// Copyright 2005 Google Inc. All Rights Reserved.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are
// met:
//
// * Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
// * Redistributions in binary form must reproduce the above
// copyright notice, this list of conditions and the following disclaimer
// in the documentation and/or other materials provided with the
// distribution.
// * Neither the name of Google Inc. nor the names of its
// contributors may be used to endorse or promote products derived from
// this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "snappy-internal.h"
#include "snappy-sinksource.h"
#include "snappy.h"
#if !
defined(SNAPPY_HAVE_BMI2)
// __BMI2__ is defined by GCC and Clang. Visual Studio doesn't target BMI2
// specifically, but it does define __AVX2__ when AVX2 support is available.
// Fortunately, AVX2 was introduced in Haswell, just like BMI2.
//
// BMI2 is not defined as a subset of AVX2 (unlike SSSE3 and AVX above). So,
// GCC and Clang can build code with AVX2 enabled but BMI2 disabled, in which
// case issuing BMI2 instructions results in a compiler error.
#if defined(__BMI2__) || (
defined(_MSC_VER) &&
defined(__AVX2__))
#define SNAPPY_HAVE_BMI2 1
#else
#define SNAPPY_HAVE_BMI2 0
#endif
#endif // !defined(SNAPPY_HAVE_BMI2)
#if !
defined(SNAPPY_HAVE_X86_CRC32)
#if defined(__SSE4_2__)
#define SNAPPY_HAVE_X86_CRC32 1
#else
#define SNAPPY_HAVE_X86_CRC32 0
#endif
#endif // !defined(SNAPPY_HAVE_X86_CRC32)
#if !
defined(SNAPPY_HAVE_NEON_CRC32)
#if SNAPPY_HAVE_NEON &&
defined(__ARM_FEATURE_CRC32)
#define SNAPPY_HAVE_NEON_CRC32 1
#else
#define SNAPPY_HAVE_NEON_CRC32 0
#endif
#endif // !defined(SNAPPY_HAVE_NEON_CRC32)
#if SNAPPY_HAVE_BMI2 || SNAPPY_HAVE_X86_CRC32
// Please do not replace with <x86intrin.h>. or with headers that assume more
// advanced SSE versions without checking with all the OWNERS.
#include <immintrin.h>
#elif SNAPPY_HAVE_NEON_CRC32
#include <arm_acle.h>
#endif
#include <algorithm>
#include <array>
#include <cstddef>
#include <cstdint>
#include <cstdio>
#include <cstring>
#include <memory>
#include <string>
#include <utility>
#include <vector>
namespace snappy {
namespace {
// The amount of slop bytes writers are using for unconditional copies.
constexpr
int kSlopBytes = 64;
using internal::char_table;
using internal::COPY_1_BYTE_OFFSET;
using internal::COPY_2_BYTE_OFFSET;
using internal::COPY_4_BYTE_OFFSET;
using internal::kMaximumTagLength;
using internal::LITERAL;
#if SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE
using internal::V128;
using internal::V128_Load;
using internal::V128_LoadU;
using internal::V128_Shuffle;
using internal::V128_StoreU;
using internal::V128_DupChar;
#endif
// We translate the information encoded in a tag through a lookup table to a
// format that requires fewer instructions to decode. Effectively we store
// the length minus the tag part of the offset. The lowest significant byte
// thus stores the length. While total length - offset is given by
// entry - ExtractOffset(type). The nice thing is that the subtraction
// immediately sets the flags for the necessary check that offset >= length.
// This folds the cmp with sub. We engineer the long literals and copy-4 to
// always fail this check, so their presence doesn't affect the fast path.
// To prevent literals from triggering the guard against offset < length (offset
// does not apply to literals) the table is giving them a spurious offset of
// 256.
inline constexpr int16_t MakeEntry(int16_t len, int16_t offset) {
return len - (offset << 8);
}
inline constexpr int16_t LengthMinusOffset(
int data,
int type) {
return type == 3 ? 0xFF
// copy-4 (or type == 3)
: type == 2 ? MakeEntry(data + 1, 0)
// copy-2
: type == 1 ? MakeEntry((data & 7) + 4, data >> 3)
// copy-1
: data < 60 ? MakeEntry(data + 1, 1)
// note spurious offset.
: 0xFF;
// long literal
}
inline constexpr int16_t LengthMinusOffset(uint8_t tag) {
return LengthMinusOffset(tag >> 2, tag & 3);
}
template <size_t... Ints>
struct index_sequence {};
template <std::size_t N, size_t... Is>
struct make_index_sequence : make_index_sequence<N - 1, N - 1, Is...> {};
template <size_t... Is>
struct make_index_sequence<0, Is...> : index_sequence<Is...> {};
template <size_t... seq>
constexpr std::array<int16_t, 256> MakeTable(index_sequence<seq...>) {
return std::array<int16_t, 256>{LengthMinusOffset(seq)...};
}
alignas(64)
const std::array<int16_t, 256> kLengthMinusOffset =
MakeTable(make_index_sequence<256>{});
// Given a table of uint16_t whose size is mask / 2 + 1, return a pointer to the
// relevant entry, if any, for the given bytes. Any hash function will do,
// but a good hash function reduces the number of collisions and thus yields
// better compression for compressible input.
//
// REQUIRES: mask is 2 * (table_size - 1), and table_size is a power of two.
inline uint16_t* TableEntry(uint16_t* table, uint32_t bytes, uint32_t mask) {
// Our choice is quicker-and-dirtier than the typical hash function;
// empirically, that seems beneficial. The upper bits of kMagic * bytes are a
// higher-quality hash than the lower bits, so when using kMagic * bytes we
// also shift right to get a higher-quality end result. There's no similar
// issue with a CRC because all of the output bits of a CRC are equally good
// "hashes." So, a CPU instruction for CRC, if available, tends to be a good
// choice.
#if SNAPPY_HAVE_NEON_CRC32
// We use mask as the second arg to the CRC function, as it's about to
// be used anyway; it'd be equally correct to use 0 or some constant.
// Mathematically, _mm_crc32_u32 (or similar) is a function of the
// xor of its arguments.
const uint32_t hash = __crc32cw(bytes, mask);
#elif SNAPPY_HAVE_X86_CRC32
const uint32_t hash = _mm_crc32_u32(bytes, mask);
#else
constexpr uint32_t kMagic = 0x1e35a7bd;
const uint32_t hash = (kMagic * bytes) >> (31 - kMaxHashTableBits);
#endif
return reinterpret_cast<uint16_t*>(
reinterpret_cast<uintptr_t>(table) +
(hash & mask));
}
inline uint16_t* TableEntry4ByteMatch(uint16_t* table, uint32_t bytes,
uint32_t mask) {
constexpr uint32_t kMagic = 2654435761U;
const uint32_t hash = (kMagic * bytes) >> (32 - kMaxHashTableBits);
return reinterpret_cast<uint16_t*>(
reinterpret_cast<uintptr_t>(table) +
(hash & mask));
}
inline uint16_t* TableEntry8ByteMatch(uint16_t* table, uint64_t bytes,
uint32_t mask) {
constexpr uint64_t kMagic = 58295818150454627ULL;
const uint32_t hash = (kMagic * bytes) >> (64 - kMaxHashTableBits);
return reinterpret_cast<uint16_t*>(
reinterpret_cast<uintptr_t>(table) +
(hash & mask));
}
}
// namespace
size_t MaxCompressedLength(size_t source_bytes) {
// Compressed data can be defined as:
// compressed := item* literal*
// item := literal* copy
//
// The trailing literal sequence has a space blowup of at most 62/60
// since a literal of length 60 needs one tag byte + one extra byte
// for length information.
//
// Item blowup is trickier to measure. Suppose the "copy" op copies
// 4 bytes of data. Because of a special check in the encoding code,
// we produce a 4-byte copy only if the offset is < 65536. Therefore
// the copy op takes 3 bytes to encode, and this type of item leads
// to at most the 62/60 blowup for representing literals.
//
// Suppose the "copy" op copies 5 bytes of data. If the offset is big
// enough, it will take 5 bytes to encode the copy op. Therefore the
// worst case here is a one-byte literal followed by a five-byte copy.
