nghttp2_hd: Rewrite huffman decoding

lib/nghttp2_hd.h

@@ -371,19 +371,6 @@ ssize_t nghttp2_hd_huff_encode(uint8_t *dest, size_t destlen,
                                const uint8_t *src, size_t srclen,
                                nghttp2_hd_side side);
 
-/*
- * Counts the number of required bytes to decode |src| with length
- * |srclen|. The given input must be padded with the prefix of
- * terminal code. If |side| is NGHTTP2_HD_SIDE_REQUEST, the request
- * huffman code table is used. Otherwise, the response code table is
- * used.
- *
- * This function returns the number of required bytes to decode given
- * data if it succeeds, or -1.
- */
-ssize_t nghttp2_hd_huff_decode_count(const uint8_t *src, size_t srclen,
-                                     nghttp2_hd_side side);
-
 /*
  * Decodes the given data |src| with length |srclen|. This function
  * allocates memory to store the result and assigns the its pointer to

lib/nghttp2_hd_huffman.c

@@ -31,64 +31,11 @@
 #include "nghttp2_hd.h"
 
 extern const nghttp2_huff_sym req_huff_sym_table[];
-extern const int16_t req_huff_decode_table[][256];
+extern const nghttp2_huff_decode req_huff_decode_table[][16];
 
 extern const nghttp2_huff_sym res_huff_sym_table[];
-extern const int16_t res_huff_decode_table[][256];
+extern const nghttp2_huff_decode res_huff_decode_table[][16];
 
-/*
- * Returns next 8 bits of data from |in|, starting |bitoff| bits
- * offset. If there are fewer bits left than |bitoff|, the left bits
- * with padded with 0 are returned. The |bitoff| must be strictly less
- * than 8.
- */
-static uint8_t get_prefix_byte(const uint8_t *in, size_t len, size_t bitoff)
-{
-  uint8_t b;
-  if(bitoff == 0) {
-    return *in;
-  }
-  b = *in << bitoff;
-  if(len > 1) {
-    b |= *(in + 1) >> (8 - bitoff);
-  }
-  return b;
-}
-
-/*
- * Decodes next byte from input |in| with length |len|, starting
- * |bitoff| bit offset.
- *
- * This function returns the decoded symbol number (0-255 and 256 for
- * special terminal symbol) if it succeeds, or -1.
- */
-static int huff_decode(const uint8_t *in, size_t len, size_t bitoff,
-                       const nghttp2_huff_sym *huff_sym_table,
-                       const huff_decode_table_type *huff_decode_table)
-{
-  int rv = 0;
-  size_t len_orig = len;
-  if(len == 0) {
-    return -1;
-  }
-  for(;;) {
-    rv = huff_decode_table[rv][get_prefix_byte(in, len, bitoff)];
-    if(rv >= 0) {
-      break;
-    }
-    /* Negative return value means we need to lookup next table. */
-    rv = -rv;
-    ++in;
-    --len;
-    if(len == 0) {
-      return -1;
-    }
-  }
-  if(bitoff + huff_sym_table[rv].nbits > len_orig * 8) {
-    return -1;
-  }
-  return rv;
-}
 /*
  * Encodes huffman code |sym| into |*dest_ptr|, whose least |rembits|
  * bits are not filled yet.  The |rembits| must be in range [1, 8],
@@ -167,89 +114,36 @@ ssize_t nghttp2_hd_huff_encode(uint8_t *dest, size_t destlen,
   return dest - dest_first;
 }
 
-static int check_last_byte(const uint8_t *src, size_t srclen, size_t idx,
-                           size_t bitoff)
-{
-  uint8_t last_mask = (1 << (8 - bitoff)) - 1;
-  return idx + 1 == srclen && bitoff > 0 &&
-    (src[idx] & last_mask) == last_mask;
-}
-
-ssize_t nghttp2_hd_huff_decode_count(const uint8_t *src, size_t srclen,
-                                     nghttp2_hd_side side)
-{
-  size_t bitoff = 0;
-  size_t i, j;
-  const nghttp2_huff_sym *huff_sym_table;
-  const huff_decode_table_type *huff_decode_table;
-
-  if(side == NGHTTP2_HD_SIDE_REQUEST) {
-    huff_sym_table = req_huff_sym_table;
-    huff_decode_table = req_huff_decode_table;
-  } else {
-    huff_sym_table = res_huff_sym_table;
-    huff_decode_table = res_huff_decode_table;
-  }
-  j = 0;
-  for(i = 0; i < srclen;) {
-    int rv = huff_decode(src + i, srclen - i, bitoff,
-                         huff_sym_table, huff_decode_table);
-    if(rv == -1) {
-      if(check_last_byte(src, srclen, i, bitoff)) {
-        break;
-      }
-      return -1;
-    }
-    if(rv == 256) {
-      /* 256 is special terminal symbol and it should not encoded in
-         byte string. */
-      return -1;
-    }
-    j++;
-    bitoff += huff_sym_table[rv].nbits;
-    i += bitoff / 8;
-    bitoff &= 0x7;
-  }
-  return j;
-}
-
 ssize_t nghttp2_hd_huff_decode(uint8_t **dest_ptr,
                                const uint8_t *src, size_t srclen,
                                nghttp2_hd_side side)
 {
-  size_t bitoff = 0;
-  size_t i, j;
-  const nghttp2_huff_sym *huff_sym_table;
+  size_t i, j, k;
   const huff_decode_table_type *huff_decode_table;
   uint8_t *dest = NULL;
   size_t destlen = 0;
   int rv;
+  int16_t state = 0;
+  const nghttp2_huff_decode *t = NULL;
 
