diff options
Diffstat (limited to '')
| -rw-r--r-- | src/liblzma/common/memcmplen.h | 37 |
1 files changed, 26 insertions, 11 deletions
diff --git a/src/liblzma/common/memcmplen.h b/src/liblzma/common/memcmplen.h index 99d9c51..394a485 100644 --- a/src/liblzma/common/memcmplen.h +++ b/src/liblzma/common/memcmplen.h @@ -1,3 +1,5 @@ +// SPDX-License-Identifier: 0BSD + /////////////////////////////////////////////////////////////////////////////// // /// \file memcmplen.h @@ -5,9 +7,6 @@ // // 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 @@ -24,7 +23,8 @@ // can use the intrinsics without the header file. #if defined(TUKLIB_FAST_UNALIGNED_ACCESS) \ && defined(_MSC_VER) \ - && defined(_M_X64) \ + && (defined(_M_X64) \ + || defined(_M_ARM64) || defined(_M_ARM64EC)) \ && !defined(__INTEL_COMPILER) # include <intrin.h> #endif @@ -57,20 +57,35 @@ lzma_memcmplen(const uint8_t *buf1, const uint8_t *buf2, assert(limit <= UINT32_MAX / 2); #if defined(TUKLIB_FAST_UNALIGNED_ACCESS) \ - && ((TUKLIB_GNUC_REQ(3, 4) && defined(__x86_64__)) \ + && (((TUKLIB_GNUC_REQ(3, 4) || defined(__clang__)) \ + && (defined(__x86_64__) \ + || defined(__aarch64__))) \ || (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(). + || (defined(_MSC_VER) && (defined(_M_X64) \ + || defined(_M_ARM64) || defined(_M_ARM64EC)))) + // This is only for x86-64 and ARM64 for now. This might be fine on + // other 64-bit processors too. On big endian one should use xor + // instead of subtraction and switch to __builtin_clzll(). + // + // Reasons to use subtraction instead of xor: + // + // - On some x86-64 processors (Intel Sandy Bridge to Tiger Lake), + // sub+jz and sub+jnz can be fused but xor+jz or xor+jnz cannot. + // Thus using subtraction has potential to be a tiny amount faster + // since the code checks if the quotient is non-zero. + // + // - Some processors (Intel Pentium 4) used to have more ALU + // resources for add/sub instructions than and/or/xor. + // + // The processor info is based on Agner Fog's microarchitecture.pdf + // version 2023-05-26. https://www.agner.org/optimize/ #define LZMA_MEMCMPLEN_EXTRA 8 while (len < limit) { const uint64_t x = read64ne(buf1 + len) - read64ne(buf2 + len); if (x != 0) { // MSVC or Intel C compiler on Windows -# if (defined(_MSC_VER) || defined(__INTEL_COMPILER)) && defined(_M_X64) +# if defined(_MSC_VER) || defined(__INTEL_COMPILER) unsigned long tmp; _BitScanForward64(&tmp, x); len += (uint32_t)tmp >> 3; |