1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
|
/*
* Distributed under the Boost Software License, Version 1.0.
* (See accompanying file LICENSE_1_0.txt or copy at
* http://www.boost.org/LICENSE_1_0.txt)
*
* Copyright (c) 2020 Andrey Semashev
*/
/*!
* \file find_address_sse2.cpp
*
* This file contains SSE2 implementation of the \c find_address algorithm
*/
#include <boost/predef/architecture/x86.h>
#include <boost/atomic/detail/int_sizes.hpp>
#if BOOST_ARCH_X86 && defined(BOOST_ATOMIC_DETAIL_SIZEOF_POINTER) && (BOOST_ATOMIC_DETAIL_SIZEOF_POINTER == 8 || BOOST_ATOMIC_DETAIL_SIZEOF_POINTER == 4)
#include <cstddef>
#include <emmintrin.h>
#include <boost/cstdint.hpp>
#include <boost/atomic/detail/config.hpp>
#include <boost/atomic/detail/intptr.hpp>
#include "find_address.hpp"
#include "x86_vector_tools.hpp"
#include "bit_operation_tools.hpp"
#include <boost/atomic/detail/header.hpp>
namespace boost {
namespace atomics {
namespace detail {
#if BOOST_ATOMIC_DETAIL_SIZEOF_POINTER == 8
namespace {
BOOST_FORCEINLINE __m128i mm_pand_si128(__m128i mm1, __m128i mm2)
{
// As of 2020, gcc, clang and icc prefer to generate andps instead of pand if the surrounding
// instructions pertain to FP domain, even if we use the _mm_and_si128 intrinsic. In our
// algorithm implementation, the FP instruction happen to be shufps, which is not actually
// restricted to FP domain (it is actually implemented in a separate MMX EU in Pentium 4 or
// a shuffle EU in INT domain in Core 2; on AMD K8/K10 all SSE instructions are implemented in
// FADD, FMUL and FMISC EUs regardless of INT/FP data types, and shufps is implemented in FADD/FMUL).
// In other words, there should be no domain bypass penalty between shufps and pand.
//
// This would usually not pose a problem since andps and pand have the same latency and throughput
// on most architectures of that age (before SSE4.1). However, it is possible that a newer architecture
// runs the SSE2 code path (e.g. because some weird compiler doesn't support SSE4.1 or because
// a hypervisor blocks SSE4.1 detection), and there pand may have a better throughput. For example,
// Sandy Bridge can execute 3 pand instructions per cycle, but only one andps. For this reason
// we prefer to generate pand and not andps.
#if defined(__GNUC__)
__asm__("pand %1, %0\n\t" : "+x" (mm1) : "x" (mm2));
#else
mm1 = _mm_and_si128(mm1, mm2);
#endif
return mm1;
}
} // namespace
#endif // BOOST_ATOMIC_DETAIL_SIZEOF_POINTER == 8
//! SSE2 implementation of the \c find_address algorithm
std::size_t find_address_sse2(const volatile void* addr, const volatile void* const* addrs, std::size_t size)
{
#if BOOST_ATOMIC_DETAIL_SIZEOF_POINTER == 8
if (size < 12u)
return find_address_generic(addr, addrs, size);
const __m128i mm_addr = mm_set1_epiptr((uintptr_t)addr);
std::size_t pos = 0u;
const std::size_t n = (size + 1u) & ~static_cast< std::size_t >(1u);
for (std::size_t m = n & ~static_cast< std::size_t >(15u); pos < m; pos += 16u)
{
__m128i mm1 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos));
__m128i mm2 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos + 2u));
__m128i mm3 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos + 4u));
__m128i mm4 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos + 6u));
__m128i mm5 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos + 8u));
__m128i mm6 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos + 10u));
__m128i mm7 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos + 12u));
__m128i mm8 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos + 14u));
mm1 = _mm_cmpeq_epi32(mm1, mm_addr);
