xref: /freebsd/contrib/xz/src/liblzma/common/memcmplen.h (revision d9a42747950146bf03cda7f6e25d219253f8a57a)
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 
23 /// Find out how many equal bytes the two buffers have.
24 ///
25 /// \param      buf1    First buffer
26 /// \param      buf2    Second buffer
27 /// \param      len     How many bytes have already been compared and will
28 ///                     be assumed to match
29 /// \param      limit   How many bytes to compare at most, including the
30 ///                     already-compared bytes. This must be significantly
31 ///                     smaller than UINT32_MAX to avoid integer overflows.
32 ///                     Up to LZMA_MEMCMPLEN_EXTRA bytes may be read past
33 ///                     the specified limit from both buf1 and buf2.
34 ///
35 /// \return     Number of equal bytes in the buffers is returned.
36 ///             This is always at least len and at most limit.
37 ///
38 /// \note       LZMA_MEMCMPLEN_EXTRA defines how many extra bytes may be read.
39 ///             It's rounded up to 2^n. This extra amount needs to be
40 ///             allocated in the buffers being used. It needs to be
41 ///             initialized too to keep Valgrind quiet.
42 static inline uint32_t lzma_attribute((__always_inline__))
43 lzma_memcmplen(const uint8_t *buf1, const uint8_t *buf2,
44 		uint32_t len, uint32_t limit)
45 {
46 	assert(len <= limit);
47 	assert(limit <= UINT32_MAX / 2);
48 
49 #if defined(TUKLIB_FAST_UNALIGNED_ACCESS) \
50 		&& ((TUKLIB_GNUC_REQ(3, 4) && defined(__x86_64__)) \
51 			|| (defined(__INTEL_COMPILER) && defined(__x86_64__)) \
52 			|| (defined(__INTEL_COMPILER) && defined(_M_X64)) \
53 			|| (defined(_MSC_VER) && defined(_M_X64)))
54 	// I keep this x86-64 only for now since that's where I know this
55 	// to be a good method. This may be fine on other 64-bit CPUs too.
56 	// On big endian one should use xor instead of subtraction and switch
57 	// to __builtin_clzll().
58 #define LZMA_MEMCMPLEN_EXTRA 8
59 	while (len < limit) {
60 		const uint64_t x = read64ne(buf1 + len) - read64ne(buf2 + len);
61 		if (x != 0) {
62 #	if defined(_M_X64) // MSVC or Intel C compiler on Windows
63 			unsigned long tmp;
64 			_BitScanForward64(&tmp, x);
65 			len += (uint32_t)tmp >> 3;
66 #	else // GCC, clang, or Intel C compiler
67 			len += (uint32_t)__builtin_ctzll(x) >> 3;
68 #	endif
69 			return my_min(len, limit);
70 		}
71 
72 		len += 8;
73 	}
74 
75 	return limit;
76 
77 #elif defined(TUKLIB_FAST_UNALIGNED_ACCESS) \
78 		&& defined(HAVE__MM_MOVEMASK_EPI8) \
79 		&& (defined(__SSE2__) \
80 			|| (defined(_MSC_VER) && defined(_M_IX86_FP) \
81 				&& _M_IX86_FP >= 2))
82 	// NOTE: This will use 128-bit unaligned access which
83 	// TUKLIB_FAST_UNALIGNED_ACCESS wasn't meant to permit,
84 	// but it's convenient here since this is x86-only.
85 	//
86 	// SSE2 version for 32-bit and 64-bit x86. On x86-64 the above
87 	// version is sometimes significantly faster and sometimes
88 	// slightly slower than this SSE2 version, so this SSE2
89 	// version isn't used on x86-64.
90 #	define LZMA_MEMCMPLEN_EXTRA 16
91 	while (len < limit) {
92 		const uint32_t x = 0xFFFF ^ _mm_movemask_epi8(_mm_cmpeq_epi8(
93 			_mm_loadu_si128((const __m128i *)(buf1 + len)),
94 			_mm_loadu_si128((const __m128i *)(buf2 + len))));
95 
96 		if (x != 0) {
97 			len += ctz32(x);
98 			return my_min(len, limit);
99 		}
100 
101 		len += 16;
102 	}
103 
104 	return limit;
105 
106 #elif defined(TUKLIB_FAST_UNALIGNED_ACCESS) && !defined(WORDS_BIGENDIAN)
107 	// Generic 32-bit little endian method
108 #	define LZMA_MEMCMPLEN_EXTRA 4
109 	while (len < limit) {
110 		uint32_t x = read32ne(buf1 + len) - read32ne(buf2 + len);
111 		if (x != 0) {
112 			if ((x & 0xFFFF) == 0) {
113 				len += 2;
114 				x >>= 16;
115 			}
116 
117 			if ((x & 0xFF) == 0)
118 				++len;
119 
120 			return my_min(len, limit);
121 		}
122 
123 		len += 4;
124 	}
125 
126 	return limit;
127 
128 #elif defined(TUKLIB_FAST_UNALIGNED_ACCESS) && defined(WORDS_BIGENDIAN)
129 	// Generic 32-bit big endian method
130 #	define LZMA_MEMCMPLEN_EXTRA 4
131 	while (len < limit) {
132 		uint32_t x = read32ne(buf1 + len) ^ read32ne(buf2 + len);
133 		if (x != 0) {
134 			if ((x & 0xFFFF0000) == 0) {
135 				len += 2;
136 				x <<= 16;
137 			}
138 
139 			if ((x & 0xFF000000) == 0)
140 				++len;
141 
142 			return my_min(len, limit);
143 		}
144 
145 		len += 4;
146 	}
147 
148 	return limit;
149 
150 #else
151 	// Simple portable version that doesn't use unaligned access.
152 #	define LZMA_MEMCMPLEN_EXTRA 0
153 	while (len < limit && buf1[len] == buf2[len])
154 		++len;
155 
156 	return len;
157 #endif
158 }
159 
160 #endif
161