GDScript: Pool temporary values by type on the stack

So the stack slots perform less type changes, which is useful for
future optimizations.

5 files changed, 115 insertions, 49 deletions

diff --git a/modules/gdscript/gdscript_byte_codegen.cpp b/modules/gdscript/gdscript_byte_codegen.cpp
index b553dcede3..0aa35e10e7 100644
--- a/modules/gdscript/gdscript_byte_codegen.cpp
+++ b/modules/gdscript/gdscript_byte_codegen.cpp
@@ -47,7 +47,8 @@ uint32_t GDScriptByteCodeGenerator::add_parameter(const StringName &p_name, bool
 }
 
 uint32_t GDScriptByteCodeGenerator::add_local(const StringName &p_name, const GDScriptDataType &p_type) {
-	int stack_pos = increase_stack();
+	int stack_pos = locals.size() + RESERVED_STACK;
+	locals.push_back(StackSlot(p_type.builtin_type));
 	add_stack_identifier(p_name, stack_pos);
 	return stack_pos;
 }
@@ -66,25 +67,80 @@ uint32_t GDScriptByteCodeGenerator::add_or_get_name(const StringName &p_name) {
 	return get_name_map_pos(p_name);
 }
 
-uint32_t GDScriptByteCodeGenerator::add_temporary() {
-	current_temporaries++;
-	int idx = increase_stack();
-#ifdef DEBUG_ENABLED
-	temp_stack.push_back(idx);
-#endif
-	return idx;
+uint32_t GDScriptByteCodeGenerator::add_temporary(const GDScriptDataType &p_type) {
+	Variant::Type temp_type = Variant::NIL;
+	if (p_type.has_type) {
+		if (p_type.kind == GDScriptDataType::BUILTIN) {
+			switch (p_type.builtin_type) {
+				case Variant::NIL:
+				case Variant::BOOL:
+				case Variant::INT:
+				case Variant::FLOAT:
+				case Variant::STRING:
+				case Variant::VECTOR2:
+				case Variant::VECTOR2I:
+				case Variant::RECT2:
+				case Variant::RECT2I:
+				case Variant::VECTOR3:
+				case Variant::VECTOR3I:
+				case Variant::TRANSFORM2D:
+				case Variant::PLANE:
+				case Variant::QUAT:
+				case Variant::AABB:
+				case Variant::BASIS:
+				case Variant::TRANSFORM:
+				case Variant::COLOR:
+				case Variant::STRING_NAME:
+				case Variant::NODE_PATH:
+				case Variant::RID:
+				case Variant::OBJECT:
+				case Variant::CALLABLE:
+				case Variant::SIGNAL:
+				case Variant::DICTIONARY:
+				case Variant::ARRAY:
+					temp_type = p_type.builtin_type;
+					break;
+				case Variant::PACKED_BYTE_ARRAY:
+				case Variant::PACKED_INT32_ARRAY:
+				case Variant::PACKED_INT64_ARRAY:
+				case Variant::PACKED_FLOAT32_ARRAY:
+				case Variant::PACKED_FLOAT64_ARRAY:
+				case Variant::PACKED_STRING_ARRAY:
+				case Variant::PACKED_VECTOR2_ARRAY:
+				case Variant::PACKED_VECTOR3_ARRAY:
+				case Variant::PACKED_COLOR_ARRAY:
+				case Variant::VARIANT_MAX:
+					// Packed arrays are reference counted, so we don't use the pool for them.
+					temp_type = Variant::NIL;
+					break;
+			}
+		} else {
+			temp_type = Variant::OBJECT;
+		}
+	}
+
+	if (!temporaries_pool.has(temp_type)) {
+		temporaries_pool[temp_type] = List<int>();
+	}
+
+	List<int> &pool = temporaries_pool[temp_type];
+	if (pool.is_empty()) {
+		StackSlot new_temp(temp_type);
+		pool.push_back(temporaries.size());
+		temporaries.push_back(new_temp);
+	}
+	int slot = pool.front()->get();
+	pool.pop_front();
+	used_temporaries.push_back(slot);
+	return slot;
 }
 
