diff --git a/src/enc/backward_references.c b/src/enc/backward_references.c
index 9c1d240f..8e489a62 100644
--- a/src/enc/backward_references.c
+++ b/src/enc/backward_references.c
@@ -21,8 +21,6 @@
 
 #define VALUES_IN_BYTE 256
 
-#define HASH_BITS 18
-#define HASH_SIZE (1 << HASH_BITS)
 #define HASH_MULTIPLIER (0xc6a4a7935bd1e995ULL)
 
 #define MAX_ENTROPY    (1e30f)
@@ -34,14 +32,6 @@
 #define MIN_LENGTH 2
 #define MAX_LENGTH 4096
 
-typedef struct {
-  // Stores the most recently added position with the given hash value.
-  int32_t hash_to_first_index_[HASH_SIZE];
-  // chain_[pos] stores the previous position with the same hash value
-  // for every pixel in the image.
-  int32_t* chain_;
-} HashChain;
-
 // -----------------------------------------------------------------------------
 
 static const uint8_t plane_to_code_lut[128] = {
@@ -79,29 +69,44 @@ static WEBP_INLINE int FindMatchLength(const uint32_t* const array1,
 // -----------------------------------------------------------------------------
 //  VP8LBackwardRefs
 
-void VP8LInitBackwardRefs(VP8LBackwardRefs* const refs) {
-  if (refs != NULL) {
-    refs->refs = NULL;
-    refs->size = 0;
-    refs->max_size = 0;
-  }
-}
-
-void VP8LClearBackwardRefs(VP8LBackwardRefs* const refs) {
-  if (refs != NULL) {
-    WebPSafeFree(refs->refs);
-    VP8LInitBackwardRefs(refs);
-  }
-}
-
-int VP8LBackwardRefsAlloc(VP8LBackwardRefs* const refs, int max_size) {
+static void ClearBackwardRefs(VP8LBackwardRefs* const refs) {
   assert(refs != NULL);
   refs->size = 0;
+}
+
+void VP8LBackwardRefsDelete(VP8LBackwardRefs* const refs) {
+  if (refs != NULL) {
+    WebPSafeFree(refs->refs);
+    WebPSafeFree(refs);
+  }
+}
+
+VP8LBackwardRefs* VP8LBackwardRefsNew(int max_size) {
+  VP8LBackwardRefs* const refs =
+      (VP8LBackwardRefs*)WebPSafeMalloc(1ULL, sizeof(*refs));
+  if (refs == NULL) {
+    return NULL;
+  }
+  ClearBackwardRefs(refs);
   refs->max_size = 0;
   refs->refs = (PixOrCopy*)WebPSafeMalloc((uint64_t)max_size,
-                                          sizeof(*refs->refs));
-  if (refs->refs == NULL) return 0;
+                                         sizeof(*refs->refs));
+  if (refs->refs == NULL) {
+    WebPSafeFree(refs);
+    return NULL;
+  }
   refs->max_size = max_size;
+  return refs;
+}
+
+int VP8LBackwardRefsCopy(const VP8LBackwardRefs* const src,
+                         VP8LBackwardRefs* const dst) {
+  assert(src != NULL && dst != NULL);
+  if (dst->max_size != src->max_size) {
+    return 0;
+  }
+  dst->size = src->size;
+  memcpy(dst->refs, src->refs, src->size * sizeof(*src->refs));
   return 1;
 }
 
@@ -114,9 +119,19 @@ static WEBP_INLINE uint64_t GetPixPairHash64(const uint32_t* const argb) {
   return key;
 }
 
-static HashChain* HashChainNew(int size) {
+static void HashChainInit(VP8LHashChain* const p) {
   int i;
-  HashChain* const p = (HashChain*)WebPSafeMalloc(1ULL, sizeof(*p));
+  assert(p != NULL);
+  for (i = 0; i < p->size_; ++i) {
+    p->chain_[i] = -1;
+  }
+  for (i = 0; i < HASH_SIZE; ++i) {
+    p->hash_to_first_index_[i] = -1;
+  }
+}
+
+VP8LHashChain* VP8LHashChainNew(int size) {
+  VP8LHashChain* const p = (VP8LHashChain*)WebPSafeMalloc(1ULL, sizeof(*p));
   if (p == NULL) {
     return NULL;
   }
@@ -125,16 +140,12 @@ static HashChain* HashChainNew(int size) {
     WebPSafeFree(p);
     return NULL;
   }
-  for (i = 0; i < size; ++i) {
-    p->chain_[i] = -1;
-  }
-  for (i = 0; i < HASH_SIZE; ++i) {
-    p->hash_to_first_index_[i] = -1;
-  }
+  p->size_ = size;
+  HashChainInit(p);
   return p;
 }
 
-static void HashChainDelete(HashChain* const p) {
+void VP8LHashChainDelete(VP8LHashChain* const p) {
   if (p != NULL) {
     WebPSafeFree(p->chain_);
     WebPSafeFree(p);
@@ -142,7 +153,7 @@ static void HashChainDelete(HashChain* const p) {
 }
 
 // Insertion of two pixels at a time.
-static void HashChainInsert(HashChain* const p,
+static void HashChainInsert(VP8LHashChain* const p,
                             const uint32_t* const argb, int pos) {
   const uint64_t hash_code = GetPixPairHash64(argb);
   p->chain_[pos] = p->hash_to_first_index_[hash_code];
@@ -167,7 +178,7 @@ static void GetParamsForHashChainFindCopy(int quality, int xsize,
   *iter_limit = (cache_bits > 0) ? iter_neg : iter_neg / 2;
 }
 
-static int HashChainFindCopy(const HashChain* const p,
+static int HashChainFindCopy(const VP8LHashChain* const p,
                              int base_position, int xsize_signed,
                              const uint32_t* const argb, int max_len,
                              int window_size, int iter_pos, int iter_limit,
@@ -260,7 +271,7 @@ static void BackwardReferencesRle(int xsize, int ysize,
   const int pix_count = xsize * ysize;
   int match_len = 0;
   int i;
-  refs->size = 0;
+  ClearBackwardRefs(refs);
   PushBackCopy(refs, match_len);    // i=0 case
   refs->refs[refs->size++] = PixOrCopyCreateLiteral(argb[0]);
   for (i = 1; i < pix_count; ++i) {
@@ -278,28 +289,27 @@ static void BackwardReferencesRle(int xsize, int ysize,
 static int BackwardReferencesHashChain(int xsize, int ysize,
                                        const uint32_t* const argb,
                                        int cache_bits, int quality,
+                                       VP8LHashChain* const hash_chain,
                                        VP8LBackwardRefs* const refs) {
   int i;
   int ok = 0;
   int cc_init = 0;
   const int use_color_cache = (cache_bits > 0);
   const int pix_count = xsize * ysize;
-  HashChain* const hash_chain = HashChainNew(pix_count);
   VP8LColorCache hashers;
   int window_size = WINDOW_SIZE;
   int iter_pos = 1;
   int iter_limit = -1;
 
-  if (hash_chain == NULL) return 0;
   if (use_color_cache) {
     cc_init = VP8LColorCacheInit(&hashers, cache_bits);
     if (!cc_init) goto Error;
   }
 
-
-  refs->size = 0;
+  ClearBackwardRefs(refs);
   GetParamsForHashChainFindCopy(quality, xsize, cache_bits,
                                 &window_size, &iter_pos, &iter_limit);
+  HashChainInit(hash_chain);
   for (i = 0; i < pix_count; ) {
     // Alternative#1: Code the pixels starting at 'i' using backward reference.
     int offset = 0;
@@ -375,7 +385,6 @@ static int BackwardReferencesHashChain(int xsize, int ysize,
   ok = 1;
 Error:
   if (cc_init) VP8LColorCacheClear(&hashers);
-  HashChainDelete(hash_chain);
   return ok;
 }
 
@@ -392,6 +401,7 @@ typedef struct {
 static int BackwardReferencesTraceBackwards(
     int xsize, int ysize, int recursive_cost_model,
     const uint32_t* const argb, int quality, int cache_bits,
+    VP8LHashChain* const hash_chain,
     VP8LBackwardRefs* const refs);
 
 static void ConvertPopulationCountTableToBitEstimates(
@@ -417,40 +427,45 @@ static void ConvertPopulationCountTableToBitEstimates(
 
 static int CostModelBuild(CostModel* const m, int xsize, int ysize,
                           int recursion_level, const uint32_t* const argb,
-                          int quality, int cache_bits) {
+                          int quality, int cache_bits,
+                          VP8LHashChain* const hash_chain,
+                          VP8LBackwardRefs* const refs) {
   int ok = 0;
-  VP8LHistogram histo;
-  VP8LBackwardRefs refs;
-
-  if (!VP8LBackwardRefsAlloc(&refs, xsize * ysize)) goto Error;
+  VP8LHistogram* histo = NULL;
 
+  ClearBackwardRefs(refs);
   if (recursion_level > 0) {
     if (!BackwardReferencesTraceBackwards(xsize, ysize, recursion_level - 1,
-                                          argb, quality, cache_bits, &refs)) {
+                                          argb, quality, cache_bits, hash_chain,
+                                          refs)) {
       goto Error;
     }
   } else {
     if (!BackwardReferencesHashChain(xsize, ysize, argb, cache_bits, quality,
-                                     &refs)) {
+                                     hash_chain, refs)) {
       goto Error;
     }
   }
-  VP8LHistogramCreate(&histo, &refs, cache_bits);
+  histo = VP8LAllocateHistogram(cache_bits);
+  if (histo == NULL) goto Error;
+
+  VP8LHistogramCreate(histo, refs, cache_bits);
+
   ConvertPopulationCountTableToBitEstimates(
-      VP8LHistogramNumCodes(histo.palette_code_bits_),
-      histo.literal_, m->literal_);
+      VP8LHistogramNumCodes(histo->palette_code_bits_),
+      histo->literal_, m->literal_);
   ConvertPopulationCountTableToBitEstimates(
-      VALUES_IN_BYTE, histo.red_, m->red_);
+      VALUES_IN_BYTE, histo->red_, m->red_);
   ConvertPopulationCountTableToBitEstimates(
-      VALUES_IN_BYTE, histo.blue_, m->blue_);
+      VALUES_IN_BYTE, histo->blue_, m->blue_);
   ConvertPopulationCountTableToBitEstimates(
-      VALUES_IN_BYTE, histo.alpha_, m->alpha_);
+      VALUES_IN_BYTE, histo->alpha_, m->alpha_);
   ConvertPopulationCountTableToBitEstimates(
-      NUM_DISTANCE_CODES, histo.distance_, m->distance_);
+      NUM_DISTANCE_CODES, histo->distance_, m->distance_);
   ok = 1;
 
