.close(); +} + +// Write SPIR-V out to a text file with 32-bit hexadecimal words +void OutputSpvHex(const std::vector<unsigned int>& spirv, const char* baseName, const char* varName) +{ + std::ofstream out; + out.open(baseName, std::ios::binary | std::ios::out); + if (out.fail()) + printf("ERROR: Failed to open file: %s\n", baseName); + out << "\t// " << + GetSpirvGeneratorVersion() << "." << GLSLANG_MINOR_VERSION << "." << GLSLANG_PATCH_LEVEL << + std::endl; + if (varName != nullptr) { + out << "\t #pragma once" << std::endl; + out << "const uint32_t " << varName << "[] = {" << std::endl; + } + const int WORDS_PER_LINE = 8; + for (int i = 0; i < (int)spirv.size(); i += WORDS_PER_LINE) { + out << "\t"; + for (int j = 0; j < WORDS_PER_LINE && i + j < (int)spirv.size(); ++j) { + const unsigned int word = spirv[i + j]; + out << "0x" << std::hex << std::setw(8) << std::setfill('0') << word; + if (i + j + 1 < (int)spirv.size()) { + out << ","; + } + } + out << std::endl; + } + if (varName != nullptr) { + out << "};"; + } + out.close(); +} + +// +// Set up the glslang traversal +// +void GlslangToSpv(const TIntermediate& intermediate, std::vector<unsigned int>& spirv, SpvOptions* options) +{ + spv::SpvBuildLogger logger; + GlslangToSpv(intermediate, spirv, &logger, options); +} + +void GlslangToSpv(const TIntermediate& intermediate, std::vector<unsigned int>& spirv, + spv::SpvBuildLogger* logger, SpvOptions* options) +{ + TIntermNode* root = intermediate.getTreeRoot(); + + if (root == 0) + return; + + SpvOptions defaultOptions; + if (options == nullptr) + options = &defaultOptions; + + GetThreadPoolAllocator().push(); + + TGlslangToSpvTraverser it(intermediate.getSpv().spv, &intermediate, logger, *options); + root->traverse(&it); + it.finishSpv(); + it.dumpSpv(spirv); + +#if ENABLE_OPT + // If from HLSL, run spirv-opt to "legalize" the SPIR-V for Vulkan + // eg. forward and remove memory writes of opaque types. + if ((intermediate.getSource() == EShSourceHlsl || options->optimizeSize) && !options->disableOptimizer) + SpirvToolsLegalize(intermediate, spirv, logger, options); + + if (options->validate) + SpirvToolsValidate(intermediate, spirv, logger); + + if (options->disassemble) + SpirvToolsDisassemble(std::cout, spirv); + +#endif + + GetThreadPoolAllocator().pop(); +} + +}; // end namespace glslang diff --git a/thirdparty/glslang/SPIRV/GlslangToSpv.h b/thirdparty/glslang/SPIRV/GlslangToSpv.h new file mode 100644 index 0000000000..86e1c23bf6 --- /dev/null +++ b/thirdparty/glslang/SPIRV/GlslangToSpv.h @@ -0,0 +1,61 @@ +// +// Copyright (C) 2014 LunarG, Inc. +// Copyright (C) 2015-2018 Google, Inc. +// +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions +// are met: +// +// Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// +// Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following +// disclaimer in the documentation and/or other materials provided +// with the distribution. +// +// Neither the name of 3Dlabs Inc. Ltd. nor the names of its +// contributors may be used to endorse or promote products derived +// from this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS +// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE +// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, +// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, +// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; +// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN +// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +// POSSIBILITY OF SUCH DAMAGE. + +#pragma once + +#if defined(_MSC_VER) && _MSC_VER >= 1900 + #pragma warning(disable : 4464) // relative include path contains '..' +#endif + +#include "SpvTools.h" +#include "../glslang/Include/intermediate.h" + +#include <string> +#include <vector> + +#include "Logger.h" + +namespace glslang { + +void GetSpirvVersion(std::string&); +int GetSpirvGeneratorVersion(); +void GlslangToSpv(const glslang::TIntermediate& intermediate, std::vector<unsigned int>& spirv, + SpvOptions* options = nullptr); +void GlslangToSpv(const glslang::TIntermediate& intermediate, std::vector<unsigned int>& spirv, + spv::SpvBuildLogger* logger, SpvOptions* options = nullptr); +void OutputSpvBin(const std::vector<unsigned int>& spirv, const char* baseName); +void OutputSpvHex(const std::vector<unsigned int>& spirv, const char* baseName, const char* varName); + +} diff --git a/thirdparty/glslang/SPIRV/InReadableOrder.cpp b/thirdparty/glslang/SPIRV/InReadableOrder.cpp new file mode 100644 index 0000000000..52b29613a4 --- /dev/null +++ b/thirdparty/glslang/SPIRV/InReadableOrder.cpp @@ -0,0 +1,113 @@ +// +// Copyright (C) 2016 Google, Inc. +// +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions +// are met: +// +// Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// +// Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following +// disclaimer in the documentation and/or other materials provided +// with the distribution. +// +// Neither the name of 3Dlabs Inc. Ltd. nor the names of its +// contributors may be used to endorse or promote products derived +// from this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS +// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE +// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, +// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, +// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; +// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN +// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +// POSSIBILITY OF SUCH DAMAGE. + +// The SPIR-V spec requires code blocks to appear in an order satisfying the +// dominator-tree direction (ie, dominator before the dominated). This is, +// actually, easy to achieve: any pre-order CFG traversal algorithm will do it. +// Because such algorithms visit a block only after traversing some path to it +// from the root, they necessarily visit the block's idom first. +// +// But not every graph-traversal algorithm outputs blocks in an order that +// appears logical to human readers. The problem is that unrelated branches may +// be interspersed with each other, and merge blocks may come before some of the +// branches being merged. +// +// A good, human-readable order of blocks may be achieved by performing +// depth-first search but delaying merge nodes until after all their branches +// have been visited. This is implemented below by the inReadableOrder() +// function. + +#include "spvIR.h" + +#include <cassert> +#include <unordered_set> + +using spv::Block; +using spv::Id; + +namespace { +// Traverses CFG in a readable order, invoking a pre-set callback on each block. +// Use by calling visit() on the root block. +class ReadableOrderTraverser { +public: + explicit ReadableOrderTraverser(std::function<void(Block*)> callback) : callback_(callback) {} + // Visits the block if it hasn't been visited already and isn't currently + // being delayed. Invokes callback(block), then descends into its + // successors. Delays merge-block and continue-block processing until all + // the branches have been completed. + void visit(Block* block) + { + assert(block); + if (visited_.count(block) || delayed_.count(block)) + return; + callback_(block); + visited_.insert(block); + Block* mergeBlock = nullptr; + Block* continueBlock = nullptr; + auto mergeInst = block->getMergeInstruction(); + if (mergeInst) { + Id mergeId = mergeInst->getIdOperand(0); + mergeBlock = block->getParent().getParent().getInstruction(mergeId)->getBlock(); + delayed_.insert(mergeBlock); + if (mergeInst->getOpCode() == spv::OpLoopMerge) { + Id continueId = mergeInst->getIdOperand(1); + continueBlock = + block->getParent().getParent().getInstruction(continueId)->getBlock(); + delayed_.insert(continueBlock); + } + } + const auto successors = block->getSuccessors(); + for (auto it = successors.cbegin(); it != successors.cend(); ++it) + visit(*it); + if (continueBlock) { + delayed_.erase(continueBlock); + visit(continueBlock); + } + if (mergeBlock) { + delayed_.erase(mergeBlock); + visit(mergeBlock); + } + } + +private: + std::function<void(Block*)> callback_; + // Whether a block has already been visited or is being delayed. + std::unordered_set<Block *> visited_, delayed_; +}; +} + +void spv::inReadableOrder(Block* root, std::function<void(Block*)> callback) +{ + ReadableOrderTraverser(callback).visit(root); +} diff --git a/thirdparty/glslang/SPIRV/Logger.cpp b/thirdparty/glslang/SPIRV/Logger.cpp new file mode 100644 index 0000000000..48bd4e3ade --- /dev/null +++ b/thirdparty/glslang/SPIRV/Logger.cpp @@ -0,0 +1,68 @@ +// +// Copyright (C) 2016 Google, Inc. +// +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions +// are met: +// +// Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// +// Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following +// disclaimer in the documentation and/or other materials provided +// with the distribution. +// +// Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived +// from this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS +// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE +// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, +// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, +// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; +// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN +// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +// POSSIBILITY OF SUCH DAMAGE. + +#include "Logger.h" + +#include <algorithm> +#include <iterator> +#include <sstream> + +namespace spv { + +void SpvBuildLogger::tbdFunctionality(const std::string& f) +{ + if (std::find(std::begin(tbdFeatures), std::end(tbdFeatures), f) == std::end(tbdFeatures)) + tbdFeatures.push_back(f); +} + +void SpvBuildLogger::missingFunctionality(const std::string& f) +{ + if (std::find(std::begin(missingFeatures), std::end(missingFeatures), f) == std::end(missingFeatures)) + missingFeatures.push_back(f); +} + +std::string SpvBuildLogger::getAllMessages() const { + std::ostringstream messages; + for (auto it = tbdFeatures.cbegin(); it != tbdFeatures.cend(); ++it) + messages << "TBD functionality: " << *it << "\n"; + for (auto it = missingFeatures.cbegin(); it != missingFeatures.cend(); ++it) + messages << "Missing functionality: " << *it << "\n"; + for (auto it = warnings.cbegin(); it != warnings.cend(); ++it) + messages << "warning: " << *it << "\n"; + for (auto it = errors.cbegin(); it != errors.cend(); ++it) + messages << "error: " << *it << "\n"; + return messages.str(); +} + +} // end spv namespace diff --git a/thirdparty/glslang/SPIRV/Logger.h b/thirdparty/glslang/SPIRV/Logger.h new file mode 100644 index 0000000000..2e4ddaf517 --- /dev/null +++ b/thirdparty/glslang/SPIRV/Logger.h @@ -0,0 +1,74 @@ +// +// Copyright (C) 2016 Google, Inc. +// +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions +// are met: +// +// Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// +// Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following +// disclaimer in the documentation and/or other materials provided +// with the distribution. +// +// Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived +// from this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS +// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE +// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, +// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, +// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; +// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN +// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +// POSSIBILITY OF SUCH DAMAGE. + +#ifndef GLSLANG_SPIRV_LOGGER_H +#define GLSLANG_SPIRV_LOGGER_H + +#include <string> +#include <vector> + +namespace spv { + +// A class for holding all SPIR-V build status messages, including +// missing/TBD functionalities, warnings, and errors. +class SpvBuildLogger { +public: + SpvBuildLogger() {} + + // Registers a TBD functionality. + void tbdFunctionality(const std::string& f); + // Registers a missing functionality. + void missingFunctionality(const std::string& f); + + // Logs a warning. + void warning(const std::string& w) { warnings.push_back(w); } + // Logs an error. + void error(const std::string& e) { errors.push_back(e); } + + // Returns all messages accumulated in the order of: + // TBD functionalities, missing functionalities, warnings, errors. + std::string getAllMessages() const; + +private: + SpvBuildLogger(const SpvBuildLogger&); + + std::vector<std::string> tbdFeatures; + std::vector<std::string> missingFeatures; + std::vector<std::string> warnings; + std::vector<std::string> errors; +}; + +} // end spv namespace + +#endif // GLSLANG_SPIRV_LOGGER_H diff --git a/thirdparty/glslang/SPIRV/SPVRemapper.cpp b/thirdparty/glslang/SPIRV/SPVRemapper.cpp new file mode 100644 index 0000000000..fd0bb8950c --- /dev/null +++ b/thirdparty/glslang/SPIRV/SPVRemapper.cpp @@ -0,0 +1,1487 @@ +// +// Copyright (C) 2015 LunarG, Inc. +// +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions +// are met: +// +// Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// +// Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following +// disclaimer in the documentation and/or other materials provided +// with the distribution. +// +// Neither the name of 3Dlabs Inc. Ltd. nor the names of its +// contributors may be used to endorse or promote products derived +// from this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS +// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE +// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, +// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, +// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; +// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN +// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +// POSSIBILITY OF SUCH DAMAGE. +// + +#include "SPVRemapper.h" +#include "doc.h" + +#if !defined (use_cpp11) +// ... not supported before C++11 +#else // defined (use_cpp11) + +#include <algorithm> +#include <cassert> +#include "../glslang/Include/Common.h" + +namespace spv { + + // By default, just abort on error. Can be overridden via RegisterErrorHandler + spirvbin_t::errorfn_t spirvbin_t::errorHandler = [](const std::string&) { exit(5); }; + // By default, eat log messages. Can be overridden via RegisterLogHandler + spirvbin_t::logfn_t spirvbin_t::logHandler = [](const std::string&) { }; + + // This can be overridden to provide other message behavior if needed + void spirvbin_t::msg(int minVerbosity, int indent, const std::string& txt) const + { + if (verbose >= minVerbosity) + logHandler(std::string(indent, ' ') + txt); + } + + // hash opcode, with special handling for OpExtInst + std::uint32_t spirvbin_t::asOpCodeHash(unsigned word) + { + const spv::Op opCode = asOpCode(word); + + std::uint32_t offset = 0; + + switch (opCode) { + case spv::OpExtInst: + offset += asId(word + 4); break; + default: + break; + } + + return opCode * 19 + offset; // 19 = small prime + } + + spirvbin_t::range_t spirvbin_t::literalRange(spv::Op opCode) const + { + static const int maxCount = 1<<30; + + switch (opCode) { + case spv::OpTypeFloat: // fall through... + case spv::OpTypePointer: return range_t(2, 3); + case spv::OpTypeInt: return range_t(2, 4); + // TODO: case spv::OpTypeImage: + // TODO: case spv::OpTypeSampledImage: + case spv::OpTypeSampler: return range_t(3, 8); + case spv::OpTypeVector: // fall through + case spv::OpTypeMatrix: // ... + case spv::OpTypePipe: return range_t(3, 4); + case spv::OpConstant: return range_t(3, maxCount); + default: return range_t(0, 0); + } + } + + spirvbin_t::range_t spirvbin_t::typeRange(spv::Op opCode) const + { + static const int maxCount = 1<<30; + + if (isConstOp(opCode)) + return range_t(1, 2); + + switch (opCode) { + case spv::OpTypeVector: // fall through + case spv::OpTypeMatrix: // ... + case spv::OpTypeSampler: // ... + case spv::OpTypeArray: // ... + case spv::OpTypeRuntimeArray: // ... + case spv::OpTypePipe: return range_t(2, 3); + case spv::OpTypeStruct: // fall through + case spv::OpTypeFunction: return range_t(2, maxCount); + case spv::OpTypePointer: return range_t(3, 4); + default: return range_t(0, 0); + } + } + + spirvbin_t::range_t spirvbin_t::constRange(spv::Op opCode) const + { + static const int maxCount = 1<<30; + + switch (opCode) { + case spv::OpTypeArray: // fall through... + case spv::OpTypeRuntimeArray: return range_t(3, 4); + case spv::OpConstantComposite: return range_t(3, maxCount); + default: return range_t(0, 0); + } + } + + // Return the size of a type in 32-bit words. This currently only + // handles ints and floats, and is only invoked by queries which must be + // integer types. If ever needed, it can be generalized. + unsigned spirvbin_t::typeSizeInWords(spv::Id id) const + { + const unsigned typeStart = idPos(id); + const spv::Op opCode = asOpCode(typeStart); + + if (errorLatch) + return 0; + + switch (opCode) { + case spv::OpTypeInt: // fall through... + case spv::OpTypeFloat: return (spv[typeStart+2]+31)/32; + default: + return 0; + } + } + + // Looks up the type of a given const or variable ID, and + // returns its size in 32-bit words. + unsigned spirvbin_t::idTypeSizeInWords(spv::Id id) const + { + const auto tid_it = idTypeSizeMap.find(id); + if (tid_it == idTypeSizeMap.end()) { + error("type size for ID not found"); + return 0; + } + + return tid_it->second; + } + + // Is this an opcode we should remove when using --strip? + bool spirvbin_t::isStripOp(spv::Op opCode) const + { + switch (opCode) { + case spv::OpSource: + case spv::OpSourceExtension: + case spv::OpName: + case spv::OpMemberName: + case spv::OpLine: return true; + default: return false; + } + } + + // Return true if this opcode is flow control + bool spirvbin_t::isFlowCtrl(spv::Op opCode) const + { + switch (opCode) { + case spv::OpBranchConditional: + case spv::OpBranch: + case spv::OpSwitch: + case spv::OpLoopMerge: + case spv::OpSelectionMerge: + case spv::OpLabel: + case spv::OpFunction: + case spv::OpFunctionEnd: return true; + default: return false; + } + } + + // Return true if this opcode defines a type + bool spirvbin_t::isTypeOp(spv::Op opCode) const + { + switch (opCode) { + case spv::OpTypeVoid: + case spv::OpTypeBool: + case spv::OpTypeInt: + case spv::OpTypeFloat: + case spv::OpTypeVector: + case spv::OpTypeMatrix: + case spv::OpTypeImage: + case spv::OpTypeSampler: + case spv::OpTypeArray: + case spv::OpTypeRuntimeArray: + case spv::OpTypeStruct: + case spv::OpTypeOpaque: + case spv::OpTypePointer: + case spv::OpTypeFunction: + case spv::OpTypeEvent: + case spv::OpTypeDeviceEvent: + case spv::OpTypeReserveId: + case spv::OpTypeQueue: + case spv::OpTypeSampledImage: + case spv::OpTypePipe: return true; + default: return false; + } + } + + // Return true if this opcode defines a constant + bool spirvbin_t::isConstOp(spv::Op opCode) const + { + switch (opCode) { + case spv::OpConstantSampler: + error("unimplemented constant type"); + return true; + + case spv::OpConstantNull: + case spv::OpConstantTrue: + case spv::OpConstantFalse: + case spv::OpConstantComposite: + case spv::OpConstant: + return true; + + default: + return false; + } + } + + const auto inst_fn_nop = [](spv::Op, unsigned) { return false; }; + const auto op_fn_nop = [](spv::Id&) { }; + + // g++ doesn't like these defined in the class proper in an anonymous namespace. + // Dunno why. Also MSVC doesn't like the constexpr keyword. Also dunno why. + // Defining them externally seems to please both compilers, so, here they are. + const spv::Id spirvbin_t::unmapped = spv::Id(-10000); + const spv::Id spirvbin_t::unused = spv::Id(-10001); + const int spirvbin_t::header_size = 5; + + spv::Id spirvbin_t::nextUnusedId(spv::Id id) + { + while (isNewIdMapped(id)) // search for an unused ID + ++id; + + return id; + } + + spv::Id spirvbin_t::localId(spv::Id id, spv::Id newId) + { + //assert(id != spv::NoResult && newId != spv::NoResult); + + if (id > bound()) { + error(std::string("ID out of range: ") + std::to_string(id)); + return spirvbin_t::unused; + } + + if (id >= idMapL.size()) + idMapL.resize(id+1, unused); + + if (newId != unmapped && newId != unused) { + if (isOldIdUnused(id)) { + error(std::string("ID unused in module: ") + std::to_string(id)); + return spirvbin_t::unused; + } + + if (!isOldIdUnmapped(id)) { + error(std::string("ID already mapped: ") + std::to_string(id) + " -> " + + std::to_string(localId(id))); + + return spirvbin_t::unused; + } + + if (isNewIdMapped(newId)) { + error(std::string("ID already used in module: ") + std::to_string(newId)); + return spirvbin_t::unused; + } + + msg(4, 4, std::string("map: ") + std::to_string(id) + " -> " + std::to_string(newId)); + setMapped(newId); + largestNewId = std::max(largestNewId, newId); + } + + return idMapL[id] = newId; + } + + // Parse a literal string from the SPIR binary and return it as an std::string + // Due to C++11 RValue references, this doesn't copy the result string. + std::string spirvbin_t::literalString(unsigned word) const + { + std::string literal; + + literal.reserve(16); + + const char* bytes = reinterpret_cast<const char*>(spv.data() + word); + + while (bytes && *bytes) + literal += *bytes++; + + return literal; + } + + void spirvbin_t::applyMap() + { + msg(3, 2, std::string("Applying map: ")); + + // Map local IDs through the ID map + process(inst_fn_nop, // ignore instructions + [this](spv::Id& id) { + id = localId(id); + + if (errorLatch) + return; + + assert(id != unused && id != unmapped); + } + ); + } + + // Find free IDs for anything we haven't mapped + void spirvbin_t::mapRemainder() + { + msg(3, 2, std::string("Remapping remainder: ")); + + spv::Id unusedId = 1; // can't use 0: that's NoResult + spirword_t maxBound = 0; + + for (spv::Id id = 0; id < idMapL.size(); ++id) { + if (isOldIdUnused(id)) + continue; + + // Find a new mapping for any used but unmapped IDs + if (isOldIdUnmapped(id)) { + localId(id, unusedId = nextUnusedId(unusedId)); + if (errorLatch) + return; + } + + if (isOldIdUnmapped(id)) { + error(std::string("old ID not mapped: ") + std::to_string(id)); + return; + } + + // Track max bound + maxBound = std::max(maxBound, localId(id) + 1); + + if (errorLatch) + return; + } + + bound(maxBound); // reset header ID bound to as big as it now needs to be + } + + // Mark debug instructions for stripping + void spirvbin_t::stripDebug() + { + // Strip instructions in the stripOp set: debug info. + process( + [&](spv::Op opCode, unsigned start) { + // remember opcodes we want to strip later + if (isStripOp(opCode)) + stripInst(start); + return true; + }, + op_fn_nop); + } + + // Mark instructions that refer to now-removed IDs for stripping + void spirvbin_t::stripDeadRefs() + { + process( + [&](spv::Op opCode, unsigned start) { + // strip opcodes pointing to removed data + switch (opCode) { + case spv::OpName: + case spv::OpMemberName: + case spv::OpDecorate: + case spv::OpMemberDecorate: + if (idPosR.find(asId(start+1)) == idPosR.end()) + stripInst(start); + break; + default: + break; // leave it alone + } + + return true; + }, + op_fn_nop); + + strip(); + } + + // Update local maps of ID, type, etc positions + void spirvbin_t::buildLocalMaps() + { + msg(2, 2, std::string("build local maps: ")); + + mapped.clear(); + idMapL.clear(); +// preserve nameMap, so we don't clear that. + fnPos.clear(); + fnCalls.clear(); + typeConstPos.clear(); + idPosR.clear(); + entryPoint = spv::NoResult; + largestNewId = 0; + + idMapL.resize(bound(), unused); + + int fnStart = 0; + spv::Id fnRes = spv::NoResult; + + // build local Id and name maps + process( + [&](spv::Op opCode, unsigned start) { + unsigned word = start+1; + spv::Id typeId = spv::NoResult; + + if (spv::InstructionDesc[opCode].hasType()) + typeId = asId(word++); + + // If there's a result ID, remember the size of its type + if (spv::InstructionDesc[opCode].hasResult()) { + const spv::Id resultId = asId(word++); + idPosR[resultId] = start; + + if (typeId != spv::NoResult) { + const unsigned idTypeSize = typeSizeInWords(typeId); + + if (errorLatch) + return false; + + if (idTypeSize != 0) + idTypeSizeMap[resultId] = idTypeSize; + } + } + + if (opCode == spv::Op::OpName) { + const spv::Id target = asId(start+1); + const std::string name = literalString(start+2); + nameMap[name] = target; + + } else if (opCode == spv::Op::OpFunctionCall) { + ++fnCalls[asId(start + 3)]; + } else if (opCode == spv::Op::OpEntryPoint) { + entryPoint = asId(start + 2); + } else if (opCode == spv::Op::OpFunction) { + if (fnStart != 0) { + error("nested function found"); + return false; + } + + fnStart = start; + fnRes = asId(start + 2); + } else if (opCode == spv::Op::OpFunctionEnd) { + assert(fnRes != spv::NoResult); + if (fnStart == 0) { + error("function end without function start"); + return false; + } + + fnPos[fnRes] = range_t(fnStart, start + asWordCount(start)); + fnStart = 0; + } else if (isConstOp(opCode)) { + if (errorLatch) + return false; + + assert(asId(start + 2) != spv::NoResult); + typeConstPos.insert(start); + } else if (isTypeOp(opCode)) { + assert(asId(start + 1) != spv::NoResult); + typeConstPos.insert(start); + } + + return false; + }, + + [this](spv::Id& id) { localId(id, unmapped); } + ); + } + + // Validate the SPIR header + void spirvbin_t::validate() const + { + msg(2, 2, std::string("validating: ")); + + if (spv.size() < header_size) { + error("file too short: "); + return; + } + + if (magic() != spv::MagicNumber) { + error("bad magic number"); + return; + } + + // field 1 = version + // field 2 = generator magic + // field 3 = result <id> bound + + if (schemaNum() != 0) { + error("bad schema, must be 0"); + return; + } + } + + int spirvbin_t::processInstruction(unsigned word, instfn_t instFn, idfn_t idFn) + { + const auto instructionStart = word; + const unsigned wordCount = asWordCount(instructionStart); + const int nextInst = word++ + wordCount; + spv::Op opCode = asOpCode(instructionStart); + + if (nextInst > int(spv.size())) { + error("spir instruction terminated too early"); + return -1; + } + + // Base for computing number of operands; will be updated as more is learned + unsigned numOperands = wordCount - 1; + + if (instFn(opCode, instructionStart)) + return nextInst; + + // Read type and result ID from instruction desc table + if (spv::InstructionDesc[opCode].hasType()) { + idFn(asId(word++)); + --numOperands; + } + + if (spv::InstructionDesc[opCode].hasResult()) { + idFn(asId(word++)); + --numOperands; + } + + // Extended instructions: currently, assume everything is an ID. + // TODO: add whatever data we need for exceptions to that + if (opCode == spv::OpExtInst) { + word += 2; // instruction set, and instruction from set + numOperands -= 2; + + for (unsigned op=0; op < numOperands; ++op) + idFn(asId(word++)); // ID + + return nextInst; + } + + // Circular buffer so we can look back at previous unmapped values during the mapping pass. + static const unsigned idBufferSize = 4; + spv::Id idBuffer[idBufferSize]; + unsigned idBufferPos = 0; + + // Store IDs from instruction in our map + for (int op = 0; numOperands > 0; ++op, --numOperands) { + // SpecConstantOp is special: it includes the operands of another opcode which is + // given as a literal in the 3rd word. We will switch over to pretending that the + // opcode being processed is the literal opcode value of the SpecConstantOp. See the + // SPIRV spec for details. This way we will handle IDs and literals as appropriate for + // the embedded op. + if (opCode == spv::OpSpecConstantOp) { + if (op == 0) { + opCode = asOpCode(word++); // this is the opcode embedded in the SpecConstantOp. + --numOperands; + } + } + + switch (spv::InstructionDesc[opCode].operands.getClass(op)) { + case spv::OperandId: + case spv::OperandScope: + case spv::OperandMemorySemantics: + idBuffer[idBufferPos] = asId(word); + idBufferPos = (idBufferPos + 1) % idBufferSize; + idFn(asId(word++)); + break; + + case spv::OperandVariableIds: + for (unsigned i = 0; i < numOperands; ++i) + idFn(asId(word++)); + return nextInst; + + case spv::OperandVariableLiterals: + // for clarity + // if (opCode == spv::OpDecorate && asDecoration(word - 1) == spv::DecorationBuiltIn) { + // ++word; + // --numOperands; + // } + // word += numOperands; + return nextInst; + + case spv::OperandVariableLiteralId: { + if (opCode == OpSwitch) { + // word-2 is the position of the selector ID. OpSwitch Literals match its type. + // In case the IDs are currently being remapped, we get the word[-2] ID from + // the circular idBuffer. + const unsigned literalSizePos = (idBufferPos+idBufferSize-2) % idBufferSize; + const unsigned literalSize = idTypeSizeInWords(idBuffer[literalSizePos]); + const unsigned numLiteralIdPairs = (nextInst-word) / (1+literalSize); + + if (errorLatch) + return -1; + + for (unsigned arg=0; arg<numLiteralIdPairs; ++arg) { + word += literalSize; // literal + idFn(asId(word++)); // label + } + } else { + assert(0); // currentely, only OpSwitch uses OperandVariableLiteralId + } + + return nextInst; + } + + case spv::OperandLiteralString: { + const int stringWordCount = literalStringWords(literalString(word)); + word += stringWordCount; + numOperands -= (stringWordCount-1); // -1 because for() header post-decrements + break; + } + + // Execution mode might have extra literal operands. Skip them. + case spv::OperandExecutionMode: + return nextInst; + + // Single word operands we simply ignore, as they hold no IDs + case spv::OperandLiteralNumber: + case spv::OperandSource: + case spv::OperandExecutionModel: + case spv::OperandAddressing: + case spv::OperandMemory: + case spv::OperandStorage: + case spv::OperandDimensionality: + case spv::OperandSamplerAddressingMode: + case spv::OperandSamplerFilterMode: + case spv::OperandSamplerImageFormat: + case spv::OperandImageChannelOrder: + case spv::OperandImageChannelDataType: + case spv::OperandImageOperands: + case spv::OperandFPFastMath: + case spv::OperandFPRoundingMode: + case spv::OperandLinkageType: + case spv::OperandAccessQualifier: + case spv::OperandFuncParamAttr: + case spv::OperandDecoration: + case spv::OperandBuiltIn: + case spv::OperandSelect: + case spv::OperandLoop: + case spv::OperandFunction: + case spv::OperandMemoryAccess: + case spv::OperandGroupOperation: + case spv::OperandKernelEnqueueFlags: + case spv::OperandKernelProfilingInfo: + case spv::OperandCapability: + ++word; + break; + + default: + assert(0 && "Unhandled Operand Class"); + break; + } + } + + return nextInst; + } + + // Make a pass over all the instructions and process them given appropriate functions + spirvbin_t& spirvbin_t::process(instfn_t instFn, idfn_t idFn, unsigned begin, unsigned end) + { + // For efficiency, reserve name map space. It can grow if needed. + nameMap.reserve(32); + + // If begin or end == 0, use defaults + begin = (begin == 0 ? header_size : begin); + end = (end == 0 ? unsigned(spv.size()) : end); + + // basic parsing and InstructionDesc table borrowed from SpvDisassemble.cpp... + unsigned nextInst = unsigned(spv.size()); + + for (unsigned word = begin; word < end; word = nextInst) { + nextInst = processInstruction(word, instFn, idFn); + + if (errorLatch) + return *this; + } + + return *this; + } + + // Apply global name mapping to a single module + void spirvbin_t::mapNames() + { + static const std::uint32_t softTypeIdLimit = 3011; // small prime. TODO: get from options + static const std::uint32_t firstMappedID = 3019; // offset into ID space + + for (const auto& name : nameMap) { + std::uint32_t hashval = 1911; + for (const char c : name.first) + hashval = hashval * 1009 + c; + + if (isOldIdUnmapped(name.second)) { + localId(name.second, nextUnusedId(hashval % softTypeIdLimit + firstMappedID)); + if (errorLatch) + return; + } + } + } + + // Map fn contents to IDs of similar functions in other modules + void spirvbin_t::mapFnBodies() + { + static const std::uint32_t softTypeIdLimit = 19071; // small prime. TODO: get from options + static const std::uint32_t firstMappedID = 6203; // offset into ID space + + // Initial approach: go through some high priority opcodes first and assign them + // hash values. + + spv::Id fnId = spv::NoResult; + std::vector<unsigned> instPos; + instPos.reserve(unsigned(spv.size()) / 16); // initial estimate; can grow if needed. + + // Build local table of instruction start positions + process( + [&](spv::Op, unsigned start) { instPos.push_back(start); return true; }, + op_fn_nop); + + if (errorLatch) + return; + + // Window size for context-sensitive canonicalization values + // Empirical best size from a single data set. TODO: Would be a good tunable. + // We essentially perform a little convolution around each instruction, + // to capture the flavor of nearby code, to hopefully match to similar + // code in other modules. + static const unsigned windowSize = 2; + + for (unsigned entry = 0; entry < unsigned(instPos.size()); ++entry) { + const unsigned start = instPos[entry]; + const spv::Op opCode = asOpCode(start); + + if (opCode == spv::OpFunction) + fnId = asId(start + 2); + + if (opCode == spv::OpFunctionEnd) + fnId = spv::NoResult; + + if (fnId != spv::NoResult) { // if inside a function + if (spv::InstructionDesc[opCode].hasResult()) { + const unsigned word = start + (spv::InstructionDesc[opCode].hasType() ? 2 : 1); + const spv::Id resId = asId(word); + std::uint32_t hashval = fnId * 17; // small prime + + for (unsigned i = entry-1; i >= entry-windowSize; --i) { + if (asOpCode(instPos[i]) == spv::OpFunction) + break; + hashval = hashval * 30103 + asOpCodeHash(instPos[i]); // 30103 = semiarbitrary prime + } + + for (unsigned i = entry; i <= entry + windowSize; ++i) { + if (asOpCode(instPos[i]) == spv::OpFunctionEnd) + break; + hashval = hashval * 30103 + asOpCodeHash(instPos[i]); // 30103 = semiarbitrary prime + } + + if (isOldIdUnmapped(resId)) { + localId(resId, nextUnusedId(hashval % softTypeIdLimit + firstMappedID)); + if (errorLatch) + return; + } + + } + } + } + + spv::Op thisOpCode(spv::OpNop); + std::unordered_map<int, int> opCounter; + int idCounter(0); + fnId = spv::NoResult; + + process( + [&](spv::Op opCode, unsigned start) { + switch (opCode) { + case spv::OpFunction: + // Reset counters at each function + idCounter = 0; + opCounter.clear(); + fnId = asId(start + 2); + break; + + case spv::OpImageSampleImplicitLod: + case spv::OpImageSampleExplicitLod: + case spv::OpImageSampleDrefImplicitLod: + case spv::OpImageSampleDrefExplicitLod: + case spv::OpImageSampleProjImplicitLod: + case spv::OpImageSampleProjExplicitLod: + case spv::OpImageSampleProjDrefImplicitLod: + case spv::OpImageSampleProjDrefExplicitLod: + case spv::OpDot: + case spv::OpCompositeExtract: + case spv::OpCompositeInsert: + case spv::OpVectorShuffle: + case spv::OpLabel: + case spv::OpVariable: + + case spv::OpAccessChain: + case spv::OpLoad: + case spv::OpStore: + case spv::OpCompositeConstruct: + case spv::OpFunctionCall: + ++opCounter[opCode]; + idCounter = 0; + thisOpCode = opCode; + break; + default: + thisOpCode = spv::OpNop; + } + + return false; + }, + + [&](spv::Id& id) { + if (thisOpCode != spv::OpNop) { + ++idCounter; + const std::uint32_t hashval = opCounter[thisOpCode] * thisOpCode * 50047 + idCounter + fnId * 117; + + if (isOldIdUnmapped(id)) + localId(id, nextUnusedId(hashval % softTypeIdLimit + firstMappedID)); + } + }); + } + + // EXPERIMENTAL: forward IO and uniform load/stores into operands + // This produces invalid Schema-0 SPIRV + void spirvbin_t::forwardLoadStores() + { + idset_t fnLocalVars; // set of function local vars + idmap_t idMap; // Map of load result IDs to what they load + + // EXPERIMENTAL: Forward input and access chain loads into consumptions + process( + [&](spv::Op opCode, unsigned start) { + // Add inputs and uniforms to the map + if ((opCode == spv::OpVariable && asWordCount(start) == 4) && + (spv[start+3] == spv::StorageClassUniform || + spv[start+3] == spv::StorageClassUniformConstant || + spv[start+3] == spv::StorageClassInput)) + fnLocalVars.insert(asId(start+2)); + + if (opCode == spv::OpAccessChain && fnLocalVars.count(asId(start+3)) > 0) + fnLocalVars.insert(asId(start+2)); + + if (opCode == spv::OpLoad && fnLocalVars.count(asId(start+3)) > 0) { + idMap[asId(start+2)] = asId(start+3); + stripInst(start); + } + + return false; + }, + + [&](spv::Id& id) { if (idMap.find(id) != idMap.end()) id = idMap[id]; } + ); + + if (errorLatch) + return; + + // EXPERIMENTAL: Implicit output stores + fnLocalVars.clear(); + idMap.clear(); + + process( + [&](spv::Op opCode, unsigned start) { + // Add inputs and uniforms to the map + if ((opCode == spv::OpVariable && asWordCount(start) == 4) && + (spv[start+3] == spv::StorageClassOutput)) + fnLocalVars.insert(asId(start+2)); + + if (opCode == spv::OpStore && fnLocalVars.count(asId(start+1)) > 0) { + idMap[asId(start+2)] = asId(start+1); + stripInst(start); + } + + return false; + }, + op_fn_nop); + + if (errorLatch) + return; + + process( + inst_fn_nop, + [&](spv::Id& id) { if (idMap.find(id) != idMap.end()) id = idMap[id]; } + ); + + if (errorLatch) + return; + + strip(); // strip out data we decided to eliminate + } + + // optimize loads and stores + void spirvbin_t::optLoadStore() + { + idset_t fnLocalVars; // candidates for removal (only locals) + idmap_t idMap; // Map of load result IDs to what they load + blockmap_t blockMap; // Map of IDs to blocks they first appear in + int blockNum = 0; // block count, to avoid crossing flow control + + // Find all the function local pointers stored at most once, and not via access chains + process( + [&](spv::Op opCode, unsigned start) { + const int wordCount = asWordCount(start); + + // Count blocks, so we can avoid crossing flow control + if (isFlowCtrl(opCode)) + ++blockNum; + + // Add local variables to the map + if ((opCode == spv::OpVariable && spv[start+3] == spv::StorageClassFunction && asWordCount(start) == 4)) { + fnLocalVars.insert(asId(start+2)); + return true; + } + + // Ignore process vars referenced via access chain + if ((opCode == spv::OpAccessChain || opCode == spv::OpInBoundsAccessChain) && fnLocalVars.count(asId(start+3)) > 0) { + fnLocalVars.erase(asId(start+3)); + idMap.erase(asId(start+3)); + return true; + } + + if (opCode == spv::OpLoad && fnLocalVars.count(asId(start+3)) > 0) { + const spv::Id varId = asId(start+3); + + // Avoid loads before stores + if (idMap.find(varId) == idMap.end()) { + fnLocalVars.erase(varId); + idMap.erase(varId); + } + + // don't do for volatile references + if (wordCount > 4 && (spv[start+4] & spv::MemoryAccessVolatileMask)) { + fnLocalVars.erase(varId); + idMap.erase(varId); + } + + // Handle flow control + if (blockMap.find(varId) == blockMap.end()) { + blockMap[varId] = blockNum; // track block we found it in. + } else if (blockMap[varId] != blockNum) { + fnLocalVars.erase(varId); // Ignore if crosses flow control + idMap.erase(varId); + } + + return true; + } + + if (opCode == spv::OpStore && fnLocalVars.count(asId(start+1)) > 0) { + const spv::Id varId = asId(start+1); + + if (idMap.find(varId) == idMap.end()) { + idMap[varId] = asId(start+2); + } else { + // Remove if it has more than one store to the same pointer + fnLocalVars.erase(varId); + idMap.erase(varId); + } + + // don't do for volatile references + if (wordCount > 3 && (spv[start+3] & spv::MemoryAccessVolatileMask)) { + fnLocalVars.erase(asId(start+3)); + idMap.erase(asId(start+3)); + } + + // Handle flow control + if (blockMap.find(varId) == blockMap.end()) { + blockMap[varId] = blockNum; // track block we found it in. + } else if (blockMap[varId] != blockNum) { + fnLocalVars.erase(varId); // Ignore if crosses flow control + idMap.erase(varId); + } + + return true; + } + + return false; + }, + + // If local var id used anywhere else, don't eliminate + [&](spv::Id& id) { + if (fnLocalVars.count(id) > 0) { + fnLocalVars.erase(id); + idMap.erase(id); + } + } + ); + + if (errorLatch) + return; + + process( + [&](spv::Op opCode, unsigned start) { + if (opCode == spv::OpLoad && fnLocalVars.count(asId(start+3)) > 0) + idMap[asId(start+2)] = idMap[asId(start+3)]; + return false; + }, + op_fn_nop); + + if (errorLatch) + return; + + // Chase replacements to their origins, in case there is a chain such as: + // 2 = store 1 + // 3 = load 2 + // 4 = store 3 + // 5 = load 4 + // We want to replace uses of 5 with 1. + for (const auto& idPair : idMap) { + spv::Id id = idPair.first; + while (idMap.find(id) != idMap.end()) // Chase to end of chain + id = idMap[id]; + + idMap[idPair.first] = id; // replace with final result + } + + // Remove the load/store/variables for the ones we've discovered + process( + [&](spv::Op opCode, unsigned start) { + if ((opCode == spv::OpLoad && fnLocalVars.count(asId(start+3)) > 0) || + (opCode == spv::OpStore && fnLocalVars.count(asId(start+1)) > 0) || + (opCode == spv::OpVariable && fnLocalVars.count(asId(start+2)) > 0)) { + + stripInst(start); + return true; + } + + return false; + }, + + [&](spv::Id& id) { + if (idMap.find(id) != idMap.end()) id = idMap[id]; + } + ); + + if (errorLatch) + return; + + strip(); // strip out data we decided to eliminate + } + + // remove bodies of uncalled functions + void spirvbin_t::dceFuncs() + { + msg(3, 2, std::string("Removing Dead Functions: ")); + + // TODO: There are more efficient ways to do this. + bool changed = true; + + while (changed) { + changed = false; + + for (auto fn = fnPos.begin(); fn != fnPos.end(); ) { + if (fn->first == entryPoint) { // don't DCE away the entry point! + ++fn; + continue; + } + + const auto call_it = fnCalls.find(fn->first); + + if (call_it == fnCalls.end() || call_it->second == 0) { + changed = true; + stripRange.push_back(fn->second); + + // decrease counts of called functions + process( + [&](spv::Op opCode, unsigned start) { + if (opCode == spv::Op::OpFunctionCall) { + const auto call_it = fnCalls.find(asId(start + 3)); + if (call_it != fnCalls.end()) { + if (--call_it->second <= 0) + fnCalls.erase(call_it); + } + } + + return true; + }, + op_fn_nop, + fn->second.first, + fn->second.second); + + if (errorLatch) + return; + + fn = fnPos.erase(fn); + } else ++fn; + } + } + } + + // remove unused function variables + decorations + void spirvbin_t::dceVars() + { + msg(3, 2, std::string("DCE Vars: ")); + + std::unordered_map<spv::Id, int> varUseCount; + + // Count function variable use + process( + [&](spv::Op opCode, unsigned start) { + if (opCode == spv::OpVariable) { + ++varUseCount[asId(start+2)]; + return true; + } else if (opCode == spv::OpEntryPoint) { + const int wordCount = asWordCount(start); + for (int i = 4; i < wordCount; i++) { + ++varUseCount[asId(start+i)]; + } + return true; + } else + return false; + }, + + [&](spv::Id& id) { if (varUseCount[id]) ++varUseCount[id]; } + ); + + if (errorLatch) + return; + + // Remove single-use function variables + associated decorations and names + process( + [&](spv::Op opCode, unsigned start) { + spv::Id id = spv::NoResult; + if (opCode == spv::OpVariable) + id = asId(start+2); + if (opCode == spv::OpDecorate || opCode == spv::OpName) + id = asId(start+1); + + if (id != spv::NoResult && varUseCount[id] == 1) + stripInst(start); + + return true; + }, + op_fn_nop); + } + + // remove unused types + void spirvbin_t::dceTypes() + { + std::vector<bool> isType(bound(), false); + + // for speed, make O(1) way to get to type query (map is log(n)) + for (const auto typeStart : typeConstPos) + isType[asTypeConstId(typeStart)] = true; + + std::unordered_map<spv::Id, int> typeUseCount; + + // This is not the most efficient algorithm, but this is an offline tool, and + // it's easy to write this way. Can be improved opportunistically if needed. + bool changed = true; + while (changed) { + changed = false; + strip(); + typeUseCount.clear(); + + // Count total type usage + process(inst_fn_nop, + [&](spv::Id& id) { if (isType[id]) ++typeUseCount[id]; } + ); + + if (errorLatch) + return; + + // Remove single reference types + for (const auto typeStart : typeConstPos) { + const spv::Id typeId = asTypeConstId(typeStart); + if (typeUseCount[typeId] == 1) { + changed = true; + --typeUseCount[typeId]; + stripInst(typeStart); + } + } + + if (errorLatch) + return; + } + } + +#ifdef NOTDEF + bool spirvbin_t::matchType(const spirvbin_t::globaltypes_t& globalTypes, spv::Id lt, spv::Id gt) const + { + // Find the local type id "lt" and global type id "gt" + const auto lt_it = typeConstPosR.find(lt); + if (lt_it == typeConstPosR.end()) + return false; + + const auto typeStart = lt_it->second; + + // Search for entry in global table + const auto gtype = globalTypes.find(gt); + if (gtype == globalTypes.end()) + return false; + + const auto& gdata = gtype->second; + + // local wordcount and opcode + const int wordCount = asWordCount(typeStart); + const spv::Op opCode = asOpCode(typeStart); + + // no type match if opcodes don't match, or operand count doesn't match + if (opCode != opOpCode(gdata[0]) || wordCount != opWordCount(gdata[0])) + return false; + + const unsigned numOperands = wordCount - 2; // all types have a result + + const auto cmpIdRange = [&](range_t range) { + for (int x=range.first; x<std::min(range.second, wordCount); ++x) + if (!matchType(globalTypes, asId(typeStart+x), gdata[x])) + return false; + return true; + }; + + const auto cmpConst = [&]() { return cmpIdRange(constRange(opCode)); }; + const auto cmpSubType = [&]() { return cmpIdRange(typeRange(opCode)); }; + + // Compare literals in range [start,end) + const auto cmpLiteral = [&]() { + const auto range = literalRange(opCode); + return std::equal(spir.begin() + typeStart + range.first, + spir.begin() + typeStart + std::min(range.second, wordCount), + gdata.begin() + range.first); + }; + + assert(isTypeOp(opCode) || isConstOp(opCode)); + + switch (opCode) { + case spv::OpTypeOpaque: // TODO: disable until we compare the literal strings. + case spv::OpTypeQueue: return false; + case spv::OpTypeEvent: // fall through... + case spv::OpTypeDeviceEvent: // ... + case spv::OpTypeReserveId: return false; + // for samplers, we don't handle the optional parameters yet + case spv::OpTypeSampler: return cmpLiteral() && cmpConst() && cmpSubType() && wordCount == 8; + default: return cmpLiteral() && cmpConst() && cmpSubType(); + } + } + + // Look for an equivalent type in the globalTypes map + spv::Id spirvbin_t::findType(const spirvbin_t::globaltypes_t& globalTypes, spv::Id lt) const + { + // Try a recursive type match on each in turn, and return a match if we find one + for (const auto& gt : globalTypes) + if (matchType(globalTypes, lt, gt.first)) + return gt.first; + + return spv::NoType; + } +#endif // NOTDEF + + // Return start position in SPV of given Id. error if not found. + unsigned spirvbin_t::idPos(spv::Id id) const + { + const auto tid_it = idPosR.find(id); + if (tid_it == idPosR.end()) { + error("ID not found"); + return 0; + } + + return tid_it->second; + } + + // Hash types to canonical values. This can return ID collisions (it's a bit + // inevitable): it's up to the caller to handle that gracefully. + std::uint32_t spirvbin_t::hashType(unsigned typeStart) const + { + const unsigned wordCount = asWordCount(typeStart); + const spv::Op opCode = asOpCode(typeStart); + + switch (opCode) { + case spv::OpTypeVoid: return 0; + case spv::OpTypeBool: return 1; + case spv::OpTypeInt: return 3 + (spv[typeStart+3]); + case spv::OpTypeFloat: return 5; + case spv::OpTypeVector: + return 6 + hashType(idPos(spv[typeStart+2])) * (spv[typeStart+3] - 1); + case spv::OpTypeMatrix: + return 30 + hashType(idPos(spv[typeStart+2])) * (spv[typeStart+3] - 1); + case spv::OpTypeImage: + return 120 + hashType(idPos(spv[typeStart+2])) + + spv[typeStart+3] + // dimensionality + spv[typeStart+4] * 8 * 16 + // depth + spv[typeStart+5] * 4 * 16 + // arrayed + spv[typeStart+6] * 2 * 16 + // multisampled + spv[typeStart+7] * 1 * 16; // format + case spv::OpTypeSampler: + return 500; + case spv::OpTypeSampledImage: + return 502; + case spv::OpTypeArray: + return 501 + hashType(idPos(spv[typeStart+2])) * spv[typeStart+3]; + case spv::OpTypeRuntimeArray: + return 5000 + hashType(idPos(spv[typeStart+2])); + case spv::OpTypeStruct: + { + std::uint32_t hash = 10000; + for (unsigned w=2; w < wordCount; ++w) + hash += w * hashType(idPos(spv[typeStart+w])); + return hash; + } + + case spv::OpTypeOpaque: return 6000 + spv[typeStart+2]; + case spv::OpTypePointer: return 100000 + hashType(idPos(spv[typeStart+3])); + case spv::OpTypeFunction: + { + std::uint32_t hash = 200000; + for (unsigned w=2; w < wordCount; ++w) + hash += w * hashType(idPos(spv[typeStart+w])); + return hash; + } + + case spv::OpTypeEvent: return 300000; + case spv::OpTypeDeviceEvent: return 300001; + case spv::OpTypeReserveId: return 300002; + case spv::OpTypeQueue: return 300003; + case spv::OpTypePipe: return 300004; + case spv::OpConstantTrue: return 300007; + case spv::OpConstantFalse: return 300008; + case spv::OpConstantComposite: + { + std::uint32_t hash = 300011 + hashType(idPos(spv[typeStart+1])); + for (unsigned w=3; w < wordCount; ++w) + hash += w * hashType(idPos(spv[typeStart+w])); + return hash; + } + case spv::OpConstant: + { + std::uint32_t hash = 400011 + hashType(idPos(spv[typeStart+1])); + for (unsigned w=3; w < wordCount; ++w) + hash += w * spv[typeStart+w]; + return hash; + } + case spv::OpConstantNull: + { + std::uint32_t hash = 500009 + hashType(idPos(spv[typeStart+1])); + return hash; + } + case spv::OpConstantSampler: + { + std::uint32_t hash = 600011 + hashType(idPos(spv[typeStart+1])); + for (unsigned w=3; w < wordCount; ++w) + hash += w * spv[typeStart+w]; + return hash; + } + + default: + error("unknown type opcode"); + return 0; + } + } + + void spirvbin_t::mapTypeConst() + { + globaltypes_t globalTypeMap; + + msg(3, 2, std::string("Remapping Consts & Types: ")); + + static const std::uint32_t softTypeIdLimit = 3011; // small prime. TODO: get from options + static const std::uint32_t firstMappedID = 8; // offset into ID space + + for (auto& typeStart : typeConstPos) { + const spv::Id resId = asTypeConstId(typeStart); + const std::uint32_t hashval = hashType(typeStart); + + if (errorLatch) + return; + + if (isOldIdUnmapped(resId)) { + localId(resId, nextUnusedId(hashval % softTypeIdLimit + firstMappedID)); + if (errorLatch) + return; + } + } + } + + // Strip a single binary by removing ranges given in stripRange + void spirvbin_t::strip() + { + if (stripRange.empty()) // nothing to do + return; + + // Sort strip ranges in order of traversal + std::sort(stripRange.begin(), stripRange.end()); + + // Allocate a new binary big enough to hold old binary + // We'll step this iterator through the strip ranges as we go through the binary + auto strip_it = stripRange.begin(); + + int strippedPos = 0; + for (unsigned word = 0; word < unsigned(spv.size()); ++word) { + while (strip_it != stripRange.end() && word >= strip_it->second) + ++strip_it; + + if (strip_it == stripRange.end() || word < strip_it->first || word >= strip_it->second) + spv[strippedPos++] = spv[word]; + } + + spv.resize(strippedPos); + stripRange.clear(); + + buildLocalMaps(); + } + + // Strip a single binary by removing ranges given in stripRange + void spirvbin_t::remap(std::uint32_t opts) + { + options = opts; + + // Set up opcode tables from SpvDoc + spv::Parameterize(); + + validate(); // validate header + buildLocalMaps(); // build ID maps + + msg(3, 4, std::string("ID bound: ") + std::to_string(bound())); + + if (options & STRIP) stripDebug(); + if (errorLatch) return; + + strip(); // strip out data we decided to eliminate + if (errorLatch) return; + + if (options & OPT_LOADSTORE) optLoadStore(); + if (errorLatch) return; + + if (options & OPT_FWD_LS) forwardLoadStores(); + if (errorLatch) return; + + if (options & DCE_FUNCS) dceFuncs(); + if (errorLatch) return; + + if (options & DCE_VARS) dceVars(); + if (errorLatch) return; + + if (options & DCE_TYPES) dceTypes(); + if (errorLatch) return; + + strip(); // strip out data we decided to eliminate + if (errorLatch) return; + + stripDeadRefs(); // remove references to things we DCEed + if (errorLatch) return; + + // after the last strip, we must clean any debug info referring to now-deleted data + + if (options & MAP_TYPES) mapTypeConst(); + if (errorLatch) return; + + if (options & MAP_NAMES) mapNames(); + if (errorLatch) return; + + if (options & MAP_FUNCS) mapFnBodies(); + if (errorLatch) return; + + if (options & MAP_ALL) { + mapRemainder(); // map any unmapped IDs + if (errorLatch) return; + + applyMap(); // Now remap each shader to the new IDs we've come up with + if (errorLatch) return; + } + } + + // remap from a memory image + void spirvbin_t::remap(std::vector<std::uint32_t>& in_spv, std::uint32_t opts) + { + spv.swap(in_spv); + remap(opts); + spv.swap(in_spv); + } + +} // namespace SPV + +#endif // defined (use_cpp11) + diff --git a/thirdparty/glslang/SPIRV/SPVRemapper.h b/thirdparty/glslang/SPIRV/SPVRemapper.h new file mode 100644 index 0000000000..fa61bb94d8 --- /dev/null +++ b/thirdparty/glslang/SPIRV/SPVRemapper.h @@ -0,0 +1,304 @@ +// +// Copyright (C) 2015 LunarG, Inc. +// +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions +// are met: +// +// Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// +// Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following +// disclaimer in the documentation and/or other materials provided +// with the distribution. +// +// Neither the name of 3Dlabs Inc. Ltd. nor the names of its +// contributors may be used to endorse or promote products derived +// from this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS +// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE +// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, +// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, +// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; +// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN +// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +// POSSIBILITY OF SUCH DAMAGE. +// + +#ifndef SPIRVREMAPPER_H +#define SPIRVREMAPPER_H + +#include <string> +#include <vector> +#include <cstdlib> +#include <exception> + +namespace spv { + +// MSVC defines __cplusplus as an older value, even when it supports almost all of 11. +// We handle that here by making our own symbol. +#if __cplusplus >= 201103L || (defined(_MSC_VER) && _MSC_VER >= 1700) +# define use_cpp11 1 +#endif + +class spirvbin_base_t +{ +public: + enum Options { + NONE = 0, + STRIP = (1<<0), + MAP_TYPES = (1<<1), + MAP_NAMES = (1<<2), + MAP_FUNCS = (1<<3), + DCE_FUNCS = (1<<4), + DCE_VARS = (1<<5), + DCE_TYPES = (1<<6), + OPT_LOADSTORE = (1<<7), + OPT_FWD_LS = (1<<8), // EXPERIMENTAL: PRODUCES INVALID SCHEMA-0 SPIRV + MAP_ALL = (MAP_TYPES | MAP_NAMES | MAP_FUNCS), + DCE_ALL = (DCE_FUNCS | DCE_VARS | DCE_TYPES), + OPT_ALL = (OPT_LOADSTORE), + + ALL_BUT_STRIP = (MAP_ALL | DCE_ALL | OPT_ALL), + DO_EVERYTHING = (STRIP | ALL_BUT_STRIP) + }; +}; + +} // namespace SPV + +#if !defined (use_cpp11) +#include <cstdio> +#include <cstdint> + +namespace spv { +class spirvbin_t : public spirvbin_base_t +{ +public: + spirvbin_t(int /*verbose = 0*/) { } + + void remap(std::vector<std::uint32_t>& /*spv*/, unsigned int /*opts = 0*/) + { + printf("Tool not compiled for C++11, which is required for SPIR-V remapping.\n"); + exit(5); + } +}; + +} // namespace SPV + +#else // defined (use_cpp11) + +#include <functional> +#include <cstdint> +#include <unordered_map> +#include <unordered_set> +#include <map> +#include <set> +#include <cassert> + +#include "spirv.hpp" +#include "spvIR.h" + +namespace spv { + +// class to hold SPIR-V binary data for remapping, DCE, and debug stripping +class spirvbin_t : public spirvbin_base_t +{ +public: + spirvbin_t(int verbose = 0) : entryPoint(spv::NoResult), largestNewId(0), verbose(verbose), errorLatch(false) + { } + + virtual ~spirvbin_t() { } + + // remap on an existing binary in memory + void remap(std::vector<std::uint32_t>& spv, std::uint32_t opts = DO_EVERYTHING); + + // Type for error/log handler functions + typedef std::function<void(const std::string&)> errorfn_t; + typedef std::function<void(const std::string&)> logfn_t; + + // Register error/log handling functions (can be lambda fn / functor / etc) + static void registerErrorHandler(errorfn_t handler) { errorHandler = handler; } + static void registerLogHandler(logfn_t handler) { logHandler = handler; } + +protected: + // This can be overridden to provide other message behavior if needed + virtual void msg(int minVerbosity, int indent, const std::string& txt) const; + +private: + // Local to global, or global to local ID map + typedef std::unordered_map<spv::Id, spv::Id> idmap_t; + typedef std::unordered_set<spv::Id> idset_t; + typedef std::unordered_map<spv::Id, int> blockmap_t; + + void remap(std::uint32_t opts = DO_EVERYTHING); + + // Map of names to IDs + typedef std::unordered_map<std::string, spv::Id> namemap_t; + + typedef std::uint32_t spirword_t; + + typedef std::pair<unsigned, unsigned> range_t; + typedef std::function<void(spv::Id&)> idfn_t; + typedef std::function<bool(spv::Op, unsigned start)> instfn_t; + + // Special Values for ID map: + static const spv::Id unmapped; // unchanged from default value + static const spv::Id unused; // unused ID + static const int header_size; // SPIR header = 5 words + + class id_iterator_t; + + // For mapping type entries between different shaders + typedef std::vector<spirword_t> typeentry_t; + typedef std::map<spv::Id, typeentry_t> globaltypes_t; + + // A set that preserves position order, and a reverse map + typedef std::set<int> posmap_t; + typedef std::unordered_map<spv::Id, int> posmap_rev_t; + + // Maps and ID to the size of its base type, if known. + typedef std::unordered_map<spv::Id, unsigned> typesize_map_t; + + // handle error + void error(const std::string& txt) const { errorLatch = true; errorHandler(txt); } + + bool isConstOp(spv::Op opCode) const; + bool isTypeOp(spv::Op opCode) const; + bool isStripOp(spv::Op opCode) const; + bool isFlowCtrl(spv::Op opCode) const; + range_t literalRange(spv::Op opCode) const; + range_t typeRange(spv::Op opCode) const; + range_t constRange(spv::Op opCode) const; + unsigned typeSizeInWords(spv::Id id) const; + unsigned idTypeSizeInWords(spv::Id id) const; + + spv::Id& asId(unsigned word) { return spv[word]; } + const spv::Id& asId(unsigned word) const { return spv[word]; } + spv::Op asOpCode(unsigned word) const { return opOpCode(spv[word]); } + std::uint32_t asOpCodeHash(unsigned word); + spv::Decoration asDecoration(unsigned word) const { return spv::Decoration(spv[word]); } + unsigned asWordCount(unsigned word) const { return opWordCount(spv[word]); } + spv::Id asTypeConstId(unsigned word) const { return asId(word + (isTypeOp(asOpCode(word)) ? 