libwebp-0.5.2

bug fixes, minor encoder updates

Test: aosp_* build, graphics cts pass on n6p

Change-Id: I9d3ceef65b160cfe8b2734ad9193d681fe05f9ed

32 files changed, 315 insertions, 242 deletions

diff --git a/ChangeLog b/ChangeLog
index 3b62f4f6..afcd6792 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -14,3 +14,4 @@
 - 11/2: Release version 0.4.4 (tag: v0.4.4)
 - 1/12: Release version 0.5.0 (tag: v0.5.0)
 - 7/12: Release version 0.5.1 (tag: v0.5.1)
+- 12/20: Release version 0.5.2 (tag: v0.5.2)
diff --git a/README b/README
index 90f8f10f..f2a719cc 100644
--- a/README
+++ b/README
@@ -4,7 +4,7 @@
           \__\__/\____/\_____/__/ ____  ___
                 / _/ /    \    \ /  _ \/ _/
                /  \_/   / /   \ \   __/  \__
-               \____/____/\_____/_____/____/v0.5.1
+               \____/____/\_____/_____/____/v0.5.2
 
 Description:
 ============
@@ -220,8 +220,9 @@ assumed to be a PNG, JPEG, TIFF or WebP file.
 Options:
   -h / -help ............. short help
   -H / -longhelp ......... long help
-  -q <float> ............. quality factor (0:small..100:big)
-  -alpha_q <int> ......... transparency-compression quality (0..100)
+  -q <float> ............. quality factor (0:small..100:big), default=75
+  -alpha_q <int> ......... transparency-compression quality (0..100),
+                           default=100
   -preset <string> ....... preset setting, one of:
                             default, photo, picture,
                             drawing, icon, text
@@ -229,15 +230,15 @@ Options:
   -z <int> ............... activates lossless preset with given
                            level in [0:fast, ..., 9:slowest]
 
-  -m <int> ............... compression method (0=fast, 6=slowest)
-  -segments <int> ........ number of segments to use (1..4)
+  -m <int> ............... compression method (0=fast, 6=slowest), default=4
+  -segments <int> ........ number of segments to use (1..4), default=4
   -size <int> ............ target size (in bytes)
   -psnr <float> .......... target PSNR (in dB. typically: 42)
 
   -s <int> <int> ......... input size (width x height) for YUV
-  -sns <int> ............. spatial noise shaping (0:off, 100:max)
-  -f <int> ............... filter strength (0=off..100)
-  -sharpness <int> ....... filter sharpness (0:most .. 7:least sharp)
+  -sns <int> ............. spatial noise shaping (0:off, 100:max), default=50
+  -f <int> ............... filter strength (0=off..100), default=60
+  -sharpness <int> ....... filter sharpness (0:most .. 7:least sharp), default=0
   -strong ................ use strong filter instead of simple (default)
   -nostrong .............. use simple filter instead of strong
   -partition_limit <int> . limit quality to fit the 512k limit on
@@ -252,18 +253,18 @@ Options:
   -print_ssim ............ prints averaged SSIM distortion
   -print_lsim ............ prints local-similarity distortion
   -d <file.pgm> .......... dump the compressed output (PGM file)
-  -alpha_method <int> .... transparency-compression method (0..1)
+  -alpha_method <int> .... transparency-compression method (0..1), default=1
   -alpha_filter <string> . predictive filtering for alpha plane,
                            one of: none, fast (default) or best
-  -exact ................. preserve RGB values in transparent area
+  -exact ................. preserve RGB values in transparent area, default=off
   -blend_alpha <hex> ..... blend colors against background color
                            expressed as RGB values written in
                            hexadecimal, e.g. 0xc0e0d0 for red=0xc0
                            green=0xe0 and blue=0xd0
   -noalpha ............... discard any transparency information
-  -lossless .............. encode image losslessly
+  -lossless .............. encode image losslessly, default=off
   -near_lossless <int> ... use near-lossless image
-                           preprocessing (0..100=off)
+                           preprocessing (0..100=off), default=100
   -hint <string> ......... specify image characteristics hint,
                            one of: photo, picture or graph
 
diff --git a/README.android b/README.android
index 80d22211..1fa26e5f 100644
--- a/README.android
+++ b/README.android
@@ -1,5 +1,5 @@
 URL: https://chromium.googlesource.com/webm/libwebp
-Version: v0.5.1
+Version: v0.5.2
 License: Google BSD like
 
 Local modifications:
diff --git a/README.version b/README.version
index 3f3b0410..d9b87957 100644
--- a/README.version
+++ b/README.version
@@ -1,3 +1,3 @@
-URL: http://downloads.webmproject.org/releases/webp/libwebp-0.4.1.tar.gz
-Version: 0.4.1
+URL: https://storage.googleapis.com/downloads.webmproject.org/releases/webp/libwebp-0.5.2.tar.gz
+Version: 0.5.2
 BugComponent: 20174
diff --git a/include/webp/decode.h b/include/webp/decode.h
index 7a3bed93..4c5e74ac 100644
--- a/include/webp/decode.h
+++ b/include/webp/decode.h
@@ -248,19 +248,19 @@ typedef enum VP8StatusCode {
 // picture is only partially decoded, pending additional input.
 // Code example:
 //
-//   WebPInitDecBuffer(&buffer);
-//   buffer.colorspace = mode;
+//   WebPInitDecBuffer(&output_buffer);
+//   output_buffer.colorspace = mode;
 //   ...
-//   WebPIDecoder* idec = WebPINewDecoder(&buffer);
-//   while (has_more_data) {
-//     // ... (get additional data)
+//   WebPIDecoder* idec = WebPINewDecoder(&output_buffer);
+//   while (additional_data_is_available) {
+//     // ... (get additional data in some new_data[] buffer)
 //     status = WebPIAppend(idec, new_data, new_data_size);
-//     if (status != VP8_STATUS_SUSPENDED ||
-//       break;
+//     if (status != VP8_STATUS_OK && status != VP8_STATUS_SUSPENDED) {
+//       break;    // an error occurred.
 //     }
 //
 //     // The above call decodes the current available buffer.
-//     // Part of the image can now be refreshed by calling to
+//     // Part of the image can now be refreshed by calling
 //     // WebPIDecGetRGB()/WebPIDecGetYUVA() etc.
 //   }
 //   WebPIDelete(idec);
diff --git a/include/webp/encode.h b/include/webp/encode.h
index 9291b719..b65e27e7 100644
--- a/include/webp/encode.h
+++ b/include/webp/encode.h
@@ -481,10 +481,10 @@ WEBP_EXTERN(int) WebPPictureARGBToYUVADithered(
 WEBP_EXTERN(int) WebPPictureSmartARGBToYUVA(WebPPicture* picture);
 
 // Converts picture->yuv to picture->argb and sets picture->use_argb to true.
-// The input format must be YUV_420 or YUV_420A.
-// Note that the use of this method is discouraged if one has access to the
-// raw ARGB samples, since using YUV420 is comparatively lossy. Also, the
-// conversion from YUV420 to ARGB incurs a small loss too.
+// The input format must be YUV_420 or YUV_420A. The conversion from YUV420 to
+// ARGB incurs a small loss too.
+// Note that the use of this colorspace is discouraged if one has access to the
+// raw ARGB samples, since using YUV420 is comparatively lossy.
 // Returns false in case of error.
 WEBP_EXTERN(int) WebPPictureYUVAToARGB(WebPPicture* picture);
 
diff --git a/src/dec/alpha.c b/src/dec/alpha.c
index 028eb3d5..d88f01d8 100644
--- a/src/dec/alpha.c
+++ b/src/dec/alpha.c
@@ -67,7 +67,7 @@ static int ALPHInit(ALPHDecoder* const dec, const uint8_t* data,
   }
 
   dec->method_ = (data[0] >> 0) & 0x03;
-  dec->filter_ = (data[0] >> 2) & 0x03;
+  dec->filter_ = (WEBP_FILTER_TYPE)((data[0] >> 2) & 0x03);
   dec->pre_processing_ = (data[0] >> 4) & 0x03;
   rsrv = (data[0] >> 6) & 0x03;
   if (dec->method_ < ALPHA_NO_COMPRESSION ||
diff --git a/src/dec/vp8i.h b/src/dec/vp8i.h
index 00da02ba..313d8a7b 100644
--- a/src/dec/vp8i.h
+++ b/src/dec/vp8i.h
@@ -32,7 +32,7 @@ extern "C" {
 // version numbers
 #define DEC_MAJ_VERSION 0
 #define DEC_MIN_VERSION 5
-#define DEC_REV_VERSION 1
+#define DEC_REV_VERSION 2
 