// I.e., 6 bytes of input turn into 7 bytes of "compressed" data.
//
// This last factor dominates the blowup, so the final estimate is:
return 32 + source_bytes + source_bytes / 6;
}
namespace {
void UnalignedCopy64(
const void* src,
void* dst) {
char tmp[8];
std::memcpy(tmp, src, 8);
std::memcpy(dst, tmp, 8);
}
void UnalignedCopy128(
const void* src,
void* dst) {
// std::memcpy() gets vectorized when the appropriate compiler options are
// used. For example, x86 compilers targeting SSE2+ will optimize to an SSE2
// load and store.
char tmp[16];
std::memcpy(tmp, src, 16);
std::memcpy(dst, tmp, 16);
}
template <
bool use_16bytes_chunk>
inline void ConditionalUnalignedCopy128(
const char* src,
char* dst) {
if (use_16bytes_chunk) {
UnalignedCopy128(src, dst);
}
else {
UnalignedCopy64(src, dst);
UnalignedCopy64(src + 8, dst + 8);
}
}
// Copy [src, src+(op_limit-op)) to [op, (op_limit-op)) a byte at a time. Used
// for handling COPY operations where the input and output regions may overlap.
// For example, suppose:
// src == "ab"
// op == src + 2
// op_limit == op + 20
// After IncrementalCopySlow(src, op, op_limit), the result will have eleven
// copies of "ab"
// ababababababababababab
// Note that this does not match the semantics of either std::memcpy() or
// std::memmove().
inline char* IncrementalCopySlow(
const char* src,
char* op,
char*
const op_limit) {
// TODO: Remove pragma when LLVM is aware this
// function is only called in cold regions and when cold regions don't get
// vectorized or unrolled.
#ifdef __clang__
#pragma clang loop unroll(disable)
#endif
while (op < op_limit) {
*op++ = *src++;
}
return op_limit;
}
#if SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE
// Computes the bytes for shuffle control mask (please read comments on
// 'pattern_generation_masks' as well) for the given index_offset and
// pattern_size. For example, when the 'offset' is 6, it will generate a
// repeating pattern of size 6. So, the first 16 byte indexes will correspond to
// the pattern-bytes {0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3} and the
// next 16 byte indexes will correspond to the pattern-bytes {4, 5, 0, 1, 2, 3,
// 4, 5, 0, 1, 2, 3, 4, 5, 0, 1}. These byte index sequences are generated by
// calling MakePatternMaskBytes(0, 6, index_sequence<16>()) and
// MakePatternMaskBytes(16, 6, index_sequence<16>()) respectively.
template <size_t... indexes>
inline constexpr std::array<
char,
sizeof...(indexes)> MakePatternMaskBytes(
int index_offset,
int pattern_size, index_sequence<indexes...>) {
return {
static_cast<
char>((index_offset + indexes) % pattern_size)...};
}
// Computes the shuffle control mask bytes array for given pattern-sizes and
// returns an array.
template <size_t... pattern_sizes_minus_one>
inline constexpr std::array<std::array<
char,
sizeof(V128)>,
sizeof...(pattern_sizes_minus_one)>
MakePatternMaskBytesTable(
int index_offset,
index_sequence<pattern_sizes_minus_one...>) {
return {
MakePatternMaskBytes(index_offset, pattern_sizes_minus_one + 1,
make_index_sequence<
/*indexes=*/sizeof(V128)>())...};
}
// This is an array of shuffle control masks that can be used as the source
// operand for PSHUFB to permute the contents of the destination XMM register
// into a repeating byte pattern.
alignas(16) constexpr std::array<std::array<
char,
sizeof(V128)>,
16> pattern_generation_masks =
MakePatternMaskBytesTable(
/*index_offset=*/0,
/*pattern_sizes_minus_one=*/make_index_sequence<16>());
// Similar to 'pattern_generation_masks', this table is used to "rotate" the
// pattern so that we can copy the *next 16 bytes* consistent with the pattern.
// Basically, pattern_reshuffle_masks is a continuation of
// pattern_generation_masks. It follows that, pattern_reshuffle_masks is same as
// pattern_generation_masks for offsets 1, 2, 4, 8 and 16.
alignas(16) constexpr std::array<std::array<
char,
sizeof(V128)>,
16> pattern_reshuffle_masks =
MakePatternMaskBytesTable(
/*index_offset=*/16,
/*pattern_sizes_minus_one=*/make_index_sequence<16>());
SNAPPY_ATTRIBUTE_ALWAYS_INLINE
static inline V128 LoadPattern(
const char* src,
const size_t pattern_size) {
V128 generation_mask = V128_Load(
reinterpret_cast<
const V128*>(
pattern_generation_masks[pattern_size - 1].data()));
// Uninitialized bytes are masked out by the shuffle mask.
// TODO: remove annotation and macro defs once MSan is fixed.
SNAPPY_ANNOTATE_MEMORY_IS_INITIALIZED(src + pattern_size, 16 - pattern_size);
return V128_Shuffle(V128_LoadU(
reinterpret_cast<
const V128*>(src)),
generation_mask);
}
SNAPPY_ATTRIBUTE_ALWAYS_INLINE
static inline std::pair<V128
/* pattern */, V128 /* reshuffle_mask */>
LoadPatternAndReshuffleMask(
const char* src,
const size_t pattern_size) {
V128 pattern = LoadPattern(src, pattern_size);
// This mask will generate the next 16 bytes in-place. Doing so enables us to
// write data by at most 4 V128_StoreU.
//
// For example, suppose pattern is: abcdefabcdefabcd
// Shuffling with this mask will generate: efabcdefabcdefab
// Shuffling again will generate: cdefabcdefabcdef
V128 reshuffle_mask = V128_Load(
reinterpret_cast<
const V128*>(
pattern_reshuffle_masks[pattern_size - 1].data()));
return {pattern, reshuffle_mask};
}
#endif // SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE
// Fallback for when we need to copy while extending the pattern, for example
// copying 10 bytes from 3 positions back abc -> abcabcabcabca.
//
// REQUIRES: [dst - offset, dst + 64) is a valid address range.
SNAPPY_ATTRIBUTE_ALWAYS_INLINE
static inline bool Copy64BytesWithPatternExtension(
char* dst, size_t offset) {
#if SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE
if (SNAPPY_PREDICT_TRUE(offset <= 16)) {
switch (offset) {
case 0:
return false;
case 1: {
// TODO: Ideally we should memset, move back once the
// codegen issues are fixed.
V128 pattern = V128_DupChar(dst[-1]);
for (
int i = 0; i < 4; i++) {
V128_StoreU(
reinterpret_cast<V128*>(dst + 16 * i), pattern);
}
return true;
}
case 2:
case 4:
case 8:
case 16: {
V128 pattern = LoadPattern(dst - offset, offset);
for (
int i = 0; i < 4; i++) {
V128_StoreU(
reinterpret_cast<V128*>(dst + 16 * i), pattern);
}
return true;
}
default: {
auto pattern_and_reshuffle_mask =
LoadPatternAndReshuffleMask(dst - offset, offset);
V128 pattern = pattern_and_reshuffle_mask.first;
V128 reshuffle_mask = pattern_and_reshuffle_mask.second;
for (
int i = 0; i < 4; i++) {
V128_StoreU(
reinterpret_cast<V128*>(dst + 16 * i), pattern);
pattern = V128_Shuffle(pattern, reshuffle_mask);
}
return true;
}
}
}
#else
if (SNAPPY_PREDICT_TRUE(offset < 16)) {
if (SNAPPY_PREDICT_FALSE(offset == 0))
return false;
// Extend the pattern to the first 16 bytes.
