26 files changed, 157 insertions, 109 deletions

diff --git a/thirdparty/libwebp/utils/bit_reader_inl.h b/thirdparty/libwebp/utils/bit_reader_inl_utils.h
index 99ed3137d2..fd7fb0446c 100644
--- a/thirdparty/libwebp/utils/bit_reader_inl.h
+++ b/thirdparty/libwebp/utils/bit_reader_inl_utils.h
@@ -20,13 +20,12 @@
 #include "../webp/config.h"
 #endif
 
-#ifdef WEBP_FORCE_ALIGNED
-#include <string.h>  // memcpy
-#endif
+#include <string.h>  // for memcpy
 
 #include "../dsp/dsp.h"
-#include "./bit_reader.h"
-#include "./endian_inl.h"
+#include "./bit_reader_utils.h"
+#include "./endian_inl_utils.h"
+#include "./utils.h"
 
 #ifdef __cplusplus
 extern "C" {
@@ -62,10 +61,7 @@ void VP8LoadNewBytes(VP8BitReader* const br) {
   if (br->buf_ < br->buf_max_) {
     // convert memory type to register type (with some zero'ing!)
     bit_t bits;
-#if defined(WEBP_FORCE_ALIGNED)
-    lbit_t in_bits;
-    memcpy(&in_bits, br->buf_, sizeof(in_bits));
-#elif defined(WEBP_USE_MIPS32)
+#if defined(WEBP_USE_MIPS32)
     // This is needed because of un-aligned read.
     lbit_t in_bits;
     lbit_t* p_buf_ = (lbit_t*)br->buf_;
@@ -80,7 +76,8 @@ void VP8LoadNewBytes(VP8BitReader* const br) {
       : "memory", "at"
     );
 #else
-    const lbit_t in_bits = *(const lbit_t*)br->buf_;
+    lbit_t in_bits;
+    memcpy(&in_bits, br->buf_, sizeof(in_bits));
 #endif
     br->buf_ += BITS >> 3;
 #if !defined(WORDS_BIGENDIAN)
@@ -119,37 +116,26 @@ static WEBP_INLINE int VP8GetBit(VP8BitReader* const br, int prob) {
     const int pos = br->bits_;
     const range_t split = (range * prob) >> 8;
     const range_t value = (range_t)(br->value_ >> pos);
-#if defined(__arm__) || defined(_M_ARM)      // ARM-specific
-    const int bit = ((int)(split - value) >> 31) & 1;
-    if (value > split) {
-      range -= split + 1;
-      br->value_ -= (bit_t)(split + 1) << pos;
-    } else {
-      range = split;
-    }
-#else  // faster version on x86
-    int bit;  // Don't use 'const int bit = (value > split);", it's slower.
-    if (value > split) {
-      range -= split + 1;
+    const int bit = (value > split);
+    if (bit) {
+      range -= split;
       br->value_ -= (bit_t)(split + 1) << pos;
-      bit = 1;
     } else {
-      range = split;
-      bit = 0;
+      range = split + 1;
     }
-#endif
-    if (range <= (range_t)0x7e) {
-      const int shift = kVP8Log2Range[range];
-      range = kVP8NewRange[range];
+    {
+      const int shift = 7 ^ BitsLog2Floor(range);
+      range <<= shift;
       br->bits_ -= shift;
     }
-    br->range_ = range;
+    br->range_ = range - 1;
     return bit;
   }
 }
 
 // simplified version of VP8GetBit() for prob=0x80 (note shift is always 1 here)
-static WEBP_INLINE int VP8GetSigned(VP8BitReader* const br, int v) {
+static WEBP_UBSAN_IGNORE_UNSIGNED_OVERFLOW WEBP_INLINE
+int VP8GetSigned(VP8BitReader* const br, int v) {
   if (br->bits_ < 0) {
     VP8LoadNewBytes(br);
   }
@@ -166,6 +152,37 @@ static WEBP_INLINE int VP8GetSigned(VP8BitReader* const br, int v) {
   }
 }
 
