diff options
Diffstat (limited to 'other-licenses/snappy/src')
| -rw-r--r-- | other-licenses/snappy/src/CONTRIBUTING.md | 33 | ||||
| -rw-r--r-- | other-licenses/snappy/src/NEWS | 6 | ||||
| -rw-r--r-- | other-licenses/snappy/src/README.md | 33 | ||||
| -rw-r--r-- | other-licenses/snappy/src/snappy-internal.h | 127 | ||||
| -rw-r--r-- | other-licenses/snappy/src/snappy-stubs-internal.h | 85 | ||||
| -rw-r--r-- | other-licenses/snappy/src/snappy-test.cc | 2 | ||||
| -rw-r--r-- | other-licenses/snappy/src/snappy-test.h | 18 | ||||
| -rw-r--r-- | other-licenses/snappy/src/snappy.cc | 730 | ||||
| -rw-r--r-- | other-licenses/snappy/src/snappy.h | 56 | ||||
| -rw-r--r-- | other-licenses/snappy/src/snappy_compress_fuzzer.cc | 32 | ||||
| -rw-r--r-- | other-licenses/snappy/src/snappy_unittest.cc | 86 |
11 files changed, 940 insertions, 268 deletions
diff --git a/other-licenses/snappy/src/CONTRIBUTING.md b/other-licenses/snappy/src/CONTRIBUTING.md index d0ce551527..66a60d5c86 100644 --- a/other-licenses/snappy/src/CONTRIBUTING.md +++ b/other-licenses/snappy/src/CONTRIBUTING.md @@ -3,30 +3,10 @@ We'd love to accept your patches and contributions to this project. There are just a few small guidelines you need to follow. -## Project Goals - -In addition to the aims listed at the top of the [README](README.md) Snappy -explicitly supports the following: - -1. C++11 -2. Clang (gcc and MSVC are best-effort). -3. Low level optimizations (e.g. assembly or equivalent intrinsics) for: - 1. [x86](https://en.wikipedia.org/wiki/X86) - 2. [x86-64](https://en.wikipedia.org/wiki/X86-64) - 3. ARMv7 (32-bit) - 4. ARMv8 (AArch64) -4. Supports only the Snappy compression scheme as described in - [format_description.txt](format_description.txt). -5. CMake for building - -Changes adding features or dependencies outside of the core area of focus listed -above might not be accepted. If in doubt post a message to the -[Snappy discussion mailing list](https://groups.google.com/g/snappy-compression). - ## Contributor License Agreement Contributions to this project must be accompanied by a Contributor License -Agreement. You (or your employer) retain the copyright to your contribution, +Agreement. You (or your employer) retain the copyright to your contribution; this simply gives us permission to use and redistribute your contributions as part of the project. Head over to <https://cla.developers.google.com/> to see your current agreements on file or to sign a new one. @@ -35,12 +15,17 @@ You generally only need to submit a CLA once, so if you've already submitted one (even if it was for a different project), you probably don't need to do it again. -## Code reviews +## Code Reviews All submissions, including submissions by project members, require review. We use GitHub pull requests for this purpose. Consult [GitHub Help](https://help.github.com/articles/about-pull-requests/) for more information on using pull requests. -Please make sure that all the automated checks (CLA, AppVeyor, Travis) pass for -your pull requests. Pull requests whose checks fail may be ignored. +See [the README](README.md#contributing-to-the-snappy-project) for areas +where we are likely to accept external contributions. + +## Community Guidelines + +This project follows [Google's Open Source Community +Guidelines](https://opensource.google/conduct/). diff --git a/other-licenses/snappy/src/NEWS b/other-licenses/snappy/src/NEWS index 931a5e13fd..792a578001 100644 --- a/other-licenses/snappy/src/NEWS +++ b/other-licenses/snappy/src/NEWS @@ -1,3 +1,9 @@ +Snappy v1.1.10, Mar 8th 2023: + + * Performance improvements + + * Compilation fixes for various environments + Snappy v1.1.9, May 4th 2021: * Performance improvements. diff --git a/other-licenses/snappy/src/README.md b/other-licenses/snappy/src/README.md index 8fd7dc09ed..398be7d58a 100644 --- a/other-licenses/snappy/src/README.md +++ b/other-licenses/snappy/src/README.md @@ -1,7 +1,6 @@ Snappy, a fast compressor/decompressor. -[](https://travis-ci.org/google/snappy) -[](https://ci.appveyor.com/project/pwnall/leveldb) +[](https://github.com/google/snappy/actions/workflows/build.yml) Introduction ============ @@ -90,13 +89,13 @@ your calling file, and link against the compiled library. There are many ways to call Snappy, but the simplest possible is -```cpp +```c++ snappy::Compress(input.data(), input.size(), &output); ``` and similarly -```cpp +```c++ snappy::Uncompress(input.data(), input.size(), &output); ``` @@ -132,6 +131,32 @@ should provide a reasonably balanced starting point for benchmarking. (Note that baddata[1-3].snappy are not intended as benchmarks; they are used to verify correctness in the presence of corrupted data in the unit test.) +Contributing to the Snappy Project +================================== + +In addition to the aims listed at the top of the [README](README.md) Snappy +explicitly supports the following: + +1. C++11 +2. Clang (gcc and MSVC are best-effort). +3. Low level optimizations (e.g. assembly or equivalent intrinsics) for: + 1. [x86](https://en.wikipedia.org/wiki/X86) + 2. [x86-64](https://en.wikipedia.org/wiki/X86-64) + 3. ARMv7 (32-bit) + 4. ARMv8 (AArch64) +4. Supports only the Snappy compression scheme as described in + [format_description.txt](format_description.txt). +5. CMake for building + +Changes adding features or dependencies outside of the core area of focus listed +above might not be accepted. If in doubt post a message to the +[Snappy discussion mailing list](https://groups.google.com/g/snappy-compression). + +We are unlikely to accept contributions to the build configuration files, such +as `CMakeLists.txt`. We are focused on maintaining a build configuration that +allows us to test that the project works in a few supported configurations +inside Google. We are not currently interested in supporting other requirements, +such as different operating systems, compilers, or build systems. Contact ======= diff --git a/other-licenses/snappy/src/snappy-internal.h b/other-licenses/snappy/src/snappy-internal.h index 720ccd8282..ae78247dbb 100644 --- a/other-licenses/snappy/src/snappy-internal.h +++ b/other-licenses/snappy/src/snappy-internal.h @@ -31,11 +31,88 @@ #ifndef THIRD_PARTY_SNAPPY_SNAPPY_INTERNAL_H_ #define THIRD_PARTY_SNAPPY_SNAPPY_INTERNAL_H_ +#include <utility> + #include "snappy-stubs-internal.h" +#if SNAPPY_HAVE_SSSE3 +// Please do not replace with <x86intrin.h> or with headers that assume more +// advanced SSE versions without checking with all the OWNERS. +#include <emmintrin.h> +#include <tmmintrin.h> +#endif + +#if SNAPPY_HAVE_NEON +#include <arm_neon.h> +#endif + +#if SNAPPY_HAVE_SSSE3 || SNAPPY_HAVE_NEON +#define SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE 1 +#else +#define SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE 0 +#endif + namespace snappy { namespace internal { +#if SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE +#if SNAPPY_HAVE_SSSE3 +using V128 = __m128i; +#elif SNAPPY_HAVE_NEON +using V128 = uint8x16_t; +#endif + +// Load 128 bits of integer data. `src` must be 16-byte aligned. +inline V128 V128_Load(const V128* src); + +// Load 128 bits of integer data. `src` does not need to be aligned. +inline V128 V128_LoadU(const V128* src); + +// Store 128 bits of integer data. `dst` does not need to be aligned. +inline void V128_StoreU(V128* dst, V128 val); + +// Shuffle packed 8-bit integers using a shuffle mask. +// Each packed integer in the shuffle mask must be in [0,16). +inline V128 V128_Shuffle(V128 input, V128 shuffle_mask); + +// Constructs V128 with 16 chars |c|. +inline V128 V128_DupChar(char c); + +#if SNAPPY_HAVE_SSSE3 +inline V128 V128_Load(const V128* src) { return _mm_load_si128(src); } + +inline V128 V128_LoadU(const V128* src) { return _mm_loadu_si128(src); } + +inline void V128_StoreU(V128* dst, V128 val) { _mm_storeu_si128(dst, val); } + +inline V128 V128_Shuffle(V128 input, V128 shuffle_mask) { + return _mm_shuffle_epi8(input, shuffle_mask); +} + +inline V128 V128_DupChar(char c) { return _mm_set1_epi8(c); } + +#elif SNAPPY_HAVE_NEON +inline V128 V128_Load(const V128* src) { + return vld1q_u8(reinterpret_cast<const uint8_t*>(src)); +} + +inline V128 V128_LoadU(const V128* src) { + return vld1q_u8(reinterpret_cast<const uint8_t*>(src)); +} + +inline void V128_StoreU(V128* dst, V128 val) { + vst1q_u8(reinterpret_cast<uint8_t*>(dst), val); +} + +inline V128 V128_Shuffle(V128 input, V128 shuffle_mask) { + assert(vminvq_u8(shuffle_mask) >= 0 && vmaxvq_u8(shuffle_mask) <= 15); + return vqtbl1q_u8(input, shuffle_mask); +} + +inline V128 V128_DupChar(char c) { return vdupq_n_u8(c); } +#endif +#endif // SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE + // Working memory performs a single allocation to hold all scratch space // required for compression. class WorkingMemory { @@ -95,8 +172,9 @@ char* CompressFragment(const char* input, // loading from s2 + n. // // Separate implementation for 64-bit, little-endian cpus. -#if !defined(SNAPPY_IS_BIG_ENDIAN) && \ - (defined(__x86_64__) || defined(_M_X64) || defined(ARCH_PPC) || defined(ARCH_ARM)) +#if !SNAPPY_IS_BIG_ENDIAN && \ + (defined(__x86_64__) || defined(_M_X64) || defined(ARCH_PPC) || \ + defined(ARCH_ARM)) static inline std::pair<size_t, bool> FindMatchLength(const char* s1, const char* s2, const char* s2_limit, @@ -154,8 +232,9 @@ static inline std::pair<size_t, bool> FindMatchLength(const char* s1, uint64_t xorval = a1 ^ a2; int shift = Bits::FindLSBSetNonZero64(xorval); size_t matched_bytes = shift >> 3; + uint64_t a3 = UNALIGNED_LOAD64(s2 + 4); #ifndef __x86_64__ - *data = UNALIGNED_LOAD64(s2 + matched_bytes); + a2 = static_cast<uint32_t>(xorval) == 0 ? a3 : a2; #else // Ideally this would just be // @@ -166,19 +245,21 @@ static inline std::pair<size_t, bool> FindMatchLength(const char* s1, // use a conditional move (it's tuned to cut data dependencies). In this // case there is a longer parallel chain anyway AND this will be fairly // unpredictable. - uint64_t a3 = UNALIGNED_LOAD64(s2 + 4); asm("testl %k2, %k2\n\t" "cmovzq %1, %0\n\t" : "+r"(a2) - : "r"(a3), "r"(xorval)); - *data = a2 >> (shift & (3 * 8)); + : "r"(a3), "r"(xorval) + : "cc"); #endif + *data = a2 >> (shift & (3 * 8)); return std::pair<size_t, bool>(matched_bytes, true); } else { matched = 8; s2 += 8; } } + SNAPPY_PREFETCH(s1 + 64); + SNAPPY_PREFETCH(s2 + 64); // Find out how long the match is. We loop over the data 64 bits at a // time until we find a 64-bit block that doesn't match; then we find @@ -194,16 +275,17 @@ static inline std::pair<size_t, bool> FindMatchLength(const char* s1, uint64_t xorval = a1 ^ a2; int shift = Bits::FindLSBSetNonZero64(xorval); size_t matched_bytes = shift >> 3; + uint64_t a3 = UNALIGNED_LOAD64(s2 + 4); #ifndef __x86_64__ - *data = UNALIGNED_LOAD64(s2 + matched_bytes); + a2 = static_cast<uint32_t>(xorval) == 0 ? a3 : a2; #else - uint64_t a3 = UNALIGNED_LOAD64(s2 + 4); asm("testl %k2, %k2\n\t" "cmovzq %1, %0\n\t" : "+r"(a2) - : "r"(a3), "r"(xorval)); - *data = a2 >> (shift & (3 * 8)); + : "r"(a3), "r"(xorval) + : "cc"); #endif + *data = a2 >> (shift & (3 * 8)); matched += matched_bytes; assert(matched >= 8); return std::pair<size_t, bool>(matched, false); @@ -252,6 +334,31 @@ static inline std::pair<size_t, bool> FindMatchLength(const char* s1, } #endif +static inline size_t FindMatchLengthPlain(const char* s1, const char* s2, + const char* s2_limit) { + // Implementation based on the x86-64 version, above. + assert(s2_limit >= s2); + int matched = 0; + + while (s2 <= s2_limit - 8 && + UNALIGNED_LOAD64(s2) == UNALIGNED_LOAD64(s1 + matched)) { + s2 += 8; + matched += 8; + } + if (LittleEndian::IsLittleEndian() && s2 <= s2_limit - 8) { + uint64_t x = UNALIGNED_LOAD64(s2) ^ UNALIGNED_LOAD64(s1 + matched); + int matching_bits = Bits::FindLSBSetNonZero64(x); + matched += matching_bits >> 3; + s2 += matching_bits >> 3; + } else { + while ((s2 < s2_limit) && (s1[matched] == *s2)) { + ++s2; + ++matched; + } + } + return matched; +} + // Lookup tables for decompression code. Give --snappy_dump_decompression_table // to the unit test to recompute char_table. diff --git a/other-licenses/snappy/src/snappy-stubs-internal.h b/other-licenses/snappy/src/snappy-stubs-internal.h index c2a838f38f..526c38b700 100644 --- a/other-licenses/snappy/src/snappy-stubs-internal.h +++ b/other-licenses/snappy/src/snappy-stubs-internal.h @@ -31,7 +31,7 @@ #ifndef THIRD_PARTY_SNAPPY_OPENSOURCE_SNAPPY_STUBS_INTERNAL_H_ #define THIRD_PARTY_SNAPPY_OPENSOURCE_SNAPPY_STUBS_INTERNAL_H_ -#ifdef HAVE_CONFIG_H +#if HAVE_CONFIG_H #include "config.h" #endif @@ -43,11 +43,11 @@ #include <limits> #include <string> -#ifdef HAVE_SYS_MMAN_H +#if HAVE_SYS_MMAN_H #include <sys/mman.h> #endif -#ifdef HAVE_UNISTD_H +#if HAVE_UNISTD_H #include <unistd.h> #endif @@ -90,19 +90,25 @@ #define ARRAYSIZE(a) int{sizeof(a) / sizeof(*(a))} // Static prediction hints. -#ifdef HAVE_BUILTIN_EXPECT +#if HAVE_BUILTIN_EXPECT #define SNAPPY_PREDICT_FALSE(x) (__builtin_expect(x, 0)) #define SNAPPY_PREDICT_TRUE(x) (__builtin_expect(!!