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 #define LZMA_MEMCMPLEN_EXTRA 8 71 while (len < limit) { 72 const uint64_t x = read64ne(buf1 + len) - read64ne(buf2 + len); 73 if (x != 0) { 74 // MSVC or Intel C compiler on Windows 75 # if defined(_MSC_VER) || defined(__INTEL_COMPILER) 76 unsigned long tmp; 77 _BitScanForward64(&tmp, x); 78 len += (uint32_t)tmp >> 3; 79 // GCC, Clang, or Intel C compiler 80 # else 81 len += (uint32_t)__builtin_ctzll(x) >> 3; 82 # endif 83 return my_min(len, limit); 84 } 85 86 len += 8; 87 } 88 89 return limit; 90 91 #elif defined(TUKLIB_FAST_UNALIGNED_ACCESS) \ 92 && defined(HAVE__MM_MOVEMASK_EPI8) \ 93 && (defined(__SSE2__) \ 94 || (defined(_MSC_VER) && defined(_M_IX86_FP) \ 95 && _M_IX86_FP >= 2)) 96 // NOTE: This will use 128-bit unaligned access which 97 // TUKLIB_FAST_UNALIGNED_ACCESS wasn't meant to permit, 98 // but it's convenient here since this is x86-only. 99 // 100 // SSE2 version for 32-bit and 64-bit x86. On x86-64 the above 101 // version is sometimes significantly faster and sometimes 102 // slightly slower than this SSE2 version, so this SSE2 103 // version isn't used on x86-64. 104 # define LZMA_MEMCMPLEN_EXTRA 16 105 while (len < limit) { 106 const uint32_t x = 0xFFFF ^ (uint32_t)_mm_movemask_epi8( 107 _mm_cmpeq_epi8( 108 _mm_loadu_si128((const __m128i *)(buf1 + len)), 109 _mm_loadu_si128((const __m128i *)(buf2 + len)))); 110 111 if (x != 0) { 112 len += ctz32(x); 113 return my_min(len, limit); 114 } 115 116 len += 16; 117 } 118 119 return limit; 120 121 #elif defined(TUKLIB_FAST_UNALIGNED_ACCESS) && !defined(WORDS_BIGENDIAN) 122 // Generic 32-bit little endian method 123 # define LZMA_MEMCMPLEN_EXTRA 4 124 while (len < limit) { 125 uint32_t x = read32ne(buf1 + len) - read32ne(buf2 + len); 126 if (x != 0) { 127 if ((x & 0xFFFF) == 0) { 128 len += 2; 129 x >>= 16; 130 } 131 132 if ((x & 0xFF) == 0) 133 ++len; 134 135 return my_min(len, limit); 136 } 137 138 len += 4; 139 } 140 141 return limit; 142 143 #elif defined(TUKLIB_FAST_UNALIGNED_ACCESS) && defined(WORDS_BIGENDIAN) 144 // Generic 32-bit big endian method 145 # define LZMA_MEMCMPLEN_EXTRA 4 146 while (len < limit) { 147 uint32_t x = read32ne(buf1 + len) ^ read32ne(buf2 + len); 148 if (x != 0) { 149 if ((x & 0xFFFF0000) == 0) { 150 len += 2; 151 x <<= 16; 152 } 153 154 if ((x & 0xFF000000) == 0) 155 ++len; 156 157 return my_min(len, limit); 158 } 159 160 len += 4; 161 } 162 163 return limit; 164 165 #else 166 // Simple portable version that doesn't use unaligned access. 167 # define LZMA_MEMCMPLEN_EXTRA 0 168 while (len < limit && buf1[len] == buf2[len]) 169 ++len; 170 171 return len; 172 #endif 173 } 174 175 #endif 176