  Error:
-  VP8LClearBackwardRefs(&refs);
+  VP8LFreeHistogram(histo);
   return ok;
 }
 
@@ -482,7 +497,8 @@ static WEBP_INLINE double GetDistanceCost(const CostModel* const m,
 
 static int BackwardReferencesHashChainDistanceOnly(
     int xsize, int ysize, int recursive_cost_model, const uint32_t* const argb,
-    int quality, int cache_bits, uint32_t* const dist_array) {
+    int quality, int cache_bits, VP8LHashChain* const hash_chain,
+    VP8LBackwardRefs* const refs, uint32_t* const dist_array) {
   int i;
   int ok = 0;
   int cc_init = 0;
@@ -491,7 +507,6 @@ static int BackwardReferencesHashChainDistanceOnly(
   float* const cost =
       (float*)WebPSafeMalloc((uint64_t)pix_count, sizeof(*cost));
   CostModel* cost_model = (CostModel*)WebPSafeMalloc(1ULL, sizeof(*cost_model));
-  HashChain* hash_chain = HashChainNew(pix_count);
   VP8LColorCache hashers;
   const double mul0 = (recursive_cost_model != 0) ? 1.0 : 0.68;
   const double mul1 = (recursive_cost_model != 0) ? 1.0 : 0.82;
@@ -500,7 +515,7 @@ static int BackwardReferencesHashChainDistanceOnly(
   int iter_pos = 1;
   int iter_limit = -1;
 
-  if (cost == NULL || cost_model == NULL || hash_chain == NULL) goto Error;
+  if (cost == NULL || cost_model == NULL) goto Error;
 
   if (use_color_cache) {
     cc_init = VP8LColorCacheInit(&hashers, cache_bits);
@@ -508,7 +523,7 @@ static int BackwardReferencesHashChainDistanceOnly(
   }
 
   if (!CostModelBuild(cost_model, xsize, ysize, recursive_cost_model, argb,
-                      quality, cache_bits)) {
+                      quality, cache_bits, hash_chain, refs)) {
     goto Error;
   }
 
@@ -519,6 +534,7 @@ static int BackwardReferencesHashChainDistanceOnly(
   dist_array[0] = 0;
   GetParamsForHashChainFindCopy(quality, xsize, cache_bits,
                                 &window_size, &iter_pos, &iter_limit);
+  HashChainInit(hash_chain);
   for (i = 0; i < pix_count; ++i) {
     double prev_cost = 0.0;
     int shortmax;
@@ -596,7 +612,6 @@ static int BackwardReferencesHashChainDistanceOnly(
   ok = 1;
 Error:
   if (cc_init) VP8LColorCacheClear(&hashers);
-  HashChainDelete(hash_chain);
   WebPSafeFree(cost_model);
   WebPSafeFree(cost);
   return ok;
@@ -625,6 +640,7 @@ static int BackwardReferencesHashChainFollowChosenPath(
     int xsize, int ysize, const uint32_t* const argb,
     int quality, int cache_bits,
     const uint32_t* const chosen_path, int chosen_path_size,
+    VP8LHashChain* const hash_chain,
     VP8LBackwardRefs* const refs) {
   const int pix_count = xsize * ysize;
   const int use_color_cache = (cache_bits > 0);
@@ -637,19 +653,17 @@ static int BackwardReferencesHashChainFollowChosenPath(
   int window_size = WINDOW_SIZE;
   int iter_pos = 1;
   int iter_limit = -1;
-  HashChain* hash_chain = HashChainNew(pix_count);
   VP8LColorCache hashers;
 
-  if (hash_chain == NULL) goto Error;
-
   if (use_color_cache) {
     cc_init = VP8LColorCacheInit(&hashers, cache_bits);
     if (!cc_init) goto Error;
   }
 
-  refs->size = 0;
+  ClearBackwardRefs(refs);
   GetParamsForHashChainFindCopy(quality, xsize, cache_bits,
                                 &window_size, &iter_pos, &iter_limit);
+  HashChainInit(hash_chain);
   for (ix = 0; ix < chosen_path_size; ++ix, ++size) {
     int offset = 0;
     int len = 0;
@@ -692,7 +706,6 @@ static int BackwardReferencesHashChainFollowChosenPath(
   ok = 1;
 Error:
   if (cc_init) VP8LColorCacheClear(&hashers);
-  HashChainDelete(hash_chain);
   return ok;
 }
 
@@ -701,6 +714,7 @@ static int BackwardReferencesTraceBackwards(int xsize, int ysize,
                                             int recursive_cost_model,
                                             const uint32_t* const argb,
                                             int quality, int cache_bits,
+                                            VP8LHashChain* const hash_chain,
                                             VP8LBackwardRefs* const refs) {
   int ok = 0;
   const int dist_array_size = xsize * ysize;
@@ -712,14 +726,14 @@ static int BackwardReferencesTraceBackwards(int xsize, int ysize,
   if (dist_array == NULL) goto Error;
 
   if (!BackwardReferencesHashChainDistanceOnly(
-      xsize, ysize, recursive_cost_model, argb, quality, cache_bits,
-      dist_array)) {
+      xsize, ysize, recursive_cost_model, argb, quality, cache_bits, hash_chain,
+      refs, dist_array)) {
     goto Error;
   }
   TraceBackwards(dist_array, dist_array_size, &chosen_path, &chosen_path_size);
   if (!BackwardReferencesHashChainFollowChosenPath(
       xsize, ysize, argb, quality, cache_bits, chosen_path, chosen_path_size,
-      refs)) {
+      hash_chain, refs)) {
     goto Error;
   }
   ok = 1;
@@ -740,82 +754,64 @@ static void BackwardReferences2DLocality(int xsize,
   }
 }
 
-int VP8LGetBackwardReferences(int width, int height,
-                              const uint32_t* const argb,
-                              int quality, int cache_bits, int use_2d_locality,
-                              VP8LBackwardRefs* const best) {
-  int ok = 0;
+VP8LBackwardRefs* VP8LGetBackwardReferences(
+    int width, int height, const uint32_t* const argb, int quality,
+    int cache_bits, int use_2d_locality, VP8LHashChain* const hash_chain,
+    VP8LBackwardRefs* const refs_array[2]) {
   int lz77_is_useful;
-  VP8LBackwardRefs refs_rle, refs_lz77;
   const int num_pix = width * height;
+  VP8LBackwardRefs* best = NULL;
+  VP8LBackwardRefs* const refs_lz77 = refs_array[0];
+  VP8LBackwardRefs* const refs_rle = refs_array[1];
 
-  VP8LBackwardRefsAlloc(&refs_rle, num_pix);
-  VP8LBackwardRefsAlloc(&refs_lz77, num_pix);
-  VP8LInitBackwardRefs(best);
-  if (refs_rle.refs == NULL || refs_lz77.refs == NULL) {
- Error1:
-    VP8LClearBackwardRefs(&refs_rle);
-    VP8LClearBackwardRefs(&refs_lz77);
-    goto End;
-  }
-
+  ClearBackwardRefs(refs_lz77);
   if (!BackwardReferencesHashChain(width, height, argb, cache_bits, quality,
-                                   &refs_lz77)) {
+                                   hash_chain, refs_lz77)) {
     goto End;
   }
-  // Backward Reference using RLE only.
-  BackwardReferencesRle(width, height, argb, &refs_rle);
+  ClearBackwardRefs(refs_rle);
+  BackwardReferencesRle(width, height, argb, refs_rle);
 
   {
     double bit_cost_lz77, bit_cost_rle;
-    VP8LHistogram* const histo =
-        (VP8LHistogram*)WebPSafeMalloc(1ULL, sizeof(*histo));
-    if (histo == NULL) goto Error1;
-    // Evaluate lz77 coding
-    VP8LHistogramCreate(histo, &refs_lz77, cache_bits);
+    VP8LHistogram* const histo = VP8LAllocateHistogram(cache_bits);
+    if (histo == NULL) goto End;
+    // Evaluate LZ77 coding.
+    VP8LHistogramCreate(histo, refs_lz77, cache_bits);
     bit_cost_lz77 = VP8LHistogramEstimateBits(histo);
-    // Evaluate RLE coding
-    VP8LHistogramCreate(histo, &refs_rle, cache_bits);
+    // Evaluate RLE coding.
+    VP8LHistogramCreate(histo, refs_rle, cache_bits);
     bit_cost_rle = VP8LHistogramEstimateBits(histo);
     // Decide if LZ77 is useful.
     lz77_is_useful = (bit_cost_lz77 < bit_cost_rle);
-    WebPSafeFree(histo);
+    VP8LFreeHistogram(histo);
   }
 
   // Choose appropriate backward reference.
   if (lz77_is_useful) {
     // TraceBackwards is costly. Don't execute it at lower quality.
     const int try_lz77_trace_backwards = (quality >= 25);
-    *best = refs_lz77;   // default guess: lz77 is better
-    VP8LClearBackwardRefs(&refs_rle);
+    best = refs_lz77;   // default guess: lz77 is better
     if (try_lz77_trace_backwards) {
       // Set recursion level for large images using a color cache.
       const int recursion_level =
           (num_pix < 320 * 200) && (cache_bits > 0) ? 1 : 0;
-      VP8LBackwardRefs refs_trace;
-      if (!VP8LBackwardRefsAlloc(&refs_trace, num_pix)) {
-        goto End;
-      }
+      VP8LBackwardRefs* const refs_trace = refs_array[1];
+      ClearBackwardRefs(refs_trace);
       if (BackwardReferencesTraceBackwards(width, height, recursion_level, argb,
-                                           quality, cache_bits, &refs_trace)) {
-        VP8LClearBackwardRefs(&refs_lz77);
-        *best = refs_trace;
+                                           quality, cache_bits, hash_chain,
+                                           refs_trace)) {
+        best = refs_trace;
       }
     }
   } else {
-    VP8LClearBackwardRefs(&refs_lz77);
-    *best = refs_rle;
+    best = refs_rle;
   }
 
   if (use_2d_locality) BackwardReferences2DLocality(width, best);
 
-  ok = 1;
-
  End:
-  if (!ok) {
-    VP8LClearBackwardRefs(best);
-  }
-  return ok;
+  return best;
 }
 