// The simpler formulation of `dst[i - offset]` induces undefined behavior.
for (
int i = 0; i < 16; i++) dst[i] = (dst - offset)[i];
// Find a multiple of pattern >= 16.
static std::array<uint8_t, 16> pattern_sizes = []() {
std::array<uint8_t, 16> res;
for (
int i = 1; i < 16; i++) res[i] = (16 / i + 1) * i;
return res;
}();
offset = pattern_sizes[offset];
for (
int i = 1; i < 4; i++) {
std::memcpy(dst + i * 16, dst + i * 16 - offset, 16);
}
return true;
}
#endif // SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE
// Very rare.
for (
int i = 0; i < 4; i++) {
std::memcpy(dst + i * 16, dst + i * 16 - offset, 16);
}
return true;
}
// Copy [src, src+(op_limit-op)) to [op, op_limit) but faster than
// IncrementalCopySlow. buf_limit is the address past the end of the writable
// region of the buffer.
inline char* IncrementalCopy(
const char* src,
char* op,
char*
const op_limit,
char*
const buf_limit) {
#if SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE
constexpr
int big_pattern_size_lower_bound = 16;
#else
constexpr
int big_pattern_size_lower_bound = 8;
#endif
// Terminology:
//
// slop = buf_limit - op
// pat = op - src
// len = op_limit - op
assert(src < op);
assert(op < op_limit);
assert(op_limit <= buf_limit);
// NOTE: The copy tags use 3 or 6 bits to store the copy length, so len <= 64.
assert(op_limit - op <= 64);
// NOTE: In practice the compressor always emits len >= 4, so it is ok to
// assume that to optimize this function, but this is not guaranteed by the
// compression format, so we have to also handle len < 4 in case the input
// does not satisfy these conditions.
size_t pattern_size = op - src;
// The cases are split into different branches to allow the branch predictor,
// FDO, and static prediction hints to work better. For each input we list the
// ratio of invocations that match each condition.
//
// input slop < 16 pat < 8 len > 16
// ------------------------------------------
// html|html4|cp 0% 1.01% 27.73%
// urls 0% 0.88% 14.79%
// jpg 0% 64.29% 7.14%
// pdf 0% 2.56% 58.06%
// txt[1-4] 0% 0.23% 0.97%
// pb 0% 0.96% 13.88%
// bin 0.01% 22.27% 41.17%
//
// It is very rare that we don't have enough slop for doing block copies. It
// is also rare that we need to expand a pattern. Small patterns are common
// for incompressible formats and for those we are plenty fast already.
// Lengths are normally not greater than 16 but they vary depending on the
// input. In general if we always predict len <= 16 it would be an ok
// prediction.
//
// In order to be fast we want a pattern >= 16 bytes (or 8 bytes in non-SSE)
// and an unrolled loop copying 1x 16 bytes (or 2x 8 bytes in non-SSE) at a
// time.
// Handle the uncommon case where pattern is less than 16 (or 8 in non-SSE)
// bytes.
if (pattern_size < big_pattern_size_lower_bound) {
#if SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE
// Load the first eight bytes into an 128-bit XMM register, then use PSHUFB
// to permute the register's contents in-place into a repeating sequence of
// the first "pattern_size" bytes.
// For example, suppose:
// src == "abc"
// op == op + 3
// After V128_Shuffle(), "pattern" will have five copies of "abc"
// followed by one byte of slop: abcabcabcabcabca.
//
// The non-SSE fallback implementation suffers from store-forwarding stalls
// because its loads and stores partly overlap. By expanding the pattern
// in-place, we avoid the penalty.
// Typically, the op_limit is the gating factor so try to simplify the loop
// based on that.
if (SNAPPY_PREDICT_TRUE(op_limit <= buf_limit - 15)) {
auto pattern_and_reshuffle_mask =
LoadPatternAndReshuffleMask(src, pattern_size);
V128 pattern = pattern_and_reshuffle_mask.first;
V128 reshuffle_mask = pattern_and_reshuffle_mask.second;
// There is at least one, and at most four 16-byte blocks. Writing four
// conditionals instead of a loop allows FDO to layout the code with
// respect to the actual probabilities of each length.
// TODO: Replace with loop with trip count hint.
V128_StoreU(
reinterpret_cast<V128*>(op), pattern);
if (op + 16 < op_limit) {
pattern = V128_Shuffle(pattern, reshuffle_mask);
V128_StoreU(
reinterpret_cast<V128*>(op + 16), pattern);
}
if (op + 32 < op_limit) {
pattern = V128_Shuffle(pattern, reshuffle_mask);
V128_StoreU(
reinterpret_cast<V128*>(op + 32), pattern);
}
if (op + 48 < op_limit) {
pattern = V128_Shuffle(pattern, reshuffle_mask);
V128_StoreU(
reinterpret_cast<V128*>(op + 48), pattern);
}
return op_limit;
}
char*
const op_end = buf_limit - 15;
if (SNAPPY_PREDICT_TRUE(op < op_end)) {
auto pattern_and_reshuffle_mask =
LoadPatternAndReshuffleMask(src, pattern_size);
V128 pattern = pattern_and_reshuffle_mask.first;
V128 reshuffle_mask = pattern_and_reshuffle_mask.second;
// This code path is relatively cold however so we save code size
// by avoiding unrolling and vectorizing.
//
// TODO: Remove pragma when when cold regions don't get
// vectorized or unrolled.
#ifdef __clang__
#pragma clang loop unroll(disable)
#endif
do {
V128_StoreU(
reinterpret_cast<V128*>(op), pattern);
pattern = V128_Shuffle(pattern, reshuffle_mask);
op += 16;
}
while (SNAPPY_PREDICT_TRUE(op < op_end));
}
return IncrementalCopySlow(op - pattern_size, op, op_limit);
#else // !SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE
// If plenty of buffer space remains, expand the pattern to at least 8
// bytes. The way the following loop is written, we need 8 bytes of buffer
// space if pattern_size >= 4, 11 bytes if pattern_size is 1 or 3, and 10
// bytes if pattern_size is 2. Precisely encoding that is probably not
// worthwhile; instead, invoke the slow path if we cannot write 11 bytes
// (because 11 are required in the worst case).
if (SNAPPY_PREDICT_TRUE(op <= buf_limit - 11)) {
while (pattern_size < 8) {
UnalignedCopy64(src, op);
op += pattern_size;
pattern_size *= 2;
}
if (SNAPPY_PREDICT_TRUE(op >= op_limit))
return op_limit;
}
else {
return IncrementalCopySlow(src, op, op_limit);
}
#endif // SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE
}
assert(pattern_size >= big_pattern_size_lower_bound);
constexpr
bool use_16bytes_chunk = big_pattern_size_lower_bound == 16;
// Copy 1x 16 bytes (or 2x 8 bytes in non-SSE) at a time. Because op - src can
// be < 16 in non-SSE, a single UnalignedCopy128 might overwrite data in op.
// UnalignedCopy64 is safe because expanding the pattern to at least 8 bytes
// guarantees that op - src >= 8.
//
// Typically, the op_limit is the gating factor so try to simplify the loop
// based on that.
if (SNAPPY_PREDICT_TRUE(op_limit <= buf_limit - 15)) {
// There is at least one, and at most four 16-byte blocks. Writing four
// conditionals instead of a loop allows FDO to layout the code with respect
// to the actual probabilities of each length.
// TODO: Replace with loop with trip count hint.
ConditionalUnalignedCopy128<use_16bytes_chunk>(src, op);
if (op + 16 < op_limit) {
ConditionalUnalignedCopy128<use_16bytes_chunk>(src + 16, op + 16);
}
if (op + 32 < op_limit) {
ConditionalUnalignedCopy128<use_16bytes_chunk>(src + 32, op + 32);
}
if (op + 48 < op_limit) {
ConditionalUnalignedCopy128<use_16bytes_chunk>(src + 48, op + 48);
}
return op_limit;
}
// Fall back to doing as much as we can with the available slop in the
// buffer. This code path is relatively cold however so we save code size by
// avoiding unrolling and vectorizing.
//
// TODO: Remove pragma when when cold regions don't get vectorized
// or unrolled.