+static WEBP_INLINE int VP8GetBitAlt(VP8BitReader* const br, int prob) {
+  // Don't move this declaration! It makes a big speed difference to store
+  // 'range' *before* calling VP8LoadNewBytes(), even if this function doesn't
+  // alter br->range_ value.
+  range_t range = br->range_;
+  if (br->bits_ < 0) {
+    VP8LoadNewBytes(br);
+  }
+  {
+    const int pos = br->bits_;
+    const range_t split = (range * prob) >> 8;
+    const range_t value = (range_t)(br->value_ >> pos);
+    int bit;  // Don't use 'const int bit = (value > split);", it's slower.
+    if (value > split) {
+      range -= split + 1;
+      br->value_ -= (bit_t)(split + 1) << pos;
+      bit = 1;
+    } else {
+      range = split;
+      bit = 0;
+    }
+    if (range <= (range_t)0x7e) {
+      const int shift = kVP8Log2Range[range];
+      range = kVP8NewRange[range];
+      br->bits_ -= shift;
+    }
+    br->range_ = range;
+    return bit;
+  }
+}
+
 #ifdef __cplusplus
 }    // extern "C"
 #endif
diff --git a/thirdparty/libwebp/utils/bit_reader.c b/thirdparty/libwebp/utils/bit_reader_utils.c
index 2eb46e0b4b..053b710bb8 100644
--- a/thirdparty/libwebp/utils/bit_reader.c
+++ b/thirdparty/libwebp/utils/bit_reader_utils.c
@@ -15,7 +15,7 @@
 #include "../webp/config.h"
 #endif
 
-#include "./bit_reader_inl.h"
+#include "./bit_reader_inl_utils.h"
 #include "../utils/utils.h"
 
 //------------------------------------------------------------------------------
diff --git a/thirdparty/libwebp/utils/bit_reader.h b/thirdparty/libwebp/utils/bit_reader_utils.h
index ea5c584eb4..ea5c584eb4 100644
--- a/thirdparty/libwebp/utils/bit_reader.h
+++ b/thirdparty/libwebp/utils/bit_reader_utils.h
diff --git a/thirdparty/libwebp/utils/bit_writer.c b/thirdparty/libwebp/utils/bit_writer_utils.c
index 064428691b..ab0c49dce8 100644
--- a/thirdparty/libwebp/utils/bit_writer.c
+++ b/thirdparty/libwebp/utils/bit_writer_utils.c
@@ -16,8 +16,8 @@
 #include <string.h>   // for memcpy()
 #include <stdlib.h>
 
-#include "./bit_writer.h"
-#include "./endian_inl.h"
+#include "./bit_writer_utils.h"
+#include "./endian_inl_utils.h"
 #include "./utils.h"
 
 //------------------------------------------------------------------------------
@@ -143,13 +143,13 @@ int VP8PutBitUniform(VP8BitWriter* const bw, int bit) {
 void VP8PutBits(VP8BitWriter* const bw, uint32_t value, int nb_bits) {
   uint32_t mask;
   assert(nb_bits > 0 && nb_bits < 32);
-  for (mask = 1u << (nb_bits - 1); mask; mask >>= 1)
+  for (mask = 1u << (nb_bits - 1); mask; mask >>= 1) {
     VP8PutBitUniform(bw, value & mask);
+  }
 }
 
 void VP8PutSignedBits(VP8BitWriter* const bw, int value, int nb_bits) {
-  if (!VP8PutBitUniform(bw, value != 0))
-    return;
+  if (!VP8PutBitUniform(bw, value != 0)) return;
   if (value < 0) {
     VP8PutBits(bw, ((-value) << 1) | 1, nb_bits + 1);
   } else {
diff --git a/thirdparty/libwebp/utils/bit_writer.h b/thirdparty/libwebp/utils/bit_writer_utils.h
index ef360d1dc6..9c02bbc06d 100644
--- a/thirdparty/libwebp/utils/bit_writer.h
+++ b/thirdparty/libwebp/utils/bit_writer_utils.h
@@ -54,7 +54,8 @@ int VP8BitWriterAppend(VP8BitWriter* const bw,
 
