Implement better HPACK Huffman code decoder.

net/spdy/hpack/hpack_huffman_decoder.cc

+// Copyright 2016 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+//
+// Decoder for strings encoded using the HPACK Huffman Code (see
+// https://httpwg.github.io/specs/rfc7541.html#huffman.code).
+//
+// This implementation is inspired by the One-Shift algorithm described in
+// "On the Implementation of Minimum Redundancy Prefix Codes", by Alistair
+// Moffat and Andrew Turpin, 1997.
+// See also https://en.wikipedia.org/wiki/Canonical_Huffman_code for background
+// on canonical Huffman codes.
+//
+// This decoder differs from that in .../spdy/hpack/hpack_huffman_table.cc
+// as follows:
+//   1) It decodes only the code described in RFC7541, where as the older
+//      implementation supported any canonical Huffman code provided at run
+//      time.
+//   2) It uses a fixed amount of memory allocated at build time; it doesn't
+//      construct a tree of of decoding tables based on an encoding
+//      table provided at run time.
+//   3) In benchmarks it runs from 10% to 70% faster, based on the length
+//      of the strings (faster for longer strings). Some of the improvements
+//      could be back ported, but others are fundamental to the approach.
+
+#include "net/spdy/hpack/hpack_huffman_decoder.h"
+
+#include <bitset>
+#include <limits>
+#include <utility>
+
+#include "base/logging.h"
+#include "net/spdy/hpack/hpack_input_stream.h"
+
+namespace net {
+namespace {
+
+typedef HpackHuffmanDecoder::HuffmanWord HuffmanWord;
+typedef HpackHuffmanDecoder::HuffmanCodeLength HuffmanCodeLength;
+
+const HuffmanCodeLength kHuffmanWordLength =
+    std::numeric_limits<HuffmanWord>::digits;
+
+const HuffmanCodeLength kMinCodeLength = 5;
+const HuffmanCodeLength kMaxCodeLength = 30;
+
+const HuffmanWord kInvalidLJCode = ~static_cast<HuffmanWord>(0);
+// Length of a code in bits to the first code with that length, left-justified.
+// Note that this can be computed from kLengthToFirstCanonical.
+const HuffmanWord kLengthToFirstLJCode[] = {
+    kInvalidLJCode,  // There are no codes of length 0.
+    kInvalidLJCode,  // There are no codes of length 1.
+    kInvalidLJCode,  // There are no codes of length 2.
+    kInvalidLJCode,  // There are no codes of length 3.
+    kInvalidLJCode,  // There are no codes of length 4.
+    0x00000000,      // Length 5.
+    0x50000000,      // Length 6.
+    0xb8000000,      // Length 7.
+    0xf8000000,      // Length 8.
+    kInvalidLJCode,  // There are no codes of length 9.
+    0xfe000000,      // Length 10.
+    0xff400000,      // Length 11.
+    0xffa00000,      // Length 12.
+    0xffc00000,      // Length 13.
+    0xfff00000,      // Length 14.
+    0xfff80000,      // Length 15.
+    kInvalidLJCode,  // There are no codes of length 16.
+    kInvalidLJCode,  // There are no codes of length 17.
+    kInvalidLJCode,  // There are no codes of length 18.
+    0xfffe0000,      // Length 19.
+    0xfffe6000,      // Length 20.
+    0xfffee000,      // Length 21.
+    0xffff4800,      // Length 22.
+    0xffffb000,      // Length 23.
+    0xffffea00,      // Length 24.
+    0xfffff600,      // Length 25.
+    0xfffff800,      // Length 26.
+    0xfffffbc0,      // Length 27.
+    0xfffffe20,      // Length 28.
+    kInvalidLJCode,  // There are no codes of length 29.
+    0xfffffff0,      // Length 30.
+};
+
+// TODO(jamessynge): Determine the performance impact of different types for
+// the elements of this array (i.e. a larger type uses more cache, yet might
+// better on some architectures).
+const uint8_t kInvalidCanonical = 255;
+// Maps from length of a code to the first 'canonical symbol' with that length.
+const uint8_t kLengthToFirstCanonical[] = {
+    kInvalidCanonical,  // Length 0, 0 codes.
+    kInvalidCanonical,  // Length 1, 0 codes.
+    kInvalidCanonical,  // Length 2, 0 codes.
+    kInvalidCanonical,  // Length 3, 0 codes.
+    kInvalidCanonical,  // Length 4, 0 codes.
+    0,                  // Length 5, 10 codes.
+    10,                 // Length 6, 26 codes.
+    36,                 // Length 7, 32 codes.
+    68,                 // Length 8, 6 codes.
+    kInvalidCanonical,  // Length 9, 0 codes.
