Merge pull request #112 from szabadka/master

Speedups and fixes to the decoder.

dec/bit_reader.h

@@ -191,19 +191,19 @@ static BROTLI_INLINE int BrotliReadMoreInput(BrotliBitReader* const br) {
 /* Guarantees that there are at least 24 bits in the buffer. */
 static BROTLI_INLINE void BrotliFillBitWindow(BrotliBitReader* const br) {
 #if (BROTLI_USE_64_BITS)
-  if (br->bit_pos_ >= 40) {
+  if (br->bit_pos_ >= 32) {
     /*
-     * Advances the Read buffer by 5 bytes to make room for reading next
+     * Advances the Read buffer by 4 bytes to make room for reading next
      * 24 bits.
      * The expression below needs a little-endian arch to work correctly.
      * This gives a large speedup for decoding speed.
      */
-    br->val_ >>= 40;
-    br->val_ |= *(const uint64_t*)(
-        br->buf_ + (br->pos_ & BROTLI_IBUF_MASK)) << 24;
-    br->pos_ += 5;
-    br->bit_pos_ -= 40;
-    br->bit_end_pos_ -= 40;
+    br->val_ >>= 32;
+    br->val_ |= ((uint64_t)(*(const uint32_t*)(
+        br->buf_ + (br->pos_ & BROTLI_IBUF_MASK)))) << 32;
+    br->pos_ += 4;
+    br->bit_pos_ -= 32;
+    br->bit_end_pos_ -= 32;
   }
 #else
   ShiftBytes32(br);

dec/decode.c

@@ -222,13 +222,16 @@ static BrotliResult ReadHuffmanCodeLengths(
         p += (br->val_ >> br->bit_pos_) & 31;
         br->bit_pos_ += p->bits;
         code_len = (uint8_t)p->value;
+        /* We predict that branch will be taken and write value now.
+           Even if branch is mispredicted - it works as prefetch. */
+        code_lengths[s->symbol] = code_len;
         if (code_len < kCodeLengthRepeatCode) {
           s->repeat = 0;
-          code_lengths[s->symbol++] = code_len;
           if (code_len != 0) {
             s->prev_code_len = code_len;
             s->space -= 32768 >> code_len;
           }
+          s->symbol++;
         } else {
           const int extra_bits = code_len - 14;
           int old_repeat;

dec/huffman.c

@@ -19,6 +19,7 @@
 #include <stdio.h>
 #include <string.h>
 #include "./huffman.h"
+#include "./port.h"
 #include "./safe_malloc.h"
 
 #if defined(__cplusplus) || defined(c_plusplus)
@@ -27,6 +28,9 @@ extern "C" {
 
 #define MAX_LENGTH 15
 
+/* For current format this constant equals to kNumInsertAndCopyCodes */
+#define MAX_CODE_LENGTHS_SIZE 704
+
 /* Returns reverse(reverse(key, len) + 1, len), where reverse(key, len) is the
    bit-wise reversal of the len least significant bits of key. */
 static BROTLI_INLINE int GetNextKey(int key, int len) {
@@ -78,12 +82,11 @@ int BrotliBuildHuffmanTable(HuffmanCode* root_table,
   int table_bits;      /* key length of current table */
   int table_size;      /* size of current table */
   int total_size;      /* sum of root table size and 2nd level table sizes */
-  int* sorted;         /* symbols sorted by code length */
+  int sorted[MAX_CODE_LENGTHS_SIZE];  /* symbols sorted by code length */
   int count[MAX_LENGTH + 1] = { 0 };  /* number of codes of each length */
   int offset[MAX_LENGTH + 1];  /* offsets in sorted table for each length */
 
-  sorted = (int*)malloc((size_t)code_lengths_size * sizeof(*sorted));
-  if (sorted == NULL) {
+  if (PREDICT_FALSE(code_lengths_size > MAX_CODE_LENGTHS_SIZE)) {
     return 0;
   }
 
@@ -117,7 +120,6 @@ int BrotliBuildHuffmanTable(HuffmanCode* root_table,
     for (key = 0; key < total_size; ++key) {
       table[key] = code;
     }
-    free(sorted);
     return total_size;
   }
 
@@ -154,7 +156,6 @@ int BrotliBuildHuffmanTable(HuffmanCode* root_table,
     }
   }
 
-  free(sorted);
   return total_size;
 }