 // return approximate write position (in bits)
 static WEBP_INLINE uint64_t VP8BitWriterPos(const VP8BitWriter* const bw) {
-  return (uint64_t)(bw->pos_ + bw->run_) * 8 + 8 + bw->nb_bits_;
+  const uint64_t nb_bits = 8 + bw->nb_bits_;   // bw->nb_bits_ is <= 0, note
+  return (bw->pos_ + bw->run_) * 8 + nb_bits;
 }
 
 // Returns a pointer to the internal buffer.
diff --git a/thirdparty/libwebp/utils/color_cache.c b/thirdparty/libwebp/utils/color_cache_utils.c
index c34b2e7f1a..0172590c48 100644
--- a/thirdparty/libwebp/utils/color_cache.c
+++ b/thirdparty/libwebp/utils/color_cache_utils.c
@@ -14,7 +14,7 @@
 #include <assert.h>
 #include <stdlib.h>
 #include <string.h>
-#include "./color_cache.h"
+#include "./color_cache_utils.h"
 #include "./utils.h"
 
 //------------------------------------------------------------------------------
diff --git a/thirdparty/libwebp/utils/color_cache.h b/thirdparty/libwebp/utils/color_cache_utils.h
index a9a9f64270..c373e6b361 100644
--- a/thirdparty/libwebp/utils/color_cache.h
+++ b/thirdparty/libwebp/utils/color_cache_utils.h
@@ -28,7 +28,11 @@ typedef struct {
   int hash_bits_;
 } VP8LColorCache;
 
-static const uint32_t kHashMul = 0x1e35a7bd;
+static const uint64_t kHashMul = 0x1e35a7bdull;
+
+static WEBP_INLINE int HashPix(uint32_t argb, int shift) {
+  return (int)(((argb * kHashMul) & 0xffffffffu) >> shift);
+}
 
 static WEBP_INLINE uint32_t VP8LColorCacheLookup(
     const VP8LColorCache* const cc, uint32_t key) {
@@ -44,19 +48,20 @@ static WEBP_INLINE void VP8LColorCacheSet(const VP8LColorCache* const cc,
 
 static WEBP_INLINE void VP8LColorCacheInsert(const VP8LColorCache* const cc,
                                              uint32_t argb) {
-  const uint32_t key = (kHashMul * argb) >> cc->hash_shift_;
+  const int key = HashPix(argb, cc->hash_shift_);
   cc->colors_[key] = argb;
 }
 
 static WEBP_INLINE int VP8LColorCacheGetIndex(const VP8LColorCache* const cc,
                                               uint32_t argb) {
-  return (kHashMul * argb) >> cc->hash_shift_;
+  return HashPix(argb, cc->hash_shift_);
 }
 
+// Return the key if cc contains argb, and -1 otherwise.
 static WEBP_INLINE int VP8LColorCacheContains(const VP8LColorCache* const cc,
                                               uint32_t argb) {
-  const uint32_t key = (kHashMul * argb) >> cc->hash_shift_;
-  return (cc->colors_[key] == argb);
+  const int key = HashPix(argb, cc->hash_shift_);
+  return (cc->colors_[key] == argb) ? key : -1;
 }
 
 //------------------------------------------------------------------------------
diff --git a/thirdparty/libwebp/utils/endian_inl.h b/thirdparty/libwebp/utils/endian_inl_utils.h
index e11260ff7d..e11260ff7d 100644
--- a/thirdparty/libwebp/utils/endian_inl.h
+++ b/thirdparty/libwebp/utils/endian_inl_utils.h
diff --git a/thirdparty/libwebp/utils/filters.c b/thirdparty/libwebp/utils/filters_utils.c
index 15543b1271..49c1d18a22 100644
--- a/thirdparty/libwebp/utils/filters.c
+++ b/thirdparty/libwebp/utils/filters_utils.c
@@ -11,7 +11,7 @@
 //
 // Author: Urvang (urvang@google.com)
 