 // Returns entropy for the given cache bits.
@@ -828,17 +824,17 @@ static double ComputeCacheEntropy(const uint32_t* const argb,
   uint32_t k;
   const int use_color_cache = (cache_bits > 0);
   int cc_init = 0;
-  double entropy;
+  double entropy = MAX_ENTROPY;
   const double kSmallPenaltyForLargeCache = 4.0;
   VP8LColorCache hashers;
-  VP8LHistogram histo;
+  VP8LHistogram* histo = VP8LAllocateHistogram(cache_bits);
+  if (histo == NULL) goto Error;
 
   if (use_color_cache) {
     cc_init = VP8LColorCacheInit(&hashers, cache_bits);
-    if (!cc_init) return MAX_ENTROPY;
+    if (!cc_init) goto Error;
   }
 
-  VP8LHistogramInit(&histo, cache_bits);
   for (i = 0; i < refs->size; ++i) {
     const PixOrCopy* const v = &refs->refs[i];
     if (PixOrCopyIsLiteral(v)) {
@@ -847,12 +843,12 @@ static double ComputeCacheEntropy(const uint32_t* const argb,
         // push pixel as a cache index
         const int ix = VP8LColorCacheGetIndex(&hashers, argb[pixel_index]);
         const PixOrCopy token = PixOrCopyCreateCacheIdx(ix);
-        VP8LHistogramAddSinglePixOrCopy(&histo, &token);
+        VP8LHistogramAddSinglePixOrCopy(histo, &token);
       } else {
-        VP8LHistogramAddSinglePixOrCopy(&histo, v);
+        VP8LHistogramAddSinglePixOrCopy(histo, v);
       }
     } else {
-      VP8LHistogramAddSinglePixOrCopy(&histo, v);
+      VP8LHistogramAddSinglePixOrCopy(histo, v);
     }
     if (use_color_cache) {
       for (k = 0; k < PixOrCopyLength(v); ++k) {
@@ -864,15 +860,19 @@ static double ComputeCacheEntropy(const uint32_t* const argb,
   assert(pixel_index == xsize * ysize);
   (void)xsize;  // xsize is not used in non-debug compilations otherwise.
   (void)ysize;  // ysize is not used in non-debug compilations otherwise.
-  entropy = VP8LHistogramEstimateBits(&histo) +
+  entropy = VP8LHistogramEstimateBits(histo) +
       kSmallPenaltyForLargeCache * cache_bits;
+ Error:
   if (cc_init) VP8LColorCacheClear(&hashers);
+  VP8LFreeHistogram(histo);
   return entropy;
 }
 
 // Returns how many bits are to be used for a color cache.
 int VP8LCalculateEstimateForCacheSize(const uint32_t* const argb,
                                       int xsize, int ysize, int quality,
+                                      VP8LHashChain* const hash_chain,
+                                      VP8LBackwardRefs* const refs,
                                       int* const best_cache_bits) {
   int ok = 0;
   int eval_low = 1;
@@ -881,22 +881,22 @@ int VP8LCalculateEstimateForCacheSize(const uint32_t* const argb,
   double entropy_high = MAX_ENTROPY;
   int cache_bits_low = 0;
   int cache_bits_high = MAX_COLOR_CACHE_BITS;
-  VP8LBackwardRefs refs;
 
-  if (!VP8LBackwardRefsAlloc(&refs, xsize * ysize) ||
-      !BackwardReferencesHashChain(xsize, ysize, argb, 0, quality, &refs)) {
+  ClearBackwardRefs(refs);
+  if (!BackwardReferencesHashChain(xsize, ysize, argb, 0, quality, hash_chain,
+                                   refs)) {
     goto Error;
   }
   // Do a binary search to find the optimal entropy for cache_bits.
   while (cache_bits_high - cache_bits_low > 1) {
     if (eval_low) {
       entropy_low =
-          ComputeCacheEntropy(argb, xsize, ysize, &refs, cache_bits_low);
+          ComputeCacheEntropy(argb, xsize, ysize, refs, cache_bits_low);
       eval_low = 0;
     }
     if (eval_high) {
       entropy_high =
-          ComputeCacheEntropy(argb, xsize, ysize, &refs, cache_bits_high);
+          ComputeCacheEntropy(argb, xsize, ysize, refs, cache_bits_high);
       eval_high = 0;
     }
     if (entropy_high < entropy_low) {
@@ -911,6 +911,5 @@ int VP8LCalculateEstimateForCacheSize(const uint32_t* const argb,
   }
   ok = 1;
  Error:
-  VP8LClearBackwardRefs(&refs);
   return ok;
 }
diff --git a/src/enc/backward_references.h b/src/enc/backward_references.h
index 96d0a49a..6cf33166 100644
--- a/src/enc/backward_references.h
+++ b/src/enc/backward_references.h
@@ -115,34 +115,59 @@ static WEBP_INLINE uint32_t PixOrCopyDistance(const PixOrCopy* const p) {
 // -----------------------------------------------------------------------------
 // VP8LBackwardRefs
 
-typedef struct {
+#define HASH_BITS 18
+#define HASH_SIZE (1 << HASH_BITS)
+
+typedef struct VP8LHashChain VP8LHashChain;
+struct VP8LHashChain {
+  // Stores the most recently added position with the given hash value.
+  int32_t hash_to_first_index_[HASH_SIZE];
+  // chain_[pos] stores the previous position with the same hash value
+  // for every pixel in the image.
+  int32_t* chain_;
+  // This is the maximum size of the hash_chain that can be constructed.
+  // Typically this is the pixel count (width x height) for a given image.
+  int size_;
+};
+
+VP8LHashChain* VP8LHashChainNew(int size);
+void VP8LHashChainDelete(VP8LHashChain* const p);
+
+typedef struct VP8LBackwardRefs VP8LBackwardRefs;
+struct VP8LBackwardRefs {
   PixOrCopy* refs;
   int size;      // currently used
   int max_size;  // maximum capacity
-} VP8LBackwardRefs;
+};
 
-// Initialize the object. Must be called first. 'refs' can be NULL.
-void VP8LInitBackwardRefs(VP8LBackwardRefs* const refs);
+// Release backward references. 'refs' can be NULL.
+void VP8LBackwardRefsDelete(VP8LBackwardRefs* const refs);
 
-// Release memory and re-initialize the object. 'refs' can be NULL.
-void VP8LClearBackwardRefs(VP8LBackwardRefs* const refs);
+// Allocate 'max_size' references. Returns NULL in case of memory error.
+VP8LBackwardRefs* VP8LBackwardRefsNew(int max_size);
 
-// Allocate 'max_size' references. Returns false in case of memory error.
-int VP8LBackwardRefsAlloc(VP8LBackwardRefs* const refs, int max_size);
+// Copies the 'src' backward refs to the 'dst'. Returns 0 if there's mismatch
+// in the capacity (max_size) of 'src' and 'dst' refs.
+int VP8LBackwardRefsCopy(const VP8LBackwardRefs* const src,
+                         VP8LBackwardRefs* const dst);
 
 // -----------------------------------------------------------------------------
 // Main entry points
 
 // Evaluates best possible backward references for specified quality.
 // Further optimize for 2D locality if use_2d_locality flag is set.
-int VP8LGetBackwardReferences(int width, int height,
-                              const uint32_t* const argb,
-                              int quality, int cache_bits, int use_2d_locality,
-                              VP8LBackwardRefs* const best);
+// The return value is the pointer to the best of the two backward refs viz,
+// refs[0] or refs[1].
+VP8LBackwardRefs* VP8LGetBackwardReferences(
+    int width, int height, const uint32_t* const argb, int quality,
+    int cache_bits, int use_2d_locality, VP8LHashChain* const hash_chain,
+    VP8LBackwardRefs* const refs[2]);
 
 // Produce an estimate for a good color cache size for the image.
 int VP8LCalculateEstimateForCacheSize(const uint32_t* const argb,
                                       int xsize, int ysize, int quality,
+                                      VP8LHashChain* const hash_chain,
+                                      VP8LBackwardRefs* const ref,
                                       int* const best_cache_bits);
 
 #ifdef __cplusplus
diff --git a/src/enc/histogram.c b/src/enc/histogram.c
index 683c4a5f..c0e6f02a 100644
--- a/src/enc/histogram.c
+++ b/src/enc/histogram.c
@@ -29,12 +29,35 @@
 #define BIN_SIZE (NUM_PARTITIONS * NUM_PARTITIONS * NUM_PARTITIONS)
 
 static void HistogramClear(VP8LHistogram* const p) {
-  memset(p->literal_, 0, sizeof(p->literal_));
-  memset(p->red_, 0, sizeof(p->red_));
-  memset(p->blue_, 0, sizeof(p->blue_));
-  memset(p->alpha_, 0, sizeof(p->alpha_));
-  memset(p->distance_, 0, sizeof(p->distance_));
-  p->bit_cost_ = 0;
+  int* const literal = p->literal_;
+  const int cache_bits = p->palette_code_bits_;
+  const uint64_t histo_size = VP8LGetHistogramSize(cache_bits);
+  memset(p, 0, histo_size);
+  p->palette_code_bits_ = cache_bits;
+  p->literal_ = literal;
+}
+
+static void HistogramCopy(const VP8LHistogram* const src,
+                          VP8LHistogram* const dst) {
+  int* const dst_literal = dst->literal_;
+  const int dst_cache_bits = dst->palette_code_bits_;
+  const uint64_t histo_size = VP8LGetHistogramSize(dst_cache_bits);
+  assert(src->palette_code_bits_ == dst_cache_bits);
+  memcpy(dst, src, histo_size);
+  dst->literal_ = dst_literal;
+}
+
+int VP8LGetHistogramSize(int cache_bits) {
+  const uint64_t literal_size = VP8LHistogramNumCodes(cache_bits);
+  return sizeof(VP8LHistogram) + sizeof(int) * literal_size;
+}
+
+void VP8LFreeHistogram(VP8LHistogram* const histo) {
+  WebPSafeFree(histo);
+}
+
+void VP8LFreeHistogramSet(VP8LHistogramSet* const histo) {
+  WebPSafeFree(histo);
 }
 
 void VP8LHistogramStoreRefs(const VP8LBackwardRefs* const refs,
@@ -60,13 +83,24 @@ void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits) {
   HistogramClear(p);
 }
 