1 : 2)); } + unsigned idPos(spv::Id id) const; + + static unsigned opWordCount(spirword_t data) { return data >> spv::WordCountShift; } + static spv::Op opOpCode(spirword_t data) { return spv::Op(data & spv::OpCodeMask); } + + // Header access & set methods + spirword_t magic() const { return spv[0]; } // return magic number + spirword_t bound() const { return spv[3]; } // return Id bound from header + spirword_t bound(spirword_t b) { return spv[3] = b; }; + spirword_t genmagic() const { return spv[2]; } // generator magic + spirword_t genmagic(spirword_t m) { return spv[2] = m; } + spirword_t schemaNum() const { return spv[4]; } // schema number from header + + // Mapping fns: get + spv::Id localId(spv::Id id) const { return idMapL[id]; } + + // Mapping fns: set + inline spv::Id localId(spv::Id id, spv::Id newId); + void countIds(spv::Id id); + + // Return next unused new local ID. + // NOTE: boost::dynamic_bitset would be more efficient due to find_next(), + // which std::vector<bool> doens't have. + inline spv::Id nextUnusedId(spv::Id id); + + void buildLocalMaps(); + std::string literalString(unsigned word) const; // Return literal as a std::string + int literalStringWords(const std::string& str) const { return (int(str.size())+4)/4; } + + bool isNewIdMapped(spv::Id newId) const { return isMapped(newId); } + bool isOldIdUnmapped(spv::Id oldId) const { return localId(oldId) == unmapped; } + bool isOldIdUnused(spv::Id oldId) const { return localId(oldId) == unused; } + bool isOldIdMapped(spv::Id oldId) const { return !isOldIdUnused(oldId) && !isOldIdUnmapped(oldId); } + bool isFunction(spv::Id oldId) const { return fnPos.find(oldId) != fnPos.end(); } + + // bool matchType(const globaltypes_t& globalTypes, spv::Id lt, spv::Id gt) const; + // spv::Id findType(const globaltypes_t& globalTypes, spv::Id lt) const; + std::uint32_t hashType(unsigned typeStart) const; + + spirvbin_t& process(instfn_t, idfn_t, unsigned begin = 0, unsigned end = 0); + int processInstruction(unsigned word, instfn_t, idfn_t); + + void validate() const; + void mapTypeConst(); + void mapFnBodies(); + void optLoadStore(); + void dceFuncs(); + void dceVars(); + void dceTypes(); + void mapNames(); + void foldIds(); // fold IDs to smallest space + void forwardLoadStores(); // load store forwarding (EXPERIMENTAL) + void offsetIds(); // create relative offset IDs + + void applyMap(); // remap per local name map + void mapRemainder(); // map any IDs we haven't touched yet + void stripDebug(); // strip all debug info + void stripDeadRefs(); // strips debug info for now-dead references after DCE + void strip(); // remove debug symbols + + std::vector<spirword_t> spv; // SPIR words + + namemap_t nameMap; // ID names from OpName + + // Since we want to also do binary ops, we can't use std::vector<bool>. we could use + // boost::dynamic_bitset, but we're trying to avoid a boost dependency. + typedef std::uint64_t bits_t; + std::vector<bits_t> mapped; // which new IDs have been mapped + static const int mBits = sizeof(bits_t) * 4; + + bool isMapped(spv::Id id) const { return id < maxMappedId() && ((mapped[id/mBits] & (1LL<<(id%mBits))) != 0); } + void setMapped(spv::Id id) { resizeMapped(id); mapped[id/mBits] |= (1LL<<(id%mBits)); } + void resizeMapped(spv::Id id) { if (id >= maxMappedId()) mapped.resize(id/mBits+1, 0); } + size_t maxMappedId() const { return mapped.size() * mBits; } + + // Add a strip range for a given instruction starting at 'start' + // Note: avoiding brace initializers to please older versions os MSVC. + void stripInst(unsigned start) { stripRange.push_back(range_t(start, start + asWordCount(start))); } + + // Function start and end. use unordered_map because we'll have + // many fewer functions than IDs. + std::unordered_map<spv::Id, range_t> fnPos; + + // Which functions are called, anywhere in the module, with a call count + std::unordered_map<spv::Id, int> fnCalls; + + posmap_t typeConstPos; // word positions that define types & consts (ordered) + posmap_rev_t idPosR; // reverse map from IDs to positions + typesize_map_t idTypeSizeMap; // maps each ID to its type size, if known. + + std::vector<spv::Id> idMapL; // ID {M}ap from {L}ocal to {G}lobal IDs + + spv::Id entryPoint; // module entry point + spv::Id largestNewId; // biggest new ID we have mapped anything to + + // Sections of the binary to strip, given as [begin,end) + std::vector<range_t> stripRange; + + // processing options: + std::uint32_t options; + int verbose; // verbosity level + + // Error latch: this is set if the error handler is ever executed. It would be better to + // use a try/catch block and throw, but that's not desired for certain environments, so + // this is the alternative. + mutable bool errorLatch; + + static errorfn_t errorHandler; + static logfn_t logHandler; +}; + +} // namespace SPV + +#endif // defined (use_cpp11) +#endif // SPIRVREMAPPER_H diff --git a/thirdparty/glslang/SPIRV/SpvBuilder.cpp b/thirdparty/glslang/SPIRV/SpvBuilder.cpp new file mode 100644 index 0000000000..4ef7e5fe7f --- /dev/null +++ b/thirdparty/glslang/SPIRV/SpvBuilder.cpp @@ -0,0 +1,3058 @@ +// +// Copyright (C) 2014-2015 LunarG, Inc. +// Copyright (C) 2015-2018 Google, Inc. +// +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions +// are met: +// +// Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// +// Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following +// disclaimer in the documentation and/or other materials provided +// with the distribution. +// +// Neither the name of 3Dlabs Inc. Ltd. nor the names of its +// contributors may be used to endorse or promote products derived +// from this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS +// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE +// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, +// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, +// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; +// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN +// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +// POSSIBILITY OF SUCH DAMAGE. + +// +// Helper for making SPIR-V IR. Generally, this is documented in the header +// SpvBuilder.h. +// + +#include <cassert> +#include <cstdlib> + +#include <unordered_set> +#include <algorithm> + +#include "SpvBuilder.h" + +#include "hex_float.h" + +#ifndef _WIN32 + #include <cstdio> +#endif + +namespace spv { + +Builder::Builder(unsigned int spvVersion, unsigned int magicNumber, SpvBuildLogger* buildLogger) : + spvVersion(spvVersion), + source(SourceLanguageUnknown), + sourceVersion(0), + sourceFileStringId(NoResult), + currentLine(0), + currentFile(nullptr), + emitOpLines(false), + addressModel(AddressingModelLogical), + memoryModel(MemoryModelGLSL450), + builderNumber(magicNumber), + buildPoint(0), + uniqueId(0), + entryPointFunction(0), + generatingOpCodeForSpecConst(false), + logger(buildLogger) +{ + clearAccessChain(); +} + +Builder::~Builder() +{ +} + +Id Builder::import(const char* name) +{ + Instruction* import = new Instruction(getUniqueId(), NoType, OpExtInstImport); + import->addStringOperand(name); + module.mapInstruction(import); + + imports.push_back(std::unique_ptr<Instruction>(import)); + return import->getResultId(); +} + +// Emit instruction for non-filename-based #line directives (ie. no filename +// seen yet): emit an OpLine if we've been asked to emit OpLines and the line +// number has changed since the last time, and is a valid line number. +void Builder::setLine(int lineNum) +{ + if (lineNum != 0 && lineNum != currentLine) { + currentLine = lineNum; + if (emitOpLines) + addLine(sourceFileStringId, currentLine, 0); + } +} + +// If no filename, do non-filename-based #line emit. Else do filename-based emit. +// Emit OpLine if we've been asked to emit OpLines and the line number or filename +// has changed since the last time, and line number is valid. +void Builder::setLine(int lineNum, const char* filename) +{ + if (filename == nullptr) { + setLine(lineNum); + return; + } + if ((lineNum != 0 && lineNum != currentLine) || currentFile == nullptr || + strncmp(filename, currentFile, strlen(currentFile) + 1) != 0) { + currentLine = lineNum; + currentFile = filename; + if (emitOpLines) { + spv::Id strId = getStringId(filename); + addLine(strId, currentLine, 0); + } + } +} + +void Builder::addLine(Id fileName, int lineNum, int column) +{ + Instruction* line = new Instruction(OpLine); + line->addIdOperand(fileName); + line->addImmediateOperand(lineNum); + line->addImmediateOperand(column); + buildPoint->addInstruction(std::unique_ptr<Instruction>(line)); +} + +// For creating new groupedTypes (will return old type if the requested one was already made). +Id Builder::makeVoidType() +{ + Instruction* type; + if (groupedTypes[OpTypeVoid].size() == 0) { + type = new Instruction(getUniqueId(), NoType, OpTypeVoid); + groupedTypes[OpTypeVoid].push_back(type); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(type)); + module.mapInstruction(type); + } else + type = groupedTypes[OpTypeVoid].back(); + + return type->getResultId(); +} + +Id Builder::makeBoolType() +{ + Instruction* type; + if (groupedTypes[OpTypeBool].size() == 0) { + type = new Instruction(getUniqueId(), NoType, OpTypeBool); + groupedTypes[OpTypeBool].push_back(type); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(type)); + module.mapInstruction(type); + } else + type = groupedTypes[OpTypeBool].back(); + + return type->getResultId(); +} + +Id Builder::makeSamplerType() +{ + Instruction* type; + if (groupedTypes[OpTypeSampler].size() == 0) { + type = new Instruction(getUniqueId(), NoType, OpTypeSampler); + groupedTypes[OpTypeSampler].push_back(type); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(type)); + module.mapInstruction(type); + } else + type = groupedTypes[OpTypeSampler].back(); + + return type->getResultId(); +} + +Id Builder::makePointer(StorageClass storageClass, Id pointee) +{ + // try to find it + Instruction* type; + for (int t = 0; t < (int)groupedTypes[OpTypePointer].size(); ++t) { + type = groupedTypes[OpTypePointer][t]; + if (type->getImmediateOperand(0) == (unsigned)storageClass && + type->getIdOperand(1) == pointee) + return type->getResultId(); + } + + // not found, make it + type = new Instruction(getUniqueId(), NoType, OpTypePointer); + type->addImmediateOperand(storageClass); + type->addIdOperand(pointee); + groupedTypes[OpTypePointer].push_back(type); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(type)); + module.mapInstruction(type); + + return type->getResultId(); +} + +Id Builder::makeForwardPointer(StorageClass storageClass) +{ + // Caching/uniquifying doesn't work here, because we don't know the + // pointee type and there can be multiple forward pointers of the same + // storage type. Somebody higher up in the stack must keep track. + Instruction* type = new Instruction(getUniqueId(), NoType, OpTypeForwardPointer); + type->addImmediateOperand(storageClass); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(type)); + module.mapInstruction(type); + + return type->getResultId(); +} + +Id Builder::makePointerFromForwardPointer(StorageClass storageClass, Id forwardPointerType, Id pointee) +{ + // try to find it + Instruction* type; + for (int t = 0; t < (int)groupedTypes[OpTypePointer].size(); ++t) { + type = groupedTypes[OpTypePointer][t]; + if (type->getImmediateOperand(0) == (unsigned)storageClass && + type->getIdOperand(1) == pointee) + return type->getResultId(); + } + + type = new Instruction(forwardPointerType, NoType, OpTypePointer); + type->addImmediateOperand(storageClass); + type->addIdOperand(pointee); + groupedTypes[OpTypePointer].push_back(type); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(type)); + module.mapInstruction(type); + + return type->getResultId(); +} + +Id Builder::makeIntegerType(int width, bool hasSign) +{ + // try to find it + Instruction* type; + for (int t = 0; t < (int)groupedTypes[OpTypeInt].size(); ++t) { + type = groupedTypes[OpTypeInt][t]; + if (type->getImmediateOperand(0) == (unsigned)width && + type->getImmediateOperand(1) == (hasSign ? 1u : 0u)) + return type->getResultId(); + } + + // not found, make it + type = new Instruction(getUniqueId(), NoType, OpTypeInt); + type->addImmediateOperand(width); + type->addImmediateOperand(hasSign ? 1 : 0); + groupedTypes[OpTypeInt].push_back(type); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(type)); + module.mapInstruction(type); + + // deal with capabilities + switch (width) { + case 8: + case 16: + // these are currently handled by storage-type declarations and post processing + break; + case 64: + addCapability(CapabilityInt64); + break; + default: + break; + } + + return type->getResultId(); +} + +Id Builder::makeFloatType(int width) +{ + // try to find it + Instruction* type; + for (int t = 0; t < (int)groupedTypes[OpTypeFloat].size(); ++t) { + type = groupedTypes[OpTypeFloat][t]; + if (type->getImmediateOperand(0) == (unsigned)width) + return type->getResultId(); + } + + // not found, make it + type = new Instruction(getUniqueId(), NoType, OpTypeFloat); + type->addImmediateOperand(width); + groupedTypes[OpTypeFloat].push_back(type); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(type)); + module.mapInstruction(type); + + // deal with capabilities + switch (width) { + case 16: + // currently handled by storage-type declarations and post processing + break; + case 64: + addCapability(CapabilityFloat64); + break; + default: + break; + } + + return type->getResultId(); +} + +// Make a struct without checking for duplication. +// See makeStructResultType() for non-decorated structs +// needed as the result of some instructions, which does +// check for duplicates. +Id Builder::makeStructType(const std::vector<Id>& members, const char* name) +{ + // Don't look for previous one, because in the general case, + // structs can be duplicated except for decorations. + + // not found, make it + Instruction* type = new Instruction(getUniqueId(), NoType, OpTypeStruct); + for (int op = 0; op < (int)members.size(); ++op) + type->addIdOperand(members[op]); + groupedTypes[OpTypeStruct].push_back(type); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(type)); + module.mapInstruction(type); + addName(type->getResultId(), name); + + return type->getResultId(); +} + +// Make a struct for the simple results of several instructions, +// checking for duplication. +Id Builder::makeStructResultType(Id type0, Id type1) +{ + // try to find it + Instruction* type; + for (int t = 0; t < (int)groupedTypes[OpTypeStruct].size(); ++t) { + type = groupedTypes[OpTypeStruct][t]; + if (type->getNumOperands() != 2) + continue; + if (type->getIdOperand(0) != type0 || + type->getIdOperand(1) != type1) + continue; + return type->getResultId(); + } + + // not found, make it + std::vector<spv::Id> members; + members.push_back(type0); + members.push_back(type1); + + return makeStructType(members, "ResType"); +} + +Id Builder::makeVectorType(Id component, int size) +{ + // try to find it + Instruction* type; + for (int t = 0; t < (int)groupedTypes[OpTypeVector].size(); ++t) { + type = groupedTypes[OpTypeVector][t]; + if (type->getIdOperand(0) == component && + type->getImmediateOperand(1) == (unsigned)size) + return type->getResultId(); + } + + // not found, make it + type = new Instruction(getUniqueId(), NoType, OpTypeVector); + type->addIdOperand(component); + type->addImmediateOperand(size); + groupedTypes[OpTypeVector].push_back(type); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(type)); + module.mapInstruction(type); + + return type->getResultId(); +} + +Id Builder::makeMatrixType(Id component, int cols, int rows) +{ + assert(cols <= maxMatrixSize && rows <= maxMatrixSize); + + Id column = makeVectorType(component, rows); + + // try to find it + Instruction* type; + for (int t = 0; t < (int)groupedTypes[OpTypeMatrix].size(); ++t) { + type = groupedTypes[OpTypeMatrix][t]; + if (type->getIdOperand(0) == column && + type->getImmediateOperand(1) == (unsigned)cols) + return type->getResultId(); + } + + // not found, make it + type = new Instruction(getUniqueId(), NoType, OpTypeMatrix); + type->addIdOperand(column); + type->addImmediateOperand(cols); + groupedTypes[OpTypeMatrix].push_back(type); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(type)); + module.mapInstruction(type); + + return type->getResultId(); +} + +Id Builder::makeCooperativeMatrixType(Id component, Id scope, Id rows, Id cols) +{ + // try to find it + Instruction* type; + for (int t = 0; t < (int)groupedTypes[OpTypeCooperativeMatrixNV].size(); ++t) { + type = groupedTypes[OpTypeCooperativeMatrixNV][t]; + if (type->getIdOperand(0) == component && + type->getIdOperand(1) == scope && + type->getIdOperand(2) == rows && + type->getIdOperand(3) == cols) + return type->getResultId(); + } + + // not found, make it + type = new Instruction(getUniqueId(), NoType, OpTypeCooperativeMatrixNV); + type->addIdOperand(component); + type->addIdOperand(scope); + type->addIdOperand(rows); + type->addIdOperand(cols); + groupedTypes[OpTypeCooperativeMatrixNV].push_back(type); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(type)); + module.mapInstruction(type); + + return type->getResultId(); +} + + +// TODO: performance: track arrays per stride +// If a stride is supplied (non-zero) make an array. +// If no stride (0), reuse previous array types. +// 'size' is an Id of a constant or specialization constant of the array size +Id Builder::makeArrayType(Id element, Id sizeId, int stride) +{ + Instruction* type; + if (stride == 0) { + // try to find existing type + for (int t = 0; t < (int)groupedTypes[OpTypeArray].size(); ++t) { + type = groupedTypes[OpTypeArray][t]; + if (type->getIdOperand(0) == element && + type->getIdOperand(1) == sizeId) + return type->getResultId(); + } + } + + // not found, make it + type = new Instruction(getUniqueId(), NoType, OpTypeArray); + type->addIdOperand(element); + type->addIdOperand(sizeId); + groupedTypes[OpTypeArray].push_back(type); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(type)); + module.mapInstruction(type); + + return type->getResultId(); +} + +Id Builder::makeRuntimeArray(Id element) +{ + Instruction* type = new Instruction(getUniqueId(), NoType, OpTypeRuntimeArray); + type->addIdOperand(element); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(type)); + module.mapInstruction(type); + + return type->getResultId(); +} + +Id Builder::makeFunctionType(Id returnType, const std::vector<Id>& paramTypes) +{ + // try to find it + Instruction* type; + for (int t = 0; t < (int)groupedTypes[OpTypeFunction].size(); ++t) { + type = groupedTypes[OpTypeFunction][t]; + if (type->getIdOperand(0) != returnType || (int)paramTypes.size() != type->getNumOperands() - 1) + continue; + bool mismatch = false; + for (int p = 0; p < (int)paramTypes.size(); ++p) { + if (paramTypes[p] != type->getIdOperand(p + 1)) { + mismatch = true; + break; + } + } + if (! mismatch) + return type->getResultId(); + } + + // not found, make it + type = new Instruction(getUniqueId(), NoType, OpTypeFunction); + type->addIdOperand(returnType); + for (int p = 0; p < (int)paramTypes.size(); ++p) + type->addIdOperand(paramTypes[p]); + groupedTypes[OpTypeFunction].push_back(type); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(type)); + module.mapInstruction(type); + + return type->getResultId(); +} + +Id Builder::makeImageType(Id sampledType, Dim dim, bool depth, bool arrayed, bool ms, unsigned sampled, ImageFormat format) +{ + assert(sampled == 1 || sampled == 2); + + // try to find it + Instruction* type; + for (int t = 0; t < (int)groupedTypes[OpTypeImage].size(); ++t) { + type = groupedTypes[OpTypeImage][t]; + if (type->getIdOperand(0) == sampledType && + type->getImmediateOperand(1) == (unsigned int)dim && + type->getImmediateOperand(2) == ( depth ? 1u : 0u) && + type->getImmediateOperand(3) == (arrayed ? 1u : 0u) && + type->getImmediateOperand(4) == ( ms ? 1u : 0u) && + type->getImmediateOperand(5) == sampled && + type->getImmediateOperand(6) == (unsigned int)format) + return type->getResultId(); + } + + // not found, make it + type = new Instruction(getUniqueId(), NoType, OpTypeImage); + type->addIdOperand(sampledType); + type->addImmediateOperand( dim); + type->addImmediateOperand( depth ? 1 : 0); + type->addImmediateOperand(arrayed ? 1 : 0); + type->addImmediateOperand( ms ? 1 : 0); + type->addImmediateOperand(sampled); + type->addImmediateOperand((unsigned int)format); + + groupedTypes[OpTypeImage].push_back(type); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(type)); + module.mapInstruction(type); + + // deal with capabilities + switch (dim) { + case DimBuffer: + if (sampled == 1) + addCapability(CapabilitySampledBuffer); + else + addCapability(CapabilityImageBuffer); + break; + case Dim1D: + if (sampled == 1) + addCapability(CapabilitySampled1D); + else + addCapability(CapabilityImage1D); + break; + case DimCube: + if (arrayed) { + if (sampled == 1) + addCapability(CapabilitySampledCubeArray); + else + addCapability(CapabilityImageCubeArray); + } + break; + case DimRect: + if (sampled == 1) + addCapability(CapabilitySampledRect); + else + addCapability(CapabilityImageRect); + break; + case DimSubpassData: + addCapability(CapabilityInputAttachment); + break; + default: + break; + } + + if (ms) { + if (sampled == 2) { + // Images used with subpass data are not storage + // images, so don't require the capability for them. + if (dim != Dim::DimSubpassData) + addCapability(CapabilityStorageImageMultisample); + if (arrayed) + addCapability(CapabilityImageMSArray); + } + } + + return type->getResultId(); +} + +Id Builder::makeSampledImageType(Id imageType) +{ + // try to find it + Instruction* type; + for (int t = 0; t < (int)groupedTypes[OpTypeSampledImage].size(); ++t) { + type = groupedTypes[OpTypeSampledImage][t]; + if (type->getIdOperand(0) == imageType) + return type->getResultId(); + } + + // not found, make it + type = new Instruction(getUniqueId(), NoType, OpTypeSampledImage); + type->addIdOperand(imageType); + + groupedTypes[OpTypeSampledImage].push_back(type); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(type)); + module.mapInstruction(type); + + return type->getResultId(); +} + +#ifdef NV_EXTENSIONS +Id Builder::makeAccelerationStructureNVType() +{ + Instruction *type; + if (groupedTypes[OpTypeAccelerationStructureNV].size() == 0) { + type = new Instruction(getUniqueId(), NoType, OpTypeAccelerationStructureNV); + groupedTypes[OpTypeAccelerationStructureNV].push_back(type); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(type)); + module.mapInstruction(type); + } else { + type = groupedTypes[OpTypeAccelerationStructureNV].back(); + } + + return type->getResultId(); +} +#endif +Id Builder::getDerefTypeId(Id resultId) const +{ + Id typeId = getTypeId(resultId); + assert(isPointerType(typeId)); + + return module.getInstruction(typeId)->getIdOperand(1); +} + +Op Builder::getMostBasicTypeClass(Id typeId) const +{ + Instruction* instr = module.getInstruction(typeId); + + Op typeClass = instr->getOpCode(); + switch (typeClass) + { + case OpTypeVector: + case OpTypeMatrix: + case OpTypeArray: + case OpTypeRuntimeArray: + return getMostBasicTypeClass(instr->getIdOperand(0)); + case OpTypePointer: + return getMostBasicTypeClass(instr->getIdOperand(1)); + default: + return typeClass; + } +} + +int Builder::getNumTypeConstituents(Id typeId) const +{ + Instruction* instr = module.getInstruction(typeId); + + switch (instr->getOpCode()) + { + case OpTypeBool: + case OpTypeInt: + case OpTypeFloat: + case OpTypePointer: + return 1; + case OpTypeVector: + case OpTypeMatrix: + return instr->getImmediateOperand(1); + case OpTypeArray: + { + Id lengthId = instr->getIdOperand(1); + return module.getInstruction(lengthId)->getImmediateOperand(0); + } + case OpTypeStruct: + return instr->getNumOperands(); + case OpTypeCooperativeMatrixNV: + // has only one constituent when used with OpCompositeConstruct. + return 1; + default: + assert(0); + return 1; + } +} + +// Return the lowest-level type of scalar that an homogeneous composite is made out of. +// Typically, this is just to find out if something is made out of ints or floats. +// However, it includes returning a structure, if say, it is an array of structure. +Id Builder::getScalarTypeId(Id typeId) const +{ + Instruction* instr = module.getInstruction(typeId); + + Op typeClass = instr->getOpCode(); + switch (typeClass) + { + case OpTypeVoid: + case OpTypeBool: + case OpTypeInt: + case OpTypeFloat: + case OpTypeStruct: + return instr->getResultId(); + case OpTypeVector: + case OpTypeMatrix: + case OpTypeArray: + case OpTypeRuntimeArray: + case OpTypePointer: + return getScalarTypeId(getContainedTypeId(typeId)); + default: + assert(0); + return NoResult; + } +} + +// Return the type of 'member' of a composite. +Id Builder::getContainedTypeId(Id typeId, int member) const +{ + Instruction* instr = module.getInstruction(typeId); + + Op typeClass = instr->getOpCode(); + switch (typeClass) + { + case OpTypeVector: + case OpTypeMatrix: + case OpTypeArray: + case OpTypeRuntimeArray: + case OpTypeCooperativeMatrixNV: + return instr->getIdOperand(0); + case OpTypePointer: + return instr->getIdOperand(1); + case OpTypeStruct: + return instr->getIdOperand(member); + default: + assert(0); + return NoResult; + } +} + +// Return the immediately contained type of a given composite type. +Id Builder::getContainedTypeId(Id typeId) const +{ + return getContainedTypeId(typeId, 0); +} + +// Returns true if 'typeId' is or contains a scalar type declared with 'typeOp' +// of width 'width'. The 'width' is only consumed for int and float types. +// Returns false otherwise. +bool Builder::containsType(Id typeId, spv::Op typeOp, unsigned int width) const +{ + const Instruction& instr = *module.getInstruction(typeId); + + Op typeClass = instr.getOpCode(); + switch (typeClass) + { + case OpTypeInt: + case OpTypeFloat: + return typeClass == typeOp && instr.getImmediateOperand(0) == width; + case OpTypeStruct: + for (int m = 0; m < instr.getNumOperands(); ++m) { + if (containsType(instr.getIdOperand(m), typeOp, width)) + return true; + } + return false; + case OpTypePointer: + return false; + case OpTypeVector: + case OpTypeMatrix: + case OpTypeArray: + case OpTypeRuntimeArray: + return containsType(getContainedTypeId(typeId), typeOp, width); + default: + return typeClass == typeOp; + } +} + +// return true if the type is a pointer to PhysicalStorageBufferEXT or an +// array of such pointers. These require restrict/aliased decorations. +bool Builder::containsPhysicalStorageBufferOrArray(Id typeId) const +{ + const Instruction& instr = *module.getInstruction(typeId); + + Op typeClass = instr.getOpCode(); + switch (typeClass) + { + case OpTypePointer: + return getTypeStorageClass(typeId) == StorageClassPhysicalStorageBufferEXT; + case OpTypeArray: + return containsPhysicalStorageBufferOrArray(getContainedTypeId(typeId)); + default: + return false; + } +} + +// See if a scalar constant of this type has already been created, so it +// can be reused rather than duplicated. (Required by the specification). +Id Builder::findScalarConstant(Op typeClass, Op opcode, Id typeId, unsigned value) +{ + Instruction* constant; + for (int i = 0; i < (int)groupedConstants[typeClass].size(); ++i) { + constant = groupedConstants[typeClass][i]; + if (constant->getOpCode() == opcode && + constant->getTypeId() == typeId && + constant->getImmediateOperand(0) == value) + return constant->getResultId(); + } + + return 0; +} + +// Version of findScalarConstant (see above) for scalars that take two operands (e.g. a 'double' or 'int64'). +Id Builder::findScalarConstant(Op typeClass, Op opcode, Id typeId, unsigned v1, unsigned v2) +{ + Instruction* constant; + for (int i = 0; i < (int)groupedConstants[typeClass].size(); ++i) { + constant = groupedConstants[typeClass][i]; + if (constant->getOpCode() == opcode && + constant->getTypeId() == typeId && + constant->getImmediateOperand(0) == v1 && + constant->getImmediateOperand(1) == v2) + return constant->getResultId(); + } + + return 0; +} + +// Return true if consuming 'opcode' means consuming a constant. +// "constant" here means after final transform to executable code, +// the value consumed will be a constant, so includes specialization. +bool Builder::isConstantOpCode(Op opcode) const +{ + switch (opcode) { + case OpUndef: + case OpConstantTrue: + case OpConstantFalse: + case OpConstant: + case OpConstantComposite: + case OpConstantSampler: + case OpConstantNull: + case OpSpecConstantTrue: + case OpSpecConstantFalse: + case OpSpecConstant: + case OpSpecConstantComposite: + case OpSpecConstantOp: + return true; + default: + return false; + } +} + +// Return true if consuming 'opcode' means consuming a specialization constant. +bool Builder::isSpecConstantOpCode(Op opcode) const +{ + switch (opcode) { + case OpSpecConstantTrue: + case OpSpecConstantFalse: + case OpSpecConstant: + case OpSpecConstantComposite: + case OpSpecConstantOp: + return true; + default: + return false; + } +} + +Id Builder::makeBoolConstant(bool b, bool specConstant) +{ + Id typeId = makeBoolType(); + Instruction* constant; + Op opcode = specConstant ? (b ? OpSpecConstantTrue : OpSpecConstantFalse) : (b ? OpConstantTrue : OpConstantFalse); + + // See if we already made it. Applies only to regular constants, because specialization constants + // must remain distinct for the purpose of applying a SpecId decoration. + if (! specConstant) { + Id existing = 0; + for (int i = 0; i < (int)groupedConstants[OpTypeBool].size(); ++i) { + constant = groupedConstants[OpTypeBool][i]; + if (constant->getTypeId() == typeId && constant->getOpCode() == opcode) + existing = constant->getResultId(); + } + + if (existing) + return existing; + } + + // Make it + Instruction* c = new Instruction(getUniqueId(), typeId, opcode); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(c)); + groupedConstants[OpTypeBool].push_back(c); + module.mapInstruction(c); + + return c->getResultId(); +} + +Id Builder::makeIntConstant(Id typeId, unsigned value, bool specConstant) +{ + Op opcode = specConstant ? OpSpecConstant : OpConstant; + + // See if we already made it. Applies only to regular constants, because specialization constants + // must remain distinct for the purpose of applying a SpecId decoration. + if (! specConstant) { + Id existing = findScalarConstant(OpTypeInt, opcode, typeId, value); + if (existing) + return existing; + } + + Instruction* c = new Instruction(getUniqueId(), typeId, opcode); + c->addImmediateOperand(value); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(c)); + groupedConstants[OpTypeInt].push_back(c); + module.mapInstruction(c); + + return c->getResultId(); +} + +Id Builder::makeInt64Constant(Id typeId, unsigned long long value, bool specConstant) +{ + Op opcode = specConstant ? OpSpecConstant : OpConstant; + + unsigned op1 = value & 0xFFFFFFFF; + unsigned op2 = value >> 32; + + // See if we already made it. Applies only to regular constants, because specialization constants + // must remain distinct for the purpose of applying a SpecId decoration. + if (! specConstant) { + Id existing = findScalarConstant(OpTypeInt, opcode, typeId, op1, op2); + if (existing) + return existing; + } + + Instruction* c = new Instruction(getUniqueId(), typeId, opcode); + c->addImmediateOperand(op1); + c->addImmediateOperand(op2); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(c)); + groupedConstants[OpTypeInt].push_back(c); + module.mapInstruction(c); + + return c->getResultId(); +} + +Id Builder::makeFloatConstant(float f, bool specConstant) +{ + Op opcode = specConstant ? OpSpecConstant : OpConstant; + Id typeId = makeFloatType(32); + union { float fl; unsigned int ui; } u; + u.fl = f; + unsigned value = u.ui; + + // See if we already made it. Applies only to regular constants, because specialization constants + // must remain distinct for the purpose of applying a SpecId decoration. + if (! specConstant) { + Id existing = findScalarConstant(OpTypeFloat, opcode, typeId, value); + if (existing) + return existing; + } + + Instruction* c = new Instruction(getUniqueId(), typeId, opcode); + c->addImmediateOperand(value); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(c)); + groupedConstants[OpTypeFloat].push_back(c); + module.mapInstruction(c); + + return c->getResultId(); +} + +Id Builder::makeDoubleConstant(double d, bool specConstant) +{ + Op opcode = specConstant ? OpSpecConstant : OpConstant; + Id typeId = makeFloatType(64); + union { double db; unsigned long long ull; } u; + u.db = d; + unsigned long long value = u.ull; + unsigned op1 = value & 0xFFFFFFFF; + unsigned op2 = value >> 32; + + // See if we already made it. Applies only to regular constants, because specialization constants + // must remain distinct for the purpose of applying a SpecId decoration. + if (! specConstant) { + Id existing = findScalarConstant(OpTypeFloat, opcode, typeId, op1, op2); + if (existing) + return existing; + } + + Instruction* c = new Instruction(getUniqueId(), typeId, opcode); + c->addImmediateOperand(op1); + c->addImmediateOperand(op2); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(c)); + groupedConstants[OpTypeFloat].push_back(c); + module.mapInstruction(c); + + return c->getResultId(); +} + +Id Builder::makeFloat16Constant(float f16, bool specConstant) +{ + Op opcode = specConstant ? OpSpecConstant : OpConstant; + Id typeId = makeFloatType(16); + + spvutils::HexFloat<spvutils::FloatProxy<float>> fVal(f16); + spvutils::HexFloat<spvutils::FloatProxy<spvutils::Float16>> f16Val(0); + fVal.castTo(f16Val, spvutils::kRoundToZero); + + unsigned value = f16Val.value().getAsFloat().get_value(); + + // See if we already made it. Applies only to regular constants, because specialization constants + // must remain distinct for the purpose of applying a SpecId decoration. + if (!specConstant) { + Id existing = findScalarConstant(OpTypeFloat, opcode, typeId, value); + if (existing) + return existing; + } + + Instruction* c = new Instruction(getUniqueId(), typeId, opcode); + c->addImmediateOperand(value); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(c)); + groupedConstants[OpTypeFloat].push_back(c); + module.mapInstruction(c); + + return c->getResultId(); +} + +Id Builder::makeFpConstant(Id type, double d, bool specConstant) +{ + assert(isFloatType(type)); + + switch (getScalarTypeWidth(type)) { + case 16: + return makeFloat16Constant((float)d, specConstant); + case 32: + return makeFloatConstant((float)d, specConstant); + case 64: + return makeDoubleConstant(d, specConstant); + default: + break; + } + + assert(false); + return NoResult; +} + +Id Builder::findCompositeConstant(Op typeClass, Id typeId, const std::vector<Id>& comps) +{ + Instruction* constant = 0; + bool found = false; + for (int i = 0; i < (int)groupedConstants[typeClass].size(); ++i) { + constant = groupedConstants[typeClass][i]; + + if (constant->getTypeId() != typeId) + continue; + + // same contents? + bool mismatch = false; + for (int op = 0; op < constant->getNumOperands(); ++op) { + if (constant->getIdOperand(op) != comps[op]) { + mismatch = true; + break; + } + } + if (! mismatch) { + found = true; + break; + } + } + + return found ? constant->getResultId() : NoResult; +} + +Id Builder::findStructConstant(Id typeId, const std::vector<Id>& comps) +{ + Instruction* constant = 0; + bool found = false; + for (int i = 0; i < (int)groupedStructConstants[typeId].size(); ++i) { + constant = groupedStructConstants[typeId][i]; + + // same contents? + bool mismatch = false; + for (int op = 0; op < constant->getNumOperands(); ++op) { + if (constant->getIdOperand(op) != comps[op]) { + mismatch = true; + break; + } + } + if (! mismatch) { + found = true; + break; + } + } + + return found ? constant->getResultId() : NoResult; +} + +// Comments in header +Id Builder::makeCompositeConstant(Id typeId, const std::vector<Id>& members, bool specConstant) +{ + Op opcode = specConstant ? OpSpecConstantComposite : OpConstantComposite; + assert(typeId); + Op typeClass = getTypeClass(typeId); + + switch (typeClass) { + case OpTypeVector: + case OpTypeArray: + case OpTypeMatrix: + case OpTypeCooperativeMatrixNV: + if (! specConstant) { + Id existing = findCompositeConstant(typeClass, typeId, members); + if (existing) + return existing; + } + break; + case OpTypeStruct: + if (! specConstant) { + Id existing = findStructConstant(typeId, members); + if (existing) + return existing; + } + break; + default: + assert(0); + return makeFloatConstant(0.0); + } + + Instruction* c = new Instruction(getUniqueId(), typeId, opcode); + for (int op = 0; op < (int)members.size(); ++op) + c->addIdOperand(members[op]); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(c)); + if (typeClass == OpTypeStruct) + groupedStructConstants[typeId].push_back(c); + else + groupedConstants[typeClass].push_back(c); + module.mapInstruction(c); + + return c->getResultId(); +} + +Instruction* Builder::addEntryPoint(ExecutionModel model, Function* function, const char* name) +{ + Instruction* entryPoint = new Instruction(OpEntryPoint); + entryPoint->addImmediateOperand(model); + entryPoint->addIdOperand(function->getId()); + entryPoint->addStringOperand(name); + + entryPoints.push_back(std::unique_ptr<Instruction>(entryPoint)); + + return entryPoint; +} + +// Currently relying on the fact that all 'value' of interest are small non-negative values. +void Builder::addExecutionMode(Function* entryPoint, ExecutionMode mode, int value1, int value2, int value3) +{ + Instruction* instr = new Instruction(OpExecutionMode); + instr->addIdOperand(entryPoint->getId()); + instr->addImmediateOperand(mode); + if (value1 >= 0) + instr->addImmediateOperand(value1); + if (value2 >= 0) + instr->addImmediateOperand(value2); + if (value3 >= 0) + instr->addImmediateOperand(value3); + + executionModes.push_back(std::unique_ptr<Instruction>(instr)); +} + +void Builder::addName(Id id, const char* string) +{ + Instruction* name = new Instruction(OpName); + name->addIdOperand(id); + name->addStringOperand(string); + + names.push_back(std::unique_ptr<Instruction>(name)); +} + +void Builder::addMemberName(Id id, int memberNumber, const char* string) +{ + Instruction* name = new Instruction(OpMemberName); + name->addIdOperand(id); + name->addImmediateOperand(memberNumber); + name->addStringOperand(string); + + names.push_back(std::unique_ptr<Instruction>(name)); +} + +void Builder::addDecoration(Id id, Decoration decoration, int num) +{ + if (decoration == spv::DecorationMax) + return; + + Instruction* dec = new Instruction(OpDecorate); + dec->addIdOperand(id); + dec->addImmediateOperand(decoration); + if (num >= 0) + dec->addImmediateOperand(num); + + decorations.push_back(std::unique_ptr<Instruction>(dec)); +} + +void Builder::addDecoration(Id id, Decoration decoration, const char* s) +{ + if (decoration == spv::DecorationMax) + return; + + Instruction* dec = new Instruction(OpDecorateStringGOOGLE); + dec->addIdOperand(id); + dec->addImmediateOperand(decoration); + dec->addStringOperand(s); + + decorations.push_back(std::unique_ptr<Instruction>(dec)); +} + +void Builder::addDecorationId(Id id, Decoration decoration, Id idDecoration) +{ + if (decoration == spv::DecorationMax) + return; + + Instruction* dec = new Instruction(OpDecorateId); + dec->addIdOperand(id); + dec->addImmediateOperand(decoration); + dec->addIdOperand(idDecoration); + + decorations.push_back(std::unique_ptr<Instruction>(dec)); +} + +void Builder::addMemberDecoration(Id id, unsigned int member, Decoration decoration, int num) +{ + if (decoration == spv::DecorationMax) + return; + + Instruction* dec = new Instruction(OpMemberDecorate); + dec->addIdOperand(id); + dec->addImmediateOperand(member); + dec->addImmediateOperand(decoration); + if (num >= 0) + dec->addImmediateOperand(num); + + decorations.push_back(std::unique_ptr<Instruction>(dec)); +} + +void Builder::addMemberDecoration(Id id, unsigned int member, Decoration decoration, const char *s) +{ + if (decoration == spv::DecorationMax) + return; + + Instruction* dec = new Instruction(OpMemberDecorateStringGOOGLE); + dec->addIdOperand(id); + dec->addImmediateOperand(member); + dec->addImmediateOperand(decoration); + dec->addStringOperand(s); + + decorations.push_back(std::unique_ptr<Instruction>(dec)); +} + +// Comments in header +Function* Builder::makeEntryPoint(const char* entryPoint) +{ + assert(! entryPointFunction); + + Block* entry; + std::vector<Id> params; + std::vector<std::vector<Decoration>> decorations; + + entryPointFunction = makeFunctionEntry(NoPrecision, makeVoidType(), entryPoint, params, decorations, &entry); + + return entryPointFunction; +} + +// Comments in header +Function* Builder::makeFunctionEntry(Decoration precision, Id returnType, const char* name, + const std::vector<Id>& paramTypes, const std::vector<std::vector<Decoration>>& decorations, Block **entry) +{ + // Make the function and initial instructions in it + Id typeId = makeFunctionType(returnType, paramTypes); + Id firstParamId = paramTypes.size() == 0 ? 0 : getUniqueIds((int)paramTypes.size()); + Function* function = new Function(getUniqueId(), returnType, typeId, firstParamId, module); + + // Set up the precisions + setPrecision(function->getId(), precision); + for (unsigned p = 0; p < (unsigned)decorations.size(); ++p) { + for (int d = 0; d < (int)decorations[p].size(); ++d) + addDecoration(firstParamId + p, decorations[p][d]); + } + + // CFG + if (entry) { + *entry = new Block(getUniqueId(), *function); + function->addBlock(*entry); + setBuildPoint(*entry); + } + + if (name) + addName(function->getId(), name); + + functions.push_back(std::unique_ptr<Function>(function)); + + return function; +} + +// Comments in header +void Builder::makeReturn(bool implicit, Id retVal) +{ + if (retVal) { + Instruction* inst = new Instruction(NoResult, NoType, OpReturnValue); + inst->addIdOperand(retVal); + buildPoint->addInstruction(std::unique_ptr<Instruction>(inst)); + } else + buildPoint->addInstruction(std::unique_ptr<Instruction>(new Instruction(NoResult, NoType, OpReturn))); + + if (! implicit) + createAndSetNoPredecessorBlock("post-return"); +} + +// Comments in header +void Builder::leaveFunction() +{ + Block* block = buildPoint; + Function& function = buildPoint->getParent(); + assert(block); + + // If our function did not contain a return, add a return void now. + if (! block->isTerminated()) { + if (function.getReturnType() == makeVoidType()) + makeReturn(true); + else { + makeReturn(true, createUndefined(function.getReturnType())); + } + } +} + +// Comments in header +void Builder::makeDiscard() +{ + buildPoint->addInstruction(std::unique_ptr<Instruction>(new Instruction(OpKill))); + createAndSetNoPredecessorBlock("post-discard"); +} + +// Comments in header +Id Builder::createVariable(StorageClass storageClass, Id type, const char* name, Id initializer) +{ + Id pointerType = makePointer(storageClass, type); + Instruction* inst = new Instruction(getUniqueId(), pointerType, OpVariable); + inst->addImmediateOperand(storageClass); + if (initializer != NoResult) + inst->addIdOperand(initializer); + + switch (storageClass) { + case StorageClassFunction: + // Validation rules require the declaration in the entry block + buildPoint->getParent().addLocalVariable(std::unique_ptr<Instruction>(inst)); + break; + + default: + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(inst)); + module.mapInstruction(inst); + break; + } + + if (name) + addName(inst->getResultId(), name); + + return inst->getResultId(); +} + +// Comments in header +Id Builder::createUndefined(Id type) +{ + Instruction* inst = new Instruction(getUniqueId(), type, OpUndef); + buildPoint->addInstruction(std::unique_ptr<Instruction>(inst)); + return inst->getResultId(); +} + +// av/vis/nonprivate are unnecessary and illegal for some storage classes. +spv::MemoryAccessMask Builder::sanitizeMemoryAccessForStorageClass(spv::MemoryAccessMask memoryAccess, StorageClass sc) const +{ + switch (sc) { + case spv::StorageClassUniform: + case spv::StorageClassWorkgroup: + case spv::StorageClassStorageBuffer: + case spv::StorageClassPhysicalStorageBufferEXT: + break; + default: + memoryAccess = spv::MemoryAccessMask(memoryAccess & + ~(spv::MemoryAccessMakePointerAvailableKHRMask | + spv::MemoryAccessMakePointerVisibleKHRMask | + spv::MemoryAccessNonPrivatePointerKHRMask)); + break; + } + return memoryAccess; +} + +// Comments in header +void Builder::createStore(Id rValue, Id lValue, spv::MemoryAccessMask memoryAccess, spv::Scope scope, unsigned int alignment) +{ + Instruction* store = new Instruction(OpStore); + store->addIdOperand(lValue); + store->addIdOperand(rValue); + + memoryAccess = sanitizeMemoryAccessForStorageClass(memoryAccess, getStorageClass(lValue)); + + if (memoryAccess != MemoryAccessMaskNone) { + store->addImmediateOperand(memoryAccess); + if (memoryAccess & spv::MemoryAccessAlignedMask) { + store->addImmediateOperand(alignment); + } + if (memoryAccess & spv::MemoryAccessMakePointerAvailableKHRMask) { + store->addIdOperand(makeUintConstant(scope)); + } + } + + buildPoint->addInstruction(std::unique_ptr<Instruction>(store)); +} + +// Comments in header +Id Builder::createLoad(Id lValue, spv::MemoryAccessMask memoryAccess, spv::Scope scope, unsigned int alignment) +{ + Instruction* load = new Instruction(getUniqueId(), getDerefTypeId(lValue), OpLoad); + load->addIdOperand(lValue); + + memoryAccess = sanitizeMemoryAccessForStorageClass(memoryAccess, getStorageClass(lValue)); + + if (memoryAccess != MemoryAccessMaskNone) { + load->addImmediateOperand(memoryAccess); + if (memoryAccess & spv::MemoryAccessAlignedMask) { + load->addImmediateOperand(alignment); + } + if (memoryAccess & spv::MemoryAccessMakePointerVisibleKHRMask) { + load->addIdOperand(makeUintConstant(scope)); + } + } + + buildPoint->addInstruction(std::unique_ptr<Instruction>(load)); + + return load->getResultId(); +} + +// Comments in header +Id Builder::createAccessChain(StorageClass storageClass, Id base, const std::vector<Id>& offsets) +{ + // Figure out the final resulting type. + spv::Id typeId = getTypeId(base); + assert(isPointerType(typeId) && offsets.size() > 0); + typeId = getContainedTypeId(typeId); + for (int i = 0; i < (int)offsets.size(); ++i) { + if (isStructType(typeId)) { + assert(isConstantScalar(offsets[i])); + typeId = getContainedTypeId(typeId, getConstantScalar(offsets[i])); + } else + typeId = getContainedTypeId(typeId, offsets[i]); + } + typeId = makePointer(storageClass, typeId); + + // Make the instruction + Instruction* chain = new Instruction(getUniqueId(), typeId, OpAccessChain); + chain->addIdOperand(base); + for (int i = 0; i < (int)offsets.size(); ++i) + chain->addIdOperand(offsets[i]); + buildPoint->addInstruction(std::unique_ptr<Instruction>(chain)); + + return chain->getResultId(); +} + +Id Builder::createArrayLength(Id base, unsigned int member) +{ + spv::Id intType = makeUintType(32); + Instruction* length = new Instruction(getUniqueId(), intType, OpArrayLength); + length->addIdOperand(base); + length->addImmediateOperand(member); + buildPoint->addInstruction(std::unique_ptr<Instruction>(length)); + + return length->getResultId(); +} + +Id Builder::createCooperativeMatrixLength(Id type) +{ + spv::Id intType = makeUintType(32); + + // Generate code for spec constants if in spec constant operation + // generation mode. + if (generatingOpCodeForSpecConst) { + return createSpecConstantOp(OpCooperativeMatrixLengthNV, intType, std::vector<Id>(1, type), std::vector<Id>()); + } + + Instruction* length = new Instruction(getUniqueId(), intType, OpCooperativeMatrixLengthNV); + length->addIdOperand(type); + buildPoint->addInstruction(std::unique_ptr<Instruction>(length)); + + return length->getResultId(); +} + +Id Builder::createCompositeExtract(Id composite, Id typeId, unsigned index) +{ + // Generate code for spec constants if in spec constant operation + // generation mode. + if (generatingOpCodeForSpecConst) { + return createSpecConstantOp(OpCompositeExtract, typeId, std::vector<Id>(1, composite), std::vector<Id>(1, index)); + } + Instruction* extract = new Instruction(getUniqueId(), typeId, OpCompositeExtract); + extract->addIdOperand(composite); + extract->addImmediateOperand(index); + buildPoint->addInstruction(std::unique_ptr<Instruction>(extract)); + + return extract->getResultId(); +} + +Id Builder::createCompositeExtract(Id composite, Id typeId, const std::vector<unsigned>& indexes) +{ + // Generate code for spec constants if in spec constant operation + // generation mode. + if (generatingOpCodeForSpecConst) { + return createSpecConstantOp(OpCompositeExtract, typeId, std::vector<Id>(1, composite), indexes); + } + Instruction* extract = new Instruction(getUniqueId(), typeId, OpCompositeExtract); + extract->addIdOperand(composite); + for (int i = 0; i < (int)indexes.size(); ++i) + extract->addImmediateOperand(indexes[i]); + buildPoint->addInstruction(std::unique_ptr<Instruction>(extract)); + + return extract->getResultId(); +} + +Id Builder::createCompositeInsert(Id object, Id composite, Id typeId, unsigned index) +{ + Instruction* insert = new Instruction(getUniqueId(), typeId, OpCompositeInsert); + insert->addIdOperand(object); + insert->addIdOperand(composite); + insert->addImmediateOperand(index); + buildPoint->addInstruction(std::unique_ptr<Instruction>(insert)); + + return insert->getResultId(); +} + +Id Builder::createCompositeInsert(Id object, Id composite, Id typeId, const std::vector<unsigned>& indexes) +{ + Instruction* insert = new Instruction(getUniqueId(), typeId, OpCompositeInsert); + insert->addIdOperand(object); + insert->addIdOperand(composite); + for (int i = 0; i < (int)indexes.size(); ++i) + insert->addImmediateOperand(indexes[i]); + buildPoint->addInstruction(std::unique_ptr<Instruction>(insert)); + + return insert->getResultId(); +} + +Id Builder::createVectorExtractDynamic(Id vector, Id typeId, Id componentIndex) +{ + Instruction* extract = new Instruction(getUniqueId(), typeId, OpVectorExtractDynamic); + extract->addIdOperand(vector); + extract->addIdOperand(componentIndex); + buildPoint->addInstruction(std::unique_ptr<Instruction>(extract)); + + return extract->getResultId(); +} + +Id Builder::createVectorInsertDynamic(Id vector, Id typeId, Id component, Id componentIndex) +{ + Instruction* insert = new Instruction(getUniqueId(), typeId, OpVectorInsertDynamic); + insert->addIdOperand(vector); + insert->addIdOperand(component); + insert->addIdOperand(componentIndex); + buildPoint->addInstruction(std::unique_ptr<Instruction>(insert)); + + return insert->getResultId(); +} + +// An opcode that has no operands, no result id, and no type +void Builder::createNoResultOp(Op opCode) +{ + Instruction* op = new Instruction(opCode); + buildPoint->addInstruction(std::unique_ptr<Instruction>(op)); +} + +// An opcode that has one id operand, no result id, and no type +void Builder::createNoResultOp(Op opCode, Id operand) +{ + Instruction* op = new Instruction(opCode); + op->addIdOperand(operand); + buildPoint->addInstruction(std::unique_ptr<Instruction>(op)); +} + +// An opcode that has one or more operands, no result id, and no type +void Builder::createNoResultOp(Op opCode, const std::vector<Id>& operands) +{ + Instruction* op = new Instruction(opCode); + for (auto it = operands.cbegin(); it != operands.cend(); ++it) { + op->addIdOperand(*it); + } + buildPoint->addInstruction(std::unique_ptr<Instruction>(op)); +} + +// An opcode that has multiple operands, no result id, and no type +void Builder::createNoResultOp(Op opCode, const std::vector<IdImmediate>& operands) +{ + Instruction* op = new Instruction(opCode); + for (auto it = operands.cbegin(); it != operands.cend(); ++it) { + if (it->isId) + op->addIdOperand(it->word); + else + op->addImmediateOperand(it->word); + } + buildPoint->addInstruction(std::unique_ptr<Instruction>(op)); +} + +void Builder::createControlBarrier(Scope execution, Scope memory, MemorySemanticsMask semantics) +{ + Instruction* op = new Instruction(OpControlBarrier); + op->addIdOperand(makeUintConstant(execution)); + op->addIdOperand(makeUintConstant(memory)); + op->addIdOperand(makeUintConstant(semantics)); + buildPoint->addInstruction(std::unique_ptr<Instruction>(op)); +} + +void Builder::createMemoryBarrier(unsigned executionScope, unsigned memorySemantics) +{ + Instruction* op = new Instruction(OpMemoryBarrier); + op->addIdOperand(makeUintConstant(executionScope)); + op->addIdOperand(makeUintConstant(memorySemantics)); + buildPoint->addInstruction(std::unique_ptr<Instruction>(op)); +} + +// An opcode that has one operands, a result id, and a type +Id Builder::createUnaryOp(Op opCode, Id typeId, Id operand) +{ + // Generate code for spec constants if in spec constant operation + // generation mode. + if (generatingOpCodeForSpecConst) { + return createSpecConstantOp(opCode, typeId, std::vector<Id>(1, operand), std::vector<Id>()); + } + Instruction* op = new Instruction(getUniqueId(), typeId, opCode); + op->addIdOperand(operand); + buildPoint->addInstruction(std::unique_ptr<Instruction>(op)); + + return op->getResultId(); +} + +Id Builder::createBinOp(Op opCode, Id typeId, Id left, Id right) +{ + // Generate code for spec constants if in spec constant operation + // generation mode. + if (generatingOpCodeForSpecConst) { + std::vector<Id> operands(2); + operands[0] = left; operands[1] = right; + return createSpecConstantOp(opCode, typeId, operands, std::vector<Id>()); + } + Instruction* op = new Instruction(getUniqueId(), typeId, opCode); + op->addIdOperand(left); + op->addIdOperand(right); + buildPoint->addInstruction(std::unique_ptr<Instruction>(op)); + + return op->getResultId(); +} + +Id Builder::createTriOp(Op opCode, Id typeId, Id op1, Id op2, Id op3) +{ + // Generate code for spec constants if in spec constant operation + // generation mode. + if (generatingOpCodeForSpecConst) { + std::vector<Id> operands(3); + operands[0] = op1; + operands[1] = op2; + operands[2] = op3; + return createSpecConstantOp( + opCode, typeId, operands, std::vector<Id>()); + } + Instruction* op = new Instruction(getUniqueId(), typeId, opCode); + op->addIdOperand(op1); + op->addIdOperand(op2); + op->addIdOperand(op3); + buildPoint->addInstruction(std::unique_ptr<Instruction>(op)); + + return op->getResultId(); +} + +Id Builder::createOp(Op opCode, Id typeId, const std::vector<Id>& operands) +{ + Instruction* op = new Instruction(getUniqueId(), typeId, opCode); + for (auto it = operands.cbegin(); it != operands.cend(); ++it) + op->addIdOperand(*it); + buildPoint->addInstruction(std::unique_ptr<Instruction>(op)); + + return op->getResultId(); +} + +Id Builder::createOp(Op opCode, Id typeId, const std::vector<IdImmediate>& operands) +{ + Instruction* op = new Instruction(getUniqueId(), typeId, opCode); + for (auto it = operands.cbegin(); it != operands.cend(); ++it) { + if (it->isId) + op->addIdOperand(it->word); + else + op->addImmediateOperand(it->word); + } + buildPoint->addInstruction(std::unique_ptr<Instruction>(op)); + + return op->getResultId(); +} + +Id Builder::createSpecConstantOp(Op opCode, Id typeId, const std::vector<Id>& operands, const std::vector<unsigned>& literals) +{ + Instruction* op = new Instruction(getUniqueId(), typeId, OpSpecConstantOp); + op->addImmediateOperand((unsigned) opCode); + for (auto it = operands.cbegin(); it != operands.cend(); ++it) + op->addIdOperand(*it); + for (auto it = literals.cbegin(); it != literals.cend(); ++it) + op->addImmediateOperand(*it); + module.mapInstruction(op); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(op)); + + return op->getResultId(); +} + +Id Builder::createFunctionCall(spv::Function* function, const std::vector<spv::Id>& args) +{ + Instruction* op = new Instruction(getUniqueId(), function->getReturnType(), OpFunctionCall); + op->addIdOperand(function->getId()); + for (int a = 0; a < (int)args.size(); ++a) + op->addIdOperand(args[a]); + buildPoint->addInstruction(std::unique_ptr<Instruction>(op)); + + return op->getResultId(); +} + +// Comments in header +Id Builder::createRvalueSwizzle(Decoration precision, Id typeId, Id source, const std::vector<unsigned>& channels) +{ + if (channels.size() == 1) + return setPrecision(createCompositeExtract(source, typeId, channels.front()), precision); + + if (generatingOpCodeForSpecConst) { + std::vector<Id> operands(2); + operands[0] = operands[1] = source; + return setPrecision(createSpecConstantOp(OpVectorShuffle, typeId, operands, channels), precision); + } + Instruction* swizzle = new Instruction(getUniqueId(), typeId, OpVectorShuffle); + assert(isVector(source)); + swizzle->addIdOperand(source); + swizzle->addIdOperand(source); + for (int i = 0; i < (int)channels.size(); ++i) + swizzle->addImmediateOperand(channels[i]); + buildPoint->addInstruction(std::unique_ptr<Instruction>(swizzle)); + + return setPrecision(swizzle->getResultId(), precision); +} + +// Comments in header +Id Builder::createLvalueSwizzle(Id typeId, Id target, Id source, const std::vector<unsigned>& channels) +{ + if (channels.size() == 1 && getNumComponents(source) == 1) + return createCompositeInsert(source, target, typeId, channels.front()); + + Instruction* swizzle = new Instruction(getUniqueId(), typeId, OpVectorShuffle); + + assert(isVector(target)); + swizzle->addIdOperand(target); + + assert(getNumComponents(source) == (int)channels.size()); + assert(isVector(source)); + swizzle->addIdOperand(source); + + // Set up an identity shuffle from the base value to the result value + unsigned int components[4]; + int numTargetComponents = getNumComponents(target); + for (int i = 0; i < numTargetComponents; ++i) + components[i] = i; + + // Punch in the l-value swizzle + for (int i = 0; i < (int)channels.size(); ++i) + components[channels[i]] = numTargetComponents + i; + + // finish the instruction with these components selectors + for (int i = 0; i < numTargetComponents; ++i) + swizzle->addImmediateOperand(components[i]); + buildPoint->addInstruction(std::unique_ptr<Instruction>(swizzle)); + + return swizzle->getResultId(); +} + +// Comments in header +void Builder::promoteScalar(Decoration precision, Id& left, Id& right) +{ + int direction = getNumComponents(right) - getNumComponents(left); + + if (direction > 0) + left = smearScalar(precision, left, makeVectorType(getTypeId(left), getNumComponents(right))); + else if (direction < 0) + right = smearScalar(precision, right, makeVectorType(getTypeId(right), getNumComponents(left))); + + return; +} + +// Comments in header +Id Builder::smearScalar(Decoration precision, Id scalar, Id vectorType) +{ + assert(getNumComponents(scalar) == 1); + assert(getTypeId(scalar) == getScalarTypeId(vectorType)); + + int numComponents = getNumTypeComponents(vectorType); + if (numComponents == 1) + return scalar; + + Instruction* smear = nullptr; + if (generatingOpCodeForSpecConst) { + auto members = std::vector<spv::Id>(numComponents, scalar); + // Sometime even in spec-constant-op mode, the temporary vector created by + // promoting a scalar might not be a spec constant. This should depend on + // the scalar. + // e.g.: + // const vec2 spec_const_result = a_spec_const_vec2 + a_front_end_const_scalar; + // In such cases, the temporary vector created from a_front_end_const_scalar + // is not a spec constant vector, even though the binary operation node is marked + // as 'specConstant' and we are in spec-constant-op mode. + auto result_id = makeCompositeConstant(vectorType, members, isSpecConstant(scalar)); + smear = module.getInstruction(result_id); + } else { + smear = new Instruction(getUniqueId(), vectorType, OpCompositeConstruct); + for (int c = 0; c < numComponents; ++c) + smear->addIdOperand(scalar); + buildPoint->addInstruction(std::unique_ptr<Instruction>(smear)); + } + + return setPrecision(smear->getResultId(), precision); +} + +// Comments in header +Id Builder::createBuiltinCall(Id resultType, Id builtins, int entryPoint, const std::vector<Id>& args) +{ + Instruction* inst = new Instruction(getUniqueId(), resultType, OpExtInst); + inst->addIdOperand(builtins); + inst->addImmediateOperand(entryPoint); + for (int arg = 0; arg < (int)args.size(); ++arg) + inst->addIdOperand(args[arg]); + + buildPoint->addInstruction(std::unique_ptr<Instruction>(inst)); + + return inst->getResultId(); +} + +// Accept all parameters needed to create a texture instruction. +// Create the correct instruction based on the inputs, and make the call. +Id Builder::createTextureCall(Decoration precision, Id resultType, bool sparse, bool fetch, bool proj, bool gather, + bool noImplicitLod, const TextureParameters& parameters, ImageOperandsMask signExtensionMask) +{ + static const int maxTextureArgs = 10; + Id texArgs[maxTextureArgs] = {}; + + // + // Set up the fixed arguments + // + int numArgs = 0; + bool explicitLod = false; + texArgs[numArgs++] = parameters.sampler; + texArgs[numArgs++] = parameters.coords; + if (parameters.Dref != NoResult) + texArgs[numArgs++] = parameters.Dref; + if (parameters.component != NoResult) + texArgs[numArgs++] = parameters.component; + +#ifdef NV_EXTENSIONS + if (parameters.granularity != NoResult) + texArgs[numArgs++] = parameters.granularity; + if (parameters.coarse != NoResult) + texArgs[numArgs++] = parameters.coarse; +#endif + + // + // Set up the optional arguments + // + int optArgNum = numArgs; // track which operand, if it exists, is the mask of optional arguments + ++numArgs; // speculatively make room for the mask operand + ImageOperandsMask mask = ImageOperandsMaskNone; // the mask operand + if (parameters.bias) { + mask = (ImageOperandsMask)(mask | ImageOperandsBiasMask); + texArgs[numArgs++] = parameters.bias; + } + if (parameters.lod) { + mask = (ImageOperandsMask)(mask | ImageOperandsLodMask); + texArgs[numArgs++] = parameters.lod; + explicitLod = true; + } else if (parameters.gradX) { + mask = (ImageOperandsMask)(mask | ImageOperandsGradMask); + texArgs[numArgs++] = parameters.gradX; + texArgs[numArgs++] = parameters.gradY; + explicitLod = true; + } else if (noImplicitLod && ! fetch && ! gather) { + // have to explicitly use lod of 0 if not allowed to have them be implicit, and + // we would otherwise be about to issue an implicit instruction + mask = (ImageOperandsMask)(mask | ImageOperandsLodMask); + texArgs[numArgs++] = makeFloatConstant(0.0); + explicitLod = true; + } + if (parameters.offset) { + if (isConstant(parameters.offset)) + mask = (ImageOperandsMask)(mask | ImageOperandsConstOffsetMask); + else { + addCapability(CapabilityImageGatherExtended); + mask = (ImageOperandsMask)(mask | ImageOperandsOffsetMask); + } + texArgs[numArgs++] = parameters.offset; + } + if (parameters.offsets) { + addCapability(CapabilityImageGatherExtended); + mask = (ImageOperandsMask)(mask | ImageOperandsConstOffsetsMask); + texArgs[numArgs++] = parameters.offsets; + } + if (parameters.sample) { + mask = (ImageOperandsMask)(mask | ImageOperandsSampleMask); + texArgs[numArgs++] = parameters.sample; + } + if (parameters.lodClamp) { + // capability if this bit is used + addCapability(CapabilityMinLod); + + mask = (ImageOperandsMask)(mask | ImageOperandsMinLodMask); + texArgs[numArgs++] = parameters.lodClamp; + } + if (parameters.nonprivate) { + mask = mask | ImageOperandsNonPrivateTexelKHRMask; + } + if (parameters.volatil) { + mask = mask | ImageOperandsVolatileTexelKHRMask; + } + mask = mask | signExtensionMask; + if (mask == ImageOperandsMaskNone) + --numArgs; // undo speculative reservation for the mask argument + else + texArgs[optArgNum] = mask; + + // + // Set up the instruction + // + Op opCode = OpNop; // All paths below need to set this + if (fetch) { + if (sparse) + opCode = OpImageSparseFetch; + else + opCode = OpImageFetch; +#ifdef NV_EXTENSIONS + } else if (parameters.granularity && parameters.coarse) { + opCode = OpImageSampleFootprintNV; +#endif + } else if (gather) { + if (parameters.Dref) + if (sparse) + opCode = OpImageSparseDrefGather; + else + opCode = OpImageDrefGather; + else + if (sparse) + opCode = OpImageSparseGather; + else + opCode = OpImageGather; + } else if (explicitLod) { + if (parameters.Dref) { + if (proj) + if (sparse) + opCode = OpImageSparseSampleProjDrefExplicitLod; + else + opCode = OpImageSampleProjDrefExplicitLod; + else + if (sparse) + opCode = OpImageSparseSampleDrefExplicitLod; + else + opCode = OpImageSampleDrefExplicitLod; + } else { + if (proj) + if (sparse) + opCode = OpImageSparseSampleProjExplicitLod; + else + opCode = OpImageSampleProjExplicitLod; + else + if (sparse) + opCode = OpImageSparseSampleExplicitLod; + else + opCode = OpImageSampleExplicitLod; + } + } else { + if (parameters.Dref) { + if (proj) + if (sparse) + opCode = OpImageSparseSampleProjDrefImplicitLod; + else + opCode = OpImageSampleProjDrefImplicitLod; + else + if (sparse) + opCode = OpImageSparseSampleDrefImplicitLod; + else + opCode = OpImageSampleDrefImplicitLod; + } else { + if (proj) + if (sparse) + opCode = OpImageSparseSampleProjImplicitLod; + else + opCode = OpImageSampleProjImplicitLod; + else + if (sparse) + opCode = OpImageSparseSampleImplicitLod; + else + opCode = OpImageSampleImplicitLod; + } + } + + // See if the result type is expecting a smeared result. + // This happens when a legacy shadow*() call is made, which + // gets a vec4 back instead of a float. + Id smearedType = resultType; + if (! isScalarType(resultType)) { + switch (opCode) { + case OpImageSampleDrefImplicitLod: + case OpImageSampleDrefExplicitLod: + case OpImageSampleProjDrefImplicitLod: + case OpImageSampleProjDrefExplicitLod: + resultType = getScalarTypeId(resultType); + break; + default: + break; + } + } + + Id typeId0 = 0; + Id typeId1 = 0; + + if (sparse) { + typeId0 = resultType; + typeId1 = getDerefTypeId(parameters.texelOut); + resultType = makeStructResultType(typeId0, typeId1); + } + + // Build the SPIR-V instruction + Instruction* textureInst = new Instruction(getUniqueId(), resultType, opCode); + for (int op = 0; op < optArgNum; ++op) + textureInst->addIdOperand(texArgs[op]); + if (optArgNum < numArgs) + textureInst->addImmediateOperand(texArgs[optArgNum]); + for (int op = optArgNum + 1; op < numArgs; ++op) + textureInst->addIdOperand(texArgs[op]); + setPrecision(textureInst->getResultId(), precision); + buildPoint->addInstruction(std::unique_ptr<Instruction>(textureInst)); + + Id resultId = textureInst->getResultId(); + + if (sparse) { + // set capability + addCapability(CapabilitySparseResidency); + + // Decode the return type that was a special structure + createStore(createCompositeExtract(resultId, typeId1, 1), parameters.texelOut); + resultId = createCompositeExtract(resultId, typeId0, 0); + setPrecision(resultId, precision); + } else { + // When a smear is needed, do it, as per what was computed + // above when resultType was changed to a scalar type. + if (resultType != smearedType) + resultId = smearScalar(precision, resultId, smearedType); + } + + return resultId; +} + +// Comments in header +Id Builder::createTextureQueryCall(Op opCode, const TextureParameters& parameters, bool isUnsignedResult) +{ + // Figure out the result type + Id resultType = 0; + switch (opCode) { + case OpImageQuerySize: + case OpImageQuerySizeLod: + { + int numComponents = 0; + switch (getTypeDimensionality(getImageType(parameters.sampler))) { + case Dim1D: + case DimBuffer: + numComponents = 1; + break; + case Dim2D: + case DimCube: + case DimRect: + case DimSubpassData: + numComponents = 2; + break; + case Dim3D: + numComponents = 3; + break; + + default: + assert(0); + break; + } + if (isArrayedImageType(getImageType(parameters.sampler))) + ++numComponents; + + Id intType = isUnsignedResult ? makeUintType(32) : makeIntType(32); + if (numComponents == 1) + resultType = intType; + else + resultType = makeVectorType(intType, numComponents); + + break; + } + case OpImageQueryLod: +#ifdef AMD_EXTENSIONS + resultType = makeVectorType(getScalarTypeId(getTypeId(parameters.coords)), 2); +#else + resultType = makeVectorType(makeFloatType(32), 2); +#endif + break; + case OpImageQueryLevels: + case OpImageQuerySamples: + resultType = isUnsignedResult ? makeUintType(32) : makeIntType(32); + break; + default: + assert(0); + break; + } + + Instruction* query = new Instruction(getUniqueId(), resultType, opCode); + query->addIdOperand(parameters.sampler); + if (parameters.coords) + query->addIdOperand(parameters.coords); + if (parameters.lod) + query->addIdOperand(parameters.lod); + buildPoint->addInstruction(std::unique_ptr<Instruction>(query)); + + return query->getResultId(); +} + +// External comments in header. +// Operates recursively to visit the composite's hierarchy. +Id Builder::createCompositeCompare(Decoration