#ifdef __clang__
#pragma clang loop unroll(disable)
#endif
for (
char* op_end = buf_limit - 16; op < op_end; op += 16, src += 16) {
ConditionalUnalignedCopy128<use_16bytes_chunk>(src, op);
}
if (op >= op_limit)
return op_limit;
// We only take this branch if we didn't have enough slop and we can do a
// single 8 byte copy.
if (SNAPPY_PREDICT_FALSE(op <= buf_limit - 8)) {
UnalignedCopy64(src, op);
src += 8;
op += 8;
}
return IncrementalCopySlow(src, op, op_limit);
}
}
// namespace
template <
bool allow_fast_path>
static inline char* EmitLiteral(
char* op,
const char* literal,
int len) {
// The vast majority of copies are below 16 bytes, for which a
// call to std::memcpy() is overkill. This fast path can sometimes
// copy up to 15 bytes too much, but that is okay in the
// main loop, since we have a bit to go on for both sides:
//
// - The input will always have kInputMarginBytes = 15 extra
// available bytes, as long as we're in the main loop, and
// if not, allow_fast_path = false.
// - The output will always have 32 spare bytes (see
// MaxCompressedLength).
assert(len > 0);
// Zero-length literals are disallowed
int n = len - 1;
if (allow_fast_path && len <= 16) {
// Fits in tag byte
*op++ = LITERAL | (n << 2);
UnalignedCopy128(literal, op);
return op + len;
}
if (n < 60) {
// Fits in tag byte
*op++ = LITERAL | (n << 2);
}
else {
int count = (Bits::Log2Floor(n) >> 3) + 1;
assert(count >= 1);
assert(count <= 4);
*op++ = LITERAL | ((59 + count) << 2);
// Encode in upcoming bytes.
// Write 4 bytes, though we may care about only 1 of them. The output buffer
// is guaranteed to have at least 3 more spaces left as 'len >= 61' holds
// here and there is a std::memcpy() of size 'len' below.
LittleEndian::Store32(op, n);
op += count;
}
// When allow_fast_path is true, we can overwrite up to 16 bytes.
if (allow_fast_path) {
char* destination = op;
const char* source = literal;
const char* end = destination + len;
do {
std::memcpy(destination, source, 16);
destination += 16;
source += 16;
}
while (destination < end);
}
else {
std::memcpy(op, literal, len);
}
return op + len;
}
template <
bool len_less_than_12>
static inline char* EmitCopyAtMost64(
char* op, size_t offset, size_t len) {
assert(len <= 64);
assert(len >= 4);
assert(offset < 65536);
assert(len_less_than_12 == (len < 12));
if (len_less_than_12) {
uint32_t u = (len << 2) + (offset << 8);
uint32_t copy1 = COPY_1_BYTE_OFFSET - (4 << 2) + ((offset >> 3) & 0xe0);
uint32_t copy2 = COPY_2_BYTE_OFFSET - (1 << 2);
// It turns out that offset < 2048 is a difficult to predict branch.
// `perf record` shows this is the highest percentage of branch misses in
// benchmarks. This code produces branch free code, the data dependency
// chain that bottlenecks the throughput is so long that a few extra
// instructions are completely free (IPC << 6 because of data deps).
u += offset < 2048 ? copy1 : copy2;
LittleEndian::Store32(op, u);
op += offset < 2048 ? 2 : 3;
}
else {
// Write 4 bytes, though we only care about 3 of them. The output buffer
// is required to have some slack, so the extra byte won't overrun it.
uint32_t u = COPY_2_BYTE_OFFSET + ((len - 1) << 2) + (offset << 8);
LittleEndian::Store32(op, u);
op += 3;
}
return op;
}
template <
bool len_less_than_12>
static inline char* EmitCopy(
char* op, size_t offset, size_t len) {
assert(len_less_than_12 == (len < 12));
if (len_less_than_12) {
return EmitCopyAtMost64<
/*len_less_than_12=*/true>(op, offset, len);
}
else {
// A special case for len <= 64 might help, but so far measurements suggest
// it's in the noise.
// Emit 64 byte copies but make sure to keep at least four bytes reserved.
while (SNAPPY_PREDICT_FALSE(len >= 68)) {
op = EmitCopyAtMost64<
/*len_less_than_12=*/false>(op, offset, 64);
len -= 64;
}
// One or two copies will now finish the job.
if (len > 64) {
op = EmitCopyAtMost64<
/*len_less_than_12=*/false>(op, offset, 60);
len -= 60;
}
// Emit remainder.
if (len < 12) {
op = EmitCopyAtMost64<
/*len_less_than_12=*/true>(op, offset, len);
}
else {
op = EmitCopyAtMost64<
/*len_less_than_12=*/false>(op, offset, len);
}
return op;
}
}
bool GetUncompressedLength(
const char* start, size_t n, size_t* result) {
uint32_t v = 0;
const char* limit = start + n;
if (Varint::Parse32WithLimit(start, limit, &v) != NULL) {
*result = v;
return true;
}
else {
return false;
}
}
namespace {
uint32_t CalculateTableSize(uint32_t input_size) {
static_assert(
kMaxHashTableSize >= kMinHashTableSize,
"kMaxHashTableSize should be greater or equal to kMinHashTableSize.");
if (input_size > kMaxHashTableSize) {
return kMaxHashTableSize;
}
if (input_size < kMinHashTableSize) {
return kMinHashTableSize;
}
// This is equivalent to Log2Ceiling(input_size), assuming input_size > 1.
// 2 << Log2Floor(x - 1) is equivalent to 1 << (1 + Log2Floor(x - 1)).
return 2u << Bits::Log2Floor(input_size - 1);
}
}
// namespace
namespace internal {
WorkingMemory::WorkingMemory(size_t input_size) {
const size_t max_fragment_size = std::min(input_size, kBlockSize);
const size_t table_size = CalculateTableSize(max_fragment_size);
size_ = table_size *
sizeof(*table_) + max_fragment_size +
MaxCompressedLength(max_fragment_size);
mem_ = std::allocator<
char>().allocate(size_);
table_ =
reinterpret_cast<uint16_t*>(mem_);
input_ = mem_ + table_size *
sizeof(*table_);
output_ = input_ + max_fragment_size;
}
WorkingMemory::~WorkingMemory() {
std::allocator<
char>().deallocate(mem_, size_);
}
uint16_t* WorkingMemory::GetHashTable(size_t fragment_size,
int* table_size)
const {
const size_t htsize = CalculateTableSize(fragment_size);
memset(table_, 0, htsize *
sizeof(*table_));
*table_size = htsize;
return table_;
}
}
// end namespace internal
// Flat array compression that does not emit the "uncompressed length"
// prefix. Compresses "input" string to the "*op" buffer.
//
// REQUIRES: "input" is at most "kBlockSize" bytes long.
// REQUIRES: "op" points to an array of memory that is at least
// "MaxCompressedLength(input.size())" in size.
// REQUIRES: All elements in "table[0..table_size-1]" are initialized to zero.
// REQUIRES: "table_size" is a power of two
//
// Returns an "end" pointer into "op" buffer.
// "end - op" is the compressed size of "input".
namespace internal {
char* CompressFragment(
const char* input, size_t input_size,
char* op,
uint16_t* table,
const int table_size) {
// "ip" is the input pointer, and "op" is the output pointer.
const char* ip = input;
assert(input_size <= kBlockSize);
assert((table_size & (table_size - 1)) == 0);
// table must be power of two
const uint32_t mask = 2 * (table_size - 1);
const char* ip_end = input + input_size;
const char* base_ip = ip;
const size_t kInputMarginBytes = 15;
if (SNAPPY_PREDICT_TRUE(input_size >= kInputMarginBytes)) {
const char* ip_limit = input + input_size - kInputMarginBytes;
for (uint32_t preload = LittleEndian::Load32(ip + 1);;) {
// Bytes in [next_emit, ip) will be emitted as literal bytes. Or
// [next_emit, ip_end) after the main loop.
const char* next_emit = ip++;
uint64_t data = LittleEndian::Load64(ip);
// The body of this loop calls EmitLiteral once and then EmitCopy one or
// more times. (The exception is that when we're close to exhausting
// the input we goto emit_remainder.)
//
// In the first iteration of this loop we're just starting, so
// there's nothing to copy, so calling EmitLiteral once is
// necessary. And we only start a new iteration when the
// current iteration has determined that a call to EmitLiteral will
// precede the next call to EmitCopy (if any).