+VP8LHistogram* VP8LAllocateHistogram(int cache_bits) {
+  VP8LHistogram* histo = NULL;
+  const uint64_t total_size = VP8LGetHistogramSize(cache_bits);
+  uint8_t* const memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory));
+  if (memory == NULL) return NULL;
+  histo = (VP8LHistogram*)memory;
+  // literal_ won't necessary be aligned.
+  histo->literal_ = (int*)(memory + sizeof(VP8LHistogram));
+  VP8LHistogramInit(histo, cache_bits);
+  return histo;
+}
+
 VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits) {
   int i;
   VP8LHistogramSet* set;
-  VP8LHistogram* bulk;
   const uint64_t total_size = sizeof(*set)
                             + (uint64_t)size * sizeof(*set->histograms)
-                            + (uint64_t)size * sizeof(**set->histograms);
+                            + (uint64_t)size * VP8LGetHistogramSize(cache_bits);
   uint8_t* memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory));
   if (memory == NULL) return NULL;
 
@@ -74,12 +108,15 @@ VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits) {
   memory += sizeof(*set);
   set->histograms = (VP8LHistogram**)memory;
   memory += size * sizeof(*set->histograms);
-  bulk = (VP8LHistogram*)memory;
   set->max_size = size;
   set->size = size;
   for (i = 0; i < size; ++i) {
-    set->histograms[i] = bulk + i;
+    set->histograms[i] = (VP8LHistogram*)memory;
+    // literal_ won't necessary be aligned.
+    set->histograms[i]->literal_ = (int*)(memory + sizeof(VP8LHistogram));
     VP8LHistogramInit(set->histograms[i], cache_bits);
+    // There's no padding/alignment between successive histograms.
+    memory += VP8LGetHistogramSize(cache_bits);
   }
   return set;
 }
@@ -94,12 +131,13 @@ void VP8LHistogramAddSinglePixOrCopy(VP8LHistogram* const histo,
     ++histo->literal_[PixOrCopyLiteral(v, 1)];
     ++histo->blue_[PixOrCopyLiteral(v, 0)];
   } else if (PixOrCopyIsCacheIdx(v)) {
-    int literal_ix = 256 + NUM_LENGTH_CODES + PixOrCopyCacheIdx(v);
+    int literal_ix =
+        NUM_LITERAL_CODES + NUM_LENGTH_CODES + PixOrCopyCacheIdx(v);
     ++histo->literal_[literal_ix];
   } else {
     int code, extra_bits;
     VP8LPrefixEncodeBits(PixOrCopyLength(v), &code, &extra_bits);
-    ++histo->literal_[256 + code];
+    ++histo->literal_[NUM_LITERAL_CODES + code];
     VP8LPrefixEncodeBits(PixOrCopyDistance(v), &code, &extra_bits);
     ++histo->distance_[code];
   }
@@ -224,22 +262,22 @@ static double GetCombinedEntropy(const int* const X, const int* const Y,
 double VP8LHistogramEstimateBits(const VP8LHistogram* const p) {
   return
       PopulationCost(p->literal_, VP8LHistogramNumCodes(p->palette_code_bits_))
-      + PopulationCost(p->red_, 256)
-      + PopulationCost(p->blue_, 256)
-      + PopulationCost(p->alpha_, 256)
+      + PopulationCost(p->red_, NUM_LITERAL_CODES)
+      + PopulationCost(p->blue_, NUM_LITERAL_CODES)
+      + PopulationCost(p->alpha_, NUM_LITERAL_CODES)
       + PopulationCost(p->distance_, NUM_DISTANCE_CODES)
-      + VP8LExtraCost(p->literal_ + 256, NUM_LENGTH_CODES)
+      + VP8LExtraCost(p->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES)
       + VP8LExtraCost(p->distance_, NUM_DISTANCE_CODES);
 }
 
 double VP8LHistogramEstimateBitsBulk(const VP8LHistogram* const p) {
   return
       BitsEntropy(p->literal_, VP8LHistogramNumCodes(p->palette_code_bits_))
-      + BitsEntropy(p->red_, 256)
-      + BitsEntropy(p->blue_, 256)
-      + BitsEntropy(p->alpha_, 256)
+      + BitsEntropy(p->red_, NUM_LITERAL_CODES)
+      + BitsEntropy(p->blue_, NUM_LITERAL_CODES)
+      + BitsEntropy(p->alpha_, NUM_LITERAL_CODES)
       + BitsEntropy(p->distance_, NUM_DISTANCE_CODES)
-      + VP8LExtraCost(p->literal_ + 256, NUM_LENGTH_CODES)
+      + VP8LExtraCost(p->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES)
       + VP8LExtraCost(p->distance_, NUM_DISTANCE_CODES);
 }
 
@@ -250,13 +288,15 @@ double VP8LHistogramEstimateBitsBulk(const VP8LHistogram* const p) {
 static void HistogramAdd(const VP8LHistogram* const in,
                          VP8LHistogram* const out) {
   int i;
-  for (i = 0; i < PIX_OR_COPY_CODES_MAX; ++i) {
+  int literal_size = VP8LHistogramNumCodes(out->palette_code_bits_);
+  assert(in->palette_code_bits_ == out->palette_code_bits_);
+  for (i = 0; i < literal_size; ++i) {
     out->literal_[i] += in->literal_[i];
   }
   for (i = 0; i < NUM_DISTANCE_CODES; ++i) {
     out->distance_[i] += in->distance_[i];
   }
-  for (i = 0; i < 256; ++i) {
+  for (i = 0; i < NUM_LITERAL_CODES; ++i) {
     out->red_[i] += in->red_[i];
     out->blue_[i] += in->blue_[i];
     out->alpha_[i] += in->alpha_[i];
@@ -267,22 +307,22 @@ static int GetCombinedHistogramEntropy(const VP8LHistogram* const a,
                                        const VP8LHistogram* const b,
                                        double cost_threshold,
                                        double* cost) {
-  const int palette_code_bits =
-      (a->palette_code_bits_ > b->palette_code_bits_) ? a->palette_code_bits_ :
-                                                        b->palette_code_bits_;
+  const int palette_code_bits = a->palette_code_bits_;
+  assert(a->palette_code_bits_ == b->palette_code_bits_);
   *cost += GetCombinedEntropy(a->literal_, b->literal_,
                               VP8LHistogramNumCodes(palette_code_bits));
-  *cost += VP8LExtraCostCombined(a->literal_ + 256, b->literal_ + 256,
+  *cost += VP8LExtraCostCombined(a->literal_ + NUM_LITERAL_CODES,
+                                 b->literal_ + NUM_LITERAL_CODES,
                                  NUM_LENGTH_CODES);
   if (*cost > cost_threshold) return 0;
 
-  *cost += GetCombinedEntropy(a->red_, b->red_, 256);
+  *cost += GetCombinedEntropy(a->red_, b->red_, NUM_LITERAL_CODES);
   if (*cost > cost_threshold) return 0;
 
-  *cost += GetCombinedEntropy(a->blue_, b->blue_, 256);
+  *cost += GetCombinedEntropy(a->blue_, b->blue_, NUM_LITERAL_CODES);
   if (*cost > cost_threshold) return 0;
 
-  *cost += GetCombinedEntropy(a->alpha_, b->alpha_, 256);
+  *cost += GetCombinedEntropy(a->alpha_, b->alpha_, NUM_LITERAL_CODES);
   if (*cost > cost_threshold) return 0;
 
   *cost += GetCombinedEntropy(a->distance_, b->distance_, NUM_DISTANCE_CODES);
@@ -306,22 +346,24 @@ static double HistogramAddEval(const VP8LHistogram* const a,
   double cost = 0;
   const double sum_cost = a->bit_cost_ + b->bit_cost_;
   int i;
+  assert(a->palette_code_bits_ == b->palette_code_bits_);
   cost_threshold += sum_cost;
 
   if (GetCombinedHistogramEntropy(a, b, cost_threshold, &cost)) {
-    for (i = 0; i < PIX_OR_COPY_CODES_MAX; ++i) {
+    int literal_size;
+    out->palette_code_bits_ = a->palette_code_bits_;
+    literal_size = VP8LHistogramNumCodes(out->palette_code_bits_);
+    for (i = 0; i < literal_size; ++i) {
       out->literal_[i] = a->literal_[i] + b->literal_[i];
     }
     for (i = 0; i < NUM_DISTANCE_CODES; ++i) {
       out->distance_[i] = a->distance_[i] + b->distance_[i];
     }
-    for (i = 0; i < 256; ++i) {
+    for (i = 0; i < NUM_LITERAL_CODES; ++i) {
       out->red_[i] = a->red_[i] + b->red_[i];
       out->blue_[i] = a->blue_[i] + b->blue_[i];
       out->alpha_[i] = a->alpha_[i] + b->alpha_[i];
     }
-    out->palette_code_bits_ = (a->palette_code_bits_ > b->palette_code_bits_) ?
-                              a->palette_code_bits_ : b->palette_code_bits_;
     out->bit_cost_ = cost;
   }
 
@@ -374,15 +416,16 @@ static void UpdateDominantCostRange(
 }
 
 static void UpdateHistogramCost(VP8LHistogram* const h) {
-  const double alpha_cost = PopulationCost(h->alpha_, 256);
+  const double alpha_cost = PopulationCost(h->alpha_, NUM_LITERAL_CODES);
   const double distance_cost =
       PopulationCost(h->distance_, NUM_DISTANCE_CODES) +
       VP8LExtraCost(h->distance_, NUM_DISTANCE_CODES);
   const int num_codes = VP8LHistogramNumCodes(h->palette_code_bits_);
   h->literal_cost_ = PopulationCost(h->literal_, num_codes) +
-                     VP8LExtraCost(h->literal_ + 256, NUM_LENGTH_CODES);
-  h->red_cost_ = PopulationCost(h->red_, 256);
-  h->blue_cost_ = PopulationCost(h->blue_, 256);
+                     VP8LExtraCost(h->literal_ + NUM_LITERAL_CODES,
+                                   NUM_LENGTH_CODES);
+  h->red_cost_ = PopulationCost(h->red_, NUM_LITERAL_CODES);
+  h->blue_cost_ = PopulationCost(h->blue_, NUM_LITERAL_CODES);
   h->bit_cost_ = h->literal_cost_ + h->red_cost_ + h->blue_cost_ +
                  alpha_cost + distance_cost;
 }
@@ -439,7 +482,7 @@ static void HistogramAnalyze(
     VP8LHistogram* const histo = histograms[i];
     histo->bit_cost_ = VP8LHistogramEstimateBits(histo);
     // Copy histograms from init_histo[] to histo_image[].
-    *histo_image->histograms[i] = *histo;
+    HistogramCopy(histo, histo_image->histograms[i]);
   }
 }
 