 // YUV-cache parameters. Cache is 32-bytes wide (= one cacheline).
 // Constraints are: We need to store one 16x16 block of luma samples (y),
diff --git a/src/dec/webp.c b/src/dec/webp.c
index 57af9505..d0b912f0 100644
--- a/src/dec/webp.c
+++ b/src/dec/webp.c
@@ -415,7 +415,8 @@ static VP8StatusCode ParseHeadersInternal(const uint8_t* data,
 }
 
 VP8StatusCode WebPParseHeaders(WebPHeaderStructure* const headers) {
-  VP8StatusCode status;
+  // status is marked volatile as a workaround for a clang-3.8 (aarch64) bug
+  volatile VP8StatusCode status;
   int has_animation = 0;
   assert(headers != NULL);
   // fill out headers, ignore width/height/has_alpha.
diff --git a/src/demux/anim_decode.c b/src/demux/anim_decode.c
index 1989eb4a..f1cf176e 100644
--- a/src/demux/anim_decode.c
+++ b/src/demux/anim_decode.c
@@ -112,18 +112,15 @@ WebPAnimDecoder* WebPAnimDecoderNewInternal(
   dec->info_.bgcolor = WebPDemuxGetI(dec->demux_, WEBP_FF_BACKGROUND_COLOR);
   dec->info_.frame_count = WebPDemuxGetI(dec->demux_, WEBP_FF_FRAME_COUNT);
 
-  {
-    const int canvas_bytes =
-        dec->info_.canvas_width * NUM_CHANNELS * dec->info_.canvas_height;
-    // Note: calloc() because we fill frame with zeroes as well.
-    dec->curr_frame_ = WebPSafeCalloc(1ULL, canvas_bytes);
-    if (dec->curr_frame_ == NULL) goto Error;
-    dec->prev_frame_disposed_ = WebPSafeCalloc(1ULL, canvas_bytes);
-    if (dec->prev_frame_disposed_ == NULL) goto Error;
-  }
+  // Note: calloc() because we fill frame with zeroes as well.
+  dec->curr_frame_ = (uint8_t*)WebPSafeCalloc(
+      dec->info_.canvas_width * NUM_CHANNELS, dec->info_.canvas_height);
+  if (dec->curr_frame_ == NULL) goto Error;
+  dec->prev_frame_disposed_ = (uint8_t*)WebPSafeCalloc(
+      dec->info_.canvas_width * NUM_CHANNELS, dec->info_.canvas_height);
+  if (dec->prev_frame_disposed_ == NULL) goto Error;
 
   WebPAnimDecoderReset(dec);
-
   return dec;
 
  Error:
@@ -144,9 +141,13 @@ static int IsFullFrame(int width, int height, int canvas_width,
 }
 
 // Clear the canvas to transparent.
-static void ZeroFillCanvas(uint8_t* buf, uint32_t canvas_width,
-                           uint32_t canvas_height) {
-  memset(buf, 0, canvas_width * NUM_CHANNELS * canvas_height);
+static int ZeroFillCanvas(uint8_t* buf, uint32_t canvas_width,
+                          uint32_t canvas_height) {
+  const uint64_t size =
+      (uint64_t)canvas_width * canvas_height * NUM_CHANNELS * sizeof(*buf);
+  if (size != (size_t)size) return 0;
+  memset(buf, 0, (size_t)size);
+  return 1;
 }
 
 // Clear given frame rectangle to transparent.
@@ -162,10 +163,13 @@ static void ZeroFillFrameRect(uint8_t* buf, int buf_stride, int x_offset,
 }
 
 // Copy width * height pixels from 'src' to 'dst'.
-static void CopyCanvas(const uint8_t* src, uint8_t* dst,
-                       uint32_t width, uint32_t height) {
+static int CopyCanvas(const uint8_t* src, uint8_t* dst,
+                      uint32_t width, uint32_t height) {
+  const uint64_t size = (uint64_t)width * height * NUM_CHANNELS;
+  if (size != (size_t)size) return 0;
   assert(src != NULL && dst != NULL);
-  memcpy(dst, src, width * NUM_CHANNELS * height);
+  memcpy(dst, src, (size_t)size);
+  return 1;
 }
 
 // Returns true if the current frame is a key-frame.
@@ -328,9 +332,14 @@ int WebPAnimDecoderGetNext(WebPAnimDecoder* dec,
   is_key_frame = IsKeyFrame(&iter, &dec->prev_iter_,
                             dec->prev_frame_was_keyframe_, width, height);
   if (is_key_frame) {
-    ZeroFillCanvas(dec->curr_frame_, width, height);
+    if (!ZeroFillCanvas(dec->curr_frame_, width, height)) {
+      goto Error;
+    }
   } else {
-    CopyCanvas(dec->prev_frame_disposed_, dec->curr_frame_, width, height);
+    if (!CopyCanvas(dec->prev_frame_disposed_, dec->curr_frame_,
+                    width, height)) {
+      goto Error;
+    }
   }
 
   // Decode.
@@ -393,6 +402,7 @@ int WebPAnimDecoderGetNext(WebPAnimDecoder* dec,
 
   // Update info of the previous frame and dispose it for the next iteration.
   dec->prev_frame_timestamp_ = timestamp;
+  WebPDemuxReleaseIterator(&dec->prev_iter_);
   dec->prev_iter_ = iter;
   dec->prev_frame_was_keyframe_ = is_key_frame;
   CopyCanvas(dec->curr_frame_, dec->prev_frame_disposed_, width, height);
@@ -421,6 +431,7 @@ int WebPAnimDecoderHasMoreFrames(const WebPAnimDecoder* dec) {
 void WebPAnimDecoderReset(WebPAnimDecoder* dec) {
   if (dec != NULL) {
     dec->prev_frame_timestamp_ = 0;
+    WebPDemuxReleaseIterator(&dec->prev_iter_);
     memset(&dec->prev_iter_, 0, sizeof(dec->prev_iter_));
     dec->prev_frame_was_keyframe_ = 0;
     dec->next_frame_ = 1;
@@ -434,6 +445,7 @@ const WebPDemuxer* WebPAnimDecoderGetDemuxer(const WebPAnimDecoder* dec) {
 
 void WebPAnimDecoderDelete(WebPAnimDecoder* dec) {
   if (dec != NULL) {
+    WebPDemuxReleaseIterator(&dec->prev_iter_);
     WebPDemuxDelete(dec->demux_);
     WebPSafeFree(dec->curr_frame_);
     WebPSafeFree(dec->prev_frame_disposed_);
diff --git a/src/demux/demux.c b/src/demux/demux.c
index 0d2989f6..1cb9bd57 100644
--- a/src/demux/demux.c
+++ b/src/demux/demux.c
@@ -25,7 +25,7 @@
 
 #define DMUX_MAJ_VERSION 0
 #define DMUX_MIN_VERSION 3
-#define DMUX_REV_VERSION 0
+#define DMUX_REV_VERSION 1
 
 typedef struct {
   size_t start_;        // start location of the data
diff --git a/src/dsp/dec.c b/src/dsp/dec.c
index e92d6933..49bd16d9 100644
--- a/src/dsp/dec.c
+++ b/src/dsp/dec.c
@@ -239,7 +239,7 @@ VP8PredFunc VP8PredLuma16[NUM_B_DC_MODES];
 //------------------------------------------------------------------------------
 // 4x4
 
-#define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2)
+#define AVG3(a, b, c) ((uint8_t)(((a) + 2 * (b) + (c) + 2) >> 2))
 #define AVG2(a, b) (((a) + (b) + 1) >> 1)
 
 static void VE4(uint8_t* dst) {    // vertical
diff --git a/src/dsp/dsp.h b/src/dsp/dsp.h
index 411d3e4c..e0f59ce6 100644
--- a/src/dsp/dsp.h
+++ b/src/dsp/dsp.h
@@ -74,8 +74,9 @@ extern "C" {
 
 // The intrinsics currently cause compiler errors with arm-nacl-gcc and the
 // inline assembly would need to be modified for use with Native Client.
-#if (defined(__ARM_NEON__) || defined(__aarch64__) || \
-     defined(WEBP_HAVE_NEON)) && !defined(__native_client__)
+#if (defined(__ARM_NEON__) || \
+     defined(__aarch64__) || defined(WEBP_HAVE_NEON)) && \
+    !defined(__native_client__)
 #define WEBP_USE_NEON
 #endif
 
diff --git a/src/dsp/enc.c b/src/dsp/enc.c
index f639f557..db0e9e70 100644
--- a/src/dsp/enc.c
+++ b/src/dsp/enc.c
@@ -335,7 +335,7 @@ static void Intra16Preds(uint8_t* dst,
 // luma 4x4 prediction
 