mm2 = _mm_cmpeq_epi32(mm2, mm_addr);
mm3 = _mm_cmpeq_epi32(mm3, mm_addr);
mm4 = _mm_cmpeq_epi32(mm4, mm_addr);
mm5 = _mm_cmpeq_epi32(mm5, mm_addr);
mm6 = _mm_cmpeq_epi32(mm6, mm_addr);
mm7 = _mm_cmpeq_epi32(mm7, mm_addr);
mm8 = _mm_cmpeq_epi32(mm8, mm_addr);
__m128i mm_mask1_lo = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(mm1), _mm_castsi128_ps(mm2), _MM_SHUFFLE(2, 0, 2, 0)));
__m128i mm_mask1_hi = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(mm1), _mm_castsi128_ps(mm2), _MM_SHUFFLE(3, 1, 3, 1)));
__m128i mm_mask2_lo = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(mm3), _mm_castsi128_ps(mm4), _MM_SHUFFLE(2, 0, 2, 0)));
__m128i mm_mask2_hi = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(mm3), _mm_castsi128_ps(mm4), _MM_SHUFFLE(3, 1, 3, 1)));
__m128i mm_mask3_lo = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(mm5), _mm_castsi128_ps(mm6), _MM_SHUFFLE(2, 0, 2, 0)));
__m128i mm_mask3_hi = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(mm5), _mm_castsi128_ps(mm6), _MM_SHUFFLE(3, 1, 3, 1)));
__m128i mm_mask4_lo = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(mm7), _mm_castsi128_ps(mm8), _MM_SHUFFLE(2, 0, 2, 0)));
__m128i mm_mask4_hi = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(mm7), _mm_castsi128_ps(mm8), _MM_SHUFFLE(3, 1, 3, 1)));
mm_mask1_lo = mm_pand_si128(mm_mask1_lo, mm_mask1_hi);
mm_mask2_lo = mm_pand_si128(mm_mask2_lo, mm_mask2_hi);
mm_mask3_lo = mm_pand_si128(mm_mask3_lo, mm_mask3_hi);
mm_mask4_lo = mm_pand_si128(mm_mask4_lo, mm_mask4_hi);
mm_mask1_lo = _mm_packs_epi32(mm_mask1_lo, mm_mask2_lo);
mm_mask3_lo = _mm_packs_epi32(mm_mask3_lo, mm_mask4_lo);
mm_mask1_lo = _mm_packs_epi16(mm_mask1_lo, mm_mask3_lo);
uint32_t mask = _mm_movemask_epi8(mm_mask1_lo);
if (mask)
{
pos += atomics::detail::count_trailing_zeros(mask);
goto done;
}
}
if ((n - pos) >= 8u)
{
__m128i mm1 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos));
__m128i mm2 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos + 2u));
__m128i mm3 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos + 4u));
__m128i mm4 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos + 6u));
mm1 = _mm_cmpeq_epi32(mm1, mm_addr);
mm2 = _mm_cmpeq_epi32(mm2, mm_addr);
mm3 = _mm_cmpeq_epi32(mm3, mm_addr);
mm4 = _mm_cmpeq_epi32(mm4, mm_addr);
__m128i mm_mask1_lo = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(mm1), _mm_castsi128_ps(mm2), _MM_SHUFFLE(2, 0, 2, 0)));
__m128i mm_mask1_hi = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(mm1), _mm_castsi128_ps(mm2), _MM_SHUFFLE(3, 1, 3, 1)));
__m128i mm_mask2_lo = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(mm3), _mm_castsi128_ps(mm4), _MM_SHUFFLE(2, 0, 2, 0)));
__m128i mm_mask2_hi = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(mm3), _mm_castsi128_ps(mm4), _MM_SHUFFLE(3, 1, 3, 1)));
mm_mask1_lo = mm_pand_si128(mm_mask1_lo, mm_mask1_hi);
mm_mask2_lo = mm_pand_si128(mm_mask2_lo, mm_mask2_hi);
mm_mask1_lo = _mm_packs_epi32(mm_mask1_lo, mm_mask2_lo);