@@ -460,7 +503,7 @@ static void HistogramAnalyzeBin(
     VP8LHistogram* const histo = histograms[i];
     UpdateHistogramCost(histo);
     // Copy histograms from init_histo[] to histo_image[].
-    *histo_image->histograms[i] = *histo;
+    HistogramCopy(histo, histo_image->histograms[i]);
     UpdateDominantCostRange(histo, &cost_range);
   }
 
@@ -503,7 +546,8 @@ static void HistogramCompactBins(VP8LHistogramSet* const histo_image) {
     if (start < end) {
       assert(histo_image->histograms[start] != NULL);
       assert(histo_image->histograms[end] != NULL);
-      *histo_image->histograms[start] = *histo_image->histograms[end];
+      HistogramCopy(histo_image->histograms[end],
+                    histo_image->histograms[start]);
       histo_image->histograms[end] = NULL;
       --end;
     }
@@ -533,7 +577,7 @@ static void HistogramCombineBin(VP8LHistogramSet* const histo_image,
                              histo_image->histograms[idx2],
                              cur_combo, bit_cost_thresh);
         if (curr_cost_diff < bit_cost_thresh) {
-          *histo_image->histograms[idx1] = *cur_combo;
+          HistogramCopy(cur_combo, histo_image->histograms[idx1]);
           histo_image->histograms[idx2]->bit_cost_ = 0.;
         }
       }
@@ -551,7 +595,7 @@ static uint32_t MyRand(uint32_t *seed) {
 }
 
 static void HistogramCombine(VP8LHistogramSet* const histo_image,
-                             VP8LHistogram* const histos, int quality) {
+                             VP8LHistogramSet* const histos, int quality) {
   int iter;
   uint32_t seed = 0;
   int tries_with_no_success = 0;
@@ -561,8 +605,8 @@ static void HistogramCombine(VP8LHistogramSet* const histo_image,
   const int num_pairs = histo_image_size / 2;
   const int num_tries_no_success = outer_iters / 2;
   const int min_cluster_size = 2;
-  VP8LHistogram* cur_combo = histos + 0;    // trial merged histogram
-  VP8LHistogram* best_combo = histos + 1;   // best merged histogram so far
+  VP8LHistogram* cur_combo = histos->histograms[0];   // trial histogram
+  VP8LHistogram* best_combo = histos->histograms[1];  // best histogram so far
 
   // Collapse similar histograms in 'histo_image'.
   for (iter = 0;
@@ -603,12 +647,12 @@ static void HistogramCombine(VP8LHistogramSet* const histo_image,
     }
 
     if (best_idx1 >= 0) {
-      *histo_image->histograms[best_idx1] = *best_combo;
+      HistogramCopy(best_combo, histo_image->histograms[best_idx1]);
       // swap best_idx2 slot with last one (which is now unused)
       --histo_image_size;
       if (best_idx2 != histo_image_size) {
-        histo_image->histograms[best_idx2] =
-            histo_image->histograms[histo_image_size];
+        HistogramCopy(histo_image->histograms[histo_image_size],
+                      histo_image->histograms[best_idx2]);
         histo_image->histograms[histo_image_size] = NULL;
       }
       tries_with_no_success = 0;
@@ -683,8 +727,7 @@ int VP8LGetHistoImageSymbols(int xsize, int ysize,
   // bin_map[n][num_histo + 1] ... bin_map[n][bin_depth - 1] = un-used indices.
   const int bin_depth = histo_image_raw_size + 1;
   int16_t* bin_map = NULL;
-  VP8LHistogram* const histos =
-      (VP8LHistogram*)WebPSafeMalloc(2ULL, sizeof(*histos));
+  VP8LHistogramSet* const histos = VP8LAllocateHistogramSet(2, cache_bits);
   VP8LHistogramSet* const init_histo =
       VP8LAllocateHistogramSet(histo_image_raw_size, cache_bits);
 
@@ -709,8 +752,8 @@ int VP8LGetHistoImageSymbols(int xsize, int ysize,
     const double combine_cost_factor =
         GetCombineCostFactor(histo_image_raw_size, quality);
     HistogramAnalyzeBin(init_histo, histo_image, bin_map);
-    HistogramCombineBin(histo_image, histos, bin_depth, combine_cost_factor,
-                        bin_map);
+    HistogramCombineBin(histo_image, histos->histograms[0],
+                        bin_depth, combine_cost_factor, bin_map);
   } else {
     HistogramAnalyze(init_histo, histo_image);
   }
@@ -725,7 +768,7 @@ int VP8LGetHistoImageSymbols(int xsize, int ysize,
 
  Error:
   WebPSafeFree(bin_map);
-  WebPSafeFree(init_histo);
-  WebPSafeFree(histos);
+  VP8LFreeHistogramSet(init_histo);
+  VP8LFreeHistogramSet(histos);
   return ok;
 }
diff --git a/src/enc/histogram.h b/src/enc/histogram.h
index 9d13ecd5..45b4b1e4 100644
--- a/src/enc/histogram.h
+++ b/src/enc/histogram.h
@@ -32,7 +32,7 @@ extern "C" {
 typedef struct {
   // literal_ contains green literal, palette-code and
   // copy-length-prefix histogram
-  int literal_[PIX_OR_COPY_CODES_MAX];
+  int* literal_;         // Pointer to the allocated buffer for literal.
   int red_[256];
   int blue_[256];
   int alpha_[256];
@@ -62,6 +62,9 @@ void VP8LHistogramCreate(VP8LHistogram* const p,
                          const VP8LBackwardRefs* const refs,
                          int palette_code_bits);
 
+// Return the size of the histogram for a given palette_code_bits.
+int VP8LGetHistogramSize(int palette_code_bits);
+
 // Set the palette_code_bits and reset the stats.
 void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits);
 
@@ -69,10 +72,21 @@ void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits);
 void VP8LHistogramStoreRefs(const VP8LBackwardRefs* const refs,
                             VP8LHistogram* const histo);
 
+// Free the memory allocated for the histogram.
+void VP8LFreeHistogram(VP8LHistogram* const histo);
+
+// Free the memory allocated for the histogram set.
+void VP8LFreeHistogramSet(VP8LHistogramSet* const histo);
+
 // Allocate an array of pointer to histograms, allocated and initialized
 // using 'cache_bits'. Return NULL in case of memory error.
 VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits);
 
+// Allocate and initialize histogram object with specified 'cache_bits'.
+// Returns NULL in case of memory error.
+// Special case of VP8LAllocateHistogramSet, with size equals 1.
+VP8LHistogram* VP8LAllocateHistogram(int cache_bits);
+
 // Accumulate a token 'v' into a histogram.
 void VP8LHistogramAddSinglePixOrCopy(VP8LHistogram* const histo,
                                      const PixOrCopy* const v);
@@ -86,7 +100,7 @@ double VP8LHistogramEstimateBits(const VP8LHistogram* const p);
 double VP8LHistogramEstimateBitsBulk(const VP8LHistogram* const p);
 
 static WEBP_INLINE int VP8LHistogramNumCodes(int palette_code_bits) {
-  return 256 + NUM_LENGTH_CODES +
+  return NUM_LITERAL_CODES + NUM_LENGTH_CODES +
       ((palette_code_bits > 0) ? (1 << palette_code_bits) : 0);
 }
 
diff --git a/src/enc/vp8l.c b/src/enc/vp8l.c
index c1974055..e0bb4332 100644
--- a/src/enc/vp8l.c
+++ b/src/enc/vp8l.c
@@ -106,14 +106,9 @@ static int AnalyzeEntropy(const uint32_t* argb,
   const uint32_t* last_line = NULL;
   uint32_t last_pix = argb[0];    // so we're sure that pix_diff == 0
 
-  VP8LHistogram* nonpredicted = NULL;
-  VP8LHistogram* predicted =
-      (VP8LHistogram*)WebPSafeMalloc(2ULL, sizeof(*predicted));
-  if (predicted == NULL) return 0;
-  nonpredicted = predicted + 1;
+  VP8LHistogramSet* const histo_set = VP8LAllocateHistogramSet(2, 0);
+  if (histo_set == NULL) return 0;
 
-  VP8LHistogramInit(predicted, 0);
-  VP8LHistogramInit(nonpredicted, 0);
   for (y = 0; y < height; ++y) {
     for (x = 0; x < width; ++x) {
       const uint32_t pix = argb[x];
@@ -126,21 +121,25 @@ static int AnalyzeEntropy(const uint32_t* argb,
       {
         const PixOrCopy pix_token = PixOrCopyCreateLiteral(pix);
         const PixOrCopy pix_diff_token = PixOrCopyCreateLiteral(pix_diff);
-        VP8LHistogramAddSinglePixOrCopy(nonpredicted, &pix_token);
-        VP8LHistogramAddSinglePixOrCopy(predicted, &pix_diff_token);
+        VP8LHistogramAddSinglePixOrCopy(histo_set->histograms[0], &pix_token);
+        VP8LHistogramAddSinglePixOrCopy(histo_set->histograms[1],
+                                        &pix_diff_token);
       }
     }
     last_line = argb;
     argb += argb_stride;
   }
-  *nonpredicted_bits = VP8LHistogramEstimateBitsBulk(nonpredicted);
-  *predicted_bits = VP8LHistogramEstimateBitsBulk(predicted);
-  WebPSafeFree(predicted);
+  *nonpredicted_bits = VP8LHistogramEstimateBitsBulk(histo_set->histograms[0]);
+  *predicted_bits = VP8LHistogramEstimateBitsBulk(histo_set->histograms[1]);
+  VP8LFreeHistogramSet(histo_set);
   return 1;
 }
 
-static int VP8LEncAnalyze(VP8LEncoder* const enc, WebPImageHint image_hint) {
+static int AnalyzeAndInit(VP8LEncoder* const enc, WebPImageHint image_hint) {
   const WebPPicture* const pic = enc->pic_;
+  const int width = pic->width;
+  const int height = pic->height;
+  const int pix_cnt = width * height;
   assert(pic != NULL && pic->argb != NULL);
 
   enc->use_palette_ =
@@ -158,7 +157,7 @@ static int VP8LEncAnalyze(VP8LEncoder* const enc, WebPImageHint image_hint) {
       enc->use_cross_color_ = 1;
     } else {
       double non_pred_entropy, pred_entropy;
-      if (!AnalyzeEntropy(pic->argb, pic->width, pic->height, pic->argb_stride,
+      if (!AnalyzeEntropy(pic->argb, width, height, pic->argb_stride,
                           &non_pred_entropy, &pred_entropy)) {
         return 0;
       }
@@ -168,6 +167,14 @@ static int VP8LEncAnalyze(VP8LEncoder* const enc, WebPImageHint image_hint) {
       }
     }
   }
+  enc->hash_chain_ = VP8LHashChainNew(pix_cnt);
+  if (enc->hash_chain_ == NULL) return 0;
+
+  enc->refs_[0] = VP8LBackwardRefsNew(pix_cnt);
+  enc->refs_[1] = VP8LBackwardRefsNew(pix_cnt);
+  if (enc->refs_[0] == NULL || enc->refs_[1] == NULL) {
+    return 0;
+  }
 
   return 1;
 }
@@ -176,10 +183,13 @@ static int GetHuffBitLengthsAndCodes(
     const VP8LHistogramSet* const histogram_image,
     HuffmanTreeCode* const huffman_codes) {
   int i, k;
-  int ok = 1;
+  int ok = 0;
   uint64_t total_length_size = 0;
   uint8_t* mem_buf = NULL;
   const int histogram_image_size = histogram_image->size;
+  int max_num_symbols = 0;
+  uint8_t* buf_rle = NULL;
+  HuffmanTree* huff_tree = NULL;
 