 void GDScriptByteCodeGenerator::pop_temporary() {
-	ERR_FAIL_COND(current_temporaries == 0);
-	current_stack_size--;
-#ifdef DEBUG_ENABLED
-	if (temp_stack.back()->get() != current_stack_size) {
-		ERR_PRINT("Mismatched popping of temporary value");
-	}
-	temp_stack.pop_back();
-#endif
-	current_temporaries--;
+	ERR_FAIL_COND(used_temporaries.is_empty());
+	int slot_idx = used_temporaries.back()->get();
+	const StackSlot &slot = temporaries[slot_idx];
+	temporaries_pool[slot.type].push_back(slot_idx);
+	used_temporaries.pop_back();
 }
 
 void GDScriptByteCodeGenerator::start_parameters() {
@@ -119,12 +175,18 @@ void GDScriptByteCodeGenerator::write_start(GDScript *p_script, const StringName
 
 GDScriptFunction *GDScriptByteCodeGenerator::write_end() {
 #ifdef DEBUG_ENABLED
-	if (current_temporaries != 0) {
-		ERR_PRINT("Non-zero temporary variables at end of function: " + itos(current_temporaries));
+	if (!used_temporaries.is_empty()) {
+		ERR_PRINT("Non-zero temporary variables at end of function: " + itos(used_temporaries.size()));
 	}
 #endif
 	append(GDScriptFunction::OPCODE_END, 0);
 
+	for (int i = 0; i < temporaries.size(); i++) {
+		for (int j = 0; j < temporaries[i].bytecode_indices.size(); j++) {
+			opcodes.write[temporaries[i].bytecode_indices[j]] = (i + max_locals + RESERVED_STACK) | (GDScriptFunction::ADDR_TYPE_STACK << GDScriptFunction::ADDR_BITS);
+		}
+	}
+
 	if (constant_map.size()) {
 		function->_constant_count = constant_map.size();
 		function->constants.resize(constant_map.size());
@@ -317,7 +379,7 @@ GDScriptFunction *GDScriptByteCodeGenerator::write_end() {
 	if (debug_stack) {
 		function->stack_debug = stack_debug;
 	}
-	function->_stack_size = stack_max;
+	function->_stack_size = RESERVED_STACK + max_locals + temporaries.size();
 	function->_instruction_args_size = instr_args_max;
 	function->_ptrcall_args_size = ptrcall_max;
 
diff --git a/modules/gdscript/gdscript_byte_codegen.h b/modules/gdscript/gdscript_byte_codegen.h
index 4b196ed420..909d88ee63 100644
--- a/modules/gdscript/gdscript_byte_codegen.h
+++ b/modules/gdscript/gdscript_byte_codegen.h
@@ -37,6 +37,17 @@
 #include "gdscript_utility_functions.h"
 
 class GDScriptByteCodeGenerator : public GDScriptCodeGenerator {
+	struct StackSlot {
+		Variant::Type type = Variant::NIL;
+		Vector<int> bytecode_indices;
+
+		StackSlot() = default;
+		StackSlot(Variant::Type p_type) :
+				type(p_type) {}
+	};
+
+	const static int RESERVED_STACK = 3; // For self, class, and nil.
+
 	bool ended = false;
 	GDScriptFunction *function = nullptr;
 	bool debug_stack = false;
@@ -47,15 +58,17 @@ class GDScriptByteCodeGenerator : public GDScriptCodeGenerator {
 	List<int> stack_identifiers_counts;
 	Map<StringName, int> local_constants;
 