//
// Step 1: Scan forward in the input looking for a 4-byte-long match.
// If we get close to exhausting the input then goto emit_remainder.
//
// Heuristic match skipping: If 32 bytes are scanned with no matches
// found, start looking only at every other byte. If 32 more bytes are
// scanned (or skipped), look at every third byte, etc.. When a match is
// found, immediately go back to looking at every byte. This is a small
// loss (~5% performance, ~0.1% density) for compressible data due to more
// bookkeeping, but for non-compressible data (such as JPEG) it's a huge
// win since the compressor quickly "realizes" the data is incompressible
// and doesn't bother looking for matches everywhere.
//
// The "skip" variable keeps track of how many bytes there are since the
// last match; dividing it by 32 (ie. right-shifting by five) gives the
// number of bytes to move ahead for each iteration.
uint32_t skip = 32;
const char* candidate;
if (ip_limit - ip >= 16) {
auto delta = ip - base_ip;
for (
int j = 0; j < 4; ++j) {
for (
int k = 0; k < 4; ++k) {
int i = 4 * j + k;
// These for-loops are meant to be unrolled. So we can freely
// special case the first iteration to use the value already
// loaded in preload.
uint32_t dword = i == 0 ? preload :
static_cast<uint32_t>(data);
assert(dword == LittleEndian::Load32(ip + i));
uint16_t* table_entry = TableEntry(table, dword, mask);
candidate = base_ip + *table_entry;
assert(candidate >= base_ip);
assert(candidate < ip + i);
*table_entry = delta + i;
if (SNAPPY_PREDICT_FALSE(LittleEndian::Load32(candidate) == dword)) {
*op = LITERAL | (i << 2);
UnalignedCopy128(next_emit, op + 1);
ip += i;
op = op + i + 2;
goto emit_match;
}
data >>= 8;
}
data = LittleEndian::Load64(ip + 4 * j + 4);
}
ip += 16;
skip += 16;
}
while (
true) {
assert(
static_cast<uint32_t>(data) == LittleEndian::Load32(ip));
uint16_t* table_entry = TableEntry(table, data, mask);
uint32_t bytes_between_hash_lookups = skip >> 5;
skip += bytes_between_hash_lookups;
const char* next_ip = ip + bytes_between_hash_lookups;
if (SNAPPY_PREDICT_FALSE(next_ip > ip_limit)) {
ip = next_emit;
goto emit_remainder;
}
candidate = base_ip + *table_entry;
assert(candidate >= base_ip);
assert(candidate < ip);
*table_entry = ip - base_ip;
if (SNAPPY_PREDICT_FALSE(
static_cast<uint32_t>(data) ==
LittleEndian::Load32(candidate))) {
break;
}
data = LittleEndian::Load32(next_ip);
ip = next_ip;
}
// Step 2: A 4-byte match has been found. We'll later see if more
// than 4 bytes match. But, prior to the match, input
// bytes [next_emit, ip) are unmatched. Emit them as "literal bytes."
assert(next_emit + 16 <= ip_end);
op = EmitLiteral<
/*allow_fast_path=*/true>(op, next_emit, ip - next_emit);
// Step 3: Call EmitCopy, and then see if another EmitCopy could
// be our next move. Repeat until we find no match for the
// input immediately after what was consumed by the last EmitCopy call.
//
// If we exit this loop normally then we need to call EmitLiteral next,
// though we don't yet know how big the literal will be. We handle that
// by proceeding to the next iteration of the main loop. We also can exit
// this loop via goto if we get close to exhausting the input.
emit_match:
do {
// We have a 4-byte match at ip, and no need to emit any
// "literal bytes" prior to ip.
const char* base = ip;
std::pair<size_t,
bool> p =
FindMatchLength(candidate + 4, ip + 4, ip_end, &data);
size_t matched = 4 + p.first;
ip += matched;
size_t offset = base - candidate;
assert(0 == memcmp(base, candidate, matched));
if (p.second) {
op = EmitCopy<
/*len_less_than_12=*/true>(op, offset, matched);
}
else {
op = EmitCopy<
/*len_less_than_12=*/false>(op, offset, matched);
}
if (SNAPPY_PREDICT_FALSE(ip >= ip_limit)) {
goto emit_remainder;
}
// Expect 5 bytes to match
assert((data & 0xFFFFFFFFFF) ==
(LittleEndian::Load64(ip) & 0xFFFFFFFFFF));
// We are now looking for a 4-byte match again. We read
// table[Hash(ip, mask)] for that. To improve compression,
// we also update table[Hash(ip - 1, mask)] and table[Hash(ip, mask)].
*TableEntry(table, LittleEndian::Load32(ip - 1), mask) =
ip - base_ip - 1;
uint16_t* table_entry = TableEntry(table, data, mask);
candidate = base_ip + *table_entry;
*table_entry = ip - base_ip;
// Measurements on the benchmarks have shown the following probabilities
// for the loop to exit (ie. avg. number of iterations is reciprocal).
// BM_Flat/6 txt1 p = 0.3-0.4
// BM_Flat/7 txt2 p = 0.35
// BM_Flat/8 txt3 p = 0.3-0.4
// BM_Flat/9 txt3 p = 0.34-0.4
// BM_Flat/10 pb p = 0.4
// BM_Flat/11 gaviota p = 0.1
// BM_Flat/12 cp p = 0.5
// BM_Flat/13 c p = 0.3
}
while (
static_cast<uint32_t>(data) == LittleEndian::Load32(candidate));
// Because the least significant 5 bytes matched, we can utilize data
// for the next iteration.
preload = data >> 8;
}
}
emit_remainder:
// Emit the remaining bytes as a literal
if (ip < ip_end) {
op = EmitLiteral<
/*allow_fast_path=*/false>(op, ip, ip_end - ip);
}
return op;
}
char* CompressFragmentDoubleHash(
const char* input, size_t input_size,
char* op,
uint16_t* table,
const int table_size,
uint16_t* table2,
const int table_size2) {
(
void)table_size2;
assert(table_size == table_size2);
// "ip" is the input pointer, and "op" is the output pointer.
const char* ip = input;
assert(input_size <= kBlockSize);
assert((table_size & (table_size - 1)) == 0);
// table must be power of two
const uint32_t mask = 2 * (table_size - 1);
const char* ip_end = input + input_size;
const char* base_ip = ip;
const size_t kInputMarginBytes = 15;
if (SNAPPY_PREDICT_TRUE(input_size >= kInputMarginBytes)) {
const char* ip_limit = input + input_size - kInputMarginBytes;
for (;;) {
const char* next_emit = ip++;
uint64_t data = LittleEndian::Load64(ip);
uint32_t skip = 512;
const char* candidate;
uint32_t candidate_length;
while (
true) {
assert(
static_cast<uint32_t>(data) == LittleEndian::Load32(ip));
uint16_t* table_entry2 = TableEntry8ByteMatch(table2, data, mask);
uint32_t bytes_between_hash_lookups = skip >> 9;
skip++;
const char* next_ip = ip + bytes_between_hash_lookups;
if (SNAPPY_PREDICT_FALSE(next_ip > ip_limit)) {
ip = next_emit;
goto emit_remainder;
}
candidate = base_ip + *table_entry2;
assert(candidate >= base_ip);
assert(candidate < ip);
*table_entry2 = ip - base_ip;
if (SNAPPY_PREDICT_FALSE(
static_cast<uint32_t>(data) ==
LittleEndian::Load32(candidate))) {
candidate_length =
FindMatchLengthPlain(candidate + 4, ip + 4, ip_end) + 4;
break;
}
uint16_t* table_entry = TableEntry4ByteMatch(table, data, mask);
candidate = base_ip + *table_entry;
assert(candidate >= base_ip);
assert(candidate < ip);
*table_entry = ip - base_ip;
if (SNAPPY_PREDICT_FALSE(
static_cast<uint32_t>(data) ==
LittleEndian::Load32(candidate))) {
candidate_length =
FindMatchLengthPlain(candidate + 4, ip + 4, ip_end) + 4;
table_entry2 =
TableEntry8ByteMatch(table2, LittleEndian::Load64(ip + 1), mask);
auto candidate2 = base_ip + *table_entry2;
size_t candidate_length2 =
FindMatchLengthPlain(candidate2, ip + 1, ip_end);
if (candidate_length2 > candidate_length) {
*table_entry2 = ip - base_ip;
candidate = candidate2;
candidate_length = candidate_length2;
++ip;
}
break;
}
data = LittleEndian::Load64(next_ip);
ip = next_ip;
}
// Backtrack to the point it matches fully.
while (ip > next_emit && candidate > base_ip &&
*(ip - 1) == *(candidate - 1)) {
--ip;
--candidate;
++candidate_length;
}
*TableEntry8ByteMatch(table2, LittleEndian::Load64(ip + 1), mask) =
ip - base_ip + 1;
*TableEntry8ByteMatch(table2, LittleEndian::Load64(ip + 2), mask) =
ip - base_ip + 2;
*TableEntry4ByteMatch(table, LittleEndian::Load32(ip + 1), mask) =
ip - base_ip + 1;
// Step 2: A 4-byte or 8-byte match has been found.
