Faster huffman encoding

lib/nghttp2_hd_huffman.c

@@ -29,114 +29,7 @@
 #include <stdio.h>
 
 #include "nghttp2_hd.h"
-
-/*
- * Encodes huffman code |sym| into |*dest_ptr|, whose least |rembits|
- * bits are not filled yet.  The |rembits| must be in range [1, 8],
- * inclusive.  At the end of the process, the |*dest_ptr| is updated
- * and points where next output should be placed. The number of
- * unfilled bits in the pointed location is returned.
- */
-static ssize_t huff_encode_sym(nghttp2_bufs *bufs, size_t *avail_ptr,
-                               size_t rembits, const nghttp2_huff_sym *sym) {
-  int rv;
-  size_t nbits = sym->nbits;
-  uint32_t code = sym->code;
-
-  /* We assume that sym->nbits <= 32 */
-  if (rembits > nbits) {
-    nghttp2_bufs_fast_orb_hold(bufs, (uint8_t)(code << (rembits - nbits)));
-    return (ssize_t)(rembits - nbits);
-  }
-
-  if (rembits == nbits) {
-    nghttp2_bufs_fast_orb(bufs, (uint8_t)code);
-    --*avail_ptr;
-    return 8;
-  }
-
-  nghttp2_bufs_fast_orb(bufs, (uint8_t)(code >> (nbits - rembits)));
-  --*avail_ptr;
-
-  nbits -= rembits;
-  if (nbits & 0x7) {
-    /* align code to MSB byte boundary */
-    code <<= 8 - (nbits & 0x7);
-  }
-
-  if (*avail_ptr < (nbits + 7) / 8) {
-    /* slow path */
-    if (nbits > 24) {
-      rv = nghttp2_bufs_addb(bufs, (uint8_t)(code >> 24));
-      if (rv != 0) {
-        return rv;
-      }
-      nbits -= 8;
-    }
-    if (nbits > 16) {
-      rv = nghttp2_bufs_addb(bufs, (uint8_t)(code >> 16));
-      if (rv != 0) {
-        return rv;
-      }
-      nbits -= 8;
-    }
-    if (nbits > 8) {
-      rv = nghttp2_bufs_addb(bufs, (uint8_t)(code >> 8));
-      if (rv != 0) {
-        return rv;
-      }
-      nbits -= 8;
-    }
-    if (nbits == 8) {
-      rv = nghttp2_bufs_addb(bufs, (uint8_t)code);
-      if (rv != 0) {
-        return rv;
-      }
-      *avail_ptr = nghttp2_bufs_cur_avail(bufs);
-      return 8;
-    }
-
-    rv = nghttp2_bufs_addb_hold(bufs, (uint8_t)code);
-    if (rv != 0) {
-      return rv;
-    }
-    *avail_ptr = nghttp2_bufs_cur_avail(bufs);
-    return (ssize_t)(8 - nbits);
-  }
-
-  /* fast path, since most code is less than 8 */
-  if (nbits < 8) {
-    nghttp2_bufs_fast_addb_hold(bufs, (uint8_t)code);
-    *avail_ptr = nghttp2_bufs_cur_avail(bufs);
-    return (ssize_t)(8 - nbits);
-  }
-
-  /* handle longer code path */
-  if (nbits > 24) {
-    nghttp2_bufs_fast_addb(bufs, (uint8_t)(code >> 24));
-    nbits -= 8;
-  }
-
-  if (nbits > 16) {
-    nghttp2_bufs_fast_addb(bufs, (uint8_t)(code >> 16));
-    nbits -= 8;
-  }
-
-  if (nbits > 8) {
-    nghttp2_bufs_fast_addb(bufs, (uint8_t)(code >> 8));
-    nbits -= 8;
-  }
-
-  if (nbits == 8) {
-    nghttp2_bufs_fast_addb(bufs, (uint8_t)code);
-    *avail_ptr = nghttp2_bufs_cur_avail(bufs);
-    return 8;
-  }
-
-  nghttp2_bufs_fast_addb_hold(bufs, (uint8_t)code);
-  *avail_ptr = nghttp2_bufs_cur_avail(bufs);
-  return (ssize_t)(8 - nbits);
-}
+#include "nghttp2_net.h"
 