 #define DST(x, y) dst[(x) + (y) * BPS]
-#define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2)
+#define AVG3(a, b, c) ((uint8_t)(((a) + 2 * (b) + (c) + 2) >> 2))
 #define AVG2(a, b) (((a) + (b) + 1) >> 1)
 
 static void VE4(uint8_t* dst, const uint8_t* top) {    // vertical
diff --git a/src/dsp/rescaler.c b/src/dsp/rescaler.c
index bc743d5d..f5b07756 100644
--- a/src/dsp/rescaler.c
+++ b/src/dsp/rescaler.c
@@ -173,10 +173,10 @@ void WebPRescalerExportRow(WebPRescaler* const wrk) {
       WebPRescalerExportRowExpand(wrk);
     } else if (wrk->fxy_scale) {
       WebPRescalerExportRowShrink(wrk);
-    } else {  // very special case for src = dst = 1x1
+    } else {  // special case
       int i;
+      assert(wrk->src_height == wrk->dst_height && wrk->x_add == 1);
       assert(wrk->src_width == 1 && wrk->dst_width <= 2);
-      assert(wrk->src_height == 1 && wrk->dst_height == 1);
       for (i = 0; i < wrk->num_channels * wrk->dst_width; ++i) {
         wrk->dst[i] = wrk->irow[i];
         wrk->irow[i] = 0;
diff --git a/src/enc/analysis.c b/src/enc/analysis.c
index b55128fd..136c3312 100644
--- a/src/enc/analysis.c
+++ b/src/enc/analysis.c
@@ -307,6 +307,7 @@ static int MBAnalyzeBestIntra4Mode(VP8EncIterator* const it,
 
 static int MBAnalyzeBestUVMode(VP8EncIterator* const it) {
   int best_alpha = DEFAULT_ALPHA;
+  int smallest_alpha = 0;
   int best_mode = 0;
   const int max_mode = MAX_UV_MODE;
   int mode;
@@ -322,6 +323,10 @@ static int MBAnalyzeBestUVMode(VP8EncIterator* const it) {
     alpha = GetAlpha(&histo);
     if (IS_BETTER_ALPHA(alpha, best_alpha)) {
       best_alpha = alpha;
+    }
+    // The best prediction mode tends to be the one with the smallest alpha.
+    if (mode == 0 || alpha < smallest_alpha) {
+      smallest_alpha = alpha;
       best_mode = mode;
     }
   }
diff --git a/src/enc/cost.c b/src/enc/cost.c
index ae7fe013..87f89378 100644
--- a/src/enc/cost.c
+++ b/src/enc/cost.c
@@ -281,18 +281,6 @@ int VP8GetCostUV(VP8EncIterator* const it, const VP8ModeScore* const rd) {
 //------------------------------------------------------------------------------
 // Recording of token probabilities.
 
-// Record proba context used
-static int Record(int bit, proba_t* const stats) {
-  proba_t p = *stats;
-  if (p >= 0xffff0000u) {               // an overflow is inbound.
-    p = ((p + 1u) >> 1) & 0x7fff7fffu;  // -> divide the stats by 2.
-  }
-  // record bit count (lower 16 bits) and increment total count (upper 16 bits).
-  p += 0x00010000u + bit;
-  *stats = p;
-  return bit;
-}
-
 // We keep the table-free variant around for reference, in case.
 #define USE_LEVEL_CODE_TABLE
 
@@ -303,31 +291,31 @@ int VP8RecordCoeffs(int ctx, const VP8Residual* const res) {
   // should be stats[VP8EncBands[n]], but it's equivalent for n=0 or 1
   proba_t* s = res->stats[n][ctx];
   if (res->last  < 0) {
-    Record(0, s + 0);
+    VP8RecordStats(0, s + 0);
     return 0;
   }
   while (n <= res->last) {
     int v;
-    Record(1, s + 0);  // order of record doesn't matter
+    VP8RecordStats(1, s + 0);  // order of record doesn't matter
     while ((v = res->coeffs[n++]) == 0) {
-      Record(0, s + 1);
+      VP8RecordStats(0, s + 1);
       s = res->stats[VP8EncBands[n]][0];
     }
-    Record(1, s + 1);
-    if (!Record(2u < (unsigned int)(v + 1), s + 2)) {  // v = -1 or 1
+    VP8RecordStats(1, s + 1);
+    if (!VP8RecordStats(2u < (unsigned int)(v + 1), s + 2)) {  // v = -1 or 1
       s = res->stats[VP8EncBands[n]][1];
     } else {
       v = abs(v);
 #if !defined(USE_LEVEL_CODE_TABLE)
-      if (!Record(v > 4, s + 3)) {
-        if (Record(v != 2, s + 4))
-          Record(v == 4, s + 5);
-      } else if (!Record(v > 10, s + 6)) {
-        Record(v > 6, s + 7);
-      } else if (!Record((v >= 3 + (8 << 2)), s + 8)) {
-        Record((v >= 3 + (8 << 1)), s + 9);
+      if (!VP8RecordStats(v > 4, s + 3)) {
+        if (VP8RecordStats(v != 2, s + 4))
+          VP8RecordStats(v == 4, s + 5);
+      } else if (!VP8RecordStats(v > 10, s + 6)) {
+        VP8RecordStats(v > 6, s + 7);
+      } else if (!VP8RecordStats((v >= 3 + (8 << 2)), s + 8)) {
+        VP8RecordStats((v >= 3 + (8 << 1)), s + 9);
       } else {
-        Record((v >= 3 + (8 << 3)), s + 10);
+        VP8RecordStats((v >= 3 + (8 << 3)), s + 10);
       }
 #else
       if (v > MAX_VARIABLE_LEVEL) {
@@ -340,14 +328,14 @@ int VP8RecordCoeffs(int ctx, const VP8Residual* const res) {
         int i;
         for (i = 0; (pattern >>= 1) != 0; ++i) {
           const int mask = 2 << i;
-          if (pattern & 1) Record(!!(bits & mask), s + 3 + i);
+          if (pattern & 1) VP8RecordStats(!!(bits & mask), s + 3 + i);
         }
       }
 #endif
       s = res->stats[VP8EncBands[n]][2];
     }
   }
-  if (n < 16) Record(0, s + 0);
+  if (n < 16) VP8RecordStats(0, s + 0);
   return 1;
 }
 
diff --git a/src/enc/cost.h b/src/enc/cost.h
index 20960d6d..ad7959fe 100644
--- a/src/enc/cost.h
+++ b/src/enc/cost.h
@@ -41,6 +41,20 @@ void VP8InitResidual(int first, int coeff_type,
 
 int VP8RecordCoeffs(int ctx, const VP8Residual* const res);
 
+// Record proba context used.
+static WEBP_INLINE int VP8RecordStats(int bit, proba_t* const stats) {
+  proba_t p = *stats;
+  // An overflow is inbound. Note we handle this at 0xfffe0000u instead of
+  // 0xffff0000u to make sure p + 1u does not overflow.
+  if (p >= 0xfffe0000u) {
+    p = ((p + 1u) >> 1) & 0x7fff7fffu;  // -> divide the stats by 2.
+  }
+  // record bit count (lower 16 bits) and increment total count (upper 16 bits).
+  p += 0x00010000u + bit;
+  *stats = p;
+  return bit;
+}
+
 // Cost of coding one event with probability 'proba'.
 static WEBP_INLINE int VP8BitCost(int bit, uint8_t proba) {
   return !bit ? VP8EntropyCost[proba] : VP8EntropyCost[255 - proba];
diff --git a/src/enc/frame.c b/src/enc/frame.c
index 5b7a40b9..57fc471d 100644
--- a/src/enc/frame.c
+++ b/src/enc/frame.c
@@ -185,6 +185,13 @@ static int GetProba(int a, int b) {
                       : (255 * a + total / 2) / total;  // rounded proba
 }
 