uint32_t mask = _mm_movemask_epi8(mm_mask1_lo);
if (mask)
{
pos += atomics::detail::count_trailing_zeros(mask) / 2u;
goto done;
}
pos += 8u;
}
if ((n - pos) >= 4u)
{
__m128i mm1 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos));
__m128i mm2 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos + 2u));
mm1 = _mm_cmpeq_epi32(mm1, mm_addr);
mm2 = _mm_cmpeq_epi32(mm2, mm_addr);
__m128i mm_mask1_lo = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(mm1), _mm_castsi128_ps(mm2), _MM_SHUFFLE(2, 0, 2, 0)));
__m128i mm_mask1_hi = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(mm1), _mm_castsi128_ps(mm2), _MM_SHUFFLE(3, 1, 3, 1)));
mm_mask1_lo = mm_pand_si128(mm_mask1_lo, mm_mask1_hi);
uint32_t mask = _mm_movemask_ps(_mm_castsi128_ps(mm_mask1_lo));
if (mask)
{
pos += atomics::detail::count_trailing_zeros(mask);
goto done;
}
pos += 4u;
}
if (pos < n)
{
__m128i mm1 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos));
mm1 = _mm_cmpeq_epi32(mm1, mm_addr);
__m128i mm_mask = _mm_shuffle_epi32(mm1, _MM_SHUFFLE(2, 3, 0, 1));
mm_mask = mm_pand_si128(mm_mask, mm1);
uint32_t mask = _mm_movemask_pd(_mm_castsi128_pd(mm_mask));
if (mask)
{
pos += atomics::detail::count_trailing_zeros(mask);
goto done;
}
pos += 2u;
}
done:
return pos;
#else // BOOST_ATOMIC_DETAIL_SIZEOF_POINTER == 8
if (size < 10u)
return find_address_generic(addr, addrs, size);
const __m128i mm_addr = _mm_set1_epi32((uintptr_t)addr);
std::size_t pos = 0u;
const std::size_t n = (size + 3u) & ~static_cast< std::size_t >(3u);
for (std::size_t m = n & ~static_cast< std::size_t >(15u); pos < m; pos += 16u)
{
__m128i mm1 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos));
__m128i mm2 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos + 4u));
__m128i mm3 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos + 8u));
__m128i mm4 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos + 12u));
mm1 = _mm_cmpeq_epi32(mm1, mm_addr);
mm2 = _mm_cmpeq_epi32(mm2, mm_addr);
mm3 = _mm_cmpeq_epi32(mm3, mm_addr);
mm4 = _mm_cmpeq_epi32(mm4, mm_addr);
mm1 = _mm_packs_epi32(mm1, mm2);
mm3 = _mm_packs_epi32(mm3, mm4);
mm1 = _mm_packs_epi16(mm1, mm3);
uint32_t mask = _mm_movemask_epi8(mm1);
if (mask)
{
pos += atomics::detail::count_trailing_zeros(mask);
goto done;
}
}
if ((n - pos) >= 8u)
{
__m128i mm1 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos));
__m128i mm2 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos + 4u));
mm1 = _mm_cmpeq_epi32(mm1, mm_addr);
mm2 = _mm_cmpeq_epi32(mm2, mm_addr);
mm1 = _mm_packs_epi32(mm1, mm2);
uint32_t mask = _mm_movemask_epi8(mm1);
if (mask)
{
pos += atomics::detail::count_trailing_zeros(mask) / 2u;
goto done;
}
pos += 8u;
}
if (pos < n)
{
__m128i mm1 = _mm_load_si128(reinterpret_cast< const __m128i* >(addrs + pos));
mm1 = _mm_cmpeq_epi32(mm1, mm_addr);
uint32_t mask = _mm_movemask_ps(_mm_castsi128_ps(mm1));
if (mask)
{
pos += atomics::detail::count_trailing_zeros(mask);
goto done;
}
pos += 4u;
}
done:
return pos;
#endif // BOOST_ATOMIC_DETAIL_SIZEOF_POINTER == 8
}
} // namespace detail
} // namespace atomics
} // namespace boost
#include <boost/atomic/detail/footer.hpp>
#endif // BOOST_ARCH_X86 && defined(BOOST_ATOMIC_DETAIL_SIZEOF_POINTER) && (BOOST_ATOMIC_DETAIL_SIZEOF_POINTER == 8 || BOOST_ATOMIC_DETAIL_SIZEOF_POINTER == 4)
|
