xref: /freebsd/contrib/xz/src/liblzma/common/memcmplen.h (revision 43e29d03f416d7dda52112a29600a7c82ee1a91e)
1 ///////////////////////////////////////////////////////////////////////////////
2 //
3 /// \file       memcmplen.h
4 /// \brief      Optimized comparison of two buffers
5 //
6 //  Author:     Lasse Collin
7 //
8 //  This file has been put into the public domain.
9 //  You can do whatever you want with this file.
10 //
11 ///////////////////////////////////////////////////////////////////////////////
12 
13 #ifndef LZMA_MEMCMPLEN_H
14 #define LZMA_MEMCMPLEN_H
15 
16 #include "common.h"
17 
18 #ifdef HAVE_IMMINTRIN_H
19 #	include <immintrin.h>
20 #endif
21 
22 // Only include <intrin.h> if it is needed. The header is only needed
23 // on Windows when using an MSVC compatible compiler. The Intel compiler
24 // can use the intrinsics without the header file.
25 #if defined(TUKLIB_FAST_UNALIGNED_ACCESS) \
26 		&& (defined(_MSC_VER) \
27 		&& defined(_M_X64) \
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 inline uint32_t lzma_attribute((__always_inline__))
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(__x86_64__)) \
61 			|| (defined(__INTEL_COMPILER) && defined(__x86_64__)) \
62 			|| (defined(__INTEL_COMPILER) && defined(_M_X64)) \
63 			|| (defined(_MSC_VER) && defined(_M_X64)))
64 	// I keep this x86-64 only for now since that's where I know this
65 	// to be a good method. This may be fine on other 64-bit CPUs too.
66 	// On big endian one should use xor instead of subtraction and switch
67 	// to __builtin_clzll().
68 #define LZMA_MEMCMPLEN_EXTRA 8
69 	while (len < limit) {
70 		const uint64_t x = read64ne(buf1 + len) - read64ne(buf2 + len);
71 		if (x != 0) {
72 #	if defined(_M_X64) // MSVC or Intel C compiler on Windows
73 			unsigned long tmp;
74 			_BitScanForward64(&tmp, x);
75 			len += (uint32_t)tmp >> 3;
76 #	else // GCC, clang, or Intel C compiler
77 			len += (uint32_t)__builtin_ctzll(x) >> 3;
78 #	endif
79 			return my_min(len, limit);
80 		}
81 
82 		len += 8;
83 	}
84 
85 	return limit;
86 
87 #elif defined(TUKLIB_FAST_UNALIGNED_ACCESS) \
88 		&& defined(HAVE__MM_MOVEMASK_EPI8) \
89 		&& (defined(__SSE2__) \
90 			|| (defined(_MSC_VER) && defined(_M_IX86_FP) \
91 				&& _M_IX86_FP >= 2))
92 	// NOTE: This will use 128-bit unaligned access which
93 	// TUKLIB_FAST_UNALIGNED_ACCESS wasn't meant to permit,
94 	// but it's convenient here since this is x86-only.
95 	//
96 	// SSE2 version for 32-bit and 64-bit x86. On x86-64 the above
97 	// version is sometimes significantly faster and sometimes
98 	// slightly slower than this SSE2 version, so this SSE2
99 	// version isn't used on x86-64.
100 #	define LZMA_MEMCMPLEN_EXTRA 16
101 	while (len < limit) {
102 		const uint32_t x = 0xFFFF ^ (uint32_t)_mm_movemask_epi8(
103 			_mm_cmpeq_epi8(
104 			_mm_loadu_si128((const __m128i *)(buf1 + len)),
105 			_mm_loadu_si128((const __m128i *)(buf2 + len))));
106 
107 		if (x != 0) {
108 			len += ctz32(x);
109 			return my_min(len, limit);
110 		}
111 
112 		len += 16;
113 	}
114 
115 	return limit;
116 
117 #elif defined(TUKLIB_FAST_UNALIGNED_ACCESS) && !defined(WORDS_BIGENDIAN)
118 	// Generic 32-bit little endian method
119 #	define LZMA_MEMCMPLEN_EXTRA 4
120 	while (len < limit) {
121 		uint32_t x = read32ne(buf1 + len) - read32ne(buf2 + len);
122 		if (x != 0) {
123 			if ((x & 0xFFFF) == 0) {
124 				len += 2;
125 				x >>= 16;
126 			}
127 
128 			if ((x & 0xFF) == 0)
129 				++len;
130 
131 			return my_min(len, limit);
132 		}
133 
134 		len += 4;
135 	}
136 
137 	return limit;
138 
139 #elif defined(TUKLIB_FAST_UNALIGNED_ACCESS) && defined(WORDS_BIGENDIAN)
140 	// Generic 32-bit big endian method
141 #	define LZMA_MEMCMPLEN_EXTRA 4
142 	while (len < limit) {
143 		uint32_t x = read32ne(buf1 + len) ^ read32ne(buf2 + len);
144 		if (x != 0) {
145 			if ((x & 0xFFFF0000) == 0) {
146 				len += 2;
147 				x <<= 16;
148 			}
149 
150 			if ((x & 0xFF000000) == 0)
151 				++len;
152 
153 			return my_min(len, limit);
154 		}
155 
156 		len += 4;
157 	}
158 
159 	return limit;
160 
161 #else
162 	// Simple portable version that doesn't use unaligned access.
163 #	define LZMA_MEMCMPLEN_EXTRA 0
164 	while (len < limit && buf1[len] == buf2[len])
165 		++len;
166 
167 	return len;
168 #endif
169 }
170 
171 #endif
172