+    74,                 // Length 10, 5 codes.
+    79,                 // Length 11, 3 codes.
+    82,                 // Length 12, 2 codes.
+    84,                 // Length 13, 6 codes.
+    90,                 // Length 14, 2 codes.
+    92,                 // Length 15, 3 codes.
+    kInvalidCanonical,  // Length 16, 0 codes.
+    kInvalidCanonical,  // Length 17, 0 codes.
+    kInvalidCanonical,  // Length 18, 0 codes.
+    95,                 // Length 19, 3 codes.
+    98,                 // Length 20, 8 codes.
+    106,                // Length 21, 13 codes.
+    119,                // Length 22, 26 codes.
+    145,                // Length 23, 29 codes.
+    174,                // Length 24, 12 codes.
+    186,                // Length 25, 4 codes.
+    190,                // Length 26, 15 codes.
+    205,                // Length 27, 19 codes.
+    224,                // Length 28, 29 codes.
+    kInvalidCanonical,  // Length 29, 0 codes.
+    253,                // Length 30, 4 codes.
+};
+
+// Mapping from canonical symbol (0 to 255) to actual symbol.
+// clang-format off
+const uint8_t kCanonicalToSymbol[] = {
+    '0',  '1',  '2',  'a',  'c',  'e',  'i',  'o',
+    's',  't',  0x20, '%',  '-',  '.',  '/',  '3',
+    '4',  '5',  '6',  '7',  '8',  '9',  '=',  'A',
+    '_',  'b',  'd',  'f',  'g',  'h',  'l',  'm',
+    'n',  'p',  'r',  'u',  ':',  'B',  'C',  'D',
+    'E',  'F',  'G',  'H',  'I',  'J',  'K',  'L',
+    'M',  'N',  'O',  'P',  'Q',  'R',  'S',  'T',
+    'U',  'V',  'W',  'Y',  'j',  'k',  'q',  'v',
+    'w',  'x',  'y',  'z',  '&',  '*',  ',',  ';',
+    'X',  'Z',  '!',  '\"', '(',  ')',  '?',  '\'',
+    '+',  '|',  '#',  '>',  0x00, '$',  '@',  '[',
+    ']',  '~',  '^',  '}',  '<',  '`',  '{',  '\\',
+    0xc3, 0xd0, 0x80, 0x82, 0x83, 0xa2, 0xb8, 0xc2,
+    0xe0, 0xe2, 0x99, 0xa1, 0xa7, 0xac, 0xb0, 0xb1,
+    0xb3, 0xd1, 0xd8, 0xd9, 0xe3, 0xe5, 0xe6, 0x81,
+    0x84, 0x85, 0x86, 0x88, 0x92, 0x9a, 0x9c, 0xa0,
+    0xa3, 0xa4, 0xa9, 0xaa, 0xad, 0xb2, 0xb5, 0xb9,
+    0xba, 0xbb, 0xbd, 0xbe, 0xc4, 0xc6, 0xe4, 0xe8,
+    0xe9, 0x01, 0x87, 0x89, 0x8a, 0x8b, 0x8c, 0x8d,
+    0x8f, 0x93, 0x95, 0x96, 0x97, 0x98, 0x9b, 0x9d,
+    0x9e, 0xa5, 0xa6, 0xa8, 0xae, 0xaf, 0xb4, 0xb6,
+    0xb7, 0xbc, 0xbf, 0xc5, 0xe7, 0xef, 0x09, 0x8e,
+    0x90, 0x91, 0x94, 0x9f, 0xab, 0xce, 0xd7, 0xe1,
+    0xec, 0xed, 0xc7, 0xcf, 0xea, 0xeb, 0xc0, 0xc1,
+    0xc8, 0xc9, 0xca, 0xcd, 0xd2, 0xd5, 0xda, 0xdb,
+    0xee, 0xf0, 0xf2, 0xf3, 0xff, 0xcb, 0xcc, 0xd3,
+    0xd4, 0xd6, 0xdd, 0xde, 0xdf, 0xf1, 0xf4, 0xf5,
+    0xf6, 0xf7, 0xf8, 0xfa, 0xfb, 0xfc, 0xfd, 0xfe,
+    0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08, 0x0b,
+    0x0c, 0x0e, 0x0f, 0x10, 0x11, 0x12, 0x13, 0x14,
+    0x15, 0x17, 0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d,
+    0x1e, 0x1f, 0x7f, 0xdc, 0xf9, 0x0a, 0x0d, 0x16,
+};
+// clang-format on
+
+#ifndef NDEBUG
+
+// Only used in DLOG.
+bool IsEOSPrefix(HuffmanWord bits, HuffmanCodeLength bits_available) {
+  if (bits_available == 0) {
+    return true;
+  }
+  // We expect all the bits below the high order |bits_available| bits
+  // to be cleared.