   // Iterate over all histograms and get the aggregate number of codes used.
   for (i = 0; i < histogram_image_size; ++i) {
@@ -200,10 +210,8 @@ static int GetHuffBitLengthsAndCodes(
     uint8_t* lengths;
     mem_buf = (uint8_t*)WebPSafeCalloc(total_length_size,
                                        sizeof(*lengths) + sizeof(*codes));
-    if (mem_buf == NULL) {
-      ok = 0;
-      goto End;
-    }
+    if (mem_buf == NULL) goto End;
+
     codes = (uint16_t*)mem_buf;
     lengths = (uint8_t*)&codes[total_length_size];
     for (i = 0; i < 5 * histogram_image_size; ++i) {
@@ -212,24 +220,33 @@ static int GetHuffBitLengthsAndCodes(
       huffman_codes[i].code_lengths = lengths;
       codes += bit_length;
       lengths += bit_length;
+      if (max_num_symbols < bit_length) {
+        max_num_symbols = bit_length;
+      }
     }
   }
 
+  buf_rle = (uint8_t*)WebPSafeMalloc(1ULL, max_num_symbols);
+  huff_tree = (HuffmanTree*)WebPSafeMalloc(3ULL * max_num_symbols,
+                                           sizeof(*huff_tree));
+  if (buf_rle == NULL || huff_tree == NULL) goto End;
+
   // Create Huffman trees.
-  for (i = 0; ok && (i < histogram_image_size); ++i) {
+  for (i = 0; i < histogram_image_size; ++i) {
     HuffmanTreeCode* const codes = &huffman_codes[5 * i];
     VP8LHistogram* const histo = histogram_image->histograms[i];
-    ok = ok && VP8LCreateHuffmanTree(histo->literal_, 15, codes + 0);
-    ok = ok && VP8LCreateHuffmanTree(histo->red_, 15, codes + 1);
-    ok = ok && VP8LCreateHuffmanTree(histo->blue_, 15, codes + 2);
-    ok = ok && VP8LCreateHuffmanTree(histo->alpha_, 15, codes + 3);
-    ok = ok && VP8LCreateHuffmanTree(histo->distance_, 15, codes + 4);
+    VP8LCreateHuffmanTree(histo->literal_, 15, buf_rle, huff_tree, codes + 0);
+    VP8LCreateHuffmanTree(histo->red_, 15, buf_rle, huff_tree, codes + 1);
+    VP8LCreateHuffmanTree(histo->blue_, 15, buf_rle, huff_tree, codes + 2);
+    VP8LCreateHuffmanTree(histo->alpha_, 15, buf_rle, huff_tree, codes + 3);
+    VP8LCreateHuffmanTree(histo->distance_, 15, buf_rle, huff_tree, codes + 4);
   }
-
+  ok = 1;
  End:
+  WebPSafeFree(huff_tree);
+  WebPSafeFree(buf_rle);
   if (!ok) {
     WebPSafeFree(mem_buf);
-    // If one VP8LCreateHuffmanTree() above fails, we need to clean up behind.
     memset(huffman_codes, 0, 5 * histogram_image_size * sizeof(*huffman_codes));
   }
   return ok;
@@ -296,18 +313,16 @@ static void StoreHuffmanTreeToBitMask(
   }
 }
 
-static int StoreFullHuffmanCode(VP8LBitWriter* const bw,
-                                const HuffmanTreeCode* const tree) {
-  int ok = 0;
+// 'huff_tree' and 'tokens' are pre-alloacted buffers.
+static void StoreFullHuffmanCode(VP8LBitWriter* const bw,
+                                 HuffmanTree* const huff_tree,
+                                 HuffmanTreeToken* const tokens,
+                                 const HuffmanTreeCode* const tree) {
   uint8_t code_length_bitdepth[CODE_LENGTH_CODES] = { 0 };
   uint16_t code_length_bitdepth_symbols[CODE_LENGTH_CODES] = { 0 };
   const int max_tokens = tree->num_symbols;
   int num_tokens;
   HuffmanTreeCode huffman_code;
-  HuffmanTreeToken* const tokens =
-      (HuffmanTreeToken*)WebPSafeMalloc((uint64_t)max_tokens, sizeof(*tokens));
-  if (tokens == NULL) return 0;
-
   huffman_code.num_symbols = CODE_LENGTH_CODES;
   huffman_code.code_lengths = code_length_bitdepth;
   huffman_code.codes = code_length_bitdepth_symbols;
@@ -316,14 +331,13 @@ static int StoreFullHuffmanCode(VP8LBitWriter* const bw,
   num_tokens = VP8LCreateCompressedHuffmanTree(tree, tokens, max_tokens);
   {
     int histogram[CODE_LENGTH_CODES] = { 0 };
+    uint8_t buf_rle[CODE_LENGTH_CODES] = { 0 };
     int i;
     for (i = 0; i < num_tokens; ++i) {
       ++histogram[tokens[i].code];
     }
 
-    if (!VP8LCreateHuffmanTree(histogram, 7, &huffman_code)) {
-      goto End;
-    }
+    VP8LCreateHuffmanTree(histogram, 7, buf_rle, huff_tree, &huffman_code);
   }
 
   StoreHuffmanTreeOfHuffmanTreeToBitMask(bw, code_length_bitdepth);
@@ -360,14 +374,13 @@ static int StoreFullHuffmanCode(VP8LBitWriter* const bw,
     }
     StoreHuffmanTreeToBitMask(bw, tokens, length, &huffman_code);
   }
-  ok = 1;
- End:
-  WebPSafeFree(tokens);
-  return ok;
 }
 
-static int StoreHuffmanCode(VP8LBitWriter* const bw,
-                            const HuffmanTreeCode* const huffman_code) {
+// 'huff_tree' and 'tokens' are pre-alloacted buffers.
+static void StoreHuffmanCode(VP8LBitWriter* const bw,
+                             HuffmanTree* const huff_tree,
+                             HuffmanTreeToken* const tokens,
+                             const HuffmanTreeCode* const huffman_code) {
   int i;
   int count = 0;
   int symbols[2] = { 0, 0 };
@@ -385,7 +398,6 @@ static int StoreHuffmanCode(VP8LBitWriter* const bw,
   if (count == 0) {   // emit minimal tree for empty cases
     // bits: small tree marker: 1, count-1: 0, large 8-bit code: 0, code: 0
     VP8LWriteBits(bw, 4, 0x01);
-    return 1;
   } else if (count <= 2 && symbols[0] < kMaxSymbol && symbols[1] < kMaxSymbol) {
     VP8LWriteBits(bw, 1, 1);  // Small tree marker to encode 1 or 2 symbols.
     VP8LWriteBits(bw, 1, count - 1);
@@ -399,9 +411,8 @@ static int StoreHuffmanCode(VP8LBitWriter* const bw,
     if (count == 2) {
       VP8LWriteBits(bw, 8, symbols[1]);
     }
-    return 1;
   } else {
-    return StoreFullHuffmanCode(bw, huffman_code);
+    StoreFullHuffmanCode(bw, huff_tree, tokens, huffman_code);
   }
 }
 
@@ -464,21 +475,29 @@ static void StoreImageToBitMask(
 // Special case of EncodeImageInternal() for cache-bits=0, histo_bits=31
 static int EncodeImageNoHuffman(VP8LBitWriter* const bw,
                                 const uint32_t* const argb,
+                                VP8LHashChain* const hash_chain,
+                                VP8LBackwardRefs* const refs_array[2],
                                 int width, int height, int quality) {
   int i;
   int ok = 0;
-  VP8LBackwardRefs refs;
+  int max_tokens = 0;
+  VP8LBackwardRefs* refs;
+  HuffmanTreeToken* tokens = NULL;
   HuffmanTreeCode huffman_codes[5] = { { 0, NULL, NULL } };
   const uint16_t histogram_symbols[1] = { 0 };    // only one tree, one symbol
   VP8LHistogramSet* const histogram_image = VP8LAllocateHistogramSet(1, 0);
-  if (histogram_image == NULL) return 0;
+  HuffmanTree* const huff_tree = (HuffmanTree*)WebPSafeMalloc(
+        3ULL * CODE_LENGTH_CODES, sizeof(*huff_tree));
+  if (histogram_image == NULL || huff_tree == NULL) goto Error;
 
   // Calculate backward references from ARGB image.
-  if (!VP8LGetBackwardReferences(width, height, argb, quality, 0, 1, &refs)) {
+  refs = VP8LGetBackwardReferences(width, height, argb, quality, 0, 1,
+                                   hash_chain, refs_array);
+  if (refs == NULL) {
     goto Error;
   }
   // Build histogram image and symbols from backward references.
-  VP8LHistogramStoreRefs(&refs, histogram_image->histograms[0]);
+  VP8LHistogramStoreRefs(refs, histogram_image->histograms[0]);
 
   // Create Huffman bit lengths and codes for each histogram image.
   assert(histogram_image->size == 1);
@@ -489,28 +508,41 @@ static int EncodeImageNoHuffman(VP8LBitWriter* const bw,
   // No color cache, no Huffman image.
   VP8LWriteBits(bw, 1, 0);
 