-#include "./filters.h"
+#include "./filters_utils.h"
 #include <stdlib.h>
 #include <string.h>
 
diff --git a/thirdparty/libwebp/utils/filters.h b/thirdparty/libwebp/utils/filters_utils.h
index 088b132fc5..088b132fc5 100644
--- a/thirdparty/libwebp/utils/filters.h
+++ b/thirdparty/libwebp/utils/filters_utils.h
diff --git a/thirdparty/libwebp/utils/huffman_encode.c b/thirdparty/libwebp/utils/huffman_encode_utils.c
index 4e5ef6b447..f9504658ea 100644
--- a/thirdparty/libwebp/utils/huffman_encode.c
+++ b/thirdparty/libwebp/utils/huffman_encode_utils.c
@@ -14,7 +14,7 @@
 #include <assert.h>
 #include <stdlib.h>
 #include <string.h>
-#include "./huffman_encode.h"
+#include "./huffman_encode_utils.h"
 #include "./utils.h"
 #include "../webp/format_constants.h"
 
diff --git a/thirdparty/libwebp/utils/huffman_encode.h b/thirdparty/libwebp/utils/huffman_encode_utils.h
index a157165148..a157165148 100644
--- a/thirdparty/libwebp/utils/huffman_encode.h
+++ b/thirdparty/libwebp/utils/huffman_encode_utils.h
diff --git a/thirdparty/libwebp/utils/huffman.c b/thirdparty/libwebp/utils/huffman_utils.c
index 36e5502836..008b5d746f 100644
--- a/thirdparty/libwebp/utils/huffman.c
+++ b/thirdparty/libwebp/utils/huffman_utils.c
@@ -14,7 +14,7 @@
 #include <assert.h>
 #include <stdlib.h>
 #include <string.h>
-#include "./huffman.h"
+#include "./huffman_utils.h"
 #include "./utils.h"
 #include "../webp/format_constants.h"
 
@@ -45,7 +45,7 @@ static WEBP_INLINE uint32_t GetNextKey(uint32_t key, int len) {
   while (key & step) {
     step >>= 1;
   }
-  return (key & (step - 1)) + step;
+  return step ? (key & (step - 1)) + step : key;
 }
 
 // Stores code in table[0], table[step], table[2*step], ..., table[end].
@@ -75,11 +75,13 @@ static WEBP_INLINE int NextTableBitSize(const int* const count,
   return len - root_bits;
 }
 
-int VP8LBuildHuffmanTable(HuffmanCode* const root_table, int root_bits,
-                          const int code_lengths[], int code_lengths_size) {
+// sorted[code_lengths_size] is a pre-allocated array for sorting symbols
+// by code length.
+static int BuildHuffmanTable(HuffmanCode* const root_table, int root_bits,
+                             const int code_lengths[], int code_lengths_size,
+                             uint16_t sorted[]) {
   HuffmanCode* table = root_table;  // next available space in table
   int total_size = 1 << root_bits;  // total size root table + 2nd level table
-  int* sorted = NULL;               // symbols sorted by code length
   int len;                          // current code length
   int symbol;                       // symbol index in original or sorted table
   // number of codes of each length:
@@ -114,11 +116,6 @@ int VP8LBuildHuffmanTable(HuffmanCode* const root_table, int root_bits,
     offset[len + 1] = offset[len] + count[len];
   }
 
-  sorted = (int*)WebPSafeMalloc(code_lengths_size, sizeof(*sorted));
-  if (sorted == NULL) {
-    return 0;
-  }
-
   // Sort symbols by length, by symbol order within each length.
   for (symbol = 0; symbol < code_lengths_size; ++symbol) {
     const int symbol_code_length = code_lengths[symbol];
@@ -133,7 +130,6 @@ int VP8LBuildHuffmanTable(HuffmanCode* const root_table, int root_bits,
     code.bits = 0;
     code.value = (uint16_t)sorted[0];
     ReplicateValue(table, 1, total_size, code);
-    WebPSafeFree(sorted);
     return total_size;
   }
 