+  HuffmanWord expected = HuffmanWord(0xffffffff) << (32 - bits_available);
+  return bits == expected;
+}
+
+#endif  // NDEBUG
+
+}  // namespace
+
+// TODO(jamessynge): Should we read these magic numbers from
+// kLengthToFirstLJCode? Would that reduce cache consumption? Slow decoding?
+// TODO(jamessynge): Is this being inlined by the compiler? Should we inline
+// into DecodeString the tests for code lengths 5 through 8 (> 99% of codes
+// according to the HPACK spec)?
+HpackHuffmanDecoder::HuffmanCodeLength HpackHuffmanDecoder::CodeLengthOfPrefix(
+    HpackHuffmanDecoder::HuffmanWord value) {
+  HuffmanCodeLength length;
+  if (value < 0xb8000000) {
+    if (value < 0x50000000) {
+      length = 5;
+    } else {
+      length = 6;
+    }
+  } else {
+    if (value < 0xfe000000) {
+      if (value < 0xf8000000) {
+        length = 7;
+      } else {
+        length = 8;
+      }
+    } else {
+      if (value < 0xffc00000) {
+        if (value < 0xffa00000) {
+          if (value < 0xff400000) {
+            length = 10;
+          } else {
+            length = 11;
+          }
+        } else {
+          length = 12;
+        }
+      } else {
+        if (value < 0xfffe0000) {
+          if (value < 0xfff80000) {
+            if (value < 0xfff00000) {
+              length = 13;
+            } else {
+              length = 14;
+            }
+          } else {
+            length = 15;
+          }
+        } else {
+          if (value < 0xffff4800) {
+            if (value < 0xfffee000) {
+              if (value < 0xfffe6000) {
+                length = 19;
+              } else {
+                length = 20;
+              }
+            } else {
+              length = 21;
+            }
+          } else {
+            if (value < 0xffffea00) {
+              if (value < 0xffffb000) {
+                length = 22;
+              } else {
+                length = 23;
+              }
+            } else {
+              if (value < 0xfffffbc0) {
+                if (value < 0xfffff800) {
+                  if (value < 0xfffff600) {
+                    length = 24;
+                  } else {
+                    length = 25;
+                  }
+                } else {
+                  length = 26;
+                }
+              } else {
+                if (value < 0xfffffff0) {
+                  if (value < 0xfffffe20) {
+                    length = 27;
+                  } else {
+                    length = 28;
+                  }
+                } else {
+                  length = 30;
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+  return length;
+}
+
+HuffmanWord HpackHuffmanDecoder::DecodeToCanonical(
+    HuffmanCodeLength code_length,
+    HuffmanWord bits) {
+  DCHECK_LE(kMinCodeLength, code_length);
+  DCHECK_LE(code_length, kMaxCodeLength);
+
+  // What is the first left-justified code of length |code_length|?
+  HuffmanWord first_lj_code = kLengthToFirstLJCode[code_length];
+  DCHECK_NE(kInvalidLJCode, first_lj_code);
+
+  // Which canonical symbol corresponds to the high order |code_length|
+  // bits of |first_lj_code|?
+  HuffmanWord first_canonical = kLengthToFirstCanonical[code_length];
+  DCHECK_NE(kInvalidCanonical, first_canonical);
+
+  // What is the position of the canonical symbol being decoded within
+  // the canonical symbols of length |code_length|?
+  HuffmanWord ordinal_in_length =
+      ((bits - first_lj_code) >> (kHuffmanWordLength - code_length));
+
+  // Combined these two to produce the position of the canonical symbol
+  // being decoded within all of the canonical symbols.
+  return first_canonical + ordinal_in_length;
+}
+
+char HpackHuffmanDecoder::CanonicalToSource(HuffmanWord canonical) {
+  DCHECK_LT(canonical, 256u);
+  return static_cast<char>(kCanonicalToSymbol[canonical]);
+}
+
+// TODO(jamessynge): Maybe further refactorings, including just passing in a
+// StringPiece instead of an HpackInputStream, thus avoiding the PeekBits calls,
+// and also allowing us to separate the code into portions dealing with long
+// strings, and a later portion dealing with the last few bytes of strings.
+// TODO(jamessynge): Determine if that is worth it by adding some counters to
+// measure the distribution of string sizes seen in practice.
+bool HpackHuffmanDecoder::DecodeString(HpackInputStream* in,
+                                       size_t out_capacity,
+                                       std::string* out) {
+  out->clear();
+
+  // Load |bits| with the leading bits of the input stream, left justified
+  // (i.e. the bits of the first byte are the high-order bits of |bits|,
+  // and the bits of the fourth byte are the low-order bits of |bits|).
