libwebp-0.6.0-rc1

third_party/libwebp/utils/bit_reader_inl_utils.h

Index: third_party/libwebp/utils/bit_reader_inl_utils.h
diff --git a/third_party/libwebp/utils/bit_reader_inl.h b/third_party/libwebp/utils/bit_reader_inl_utils.h
similarity index 82%
rename from third_party/libwebp/utils/bit_reader_inl.h
rename to third_party/libwebp/utils/bit_reader_inl_utils.h
index 99ed3137d2cf912acec9f605adc161eb88a4173c..fd7fb0446c71fb6ef5f2b90077b402d67533b1ee 100644
--- a/third_party/libwebp/utils/bit_reader_inl.h
+++ b/third_party/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