+	Vector<StackSlot> locals;
+	Vector<StackSlot> temporaries;
+	List<int> used_temporaries;
+	Map<Variant::Type, List<int>> temporaries_pool;
+
 	List<GDScriptFunction::StackDebug> stack_debug;
 	List<Map<StringName, int>> block_identifier_stack;
 	Map<StringName, int> block_identifiers;
 
-	int current_stack_size = 3; // First 3 spots are reserved for self, class, and nil.
-	int current_temporaries = 0;
-	int current_locals = 0;
+	int max_locals = 0;
 	int current_line = 0;
-	int stack_max = 3;
 	int instr_args_max = 0;
 	int ptrcall_max = 0;
 
@@ -102,7 +115,9 @@ class GDScriptByteCodeGenerator : public GDScriptCodeGenerator {
 	List<List<int>> match_continues_to_patch;
 
 	void add_stack_identifier(const StringName &p_id, int p_stackpos) {
-		current_locals++;
+		if (locals.size() > max_locals) {
+			max_locals = locals.size();
+		}
 		stack_identifiers[p_id] = p_stackpos;
 		if (debug_stack) {
 			block_identifiers[p_id] = p_stackpos;
@@ -116,7 +131,7 @@ class GDScriptByteCodeGenerator : public GDScriptCodeGenerator {
 	}
 
 	void push_stack_identifiers() {
-		stack_identifiers_counts.push_back(current_locals);
+		stack_identifiers_counts.push_back(locals.size());
 		stack_id_stack.push_back(stack_identifiers);
 		if (debug_stack) {
 			Map<StringName, int> block_ids(block_identifiers);
@@ -126,17 +141,16 @@ class GDScriptByteCodeGenerator : public GDScriptCodeGenerator {
 	}
 
 	void pop_stack_identifiers() {
-		current_locals = stack_identifiers_counts.back()->get();
+		int current_locals = stack_identifiers_counts.back()->get();
 		stack_identifiers_counts.pop_back();
 		stack_identifiers = stack_id_stack.back()->get();
 		stack_id_stack.pop_back();
 #ifdef DEBUG_ENABLED
-		if (current_temporaries != 0) {
-			ERR_PRINT("Leaving block with non-zero temporary variables: " + itos(current_temporaries));
+		if (!used_temporaries.is_empty()) {
+			ERR_PRINT("Leaving block with non-zero temporary variables: " + itos(used_temporaries.size()));
 		}
 #endif
-		current_stack_size = current_locals + 3; // Keep the 3 reserved slots for self, class, and nil.
-
+		locals.resize(current_locals);
 		if (debug_stack) {
 			for (Map<StringName, int>::Element *E = block_identifiers.front(); E; E = E->next()) {
 				GDScriptFunction::StackDebug sd;
@@ -279,18 +293,6 @@ class GDScriptByteCodeGenerator : public GDScriptCodeGenerator {
 		return pos;
 	}
 
-	void alloc_stack(int p_level) {
-		if (p_level >= stack_max) {
-			stack_max = p_level + 1;
-		}
-	}
-
-	int increase_stack() {
-		int top = current_stack_size++;
-		alloc_stack(current_stack_size);
-		return top;
-	}
-
 	void alloc_ptrcall(int p_params) {
 		if (p_params >= ptrcall_max) {
 			ptrcall_max = p_params;
@@ -308,9 +310,11 @@ class GDScriptByteCodeGenerator : public GDScriptCodeGenerator {
 			case Address::CONSTANT:
 				return p_address.address | (GDScriptFunction::ADDR_TYPE_CONSTANT << GDScriptFunction::ADDR_BITS);
 			case Address::LOCAL_VARIABLE:
-			case Address::TEMPORARY:
 			case Address::FUNCTION_PARAMETER:
 				return p_address.address | (GDScriptFunction::ADDR_TYPE_STACK << GDScriptFunction::ADDR_BITS);
+			case Address::TEMPORARY:
+				temporaries.write[p_address.address].bytecode_indices.push_back(opcodes.size());
+				return -1;
 			case Address::NIL:
 				return GDScriptFunction::ADDR_NIL;
 		}
@@ -392,7 +396,7 @@ public:
 	virtual uint32_t add_local_constant(const StringName &p_name, const Variant &p_constant) override;
 	virtual uint32_t add_or_get_constant(const Variant &p_constant) override;
 	virtual uint32_t add_or_get_name(const StringName &p_name) override;
-	virtual uint32_t add_temporary() override;
+	virtual uint32_t add_temporary(const GDScriptDataType &p_type) override;
 	virtual void pop_temporary() override;
 