+static void ResetSegments(VP8Encoder* const enc) {
+  int n;
+  for (n = 0; n < enc->mb_w_ * enc->mb_h_; ++n) {
+    enc->mb_info_[n].segment_ = 0;
+  }
+}
+
 static void SetSegmentProbas(VP8Encoder* const enc) {
   int p[NUM_MB_SEGMENTS] = { 0 };
   int n;
@@ -206,6 +213,7 @@ static void SetSegmentProbas(VP8Encoder* const enc) {
 
     enc->segment_hdr_.update_map_ =
         (probas[0] != 255) || (probas[1] != 255) || (probas[2] != 255);
+    if (!enc->segment_hdr_.update_map_) ResetSegments(enc);
     enc->segment_hdr_.size_ =
         p[0] * (VP8BitCost(0, probas[0]) + VP8BitCost(0, probas[1])) +
         p[1] * (VP8BitCost(0, probas[0]) + VP8BitCost(1, probas[1])) +
@@ -406,9 +414,7 @@ static int RecordTokens(VP8EncIterator* const it, const VP8ModeScore* const rd,
     VP8InitResidual(0, 1, enc, &res);
     VP8SetResidualCoeffs(rd->y_dc_levels, &res);
     it->top_nz_[8] = it->left_nz_[8] =
-        VP8RecordCoeffTokens(ctx, 1,
-                             res.first, res.last, res.coeffs, tokens);
-    VP8RecordCoeffs(ctx, &res);
+        VP8RecordCoeffTokens(ctx, &res, tokens);
     VP8InitResidual(1, 0, enc, &res);
   } else {
     VP8InitResidual(0, 3, enc, &res);
@@ -420,9 +426,7 @@ static int RecordTokens(VP8EncIterator* const it, const VP8ModeScore* const rd,
       const int ctx = it->top_nz_[x] + it->left_nz_[y];
       VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
       it->top_nz_[x] = it->left_nz_[y] =
-          VP8RecordCoeffTokens(ctx, res.coeff_type,
-                               res.first, res.last, res.coeffs, tokens);
-      VP8RecordCoeffs(ctx, &res);
+          VP8RecordCoeffTokens(ctx, &res, tokens);
     }
   }
 
@@ -434,9 +438,7 @@ static int RecordTokens(VP8EncIterator* const it, const VP8ModeScore* const rd,
         const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
         VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
         it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
-            VP8RecordCoeffTokens(ctx, 2,
-                                 res.first, res.last, res.coeffs, tokens);
-        VP8RecordCoeffs(ctx, &res);
+            VP8RecordCoeffTokens(ctx, &res, tokens);
       }
     }
   }
@@ -814,7 +816,7 @@ int VP8EncTokenLoop(VP8Encoder* const enc) {
            num_pass_left, stats.last_value, stats.value,
            stats.last_q, stats.q, stats.dq);
 #endif
-    if (size_p0 > PARTITION0_SIZE_LIMIT) {
+    if (enc->max_i4_header_bits_ > 0 && size_p0 > PARTITION0_SIZE_LIMIT) {
       ++num_pass_left;
       enc->max_i4_header_bits_ >>= 1;  // strengthen header bit limitation...
       continue;                        // ...and start over
diff --git a/src/enc/histogram.c b/src/enc/histogram.c
index 395372b2..36b7f226 100644
--- a/src/enc/histogram.c
+++ b/src/enc/histogram.c
@@ -592,8 +592,8 @@ static int HistoQueueInit(HistoQueue* const histo_queue, const int max_index) {
   histo_queue->max_size = max_index * max_index;
   // We allocate max_size + 1 because the last element at index "size" is
   // used as temporary data (and it could be up to max_size).
-  histo_queue->queue = WebPSafeMalloc(histo_queue->max_size + 1,
-                                      sizeof(*histo_queue->queue));
+  histo_queue->queue = (HistogramPair*)WebPSafeMalloc(
+      histo_queue->max_size + 1, sizeof(*histo_queue->queue));
   return histo_queue->queue != NULL;
 }
 
@@ -659,7 +659,8 @@ static int HistogramCombineGreedy(VP8LHistogramSet* const image_histo) {
   int i, j;
   VP8LHistogram** const histograms = image_histo->histograms;
   // Indexes of remaining histograms.
-  int* const clusters = WebPSafeMalloc(image_histo_size, sizeof(*clusters));
+  int* const clusters =
+      (int*)WebPSafeMalloc(image_histo_size, sizeof(*clusters));
   // Priority queue of histogram pairs.
   HistoQueue histo_queue;
 
diff --git a/src/enc/picture.c b/src/enc/picture.c
index d9befbc4..28c56cd6 100644
--- a/src/enc/picture.c
+++ b/src/enc/picture.c
@@ -88,8 +88,9 @@ int WebPPictureAllocARGB(WebPPicture* const picture, int width, int height) {
 }
 
 int WebPPictureAllocYUVA(WebPPicture* const picture, int width, int height) {
-  const WebPEncCSP uv_csp = picture->colorspace & WEBP_CSP_UV_MASK;
-  const int has_alpha = picture->colorspace & WEBP_CSP_ALPHA_BIT;
+  const WebPEncCSP uv_csp =
+      (WebPEncCSP)((int)picture->colorspace & WEBP_CSP_UV_MASK);
+  const int has_alpha = (int)picture->colorspace & WEBP_CSP_ALPHA_BIT;
   const int y_stride = width;
   const int uv_width = (width + 1) >> 1;
   const int uv_height = (height + 1) >> 1;
diff --git a/src/enc/picture_csp.c b/src/enc/picture_csp.c
index 607a6240..188a3ca5 100644
--- a/src/enc/picture_csp.c
+++ b/src/enc/picture_csp.c
@@ -381,36 +381,42 @@ static WEBP_INLINE uint8_t ConvertRGBToV(int r, int g, int b) {
   return clip_8b(128 + (v >> (YUV_FIX + SFIX)));
 }
 
-static int ConvertWRGBToYUV(const fixed_y_t* const best_y,
-                            const fixed_t* const best_uv,
+static int ConvertWRGBToYUV(const fixed_y_t* best_y, const fixed_t* best_uv,
                             WebPPicture* const picture) {
   int i, j;
+  uint8_t* dst_y = picture->y;
+  uint8_t* dst_u = picture->u;
+  uint8_t* dst_v = picture->v;
+  const fixed_t* const best_uv_base = best_uv;
   const int w = (picture->width + 1) & ~1;
   const int h = (picture->height + 1) & ~1;
   const int uv_w = w >> 1;
   const int uv_h = h >> 1;
-  for (j = 0; j < picture->height; ++j) {
+  for (best_uv = best_uv_base, j = 0; j < picture->height; ++j) {
     for (i = 0; i < picture->width; ++i) {
-      const int off = 3 * ((i >> 1) + (j >> 1) * uv_w);
-      const int off2 = i + j * picture->y_stride;
-      const int W = best_y[i + j * w];
+      const int off = 3 * (i >> 1);
+      const int W = best_y[i];
       const int r = best_uv[off + 0] + W;
       const int g = best_uv[off + 1] + W;
       const int b = best_uv[off + 2] + W;
-      picture->y[off2] = ConvertRGBToY(r, g, b);
+      dst_y[i] = ConvertRGBToY(r, g, b);
     }
+    best_y += w;
+    best_uv += (j & 1) * 3 * uv_w;
+    dst_y += picture->y_stride;
   }
-  for (j = 0; j < uv_h; ++j) {
-    uint8_t* const dst_u = picture->u + j * picture->uv_stride;
-    uint8_t* const dst_v = picture->v + j * picture->uv_stride;
+  for (best_uv = best_uv_base, j = 0; j < uv_h; ++j) {
     for (i = 0; i < uv_w; ++i) {
-      const int off = 3 * (i + j * uv_w);
+      const int off = 3 * i;
       const int r = best_uv[off + 0];
       const int g = best_uv[off + 1];
       const int b = best_uv[off + 2];
       dst_u[i] = ConvertRGBToU(r, g, b);
       dst_v[i] = ConvertRGBToV(r, g, b);
     }
+    best_uv += 3 * uv_w;
+    dst_u += picture->uv_stride;
+    dst_v += picture->uv_stride;
   }
   return 1;
 }
@@ -420,9 +426,9 @@ static int ConvertWRGBToYUV(const fixed_y_t* const best_y,
 
 #define SAFE_ALLOC(W, H, T) ((T*)WebPSafeMalloc((W) * (H), sizeof(T)))
 