+  // Find maximum number of symbols for the huffman tree-set.
+  for (i = 0; i < 5; ++i) {
+    HuffmanTreeCode* const codes = &huffman_codes[i];
+    if (max_tokens < codes->num_symbols) {
+      max_tokens = codes->num_symbols;
+    }
+  }
+
+  tokens = (HuffmanTreeToken*)WebPSafeMalloc((uint64_t)max_tokens,
+                                             sizeof(*tokens));
+  if (tokens == NULL) goto Error;
+
   // Store Huffman codes.
   for (i = 0; i < 5; ++i) {
     HuffmanTreeCode* const codes = &huffman_codes[i];
-    if (!StoreHuffmanCode(bw, codes)) {
-      goto Error;
-    }
+    StoreHuffmanCode(bw, huff_tree, tokens, codes);
     ClearHuffmanTreeIfOnlyOneSymbol(codes);
   }
 
   // Store actual literals.
-  StoreImageToBitMask(bw, width, 0, &refs, histogram_symbols, huffman_codes);
+  StoreImageToBitMask(bw, width, 0, refs, histogram_symbols, huffman_codes);
   ok = 1;
 
  Error:
-  WebPSafeFree(histogram_image);
-  VP8LClearBackwardRefs(&refs);
+  WebPSafeFree(tokens);
+  WebPSafeFree(huff_tree);
+  VP8LFreeHistogramSet(histogram_image);
   WebPSafeFree(huffman_codes[0].codes);
   return ok;
 }
 
 static int EncodeImageInternal(VP8LBitWriter* const bw,
                                const uint32_t* const argb,
+                               VP8LHashChain* const hash_chain,
+                               VP8LBackwardRefs* const refs_array[2],
                                int width, int height, int quality,
                                int cache_bits, int histogram_bits) {
   int ok = 0;
@@ -520,11 +552,14 @@ static int EncodeImageInternal(VP8LBitWriter* const bw,
       VP8LSubSampleSize(width, histogram_bits) *
       VP8LSubSampleSize(height, histogram_bits);
   VP8LHistogramSet* histogram_image =
-      VP8LAllocateHistogramSet(histogram_image_xysize, 0);
+      VP8LAllocateHistogramSet(histogram_image_xysize, cache_bits);
   int histogram_image_size = 0;
   size_t bit_array_size = 0;
+  HuffmanTree* huff_tree = NULL;
+  HuffmanTreeToken* tokens = NULL;
   HuffmanTreeCode* huffman_codes = NULL;
-  VP8LBackwardRefs refs;
+  VP8LBackwardRefs* refs = NULL;
+  VP8LBackwardRefs* best_refs;
   uint16_t* const histogram_symbols =
       (uint16_t*)WebPSafeMalloc((uint64_t)histogram_image_xysize,
                                 sizeof(*histogram_symbols));
@@ -532,18 +567,27 @@ static int EncodeImageInternal(VP8LBitWriter* const bw,
   assert(histogram_bits <= MAX_HUFFMAN_BITS);
 
   if (histogram_image == NULL || histogram_symbols == NULL) {
-    WebPSafeFree(histogram_image);
+    VP8LFreeHistogramSet(histogram_image);
     WebPSafeFree(histogram_symbols);
     return 0;
   }
 
+  refs = VP8LBackwardRefsNew(refs_array[0]->max_size);
+  if (refs == NULL) {
+    goto Error;
+  }
+
+  // 'best_refs' is the reference to the best backward refs and points to one
+  // of refs_array[0] or refs_array[1].
   // Calculate backward references from ARGB image.
-  if (!VP8LGetBackwardReferences(width, height, argb, quality, cache_bits,
-                                 use_2d_locality, &refs)) {
+  best_refs = VP8LGetBackwardReferences(width, height, argb, quality,
+                                        cache_bits, use_2d_locality,
+                                        hash_chain, refs_array);
+  if (best_refs == NULL || !VP8LBackwardRefsCopy(best_refs, refs)) {
     goto Error;
   }
   // Build histogram image and symbols from backward references.
-  if (!VP8LGetHistoImageSymbols(width, height, &refs,
+  if (!VP8LGetHistoImageSymbols(width, height, refs,
                                 quality, histogram_bits, cache_bits,
                                 histogram_image,
                                 histogram_symbols)) {
@@ -559,7 +603,7 @@ static int EncodeImageInternal(VP8LBitWriter* const bw,
     goto Error;
   }
   // Free combined histograms.
-  WebPSafeFree(histogram_image);
+  VP8LFreeHistogramSet(histogram_image);
   histogram_image = NULL;
 
   // Color Cache parameters.
@@ -589,7 +633,7 @@ static int EncodeImageInternal(VP8LBitWriter* const bw,
       histogram_image_size = max_index;
 
       VP8LWriteBits(bw, 3, histogram_bits - 2);
-      ok = EncodeImageNoHuffman(bw, histogram_argb,
+      ok = EncodeImageNoHuffman(bw, histogram_argb, hash_chain, refs_array,
                                 VP8LSubSampleSize(width, histogram_bits),
                                 VP8LSubSampleSize(height, histogram_bits),
                                 quality);
@@ -601,22 +645,37 @@ static int EncodeImageInternal(VP8LBitWriter* const bw,
   // Store Huffman codes.
   {
     int i;
+    int max_tokens = 0;
+    huff_tree = (HuffmanTree*)WebPSafeMalloc(3ULL * CODE_LENGTH_CODES,
+                                             sizeof(*huff_tree));
+    if (huff_tree == NULL) goto Error;
+    // Find maximum number of symbols for the huffman tree-set.
     for (i = 0; i < 5 * histogram_image_size; ++i) {
       HuffmanTreeCode* const codes = &huffman_codes[i];
-      if (!StoreHuffmanCode(bw, codes)) goto Error;
+      if (max_tokens < codes->num_symbols) {
+        max_tokens = codes->num_symbols;
+      }
+    }
+    tokens = (HuffmanTreeToken*)WebPSafeMalloc((uint64_t)max_tokens,
+                                               sizeof(*tokens));
+    if (tokens == NULL) goto Error;
+    for (i = 0; i < 5 * histogram_image_size; ++i) {
+      HuffmanTreeCode* const codes = &huffman_codes[i];
+      StoreHuffmanCode(bw, huff_tree, tokens, codes);
       ClearHuffmanTreeIfOnlyOneSymbol(codes);
     }
   }
 
   // Store actual literals.
-  StoreImageToBitMask(bw, width, histogram_bits, &refs,
+  StoreImageToBitMask(bw, width, histogram_bits, refs,
                       histogram_symbols, huffman_codes);
   ok = 1;
 
  Error:
-  WebPSafeFree(histogram_image);
-
-  VP8LClearBackwardRefs(&refs);
+  WebPSafeFree(tokens);
+  WebPSafeFree(huff_tree);
+  VP8LFreeHistogramSet(histogram_image);
+  VP8LBackwardRefsDelete(refs);
   if (huffman_codes != NULL) {
     WebPSafeFree(huffman_codes->codes);
     WebPSafeFree(huffman_codes);
@@ -637,11 +696,9 @@ static int EvalAndApplySubtractGreen(VP8LEncoder* const enc,
     int i;
     const uint32_t* const argb = enc->argb_;
     double bit_cost_before, bit_cost_after;
-    VP8LHistogram* const histo =
-        (VP8LHistogram*)WebPSafeMalloc(1ULL, sizeof(*histo));
+    // Allocate histogram with cache_bits = 1.
+    VP8LHistogram* const histo = VP8LAllocateHistogram(1);
     if (histo == NULL) return 0;
-
-    VP8LHistogramInit(histo, 1);
     for (i = 0; i < width * height; ++i) {
       const uint32_t c = argb[i];
       ++histo->red_[(c >> 16) & 0xff];
@@ -657,7 +714,7 @@ static int EvalAndApplySubtractGreen(VP8LEncoder* const enc,
       ++histo->blue_[((c >> 0) - green) & 0xff];
     }
     bit_cost_after = VP8LHistogramEstimateBits(histo);
-    WebPSafeFree(histo);
+    VP8LFreeHistogram(histo);
 
     // Check if subtracting green yields low entropy.
     enc->use_subtract_green_ = (bit_cost_after < bit_cost_before);
@@ -683,8 +740,9 @@ static int ApplyPredictFilter(const VP8LEncoder* const enc,
   VP8LWriteBits(bw, 2, PREDICTOR_TRANSFORM);
   assert(pred_bits >= 2);
   VP8LWriteBits(bw, 3, pred_bits - 2);
-  if (!EncodeImageNoHuffman(bw, enc->transform_data_,
-                            transform_width, transform_height, quality)) {
+  if (!EncodeImageNoHuffman(bw, enc->transform_data_, enc->hash_chain_,
+                            enc->refs_, transform_width, transform_height,
+                            quality)) {
     return 0;
   }
   return 1;
@@ -703,8 +761,9 @@ static int ApplyCrossColorFilter(const VP8LEncoder* const enc,
   VP8LWriteBits(bw, 2, CROSS_COLOR_TRANSFORM);
   assert(ccolor_transform_bits >= 2);
   VP8LWriteBits(bw, 3, ccolor_transform_bits - 2);
-  if (!EncodeImageNoHuffman(bw, enc->transform_data_,
-                            transform_width, transform_height, quality)) {
+  if (!EncodeImageNoHuffman(bw, enc->transform_data_, enc->hash_chain_,
+                            enc->refs_, transform_width, transform_height,
+                            quality)) {
     return 0;
   }
   return 1;
@@ -902,7 +961,8 @@ static WebPEncodingError EncodePalette(VP8LBitWriter* const bw,
   for (i = palette_size - 1; i >= 1; --i) {
     palette[i] = VP8LSubPixels(palette[i], palette[i - 1]);
   }
-  if (!EncodeImageNoHuffman(bw, palette, palette_size, 1, quality)) {
+  if (!EncodeImageNoHuffman(bw, palette, enc->hash_chain_, enc->refs_,
+                            palette_size, 1, quality)) {
     err = VP8_ENC_ERROR_INVALID_CONFIGURATION;
     goto Error;
   }
@@ -915,7 +975,7 @@ static WebPEncodingError EncodePalette(VP8LBitWriter* const bw,
 // -----------------------------------------------------------------------------
 
 static int GetHistoBits(int method, int use_palette, int width, int height) {
-  const uint64_t hist_size = sizeof(VP8LHistogram);
+  const uint64_t hist_size = VP8LGetHistogramSize(MAX_COLOR_CACHE_BITS);
   // Make tile size a function of encoding method (Range: 0 to 6).
   int histo_bits = (use_palette ? 9 : 7) - method;
   while (1) {
@@ -969,6 +1029,9 @@ static VP8LEncoder* VP8LEncoderNew(const WebPConfig* const config,
 
 static void VP8LEncoderDelete(VP8LEncoder* enc) {
   if (enc != NULL) {
+    VP8LHashChainDelete(enc->hash_chain_);
+    VP8LBackwardRefsDelete(enc->refs_[0]);
+    VP8LBackwardRefsDelete(enc->refs_[1]);
     WebPSafeFree(enc->argb_);
     WebPSafeFree(enc);
   }
@@ -995,7 +1058,7 @@ WebPEncodingError VP8LEncodeStream(const WebPConfig* const config,
   // ---------------------------------------------------------------------------
   // Analyze image (entropy, num_palettes etc)
 
-  if (!VP8LEncAnalyze(enc, config->image_hint)) {
+  if (!AnalyzeAndInit(enc, config->image_hint)) {
     err = VP8_ENC_ERROR_OUT_OF_MEMORY;
     goto Error;
   }
@@ -1051,8 +1114,8 @@ WebPEncodingError VP8LEncodeStream(const WebPConfig* const config,
 
   if (enc->cache_bits_ > 0) {
     if (!VP8LCalculateEstimateForCacheSize(enc->argb_, enc->current_width_,
-                                           height, quality,
-                                           &enc->cache_bits_)) {
+                                           height, quality, enc->hash_chain_,
+                                           enc->refs_[0], &enc->cache_bits_)) {
       err = VP8_ENC_ERROR_INVALID_CONFIGURATION;
       goto Error;
     }
@@ -1061,8 +1124,9 @@ WebPEncodingError VP8LEncodeStream(const WebPConfig* const config,
   // ---------------------------------------------------------------------------
   // Encode and write the transformed image.
 
