diff options
author | Daniel Baumann <daniel.baumann@progress-linux.org> | 2024-04-27 21:12:04 +0000 |
---|---|---|
committer | Daniel Baumann <daniel.baumann@progress-linux.org> | 2024-04-27 21:12:04 +0000 |
commit | eac54b7c4aec25060d7bd856f7cdc290943d6aae (patch) | |
tree | 9a6d81c9f88df4698e746d63d14ddafeddd918b8 /src/liblzma/common/memcmplen.h | |
parent | Initial commit. (diff) | |
download | xz-utils-eac54b7c4aec25060d7bd856f7cdc290943d6aae.tar.xz xz-utils-eac54b7c4aec25060d7bd856f7cdc290943d6aae.zip |
Adding upstream version 5.4.1.upstream/5.4.1upstream
Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>
Diffstat (limited to 'src/liblzma/common/memcmplen.h')
-rw-r--r-- | src/liblzma/common/memcmplen.h | 160 |
1 files changed, 160 insertions, 0 deletions
diff --git a/src/liblzma/common/memcmplen.h b/src/liblzma/common/memcmplen.h new file mode 100644 index 0000000..5a481a0 --- /dev/null +++ b/src/liblzma/common/memcmplen.h @@ -0,0 +1,160 @@ +/////////////////////////////////////////////////////////////////////////////// +// +/// \file memcmplen.h +/// \brief Optimized comparison of two buffers +// +// Author: Lasse Collin +// +// This file has been put into the public domain. +// You can do whatever you want with this file. +// +/////////////////////////////////////////////////////////////////////////////// + +#ifndef LZMA_MEMCMPLEN_H +#define LZMA_MEMCMPLEN_H + +#include "common.h" + +#ifdef HAVE_IMMINTRIN_H +# include <immintrin.h> +#endif + + +/// Find out how many equal bytes the two buffers have. +/// +/// \param buf1 First buffer +/// \param buf2 Second buffer +/// \param len How many bytes have already been compared and will +/// be assumed to match +/// \param limit How many bytes to compare at most, including the +/// already-compared bytes. This must be significantly +/// smaller than UINT32_MAX to avoid integer overflows. +/// Up to LZMA_MEMCMPLEN_EXTRA bytes may be read past +/// the specified limit from both buf1 and buf2. +/// +/// \return Number of equal bytes in the buffers is returned. +/// This is always at least len and at most limit. +/// +/// \note LZMA_MEMCMPLEN_EXTRA defines how many extra bytes may be read. +/// It's rounded up to 2^n. This extra amount needs to be +/// allocated in the buffers being used. It needs to be +/// initialized too to keep Valgrind quiet. +static inline uint32_t lzma_attribute((__always_inline__)) +lzma_memcmplen(const uint8_t *buf1, const uint8_t *buf2, + uint32_t len, uint32_t limit) +{ + assert(len <= limit); + assert(limit <= UINT32_MAX / 2); + +#if defined(TUKLIB_FAST_UNALIGNED_ACCESS) \ + && ((TUKLIB_GNUC_REQ(3, 4) && defined(__x86_64__)) \ + || (defined(__INTEL_COMPILER) && defined(__x86_64__)) \ + || (defined(__INTEL_COMPILER) && defined(_M_X64)) \ + || (defined(_MSC_VER) && defined(_M_X64))) + // I keep this x86-64 only for now since that's where I know this + // to be a good method. This may be fine on other 64-bit CPUs too. + // On big endian one should use xor instead of subtraction and switch + // to __builtin_clzll(). +#define LZMA_MEMCMPLEN_EXTRA 8 + while (len < limit) { + const uint64_t x = read64ne(buf1 + len) - read64ne(buf2 + len); + if (x != 0) { +# if defined(_M_X64) // MSVC or Intel C compiler on Windows + unsigned long tmp; + _BitScanForward64(&tmp, x); + len += (uint32_t)tmp >> 3; +# else // GCC, clang, or Intel C compiler + len += (uint32_t)__builtin_ctzll(x) >> 3; +# endif + return my_min(len, limit); + } + + len += 8; + } + + return limit; + +#elif defined(TUKLIB_FAST_UNALIGNED_ACCESS) \ + && defined(HAVE__MM_MOVEMASK_EPI8) \ + && (defined(__SSE2__) \ + || (defined(_MSC_VER) && defined(_M_IX86_FP) \ + && _M_IX86_FP >= 2)) + // NOTE: This will use 128-bit unaligned access which + // TUKLIB_FAST_UNALIGNED_ACCESS wasn't meant to permit, + // but it's convenient here since this is x86-only. + // + // SSE2 version for 32-bit and 64-bit x86. On x86-64 the above + // version is sometimes significantly faster and sometimes + // slightly slower than this SSE2 version, so this SSE2 + // version isn't used on x86-64. +# define LZMA_MEMCMPLEN_EXTRA 16 + while (len < limit) { + const uint32_t x = 0xFFFF ^ _mm_movemask_epi8(_mm_cmpeq_epi8( + _mm_loadu_si128((const __m128i *)(buf1 + len)), + _mm_loadu_si128((const __m128i *)(buf2 + len)))); + + if (x != 0) { + len += ctz32(x); + return my_min(len, limit); + } + + len += 16; + } + + return limit; + +#elif defined(TUKLIB_FAST_UNALIGNED_ACCESS) && !defined(WORDS_BIGENDIAN) + // Generic 32-bit little endian method +# define LZMA_MEMCMPLEN_EXTRA 4 + while (len < limit) { + uint32_t x = read32ne(buf1 + len) - read32ne(buf2 + len); + if (x != 0) { + if ((x & 0xFFFF) == 0) { + len += 2; + x >>= 16; + } + + if ((x & 0xFF) == 0) + ++len; + + return my_min(len, limit); + } + + len += 4; + } + + return limit; + +#elif defined(TUKLIB_FAST_UNALIGNED_ACCESS) && defined(WORDS_BIGENDIAN) + // Generic 32-bit big endian method +# define LZMA_MEMCMPLEN_EXTRA 4 + while (len < limit) { + uint32_t x = read32ne(buf1 + len) ^ read32ne(buf2 + len); + if (x != 0) { + if ((x & 0xFFFF0000) == 0) { + len += 2; + x <<= 16; + } + + if ((x & 0xFF000000) == 0) + ++len; + + return my_min(len, limit); + } + + len += 4; + } + + return limit; + +#else + // Simple portable version that doesn't use unaligned access. +# define LZMA_MEMCMPLEN_EXTRA 0 + while (len < limit && buf1[len] == buf2[len]) + ++len; + + return len; +#endif +} + +#endif |