(x), 1)) #else #define SNAPPY_PREDICT_FALSE(x) x #define SNAPPY_PREDICT_TRUE(x) x -#endif +#endif // HAVE_BUILTIN_EXPECT // Inlining hints. -#ifdef HAVE_ATTRIBUTE_ALWAYS_INLINE +#if HAVE_ATTRIBUTE_ALWAYS_INLINE #define SNAPPY_ATTRIBUTE_ALWAYS_INLINE __attribute__((always_inline)) #else #define SNAPPY_ATTRIBUTE_ALWAYS_INLINE +#endif // HAVE_ATTRIBUTE_ALWAYS_INLINE + +#if HAVE_BUILTIN_PREFETCH +#define SNAPPY_PREFETCH(ptr) __builtin_prefetch(ptr, 0, 3) +#else +#define SNAPPY_PREFETCH(ptr) (void)(ptr) #endif // Stubbed version of ABSL_FLAG. @@ -171,27 +177,42 @@ class LittleEndian { public: // Functions to do unaligned loads and stores in little-endian order. static inline uint16_t Load16(const void *ptr) { - const uint8_t* const buffer = reinterpret_cast<const uint8_t*>(ptr); - // Compiles to a single mov/str on recent clang and gcc. +#if SNAPPY_IS_BIG_ENDIAN + const uint8_t* const buffer = reinterpret_cast<const uint8_t*>(ptr); return (static_cast<uint16_t>(buffer[0])) | (static_cast<uint16_t>(buffer[1]) << 8); +#else + // memcpy() turns into a single instruction early in the optimization + // pipeline (relatively to a series of byte accesses). So, using memcpy + // instead of byte accesses may lead to better decisions in more stages of + // the optimization pipeline. + uint16_t value; + std::memcpy(&value, ptr, 2); + return value; +#endif } static inline uint32_t Load32(const void *ptr) { - const uint8_t* const buffer = reinterpret_cast<const uint8_t*>(ptr); - // Compiles to a single mov/str on recent clang and gcc. +#if SNAPPY_IS_BIG_ENDIAN + const uint8_t* const buffer = reinterpret_cast<const uint8_t*>(ptr); return (static_cast<uint32_t>(buffer[0])) | (static_cast<uint32_t>(buffer[1]) << 8) | (static_cast<uint32_t>(buffer[2]) << 16) | (static_cast<uint32_t>(buffer[3]) << 24); +#else + // See Load16() for the rationale of using memcpy(). + uint32_t value; + std::memcpy(&value, ptr, 4); + return value; +#endif } static inline uint64_t Load64(const void *ptr) { - const uint8_t* const buffer = reinterpret_cast<const uint8_t*>(ptr); - // Compiles to a single mov/str on recent clang and gcc. +#if SNAPPY_IS_BIG_ENDIAN + const uint8_t* const buffer = reinterpret_cast<const uint8_t*>(ptr); return (static_cast<uint64_t>(buffer[0])) | (static_cast<uint64_t>(buffer[1]) << 8) | (static_cast<uint64_t>(buffer[2]) << 16) | @@ -200,30 +221,44 @@ class LittleEndian { (static_cast<uint64_t>(buffer[5]) << 40) | (static_cast<uint64_t>(buffer[6]) << 48) | (static_cast<uint64_t>(buffer[7]) << 56); +#else + // See Load16() for the rationale of using memcpy(). + uint64_t value; + std::memcpy(&value, ptr, 8); + return value; +#endif } static inline void Store16(void *dst, uint16_t value) { - uint8_t* const buffer = reinterpret_cast<uint8_t*>(dst); - // Compiles to a single mov/str on recent clang and gcc. +#if SNAPPY_IS_BIG_ENDIAN + uint8_t* const buffer = reinterpret_cast<uint8_t*>(dst); buffer[0] = static_cast<uint8_t>(value); buffer[1] = static_cast<uint8_t>(value >> 8); +#else + // See Load16() for the rationale of using memcpy(). + std::memcpy(dst, &value, 2); +#endif } static void Store32(void *dst, uint32_t value) { - uint8_t* const buffer = reinterpret_cast<uint8_t*>(dst); - // Compiles to a single mov/str on recent clang and gcc. +#if SNAPPY_IS_BIG_ENDIAN + uint8_t* const buffer = reinterpret_cast<uint8_t*>(dst); buffer[0] = static_cast<uint8_t>(value); buffer[1] = static_cast<uint8_t>(value >> 8); buffer[2] = static_cast<uint8_t>(value >> 16); buffer[3] = static_cast<uint8_t>(value >> 24); +#else + // See Load16() for the rationale of using memcpy(). + std::memcpy(dst, &value, 4); +#endif } static void Store64(void* dst, uint64_t value) { - uint8_t* const buffer = reinterpret_cast<uint8_t*>(dst); - // Compiles to a single mov/str on recent clang and gcc. +#if SNAPPY_IS_BIG_ENDIAN + uint8_t* const buffer = reinterpret_cast<uint8_t*>(dst); buffer[0] = static_cast<uint8_t>(value); buffer[1] = static_cast<uint8_t>(value >> 8); buffer[2] = static_cast<uint8_t>(value >> 16); @@ -232,14 +267,18 @@ class LittleEndian { buffer[5] = static_cast<uint8_t>(value >> 40); buffer[6] = static_cast<uint8_t>(value >> 48); buffer[7] = static_cast<uint8_t>(value >> 56); +#else + // See Load16() for the rationale of using memcpy(). + std::memcpy(dst, &value, 8); +#endif } static inline constexpr bool IsLittleEndian() { -#if defined(SNAPPY_IS_BIG_ENDIAN) +#if SNAPPY_IS_BIG_ENDIAN return false; #else return true; -#endif // defined(SNAPPY_IS_BIG_ENDIAN) +#endif // SNAPPY_IS_BIG_ENDIAN } }; @@ -265,7 +304,7 @@ class Bits { void operator=(const Bits&); }; -#if defined(HAVE_BUILTIN_CTZ) +#if HAVE_BUILTIN_CTZ inline int Bits::Log2FloorNonZero(uint32_t n) { assert(n != 0); @@ -354,7 +393,7 @@ inline int Bits::FindLSBSetNonZero(uint32_t n) { #endif // End portable versions. -#if defined(HAVE_BUILTIN_CTZ) +#if HAVE_BUILTIN_CTZ inline int Bits::FindLSBSetNonZero64(uint64_t n) { assert(n != 0); @@ -388,7 +427,7 @@ inline int Bits::FindLSBSetNonZero64(uint64_t n) { } } -#endif // End portable version. +#endif // HAVE_BUILTIN_CTZ // Variable-length integer encoding. class Varint { diff --git a/other-licenses/snappy/src/snappy-test.cc b/other-licenses/snappy/src/snappy-test.cc index 7eb490ac17..aae607210b 100644 --- a/other-licenses/snappy/src/snappy-test.cc +++ b/other-licenses/snappy/src/snappy-test.cc @@ -151,7 +151,7 @@ LogMessageCrash::~LogMessageCrash() { #pragma warning(pop) #endif -#ifdef HAVE_LIBZ +#if HAVE_LIBZ ZLib::ZLib() : comp_init_(false), diff --git a/other-licenses/snappy/src/snappy-test.h b/other-licenses/snappy/src/snappy-test.h index f80d343377..65f3725744 100644 --- a/other-licenses/snappy/src/snappy-test.h +++ b/other-licenses/snappy/src/snappy-test.h @@ -31,25 +31,25 @@ #ifndef THIRD_PARTY_SNAPPY_OPENSOURCE_SNAPPY_TEST_H_ #define THIRD_PARTY_SNAPPY_OPENSOURCE_SNAPPY_TEST_H_ -#ifdef HAVE_CONFIG_H +#if HAVE_CONFIG_H #include "config.h" #endif #include "snappy-stubs-internal.h" -#ifdef HAVE_SYS_MMAN_H +#if HAVE_SYS_MMAN_H #include <sys/mman.h> #endif -#ifdef HAVE_SYS_RESOURCE_H +#if HAVE_SYS_RESOURCE_H #include <sys/resource.h> #endif -#ifdef HAVE_SYS_TIME_H +#if HAVE_SYS_TIME_H #include <sys/time.h> #endif -#ifdef HAVE_WINDOWS_H +#if HAVE_WINDOWS_H // Needed to be able to use std::max without workarounds in the source code. // https://support.microsoft.com/en-us/help/143208/prb-using-stl-in-windows-program-can-cause-min-max-conflicts #define NOMINMAX @@ -58,15 +58,15 @@ #define InitGoogle(argv0, argc, argv, remove_flags) ((void)(0)) -#ifdef HAVE_LIBZ +#if HAVE_LIBZ #include "zlib.h" #endif -#ifdef HAVE_LIBLZO2 +#if HAVE_LIBLZO2 #include "lzo/lzo1x.h" #endif -#ifdef HAVE_LIBLZ4 +#if HAVE_LIBLZ4 #include "lz4.h" #endif @@ -216,7 +216,7 @@ class LogMessageVoidify { #define CHECK_GT(a, b) CRASH_UNLESS((a) > (b)) #define CHECK_OK(cond) (cond).ok() -#ifdef HAVE_LIBZ +#if HAVE_LIBZ // Object-oriented wrapper around zlib. class ZLib { diff --git a/other-licenses/snappy/src/snappy.cc b/other-licenses/snappy/src/snappy.cc index 57df3f11fc..08c2a9889f 100644 --- a/other-licenses/snappy/src/snappy.cc +++ b/other-licenses/snappy/src/snappy.cc @@ -29,18 +29,6 @@ #include "snappy-internal.h" #include "snappy-sinksource.h" #include "snappy.h" - -#if !defined(SNAPPY_HAVE_SSSE3) -// __SSSE3__ is defined by GCC and Clang. Visual Studio doesn't target SIMD -// support between SSE2 and AVX (so SSSE3 instructions require AVX support), and -// defines __AVX__ when AVX support is available. -#if defined(__SSSE3__) || defined(__AVX__) -#define SNAPPY_HAVE_SSSE3 1 -#else -#define SNAPPY_HAVE_SSSE3 0 -#endif -#endif // !defined(SNAPPY_HAVE_SSSE3) - #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. @@ -56,16 +44,28 @@ #endif #endif // !defined(SNAPPY_HAVE_BMI2) -#if SNAPPY_HAVE_SSSE3 -// Please do not replace with <x86intrin.h>. or with headers that assume more -// advanced SSE versions without checking with all the OWNERS. -#include <tmmintrin.h> +#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 SNAPPY_HAVE_BMI2 +#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> @@ -74,6 +74,7 @@ #include <cstdint> #include <cstdio> #include <cstring> +#include <memory> #include <string> #include <utility> #include <vector> @@ -91,6 +92,14 @@ 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 @@ -133,21 +142,53 @@ constexpr std::array<int16_t, 256> MakeTable(index_sequence<seq...