-static int PreprocessARGB(const uint8_t* const r_ptr,
-                          const uint8_t* const g_ptr,
-                          const uint8_t* const b_ptr,
+static int PreprocessARGB(const uint8_t* r_ptr,
+                          const uint8_t* g_ptr,
+                          const uint8_t* b_ptr,
                           int step, int rgb_stride,
                           WebPPicture* const picture) {
   // we expand the right/bottom border if needed
@@ -435,20 +441,24 @@ static int PreprocessARGB(const uint8_t* const r_ptr,
   // TODO(skal): allocate one big memory chunk. But for now, it's easier
   // for valgrind debugging to have several chunks.
   fixed_y_t* const tmp_buffer = SAFE_ALLOC(w * 3, 2, fixed_y_t);   // scratch
-  fixed_y_t* const best_y = SAFE_ALLOC(w, h, fixed_y_t);
-  fixed_y_t* const target_y = SAFE_ALLOC(w, h, fixed_y_t);
+  fixed_y_t* const best_y_base = SAFE_ALLOC(w, h, fixed_y_t);
+  fixed_y_t* const target_y_base = SAFE_ALLOC(w, h, fixed_y_t);
   fixed_y_t* const best_rgb_y = SAFE_ALLOC(w, 2, fixed_y_t);
-  fixed_t* const best_uv = SAFE_ALLOC(uv_w * 3, uv_h, fixed_t);
-  fixed_t* const target_uv = SAFE_ALLOC(uv_w * 3, uv_h, fixed_t);
+  fixed_t* const best_uv_base = SAFE_ALLOC(uv_w * 3, uv_h, fixed_t);
+  fixed_t* const target_uv_base = SAFE_ALLOC(uv_w * 3, uv_h, fixed_t);
   fixed_t* const best_rgb_uv = SAFE_ALLOC(uv_w * 3, 1, fixed_t);
+  fixed_y_t* best_y = best_y_base;
+  fixed_y_t* target_y = target_y_base;
+  fixed_t* best_uv = best_uv_base;
+  fixed_t* target_uv = target_uv_base;
   int ok;
   int diff_sum = 0;
   const int first_diff_threshold = (int)(2.5 * w * h);
   const int min_improvement = 5;   // stop if improvement is below this %
   const int min_first_improvement = 80;
 
-  if (best_y == NULL || best_uv == NULL ||
-      target_y == NULL || target_uv == NULL ||
+  if (best_y_base == NULL || best_uv_base == NULL ||
+      target_y_base == NULL || target_uv_base == NULL ||
       best_rgb_y == NULL || best_rgb_uv == NULL ||
       tmp_buffer == NULL) {
     ok = WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
@@ -462,41 +472,47 @@ static int PreprocessARGB(const uint8_t* const r_ptr,
     const int is_last_row = (j == picture->height - 1);
     fixed_y_t* const src1 = tmp_buffer;
     fixed_y_t* const src2 = tmp_buffer + 3 * w;
-    const int off1 = j * rgb_stride;
-    const int off2 = off1 + rgb_stride;
-    const int uv_off = (j >> 1) * 3 * uv_w;
-    fixed_y_t* const dst_y = best_y + j * w;
 
     // prepare two rows of input
-    ImportOneRow(r_ptr + off1, g_ptr + off1, b_ptr + off1,
-                 step, picture->width, src1);
+    ImportOneRow(r_ptr, g_ptr, b_ptr, step, picture->width, src1);
     if (!is_last_row) {
-      ImportOneRow(r_ptr + off2, g_ptr + off2, b_ptr + off2,
+      ImportOneRow(r_ptr + rgb_stride, g_ptr + rgb_stride, b_ptr + rgb_stride,
                    step, picture->width, src2);
     } else {
       memcpy(src2, src1, 3 * w * sizeof(*src2));
     }
-    UpdateW(src1, target_y + (j + 0) * w, w);
-    UpdateW(src2, target_y + (j + 1) * w, w);
-    diff_sum += UpdateChroma(src1, src2, target_uv + uv_off, dst_y, uv_w);
-    memcpy(best_uv + uv_off, target_uv + uv_off, 3 * uv_w * sizeof(*best_uv));
-    memcpy(dst_y + w, dst_y, w * sizeof(*dst_y));
+    UpdateW(src1, target_y, w);
+    UpdateW(src2, target_y + w, w);
+    diff_sum += UpdateChroma(src1, src2, target_uv, best_y, uv_w);
+    memcpy(best_uv, target_uv, 3 * uv_w * sizeof(*best_uv));
+    memcpy(best_y + w, best_y, w * sizeof(*best_y));
+    best_y += 2 * w;
+    best_uv += 3 * uv_w;
+    target_y += 2 * w;
+    target_uv += 3 * uv_w;
+    r_ptr += 2 * rgb_stride;
+    g_ptr += 2 * rgb_stride;
+    b_ptr += 2 * rgb_stride;
   }
 
   // Iterate and resolve clipping conflicts.
   for (iter = 0; iter < kNumIterations; ++iter) {
     int k;
-    const fixed_t* cur_uv = best_uv;
-    const fixed_t* prev_uv = best_uv;
+    const fixed_t* cur_uv = best_uv_base;
+    const fixed_t* prev_uv = best_uv_base;
     const int old_diff_sum = diff_sum;
     diff_sum = 0;
+
+    best_y = best_y_base;
+    best_uv = best_uv_base;
+    target_y = target_y_base;
+    target_uv = target_uv_base;
     for (j = 0; j < h; j += 2) {
       fixed_y_t* const src1 = tmp_buffer;
       fixed_y_t* const src2 = tmp_buffer + 3 * w;
       {
         const fixed_t* const next_uv = cur_uv + ((j < h - 2) ? 3 * uv_w : 0);
-        InterpolateTwoRows(best_y + j * w, prev_uv, cur_uv, next_uv,
-                           w, src1, src2);
+        InterpolateTwoRows(best_y, prev_uv, cur_uv, next_uv, w, src1, src2);
         prev_uv = cur_uv;
         cur_uv = next_uv;
       }
@@ -507,16 +523,15 @@ static int PreprocessARGB(const uint8_t* const r_ptr,
 
       // update two rows of Y and one row of RGB
       for (i = 0; i < 2 * w; ++i) {
-        const int off = i + j * w;
-        const int diff_y = target_y[off] - best_rgb_y[i];
-        const int new_y = (int)best_y[off] + diff_y;
-        best_y[off] = clip_y(new_y);
+        const int diff_y = target_y[i] - best_rgb_y[i];
+        const int new_y = (int)best_y[i] + diff_y;
+        best_y[i] = clip_y(new_y);
       }
       for (i = 0; i < uv_w; ++i) {
-        const int off = 3 * (i + (j >> 1) * uv_w);
+        const int off = 3 * i;
         int W;
         for (k = 0; k <= 2; ++k) {
-          const int diff_uv = (int)target_uv[off + k] - best_rgb_uv[3 * i + k];
+          const int diff_uv = (int)target_uv[off + k] - best_rgb_uv[off + k];
           best_uv[off + k] += diff_uv;
         }
         W = RGBToGray(best_uv[off + 0], best_uv[off + 1], best_uv[off + 2]);
@@ -524,6 +539,10 @@ static int PreprocessARGB(const uint8_t* const r_ptr,
           best_uv[off + k] -= W;
         }
       }
+      best_y += 2 * w;
+      best_uv += 3 * uv_w;
+      target_y += 2 * w;
+      target_uv += 3 * uv_w;
     }
     // test exit condition
     if (diff_sum > 0) {
@@ -545,13 +564,13 @@ static int PreprocessARGB(const uint8_t* const r_ptr,
   }
 
   // final reconstruction
-  ok = ConvertWRGBToYUV(best_y, best_uv, picture);
+  ok = ConvertWRGBToYUV(best_y_base, best_uv_base, picture);
 
  End:
-  WebPSafeFree(best_y);
-  WebPSafeFree(best_uv);
-  WebPSafeFree(target_y);
-  WebPSafeFree(target_uv);
+  WebPSafeFree(best_y_base);
+  WebPSafeFree(best_uv_base);
+  WebPSafeFree(target_y_base);
+  WebPSafeFree(target_uv_base);
   WebPSafeFree(best_rgb_y);
   WebPSafeFree(best_rgb_uv);
   WebPSafeFree(tmp_buffer);
@@ -830,10 +849,10 @@ static WEBP_INLINE void ConvertRowsToUV(const uint16_t* rgb,
   }
 }
 