@@ -153,7 +149,6 @@ int VP8LBuildHuffmanTable(HuffmanCode* const root_table, int root_bits,
       num_nodes += num_open;
       num_open -= count[len];
       if (num_open < 0) {
-        WebPSafeFree(sorted);
         return 0;
       }
       for (; count[len] > 0; --count[len]) {
@@ -172,7 +167,6 @@ int VP8LBuildHuffmanTable(HuffmanCode* const root_table, int root_bits,
       num_nodes += num_open;
       num_open -= count[len];
       if (num_open < 0) {
-        WebPSafeFree(sorted);
         return 0;
       }
       for (; count[len] > 0; --count[len]) {
@@ -195,11 +189,35 @@ int VP8LBuildHuffmanTable(HuffmanCode* const root_table, int root_bits,
 
     // Check if tree is full.
     if (num_nodes != 2 * offset[MAX_ALLOWED_CODE_LENGTH] - 1) {
-      WebPSafeFree(sorted);
       return 0;
     }
   }
 
-  WebPSafeFree(sorted);
+  return total_size;
+}
+
+// Maximum code_lengths_size is 2328 (reached for 11-bit color_cache_bits).
+// More commonly, the value is around ~280.
+#define MAX_CODE_LENGTHS_SIZE \
+  ((1 << MAX_CACHE_BITS) + NUM_LITERAL_CODES + NUM_LENGTH_CODES)
+// Cut-off value for switching between heap and stack allocation.
+#define SORTED_SIZE_CUTOFF 512
+int VP8LBuildHuffmanTable(HuffmanCode* const root_table, int root_bits,
+                          const int code_lengths[], int code_lengths_size) {
+  int total_size;
+  assert(code_lengths_size <= MAX_CODE_LENGTHS_SIZE);
+  if (code_lengths_size <= SORTED_SIZE_CUTOFF) {
+    // use local stack-allocated array.
+    uint16_t sorted[SORTED_SIZE_CUTOFF];
+    total_size = BuildHuffmanTable(root_table, root_bits,
+                                   code_lengths, code_lengths_size, sorted);
+  } else {   // rare case. Use heap allocation.
+    uint16_t* const sorted =
+        (uint16_t*)WebPSafeMalloc(code_lengths_size, sizeof(*sorted));
+    if (sorted == NULL) return 0;
+    total_size = BuildHuffmanTable(root_table, root_bits,
+                                   code_lengths, code_lengths_size, sorted);
+    WebPSafeFree(sorted);
+  }
   return total_size;
 }
diff --git a/thirdparty/libwebp/utils/huffman.h b/thirdparty/libwebp/utils/huffman_utils.h
index c6dd6aaa45..c6dd6aaa45 100644
--- a/thirdparty/libwebp/utils/huffman.h
+++ b/thirdparty/libwebp/utils/huffman_utils.h
diff --git a/thirdparty/libwebp/utils/quant_levels_dec.c b/thirdparty/libwebp/utils/quant_levels_dec_utils.c
index ee0a3fe127..d4d23d3147 100644
--- a/thirdparty/libwebp/utils/quant_levels_dec.c
+++ b/thirdparty/libwebp/utils/quant_levels_dec_utils.c
@@ -14,7 +14,7 @@
 //
 // Author: Skal (pascal.massimino@gmail.com)
 
-#include "./quant_levels_dec.h"
+#include "./quant_levels_dec_utils.h"
 
 #include <string.h>   // for memset
 
diff --git a/thirdparty/libwebp/utils/quant_levels_dec.h b/thirdparty/libwebp/utils/quant_levels_dec_utils.h
index 59a13495d3..59a13495d3 100644
--- a/thirdparty/libwebp/utils/quant_levels_dec.h
+++ b/thirdparty/libwebp/utils/quant_levels_dec_utils.h
diff --git a/thirdparty/libwebp/utils/quant_levels.c b/thirdparty/libwebp/utils/quant_levels_utils.c
index d7c8aab922..73174e8ab9 100644
--- a/thirdparty/libwebp/utils/quant_levels.c
+++ b/thirdparty/libwebp/utils/quant_levels_utils.c
@@ -14,7 +14,7 @@
 