precision, Id value1, Id value2, bool equal) +{ + Id boolType = makeBoolType(); + Id valueType = getTypeId(value1); + + Id resultId = NoResult; + + int numConstituents = getNumTypeConstituents(valueType); + + // Scalars and Vectors + + if (isScalarType(valueType) || isVectorType(valueType)) { + assert(valueType == getTypeId(value2)); + // These just need a single comparison, just have + // to figure out what it is. + Op op; + switch (getMostBasicTypeClass(valueType)) { + case OpTypeFloat: + op = equal ? OpFOrdEqual : OpFOrdNotEqual; + break; + case OpTypeInt: + default: + op = equal ? OpIEqual : OpINotEqual; + break; + case OpTypeBool: + op = equal ? OpLogicalEqual : OpLogicalNotEqual; + precision = NoPrecision; + break; + } + + if (isScalarType(valueType)) { + // scalar + resultId = createBinOp(op, boolType, value1, value2); + } else { + // vector + resultId = createBinOp(op, makeVectorType(boolType, numConstituents), value1, value2); + setPrecision(resultId, precision); + // reduce vector compares... + resultId = createUnaryOp(equal ? OpAll : OpAny, boolType, resultId); + } + + return setPrecision(resultId, precision); + } + + // Only structs, arrays, and matrices should be left. + // They share in common the reduction operation across their constituents. + assert(isAggregateType(valueType) || isMatrixType(valueType)); + + // Compare each pair of constituents + for (int constituent = 0; constituent < numConstituents; ++constituent) { + std::vector<unsigned> indexes(1, constituent); + Id constituentType1 = getContainedTypeId(getTypeId(value1), constituent); + Id constituentType2 = getContainedTypeId(getTypeId(value2), constituent); + Id constituent1 = createCompositeExtract(value1, constituentType1, indexes); + Id constituent2 = createCompositeExtract(value2, constituentType2, indexes); + + Id subResultId = createCompositeCompare(precision, constituent1, constituent2, equal); + + if (constituent == 0) + resultId = subResultId; + else + resultId = setPrecision(createBinOp(equal ? OpLogicalAnd : OpLogicalOr, boolType, resultId, subResultId), precision); + } + + return resultId; +} + +// OpCompositeConstruct +Id Builder::createCompositeConstruct(Id typeId, const std::vector<Id>& constituents) +{ + assert(isAggregateType(typeId) || (getNumTypeConstituents(typeId) > 1 && getNumTypeConstituents(typeId) == (int)constituents.size())); + + if (generatingOpCodeForSpecConst) { + // Sometime, even in spec-constant-op mode, the constant composite to be + // constructed may not be a specialization constant. + // e.g.: + // const mat2 m2 = mat2(a_spec_const, a_front_end_const, another_front_end_const, third_front_end_const); + // The first column vector should be a spec constant one, as a_spec_const is a spec constant. + // The second column vector should NOT be spec constant, as it does not contain any spec constants. + // To handle such cases, we check the constituents of the constant vector to determine whether this + // vector should be created as a spec constant. + return makeCompositeConstant(typeId, constituents, + std::any_of(constituents.begin(), constituents.end(), + [&](spv::Id id) { return isSpecConstant(id); })); + } + + Instruction* op = new Instruction(getUniqueId(), typeId, OpCompositeConstruct); + for (int c = 0; c < (int)constituents.size(); ++c) + op->addIdOperand(constituents[c]); + buildPoint->addInstruction(std::unique_ptr<Instruction>(op)); + + return op->getResultId(); +} + +// Vector or scalar constructor +Id Builder::createConstructor(Decoration precision, const std::vector<Id>& sources, Id resultTypeId) +{ + Id result = NoResult; + unsigned int numTargetComponents = getNumTypeComponents(resultTypeId); + unsigned int targetComponent = 0; + + // Special case: when calling a vector constructor with a single scalar + // argument, smear the scalar + if (sources.size() == 1 && isScalar(sources[0]) && numTargetComponents > 1) + return smearScalar(precision, sources[0], resultTypeId); + + // accumulate the arguments for OpCompositeConstruct + std::vector<Id> constituents; + Id scalarTypeId = getScalarTypeId(resultTypeId); + + // lambda to store the result of visiting an argument component + const auto latchResult = [&](Id comp) { + if (numTargetComponents > 1) + constituents.push_back(comp); + else + result = comp; + ++targetComponent; + }; + + // lambda to visit a vector argument's components + const auto accumulateVectorConstituents = [&](Id sourceArg) { + unsigned int sourceSize = getNumComponents(sourceArg); + unsigned int sourcesToUse = sourceSize; + if (sourcesToUse + targetComponent > numTargetComponents) + sourcesToUse = numTargetComponents - targetComponent; + + for (unsigned int s = 0; s < sourcesToUse; ++s) { + std::vector<unsigned> swiz; + swiz.push_back(s); + latchResult(createRvalueSwizzle(precision, scalarTypeId, sourceArg, swiz)); + } + }; + + // lambda to visit a matrix argument's components + const auto accumulateMatrixConstituents = [&](Id sourceArg) { + unsigned int sourceSize = getNumColumns(sourceArg) * getNumRows(sourceArg); + unsigned int sourcesToUse = sourceSize; + if (sourcesToUse + targetComponent > numTargetComponents) + sourcesToUse = numTargetComponents - targetComponent; + + int col = 0; + int row = 0; + for (unsigned int s = 0; s < sourcesToUse; ++s) { + if (row >= getNumRows(sourceArg)) { + row = 0; + col++; + } + std::vector<Id> indexes; + indexes.push_back(col); + indexes.push_back(row); + latchResult(createCompositeExtract(sourceArg, scalarTypeId, indexes)); + row++; + } + }; + + // Go through the source arguments, each one could have either + // a single or multiple components to contribute. + for (unsigned int i = 0; i < sources.size(); ++i) { + + if (isScalar(sources[i]) || isPointer(sources[i])) + latchResult(sources[i]); + else if (isVector(sources[i])) + accumulateVectorConstituents(sources[i]); + else if (isMatrix(sources[i])) + accumulateMatrixConstituents(sources[i]); + else + assert(0); + + if (targetComponent >= numTargetComponents) + break; + } + + // If the result is a vector, make it from the gathered constituents. + if (constituents.size() > 0) + result = createCompositeConstruct(resultTypeId, constituents); + + return setPrecision(result, precision); +} + +// Comments in header +Id Builder::createMatrixConstructor(Decoration precision, const std::vector<Id>& sources, Id resultTypeId) +{ + Id componentTypeId = getScalarTypeId(resultTypeId); + int numCols = getTypeNumColumns(resultTypeId); + int numRows = getTypeNumRows(resultTypeId); + + Instruction* instr = module.getInstruction(componentTypeId); + unsigned bitCount = instr->getImmediateOperand(0); + + // Optimize matrix constructed from a bigger matrix + if (isMatrix(sources[0]) && getNumColumns(sources[0]) >= numCols && getNumRows(sources[0]) >= numRows) { + // To truncate the matrix to a smaller number of rows/columns, we need to: + // 1. For each column, extract the column and truncate it to the required size using shuffle + // 2. Assemble the resulting matrix from all columns + Id matrix = sources[0]; + Id columnTypeId = getContainedTypeId(resultTypeId); + Id sourceColumnTypeId = getContainedTypeId(getTypeId(matrix)); + + std::vector<unsigned> channels; + for (int row = 0; row < numRows; ++row) + channels.push_back(row); + + std::vector<Id> matrixColumns; + for (int col = 0; col < numCols; ++col) { + std::vector<unsigned> indexes; + indexes.push_back(col); + Id colv = createCompositeExtract(matrix, sourceColumnTypeId, indexes); + setPrecision(colv, precision); + + if (numRows != getNumRows(matrix)) { + matrixColumns.push_back(createRvalueSwizzle(precision, columnTypeId, colv, channels)); + } else { + matrixColumns.push_back(colv); + } + } + + return setPrecision(createCompositeConstruct(resultTypeId, matrixColumns), precision); + } + + // Otherwise, will use a two step process + // 1. make a compile-time 2D array of values + // 2. construct a matrix from that array + + // Step 1. + + // initialize the array to the identity matrix + Id ids[maxMatrixSize][maxMatrixSize]; + Id one = (bitCount == 64 ? makeDoubleConstant(1.0) : makeFloatConstant(1.0)); + Id zero = (bitCount == 64 ? makeDoubleConstant(0.0) : makeFloatConstant(0.0)); + for (int col = 0; col < 4; ++col) { + for (int row = 0; row < 4; ++row) { + if (col == row) + ids[col][row] = one; + else + ids[col][row] = zero; + } + } + + // modify components as dictated by the arguments + if (sources.size() == 1 && isScalar(sources[0])) { + // a single scalar; resets the diagonals + for (int col = 0; col < 4; ++col) + ids[col][col] = sources[0]; + } else if (isMatrix(sources[0])) { + // constructing from another matrix; copy over the parts that exist in both the argument and constructee + Id matrix = sources[0]; + int minCols = std::min(numCols, getNumColumns(matrix)); + int minRows = std::min(numRows, getNumRows(matrix)); + for (int col = 0; col < minCols; ++col) { + std::vector<unsigned> indexes; + indexes.push_back(col); + for (int row = 0; row < minRows; ++row) { + indexes.push_back(row); + ids[col][row] = createCompositeExtract(matrix, componentTypeId, indexes); + indexes.pop_back(); + setPrecision(ids[col][row], precision); + } + } + } else { + // fill in the matrix in column-major order with whatever argument components are available + int row = 0; + int col = 0; + + for (int arg = 0; arg < (int)sources.size(); ++arg) { + Id argComp = sources[arg]; + for (int comp = 0; comp < getNumComponents(sources[arg]); ++comp) { + if (getNumComponents(sources[arg]) > 1) { + argComp = createCompositeExtract(sources[arg], componentTypeId, comp); + setPrecision(argComp, precision); + } + ids[col][row++] = argComp; + if (row == numRows) { + row = 0; + col++; + } + } + } + } + + // Step 2: Construct a matrix from that array. + // First make the column vectors, then make the matrix. + + // make the column vectors + Id columnTypeId = getContainedTypeId(resultTypeId); + std::vector<Id> matrixColumns; + for (int col = 0; col < numCols; ++col) { + std::vector<Id> vectorComponents; + for (int row = 0; row < numRows; ++row) + vectorComponents.push_back(ids[col][row]); + Id column = createCompositeConstruct(columnTypeId, vectorComponents); + setPrecision(column, precision); + matrixColumns.push_back(column); + } + + // make the matrix + return setPrecision(createCompositeConstruct(resultTypeId, matrixColumns), precision); +} + +// Comments in header +Builder::If::If(Id cond, unsigned int ctrl, Builder& gb) : + builder(gb), + condition(cond), + control(ctrl), + elseBlock(0) +{ + function = &builder.getBuildPoint()->getParent(); + + // make the blocks, but only put the then-block into the function, + // the else-block and merge-block will be added later, in order, after + // earlier code is emitted + thenBlock = new Block(builder.getUniqueId(), *function); + mergeBlock = new Block(builder.getUniqueId(), *function); + + // Save the current block, so that we can add in the flow control split when + // makeEndIf is called. + headerBlock = builder.getBuildPoint(); + + function->addBlock(thenBlock); + builder.setBuildPoint(thenBlock); +} + +// Comments in header +void Builder::If::makeBeginElse() +{ + // Close out the "then" by having it jump to the mergeBlock + builder.createBranch(mergeBlock); + + // Make the first else block and add it to the function + elseBlock = new Block(builder.getUniqueId(), *function); + function->addBlock(elseBlock); + + // Start building the else block + builder.setBuildPoint(elseBlock); +} + +// Comments in header +void Builder::If::makeEndIf() +{ + // jump to the merge block + builder.createBranch(mergeBlock); + + // Go back to the headerBlock and make the flow control split + builder.setBuildPoint(headerBlock); + builder.createSelectionMerge(mergeBlock, control); + if (elseBlock) + builder.createConditionalBranch(condition, thenBlock, elseBlock); + else + builder.createConditionalBranch(condition, thenBlock, mergeBlock); + + // add the merge block to the function + function->addBlock(mergeBlock); + builder.setBuildPoint(mergeBlock); +} + +// Comments in header +void Builder::makeSwitch(Id selector, unsigned int control, int numSegments, const std::vector<int>& caseValues, + const std::vector<int>& valueIndexToSegment, int defaultSegment, + std::vector<Block*>& segmentBlocks) +{ + Function& function = buildPoint->getParent(); + + // make all the blocks + for (int s = 0; s < numSegments; ++s) + segmentBlocks.push_back(new Block(getUniqueId(), function)); + + Block* mergeBlock = new Block(getUniqueId(), function); + + // make and insert the switch's selection-merge instruction + createSelectionMerge(mergeBlock, control); + + // make the switch instruction + Instruction* switchInst = new Instruction(NoResult, NoType, OpSwitch); + switchInst->addIdOperand(selector); + auto defaultOrMerge = (defaultSegment >= 0) ? segmentBlocks[defaultSegment] : mergeBlock; + switchInst->addIdOperand(defaultOrMerge->getId()); + defaultOrMerge->addPredecessor(buildPoint); + for (int i = 0; i < (int)caseValues.size(); ++i) { + switchInst->addImmediateOperand(caseValues[i]); + switchInst->addIdOperand(segmentBlocks[valueIndexToSegment[i]]->getId()); + segmentBlocks[valueIndexToSegment[i]]->addPredecessor(buildPoint); + } + buildPoint->addInstruction(std::unique_ptr<Instruction>(switchInst)); + + // push the merge block + switchMerges.push(mergeBlock); +} + +// Comments in header +void Builder::addSwitchBreak() +{ + // branch to the top of the merge block stack + createBranch(switchMerges.top()); + createAndSetNoPredecessorBlock("post-switch-break"); +} + +// Comments in header +void Builder::nextSwitchSegment(std::vector<Block*>& segmentBlock, int nextSegment) +{ + int lastSegment = nextSegment - 1; + if (lastSegment >= 0) { + // Close out previous segment by jumping, if necessary, to next segment + if (! buildPoint->isTerminated()) + createBranch(segmentBlock[nextSegment]); + } + Block* block = segmentBlock[nextSegment]; + block->getParent().addBlock(block); + setBuildPoint(block); +} + +// Comments in header +void Builder::endSwitch(std::vector<Block*>& /*segmentBlock*/) +{ + // Close out previous segment by jumping, if necessary, to next segment + if (! buildPoint->isTerminated()) + addSwitchBreak(); + + switchMerges.top()->getParent().addBlock(switchMerges.top()); + setBuildPoint(switchMerges.top()); + + switchMerges.pop(); +} + +Block& Builder::makeNewBlock() +{ + Function& function = buildPoint->getParent(); + auto block = new Block(getUniqueId(), function); + function.addBlock(block); + return *block; +} + +Builder::LoopBlocks& Builder::makeNewLoop() +{ + // This verbosity is needed to simultaneously get the same behavior + // everywhere (id's in the same order), have a syntax that works + // across lots of versions of C++, have no warnings from pedantic + // compilation modes, and leave the rest of the code alone. + Block& head = makeNewBlock(); + Block& body = makeNewBlock(); + Block& merge = makeNewBlock(); + Block& continue_target = makeNewBlock(); + LoopBlocks blocks(head, body, merge, continue_target); + loops.push(blocks); + return loops.top(); +} + +void Builder::createLoopContinue() +{ + createBranch(&loops.top().continue_target); + // Set up a block for dead code. + createAndSetNoPredecessorBlock("post-loop-continue"); +} + +void Builder::createLoopExit() +{ + createBranch(&loops.top().merge); + // Set up a block for dead code. + createAndSetNoPredecessorBlock("post-loop-break"); +} + +void Builder::closeLoop() +{ + loops.pop(); +} + +void Builder::clearAccessChain() +{ + accessChain.base = NoResult; + accessChain.indexChain.clear(); + accessChain.instr = NoResult; + accessChain.swizzle.clear(); + accessChain.component = NoResult; + accessChain.preSwizzleBaseType = NoType; + accessChain.isRValue = false; + accessChain.coherentFlags.clear(); + accessChain.alignment = 0; +} + +// Comments in header +void Builder::accessChainPushSwizzle(std::vector<unsigned>& swizzle, Id preSwizzleBaseType, AccessChain::CoherentFlags coherentFlags, unsigned int alignment) +{ + accessChain.coherentFlags |= coherentFlags; + accessChain.alignment |= alignment; + + // swizzles can be stacked in GLSL, but simplified to a single + // one here; the base type doesn't change + if (accessChain.preSwizzleBaseType == NoType) + accessChain.preSwizzleBaseType = preSwizzleBaseType; + + // if needed, propagate the swizzle for the current access chain + if (accessChain.swizzle.size() > 0) { + std::vector<unsigned> oldSwizzle = accessChain.swizzle; + accessChain.swizzle.resize(0); + for (unsigned int i = 0; i < swizzle.size(); ++i) { + assert(swizzle[i] < oldSwizzle.size()); + accessChain.swizzle.push_back(oldSwizzle[swizzle[i]]); + } + } else + accessChain.swizzle = swizzle; + + // determine if we need to track this swizzle anymore + simplifyAccessChainSwizzle(); +} + +// Comments in header +void Builder::accessChainStore(Id rvalue, spv::MemoryAccessMask memoryAccess, spv::Scope scope, unsigned int alignment) +{ + assert(accessChain.isRValue == false); + + transferAccessChainSwizzle(true); + Id base = collapseAccessChain(); + Id source = rvalue; + + // dynamic component should be gone + assert(accessChain.component == NoResult); + + // If swizzle still exists, it is out-of-order or not full, we must load the target vector, + // extract and insert elements to perform writeMask and/or swizzle. + if (accessChain.swizzle.size() > 0) { + Id tempBaseId = createLoad(base); + source = createLvalueSwizzle(getTypeId(tempBaseId), tempBaseId, source, accessChain.swizzle); + } + + // take LSB of alignment + alignment = alignment & ~(alignment & (alignment-1)); + if (getStorageClass(base) == StorageClassPhysicalStorageBufferEXT) { + memoryAccess = (spv::MemoryAccessMask)(memoryAccess | spv::MemoryAccessAlignedMask); + } + + createStore(source, base, memoryAccess, scope, alignment); +} + +// Comments in header +Id Builder::accessChainLoad(Decoration precision, Decoration nonUniform, Id resultType, spv::MemoryAccessMask memoryAccess, spv::Scope scope, unsigned int alignment) +{ + Id id; + + if (accessChain.isRValue) { + // transfer access chain, but try to stay in registers + transferAccessChainSwizzle(false); + if (accessChain.indexChain.size() > 0) { + Id swizzleBase = accessChain.preSwizzleBaseType != NoType ? accessChain.preSwizzleBaseType : resultType; + + // if all the accesses are constants, we can use OpCompositeExtract + std::vector<unsigned> indexes; + bool constant = true; + for (int i = 0; i < (int)accessChain.indexChain.size(); ++i) { + if (isConstantScalar(accessChain.indexChain[i])) + indexes.push_back(getConstantScalar(accessChain.indexChain[i])); + else { + constant = false; + break; + } + } + + if (constant) { + id = createCompositeExtract(accessChain.base, swizzleBase, indexes); + } else { + Id lValue = NoResult; + if (spvVersion >= Spv_1_4) { + // make a new function variable for this r-value, using an initializer, + // and mark it as NonWritable so that downstream it can be detected as a lookup + // table + lValue = createVariable(StorageClassFunction, getTypeId(accessChain.base), "indexable", + accessChain.base); + addDecoration(lValue, DecorationNonWritable); + } else { + lValue = createVariable(StorageClassFunction, getTypeId(accessChain.base), "indexable"); + // store into it + createStore(accessChain.base, lValue); + } + // move base to the new variable + accessChain.base = lValue; + accessChain.isRValue = false; + + // load through the access chain + id = createLoad(collapseAccessChain()); + } + setPrecision(id, precision); + } else + id = accessChain.base; // no precision, it was set when this was defined + } else { + transferAccessChainSwizzle(true); + + // take LSB of alignment + alignment = alignment & ~(alignment & (alignment-1)); + if (getStorageClass(accessChain.base) == StorageClassPhysicalStorageBufferEXT) { + memoryAccess = (spv::MemoryAccessMask)(memoryAccess | spv::MemoryAccessAlignedMask); + } + + // load through the access chain + id = createLoad(collapseAccessChain(), memoryAccess, scope, alignment); + setPrecision(id, precision); + addDecoration(id, nonUniform); + } + + // Done, unless there are swizzles to do + if (accessChain.swizzle.size() == 0 && accessChain.component == NoResult) + return id; + + // Do remaining swizzling + + // Do the basic swizzle + if (accessChain.swizzle.size() > 0) { + Id swizzledType = getScalarTypeId(getTypeId(id)); + if (accessChain.swizzle.size() > 1) + swizzledType = makeVectorType(swizzledType, (int)accessChain.swizzle.size()); + id = createRvalueSwizzle(precision, swizzledType, id, accessChain.swizzle); + } + + // Do the dynamic component + if (accessChain.component != NoResult) + id = setPrecision(createVectorExtractDynamic(id, resultType, accessChain.component), precision); + + addDecoration(id, nonUniform); + return id; +} + +Id Builder::accessChainGetLValue() +{ + assert(accessChain.isRValue == false); + + transferAccessChainSwizzle(true); + Id lvalue = collapseAccessChain(); + + // If swizzle exists, it is out-of-order or not full, we must load the target vector, + // extract and insert elements to perform writeMask and/or swizzle. This does not + // go with getting a direct l-value pointer. + assert(accessChain.swizzle.size() == 0); + assert(accessChain.component == NoResult); + + return lvalue; +} + +// comment in header +Id Builder::accessChainGetInferredType() +{ + // anything to operate on? + if (accessChain.base == NoResult) + return NoType; + Id type = getTypeId(accessChain.base); + + // do initial dereference + if (! accessChain.isRValue) + type = getContainedTypeId(type); + + // dereference each index + for (auto it = accessChain.indexChain.cbegin(); it != accessChain.indexChain.cend(); ++it) { + if (isStructType(type)) + type = getContainedTypeId(type, getConstantScalar(*it)); + else + type = getContainedTypeId(type); + } + + // dereference swizzle + if (accessChain.swizzle.size() == 1) + type = getContainedTypeId(type); + else if (accessChain.swizzle.size() > 1) + type = makeVectorType(getContainedTypeId(type), (int)accessChain.swizzle.size()); + + // dereference component selection + if (accessChain.component) + type = getContainedTypeId(type); + + return type; +} + +void Builder::dump(std::vector<unsigned int>& out) const +{ + // Header, before first instructions: + out.push_back(MagicNumber); + out.push_back(spvVersion); + out.push_back(builderNumber); + out.push_back(uniqueId + 1); + out.push_back(0); + + // Capabilities + for (auto it = capabilities.cbegin(); it != capabilities.cend(); ++it) { + Instruction capInst(0, 0, OpCapability); + capInst.addImmediateOperand(*it); + capInst.dump(out); + } + + for (auto it = extensions.cbegin(); it != extensions.cend(); ++it) { + Instruction extInst(0, 0, OpExtension); + extInst.addStringOperand(it->c_str()); + extInst.dump(out); + } + + dumpInstructions(out, imports); + Instruction memInst(0, 0, OpMemoryModel); + memInst.addImmediateOperand(addressModel); + memInst.addImmediateOperand(memoryModel); + memInst.dump(out); + + // Instructions saved up while building: + dumpInstructions(out, entryPoints); + dumpInstructions(out, executionModes); + + // Debug instructions + dumpInstructions(out, strings); + dumpSourceInstructions(out); + for (int e = 0; e < (int)sourceExtensions.size(); ++e) { + Instruction sourceExtInst(0, 0, OpSourceExtension); + sourceExtInst.addStringOperand(sourceExtensions[e]); + sourceExtInst.dump(out); + } + dumpInstructions(out, names); + dumpModuleProcesses(out); + + // Annotation instructions + dumpInstructions(out, decorations); + + dumpInstructions(out, constantsTypesGlobals); + dumpInstructions(out, externals); + + // The functions + module.dump(out); +} + +// +// Protected methods. +// + +// Turn the described access chain in 'accessChain' into an instruction(s) +// computing its address. This *cannot* include complex swizzles, which must +// be handled after this is called. +// +// Can generate code. +Id Builder::collapseAccessChain() +{ + assert(accessChain.isRValue == false); + + // did we already emit an access chain for this? + if (accessChain.instr != NoResult) + return accessChain.instr; + + // If we have a dynamic component, we can still transfer + // that into a final operand to the access chain. We need to remap the + // dynamic component through the swizzle to get a new dynamic component to + // update. + // + // This was not done in transferAccessChainSwizzle() because it might + // generate code. + remapDynamicSwizzle(); + if (accessChain.component != NoResult) { + // transfer the dynamic component to the access chain + accessChain.indexChain.push_back(accessChain.component); + accessChain.component = NoResult; + } + + // note that non-trivial swizzling is left pending + + // do we have an access chain? + if (accessChain.indexChain.size() == 0) + return accessChain.base; + + // emit the access chain + StorageClass storageClass = (StorageClass)module.getStorageClass(getTypeId(accessChain.base)); + accessChain.instr = createAccessChain(storageClass, accessChain.base, accessChain.indexChain); + + return accessChain.instr; +} + +// For a dynamic component selection of a swizzle. +// +// Turn the swizzle and dynamic component into just a dynamic component. +// +// Generates code. +void Builder::remapDynamicSwizzle() +{ + // do we have a swizzle to remap a dynamic component through? + if (accessChain.component != NoResult && accessChain.swizzle.size() > 1) { + // build a vector of the swizzle for the component to map into + std::vector<Id> components; + for (int c = 0; c < (int)accessChain.swizzle.size(); ++c) + components.push_back(makeUintConstant(accessChain.swizzle[c])); + Id mapType = makeVectorType(makeUintType(32), (int)accessChain.swizzle.size()); + Id map = makeCompositeConstant(mapType, components); + + // use it + accessChain.component = createVectorExtractDynamic(map, makeUintType(32), accessChain.component); + accessChain.swizzle.clear(); + } +} + +// clear out swizzle if it is redundant, that is reselecting the same components +// that would be present without the swizzle. +void Builder::simplifyAccessChainSwizzle() +{ + // If the swizzle has fewer components than the vector, it is subsetting, and must stay + // to preserve that fact. + if (getNumTypeComponents(accessChain.preSwizzleBaseType) > (int)accessChain.swizzle.size()) + return; + + // if components are out of order, it is a swizzle + for (unsigned int i = 0; i < accessChain.swizzle.size(); ++i) { + if (i != accessChain.swizzle[i]) + return; + } + + // otherwise, there is no need to track this swizzle + accessChain.swizzle.clear(); + if (accessChain.component == NoResult) + accessChain.preSwizzleBaseType = NoType; +} + +// To the extent any swizzling can become part of the chain +// of accesses instead of a post operation, make it so. +// If 'dynamic' is true, include transferring the dynamic component, +// otherwise, leave it pending. +// +// Does not generate code. just updates the access chain. +void Builder::transferAccessChainSwizzle(bool dynamic) +{ + // non existent? + if (accessChain.swizzle.size() == 0 && accessChain.component == NoResult) + return; + + // too complex? + // (this requires either a swizzle, or generating code for a dynamic component) + if (accessChain.swizzle.size() > 1) + return; + + // single component, either in the swizzle and/or dynamic component + if (accessChain.swizzle.size() == 1) { + assert(accessChain.component == NoResult); + // handle static component selection + accessChain.indexChain.push_back(makeUintConstant(accessChain.swizzle.front())); + accessChain.swizzle.clear(); + accessChain.preSwizzleBaseType = NoType; + } else if (dynamic && accessChain.component != NoResult) { + assert(accessChain.swizzle.size() == 0); + // handle dynamic component + accessChain.indexChain.push_back(accessChain.component); + accessChain.preSwizzleBaseType = NoType; + accessChain.component = NoResult; + } +} + +// Utility method for creating a new block and setting the insert point to +// be in it. This is useful for flow-control operations that need a "dummy" +// block proceeding them (e.g. instructions after a discard, etc). +void Builder::createAndSetNoPredecessorBlock(const char* /*name*/) +{ + Block* block = new Block(getUniqueId(), buildPoint->getParent()); + block->setUnreachable(); + buildPoint->getParent().addBlock(block); + setBuildPoint(block); + + // if (name) + // addName(block->getId(), name); +} + +// Comments in header +void Builder::createBranch(Block* block) +{ + Instruction* branch = new Instruction(OpBranch); + branch->addIdOperand(block->getId()); + buildPoint->addInstruction(std::unique_ptr<Instruction>(branch)); + block->addPredecessor(buildPoint); +} + +void Builder::createSelectionMerge(Block* mergeBlock, unsigned int control) +{ + Instruction* merge = new Instruction(OpSelectionMerge); + merge->addIdOperand(mergeBlock->getId()); + merge->addImmediateOperand(control); + buildPoint->addInstruction(std::unique_ptr<Instruction>(merge)); +} + +void Builder::createLoopMerge(Block* mergeBlock, Block* continueBlock, unsigned int control, + const std::vector<unsigned int>& operands) +{ + Instruction* merge = new Instruction(OpLoopMerge); + merge->addIdOperand(mergeBlock->getId()); + merge->addIdOperand(continueBlock->getId()); + merge->addImmediateOperand(control); + for (int op = 0; op < (int)operands.size(); ++op) + merge->addImmediateOperand(operands[op]); + buildPoint->addInstruction(std::unique_ptr<Instruction>(merge)); +} + +void Builder::createConditionalBranch(Id condition, Block* thenBlock, Block* elseBlock) +{ + Instruction* branch = new Instruction(OpBranchConditional); + branch->addIdOperand(condition); + branch->addIdOperand(thenBlock->getId()); + branch->addIdOperand(elseBlock->getId()); + buildPoint->addInstruction(std::unique_ptr<Instruction>(branch)); + thenBlock->addPredecessor(buildPoint); + elseBlock->addPredecessor(buildPoint); +} + +// OpSource +// [OpSourceContinued] +// ... +void Builder::dumpSourceInstructions(const spv::Id fileId, const std::string& text, + std::vector<unsigned int>& out) const +{ + const int maxWordCount = 0xFFFF; + const int opSourceWordCount = 4; + const int nonNullBytesPerInstruction = 4 * (maxWordCount - opSourceWordCount) - 1; + + if (source != SourceLanguageUnknown) { + // OpSource Language Version File Source + Instruction sourceInst(NoResult, NoType, OpSource); + sourceInst.addImmediateOperand(source); + sourceInst.addImmediateOperand(sourceVersion); + // File operand + if (fileId != NoResult) { + sourceInst.addIdOperand(fileId); + // Source operand + if (text.size() > 0) { + int nextByte = 0; + std::string subString; + while ((int)text.size() - nextByte > 0) { + subString = text.substr(nextByte, nonNullBytesPerInstruction); + if (nextByte == 0) { + // OpSource + sourceInst.addStringOperand(subString.c_str()); + sourceInst.dump(out); + } else { + // OpSourcContinued + Instruction sourceContinuedInst(OpSourceContinued); + sourceContinuedInst.addStringOperand(subString.c_str()); + sourceContinuedInst.dump(out); + } + nextByte += nonNullBytesPerInstruction; + } + } else + sourceInst.dump(out); + } else + sourceInst.dump(out); + } +} + +// Dump an OpSource[Continued] sequence for the source and every include file +void Builder::dumpSourceInstructions(std::vector<unsigned int>& out) const +{ + dumpSourceInstructions(sourceFileStringId, sourceText, out); + for (auto iItr = includeFiles.begin(); iItr != includeFiles.end(); ++iItr) + dumpSourceInstructions(iItr->first, *iItr->second, out); +} + +void Builder::dumpInstructions(std::vector<unsigned int>& out, const std::vector<std::unique_ptr<Instruction> >& instructions) const +{ + for (int i = 0; i < (int)instructions.size(); ++i) { + instructions[i]->dump(out); + } +} + +void Builder::dumpModuleProcesses(std::vector<unsigned int>& out) const +{ + for (int i = 0; i < (int)moduleProcesses.size(); ++i) { + Instruction moduleProcessed(OpModuleProcessed); + moduleProcessed.addStringOperand(moduleProcesses[i]); + moduleProcessed.dump(out); + } +} + +}; // end spv namespace diff --git a/thirdparty/glslang/SPIRV/SpvBuilder.h b/thirdparty/glslang/SPIRV/SpvBuilder.h new file mode 100644 index 0000000000..faed8e8230 --- /dev/null +++ b/thirdparty/glslang/SPIRV/SpvBuilder.h @@ -0,0 +1,758 @@ +// +// Copyright (C) 2014-2015 LunarG, Inc. +// Copyright (C) 2015-2018 Google, Inc. +// Copyright (C) 2017 ARM Limited. +// +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions +// are met: +// +// Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// +// Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following +// disclaimer in the documentation and/or other materials provided +// with the distribution. +// +// Neither the name of 3Dlabs Inc. Ltd. nor the names of its +// contributors may be used to endorse or promote products derived +// from this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS +// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE +// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, +// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, +// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; +// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN +// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +// POSSIBILITY OF SUCH DAMAGE. + +// +// "Builder" is an interface to fully build SPIR-V IR. Allocate one of +// these to build (a thread safe) internal SPIR-V representation (IR), +// and then dump it as a binary stream according to the SPIR-V specification. +// +// A Builder has a 1:1 relationship with a SPIR-V module. +// + +#pragma once +#ifndef SpvBuilder_H +#define SpvBuilder_H + +#include "Logger.h" +#include "spirv.hpp" +#include "spvIR.h" + +#include <algorithm> +#include <map> +#include <memory> +#include <set> +#include <sstream> +#include <stack> +#include <unordered_map> +#include <map> + +namespace spv { + +typedef enum { + Spv_1_0 = (1 << 16), + Spv_1_1 = (1 << 16) | (1 << 8), + Spv_1_2 = (1 << 16) | (2 << 8), + Spv_1_3 = (1 << 16) | (3 << 8), + Spv_1_4 = (1 << 16) | (4 << 8), +} SpvVersion; + +class Builder { +public: + Builder(unsigned int spvVersion, unsigned int userNumber, SpvBuildLogger* logger); + virtual ~Builder(); + + static const int maxMatrixSize = 4; + + unsigned int getSpvVersion() const { return spvVersion; } + + void setSource(spv::SourceLanguage lang, int version) + { + source = lang; + sourceVersion = version; + } + spv::Id getStringId(const std::string& str) + { + auto sItr = stringIds.find(str); + if (sItr != stringIds.end()) + return sItr->second; + spv::Id strId = getUniqueId(); + Instruction* fileString = new Instruction(strId, NoType, OpString); + const char* file_c_str = str.c_str(); + fileString->addStringOperand(file_c_str); + strings.push_back(std::unique_ptr<Instruction>(fileString)); + stringIds[file_c_str] = strId; + return strId; + } + void setSourceFile(const std::string& file) + { + sourceFileStringId = getStringId(file); + } + void setSourceText(const std::string& text) { sourceText = text; } + void addSourceExtension(const char* ext) { sourceExtensions.push_back(ext); } + void addModuleProcessed(const std::string& p) { moduleProcesses.push_back(p.c_str()); } + void setEmitOpLines() { emitOpLines = true; } + void addExtension(const char* ext) { extensions.insert(ext); } + void addInclude(const std::string& name, const std::string& text) + { + spv::Id incId = getStringId(name); + includeFiles[incId] = &text; + } + Id import(const char*); + void setMemoryModel(spv::AddressingModel addr, spv::MemoryModel mem) + { + addressModel = addr; + memoryModel = mem; + } + + void addCapability(spv::Capability cap) { capabilities.insert(cap); } + + // To get a new <id> for anything needing a new one. + Id getUniqueId() { return ++uniqueId; } + + // To get a set of new <id>s, e.g., for a set of function parameters + Id getUniqueIds(int numIds) + { + Id id = uniqueId + 1; + uniqueId += numIds; + return id; + } + + // Generate OpLine for non-filename-based #line directives (ie no filename + // seen yet): Log the current line, and if different than the last one, + // issue a new OpLine using the new line and current source file name. + void setLine(int line); + + // If filename null, generate OpLine for non-filename-based line directives, + // else do filename-based: Log the current line and file, and if different + // than the last one, issue a new OpLine using the new line and file + // name. + void setLine(int line, const char* filename); + // Low-level OpLine. See setLine() for a layered helper. + void addLine(Id fileName, int line, int column); + + // For creating new types (will return old type if the requested one was already made). + Id makeVoidType(); + Id makeBoolType(); + Id makePointer(StorageClass, Id pointee); + Id makeForwardPointer(StorageClass); + Id makePointerFromForwardPointer(StorageClass, Id forwardPointerType, Id pointee); + Id makeIntegerType(int width, bool hasSign); // generic + Id makeIntType(int width) { return makeIntegerType(width, true); } + Id makeUintType(int width) { return makeIntegerType(width, false); } + Id makeFloatType(int width); + Id makeStructType(const std::vector<Id>& members, const char*); + Id makeStructResultType(Id type0, Id type1); + Id makeVectorType(Id component, int size); + Id makeMatrixType(Id component, int cols, int rows); + Id makeArrayType(Id element, Id sizeId, int stride); // 0 stride means no stride decoration + Id makeRuntimeArray(Id element); + Id makeFunctionType(Id returnType, const std::vector<Id>& paramTypes); + Id makeImageType(Id sampledType, Dim, bool depth, bool arrayed, bool ms, unsigned sampled, ImageFormat format); + Id makeSamplerType(); + Id makeSampledImageType(Id imageType); + Id makeCooperativeMatrixType(Id component, Id scope, Id rows, Id cols); + + // accelerationStructureNV type + Id makeAccelerationStructureNVType(); + + // For querying about types. + Id getTypeId(Id resultId) const { return module.getTypeId(resultId); } + Id getDerefTypeId(Id resultId) const; + Op getOpCode(Id id) const { return module.getInstruction(id)->getOpCode(); } + Op getTypeClass(Id typeId) const { return getOpCode(typeId); } + Op getMostBasicTypeClass(Id typeId) const; + int getNumComponents(Id resultId) const { return getNumTypeComponents(getTypeId(resultId)); } + int getNumTypeConstituents(Id typeId) const; + int getNumTypeComponents(Id typeId) const { return getNumTypeConstituents(typeId); } + Id getScalarTypeId(Id typeId) const; + Id getContainedTypeId(Id typeId) const; + Id getContainedTypeId(Id typeId, int) const; + StorageClass getTypeStorageClass(Id typeId) const { return module.getStorageClass(typeId); } + ImageFormat getImageTypeFormat(Id typeId) const { return (ImageFormat)module.getInstruction(typeId)->getImmediateOperand(6); } + + bool isPointer(Id resultId) const { return isPointerType(getTypeId(resultId)); } + bool isScalar(Id resultId) const { return isScalarType(getTypeId(resultId)); } + bool isVector(Id resultId) const { return isVectorType(getTypeId(resultId)); } + bool isMatrix(Id resultId) const { return isMatrixType(getTypeId(resultId)); } + bool isCooperativeMatrix(Id resultId)const { return isCooperativeMatrixType(getTypeId(resultId)); } + bool isAggregate(Id resultId) const { return isAggregateType(getTypeId(resultId)); } + bool isSampledImage(Id resultId) const { return isSampledImageType(getTypeId(resultId)); } + + bool isBoolType(Id typeId) { return groupedTypes[OpTypeBool].size() > 0 && typeId == groupedTypes[OpTypeBool].back()->getResultId(); } + bool isIntType(Id typeId) const { return getTypeClass(typeId) == OpTypeInt && module.getInstruction(typeId)->getImmediateOperand(1) != 0; } + bool isUintType(Id typeId) const { return getTypeClass(typeId) == OpTypeInt && module.getInstruction(typeId)->getImmediateOperand(1) == 0; } + bool isFloatType(Id typeId) const { return getTypeClass(typeId) == OpTypeFloat; } + bool isPointerType(Id typeId) const { return getTypeClass(typeId) == OpTypePointer; } + bool isScalarType(Id typeId) const { return getTypeClass(typeId) == OpTypeFloat || getTypeClass(typeId) == OpTypeInt || getTypeClass(typeId) == OpTypeBool; } + bool isVectorType(Id typeId) const { return getTypeClass(typeId) == OpTypeVector; } + bool isMatrixType(Id typeId) const { return getTypeClass(typeId) == OpTypeMatrix; } + bool isStructType(Id typeId) const { return getTypeClass(typeId) == OpTypeStruct; } + bool isArrayType(Id typeId) const { return getTypeClass(typeId) == OpTypeArray; } + bool isCooperativeMatrixType(Id typeId)const { return getTypeClass(typeId) == OpTypeCooperativeMatrixNV; } + bool isAggregateType(Id typeId) const { return isArrayType(typeId) || isStructType(typeId) || isCooperativeMatrixType(typeId); } + bool isImageType(Id typeId) const { return getTypeClass(typeId) == OpTypeImage; } + bool isSamplerType(Id typeId) const { return getTypeClass(typeId) == OpTypeSampler; } + bool isSampledImageType(Id typeId) const { return getTypeClass(typeId) == OpTypeSampledImage; } + bool containsType(Id typeId, Op typeOp, unsigned int width) const; + bool containsPhysicalStorageBufferOrArray(Id typeId) const; + + bool isConstantOpCode(Op opcode) const; + bool isSpecConstantOpCode(Op opcode) const; + bool isConstant(Id resultId) const { return isConstantOpCode(getOpCode(resultId)); } + bool isConstantScalar(Id resultId) const { return getOpCode(resultId) == OpConstant; } + bool isSpecConstant(Id resultId) const { return isSpecConstantOpCode(getOpCode(resultId)); } + unsigned int getConstantScalar(Id resultId) const { return module.getInstruction(resultId)->getImmediateOperand(0); } + StorageClass getStorageClass(Id resultId) const { return getTypeStorageClass(getTypeId(resultId)); } + + int getScalarTypeWidth(Id typeId) const + { + Id scalarTypeId = getScalarTypeId(typeId); + assert(getTypeClass(scalarTypeId) == OpTypeInt || getTypeClass(scalarTypeId) == OpTypeFloat); + return module.getInstruction(scalarTypeId)->getImmediateOperand(0); + } + + int getTypeNumColumns(Id typeId) const + { + assert(isMatrixType(typeId)); + return getNumTypeConstituents(typeId); + } + int getNumColumns(Id resultId) const { return getTypeNumColumns(getTypeId(resultId)); } + int getTypeNumRows(Id typeId) const + { + assert(isMatrixType(typeId)); + return getNumTypeComponents(getContainedTypeId(typeId)); + } + int getNumRows(Id resultId) const { return getTypeNumRows(getTypeId(resultId)); } + + Dim getTypeDimensionality(Id typeId) const + { + assert(isImageType(typeId)); + return (Dim)module.getInstruction(typeId)->getImmediateOperand(1); + } + Id getImageType(Id resultId) const + { + Id typeId = getTypeId(resultId); + assert(isImageType(typeId) || isSampledImageType(typeId)); + return isSampledImageType(typeId) ? module.getInstruction(typeId)->getIdOperand(0) : typeId; + } + bool isArrayedImageType(Id typeId) const + { + assert(isImageType(typeId)); + return module.getInstruction(typeId)->getImmediateOperand(3) != 0; + } + + // For making new constants (will return old constant if the requested one was already made). + Id makeBoolConstant(bool b, bool specConstant = false); + Id makeInt8Constant(int i, bool specConstant = false) { return makeIntConstant(makeIntType(8), (unsigned)i, specConstant); } + Id makeUint8Constant(unsigned u, bool specConstant = false) { return makeIntConstant(makeUintType(8), u, specConstant); } + Id makeInt16Constant(int i, bool specConstant = false) { return makeIntConstant(makeIntType(16), (unsigned)i, specConstant); } + Id makeUint16Constant(unsigned u, bool specConstant = false) { return makeIntConstant(makeUintType(16), u, specConstant); } + Id makeIntConstant(int i, bool specConstant = false) { return makeIntConstant(makeIntType(32), (unsigned)i, specConstant); } + Id makeUintConstant(unsigned u, bool specConstant = false) { return makeIntConstant(makeUintType(32), u, specConstant); } + Id makeInt64Constant(long long i, bool specConstant = false) { return makeInt64Constant(makeIntType(64), (unsigned long long)i, specConstant); } + Id makeUint64Constant(unsigned long long u, bool specConstant = false) { return makeInt64Constant(makeUintType(64), u, specConstant); } + Id makeFloatConstant(float f, bool specConstant = false); + Id makeDoubleConstant(double d, bool specConstant = false); + Id makeFloat16Constant(float f16, bool specConstant = false); + Id makeFpConstant(Id type, double d, bool specConstant = false); + + // Turn the array of constants into a proper spv constant of the requested type. + Id makeCompositeConstant(Id type, const std::vector<Id>& comps, bool specConst = false); + + // Methods for adding information outside the CFG. + Instruction* addEntryPoint(ExecutionModel, Function*, const char* name); + void addExecutionMode(Function*, ExecutionMode mode, int value1 = -1, int value2 = -1, int value3 = -1); + void addName(Id, const char* name); + void addMemberName(Id, int member, const char* name); + void addDecoration(Id, Decoration, int num = -1); + void addDecoration(Id, Decoration, const char*); + void addDecorationId(Id id, Decoration, Id idDecoration); + void addMemberDecoration(Id, unsigned int member, Decoration, int num = -1); + void addMemberDecoration(Id, unsigned int member, Decoration, const char*); + + // At the end of what block do the next create*() instructions go? + void setBuildPoint(Block* bp) { buildPoint = bp; } + Block* getBuildPoint() const { return buildPoint; } + + // Make the entry-point function. The returned pointer is only valid + // for the lifetime of this builder. + Function* makeEntryPoint(const char*); + + // Make a shader-style function, and create its entry block if entry is non-zero. + // Return the function, pass back the entry. + // The returned pointer is only valid for the lifetime of this builder. + Function* makeFunctionEntry(Decoration precision, Id returnType, const char* name, const std::vector<Id>& paramTypes, + const std::vector<std::vector<Decoration>>& precisions, Block **entry = 0); + + // Create a return. An 'implicit' return is one not appearing in the source + // code. In the case of an implicit return, no post-return block is inserted. + void makeReturn(bool implicit, Id retVal = 0); + + // Generate all the code needed to finish up a function. + void leaveFunction(); + + // Create a discard. + void makeDiscard(); + + // Create a global or function local or IO variable. + Id createVariable(StorageClass, Id type, const char* name = 0, Id initializer = NoResult); + + // Create an intermediate with an undefined value. + Id createUndefined(Id type); + + // Store into an Id and return the l-value + void createStore(Id rValue, Id lValue, spv::MemoryAccessMask memoryAccess = spv::MemoryAccessMaskNone, spv::Scope scope = spv::ScopeMax, unsigned int alignment = 0); + + // Load from an Id and return it + Id createLoad(Id lValue, spv::MemoryAccessMask memoryAccess = spv::MemoryAccessMaskNone, spv::Scope scope = spv::ScopeMax, unsigned int alignment = 0); + + // Create an OpAccessChain instruction + Id createAccessChain(StorageClass, Id base, const std::vector<Id>& offsets); + + // Create an OpArrayLength instruction + Id createArrayLength(Id base, unsigned int member); + + // Create an OpCooperativeMatrixLengthNV instruction + Id createCooperativeMatrixLength(Id type); + + // Create an OpCompositeExtract instruction + Id createCompositeExtract(Id composite, Id typeId, unsigned index); + Id createCompositeExtract(Id composite, Id typeId, const std::vector<unsigned>& indexes); + Id createCompositeInsert(Id object, Id composite, Id typeId, unsigned index); + Id createCompositeInsert(Id object, Id composite, Id typeId, const std::vector<unsigned>& indexes); + + Id createVectorExtractDynamic(Id vector, Id typeId, Id componentIndex); + Id createVectorInsertDynamic(Id vector, Id typeId, Id component, Id componentIndex); + + void createNoResultOp(Op); + void createNoResultOp(Op, Id operand); + void createNoResultOp(Op, const std::vector<Id>& operands); + void createNoResultOp(Op, const std::vector<IdImmediate>& operands); + void createControlBarrier(Scope execution, Scope memory, MemorySemanticsMask); + void createMemoryBarrier(unsigned executionScope, unsigned memorySemantics); + Id createUnaryOp(Op, Id typeId, Id operand); + Id createBinOp(Op, Id typeId, Id operand1, Id operand2); + Id createTriOp(Op, Id typeId, Id operand1, Id operand2, Id operand3); + Id createOp(Op, Id typeId, const std::vector<Id>& operands); + Id createOp(Op, Id typeId, const std::vector<IdImmediate>& operands); + Id createFunctionCall(spv::Function*, const std::vector<spv::Id>&); + Id createSpecConstantOp(Op, Id typeId, const std::vector<spv::Id>& operands, const std::vector<unsigned>& literals); + + // Take an rvalue (source) and a set of channels to extract from it to + // make a new rvalue, which is returned. + Id createRvalueSwizzle(Decoration precision, Id typeId, Id source, const std::vector<unsigned>& channels); + + // Take a copy of an lvalue (target) and a source of components, and set the + // source components into the lvalue where the 'channels' say to put them. + // An updated version of the target is returned. + // (No true lvalue or stores are used.) + Id createLvalueSwizzle(Id typeId, Id target, Id source, const std::vector<unsigned>& channels); + + // If both the id and precision are valid, the id + // gets tagged with the requested precision. + // The passed in id is always the returned id, to simplify use patterns. + Id setPrecision(Id id, Decoration precision) + { + if (precision != NoPrecision && id != NoResult) + addDecoration(id, precision); + + return id; + } + + // Can smear a scalar to a vector for the following forms: + // - promoteScalar(scalar, vector) // smear scalar to width of vector + // - promoteScalar(vector, scalar) // smear scalar to width of vector + // - promoteScalar(pointer, scalar) // smear scalar to width of what pointer points to + // - promoteScalar(scalar, scalar) // do nothing + // Other forms are not allowed. + // + // Generally, the type of 'scalar' does not need to be the same type as the components in 'vector'. + // The type of the created vector is a vector of components of the same type as the scalar. + // + // Note: One of the arguments will change, with the result coming back that way rather than + // through the return value. + void promoteScalar(Decoration precision, Id& left, Id& right); + + // Make a value by smearing the scalar to fill the type. + // vectorType should be the correct type for making a vector of scalarVal. + // (No conversions are done.) + Id smearScalar(Decoration precision, Id scalarVal, Id vectorType); + + // Create a call to a built-in function. + Id createBuiltinCall(Id resultType, Id builtins, int entryPoint, const std::vector<Id>& args); + + // List of parameters used to create a texture operation + struct TextureParameters { + Id sampler; + Id coords; + Id bias; + Id lod; + Id Dref; + Id offset; + Id offsets; + Id gradX; + Id gradY; + Id sample; + Id component; + Id texelOut; + Id lodClamp; + Id granularity; + Id coarse; + bool nonprivate; + bool volatil; + }; + + // Select the correct texture operation based on all inputs, and emit the correct instruction + Id createTextureCall(Decoration precision, Id resultType, bool sparse, bool fetch, bool proj, bool gather, + bool noImplicit, const TextureParameters&, ImageOperandsMask); + + // Emit the OpTextureQuery* instruction that was passed in. + // Figure out the right return value and type, and return it. + Id createTextureQueryCall(Op, const TextureParameters&, bool isUnsignedResult); + + Id createSamplePositionCall(Decoration precision, Id, Id); + + Id createBitFieldExtractCall(Decoration precision, Id, Id, Id, bool isSigned); + Id createBitFieldInsertCall(Decoration precision, Id, Id, Id, Id); + + // Reduction comparison for composites: For equal and not-equal resulting in a scalar. + Id createCompositeCompare(Decoration precision, Id, Id, bool /* true if for equal, false if for not-equal */); + + // OpCompositeConstruct + Id createCompositeConstruct(Id typeId, const std::vector<Id>& constituents); + + // vector or scalar constructor + Id createConstructor(Decoration precision, const std::vector<Id>& sources, Id resultTypeId); + + // matrix constructor + Id createMatrixConstructor(Decoration precision, const std::vector<Id>& sources, Id constructee); + + // Helper to use for building nested control flow with if-then-else. + class If { + public: + If(Id condition, unsigned int ctrl, Builder& builder); + ~If() {} + + void makeBeginElse(); + void makeEndIf(); + + private: + If(const If&); + If& operator=(If&); + + Builder& builder; + Id condition; + unsigned int control; + Function* function; + Block* headerBlock; + Block* thenBlock; + Block* elseBlock; + Block* mergeBlock; + }; + + // Make a switch statement. A switch has 'numSegments' of pieces of code, not containing + // any case/default labels, all separated by one or more case/default labels. Each possible + // case value v is a jump to the caseValues[v] segment. The defaultSegment is also in this + // number space. How to compute the value is given by 'condition', as in switch(condition). + // + // The SPIR-V Builder will maintain the stack of post-switch merge blocks for nested switches. + // + // Use a defaultSegment < 0 if there is no default segment (to branch to post switch). + // + // Returns the right set of basic blocks to start each code segment with, so that the caller's + // recursion stack can hold the memory for it. + // + void makeSwitch(Id condition, unsigned int control, int numSegments, const std::vector<int>& caseValues, + const std::vector<int>& valueToSegment, int defaultSegment, std::vector<Block*>& segmentBB); // return argument + + // Add a branch to the innermost switch's merge block. + void addSwitchBreak(); + + // Move to the next code segment, passing in the return argument in makeSwitch() + void nextSwitchSegment(std::vector<Block*>& segmentBB, int segment); + + // Finish off the innermost switch. + void endSwitch(std::vector<Block*>& segmentBB); + + struct LoopBlocks { + LoopBlocks(Block& head, Block& body, Block& merge, Block& continue_target) : + head(head), body(body), merge(merge), continue_target(continue_target) { } + Block &head, &body, &merge, &continue_target; + private: + LoopBlocks(); + LoopBlocks& operator=(const LoopBlocks&); + }; + + // Start a new loop and prepare the builder to generate code for it. Until + // closeLoop() is called for this loop, createLoopContinue() and + // createLoopExit() will target its corresponding blocks. + LoopBlocks& makeNewLoop(); + + // Create a new block in the function containing the build point. Memory is + // owned by the function object. + Block& makeNewBlock(); + + // Add a branch to the continue_target of the current (innermost) loop. + void createLoopContinue(); + + // Add an exit (e.g. "break") from the innermost loop that we're currently + // in. + void createLoopExit(); + + // Close the innermost loop that you're in + void closeLoop(); + + // + // Access chain design for an R-Value vs. L-Value: + // + // There is a single access chain the builder is building at + // any particular time. Such a chain can be used to either to a load or + // a store, when desired. + // + // Expressions can be r-values, l-values, or both, or only r-values: + // a[b.c].d = .... // l-value + // ... = a[b.c].d; // r-value, that also looks like an l-value + // ++a[b.c].d; // r-value and l-value + // (x + y)[2]; // r-value only, can't possibly be l-value + // + // Computing an r-value means generating code. Hence, + // r-values should only be computed when they are needed, not speculatively. + // + // Computing an l-value means saving away information for later use in the compiler, + // no code is generated until the l-value is later dereferenced. It is okay + // to speculatively generate an l-value, just not okay to speculatively dereference it. + // + // The base of the access chain (the left-most variable or expression + // from which everything is based) can be set either as an l-value + // or as an r-value. Most efficient would be to set an l-value if one + // is available. If an expression was evaluated, the resulting r-value + // can be set as the chain base. + // + // The users of this single access chain can save and restore if they + // want to nest or manage multiple chains. + // + + struct AccessChain { + Id base; // for l-values, pointer to the base object, for r-values, the base object + std::vector<Id> indexChain; + Id instr; // cache the instruction that generates this access chain + std::vector<unsigned> swizzle; // each std::vector element selects the next GLSL component number + Id component; // a dynamic component index, can coexist with a swizzle, done after the swizzle, NoResult if not present + Id preSwizzleBaseType; // dereferenced type, before swizzle or component is applied; NoType unless a swizzle or component is present + bool isRValue; // true if 'base' is an r-value, otherwise, base is an l-value + unsigned int alignment; // bitwise OR of alignment values passed in. Accumulates worst alignment. Only tracks base and (optional) component selection alignment. + + // Accumulate whether anything in the chain of structures has coherent decorations. + struct CoherentFlags { + unsigned coherent : 1; + unsigned devicecoherent : 1; + unsigned queuefamilycoherent : 1; + unsigned workgroupcoherent : 1; + unsigned subgroupcoherent : 1; + unsigned nonprivate : 1; + unsigned volatil : 1; + unsigned isImage : 1; + + void clear() { + coherent = 0; + devicecoherent = 0; + queuefamilycoherent = 0; + workgroupcoherent = 0; + subgroupcoherent = 0; + nonprivate = 0; + volatil = 0; + isImage = 0; + } + + CoherentFlags() { clear(); } + CoherentFlags operator |=(const CoherentFlags &other) { + coherent |= other.coherent; + devicecoherent |= other.devicecoherent; + queuefamilycoherent |= other.queuefamilycoherent; + workgroupcoherent |= other.workgroupcoherent; + subgroupcoherent |= other.subgroupcoherent; + nonprivate |= other.nonprivate; + volatil |= other.volatil; + isImage |= other.isImage; + return *this; + } + }; + CoherentFlags coherentFlags; + }; + + // + // the SPIR-V builder maintains a single active chain that + // the following methods operate on + // + + // for external save and restore + AccessChain getAccessChain() { return accessChain; } + void setAccessChain(AccessChain newChain) { accessChain = newChain; } + + // clear accessChain + void clearAccessChain(); + + // set new base as an l-value base + void setAccessChainLValue(Id lValue) + { + assert(isPointer(lValue)); + accessChain.base = lValue; + } + + // set new base value as an r-value + void setAccessChainRValue(Id rValue) + { + accessChain.isRValue = true; + accessChain.base = rValue; + } + + // push offset onto the end of the chain + void accessChainPush(Id offset, AccessChain::CoherentFlags coherentFlags, unsigned int alignment) + { + accessChain.indexChain.push_back(offset); + accessChain.coherentFlags |= coherentFlags; + accessChain.alignment |= alignment; + } + + // push new swizzle onto the end of any existing swizzle, merging into a single swizzle + void accessChainPushSwizzle(std::vector<unsigned>& swizzle, Id preSwizzleBaseType, AccessChain::CoherentFlags coherentFlags, unsigned int alignment); + + // push a dynamic component selection onto the access chain, only applicable with a + // non-trivial swizzle or no swizzle + void accessChainPushComponent(Id component, Id preSwizzleBaseType, AccessChain::CoherentFlags coherentFlags, unsigned int alignment) + { + if (accessChain.swizzle.size() != 1) { + accessChain.component = component; + if (accessChain.preSwizzleBaseType == NoType) + accessChain.preSwizzleBaseType = preSwizzleBaseType; + } + accessChain.coherentFlags |= coherentFlags; + accessChain.alignment |= alignment; + } + + // use accessChain and swizzle to store value + void accessChainStore(Id rvalue, spv::MemoryAccessMask memoryAccess = spv::MemoryAccessMaskNone, spv::Scope scope = spv::ScopeMax, unsigned int alignment = 0); + + // use accessChain and swizzle to load an r-value + Id accessChainLoad(Decoration precision, Decoration nonUniform, Id ResultType, spv::MemoryAccessMask memoryAccess = spv::MemoryAccessMaskNone, spv::Scope scope = spv::ScopeMax, unsigned int alignment = 0); + + // get the direct pointer for an l-value + Id accessChainGetLValue(); + + // Get the inferred SPIR-V type of the result of the current access chain, + // based on the type of the base and the chain of dereferences. + Id accessChainGetInferredType(); + + // Add capabilities, extensions, remove unneeded decorations, etc., + // based on the resulting SPIR-V. + void postProcess(); + + // Hook to visit each instruction in a block in a function + void postProcess(Instruction&); + // Hook to visit each instruction in a reachable block in a function. + void postProcessReachable(const Instruction&); + // Hook to visit each non-32-bit sized float/int operation in a block. + void postProcessType(const Instruction&, spv::Id typeId); + + void dump(std::vector<unsigned int>&) const; + + void createBranch(Block* block); + void createConditionalBranch(Id condition, Block* thenBlock, Block* elseBlock); + void createLoopMerge(Block* mergeBlock, Block* continueBlock, unsigned int control, const std::vector<unsigned int>& operands); + + // Sets to generate opcode for specialization constants. + void setToSpecConstCodeGenMode() { generatingOpCodeForSpecConst = true; } + // Sets to generate opcode for non-specialization constants (normal mode). + void setToNormalCodeGenMode() { generatingOpCodeForSpecConst = false; } + // Check if the builder is generating code for spec constants. + bool isInSpecConstCodeGenMode() { return generatingOpCodeForSpecConst; } + + protected: + Id makeIntConstant(Id typeId, unsigned value, bool specConstant); + Id makeInt64Constant(Id typeId, unsigned long long value, bool specConstant); + Id findScalarConstant(Op typeClass, Op opcode, Id typeId, unsigned value); + Id findScalarConstant(Op typeClass, Op opcode, Id typeId, unsigned v1, unsigned v2); + Id findCompositeConstant(Op typeClass, Id typeId, const std::vector<Id>& comps); + Id findStructConstant(Id typeId, const std::vector<Id>& comps); + Id collapseAccessChain(); + void remapDynamicSwizzle(); + void transferAccessChainSwizzle(bool dynamic); + void simplifyAccessChainSwizzle(); + void createAndSetNoPredecessorBlock(const char*); + void createSelectionMerge(Block* mergeBlock, unsigned int control); + void dumpSourceInstructions(std::vector<unsigned int>&) const; + void dumpSourceInstructions(const spv::Id fileId, const std::string& text, std::vector<unsigned int>&) const; + void dumpInstructions(std::vector<unsigned int>&, const std::vector<std::unique_ptr<Instruction> >&) const; + void dumpModuleProcesses(std::vector<unsigned int>&) const; + spv::MemoryAccessMask sanitizeMemoryAccessForStorageClass(spv::MemoryAccessMask memoryAccess, StorageClass sc) const; + + unsigned int spvVersion; // the version of SPIR-V to emit in the header + SourceLanguage source; + int sourceVersion; + spv::Id sourceFileStringId; + std::string sourceText; + int currentLine; + const char* currentFile; + bool emitOpLines; + std::set<std::string> extensions; + std::vector<const char*> sourceExtensions; + std::vector<const char*> moduleProcesses; + AddressingModel addressModel; + MemoryModel memoryModel; + std::set<spv::Capability> capabilities; + int builderNumber; + Module module; + Block* buildPoint; + Id uniqueId; + Function* entryPointFunction; + bool generatingOpCodeForSpecConst; + AccessChain accessChain; + + // special blocks of instructions for output + std::vector<std::unique_ptr<Instruction> > strings; + std::vector<std::unique_ptr<Instruction> > imports; + std::vector<std::unique_ptr<Instruction> > entryPoints; + std::vector<std::unique_ptr<Instruction> > executionModes; + std::vector<std::unique_ptr<Instruction> > names; + std::vector<std::unique_ptr<Instruction> > decorations; + std::vector<std::unique_ptr<Instruction> > constantsTypesGlobals; + std::vector<std::unique_ptr<Instruction> > externals; + std::vector<std::unique_ptr<Function> > functions; + + // not output, internally used for quick & dirty canonical (unique) creation + std::unordered_map<unsigned int, std::vector<Instruction*>> groupedConstants; // map type opcodes to constant inst. + std::unordered_map<unsigned int, std::vector<Instruction*>> groupedStructConstants; // map struct-id to constant instructions + std::unordered_map<unsigned int, std::vector<Instruction*>> groupedTypes; // map type opcodes to type instructions + + // stack of switches + std::stack<Block*> switchMerges; + + // Our loop stack. + std::stack<LoopBlocks> loops; + + // map from strings to their string ids + std::unordered_map<std::string, spv::Id> stringIds; + + // map from include file name ids to their contents + std::map<spv::Id, const std::string*> includeFiles; + + // The stream for outputting warnings and errors. + SpvBuildLogger* logger; +}; // end Builder class + +}; // end spv namespace + +#endif // SpvBuilder_H diff --git a/thirdparty/glslang/SPIRV/SpvPostProcess.cpp b/thirdparty/glslang/SPIRV/SpvPostProcess.cpp new file mode 100644 index 0000000000..6e1f7cf61f --- /dev/null +++ b/thirdparty/glslang/SPIRV/SpvPostProcess.cpp @@ -0,0 +1,426 @@ +// +// Copyright (C) 2018 Google, Inc. +// +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions +// are met: +// +// Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// +// Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following +// disclaimer in the documentation and/or other materials provided +// with the distribution. +// +// Neither the name of 3Dlabs Inc. Ltd. nor the names of its +// contributors may be used to endorse or promote products derived +// from this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS +// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE +// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, +// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, +// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; +// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN +// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +// POSSIBILITY OF SUCH DAMAGE. + +// +// Post-processing for SPIR-V IR, in internal form, not standard binary form. +// + +#include <cassert> +#include <cstdlib> + +#include <unordered_set> +#include <algorithm> + +#include "SpvBuilder.h" + +#include "spirv.hpp" +#include "GlslangToSpv.h" +#include "SpvBuilder.h" +namespace spv { + #include "GLSL.std.450.h" + #include "GLSL.ext.KHR.h" + #include "GLSL.ext.EXT.h" +#ifdef AMD_EXTENSIONS + #include "GLSL.ext.AMD.h" +#endif +#ifdef NV_EXTENSIONS + #include "GLSL.ext.NV.h" +#endif +} + +namespace spv { + +// Hook to visit each operand type and result type of an instruction. +// Will be called multiple times for one instruction, once for each typed +// operand and the result. +void Builder::postProcessType(const Instruction& inst, Id typeId) +{ + // Characterize the type being questioned + Id basicTypeOp = getMostBasicTypeClass(typeId); + int width = 0; + if (basicTypeOp == OpTypeFloat || basicTypeOp == OpTypeInt) + width = getScalarTypeWidth(typeId); + + // Do opcode-specific checks + switch (inst.getOpCode()) { + case OpLoad: + case OpStore: + if (basicTypeOp == OpTypeStruct) { + if (containsType(typeId, OpTypeInt, 8)) + addCapability(CapabilityInt8); + if (containsType(typeId, OpTypeInt, 16)) + addCapability(CapabilityInt16); + if (containsType(typeId, OpTypeFloat, 16)) + addCapability(CapabilityFloat16); + } else { + StorageClass storageClass = getStorageClass(inst.getIdOperand(0)); + if (width == 8) { + switch (storageClass) { + case StorageClassPhysicalStorageBufferEXT: + case StorageClassUniform: + case StorageClassStorageBuffer: + case StorageClassPushConstant: + break; + default: + addCapability(CapabilityInt8); + break; + } + } else if (width == 16) { + switch (storageClass) { + case StorageClassPhysicalStorageBufferEXT: + case StorageClassUniform: + case StorageClassStorageBuffer: + case StorageClassPushConstant: + case StorageClassInput: + case StorageClassOutput: + break; + default: + if (basicTypeOp == OpTypeInt) + addCapability(CapabilityInt16); + if (basicTypeOp == OpTypeFloat) + addCapability(CapabilityFloat16); + break; + } + } + } + break; + case OpAccessChain: + case OpPtrAccessChain: + case OpCopyObject: + break; + case OpFConvert: + case OpSConvert: + case OpUConvert: + // Look for any 8/16-bit storage capabilities. If there are none, assume that + // the convert instruction requires the Float16/Int8/16 capability. + if (containsType(typeId, OpTypeFloat, 16) || containsType(typeId, OpTypeInt, 16)) { + bool foundStorage = false; + for (auto it = capabilities.begin(); it != capabilities.end(); ++it) { + spv::Capability cap = *it; + if (cap == spv::CapabilityStorageInputOutput16 || + cap == spv::CapabilityStoragePushConstant16 || + cap == spv::CapabilityStorageUniformBufferBlock16 || + cap == spv::CapabilityStorageUniform16) { + foundStorage = true; + break; + } + } + if (!foundStorage) { + if (containsType(typeId, OpTypeFloat, 16)) + addCapability(CapabilityFloat16); + if (containsType(typeId, OpTypeInt, 16)) + addCapability(CapabilityInt16); + } + } + if (containsType(typeId, OpTypeInt, 8)) { + bool foundStorage = false; + for (auto it = capabilities.begin(); it != capabilities.end(); ++it) { + spv::Capability cap = *it; + if (cap == spv::CapabilityStoragePushConstant8 || + cap == spv::CapabilityUniformAndStorageBuffer8BitAccess || + cap == spv::CapabilityStorageBuffer8BitAccess) { + foundStorage = true; + break; + } + } + if (!foundStorage) { + addCapability(CapabilityInt8); + } + } + break; + case OpExtInst: +#if AMD_EXTENSIONS + switch (inst.getImmediateOperand(1)) { + case GLSLstd450Frexp: + case GLSLstd450FrexpStruct: + if (getSpvVersion() < glslang::EShTargetSpv_1_3 && containsType(typeId, OpTypeInt, 16)) + addExtension(spv::E_SPV_AMD_gpu_shader_int16); + break; + case GLSLstd450InterpolateAtCentroid: + case GLSLstd450InterpolateAtSample: + case GLSLstd450InterpolateAtOffset: + if (getSpvVersion() < glslang::EShTargetSpv_1_3 && containsType(typeId, OpTypeFloat, 16)) + addExtension(spv::E_SPV_AMD_gpu_shader_half_float); + break; + default: + break; + } +#endif + break; + default: + if (basicTypeOp == OpTypeFloat && width == 16) + addCapability(CapabilityFloat16); + if (basicTypeOp == OpTypeInt && width == 16) + addCapability(CapabilityInt16); + if (basicTypeOp == OpTypeInt && width == 8) + addCapability(CapabilityInt8); + break; + } +} + +// Called for each instruction that resides in a block. +void Builder::postProcess(Instruction& inst) +{ + // Add capabilities based simply on the opcode. + switch (inst.getOpCode()) { + case OpExtInst: + switch (inst.getImmediateOperand(1)) { + case GLSLstd450InterpolateAtCentroid: + case GLSLstd450InterpolateAtSample: + case GLSLstd450InterpolateAtOffset: + addCapability(CapabilityInterpolationFunction); + break; + default: + break; + } + break; + case OpDPdxFine: + case OpDPdyFine: + case OpFwidthFine: + case OpDPdxCoarse: + case OpDPdyCoarse: + case OpFwidthCoarse: + addCapability(CapabilityDerivativeControl); + break; + + case OpImageQueryLod: + case OpImageQuerySize: + case OpImageQuerySizeLod: + case OpImageQuerySamples: + case OpImageQueryLevels: + addCapability(CapabilityImageQuery); + break; + +#ifdef NV_EXTENSIONS + case OpGroupNonUniformPartitionNV: + addExtension(E_SPV_NV_shader_subgroup_partitioned); + addCapability(CapabilityGroupNonUniformPartitionedNV); + break; +#endif + + case OpLoad: + case OpStore: + { + // For any load/store to a PhysicalStorageBufferEXT, walk the accesschain + // index list to compute the misalignment. The pre-existing alignment value + // (set via Builder::AccessChain::alignment) only accounts for the base of + // the reference type and any scalar component selection in the accesschain, + // and this function computes the rest from the SPIR-V Offset decorations. + Instruction *accessChain = module.getInstruction(inst.getIdOperand(0)); + if (accessChain->getOpCode() == OpAccessChain) { + Instruction *base = module.getInstruction(accessChain->getIdOperand(0)); + // Get the type of the base of the access chain. It must be a pointer type. + Id typeId = base->getTypeId(); + Instruction *type = module.getInstruction(typeId); + assert(type->getOpCode() == OpTypePointer); + if (type->getImmediateOperand(0) != StorageClassPhysicalStorageBufferEXT) { + break; + } + // Get the pointee type. + typeId = type->getIdOperand(1); + type = module.getInstruction(typeId); + // Walk the index list for the access chain. For each index, find any + // misalignment that can apply when accessing the member/element via + // Offset/ArrayStride/MatrixStride decorations, and bitwise OR them all + // together. + int alignment = 0; + for (int i = 1; i < accessChain->getNumOperands(); ++i) { + Instruction *idx = module.getInstruction(accessChain->getIdOperand(i)); + if (type->getOpCode() == OpTypeStruct) { + assert(idx->getOpCode() == OpConstant); + unsigned int c = idx->getImmediateOperand(0); + + const auto function = [&](const std::unique_ptr<Instruction>& decoration) { + if (decoration.get()->getOpCode() == OpMemberDecorate && + decoration.get()->getIdOperand(0) == typeId && + decoration.get()->getImmediateOperand(1) == c && + (decoration.get()->getImmediateOperand(2) == DecorationOffset || + decoration.get()->getImmediateOperand(2) == DecorationMatrixStride)) { + alignment |= decoration.get()->getImmediateOperand(3); + } + }; + std::for_each(decorations.begin(), decorations.end(), function); + // get the next member type + typeId = type->getIdOperand(c); + type = module.getInstruction(typeId); + } else if (type->getOpCode() == OpTypeArray || + type->getOpCode() == OpTypeRuntimeArray) { + const auto function = [&](const std::unique_ptr<Instruction>& decoration) { + if (decoration.get()->getOpCode() == OpDecorate && + decoration.get()->getIdOperand(0) == typeId && + decoration.get()->getImmediateOperand(1) == DecorationArrayStride) { + alignment |= decoration.get()->getImmediateOperand(2); + } + }; + std::for_each(decorations.begin(), decorations.end(), function); + // Get the element type + typeId = type->getIdOperand(0); + type = module.getInstruction(typeId); + } else { + // Once we get to any non-aggregate type, we're done. + break; + } + } + assert(inst.getNumOperands() >= 3); + unsigned int memoryAccess = inst.getImmediateOperand((inst.getOpCode() == OpStore) ? 2 : 1); + assert(memoryAccess & MemoryAccessAlignedMask); + static_cast<void>(memoryAccess); + // Compute the index of the alignment operand. + int alignmentIdx = 2; + if (inst.getOpCode() == OpStore) + alignmentIdx++; + // Merge new and old (mis)alignment + alignment |= inst.getImmediateOperand(alignmentIdx); + // Pick the LSB + alignment = alignment & ~(alignment & (alignment-1)); + // update the Aligned operand + inst.setImmediateOperand(alignmentIdx, alignment); + } + break; + } + + default: + break; + } + + // Checks based on type + if (inst.getTypeId() != NoType) + postProcessType(inst, inst.getTypeId()); + for (int op = 0; op < inst.getNumOperands(); ++op) { + if (inst.isIdOperand(op)) { + // In blocks, these are always result ids, but we are relying on + // getTypeId() to return NoType for things like OpLabel. + if (getTypeId(inst.getIdOperand(op)) != NoType) + postProcessType(inst, getTypeId(inst.getIdOperand(op))); + } + } +} + +// Called for each instruction in a reachable block. +void Builder::postProcessReachable(const Instruction&) +{ + // did have code here, but questionable to do so without deleting the instructions +} + +// comment in header +void Builder::postProcess() +{ + std::unordered_set<const Block*> reachableBlocks; + std::unordered_set<Id> unreachableDefinitions; + // Collect IDs defined in unreachable blocks. For each function, label the + // reachable blocks first. Then for each unreachable block, collect the + // result IDs of the instructions in it. + for (auto fi = module.getFunctions().cbegin(); fi != module.getFunctions().cend(); fi++) { + Function* f = *fi; + Block* entry = f->getEntryBlock(); + inReadableOrder(entry, [&reachableBlocks](const Block* b) { reachableBlocks.insert(b); }); + for (auto bi = f->getBlocks().cbegin(); bi != f->getBlocks().cend(); bi++) { + Block* b = *bi; + if (reachableBlocks.count(b) == 0) { + for (auto ii = b->getInstructions().cbegin(); ii != b->getInstructions().cend(); ii++) + unreachableDefinitions.insert(ii->get()->getResultId()); + } + } + } + + // Remove unneeded decorations, for unreachable instructions + decorations.erase(std::remove_if(decorations.begin(), decorations.end(), + [&unreachableDefinitions](std::unique_ptr<Instruction>& I) -> bool { + Id decoration_id = I.get()->getIdOperand(0); + return unreachableDefinitions.count(decoration_id) != 0; + }), + decorations.end()); + + // Add per-instruction capabilities, extensions, etc., + + // Look for any 8/16 bit type in physical storage buffer class, and set the + // appropriate capability. This happens in createSpvVariable for other storage + // classes, but there isn't always a variable for physical storage buffer. + for (int t = 0; t < (int)groupedTypes[OpTypePointer].size(); ++t) { + Instruction* type = groupedTypes[OpTypePointer][t]; + if (type->getImmediateOperand(0) == (unsigned)StorageClassPhysicalStorageBufferEXT) { + if (containsType(type->getIdOperand(1), OpTypeInt, 8)) { + addExtension(spv::E_SPV_KHR_8bit_storage); + addCapability(spv::CapabilityStorageBuffer8BitAccess); + } + if (containsType(type->getIdOperand(1), OpTypeInt, 16) || + containsType(type->getIdOperand(1), OpTypeFloat, 16)) { + addExtension(spv::E_SPV_KHR_16bit_storage); + addCapability(spv::CapabilityStorageBuffer16BitAccess); + } + } + } + + // process all reachable instructions... + for (auto bi = reachableBlocks.cbegin(); bi != reachableBlocks.cend(); ++bi) { + const Block* block = *bi; + const auto function = [this](const std::unique_ptr<Instruction>& inst) { postProcessReachable(*inst.get()); }; + std::for_each(block->getInstructions().begin(), block->getInstructions().end(), function); + } + + // process all block-contained instructions + for (auto fi = module.getFunctions().cbegin(); fi != module.getFunctions().cend(); fi++) { + Function* f = *fi; + for (auto bi = f->getBlocks().cbegin(); bi != f->getBlocks().cend(); bi++) { + Block* b = *bi; + for (auto ii = b->getInstructions().cbegin(); ii != b->getInstructions().cend(); ii++) + postProcess(*ii->get()); + + // For all local variables that contain pointers to PhysicalStorageBufferEXT, check whether + // there is an existing restrict/aliased decoration. If we don't find one, add Aliased as the + // default. + for (auto vi = b->getLocalVariables().cbegin(); vi != b->getLocalVariables().cend(); vi++) { + const Instruction& inst = *vi->get(); + Id resultId = inst.getResultId(); + if (containsPhysicalStorageBufferOrArray(getDerefTypeId(resultId))) { + bool foundDecoration = false; + const auto function = [&](const std::unique_ptr<Instruction>& decoration) { + if (decoration.get()->getIdOperand(0) == resultId && + decoration.get()->getOpCode() == OpDecorate && + (decoration.get()->getImmediateOperand(1) == spv::DecorationAliasedPointerEXT || + decoration.get()->getImmediateOperand(1) == spv::DecorationRestrictPointerEXT)) { + foundDecoration = true; + } + }; + std::for_each(decorations.begin(), decorations.end(), function); + if (!foundDecoration) { + addDecoration(resultId, spv::DecorationAliasedPointerEXT); + } + } + } + } + } +} + +}; // end spv namespace diff --git a/thirdparty/glslang/SPIRV/SpvTools.cpp b/thirdparty/glslang/SPIRV/SpvTools.cpp new file mode 100644 index 0000000000..db26d59089 --- /dev/null +++ b/thirdparty/glslang/SPIRV/SpvTools.cpp @@ -0,0 +1,214 @@ +// +// Copyright (C) 2014-2016 LunarG, Inc. +// Copyright (C) 2018 Google, Inc. +// +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions +// are met: +// +// Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// +// Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following +// disclaimer in the documentation and/or other materials provided +// with the distribution. +// +// Neither the name of 3Dlabs Inc. Ltd. nor the names of its +// contributors may be used to endorse or promote products derived +// from this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS +// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE +// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, +// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, +// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; +// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN +// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +// POSSIBILITY OF SUCH DAMAGE. + +// +// Call into SPIRV-Tools to disassemble, validate, and optimize. +// + +#if ENABLE_OPT + +#include <cstdio> +#include <iostream> + +#include "SpvTools.h" +#include "spirv-tools/optimizer.hpp" +#include "spirv-tools/libspirv.h" + +namespace glslang { + +// Translate glslang's view of target versioning to what SPIRV-Tools uses. +spv_target_env MapToSpirvToolsEnv(const SpvVersion& spvVersion, spv::SpvBuildLogger* logger) +{ + switch (spvVersion.vulkan) { + case glslang::EShTargetVulkan_1_0: + return spv_target_env::SPV_ENV_VULKAN_1_0; + case glslang::EShTargetVulkan_1_1: + switch (spvVersion.spv) { + case EShTargetSpv_1_0: + case EShTargetSpv_1_1: + case EShTargetSpv_1_2: + case EShTargetSpv_1_3: + return spv_target_env::SPV_ENV_VULKAN_1_1; + case EShTargetSpv_1_4: + return spv_target_env::SPV_ENV_VULKAN_1_1_SPIRV_1_4; + default: + logger->missingFunctionality("Target version for SPIRV-Tools validator"); + return spv_target_env::SPV_ENV_VULKAN_1_1; + } + default: + break; + } + + if (spvVersion.openGl > 0) + return spv_target_env::SPV_ENV_OPENGL_4_5; + + logger->missingFunctionality("Target version for SPIRV-Tools validator"); + return spv_target_env::SPV_ENV_UNIVERSAL_1_0; +} + + +// Use the SPIRV-Tools disassembler to print SPIR-V. +void SpirvToolsDisassemble(std::ostream& out, const std::vector<unsigned int>& spirv) +{ + // disassemble + spv_context context = spvContextCreate(SPV_ENV_UNIVERSAL_1_3); + spv_text text; + spv_diagnostic diagnostic = nullptr; + spvBinaryToText(context, spirv.data(), spirv.size(), + SPV_BINARY_TO_TEXT_OPTION_FRIENDLY_NAMES | SPV_BINARY_TO_TEXT_OPTION_INDENT, + &text, &diagnostic); + + // dump + if (diagnostic == nullptr) + out << text->str; + else + spvDiagnosticPrint(diagnostic); + + // teardown + spvDiagnosticDestroy(diagnostic); + spvContextDestroy(context); +} + +// Apply the SPIRV-Tools validator to generated SPIR-V. +void SpirvToolsValidate(const glslang::TIntermediate& intermediate, std::vector<unsigned int>& spirv, + spv::SpvBuildLogger* logger) +{ + // validate + spv_context context = spvContextCreate(MapToSpirvToolsEnv(intermediate.getSpv(), logger)); + spv_const_binary_t binary = { spirv.data(), spirv.size() }; + spv_diagnostic diagnostic = nullptr; + spv_validator_options options = spvValidatorOptionsCreate(); + spvValidatorOptionsSetRelaxBlockLayout(options, intermediate.usingHlslOffsets()); + spvValidateWithOptions(context, options, &binary, &diagnostic); + + // report + if (diagnostic != nullptr) { + logger->error("SPIRV-Tools Validation Errors"); + logger->error(diagnostic->error); + } + + // tear down + spvValidatorOptionsDestroy(options); + spvDiagnosticDestroy(diagnostic); + spvContextDestroy(context); +} + +// Apply the SPIRV-Tools optimizer to generated SPIR-V, for the purpose of +// legalizing HLSL SPIR-V. +void SpirvToolsLegalize(const glslang::TIntermediate&, std::vector<unsigned int>& spirv, + spv::SpvBuildLogger*, const SpvOptions* options) +{ + spv_target_env target_env = SPV_ENV_UNIVERSAL_1_2; + + spvtools::Optimizer optimizer(target_env); + optimizer.SetMessageConsumer( + [](spv_message_level_t level, const char *source, const spv_position_t &position, const char *message) { + auto &out = std::cerr; + switch (level) + { + case SPV_MSG_FATAL: + case SPV_MSG_INTERNAL_ERROR: + case SPV_MSG_ERROR: + out << "error: "; + break; + case SPV_MSG_WARNING: + out << "warning: "; + break; + case SPV_MSG_INFO: + case SPV_MSG_DEBUG: + out << "info: "; + break; + default: + break; + } + if (source) + { + out << source << ":"; + } + out << position.line << ":" << position.column << ":" << position.index << ":"; + if (message) + { + out << " " << message; + } + out << std::endl; + }); + + // If debug (specifically source line info) is being generated, propagate + // line information into all SPIR-V instructions. This avoids loss of + // information when instructions are deleted or moved. Later, remove + // redundant information to minimize final SPRIR-V size. + if (options->generateDebugInfo) { + optimizer.RegisterPass(spvtools::CreatePropagateLineInfoPass()); + } + optimizer.RegisterPass(spvtools::CreateDeadBranchElimPass()); + optimizer.RegisterPass(spvtools::CreateMergeReturnPass()); + optimizer.RegisterPass(spvtools::CreateInlineExhaustivePass()); + optimizer.RegisterPass(spvtools::CreateEliminateDeadFunctionsPass()); + optimizer.RegisterPass(spvtools::CreateScalarReplacementPass()); + optimizer.RegisterPass(spvtools::CreateLocalAccessChainConvertPass()); + optimizer.RegisterPass(spvtools::CreateLocalSingleBlockLoadStoreElimPass()); + optimizer.RegisterPass(spvtools::CreateLocalSingleStoreElimPass()); + optimizer.RegisterPass(spvtools::CreateSimplificationPass()); + optimizer.RegisterPass(spvtools::CreateAggressiveDCEPass()); + optimizer.RegisterPass(spvtools::CreateVectorDCEPass()); + optimizer.RegisterPass(spvtools::CreateDeadInsertElimPass()); + optimizer.RegisterPass(spvtools::CreateAggressiveDCEPass()); + optimizer.RegisterPass(spvtools::CreateDeadBranchElimPass()); + optimizer.RegisterPass(spvtools::CreateBlockMergePass()); + optimizer.RegisterPass(spvtools::CreateLocalMultiStoreElimPass()); + optimizer.RegisterPass(spvtools::CreateIfConversionPass()); + optimizer.RegisterPass(spvtools::CreateSimplificationPass()); + optimizer.RegisterPass(spvtools::CreateAggressiveDCEPass()); + optimizer.RegisterPass(spvtools::CreateVectorDCEPass()); + optimizer.RegisterPass(spvtools::CreateDeadInsertElimPass()); + if (options->optimizeSize) { + optimizer.RegisterPass(spvtools::CreateRedundancyEliminationPass()); + // TODO(greg-lunarg): Add this when AMD driver issues are resolved + // optimizer.RegisterPass(CreateCommonUniformElimPass()); + } + optimizer.RegisterPass(spvtools::CreateAggressiveDCEPass()); + optimizer.RegisterPass(spvtools::CreateCFGCleanupPass()); + if (options->generateDebugInfo) { + optimizer.RegisterPass(spvtools::CreateRedundantLineInfoElimPass()); + } + + spvtools::OptimizerOptions spvOptOptions; + spvOptOptions.set_run_validator(false); // The validator may run as a seperate step later on + optimizer.Run(spirv.data(), spirv.size(), &spirv, spvOptOptions); +} + +}; // end namespace glslang + +#endif diff --git a/thirdparty/glslang/SPIRV/SpvTools.h b/thirdparty/glslang/SPIRV/SpvTools.h new file mode 100644 index 0000000000..7e49ae0b30 --- /dev/null +++ b/thirdparty/glslang/SPIRV/SpvTools.h @@ -0,0 +1,80 @@ +// +// Copyright (C) 2014-2016 LunarG, Inc. +// Copyright (C) 2018 Google, Inc. +// +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions +// are met: +// +// Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// +// Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following +// disclaimer in the documentation and/or other materials provided +// with the distribution. +// +// Neither the name of 3Dlabs Inc. Ltd. nor the names of its +// contributors may be used to endorse or promote products derived +// from this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS +// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE +// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, +// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, +// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; +// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN +// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +// POSSIBILITY OF SUCH DAMAGE. + +// +// Call into SPIRV-Tools to disassemble, validate, and optimize. +// + +#pragma once +#ifndef GLSLANG_SPV_TOOLS_H +#define GLSLANG_SPV_TOOLS_H + +#include <vector> +#include <ostream> + +#include "../glslang/MachineIndependent/localintermediate.h" +#include "Logger.h" + +namespace glslang { + +struct SpvOptions { + SpvOptions() : generateDebugInfo(false), disableOptimizer(true), + optimizeSize(false), disassemble(false), validate(false) { } + bool generateDebugInfo; + bool disableOptimizer; + bool optimizeSize; + bool disassemble; + bool validate; +}; + +#if ENABLE_OPT + +// Use the SPIRV-Tools disassembler to print SPIR-V. +void SpirvToolsDisassemble(std::ostream& out, const std::vector<unsigned int>& spirv); + +// Apply the SPIRV-Tools validator to generated SPIR-V. +void SpirvToolsValidate(const glslang::TIntermediate& intermediate, std::vector<unsigned int>& spirv, + spv::SpvBuildLogger*); + +// Apply the SPIRV-Tools optimizer to generated SPIR-V, for the purpose of +// legalizing HLSL SPIR-V. +void SpirvToolsLegalize(const glslang::TIntermediate& intermediate, std::vector<unsigned int>& spirv, + spv::SpvBuildLogger*, const SpvOptions*); + +#endif + +} // end namespace glslang + +#endif // GLSLANG_SPV_TOOLS_H diff --git a/thirdparty/glslang/SPIRV/bitutils.h b/thirdparty/glslang/SPIRV/bitutils.h new file mode 100644 index 0000000000..22e44cec26 --- /dev/null +++ b/thirdparty/glslang/SPIRV/bitutils.h @@ -0,0 +1,81 @@ +// Copyright (c) 2015-2016 The Khronos Group Inc. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#ifndef LIBSPIRV_UTIL_BITUTILS_H_ +#define LIBSPIRV_UTIL_BITUTILS_H_ + +#include <cstdint> +#include <cstring> + +namespace spvutils { + +// Performs a bitwise copy of source to the destination type Dest. +template <typename Dest, typename Src> +Dest BitwiseCast(Src source) { + Dest dest; + static_assert(sizeof(source) == sizeof(dest), + "BitwiseCast: Source and destination must have the same size"); + std::memcpy(static_cast<void*>(&dest), &source, sizeof(dest)); + return dest; +} + +// SetBits<T, First, Num> returns an integer of type <T> with bits set +// for position <First> through <First + Num - 1>, counting from the least +// significant bit. In particular when Num == 0, no positions are set to 1. +// A static assert will be triggered if First + Num > sizeof(T) * 8, that is, +// a bit that will not fit in the underlying type is set. +template <typename T, size_t First = 0, size_t Num = 0> +struct SetBits { + static_assert(First < sizeof(T) * 8, + "Tried to set a bit that is shifted too far."); + const static T get = (T(1) << First) | SetBits<T, First + 1, Num - 1>::get; +}; + +template <typename T, size_t Last> +struct SetBits<T, Last, 0> { + const static T get = T(0); +}; + +// This is all compile-time so we can put our tests right here. +static_assert(SetBits<uint32_t, 0, 0>::get == uint32_t(0x00000000), + "SetBits failed"); +static_assert(SetBits<uint32_t, 0, 1>::get == uint32_t(0x00000001), + "SetBits failed"); +static_assert(SetBits<uint32_t, 31, 1>::get == uint32_t(0x80000000), + "SetBits failed"); +static_assert(SetBits<uint32_t, 1, 2>::get == uint32_t(0x00000006), + "SetBits failed"); +static_assert(SetBits<uint32_t, 30, 2>::get == uint32_t(0xc0000000), + "SetBits failed"); +static_assert(SetBits<uint32_t, 0, 31>::get == uint32_t(0x7FFFFFFF), + "SetBits failed"); +static_assert(SetBits<uint32_t, 0, 32>::get == uint32_t(0xFFFFFFFF), + "SetBits failed"); +static_assert(SetBits<uint32_t, 16, 16>::get == uint32_t(0xFFFF0000), + "SetBits failed"); + +static_assert(SetBits<uint64_t, 0, 1>::get == uint64_t(0x0000000000000001LL), + "SetBits failed"); +static_assert(SetBits<uint64_t, 63, 1>::get == uint64_t(0x8000000000000000LL), + "SetBits failed"); +static_assert(SetBits<uint64_t, 62, 2>::get == uint64_t(0xc000000000000000LL), + "SetBits failed"); +static_assert(SetBits<uint64_t, 31, 1>::get == uint64_t(0x0000000080000000LL), + "SetBits failed"); +static_assert(SetBits<uint64_t, 16, 16>::get == uint64_t(0x00000000FFFF0000LL), + "SetBits failed"); + +} // namespace spvutils + +#endif // LIBSPIRV_UTIL_BITUTILS_H_ diff --git a/thirdparty/glslang/SPIRV/disassemble.cpp b/thirdparty/glslang/SPIRV/disassemble.cpp new file mode 100644 index 0000000000..631173c0ec --- /dev/null +++ b/thirdparty/glslang/SPIRV/disassemble.cpp @@ -0,0 +1,759 @@ +// +// Copyright (C) 2014-2015 LunarG, Inc. +// +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions +// are met: +// +// Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// +// Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following +// disclaimer in the documentation and/or other materials provided +// with the distribution. +// +// Neither the name of 3Dlabs Inc. Ltd. nor the names of its +// contributors may be used to endorse or promote products derived +// from this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS +// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE +// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, +// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, +// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; +// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN +// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +// POSSIBILITY OF SUCH DAMAGE. + +// +// Disassembler for SPIR-V. +// + +#include <cstdlib> +#include <cstring> +#include <cassert> +#include <iomanip> +#include <stack> +#include <sstream> +#include <cstring> + +#include "disassemble.h" +#include "doc.h" +#include "SpvTools.h" + +namespace spv { + extern "C" { + // Include C-based headers that don't have a namespace + #include "GLSL.std.450.h" +#ifdef AMD_EXTENSIONS + #include "GLSL.ext.AMD.h" +#endif + +#ifdef NV_EXTENSIONS + #include "GLSL.ext.NV.h" +#endif + } +} +const char* GlslStd450DebugNames[spv::GLSLstd450Count]; + +namespace spv { + +#ifdef AMD_EXTENSIONS +static const char* GLSLextAMDGetDebugNames(const char*, unsigned); +#endif + +#ifdef NV_EXTENSIONS +static const char* GLSLextNVGetDebugNames(const char*, unsigned); +#endif + +static void Kill(std::ostream& out, const char* message) +{ + out << std::endl << "Disassembly failed: " << message << std::endl; + exit(1); +} + +// used to identify the extended instruction library imported when printing +enum ExtInstSet { + GLSL450Inst, + +#ifdef AMD_EXTENSIONS + GLSLextAMDInst, +#endif + +#ifdef NV_EXTENSIONS + GLSLextNVInst, +#endif + + OpenCLExtInst, +}; + +// Container class for a single instance of a SPIR-V stream, with methods for disassembly. +class SpirvStream { +public: + SpirvStream(std::ostream& out, const std::vector<unsigned int>& stream) : out(out), stream(stream), word(0), nextNestedControl(0) { } + virtual ~SpirvStream() { } + + void validate(); + void processInstructions(); + +protected: + SpirvStream(const SpirvStream&); + SpirvStream& operator=(const SpirvStream&); + Op getOpCode(int id) const { return idInstruction[id] ? (Op)(stream[idInstruction[id]] & OpCodeMask) : OpNop; } + + // Output methods + void outputIndent(); + void formatId(Id id, std::stringstream&); + void outputResultId(Id id); + void outputTypeId(Id id); + void outputId(Id id); + void outputMask(OperandClass operandClass, unsigned mask); + void disassembleImmediates(int numOperands); + void disassembleIds(int numOperands); + int disassembleString(); + void disassembleInstruction(Id resultId, Id typeId, Op opCode, int numOperands); + + // Data + std::ostream& out; // where to write the disassembly + const std::vector<unsigned int>& stream; // the actual word stream + int size; // the size of the word stream + int word; // the next word of the stream to read + + // map each <id> to the instruction that created it + Id bound; + std::vector<unsigned int> idInstruction; // the word offset into the stream where the instruction for result [id] starts; 0 if not yet seen (forward reference or function parameter) + + std::vector<std::string> idDescriptor; // the best text string known for explaining the <id> + + // schema + unsigned int schema; + + // stack of structured-merge points + std::stack<Id> nestedControl; + Id nextNestedControl; // need a slight delay for when we are nested +}; + +void SpirvStream::validate() +{ + size = (int)stream.size(); + if (size < 4) + Kill(out, "stream is too short"); + + // Magic number + if (stream[word++] != MagicNumber) { + out << "Bad magic number"; + return; + } + + // Version + out << "// Module Version " << std::hex << stream[word++] << std::endl; + + // Generator's magic number + out << "// Generated by (magic number): " << std::hex << stream[word++] << std::dec << std::endl; + + // Result <id> bound + bound = stream[word++]; + idInstruction.resize(bound); + idDescriptor.resize(bound); + out << "// Id's are bound by " << bound << std::endl; + out << std::endl; + + // Reserved schema, must be 0 for now + schema = stream[word++]; + if (schema != 0) + Kill(out, "bad schema, must be 0"); +} + +// Loop over all the instructions, in order, processing each. +// Boiler plate for each is handled here directly, the rest is dispatched. +void SpirvStream::processInstructions() +{ + // Instructions + while (word < size) { + int instructionStart = word; + + // Instruction wordCount and opcode + unsigned int firstWord = stream[word]; + unsigned wordCount = firstWord >> WordCountShift; + Op opCode = (Op)(firstWord & OpCodeMask); + int nextInst = word + wordCount; + ++word; + + // Presence of full instruction + if (nextInst > size) + Kill(out, "stream instruction terminated too early"); + + // Base for computing number of operands; will be updated as more is learned + unsigned numOperands = wordCount - 1; + + // Type <id> + Id typeId = 0; + if (InstructionDesc[opCode].hasType()) { + typeId = stream[word++]; + --numOperands; + } + + // Result <id> + Id resultId = 0; + if (InstructionDesc[opCode].hasResult()) { + resultId = stream[word++]; + --numOperands; + + // save instruction for future reference + idInstruction[resultId] = instructionStart; + } + + outputResultId(resultId); + outputTypeId(typeId); + outputIndent(); + + // Hand off the Op and all its operands + disassembleInstruction(resultId, typeId, opCode, numOperands); + if (word != nextInst) { + out << " ERROR, incorrect number of operands consumed. At " << word << " instead of " << nextInst << " instruction start was " << instructionStart; + word = nextInst; + } + out << std::endl; + } +} + +void SpirvStream::outputIndent() +{ + for (int i = 0; i < (int)nestedControl.size(); ++i) + out << " "; +} + +void SpirvStream::formatId(Id id, std::stringstream& idStream) +{ + if (id != 0) { + // On instructions with no IDs, this is called with "0", which does not + // have to be within ID bounds on null shaders. + if (id >= bound) + Kill(out, "Bad <id>"); + + idStream << id; + if (idDescriptor[id].size() > 0) + idStream << "(" << idDescriptor[id] << ")"; + } +} + +void SpirvStream::outputResultId(Id id) +{ + const int width = 16; + std::stringstream idStream; + formatId(id, idStream); + out << std::setw(width) << std::right << idStream.str(); + if (id != 0) + out << ":"; + else + out << " "; + + if (nestedControl.size() && id == nestedControl.top()) + nestedControl.pop(); +} + +void SpirvStream::outputTypeId(Id id) +{ + const int width = 12; + std::stringstream idStream; + formatId(id, idStream); + out << std::setw(width) << std::right << idStream.str() << " "; +} + +void SpirvStream::outputId(Id id) +{ + if (id >= bound) + Kill(out, "Bad <id>"); + + out << id; + if (idDescriptor[id].size() > 0) + out << "(" << idDescriptor[id] << ")"; +} + +void SpirvStream::outputMask(OperandClass operandClass, unsigned mask) +{ + if (mask == 0) + out << "None"; + else { + for (int m = 0; m < OperandClassParams[operandClass].ceiling; ++m) { + if (mask & (1 << m)) + out << OperandClassParams[operandClass].getName(m) << " "; + } + } +} + +void SpirvStream::disassembleImmediates(int numOperands) +{ + for (int i = 0; i < numOperands; ++i) { + out << stream[word++]; + if (i < numOperands - 1) + out << " "; + } +} + +void SpirvStream::disassembleIds(int numOperands) +{ + for (int i = 0; i < numOperands; ++i) { + outputId(stream[word++]); + if (i < numOperands - 1) + out << " "; + } +} + +// return the number of operands consumed by the string +int SpirvStream::disassembleString() +{ + int startWord = word; + + out << " \""; + + const char* wordString; + bool done = false; + do { + unsigned int content = stream[word]; + wordString = (const char*)&content; + for (int charCount = 0; charCount < 4; ++charCount) { + if (*wordString == 0) { + done = true; + break; + } + out << *(wordString++); + } + ++word; + } while (! done); + + out << "\""; + + return word - startWord; +} + +void SpirvStream::disassembleInstruction(Id resultId, Id /*typeId*/, Op opCode, int numOperands) +{ + // Process the opcode + + out << (OpcodeString(opCode) + 2); // leave out the "Op" + + if (opCode == OpLoopMerge || opCode == OpSelectionMerge) + nextNestedControl = stream[word]; + else if (opCode == OpBranchConditional || opCode == OpSwitch) { + if (nextNestedControl) { + nestedControl.push(nextNestedControl); + nextNestedControl = 0; + } + } else if (opCode == OpExtInstImport) { + idDescriptor[resultId] = (const char*)(&stream[word]); + } + else { + if (resultId != 0 && idDescriptor[resultId].size() == 0) { + switch (opCode) { + case OpTypeInt: + switch (stream[word]) { + case 8: idDescriptor[resultId] = "int8_t"; break; + case 16: idDescriptor[resultId] = "int16_t"; break; + default: assert(0); // fallthrough + case 32: idDescriptor[resultId] = "int"; break; + case 64: idDescriptor[resultId] = "int64_t"; break; + } + break; + case OpTypeFloat: + switch (stream[word]) { + case 16: idDescriptor[resultId] = "float16_t"; break; + default: assert(0); // fallthrough + case 32: idDescriptor[resultId] = "float"; break; + case 64: idDescriptor[resultId] = "float64_t"; break; + } + break; + case OpTypeBool: + idDescriptor[resultId] = "bool"; + break; + case OpTypeStruct: + idDescriptor[resultId] = "struct"; + break; + case OpTypePointer: + idDescriptor[resultId] = "ptr"; + break; + case OpTypeVector: + if (idDescriptor[stream[word]].size() > 0) { + idDescriptor[resultId].append(idDescriptor[stream[word]].begin(), idDescriptor[stream[word]].begin() + 1); + if (strstr(idDescriptor[stream[word]].c_str(), "8")) { + idDescriptor[resultId].append("8"); + } + if (strstr(idDescriptor[stream[word]].c_str(), "16")) { + idDescriptor[resultId].append("16"); + } + if (strstr(idDescriptor[stream[word]].c_str(), "64")) { + idDescriptor[resultId].append("64"); + } + } + idDescriptor[resultId].append("vec"); + switch (stream[word + 1]) { + case 2: idDescriptor[resultId].append("2"); break; + case 3: idDescriptor[resultId].append("3"); break; + case 4: idDescriptor[resultId].append("4"); break; + case 8: idDescriptor[resultId].append("8"); break; + case 16: idDescriptor[resultId].append("16"); break; + case 32: idDescriptor[resultId].append("32"); break; + default: break; + } + break; + default: + break; + } + } + } + + // Process the operands. Note, a new context-dependent set could be + // swapped in mid-traversal. + + // Handle images specially, so can put out helpful strings. + if (opCode == OpTypeImage) { + out << " "; + disassembleIds(1); + out << " " << DimensionString((Dim)stream[word++]); + out << (stream[word++] != 0 ? " depth" : ""); + out << (stream[word++] != 0 ? " array" : ""); + out << (stream[word++] != 0 ? " multi-sampled" : ""); + switch (stream[word++]) { + case 0: out << " runtime"; break; + case 1: out << " sampled"; break; + case 2: out << " nonsampled"; break; + } + out << " format:" << ImageFormatString((ImageFormat)stream[word++]); + + if (numOperands == 8) { + out << " " << AccessQualifierString(stream[word++]); + } + return; + } + + // Handle all the parameterized operands + for (int op = 0; op < InstructionDesc[opCode].operands.getNum() && numOperands > 0; ++op) { + out << " "; + OperandClass operandClass = InstructionDesc[opCode].operands.getClass(op); + switch (operandClass) { + case OperandId: + case OperandScope: + case OperandMemorySemantics: + disassembleIds(1); + --numOperands; + // Get names for printing "(XXX)" for readability, *after* this id + if (opCode == OpName) + idDescriptor[stream[word - 1]] = (const char*)(&stream[word]); + break; + case OperandVariableIds: + disassembleIds(numOperands); + return; + case OperandImageOperands: + outputMask(OperandImageOperands, stream[word++]); + --numOperands; + disassembleIds(numOperands); + return; + case OperandOptionalLiteral: + case OperandVariableLiterals: + if ((opCode == OpDecorate && stream[word - 1] == DecorationBuiltIn) || + (opCode == OpMemberDecorate && stream[word - 1] == DecorationBuiltIn)) { + out << BuiltInString(stream[word++]); + --numOperands; + ++op; + } + disassembleImmediates(numOperands); + return; + case OperandVariableIdLiteral: + while (numOperands > 0) { + out << std::endl; + outputResultId(0); + outputTypeId(0); + outputIndent(); + out << " Type "; + disassembleIds(1); + out << ", member "; + disassembleImmediates(1); + numOperands -= 2; + } + return; + case OperandVariableLiteralId: + while (numOperands > 0) { + out << std::endl; + outputResultId(0); + outputTypeId(0); + outputIndent(); + out << " case "; + disassembleImmediates(1); + out << ": "; + disassembleIds(1); + numOperands -= 2; + } + return; + case OperandLiteralNumber: + disassembleImmediates(1); + --numOperands; + if (opCode == OpExtInst) { + ExtInstSet extInstSet = GLSL450Inst; + const char* name = idDescriptor[stream[word - 2]].c_str(); + if (0 == memcmp("OpenCL", name, 6)) { + extInstSet = OpenCLExtInst; +#ifdef AMD_EXTENSIONS + } else if (strcmp(spv::E_SPV_AMD_shader_ballot, name) == 0 || + strcmp(spv::E_SPV_AMD_shader_trinary_minmax, name) == 0 || + strcmp(spv::E_SPV_AMD_shader_explicit_vertex_parameter, name) == 0 || + strcmp(spv::E_SPV_AMD_gcn_shader, name) == 0) { + extInstSet = GLSLextAMDInst; +#endif +#ifdef NV_EXTENSIONS + }else if (strcmp(spv::E_SPV_NV_sample_mask_override_coverage, name) == 0 || + strcmp(spv::E_SPV_NV_geometry_shader_passthrough, name) == 0 || + strcmp(spv::E_SPV_NV_viewport_array2, name) == 0 || + strcmp(spv::E_SPV_NVX_multiview_per_view_attributes, name) == 0 || + strcmp(spv::E_SPV_NV_fragment_shader_barycentric, name) == 0 || + strcmp(spv::E_SPV_NV_mesh_shader, name) == 0) { + extInstSet = GLSLextNVInst; +#endif + } + unsigned entrypoint = stream[word - 1]; + if (extInstSet == GLSL450Inst) { + if (entrypoint < GLSLstd450Count) { + out << "(" << GlslStd450DebugNames[entrypoint] << ")"; + } +#ifdef AMD_EXTENSIONS + } else if (extInstSet == GLSLextAMDInst) { + out << "(" << GLSLextAMDGetDebugNames(name, entrypoint) << ")"; +#endif +#ifdef NV_EXTENSIONS + } + else if (extInstSet == GLSLextNVInst) { + out << "(" << GLSLextNVGetDebugNames(name, entrypoint) << ")"; +#endif + } + } + break; + case OperandOptionalLiteralString: + case OperandLiteralString: + numOperands -= disassembleString(); + break; + case OperandMemoryAccess: + outputMask(OperandMemoryAccess, stream[word++]); + --numOperands; + // Aligned is the only memory access operand that uses an immediate + // value, and it is also the first operand that uses a value at all. + if (stream[word-1] & MemoryAccessAlignedMask) { + disassembleImmediates(1); + numOperands--; + if (numOperands) + out << " "; + } + disassembleIds(numOperands); + return; + default: + assert(operandClass >= OperandSource && operandClass < OperandOpcode); + + if (OperandClassParams[operandClass].bitmask) + outputMask(operandClass, stream[word++]); + else + out << OperandClassParams[operandClass].getName(stream[word++]); + --numOperands; + + break; + } + } + + return; +} + +static void GLSLstd450GetDebugNames(const char** names) +{ + for (int i = 0; i < GLSLstd450Count; ++i) + names[i] = "Unknown"; + + names[GLSLstd450Round] = "Round"; + names[GLSLstd450RoundEven] = "RoundEven"; + names[GLSLstd450Trunc] = "Trunc"; + names[GLSLstd450FAbs] = "FAbs"; + names[GLSLstd450SAbs] = "SAbs"; + names[GLSLstd450FSign] = "FSign"; + names[GLSLstd450SSign] = "SSign"; + names[GLSLstd450Floor] = "Floor"; + names[GLSLstd450Ceil] = "Ceil"; + names[GLSLstd450Fract] = "Fract"; + names[GLSLstd450Radians] = "Radians"; + names[GLSLstd450Degrees] = "Degrees"; + names[GLSLstd450Sin] = "Sin"; + names[GLSLstd450Cos] = "Cos"; + names[GLSLstd450Tan] = "Tan"; + names[GLSLstd450Asin] = "Asin"; + names[GLSLstd450Acos] = "Acos"; + names[GLSLstd450Atan] = "Atan"; + names[GLSLstd450Sinh] = "Sinh"; + names[GLSLstd450Cosh] = "Cosh"; + names[GLSLstd450Tanh] = "Tanh"; + names[GLSLstd450Asinh] = "Asinh"; + names[GLSLstd450Acosh] = "Acosh"; + names[GLSLstd450Atanh] = "Atanh"; + names[GLSLstd450Atan2] = "Atan2"; + names[GLSLstd450Pow] = "Pow"; + names[GLSLstd450Exp] = "Exp"; + names[GLSLstd450Log] = "Log"; + names[GLSLstd450Exp2] = "Exp2"; + names[GLSLstd450Log2] = "Log2"; + names[GLSLstd450Sqrt] = "Sqrt"; + names[GLSLstd450InverseSqrt] = "InverseSqrt"; + names[GLSLstd450Determinant] = "Determinant"; + names[GLSLstd450MatrixInverse] = "MatrixInverse"; + names[GLSLstd450Modf] = "Modf"; + names[GLSLstd450ModfStruct] = "ModfStruct"; + names[GLSLstd450FMin] = "FMin"; + names[GLSLstd450SMin] = "SMin"; + names[GLSLstd450UMin] = "UMin"; + names[GLSLstd450FMax] = "FMax"; + names[GLSLstd450SMax] = "SMax"; + names[GLSLstd450UMax] = "UMax"; + names[GLSLstd450FClamp] = "FClamp"; + names[GLSLstd450SClamp] = "SClamp"; + names[GLSLstd450UClamp] = "UClamp"; + names[GLSLstd450FMix] = "FMix"; + names[GLSLstd450Step] = "Step"; + names[GLSLstd450SmoothStep] = "SmoothStep"; + names[GLSLstd450Fma] = "Fma"; + names[GLSLstd450Frexp] = "Frexp"; + names[GLSLstd450FrexpStruct] = "FrexpStruct"; + names[GLSLstd450Ldexp] = "Ldexp"; + names[GLSLstd450PackSnorm4x8] = "PackSnorm4x8"; + names[GLSLstd450PackUnorm4x8] = "PackUnorm4x8"; + names[GLSLstd450PackSnorm2x16] = "PackSnorm2x16"; + names[GLSLstd450PackUnorm2x16] = "PackUnorm2x16"; + names[GLSLstd450PackHalf2x16] = "PackHalf2x16"; + names[GLSLstd450PackDouble2x32] = "PackDouble2x32"; + names[GLSLstd450UnpackSnorm2x16] = "UnpackSnorm2x16"; + names[GLSLstd450UnpackUnorm2x16] = "UnpackUnorm2x16"; + names[GLSLstd450UnpackHalf2x16] = "UnpackHalf2x16"; + names[GLSLstd450UnpackSnorm4x8] = "UnpackSnorm4x8"; + names[GLSLstd450UnpackUnorm4x8] = "UnpackUnorm4x8"; + names[GLSLstd450UnpackDouble2x32] = "UnpackDouble2x32"; + names[GLSLstd450Length] = "Length"; + names[GLSLstd450Distance] = "Distance"; + names[GLSLstd450Cross] = "Cross"; + names[GLSLstd450Normalize] = "Normalize"; + names[GLSLstd450FaceForward] = "FaceForward"; + names[GLSLstd450Reflect] = "Reflect"; + names[GLSLstd450Refract] = "Refract"; + names[GLSLstd450FindILsb] = "FindILsb"; + names[GLSLstd450FindSMsb] = "FindSMsb"; + names[GLSLstd450FindUMsb] = "FindUMsb"; + names[GLSLstd450InterpolateAtCentroid] = "InterpolateAtCentroid"; + names[GLSLstd450InterpolateAtSample] = "InterpolateAtSample"; + names[GLSLstd450InterpolateAtOffset] = "InterpolateAtOffset"; +} + +#ifdef AMD_EXTENSIONS +static const char* GLSLextAMDGetDebugNames(const char* name, unsigned entrypoint) +{ + if (strcmp(name, spv::E_SPV_AMD_shader_ballot) == 0) { + switch (entrypoint) { + case SwizzleInvocationsAMD: return "SwizzleInvocationsAMD"; + case SwizzleInvocationsMaskedAMD: return "SwizzleInvocationsMaskedAMD"; + case WriteInvocationAMD: return "WriteInvocationAMD"; + case MbcntAMD: return "MbcntAMD"; + default: return "Bad"; + } + } else if (strcmp(name, spv::E_SPV_AMD_shader_trinary_minmax) == 0) { + switch (entrypoint) { + case FMin3AMD: return "FMin3AMD"; + case UMin3AMD: return "UMin3AMD"; + case SMin3AMD: return "SMin3AMD"; + case FMax3AMD: return "FMax3AMD"; + case UMax3AMD: return "UMax3AMD"; + case SMax3AMD: return "SMax3AMD"; + case FMid3AMD: return "FMid3AMD"; + case UMid3AMD: return "UMid3AMD"; + case SMid3AMD: return "SMid3AMD"; + default: return "Bad"; + } + } else if (strcmp(name, spv::E_SPV_AMD_shader_explicit_vertex_parameter) == 0) { + switch (entrypoint) { + case InterpolateAtVertexAMD: return "InterpolateAtVertexAMD"; + default: return "Bad"; + } + } + else if (strcmp(name, spv::E_SPV_AMD_gcn_shader) == 0) { + switch (entrypoint) { + case CubeFaceIndexAMD: return "CubeFaceIndexAMD"; + case CubeFaceCoordAMD: return "CubeFaceCoordAMD"; + case TimeAMD: return "TimeAMD"; + default: + break; + } + } + + return "Bad"; +} +#endif + +#ifdef NV_EXTENSIONS +static const char* GLSLextNVGetDebugNames(const char* name, unsigned entrypoint) +{ + if (strcmp(name, spv::E_SPV_NV_sample_mask_override_coverage) == 0 || + strcmp(name, spv::E_SPV_NV_geometry_shader_passthrough) == 0 || + strcmp(name, spv::E_ARB_shader_viewport_layer_array) == 0 || + strcmp(name, spv::E_SPV_NV_viewport_array2) == 0 || + strcmp(spv::E_SPV_NVX_multiview_per_view_attributes, name) == 0 || + strcmp(spv::E_SPV_NV_fragment_shader_barycentric, name) == 0 || + strcmp(name, spv::E_SPV_NV_mesh_shader) == 0) { + switch (entrypoint) { + // NV builtins + case BuiltInViewportMaskNV: return "ViewportMaskNV"; + case BuiltInSecondaryPositionNV: return "SecondaryPositionNV"; + case BuiltInSecondaryViewportMaskNV: return "SecondaryViewportMaskNV"; + case BuiltInPositionPerViewNV: return "PositionPerViewNV"; + case BuiltInViewportMaskPerViewNV: return "ViewportMaskPerViewNV"; + case BuiltInBaryCoordNV: return "BaryCoordNV"; + case BuiltInBaryCoordNoPerspNV: return "BaryCoordNoPerspNV"; + case BuiltInTaskCountNV: return "TaskCountNV"; + case BuiltInPrimitiveCountNV: return "PrimitiveCountNV"; + case BuiltInPrimitiveIndicesNV: return "PrimitiveIndicesNV"; + case BuiltInClipDistancePerViewNV: return "ClipDistancePerViewNV"; + case BuiltInCullDistancePerViewNV: return "CullDistancePerViewNV"; + case BuiltInLayerPerViewNV: return "LayerPerViewNV"; + case BuiltInMeshViewCountNV: return "MeshViewCountNV"; + case BuiltInMeshViewIndicesNV: return "MeshViewIndicesNV"; + + // NV Capabilities + case CapabilityGeometryShaderPassthroughNV: return "GeometryShaderPassthroughNV"; + case CapabilityShaderViewportMaskNV: return "ShaderViewportMaskNV"; + case CapabilityShaderStereoViewNV: return "ShaderStereoViewNV"; + case CapabilityPerViewAttributesNV: return "PerViewAttributesNV"; + case CapabilityFragmentBarycentricNV: return "FragmentBarycentricNV"; + case CapabilityMeshShadingNV: return "MeshShadingNV"; + + // NV Decorations + case DecorationOverrideCoverageNV: return "OverrideCoverageNV"; + case DecorationPassthroughNV: return "PassthroughNV"; + case DecorationViewportRelativeNV: return "ViewportRelativeNV"; + case DecorationSecondaryViewportRelativeNV: return "SecondaryViewportRelativeNV"; + case DecorationPerVertexNV: return "PerVertexNV"; + case DecorationPerPrimitiveNV: return "PerPrimitiveNV"; + case DecorationPerViewNV: return "PerViewNV"; + case DecorationPerTaskNV: return "PerTaskNV"; + + default: return "Bad"; + } + } + return "Bad"; +} +#endif + +void Disassemble(std::ostream& out, const std::vector<unsigned int>& stream) +{ + SpirvStream SpirvStream(out, stream); + spv::Parameterize(); + GLSLstd450GetDebugNames(GlslStd450DebugNames); + SpirvStream.validate(); + SpirvStream.processInstructions(); +} + +}; // end namespace spv diff --git a/thirdparty/glslang/SPIRV/disassemble.h b/thirdparty/glslang/SPIRV/disassemble.h new file mode 100644 index 0000000000..b6a4635775 --- /dev/null +++ b/thirdparty/glslang/SPIRV/disassemble.h @@ -0,0 +1,53 @@ +// +// Copyright (C) 2014-2015 LunarG, Inc. +// +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions +// are met: +// +// Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// +// Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following +// disclaimer in the documentation and/or other materials provided +// with the distribution. +// +// Neither the name of 3Dlabs Inc. Ltd. nor the names of its +// contributors may be used to endorse or promote products derived +// from this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS +// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE +// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, +// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, +// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; +// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN +// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +// POSSIBILITY OF SUCH DAMAGE. + +// +// Disassembler for SPIR-V. +// + +#pragma once +#ifndef disassembler_H +#define disassembler_H + +#include <iostream> +#include <vector> + +namespace spv { + + // disassemble with glslang custom disassembler + void Disassemble(std::ostream& out, const std::vector<unsigned int>&); + +} // end namespace spv + +#endif // disassembler_H diff --git a/thirdparty/glslang/SPIRV/doc.cpp b/thirdparty/glslang/SPIRV/doc.cpp new file mode 100644 index 0000000000..3b85767216 --- /dev/null +++ b/thirdparty/glslang/SPIRV/doc.cpp @@ -0,0 +1,2767 @@ +// +// Copyright (C) 2014-2015 LunarG, Inc. +// +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions +// are met: +// +// Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// +// Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following +// disclaimer in the documentation and/or other materials provided +// with the distribution. +// +// Neither the name of 3Dlabs Inc. Ltd. nor the names of its +// contributors may be used to endorse or promote products derived +// from this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS +// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE +// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, +// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, +// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; +// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN +// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +// POSSIBILITY OF SUCH DAMAGE. + +// +// 1) Programmatically fill in instruction/operand information. +// This can be used for disassembly, printing documentation, etc. +// +// 2) Print documentation from this parameterization. +// + +#include "doc.h" + +#include <cstdio> +#include <cstring> +#include <algorithm> + +namespace spv { + extern "C" { + // Include C-based headers that don't have a namespace + #include "GLSL.ext.KHR.h" + #include "GLSL.ext.EXT.h" +#ifdef AMD_EXTENSIONS + #include "GLSL.ext.AMD.h" +#endif +#ifdef NV_EXTENSIONS + #include "GLSL.ext.NV.h" +#endif + } +} + +namespace spv { + +// +// Whole set of functions that translate enumerants to their text strings for +// the specification (or their sanitized versions for auto-generating the +// spirv headers. +// +// Also, for masks the ceilings are declared next to these, to help keep them in sync. +// Ceilings should be +// - one more than the maximum value an enumerant takes on, for non-mask enumerants +// (for non-sparse enums, this is the number of enumerants) +// - the number of bits consumed by the set of masks +// (for non-sparse mask enums, this is the number of enumerants) +// + +const char* SourceString(int source) +{ + switch (source) { + case 0: return "Unknown"; + case 1: return "ESSL"; + case 2: return "GLSL"; + case 3: return "OpenCL_C"; + case 4: return "OpenCL_CPP"; + case 5: return "HLSL"; + + default: return "Bad"; + } +} + +const char* ExecutionModelString(int model) +{ + switch (model) { + case 0: return "Vertex"; + case 1: return "TessellationControl"; + case 2: return "TessellationEvaluation"; + case 3: return "Geometry"; + case 4: return "Fragment"; + case 5: return "GLCompute"; + case 6: return "Kernel"; +#ifdef NV_EXTENSIONS + case ExecutionModelTaskNV: return "TaskNV"; + case ExecutionModelMeshNV: return "MeshNV"; +#endif + + default: return "Bad"; + +#ifdef NV_EXTENSIONS + case ExecutionModelRayGenerationNV: return "RayGenerationNV"; + case ExecutionModelIntersectionNV: return "IntersectionNV"; + case ExecutionModelAnyHitNV: return "AnyHitNV"; + case ExecutionModelClosestHitNV: return "ClosestHitNV"; + case ExecutionModelMissNV: return "MissNV"; + case ExecutionModelCallableNV: return "CallableNV"; +#endif + + } +} + +const char* AddressingString(int addr) +{ + switch (addr) { + case 0: return "Logical"; + case 1: return "Physical32"; + case 2: return "Physical64"; + + case AddressingModelPhysicalStorageBuffer64EXT: return "PhysicalStorageBuffer64EXT"; + + default: return "Bad"; + } +} + +const char* MemoryString(int mem) +{ + switch (mem) { + case MemoryModelSimple: return "Simple"; + case MemoryModelGLSL450: return "GLSL450"; + case MemoryModelOpenCL: return "OpenCL"; + case MemoryModelVulkanKHR: return "VulkanKHR"; + + default: return "Bad"; + } +} + +const int ExecutionModeCeiling = 33; + +const char* ExecutionModeString(int mode) +{ + switch (mode) { + case 0: return "Invocations"; + case 1: return "SpacingEqual"; + case 2: return "SpacingFractionalEven"; + case 3: return "SpacingFractionalOdd"; + case 4: return "VertexOrderCw"; + case 5: return "VertexOrderCcw"; + case 6: return "PixelCenterInteger"; + case 7: return "OriginUpperLeft"; + case 8: return "OriginLowerLeft"; + case 9: return "EarlyFragmentTests"; + case 10: return "PointMode"; + case 11: return "Xfb"; + case 12: return "DepthReplacing"; + case 13: return "Bad"; + case 14: return "DepthGreater"; + case 15: return "DepthLess"; + case 16: return "DepthUnchanged"; + case 17: return "LocalSize"; + case 18: return "LocalSizeHint"; + case 19: return "InputPoints"; + case 20: return "InputLines"; + case 21: return "InputLinesAdjacency"; + case 22: return "Triangles"; + case 23: return "InputTrianglesAdjacency"; + case 24: return "Quads"; + case 25: return "Isolines"; + case 26: return "OutputVertices"; + case 27: return "OutputPoints"; + case 28: return "OutputLineStrip"; + case 29: return "OutputTriangleStrip"; + case 30: return "VecTypeHint"; + case 31: return "ContractionOff"; + case 32: return "Bad"; + + case 4446: return "PostDepthCoverage"; + +#ifdef NV_EXTENSIONS + case ExecutionModeOutputLinesNV: return "OutputLinesNV"; + case ExecutionModeOutputPrimitivesNV: return "OutputPrimitivesNV"; + case ExecutionModeOutputTrianglesNV: return "OutputTrianglesNV"; + case ExecutionModeDerivativeGroupQuadsNV: return "DerivativeGroupQuadsNV"; + case ExecutionModeDerivativeGroupLinearNV: return "DerivativeGroupLinearNV"; +#endif + + case ExecutionModeCeiling: + default: return "Bad"; + } +} + +const char* StorageClassString(int StorageClass) +{ + switch (StorageClass) { + case 0: return "UniformConstant"; + case 1: return "Input"; + case 2: return "Uniform"; + case 3: return "Output"; + case 4: return "Workgroup"; + case 5: return "CrossWorkgroup"; + case 6: return "Private"; + case 7: return "Function"; + case 8: return "Generic"; + case 9: return "PushConstant"; + case 10: return "AtomicCounter"; + case 11: return "Image"; + case 12: return "StorageBuffer"; + +#ifdef NV_EXTENSIONS + case StorageClassRayPayloadNV: return "RayPayloadNV"; + case StorageClassHitAttributeNV: return "HitAttributeNV"; + case StorageClassIncomingRayPayloadNV: return "IncomingRayPayloadNV"; + case StorageClassShaderRecordBufferNV: return "ShaderRecordBufferNV"; + case StorageClassCallableDataNV: return "CallableDataNV"; + case StorageClassIncomingCallableDataNV: return "IncomingCallableDataNV"; +#endif + + case StorageClassPhysicalStorageBufferEXT: return "PhysicalStorageBufferEXT"; + + default: return "Bad"; + } +} + +const int DecorationCeiling = 45; + +const char* DecorationString(int decoration) +{ + switch (decoration) { + case 0: return "RelaxedPrecision"; + case 1: return "SpecId"; + case 2: return "Block"; + case 3: return "BufferBlock"; + case 4: return "RowMajor"; + case 5: return "ColMajor"; + case 6: return "ArrayStride"; + case 7: return "MatrixStride"; + case 8: return "GLSLShared"; + case 9: return "GLSLPacked"; + case 10: return "CPacked"; + case 11: return "BuiltIn"; + case 12: return "Bad"; + case 13: return "NoPerspective"; + case 14: return "Flat"; + case 15: return "Patch"; + case 16: return "Centroid"; + case 17: return "Sample"; + case 18: return "Invariant"; + case 19: return "Restrict"; + case 20: return "Aliased"; + case 21: return "Volatile"; + case 22: return "Constant"; + case 23: return "Coherent"; + case 24: return "NonWritable"; + case 25: return "NonReadable"; + case 26: return "Uniform"; + case 27: return "Bad"; + case 28: return "SaturatedConversion"; + case 29: return "Stream"; + case 30: return "Location"; + case 31: return "Component"; + case 32: return "Index"; + case 33: return "Binding"; + case 34: return "DescriptorSet"; + case 35: return "Offset"; + case 36: return "XfbBuffer"; + case 37: return "XfbStride"; + case 38: return "FuncParamAttr"; + case 39: return "FP Rounding Mode"; + case 40: return "FP Fast Math Mode"; + case 41: return "Linkage Attributes"; + case 42: return "NoContraction"; + case 43: return "InputAttachmentIndex"; + case 44: return "Alignment"; + + case DecorationCeiling: + default: return "Bad"; + +#ifdef AMD_EXTENSIONS + case DecorationExplicitInterpAMD: return "ExplicitInterpAMD"; +#endif +#ifdef NV_EXTENSIONS + case DecorationOverrideCoverageNV: return "OverrideCoverageNV"; + case DecorationPassthroughNV: return "PassthroughNV"; + case DecorationViewportRelativeNV: return "ViewportRelativeNV"; + case DecorationSecondaryViewportRelativeNV: return "SecondaryViewportRelativeNV"; + case DecorationPerPrimitiveNV: return "PerPrimitiveNV"; + case DecorationPerViewNV: return "PerViewNV"; + case DecorationPerTaskNV: return "PerTaskNV"; + case DecorationPerVertexNV: return "PerVertexNV"; +#endif + + case DecorationNonUniformEXT: return "DecorationNonUniformEXT"; + case DecorationHlslCounterBufferGOOGLE: return "DecorationHlslCounterBufferGOOGLE"; + case DecorationHlslSemanticGOOGLE: return "DecorationHlslSemanticGOOGLE"; + case DecorationRestrictPointerEXT: return "DecorationRestrictPointerEXT"; + case DecorationAliasedPointerEXT: return "DecorationAliasedPointerEXT"; + } +} + +const char* BuiltInString(int builtIn) +{ + switch (builtIn) { + case 0: return "Position"; + case 1: return "PointSize"; + case 2: return "Bad"; + case 3: return "ClipDistance"; + case 4: return "CullDistance"; + case 5: return "VertexId"; + case 6: return "InstanceId"; + case 7: return "PrimitiveId"; + case 8: return "InvocationId"; + case 9: return "Layer"; + case 10: return "ViewportIndex"; + case 11: return "TessLevelOuter"; + case 12: return "TessLevelInner"; + case 13: return "TessCoord"; + case 14: return "PatchVertices"; + case 15: return "FragCoord"; + case 16: return "PointCoord"; + case 17: return "FrontFacing"; + case 18: return "SampleId"; + case 19: return "SamplePosition"; + case 20: return "SampleMask"; + case 21: return "Bad"; + case 22: return "FragDepth"; + case 23: return "HelperInvocation"; + case 24: return "NumWorkgroups"; + case 25: return "WorkgroupSize"; + case 26: return "WorkgroupId"; + case 27: return "LocalInvocationId"; + case 28: return "GlobalInvocationId"; + case 29: return "LocalInvocationIndex"; + case 30: return "WorkDim"; + case 31: return "GlobalSize"; + case 32: return "EnqueuedWorkgroupSize"; + case 33: return "GlobalOffset"; + case 34: return "GlobalLinearId"; + case 35: return "Bad"; + case 36: return "SubgroupSize"; + case 37: return "SubgroupMaxSize"; + case 38: return "NumSubgroups"; + case 39: return "NumEnqueuedSubgroups"; + case 40: return "SubgroupId"; + case 41: return "SubgroupLocalInvocationId"; + case 42: return "VertexIndex"; // TBD: put next to VertexId? + case 43: return "InstanceIndex"; // TBD: put next to InstanceId? + + case 