1 // SPDX-License-Identifier: 0BSD 2 3 /////////////////////////////////////////////////////////////////////////////// 4 // 5 /// \file memcmplen.h 6 /// \brief Optimized comparison of two buffers 7 // 8 // Author: Lasse Collin 9 // 10 /////////////////////////////////////////////////////////////////////////////// 11 12 #ifndef LZMA_MEMCMPLEN_H 13 #define LZMA_MEMCMPLEN_H 14 15 #include "common.h" 16 17 #ifdef HAVE_IMMINTRIN_H 18 # include <immintrin.h> 19 #endif 20 21 // Only include <intrin.h> if it is needed. The header is only needed 22 // on Windows when using an MSVC compatible compiler. The Intel compiler 23 // can use the intrinsics without the header file. 24 #if defined(TUKLIB_FAST_UNALIGNED_ACCESS) \ 25 && defined(_MSC_VER) \ 26 && (defined(_M_X64) \ 27 || defined(_M_ARM64) || defined(_M_ARM64EC)) \ 28 && !defined(__INTEL_COMPILER) 29 # include <intrin.h> 30 #endif 31 32 33 /// Find out how many equal bytes the two buffers have. 34 /// 35 /// \param buf1 First buffer 36 /// \param buf2 Second buffer 37 /// \param len How many bytes have already been compared and will 38 /// be assumed to match 39 /// \param limit How many bytes to compare at most, including the 40 /// already-compared bytes. This must be significantly 41 /// smaller than UINT32_MAX to avoid integer overflows. 42 /// Up to LZMA_MEMCMPLEN_EXTRA bytes may be read past 43 /// the specified limit from both buf1 and buf2. 44 /// 45 /// \return Number of equal bytes in the buffers is returned. 46 /// This is always at least len and at most limit. 47 /// 48 /// \note LZMA_MEMCMPLEN_EXTRA defines how many extra bytes may be read. 49 /// It's rounded up to 2^n. This extra amount needs to be 50 /// allocated in the buffers being used. It needs to be 51 /// initialized too to keep Valgrind quiet. 52 static lzma_always_inline uint32_t 53 lzma_memcmplen(const uint8_t *buf1, const uint8_t *buf2, 54 uint32_t len, uint32_t limit) 55 { 56 assert(len <= limit); 57 assert(limit <= UINT32_MAX / 2); 58 59 #if defined(TUKLIB_FAST_UNALIGNED_ACCESS) \ 60 && (((TUKLIB_GNUC_REQ(3, 4) || defined(__clang__)) \ 61 && (defined(__x86_64__) \ 62 || defined(__aarch64__))) \ 63 || (defined(__INTEL_COMPILER) && defined(__x86_64__)) \ 64 || (defined(__INTEL_COMPILER) && defined(_M_X64)) \ 65 || (defined(_MSC_VER) && (defined(_M_X64) \ 66 || defined(_M_ARM64) || defined(_M_ARM64EC)))) 67 // This is only for x86-64 and ARM64 for now. This might be fine on 68 // other 64-bit processors too. On big endian one should use xor 69 // instead of subtraction and switch to __builtin_clzll(). 70 // 71 // Reasons to use subtraction instead of xor: 72 // 73 // - On some x86-64 processors (Intel Sandy Bridge to Tiger Lake), 74 // sub+jz and sub+jnz can be fused but xor+jz or xor+jnz cannot. 75 // Thus using subtraction has potential to be a tiny amount faster 76 // since the code checks if the quotient is non-zero. 77 // 78 // - Some processors (Intel Pentium 4) used to have more ALU 79 // resources for add/sub instructions than and/or/xor. 80 // 81 // The processor info is based on Agner Fog's microarchitecture.pdf 82 // version 2023-05-26. https://www.agner.org/optimize/ 83 #define LZMA_MEMCMPLEN_EXTRA 8 84 while (len < limit) { 85 const uint64_t x = read64ne(buf1 + len) - read64ne(buf2 + len); 86 if (x != 0) { 87 // MSVC or Intel C compiler on Windows 88 # if defined(_MSC_VER) || defined(__INTEL_COMPILER) 89 unsigned long tmp; 90 _BitScanForward64(&tmp, x); 91 len += (uint32_t)tmp >> 3; 92 // GCC, Clang, or Intel C compiler 93 # else 94 len += (uint32_t)__builtin_ctzll(x) >> 3; 95 # endif 96 return my_min(len, limit); 97 } 98 99 len += 8; 100 } 101 102 return limit; 103 104 #elif defined(TUKLIB_FAST_UNALIGNED_ACCESS) \ 105 && defined(HAVE__MM_MOVEMASK_EPI8) \ 106 && (defined(__SSE2__) \ 107 || (defined(_MSC_VER) && defined(_M_IX86_FP) \ 108 && _M_IX86_FP >= 2)) 109 // NOTE: This will use 128-bit unaligned access which 110 // TUKLIB_FAST_UNALIGNED_ACCESS wasn't meant to permit, 111 // but it's convenient here since this is x86-only. 112 // 113 // SSE2 version for 32-bit and 64-bit x86. On x86-64 the above 114 // version is sometimes significantly faster and sometimes 115 // slightly slower than this SSE2 version, so this SSE2 116 // version isn't used on x86-64. 117 # define LZMA_MEMCMPLEN_EXTRA 16 118 while (len < limit) { 119 const uint32_t x = 0xFFFF ^ (uint32_t)_mm_movemask_epi8( 120 _mm_cmpeq_epi8( 121 _mm_loadu_si128((const __m128i *)(buf1 + len)), 122 _mm_loadu_si128((const __m128i *)(buf2 + len)))); 123 124 if (x != 0) { 125 len += ctz32(x); 126 return my_min(len, limit); 127 } 128 129 len += 16; 130 } 131 132 return limit; 133 134 #elif defined(TUKLIB_FAST_UNALIGNED_ACCESS) && !defined(WORDS_BIGENDIAN) 135 // Generic 32-bit little endian method 136 # define LZMA_MEMCMPLEN_EXTRA 4 137 while (len < limit) { 138 uint32_t x = read32ne(buf1 + len) - read32ne(buf2 + len); 139 if (x != 0) { 140 if ((x & 0xFFFF) == 0) { 141 len += 2; 142 x >>= 16; 143 } 144 145 if ((x & 0xFF) == 0) 146 ++len; 147 148 return my_min(len, limit); 149 } 150 151 len += 4; 152 } 153 154 return limit; 155 156 #elif defined(TUKLIB_FAST_UNALIGNED_ACCESS) && defined(WORDS_BIGENDIAN) 157 // Generic 32-bit big endian method 158 # define LZMA_MEMCMPLEN_EXTRA 4 159 while (len < limit) { 160 uint32_t x = read32ne(buf1 + len) ^ read32ne(buf2 + len); 161 if (x != 0) { 162 if ((x & 0xFFFF0000) == 0) { 163 len += 2; 164 x <<= 16; 165 } 166 167 if ((x & 0xFF000000) == 0) 168 ++len; 169 170 return my_min(len, limit); 171 } 172 173 len += 4; 174 } 175 176 return limit; 177 178 #else 179 // Simple portable version that doesn't use unaligned access. 180 # define LZMA_MEMCMPLEN_EXTRA 0 181 while (len < limit && buf1[len] == buf2[len]) 182 ++len; 183 184 return len; 185 #endif 186 } 187 188 #endif 189