// We'll later see if more than 4 bytes match. But, prior to the match,
// input bytes [next_emit, ip) are unmatched. Emit them as
// "literal bytes."
assert(next_emit + 16 <= ip_end);
if (ip - next_emit > 0) {
op = EmitLiteral<
/*allow_fast_path=*/true>(op, next_emit,
ip - next_emit);
}
// Step 3: Call EmitCopy, and then see if another EmitCopy could
// be our next move. Repeat until we find no match for the
// input immediately after what was consumed by the last EmitCopy call.
//
// If we exit this loop normally then we need to call EmitLiteral next,
// though we don't yet know how big the literal will be. We handle that
// by proceeding to the next iteration of the main loop. We also can exit
// this loop via goto if we get close to exhausting the input.
do {
// We have a 4-byte match at ip, and no need to emit any
// "literal bytes" prior to ip.
const char* base = ip;
ip += candidate_length;
size_t offset = base - candidate;
if (candidate_length < 12) {
op =
EmitCopy<
/*len_less_than_12=*/true>(op, offset, candidate_length);
}
else {
op = EmitCopy<
/*len_less_than_12=*/false>(op, offset,
candidate_length);
}
if (SNAPPY_PREDICT_FALSE(ip >= ip_limit)) {
goto emit_remainder;
}
// We are now looking for a 4-byte match again. We read
// table[Hash(ip, mask)] for that. To improve compression,
// we also update several previous table entries.
if (ip - base_ip > 7) {
*TableEntry8ByteMatch(table2, LittleEndian::Load64(ip - 7), mask) =
ip - base_ip - 7;
*TableEntry8ByteMatch(table2, LittleEndian::Load64(ip - 4), mask) =
ip - base_ip - 4;
}
*TableEntry8ByteMatch(table2, LittleEndian::Load64(ip - 3), mask) =
ip - base_ip - 3;
*TableEntry8ByteMatch(table2, LittleEndian::Load64(ip - 2), mask) =
ip - base_ip - 2;
*TableEntry4ByteMatch(table, LittleEndian::Load32(ip - 2), mask) =
ip - base_ip - 2;
*TableEntry4ByteMatch(table, LittleEndian::Load32(ip - 1), mask) =
ip - base_ip - 1;
uint16_t* table_entry =
TableEntry8ByteMatch(table2, LittleEndian::Load64(ip), mask);
candidate = base_ip + *table_entry;
*table_entry = ip - base_ip;
if (LittleEndian::Load32(ip) == LittleEndian::Load32(candidate)) {
candidate_length =
FindMatchLengthPlain(candidate + 4, ip + 4, ip_end) + 4;
continue;
}
table_entry =
TableEntry4ByteMatch(table, LittleEndian::Load32(ip), mask);
candidate = base_ip + *table_entry;
*table_entry = ip - base_ip;
if (LittleEndian::Load32(ip) == LittleEndian::Load32(candidate)) {
candidate_length =
FindMatchLengthPlain(candidate + 4, ip + 4, ip_end) + 4;
continue;
}
break;
}
while (
true);
}
}
emit_remainder:
// Emit the remaining bytes as a literal
if (ip < ip_end) {
op = EmitLiteral<
/*allow_fast_path=*/false>(op, ip, ip_end - ip);
}
return op;
}
}
// end namespace internal
static inline void Report(
int token,
const char *algorithm, size_t
compressed_size, size_t uncompressed_size) {
// TODO: Switch to [[maybe_unused]] when we can assume C++17.
(
void)token;
(
void)algorithm;
(
void)compressed_size;
(
void)uncompressed_size;
}
// Signature of output types needed by decompression code.
// The decompression code is templatized on a type that obeys this
// signature so that we do not pay virtual function call overhead in
// the middle of a tight decompression loop.
//
// class DecompressionWriter {
// public:
// // Called before decompression
// void SetExpectedLength(size_t length);
//
// // For performance a writer may choose to donate the cursor variable to the
// // decompression function. The decompression will inject it in all its
// // function calls to the writer. Keeping the important output cursor as a
// // function local stack variable allows the compiler to keep it in
// // register, which greatly aids performance by avoiding loads and stores of
// // this variable in the fast path loop iterations.
// T GetOutputPtr() const;
//
// // At end of decompression the loop donates the ownership of the cursor
// // variable back to the writer by calling this function.
// void SetOutputPtr(T op);
//
// // Called after decompression
// bool CheckLength() const;
//
// // Called repeatedly during decompression
// // Each function get a pointer to the op (output pointer), that the writer
// // can use and update. Note it's important that these functions get fully
// // inlined so that no actual address of the local variable needs to be
// // taken.
// bool Append(const char* ip, size_t length, T* op);
// bool AppendFromSelf(uint32_t offset, size_t length, T* op);
//
// // The rules for how TryFastAppend differs from Append are somewhat
// // convoluted:
// //
// // - TryFastAppend is allowed to decline (return false) at any
// // time, for any reason -- just "return false" would be
// // a perfectly legal implementation of TryFastAppend.
// // The intention is for TryFastAppend to allow a fast path
// // in the common case of a small append.
// // - TryFastAppend is allowed to read up to <available> bytes
// // from the input buffer, whereas Append is allowed to read
// // <length>. However, if it returns true, it must leave
// // at least five (kMaximumTagLength) bytes in the input buffer
// // afterwards, so that there is always enough space to read the
// // next tag without checking for a refill.
// // - TryFastAppend must always return decline (return false)
// // if <length> is 61 or more, as in this case the literal length is not
// // decoded fully. In practice, this should not be a big problem,
// // as it is unlikely that one would implement a fast path accepting
// // this much data.
// //
// bool TryFastAppend(const char* ip, size_t available, size_t length, T* op);
// };
static inline uint32_t ExtractLowBytes(
const uint32_t& v,
int n) {
assert(n >= 0);
assert(n <= 4);
#if SNAPPY_HAVE_BMI2
return _bzhi_u32(v, 8 * n);
#else
// This needs to be wider than uint32_t otherwise `mask << 32` will be
// undefined.
uint64_t mask = 0xffffffff;
return v & ~(mask << (8 * n));
#endif
}
static inline bool LeftShiftOverflows(uint8_t value, uint32_t shift) {
assert(shift < 32);
static const uint8_t masks[] = {
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
//
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
//
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
//
0x00, 0x80, 0xc0, 0xe0, 0xf0, 0xf8, 0xfc, 0xfe};
return (value & masks[shift]) != 0;
}
inline bool Copy64BytesWithPatternExtension(ptrdiff_t dst, size_t offset) {
// TODO: Switch to [[maybe_unused]] when we can assume C++17.