4416: return "SubgroupEqMaskKHR"; + case 4417: return "SubgroupGeMaskKHR"; + case 4418: return "SubgroupGtMaskKHR"; + case 4419: return "SubgroupLeMaskKHR"; + case 4420: return "SubgroupLtMaskKHR"; + case 4438: return "DeviceIndex"; + case 4440: return "ViewIndex"; + case 4424: return "BaseVertex"; + case 4425: return "BaseInstance"; + case 4426: return "DrawIndex"; + case 5014: return "FragStencilRefEXT"; + +#ifdef AMD_EXTENSIONS + case 4992: return "BaryCoordNoPerspAMD"; + case 4993: return "BaryCoordNoPerspCentroidAMD"; + case 4994: return "BaryCoordNoPerspSampleAMD"; + case 4995: return "BaryCoordSmoothAMD"; + case 4996: return "BaryCoordSmoothCentroidAMD"; + case 4997: return "BaryCoordSmoothSampleAMD"; + case 4998: return "BaryCoordPullModelAMD"; +#endif + +#ifdef NV_EXTENSIONS + case BuiltInLaunchIdNV: return "LaunchIdNV"; + case BuiltInLaunchSizeNV: return "LaunchSizeNV"; + case BuiltInWorldRayOriginNV: return "WorldRayOriginNV"; + case BuiltInWorldRayDirectionNV: return "WorldRayDirectionNV"; + case BuiltInObjectRayOriginNV: return "ObjectRayOriginNV"; + case BuiltInObjectRayDirectionNV: return "ObjectRayDirectionNV"; + case BuiltInRayTminNV: return "RayTminNV"; + case BuiltInRayTmaxNV: return "RayTmaxNV"; + case BuiltInInstanceCustomIndexNV: return "InstanceCustomIndexNV"; + case BuiltInObjectToWorldNV: return "ObjectToWorldNV"; + case BuiltInWorldToObjectNV: return "WorldToObjectNV"; + case BuiltInHitTNV: return "HitTNV"; + case BuiltInHitKindNV: return "HitKindNV"; + case BuiltInIncomingRayFlagsNV: return "IncomingRayFlagsNV"; + case BuiltInViewportMaskNV: return "ViewportMaskNV"; + case BuiltInSecondaryPositionNV: return "SecondaryPositionNV"; + case BuiltInSecondaryViewportMaskNV: return "SecondaryViewportMaskNV"; + case BuiltInPositionPerViewNV: return "PositionPerViewNV"; + case BuiltInViewportMaskPerViewNV: return "ViewportMaskPerViewNV"; +// case BuiltInFragmentSizeNV: return "FragmentSizeNV"; // superseded by BuiltInFragSizeEXT +// case BuiltInInvocationsPerPixelNV: return "InvocationsPerPixelNV"; // superseded by BuiltInFragInvocationCountEXT + case BuiltInBaryCoordNV: return "BaryCoordNV"; + case BuiltInBaryCoordNoPerspNV: return "BaryCoordNoPerspNV"; +#endif + + case BuiltInFragSizeEXT: return "FragSizeEXT"; + case BuiltInFragInvocationCountEXT: return "FragInvocationCountEXT"; + + case 5264: return "FullyCoveredEXT"; + + +#ifdef NV_EXTENSIONS + case BuiltInTaskCountNV: return "TaskCountNV"; + case BuiltInPrimitiveCountNV: return "PrimitiveCountNV"; + case BuiltInPrimitiveIndicesNV: return "PrimitiveIndicesNV"; + case BuiltInClipDistancePerViewNV: return "ClipDistancePerViewNV"; + case BuiltInCullDistancePerViewNV: return "CullDistancePerViewNV"; + case BuiltInLayerPerViewNV: return "LayerPerViewNV"; + case BuiltInMeshViewCountNV: return "MeshViewCountNV"; + case BuiltInMeshViewIndicesNV: return "MeshViewIndicesNV"; +#endif + + default: return "Bad"; + } +} + +const char* DimensionString(int dim) +{ + switch (dim) { + case 0: return "1D"; + case 1: return "2D"; + case 2: return "3D"; + case 3: return "Cube"; + case 4: return "Rect"; + case 5: return "Buffer"; + case 6: return "SubpassData"; + + default: return "Bad"; + } +} + +const char* SamplerAddressingModeString(int mode) +{ + switch (mode) { + case 0: return "None"; + case 1: return "ClampToEdge"; + case 2: return "Clamp"; + case 3: return "Repeat"; + case 4: return "RepeatMirrored"; + + default: return "Bad"; + } +} + +const char* SamplerFilterModeString(int mode) +{ + switch (mode) { + case 0: return "Nearest"; + case 1: return "Linear"; + + default: return "Bad"; + } +} + +const char* ImageFormatString(int format) +{ + switch (format) { + case 0: return "Unknown"; + + // ES/Desktop float + case 1: return "Rgba32f"; + case 2: return "Rgba16f"; + case 3: return "R32f"; + case 4: return "Rgba8"; + case 5: return "Rgba8Snorm"; + + // Desktop float + case 6: return "Rg32f"; + case 7: return "Rg16f"; + case 8: return "R11fG11fB10f"; + case 9: return "R16f"; + case 10: return "Rgba16"; + case 11: return "Rgb10A2"; + case 12: return "Rg16"; + case 13: return "Rg8"; + case 14: return "R16"; + case 15: return "R8"; + case 16: return "Rgba16Snorm"; + case 17: return "Rg16Snorm"; + case 18: return "Rg8Snorm"; + case 19: return "R16Snorm"; + case 20: return "R8Snorm"; + + // ES/Desktop int + case 21: return "Rgba32i"; + case 22: return "Rgba16i"; + case 23: return "Rgba8i"; + case 24: return "R32i"; + + // Desktop int + case 25: return "Rg32i"; + case 26: return "Rg16i"; + case 27: return "Rg8i"; + case 28: return "R16i"; + case 29: return "R8i"; + + // ES/Desktop uint + case 30: return "Rgba32ui"; + case 31: return "Rgba16ui"; + case 32: return "Rgba8ui"; + case 33: return "R32ui"; + + // Desktop uint + case 34: return "Rgb10a2ui"; + case 35: return "Rg32ui"; + case 36: return "Rg16ui"; + case 37: return "Rg8ui"; + case 38: return "R16ui"; + case 39: return "R8ui"; + + default: + return "Bad"; + } +} + +const char* ImageChannelOrderString(int format) +{ + switch (format) { + case 0: return "R"; + case 1: return "A"; + case 2: return "RG"; + case 3: return "RA"; + case 4: return "RGB"; + case 5: return "RGBA"; + case 6: return "BGRA"; + case 7: return "ARGB"; + case 8: return "Intensity"; + case 9: return "Luminance"; + case 10: return "Rx"; + case 11: return "RGx"; + case 12: return "RGBx"; + case 13: return "Depth"; + case 14: return "DepthStencil"; + case 15: return "sRGB"; + case 16: return "sRGBx"; + case 17: return "sRGBA"; + case 18: return "sBGRA"; + + default: + return "Bad"; + } +} + +const char* ImageChannelDataTypeString(int type) +{ + switch (type) + { + case 0: return "SnormInt8"; + case 1: return "SnormInt16"; + case 2: return "UnormInt8"; + case 3: return "UnormInt16"; + case 4: return "UnormShort565"; + case 5: return "UnormShort555"; + case 6: return "UnormInt101010"; + case 7: return "SignedInt8"; + case 8: return "SignedInt16"; + case 9: return "SignedInt32"; + case 10: return "UnsignedInt8"; + case 11: return "UnsignedInt16"; + case 12: return "UnsignedInt32"; + case 13: return "HalfFloat"; + case 14: return "Float"; + case 15: return "UnormInt24"; + case 16: return "UnormInt101010_2"; + + default: + return "Bad"; + } +} + +const int ImageOperandsCeiling = 14; + +const char* ImageOperandsString(int format) +{ + switch (format) { + case ImageOperandsBiasShift: return "Bias"; + case ImageOperandsLodShift: return "Lod"; + case ImageOperandsGradShift: return "Grad"; + case ImageOperandsConstOffsetShift: return "ConstOffset"; + case ImageOperandsOffsetShift: return "Offset"; + case ImageOperandsConstOffsetsShift: return "ConstOffsets"; + case ImageOperandsSampleShift: return "Sample"; + case ImageOperandsMinLodShift: return "MinLod"; + case ImageOperandsMakeTexelAvailableKHRShift: return "MakeTexelAvailableKHR"; + case ImageOperandsMakeTexelVisibleKHRShift: return "MakeTexelVisibleKHR"; + case ImageOperandsNonPrivateTexelKHRShift: return "NonPrivateTexelKHR"; + case ImageOperandsVolatileTexelKHRShift: return "VolatileTexelKHR"; + case ImageOperandsSignExtendShift: return "SignExtend"; + case ImageOperandsZeroExtendShift: return "ZeroExtend"; + + case ImageOperandsCeiling: + default: + return "Bad"; + } +} + +const char* FPFastMathString(int mode) +{ + switch (mode) { + case 0: return "NotNaN"; + case 1: return "NotInf"; + case 2: return "NSZ"; + case 3: return "AllowRecip"; + case 4: return "Fast"; + + default: return "Bad"; + } +} + +const char* FPRoundingModeString(int mode) +{ + switch (mode) { + case 0: return "RTE"; + case 1: return "RTZ"; + case 2: return "RTP"; + case 3: return "RTN"; + + default: return "Bad"; + } +} + +const char* LinkageTypeString(int type) +{ + switch (type) { + case 0: return "Export"; + case 1: return "Import"; + + default: return "Bad"; + } +} + +const char* FuncParamAttrString(int attr) +{ + switch (attr) { + case 0: return "Zext"; + case 1: return "Sext"; + case 2: return "ByVal"; + case 3: return "Sret"; + case 4: return "NoAlias"; + case 5: return "NoCapture"; + case 6: return "NoWrite"; + case 7: return "NoReadWrite"; + + default: return "Bad"; + } +} + +const char* AccessQualifierString(int attr) +{ + switch (attr) { + case 0: return "ReadOnly"; + case 1: return "WriteOnly"; + case 2: return "ReadWrite"; + + default: return "Bad"; + } +} + +const int SelectControlCeiling = 2; + +const char* SelectControlString(int cont) +{ + switch (cont) { + case 0: return "Flatten"; + case 1: return "DontFlatten"; + + case SelectControlCeiling: + default: return "Bad"; + } +} + +const int LoopControlCeiling = LoopControlPartialCountShift + 1; + +const char* LoopControlString(int cont) +{ + switch (cont) { + case LoopControlUnrollShift: return "Unroll"; + case LoopControlDontUnrollShift: return "DontUnroll"; + case LoopControlDependencyInfiniteShift: return "DependencyInfinite"; + case LoopControlDependencyLengthShift: return "DependencyLength"; + case LoopControlMinIterationsShift: return "MinIterations"; + case LoopControlMaxIterationsShift: return "MaxIterations"; + case LoopControlIterationMultipleShift: return "IterationMultiple"; + case LoopControlPeelCountShift: return "PeelCount"; + case LoopControlPartialCountShift: return "PartialCount"; + + case LoopControlCeiling: + default: return "Bad"; + } +} + +const int FunctionControlCeiling = 4; + +const char* FunctionControlString(int cont) +{ + switch (cont) { + case 0: return "Inline"; + case 1: return "DontInline"; + case 2: return "Pure"; + case 3: return "Const"; + + case FunctionControlCeiling: + default: return "Bad"; + } +} + +const char* MemorySemanticsString(int mem) +{ + // Note: No bits set (None) means "Relaxed" + switch (mem) { + case 0: return "Bad"; // Note: this is a placeholder for 'Consume' + case 1: return "Acquire"; + case 2: return "Release"; + case 3: return "AcquireRelease"; + case 4: return "SequentiallyConsistent"; + case 5: return "Bad"; // Note: reserved for future expansion + case 6: return "UniformMemory"; + case 7: return "SubgroupMemory"; + case 8: return "WorkgroupMemory"; + case 9: return "CrossWorkgroupMemory"; + case 10: return "AtomicCounterMemory"; + case 11: return "ImageMemory"; + + default: return "Bad"; + } +} + +const int MemoryAccessCeiling = 6; + +const char* MemoryAccessString(int mem) +{ + switch (mem) { + case MemoryAccessVolatileShift: return "Volatile"; + case MemoryAccessAlignedShift: return "Aligned"; + case MemoryAccessNontemporalShift: return "Nontemporal"; + case MemoryAccessMakePointerAvailableKHRShift: return "MakePointerAvailableKHR"; + case MemoryAccessMakePointerVisibleKHRShift: return "MakePointerVisibleKHR"; + case MemoryAccessNonPrivatePointerKHRShift: return "NonPrivatePointerKHR"; + + default: return "Bad"; + } +} + +const char* ScopeString(int mem) +{ + switch (mem) { + case 0: return "CrossDevice"; + case 1: return "Device"; + case 2: return "Workgroup"; + case 3: return "Subgroup"; + case 4: return "Invocation"; + + default: return "Bad"; + } +} + +const char* GroupOperationString(int gop) +{ + + switch (gop) + { + case GroupOperationReduce: return "Reduce"; + case GroupOperationInclusiveScan: return "InclusiveScan"; + case GroupOperationExclusiveScan: return "ExclusiveScan"; + case GroupOperationClusteredReduce: return "ClusteredReduce"; +#ifdef NV_EXTENSIONS + case GroupOperationPartitionedReduceNV: return "PartitionedReduceNV"; + case GroupOperationPartitionedInclusiveScanNV: return "PartitionedInclusiveScanNV"; + case GroupOperationPartitionedExclusiveScanNV: return "PartitionedExclusiveScanNV"; +#endif + + default: return "Bad"; + } +} + +const char* KernelEnqueueFlagsString(int flag) +{ + switch (flag) + { + case 0: return "NoWait"; + case 1: return "WaitKernel"; + case 2: return "WaitWorkGroup"; + + default: return "Bad"; + } +} + +const char* KernelProfilingInfoString(int info) +{ + switch (info) + { + case 0: return "CmdExecTime"; + + default: return "Bad"; + } +} + +const char* CapabilityString(int info) +{ + switch (info) + { + case 0: return "Matrix"; + case 1: return "Shader"; + case 2: return "Geometry"; + case 3: return "Tessellation"; + case 4: return "Addresses"; + case 5: return "Linkage"; + case 6: return "Kernel"; + case 7: return "Vector16"; + case 8: return "Float16Buffer"; + case 9: return "Float16"; + case 10: return "Float64"; + case 11: return "Int64"; + case 12: return "Int64Atomics"; + case 13: return "ImageBasic"; + case 14: return "ImageReadWrite"; + case 15: return "ImageMipmap"; + case 16: return "Bad"; + case 17: return "Pipes"; + case 18: return "Groups"; + case 19: return "DeviceEnqueue"; + case 20: return "LiteralSampler"; + case 21: return "AtomicStorage"; + case 22: return "Int16"; + case 23: return "TessellationPointSize"; + case 24: return "GeometryPointSize"; + case 25: return "ImageGatherExtended"; + case 26: return "Bad"; + case 27: return "StorageImageMultisample"; + case 28: return "UniformBufferArrayDynamicIndexing"; + case 29: return "SampledImageArrayDynamicIndexing"; + case 30: return "StorageBufferArrayDynamicIndexing"; + case 31: return "StorageImageArrayDynamicIndexing"; + case 32: return "ClipDistance"; + case 33: return "CullDistance"; + case 34: return "ImageCubeArray"; + case 35: return "SampleRateShading"; + case 36: return "ImageRect"; + case 37: return "SampledRect"; + case 38: return "GenericPointer"; + case 39: return "Int8"; + case 40: return "InputAttachment"; + case 41: return "SparseResidency"; + case 42: return "MinLod"; + case 43: return "Sampled1D"; + case 44: return "Image1D"; + case 45: return "SampledCubeArray"; + case 46: return "SampledBuffer"; + case 47: return "ImageBuffer"; + case 48: return "ImageMSArray"; + case 49: return "StorageImageExtendedFormats"; + case 50: return "ImageQuery"; + case 51: return "DerivativeControl"; + case 52: return "InterpolationFunction"; + case 53: return "TransformFeedback"; + case 54: return "GeometryStreams"; + case 55: return "StorageImageReadWithoutFormat"; + case 56: return "StorageImageWriteWithoutFormat"; + case 57: return "MultiViewport"; + case 61: return "GroupNonUniform"; + case 62: return "GroupNonUniformVote"; + case 63: return "GroupNonUniformArithmetic"; + case 64: return "GroupNonUniformBallot"; + case 65: return "GroupNonUniformShuffle"; + case 66: return "GroupNonUniformShuffleRelative"; + case 67: return "GroupNonUniformClustered"; + case 68: return "GroupNonUniformQuad"; + + case CapabilitySubgroupBallotKHR: return "SubgroupBallotKHR"; + case CapabilityDrawParameters: return "DrawParameters"; + case CapabilitySubgroupVoteKHR: return "SubgroupVoteKHR"; + + case CapabilityStorageUniformBufferBlock16: return "StorageUniformBufferBlock16"; + case CapabilityStorageUniform16: return "StorageUniform16"; + case CapabilityStoragePushConstant16: return "StoragePushConstant16"; + case CapabilityStorageInputOutput16: return "StorageInputOutput16"; + + case CapabilityStorageBuffer8BitAccess: return "CapabilityStorageBuffer8BitAccess"; + case CapabilityUniformAndStorageBuffer8BitAccess: return "CapabilityUniformAndStorageBuffer8BitAccess"; + case CapabilityStoragePushConstant8: return "CapabilityStoragePushConstant8"; + + case CapabilityDeviceGroup: return "DeviceGroup"; + case CapabilityMultiView: return "MultiView"; + + case CapabilityStencilExportEXT: return "StencilExportEXT"; + +#ifdef AMD_EXTENSIONS + case CapabilityFloat16ImageAMD: return "Float16ImageAMD"; + case CapabilityImageGatherBiasLodAMD: return "ImageGatherBiasLodAMD"; + case CapabilityFragmentMaskAMD: return "FragmentMaskAMD"; + case CapabilityImageReadWriteLodAMD: return "ImageReadWriteLodAMD"; +#endif + + case CapabilityAtomicStorageOps: return "AtomicStorageOps"; + + case CapabilitySampleMaskPostDepthCoverage: return "SampleMaskPostDepthCoverage"; +#ifdef NV_EXTENSIONS + case CapabilityGeometryShaderPassthroughNV: return "GeometryShaderPassthroughNV"; + case CapabilityShaderViewportIndexLayerNV: return "ShaderViewportIndexLayerNV"; + case CapabilityShaderViewportMaskNV: return "ShaderViewportMaskNV"; + case CapabilityShaderStereoViewNV: return "ShaderStereoViewNV"; + case CapabilityPerViewAttributesNV: return "PerViewAttributesNV"; + case CapabilityGroupNonUniformPartitionedNV: return "GroupNonUniformPartitionedNV"; + case CapabilityRayTracingNV: return "RayTracingNV"; + case CapabilityComputeDerivativeGroupQuadsNV: return "ComputeDerivativeGroupQuadsNV"; + case CapabilityComputeDerivativeGroupLinearNV: return "ComputeDerivativeGroupLinearNV"; + case CapabilityFragmentBarycentricNV: return "FragmentBarycentricNV"; + case CapabilityMeshShadingNV: return "MeshShadingNV"; +// case CapabilityShadingRateNV: return "ShadingRateNV"; // superseded by CapabilityFragmentDensityEXT +#endif + case CapabilityFragmentDensityEXT: return "FragmentDensityEXT"; + + case CapabilityFragmentFullyCoveredEXT: return "FragmentFullyCoveredEXT"; + + case CapabilityShaderNonUniformEXT: return "CapabilityShaderNonUniformEXT"; + case CapabilityRuntimeDescriptorArrayEXT: return "CapabilityRuntimeDescriptorArrayEXT"; + case CapabilityInputAttachmentArrayDynamicIndexingEXT: return "CapabilityInputAttachmentArrayDynamicIndexingEXT"; + case CapabilityUniformTexelBufferArrayDynamicIndexingEXT: return "CapabilityUniformTexelBufferArrayDynamicIndexingEXT"; + case CapabilityStorageTexelBufferArrayDynamicIndexingEXT: return "CapabilityStorageTexelBufferArrayDynamicIndexingEXT"; + case CapabilityUniformBufferArrayNonUniformIndexingEXT: return "CapabilityUniformBufferArrayNonUniformIndexingEXT"; + case CapabilitySampledImageArrayNonUniformIndexingEXT: return "CapabilitySampledImageArrayNonUniformIndexingEXT"; + case CapabilityStorageBufferArrayNonUniformIndexingEXT: return "CapabilityStorageBufferArrayNonUniformIndexingEXT"; + case CapabilityStorageImageArrayNonUniformIndexingEXT: return "CapabilityStorageImageArrayNonUniformIndexingEXT"; + case CapabilityInputAttachmentArrayNonUniformIndexingEXT: return "CapabilityInputAttachmentArrayNonUniformIndexingEXT"; + case CapabilityUniformTexelBufferArrayNonUniformIndexingEXT: return "CapabilityUniformTexelBufferArrayNonUniformIndexingEXT"; + case CapabilityStorageTexelBufferArrayNonUniformIndexingEXT: return "CapabilityStorageTexelBufferArrayNonUniformIndexingEXT"; + + case CapabilityVulkanMemoryModelKHR: return "CapabilityVulkanMemoryModelKHR"; + case CapabilityVulkanMemoryModelDeviceScopeKHR: return "CapabilityVulkanMemoryModelDeviceScopeKHR"; + + case CapabilityPhysicalStorageBufferAddressesEXT: return "CapabilityPhysicalStorageBufferAddressesEXT"; + + case CapabilityVariablePointers: return "CapabilityVariablePointers"; + + case CapabilityCooperativeMatrixNV: return "CapabilityCooperativeMatrixNV"; + + default: return "Bad"; + } +} + +const char* OpcodeString(int op) +{ + switch (op) { + case 0: return "OpNop"; + case 1: return "OpUndef"; + case 2: return "OpSourceContinued"; + case 3: return "OpSource"; + case 4: return "OpSourceExtension"; + case 5: return "OpName"; + case 6: return "OpMemberName"; + case 7: return "OpString"; + case 8: return "OpLine"; + case 9: return "Bad"; + case 10: return "OpExtension"; + case 11: return "OpExtInstImport"; + case 12: return "OpExtInst"; + case 13: return "Bad"; + case 14: return "OpMemoryModel"; + case 15: return "OpEntryPoint"; + case 16: return "OpExecutionMode"; + case 17: return "OpCapability"; + case 18: return "Bad"; + case 19: return "OpTypeVoid"; + case 20: return "OpTypeBool"; + case 21: return "OpTypeInt"; + case 22: return "OpTypeFloat"; + case 23: return "OpTypeVector"; + case 24: return "OpTypeMatrix"; + case 25: return "OpTypeImage"; + case 26: return "OpTypeSampler"; + case 27: return "OpTypeSampledImage"; + case 28: return "OpTypeArray"; + case 29: return "OpTypeRuntimeArray"; + case 30: return "OpTypeStruct"; + case 31: return "OpTypeOpaque"; + case 32: return "OpTypePointer"; + case 33: return "OpTypeFunction"; + case 34: return "OpTypeEvent"; + case 35: return "OpTypeDeviceEvent"; + case 36: return "OpTypeReserveId"; + case 37: return "OpTypeQueue"; + case 38: return "OpTypePipe"; + case 39: return "OpTypeForwardPointer"; + case 40: return "Bad"; + case 41: return "OpConstantTrue"; + case 42: return "OpConstantFalse"; + case 43: return "OpConstant"; + case 44: return "OpConstantComposite"; + case 45: return "OpConstantSampler"; + case 46: return "OpConstantNull"; + case 47: return "Bad"; + case 48: return "OpSpecConstantTrue"; + case 49: return "OpSpecConstantFalse"; + case 50: return "OpSpecConstant"; + case 51: return "OpSpecConstantComposite"; + case 52: return "OpSpecConstantOp"; + case 53: return "Bad"; + case 54: return "OpFunction"; + case 55: return "OpFunctionParameter"; + case 56: return "OpFunctionEnd"; + case 57: return "OpFunctionCall"; + case 58: return "Bad"; + case 59: return "OpVariable"; + case 60: return "OpImageTexelPointer"; + case 61: return "OpLoad"; + case 62: return "OpStore"; + case 63: return "OpCopyMemory"; + case 64: return "OpCopyMemorySized"; + case 65: return "OpAccessChain"; + case 66: return "OpInBoundsAccessChain"; + case 67: return "OpPtrAccessChain"; + case 68: return "OpArrayLength"; + case 69: return "OpGenericPtrMemSemantics"; + case 70: return "OpInBoundsPtrAccessChain"; + case 71: return "OpDecorate"; + case 72: return "OpMemberDecorate"; + case 73: return "OpDecorationGroup"; + case 74: return "OpGroupDecorate"; + case 75: return "OpGroupMemberDecorate"; + case 76: return "Bad"; + case 77: return "OpVectorExtractDynamic"; + case 78: return "OpVectorInsertDynamic"; + case 79: return "OpVectorShuffle"; + case 80: return "OpCompositeConstruct"; + case 81: return "OpCompositeExtract"; + case 82: return "OpCompositeInsert"; + case 83: return "OpCopyObject"; + case 84: return "OpTranspose"; + case OpCopyLogical: return "OpCopyLogical"; + case 85: return "Bad"; + case 86: return "OpSampledImage"; + case 87: return "OpImageSampleImplicitLod"; + case 88: return "OpImageSampleExplicitLod"; + case 89: return "OpImageSampleDrefImplicitLod"; + case 90: return "OpImageSampleDrefExplicitLod"; + case 91: return "OpImageSampleProjImplicitLod"; + case 92: return "OpImageSampleProjExplicitLod"; + case 93: return "OpImageSampleProjDrefImplicitLod"; + case 94: return "OpImageSampleProjDrefExplicitLod"; + case 95: return "OpImageFetch"; + case 96: return "OpImageGather"; + case 97: return "OpImageDrefGather"; + case 98: return "OpImageRead"; + case 99: return "OpImageWrite"; + case 100: return "OpImage"; + case 101: return "OpImageQueryFormat"; + case 102: return "OpImageQueryOrder"; + case 103: return "OpImageQuerySizeLod"; + case 104: return "OpImageQuerySize"; + case 105: return "OpImageQueryLod"; + case 106: return "OpImageQueryLevels"; + case 107: return "OpImageQuerySamples"; + case 108: return "Bad"; + case 109: return "OpConvertFToU"; + case 110: return "OpConvertFToS"; + case 111: return "OpConvertSToF"; + case 112: return "OpConvertUToF"; + case 113: return "OpUConvert"; + case 114: return "OpSConvert"; + case 115: return "OpFConvert"; + case 116: return "OpQuantizeToF16"; + case 117: return "OpConvertPtrToU"; + case 118: return "OpSatConvertSToU"; + case 119: return "OpSatConvertUToS"; + case 120: return "OpConvertUToPtr"; + case 121: return "OpPtrCastToGeneric"; + case 122: return "OpGenericCastToPtr"; + case 123: return "OpGenericCastToPtrExplicit"; + case 124: return "OpBitcast"; + case 125: return "Bad"; + case 126: return "OpSNegate"; + case 127: return "OpFNegate"; + case 128: return "OpIAdd"; + case 129: return "OpFAdd"; + case 130: return "OpISub"; + case 131: return "OpFSub"; + case 132: return "OpIMul"; + case 133: return "OpFMul"; + case 134: return "OpUDiv"; + case 135: return "OpSDiv"; + case 136: return "OpFDiv"; + case 137: return "OpUMod"; + case 138: return "OpSRem"; + case 139: return "OpSMod"; + case 140: return "OpFRem"; + case 141: return "OpFMod"; + case 142: return "OpVectorTimesScalar"; + case 143: return "OpMatrixTimesScalar"; + case 144: return "OpVectorTimesMatrix"; + case 145: return "OpMatrixTimesVector"; + case 146: return "OpMatrixTimesMatrix"; + case 147: return "OpOuterProduct"; + case 148: return "OpDot"; + case 149: return "OpIAddCarry"; + case 150: return "OpISubBorrow"; + case 151: return "OpUMulExtended"; + case 152: return "OpSMulExtended"; + case 153: return "Bad"; + case 154: return "OpAny"; + case 155: return "OpAll"; + case 156: return "OpIsNan"; + case 157: return "OpIsInf"; + case 158: return "OpIsFinite"; + case 159: return "OpIsNormal"; + case 160: return "OpSignBitSet"; + case 161: return "OpLessOrGreater"; + case 162: return "OpOrdered"; + case 163: return "OpUnordered"; + case 164: return "OpLogicalEqual"; + case 165: return "OpLogicalNotEqual"; + case 166: return "OpLogicalOr"; + case 167: return "OpLogicalAnd"; + case 168: return "OpLogicalNot"; + case 169: return "OpSelect"; + case 170: return "OpIEqual"; + case 171: return "OpINotEqual"; + case 172: return "OpUGreaterThan"; + case 173: return "OpSGreaterThan"; + case 174: return "OpUGreaterThanEqual"; + case 175: return "OpSGreaterThanEqual"; + case 176: return "OpULessThan"; + case 177: return "OpSLessThan"; + case 178: return "OpULessThanEqual"; + case 179: return "OpSLessThanEqual"; + case 180: return "OpFOrdEqual"; + case 181: return "OpFUnordEqual"; + case 182: return "OpFOrdNotEqual"; + case 183: return "OpFUnordNotEqual"; + case 184: return "OpFOrdLessThan"; + case 185: return "OpFUnordLessThan"; + case 186: return "OpFOrdGreaterThan"; + case 187: return "OpFUnordGreaterThan"; + case 188: return "OpFOrdLessThanEqual"; + case 189: return "OpFUnordLessThanEqual"; + case 190: return "OpFOrdGreaterThanEqual"; + case 191: return "OpFUnordGreaterThanEqual"; + case 192: return "Bad"; + case 193: return "Bad"; + case 194: return "OpShiftRightLogical"; + case 195: return "OpShiftRightArithmetic"; + case 196: return "OpShiftLeftLogical"; + case 197: return "OpBitwiseOr"; + case 198: return "OpBitwiseXor"; + case 199: return "OpBitwiseAnd"; + case 200: return "OpNot"; + case 201: return "OpBitFieldInsert"; + case 202: return "OpBitFieldSExtract"; + case 203: return "OpBitFieldUExtract"; + case 204: return "OpBitReverse"; + case 205: return "OpBitCount"; + case 206: return "Bad"; + case 207: return "OpDPdx"; + case 208: return "OpDPdy"; + case 209: return "OpFwidth"; + case 210: return "OpDPdxFine"; + case 211: return "OpDPdyFine"; + case 212: return "OpFwidthFine"; + case 213: return "OpDPdxCoarse"; + case 214: return "OpDPdyCoarse"; + case 215: return "OpFwidthCoarse"; + case 216: return "Bad"; + case 217: return "Bad"; + case 218: return "OpEmitVertex"; + case 219: return "OpEndPrimitive"; + case 220: return "OpEmitStreamVertex"; + case 221: return "OpEndStreamPrimitive"; + case 222: return "Bad"; + case 223: return "Bad"; + case 224: return "OpControlBarrier"; + case 225: return "OpMemoryBarrier"; + case 226: return "Bad"; + case 227: return "OpAtomicLoad"; + case 228: return "OpAtomicStore"; + case 229: return "OpAtomicExchange"; + case 230: return "OpAtomicCompareExchange"; + case 231: return "OpAtomicCompareExchangeWeak"; + case 232: return "OpAtomicIIncrement"; + case 233: return "OpAtomicIDecrement"; + case 234: return "OpAtomicIAdd"; + case 235: return "OpAtomicISub"; + case 236: return "OpAtomicSMin"; + case 237: return "OpAtomicUMin"; + case 238: return "OpAtomicSMax"; + case 239: return "OpAtomicUMax"; + case 240: return "OpAtomicAnd"; + case 241: return "OpAtomicOr"; + case 242: return "OpAtomicXor"; + case 243: return "Bad"; + case 244: return "Bad"; + case 245: return "OpPhi"; + case 246: return "OpLoopMerge"; + case 247: return "OpSelectionMerge"; + case 248: return "OpLabel"; + case 249: return "OpBranch"; + case 250: return "OpBranchConditional"; + case 251: return "OpSwitch"; + case 252: return "OpKill"; + case 253: return "OpReturn"; + case 254: return "OpReturnValue"; + case 255: return "OpUnreachable"; + case 256: return "OpLifetimeStart"; + case 257: return "OpLifetimeStop"; + case 258: return "Bad"; + case 259: return "OpGroupAsyncCopy"; + case 260: return "OpGroupWaitEvents"; + case 261: return "OpGroupAll"; + case 262: return "OpGroupAny"; + case 263: return "OpGroupBroadcast"; + case 264: return "OpGroupIAdd"; + case 265: return "OpGroupFAdd"; + case 266: return "OpGroupFMin"; + case 267: return "OpGroupUMin"; + case 268: return "OpGroupSMin"; + case 269: return "OpGroupFMax"; + case 270: return "OpGroupUMax"; + case 271: return "OpGroupSMax"; + case 272: return "Bad"; + case 273: return "Bad"; + case 274: return "OpReadPipe"; + case 275: return "OpWritePipe"; + case 276: return "OpReservedReadPipe"; + case 277: return "OpReservedWritePipe"; + case 278: return "OpReserveReadPipePackets"; + case 279: return "OpReserveWritePipePackets"; + case 280: return "OpCommitReadPipe"; + case 281: return "OpCommitWritePipe"; + case 282: return "OpIsValidReserveId"; + case 283: return "OpGetNumPipePackets"; + case 284: return "OpGetMaxPipePackets"; + case 285: return "OpGroupReserveReadPipePackets"; + case 286: return "OpGroupReserveWritePipePackets"; + case 287: return "OpGroupCommitReadPipe"; + case 288: return "OpGroupCommitWritePipe"; + case 289: return "Bad"; + case 290: return "Bad"; + case 291: return "OpEnqueueMarker"; + case 292: return "OpEnqueueKernel"; + case 293: return "OpGetKernelNDrangeSubGroupCount"; + case 294: return "OpGetKernelNDrangeMaxSubGroupSize"; + case 295: return "OpGetKernelWorkGroupSize"; + case 296: return "OpGetKernelPreferredWorkGroupSizeMultiple"; + case 297: return "OpRetainEvent"; + case 298: return "OpReleaseEvent"; + case 299: return "OpCreateUserEvent"; + case 300: return "OpIsValidEvent"; + case 301: return "OpSetUserEventStatus"; + case 302: return "OpCaptureEventProfilingInfo"; + case 303: return "OpGetDefaultQueue"; + case 304: return "OpBuildNDRange"; + case 305: return "OpImageSparseSampleImplicitLod"; + case 306: return "OpImageSparseSampleExplicitLod"; + case 307: return "OpImageSparseSampleDrefImplicitLod"; + case 308: return "OpImageSparseSampleDrefExplicitLod"; + case 309: return "OpImageSparseSampleProjImplicitLod"; + case 310: return "OpImageSparseSampleProjExplicitLod"; + case 311: return "OpImageSparseSampleProjDrefImplicitLod"; + case 312: return "OpImageSparseSampleProjDrefExplicitLod"; + case 313: return "OpImageSparseFetch"; + case 314: return "OpImageSparseGather"; + case 315: return "OpImageSparseDrefGather"; + case 316: return "OpImageSparseTexelsResident"; + case 317: return "OpNoLine"; + case 318: return "OpAtomicFlagTestAndSet"; + case 319: return "OpAtomicFlagClear"; + case 320: return "OpImageSparseRead"; + + case OpModuleProcessed: return "OpModuleProcessed"; + case OpDecorateId: return "OpDecorateId"; + + case 333: return "OpGroupNonUniformElect"; + case 334: return "OpGroupNonUniformAll"; + case 335: return "OpGroupNonUniformAny"; + case 336: return "OpGroupNonUniformAllEqual"; + case 337: return "OpGroupNonUniformBroadcast"; + case 338: return "OpGroupNonUniformBroadcastFirst"; + case 339: return "OpGroupNonUniformBallot"; + case 340: return "OpGroupNonUniformInverseBallot"; + case 341: return "OpGroupNonUniformBallotBitExtract"; + case 342: return "OpGroupNonUniformBallotBitCount"; + case 343: return "OpGroupNonUniformBallotFindLSB"; + case 344: return "OpGroupNonUniformBallotFindMSB"; + case 345: return "OpGroupNonUniformShuffle"; + case 346: return "OpGroupNonUniformShuffleXor"; + case 347: return "OpGroupNonUniformShuffleUp"; + case 348: return "OpGroupNonUniformShuffleDown"; + case 349: return "OpGroupNonUniformIAdd"; + case 350: return "OpGroupNonUniformFAdd"; + case 351: return "OpGroupNonUniformIMul"; + case 352: return "OpGroupNonUniformFMul"; + case 353: return "OpGroupNonUniformSMin"; + case 354: return "OpGroupNonUniformUMin"; + case 355: return "OpGroupNonUniformFMin"; + case 356: return "OpGroupNonUniformSMax"; + case 357: return "OpGroupNonUniformUMax"; + case 358: return "OpGroupNonUniformFMax"; + case 359: return "OpGroupNonUniformBitwiseAnd"; + case 360: return "OpGroupNonUniformBitwiseOr"; + case 361: return "OpGroupNonUniformBitwiseXor"; + case 362: return "OpGroupNonUniformLogicalAnd"; + case 363: return "OpGroupNonUniformLogicalOr"; + case 364: return "OpGroupNonUniformLogicalXor"; + case 365: return "OpGroupNonUniformQuadBroadcast"; + case 366: return "OpGroupNonUniformQuadSwap"; + + case 4421: return "OpSubgroupBallotKHR"; + case 4422: return "OpSubgroupFirstInvocationKHR"; + case 4428: return "OpSubgroupAllKHR"; + case 4429: return "OpSubgroupAnyKHR"; + case 4430: return "OpSubgroupAllEqualKHR"; + case 4432: return "OpSubgroupReadInvocationKHR"; + +#ifdef AMD_EXTENSIONS + case 5000: return "OpGroupIAddNonUniformAMD"; + case 5001: return "OpGroupFAddNonUniformAMD"; + case 5002: return "OpGroupFMinNonUniformAMD"; + case 5003: return "OpGroupUMinNonUniformAMD"; + case 5004: return "OpGroupSMinNonUniformAMD"; + case 5005: return "OpGroupFMaxNonUniformAMD"; + case 5006: return "OpGroupUMaxNonUniformAMD"; + case 5007: return "OpGroupSMaxNonUniformAMD"; + + case 5011: return "OpFragmentMaskFetchAMD"; + case 5012: return "OpFragmentFetchAMD"; +#endif + + case OpDecorateStringGOOGLE: return "OpDecorateStringGOOGLE"; + case OpMemberDecorateStringGOOGLE: return "OpMemberDecorateStringGOOGLE"; + +#ifdef NV_EXTENSIONS + case OpGroupNonUniformPartitionNV: return "OpGroupNonUniformPartitionNV"; + case OpReportIntersectionNV: return "OpReportIntersectionNV"; + case OpIgnoreIntersectionNV: return "OpIgnoreIntersectionNV"; + case OpTerminateRayNV: return "OpTerminateRayNV"; + case OpTraceNV: return "OpTraceNV"; + case OpTypeAccelerationStructureNV: return "OpTypeAccelerationStructureNV"; + case OpExecuteCallableNV: return "OpExecuteCallableNV"; + case OpImageSampleFootprintNV: return "OpImageSampleFootprintNV"; + case OpWritePackedPrimitiveIndices4x8NV: return "OpWritePackedPrimitiveIndices4x8NV"; +#endif + + case OpTypeCooperativeMatrixNV: return "OpTypeCooperativeMatrixNV"; + case OpCooperativeMatrixLoadNV: return "OpCooperativeMatrixLoadNV"; + case OpCooperativeMatrixStoreNV: return "OpCooperativeMatrixStoreNV"; + case OpCooperativeMatrixMulAddNV: return "OpCooperativeMatrixMulAddNV"; + case OpCooperativeMatrixLengthNV: return "OpCooperativeMatrixLengthNV"; + + default: + return "Bad"; + } +} + +// The set of objects that hold all the instruction/operand +// parameterization information. +InstructionParameters InstructionDesc[OpCodeMask + 1]; +OperandParameters ExecutionModeOperands[ExecutionModeCeiling]; +OperandParameters DecorationOperands[DecorationCeiling]; + +EnumDefinition OperandClassParams[OperandCount]; +EnumParameters ExecutionModeParams[ExecutionModeCeiling]; +EnumParameters ImageOperandsParams[ImageOperandsCeiling]; +EnumParameters DecorationParams[DecorationCeiling]; +EnumParameters LoopControlParams[FunctionControlCeiling]; +EnumParameters SelectionControlParams[SelectControlCeiling]; +EnumParameters FunctionControlParams[FunctionControlCeiling]; +EnumParameters MemoryAccessParams[MemoryAccessCeiling]; + +// Set up all the parameterizing descriptions of the opcodes, operands, etc. +void Parameterize() +{ + // only do this once. + static bool initialized = false; + if (initialized) + return; + initialized = true; + + // Exceptions to having a result <id> and a resulting type <id>. + // (Everything is initialized to have both). + + InstructionDesc[OpNop].setResultAndType(false, false); + InstructionDesc[OpSource].setResultAndType(false, false); + InstructionDesc[OpSourceContinued].setResultAndType(false, false); + InstructionDesc[OpSourceExtension].setResultAndType(false, false); + InstructionDesc[OpExtension].setResultAndType(false, false); + InstructionDesc[OpExtInstImport].setResultAndType(true, false); + InstructionDesc[OpCapability].setResultAndType(false, false); + InstructionDesc[OpMemoryModel].setResultAndType(false, false); + InstructionDesc[OpEntryPoint].setResultAndType(false, false); + InstructionDesc[OpExecutionMode].setResultAndType(false, false); + InstructionDesc[OpTypeVoid].setResultAndType(true, false); + InstructionDesc[OpTypeBool].setResultAndType(true, false); + InstructionDesc[OpTypeInt].setResultAndType(true, false); + InstructionDesc[OpTypeFloat].setResultAndType(true, false); + InstructionDesc[OpTypeVector].setResultAndType(true, false); + InstructionDesc[OpTypeMatrix].setResultAndType(true, false); + InstructionDesc[OpTypeImage].setResultAndType(true, false); + InstructionDesc[OpTypeSampler].setResultAndType(true, false); + InstructionDesc[OpTypeSampledImage].setResultAndType(true, false); + InstructionDesc[OpTypeArray].setResultAndType(true, false); + InstructionDesc[OpTypeRuntimeArray].setResultAndType(true, false); + InstructionDesc[OpTypeStruct].setResultAndType(true, false); + InstructionDesc[OpTypeOpaque].setResultAndType(true, false); + InstructionDesc[OpTypePointer].setResultAndType(true, false); + InstructionDesc[OpTypeForwardPointer].setResultAndType(false, false); + InstructionDesc[OpTypeFunction].setResultAndType(true, false); + InstructionDesc[OpTypeEvent].setResultAndType(true, false); + InstructionDesc[OpTypeDeviceEvent].setResultAndType(true, false); + InstructionDesc[OpTypeReserveId].setResultAndType(true, false); + InstructionDesc[OpTypeQueue].setResultAndType(true, false); + InstructionDesc[OpTypePipe].setResultAndType(true, false); + InstructionDesc[OpFunctionEnd].setResultAndType(false, false); + InstructionDesc[OpStore].setResultAndType(false, false); + InstructionDesc[OpImageWrite].setResultAndType(false, false); + InstructionDesc[OpDecorationGroup].setResultAndType(true, false); + InstructionDesc[OpDecorate].setResultAndType(false, false); + InstructionDesc[OpDecorateId].setResultAndType(false, false); + InstructionDesc[OpDecorateStringGOOGLE].setResultAndType(false, false); + InstructionDesc[OpMemberDecorate].setResultAndType(false, false); + InstructionDesc[OpMemberDecorateStringGOOGLE].setResultAndType(false, false); + InstructionDesc[OpGroupDecorate].setResultAndType(false, false); + InstructionDesc[OpGroupMemberDecorate].setResultAndType(false, false); + InstructionDesc[OpName].setResultAndType(false, false); + InstructionDesc[OpMemberName].setResultAndType(false, false); + InstructionDesc[OpString].setResultAndType(true, false); + InstructionDesc[OpLine].setResultAndType(false, false); + InstructionDesc[OpNoLine].setResultAndType(false, false); + InstructionDesc[OpCopyMemory].setResultAndType(false, false); + InstructionDesc[OpCopyMemorySized].setResultAndType(false, false); + InstructionDesc[OpEmitVertex].setResultAndType(false, false); + InstructionDesc[OpEndPrimitive].setResultAndType(false, false); + InstructionDesc[OpEmitStreamVertex].setResultAndType(false, false); + InstructionDesc[OpEndStreamPrimitive].setResultAndType(false, false); + InstructionDesc[OpControlBarrier].setResultAndType(false, false); + InstructionDesc[OpMemoryBarrier].setResultAndType(false, false); + InstructionDesc[OpAtomicStore].setResultAndType(false, false); + InstructionDesc[OpLoopMerge].setResultAndType(false, false); + InstructionDesc[OpSelectionMerge].setResultAndType(false, false); + InstructionDesc[OpLabel].setResultAndType(true, false); + InstructionDesc[OpBranch].setResultAndType(false, false); + InstructionDesc[OpBranchConditional].setResultAndType(false, false); + InstructionDesc[OpSwitch].setResultAndType(false, false); + InstructionDesc[OpKill].setResultAndType(false, false); + InstructionDesc[OpReturn].setResultAndType(false, false); + InstructionDesc[OpReturnValue].setResultAndType(false, false); + InstructionDesc[OpUnreachable].setResultAndType(false, false); + InstructionDesc[OpLifetimeStart].setResultAndType(false, false); + InstructionDesc[OpLifetimeStop].setResultAndType(false, false); + InstructionDesc[OpCommitReadPipe].setResultAndType(false, false); + InstructionDesc[OpCommitWritePipe].setResultAndType(false, false); + InstructionDesc[OpGroupCommitWritePipe].setResultAndType(false, false); + InstructionDesc[OpGroupCommitReadPipe].setResultAndType(false, false); + InstructionDesc[OpCaptureEventProfilingInfo].setResultAndType(false, false); + InstructionDesc[OpSetUserEventStatus].setResultAndType(false, false); + InstructionDesc[OpRetainEvent].setResultAndType(false, false); + InstructionDesc[OpReleaseEvent].setResultAndType(false, false); + InstructionDesc[OpGroupWaitEvents].setResultAndType(false, false); + InstructionDesc[OpAtomicFlagClear].setResultAndType(false, false); + InstructionDesc[OpModuleProcessed].setResultAndType(false, false); + InstructionDesc[OpTypeCooperativeMatrixNV].setResultAndType(true, false); + InstructionDesc[OpCooperativeMatrixStoreNV].setResultAndType(false, false); + + // Specific additional context-dependent operands + + ExecutionModeOperands[ExecutionModeInvocations].push(OperandLiteralNumber, "'Number of <<Invocation,invocations>>'"); + + ExecutionModeOperands[ExecutionModeLocalSize].push(OperandLiteralNumber, "'x size'"); + ExecutionModeOperands[ExecutionModeLocalSize].push(OperandLiteralNumber, "'y size'"); + ExecutionModeOperands[ExecutionModeLocalSize].push(OperandLiteralNumber, "'z size'"); + + ExecutionModeOperands[ExecutionModeLocalSizeHint].push(OperandLiteralNumber, "'x size'"); + ExecutionModeOperands[ExecutionModeLocalSizeHint].push(OperandLiteralNumber, "'y size'"); + ExecutionModeOperands[ExecutionModeLocalSizeHint].push(OperandLiteralNumber, "'z size'"); + + ExecutionModeOperands[ExecutionModeOutputVertices].push(OperandLiteralNumber, "'Vertex count'"); + ExecutionModeOperands[ExecutionModeVecTypeHint].push(OperandLiteralNumber, "'Vector type'"); + + DecorationOperands[DecorationStream].push(OperandLiteralNumber, "'Stream Number'"); + DecorationOperands[DecorationLocation].push(OperandLiteralNumber, "'Location'"); + DecorationOperands[DecorationComponent].push(OperandLiteralNumber, "'Component'"); + DecorationOperands[DecorationIndex].push(OperandLiteralNumber, "'Index'"); + DecorationOperands[DecorationBinding].push(OperandLiteralNumber, "'Binding Point'"); + DecorationOperands[DecorationDescriptorSet].push(OperandLiteralNumber, "'Descriptor Set'"); + DecorationOperands[DecorationOffset].push(OperandLiteralNumber, "'Byte Offset'"); + DecorationOperands[DecorationXfbBuffer].push(OperandLiteralNumber, "'XFB Buffer Number'"); + DecorationOperands[DecorationXfbStride].push(OperandLiteralNumber, "'XFB Stride'"); + DecorationOperands[DecorationArrayStride].push(OperandLiteralNumber, "'Array Stride'"); + DecorationOperands[DecorationMatrixStride].push(OperandLiteralNumber, "'Matrix Stride'"); + DecorationOperands[DecorationBuiltIn].push(OperandLiteralNumber, "See <<BuiltIn,*BuiltIn*>>"); + DecorationOperands[DecorationFPRoundingMode].push(OperandFPRoundingMode, "'Floating-Point Rounding Mode'"); + DecorationOperands[DecorationFPFastMathMode].push(OperandFPFastMath, "'Fast-Math Mode'"); + DecorationOperands[DecorationLinkageAttributes].push(OperandLiteralString, "'Name'"); + DecorationOperands[DecorationLinkageAttributes].push(OperandLinkageType, "'Linkage Type'"); + DecorationOperands[DecorationFuncParamAttr].push(OperandFuncParamAttr, "'Function Parameter Attribute'"); + DecorationOperands[DecorationSpecId].push(OperandLiteralNumber, "'Specialization Constant ID'"); + DecorationOperands[DecorationInputAttachmentIndex].push(OperandLiteralNumber, "'Attachment Index'"); + DecorationOperands[DecorationAlignment].push(OperandLiteralNumber, "'Alignment'"); + + OperandClassParams[OperandSource].set(0, SourceString, 0); + OperandClassParams[OperandExecutionModel].set(0, ExecutionModelString, nullptr); + OperandClassParams[OperandAddressing].set(0, AddressingString, nullptr); + OperandClassParams[OperandMemory].set(0, MemoryString, nullptr); + OperandClassParams[OperandExecutionMode].set(ExecutionModeCeiling, ExecutionModeString, ExecutionModeParams); + OperandClassParams[OperandExecutionMode].setOperands(ExecutionModeOperands); + OperandClassParams[OperandStorage].set(0, StorageClassString, nullptr); + OperandClassParams[OperandDimensionality].set(0, DimensionString, nullptr); + OperandClassParams[OperandSamplerAddressingMode].set(0, SamplerAddressingModeString, nullptr); + OperandClassParams[OperandSamplerFilterMode].set(0, SamplerFilterModeString, nullptr); + OperandClassParams[OperandSamplerImageFormat].set(0, ImageFormatString, nullptr); + OperandClassParams[OperandImageChannelOrder].set(0, ImageChannelOrderString, nullptr); + OperandClassParams[OperandImageChannelDataType].set(0, ImageChannelDataTypeString, nullptr); + OperandClassParams[OperandImageOperands].set(ImageOperandsCeiling, ImageOperandsString, ImageOperandsParams, true); + OperandClassParams[OperandFPFastMath].set(0, FPFastMathString, nullptr, true); + OperandClassParams[OperandFPRoundingMode].set(0, FPRoundingModeString, nullptr); + OperandClassParams[OperandLinkageType].set(0, LinkageTypeString, nullptr); + OperandClassParams[OperandFuncParamAttr].set(0, FuncParamAttrString, nullptr); + OperandClassParams[OperandAccessQualifier].set(0, AccessQualifierString, nullptr); + OperandClassParams[OperandDecoration].set(DecorationCeiling, DecorationString, DecorationParams); + OperandClassParams[OperandDecoration].setOperands(DecorationOperands); + OperandClassParams[OperandBuiltIn].set(0, BuiltInString, nullptr); + OperandClassParams[OperandSelect].set(SelectControlCeiling, SelectControlString, SelectionControlParams, true); + OperandClassParams[OperandLoop].set(LoopControlCeiling, LoopControlString, LoopControlParams, true); + OperandClassParams[OperandFunction].set(FunctionControlCeiling, FunctionControlString, FunctionControlParams, true); + OperandClassParams[OperandMemorySemantics].set(0, MemorySemanticsString, nullptr, true); + OperandClassParams[OperandMemoryAccess].set(MemoryAccessCeiling, MemoryAccessString, MemoryAccessParams, true); + OperandClassParams[OperandScope].set(0, ScopeString, nullptr); + OperandClassParams[OperandGroupOperation].set(0, GroupOperationString, nullptr); + OperandClassParams[OperandKernelEnqueueFlags].set(0, KernelEnqueueFlagsString, nullptr); + OperandClassParams[OperandKernelProfilingInfo].set(0, KernelProfilingInfoString, nullptr, true); + OperandClassParams[OperandCapability].set(0, CapabilityString, nullptr); + OperandClassParams[OperandOpcode].set(OpCodeMask + 1, OpcodeString, 0); + + // set name of operator, an initial set of <id> style operands, and the description + + InstructionDesc[OpSource].operands.push(OperandSource, ""); + InstructionDesc[OpSource].operands.push(OperandLiteralNumber, "'Version'"); + InstructionDesc[OpSource].operands.push(OperandId, "'File'", true); + InstructionDesc[OpSource].operands.push(OperandLiteralString, "'Source'", true); + + InstructionDesc[OpSourceContinued].operands.push(OperandLiteralString, "'Continued Source'"); + + InstructionDesc[OpSourceExtension].operands.push(OperandLiteralString, "'Extension'"); + + InstructionDesc[OpName].operands.push(OperandId, "'Target'"); + InstructionDesc[OpName].operands.push(OperandLiteralString, "'Name'"); + + InstructionDesc[OpMemberName].operands.push(OperandId, "'Type'"); + InstructionDesc[OpMemberName].operands.push(OperandLiteralNumber, "'Member'"); + InstructionDesc[OpMemberName].operands.push(OperandLiteralString, "'Name'"); + + InstructionDesc[OpString].operands.push(OperandLiteralString, "'String'"); + + InstructionDesc[OpLine].operands.push(OperandId, "'File'"); + InstructionDesc[OpLine].operands.push(OperandLiteralNumber, "'Line'"); + InstructionDesc[OpLine].operands.push(OperandLiteralNumber, "'Column'"); + + InstructionDesc[OpExtension].operands.push(OperandLiteralString, "'Name'"); + + InstructionDesc[OpExtInstImport].operands.push(OperandLiteralString, "'Name'"); + + InstructionDesc[OpCapability].operands.push(OperandCapability, "'Capability'"); + + InstructionDesc[OpMemoryModel].operands.push(OperandAddressing, ""); + InstructionDesc[OpMemoryModel].operands.push(OperandMemory, ""); + + InstructionDesc[OpEntryPoint].operands.push(OperandExecutionModel, ""); + InstructionDesc[OpEntryPoint].operands.push(OperandId, "'Entry Point'"); + InstructionDesc[OpEntryPoint].operands.push(OperandLiteralString, "'Name'"); + InstructionDesc[OpEntryPoint].operands.push(OperandVariableIds, "'Interface'"); + + InstructionDesc[OpExecutionMode].operands.push(OperandId, "'Entry Point'"); + InstructionDesc[OpExecutionMode].operands.push(OperandExecutionMode, "'Mode'"); + InstructionDesc[OpExecutionMode].operands.push(OperandOptionalLiteral, "See <<Execution_Mode,Execution Mode>>"); + + InstructionDesc[OpTypeInt].operands.push(OperandLiteralNumber, "'Width'"); + InstructionDesc[OpTypeInt].operands.push(OperandLiteralNumber, "'Signedness'"); + + InstructionDesc[OpTypeFloat].operands.push(OperandLiteralNumber, "'Width'"); + + InstructionDesc[OpTypeVector].operands.push(OperandId, "'Component Type'"); + InstructionDesc[OpTypeVector].operands.push(OperandLiteralNumber, "'Component Count'"); + + InstructionDesc[OpTypeMatrix].operands.push(OperandId, "'Column Type'"); + InstructionDesc[OpTypeMatrix].operands.push(OperandLiteralNumber, "'Column Count'"); + + InstructionDesc[OpTypeImage].operands.push(OperandId, "'Sampled Type'"); + InstructionDesc[OpTypeImage].operands.push(OperandDimensionality, ""); + InstructionDesc[OpTypeImage].operands.push(OperandLiteralNumber, "'Depth'"); + InstructionDesc[OpTypeImage].operands.push(OperandLiteralNumber, "'Arrayed'"); + InstructionDesc[OpTypeImage].operands.push(OperandLiteralNumber, "'MS'"); + InstructionDesc[OpTypeImage].operands.push(OperandLiteralNumber, "'Sampled'"); + InstructionDesc[OpTypeImage].operands.push(OperandSamplerImageFormat, ""); + InstructionDesc[OpTypeImage].operands.push(OperandAccessQualifier, "", true); + + InstructionDesc[OpTypeSampledImage].operands.push(OperandId, "'Image Type'"); + + InstructionDesc[OpTypeArray].operands.push(OperandId, "'Element Type'"); + InstructionDesc[OpTypeArray].operands.push(OperandId, "'Length'"); + + InstructionDesc[OpTypeRuntimeArray].operands.push(OperandId, "'Element Type'"); + + InstructionDesc[OpTypeStruct].operands.push(OperandVariableIds, "'Member 0 type', +\n'member 1 type', +\n..."); + + InstructionDesc[OpTypeOpaque].operands.push(OperandLiteralString, "The name of the opaque type."); + + InstructionDesc[OpTypePointer].operands.push(OperandStorage, ""); + InstructionDesc[OpTypePointer].operands.push(OperandId, "'Type'"); + + InstructionDesc[OpTypeForwardPointer].operands.push(OperandId, "'Pointer Type'"); + InstructionDesc[OpTypeForwardPointer].operands.push(OperandStorage, ""); + + InstructionDesc[OpTypePipe].operands.push(OperandAccessQualifier, "'Qualifier'"); + + InstructionDesc[OpTypeFunction].operands.push(OperandId, "'Return Type'"); + InstructionDesc[OpTypeFunction].operands.push(OperandVariableIds, "'Parameter 0 Type', +\n'Parameter 1 Type', +\n..."); + + InstructionDesc[OpConstant].operands.push(OperandVariableLiterals, "'Value'"); + + InstructionDesc[OpConstantComposite].operands.push(OperandVariableIds, "'Constituents'"); + + InstructionDesc[OpConstantSampler].operands.push(OperandSamplerAddressingMode, ""); + InstructionDesc[OpConstantSampler].operands.push(OperandLiteralNumber, "'Param'"); + InstructionDesc[OpConstantSampler].operands.push(OperandSamplerFilterMode, ""); + + InstructionDesc[OpSpecConstant].operands.push(OperandVariableLiterals, "'Value'"); + + InstructionDesc[OpSpecConstantComposite].operands.push(OperandVariableIds, "'Constituents'"); + + InstructionDesc[OpSpecConstantOp].operands.push(OperandLiteralNumber, "'Opcode'"); + InstructionDesc[OpSpecConstantOp].operands.push(OperandVariableIds, "'Operands'"); + + InstructionDesc[OpVariable].operands.push(OperandStorage, ""); + InstructionDesc[OpVariable].operands.push(OperandId, "'Initializer'", true); + + InstructionDesc[OpFunction].operands.push(OperandFunction, ""); + InstructionDesc[OpFunction].operands.push(OperandId, "'Function Type'"); + + InstructionDesc[OpFunctionCall].operands.push(OperandId, "'Function'"); + InstructionDesc[OpFunctionCall].operands.push(OperandVariableIds, "'Argument 0', +\n'Argument 1', +\n..."); + + InstructionDesc[OpExtInst].operands.push(OperandId, "'Set'"); + InstructionDesc[OpExtInst].operands.push(OperandLiteralNumber, "'Instruction'"); + InstructionDesc[OpExtInst].operands.push(OperandVariableIds, "'Operand 1', +\n'Operand 2', +\n..."); + + InstructionDesc[OpLoad].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpLoad].operands.push(OperandMemoryAccess, "", true); + InstructionDesc[OpLoad].operands.push(OperandLiteralNumber, "", true); + InstructionDesc[OpLoad].operands.push(OperandId, "", true); + + InstructionDesc[OpStore].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpStore].operands.push(OperandId, "'Object'"); + InstructionDesc[OpStore].operands.push(OperandMemoryAccess, "", true); + InstructionDesc[OpStore].operands.push(OperandLiteralNumber, "", true); + InstructionDesc[OpStore].operands.push(OperandId, "", true); + + InstructionDesc[OpPhi].operands.push(OperandVariableIds, "'Variable, Parent, ...'"); + + InstructionDesc[OpDecorate].operands.push(OperandId, "'Target'"); + InstructionDesc[OpDecorate].operands.push(OperandDecoration, ""); + InstructionDesc[OpDecorate].operands.push(OperandVariableLiterals, "See <<Decoration,'Decoration'>>."); + + InstructionDesc[OpDecorateId].operands.push(OperandId, "'Target'"); + InstructionDesc[OpDecorateId].operands.push(OperandDecoration, ""); + InstructionDesc[OpDecorateId].operands.push(OperandVariableIds, "See <<Decoration,'Decoration'>>."); + + InstructionDesc[OpDecorateStringGOOGLE].operands.push(OperandId, "'Target'"); + InstructionDesc[OpDecorateStringGOOGLE].operands.push(OperandDecoration, ""); + InstructionDesc[OpDecorateStringGOOGLE].operands.push(OperandLiteralString, "'Literal String'"); + + InstructionDesc[OpMemberDecorate].operands.push(OperandId, "'Structure Type'"); + InstructionDesc[OpMemberDecorate].operands.push(OperandLiteralNumber, "'Member'"); + InstructionDesc[OpMemberDecorate].operands.push(OperandDecoration, ""); + InstructionDesc[OpMemberDecorate].operands.push(OperandVariableLiterals, "See <<Decoration,'Decoration'>>."); + + InstructionDesc[OpMemberDecorateStringGOOGLE].operands.push(OperandId, "'Structure Type'"); + InstructionDesc[OpMemberDecorateStringGOOGLE].operands.push(OperandLiteralNumber, "'Member'"); + InstructionDesc[OpMemberDecorateStringGOOGLE].operands.push(OperandDecoration, ""); + InstructionDesc[OpMemberDecorateStringGOOGLE].operands.push(OperandLiteralString, "'Literal String'"); + + InstructionDesc[OpGroupDecorate].operands.push(OperandId, "'Decoration Group'"); + InstructionDesc[OpGroupDecorate].operands.push(OperandVariableIds, "'Targets'"); + + InstructionDesc[OpGroupMemberDecorate].operands.push(OperandId, "'Decoration Group'"); + InstructionDesc[OpGroupMemberDecorate].operands.push(OperandVariableIdLiteral, "'Targets'"); + + InstructionDesc[OpVectorExtractDynamic].operands.push(OperandId, "'Vector'"); + InstructionDesc[OpVectorExtractDynamic].operands.push(OperandId, "'Index'"); + + InstructionDesc[OpVectorInsertDynamic].operands.push(OperandId, "'Vector'"); + InstructionDesc[OpVectorInsertDynamic].operands.push(OperandId, "'Component'"); + InstructionDesc[OpVectorInsertDynamic].operands.push(OperandId, "'Index'"); + + InstructionDesc[OpVectorShuffle].operands.push(OperandId, "'Vector 1'"); + InstructionDesc[OpVectorShuffle].operands.push(OperandId, "'Vector 2'"); + InstructionDesc[OpVectorShuffle].operands.push(OperandVariableLiterals, "'Components'"); + + InstructionDesc[OpCompositeConstruct].operands.push(OperandVariableIds, "'Constituents'"); + + InstructionDesc[OpCompositeExtract].operands.push(OperandId, "'Composite'"); + InstructionDesc[OpCompositeExtract].operands.push(OperandVariableLiterals, "'Indexes'"); + + InstructionDesc[OpCompositeInsert].operands.push(OperandId, "'Object'"); + InstructionDesc[OpCompositeInsert].operands.push(OperandId, "'Composite'"); + InstructionDesc[OpCompositeInsert].operands.push(OperandVariableLiterals, "'Indexes'"); + + InstructionDesc[OpCopyObject].operands.push(OperandId, "'Operand'"); + + InstructionDesc[OpCopyMemory].operands.push(OperandId, "'Target'"); + InstructionDesc[OpCopyMemory].operands.push(OperandId, "'Source'"); + InstructionDesc[OpCopyMemory].operands.push(OperandMemoryAccess, "", true); + + InstructionDesc[OpCopyMemorySized].operands.push(OperandId, "'Target'"); + InstructionDesc[OpCopyMemorySized].operands.push(OperandId, "'Source'"); + InstructionDesc[OpCopyMemorySized].operands.push(OperandId, "'Size'"); + InstructionDesc[OpCopyMemorySized].operands.push(OperandMemoryAccess, "", true); + + InstructionDesc[OpSampledImage].operands.push(OperandId, "'Image'"); + InstructionDesc[OpSampledImage].operands.push(OperandId, "'Sampler'"); + + InstructionDesc[OpImage].operands.push(OperandId, "'Sampled Image'"); + + InstructionDesc[OpImageRead].operands.push(OperandId, "'Image'"); + InstructionDesc[OpImageRead].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageRead].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageRead].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageWrite].operands.push(OperandId, "'Image'"); + InstructionDesc[OpImageWrite].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageWrite].operands.push(OperandId, "'Texel'"); + InstructionDesc[OpImageWrite].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageWrite].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSampleImplicitLod].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageSampleImplicitLod].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSampleImplicitLod].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSampleImplicitLod].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSampleExplicitLod].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageSampleExplicitLod].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSampleExplicitLod].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSampleExplicitLod].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSampleDrefImplicitLod].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageSampleDrefImplicitLod].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSampleDrefImplicitLod].operands.push(OperandId, "'D~ref~'"); + InstructionDesc[OpImageSampleDrefImplicitLod].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSampleDrefImplicitLod].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSampleDrefExplicitLod].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageSampleDrefExplicitLod].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSampleDrefExplicitLod].operands.push(OperandId, "'D~ref~'"); + InstructionDesc[OpImageSampleDrefExplicitLod].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSampleDrefExplicitLod].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSampleProjImplicitLod].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageSampleProjImplicitLod].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSampleProjImplicitLod].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSampleProjImplicitLod].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSampleProjExplicitLod].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageSampleProjExplicitLod].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSampleProjExplicitLod].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSampleProjExplicitLod].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSampleProjDrefImplicitLod].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageSampleProjDrefImplicitLod].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSampleProjDrefImplicitLod].operands.push(OperandId, "'D~ref~'"); + InstructionDesc[OpImageSampleProjDrefImplicitLod].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSampleProjDrefImplicitLod].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSampleProjDrefExplicitLod].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageSampleProjDrefExplicitLod].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSampleProjDrefExplicitLod].operands.push(OperandId, "'D~ref~'"); + InstructionDesc[OpImageSampleProjDrefExplicitLod].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSampleProjDrefExplicitLod].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageFetch].operands.push(OperandId, "'Image'"); + InstructionDesc[OpImageFetch].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageFetch].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageFetch].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageGather].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageGather].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageGather].operands.push(OperandId, "'Component'"); + InstructionDesc[OpImageGather].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageGather].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageDrefGather].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageDrefGather].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageDrefGather].operands.push(OperandId, "'D~ref~'"); + InstructionDesc[OpImageDrefGather].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageDrefGather].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSparseSampleImplicitLod].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageSparseSampleImplicitLod].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSparseSampleImplicitLod].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSparseSampleImplicitLod].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSparseSampleExplicitLod].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageSparseSampleExplicitLod].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSparseSampleExplicitLod].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSparseSampleExplicitLod].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSparseSampleDrefImplicitLod].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageSparseSampleDrefImplicitLod].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSparseSampleDrefImplicitLod].operands.push(OperandId, "'D~ref~'"); + InstructionDesc[OpImageSparseSampleDrefImplicitLod].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSparseSampleDrefImplicitLod].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSparseSampleDrefExplicitLod].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageSparseSampleDrefExplicitLod].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSparseSampleDrefExplicitLod].operands.push(OperandId, "'D~ref~'"); + InstructionDesc[OpImageSparseSampleDrefExplicitLod].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSparseSampleDrefExplicitLod].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSparseSampleProjImplicitLod].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageSparseSampleProjImplicitLod].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSparseSampleProjImplicitLod].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSparseSampleProjImplicitLod].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSparseSampleProjExplicitLod].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageSparseSampleProjExplicitLod].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSparseSampleProjExplicitLod].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSparseSampleProjExplicitLod].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSparseSampleProjDrefImplicitLod].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageSparseSampleProjDrefImplicitLod].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSparseSampleProjDrefImplicitLod].operands.push(OperandId, "'D~ref~'"); + InstructionDesc[OpImageSparseSampleProjDrefImplicitLod].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSparseSampleProjDrefImplicitLod].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSparseSampleProjDrefExplicitLod].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageSparseSampleProjDrefExplicitLod].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSparseSampleProjDrefExplicitLod].operands.push(OperandId, "'D~ref~'"); + InstructionDesc[OpImageSparseSampleProjDrefExplicitLod].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSparseSampleProjDrefExplicitLod].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSparseFetch].operands.push(OperandId, "'Image'"); + InstructionDesc[OpImageSparseFetch].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSparseFetch].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSparseFetch].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSparseGather].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageSparseGather].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSparseGather].operands.push(OperandId, "'Component'"); + InstructionDesc[OpImageSparseGather].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSparseGather].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSparseDrefGather].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageSparseDrefGather].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSparseDrefGather].operands.push(OperandId, "'D~ref~'"); + InstructionDesc[OpImageSparseDrefGather].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSparseDrefGather].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSparseRead].operands.push(OperandId, "'Image'"); + InstructionDesc[OpImageSparseRead].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSparseRead].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSparseRead].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpImageSparseTexelsResident].operands.push(OperandId, "'Resident Code'"); + + InstructionDesc[OpImageQuerySizeLod].operands.push(OperandId, "'Image'"); + InstructionDesc[OpImageQuerySizeLod].operands.push(OperandId, "'Level of Detail'"); + + InstructionDesc[OpImageQuerySize].operands.push(OperandId, "'Image'"); + + InstructionDesc[OpImageQueryLod].operands.push(OperandId, "'Image'"); + InstructionDesc[OpImageQueryLod].operands.push(OperandId, "'Coordinate'"); + + InstructionDesc[OpImageQueryLevels].operands.push(OperandId, "'Image'"); + + InstructionDesc[OpImageQuerySamples].operands.push(OperandId, "'Image'"); + + InstructionDesc[OpImageQueryFormat].operands.push(OperandId, "'Image'"); + + InstructionDesc[OpImageQueryOrder].operands.push(OperandId, "'Image'"); + + InstructionDesc[OpAccessChain].operands.push(OperandId, "'Base'"); + InstructionDesc[OpAccessChain].operands.push(OperandVariableIds, "'Indexes'"); + + InstructionDesc[OpInBoundsAccessChain].operands.push(OperandId, "'Base'"); + InstructionDesc[OpInBoundsAccessChain].operands.push(OperandVariableIds, "'Indexes'"); + + InstructionDesc[OpPtrAccessChain].operands.push(OperandId, "'Base'"); + InstructionDesc[OpPtrAccessChain].operands.push(OperandId, "'Element'"); + InstructionDesc[OpPtrAccessChain].operands.push(OperandVariableIds, "'Indexes'"); + + InstructionDesc[OpInBoundsPtrAccessChain].operands.push(OperandId, "'Base'"); + InstructionDesc[OpInBoundsPtrAccessChain].operands.push(OperandId, "'Element'"); + InstructionDesc[OpInBoundsPtrAccessChain].operands.push(OperandVariableIds, "'Indexes'"); + + InstructionDesc[OpSNegate].operands.push(OperandId, "'Operand'"); + + InstructionDesc[OpFNegate].operands.push(OperandId, "'Operand'"); + + InstructionDesc[OpNot].operands.push(OperandId, "'Operand'"); + + InstructionDesc[OpAny].operands.push(OperandId, "'Vector'"); + + InstructionDesc[OpAll].operands.push(OperandId, "'Vector'"); + + InstructionDesc[OpConvertFToU].operands.push(OperandId, "'Float Value'"); + + InstructionDesc[OpConvertFToS].operands.push(OperandId, "'Float Value'"); + + InstructionDesc[OpConvertSToF].operands.push(OperandId, "'Signed Value'"); + + InstructionDesc[OpConvertUToF].operands.push(OperandId, "'Unsigned Value'"); + + InstructionDesc[OpUConvert].operands.push(OperandId, "'Unsigned Value'"); + + InstructionDesc[OpSConvert].operands.push(OperandId, "'Signed Value'"); + + InstructionDesc[OpFConvert].operands.push(OperandId, "'Float Value'"); + + InstructionDesc[OpSatConvertSToU].operands.push(OperandId, "'Signed Value'"); + + InstructionDesc[OpSatConvertUToS].operands.push(OperandId, "'Unsigned Value'"); + + InstructionDesc[OpConvertPtrToU].operands.push(OperandId, "'Pointer'"); + + InstructionDesc[OpConvertUToPtr].operands.push(OperandId, "'Integer Value'"); + + InstructionDesc[OpPtrCastToGeneric].operands.push(OperandId, "'Pointer'"); + + InstructionDesc[OpGenericCastToPtr].operands.push(OperandId, "'Pointer'"); + + InstructionDesc[OpGenericCastToPtrExplicit].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpGenericCastToPtrExplicit].operands.push(OperandStorage, "'Storage'"); + + InstructionDesc[OpGenericPtrMemSemantics].operands.push(OperandId, "'Pointer'"); + + InstructionDesc[OpBitcast].operands.push(OperandId, "'Operand'"); + + InstructionDesc[OpQuantizeToF16].operands.push(OperandId, "'Value'"); + + InstructionDesc[OpTranspose].operands.push(OperandId, "'Matrix'"); + + InstructionDesc[OpCopyLogical].operands.push(OperandId, "'Operand'"); + + InstructionDesc[OpIsNan].operands.push(OperandId, "'x'"); + + InstructionDesc[OpIsInf].operands.push(OperandId, "'x'"); + + InstructionDesc[OpIsFinite].operands.push(OperandId, "'x'"); + + InstructionDesc[OpIsNormal].operands.push(OperandId, "'x'"); + + InstructionDesc[OpSignBitSet].operands.push(OperandId, "'x'"); + + InstructionDesc[OpLessOrGreater].operands.push(OperandId, "'x'"); + InstructionDesc[OpLessOrGreater].operands.push(OperandId, "'y'"); + + InstructionDesc[OpOrdered].operands.push(OperandId, "'x'"); + InstructionDesc[OpOrdered].operands.push(OperandId, "'y'"); + + InstructionDesc[OpUnordered].operands.push(OperandId, "'x'"); + InstructionDesc[OpUnordered].operands.push(OperandId, "'y'"); + + InstructionDesc[OpArrayLength].operands.push(OperandId, "'Structure'"); + InstructionDesc[OpArrayLength].operands.push(OperandLiteralNumber, "'Array member'"); + + InstructionDesc[OpIAdd].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpIAdd].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpFAdd].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpFAdd].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpISub].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpISub].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpFSub].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpFSub].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpIMul].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpIMul].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpFMul].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpFMul].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpUDiv].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpUDiv].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpSDiv].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpSDiv].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpFDiv].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpFDiv].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpUMod].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpUMod].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpSRem].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpSRem].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpSMod].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpSMod].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpFRem].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpFRem].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpFMod].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpFMod].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpVectorTimesScalar].operands.push(OperandId, "'Vector'"); + InstructionDesc[OpVectorTimesScalar].operands.push(OperandId, "'Scalar'"); + + InstructionDesc[OpMatrixTimesScalar].operands.push(OperandId, "'Matrix'"); + InstructionDesc[OpMatrixTimesScalar].operands.push(OperandId, "'Scalar'"); + + InstructionDesc[OpVectorTimesMatrix].operands.push(OperandId, "'Vector'"); + InstructionDesc[OpVectorTimesMatrix].operands.push(OperandId, "'Matrix'"); + + InstructionDesc[OpMatrixTimesVector].operands.push(OperandId, "'Matrix'"); + InstructionDesc[OpMatrixTimesVector].operands.push(OperandId, "'Vector'"); + + InstructionDesc[OpMatrixTimesMatrix].operands.push(OperandId, "'LeftMatrix'"); + InstructionDesc[OpMatrixTimesMatrix].operands.push(OperandId, "'RightMatrix'"); + + InstructionDesc[OpOuterProduct].operands.push(OperandId, "'Vector 1'"); + InstructionDesc[OpOuterProduct].operands.push(OperandId, "'Vector 2'"); + + InstructionDesc[OpDot].operands.push(OperandId, "'Vector 1'"); + InstructionDesc[OpDot].operands.push(OperandId, "'Vector 2'"); + + InstructionDesc[OpIAddCarry].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpIAddCarry].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpISubBorrow].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpISubBorrow].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpUMulExtended].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpUMulExtended].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpSMulExtended].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpSMulExtended].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpShiftRightLogical].operands.push(OperandId, "'Base'"); + InstructionDesc[OpShiftRightLogical].operands.push(OperandId, "'Shift'"); + + InstructionDesc[OpShiftRightArithmetic].operands.push(OperandId, "'Base'"); + InstructionDesc[OpShiftRightArithmetic].operands.push(OperandId, "'Shift'"); + + InstructionDesc[OpShiftLeftLogical].operands.push(OperandId, "'Base'"); + InstructionDesc[OpShiftLeftLogical].operands.push(OperandId, "'Shift'"); + + InstructionDesc[OpLogicalOr].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpLogicalOr].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpLogicalAnd].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpLogicalAnd].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpLogicalEqual].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpLogicalEqual].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpLogicalNotEqual].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpLogicalNotEqual].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpLogicalNot].operands.push(OperandId, "'Operand'"); + + InstructionDesc[OpBitwiseOr].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpBitwiseOr].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpBitwiseXor].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpBitwiseXor].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpBitwiseAnd].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpBitwiseAnd].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpBitFieldInsert].operands.push(OperandId, "'Base'"); + InstructionDesc[OpBitFieldInsert].operands.push(OperandId, "'Insert'"); + InstructionDesc[OpBitFieldInsert].operands.push(OperandId, "'Offset'"); + InstructionDesc[OpBitFieldInsert].operands.push(OperandId, "'Count'"); + + InstructionDesc[OpBitFieldSExtract].operands.push(OperandId, "'Base'"); + InstructionDesc[OpBitFieldSExtract].operands.push(OperandId, "'Offset'"); + InstructionDesc[OpBitFieldSExtract].operands.push(OperandId, "'Count'"); + + InstructionDesc[OpBitFieldUExtract].operands.push(OperandId, "'Base'"); + InstructionDesc[OpBitFieldUExtract].operands.push(OperandId, "'Offset'"); + InstructionDesc[OpBitFieldUExtract].operands.push(OperandId, "'Count'"); + + InstructionDesc[OpBitReverse].operands.push(OperandId, "'Base'"); + + InstructionDesc[OpBitCount].operands.push(OperandId, "'Base'"); + + InstructionDesc[OpSelect].operands.push(OperandId, "'Condition'"); + InstructionDesc[OpSelect].operands.push(OperandId, "'Object 1'"); + InstructionDesc[OpSelect].operands.push(OperandId, "'Object 2'"); + + InstructionDesc[OpIEqual].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpIEqual].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpFOrdEqual].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpFOrdEqual].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpFUnordEqual].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpFUnordEqual].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpINotEqual].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpINotEqual].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpFOrdNotEqual].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpFOrdNotEqual].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpFUnordNotEqual].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpFUnordNotEqual].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpULessThan].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpULessThan].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpSLessThan].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpSLessThan].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpFOrdLessThan].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpFOrdLessThan].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpFUnordLessThan].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpFUnordLessThan].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpUGreaterThan].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpUGreaterThan].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpSGreaterThan].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpSGreaterThan].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpFOrdGreaterThan].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpFOrdGreaterThan].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpFUnordGreaterThan].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpFUnordGreaterThan].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpULessThanEqual].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpULessThanEqual].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpSLessThanEqual].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpSLessThanEqual].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpFOrdLessThanEqual].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpFOrdLessThanEqual].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpFUnordLessThanEqual].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpFUnordLessThanEqual].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpUGreaterThanEqual].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpUGreaterThanEqual].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpSGreaterThanEqual].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpSGreaterThanEqual].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpFOrdGreaterThanEqual].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpFOrdGreaterThanEqual].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpFUnordGreaterThanEqual].operands.push(OperandId, "'Operand 1'"); + InstructionDesc[OpFUnordGreaterThanEqual].operands.push(OperandId, "'Operand 2'"); + + InstructionDesc[OpDPdx].operands.push(OperandId, "'P'"); + + InstructionDesc[OpDPdy].operands.push(OperandId, "'P'"); + + InstructionDesc[OpFwidth].operands.push(OperandId, "'P'"); + + InstructionDesc[OpDPdxFine].operands.push(OperandId, "'P'"); + + InstructionDesc[OpDPdyFine].operands.push(OperandId, "'P'"); + + InstructionDesc[OpFwidthFine].operands.push(OperandId, "'P'"); + + InstructionDesc[OpDPdxCoarse].operands.push(OperandId, "'P'"); + + InstructionDesc[OpDPdyCoarse].operands.push(OperandId, "'P'"); + + InstructionDesc[OpFwidthCoarse].operands.push(OperandId, "'P'"); + + InstructionDesc[OpEmitStreamVertex].operands.push(OperandId, "'Stream'"); + + InstructionDesc[OpEndStreamPrimitive].operands.push(OperandId, "'Stream'"); + + InstructionDesc[OpControlBarrier].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpControlBarrier].operands.push(OperandScope, "'Memory'"); + InstructionDesc[OpControlBarrier].operands.push(OperandMemorySemantics, "'Semantics'"); + + InstructionDesc[OpMemoryBarrier].operands.push(OperandScope, "'Memory'"); + InstructionDesc[OpMemoryBarrier].operands.push(OperandMemorySemantics, "'Semantics'"); + + InstructionDesc[OpImageTexelPointer].operands.push(OperandId, "'Image'"); + InstructionDesc[OpImageTexelPointer].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageTexelPointer].operands.push(OperandId, "'Sample'"); + + InstructionDesc[OpAtomicLoad].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicLoad].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicLoad].operands.push(OperandMemorySemantics, "'Semantics'"); + + InstructionDesc[OpAtomicStore].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicStore].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicStore].operands.push(OperandMemorySemantics, "'Semantics'"); + InstructionDesc[OpAtomicStore].operands.push(OperandId, "'Value'"); + + InstructionDesc[OpAtomicExchange].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicExchange].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicExchange].operands.push(OperandMemorySemantics, "'Semantics'"); + InstructionDesc[OpAtomicExchange].operands.push(OperandId, "'Value'"); + + InstructionDesc[OpAtomicCompareExchange].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicCompareExchange].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicCompareExchange].operands.push(OperandMemorySemantics, "'Equal'"); + InstructionDesc[OpAtomicCompareExchange].operands.push(OperandMemorySemantics, "'Unequal'"); + InstructionDesc[OpAtomicCompareExchange].operands.push(OperandId, "'Value'"); + InstructionDesc[OpAtomicCompareExchange].operands.push(OperandId, "'Comparator'"); + + InstructionDesc[OpAtomicCompareExchangeWeak].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicCompareExchangeWeak].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicCompareExchangeWeak].operands.push(OperandMemorySemantics, "'Equal'"); + InstructionDesc[OpAtomicCompareExchangeWeak].operands.push(OperandMemorySemantics, "'Unequal'"); + InstructionDesc[OpAtomicCompareExchangeWeak].operands.push(OperandId, "'Value'"); + InstructionDesc[OpAtomicCompareExchangeWeak].operands.push(OperandId, "'Comparator'"); + + InstructionDesc[OpAtomicIIncrement].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicIIncrement].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicIIncrement].operands.push(OperandMemorySemantics, "'Semantics'"); + + InstructionDesc[OpAtomicIDecrement].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicIDecrement].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicIDecrement].operands.push(OperandMemorySemantics, "'Semantics'"); + + InstructionDesc[OpAtomicIAdd].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicIAdd].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicIAdd].operands.push(OperandMemorySemantics, "'Semantics'"); + InstructionDesc[OpAtomicIAdd].operands.push(OperandId, "'Value'"); + + InstructionDesc[OpAtomicISub].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicISub].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicISub].operands.push(OperandMemorySemantics, "'Semantics'"); + InstructionDesc[OpAtomicISub].operands.push(OperandId, "'Value'"); + + InstructionDesc[OpAtomicUMin].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicUMin].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicUMin].operands.push(OperandMemorySemantics, "'Semantics'"); + InstructionDesc[OpAtomicUMin].operands.push(OperandId, "'Value'"); + + InstructionDesc[OpAtomicUMax].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicUMax].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicUMax].operands.push(OperandMemorySemantics, "'Semantics'"); + InstructionDesc[OpAtomicUMax].operands.push(OperandId, "'Value'"); + + InstructionDesc[OpAtomicSMin].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicSMin].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicSMin].operands.push(OperandMemorySemantics, "'Semantics'"); + InstructionDesc[OpAtomicSMin].operands.push(OperandId, "'Value'"); + + InstructionDesc[OpAtomicSMax].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicSMax].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicSMax].operands.push(OperandMemorySemantics, "'Semantics'"); + InstructionDesc[OpAtomicSMax].operands.push(OperandId, "'Value'"); + + InstructionDesc[OpAtomicAnd].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicAnd].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicAnd].operands.push(OperandMemorySemantics, "'Semantics'"); + InstructionDesc[OpAtomicAnd].operands.push(OperandId, "'Value'"); + + InstructionDesc[OpAtomicOr].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicOr].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicOr].operands.push(OperandMemorySemantics, "'Semantics'"); + InstructionDesc[OpAtomicOr].operands.push(OperandId, "'Value'"); + + InstructionDesc[OpAtomicXor].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicXor].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicXor].operands.push(OperandMemorySemantics, "'Semantics'"); + InstructionDesc[OpAtomicXor].operands.push(OperandId, "'Value'"); + + InstructionDesc[OpAtomicFlagTestAndSet].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicFlagTestAndSet].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicFlagTestAndSet].operands.push(OperandMemorySemantics, "'Semantics'"); + + InstructionDesc[OpAtomicFlagClear].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicFlagClear].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicFlagClear].operands.push(OperandMemorySemantics, "'Semantics'"); + + InstructionDesc[OpLoopMerge].operands.push(OperandId, "'Merge Block'"); + InstructionDesc[OpLoopMerge].operands.push(OperandId, "'Continue Target'"); + InstructionDesc[OpLoopMerge].operands.push(OperandLoop, ""); + InstructionDesc[OpLoopMerge].operands.push(OperandOptionalLiteral, ""); + + InstructionDesc[OpSelectionMerge].operands.push(OperandId, "'Merge Block'"); + InstructionDesc[OpSelectionMerge].operands.push(OperandSelect, ""); + + InstructionDesc[OpBranch].operands.push(OperandId, "'Target Label'"); + + InstructionDesc[OpBranchConditional].operands.push(OperandId, "'Condition'"); + InstructionDesc[OpBranchConditional].operands.push(OperandId, "'True Label'"); + InstructionDesc[OpBranchConditional].operands.push(OperandId, "'False Label'"); + InstructionDesc[OpBranchConditional].operands.push(OperandVariableLiterals, "'Branch weights'"); + + InstructionDesc[OpSwitch].operands.push(OperandId, "'Selector'"); + InstructionDesc[OpSwitch].operands.push(OperandId, "'Default'"); + InstructionDesc[OpSwitch].operands.push(OperandVariableLiteralId, "'Target'"); + + + InstructionDesc[OpReturnValue].operands.push(OperandId, "'Value'"); + + InstructionDesc[OpLifetimeStart].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpLifetimeStart].operands.push(OperandLiteralNumber, "'Size'"); + + InstructionDesc[OpLifetimeStop].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpLifetimeStop].operands.push(OperandLiteralNumber, "'Size'"); + + InstructionDesc[OpGroupAsyncCopy].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupAsyncCopy].operands.push(OperandId, "'Destination'"); + InstructionDesc[OpGroupAsyncCopy].operands.push(OperandId, "'Source'"); + InstructionDesc[OpGroupAsyncCopy].operands.push(OperandId, "'Num Elements'"); + InstructionDesc[OpGroupAsyncCopy].operands.push(OperandId, "'Stride'"); + InstructionDesc[OpGroupAsyncCopy].operands.push(OperandId, "'Event'"); + + InstructionDesc[OpGroupWaitEvents].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupWaitEvents].operands.push(OperandId, "'Num Events'"); + InstructionDesc[OpGroupWaitEvents].operands.push(OperandId, "'Events List'"); + + InstructionDesc[OpGroupAll].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupAll].operands.push(OperandId, "'Predicate'"); + + InstructionDesc[OpGroupAny].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupAny].operands.push(OperandId, "'Predicate'"); + + InstructionDesc[OpGroupBroadcast].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupBroadcast].operands.push(OperandId, "'Value'"); + InstructionDesc[OpGroupBroadcast].operands.push(OperandId, "'LocalId'"); + + InstructionDesc[OpGroupIAdd].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupIAdd].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupIAdd].operands.push(OperandId, "'X'"); + + InstructionDesc[OpGroupFAdd].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupFAdd].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupFAdd].operands.push(OperandId, "'X'"); + + InstructionDesc[OpGroupUMin].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupUMin].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupUMin].operands.push(OperandId, "'X'"); + + InstructionDesc[OpGroupSMin].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupSMin].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupSMin].operands.push(OperandId, "X"); + + InstructionDesc[OpGroupFMin].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupFMin].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupFMin].operands.push(OperandId, "X"); + + InstructionDesc[OpGroupUMax].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupUMax].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupUMax].operands.push(OperandId, "X"); + + InstructionDesc[OpGroupSMax].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupSMax].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupSMax].operands.push(OperandId, "X"); + + InstructionDesc[OpGroupFMax].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupFMax].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupFMax].operands.push(OperandId, "X"); + + InstructionDesc[OpReadPipe].operands.push(OperandId, "'Pipe'"); + InstructionDesc[OpReadPipe].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpReadPipe].operands.push(OperandId, "'Packet Size'"); + InstructionDesc[OpReadPipe].operands.push(OperandId, "'Packet Alignment'"); + + InstructionDesc[OpWritePipe].operands.push(OperandId, "'Pipe'"); + InstructionDesc[OpWritePipe].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpWritePipe].operands.push(OperandId, "'Packet Size'"); + InstructionDesc[OpWritePipe].operands.push(OperandId, "'Packet Alignment'"); + + InstructionDesc[OpReservedReadPipe].operands.push(OperandId, "'Pipe'"); + InstructionDesc[OpReservedReadPipe].operands.push(OperandId, "'Reserve Id'"); + InstructionDesc[OpReservedReadPipe].operands.push(OperandId, "'Index'"); + InstructionDesc[OpReservedReadPipe].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpReservedReadPipe].operands.push(OperandId, "'Packet Size'"); + InstructionDesc[OpReservedReadPipe].operands.push(OperandId, "'Packet Alignment'"); + + InstructionDesc[OpReservedWritePipe].operands.push(OperandId, "'Pipe'"); + InstructionDesc[OpReservedWritePipe].operands.push(OperandId, "'Reserve Id'"); + InstructionDesc[OpReservedWritePipe].operands.push(OperandId, "'Index'"); + InstructionDesc[OpReservedWritePipe].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpReservedWritePipe].operands.push(OperandId, "'Packet Size'"); + InstructionDesc[OpReservedWritePipe].operands.push(OperandId, "'Packet Alignment'"); + + InstructionDesc[OpReserveReadPipePackets].operands.push(OperandId, "'Pipe'"); + InstructionDesc[OpReserveReadPipePackets].operands.push(OperandId, "'Num Packets'"); + InstructionDesc[OpReserveReadPipePackets].operands.push(OperandId, "'Packet Size'"); + InstructionDesc[OpReserveReadPipePackets].operands.push(OperandId, "'Packet Alignment'"); + + InstructionDesc[OpReserveWritePipePackets].operands.push(OperandId, "'Pipe'"); + InstructionDesc[OpReserveWritePipePackets].operands.push(OperandId, "'Num Packets'"); + InstructionDesc[OpReserveWritePipePackets].operands.push(OperandId, "'Packet Size'"); + InstructionDesc[OpReserveWritePipePackets].operands.push(OperandId, "'Packet Alignment'"); + + InstructionDesc[OpCommitReadPipe].operands.push(OperandId, "'Pipe'"); + InstructionDesc[OpCommitReadPipe].operands.push(OperandId, "'Reserve Id'"); + InstructionDesc[OpCommitReadPipe].operands.push(OperandId, "'Packet Size'"); + InstructionDesc[OpCommitReadPipe].operands.push(OperandId, "'Packet Alignment'"); + + InstructionDesc[OpCommitWritePipe].operands.push(OperandId, "'Pipe'"); + InstructionDesc[OpCommitWritePipe].operands.push(OperandId, "'Reserve Id'"); + InstructionDesc[OpCommitWritePipe].operands.push(OperandId, "'Packet Size'"); + InstructionDesc[OpCommitWritePipe].operands.push(OperandId, "'Packet Alignment'"); + + InstructionDesc[OpIsValidReserveId].operands.push(OperandId, "'Reserve Id'"); + + InstructionDesc[OpGetNumPipePackets].operands.push(OperandId, "'Pipe'"); + InstructionDesc[OpGetNumPipePackets].operands.push(OperandId, "'Packet Size'"); + InstructionDesc[OpGetNumPipePackets].operands.push(OperandId, "'Packet Alignment'"); + + InstructionDesc[OpGetMaxPipePackets].operands.push(OperandId, "'Pipe'"); + InstructionDesc[OpGetMaxPipePackets].operands.push(OperandId, "'Packet Size'"); + InstructionDesc[OpGetMaxPipePackets].operands.push(OperandId, "'Packet Alignment'"); + + InstructionDesc[OpGroupReserveReadPipePackets].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupReserveReadPipePackets].operands.push(OperandId, "'Pipe'"); + InstructionDesc[OpGroupReserveReadPipePackets].operands.push(OperandId, "'Num Packets'"); + InstructionDesc[OpGroupReserveReadPipePackets].operands.push(OperandId, "'Packet Size'"); + InstructionDesc[OpGroupReserveReadPipePackets].operands.push(OperandId, "'Packet Alignment'"); + + InstructionDesc[OpGroupReserveWritePipePackets].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupReserveWritePipePackets].operands.push(OperandId, "'Pipe'"); + InstructionDesc[OpGroupReserveWritePipePackets].operands.push(OperandId, "'Num Packets'"); + InstructionDesc[OpGroupReserveWritePipePackets].operands.push(OperandId, "'Packet Size'"); + InstructionDesc[OpGroupReserveWritePipePackets].operands.push(OperandId, "'Packet Alignment'"); + + InstructionDesc[OpGroupCommitReadPipe].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupCommitReadPipe].operands.push(OperandId, "'Pipe'"); + InstructionDesc[OpGroupCommitReadPipe].operands.push(OperandId, "'Reserve Id'"); + InstructionDesc[OpGroupCommitReadPipe].operands.push(OperandId, "'Packet Size'"); + InstructionDesc[OpGroupCommitReadPipe].operands.push(OperandId, "'Packet Alignment'"); + + InstructionDesc[OpGroupCommitWritePipe].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupCommitWritePipe].operands.push(OperandId, "'Pipe'"); + InstructionDesc[OpGroupCommitWritePipe].operands.push(OperandId, "'Reserve Id'"); + InstructionDesc[OpGroupCommitWritePipe].operands.push(OperandId, "'Packet Size'"); + InstructionDesc[OpGroupCommitWritePipe].operands.push(OperandId, "'Packet Alignment'"); + + InstructionDesc[OpBuildNDRange].operands.push(OperandId, "'GlobalWorkSize'"); + InstructionDesc[OpBuildNDRange].operands.push(OperandId, "'LocalWorkSize'"); + InstructionDesc[OpBuildNDRange].operands.push(OperandId, "'GlobalWorkOffset'"); + + InstructionDesc[OpCaptureEventProfilingInfo].operands.push(OperandId, "'Event'"); + InstructionDesc[OpCaptureEventProfilingInfo].operands.push(OperandId, "'Profiling Info'"); + InstructionDesc[OpCaptureEventProfilingInfo].operands.push(OperandId, "'Value'"); + + InstructionDesc[OpSetUserEventStatus].operands.push(OperandId, "'Event'"); + InstructionDesc[OpSetUserEventStatus].operands.push(OperandId, "'Status'"); + + InstructionDesc[OpIsValidEvent].operands.push(OperandId, "'Event'"); + + InstructionDesc[OpRetainEvent].operands.push(OperandId, "'Event'"); + + InstructionDesc[OpReleaseEvent].operands.push(OperandId, "'Event'"); + + InstructionDesc[OpGetKernelWorkGroupSize].operands.push(OperandId, "'Invoke'"); + InstructionDesc[OpGetKernelWorkGroupSize].operands.push(OperandId, "'Param'"); + InstructionDesc[OpGetKernelWorkGroupSize].operands.push(OperandId, "'Param Size'"); + InstructionDesc[OpGetKernelWorkGroupSize].operands.push(OperandId, "'Param Align'"); + + InstructionDesc[OpGetKernelPreferredWorkGroupSizeMultiple].operands.push(OperandId, "'Invoke'"); + InstructionDesc[OpGetKernelPreferredWorkGroupSizeMultiple].operands.push(OperandId, "'Param'"); + InstructionDesc[OpGetKernelPreferredWorkGroupSizeMultiple].operands.push(OperandId, "'Param Size'"); + InstructionDesc[OpGetKernelPreferredWorkGroupSizeMultiple].operands.push(OperandId, "'Param Align'"); + + InstructionDesc[OpGetKernelNDrangeSubGroupCount].operands.push(OperandId, "'ND Range'"); + InstructionDesc[OpGetKernelNDrangeSubGroupCount].operands.push(OperandId, "'Invoke'"); + InstructionDesc[OpGetKernelNDrangeSubGroupCount].operands.push(OperandId, "'Param'"); + InstructionDesc[OpGetKernelNDrangeSubGroupCount].operands.push(OperandId, "'Param Size'"); + InstructionDesc[OpGetKernelNDrangeSubGroupCount].operands.push(OperandId, "'Param Align'"); + + InstructionDesc[OpGetKernelNDrangeMaxSubGroupSize].operands.push(OperandId, "'ND Range'"); + InstructionDesc[OpGetKernelNDrangeMaxSubGroupSize].operands.push(OperandId, "'Invoke'"); + InstructionDesc[OpGetKernelNDrangeMaxSubGroupSize].operands.push(OperandId, "'Param'"); + InstructionDesc[OpGetKernelNDrangeMaxSubGroupSize].operands.push(OperandId, "'Param Size'"); + InstructionDesc[OpGetKernelNDrangeMaxSubGroupSize].operands.push(OperandId, "'Param Align'"); + + InstructionDesc[OpEnqueueKernel].operands.push(OperandId, "'Queue'"); + InstructionDesc[OpEnqueueKernel].operands.push(OperandId, "'Flags'"); + InstructionDesc[OpEnqueueKernel].operands.push(OperandId, "'ND Range'"); + InstructionDesc[OpEnqueueKernel].operands.push(OperandId, "'Num Events'"); + InstructionDesc[OpEnqueueKernel].operands.push(OperandId, "'Wait Events'"); + InstructionDesc[OpEnqueueKernel].operands.push(OperandId, "'Ret Event'"); + InstructionDesc[OpEnqueueKernel].operands.push(OperandId, "'Invoke'"); + InstructionDesc[OpEnqueueKernel].operands.push(OperandId, "'Param'"); + InstructionDesc[OpEnqueueKernel].operands.push(OperandId, "'Param Size'"); + InstructionDesc[OpEnqueueKernel].operands.push(OperandId, "'Param Align'"); + InstructionDesc[OpEnqueueKernel].operands.push(OperandVariableIds, "'Local Size'"); + + InstructionDesc[OpEnqueueMarker].operands.push(OperandId, "'Queue'"); + InstructionDesc[OpEnqueueMarker].operands.push(OperandId, "'Num Events'"); + InstructionDesc[OpEnqueueMarker].operands.push(OperandId, "'Wait Events'"); + InstructionDesc[OpEnqueueMarker].operands.push(OperandId, "'Ret Event'"); + + InstructionDesc[OpGroupNonUniformElect].operands.push(OperandScope, "'Execution'"); + + InstructionDesc[OpGroupNonUniformAll].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformAll].operands.push(OperandId, "X"); + + InstructionDesc[OpGroupNonUniformAny].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformAny].operands.push(OperandId, "X"); + + InstructionDesc[OpGroupNonUniformAllEqual].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformAllEqual].operands.push(OperandId, "X"); + + InstructionDesc[OpGroupNonUniformBroadcast].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformBroadcast].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformBroadcast].operands.push(OperandId, "ID"); + + InstructionDesc[OpGroupNonUniformBroadcastFirst].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformBroadcastFirst].operands.push(OperandId, "X"); + + InstructionDesc[OpGroupNonUniformBallot].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformBallot].operands.push(OperandId, "X"); + + InstructionDesc[OpGroupNonUniformInverseBallot].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformInverseBallot].operands.push(OperandId, "X"); + + InstructionDesc[OpGroupNonUniformBallotBitExtract].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformBallotBitExtract].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformBallotBitExtract].operands.push(OperandId, "Bit"); + + InstructionDesc[OpGroupNonUniformBallotBitCount].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformBallotBitCount].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupNonUniformBallotBitCount].operands.push(OperandId, "X"); + + InstructionDesc[OpGroupNonUniformBallotFindLSB].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformBallotFindLSB].operands.push(OperandId, "X"); + + InstructionDesc[OpGroupNonUniformBallotFindMSB].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformBallotFindMSB].operands.push(OperandId, "X"); + + InstructionDesc[OpGroupNonUniformShuffle].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformShuffle].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformShuffle].operands.push(OperandId, "'Id'"); + + InstructionDesc[OpGroupNonUniformShuffleXor].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformShuffleXor].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformShuffleXor].operands.push(OperandId, "Mask"); + + InstructionDesc[OpGroupNonUniformShuffleUp].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformShuffleUp].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformShuffleUp].operands.push(OperandId, "Offset"); + + InstructionDesc[OpGroupNonUniformShuffleDown].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformShuffleDown].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformShuffleDown].operands.push(OperandId, "Offset"); + + InstructionDesc[OpGroupNonUniformIAdd].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformIAdd].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupNonUniformIAdd].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformIAdd].operands.push(OperandId, "'ClusterSize'", true); + + InstructionDesc[OpGroupNonUniformFAdd].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformFAdd].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupNonUniformFAdd].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformFAdd].operands.push(OperandId, "'ClusterSize'", true); + + InstructionDesc[OpGroupNonUniformIMul].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformIMul].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupNonUniformIMul].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformIMul].operands.push(OperandId, "'ClusterSize'", true); + + InstructionDesc[OpGroupNonUniformFMul].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformFMul].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupNonUniformFMul].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformFMul].operands.push(OperandId, "'ClusterSize'", true); + + InstructionDesc[OpGroupNonUniformSMin].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformSMin].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupNonUniformSMin].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformSMin].operands.push(OperandId, "'ClusterSize'", true); + + InstructionDesc[OpGroupNonUniformUMin].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformUMin].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupNonUniformUMin].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformUMin].operands.push(OperandId, "'ClusterSize'", true); + + InstructionDesc[OpGroupNonUniformFMin].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformFMin].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupNonUniformFMin].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformFMin].operands.push(OperandId, "'ClusterSize'", true); + + InstructionDesc[OpGroupNonUniformSMax].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformSMax].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupNonUniformSMax].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformSMax].operands.push(OperandId, "'ClusterSize'", true); + + InstructionDesc[OpGroupNonUniformUMax].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformUMax].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupNonUniformUMax].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformUMax].operands.push(OperandId, "'ClusterSize'", true); + + InstructionDesc[OpGroupNonUniformFMax].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformFMax].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupNonUniformFMax].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformFMax].operands.push(OperandId, "'ClusterSize'", true); + + InstructionDesc[OpGroupNonUniformBitwiseAnd].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformBitwiseAnd].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupNonUniformBitwiseAnd].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformBitwiseAnd].operands.push(OperandId, "'ClusterSize'", true); + + InstructionDesc[OpGroupNonUniformBitwiseOr].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformBitwiseOr].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupNonUniformBitwiseOr].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformBitwiseOr].operands.push(OperandId, "'ClusterSize'", true); + + InstructionDesc[OpGroupNonUniformBitwiseXor].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformBitwiseXor].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupNonUniformBitwiseXor].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformBitwiseXor].operands.push(OperandId, "'ClusterSize'", true); + + InstructionDesc[OpGroupNonUniformLogicalAnd].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformLogicalAnd].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupNonUniformLogicalAnd].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformLogicalAnd].operands.push(OperandId, "'ClusterSize'", true); + + InstructionDesc[OpGroupNonUniformLogicalOr].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformLogicalOr].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupNonUniformLogicalOr].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformLogicalOr].operands.push(OperandId, "'ClusterSize'", true); + + InstructionDesc[OpGroupNonUniformLogicalXor].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformLogicalXor].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupNonUniformLogicalXor].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformLogicalXor].operands.push(OperandId, "'ClusterSize'", true); + + InstructionDesc[OpGroupNonUniformQuadBroadcast].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformQuadBroadcast].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformQuadBroadcast].operands.push(OperandId, "'Id'"); + + InstructionDesc[OpGroupNonUniformQuadSwap].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupNonUniformQuadSwap].operands.push(OperandId, "X"); + InstructionDesc[OpGroupNonUniformQuadSwap].operands.push(OperandLiteralNumber, "'Direction'"); + + InstructionDesc[OpSubgroupBallotKHR].operands.push(OperandId, "'Predicate'"); + + InstructionDesc[OpSubgroupFirstInvocationKHR].operands.push(OperandId, "'Value'"); + + InstructionDesc[OpSubgroupAnyKHR].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpSubgroupAnyKHR].operands.push(OperandId, "'Predicate'"); + + InstructionDesc[OpSubgroupAllKHR].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpSubgroupAllKHR].operands.push(OperandId, "'Predicate'"); + + InstructionDesc[OpSubgroupAllEqualKHR].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpSubgroupAllEqualKHR].operands.push(OperandId, "'Predicate'"); + + InstructionDesc[OpSubgroupReadInvocationKHR].operands.push(OperandId, "'Value'"); + InstructionDesc[OpSubgroupReadInvocationKHR].operands.push(OperandId, "'Index'"); + + InstructionDesc[OpModuleProcessed].operands.push(OperandLiteralString, "'process'"); + +#ifdef AMD_EXTENSIONS + InstructionDesc[OpGroupIAddNonUniformAMD].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupIAddNonUniformAMD].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupIAddNonUniformAMD].operands.push(OperandId, "'X'"); + + InstructionDesc[OpGroupFAddNonUniformAMD].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupFAddNonUniformAMD].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupFAddNonUniformAMD].operands.push(OperandId, "'X'"); + + InstructionDesc[OpGroupUMinNonUniformAMD].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupUMinNonUniformAMD].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupUMinNonUniformAMD].operands.push(OperandId, "'X'"); + + InstructionDesc[OpGroupSMinNonUniformAMD].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupSMinNonUniformAMD].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupSMinNonUniformAMD].operands.push(OperandId, "X"); + + InstructionDesc[OpGroupFMinNonUniformAMD].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupFMinNonUniformAMD].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupFMinNonUniformAMD].operands.push(OperandId, "X"); + + InstructionDesc[OpGroupUMaxNonUniformAMD].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupUMaxNonUniformAMD].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupUMaxNonUniformAMD].operands.push(OperandId, "X"); + + InstructionDesc[OpGroupSMaxNonUniformAMD].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupSMaxNonUniformAMD].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupSMaxNonUniformAMD].operands.push(OperandId, "X"); + + InstructionDesc[OpGroupFMaxNonUniformAMD].operands.push(OperandScope, "'Execution'"); + InstructionDesc[OpGroupFMaxNonUniformAMD].operands.push(OperandGroupOperation, "'Operation'"); + InstructionDesc[OpGroupFMaxNonUniformAMD].operands.push(OperandId, "X"); + + InstructionDesc[OpFragmentMaskFetchAMD].operands.push(OperandId, "'Image'"); + InstructionDesc[OpFragmentMaskFetchAMD].operands.push(OperandId, "'Coordinate'"); + + InstructionDesc[OpFragmentFetchAMD].operands.push(OperandId, "'Image'"); + InstructionDesc[OpFragmentFetchAMD].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpFragmentFetchAMD].operands.push(OperandId, "'Fragment Index'"); +#endif + +#ifdef NV_EXTENSIONS + InstructionDesc[OpGroupNonUniformPartitionNV].operands.push(OperandId, "X"); + + InstructionDesc[OpTypeAccelerationStructureNV].setResultAndType(true, false); + + InstructionDesc[OpTraceNV].operands.push(OperandId, "'NV Acceleration Structure'"); + InstructionDesc[OpTraceNV].operands.push(OperandId, "'Ray Flags'"); + InstructionDesc[OpTraceNV].operands.push(OperandId, "'Cull Mask'"); + InstructionDesc[OpTraceNV].operands.push(OperandId, "'SBT Record Offset'"); + InstructionDesc[OpTraceNV].operands.push(OperandId, "'SBT Record Stride'"); + InstructionDesc[OpTraceNV].operands.push(OperandId, "'Miss Index'"); + InstructionDesc[OpTraceNV].operands.push(OperandId, "'Ray Origin'"); + InstructionDesc[OpTraceNV].operands.push(OperandId, "'TMin'"); + InstructionDesc[OpTraceNV].operands.push(OperandId, "'Ray Direction'"); + InstructionDesc[OpTraceNV].operands.push(OperandId, "'TMax'"); + InstructionDesc[OpTraceNV].operands.push(OperandId, "'Payload'"); + InstructionDesc[OpTraceNV].setResultAndType(false, false); + + InstructionDesc[OpReportIntersectionNV].operands.push(OperandId, "'Hit Parameter'"); + InstructionDesc[OpReportIntersectionNV].operands.push(OperandId, "'Hit Kind'"); + + InstructionDesc[OpIgnoreIntersectionNV].setResultAndType(false, false); + + InstructionDesc[OpTerminateRayNV].setResultAndType(false, false); + + InstructionDesc[OpExecuteCallableNV].operands.push(OperandId, "SBT Record Index"); + InstructionDesc[OpExecuteCallableNV].operands.push(OperandId, "CallableData ID"); + InstructionDesc[OpExecuteCallableNV].setResultAndType(false, false); + + InstructionDesc[OpImageSampleFootprintNV].operands.push(OperandId, "'Sampled Image'"); + InstructionDesc[OpImageSampleFootprintNV].operands.push(OperandId, "'Coordinate'"); + InstructionDesc[OpImageSampleFootprintNV].operands.push(OperandId, "'Granularity'"); + InstructionDesc[OpImageSampleFootprintNV].operands.push(OperandId, "'Coarse'"); + InstructionDesc[OpImageSampleFootprintNV].operands.push(OperandImageOperands, "", true); + InstructionDesc[OpImageSampleFootprintNV].operands.push(OperandVariableIds, "", true); + + InstructionDesc[OpWritePackedPrimitiveIndices4x8NV].operands.push(OperandId, "'Index Offset'"); + InstructionDesc[OpWritePackedPrimitiveIndices4x8NV].operands.push(OperandId, "'Packed Indices'"); +#endif + + InstructionDesc[OpTypeCooperativeMatrixNV].operands.push(OperandId, "'Component Type'"); + InstructionDesc[OpTypeCooperativeMatrixNV].operands.push(OperandId, "'Scope'"); + InstructionDesc[OpTypeCooperativeMatrixNV].operands.push(OperandId, "'Rows'"); + InstructionDesc[OpTypeCooperativeMatrixNV].operands.push(OperandId, "'Columns'"); + + InstructionDesc[OpCooperativeMatrixLoadNV].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpCooperativeMatrixLoadNV].operands.push(OperandId, "'Stride'"); + InstructionDesc[OpCooperativeMatrixLoadNV].operands.push(OperandId, "'Column Major'"); + InstructionDesc[OpCooperativeMatrixLoadNV].operands.push(OperandMemoryAccess, "'Memory Access'"); + InstructionDesc[OpCooperativeMatrixLoadNV].operands.push(OperandLiteralNumber, "", true); + InstructionDesc[OpCooperativeMatrixLoadNV].operands.push(OperandId, "", true); + + InstructionDesc[OpCooperativeMatrixStoreNV].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpCooperativeMatrixStoreNV].operands.push(OperandId, "'Object'"); + InstructionDesc[OpCooperativeMatrixStoreNV].operands.push(OperandId, "'Stride'"); + InstructionDesc[OpCooperativeMatrixStoreNV].operands.push(OperandId, "'Column Major'"); + InstructionDesc[OpCooperativeMatrixStoreNV].operands.push(OperandMemoryAccess, "'Memory Access'"); + InstructionDesc[OpCooperativeMatrixStoreNV].operands.push(OperandLiteralNumber, "", true); + InstructionDesc[OpCooperativeMatrixStoreNV].operands.push(OperandId, "", true); + + InstructionDesc[OpCooperativeMatrixMulAddNV].operands.push(OperandId, "'A'"); + InstructionDesc[OpCooperativeMatrixMulAddNV].operands.push(OperandId, "'B'"); + InstructionDesc[OpCooperativeMatrixMulAddNV].operands.push(OperandId, "'C'"); + + InstructionDesc[OpCooperativeMatrixLengthNV].operands.push(OperandId, "'Type'"); +} + +}; // end spv namespace diff --git a/thirdparty/glslang/SPIRV/doc.h b/thirdparty/glslang/SPIRV/doc.h new file mode 100644 index 0000000000..293256a2c6 --- /dev/null +++ b/thirdparty/glslang/SPIRV/doc.h @@ -0,0 +1,258 @@ +// +// Copyright (C) 2014-2015 LunarG, Inc. +// +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions +// are met: +// +// Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// +// Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following +// disclaimer in the documentation and/or other materials provided +// with the distribution. +// +// Neither the name of 3Dlabs Inc. Ltd. nor the names of its +// contributors may be used to endorse or promote products derived +// from this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS +// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE +// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, +// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, +// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; +// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN +// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +// POSSIBILITY OF SUCH DAMAGE. + +// +// Parameterize the SPIR-V enumerants. +// + +#pragma once + +#include "spirv.hpp" + +#include <vector> + +namespace spv { + +// Fill in all the parameters +void Parameterize(); + +// Return the English names of all the enums. +const char* SourceString(int); +const char* AddressingString(int); +const char* MemoryString(int); +const char* ExecutionModelString(int); +const char* ExecutionModeString(int); +const char* StorageClassString(int); +const char* DecorationString(int); +const char* BuiltInString(int); +const char* DimensionString(int); +const char* SelectControlString(int); +const char* LoopControlString(int); +const char* FunctionControlString(int); +const char* SamplerAddressingModeString(int); +const char* SamplerFilterModeString(int); +const char* ImageFormatString(int); +const char* ImageChannelOrderString(int); +const char* ImageChannelTypeString(int); +const char* ImageChannelDataTypeString(int type); +const char* ImageOperandsString(int format); +const char* ImageOperands(int); +const char* FPFastMathString(int); +const char* FPRoundingModeString(int); +const char* LinkageTypeString(int); +const char* FuncParamAttrString(int); +const char* AccessQualifierString(int); +const char* MemorySemanticsString(int); +const char* MemoryAccessString(int); +const char* ExecutionScopeString(int); +const char* GroupOperationString(int); +const char* KernelEnqueueFlagsString(int); +const char* KernelProfilingInfoString(int); +const char* CapabilityString(int); +const char* OpcodeString(int); +const char* ScopeString(int mem); + +// For grouping opcodes into subsections +enum OpcodeClass { + OpClassMisc, + OpClassDebug, + OpClassAnnotate, + OpClassExtension, + OpClassMode, + OpClassType, + OpClassConstant, + OpClassMemory, + OpClassFunction, + OpClassImage, + OpClassConvert, + OpClassComposite, + OpClassArithmetic, + OpClassBit, + OpClassRelationalLogical, + OpClassDerivative, + OpClassFlowControl, + OpClassAtomic, + OpClassPrimitive, + OpClassBarrier, + OpClassGroup, + OpClassDeviceSideEnqueue, + OpClassPipe, + + OpClassCount, + OpClassMissing // all instructions start out as missing +}; + +// For parameterizing operands. +enum OperandClass { + OperandNone, + OperandId, + OperandVariableIds, + OperandOptionalLiteral, + OperandOptionalLiteralString, + OperandVariableLiterals, + OperandVariableIdLiteral, + OperandVariableLiteralId, + OperandLiteralNumber, + OperandLiteralString, + OperandSource, + OperandExecutionModel, + OperandAddressing, + OperandMemory, + OperandExecutionMode, + OperandStorage, + OperandDimensionality, + OperandSamplerAddressingMode, + OperandSamplerFilterMode, + OperandSamplerImageFormat, + OperandImageChannelOrder, + OperandImageChannelDataType, + OperandImageOperands, + OperandFPFastMath, + OperandFPRoundingMode, + OperandLinkageType, + OperandAccessQualifier, + OperandFuncParamAttr, + OperandDecoration, + OperandBuiltIn, + OperandSelect, + OperandLoop, + OperandFunction, + OperandMemorySemantics, + OperandMemoryAccess, + OperandScope, + OperandGroupOperation, + OperandKernelEnqueueFlags, + OperandKernelProfilingInfo, + OperandCapability, + + OperandOpcode, + + OperandCount +}; + +// Any specific enum can have a set of capabilities that allow it: +typedef std::vector<Capability> EnumCaps; + +// Parameterize a set of operands with their OperandClass(es) and descriptions. +class OperandParameters { +public: + OperandParameters() { } + void push(OperandClass oc, const char* d, bool opt = false) + { + opClass.push_back(oc); + desc.push_back(d); + optional.push_back(opt); + } + void setOptional(); + OperandClass getClass(int op) const { return opClass[op]; } + const char* getDesc(int op) const { return desc[op]; } + bool isOptional(int op) const { return optional[op]; } + int getNum() const { return (int)opClass.size(); } + +protected: + std::vector<OperandClass> opClass; + std::vector<const char*> desc; + std::vector<bool> optional; +}; + +// Parameterize an enumerant +class EnumParameters { +public: + EnumParameters() : desc(0) { } + const char* desc; +}; + +// Parameterize a set of enumerants that form an enum +class EnumDefinition : public EnumParameters { +public: + EnumDefinition() : + ceiling(0), bitmask(false), getName(0), enumParams(0), operandParams(0) { } + void set(int ceil, const char* (*name)(int), EnumParameters* ep, bool mask = false) + { + ceiling = ceil; + getName = name; + bitmask = mask; + enumParams = ep; + } + void setOperands(OperandParameters* op) { operandParams = op; } + int ceiling; // ceiling of enumerants + bool bitmask; // true if these enumerants combine into a bitmask + const char* (*getName)(int); // a function that returns the name for each enumerant value (or shift) + EnumParameters* enumParams; // parameters for each individual enumerant + OperandParameters* operandParams; // sets of operands +}; + +// Parameterize an instruction's logical format, including its known set of operands, +// per OperandParameters above. +class InstructionParameters { +public: + InstructionParameters() : + opDesc("TBD"), + opClass(OpClassMissing), + typePresent(true), // most normal, only exceptions have to be spelled out + resultPresent(true) // most normal, only exceptions have to be spelled out + { } + + void setResultAndType(bool r, bool t) + { + resultPresent = r; + typePresent = t; + } + + bool hasResult() const { return resultPresent != 0; } + bool hasType() const { return typePresent != 0; } + + const char* opDesc; + OpcodeClass opClass; + OperandParameters operands; + +protected: + int typePresent : 1; + int resultPresent : 1; +}; + +// The set of objects that hold all the instruction/operand +// parameterization information. +extern InstructionParameters InstructionDesc[]; + +// These hold definitions of the enumerants used for operands +extern EnumDefinition OperandClassParams[]; + +const char* GetOperandDesc(OperandClass operand); +void PrintImmediateRow(int imm, const char* name, const EnumParameters* enumParams, bool caps, bool hex = false); +const char* AccessQualifierString(int attr); + +void PrintOperands(const OperandParameters& operands, int reservedOperands); + +} // end namespace spv diff --git a/thirdparty/glslang/SPIRV/hex_float.h b/thirdparty/glslang/SPIRV/hex_float.h new file mode 100644 index 0000000000..905b21a45a --- /dev/null +++ b/thirdparty/glslang/SPIRV/hex_float.h @@ -0,0 +1,1078 @@ +// Copyright (c) 2015-2016 The Khronos Group Inc. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#ifndef LIBSPIRV_UTIL_HEX_FLOAT_H_ +#define LIBSPIRV_UTIL_HEX_FLOAT_H_ + +#include <cassert> +#include <cctype> +#include <cmath> +#include <cstdint> +#include <iomanip> +#include <limits> +#include <sstream> + +#if defined(_MSC_VER) && _MSC_VER < 1800 +namespace std { +bool isnan(double f) +{ + return ::_isnan(f) != 0; +} +bool isinf(double f) +{ + return ::_finite(f) == 0; +} +} +#endif + +#include "bitutils.h" + +namespace spvutils { + +class Float16 { + public: + Float16(uint16_t v) : val(v) {} + Float16() {} + static bool isNan(const Float16& val) { + return ((val.val & 0x7C00) == 0x7C00) && ((val.val & 0x3FF) != 0); + } + // Returns true if the given value is any kind of infinity. + static bool isInfinity(const Float16& val) { + return ((val.val & 0x7C00) == 0x7C00) && ((val.val & 0x3FF) == 0); + } + Float16(const Float16& other) { val = other.val; } + uint16_t get_value() const { return val; } + + // Returns the maximum normal value. + static Float16 max() { return Float16(0x7bff); } + // Returns the lowest normal value. + static Float16 lowest() { return Float16(0xfbff); } + + private: + uint16_t val; +}; + +// To specialize this type, you must override uint_type to define +// an unsigned integer that can fit your floating point type. +// You must also add a isNan function that returns true if +// a value is Nan. +template <typename T> +struct FloatProxyTraits { + typedef void uint_type; +}; + +template <> +struct FloatProxyTraits<float> { + typedef uint32_t uint_type; + static bool isNan(float f) { return std::isnan(f); } + // Returns true if the given value is any kind of infinity. + static bool isInfinity(float f) { return std::isinf(f); } + // Returns the maximum normal value. + static float max() { return std::numeric_limits<float>::max(); } + // Returns the lowest normal value. + static float lowest() { return std::numeric_limits<float>::lowest(); } +}; + +template <> +struct FloatProxyTraits<double> { + typedef uint64_t uint_type; + static bool isNan(double f) { return std::isnan(f); } + // Returns true if the given value is any kind of infinity. + static bool isInfinity(double f) { return std::isinf(f); } + // Returns the maximum normal value. + static double max() { return std::numeric_limits<double>::max(); } + // Returns the lowest normal value. + static double lowest() { return std::numeric_limits<double>::lowest(); } +}; + +template <> +struct FloatProxyTraits<Float16> { + typedef uint16_t uint_type; + static bool isNan(Float16 f) { return Float16::isNan(f); } + // Returns true if the given value is any kind of infinity. + static bool isInfinity(Float16 f) { return Float16::isInfinity(f); } + // Returns the maximum normal value. + static Float16 max() { return Float16::max(); } + // Returns the lowest normal value. + static Float16 lowest() { return Float16::lowest(); } +}; + +// Since copying a floating point number (especially if it is NaN) +// does not guarantee that bits are preserved, this class lets us +// store the type and use it as a float when necessary. +template <typename T> +class FloatProxy { + public: + typedef typename FloatProxyTraits<T>::uint_type uint_type; + + // Since this is to act similar to the normal floats, + // do not initialize the data by default. + FloatProxy() {} + + // Intentionally non-explicit. This is a proxy type so + // implicit conversions allow us to use it more transparently. + FloatProxy(T val) { data_ = BitwiseCast<uint_type>(val); } + + // Intentionally non-explicit. This is a proxy type so + // implicit conversions allow us to use it more transparently. + FloatProxy(uint_type val) { data_ = val; } + + // This is helpful to have and is guaranteed not to stomp bits. + FloatProxy<T> operator-() const { + return static_cast<uint_type>(data_ ^ + (uint_type(0x1) << (sizeof(T) * 8 - 1))); + } + + // Returns the data as a floating point value. + T getAsFloat() const { return BitwiseCast<T>(data_); } + + // Returns the raw data. + uint_type data() const { return data_; } + + // Returns true if the value represents any type of NaN. + bool isNan() { return FloatProxyTraits<T>::isNan(getAsFloat()); } + // Returns true if the value represents any type of infinity. + bool isInfinity() { return FloatProxyTraits<T>::isInfinity(getAsFloat()); } + + // Returns the maximum normal value. + static FloatProxy<T> max() { + return FloatProxy<T>(FloatProxyTraits<T>::max()); + } + // Returns the lowest normal value. + static FloatProxy<T> lowest() { + return FloatProxy<T>(FloatProxyTraits<T>::lowest()); + } + + private: + uint_type data_; +}; + +template <typename T> +bool operator==(const FloatProxy<T>& first, const FloatProxy<T>& second) { + return first.data() == second.data(); +} + +// Reads a FloatProxy value as a normal float from a stream. +template <typename T> +std::istream& operator>>(std::istream& is, FloatProxy<T>& value) { + T float_val; + is >> float_val; + value = FloatProxy<T>(float_val); + return is; +} + +// This is an example traits. It is not meant to be used in practice, but will +// be the default for any non-specialized type. +template <typename T> +struct HexFloatTraits { + // Integer type that can store this hex-float. + typedef void uint_type; + // Signed integer type that can store this hex-float. + typedef void int_type; + // The numerical type that this HexFloat represents. + typedef void underlying_type; + // The type needed to construct the underlying type. + typedef void native_type; + // The number of bits that are actually relevant in the uint_type. + // This allows us to deal with, for example, 24-bit values in a 32-bit + // integer. + static const uint32_t num_used_bits = 0; + // Number of bits that represent the exponent. + static const uint32_t num_exponent_bits = 0; + // Number of bits that represent the fractional part. + static const uint32_t num_fraction_bits = 0; + // The bias of the exponent. (How much we need to subtract from the stored + // value to get the correct value.) + static const uint32_t exponent_bias = 0; +}; + +// Traits for IEEE float. +// 1 sign bit, 8 exponent bits, 23 fractional bits. +template <> +struct HexFloatTraits<FloatProxy<float>> { + typedef uint32_t uint_type; + typedef int32_t int_type; + typedef FloatProxy<float> underlying_type; + typedef float native_type; + static const uint_type num_used_bits = 32; + static const uint_type num_exponent_bits = 8; + static const uint_type num_fraction_bits = 23; + static const uint_type exponent_bias = 127; +}; + +// Traits for IEEE double. +// 1 sign bit, 11 exponent bits, 52 fractional bits. +template <> +struct HexFloatTraits<FloatProxy<double>> { + typedef uint64_t uint_type; + typedef int64_t int_type; + typedef FloatProxy<double> underlying_type; + typedef double native_type; + static const uint_type num_used_bits = 64; + static const uint_type num_exponent_bits = 11; + static const uint_type num_fraction_bits = 52; + static const uint_type exponent_bias = 1023; +}; + +// Traits for IEEE half. +// 1 sign bit, 5 exponent bits, 10 fractional bits. +template <> +struct HexFloatTraits<FloatProxy<Float16>> { + typedef uint16_t uint_type; + typedef int16_t int_type; + typedef uint16_t underlying_type; + typedef uint16_t native_type; + static const uint_type num_used_bits = 16; + static const uint_type num_exponent_bits = 5; + static const uint_type num_fraction_bits = 10; + static const uint_type exponent_bias = 15; +}; + +enum round_direction { + kRoundToZero, + kRoundToNearestEven, + kRoundToPositiveInfinity, + kRoundToNegativeInfinity +}; + +// Template class that houses a floating pointer number. +// It exposes a number of constants based on the provided traits to +// assist in interpreting the bits of the value. +template <typename T, typename Traits = HexFloatTraits<T>> +class HexFloat { + public: + typedef typename Traits::uint_type uint_type; + typedef typename Traits::int_type int_type; + typedef typename Traits::underlying_type underlying_type; + typedef typename Traits::native_type native_type; + + explicit HexFloat(T f) : value_(f) {} + + T value() const { return value_; } + void set_value(T f) { value_ = f; } + + // These are all written like this because it is convenient to have + // compile-time constants for all of these values. + + // Pass-through values to save typing. + static const uint32_t num_used_bits = Traits::num_used_bits; + static const uint32_t exponent_bias = Traits::exponent_bias; + static const uint32_t num_exponent_bits = Traits::num_exponent_bits; + static const uint32_t num_fraction_bits = Traits::num_fraction_bits; + + // Number of bits to shift left to set the highest relevant bit. + static const uint32_t top_bit_left_shift = num_used_bits - 1; + // How many nibbles (hex characters) the fractional part takes up. + static const uint32_t fraction_nibbles = (num_fraction_bits + 3) / 4; + // If the fractional part does not fit evenly into a hex character (4-bits) + // then we have to left-shift to get rid of leading 0s. This is the amount + // we have to shift (might be 0). + static const uint32_t num_overflow_bits = + fraction_nibbles * 4 - num_fraction_bits; + + // The representation of the fraction, not the actual bits. This + // includes the leading bit that is usually implicit. + static const uint_type fraction_represent_mask = + spvutils::SetBits<uint_type, 0, + num_fraction_bits + num_overflow_bits>::get; + + // The topmost bit in the nibble-aligned fraction. + static const uint_type fraction_top_bit = + uint_type(1) << (num_fraction_bits + num_overflow_bits - 1); + + // The least significant bit in the exponent, which is also the bit + // immediately to the left of the significand. + static const uint_type first_exponent_bit = uint_type(1) + << (num_fraction_bits); + + // The mask for the encoded fraction. It does not include the + // implicit bit. + static const uint_type fraction_encode_mask = + spvutils::SetBits<uint_type, 0, num_fraction_bits>::get; + + // The bit that is used as a sign. + static const uint_type sign_mask = uint_type(1) << top_bit_left_shift; + + // The bits that represent the exponent. + static const uint_type exponent_mask = + spvutils::SetBits<uint_type, num_fraction_bits, num_exponent_bits>::get; + + // How far left the exponent is shifted. + static const uint32_t exponent_left_shift = num_fraction_bits; + + // How far from the right edge the fraction is shifted. + static const uint32_t fraction_right_shift = + static_cast<uint32_t>(sizeof(uint_type) * 8) - num_fraction_bits; + + // The maximum representable unbiased exponent. + static const int_type max_exponent = + (exponent_mask >> num_fraction_bits) - exponent_bias; + // The minimum representable exponent for normalized numbers. + static const int_type min_exponent = -static_cast<int_type>(exponent_bias); + + // Returns the bits associated with the value. + uint_type getBits() const { return spvutils::BitwiseCast<uint_type>(value_); } + + // Returns the bits associated with the value, without the leading sign bit. + uint_type getUnsignedBits() const { + return static_cast<uint_type>(spvutils::BitwiseCast<uint_type>(value_) & + ~sign_mask); + } + + // Returns the bits associated with the exponent, shifted to start at the + // lsb of the type. + const uint_type getExponentBits() const { + return static_cast<uint_type>((getBits() & exponent_mask) >> + num_fraction_bits); + } + + // Returns the exponent in unbiased form. This is the exponent in the + // human-friendly form. + const int_type getUnbiasedExponent() const { + return static_cast<int_type>(getExponentBits() - exponent_bias); + } + + // Returns just the significand bits from the value. + const uint_type getSignificandBits() const { + return getBits() & fraction_encode_mask; + } + + // If the number was normalized, returns the unbiased exponent. + // If the number was denormal, normalize the exponent first. + const int_type getUnbiasedNormalizedExponent() const { + if ((getBits() & ~sign_mask) == 0) { // special case if everything is 0 + return 0; + } + int_type exp = getUnbiasedExponent(); + if (exp == min_exponent) { // We are in denorm land. + uint_type significand_bits = getSignificandBits(); + while ((significand_bits & (first_exponent_bit >> 1)) == 0) { + significand_bits = static_cast<uint_type>(significand_bits << 1); + exp = static_cast<int_type>(exp - 1); + } + significand_bits &= fraction_encode_mask; + } + return exp; + } + + // Returns the signficand after it has been normalized. + const uint_type getNormalizedSignificand() const { + int_type unbiased_exponent = getUnbiasedNormalizedExponent(); + uint_type significand = getSignificandBits(); + for (int_type i = unbiased_exponent; i <= min_exponent; ++i) { + significand = static_cast<uint_type>(significand << 1); + } + significand &= fraction_encode_mask; + return significand; + } + + // Returns true if this number represents a negative value. + bool isNegative() const { return (getBits() & sign_mask) != 0; } + + // Sets this HexFloat from the individual components. + // Note this assumes EVERY significand is normalized, and has an implicit + // leading one. This means that the only way that this method will set 0, + // is if you set a number so denormalized that it underflows. + // Do not use this method with raw bits extracted from a subnormal number, + // since subnormals do not have an implicit leading 1 in the significand. + // The significand is also expected to be in the + // lowest-most num_fraction_bits of the uint_type. + // The exponent is expected to be unbiased, meaning an exponent of + // 0 actually means 0. + // If underflow_round_up is set, then on underflow, if a number is non-0 + // and would underflow, we round up to the smallest denorm. + void setFromSignUnbiasedExponentAndNormalizedSignificand( + bool negative, int_type exponent, uint_type significand, + bool round_denorm_up) { + bool significand_is_zero = significand == 0; + + if (exponent <= min_exponent) { + // If this was denormalized, then we have to shift the bit on, meaning + // the significand is not zero. + significand_is_zero = false; + significand |= first_exponent_bit; + significand = static_cast<uint_type>(significand >> 1); + } + + while (exponent < min_exponent) { + significand = static_cast<uint_type>(significand >> 1); + ++exponent; + } + + if (exponent == min_exponent) { + if (significand == 0 && !significand_is_zero && round_denorm_up) { + significand = static_cast<uint_type>(0x1); + } + } + + uint_type new_value = 0; + if (negative) { + new_value = static_cast<uint_type>(new_value | sign_mask); + } + exponent = static_cast<int_type>(exponent + exponent_bias); + assert(exponent >= 0); + + // put it all together + exponent = static_cast<uint_type>((exponent << exponent_left_shift) & + exponent_mask); + significand = static_cast<uint_type>(significand & fraction_encode_mask); + new_value = static_cast<uint_type>(new_value | (exponent | significand)); + value_ = BitwiseCast<T>(new_value); + } + + // Increments the significand of this number by the given amount. + // If this would spill the significand into the implicit bit, + // carry is set to true and the significand is shifted to fit into + // the correct location, otherwise carry is set to false. + // All significands and to_increment are assumed to be within the bounds + // for a valid significand. + static uint_type incrementSignificand(uint_type significand, + uint_type to_increment, bool* carry) { + significand = static_cast<uint_type>(significand + to_increment); + *carry = false; + if (significand & first_exponent_bit) { + *carry = true; + // The implicit 1-bit will have carried, so we should zero-out the + // top bit and shift back. + significand = static_cast<uint_type>(significand & ~first_exponent_bit); + significand = static_cast<uint_type>(significand >> 1); + } + return significand; + } + + // These exist because MSVC throws warnings on negative right-shifts + // even if they are not going to be executed. Eg: + // constant_number < 0? 0: constant_number + // These convert the negative left-shifts into right shifts. + + template <typename int_type> + uint_type negatable_left_shift(int_type N, uint_type val) + { + if(N >= 0) + return val << N; + + return val >> -N; + } + + template <typename int_type> + uint_type negatable_right_shift(int_type N, uint_type val) + { + if(N >= 0) + return val >> N; + + return val << -N; + } + + // Returns the significand, rounded to fit in a significand in + // other_T. This is shifted so that the most significant + // bit of the rounded number lines up with the most significant bit + // of the returned significand. + template <typename other_T> + typename other_T::uint_type getRoundedNormalizedSignificand( + round_direction dir, bool* carry_bit) { + typedef typename other_T::uint_type other_uint_type; + static const int_type num_throwaway_bits = + static_cast<int_type>(num_fraction_bits) - + static_cast<int_type>(other_T::num_fraction_bits); + + static const uint_type last_significant_bit = + (num_throwaway_bits < 0) + ? 0 + : negatable_left_shift(num_throwaway_bits, 1u); + static const uint_type first_rounded_bit = + (num_throwaway_bits < 1) + ? 0 + : negatable_left_shift(num_throwaway_bits - 1, 1u); + + static const uint_type throwaway_mask_bits = + num_throwaway_bits > 0 ? num_throwaway_bits : 0; + static const uint_type throwaway_mask = + spvutils::SetBits<uint_type, 0, throwaway_mask_bits>::get; + + *carry_bit = false; + other_uint_type out_val = 0; + uint_type significand = getNormalizedSignificand(); + // If we are up-casting, then we just have to shift to the right location. + if (num_throwaway_bits <= 0) { + out_val = static_cast<other_uint_type>(significand); + uint_type shift_amount = static_cast<uint_type>(-num_throwaway_bits); + out_val = static_cast<other_uint_type>(out_val << shift_amount); + return out_val; + } + + // If every non-representable bit is 0, then we don't have any casting to + // do. + if ((significand & throwaway_mask) == 0) { + return static_cast<other_uint_type>( + negatable_right_shift(num_throwaway_bits, significand)); + } + + bool round_away_from_zero = false; + // We actually have to narrow the significand here, so we have to follow the + // rounding rules. + switch (dir) { + case kRoundToZero: + break; + case kRoundToPositiveInfinity: + round_away_from_zero = !isNegative(); + break; + case kRoundToNegativeInfinity: + round_away_from_zero = isNegative(); + break; + case kRoundToNearestEven: + // Have to round down, round bit is 0 + if ((first_rounded_bit & significand) == 0) { + break; + } + if (((significand & throwaway_mask) & ~first_rounded_bit) != 0) { + // If any subsequent bit of the rounded portion is non-0 then we round + // up. + round_away_from_zero = true; + break; + } + // We are exactly half-way between 2 numbers, pick even. + if ((significand & last_significant_bit) != 0) { + // 1 for our last bit, round up. + round_away_from_zero = true; + break; + } + break; + } + + if (round_away_from_zero) { + return static_cast<other_uint_type>( + negatable_right_shift(num_throwaway_bits, incrementSignificand( + significand, last_significant_bit, carry_bit))); + } else { + return static_cast<other_uint_type>( + negatable_right_shift(num_throwaway_bits, significand)); + } + } + + // Casts this value to another HexFloat. If the cast is widening, + // then round_dir is ignored. If the cast is narrowing, then + // the result is rounded in the direction specified. + // This number will retain Nan and Inf values. + // It will also saturate to Inf if the number overflows, and + // underflow to (0 or min depending on rounding) if the number underflows. + template <typename other_T> + void castTo(other_T& other, round_direction round_dir) { + other = other_T(static_cast<typename other_T::native_type>(0)); + bool negate = isNegative(); + if (getUnsignedBits() == 0) { + if (negate) { + other.set_value(-other.value()); + } + return; + } + uint_type significand = getSignificandBits(); + bool carried = false; + typename other_T::uint_type rounded_significand = + getRoundedNormalizedSignificand<other_T>(round_dir, &carried); + + int_type exponent = getUnbiasedExponent(); + if (exponent == min_exponent) { + // If we are denormal, normalize the exponent, so that we can encode + // easily. + exponent = static_cast<int_type>(exponent + 1); + for (uint_type check_bit = first_exponent_bit >> 1; check_bit != 0; + check_bit = static_cast<uint_type>(check_bit >> 1)) { + exponent = static_cast<int_type>(exponent - 1); + if (check_bit & significand) break; + } + } + + bool is_nan = + (getBits() & exponent_mask) == exponent_mask && significand != 0; + bool is_inf = + !is_nan && + ((exponent + carried) > static_cast<int_type>(other_T::exponent_bias) || + (significand == 0 && (getBits() & exponent_mask) == exponent_mask)); + + // If we are Nan or Inf we should pass that through. + if (is_inf) { + other.set_value(BitwiseCast<typename other_T::underlying_type>( + static_cast<typename other_T::uint_type>( + (negate ? other_T::sign_mask : 0) | other_T::exponent_mask))); + return; + } + if (is_nan) { + typename other_T::uint_type shifted_significand; + shifted_significand = static_cast<typename other_T::uint_type>( + negatable_left_shift( + static_cast<int_type>(other_T::num_fraction_bits) - + static_cast<int_type>(num_fraction_bits), significand)); + + // We are some sort of Nan. We try to keep the bit-pattern of the Nan + // as close as possible. If we had to shift off bits so we are 0, then we + // just set the last bit. + other.set_value(BitwiseCast<typename other_T::underlying_type>( + static_cast<typename other_T::uint_type>( + (negate ? other_T::sign_mask : 0) | other_T::exponent_mask | + (shifted_significand == 0 ? 0x1 : shifted_significand)))); + return; + } + + bool round_underflow_up = + isNegative() ? round_dir == kRoundToNegativeInfinity + : round_dir == kRoundToPositiveInfinity; + typedef typename other_T::int_type other_int_type; + // setFromSignUnbiasedExponentAndNormalizedSignificand will + // zero out any underflowing value (but retain the sign). + other.setFromSignUnbiasedExponentAndNormalizedSignificand( + negate, static_cast<other_int_type>(exponent), rounded_significand, + round_underflow_up); + return; + } + + private: + T value_; + + static_assert(num_used_bits == + Traits::num_exponent_bits + Traits::num_fraction_bits + 1, + "The number of bits do not fit"); + static_assert(sizeof(T) == sizeof(uint_type), "The type sizes do not match"); +}; + +// Returns 4 bits represented by the hex character. +inline uint8_t get_nibble_from_character(int character) { + const char* dec = "0123456789"; + const char* lower = "abcdef"; + const char* upper = "ABCDEF"; + const char* p = nullptr; + if ((p = strchr(dec, character))) { + return static_cast<uint8_t>(p - dec); + } else if ((p = strchr(lower, character))) { + return static_cast<uint8_t>(p - lower + 0xa); + } else if ((p = strchr(upper, character))) { + return static_cast<uint8_t>(p - upper + 0xa); + } + + assert(false && "This was called with a non-hex character"); + return 0; +} + +// Outputs the given HexFloat to the stream. +template <typename T, typename Traits> +std::ostream& operator<<(std::ostream& os, const HexFloat<T, Traits>& value) { + typedef HexFloat<T, Traits> HF; + typedef typename HF::uint_type uint_type; + typedef typename HF::int_type int_type; + + static_assert(HF::num_used_bits != 0, + "num_used_bits must be non-zero for a valid float"); + static_assert(HF::num_exponent_bits != 0, + "num_exponent_bits must be non-zero for a valid float"); + static_assert(HF::num_fraction_bits != 0, + "num_fractin_bits must be non-zero for a valid float"); + + const uint_type bits = spvutils::BitwiseCast<uint_type>(value.value()); + const char* const sign = (bits & HF::sign_mask) ? "-" : ""; + const uint_type exponent = static_cast<uint_type>( + (bits & HF::exponent_mask) >> HF::num_fraction_bits); + + uint_type fraction = static_cast<uint_type>((bits & HF::fraction_encode_mask) + << HF::num_overflow_bits); + + const bool is_zero = exponent == 0 && fraction == 0; + const bool is_denorm = exponent == 0 && !is_zero; + + // exponent contains the biased exponent we have to convert it back into + // the normal range. + int_type int_exponent = static_cast<int_type>(exponent - HF::exponent_bias); + // If the number is all zeros, then we actually have to NOT shift the + // exponent. + int_exponent = is_zero ? 0 : int_exponent; + + // If we are denorm, then start shifting, and decreasing the exponent until + // our leading bit is 1. + + if (is_denorm) { + while ((fraction & HF::fraction_top_bit) == 0) { + fraction = static_cast<uint_type>(fraction << 1); + int_exponent = static_cast<int_type>(int_exponent - 1); + } + // Since this is denormalized, we have to consume the leading 1 since it + // will end up being implicit. + fraction = static_cast<uint_type>(fraction << 1); // eat the leading 1 + fraction &= HF::fraction_represent_mask; + } + + uint_type fraction_nibbles = HF::fraction_nibbles; + // We do not have to display any trailing 0s, since this represents the + // fractional part. + while (fraction_nibbles > 0 && (fraction & 0xF) == 0) { + // Shift off any trailing values; + fraction = static_cast<uint_type>(fraction >> 4); + --fraction_nibbles; + } + + const auto saved_flags = os.flags(); + const auto saved_fill = os.fill(); + + os << sign << "0x" << (is_zero ? '0' : '1'); + if (fraction_nibbles) { + // Make sure to keep the leading 0s in place, since this is the fractional + // part. + os << "." << std::setw(static_cast<int>(fraction_nibbles)) + << std::setfill('0') << std::hex << fraction; + } + os << "p" << std::dec << (int_exponent >= 0 ? "+" : "") << int_exponent; + + os.flags(saved_flags); + os.fill(saved_fill); + + return os; +} + +// Returns true if negate_value is true and the next character on the +// input stream is a plus or minus sign. In that case we also set the fail bit +// on the stream and set the value to the zero value for its type. +template <typename T, typename Traits> +inline bool RejectParseDueToLeadingSign(std::istream& is, bool negate_value, + HexFloat<T, Traits>& value) { + if (negate_value) { + auto next_char = is.peek(); + if (next_char == '-' || next_char == '+') { + // Fail the parse. Emulate standard behaviour by setting the value to + // the zero value, and set the fail bit on the stream. + value = HexFloat<T, Traits>(typename HexFloat<T, Traits>::uint_type(0)); + is.setstate(std::ios_base::failbit); + return true; + } + } + return false; +} + +// Parses a floating point number from the given stream and stores it into the +// value parameter. +// If negate_value is true then the number may not have a leading minus or +// plus, and if it successfully parses, then the number is negated before +// being stored into the value parameter. +// If the value cannot be correctly parsed or overflows the target floating +// point type, then set the fail bit on the stream. +// TODO(dneto): Promise C++11 standard behavior in how the value is set in +// the error case, but only after all target platforms implement it correctly. +// In particular, the Microsoft C++ runtime appears to be out of spec. +template <typename T, typename Traits> +inline std::istream& ParseNormalFloat(std::istream& is, bool negate_value, + HexFloat<T, Traits>& value) { + if (RejectParseDueToLeadingSign(is, negate_value, value)) { + return is; + } + T val; + is >> val; + if (negate_value) { + val = -val; + } + value.set_value(val); + // In the failure case, map -0.0 to 0.0. + if (is.fail() && value.getUnsignedBits() == 0u) { + value = HexFloat<T, Traits>(typename HexFloat<T, Traits>::uint_type(0)); + } + if (val.isInfinity()) { + // Fail the parse. Emulate standard behaviour by setting the value to + // the closest normal value, and set the fail bit on the stream. + value.set_value((value.isNegative() | negate_value) ? T::lowest() + : T::max()); + is.setstate(std::ios_base::failbit); + } + return is; +} + +// Specialization of ParseNormalFloat for FloatProxy<Float16> values. +// This will parse the float as it were a 32-bit floating point number, +// and then round it down to fit into a Float16 value. +// The number is rounded towards zero. +// If negate_value is true then the number may not have a leading minus or +// plus, and if it successfully parses, then the number is negated before +// being stored into the value parameter. +// If the value cannot be correctly parsed or overflows the target floating +// point type, then set the fail bit on the stream. +// TODO(dneto): Promise C++11 standard behavior in how the value is set in +// the error case, but only after all target platforms implement it correctly. +// In particular, the Microsoft C++ runtime appears to be out of spec. +template <> +inline std::istream& +ParseNormalFloat<FloatProxy<Float16>, HexFloatTraits<FloatProxy<Float16>>>( + std::istream& is, bool negate_value, + HexFloat<FloatProxy<Float16>, HexFloatTraits<FloatProxy<Float16>>>& value) { + // First parse as a 32-bit float. + HexFloat<FloatProxy<float>> float_val(0.0f); + ParseNormalFloat(is, negate_value, float_val); + + // Then convert to 16-bit float, saturating at infinities, and + // rounding toward zero. + float_val.castTo(value, kRoundToZero); + + // Overflow on 16-bit behaves the same as for 32- and 64-bit: set the + // fail bit and set the lowest or highest value. + if (Float16::isInfinity(value.value().getAsFloat())) { + value.set_value(value.isNegative() ? Float16::lowest() : Float16::max()); + is.setstate(std::ios_base::failbit); + } + return is; +} + +// Reads a HexFloat from the given stream. +// If the float is not encoded as a hex-float then it will be parsed +// as a regular float. +// This may fail if your stream does not support at least one unget. +// Nan values can be encoded with "0x1.<not zero>p+exponent_bias". +// This would normally overflow a float and round to +// infinity but this special pattern is the exact representation for a NaN, +// and therefore is actually encoded as the correct NaN. To encode inf, +// either 0x0p+exponent_bias can be specified or any exponent greater than +// exponent_bias. +// Examples using IEEE 32-bit float encoding. +// 0x1.0p+128 (+inf) +// -0x1.0p-128 (-inf) +// +// 0x1.1p+128 (+Nan) +// -0x1.1p+128 (-Nan) +// +// 0x1p+129 (+inf) +// -0x1p+129 (-inf) +template <typename T, typename Traits> +std::istream& operator>>(std::istream& is, HexFloat<T, Traits>& value) { + using HF = HexFloat<T, Traits>; + using uint_type = typename HF::uint_type; + using int_type = typename HF::int_type; + + value.set_value(static_cast<typename HF::native_type>(0.f)); + + if (is.flags() & std::ios::skipws) { + // If the user wants to skip whitespace , then we should obey that. + while (std::isspace(is.peek())) { + is.get(); + } + } + + auto next_char = is.peek(); + bool negate_value = false; + + if (next_char != '-' && next_char != '0') { + return ParseNormalFloat(is, negate_value, value); + } + + if (next_char == '-') { + negate_value = true; + is.get(); + next_char = is.peek(); + } + + if (next_char == '0') { + is.get(); // We may have to unget this. + auto maybe_hex_start = is.peek(); + if (maybe_hex_start != 'x' && maybe_hex_start != 'X') { + is.unget(); + return ParseNormalFloat(is, negate_value, value); + } else { + is.get(); // Throw away the 'x'; + } + } else { + return ParseNormalFloat(is, negate_value, value); + } + + // This "looks" like a hex-float so treat it as one. + bool seen_p = false; + bool seen_dot = false; + uint_type fraction_index = 0; + + uint_type fraction = 0; + int_type exponent = HF::exponent_bias; + + // Strip off leading zeros so we don't have to special-case them later. + while ((next_char = is.peek()) == '0') { + is.get(); + } + + bool is_denorm = + true; // Assume denorm "representation" until we hear otherwise. + // NB: This does not mean the value is actually denorm, + // it just means that it was written 0. + bool bits_written = false; // Stays false until we write a bit. + while (!seen_p && !seen_dot) { + // Handle characters that are left of the fractional part. + if (next_char == '.') { + seen_dot = true; + } else if (next_char == 'p') { + seen_p = true; + } else if (::isxdigit(next_char)) { + // We know this is not denormalized since we have stripped all leading + // zeroes and we are not a ".". + is_denorm = false; + int number = get_nibble_from_character(next_char); + for (int i = 0; i < 4; ++i, number <<= 1) { + uint_type write_bit = (number & 0x8) ? 0x1 : 0x0; + if (bits_written) { + // If we are here the bits represented belong in the fractional + // part of the float, and we have to adjust the exponent accordingly. + fraction = static_cast<uint_type>( + fraction | + static_cast<uint_type>( + write_bit << (HF::top_bit_left_shift - fraction_index++))); + exponent = static_cast<int_type>(exponent + 1); + } + bits_written |= write_bit != 0; + } + } else { + // We have not found our exponent yet, so we have to fail. + is.setstate(std::ios::failbit); + return is; + } + is.get(); + next_char = is.peek(); + } + bits_written = false; + while (seen_dot && !seen_p) { + // Handle only fractional parts now. + if (next_char == 'p') { + seen_p = true; + } else if (::isxdigit(next_char)) { + int number = get_nibble_from_character(next_char); + for (int i = 0; i < 4; ++i, number <<= 1) { + uint_type write_bit = (number & 0x8) ? 0x01 : 0x00; + bits_written |= write_bit != 0; + if (is_denorm && !bits_written) { + // Handle modifying the exponent here this way we can handle + // an arbitrary number of hex values without overflowing our + // integer. + exponent = static_cast<int_type>(exponent - 1); + } else { + fraction = static_cast<uint_type>( + fraction | + static_cast<uint_type>( + write_bit << (HF::top_bit_left_shift - fraction_index++))); + } + } + } else { + // We still have not found our 'p' exponent yet, so this is not a valid + // hex-float. + is.setstate(std::ios::failbit); + return is; + } + is.get(); + next_char = is.peek(); + } + + bool seen_sign = false; + int8_t exponent_sign = 1; + int_type written_exponent = 0; + while (true) { + if ((next_char == '-' || next_char == '+')) { + if (seen_sign) { + is.setstate(std::ios::failbit); + return is; + } + seen_sign = true; + exponent_sign = (next_char == '-') ? -1 : 1; + } else if (::isdigit(next_char)) { + // Hex-floats express their exponent as decimal. + written_exponent = static_cast<int_type>(written_exponent * 10); + written_exponent = + static_cast<int_type>(written_exponent + (next_char - '0')); + } else { + break; + } + is.get(); + next_char = is.peek(); + } + + written_exponent = static_cast<int_type>(written_exponent * exponent_sign); + exponent = static_cast<int_type>(exponent + written_exponent); + + bool is_zero = is_denorm && (fraction == 0); + if (is_denorm && !is_zero) { + fraction = static_cast<uint_type>(fraction << 1); + exponent = static_cast<int_type>(exponent - 1); + } else if (is_zero) { + exponent = 0; + } + + if (exponent <= 0 && !is_zero) { + fraction = static_cast<uint_type>(fraction >> 1); + fraction |= static_cast<uint_type>(1) << HF::top_bit_left_shift; + } + + fraction = (fraction >> HF::fraction_right_shift) & HF::fraction_encode_mask; + + const int_type max_exponent = + SetBits<uint_type, 0, HF::num_exponent_bits>::get; + + // Handle actual denorm numbers + while (exponent < 0 && !is_zero) { + fraction = static_cast<uint_type>(fraction >> 1); + exponent = static_cast<int_type>(exponent + 1); + + fraction &= HF::fraction_encode_mask; + if (fraction == 0) { + // We have underflowed our fraction. We should clamp to zero. + is_zero = true; + exponent = 0; + } + } + + // We have overflowed so we should be inf/-inf. + if (exponent > max_exponent) { + exponent = max_exponent; + fraction = 0; + } + + uint_type output_bits = static_cast<uint_type>( + static_cast<uint_type>(negate_value ? 1 : 0) << HF::top_bit_left_shift); + output_bits |= fraction; + + uint_type shifted_exponent = static_cast<uint_type>( + static_cast<uint_type>(exponent << HF::exponent_left_shift) & + HF::exponent_mask); + output_bits |= shifted_exponent; + + T output_float = spvutils::BitwiseCast<T>(output_bits); + value.set_value(output_float); + + return is; +} + +// Writes a FloatProxy value to a stream. +// Zero and normal numbers are printed in the usual notation, but with +// enough digits to fully reproduce the value. Other values (subnormal, +// NaN, and infinity) are printed as a hex float. +template <typename T> +std::ostream& operator<<(std::ostream& os, const FloatProxy<T>& value) { + auto float_val = value.getAsFloat(); + switch (std::fpclassify(float_val)) { + case FP_ZERO: + case FP_NORMAL: { + auto saved_precision = os.precision(); + os.precision(std::numeric_limits<T>::digits10); + os << float_val; + os.precision(saved_precision); + } break; + default: + os << HexFloat<FloatProxy<T>>(value); + break; + } + return os; +} + +template <> +inline std::ostream& operator<<<Float16>(std::ostream& os, + const FloatProxy<Float16>& value) { + os << HexFloat<FloatProxy<Float16>>(value); + return os; +} +} + +#endif // LIBSPIRV_UTIL_HEX_FLOAT_H_ diff --git a/thirdparty/glslang/SPIRV/spirv.hpp b/thirdparty/glslang/SPIRV/spirv.hpp new file mode 100644 index 0000000000..5297fd3902 --- /dev/null +++ b/thirdparty/glslang/SPIRV/spirv.hpp @@ -0,0 +1,1881 @@ +// Copyright (c) 2014-2019 The Khronos Group Inc. +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and/or associated documentation files (the "Materials"), +// to deal in the Materials without restriction, including without limitation +// the rights to use, copy, modify, merge, publish, distribute, sublicense, +// and/or sell copies of the Materials, and to permit persons to whom the +// Materials are furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Materials. +// +// MODIFICATIONS TO THIS FILE MAY MEAN IT NO LONGER ACCURATELY REFLECTS KHRONOS +// STANDARDS. THE UNMODIFIED, NORMATIVE VERSIONS OF KHRONOS SPECIFICATIONS AND +// HEADER INFORMATION ARE LOCATED AT https://www.khronos.org/registry/ +// +// THE MATERIALS ARE PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS +// OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL +// THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +// FROM,OUT OF OR IN CONNECTION WITH THE MATERIALS OR THE USE OR OTHER DEALINGS +// IN THE MATERIALS. + +// This header is automatically generated by the same tool that creates +// the Binary Section of the SPIR-V specification. + +// Enumeration tokens for SPIR-V, in various styles: +// C, C++, C++11, JSON, Lua, Python, C#, D +// +// - C will have tokens with a "Spv" prefix, e.g.: SpvSourceLanguageGLSL +// - C++ will have tokens in the "spv" name space, e.g.: spv::SourceLanguageGLSL +// - C++11 will use enum classes in the spv namespace, e.g.: spv::SourceLanguage::GLSL +// - Lua will use tables, e.g.: spv.SourceLanguage.GLSL +// - Python will use dictionaries, e.g.: spv['SourceLanguage']['GLSL'] +// - C# will use enum classes in the Specification class located in the "Spv" namespace, +// e.g.: Spv.Specification.SourceLanguage.GLSL +// - D will have tokens under the "spv" module, e.g: spv.SourceLanguage.GLSL +// +// Some tokens act like mask values, which can be OR'd together, +// while others are mutually exclusive. The mask-like ones have +// "Mask" in their name, and a parallel enum that has the shift +// amount (1 << x) for each corresponding enumerant. + +#ifndef spirv_HPP +#define spirv_HPP + +namespace spv { + +typedef unsigned int Id; + +#define SPV_VERSION 0x10400 +#define SPV_REVISION 1 + +static const unsigned int MagicNumber = 0x07230203; +static const unsigned int Version = 0x00010400; +static const unsigned int Revision = 1; +static const unsigned int OpCodeMask = 0xffff; +static const unsigned int WordCountShift = 16; + +enum SourceLanguage { + SourceLanguageUnknown = 0, + SourceLanguageESSL = 1, + SourceLanguageGLSL = 2, + SourceLanguageOpenCL_C = 3, + SourceLanguageOpenCL_CPP = 4, + SourceLanguageHLSL = 5, + SourceLanguageMax = 0x7fffffff, +}; + +enum ExecutionModel { + ExecutionModelVertex = 0, + ExecutionModelTessellationControl = 1, + ExecutionModelTessellationEvaluation = 2, + ExecutionModelGeometry = 3, + ExecutionModelFragment = 4, + ExecutionModelGLCompute = 5, + ExecutionModelKernel = 6, + ExecutionModelTaskNV = 5267, + ExecutionModelMeshNV = 5268, + ExecutionModelRayGenerationNV = 5313, + ExecutionModelIntersectionNV = 5314, + ExecutionModelAnyHitNV = 5315, + ExecutionModelClosestHitNV = 5316, + ExecutionModelMissNV = 5317, + ExecutionModelCallableNV = 5318, + ExecutionModelMax = 0x7fffffff, +}; + +enum AddressingModel { + AddressingModelLogical = 0, + AddressingModelPhysical32 = 1, + AddressingModelPhysical64 = 2, + AddressingModelPhysicalStorageBuffer64EXT = 5348, + AddressingModelMax = 0x7fffffff, +}; + +enum MemoryModel { + MemoryModelSimple = 0, + MemoryModelGLSL450 = 1, + MemoryModelOpenCL = 2, + MemoryModelVulkanKHR = 3, + MemoryModelMax = 0x7fffffff, +}; + +enum ExecutionMode { + ExecutionModeInvocations = 0, + ExecutionModeSpacingEqual = 1, + ExecutionModeSpacingFractionalEven = 2, + ExecutionModeSpacingFractionalOdd = 3, + ExecutionModeVertexOrderCw = 4, + ExecutionModeVertexOrderCcw = 5, + ExecutionModePixelCenterInteger = 6, + ExecutionModeOriginUpperLeft = 7, + ExecutionModeOriginLowerLeft = 8, + ExecutionModeEarlyFragmentTests = 9, + ExecutionModePointMode = 10, + ExecutionModeXfb = 11, + ExecutionModeDepthReplacing = 12, + ExecutionModeDepthGreater = 14, + ExecutionModeDepthLess = 15, + ExecutionModeDepthUnchanged = 16, + ExecutionModeLocalSize = 17, + ExecutionModeLocalSizeHint = 18, + ExecutionModeInputPoints = 19, + ExecutionModeInputLines = 20, + ExecutionModeInputLinesAdjacency = 21, + ExecutionModeTriangles = 22, + ExecutionModeInputTrianglesAdjacency = 23, + ExecutionModeQuads = 24, + ExecutionModeIsolines = 25, + ExecutionModeOutputVertices = 26, + ExecutionModeOutputPoints = 27, + ExecutionModeOutputLineStrip = 28, + ExecutionModeOutputTriangleStrip = 29, + ExecutionModeVecTypeHint = 30, + ExecutionModeContractionOff = 31, + ExecutionModeInitializer = 33, + ExecutionModeFinalizer = 34, + ExecutionModeSubgroupSize = 35, + ExecutionModeSubgroupsPerWorkgroup = 36, + ExecutionModeSubgroupsPerWorkgroupId = 37, + ExecutionModeLocalSizeId = 38, + ExecutionModeLocalSizeHintId = 39, + ExecutionModePostDepthCoverage = 4446, + ExecutionModeDenormPreserve = 4459, + ExecutionModeDenormFlushToZero = 4460, + ExecutionModeSignedZeroInfNanPreserve = 4461, + ExecutionModeRoundingModeRTE = 4462, + ExecutionModeRoundingModeRTZ = 4463, + ExecutionModeStencilRefReplacingEXT = 5027, + ExecutionModeOutputLinesNV = 5269, + ExecutionModeOutputPrimitivesNV = 5270, + ExecutionModeDerivativeGroupQuadsNV = 5289, + ExecutionModeDerivativeGroupLinearNV = 5290, + ExecutionModeOutputTrianglesNV = 5298, + ExecutionModeMax = 0x7fffffff, +}; + +enum StorageClass { + StorageClassUniformConstant = 0, + StorageClassInput = 1, + StorageClassUniform = 2, + StorageClassOutput = 3, + StorageClassWorkgroup = 4, + StorageClassCrossWorkgroup = 5, + StorageClassPrivate = 6, + StorageClassFunction = 7, + StorageClassGeneric = 8, + StorageClassPushConstant = 9, + StorageClassAtomicCounter = 10, + StorageClassImage = 11, + StorageClassStorageBuffer = 12, + StorageClassCallableDataNV = 5328, + StorageClassIncomingCallableDataNV = 5329, + StorageClassRayPayloadNV = 5338, + StorageClassHitAttributeNV = 5339, + StorageClassIncomingRayPayloadNV = 5342, + StorageClassShaderRecordBufferNV = 5343, + StorageClassPhysicalStorageBufferEXT = 5349, + StorageClassMax = 0x7fffffff, +}; + +enum Dim { + Dim1D = 0, + Dim2D = 1, + Dim3D = 2, + DimCube = 3, + DimRect = 4, + DimBuffer = 5, + DimSubpassData = 6, + DimMax = 0x7fffffff, +}; + +enum SamplerAddressingMode { + SamplerAddressingModeNone = 0, + SamplerAddressingModeClampToEdge = 1, + SamplerAddressingModeClamp = 2, + SamplerAddressingModeRepeat = 3, + SamplerAddressingModeRepeatMirrored = 4, + SamplerAddressingModeMax = 0x7fffffff, +}; + +enum SamplerFilterMode { + SamplerFilterModeNearest = 0, + SamplerFilterModeLinear = 1, + SamplerFilterModeMax = 0x7fffffff, +}; + +enum ImageFormat { + ImageFormatUnknown = 0, + ImageFormatRgba32f = 1, + ImageFormatRgba16f = 2, + ImageFormatR32f = 3, + ImageFormatRgba8 = 4, + ImageFormatRgba8Snorm = 5, + ImageFormatRg32f = 6, + ImageFormatRg16f = 7, + ImageFormatR11fG11fB10f = 8, + ImageFormatR16f = 9, + ImageFormatRgba16 = 10, + ImageFormatRgb10A2 = 11, + ImageFormatRg16 = 12, + ImageFormatRg8 = 13, + ImageFormatR16 = 14, + ImageFormatR8 = 15, + ImageFormatRgba16Snorm = 16, + ImageFormatRg16Snorm = 17, + ImageFormatRg8Snorm = 18, + ImageFormatR16Snorm = 19, + ImageFormatR8Snorm = 20, + ImageFormatRgba32i = 21, + ImageFormatRgba16i = 22, + ImageFormatRgba8i = 23, + ImageFormatR32i = 24, + ImageFormatRg32i = 25, + ImageFormatRg16i = 26, + ImageFormatRg8i = 27, + ImageFormatR16i = 28, + ImageFormatR8i = 29, + ImageFormatRgba32ui = 30, + ImageFormatRgba16ui = 31, + ImageFormatRgba8ui = 32, + ImageFormatR32ui = 33, + ImageFormatRgb10a2ui = 34, + ImageFormatRg32ui = 35, + ImageFormatRg16ui = 36, + ImageFormatRg8ui = 37, + ImageFormatR16ui = 38, + ImageFormatR8ui = 39, + ImageFormatMax = 0x7fffffff, +}; + +enum ImageChannelOrder { + ImageChannelOrderR = 0, + ImageChannelOrderA = 1, + ImageChannelOrderRG = 2, + ImageChannelOrderRA = 3, + ImageChannelOrderRGB = 4, + ImageChannelOrderRGBA = 5, + ImageChannelOrderBGRA = 6, + ImageChannelOrderARGB = 7, + ImageChannelOrderIntensity = 8, + ImageChannelOrderLuminance = 9, + ImageChannelOrderRx = 10, + ImageChannelOrderRGx = 11, + ImageChannelOrderRGBx = 12, + ImageChannelOrderDepth = 13, + ImageChannelOrderDepthStencil = 14, + ImageChannelOrdersRGB = 15, + ImageChannelOrdersRGBx = 16, + ImageChannelOrdersRGBA = 17, + ImageChannelOrdersBGRA = 18, + ImageChannelOrderABGR = 19, + ImageChannelOrderMax = 0x7fffffff, +}; + +enum ImageChannelDataType { + ImageChannelDataTypeSnormInt8 = 0, + ImageChannelDataTypeSnormInt16 = 1, + ImageChannelDataTypeUnormInt8 = 2, + ImageChannelDataTypeUnormInt16 = 3, + ImageChannelDataTypeUnormShort565 = 4, + ImageChannelDataTypeUnormShort555 = 5, + ImageChannelDataTypeUnormInt101010 = 6, + ImageChannelDataTypeSignedInt8 = 7, + ImageChannelDataTypeSignedInt16 = 8, + ImageChannelDataTypeSignedInt32 = 9, + ImageChannelDataTypeUnsignedInt8 = 10, + ImageChannelDataTypeUnsignedInt16 = 11, + ImageChannelDataTypeUnsignedInt32 = 12, + ImageChannelDataTypeHalfFloat = 13, + ImageChannelDataTypeFloat = 14, + ImageChannelDataTypeUnormInt24 = 15, + ImageChannelDataTypeUnormInt101010_2 = 16, + ImageChannelDataTypeMax = 0x7fffffff, +}; + +enum ImageOperandsShift { + ImageOperandsBiasShift = 0, + ImageOperandsLodShift = 1, + ImageOperandsGradShift = 2, + ImageOperandsConstOffsetShift = 3, + ImageOperandsOffsetShift = 4, + ImageOperandsConstOffsetsShift = 5, + ImageOperandsSampleShift = 6, + ImageOperandsMinLodShift = 7, + ImageOperandsMakeTexelAvailableKHRShift = 8, + ImageOperandsMakeTexelVisibleKHRShift = 9, + ImageOperandsNonPrivateTexelKHRShift = 10, + ImageOperandsVolatileTexelKHRShift = 11, + ImageOperandsSignExtendShift = 12, + ImageOperandsZeroExtendShift = 13, + ImageOperandsMax = 0x7fffffff, +}; + +enum ImageOperandsMask { + ImageOperandsMaskNone = 0, + ImageOperandsBiasMask = 0x00000001, + ImageOperandsLodMask = 0x00000002, + ImageOperandsGradMask = 0x00000004, + ImageOperandsConstOffsetMask = 0x00000008, + ImageOperandsOffsetMask = 0x00000010, + ImageOperandsConstOffsetsMask = 0x00000020, + ImageOperandsSampleMask = 0x00000040, + ImageOperandsMinLodMask = 0x00000080, + ImageOperandsMakeTexelAvailableKHRMask = 0x00000100, + ImageOperandsMakeTexelVisibleKHRMask = 0x00000200, + ImageOperandsNonPrivateTexelKHRMask = 0x00000400, + ImageOperandsVolatileTexelKHRMask = 0x00000800, + ImageOperandsSignExtendMask = 0x00001000, + ImageOperandsZeroExtendMask = 0x00002000, +}; + +enum FPFastMathModeShift { + FPFastMathModeNotNaNShift = 0, + FPFastMathModeNotInfShift = 1, + FPFastMathModeNSZShift = 2, + FPFastMathModeAllowRecipShift = 3, + FPFastMathModeFastShift = 4, + FPFastMathModeMax = 0x7fffffff, +}; + +enum FPFastMathModeMask { + FPFastMathModeMaskNone = 0, + FPFastMathModeNotNaNMask = 0x00000001, + FPFastMathModeNotInfMask = 0x00000002, + FPFastMathModeNSZMask = 0x00000004, + FPFastMathModeAllowRecipMask = 0x00000008, + FPFastMathModeFastMask = 0x00000010, +}; + +enum FPRoundingMode { + FPRoundingModeRTE = 0, + FPRoundingModeRTZ = 1, + FPRoundingModeRTP = 2, + FPRoundingModeRTN = 3, + FPRoundingModeMax = 0x7fffffff, +}; + +enum LinkageType { + LinkageTypeExport = 0, + LinkageTypeImport = 1, + LinkageTypeMax = 0x7fffffff, +}; + +enum AccessQualifier { + AccessQualifierReadOnly = 0, + AccessQualifierWriteOnly = 1, + AccessQualifierReadWrite = 2, + AccessQualifierMax = 0x7fffffff, +}; + +enum FunctionParameterAttribute { + FunctionParameterAttributeZext = 0, + FunctionParameterAttributeSext = 1, + FunctionParameterAttributeByVal = 2, + FunctionParameterAttributeSret = 3, + FunctionParameterAttributeNoAlias = 4, + FunctionParameterAttributeNoCapture = 5, + FunctionParameterAttributeNoWrite = 6, + FunctionParameterAttributeNoReadWrite = 7, + FunctionParameterAttributeMax = 0x7fffffff, +}; + +enum Decoration { + DecorationRelaxedPrecision = 0, + DecorationSpecId = 1, + DecorationBlock = 2, + DecorationBufferBlock = 3, + DecorationRowMajor = 4, + DecorationColMajor = 5, + DecorationArrayStride = 6, + DecorationMatrixStride = 7, + DecorationGLSLShared = 8, + DecorationGLSLPacked = 9, + DecorationCPacked = 10, + DecorationBuiltIn = 11, + DecorationNoPerspective = 13, + DecorationFlat = 14, + DecorationPatch = 15, + DecorationCentroid = 16, + DecorationSample = 17, + DecorationInvariant = 18, + DecorationRestrict = 19, + DecorationAliased = 20, + DecorationVolatile = 21, + DecorationConstant = 22, + DecorationCoherent = 23, + DecorationNonWritable = 24, + DecorationNonReadable = 25, + DecorationUniform = 26, + DecorationUniformId = 27, + DecorationSaturatedConversion = 28, + DecorationStream = 29, + DecorationLocation = 30, + DecorationComponent = 31, + DecorationIndex = 32, + DecorationBinding = 33, + DecorationDescriptorSet = 34, + DecorationOffset = 35, + DecorationXfbBuffer = 36, + DecorationXfbStride = 37, + DecorationFuncParamAttr = 38, + DecorationFPRoundingMode = 39, + DecorationFPFastMathMode = 40, + DecorationLinkageAttributes = 41, + DecorationNoContraction = 42, + DecorationInputAttachmentIndex = 43, + DecorationAlignment = 44, + DecorationMaxByteOffset = 45, + DecorationAlignmentId = 46, + DecorationMaxByteOffsetId = 47, + DecorationNoSignedWrap = 4469, + DecorationNoUnsignedWrap = 4470, + DecorationExplicitInterpAMD = 4999, + DecorationOverrideCoverageNV = 5248, + DecorationPassthroughNV = 5250, + DecorationViewportRelativeNV = 5252, + DecorationSecondaryViewportRelativeNV = 5256, + DecorationPerPrimitiveNV = 5271, + DecorationPerViewNV = 5272, + DecorationPerTaskNV = 5273, + DecorationPerVertexNV = 5285, + DecorationNonUniformEXT = 5300, + DecorationRestrictPointerEXT = 5355, + DecorationAliasedPointerEXT = 5356, + DecorationCounterBuffer = 5634, + DecorationHlslCounterBufferGOOGLE = 5634, + DecorationHlslSemanticGOOGLE = 5635, + DecorationUserSemantic = 5635, + DecorationMax = 0x7fffffff, +}; + +enum BuiltIn { + BuiltInPosition = 0, + BuiltInPointSize = 1, + BuiltInClipDistance = 3, + BuiltInCullDistance = 4, + BuiltInVertexId = 5, + BuiltInInstanceId = 6, + BuiltInPrimitiveId = 7, + BuiltInInvocationId = 8, + BuiltInLayer = 9, + BuiltInViewportIndex = 10, + BuiltInTessLevelOuter = 11, + BuiltInTessLevelInner = 12, + BuiltInTessCoord = 13, + BuiltInPatchVertices = 14, + BuiltInFragCoord = 15, + BuiltInPointCoord = 16, + BuiltInFrontFacing = 17, + BuiltInSampleId = 18, + BuiltInSamplePosition = 19, + BuiltInSampleMask = 20, + BuiltInFragDepth = 22, + BuiltInHelperInvocation = 23, + BuiltInNumWorkgroups = 24, + BuiltInWorkgroupSize = 25, + BuiltInWorkgroupId = 26, + BuiltInLocalInvocationId = 27, + BuiltInGlobalInvocationId = 28, + BuiltInLocalInvocationIndex = 29, + BuiltInWorkDim = 30, + BuiltInGlobalSize = 31, + BuiltInEnqueuedWorkgroupSize = 32, + BuiltInGlobalOffset = 33, + BuiltInGlobalLinearId = 34, + BuiltInSubgroupSize = 36, + BuiltInSubgroupMaxSize = 37, + BuiltInNumSubgroups = 38, + BuiltInNumEnqueuedSubgroups = 39, + BuiltInSubgroupId = 40, + BuiltInSubgroupLocalInvocationId = 41, + BuiltInVertexIndex = 42, + BuiltInInstanceIndex = 43, + BuiltInSubgroupEqMask = 4416, + BuiltInSubgroupEqMaskKHR = 4416, + BuiltInSubgroupGeMask = 4417, + BuiltInSubgroupGeMaskKHR = 4417, + BuiltInSubgroupGtMask = 4418, + BuiltInSubgroupGtMaskKHR = 4418, + BuiltInSubgroupLeMask = 4419, + BuiltInSubgroupLeMaskKHR = 4419, + BuiltInSubgroupLtMask = 4420, + BuiltInSubgroupLtMaskKHR = 4420, + BuiltInBaseVertex = 4424, + BuiltInBaseInstance = 4425, + BuiltInDrawIndex = 4426, + BuiltInDeviceIndex = 4438, + BuiltInViewIndex = 4440, + BuiltInBaryCoordNoPerspAMD = 4992, + BuiltInBaryCoordNoPerspCentroidAMD = 4993, + BuiltInBaryCoordNoPerspSampleAMD = 4994, + BuiltInBaryCoordSmoothAMD = 4995, + BuiltInBaryCoordSmoothCentroidAMD = 4996, + BuiltInBaryCoordSmoothSampleAMD = 4997, + BuiltInBaryCoordPullModelAMD = 4998, + BuiltInFragStencilRefEXT = 5014, + BuiltInViewportMaskNV = 5253, + BuiltInSecondaryPositionNV = 5257, + BuiltInSecondaryViewportMaskNV = 5258, + BuiltInPositionPerViewNV = 5261, + BuiltInViewportMaskPerViewNV = 5262, + BuiltInFullyCoveredEXT = 5264, + BuiltInTaskCountNV = 5274, + BuiltInPrimitiveCountNV = 5275, + BuiltInPrimitiveIndicesNV = 5276, + BuiltInClipDistancePerViewNV = 5277, + BuiltInCullDistancePerViewNV = 5278, + BuiltInLayerPerViewNV = 5279, + BuiltInMeshViewCountNV = 5280, + BuiltInMeshViewIndicesNV = 5281, + BuiltInBaryCoordNV = 5286, + BuiltInBaryCoordNoPerspNV = 5287, + BuiltInFragSizeEXT = 5292, + BuiltInFragmentSizeNV = 5292, + BuiltInFragInvocationCountEXT = 5293, + BuiltInInvocationsPerPixelNV = 5293, + BuiltInLaunchIdNV = 5319, + BuiltInLaunchSizeNV = 5320, + BuiltInWorldRayOriginNV = 5321, + BuiltInWorldRayDirectionNV = 5322, + BuiltInObjectRayOriginNV = 5323, + BuiltInObjectRayDirectionNV = 5324, + BuiltInRayTminNV = 5325, + BuiltInRayTmaxNV = 5326, + BuiltInInstanceCustomIndexNV = 5327, + BuiltInObjectToWorldNV = 5330, + BuiltInWorldToObjectNV = 5331, + BuiltInHitTNV = 5332, + BuiltInHitKindNV = 5333, + BuiltInIncomingRayFlagsNV = 5351, + BuiltInMax = 0x7fffffff, +}; + +enum SelectionControlShift { + SelectionControlFlattenShift = 0, + SelectionControlDontFlattenShift = 1, + SelectionControlMax = 0x7fffffff, +}; + +enum SelectionControlMask { + SelectionControlMaskNone = 0, + SelectionControlFlattenMask = 0x00000001, + SelectionControlDontFlattenMask = 0x00000002, +}; + +enum LoopControlShift { + LoopControlUnrollShift = 0, + LoopControlDontUnrollShift = 1, + LoopControlDependencyInfiniteShift = 2, + LoopControlDependencyLengthShift = 3, + LoopControlMinIterationsShift = 4, + LoopControlMaxIterationsShift = 5, + LoopControlIterationMultipleShift = 6, + LoopControlPeelCountShift = 7, + LoopControlPartialCountShift = 8, + LoopControlMax = 0x7fffffff, +}; + +enum LoopControlMask { + LoopControlMaskNone = 0, + LoopControlUnrollMask = 0x00000001, + LoopControlDontUnrollMask = 0x00000002, + LoopControlDependencyInfiniteMask = 0x00000004, + LoopControlDependencyLengthMask = 0x00000008, + LoopControlMinIterationsMask = 0x00000010, + LoopControlMaxIterationsMask = 0x00000020, + LoopControlIterationMultipleMask = 0x00000040, + LoopControlPeelCountMask = 0x00000080, + LoopControlPartialCountMask = 0x00000100, +}; + +enum FunctionControlShift { + FunctionControlInlineShift = 0, + FunctionControlDontInlineShift = 1, + FunctionControlPureShift = 2, + FunctionControlConstShift = 3, + FunctionControlMax = 0x7fffffff, +}; + +enum FunctionControlMask { + FunctionControlMaskNone = 0, + FunctionControlInlineMask = 0x00000001, + FunctionControlDontInlineMask = 0x00000002, + FunctionControlPureMask = 0x00000004, + FunctionControlConstMask = 0x00000008, +}; + +enum MemorySemanticsShift { + MemorySemanticsAcquireShift = 1, + MemorySemanticsReleaseShift = 2, + MemorySemanticsAcquireReleaseShift = 3, + MemorySemanticsSequentiallyConsistentShift = 4, + MemorySemanticsUniformMemoryShift = 6, + MemorySemanticsSubgroupMemoryShift = 7, + MemorySemanticsWorkgroupMemoryShift = 8, + MemorySemanticsCrossWorkgroupMemoryShift = 9, + MemorySemanticsAtomicCounterMemoryShift = 10, + MemorySemanticsImageMemoryShift = 11, + MemorySemanticsOutputMemoryKHRShift = 12, + MemorySemanticsMakeAvailableKHRShift = 13, + MemorySemanticsMakeVisibleKHRShift = 14, + MemorySemanticsMax = 0x7fffffff, +}; + +enum MemorySemanticsMask { + MemorySemanticsMaskNone = 0, + MemorySemanticsAcquireMask = 0x00000002, + MemorySemanticsReleaseMask = 0x00000004, + MemorySemanticsAcquireReleaseMask = 0x00000008, + MemorySemanticsSequentiallyConsistentMask = 0x00000010, + MemorySemanticsUniformMemoryMask = 0x00000040, + MemorySemanticsSubgroupMemoryMask = 0x00000080, + MemorySemanticsWorkgroupMemoryMask = 0x00000100, + MemorySemanticsCrossWorkgroupMemoryMask = 0x00000200, + MemorySemanticsAtomicCounterMemoryMask = 0x00000400, + MemorySemanticsImageMemoryMask = 0x00000800, + MemorySemanticsOutputMemoryKHRMask = 0x00001000, + MemorySemanticsMakeAvailableKHRMask = 0x00002000, + MemorySemanticsMakeVisibleKHRMask = 0x00004000, +}; + +enum MemoryAccessShift { + MemoryAccessVolatileShift = 0, + MemoryAccessAlignedShift = 1, + MemoryAccessNontemporalShift = 2, + MemoryAccessMakePointerAvailableKHRShift = 3, + MemoryAccessMakePointerVisibleKHRShift = 4, + MemoryAccessNonPrivatePointerKHRShift = 5, + MemoryAccessMax = 0x7fffffff, +}; + +enum MemoryAccessMask { + MemoryAccessMaskNone = 0, + MemoryAccessVolatileMask = 0x00000001, + MemoryAccessAlignedMask = 0x00000002, + MemoryAccessNontemporalMask = 0x00000004, + MemoryAccessMakePointerAvailableKHRMask = 0x00000008, + MemoryAccessMakePointerVisibleKHRMask = 0x00000010, + MemoryAccessNonPrivatePointerKHRMask = 0x00000020, +}; + +enum Scope { + ScopeCrossDevice = 0, + ScopeDevice = 1, + ScopeWorkgroup = 2, + ScopeSubgroup = 3, + ScopeInvocation = 4, + ScopeQueueFamilyKHR = 5, + ScopeMax = 0x7fffffff, +}; + +enum GroupOperation { + GroupOperationReduce = 0, + GroupOperationInclusiveScan = 1, + GroupOperationExclusiveScan = 2, + GroupOperationClusteredReduce = 3, + GroupOperationPartitionedReduceNV = 6, + GroupOperationPartitionedInclusiveScanNV = 7, + GroupOperationPartitionedExclusiveScanNV = 8, + GroupOperationMax = 0x7fffffff, +}; + +enum KernelEnqueueFlags { + KernelEnqueueFlagsNoWait = 0, + KernelEnqueueFlagsWaitKernel = 1, + KernelEnqueueFlagsWaitWorkGroup = 2, + KernelEnqueueFlagsMax = 0x7fffffff, +}; + +enum KernelProfilingInfoShift { + KernelProfilingInfoCmdExecTimeShift = 0, + KernelProfilingInfoMax = 0x7fffffff, +}; + +enum KernelProfilingInfoMask { + KernelProfilingInfoMaskNone = 0, + KernelProfilingInfoCmdExecTimeMask = 0x00000001, +}; + +enum Capability { + CapabilityMatrix = 0, + CapabilityShader = 1, + CapabilityGeometry = 2, + CapabilityTessellation = 3, + CapabilityAddresses = 4, + CapabilityLinkage = 5, + CapabilityKernel = 6, + CapabilityVector16 = 7, + CapabilityFloat16Buffer = 8, + CapabilityFloat16 = 9, + CapabilityFloat64 = 10, + CapabilityInt64 = 11, + CapabilityInt64Atomics = 12, + CapabilityImageBasic = 13, + CapabilityImageReadWrite = 14, + CapabilityImageMipmap = 15, + CapabilityPipes = 17, + CapabilityGroups = 18, + CapabilityDeviceEnqueue = 19, + CapabilityLiteralSampler = 20, + CapabilityAtomicStorage = 21, + CapabilityInt16 = 22, + CapabilityTessellationPointSize = 23, + CapabilityGeometryPointSize = 24, + CapabilityImageGatherExtended = 25, + CapabilityStorageImageMultisample = 27, + CapabilityUniformBufferArrayDynamicIndexing = 28, + CapabilitySampledImageArrayDynamicIndexing = 29, + CapabilityStorageBufferArrayDynamicIndexing = 30, + CapabilityStorageImageArrayDynamicIndexing = 31, + CapabilityClipDistance = 32, + CapabilityCullDistance = 33, + CapabilityImageCubeArray = 34, + CapabilitySampleRateShading = 35, + CapabilityImageRect = 36, + CapabilitySampledRect = 37, + CapabilityGenericPointer = 38, + CapabilityInt8 = 39, + CapabilityInputAttachment = 40, + CapabilitySparseResidency = 41, + CapabilityMinLod = 42, + CapabilitySampled1D = 43, + CapabilityImage1D = 44, + CapabilitySampledCubeArray = 45, + CapabilitySampledBuffer = 46, + CapabilityImageBuffer = 47, + CapabilityImageMSArray = 48, + CapabilityStorageImageExtendedFormats = 49, + CapabilityImageQuery = 50, + CapabilityDerivativeControl = 51, + CapabilityInterpolationFunction = 52, + CapabilityTransformFeedback = 53, + CapabilityGeometryStreams = 54, + CapabilityStorageImageReadWithoutFormat = 55, + CapabilityStorageImageWriteWithoutFormat = 56, + CapabilityMultiViewport = 57, + CapabilitySubgroupDispatch = 58, + CapabilityNamedBarrier = 59, + CapabilityPipeStorage = 60, + CapabilityGroupNonUniform = 61, + CapabilityGroupNonUniformVote = 62, + CapabilityGroupNonUniformArithmetic = 63, + CapabilityGroupNonUniformBallot = 64, + CapabilityGroupNonUniformShuffle = 65, + CapabilityGroupNonUniformShuffleRelative = 66, + CapabilityGroupNonUniformClustered = 67, + CapabilityGroupNonUniformQuad = 68, + CapabilitySubgroupBallotKHR = 4423, + CapabilityDrawParameters = 4427, + CapabilitySubgroupVoteKHR = 4431, + CapabilityStorageBuffer16BitAccess = 4433, + CapabilityStorageUniformBufferBlock16 = 4433, + CapabilityStorageUniform16 = 4434, + CapabilityUniformAndStorageBuffer16BitAccess = 4434, + CapabilityStoragePushConstant16 = 4435, + CapabilityStorageInputOutput16 = 4436, + CapabilityDeviceGroup = 4437, + CapabilityMultiView = 4439, + CapabilityVariablePointersStorageBuffer = 4441, + CapabilityVariablePointers = 4442, + CapabilityAtomicStorageOps = 4445, + CapabilitySampleMaskPostDepthCoverage = 4447, + CapabilityStorageBuffer8BitAccess = 4448, + CapabilityUniformAndStorageBuffer8BitAccess = 4449, + CapabilityStoragePushConstant8 = 4450, + CapabilityDenormPreserve = 4464, + CapabilityDenormFlushToZero = 4465, + CapabilitySignedZeroInfNanPreserve = 4466, + CapabilityRoundingModeRTE = 4467, + CapabilityRoundingModeRTZ = 4468, + CapabilityFloat16ImageAMD = 5008, + CapabilityImageGatherBiasLodAMD = 5009, + CapabilityFragmentMaskAMD = 5010, + CapabilityStencilExportEXT = 5013, + CapabilityImageReadWriteLodAMD = 5015, + CapabilitySampleMaskOverrideCoverageNV = 5249, + CapabilityGeometryShaderPassthroughNV = 5251, + CapabilityShaderViewportIndexLayerEXT = 5254, + CapabilityShaderViewportIndexLayerNV = 5254, + CapabilityShaderViewportMaskNV = 5255, + CapabilityShaderStereoViewNV = 5259, + CapabilityPerViewAttributesNV = 5260, + CapabilityFragmentFullyCoveredEXT = 5265, + CapabilityMeshShadingNV = 5266, + CapabilityImageFootprintNV = 5282, + CapabilityFragmentBarycentricNV = 5284, + CapabilityComputeDerivativeGroupQuadsNV = 5288, + CapabilityFragmentDensityEXT = 5291, + CapabilityShadingRateNV = 5291, + CapabilityGroupNonUniformPartitionedNV = 5297, + CapabilityShaderNonUniformEXT = 5301, + CapabilityRuntimeDescriptorArrayEXT = 5302, + CapabilityInputAttachmentArrayDynamicIndexingEXT = 5303, + CapabilityUniformTexelBufferArrayDynamicIndexingEXT = 5304, + CapabilityStorageTexelBufferArrayDynamicIndexingEXT = 5305, + CapabilityUniformBufferArrayNonUniformIndexingEXT = 5306, + CapabilitySampledImageArrayNonUniformIndexingEXT = 5307, + CapabilityStorageBufferArrayNonUniformIndexingEXT = 5308, + CapabilityStorageImageArrayNonUniformIndexingEXT = 5309, + CapabilityInputAttachmentArrayNonUniformIndexingEXT = 5310, + CapabilityUniformTexelBufferArrayNonUniformIndexingEXT = 5311, + CapabilityStorageTexelBufferArrayNonUniformIndexingEXT = 5312, + CapabilityRayTracingNV = 5340, + CapabilityVulkanMemoryModelKHR = 5345, + CapabilityVulkanMemoryModelDeviceScopeKHR = 5346, + CapabilityPhysicalStorageBufferAddressesEXT = 5347, + CapabilityComputeDerivativeGroupLinearNV = 5350, + CapabilityCooperativeMatrixNV = 5357, + CapabilitySubgroupShuffleINTEL = 5568, + CapabilitySubgroupBufferBlockIOINTEL = 5569, + CapabilitySubgroupImageBlockIOINTEL = 5570, + CapabilitySubgroupImageMediaBlockIOINTEL = 5579, + CapabilitySubgroupAvcMotionEstimationINTEL = 5696, + CapabilitySubgroupAvcMotionEstimationIntraINTEL = 5697, + CapabilitySubgroupAvcMotionEstimationChromaINTEL = 5698, + CapabilityMax = 0x7fffffff, +}; + +enum Op { + OpNop = 0, + OpUndef = 1, + OpSourceContinued = 2, + OpSource = 3, + OpSourceExtension = 4, + OpName = 5, + OpMemberName = 6, + OpString = 7, + OpLine = 8, + OpExtension = 10, + OpExtInstImport = 11, + OpExtInst = 12, + OpMemoryModel = 14, + OpEntryPoint = 15, + OpExecutionMode = 16, + OpCapability = 17, + OpTypeVoid = 19, + OpTypeBool = 20, + OpTypeInt = 21, + OpTypeFloat = 22, + OpTypeVector = 23, + OpTypeMatrix = 24, + OpTypeImage = 25, + OpTypeSampler = 26, + OpTypeSampledImage = 27, + OpTypeArray = 28, + OpTypeRuntimeArray = 29, + OpTypeStruct = 30, + OpTypeOpaque = 31, + OpTypePointer = 32, + OpTypeFunction = 33, + OpTypeEvent = 34, + OpTypeDeviceEvent = 35, + OpTypeReserveId = 36, + OpTypeQueue = 37, + OpTypePipe = 38, + OpTypeForwardPointer = 39, + OpConstantTrue = 41, + OpConstantFalse = 42, + OpConstant = 43, + OpConstantComposite = 44, + OpConstantSampler = 45, + OpConstantNull = 46, + OpSpecConstantTrue = 48, + OpSpecConstantFalse = 49, + OpSpecConstant = 50, + OpSpecConstantComposite = 51, + OpSpecConstantOp = 52, + OpFunction = 54, + OpFunctionParameter = 55, + OpFunctionEnd = 56, + OpFunctionCall = 57, + OpVariable = 59, + OpImageTexelPointer = 60, + OpLoad = 61, + OpStore = 62, + OpCopyMemory = 63, + OpCopyMemorySized = 64, + OpAccessChain = 65, + OpInBoundsAccessChain = 66, + OpPtrAccessChain = 67, + OpArrayLength = 68, + OpGenericPtrMemSemantics = 69, + OpInBoundsPtrAccessChain = 70, + OpDecorate = 71, + OpMemberDecorate = 72, + OpDecorationGroup = 73, + OpGroupDecorate = 74, + OpGroupMemberDecorate = 75, + OpVectorExtractDynamic = 77, + OpVectorInsertDynamic = 78, + OpVectorShuffle = 79, + OpCompositeConstruct = 80, + OpCompositeExtract = 81, + OpCompositeInsert = 82, + OpCopyObject = 83, + OpTranspose = 84, + OpSampledImage = 86, + OpImageSampleImplicitLod = 87, + OpImageSampleExplicitLod = 88, + OpImageSampleDrefImplicitLod = 89, + OpImageSampleDrefExplicitLod = 90, + OpImageSampleProjImplicitLod = 91, + OpImageSampleProjExplicitLod = 92, + OpImageSampleProjDrefImplicitLod = 93, + OpImageSampleProjDrefExplicitLod = 94, + OpImageFetch = 95, + OpImageGather = 96, + OpImageDrefGather = 97, + OpImageRead = 98, + OpImageWrite = 99, + OpImage = 100, + OpImageQueryFormat = 101, + OpImageQueryOrder = 102, + OpImageQuerySizeLod = 103, + OpImageQuerySize = 104, + OpImageQueryLod = 105, + OpImageQueryLevels = 106, + OpImageQuerySamples = 107, + OpConvertFToU = 109, + OpConvertFToS = 110, + OpConvertSToF = 111, + OpConvertUToF = 112, + OpUConvert = 113, + OpSConvert = 114, + OpFConvert = 115, + OpQuantizeToF16 = 116, + OpConvertPtrToU = 117, + OpSatConvertSToU = 118, + OpSatConvertUToS = 119, + OpConvertUToPtr = 120, + OpPtrCastToGeneric = 121, + OpGenericCastToPtr = 122, + OpGenericCastToPtrExplicit = 123, + OpBitcast = 124, + OpSNegate = 126, + OpFNegate = 127, + OpIAdd = 128, + OpFAdd = 129, + OpISub = 130, + OpFSub = 131, + OpIMul = 132, + OpFMul = 133, + OpUDiv = 134, + OpSDiv = 135, + OpFDiv = 136, + OpUMod = 137, + OpSRem = 138, + OpSMod = 139, + OpFRem = 140, + OpFMod = 141, + OpVectorTimesScalar = 142, + OpMatrixTimesScalar = 143, + OpVectorTimesMatrix = 144, + OpMatrixTimesVector = 145, + OpMatrixTimesMatrix = 146, + OpOuterProduct = 147, + OpDot = 148, + OpIAddCarry = 149, + OpISubBorrow = 150, + OpUMulExtended = 151, + OpSMulExtended = 152, + OpAny = 154, + OpAll = 155, + OpIsNan = 156, + OpIsInf = 157, + OpIsFinite = 158, + OpIsNormal = 159, + OpSignBitSet = 160, + OpLessOrGreater = 161, + OpOrdered = 162, + OpUnordered = 163, + OpLogicalEqual = 164, + OpLogicalNotEqual = 165, + OpLogicalOr = 166, + OpLogicalAnd = 167, + OpLogicalNot = 168, + OpSelect = 169, + OpIEqual = 170, + OpINotEqual = 171, + OpUGreaterThan = 172, + OpSGreaterThan = 173, + OpUGreaterThanEqual = 174, + OpSGreaterThanEqual = 175, + OpULessThan = 176, + OpSLessThan = 177, + OpULessThanEqual = 178, + OpSLessThanEqual = 179, + OpFOrdEqual = 180, + OpFUnordEqual = 181, + OpFOrdNotEqual = 182, + OpFUnordNotEqual = 183, + OpFOrdLessThan = 184, + OpFUnordLessThan = 185, + OpFOrdGreaterThan = 186, + OpFUnordGreaterThan = 187, + OpFOrdLessThanEqual = 188, + OpFUnordLessThanEqual = 189, + OpFOrdGreaterThanEqual = 190, + OpFUnordGreaterThanEqual = 191, + OpShiftRightLogical = 194, + OpShiftRightArithmetic = 195, + OpShiftLeftLogical = 196, + OpBitwiseOr = 197, + OpBitwiseXor = 198, + OpBitwiseAnd = 199, + OpNot = 200, + OpBitFieldInsert = 201, + OpBitFieldSExtract = 202, + OpBitFieldUExtract = 203, + OpBitReverse = 204, + OpBitCount = 205, + OpDPdx = 207, + OpDPdy = 208, + OpFwidth = 209, + OpDPdxFine = 210, + OpDPdyFine = 211, + OpFwidthFine = 212, + OpDPdxCoarse = 213, + OpDPdyCoarse = 214, + OpFwidthCoarse = 215, + OpEmitVertex = 218, + OpEndPrimitive = 219, + OpEmitStreamVertex = 220, + OpEndStreamPrimitive = 221, + OpControlBarrier = 224, + OpMemoryBarrier = 225, + OpAtomicLoad = 227, + OpAtomicStore = 228, + OpAtomicExchange = 229, + OpAtomicCompareExchange = 230, + OpAtomicCompareExchangeWeak = 231, + OpAtomicIIncrement = 232, + OpAtomicIDecrement = 233, + OpAtomicIAdd = 234, + OpAtomicISub = 235, + OpAtomicSMin = 236, + OpAtomicUMin = 237, + OpAtomicSMax = 238, + OpAtomicUMax = 239, + OpAtomicAnd = 240, + OpAtomicOr = 241, + OpAtomicXor = 242, + OpPhi = 245, + OpLoopMerge = 246, + OpSelectionMerge = 247, + OpLabel = 248, + OpBranch = 249, + OpBranchConditional = 250, + OpSwitch = 251, + OpKill = 252, + OpReturn = 253, + OpReturnValue = 254, + OpUnreachable = 255, + OpLifetimeStart = 256, + OpLifetimeStop = 257, + OpGroupAsyncCopy = 259, + OpGroupWaitEvents = 260, + OpGroupAll = 261, + OpGroupAny = 262, + OpGroupBroadcast = 263, + OpGroupIAdd = 264, + OpGroupFAdd = 265, + OpGroupFMin = 266, + OpGroupUMin = 267, + OpGroupSMin = 268, + OpGroupFMax = 269, + OpGroupUMax = 270, + OpGroupSMax = 271, + OpReadPipe = 