 	virtual void start_parameters() override;
diff --git a/modules/gdscript/gdscript_codegen.h b/modules/gdscript/gdscript_codegen.h
index cce4e856c7..214d0fefaa 100644
--- a/modules/gdscript/gdscript_codegen.h
+++ b/modules/gdscript/gdscript_codegen.h
@@ -71,7 +71,7 @@ public:
 	virtual uint32_t add_local_constant(const StringName &p_name, const Variant &p_constant) = 0;
 	virtual uint32_t add_or_get_constant(const Variant &p_constant) = 0;
 	virtual uint32_t add_or_get_name(const StringName &p_name) = 0;
-	virtual uint32_t add_temporary() = 0;
+	virtual uint32_t add_temporary(const GDScriptDataType &p_type) = 0;
 	virtual void pop_temporary() = 0;
 
 	virtual void start_parameters() = 0;
diff --git a/modules/gdscript/gdscript_compiler.cpp b/modules/gdscript/gdscript_compiler.cpp
index abbca899bd..7429e3cc0b 100644
--- a/modules/gdscript/gdscript_compiler.cpp
+++ b/modules/gdscript/gdscript_compiler.cpp
@@ -711,7 +711,7 @@ GDScriptCodeGenerator::Address GDScriptCompiler::_parse_expression(CodeGen &code
 		case GDScriptParser::Node::UNARY_OPERATOR: {
 			const GDScriptParser::UnaryOpNode *unary = static_cast<const GDScriptParser::UnaryOpNode *>(p_expression);
 
-			GDScriptCodeGenerator::Address result = codegen.add_temporary();
+			GDScriptCodeGenerator::Address result = codegen.add_temporary(_gdtype_from_datatype(unary->get_datatype()));
 
 			GDScriptCodeGenerator::Address operand = _parse_expression(codegen, r_error, unary->operand);
 			if (r_error) {
@@ -729,7 +729,7 @@ GDScriptCodeGenerator::Address GDScriptCompiler::_parse_expression(CodeGen &code
 		case GDScriptParser::Node::BINARY_OPERATOR: {
 			const GDScriptParser::BinaryOpNode *binary = static_cast<const GDScriptParser::BinaryOpNode *>(p_expression);
 
-			GDScriptCodeGenerator::Address result = codegen.add_temporary();
+			GDScriptCodeGenerator::Address result = codegen.add_temporary(_gdtype_from_datatype(binary->get_datatype()));
 
 			switch (binary->operation) {
 				case GDScriptParser::BinaryOpNode::OP_LOGIC_AND: {
diff --git a/modules/gdscript/gdscript_compiler.h b/modules/gdscript/gdscript_compiler.h
index 1b0beec0d4..c405eadb07 100644
--- a/modules/gdscript/gdscript_compiler.h
+++ b/modules/gdscript/gdscript_compiler.h
@@ -66,7 +66,7 @@ class GDScriptCompiler {
 		}
 
 		GDScriptCodeGenerator::Address add_temporary(const GDScriptDataType &p_type = GDScriptDataType()) {
-			uint32_t addr = generator->add_temporary();
+			uint32_t addr = generator->add_temporary(p_type);
 			return GDScriptCodeGenerator::Address(GDScriptCodeGenerator::Address::TEMPORARY, addr, p_type);
 		}