(
void)dst;
return offset != 0;
}
// Copies between size bytes and 64 bytes from src to dest. size cannot exceed
// 64. More than size bytes, but never exceeding 64, might be copied if doing
// so gives better performance. [src, src + size) must not overlap with
// [dst, dst + size), but [src, src + 64) may overlap with [dst, dst + 64).
void MemCopy64(
char* dst,
const void* src, size_t size) {
// Always copy this many bytes. If that's below size then copy the full 64.
constexpr
int kShortMemCopy = 32;
assert(size <= 64);
assert(std::less_equal<
const void*>()(
static_cast<
const char*>(src) + size,
dst) ||
std::less_equal<
const void*>()(dst + size, src));
// We know that src and dst are at least size bytes apart. However, because we
// might copy more than size bytes the copy still might overlap past size.
// E.g. if src and dst appear consecutively in memory (src + size >= dst).
// TODO: Investigate wider copies on other platforms.
#if defined(__x86_64__) &&
defined(__AVX__)
assert(kShortMemCopy <= 32);
__m256i data = _mm256_lddqu_si256(
static_cast<
const __m256i *>(src));
_mm256_storeu_si256(
reinterpret_cast<__m256i *>(dst), data);
// Profiling shows that nearly all copies are short.
if (SNAPPY_PREDICT_FALSE(size > kShortMemCopy)) {
data = _mm256_lddqu_si256(
static_cast<
const __m256i *>(src) + 1);
_mm256_storeu_si256(
reinterpret_cast<__m256i *>(dst) + 1, data);
}
#else
std::memmove(dst, src, kShortMemCopy);
// Profiling shows that nearly all copies are short.
if (SNAPPY_PREDICT_FALSE(size > kShortMemCopy)) {
std::memmove(dst + kShortMemCopy,
static_cast<
const uint8_t*>(src) + kShortMemCopy,
64 - kShortMemCopy);
}
#endif
}
void MemCopy64(ptrdiff_t dst,
const void* src, size_t size) {
// TODO: Switch to [[maybe_unused]] when we can assume C++17.
(
void)dst;
(
void)src;
(
void)size;
}
void ClearDeferred(
const void** deferred_src, size_t* deferred_length,
uint8_t* safe_source) {
*deferred_src = safe_source;
*deferred_length = 0;
}
void DeferMemCopy(
const void** deferred_src, size_t* deferred_length,
const void* src, size_t length) {
*deferred_src = src;
*deferred_length = length;
}
SNAPPY_ATTRIBUTE_ALWAYS_INLINE
inline size_t AdvanceToNextTagARMOptimized(
const uint8_t** ip_p, size_t* tag) {
const uint8_t*& ip = *ip_p;
// This section is crucial for the throughput of the decompression loop.
// The latency of an iteration is fundamentally constrained by the
// following data chain on ip.
// ip -> c = Load(ip) -> delta1 = (c & 3) -> ip += delta1 or delta2
// delta2 = ((c >> 2) + 1) ip++
// This is different from X86 optimizations because ARM has conditional add
// instruction (csinc) and it removes several register moves.
const size_t tag_type = *tag & 3;
const bool is_literal = (tag_type == 0);
if (is_literal) {
size_t next_literal_tag = (*tag >> 2) + 1;
*tag = ip[next_literal_tag];
ip += next_literal_tag + 1;
}
else {
*tag = ip[tag_type];
ip += tag_type + 1;
}
return tag_type;
}
SNAPPY_ATTRIBUTE_ALWAYS_INLINE
inline size_t AdvanceToNextTagX86Optimized(
const uint8_t** ip_p, size_t* tag) {
const uint8_t*& ip = *ip_p;
// This section is crucial for the throughput of the decompression loop.
// The latency of an iteration is fundamentally constrained by the
// following data chain on ip.
// ip -> c = Load(ip) -> ip1 = ip + 1 + (c & 3) -> ip = ip1 or ip2
// ip2 = ip + 2 + (c >> 2)
// This amounts to 8 cycles.
// 5 (load) + 1 (c & 3) + 1 (lea ip1, [ip + (c & 3) + 1]) + 1 (cmov)
size_t literal_len = *tag >> 2;
size_t tag_type = *tag;
bool is_literal;
#if defined(__GCC_ASM_FLAG_OUTPUTS__) &&
defined(__x86_64__)
// TODO clang misses the fact that the (c & 3) already correctly
// sets the zero flag.
asm(
"and $3, %k[tag_type]\n\t"
: [tag_type]
"+r"(tag_type),
"=@ccz"(is_literal)
::
"cc");
#else
tag_type &= 3;
is_literal = (tag_type == 0);
#endif
// TODO
// This is code is subtle. Loading the values first and then cmov has less
// latency then cmov ip and then load. However clang would move the loads
// in an optimization phase, volatile prevents this transformation.
// Note that we have enough slop bytes (64) that the loads are always valid.
size_t tag_literal =
static_cast<
const volatile uint8_t*>(ip)[1 + literal_len];
size_t tag_copy =
static_cast<
const volatile uint8_t*>(ip)[tag_type];
*tag = is_literal ? tag_literal : tag_copy;
const uint8_t* ip_copy = ip + 1 + tag_type;
const uint8_t* ip_literal = ip + 2 + literal_len;
ip = is_literal ? ip_literal : ip_copy;
#if defined(__GNUC__) &&
defined(__x86_64__)
// TODO Clang is "optimizing" zero-extension (a totally free
// operation) this means that after the cmov of tag, it emits another movzb
// tag, byte(tag). It really matters as it's on the core chain. This dummy
// asm, persuades clang to do the zero-extension at the load (it's automatic)
// removing the expensive movzb.
asm(
"" ::
"r"(tag_copy));
#endif
return tag_type;
}
// Extract the offset for copy-1 and copy-2 returns 0 for literals or copy-4.
inline uint32_t ExtractOffset(uint32_t val, size_t tag_type) {
// For x86 non-static storage works better. For ARM static storage is better.
// TODO: Once the array is recognized as a register, improve the
// readability for x86.
#if defined(__x86_64__)
constexpr uint64_t kExtractMasksCombined = 0x0000FFFF00FF0000ull;
uint16_t result;
memcpy(&result,
reinterpret_cast<
const char*>(&kExtractMasksCombined) + 2 * tag_type,
sizeof(result));
return val & result;
#elif defined(__aarch64__)
constexpr uint64_t kExtractMasksCombined = 0x0000FFFF00FF0000ull;
return val &
static_cast<uint32_t>(
(kExtractMasksCombined >> (tag_type * 16)) & 0xFFFF);
#else
static constexpr uint32_t kExtractMasks[4] = {0, 0xFF, 0xFFFF, 0};
return val & kExtractMasks[tag_type];
#endif
};
// Core decompression loop, when there is enough data available.
// Decompresses the input buffer [ip, ip_limit) into the output buffer
// [op, op_limit_min_slop). Returning when either we are too close to the end
// of the input buffer, or we exceed op_limit_min_slop or when a exceptional
// tag is encountered (literal of length > 60) or a copy-4.
// Returns {ip, op} at the points it stopped decoding.
// TODO This function probably does not need to be inlined, as it
// should decode large chunks at a time. This allows runtime dispatch to
// implementations based on CPU capability (BMI2 / perhaps 32 / 64 byte memcpy).
template <
typename T>
std::pair<
const uint8_t*, ptrdiff_t> DecompressBranchless(
const uint8_t* ip,
const uint8_t* ip_limit, ptrdiff_t op, T op_base,
ptrdiff_t op_limit_min_slop) {
// If deferred_src is invalid point it here.
uint8_t safe_source[64];
const void* deferred_src;
size_t deferred_length;
ClearDeferred(&deferred_src, &deferred_length, safe_source);
// We unroll the inner loop twice so we need twice the spare room.
op_limit_min_slop -= kSlopBytes;
if (2 * (kSlopBytes + 1) < ip_limit - ip && op < op_limit_min_slop) {
const uint8_t*
const ip_limit_min_slop = ip_limit - 2 * kSlopBytes - 1;
ip++;
// ip points just past the tag and we are touching at maximum kSlopBytes
// in an iteration.
size_t tag = ip[-1];
#if defined(__clang__) &&
defined(__aarch64__)
// Workaround for https://bugs.llvm.org/show_bug.cgi?id=51317
// when loading 1 byte, clang for aarch64 doesn't realize that it(ldrb)
// comes with free zero-extension, so clang generates another
// 'and xn, xm, 0xff' before it use that as the offset. This 'and' is
// redundant and can be removed by adding this dummy asm, which gives
// clang a hint that we're doing the zero-extension at the load.
asm(
"" ::
"r"(tag));
#endif
do {
// The throughput is limited by instructions, unrolling the inner loop
// twice reduces the amount of instructions checking limits and also
// leads to reduced mov's.