+  /* We use the decoding algorithm described in
+     http://graphics.ics.uci.edu/pub/Prefix.pdf */
   if(side == NGHTTP2_HD_SIDE_REQUEST) {
-    huff_sym_table = req_huff_sym_table;
     huff_decode_table = req_huff_decode_table;
   } else {
-    huff_sym_table = res_huff_sym_table;
     huff_decode_table = res_huff_decode_table;
   }
   j = 0;
-  for(i = 0; i < srclen;) {
-    rv = huff_decode(src + i, srclen - i, bitoff,
-                     huff_sym_table, huff_decode_table);
-    if(rv == -1) {
-      if(check_last_byte(src, srclen, i, bitoff)) {
-        break;
-      }
-      rv = NGHTTP2_ERR_HEADER_COMP;
-      goto fail;
-    }
-    if(rv == 256) {
-      /* 256 is special terminal symbol and it should not encoded in
-         byte string. */
+  for(i = 0; i < srclen; ++i) {
+    uint8_t in = src[i] >> 4;
+    for(k = 0; k < 2; ++k) {
+      t = &huff_decode_table[state][in];
+      if(t->state == -1) {
         rv = NGHTTP2_ERR_HEADER_COMP;
         goto fail;
       }
-    if(j == destlen) {
+      if(t->flags & NGHTTP2_HUFF_SYM) {
+        if(destlen == j) {
           size_t new_len = j == 0 ? 32 : j * 2;
           uint8_t *new_dest = realloc(dest, new_len);
           if(new_dest == NULL) {
@@ -259,10 +153,15 @@ ssize_t nghttp2_hd_huff_decode(uint8_t **dest_ptr,
           dest = new_dest;
           destlen = new_len;
         }
-    dest[j++] = rv;
-    bitoff += huff_sym_table[rv].nbits;
-    i += bitoff / 8;
-    bitoff &= 0x7;
+        dest[j++] = t->sym;
+      }
+      state = t->state;
+      in = src[i] & 0xf;
+    }
+  }
+  if(srclen && (t->flags & NGHTTP2_HUFF_ACCEPTED) == 0) {
+    rv = NGHTTP2_ERR_HEADER_COMP;
+    goto fail;
   }
   *dest_ptr = dest;
   return j;

lib/nghttp2_hd_huffman.h

@@ -31,7 +31,21 @@
 
 #include <nghttp2/nghttp2.h>
 
-typedef int16_t huff_decode_table_type[256];
+enum {
+  /* FSA accepts this state as the end of huffman encoding
+     sequence. */
+  NGHTTP2_HUFF_ACCEPTED = 1,
+  /* This state emits symbol */
+  NGHTTP2_HUFF_SYM = (1 << 1)
+} nghttp2_huff_decode_flag;
+
+typedef struct {
+  int16_t state;
+  uint8_t flags;
+  uint8_t sym;
+} nghttp2_huff_decode;
+
+typedef nghttp2_huff_decode huff_decode_table_type[16];
 
 typedef struct {
   /* The number of bits in this code */

mkhufftbl.py

@@ -3,9 +3,13 @@ import re
 import sys
 
 class Node:
-    def __init__(self, depth):
-        self.depth = depth
-        self.children = {}
+    def __init__(self, term = None):
+        self.term = term
+        self.left = None
+        self.right = None
+        self.trans = []
+        self.id = None
+        self.accept = False
 
 def to_bin(s):
     res = []
@@ -21,27 +25,99 @@ def to_bin(s):
 
 nodes = []
 