-static int ImportYUVAFromRGBA(const uint8_t* const r_ptr,
-                              const uint8_t* const g_ptr,
-                              const uint8_t* const b_ptr,
-                              const uint8_t* const a_ptr,
+static int ImportYUVAFromRGBA(const uint8_t* r_ptr,
+                              const uint8_t* g_ptr,
+                              const uint8_t* b_ptr,
+                              const uint8_t* a_ptr,
                               int step,         // bytes per pixel
                               int rgb_stride,   // bytes per scanline
                               float dithering,
@@ -900,36 +919,34 @@ static int ImportYUVAFromRGBA(const uint8_t* const r_ptr,
     // Downsample Y/U/V planes, two rows at a time
     for (y = 0; y < (height >> 1); ++y) {
       int rows_have_alpha = has_alpha;
-      const int off1 = (2 * y + 0) * rgb_stride;
-      const int off2 = (2 * y + 1) * rgb_stride;
       if (use_dsp) {
         if (is_rgb) {
-          WebPConvertRGB24ToY(r_ptr + off1, dst_y, width);
-          WebPConvertRGB24ToY(r_ptr + off2, dst_y + picture->y_stride, width);
+          WebPConvertRGB24ToY(r_ptr, dst_y, width);
+          WebPConvertRGB24ToY(r_ptr + rgb_stride,
+                              dst_y + picture->y_stride, width);
         } else {
-          WebPConvertBGR24ToY(b_ptr + off1, dst_y, width);
-          WebPConvertBGR24ToY(b_ptr + off2, dst_y + picture->y_stride, width);
+          WebPConvertBGR24ToY(b_ptr, dst_y, width);
+          WebPConvertBGR24ToY(b_ptr + rgb_stride,
+                              dst_y + picture->y_stride, width);
         }
       } else {
-        ConvertRowToY(r_ptr + off1, g_ptr + off1, b_ptr + off1, step,
-                      dst_y, width, rg);
-        ConvertRowToY(r_ptr + off2, g_ptr + off2, b_ptr + off2, step,
+        ConvertRowToY(r_ptr, g_ptr, b_ptr, step, dst_y, width, rg);
+        ConvertRowToY(r_ptr + rgb_stride,
+                      g_ptr + rgb_stride,
+                      b_ptr + rgb_stride, step,
                       dst_y + picture->y_stride, width, rg);
       }
       dst_y += 2 * picture->y_stride;
       if (has_alpha) {
-        rows_have_alpha &= !WebPExtractAlpha(a_ptr + off1, rgb_stride,
-                                             width, 2,
+        rows_have_alpha &= !WebPExtractAlpha(a_ptr, rgb_stride, width, 2,
                                              dst_a, picture->a_stride);
         dst_a += 2 * picture->a_stride;
       }
       // Collect averaged R/G/B(/A)
       if (!rows_have_alpha) {
-        AccumulateRGB(r_ptr + off1, g_ptr + off1, b_ptr + off1,
-                      step, rgb_stride, tmp_rgb, width);
+        AccumulateRGB(r_ptr, g_ptr, b_ptr, step, rgb_stride, tmp_rgb, width);
       } else {
-        AccumulateRGBA(r_ptr + off1, g_ptr + off1, b_ptr + off1, a_ptr + off1,
-                       rgb_stride, tmp_rgb, width);
+        AccumulateRGBA(r_ptr, g_ptr, b_ptr, a_ptr, rgb_stride, tmp_rgb, width);
       }
       // Convert to U/V
       if (rg == NULL) {
@@ -939,31 +956,33 @@ static int ImportYUVAFromRGBA(const uint8_t* const r_ptr,
       }
       dst_u += picture->uv_stride;
       dst_v += picture->uv_stride;
+      r_ptr += 2 * rgb_stride;
+      b_ptr += 2 * rgb_stride;
+      g_ptr += 2 * rgb_stride;
+      if (has_alpha) a_ptr += 2 * rgb_stride;
     }
     if (height & 1) {    // extra last row
-      const int off = 2 * y * rgb_stride;
       int row_has_alpha = has_alpha;
       if (use_dsp) {
         if (r_ptr < b_ptr) {
-          WebPConvertRGB24ToY(r_ptr + off, dst_y, width);
+          WebPConvertRGB24ToY(r_ptr, dst_y, width);
         } else {
-          WebPConvertBGR24ToY(b_ptr + off, dst_y, width);
+          WebPConvertBGR24ToY(b_ptr, dst_y, width);
         }
       } else {
-        ConvertRowToY(r_ptr + off, g_ptr + off, b_ptr + off, step,
-                      dst_y, width, rg);
+        ConvertRowToY(r_ptr, g_ptr, b_ptr, step, dst_y, width, rg);
       }
       if (row_has_alpha) {
-        row_has_alpha &= !WebPExtractAlpha(a_ptr + off, 0, width, 1, dst_a, 0);
+        row_has_alpha &= !WebPExtractAlpha(a_ptr, 0, width, 1, dst_a, 0);
       }
       // Collect averaged R/G/B(/A)
       if (!row_has_alpha) {
         // Collect averaged R/G/B
-        AccumulateRGB(r_ptr + off, g_ptr + off, b_ptr + off,
-                      step, /* rgb_stride = */ 0, tmp_rgb, width);
+        AccumulateRGB(r_ptr, g_ptr, b_ptr, step, /* rgb_stride = */ 0,
+                      tmp_rgb, width);
       } else {
-        AccumulateRGBA(r_ptr + off, g_ptr + off, b_ptr + off, a_ptr + off,
-                       /* rgb_stride = */ 0, tmp_rgb, width);
+        AccumulateRGBA(r_ptr, g_ptr, b_ptr, a_ptr, /* rgb_stride = */ 0,
+                       tmp_rgb, width);
       }
       if (rg == NULL) {
         WebPConvertRGBA32ToUV(tmp_rgb, dst_u, dst_v, uv_width);
@@ -1086,10 +1105,10 @@ static int Import(WebPPicture* const picture,
                   const uint8_t* const rgb, int rgb_stride,
                   int step, int swap_rb, int import_alpha) {
   int y;
-  const uint8_t* const r_ptr = rgb + (swap_rb ? 2 : 0);
-  const uint8_t* const g_ptr = rgb + 1;
-  const uint8_t* const b_ptr = rgb + (swap_rb ? 0 : 2);
-  const uint8_t* const a_ptr = import_alpha ? rgb + 3 : NULL;
+  const uint8_t* r_ptr = rgb + (swap_rb ? 2 : 0);
+  const uint8_t* g_ptr = rgb + 1;
+  const uint8_t* b_ptr = rgb + (swap_rb ? 0 : 2);
+  const uint8_t* a_ptr = import_alpha ? rgb + 3 : NULL;
   const int width = picture->width;
   const int height = picture->height;
 
@@ -1102,20 +1121,25 @@ static int Import(WebPPicture* const picture,
   VP8EncDspARGBInit();
 
   if (import_alpha) {
+    uint32_t* dst = picture->argb;
     assert(step == 4);
     for (y = 0; y < height; ++y) {
-      uint32_t* const dst = &picture->argb[y * picture->argb_stride];
-      const int offset = y * rgb_stride;
-      VP8PackARGB(a_ptr + offset, r_ptr + offset, g_ptr + offset,
-                  b_ptr + offset, width, dst);
+      VP8PackARGB(a_ptr, r_ptr, g_ptr, b_ptr, width, dst);
+      a_ptr += rgb_stride;
+      r_ptr += rgb_stride;
+      g_ptr += rgb_stride;
+      b_ptr += rgb_stride;
+      dst += picture->argb_stride;
     }
   } else {
+    uint32_t* dst = picture->argb;
     assert(step >= 3);
     for (y = 0; y < height; ++y) {
-      uint32_t* const dst = &picture->argb[y * picture->argb_stride];
-      const int offset = y * rgb_stride;
-      VP8PackRGB(r_ptr + offset, g_ptr + offset, b_ptr + offset,
-                 width, step, dst);
+      VP8PackRGB(r_ptr, g_ptr, b_ptr, width, step, dst);
+      r_ptr += rgb_stride;
+      g_ptr += rgb_stride;
+      b_ptr += rgb_stride;
+      dst += picture->argb_stride;
     }
   }
   return 1;
diff --git a/src/enc/picture_psnr.c b/src/enc/picture_psnr.c
index 81ab1b5c..329757de 100644
--- a/src/enc/picture_psnr.c
+++ b/src/enc/picture_psnr.c
@@ -110,7 +110,7 @@ int WebPPictureDistortion(const WebPPicture* src, const WebPPicture* ref,
           VP8SSIMAccumulatePlane(tmp1, w, tmp2, w, w, h, &stats[c]);
         }
       }
-      free(tmp_plane);
+      WebPSafeFree(tmp_plane);
     }
   } else {
     int has_alpha, uv_w, uv_h;
diff --git a/src/enc/quant.c b/src/enc/quant.c
index 549ad26f..07ffaf0a 100644
--- a/src/enc/quant.c
+++ b/src/enc/quant.c
@@ -278,7 +278,7 @@ static void SetupMatrices(VP8Encoder* enc) {
     CheckLambdaValue(&m->lambda_trellis_uv_);
     CheckLambdaValue(&m->tlambda_);
 