>) { return std::array<int16_t, 256>{LengthMinusOffset(seq)...}; } -// We maximally co-locate the two tables so that only one register needs to be -// reserved for the table address. -struct { - alignas(64) const std::array<int16_t, 256> length_minus_offset; - uint32_t extract_masks[4]; // Used for extracting offset based on tag type. -} table = {MakeTable(make_index_sequence<256>{}), {0, 0xFF, 0xFFFF, 0}}; - -// Any hash function will produce a valid compressed bitstream, but a good -// hash function reduces the number of collisions and thus yields better -// compression for compressible input, and more speed for incompressible -// input. Of course, it doesn't hurt if the hash function is reasonably fast -// either, as it gets called a lot. -inline uint32_t HashBytes(uint32_t bytes, uint32_t mask) { +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; - return ((kMagic * bytes) >> (32 - kMaxHashTableBits)) & mask; + 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 @@ -228,7 +269,7 @@ inline char* IncrementalCopySlow(const char* src, char* op, return op_limit; } -#if SNAPPY_HAVE_SSSE3 +#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 @@ -248,19 +289,19 @@ inline constexpr std::array<char, sizeof...(indexes)> MakePatternMaskBytes( // 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(__m128i)>, +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(__m128i)>())...}; + 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(__m128i)>, +alignas(16) constexpr std::array<std::array<char, sizeof(V128)>, 16> pattern_generation_masks = MakePatternMaskBytesTable( /*index_offset=*/0, @@ -271,40 +312,40 @@ alignas(16) constexpr std::array<std::array<char, sizeof(__m128i)>, // 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(__m128i)>, +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 __m128i LoadPattern(const char* src, const size_t pattern_size) { - __m128i generation_mask = _mm_load_si128(reinterpret_cast<const __m128i*>( +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 _mm_shuffle_epi8( - _mm_loadu_si128(reinterpret_cast<const __m128i*>(src)), generation_mask); + return V128_Shuffle(V128_LoadU(reinterpret_cast<const V128*>(src)), + generation_mask); } SNAPPY_ATTRIBUTE_ALWAYS_INLINE -static inline std::pair<__m128i /* pattern */, __m128i /* reshuffle_mask */> +static inline std::pair<V128 /* pattern */, V128 /* reshuffle_mask */> LoadPatternAndReshuffleMask(const char* src, const size_t pattern_size) { - __m128i pattern = LoadPattern(src, 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 _mm_storeu_si128. + // 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 - __m128i reshuffle_mask = _mm_load_si128(reinterpret_cast<const __m128i*>( + V128 reshuffle_mask = V128_Load(reinterpret_cast<const V128*>( pattern_reshuffle_masks[pattern_size - 1].data())); return {pattern, reshuffle_mask}; } -#endif // SNAPPY_HAVE_SSSE3 +#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. @@ -312,33 +353,38 @@ LoadPatternAndReshuffleMask(const char* src, const size_t pattern_size) { // 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_SSSE3 +#if SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE if (SNAPPY_PREDICT_TRUE(offset <= 16)) { switch (offset) { case 0: return false; case 1: { - std::memset(dst, dst[-1], 64); + // 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: { - __m128i pattern = LoadPattern(dst - offset, offset); + V128 pattern = LoadPattern(dst - offset, offset); for (int i = 0; i < 4; i++) { - _mm_storeu_si128(reinterpret_cast<__m128i*>(dst + 16 * i), pattern); + V128_StoreU(reinterpret_cast<V128*>(dst + 16 * i), pattern); } return true; } default: { auto pattern_and_reshuffle_mask = LoadPatternAndReshuffleMask(dst - offset, offset); - __m128i pattern = pattern_and_reshuffle_mask.first; - __m128i reshuffle_mask = pattern_and_reshuffle_mask.second; + V128 pattern = pattern_and_reshuffle_mask.first; + V128 reshuffle_mask = pattern_and_reshuffle_mask.second; for (int i = 0; i < 4; i++) { - _mm_storeu_si128(reinterpret_cast<__m128i*>(dst + 16 * i), pattern); - pattern = _mm_shuffle_epi8(pattern, reshuffle_mask); + V128_StoreU(reinterpret_cast<V128*>(dst + 16 * i), pattern); + pattern = V128_Shuffle(pattern, reshuffle_mask); } return true; } @@ -348,6 +394,7 @@ static inline bool Copy64BytesWithPatternExtension(char* dst, size_t offset) { 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 = []() { @@ -361,7 +408,7 @@ static inline bool Copy64BytesWithPatternExtension(char* dst, size_t offset) { } return true; } -#endif // SNAPPY_HAVE_SSSE3 +#endif // SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE // Very rare. for (int i = 0; i < 4; i++) { @@ -375,7 +422,7 @@ static inline bool Copy64BytesWithPatternExtension(char* dst, size_t offset) { // region of the buffer. inline char* IncrementalCopy(const char* src, char* op, char* const op_limit, char* const buf_limit) { -#if SNAPPY_HAVE_SSSE3 +#if SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE constexpr int big_pattern_size_lower_bound = 16; #else constexpr int big_pattern_size_lower_bound = 8; @@ -425,14 +472,14 @@ inline char* IncrementalCopy(const char* src, char* op, char* const op_limit, // 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_SSSE3 +#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 _mm_shuffle_epi8(), "pattern" will have five copies of "abc" + // 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 @@ -444,26 +491,26 @@ inline char* IncrementalCopy(const char* src, char* op, char* const op_limit, if (SNAPPY_PREDICT_TRUE(op_limit <= buf_limit - 15)) { auto pattern_and_reshuffle_mask = LoadPatternAndReshuffleMask(src, pattern_size); - __m128i pattern = pattern_and_reshuffle_mask.first; - __m128i reshuffle_mask = pattern_and_reshuffle_mask.second; + 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. - _mm_storeu_si128(reinterpret_cast<__m128i*>(op), pattern); + V128_StoreU(reinterpret_cast<V128*>(op), pattern); if (op + 16 < op_limit) { - pattern = _mm_shuffle_epi8(pattern, reshuffle_mask); - _mm_storeu_si128(reinterpret_cast<__m128i*>(op + 16), pattern); + pattern = V128_Shuffle(pattern, reshuffle_mask); + V128_StoreU(reinterpret_cast<V128*>(op + 16), pattern); } if (op + 32 < op_limit) { - pattern = _mm_shuffle_epi8(pattern, reshuffle_mask); - _mm_storeu_si128(reinterpret_cast<__m128i*>(op + 32), pattern); + pattern = V128_Shuffle(pattern, reshuffle_mask); + V128_StoreU(reinterpret_cast<V128*>(op + 32), pattern); } if (op + 48 < op_limit) { - pattern = _mm_shuffle_epi8(pattern, reshuffle_mask); - _mm_storeu_si128(reinterpret_cast<__m128i*>(op + 48), pattern); + pattern = V128_Shuffle(pattern, reshuffle_mask); + V128_StoreU(reinterpret_cast<V128*>(op + 48), pattern); } return op_limit; } @@ -471,8 +518,8 @@ inline char* IncrementalCopy(const char* src, char* op, char* const op_limit, if (SNAPPY_PREDICT_TRUE(op < op_end)) { auto pattern_and_reshuffle_mask = LoadPatternAndReshuffleMask(src, pattern_size); - __m128i pattern = pattern_and_reshuffle_mask.first; - __m128i reshuffle_mask = pattern_and_reshuffle_mask.second; + 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. @@ -483,13 +530,13 @@ inline char* IncrementalCopy(const char* src, char* op, char* const op_limit, #pragma clang loop unroll(disable) #endif do { - _mm_storeu_si128(reinterpret_cast<__m128i*>(op), pattern); - pattern = _mm_shuffle_epi8(pattern, reshuffle_mask); + 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_SSSE3 +#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 @@ -506,7 +553,7 @@ inline char* IncrementalCopy(const char* src, char* op, char* const op_limit, } else { return IncrementalCopySlow(src, op, op_limit); } -#endif // SNAPPY_HAVE_SSSE3 +#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; @@ -599,7 +646,19 @@ static inline char* EmitLiteral(char* op, const char* literal, int len) { LittleEndian::Store32(op, n); op += count; } - std::memcpy(op, literal, len); + // 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; } @@ -734,7 +793,7 @@ char* CompressFragment(const char* input, size_t input_size, char* op, 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 = table_size - 1; + const uint32_t mask = 2 * (table_size - 1); const char* ip_end = input + input_size; const char* base_ip = ip; @@ -785,11 +844,11 @@ char* CompressFragment(const char* input, size_t input_size, char* op, // loaded in preload. uint32_t dword = i == 0 ? preload : static_cast<uint32_t>(data); assert(dword == LittleEndian::Load32(ip + i)); - uint32_t hash = HashBytes(dword, mask); - candidate = base_ip + table[hash]; + uint16_t* table_entry = TableEntry(table, dword, mask); + candidate = base_ip + *table_entry; assert(candidate >= base_ip); assert(candidate < ip + i); - table[hash] = delta + i; + *table_entry = delta + i; if (SNAPPY_PREDICT_FALSE(LittleEndian::Load32(candidate) == dword)) { *op = LITERAL | (i << 2); UnalignedCopy128(next_emit, op + 1); @@ -806,7 +865,7 @@ char* CompressFragment(const char* input, size_t input_size, char* op, } while (true) { assert(static_cast<uint32_t>(data) == LittleEndian::Load32(ip)); - uint32_t hash = HashBytes(data, mask); + 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; @@ -814,11 +873,11 @@ char* CompressFragment(const char* input, size_t input_size, char* op, ip = next_emit; goto emit_remainder; } - candidate = base_ip + table[hash]; + candidate = base_ip + *table_entry; assert(candidate >= base_ip); assert(candidate < ip); - table[hash] = ip - base_ip; + *table_entry = ip - base_ip; if (SNAPPY_PREDICT_FALSE(static_cast<uint32_t>(data) == LittleEndian::Load32(candidate))) { break; @@ -864,12 +923,13 @@ char* CompressFragment(const char* input, size_t input_size, char* op, assert((data & 0xFFFFFFFFFF) == (LittleEndian::Load64(ip) & 0xFFFFFFFFFF)); // We are now looking for a 4-byte match again. We read - // table[Hash(ip, shift)] for that. To improve compression, + // table[Hash(ip, mask)] for that. To improve compression, // we also update table[Hash(ip - 1, mask)] and table[Hash(ip, mask)]. - table[HashBytes(LittleEndian::Load32(ip - 1), mask)] = ip - base_ip - 1; - uint32_t hash = HashBytes(data, mask); - candidate = base_ip + table[hash]; - table[hash] = ip - base_ip; + *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 @@ -895,12 +955,180 @@ emit_remainder: 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 -// Called back at avery compression call to trace parameters and sizes. -static inline void Report(const char *algorithm, size_t compressed_size, - size_t uncompressed_size) { +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; @@ -962,7 +1190,7 @@ static inline void Report(const char *algorithm, size_t compressed_size, // bool TryFastAppend(const char* ip, size_t available, size_t length, T* op); // }; -static inline uint32_t ExtractLowBytes(uint32_t v, int n) { +static inline uint32_t ExtractLowBytes(const uint32_t& v, int n) { assert(n >= 0); assert(n <= 4); #if SNAPPY_HAVE_BMI2 @@ -991,30 +1219,87 @@ inline bool Copy64BytesWithPatternExtension(ptrdiff_t dst, size_t offset) { return offset != 0; } -void MemCopy(char* dst, const uint8_t* src, size_t size) { - std::memcpy(dst, src, size); +// 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 MemCopy(ptrdiff_t dst, const uint8_t* src, size_t size) { +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 MemMove(char* dst, const void* src, size_t size) { - std::memmove(dst, src, size); +void ClearDeferred(const void** deferred_src, size_t* deferred_length, + uint8_t* safe_source) { + *deferred_src = safe_source; + *deferred_length = 0; } -void MemMove(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 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 -size_t AdvanceToNextTag(const uint8_t** ip_p, size_t* tag) { +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 @@ -1026,11 +1311,12 @@ size_t AdvanceToNextTag(const uint8_t** ip_p, size_t* tag) { size_t literal_len = *tag >> 2; size_t tag_type = *tag; bool is_literal; -#if defined(__GNUC__) && defined(__x86_64__) && defined(__GCC_ASM_FLAG_OUTPUTS__) +#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)); + : [tag_type] "+r"(tag_type), "=@ccz"(is_literal) + :: "cc"); #else tag_type &= 3; is_literal = (tag_type == 0); @@ -1060,7 +1346,24 @@ size_t AdvanceToNextTag(const uint8_t** ip_p, size_t* tag) { // 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) { - return val & table.extract_masks[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. @@ -1076,6 +1379,12 @@ 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) { @@ -1084,21 +1393,41 @@ std::pair<const uint8_t*, ptrdiff_t> DecompressBranchless( // 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_min_offset = table.length_minus_offset[tag]; - size_t tag_type = AdvanceToNextTag(&ip, &tag); - uint32_t next = LittleEndian::Load32(old_ip); - size_t len = len_min_offset & 0xFF; - len_min_offset -= ExtractOffset(next, tag_type); - if (SNAPPY_PREDICT_FALSE(len_min_offset > 0)) { + 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 @@ -1110,39 +1439,29 @@ std::pair<const uint8_t*, ptrdiff_t> DecompressBranchless( } // Only copy-1 or copy-2 