-def insert(node, sym, binpat, nbits, pidx):
-    if pidx == len(binpat) - 1:
-        #assert(binpat[pidx] not in node.children)
-        mx = (8 - (nbits & 0x7)) & 0x7;
-        #print "last", bin(binpat[pidx]), mx
-        for i in range(1 << mx):
-            node.children[binpat[pidx] + i] = sym
+def insert(node, sym, bits):
+    if len(bits) == 0:
+        node.term = sym
+        return
     else:
-        if binpat[pidx] not in node.children:
-            node.children[binpat[pidx]] = -len(nodes)
-            nextnode = Node(pidx + 1)
-            nodes.append(nextnode)
+        if bits[0] == '0':
+            if node.left is None:
+                node.left = Node()
+            child = node.left
         else:
-            nextnode = nodes[-node.children[binpat[pidx]]]
-        insert(nextnode, sym, binpat, nbits, pidx + 1)
+            if node.right is None:
+                node.right = Node()
+            child = node.right
+        insert(child, sym, bits[1:])
+
+def traverse(node, bits, syms, start_node, root, depth):
+    if depth == 4:
+        if 256 in syms:
+            syms = []
+            node = None
+        start_node.trans.append((node, bits, syms))
+        return
+
+    if node.term is not None:
+        node = root
+
+    def go(node, bit):
+        nbits = list(bits)
+        nbits.append(bit)
+        nsyms = list(syms)
+        if node.term is not None:
+            nsyms.append(node.term)
+        traverse(node, nbits, nsyms, start_node, root, depth + 1)
+
+    go(node.left, 0)
+    go(node.right, 1)
+
+idseed = 0
+
+def dfs_setid(node, prefix):
+    if node.term is not None:
+        return
+    if len(prefix) <= 7 and [1] * len(prefix) == prefix:
+        node.accept = True
+    global idseed
+    node.id = idseed
+    idseed += 1
+    dfs_setid(node.left, prefix + [0])
+    dfs_setid(node.right, prefix + [1])
+
+def dfs(node, root):
+    if node is None:
+        return
+    traverse(node, [], [], node, root, 0)
+    dfs(node.left, root)
+    dfs(node.right, root)
+
+NGHTTP2_HUFF_ACCEPTED = 1
+NGHTTP2_HUFF_SYM = 1 << 1
+
+def dfs_print(node):
+    if node.term is not None:
+        return
+    print '/* {} */'.format(node.id)
+    print '{'
+    for nd, bits, syms in node.trans:
+        outlen = len(syms)
+        flags = 0
+        if outlen == 0:
+            out = 0
+        else:
+            assert(outlen == 1)
+            out = syms[0]
+            flags |= NGHTTP2_HUFF_SYM
+        if nd is None:
+            id = -1
+        else:
+            id = nd.id
+            if id is None:
+                # if nd.id is None, it is a leaf node
+                id = 0
+                flags |= NGHTTP2_HUFF_ACCEPTED
+            elif nd.accept:
+                flags |= NGHTTP2_HUFF_ACCEPTED
+        print '  {{{}, 0x{:02x}, {}}},'.format(id, flags, out)
+    print '},'
+    dfs_print(node.left)
+    dfs_print(node.right)
 
 symbol_tbl = [(None, 0) for i in range(257)]
 tables = {}
 
-root = Node(0)
-nodes.append(root)
+root = Node()
 
 for line in sys.stdin:
     m = re.match(r'.*\(\s*(\d+)\) ([|01]+) \[(\d+)\]\s+(\S+).*', line)
@@ -50,14 +126,17 @@ for line in sys.stdin:
         if len(m.group(4)) > 8:
             raise Error('Code is more than 4 bytes long')
         sym = int(m.group(1))
-        pat = re.sub(r'\|', '', m.group(2))
+        bits = re.sub(r'\|', '', m.group(2))
         nbits = int(m.group(3))
-        assert(len(pat) == nbits)
-        binpat = to_bin(pat)
+        assert(len(bits) == nbits)
+        binpat = to_bin(bits)
         assert(len(binpat) == (nbits+7)/8)
         symbol_tbl[sym] = (binpat, nbits, m.group(4))
         #print "Inserting", sym
-        insert(root, sym, binpat, nbits, 0)
+        insert(root, sym, bits)
+
+dfs_setid(root, [])
+dfs(root, root)
 
 print '''\
 typedef struct {
@@ -67,7 +146,7 @@ typedef struct {
 '''
 
 print '''\
-nghttp2_huff_sym huff_sym_table[] = {'''
+const nghttp2_huff_sym huff_sym_table[] = {'''
 for i in range(257):
     pat = list(symbol_tbl[i][0])
     pat += [0]*(4 - len(pat))
@@ -77,22 +156,22 @@ for i in range(257):
 print '};'
 print ''
 
-print '''int16_t huff_decode_table[][256] = {'''
-for j in range(len(nodes)):
-    node = nodes[j]
-    print '/* {} */'.format(j)
-    print '{'
-    for i in range(256):
-        if i in node.children:
-            sys.stdout.write('''\
- {}{}'''.format(node.children[i], ',' if i < 255 else ''))
-        else:
-            sys.stdout.write(''' NGHTTP2_HD_HUFF_NO_ENT,''')
-        if (i+1)&0x7 == 0:
-            print ''
-    sys.stdout.write('}')
-    if j == len(nodes) - 1:
-        print ''
-    else:
-        print ','
+print '''\
+enum {{
+  NGHTTP2_HUFF_ACCEPTED = {},
+  NGHTTP2_HUFF_SYM = {}
+}} nghttp2_huff_decode_flag;
+'''.format(NGHTTP2_HUFF_ACCEPTED, NGHTTP2_HUFF_SYM)
+
+print '''\
+typedef struct {
+  int16_t state;
+  uint8_t flags;
+  uint8_t sym;
+} nghttp2_huff_decode;
+'''
+
+print '''\
+const nghttp2_huff_decode huff_decode_table[][16] = {'''
+dfs_print(root)
 print '};'