-    m->min_disto_ = 10 * m->y1_.q_[0];   // quantization-aware min disto
+    m->min_disto_ = 20 * m->y1_.q_[0];   // quantization-aware min disto
     m->max_edge_  = 0;
 
     m->i4_penalty_ = 1000 * q_i4 * q_i4;
@@ -874,9 +874,9 @@ static void StoreMaxDelta(VP8SegmentInfo* const dqm, const int16_t DCs[16]) {
   // We look at the first three AC coefficients to determine what is the average
   // delta between each sub-4x4 block.
   const int v0 = abs(DCs[1]);
-  const int v1 = abs(DCs[4]);
-  const int v2 = abs(DCs[5]);
-  int max_v = (v0 > v1) ? v1 : v0;
+  const int v1 = abs(DCs[2]);
+  const int v2 = abs(DCs[4]);
+  int max_v = (v1 > v0) ? v1 : v0;
   max_v = (v2 > max_v) ? v2 : max_v;
   if (max_v > dqm->max_edge_) dqm->max_edge_ = max_v;
 }
@@ -957,7 +957,7 @@ static void PickBestIntra16(VP8EncIterator* const it, VP8ModeScore* rd) {
   // we have a blocky macroblock (only DCs are non-zero) with fairly high
   // distortion, record max delta so we can later adjust the minimal filtering
   // strength needed to smooth these blocks out.
-  if ((rd->nz & 0xffff) == 0 && rd->D > dqm->min_disto_) {
+  if ((rd->nz & 0x100ffff) == 0x1000000 && rd->D > dqm->min_disto_) {
     StoreMaxDelta(dqm, rd->y_dc_levels);
   }
 }
@@ -1155,7 +1155,8 @@ static void RefineUsingDistortion(VP8EncIterator* const it,
   const int lambda_d_uv = 120;
   score_t score_i4 = dqm->i4_penalty_;
   score_t i4_bit_sum = 0;
-  const score_t bit_limit = it->enc_->mb_header_limit_;
+  const score_t bit_limit = try_both_modes ? it->enc_->mb_header_limit_
+                                           : MAX_COST;  // no early-out allowed
 
   if (is_i16) {   // First, evaluate Intra16 distortion
     int best_mode = -1;
diff --git a/src/enc/token.c b/src/enc/token.c
index e73256b3..087940e5 100644
--- a/src/enc/token.c
+++ b/src/enc/token.c
@@ -87,14 +87,16 @@ static int TBufferNewPage(VP8TBuffer* const b) {
 #define TOKEN_ID(t, b, ctx) \
     (NUM_PROBAS * ((ctx) + NUM_CTX * ((b) + NUM_BANDS * (t))))
 
-static WEBP_INLINE uint32_t AddToken(VP8TBuffer* const b,
-                                     uint32_t bit, uint32_t proba_idx) {
+static WEBP_INLINE uint32_t AddToken(VP8TBuffer* const b, uint32_t bit,
+                                     uint32_t proba_idx,
+                                     proba_t* const stats) {
   assert(proba_idx < FIXED_PROBA_BIT);
   assert(bit <= 1);
   if (b->left_ > 0 || TBufferNewPage(b)) {
     const int slot = --b->left_;
     b->tokens_[slot] = (bit << 15) | proba_idx;
   }
+  VP8RecordStats(bit, stats);
   return bit;
 }
 
@@ -108,13 +110,16 @@ static WEBP_INLINE void AddConstantToken(VP8TBuffer* const b,
   }
 }
 
-int VP8RecordCoeffTokens(const int ctx, const int coeff_type,
-                         int first, int last,
-                         const int16_t* const coeffs,
+int VP8RecordCoeffTokens(int ctx, const struct VP8Residual* const res,
                          VP8TBuffer* const tokens) {
-  int n = first;
+  const int16_t* const coeffs = res->coeffs;
+  const int coeff_type = res->coeff_type;
+  const int last = res->last;
+  int n = res->first;
   uint32_t base_id = TOKEN_ID(coeff_type, n, ctx);
-  if (!AddToken(tokens, last >= 0, base_id + 0)) {
+  // should be stats[VP8EncBands[n]], but it's equivalent for n=0 or 1
+  proba_t* s = res->stats[n][ctx];
+  if (!AddToken(tokens, last >= 0, base_id + 0, s + 0)) {
     return 0;
   }
 
@@ -122,18 +127,20 @@ int VP8RecordCoeffTokens(const int ctx, const int coeff_type,
     const int c = coeffs[n++];
     const int sign = c < 0;
     const uint32_t v = sign ? -c : c;
-    if (!AddToken(tokens, v != 0, base_id + 1)) {
+    if (!AddToken(tokens, v != 0, base_id + 1, s + 1)) {
       base_id = TOKEN_ID(coeff_type, VP8EncBands[n], 0);  // ctx=0
+      s = res->stats[VP8EncBands[n]][0];
       continue;
     }
-    if (!AddToken(tokens, v > 1, base_id + 2)) {
+    if (!AddToken(tokens, v > 1, base_id + 2, s + 2)) {
       base_id = TOKEN_ID(coeff_type, VP8EncBands[n], 1);  // ctx=1
+      s = res->stats[VP8EncBands[n]][1];
     } else {
-      if (!AddToken(tokens, v > 4, base_id + 3)) {
-        if (AddToken(tokens, v != 2, base_id + 4))
-          AddToken(tokens, v == 4, base_id + 5);
-      } else if (!AddToken(tokens, v > 10, base_id + 6)) {
-        if (!AddToken(tokens, v > 6, base_id + 7)) {
+      if (!AddToken(tokens, v > 4, base_id + 3, s + 3)) {
+        if (AddToken(tokens, v != 2, base_id + 4, s + 4))
+          AddToken(tokens, v == 4, base_id + 5, s + 5);
+      } else if (!AddToken(tokens, v > 10, base_id + 6, s + 6)) {
+        if (!AddToken(tokens, v > 6, base_id + 7, s + 7)) {
           AddConstantToken(tokens, v == 6, 159);
         } else {
           AddConstantToken(tokens, v >= 9, 165);
@@ -144,26 +151,26 @@ int VP8RecordCoeffTokens(const int ctx, const int coeff_type,
         const uint8_t* tab;
         uint32_t residue = v - 3;
         if (residue < (8 << 1)) {          // VP8Cat3  (3b)
-          AddToken(tokens, 0, base_id + 8);
-          AddToken(tokens, 0, base_id + 9);
+          AddToken(tokens, 0, base_id + 8, s + 8);
+          AddToken(tokens, 0, base_id + 9, s + 9);
           residue -= (8 << 0);
           mask = 1 << 2;
           tab = VP8Cat3;
         } else if (residue < (8 << 2)) {   // VP8Cat4  (4b)
-          AddToken(tokens, 0, base_id + 8);
-          AddToken(tokens, 1, base_id + 9);
+          AddToken(tokens, 0, base_id + 8, s + 8);
+          AddToken(tokens, 1, base_id + 9, s + 9);
           residue -= (8 << 1);
           mask = 1 << 3;
           tab = VP8Cat4;
         } else if (residue < (8 << 3)) {   // VP8Cat5  (5b)
-          AddToken(tokens, 1, base_id + 8);
-          AddToken(tokens, 0, base_id + 10);
+          AddToken(tokens, 1, base_id + 8, s + 8);
+          AddToken(tokens, 0, base_id + 10, s + 9);
           residue -= (8 << 2);
           mask = 1 << 4;
           tab = VP8Cat5;
         } else {                         // VP8Cat6 (11b)
-          AddToken(tokens, 1, base_id + 8);
-          AddToken(tokens, 1, base_id + 10);
+          AddToken(tokens, 1, base_id + 8, s + 8);
+          AddToken(tokens, 1, base_id + 10, s + 9);
           residue -= (8 << 3);
           mask = 1 << 10;
           tab = VP8Cat6;
@@ -174,9 +181,10 @@ int VP8RecordCoeffTokens(const int ctx, const int coeff_type,
         }
       }
       base_id = TOKEN_ID(coeff_type, VP8EncBands[n], 2);  // ctx=2
+      s = res->stats[VP8EncBands[n]][2];
     }
     AddConstantToken(tokens, sign, 128);
-    if (n == 16 || !AddToken(tokens, n <= last, base_id + 0)) {
+    if (n == 16 || !AddToken(tokens, n <= last, base_id + 0, s + 0)) {
       return 1;   // EOB
     }
   }
diff --git a/src/enc/vp8enci.h b/src/enc/vp8enci.h
index c1fbd764..5b4e162a 100644
--- a/src/enc/vp8enci.h
+++ b/src/enc/vp8enci.h
@@ -32,7 +32,7 @@ extern "C" {
 // version numbers
 #define ENC_MAJ_VERSION 0
 #define ENC_MIN_VERSION 5
-#define ENC_REV_VERSION 1
+#define ENC_REV_VERSION 2
 
 enum { MAX_LF_LEVELS = 64,       // Maximum loop filter level
        MAX_VARIABLE_LEVEL = 67,  // last (inclusive) level with variable cost
@@ -325,9 +325,7 @@ int VP8EmitTokens(VP8TBuffer* const b, VP8BitWriter* const bw,
                   const uint8_t* const probas, int final_pass);
 