SNAPPY_PREFETCH(ip + 128);
for (
int i = 0; i < 2; i++) {
const uint8_t* old_ip = ip;
assert(tag == ip[-1]);
// For literals tag_type = 0, hence we will always obtain 0 from
// ExtractLowBytes. For literals offset will thus be kLiteralOffset.
ptrdiff_t len_minus_offset = kLengthMinusOffset[tag];
uint32_t next;
#if defined(__aarch64__)
size_t tag_type = AdvanceToNextTagARMOptimized(&ip, &tag);
// We never need more than 16 bits. Doing a Load16 allows the compiler
// to elide the masking operation in ExtractOffset.
next = LittleEndian::Load16(old_ip);
#else
size_t tag_type = AdvanceToNextTagX86Optimized(&ip, &tag);
next = LittleEndian::Load32(old_ip);
#endif
size_t len = len_minus_offset & 0xFF;
ptrdiff_t extracted = ExtractOffset(next, tag_type);
ptrdiff_t len_min_offset = len_minus_offset - extracted;
if (SNAPPY_PREDICT_FALSE(len_minus_offset > extracted)) {
if (SNAPPY_PREDICT_FALSE(len & 0x80)) {
// Exceptional case (long literal or copy 4).
// Actually doing the copy here is negatively impacting the main
// loop due to compiler incorrectly allocating a register for
// this fallback. Hence we just break.
break_loop:
ip = old_ip;
goto exit;
}
// Only copy-1 or copy-2 tags can get here.
assert(tag_type == 1 || tag_type == 2);
std::ptrdiff_t delta = (op + deferred_length) + len_min_offset - len;
// Guard against copies before the buffer start.
// Execute any deferred MemCopy since we write to dst here.
MemCopy64(op_base + op, deferred_src, deferred_length);
op += deferred_length;
ClearDeferred(&deferred_src, &deferred_length, safe_source);
if (SNAPPY_PREDICT_FALSE(delta < 0 ||
!Copy64BytesWithPatternExtension(
op_base + op, len - len_min_offset))) {
goto break_loop;
}
// We aren't deferring this copy so add length right away.
op += len;
continue;
}
std::ptrdiff_t delta = (op + deferred_length) + len_min_offset - len;
if (SNAPPY_PREDICT_FALSE(delta < 0)) {
// Due to the spurious offset in literals have this will trigger
// at the start of a block when op is still smaller than 256.
if (tag_type != 0)
goto break_loop;
MemCopy64(op_base + op, deferred_src, deferred_length);
op += deferred_length;
DeferMemCopy(&deferred_src, &deferred_length, old_ip, len);
continue;
}
// For copies we need to copy from op_base + delta, for literals
// we need to copy from ip instead of from the stream.
const void* from =
tag_type ?
reinterpret_cast<
void*>(op_base + delta) : old_ip;
MemCopy64(op_base + op, deferred_src, deferred_length);
op += deferred_length;
DeferMemCopy(&deferred_src, &deferred_length, from, len);
}
}
while (ip < ip_limit_min_slop &&
static_cast<ptrdiff_t>(op + deferred_length) < op_limit_min_slop);
exit:
ip--;
assert(ip <= ip_limit);
}
// If we deferred a copy then we can perform. If we are up to date then we
// might not have enough slop bytes and could run past the end.
if (deferred_length) {
MemCopy64(op_base + op, deferred_src, deferred_length);
op += deferred_length;
ClearDeferred(&deferred_src, &deferred_length, safe_source);
}
return {ip, op};
}
// Helper class for decompression
class SnappyDecompressor {
private:
Source* reader_;
// Underlying source of bytes to decompress
const char* ip_;
// Points to next buffered byte
const char* ip_limit_;
// Points just past buffered bytes
// If ip < ip_limit_min_maxtaglen_ it's safe to read kMaxTagLength from
// buffer.
const char* ip_limit_min_maxtaglen_;
uint32_t peeked_;
// Bytes peeked from reader (need to skip)
bool eof_;
// Hit end of input without an error?
char scratch_[kMaximumTagLength];
// See RefillTag().
// Ensure that all of the tag metadata for the next tag is available
// in [ip_..ip_limit_-1]. Also ensures that [ip,ip+4] is readable even
// if (ip_limit_ - ip_ < 5).
//
// Returns true on success, false on error or end of input.
bool RefillTag();
void ResetLimit(
const char* ip) {
ip_limit_min_maxtaglen_ =
ip_limit_ - std::min<ptrdiff_t>(ip_limit_ - ip, kMaximumTagLength - 1);
}
public:
explicit SnappyDecompressor(Source* reader)
: reader_(reader), ip_(NULL), ip_limit_(NULL), peeked_(0), eof_(
false) {}
~SnappyDecompressor() {
// Advance past any bytes we peeked at from the reader
reader_->Skip(peeked_);
}
// Returns true iff we have hit the end of the input without an error.
bool eof()
const {
return eof_; }
// Read the uncompressed length stored at the start of the compressed data.
// On success, stores the length in *result and returns true.
// On failure, returns false.
bool ReadUncompressedLength(uint32_t* result) {
assert(ip_ == NULL);
// Must not have read anything yet
// Length is encoded in 1..5 bytes
*result = 0;
uint32_t shift = 0;
while (
true) {
if (shift >= 32)
return false;
size_t n;
const char* ip = reader_->Peek(&n);
if (n == 0)
return false;
const unsigned char c = *(
reinterpret_cast<
const unsigned char*>(ip));
reader_->Skip(1);
uint32_t val = c & 0x7f;
if (LeftShiftOverflows(
static_cast<uint8_t>(val), shift))
return false;
*result |= val << shift;
if (c < 128) {
break;
}
shift += 7;
}
return true;
}
// Process the next item found in the input.
// Returns true if successful, false on error or end of input.
template <
class Writer>
#if defined(__GNUC__) &&
defined(__x86_64__)
__attribute__((aligned(32)))
#endif
void
DecompressAllTags(Writer* writer) {
const char* ip = ip_;
ResetLimit(ip);
auto op = writer->GetOutputPtr();
// We could have put this refill fragment only at the beginning of the loop.
// However, duplicating it at the end of each branch gives the compiler more
// scope to optimize the <ip_limit_ - ip> expression based on the local
// context, which overall increases speed.
#define MAYBE_REFILL() \
if (SNAPPY_PREDICT_FALSE(ip >= ip_limit_min_maxtaglen_)) { \
ip_ = ip; \
if (SNAPPY_PREDICT_FALSE(!RefillTag()))
goto exit; \
ip = ip_; \
ResetLimit(ip); \
} \
preload =
static_cast<uint8_t>(*ip)
// At the start of the for loop below the least significant byte of preload
// contains the tag.
uint32_t preload;
MAYBE_REFILL();
for (;;) {
{
ptrdiff_t op_limit_min_slop;
auto op_base = writer->GetBase(&op_limit_min_slop);
if (op_base) {
auto res =
DecompressBranchless(
reinterpret_cast<
const uint8_t*>(ip),
reinterpret_cast<
const uint8_t*>(ip_limit_),
op - op_base, op_base, op_limit_min_slop);
ip =
reinterpret_cast<
const char*>(res.first);
op = op_base + res.second;
MAYBE_REFILL();
}
}
const uint8_t c =
static_cast<uint8_t>(preload);
ip++;
// Ratio of iterations that have LITERAL vs non-LITERAL for different
// inputs.
//
// input LITERAL NON_LITERAL
// -----------------------------------
// html|html4|cp 23% 77%
// urls 36% 64%
// jpg 47% 53%
--> --------------------
--> maximum size reached
--> --------------------