+  // |peeked_success| if there are more bits in |*in| (i.e. the encoding
+  // of the string to be decoded is more than 4 bytes).
+
+  auto bits_available_and_bits = in->InitializePeekBits();
+  HuffmanCodeLength bits_available = bits_available_and_bits.first;
+  HuffmanWord bits = bits_available_and_bits.second;
+
+  // |peeked_success| tracks whether the previous PeekBits call was able to
+  // store any new bits into |bits|. For the first pass through the loop below
+  // the value false is appropriate:
+  //     If we have 32 bits (i.e. the input has at least 4 bytes), then:
+  //         |peeked_sucess| is not examined because |code_length| is
+  //         at most 30 in the HPACK Huffman Code.
+  //     If we have at most 24 bits (i.e. the input has at most 3 bytes), then:
+  //         It is possible that the very first |code_length| is greater than
+  //         |bits_available|, in which case we need to read peeked_success to
+  //         determine whether we should try to read more input, or have already
+  //         loaded |bits| with the final bits of the input.
+  // After the first loop |peeked_success| has been set by a call to PeekBits.
+  bool peeked_success = false;
+
+  while (true) {
+    const HuffmanCodeLength code_length = CodeLengthOfPrefix(bits);
+    DCHECK_LE(kMinCodeLength, code_length);
+    DCHECK_LE(code_length, kMaxCodeLength);
+    DVLOG(1) << "bits: 0b" << std::bitset<32>(bits)
+             << " (avail=" << bits_available << ")"
+             << "    prefix length: " << code_length
+             << (code_length > bits_available ? "      *****" : "");
+    if (code_length > bits_available) {
+      if (!peeked_success) {
+        // Unable to read enough input for a match. If only a portion of
+        // the last byte remains, this is a successful EOS condition.
+        // Note that this does NOT check whether the available bits are all
+        // set to 1, which the encoder is required to set at EOS, and the
+        // decoder is required to check.
+        // TODO(jamessynge): Discuss whether we should enforce this check,
+        // as required by the RFC, presumably flag guarded so that we can
+        // disable it should it occur a lot. From my testing it appears that
+        // our encoder may be doing this wrong. Sigh.
+        // TODO(jamessynge): Add a counter for how often the remaining bits
+        // are non-zero.
+        in->ConsumeByteRemainder();
+        DLOG_IF(WARNING,
+                (in->HasMoreData() || !IsEOSPrefix(bits, bits_available)))
+            << "bits: 0b" << std::bitset<32>(bits)
+            << " (avail=" << bits_available << ")"
+            << "    prefix length: " << code_length
+            << "    HasMoreData: " << in->HasMoreData();
+        return !in->HasMoreData();
+      }
+      // We're dealing with a long code. It *might* be useful to add a special
+      // method to HpackInputStream for getting more than "at most 8" bits
+      // at a time.
+      do {
+        peeked_success = in->PeekBits(&bits_available, &bits);
+      } while (peeked_success && bits_available < 32);
+    } else {
+      if (out->size() == out_capacity) {
+        // This code would cause us to overflow |out_capacity|.
+        // TODO(jamessynge) Avoid this case by pre-allocating out based on
+        // scaling up the encoded size by 8/5 (shortest codes are 5 bits).
+        DLOG(WARNING) << "Output size too large: " << out_capacity;
+        return false;
+      }
+
+      // Convert from the prefix code of length |code_length| to the
+      // canonical symbol (i.e. where the input symbols (bytes) are ordered by
+      // increasing code length and then by their increasing uint8 value).
+      HuffmanWord canonical = DecodeToCanonical(code_length, bits);
+
+      bits = bits << code_length;
+      bits_available -= code_length;
+      in->ConsumeBits(code_length);
+
+      if (canonical < 256) {
+        out->push_back(CanonicalToSource(canonical));
+      } else {
+        // Encoder is not supposed to explicity encode the EOS symbol (30
+        // 1-bits).
+        // TODO(jamessynge): Discuss returning false here, as required by HPACK.
+        DCHECK(false) << "EOS explicitly encoded!\n"
+                      << "bits: 0b" << std::bitset<32>(bits)
+                      << " (avail=" << bits_available << ")"
+                      << " prefix length: " << code_length
+                      << " canonical: " << canonical;
+      }
+      // Get some more bits for decoding (up to 8). |peeked_success| is true
+      // if we got any bits.
+      peeked_success = in->PeekBits(&bits_available, &bits);
+    }
+    DLOG_IF(WARNING, (VLOG_IS_ON(1) && bits_available < 32 && !peeked_success))
+        << "no more peeking possible";
+  }
+}
+
+}  // namespace net