 // record the coding of coefficients without knowing the probabilities yet
-int VP8RecordCoeffTokens(const int ctx, const int coeff_type,
-                         int first, int last,
-                         const int16_t* const coeffs,
+int VP8RecordCoeffTokens(int ctx, const struct VP8Residual* const res,
                          VP8TBuffer* const tokens);
 
 // Estimate the final coded size given a set of 'probas'.
diff --git a/src/enc/vp8l.c b/src/enc/vp8l.c
index c16e2560..e4ad2959 100644
--- a/src/enc/vp8l.c
+++ b/src/enc/vp8l.c
@@ -34,8 +34,8 @@
 // Palette reordering for smaller sum of deltas (and for smaller storage).
 
 static int PaletteCompareColorsForQsort(const void* p1, const void* p2) {
-  const uint32_t a = WebPMemToUint32(p1);
-  const uint32_t b = WebPMemToUint32(p2);
+  const uint32_t a = WebPMemToUint32((uint8_t*)p1);
+  const uint32_t b = WebPMemToUint32((uint8_t*)p2);
   assert(a != b);
   return (a < b) ? -1 : 1;
 }
@@ -224,9 +224,8 @@ static int AnalyzeEntropy(const uint32_t* argb,
     {
       double entropy_comp[kHistoTotal];
       double entropy[kNumEntropyIx];
-      EntropyIx k;
-      EntropyIx last_mode_to_analyze =
-          use_palette ? kPalette : kSpatialSubGreen;
+      int k;
+      int last_mode_to_analyze = use_palette ? kPalette : kSpatialSubGreen;
       int j;
       // Let's add one zero to the predicted histograms. The zeros are removed
       // too efficiently by the pix_diff == 0 comparison, at least one of the
@@ -263,7 +262,7 @@ static int AnalyzeEntropy(const uint32_t* argb,
       *min_entropy_ix = kDirect;
       for (k = kDirect + 1; k <= last_mode_to_analyze; ++k) {
         if (entropy[*min_entropy_ix] > entropy[k]) {
-          *min_entropy_ix = k;
+          *min_entropy_ix = (EntropyIx)k;
         }
       }
       *red_and_blue_always_zero = 1;
diff --git a/src/utils/rescaler.c b/src/utils/rescaler.c
index 00c9300b..d2278a52 100644
--- a/src/utils/rescaler.c
+++ b/src/utils/rescaler.c
@@ -48,11 +48,15 @@ void WebPRescalerInit(WebPRescaler* const wrk, int src_width, int src_height,
   wrk->y_sub = wrk->y_expand ? y_sub - 1 : y_sub;
   wrk->y_accum = wrk->y_expand ? wrk->y_sub : wrk->y_add;
   if (!wrk->y_expand) {
-    // this is WEBP_RESCALER_FRAC(dst_height, x_add * y_add) without the cast.
+    // This is WEBP_RESCALER_FRAC(dst_height, x_add * y_add) without the cast.
+    // Its value is <= WEBP_RESCALER_ONE, because dst_height <= wrk->y_add, and
+    // wrk->x_add >= 1;
     const uint64_t ratio =
         (uint64_t)dst_height * WEBP_RESCALER_ONE / (wrk->x_add * wrk->y_add);
     if (ratio != (uint32_t)ratio) {
-      // We can't represent the ratio with the current fixed-point precision.
+      // When ratio == WEBP_RESCALER_ONE, we can't represent the ratio with the
+      // current fixed-point precision. This happens when src_height ==
+      // wrk->y_add (which == src_height), and wrk->x_add == 1.
       // => We special-case fxy_scale = 0, in WebPRescalerExportRow().
       wrk->fxy_scale = 0;
     } else {
diff --git a/src/utils/utils.c b/src/utils/utils.c
index 2602ca3c..82dbf8d5 100644
--- a/src/utils/utils.c
+++ b/src/utils/utils.c
@@ -175,8 +175,12 @@ static int CheckSizeArgumentsOverflow(uint64_t nmemb, size_t size) {
   }
 #endif
 #if defined(MALLOC_LIMIT)
-  if (mem_limit > 0 && total_mem + total_size >= mem_limit) {
-    return 0;   // fake fail!
+  if (mem_limit > 0) {
+    const uint64_t new_total_mem = (uint64_t)total_mem + total_size;
+    if (new_total_mem != (size_t)new_total_mem ||
+        new_total_mem > mem_limit) {
+      return 0;   // fake fail!
+    }
   }
 #endif
 
diff --git a/src/utils/utils.h b/src/utils/utils.h
index e0a81126..3a5d4e6a 100644
--- a/src/utils/utils.h
+++ b/src/utils/utils.h
@@ -20,6 +20,7 @@
 #endif
 
 #include <assert.h>
+#include <limits.h>
 
 #include "../dsp/dsp.h"
 #include "../webp/types.h"
@@ -32,7 +33,14 @@ extern "C" {
 // Memory allocation
 
 // This is the maximum memory amount that libwebp will ever try to allocate.
-#define WEBP_MAX_ALLOCABLE_MEMORY (1ULL << 40)
+#ifndef WEBP_MAX_ALLOCABLE_MEMORY
+#if SIZE_MAX > (1ULL << 34)
+#define WEBP_MAX_ALLOCABLE_MEMORY (1ULL << 34)
+#else
+// For 32-bit targets keep this below INT_MAX to avoid valgrind warnings.
+#define WEBP_MAX_ALLOCABLE_MEMORY ((1ULL << 31) - (1 << 16))
+#endif
+#endif  // WEBP_MAX_ALLOCABLE_MEMORY
 
 // size-checking safe malloc/calloc: verify that the requested size is not too
 // large, or return NULL. You don't need to call these for constructs like
diff --git a/src/webp/decode.h b/src/webp/decode.h
index 7a3bed93..4c5e74ac 100644
--- a/src/webp/decode.h
+++ b/src/webp/decode.h
@@ -248,19 +248,19 @@ typedef enum VP8StatusCode {
 // picture is only partially decoded, pending additional input.
 // Code example:
 //
-//   WebPInitDecBuffer(&buffer);
-//   buffer.colorspace = mode;
+//   WebPInitDecBuffer(&output_buffer);
+//   output_buffer.colorspace = mode;
 //   ...
-//   WebPIDecoder* idec = WebPINewDecoder(&buffer);
-//   while (has_more_data) {
-//     // ... (get additional data)
+//   WebPIDecoder* idec = WebPINewDecoder(&output_buffer);
+//   while (additional_data_is_available) {
+//     // ... (get additional data in some new_data[] buffer)
 //     status = WebPIAppend(idec, new_data, new_data_size);
-//     if (status != VP8_STATUS_SUSPENDED ||
-//       break;
+//     if (status != VP8_STATUS_OK && status != VP8_STATUS_SUSPENDED) {
+//       break;    // an error occurred.
 //     }
 //
 //     // The above call decodes the current available buffer.
-//     // Part of the image can now be refreshed by calling to
+//     // Part of the image can now be refreshed by calling
 //     // WebPIDecGetRGB()/WebPIDecGetYUVA() etc.
 //   }
 //   WebPIDelete(idec);
diff --git a/src/webp/encode.h b/src/webp/encode.h
index 9291b719..b65e27e7 100644
--- a/src/webp/encode.h
+++ b/src/webp/encode.h
@@ -481,10 +481,10 @@ WEBP_EXTERN(int) WebPPictureARGBToYUVADithered(
 WEBP_EXTERN(int) WebPPictureSmartARGBToYUVA(WebPPicture* picture);
 
 // Converts picture->yuv to picture->argb and sets picture->use_argb to true.
-// The input format must be YUV_420 or YUV_420A.
-// Note that the use of this method is discouraged if one has access to the
-// raw ARGB samples, since using YUV420 is comparatively lossy. Also, the
-// conversion from YUV420 to ARGB incurs a small loss too.
+// The input format must be YUV_420 or YUV_420A. The conversion from YUV420 to
+// ARGB incurs a small loss too.
+// Note that the use of this colorspace is discouraged if one has access to the
+// raw ARGB samples, since using YUV420 is comparatively lossy.
 // Returns false in case of error.
 WEBP_EXTERN(int) WebPPictureYUVAToARGB(WebPPicture* picture);