quiche / quiche / 268e4207c5b6f33211d23f7dc79721dd3b89358d / . / http2 / hpack / huffman / hpack_huffman_encoder.cc

// Copyright (c) 2018 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. | |

#include "net/third_party/quiche/src/http2/hpack/huffman/hpack_huffman_encoder.h" | |

#include "net/third_party/quiche/src/http2/hpack/huffman/huffman_spec_tables.h" | |

#include "net/third_party/quiche/src/http2/platform/api/http2_logging.h" | |

// TODO(jamessynge): Remove use of binary literals, that is a C++ 14 feature. | |

namespace http2 { | |

size_t ExactHuffmanSize(Http2StringPiece plain) { | |

size_t bits = 0; | |

for (const uint8_t c : plain) { | |

bits += HuffmanSpecTables::kCodeLengths[c]; | |

} | |

return (bits + 7) / 8; | |

} | |

size_t BoundedHuffmanSize(Http2StringPiece plain) { | |

// TODO(jamessynge): Determine whether we should set the min size for Huffman | |

// encoding much higher (i.e. if less than N, then the savings isn't worth | |

// the cost of encoding and decoding). Of course, we need to decide on a | |

// value function, which might be throughput on a full load test, or a | |

// microbenchmark of the time to encode and then decode a HEADERS frame, | |

// possibly with the cost of crypto included (i.e. crypto is going to have | |

// a fairly constant per-byte cost, so reducing the number of bytes in-transit | |

// reduces the number that must be encrypted and later decrypted). | |

if (plain.size() < 3) { | |

// Huffman encoded string can't be smaller than the plain size for very | |

// short strings. | |

return plain.size(); | |

} | |

// TODO(jamessynge): Measure whether this can be done more efficiently with | |

// nested loops (e.g. make exact measurement of 8 bytes, then check if min | |

// remaining is too long). | |

// Compute the number of bits in an encoding that is shorter than the plain | |

// string (i.e. the number of bits in a string 1 byte shorter than plain), | |

// and use this as the limit of the size of the encoding. | |

const size_t limit_bits = (plain.size() - 1) * 8; | |

// The shortest code length in the Huffman table of the HPACK spec has 5 bits | |

// (e.g. for 0, 1, a and e). | |

const size_t min_code_length = 5; | |

// We can therefore say that all plain text bytes whose code length we've not | |

// yet looked up will take at least 5 bits. | |

size_t min_bits_remaining = plain.size() * min_code_length; | |

size_t bits = 0; | |

for (const uint8_t c : plain) { | |

bits += HuffmanSpecTables::kCodeLengths[c]; | |

min_bits_remaining -= min_code_length; | |

// If our minimum estimate of the total number of bits won't yield an | |

// encoding shorter the plain text, let's bail. | |

const size_t minimum_bits_total = bits + min_bits_remaining; | |

if (minimum_bits_total > limit_bits) { | |

bits += min_bits_remaining; | |

break; | |

} | |

} | |

return (bits + 7) / 8; | |

} | |

void HuffmanEncode(Http2StringPiece plain, Http2String* huffman) { | |

DCHECK(huffman != nullptr); | |

huffman->clear(); // Note that this doesn't release memory. | |

uint64_t bit_buffer = 0; // High-bit is next bit to output. Not clear if that | |

// is more performant than having the low-bit be the | |

// last to be output. | |

size_t bits_unused = 64; // Number of bits available for the next code. | |

for (uint8_t c : plain) { | |

size_t code_length = HuffmanSpecTables::kCodeLengths[c]; | |

if (bits_unused < code_length) { | |

// There isn't enough room in bit_buffer for the code of c. | |

// Flush until bits_unused > 56 (i.e. 64 - 8). | |

do { | |

char h = static_cast<char>(bit_buffer >> 56); | |

bit_buffer <<= 8; | |

bits_unused += 8; | |

// Perhaps would be more efficient if we populated an array of chars, | |

// so we don't have to call push_back each time. Reconsider if used | |

// for production. | |

huffman->push_back(h); | |

} while (bits_unused <= 56); | |

} | |

uint64_t code = HuffmanSpecTables::kRightCodes[c]; | |

size_t shift_by = bits_unused - code_length; | |

bit_buffer |= (code << shift_by); | |

bits_unused -= code_length; | |

} | |

// bit_buffer contains (64-bits_unused) bits that still need to be flushed. | |

// Output whole bytes until we don't have any whole bytes left. | |

size_t bits_used = 64 - bits_unused; | |

while (bits_used >= 8) { | |

char h = static_cast<char>(bit_buffer >> 56); | |

bit_buffer <<= 8; | |

bits_used -= 8; | |

huffman->push_back(h); | |

} | |

if (bits_used > 0) { | |

// We have less than a byte left to output. The spec calls for padding out | |

// the final byte with the leading bits of the EOS symbol (30 1-bits). | |

constexpr uint64_t leading_eos_bits = 0b11111111; | |

bit_buffer |= (leading_eos_bits << (56 - bits_used)); | |

char h = static_cast<char>(bit_buffer >> 56); | |

huffman->push_back(h); | |

} | |

} | |

} // namespace http2 |