 #include <assert.h>
 
-#include "./quant_levels.h"
+#include "./quant_levels_utils.h"
 
 #define NUM_SYMBOLS     256
 
diff --git a/thirdparty/libwebp/utils/quant_levels.h b/thirdparty/libwebp/utils/quant_levels_utils.h
index 1cb5a32cae..1cb5a32cae 100644
--- a/thirdparty/libwebp/utils/quant_levels.h
+++ b/thirdparty/libwebp/utils/quant_levels_utils.h
diff --git a/thirdparty/libwebp/utils/random.c b/thirdparty/libwebp/utils/random_utils.c
index 24e96ad648..9f1e4154a6 100644
--- a/thirdparty/libwebp/utils/random.c
+++ b/thirdparty/libwebp/utils/random_utils.c
@@ -12,7 +12,7 @@
 // Author: Skal (pascal.massimino@gmail.com)
 
 #include <string.h>
-#include "./random.h"
+#include "./random_utils.h"
 
 //------------------------------------------------------------------------------
 
diff --git a/thirdparty/libwebp/utils/random.h b/thirdparty/libwebp/utils/random_utils.h
index c392a615ca..c392a615ca 100644
--- a/thirdparty/libwebp/utils/random.h
+++ b/thirdparty/libwebp/utils/random_utils.h
diff --git a/thirdparty/libwebp/utils/rescaler.c b/thirdparty/libwebp/utils/rescaler_utils.c
index d2278a52ff..0d1f80da24 100644
--- a/thirdparty/libwebp/utils/rescaler.c
+++ b/thirdparty/libwebp/utils/rescaler_utils.c
@@ -15,7 +15,7 @@
 #include <stdlib.h>
 #include <string.h>
 #include "../dsp/dsp.h"
-#include "./rescaler.h"
+#include "./rescaler_utils.h"
 
 //------------------------------------------------------------------------------
 
diff --git a/thirdparty/libwebp/utils/rescaler.h b/thirdparty/libwebp/utils/rescaler_utils.h
index 98b01a76d0..98b01a76d0 100644
--- a/thirdparty/libwebp/utils/rescaler.h
+++ b/thirdparty/libwebp/utils/rescaler_utils.h
diff --git a/thirdparty/libwebp/utils/thread.c b/thirdparty/libwebp/utils/thread_utils.c
index 93f7622797..1729060c70 100644
--- a/thirdparty/libwebp/utils/thread.c
+++ b/thirdparty/libwebp/utils/thread_utils.c
@@ -13,7 +13,7 @@
 
 #include <assert.h>
 #include <string.h>   // for memset()
-#include "./thread.h"
+#include "./thread_utils.h"
 #include "./utils.h"
 
 #ifdef WEBP_USE_THREAD
@@ -183,8 +183,7 @@ static int pthread_cond_wait(pthread_cond_t* const condition,
 #else
   // note that there is a consumer available so the signal isn't dropped in
   // pthread_cond_signal
-  if (!ReleaseSemaphore(condition->waiting_sem_, 1, NULL))
-    return 1;
+  if (!ReleaseSemaphore(condition->waiting_sem_, 1, NULL)) return 1;
   // now unlock the mutex so pthread_cond_signal may be issued
   pthread_mutex_unlock(mutex);
   ok = (WaitForSingleObject(condition->signal_event_, INFINITE) ==
@@ -226,8 +225,7 @@ static THREADFN ThreadLoop(void* ptr) {
 }
 