274, + OpWritePipe = 275, + OpReservedReadPipe = 276, + OpReservedWritePipe = 277, + OpReserveReadPipePackets = 278, + OpReserveWritePipePackets = 279, + OpCommitReadPipe = 280, + OpCommitWritePipe = 281, + OpIsValidReserveId = 282, + OpGetNumPipePackets = 283, + OpGetMaxPipePackets = 284, + OpGroupReserveReadPipePackets = 285, + OpGroupReserveWritePipePackets = 286, + OpGroupCommitReadPipe = 287, + OpGroupCommitWritePipe = 288, + OpEnqueueMarker = 291, + OpEnqueueKernel = 292, + OpGetKernelNDrangeSubGroupCount = 293, + OpGetKernelNDrangeMaxSubGroupSize = 294, + OpGetKernelWorkGroupSize = 295, + OpGetKernelPreferredWorkGroupSizeMultiple = 296, + OpRetainEvent = 297, + OpReleaseEvent = 298, + OpCreateUserEvent = 299, + OpIsValidEvent = 300, + OpSetUserEventStatus = 301, + OpCaptureEventProfilingInfo = 302, + OpGetDefaultQueue = 303, + OpBuildNDRange = 304, + OpImageSparseSampleImplicitLod = 305, + OpImageSparseSampleExplicitLod = 306, + OpImageSparseSampleDrefImplicitLod = 307, + OpImageSparseSampleDrefExplicitLod = 308, + OpImageSparseSampleProjImplicitLod = 309, + OpImageSparseSampleProjExplicitLod = 310, + OpImageSparseSampleProjDrefImplicitLod = 311, + OpImageSparseSampleProjDrefExplicitLod = 312, + OpImageSparseFetch = 313, + OpImageSparseGather = 314, + OpImageSparseDrefGather = 315, + OpImageSparseTexelsResident = 316, + OpNoLine = 317, + OpAtomicFlagTestAndSet = 318, + OpAtomicFlagClear = 319, + OpImageSparseRead = 320, + OpSizeOf = 321, + OpTypePipeStorage = 322, + OpConstantPipeStorage = 323, + OpCreatePipeFromPipeStorage = 324, + OpGetKernelLocalSizeForSubgroupCount = 325, + OpGetKernelMaxNumSubgroups = 326, + OpTypeNamedBarrier = 327, + OpNamedBarrierInitialize = 328, + OpMemoryNamedBarrier = 329, + OpModuleProcessed = 330, + OpExecutionModeId = 331, + OpDecorateId = 332, + OpGroupNonUniformElect = 333, + OpGroupNonUniformAll = 334, + OpGroupNonUniformAny = 335, + OpGroupNonUniformAllEqual = 336, + OpGroupNonUniformBroadcast = 337, + OpGroupNonUniformBroadcastFirst = 338, + OpGroupNonUniformBallot = 339, + OpGroupNonUniformInverseBallot = 340, + OpGroupNonUniformBallotBitExtract = 341, + OpGroupNonUniformBallotBitCount = 342, + OpGroupNonUniformBallotFindLSB = 343, + OpGroupNonUniformBallotFindMSB = 344, + OpGroupNonUniformShuffle = 345, + OpGroupNonUniformShuffleXor = 346, + OpGroupNonUniformShuffleUp = 347, + OpGroupNonUniformShuffleDown = 348, + OpGroupNonUniformIAdd = 349, + OpGroupNonUniformFAdd = 350, + OpGroupNonUniformIMul = 351, + OpGroupNonUniformFMul = 352, + OpGroupNonUniformSMin = 353, + OpGroupNonUniformUMin = 354, + OpGroupNonUniformFMin = 355, + OpGroupNonUniformSMax = 356, + OpGroupNonUniformUMax = 357, + OpGroupNonUniformFMax = 358, + OpGroupNonUniformBitwiseAnd = 359, + OpGroupNonUniformBitwiseOr = 360, + OpGroupNonUniformBitwiseXor = 361, + OpGroupNonUniformLogicalAnd = 362, + OpGroupNonUniformLogicalOr = 363, + OpGroupNonUniformLogicalXor = 364, + OpGroupNonUniformQuadBroadcast = 365, + OpGroupNonUniformQuadSwap = 366, + OpCopyLogical = 400, + OpPtrEqual = 401, + OpPtrNotEqual = 402, + OpPtrDiff = 403, + OpSubgroupBallotKHR = 4421, + OpSubgroupFirstInvocationKHR = 4422, + OpSubgroupAllKHR = 4428, + OpSubgroupAnyKHR = 4429, + OpSubgroupAllEqualKHR = 4430, + OpSubgroupReadInvocationKHR = 4432, + OpGroupIAddNonUniformAMD = 5000, + OpGroupFAddNonUniformAMD = 5001, + OpGroupFMinNonUniformAMD = 5002, + OpGroupUMinNonUniformAMD = 5003, + OpGroupSMinNonUniformAMD = 5004, + OpGroupFMaxNonUniformAMD = 5005, + OpGroupUMaxNonUniformAMD = 5006, + OpGroupSMaxNonUniformAMD = 5007, + OpFragmentMaskFetchAMD = 5011, + OpFragmentFetchAMD = 5012, + OpImageSampleFootprintNV = 5283, + OpGroupNonUniformPartitionNV = 5296, + OpWritePackedPrimitiveIndices4x8NV = 5299, + OpReportIntersectionNV = 5334, + OpIgnoreIntersectionNV = 5335, + OpTerminateRayNV = 5336, + OpTraceNV = 5337, + OpTypeAccelerationStructureNV = 5341, + OpExecuteCallableNV = 5344, + OpTypeCooperativeMatrixNV = 5358, + OpCooperativeMatrixLoadNV = 5359, + OpCooperativeMatrixStoreNV = 5360, + OpCooperativeMatrixMulAddNV = 5361, + OpCooperativeMatrixLengthNV = 5362, + OpSubgroupShuffleINTEL = 5571, + OpSubgroupShuffleDownINTEL = 5572, + OpSubgroupShuffleUpINTEL = 5573, + OpSubgroupShuffleXorINTEL = 5574, + OpSubgroupBlockReadINTEL = 5575, + OpSubgroupBlockWriteINTEL = 5576, + OpSubgroupImageBlockReadINTEL = 5577, + OpSubgroupImageBlockWriteINTEL = 5578, + OpSubgroupImageMediaBlockReadINTEL = 5580, + OpSubgroupImageMediaBlockWriteINTEL = 5581, + OpDecorateString = 5632, + OpDecorateStringGOOGLE = 5632, + OpMemberDecorateString = 5633, + OpMemberDecorateStringGOOGLE = 5633, + OpVmeImageINTEL = 5699, + OpTypeVmeImageINTEL = 5700, + OpTypeAvcImePayloadINTEL = 5701, + OpTypeAvcRefPayloadINTEL = 5702, + OpTypeAvcSicPayloadINTEL = 5703, + OpTypeAvcMcePayloadINTEL = 5704, + OpTypeAvcMceResultINTEL = 5705, + OpTypeAvcImeResultINTEL = 5706, + OpTypeAvcImeResultSingleReferenceStreamoutINTEL = 5707, + OpTypeAvcImeResultDualReferenceStreamoutINTEL = 5708, + OpTypeAvcImeSingleReferenceStreaminINTEL = 5709, + OpTypeAvcImeDualReferenceStreaminINTEL = 5710, + OpTypeAvcRefResultINTEL = 5711, + OpTypeAvcSicResultINTEL = 5712, + OpSubgroupAvcMceGetDefaultInterBaseMultiReferencePenaltyINTEL = 5713, + OpSubgroupAvcMceSetInterBaseMultiReferencePenaltyINTEL = 5714, + OpSubgroupAvcMceGetDefaultInterShapePenaltyINTEL = 5715, + OpSubgroupAvcMceSetInterShapePenaltyINTEL = 5716, + OpSubgroupAvcMceGetDefaultInterDirectionPenaltyINTEL = 5717, + OpSubgroupAvcMceSetInterDirectionPenaltyINTEL = 5718, + OpSubgroupAvcMceGetDefaultIntraLumaShapePenaltyINTEL = 5719, + OpSubgroupAvcMceGetDefaultInterMotionVectorCostTableINTEL = 5720, + OpSubgroupAvcMceGetDefaultHighPenaltyCostTableINTEL = 5721, + OpSubgroupAvcMceGetDefaultMediumPenaltyCostTableINTEL = 5722, + OpSubgroupAvcMceGetDefaultLowPenaltyCostTableINTEL = 5723, + OpSubgroupAvcMceSetMotionVectorCostFunctionINTEL = 5724, + OpSubgroupAvcMceGetDefaultIntraLumaModePenaltyINTEL = 5725, + OpSubgroupAvcMceGetDefaultNonDcLumaIntraPenaltyINTEL = 5726, + OpSubgroupAvcMceGetDefaultIntraChromaModeBasePenaltyINTEL = 5727, + OpSubgroupAvcMceSetAcOnlyHaarINTEL = 5728, + OpSubgroupAvcMceSetSourceInterlacedFieldPolarityINTEL = 5729, + OpSubgroupAvcMceSetSingleReferenceInterlacedFieldPolarityINTEL = 5730, + OpSubgroupAvcMceSetDualReferenceInterlacedFieldPolaritiesINTEL = 5731, + OpSubgroupAvcMceConvertToImePayloadINTEL = 5732, + OpSubgroupAvcMceConvertToImeResultINTEL = 5733, + OpSubgroupAvcMceConvertToRefPayloadINTEL = 5734, + OpSubgroupAvcMceConvertToRefResultINTEL = 5735, + OpSubgroupAvcMceConvertToSicPayloadINTEL = 5736, + OpSubgroupAvcMceConvertToSicResultINTEL = 5737, + OpSubgroupAvcMceGetMotionVectorsINTEL = 5738, + OpSubgroupAvcMceGetInterDistortionsINTEL = 5739, + OpSubgroupAvcMceGetBestInterDistortionsINTEL = 5740, + OpSubgroupAvcMceGetInterMajorShapeINTEL = 5741, + OpSubgroupAvcMceGetInterMinorShapeINTEL = 5742, + OpSubgroupAvcMceGetInterDirectionsINTEL = 5743, + OpSubgroupAvcMceGetInterMotionVectorCountINTEL = 5744, + OpSubgroupAvcMceGetInterReferenceIdsINTEL = 5745, + OpSubgroupAvcMceGetInterReferenceInterlacedFieldPolaritiesINTEL = 5746, + OpSubgroupAvcImeInitializeINTEL = 5747, + OpSubgroupAvcImeSetSingleReferenceINTEL = 5748, + OpSubgroupAvcImeSetDualReferenceINTEL = 5749, + OpSubgroupAvcImeRefWindowSizeINTEL = 5750, + OpSubgroupAvcImeAdjustRefOffsetINTEL = 5751, + OpSubgroupAvcImeConvertToMcePayloadINTEL = 5752, + OpSubgroupAvcImeSetMaxMotionVectorCountINTEL = 5753, + OpSubgroupAvcImeSetUnidirectionalMixDisableINTEL = 5754, + OpSubgroupAvcImeSetEarlySearchTerminationThresholdINTEL = 5755, + OpSubgroupAvcImeSetWeightedSadINTEL = 5756, + OpSubgroupAvcImeEvaluateWithSingleReferenceINTEL = 5757, + OpSubgroupAvcImeEvaluateWithDualReferenceINTEL = 5758, + OpSubgroupAvcImeEvaluateWithSingleReferenceStreaminINTEL = 5759, + OpSubgroupAvcImeEvaluateWithDualReferenceStreaminINTEL = 5760, + OpSubgroupAvcImeEvaluateWithSingleReferenceStreamoutINTEL = 5761, + OpSubgroupAvcImeEvaluateWithDualReferenceStreamoutINTEL = 5762, + OpSubgroupAvcImeEvaluateWithSingleReferenceStreaminoutINTEL = 5763, + OpSubgroupAvcImeEvaluateWithDualReferenceStreaminoutINTEL = 5764, + OpSubgroupAvcImeConvertToMceResultINTEL = 5765, + OpSubgroupAvcImeGetSingleReferenceStreaminINTEL = 5766, + OpSubgroupAvcImeGetDualReferenceStreaminINTEL = 5767, + OpSubgroupAvcImeStripSingleReferenceStreamoutINTEL = 5768, + OpSubgroupAvcImeStripDualReferenceStreamoutINTEL = 5769, + OpSubgroupAvcImeGetStreamoutSingleReferenceMajorShapeMotionVectorsINTEL = 5770, + OpSubgroupAvcImeGetStreamoutSingleReferenceMajorShapeDistortionsINTEL = 5771, + OpSubgroupAvcImeGetStreamoutSingleReferenceMajorShapeReferenceIdsINTEL = 5772, + OpSubgroupAvcImeGetStreamoutDualReferenceMajorShapeMotionVectorsINTEL = 5773, + OpSubgroupAvcImeGetStreamoutDualReferenceMajorShapeDistortionsINTEL = 5774, + OpSubgroupAvcImeGetStreamoutDualReferenceMajorShapeReferenceIdsINTEL = 5775, + OpSubgroupAvcImeGetBorderReachedINTEL = 5776, + OpSubgroupAvcImeGetTruncatedSearchIndicationINTEL = 5777, + OpSubgroupAvcImeGetUnidirectionalEarlySearchTerminationINTEL = 5778, + OpSubgroupAvcImeGetWeightingPatternMinimumMotionVectorINTEL = 5779, + OpSubgroupAvcImeGetWeightingPatternMinimumDistortionINTEL = 5780, + OpSubgroupAvcFmeInitializeINTEL = 5781, + OpSubgroupAvcBmeInitializeINTEL = 5782, + OpSubgroupAvcRefConvertToMcePayloadINTEL = 5783, + OpSubgroupAvcRefSetBidirectionalMixDisableINTEL = 5784, + OpSubgroupAvcRefSetBilinearFilterEnableINTEL = 5785, + OpSubgroupAvcRefEvaluateWithSingleReferenceINTEL = 5786, + OpSubgroupAvcRefEvaluateWithDualReferenceINTEL = 5787, + OpSubgroupAvcRefEvaluateWithMultiReferenceINTEL = 5788, + OpSubgroupAvcRefEvaluateWithMultiReferenceInterlacedINTEL = 5789, + OpSubgroupAvcRefConvertToMceResultINTEL = 5790, + OpSubgroupAvcSicInitializeINTEL = 5791, + OpSubgroupAvcSicConfigureSkcINTEL = 5792, + OpSubgroupAvcSicConfigureIpeLumaINTEL = 5793, + OpSubgroupAvcSicConfigureIpeLumaChromaINTEL = 5794, + OpSubgroupAvcSicGetMotionVectorMaskINTEL = 5795, + OpSubgroupAvcSicConvertToMcePayloadINTEL = 5796, + OpSubgroupAvcSicSetIntraLumaShapePenaltyINTEL = 5797, + OpSubgroupAvcSicSetIntraLumaModeCostFunctionINTEL = 5798, + OpSubgroupAvcSicSetIntraChromaModeCostFunctionINTEL = 5799, + OpSubgroupAvcSicSetBilinearFilterEnableINTEL = 5800, + OpSubgroupAvcSicSetSkcForwardTransformEnableINTEL = 5801, + OpSubgroupAvcSicSetBlockBasedRawSkipSadINTEL = 5802, + OpSubgroupAvcSicEvaluateIpeINTEL = 5803, + OpSubgroupAvcSicEvaluateWithSingleReferenceINTEL = 5804, + OpSubgroupAvcSicEvaluateWithDualReferenceINTEL = 5805, + OpSubgroupAvcSicEvaluateWithMultiReferenceINTEL = 5806, + OpSubgroupAvcSicEvaluateWithMultiReferenceInterlacedINTEL = 5807, + OpSubgroupAvcSicConvertToMceResultINTEL = 5808, + OpSubgroupAvcSicGetIpeLumaShapeINTEL = 5809, + OpSubgroupAvcSicGetBestIpeLumaDistortionINTEL = 5810, + OpSubgroupAvcSicGetBestIpeChromaDistortionINTEL = 5811, + OpSubgroupAvcSicGetPackedIpeLumaModesINTEL = 5812, + OpSubgroupAvcSicGetIpeChromaModeINTEL = 5813, + OpSubgroupAvcSicGetPackedSkcLumaCountThresholdINTEL = 5814, + OpSubgroupAvcSicGetPackedSkcLumaSumThresholdINTEL = 5815, + OpSubgroupAvcSicGetInterRawSadsINTEL = 5816, + OpMax = 0x7fffffff, +}; + +#ifdef SPV_ENABLE_UTILITY_CODE +inline void HasResultAndType(Op opcode, bool *hasResult, bool *hasResultType) { + *hasResult = *hasResultType = false; + switch (opcode) { + default: /* unknown opcode */ break; + case OpNop: *hasResult = false; *hasResultType = false; break; + case OpUndef: *hasResult = true; *hasResultType = true; break; + case OpSourceContinued: *hasResult = false; *hasResultType = false; break; + case OpSource: *hasResult = false; *hasResultType = false; break; + case OpSourceExtension: *hasResult = false; *hasResultType = false; break; + case OpName: *hasResult = false; *hasResultType = false; break; + case OpMemberName: *hasResult = false; *hasResultType = false; break; + case OpString: *hasResult = true; *hasResultType = false; break; + case OpLine: *hasResult = false; *hasResultType = false; break; + case OpExtension: *hasResult = false; *hasResultType = false; break; + case OpExtInstImport: *hasResult = true; *hasResultType = false; break; + case OpExtInst: *hasResult = true; *hasResultType = true; break; + case OpMemoryModel: *hasResult = false; *hasResultType = false; break; + case OpEntryPoint: *hasResult = false; *hasResultType = false; break; + case OpExecutionMode: *hasResult = false; *hasResultType = false; break; + case OpCapability: *hasResult = false; *hasResultType = false; break; + case OpTypeVoid: *hasResult = true; *hasResultType = false; break; + case OpTypeBool: *hasResult = true; *hasResultType = false; break; + case OpTypeInt: *hasResult = true; *hasResultType = false; break; + case OpTypeFloat: *hasResult = true; *hasResultType = false; break; + case OpTypeVector: *hasResult = true; *hasResultType = false; break; + case OpTypeMatrix: *hasResult = true; *hasResultType = false; break; + case OpTypeImage: *hasResult = true; *hasResultType = false; break; + case OpTypeSampler: *hasResult = true; *hasResultType = false; break; + case OpTypeSampledImage: *hasResult = true; *hasResultType = false; break; + case OpTypeArray: *hasResult = true; *hasResultType = false; break; + case OpTypeRuntimeArray: *hasResult = true; *hasResultType = false; break; + case OpTypeStruct: *hasResult = true; *hasResultType = false; break; + case OpTypeOpaque: *hasResult = true; *hasResultType = false; break; + case OpTypePointer: *hasResult = true; *hasResultType = false; break; + case OpTypeFunction: *hasResult = true; *hasResultType = false; break; + case OpTypeEvent: *hasResult = true; *hasResultType = false; break; + case OpTypeDeviceEvent: *hasResult = true; *hasResultType = false; break; + case OpTypeReserveId: *hasResult = true; *hasResultType = false; break; + case OpTypeQueue: *hasResult = true; *hasResultType = false; break; + case OpTypePipe: *hasResult = true; *hasResultType = false; break; + case OpTypeForwardPointer: *hasResult = false; *hasResultType = false; break; + case OpConstantTrue: *hasResult = true; *hasResultType = true; break; + case OpConstantFalse: *hasResult = true; *hasResultType = true; break; + case OpConstant: *hasResult = true; *hasResultType = true; break; + case OpConstantComposite: *hasResult = true; *hasResultType = true; break; + case OpConstantSampler: *hasResult = true; *hasResultType = true; break; + case OpConstantNull: *hasResult = true; *hasResultType = true; break; + case OpSpecConstantTrue: *hasResult = true; *hasResultType = true; break; + case OpSpecConstantFalse: *hasResult = true; *hasResultType = true; break; + case OpSpecConstant: *hasResult = true; *hasResultType = true; break; + case OpSpecConstantComposite: *hasResult = true; *hasResultType = true; break; + case OpSpecConstantOp: *hasResult = true; *hasResultType = true; break; + case OpFunction: *hasResult = true; *hasResultType = true; break; + case OpFunctionParameter: *hasResult = true; *hasResultType = true; break; + case OpFunctionEnd: *hasResult = false; *hasResultType = false; break; + case OpFunctionCall: *hasResult = true; *hasResultType = true; break; + case OpVariable: *hasResult = true; *hasResultType = true; break; + case OpImageTexelPointer: *hasResult = true; *hasResultType = true; break; + case OpLoad: *hasResult = true; *hasResultType = true; break; + case OpStore: *hasResult = false; *hasResultType = false; break; + case OpCopyMemory: *hasResult = false; *hasResultType = false; break; + case OpCopyMemorySized: *hasResult = false; *hasResultType = false; break; + case OpAccessChain: *hasResult = true; *hasResultType = true; break; + case OpInBoundsAccessChain: *hasResult = true; *hasResultType = true; break; + case OpPtrAccessChain: *hasResult = true; *hasResultType = true; break; + case OpArrayLength: *hasResult = true; *hasResultType = true; break; + case OpGenericPtrMemSemantics: *hasResult = true; *hasResultType = true; break; + case OpInBoundsPtrAccessChain: *hasResult = true; *hasResultType = true; break; + case OpDecorate: *hasResult = false; *hasResultType = false; break; + case OpMemberDecorate: *hasResult = false; *hasResultType = false; break; + case OpDecorationGroup: *hasResult = true; *hasResultType = false; break; + case OpGroupDecorate: *hasResult = false; *hasResultType = false; break; + case OpGroupMemberDecorate: *hasResult = false; *hasResultType = false; break; + case OpVectorExtractDynamic: *hasResult = true; *hasResultType = true; break; + case OpVectorInsertDynamic: *hasResult = true; *hasResultType = true; break; + case OpVectorShuffle: *hasResult = true; *hasResultType = true; break; + case OpCompositeConstruct: *hasResult = true; *hasResultType = true; break; + case OpCompositeExtract: *hasResult = true; *hasResultType = true; break; + case OpCompositeInsert: *hasResult = true; *hasResultType = true; break; + case OpCopyObject: *hasResult = true; *hasResultType = true; break; + case OpTranspose: *hasResult = true; *hasResultType = true; break; + case OpSampledImage: *hasResult = true; *hasResultType = true; break; + case OpImageSampleImplicitLod: *hasResult = true; *hasResultType = true; break; + case OpImageSampleExplicitLod: *hasResult = true; *hasResultType = true; break; + case OpImageSampleDrefImplicitLod: *hasResult = true; *hasResultType = true; break; + case OpImageSampleDrefExplicitLod: *hasResult = true; *hasResultType = true; break; + case OpImageSampleProjImplicitLod: *hasResult = true; *hasResultType = true; break; + case OpImageSampleProjExplicitLod: *hasResult = true; *hasResultType = true; break; + case OpImageSampleProjDrefImplicitLod: *hasResult = true; *hasResultType = true; break; + case OpImageSampleProjDrefExplicitLod: *hasResult = true; *hasResultType = true; break; + case OpImageFetch: *hasResult = true; *hasResultType = true; break; + case OpImageGather: *hasResult = true; *hasResultType = true; break; + case OpImageDrefGather: *hasResult = true; *hasResultType = true; break; + case OpImageRead: *hasResult = true; *hasResultType = true; break; + case OpImageWrite: *hasResult = false; *hasResultType = false; break; + case OpImage: *hasResult = true; *hasResultType = true; break; + case OpImageQueryFormat: *hasResult = true; *hasResultType = true; break; + case OpImageQueryOrder: *hasResult = true; *hasResultType = true; break; + case OpImageQuerySizeLod: *hasResult = true; *hasResultType = true; break; + case OpImageQuerySize: *hasResult = true; *hasResultType = true; break; + case OpImageQueryLod: *hasResult = true; *hasResultType = true; break; + case OpImageQueryLevels: *hasResult = true; *hasResultType = true; break; + case OpImageQuerySamples: *hasResult = true; *hasResultType = true; break; + case OpConvertFToU: *hasResult = true; *hasResultType = true; break; + case OpConvertFToS: *hasResult = true; *hasResultType = true; break; + case OpConvertSToF: *hasResult = true; *hasResultType = true; break; + case OpConvertUToF: *hasResult = true; *hasResultType = true; break; + case OpUConvert: *hasResult = true; *hasResultType = true; break; + case OpSConvert: *hasResult = true; *hasResultType = true; break; + case OpFConvert: *hasResult = true; *hasResultType = true; break; + case OpQuantizeToF16: *hasResult = true; *hasResultType = true; break; + case OpConvertPtrToU: *hasResult = true; *hasResultType = true; break; + case OpSatConvertSToU: *hasResult = true; *hasResultType = true; break; + case OpSatConvertUToS: *hasResult = true; *hasResultType = true; break; + case OpConvertUToPtr: *hasResult = true; *hasResultType = true; break; + case OpPtrCastToGeneric: *hasResult = true; *hasResultType = true; break; + case OpGenericCastToPtr: *hasResult = true; *hasResultType = true; break; + case OpGenericCastToPtrExplicit: *hasResult = true; *hasResultType = true; break; + case OpBitcast: *hasResult = true; *hasResultType = true; break; + case OpSNegate: *hasResult = true; *hasResultType = true; break; + case OpFNegate: *hasResult = true; *hasResultType = true; break; + case OpIAdd: *hasResult = true; *hasResultType = true; break; + case OpFAdd: *hasResult = true; *hasResultType = true; break; + case OpISub: *hasResult = true; *hasResultType = true; break; + case OpFSub: *hasResult = true; *hasResultType = true; break; + case OpIMul: *hasResult = true; *hasResultType = true; break; + case OpFMul: *hasResult = true; *hasResultType = true; break; + case OpUDiv: *hasResult = true; *hasResultType = true; break; + case OpSDiv: *hasResult = true; *hasResultType = true; break; + case OpFDiv: *hasResult = true; *hasResultType = true; break; + case OpUMod: *hasResult = true; *hasResultType = true; break; + case OpSRem: *hasResult = true; *hasResultType = true; break; + case OpSMod: *hasResult = true; *hasResultType = true; break; + case OpFRem: *hasResult = true; *hasResultType = true; break; + case OpFMod: *hasResult = true; *hasResultType = true; break; + case OpVectorTimesScalar: *hasResult = true; *hasResultType = true; break; + case OpMatrixTimesScalar: *hasResult = true; *hasResultType = true; break; + case OpVectorTimesMatrix: *hasResult = true; *hasResultType = true; break; + case OpMatrixTimesVector: *hasResult = true; *hasResultType = true; break; + case OpMatrixTimesMatrix: *hasResult = true; *hasResultType = true; break; + case OpOuterProduct: *hasResult = true; *hasResultType = true; break; + case OpDot: *hasResult = true; *hasResultType = true; break; + case OpIAddCarry: *hasResult = true; *hasResultType = true; break; + case OpISubBorrow: *hasResult = true; *hasResultType = true; break; + case OpUMulExtended: *hasResult = true; *hasResultType = true; break; + case OpSMulExtended: *hasResult = true; *hasResultType = true; break; + case OpAny: *hasResult = true; *hasResultType = true; break; + case OpAll: *hasResult = true; *hasResultType = true; break; + case OpIsNan: *hasResult = true; *hasResultType = true; break; + case OpIsInf: *hasResult = true; *hasResultType = true; break; + case OpIsFinite: *hasResult = true; *hasResultType = true; break; + case OpIsNormal: *hasResult = true; *hasResultType = true; break; + case OpSignBitSet: *hasResult = true; *hasResultType = true; break; + case OpLessOrGreater: *hasResult = true; *hasResultType = true; break; + case OpOrdered: *hasResult = true; *hasResultType = true; break; + case OpUnordered: *hasResult = true; *hasResultType = true; break; + case OpLogicalEqual: *hasResult = true; *hasResultType = true; break; + case OpLogicalNotEqual: *hasResult = true; *hasResultType = true; break; + case OpLogicalOr: *hasResult = true; *hasResultType = true; break; + case OpLogicalAnd: *hasResult = true; *hasResultType = true; break; + case OpLogicalNot: *hasResult = true; *hasResultType = true; break; + case OpSelect: *hasResult = true; *hasResultType = true; break; + case OpIEqual: *hasResult = true; *hasResultType = true; break; + case OpINotEqual: *hasResult = true; *hasResultType = true; break; + case OpUGreaterThan: *hasResult = true; *hasResultType = true; break; + case OpSGreaterThan: *hasResult = true; *hasResultType = true; break; + case OpUGreaterThanEqual: *hasResult = true; *hasResultType = true; break; + case OpSGreaterThanEqual: *hasResult = true; *hasResultType = true; break; + case OpULessThan: *hasResult = true; *hasResultType = true; break; + case OpSLessThan: *hasResult = true; *hasResultType = true; break; + case OpULessThanEqual: *hasResult = true; *hasResultType = true; break; + case OpSLessThanEqual: *hasResult = true; *hasResultType = true; break; + case OpFOrdEqual: *hasResult = true; *hasResultType = true; break; + case OpFUnordEqual: *hasResult = true; *hasResultType = true; break; + case OpFOrdNotEqual: *hasResult = true; *hasResultType = true; break; + case OpFUnordNotEqual: *hasResult = true; *hasResultType = true; break; + case OpFOrdLessThan: *hasResult = true; *hasResultType = true; break; + case OpFUnordLessThan: *hasResult = true; *hasResultType = true; break; + case OpFOrdGreaterThan: *hasResult = true; *hasResultType = true; break; + case OpFUnordGreaterThan: *hasResult = true; *hasResultType = true; break; + case OpFOrdLessThanEqual: *hasResult = true; *hasResultType = true; break; + case OpFUnordLessThanEqual: *hasResult = true; *hasResultType = true; break; + case OpFOrdGreaterThanEqual: *hasResult = true; *hasResultType = true; break; + case OpFUnordGreaterThanEqual: *hasResult = true; *hasResultType = true; break; + case OpShiftRightLogical: *hasResult = true; *hasResultType = true; break; + case OpShiftRightArithmetic: *hasResult = true; *hasResultType = true; break; + case OpShiftLeftLogical: *hasResult = true; *hasResultType = true; break; + case OpBitwiseOr: *hasResult = true; *hasResultType = true; break; + case OpBitwiseXor: *hasResult = true; *hasResultType = true; break; + case OpBitwiseAnd: *hasResult = true; *hasResultType = true; break; + case OpNot: *hasResult = true; *hasResultType = true; break; + case OpBitFieldInsert: *hasResult = true; *hasResultType = true; break; + case OpBitFieldSExtract: *hasResult = true; *hasResultType = true; break; + case OpBitFieldUExtract: *hasResult = true; *hasResultType = true; break; + case OpBitReverse: *hasResult = true; *hasResultType = true; break; + case OpBitCount: *hasResult = true; *hasResultType = true; break; + case OpDPdx: *hasResult = true; *hasResultType = true; break; + case OpDPdy: *hasResult = true; *hasResultType = true; break; + case OpFwidth: *hasResult = true; *hasResultType = true; break; + case OpDPdxFine: *hasResult = true; *hasResultType = true; break; + case OpDPdyFine: *hasResult = true; *hasResultType = true; break; + case OpFwidthFine: *hasResult = true; *hasResultType = true; break; + case OpDPdxCoarse: *hasResult = true; *hasResultType = true; break; + case OpDPdyCoarse: *hasResult = true; *hasResultType = true; break; + case OpFwidthCoarse: *hasResult = true; *hasResultType = true; break; + case OpEmitVertex: *hasResult = false; *hasResultType = false; break; + case OpEndPrimitive: *hasResult = false; *hasResultType = false; break; + case OpEmitStreamVertex: *hasResult = false; *hasResultType = false; break; + case OpEndStreamPrimitive: *hasResult = false; *hasResultType = false; break; + case OpControlBarrier: *hasResult = false; *hasResultType = false; break; + case OpMemoryBarrier: *hasResult = false; *hasResultType = false; break; + case OpAtomicLoad: *hasResult = true; *hasResultType = true; break; + case OpAtomicStore: *hasResult = false; *hasResultType = false; break; + case OpAtomicExchange: *hasResult = true; *hasResultType = true; break; + case OpAtomicCompareExchange: *hasResult = true; *hasResultType = true; break; + case OpAtomicCompareExchangeWeak: *hasResult = true; *hasResultType = true; break; + case OpAtomicIIncrement: *hasResult = true; *hasResultType = true; break; + case OpAtomicIDecrement: *hasResult = true; *hasResultType = true; break; + case OpAtomicIAdd: *hasResult = true; *hasResultType = true; break; + case OpAtomicISub: *hasResult = true; *hasResultType = true; break; + case OpAtomicSMin: *hasResult = true; *hasResultType = true; break; + case OpAtomicUMin: *hasResult = true; *hasResultType = true; break; + case OpAtomicSMax: *hasResult = true; *hasResultType = true; break; + case OpAtomicUMax: *hasResult = true; *hasResultType = true; break; + case OpAtomicAnd: *hasResult = true; *hasResultType = true; break; + case OpAtomicOr: *hasResult = true; *hasResultType = true; break; + case OpAtomicXor: *hasResult = true; *hasResultType = true; break; + case OpPhi: *hasResult = true; *hasResultType = true; break; + case OpLoopMerge: *hasResult = false; *hasResultType = false; break; + case OpSelectionMerge: *hasResult = false; *hasResultType = false; break; + case OpLabel: *hasResult = true; *hasResultType = false; break; + case OpBranch: *hasResult = false; *hasResultType = false; break; + case OpBranchConditional: *hasResult = false; *hasResultType = false; break; + case OpSwitch: *hasResult = false; *hasResultType = false; break; + case OpKill: *hasResult = false; *hasResultType = false; break; + case OpReturn: *hasResult = false; *hasResultType = false; break; + case OpReturnValue: *hasResult = false; *hasResultType = false; break; + case OpUnreachable: *hasResult = false; *hasResultType = false; break; + case OpLifetimeStart: *hasResult = false; *hasResultType = false; break; + case OpLifetimeStop: *hasResult = false; *hasResultType = false; break; + case OpGroupAsyncCopy: *hasResult = true; *hasResultType = true; break; + case OpGroupWaitEvents: *hasResult = false; *hasResultType = false; break; + case OpGroupAll: *hasResult = true; *hasResultType = true; break; + case OpGroupAny: *hasResult = true; *hasResultType = true; break; + case OpGroupBroadcast: *hasResult = true; *hasResultType = true; break; + case OpGroupIAdd: *hasResult = true; *hasResultType = true; break; + case OpGroupFAdd: *hasResult = true; *hasResultType = true; break; + case OpGroupFMin: *hasResult = true; *hasResultType = true; break; + case OpGroupUMin: *hasResult = true; *hasResultType = true; break; + case OpGroupSMin: *hasResult = true; *hasResultType = true; break; + case OpGroupFMax: *hasResult = true; *hasResultType = true; break; + case OpGroupUMax: *hasResult = true; *hasResultType = true; break; + case OpGroupSMax: *hasResult = true; *hasResultType = true; break; + case OpReadPipe: *hasResult = true; *hasResultType = true; break; + case OpWritePipe: *hasResult = true; *hasResultType = true; break; + case OpReservedReadPipe: *hasResult = true; *hasResultType = true; break; + case OpReservedWritePipe: *hasResult = true; *hasResultType = true; break; + case OpReserveReadPipePackets: *hasResult = true; *hasResultType = true; break; + case OpReserveWritePipePackets: *hasResult = true; *hasResultType = true; break; + case OpCommitReadPipe: *hasResult = false; *hasResultType = false; break; + case OpCommitWritePipe: *hasResult = false; *hasResultType = false; break; + case OpIsValidReserveId: *hasResult = true; *hasResultType = true; break; + case OpGetNumPipePackets: *hasResult = true; *hasResultType = true; break; + case OpGetMaxPipePackets: *hasResult = true; *hasResultType = true; break; + case OpGroupReserveReadPipePackets: *hasResult = true; *hasResultType = true; break; + case OpGroupReserveWritePipePackets: *hasResult = true; *hasResultType = true; break; + case OpGroupCommitReadPipe: *hasResult = false; *hasResultType = false; break; + case OpGroupCommitWritePipe: *hasResult = false; *hasResultType = false; break; + case OpEnqueueMarker: *hasResult = true; *hasResultType = true; break; + case OpEnqueueKernel: *hasResult = true; *hasResultType = true; break; + case OpGetKernelNDrangeSubGroupCount: *hasResult = true; *hasResultType = true; break; + case OpGetKernelNDrangeMaxSubGroupSize: *hasResult = true; *hasResultType = true; break; + case OpGetKernelWorkGroupSize: *hasResult = true; *hasResultType = true; break; + case OpGetKernelPreferredWorkGroupSizeMultiple: *hasResult = true; *hasResultType = true; break; + case OpRetainEvent: *hasResult = false; *hasResultType = false; break; + case OpReleaseEvent: *hasResult = false; *hasResultType = false; break; + case OpCreateUserEvent: *hasResult = true; *hasResultType = true; break; + case OpIsValidEvent: *hasResult = true; *hasResultType = true; break; + case OpSetUserEventStatus: *hasResult = false; *hasResultType = false; break; + case OpCaptureEventProfilingInfo: *hasResult = false; *hasResultType = false; break; + case OpGetDefaultQueue: *hasResult = true; *hasResultType = true; break; + case OpBuildNDRange: *hasResult = true; *hasResultType = true; break; + case OpImageSparseSampleImplicitLod: *hasResult = true; *hasResultType = true; break; + case OpImageSparseSampleExplicitLod: *hasResult = true; *hasResultType = true; break; + case OpImageSparseSampleDrefImplicitLod: *hasResult = true; *hasResultType = true; break; + case OpImageSparseSampleDrefExplicitLod: *hasResult = true; *hasResultType = true; break; + case OpImageSparseSampleProjImplicitLod: *hasResult = true; *hasResultType = true; break; + case OpImageSparseSampleProjExplicitLod: *hasResult = true; *hasResultType = true; break; + case OpImageSparseSampleProjDrefImplicitLod: *hasResult = true; *hasResultType = true; break; + case OpImageSparseSampleProjDrefExplicitLod: *hasResult = true; *hasResultType = true; break; + case OpImageSparseFetch: *hasResult = true; *hasResultType = true; break; + case OpImageSparseGather: *hasResult = true; *hasResultType = true; break; + case OpImageSparseDrefGather: *hasResult = true; *hasResultType = true; break; + case OpImageSparseTexelsResident: *hasResult = true; *hasResultType = true; break; + case OpNoLine: *hasResult = false; *hasResultType = false; break; + case OpAtomicFlagTestAndSet: *hasResult = true; *hasResultType = true; break; + case OpAtomicFlagClear: *hasResult = false; *hasResultType = false; break; + case OpImageSparseRead: *hasResult = true; *hasResultType = true; break; + case OpSizeOf: *hasResult = true; *hasResultType = true; break; + case OpTypePipeStorage: *hasResult = true; *hasResultType = false; break; + case OpConstantPipeStorage: *hasResult = true; *hasResultType = true; break; + case OpCreatePipeFromPipeStorage: *hasResult = true; *hasResultType = true; break; + case OpGetKernelLocalSizeForSubgroupCount: *hasResult = true; *hasResultType = true; break; + case OpGetKernelMaxNumSubgroups: *hasResult = true; *hasResultType = true; break; + case OpTypeNamedBarrier: *hasResult = true; *hasResultType = false; break; + case OpNamedBarrierInitialize: *hasResult = true; *hasResultType = true; break; + case OpMemoryNamedBarrier: *hasResult = false; *hasResultType = false; break; + case OpModuleProcessed: *hasResult = false; *hasResultType = false; break; + case OpExecutionModeId: *hasResult = false; *hasResultType = false; break; + case OpDecorateId: *hasResult = false; *hasResultType = false; break; + case OpGroupNonUniformElect: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformAll: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformAny: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformAllEqual: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformBroadcast: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformBroadcastFirst: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformBallot: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformInverseBallot: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformBallotBitExtract: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformBallotBitCount: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformBallotFindLSB: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformBallotFindMSB: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformShuffle: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformShuffleXor: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformShuffleUp: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformShuffleDown: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformIAdd: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformFAdd: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformIMul: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformFMul: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformSMin: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformUMin: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformFMin: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformSMax: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformUMax: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformFMax: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformBitwiseAnd: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformBitwiseOr: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformBitwiseXor: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformLogicalAnd: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformLogicalOr: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformLogicalXor: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformQuadBroadcast: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformQuadSwap: *hasResult = true; *hasResultType = true; break; + case OpCopyLogical: *hasResult = true; *hasResultType = true; break; + case OpPtrEqual: *hasResult = true; *hasResultType = true; break; + case OpPtrNotEqual: *hasResult = true; *hasResultType = true; break; + case OpPtrDiff: *hasResult = true; *hasResultType = true; break; + case OpSubgroupBallotKHR: *hasResult = true; *hasResultType = true; break; + case OpSubgroupFirstInvocationKHR: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAllKHR: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAnyKHR: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAllEqualKHR: *hasResult = true; *hasResultType = true; break; + case OpSubgroupReadInvocationKHR: *hasResult = true; *hasResultType = true; break; + case OpGroupIAddNonUniformAMD: *hasResult = true; *hasResultType = true; break; + case OpGroupFAddNonUniformAMD: *hasResult = true; *hasResultType = true; break; + case OpGroupFMinNonUniformAMD: *hasResult = true; *hasResultType = true; break; + case OpGroupUMinNonUniformAMD: *hasResult = true; *hasResultType = true; break; + case OpGroupSMinNonUniformAMD: *hasResult = true; *hasResultType = true; break; + case OpGroupFMaxNonUniformAMD: *hasResult = true; *hasResultType = true; break; + case OpGroupUMaxNonUniformAMD: *hasResult = true; *hasResultType = true; break; + case OpGroupSMaxNonUniformAMD: *hasResult = true; *hasResultType = true; break; + case OpFragmentMaskFetchAMD: *hasResult = true; *hasResultType = true; break; + case OpFragmentFetchAMD: *hasResult = true; *hasResultType = true; break; + case OpImageSampleFootprintNV: *hasResult = true; *hasResultType = true; break; + case OpGroupNonUniformPartitionNV: *hasResult = true; *hasResultType = true; break; + case OpWritePackedPrimitiveIndices4x8NV: *hasResult = false; *hasResultType = false; break; + case OpReportIntersectionNV: *hasResult = true; *hasResultType = true; break; + case OpIgnoreIntersectionNV: *hasResult = false; *hasResultType = false; break; + case OpTerminateRayNV: *hasResult = false; *hasResultType = false; break; + case OpTraceNV: *hasResult = false; *hasResultType = false; break; + case OpTypeAccelerationStructureNV: *hasResult = true; *hasResultType = false; break; + case OpExecuteCallableNV: *hasResult = false; *hasResultType = false; break; + case OpTypeCooperativeMatrixNV: *hasResult = true; *hasResultType = false; break; + case OpCooperativeMatrixLoadNV: *hasResult = true; *hasResultType = true; break; + case OpCooperativeMatrixStoreNV: *hasResult = false; *hasResultType = false; break; + case OpCooperativeMatrixMulAddNV: *hasResult = true; *hasResultType = true; break; + case OpCooperativeMatrixLengthNV: *hasResult = true; *hasResultType = true; break; + case OpSubgroupShuffleINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupShuffleDownINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupShuffleUpINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupShuffleXorINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupBlockReadINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupBlockWriteINTEL: *hasResult = false; *hasResultType = false; break; + case OpSubgroupImageBlockReadINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupImageBlockWriteINTEL: *hasResult = false; *hasResultType = false; break; + case OpSubgroupImageMediaBlockReadINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupImageMediaBlockWriteINTEL: *hasResult = false; *hasResultType = false; break; + case OpDecorateString: *hasResult = false; *hasResultType = false; break; + case OpDecorateStringGOOGLE: *hasResult = false; *hasResultType = false; break; + case OpMemberDecorateString: *hasResult = false; *hasResultType = false; break; + case OpMemberDecorateStringGOOGLE: *hasResult = false; *hasResultType = false; break; + case OpVmeImageINTEL: *hasResult = true; *hasResultType = true; break; + case OpTypeVmeImageINTEL: *hasResult = true; *hasResultType = false; break; + case OpTypeAvcImePayloadINTEL: *hasResult = true; *hasResultType = false; break; + case OpTypeAvcRefPayloadINTEL: *hasResult = true; *hasResultType = false; break; + case OpTypeAvcSicPayloadINTEL: *hasResult = true; *hasResultType = false; break; + case OpTypeAvcMcePayloadINTEL: *hasResult = true; *hasResultType = false; break; + case OpTypeAvcMceResultINTEL: *hasResult = true; *hasResultType = false; break; + case OpTypeAvcImeResultINTEL: *hasResult = true; *hasResultType = false; break; + case OpTypeAvcImeResultSingleReferenceStreamoutINTEL: *hasResult = true; *hasResultType = false; break; + case OpTypeAvcImeResultDualReferenceStreamoutINTEL: *hasResult = true; *hasResultType = false; break; + case OpTypeAvcImeSingleReferenceStreaminINTEL: *hasResult = true; *hasResultType = false; break; + case OpTypeAvcImeDualReferenceStreaminINTEL: *hasResult = true; *hasResultType = false; break; + case OpTypeAvcRefResultINTEL: *hasResult = true; *hasResultType = false; break; + case OpTypeAvcSicResultINTEL: *hasResult = true; *hasResultType = false; break; + case OpSubgroupAvcMceGetDefaultInterBaseMultiReferencePenaltyINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceSetInterBaseMultiReferencePenaltyINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceGetDefaultInterShapePenaltyINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceSetInterShapePenaltyINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceGetDefaultInterDirectionPenaltyINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceSetInterDirectionPenaltyINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceGetDefaultIntraLumaShapePenaltyINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceGetDefaultInterMotionVectorCostTableINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceGetDefaultHighPenaltyCostTableINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceGetDefaultMediumPenaltyCostTableINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceGetDefaultLowPenaltyCostTableINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceSetMotionVectorCostFunctionINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceGetDefaultIntraLumaModePenaltyINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceGetDefaultNonDcLumaIntraPenaltyINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceGetDefaultIntraChromaModeBasePenaltyINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceSetAcOnlyHaarINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceSetSourceInterlacedFieldPolarityINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceSetSingleReferenceInterlacedFieldPolarityINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceSetDualReferenceInterlacedFieldPolaritiesINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceConvertToImePayloadINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceConvertToImeResultINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceConvertToRefPayloadINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceConvertToRefResultINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceConvertToSicPayloadINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceConvertToSicResultINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceGetMotionVectorsINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceGetInterDistortionsINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceGetBestInterDistortionsINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceGetInterMajorShapeINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceGetInterMinorShapeINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceGetInterDirectionsINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceGetInterMotionVectorCountINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceGetInterReferenceIdsINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcMceGetInterReferenceInterlacedFieldPolaritiesINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeInitializeINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeSetSingleReferenceINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeSetDualReferenceINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeRefWindowSizeINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeAdjustRefOffsetINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeConvertToMcePayloadINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeSetMaxMotionVectorCountINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeSetUnidirectionalMixDisableINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeSetEarlySearchTerminationThresholdINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeSetWeightedSadINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeEvaluateWithSingleReferenceINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeEvaluateWithDualReferenceINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeEvaluateWithSingleReferenceStreaminINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeEvaluateWithDualReferenceStreaminINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeEvaluateWithSingleReferenceStreamoutINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeEvaluateWithDualReferenceStreamoutINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeEvaluateWithSingleReferenceStreaminoutINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeEvaluateWithDualReferenceStreaminoutINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeConvertToMceResultINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeGetSingleReferenceStreaminINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeGetDualReferenceStreaminINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeStripSingleReferenceStreamoutINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeStripDualReferenceStreamoutINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeGetStreamoutSingleReferenceMajorShapeMotionVectorsINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeGetStreamoutSingleReferenceMajorShapeDistortionsINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeGetStreamoutSingleReferenceMajorShapeReferenceIdsINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeGetStreamoutDualReferenceMajorShapeMotionVectorsINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeGetStreamoutDualReferenceMajorShapeDistortionsINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeGetStreamoutDualReferenceMajorShapeReferenceIdsINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeGetBorderReachedINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeGetTruncatedSearchIndicationINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeGetUnidirectionalEarlySearchTerminationINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeGetWeightingPatternMinimumMotionVectorINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcImeGetWeightingPatternMinimumDistortionINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcFmeInitializeINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcBmeInitializeINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcRefConvertToMcePayloadINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcRefSetBidirectionalMixDisableINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcRefSetBilinearFilterEnableINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcRefEvaluateWithSingleReferenceINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcRefEvaluateWithDualReferenceINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcRefEvaluateWithMultiReferenceINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcRefEvaluateWithMultiReferenceInterlacedINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcRefConvertToMceResultINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicInitializeINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicConfigureSkcINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicConfigureIpeLumaINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicConfigureIpeLumaChromaINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicGetMotionVectorMaskINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicConvertToMcePayloadINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicSetIntraLumaShapePenaltyINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicSetIntraLumaModeCostFunctionINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicSetIntraChromaModeCostFunctionINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicSetBilinearFilterEnableINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicSetSkcForwardTransformEnableINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicSetBlockBasedRawSkipSadINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicEvaluateIpeINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicEvaluateWithSingleReferenceINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicEvaluateWithDualReferenceINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicEvaluateWithMultiReferenceINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicEvaluateWithMultiReferenceInterlacedINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicConvertToMceResultINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicGetIpeLumaShapeINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicGetBestIpeLumaDistortionINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicGetBestIpeChromaDistortionINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicGetPackedIpeLumaModesINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicGetIpeChromaModeINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicGetPackedSkcLumaCountThresholdINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicGetPackedSkcLumaSumThresholdINTEL: *hasResult = true; *hasResultType = true; break; + case OpSubgroupAvcSicGetInterRawSadsINTEL: *hasResult = true; *hasResultType = true; break; + } +} +#endif /* SPV_ENABLE_UTILITY_CODE */ + +// Overload operator| for mask bit combining + +inline ImageOperandsMask operator|(ImageOperandsMask a, ImageOperandsMask b) { return ImageOperandsMask(unsigned(a) | unsigned(b)); } +inline FPFastMathModeMask operator|(FPFastMathModeMask a, FPFastMathModeMask b) { return FPFastMathModeMask(unsigned(a) | unsigned(b)); } +inline SelectionControlMask operator|(SelectionControlMask a, SelectionControlMask b) { return SelectionControlMask(unsigned(a) | unsigned(b)); } +inline LoopControlMask operator|(LoopControlMask a, LoopControlMask b) { return LoopControlMask(unsigned(a) | unsigned(b)); } +inline FunctionControlMask operator|(FunctionControlMask a, FunctionControlMask b) { return FunctionControlMask(unsigned(a) | unsigned(b)); } +inline MemorySemanticsMask operator|(MemorySemanticsMask a, MemorySemanticsMask b) { return MemorySemanticsMask(unsigned(a) | unsigned(b)); } +inline MemoryAccessMask operator|(MemoryAccessMask a, MemoryAccessMask b) { return MemoryAccessMask(unsigned(a) | unsigned(b)); } +inline KernelProfilingInfoMask operator|(KernelProfilingInfoMask a, KernelProfilingInfoMask b) { return KernelProfilingInfoMask(unsigned(a) | unsigned(b)); } + +} // end namespace spv + +#endif // #ifndef spirv_HPP + diff --git a/thirdparty/glslang/SPIRV/spvIR.h b/thirdparty/glslang/SPIRV/spvIR.h new file mode 100644 index 0000000000..b3cd0b0613 --- /dev/null +++ b/thirdparty/glslang/SPIRV/spvIR.h @@ -0,0 +1,441 @@ +// +// Copyright (C) 2014 LunarG, Inc. +// Copyright (C) 2015-2018 Google, Inc. +// +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions +// are met: +// +// Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// +// Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following +// disclaimer in the documentation and/or other materials provided +// with the distribution. +// +// Neither the name of 3Dlabs Inc. Ltd. nor the names of its +// contributors may be used to endorse or promote products derived +// from this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS +// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE +// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, +// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, +// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; +// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN +// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +// POSSIBILITY OF SUCH DAMAGE. + +// SPIRV-IR +// +// Simple in-memory representation (IR) of SPIRV. Just for holding +// Each function's CFG of blocks. Has this hierarchy: +// - Module, which is a list of +// - Function, which is a list of +// - Block, which is a list of +// - Instruction +// + +#pragma once +#ifndef spvIR_H +#define spvIR_H + +#include "spirv.hpp" + +#include <algorithm> +#include <cassert> +#include <functional> +#include <iostream> +#include <memory> +#include <vector> + +namespace spv { + +class Block; +class Function; +class Module; + +const Id NoResult = 0; +const Id NoType = 0; + +const Decoration NoPrecision = DecorationMax; + +#ifdef __GNUC__ +# define POTENTIALLY_UNUSED __attribute__((unused)) +#else +# define POTENTIALLY_UNUSED +#endif + +POTENTIALLY_UNUSED +const MemorySemanticsMask MemorySemanticsAllMemory = + (MemorySemanticsMask)(MemorySemanticsUniformMemoryMask | + MemorySemanticsWorkgroupMemoryMask | + MemorySemanticsAtomicCounterMemoryMask | + MemorySemanticsImageMemoryMask); + +struct IdImmediate { + bool isId; // true if word is an Id, false if word is an immediate + unsigned word; + IdImmediate(bool i, unsigned w) : isId(i), word(w) {} +}; + +// +// SPIR-V IR instruction. +// + +class Instruction { +public: + Instruction(Id resultId, Id typeId, Op opCode) : resultId(resultId), typeId(typeId), opCode(opCode), block(nullptr) { } + explicit Instruction(Op opCode) : resultId(NoResult), typeId(NoType), opCode(opCode), block(nullptr) { } + virtual ~Instruction() {} + void addIdOperand(Id id) { + operands.push_back(id); + idOperand.push_back(true); + } + void addImmediateOperand(unsigned int immediate) { + operands.push_back(immediate); + idOperand.push_back(false); + } + void setImmediateOperand(unsigned idx, unsigned int immediate) { + assert(!idOperand[idx]); + operands[idx] = immediate; + } + + void addStringOperand(const char* str) + { + unsigned int word; + char* wordString = (char*)&word; + char* wordPtr = wordString; + int charCount = 0; + char c; + do { + c = *(str++); + *(wordPtr++) = c; + ++charCount; + if (charCount == 4) { + addImmediateOperand(word); + wordPtr = wordString; + charCount = 0; + } + } while (c != 0); + + // deal with partial last word + if (charCount > 0) { + // pad with 0s + for (; charCount < 4; ++charCount) + *(wordPtr++) = 0; + addImmediateOperand(word); + } + } + bool isIdOperand(int op) const { return idOperand[op]; } + void setBlock(Block* b) { block = b; } + Block* getBlock() const { return block; } + Op getOpCode() const { return opCode; } + int getNumOperands() const + { + assert(operands.size() == idOperand.size()); + return (int)operands.size(); + } + Id getResultId() const { return resultId; } + Id getTypeId() const { return typeId; } + Id getIdOperand(int op) const { + assert(idOperand[op]); + return operands[op]; + } + unsigned int getImmediateOperand(int op) const { + assert(!idOperand[op]); + return operands[op]; + } + + // Write out the binary form. + void dump(std::vector<unsigned int>& out) const + { + // Compute the wordCount + unsigned int wordCount = 1; + if (typeId) + ++wordCount; + if (resultId) + ++wordCount; + wordCount += (unsigned int)operands.size(); + + // Write out the beginning of the instruction + out.push_back(((wordCount) << WordCountShift) | opCode); + if (typeId) + out.push_back(typeId); + if (resultId) + out.push_back(resultId); + + // Write out the operands + for (int op = 0; op < (int)operands.size(); ++op) + out.push_back(operands[op]); + } + +protected: + Instruction(const Instruction&); + Id resultId; + Id typeId; + Op opCode; + std::vector<Id> operands; // operands, both <id> and immediates (both are unsigned int) + std::vector<bool> idOperand; // true for operands that are <id>, false for immediates + Block* block; +}; + +// +// SPIR-V IR block. +// + +class Block { +public: + Block(Id id, Function& parent); + virtual ~Block() + { + } + + Id getId() { return instructions.front()->getResultId(); } + + Function& getParent() const { return parent; } + void addInstruction(std::unique_ptr<Instruction> inst); + void addPredecessor(Block* pred) { predecessors.push_back(pred); pred->successors.push_back(this);} + void addLocalVariable(std::unique_ptr<Instruction> inst) { localVariables.push_back(std::move(inst)); } + const std::vector<Block*>& getPredecessors() const { return predecessors; } + const std::vector<Block*>& getSuccessors() const { return successors; } + const std::vector<std::unique_ptr<Instruction> >& getInstructions() const { + return instructions; + } + const std::vector<std::unique_ptr<Instruction> >& getLocalVariables() const { return localVariables; } + void setUnreachable() { unreachable = true; } + bool isUnreachable() const { return unreachable; } + // Returns the block's merge instruction, if one exists (otherwise null). + const Instruction* getMergeInstruction() const { + if (instructions.size() < 2) return nullptr; + const Instruction* nextToLast = (instructions.cend() - 2)->get(); + switch (nextToLast->getOpCode()) { + case OpSelectionMerge: + case OpLoopMerge: + return nextToLast; + default: + return nullptr; + } + return nullptr; + } + + bool isTerminated() const + { + switch (instructions.back()->getOpCode()) { + case OpBranch: + case OpBranchConditional: + case OpSwitch: + case OpKill: + case OpReturn: + case OpReturnValue: + return true; + default: + return false; + } + } + + void dump(std::vector<unsigned int>& out) const + { + instructions[0]->dump(out); + for (int i = 0; i < (int)localVariables.size(); ++i) + localVariables[i]->dump(out); + for (int i = 1; i < (int)instructions.size(); ++i) + instructions[i]->dump(out); + } + +protected: + Block(const Block&); + Block& operator=(Block&); + + // To enforce keeping parent and ownership in sync: + friend Function; + + std::vector<std::unique_ptr<Instruction> > instructions; + std::vector<Block*> predecessors, successors; + std::vector<std::unique_ptr<Instruction> > localVariables; + Function& parent; + + // track whether this block is known to be uncreachable (not necessarily + // true for all unreachable blocks, but should be set at least + // for the extraneous ones introduced by the builder). + bool unreachable; +}; + +// Traverses the control-flow graph rooted at root in an order suited for +// readable code generation. Invokes callback at every node in the traversal +// order. +void inReadableOrder(Block* root, std::function<void(Block*)> callback); + +// +// SPIR-V IR Function. +// + +class Function { +public: + Function(Id id, Id resultType, Id functionType, Id firstParam, Module& parent); + virtual ~Function() + { + for (int i = 0; i < (int)parameterInstructions.size(); ++i) + delete parameterInstructions[i]; + + for (int i = 0; i < (int)blocks.size(); ++i) + delete blocks[i]; + } + Id getId() const { return functionInstruction.getResultId(); } + Id getParamId(int p) const { return parameterInstructions[p]->getResultId(); } + Id getParamType(int p) const { return parameterInstructions[p]->getTypeId(); } + + void addBlock(Block* block) { blocks.push_back(block); } + void removeBlock(Block* block) + { + auto found = find(blocks.begin(), blocks.end(), block); + assert(found != blocks.end()); + blocks.erase(found); + delete block; + } + + Module& getParent() const { return parent; } + Block* getEntryBlock() const { return blocks.front(); } + Block* getLastBlock() const { return blocks.back(); } + const std::vector<Block*>& getBlocks() const { return blocks; } + void addLocalVariable(std::unique_ptr<Instruction> inst); + Id getReturnType() const { return functionInstruction.getTypeId(); } + + void setImplicitThis() { implicitThis = true; } + bool hasImplicitThis() const { return implicitThis; } + + void dump(std::vector<unsigned int>& out) const + { + // OpFunction + functionInstruction.dump(out); + + // OpFunctionParameter + for (int p = 0; p < (int)parameterInstructions.size(); ++p) + parameterInstructions[p]->dump(out); + + // Blocks + inReadableOrder(blocks[0], [&out](const Block* b) { b->dump(out); }); + Instruction end(0, 0, OpFunctionEnd); + end.dump(out); + } + +protected: + Function(const Function&); + Function& operator=(Function&); + + Module& parent; + Instruction functionInstruction; + std::vector<Instruction*> parameterInstructions; + std::vector<Block*> blocks; + bool implicitThis; // true if this is a member function expecting to be passed a 'this' as the first argument +}; + +// +// SPIR-V IR Module. +// + +class Module { +public: + Module() {} + virtual ~Module() + { + // TODO delete things + } + + void addFunction(Function *fun) { functions.push_back(fun); } + + void mapInstruction(Instruction *instruction) + { + spv::Id resultId = instruction->getResultId(); + // map the instruction's result id + if (resultId >= idToInstruction.size()) + idToInstruction.resize(resultId + 16); + idToInstruction[resultId] = instruction; + } + + Instruction* getInstruction(Id id) const { return idToInstruction[id]; } + const std::vector<Function*>& getFunctions() const { return functions; } + spv::Id getTypeId(Id resultId) const { + return idToInstruction[resultId] == nullptr ? NoType : idToInstruction[resultId]->getTypeId(); + } + StorageClass getStorageClass(Id typeId) const + { + assert(idToInstruction[typeId]->getOpCode() == spv::OpTypePointer); + return (StorageClass)idToInstruction[typeId]->getImmediateOperand(0); + } + + void dump(std::vector<unsigned int>& out) const + { + for (int f = 0; f < (int)functions.size(); ++f) + functions[f]->dump(out); + } + +protected: + Module(const Module&); + std::vector<Function*> functions; + + // map from result id to instruction having that result id + std::vector<Instruction*> idToInstruction; + + // map from a result id to its type id +}; + +// +// Implementation (it's here due to circular type definitions). +// + +// Add both +// - the OpFunction instruction +// - all the OpFunctionParameter instructions +__inline Function::Function(Id id, Id resultType, Id functionType, Id firstParamId, Module& parent) + : parent(parent), functionInstruction(id, resultType, OpFunction), implicitThis(false) +{ + // OpFunction + functionInstruction.addImmediateOperand(FunctionControlMaskNone); + functionInstruction.addIdOperand(functionType); + parent.mapInstruction(&functionInstruction); + parent.addFunction(this); + + // OpFunctionParameter + Instruction* typeInst = parent.getInstruction(functionType); + int numParams = typeInst->getNumOperands() - 1; + for (int p = 0; p < numParams; ++p) { + Instruction* param = new Instruction(firstParamId + p, typeInst->getIdOperand(p + 1), OpFunctionParameter); + parent.mapInstruction(param); + parameterInstructions.push_back(param); + } +} + +__inline void Function::addLocalVariable(std::unique_ptr<Instruction> inst) +{ + Instruction* raw_instruction = inst.get(); + blocks[0]->addLocalVariable(std::move(inst)); + parent.mapInstruction(raw_instruction); +} + +__inline Block::Block(Id id, Function& parent) : parent(parent), unreachable(false) +{ + instructions.push_back(std::unique_ptr<Instruction>(new Instruction(id, NoType, OpLabel))); + instructions.back()->setBlock(this); + parent.getParent().mapInstruction(instructions.back().get()); +} + +__inline void Block::addInstruction(std::unique_ptr<Instruction> inst) +{ + Instruction* raw_instruction = inst.get(); + instructions.push_back(std::move(inst)); + raw_instruction->setBlock(this); + if (raw_instruction->getResultId()) + parent.getParent().mapInstruction(raw_instruction); +} + +}; // end spv namespace + +#endif // spvIR_H |