tags can get here. assert(tag_type == 1 || tag_type == 2); - std::ptrdiff_t delta = op + len_min_offset - len; + 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 + len_min_offset - len; + std::ptrdiff_t delta = (op + deferred_length) + len_min_offset - len; if (SNAPPY_PREDICT_FALSE(delta < 0)) { -#if defined(__GNUC__) && defined(__x86_64__) - // TODO - // When validating, both code path reduced to `op += len`. Ie. this - // becomes effectively - // - // if (delta < 0) if (tag_type != 0) goto break_loop; - // op += len; - // - // The compiler interchanges the predictable and almost always false - // first if-statement with the completely unpredictable second - // if-statement, putting an unpredictable branch on every iteration. - // This empty asm is worth almost 2x, which I think qualifies for an - // award for the most load-bearing empty statement. - asm(""); -#endif - // 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; - MemCopy(op_base + op, old_ip, 64); - op += len; + MemCopy64(op_base + op, deferred_src, deferred_length); + op += deferred_length; + DeferMemCopy(&deferred_src, &deferred_length, old_ip, len); continue; } @@ -1150,14 +1469,23 @@ std::pair<const uint8_t*, ptrdiff_t> DecompressBranchless( // we need to copy from ip instead of from the stream. const void* from = tag_type ? reinterpret_cast<void*>(op_base + delta) : old_ip; - MemMove(op_base + op, from, 64); - op += len; + MemCopy64(op_base + op, deferred_src, deferred_length); + op += deferred_length; + DeferMemCopy(&deferred_src, &deferred_length, from, len); } - } while (ip < ip_limit_min_slop && op < op_limit_min_slop); + } 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}; } @@ -1325,7 +1653,7 @@ class SnappyDecompressor { if (!writer->AppendFromSelf(copy_offset, length, &op)) goto exit; } else { - const ptrdiff_t entry = table.length_minus_offset[c]; + const ptrdiff_t entry = kLengthMinusOffset[c]; preload = LittleEndian::Load32(ip); const uint32_t trailer = ExtractLowBytes(preload, c & 3); const uint32_t length = entry & 0xff; @@ -1448,7 +1776,8 @@ template <typename Writer> static bool InternalUncompressAllTags(SnappyDecompressor* decompressor, Writer* writer, uint32_t compressed_len, uint32_t uncompressed_len) { - Report("snappy_uncompress", compressed_len, uncompressed_len); + int token = 0; + Report(token, "snappy_uncompress", compressed_len, uncompressed_len); writer->SetExpectedLength(uncompressed_len); @@ -1463,7 +1792,9 @@ bool GetUncompressedLength(Source* source, uint32_t* result) { return decompressor.ReadUncompressedLength(result); } -size_t Compress(Source* reader, Sink* writer) { +size_t Compress(Source* reader, Sink* writer, CompressionOptions options) { + assert(options.level == 1 || options.level == 2); + int token = 0; size_t written = 0; size_t N = reader->Available(); const size_t uncompressed_size = N; @@ -1510,17 +1841,23 @@ size_t Compress(Source* reader, Sink* writer) { uint16_t* table = wmem.GetHashTable(num_to_read, &table_size); // Compress input_fragment and append to dest - const int max_output = MaxCompressedLength(num_to_read); - - // Need a scratch buffer for the output, in case the byte sink doesn't - // have room for us directly. + int max_output = MaxCompressedLength(num_to_read); // Since we encode kBlockSize regions followed by a region // which is <= kBlockSize in length, a previously allocated // scratch_output[] region is big enough for this iteration. + // Need a scratch buffer for the output, in case the byte sink doesn't + // have room for us directly. char* dest = writer->GetAppendBuffer(max_output, wmem.GetScratchOutput()); - char* end = internal::CompressFragment(fragment, fragment_size, dest, table, - table_size); + char* end = nullptr; + if (options.level == 1) { + end = internal::CompressFragment(fragment, fragment_size, dest, table, + table_size); + } else if (options.level == 2) { + end = internal::CompressFragmentDoubleHash( + fragment, fragment_size, dest, table, table_size >> 1, + table + (table_size >> 1), table_size >> 1); + } writer->Append(dest, end - dest); written += (end - dest); @@ -1528,8 +1865,7 @@ size_t Compress(Source* reader, Sink* writer) { reader->Skip(pending_advance); } - Report("snappy_compress", written, uncompressed_size); - + Report(token, "snappy_compress", written, uncompressed_size); return written; } @@ -1537,6 +1873,67 @@ size_t Compress(Source* reader, Sink* writer) { // IOVec interfaces // ----------------------------------------------------------------------- +// A `Source` implementation that yields the contents of an `iovec` array. Note +// that `total_size` is the total number of bytes to be read from the elements +// of `iov` (_not_ the total number of elements in `iov`). +class SnappyIOVecReader : public Source { + public: + SnappyIOVecReader(const struct iovec* iov, size_t total_size) + : curr_iov_(iov), + curr_pos_(total_size > 0 ? reinterpret_cast<const char*>(iov->iov_base) + : nullptr), + curr_size_remaining_(total_size > 0 ? iov->iov_len : 0), + total_size_remaining_(total_size) { + // Skip empty leading `iovec`s. + if (total_size > 0 && curr_size_remaining_ == 0) Advance(); + } + + ~SnappyIOVecReader() override = default; + + size_t Available() const override { return total_size_remaining_; } + + const char* Peek(size_t* len) override { + *len = curr_size_remaining_; + return curr_pos_; + } + + void Skip(size_t n) override { + while (n >= curr_size_remaining_ && n > 0) { + n -= curr_size_remaining_; + Advance(); + } + curr_size_remaining_ -= n; + total_size_remaining_ -= n; + curr_pos_ += n; + } + + private: + // Advances to the next nonempty `iovec` and updates related variables. + void Advance() { + do { + assert(total_size_remaining_ >= curr_size_remaining_); + total_size_remaining_ -= curr_size_remaining_; + if (total_size_remaining_ == 0) { + curr_pos_ = nullptr; + curr_size_remaining_ = 0; + return; + } + ++curr_iov_; + curr_pos_ = reinterpret_cast<const char*>(curr_iov_->iov_base); + curr_size_remaining_ = curr_iov_->iov_len; + } while (curr_size_remaining_ == 0); + } + + // The `iovec` currently being read. + const struct iovec* curr_iov_; + // The location in `curr_iov_` currently being read. + const char* curr_pos_; + // The amount of unread data in `curr_iov_`. + size_t curr_size_remaining_; + // The amount of unread data in the entire input array. + size_t total_size_remaining_; +}; + // A type that writes to an iovec. // Note that this is not a "ByteSink", but a type that matches the // Writer template argument to SnappyDecompressor::DecompressAllTags(). @@ -1902,24 +2299,54 @@ bool IsValidCompressed(Source* compressed) { } void RawCompress(const char* input, size_t input_length, char* compressed, - size_t* compressed_length) { + size_t* compressed_length, CompressionOptions options) { ByteArraySource reader(input, input_length); UncheckedByteArraySink