 // main thread state control
-static void ChangeState(WebPWorker* const worker,
-                        WebPWorkerStatus new_status) {
+static void ChangeState(WebPWorker* const worker, WebPWorkerStatus new_status) {
   // No-op when attempting to change state on a thread that didn't come up.
   // Checking status_ without acquiring the lock first would result in a data
   // race.
diff --git a/thirdparty/libwebp/utils/thread.h b/thirdparty/libwebp/utils/thread_utils.h
index 8408311855..8408311855 100644
--- a/thirdparty/libwebp/utils/thread.h
+++ b/thirdparty/libwebp/utils/thread_utils.h
diff --git a/thirdparty/libwebp/utils/utils.c b/thirdparty/libwebp/utils/utils.c
index 82dbf8d5e5..504d924b60 100644
--- a/thirdparty/libwebp/utils/utils.c
+++ b/thirdparty/libwebp/utils/utils.c
@@ -25,7 +25,7 @@
 //    http://valgrind.org/docs/manual/ms-manual.html
 // Here is an example command line:
 /*    valgrind --tool=massif --massif-out-file=massif.out \
-               --stacks=yes --alloc-fn=WebPSafeAlloc --alloc-fn=WebPSafeCalloc
+               --stacks=yes --alloc-fn=WebPSafeMalloc --alloc-fn=WebPSafeCalloc
       ms_print massif.out
 */
 // In addition:
@@ -243,8 +243,7 @@ void WebPCopyPixels(const WebPPicture* const src, WebPPicture* const dst) {
 
 //------------------------------------------------------------------------------
 
-#define MAX_COLOR_COUNT         MAX_PALETTE_SIZE
-#define COLOR_HASH_SIZE         (MAX_COLOR_COUNT * 4)
+#define COLOR_HASH_SIZE         (MAX_PALETTE_SIZE * 4)
 #define COLOR_HASH_RIGHT_SHIFT  22  // 32 - log2(COLOR_HASH_SIZE).
 
 int WebPGetColorPalette(const WebPPicture* const pic, uint32_t* const palette) {
@@ -253,7 +252,7 @@ int WebPGetColorPalette(const WebPPicture* const pic, uint32_t* const palette) {
   int num_colors = 0;
   uint8_t in_use[COLOR_HASH_SIZE] = { 0 };
   uint32_t colors[COLOR_HASH_SIZE];
-  static const uint32_t kHashMul = 0x1e35a7bdU;
+  static const uint64_t kHashMul = 0x1e35a7bdull;
   const uint32_t* argb = pic->argb;
   const int width = pic->width;
   const int height = pic->height;
@@ -268,14 +267,14 @@ int WebPGetColorPalette(const WebPPicture* const pic, uint32_t* const palette) {
         continue;
       }
       last_pix = argb[x];
-      key = (kHashMul * last_pix) >> COLOR_HASH_RIGHT_SHIFT;
+      key = ((last_pix * kHashMul) & 0xffffffffu) >> COLOR_HASH_RIGHT_SHIFT;
       while (1) {
         if (!in_use[key]) {
           colors[key] = last_pix;
           in_use[key] = 1;
           ++num_colors;
-          if (num_colors > MAX_COLOR_COUNT) {
-            return MAX_COLOR_COUNT + 1;  // Exact count not needed.
+          if (num_colors > MAX_PALETTE_SIZE) {
+            return MAX_PALETTE_SIZE + 1;  // Exact count not needed.
           }
           break;
         } else if (colors[key] == last_pix) {
@@ -302,8 +301,30 @@ int WebPGetColorPalette(const WebPPicture* const pic, uint32_t* const palette) {
   return num_colors;
 }
 
-#undef MAX_COLOR_COUNT
 #undef COLOR_HASH_SIZE
 #undef COLOR_HASH_RIGHT_SHIFT
 
 //------------------------------------------------------------------------------
+
+#if defined(WEBP_NEED_LOG_TABLE_8BIT)
+const uint8_t WebPLogTable8bit[256] = {   // 31 ^ clz(i)
+  0, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3,
+  4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
+  6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
+  6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
+  6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7
+};
+#endif
+
+//------------------------------------------------------------------------------
diff --git a/thirdparty/libwebp/utils/utils.h b/thirdparty/libwebp/utils/utils.h
index 3a5d4e6a78..3ab459050a 100644
--- a/thirdparty/libwebp/utils/utils.h
+++ b/thirdparty/libwebp/utils/utils.h
@@ -62,7 +62,6 @@ WEBP_EXTERN(void) WebPSafeFree(void* const ptr);
 #define WEBP_ALIGN_CST 31
 #define WEBP_ALIGN(PTR) (((uintptr_t)(PTR) + WEBP_ALIGN_CST) & ~WEBP_ALIGN_CST)
 