 size_t nghttp2_hd_huff_encode_count(const uint8_t *src, size_t len) {
   size_t i;
@@ -151,40 +44,60 @@ size_t nghttp2_hd_huff_encode_count(const uint8_t *src, size_t len) {
 
 int nghttp2_hd_huff_encode(nghttp2_bufs *bufs, const uint8_t *src,
                            size_t srclen) {
-  int rv;
-  ssize_t rembits = 8;
-  size_t i;
+  const nghttp2_huff_sym *sym;
+  const uint8_t *end = src + srclen;
+  uint64_t code = 0;
+  uint32_t x;
+  size_t nbits = 0;
   size_t avail;
+  int rv;
 
   avail = nghttp2_bufs_cur_avail(bufs);
 
-  for (i = 0; i < srclen; ++i) {
-    const nghttp2_huff_sym *sym = &huff_sym_table[src[i]];
-    if (rembits == 8) {
-      if (avail) {
-        nghttp2_bufs_fast_addb_hold(bufs, 0);
-      } else {
-        rv = nghttp2_bufs_addb_hold(bufs, 0);
+  for (; src != end;) {
+    sym = &huff_sym_table[*src++];
+    code |= (uint64_t)sym->code << (32 - nbits);
+    nbits += sym->nbits;
+    if (nbits < 32) {
+      continue;
+    }
+    if (avail >= 4) {
+      x = htonl((uint32_t)(code >> 32));
+      memcpy(bufs->cur->buf.last, &x, 4);
+      bufs->cur->buf.last += 4;
+      avail -= 4;
+      code <<= 32;
+      nbits -= 32;
+      continue;
+    }
+
+    for (; nbits >= 8;) {
+      rv = nghttp2_bufs_addb(bufs, (uint8_t)(code >> 56));
       if (rv != 0) {
         return rv;
       }
+      code <<= 8;
+      nbits -= 8;
+    }
+
     avail = nghttp2_bufs_cur_avail(bufs);
   }
+
+  for (; nbits >= 8;) {
+    rv = nghttp2_bufs_addb(bufs, (uint8_t)(code >> 56));
+    if (rv != 0) {
+      return rv;
     }
-    rembits = huff_encode_sym(bufs, &avail, (size_t)rembits, sym);
-    if (rembits < 0) {
-      return (int)rembits;
+    code <<= 8;
+    nbits -= 8;
   }
+
+  if (nbits) {
+    rv = nghttp2_bufs_addb(
+        bufs, (uint8_t)((uint8_t)(code >> 56) | ((1 << (8 - nbits)) - 1)));
+    if (rv != 0) {
+      return rv;
     }
-  /* 256 is special terminal symbol, pad with its prefix */
-  if (rembits < 8) {
-    /* if rembits < 8, we should have at least 1 buffer space
-       available */
-    const nghttp2_huff_sym *sym = &huff_sym_table[256];
-    assert(avail);
-    /* Caution we no longer adjust avail here */
-    nghttp2_bufs_fast_orb(
-        bufs, (uint8_t)(sym->code >> (sym->nbits - (size_t)rembits)));
   }
 
   return 0;

mkhufftbl.py

@@ -423,9 +423,12 @@ typedef struct {
     print '''\
 const nghttp2_huff_sym huff_sym_table[] = {'''
     for i in range(257):
+        nbits = symbol_tbl[i][0]
+        k = int(symbol_tbl[i][1], 16)
+        k = k << (32 - nbits)
         print '''\
   {{ {}, 0x{}u }}{}\
-'''.format(symbol_tbl[i][0], symbol_tbl[i][1], ',' if i < 256 else '')
+'''.format(symbol_tbl[i][0], hex(k)[2:], ',' if i < 256 else '')
     print '};'
     print ''