writer(compressed); - Compress(&reader, &writer); + Compress(&reader, &writer, options); // Compute how many bytes were added *compressed_length = (writer.CurrentDestination() - compressed); } -size_t Compress(const char* input, size_t input_length, - std::string* compressed) { +void RawCompressFromIOVec(const struct iovec* iov, size_t uncompressed_length, + char* compressed, size_t* compressed_length, + CompressionOptions options) { + SnappyIOVecReader reader(iov, uncompressed_length); + UncheckedByteArraySink writer(compressed); + Compress(&reader, &writer, options); + + // Compute how many bytes were added. + *compressed_length = writer.CurrentDestination() - compressed; +} + +size_t Compress(const char* input, size_t input_length, std::string* compressed, + CompressionOptions options) { // Pre-grow the buffer to the max length of the compressed output STLStringResizeUninitialized(compressed, MaxCompressedLength(input_length)); size_t compressed_length; RawCompress(input, input_length, string_as_array(compressed), - &compressed_length); - compressed->resize(compressed_length); + &compressed_length, options); + compressed->erase(compressed_length); + return compressed_length; +} + +size_t CompressFromIOVec(const struct iovec* iov, size_t iov_cnt, + std::string* compressed, CompressionOptions options) { + // Compute the number of bytes to be compressed. + size_t uncompressed_length = 0; + for (size_t i = 0; i < iov_cnt; ++i) { + uncompressed_length += iov[i].iov_len; + } + + // Pre-grow the buffer to the max length of the compressed output. + STLStringResizeUninitialized(compressed, MaxCompressedLength( + uncompressed_length)); + + size_t compressed_length; + RawCompressFromIOVec(iov, uncompressed_length, string_as_array(compressed), + &compressed_length, options); + compressed->erase(compressed_length); return compressed_length; } @@ -2108,7 +2535,6 @@ bool SnappyScatteredWriter<Allocator>::SlowAppendFromSelf(size_t offset, class SnappySinkAllocator { public: explicit SnappySinkAllocator(Sink* dest) : dest_(dest) {} - ~SnappySinkAllocator() {} char* Allocate(int size) { Datablock block(new char[size], size); diff --git a/other-licenses/snappy/src/snappy.h b/other-licenses/snappy/src/snappy.h index e4fdad3354..9e0572dcc8 100644 --- a/other-licenses/snappy/src/snappy.h +++ b/other-licenses/snappy/src/snappy.h @@ -50,13 +50,36 @@ namespace snappy { class Source; class Sink; + struct CompressionOptions { + // Compression level. + // Level 1 is the fastest + // Level 2 is a little slower but provides better compression. Level 2 is + // **EXPERIMENTAL** for the time being. It might happen that we decide to + // fall back to level 1 in the future. + // Levels 3+ are currently not supported. We plan to support levels up to + // 9 in the future. + // If you played with other compression algorithms, level 1 is equivalent to + // fast mode (level 1) of LZ4, level 2 is equivalent to LZ4's level 2 mode + // and compresses somewhere around zstd:-3 and zstd:-2 but generally with + // faster decompression speeds than snappy:1 and zstd:-3. + int level = DefaultCompressionLevel(); + + constexpr CompressionOptions() = default; + constexpr explicit CompressionOptions(int compression_level) + : level(compression_level) {} + static constexpr int MinCompressionLevel() { return 1; } + static constexpr int MaxCompressionLevel() { return 2; } + static constexpr int DefaultCompressionLevel() { return 1; } + }; + // ------------------------------------------------------------------------ // Generic compression/decompression routines. // ------------------------------------------------------------------------ - // Compress the bytes read from "*source" and append to "*sink". Return the + // Compress the bytes read from "*reader" and append to "*writer". Return the // number of bytes written. - size_t Compress(Source* source, Sink* sink); + size_t Compress(Source* reader, Sink* writer, + CompressionOptions options = {}); // Find the uncompressed length of the given stream, as given by the header. // Note that the true length could deviate from this; the stream could e.g. @@ -71,14 +94,22 @@ namespace snappy { // Higher-level string based routines (should be sufficient for most users) // ------------------------------------------------------------------------ - // Sets "*compressed" to the compressed version of "input[0,input_length-1]". + // Sets "*compressed" to the compressed version of "input[0..input_length-1]". // Original contents of *compressed are lost. // // REQUIRES: "input[]" is not an alias of "*compressed". size_t Compress(const char* input, size_t input_length, - std::string* compressed); + std::string* compressed, CompressionOptions options = {}); + + // Same as `Compress` above but taking an `iovec` array as input. Note that + // this function preprocesses the inputs to compute the sum of + // `iov[0..iov_cnt-1].iov_len` before reading. To avoid this, use + // `RawCompressFromIOVec` below. + size_t CompressFromIOVec(const struct iovec* iov, size_t iov_cnt, + std::string* compressed, + CompressionOptions options = {}); - // Decompresses "compressed[0,compressed_length-1]" to "*uncompressed". + // Decompresses "compressed[0..compressed_length-1]" to "*uncompressed". // Original contents of "*uncompressed" are lost. // // REQUIRES: "compressed[]" is not an alias of "*uncompressed". @@ -119,10 +150,15 @@ namespace snappy { // RawCompress(input, input_length, output, &output_length); // ... Process(output, output_length) ... // delete [] output; - void RawCompress(const char* input, - size_t input_length, - char* compressed, - size_t* compressed_length); + void RawCompress(const char* input, size_t input_length, char* compressed, + size_t* compressed_length, CompressionOptions options = {}); + + // Same as `RawCompress` above but taking an `iovec` array as input. Note that + // `uncompressed_length` is the total number of bytes to be read from the + // elements of `iov` (_not_ the number of elements in `iov`). + void RawCompressFromIOVec(const struct iovec* iov, size_t uncompressed_length, + char* compressed, size_t* compressed_length, + CompressionOptions options = {}); // Given data in "compressed[0..compressed_length-1]" generated by // calling the Snappy::Compress routine, this routine @@ -202,7 +238,7 @@ namespace snappy { static constexpr int kMinHashTableBits = 8; static constexpr size_t kMinHashTableSize = 1 << kMinHashTableBits; - static constexpr int kMaxHashTableBits = 14; + static constexpr int kMaxHashTableBits = 15; static constexpr size_t kMaxHashTableSize = 1 << kMaxHashTableBits; } // end namespace snappy diff --git a/other-licenses/snappy/src/snappy_compress_fuzzer.cc b/other-licenses/snappy/src/snappy_compress_fuzzer.cc index 1d4429a8c1..93254a28b1 100644 --- a/other-licenses/snappy/src/snappy_compress_fuzzer.cc +++ b/other-licenses/snappy/src/snappy_compress_fuzzer.cc @@ -39,22 +39,26 @@ // Entry point for LibFuzzer. extern "C" int LLVMFuzzerTestOneInput(const uint8_t* data, size_t size) { std::string input(reinterpret_cast<const char*>(data), size); + for (int level = snappy::CompressionOptions::MinCompressionLevel(); + level <= snappy::CompressionOptions::MaxCompressionLevel(); ++level) { + std::string compressed; + size_t compressed_size = + snappy::Compress(input.data(), input.size(), &compressed, + snappy::CompressionOptions{/*level=*/level}); - std::string compressed; - size_t compressed_size = - snappy::Compress(input.data(), input.size(), &compressed); + (void)compressed_size; // Variable only used in debug builds. + assert(compressed_size == compressed.size()); + assert(compressed.size() <= snappy::MaxCompressedLength(input.size())); + assert( + snappy::IsValidCompressedBuffer(compressed.data(), compressed.size())); - (void)compressed_size; // Variable only used in debug builds. - assert(compressed_size == compressed.size()); - assert(compressed.size() <= snappy::MaxCompressedLength(input.size())); - assert(snappy::IsValidCompressedBuffer(compressed.data(), compressed.size())); + std::string uncompressed_after_compress; + bool uncompress_succeeded = snappy::Uncompress( + compressed.data(), compressed.size(), &uncompressed_after_compress); - std::string uncompressed_after_compress; - bool uncompress_succeeded = snappy::Uncompress( - compressed.data(), compressed.size(), &uncompressed_after_compress); - - (void)uncompress_succeeded; // Variable only used in debug builds. - assert(uncompress_succeeded); - assert(input == uncompressed_after_compress); + (void)uncompress_succeeded; // Variable only used in debug builds. + assert(uncompress_succeeded); + assert(input == uncompressed_after_compress); + } return 0; } diff --git a/other-licenses/snappy/src/snappy_unittest.cc b/other-licenses/snappy/src/snappy_unittest.cc index 7a85635d73..e57b13d0e8 100644 --- a/other-licenses/snappy/src/snappy_unittest.cc +++ b/other-licenses/snappy/src/snappy_unittest.cc @@ -50,7 +50,7 @@ namespace snappy { namespace { -#if defined(HAVE_FUNC_MMAP) && defined(HAVE_FUNC_SYSCONF) +#if HAVE_FUNC_MMAP && HAVE_FUNC_SYSCONF // To test against code that reads beyond its input, this class copies a // string to a newly allocated group of pages, the last of which @@ -96,7 +96,7 @@ class DataEndingAtUnreadablePage { size_t size_; }; -#else // defined(HAVE_FUNC_MMAP) && defined(HAVE_FUNC_SYSCONF) +#else // HAVE_FUNC_MMAP) && HAVE_FUNC_SYSCONF // Fallback for systems without mmap. using DataEndingAtUnreadablePage = std::string; @@ -137,21 +137,10 @@ void VerifyStringSink(const std::string& input) { CHECK_EQ(uncompressed, input); } -void VerifyIOVec(const std::string& input) { - std::string compressed; - DataEndingAtUnreadablePage i(input); - const size_t written = snappy::Compress(i.data(), i.size(), &compressed); - CHECK_EQ(written, compressed.size()); - CHECK_LE(compressed.size(), - snappy::MaxCompressedLength(input.size())); - CHECK(snappy::IsValidCompressedBuffer(compressed.data(), compressed.size())); - - // Try uncompressing into an iovec containing a random number of entries - // ranging from 1 to 10. - char* buf = new char[input.size()]; +struct iovec* GetIOVec(const std::string& input, char*& buf, size_t& num) { std::minstd_rand0 rng(input.size()); std::uniform_int_distribution<size_t> uniform_1_to_10(1, 10); - size_t num = uniform_1_to_10(rng); + num = uniform_1_to_10(rng); if (input.size() < num) { num = input.size(); } @@ -175,8 +164,40 @@ void VerifyIOVec(const std::string& input) { } used_so_far += iov[i].iov_len; } - CHECK(snappy::RawUncompressToIOVec( - compressed.data(), compressed.size(), iov, num)); + return iov; +} + +int VerifyIOVecSource(const std::string& input) { + std::string compressed; + std::string copy = input; + char* buf = const_cast<char*>(copy.data()); + size_t num = 0; + struct iovec* iov = GetIOVec(input, buf, num); + const size_t written = snappy::CompressFromIOVec(iov, num, &compressed); + CHECK_EQ(written, compressed.size()); + CHECK_LE(compressed.size(), snappy::MaxCompressedLength(input.size())); + CHECK(snappy::IsValidCompressedBuffer(compressed.data(), compressed.size())); + + std::string uncompressed; + DataEndingAtUnreadablePage c(compressed); + CHECK(snappy::Uncompress(c.data(), c.size(), &uncompressed)); + CHECK_EQ(uncompressed, input); + delete[] iov; + return uncompressed.size(); +} + +void VerifyIOVecSink(const std::string& input) { + std::string compressed; + DataEndingAtUnreadablePage i(input); + const size_t written = snappy::Compress(i.data(), i.size(), &compressed); + CHECK_EQ(written, compressed.size()); + CHECK_LE(compressed.size(), snappy::MaxCompressedLength(input.size())); + CHECK(snappy::IsValidCompressedBuffer(compressed.data(), compressed.size())); + char* buf = new char[input.size()]; + size_t num = 0; + struct iovec* iov = GetIOVec(input, buf, num); + CHECK(snappy::RawUncompressToIOVec(compressed.data(), compressed.size(), iov, + num)); CHECK(!memcmp(buf, input.data(), input.size())); delete[] iov; delete[] buf; @@ -252,15 +273,18 @@ int Verify(const std::string& input) { // Compress using string based routines const int result = VerifyString(input); + // Compress using `iovec`-based routines. + CHECK_EQ(VerifyIOVecSource(input), result); + // Verify using sink based routines VerifyStringSink(input); VerifyNonBlockedCompression(input); - VerifyIOVec(input); + VerifyIOVecSink(input); if (!input.empty()) { const std::string expanded = Expand(input); VerifyNonBlockedCompression(expanded); - VerifyIOVec(input); + VerifyIOVecSink(input); } return result; @@ -540,7 +564,27 @@ TEST(Snappy, FourByteOffset) { CHECK_EQ(uncompressed, src); } -TEST(Snappy, IOVecEdgeCases) { +TEST(Snappy, IOVecSourceEdgeCases) { + // Validate that empty leading, trailing, and in-between iovecs are handled: + // [] [] ['a'] [] ['b'] []. + std::string data = "ab"; + char* buf = const_cast<char*>(data.data()); + size_t used_so_far = 0; + static const int kLengths[] = {0, 0, 1, 0, 1, 0}; + struct iovec iov[ARRAYSIZE(kLengths)]; + for (int i = 0; i < ARRAYSIZE(kLengths); ++i) { + iov[i].iov_base = buf + used_so_far; + iov[i].iov_len = kLengths[i]; + used_so_far += kLengths[i]; + } + std::string compressed; + snappy::CompressFromIOVec(iov, ARRAYSIZE(kLengths), &compressed); + std::string uncompressed; + snappy::Uncompress(compressed.data(), compressed.size(), &uncompressed); + CHECK_EQ(data, uncompressed); +} + +TEST(Snappy, IOVecSinkEdgeCases) { // Test some tricky edge cases in the iovec output that are not necessarily // exercised by random tests. @@ -905,7 +949,7 @@ TEST(Snappy, VerifyCharTable) { // COPY_1_BYTE_OFFSET. // // The tag byte in the compressed data stores len-4 in 3 bits, and - // offset/256 in 5 bits. offset%256 is stored in the next byte. + // offset/256 in 3 bits. offset%256 is stored in the next byte. // // This format is used for length in range [4..11] and offset in // range [0..2047] |