-#if defined(WEBP_FORCE_ALIGNED)
 #include <string.h>
 // memcpy() is the safe way of moving potentially unaligned 32b memory.
 static WEBP_INLINE uint32_t WebPMemToUint32(const uint8_t* const ptr) {
@@ -73,16 +72,6 @@ static WEBP_INLINE uint32_t WebPMemToUint32(const uint8_t* const ptr) {
 static WEBP_INLINE void WebPUint32ToMem(uint8_t* const ptr, uint32_t val) {
   memcpy(ptr, &val, sizeof(val));
 }
-#else
-static WEBP_UBSAN_IGNORE_UNDEF WEBP_INLINE
-uint32_t WebPMemToUint32(const uint8_t* const ptr) {
-  return *(const uint32_t*)ptr;
-}
-static WEBP_UBSAN_IGNORE_UNDEF WEBP_INLINE
-void WebPUint32ToMem(uint8_t* const ptr, uint32_t val) {
-  *(uint32_t*)ptr = val;
-}
-#endif
 
 //------------------------------------------------------------------------------
 // Reading/writing data.
@@ -118,6 +107,19 @@ static WEBP_INLINE void PutLE32(uint8_t* const data, uint32_t val) {
   PutLE16(data + 2, (int)(val >> 16));
 }
 
+// Returns 31 ^ clz(n) = log2(n). This is the default C-implementation, either
+// based on table or not. Can be used as fallback if clz() is not available.
+#define WEBP_NEED_LOG_TABLE_8BIT
+extern const uint8_t WebPLogTable8bit[256];
+static WEBP_INLINE int WebPLog2FloorC(uint32_t n) {
+  int log = 0;
+  while (n >= 256) {
+    log += 8;
+    n >>= 8;
+  }
+  return log + WebPLogTable8bit[n];
+}
+
 // Returns (int)floor(log2(n)). n must be > 0.
 // use GNU builtins where available.
 #if defined(__GNUC__) && \
@@ -135,22 +137,8 @@ static WEBP_INLINE int BitsLog2Floor(uint32_t n) {
   _BitScanReverse(&first_set_bit, n);
   return first_set_bit;
 }
-#else
-static WEBP_INLINE int BitsLog2Floor(uint32_t n) {
-  int log = 0;
-  uint32_t value = n;
-  int i;
-
-  for (i = 4; i >= 0; --i) {
-    const int shift = (1 << i);
-    const uint32_t x = value >> shift;
-    if (x != 0) {
-      value = x;
-      log += shift;
-    }
-  }
-  return log;
-}
+#else   // default: use the C-version.
+static WEBP_INLINE int BitsLog2Floor(uint32_t n) { return WebPLog2FloorC(n); }
 #endif
 
 //------------------------------------------------------------------------------
@@ -172,12 +160,12 @@ WEBP_EXTERN(void) WebPCopyPixels(const struct WebPPicture* const src,
 // Unique colors.
 
 // Returns count of unique colors in 'pic', assuming pic->use_argb is true.
-// If the unique color count is more than MAX_COLOR_COUNT, returns
-// MAX_COLOR_COUNT+1.
+// If the unique color count is more than MAX_PALETTE_SIZE, returns
+// MAX_PALETTE_SIZE+1.
 // If 'palette' is not NULL and number of unique colors is less than or equal to
-// MAX_COLOR_COUNT, also outputs the actual unique colors into 'palette'.
+// MAX_PALETTE_SIZE, also outputs the actual unique colors into 'palette'.
 // Note: 'palette' is assumed to be an array already allocated with at least
-// MAX_COLOR_COUNT elements.
+// MAX_PALETTE_SIZE elements.
 WEBP_EXTERN(int) WebPGetColorPalette(const struct WebPPicture* const pic,
                                      uint32_t* const palette);