-  if (!EncodeImageInternal(bw, enc->argb_, enc->current_width_, height,
-                           quality, enc->cache_bits_, enc->histo_bits_)) {
+  if (!EncodeImageInternal(bw, enc->argb_, enc->hash_chain_, enc->refs_,
+                           enc->current_width_, height, quality,
+                           enc->cache_bits_, enc->histo_bits_)) {
     err = VP8_ENC_ERROR_OUT_OF_MEMORY;
     goto Error;
   }
@@ -1092,6 +1156,7 @@ int VP8LEncodeImage(const WebPConfig* const config,
   int has_alpha;
   size_t coded_size;
   int percent = 0;
+  int initial_size;
   WebPEncodingError err = VP8_ENC_OK;
   VP8LBitWriter bw;
 
@@ -1105,8 +1170,11 @@ int VP8LEncodeImage(const WebPConfig* const config,
 
   width = picture->width;
   height = picture->height;
-  // Initialize BitWriter with size corresponding to 8bpp.
-  if (!VP8LBitWriterInit(&bw, width * height)) {
+  // Initialize BitWriter with size corresponding to 16 bpp to photo images and
+  // 8 bpp for graphical images.
+  initial_size = (config->image_hint == WEBP_HINT_GRAPH) ?
+                 width * height : width * height * 2;
+  if (!VP8LBitWriterInit(&bw, initial_size)) {
     err = VP8_ENC_ERROR_OUT_OF_MEMORY;
     goto Error;
   }
diff --git a/src/enc/vp8li.h b/src/enc/vp8li.h
index 96d6faed..7db49f7a 100644
--- a/src/enc/vp8li.h
+++ b/src/enc/vp8li.h
@@ -23,6 +23,8 @@
 extern "C" {
 #endif
 
+struct VP8LHashChain;  // Defined in backward_references.h
+
 typedef struct {
   const WebPConfig* config_;    // user configuration and parameters
   const WebPPicture* pic_;      // input picture.
@@ -33,6 +35,11 @@ typedef struct {
   uint32_t* transform_data_;    // Scratch memory for transform data.
   int       current_width_;     // Corresponds to packed image width.
 
+  struct VP8LHashChain* hash_chain_;  // HashChain data for constructing
+                                      // backward references.
+  struct VP8LBackwardRefs* refs_[2];  // Backward Refs array corresponding to
+                                      // LZ77 & RLE coding.
+
   // Encoding parameters derived from quality parameter.
   int histo_bits_;
   int transform_bits_;
diff --git a/src/utils/huffman_encode.c b/src/utils/huffman_encode.c
index cf4e0504..6cb433df 100644
--- a/src/utils/huffman_encode.c
+++ b/src/utils/huffman_encode.c
@@ -28,13 +28,13 @@ static int ValuesShouldBeCollapsedToStrideAverage(int a, int b) {
 
 // Change the population counts in a way that the consequent
 // Huffman tree compression, especially its RLE-part, give smaller output.
-static int OptimizeHuffmanForRle(int length, int* const counts) {
-  uint8_t* good_for_rle;
+static void OptimizeHuffmanForRle(int length, uint8_t* const good_for_rle,
+                                  int* const counts) {
   // 1) Let's make the Huffman code more compatible with rle encoding.
   int i;
   for (; length >= 0; --length) {
     if (length == 0) {
-      return 1;  // All zeros.
+      return;  // All zeros.
     }
     if (counts[length - 1] != 0) {
       // Now counts[0..length - 1] does not have trailing zeros.
@@ -43,10 +43,6 @@ static int OptimizeHuffmanForRle(int length, int* const counts) {
   }
   // 2) Let's mark all population counts that already can be encoded
   // with an rle code.
-  good_for_rle = (uint8_t*)WebPSafeCalloc(1ULL, length);
-  if (good_for_rle == NULL) {
-    return 0;
-  }
   {
     // Let's not spoil any of the existing good rle codes.
     // Mark any seq of 0's that is longer as 5 as a good_for_rle.
@@ -119,17 +115,8 @@ static int OptimizeHuffmanForRle(int length, int* const counts) {
       }
     }
   }
-  WebPSafeFree(good_for_rle);
-  return 1;
 }
 
-typedef struct {
-  int total_count_;
-  int value_;
-  int pool_index_left_;
-  int pool_index_right_;
-} HuffmanTree;
-
 // A comparer function for two Huffman trees: sorts first by 'total count'
 // (more comes first), and then by 'value' (more comes first).
 static int CompareHuffmanTrees(const void* ptr1, const void* ptr2) {
@@ -175,12 +162,11 @@ static void SetBitDepths(const HuffmanTree* const tree,
 // we are not planning to use this with extremely long blocks.
 //
 // See http://en.wikipedia.org/wiki/Huffman_coding
-static int GenerateOptimalTree(const int* const histogram, int histogram_size,
-                               int tree_depth_limit,
-                               uint8_t* const bit_depths) {
+static void GenerateOptimalTree(const int* const histogram, int histogram_size,
+                                HuffmanTree* tree, int tree_depth_limit,
+                                uint8_t* const bit_depths) {
   int count_min;
   HuffmanTree* tree_pool;
-  HuffmanTree* tree;
   int tree_size_orig = 0;
   int i;
 
@@ -191,15 +177,9 @@ static int GenerateOptimalTree(const int* const histogram, int histogram_size,
   }
 
   if (tree_size_orig == 0) {   // pretty optimal already!
-    return 1;
+    return;
   }
 
-  // 3 * tree_size is enough to cover all the nodes representing a
-  // population and all the inserted nodes combining two existing nodes.
-  // The tree pool needs 2 * (tree_size_orig - 1) entities, and the
-  // tree needs exactly tree_size_orig entities.
-  tree = (HuffmanTree*)WebPSafeMalloc(3ULL * tree_size_orig, sizeof(*tree));
-  if (tree == NULL) return 0;
   tree_pool = tree + tree_size_orig;
 
   // For block sizes with less than 64k symbols we never need to do a
@@ -272,8 +252,6 @@ static int GenerateOptimalTree(const int* const histogram, int histogram_size,
       }
     }
   }
-  WebPSafeFree(tree);
-  return 1;
 }
 
 // -----------------------------------------------------------------------------
@@ -424,17 +402,15 @@ static void ConvertBitDepthsToSymbols(HuffmanTreeCode* const tree) {
 // -----------------------------------------------------------------------------
 // Main entry point
 
-int VP8LCreateHuffmanTree(int* const histogram, int tree_depth_limit,
-                          HuffmanTreeCode* const tree) {
-  const int num_symbols = tree->num_symbols;
-  if (!OptimizeHuffmanForRle(num_symbols, histogram)) {
-    return 0;
-  }
-  if (!GenerateOptimalTree(histogram, num_symbols,
-                           tree_depth_limit, tree->code_lengths)) {
-    return 0;
-  }
+void VP8LCreateHuffmanTree(int* const histogram, int tree_depth_limit,
+                           uint8_t* const buf_rle,
+                           HuffmanTree* const huff_tree,
+                           HuffmanTreeCode* const huff_code) {
+  const int num_symbols = huff_code->num_symbols;
+  memset(buf_rle, 0, num_symbols * sizeof(*buf_rle));
+  OptimizeHuffmanForRle(num_symbols, buf_rle, histogram);
+  GenerateOptimalTree(histogram, num_symbols, huff_tree, tree_depth_limit,
+                      huff_code->code_lengths);
   // Create the actual bit codes for the bit lengths.
-  ConvertBitDepthsToSymbols(tree);
-  return 1;
+  ConvertBitDepthsToSymbols(huff_code);
 }
diff --git a/src/utils/huffman_encode.h b/src/utils/huffman_encode.h
index ee51c68c..44e09128 100644
--- a/src/utils/huffman_encode.h
+++ b/src/utils/huffman_encode.h
@@ -33,14 +33,26 @@ typedef struct {
   uint16_t* codes;         // Symbol Codes.
 } HuffmanTreeCode;
 
+// Struct to represent the Huffman tree.
+// TODO(vikasa): Add comment for the fields of the Struct.
+typedef struct {
+  int total_count_;
+  int value_;
+  int pool_index_left_;    // Index for the left sub-tree.
+  int pool_index_right_;   // Index for the right sub-tree.
+} HuffmanTree;
+
 // Turn the Huffman tree into a token sequence.
 // Returns the number of tokens used.
 int VP8LCreateCompressedHuffmanTree(const HuffmanTreeCode* const tree,
                                     HuffmanTreeToken* tokens, int max_tokens);
 
 // Create an optimized tree, and tokenize it.
-int VP8LCreateHuffmanTree(int* const histogram, int tree_depth_limit,
-                          HuffmanTreeCode* const tree);
+// 'buf_rle' and 'huff_tree' are pre-allocated and the 'tree' is the constructed
+// huffman code tree.
+void VP8LCreateHuffmanTree(int* const histogram, int tree_depth_limit,
+                           uint8_t* const buf_rle, HuffmanTree* const huff_tree,
+                           HuffmanTreeCode* const tree);
 
 #ifdef __cplusplus
 }