libstdc++
simd_x86.h
1// Simd x86 specific implementations -*- C++ -*-
2
3// Copyright (C) 2020-2024 Free Software Foundation, Inc.
4//
5// This file is part of the GNU ISO C++ Library. This library is free
6// software; you can redistribute it and/or modify it under the
7// terms of the GNU General Public License as published by the
8// Free Software Foundation; either version 3, or (at your option)
9// any later version.
10
11// This library is distributed in the hope that it will be useful,
12// but WITHOUT ANY WARRANTY; without even the implied warranty of
13// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14// GNU General Public License for more details.
15
16// Under Section 7 of GPL version 3, you are granted additional
17// permissions described in the GCC Runtime Library Exception, version
18// 3.1, as published by the Free Software Foundation.
19
20// You should have received a copy of the GNU General Public License and
21// a copy of the GCC Runtime Library Exception along with this program;
22// see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
23// <http://www.gnu.org/licenses/>.
24
25#ifndef _GLIBCXX_EXPERIMENTAL_SIMD_X86_H_
26#define _GLIBCXX_EXPERIMENTAL_SIMD_X86_H_
27
28#if __cplusplus >= 201703L
29
30#if !_GLIBCXX_SIMD_X86INTRIN
31#error \
32 "simd_x86.h may only be included when MMX or SSE on x86(_64) are available"
33#endif
34
35_GLIBCXX_SIMD_BEGIN_NAMESPACE
36
37// __to_masktype {{{
38// Given <T, N> return <__int_for_sizeof_t<T>, N>. For _SimdWrapper and
39// __vector_type_t.
40template <typename _Tp, size_t _Np>
41 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper<__int_for_sizeof_t<_Tp>, _Np>
42 __to_masktype(_SimdWrapper<_Tp, _Np> __x)
43 { return reinterpret_cast<__vector_type_t<__int_for_sizeof_t<_Tp>, _Np>>(__x._M_data); }
44
45template <typename _TV,
46 typename _TVT
47 = enable_if_t<__is_vector_type_v<_TV>, _VectorTraits<_TV>>,
48 typename _Up = __int_for_sizeof_t<typename _TVT::value_type>>
49 _GLIBCXX_SIMD_INTRINSIC constexpr __vector_type_t<_Up, _TVT::_S_full_size>
50 __to_masktype(_TV __x)
51 { return reinterpret_cast<__vector_type_t<_Up, _TVT::_S_full_size>>(__x); }
52
53// }}}
54// __interleave128_lo {{{
55template <typename _Ap, typename _Bp, typename _Tp = common_type_t<_Ap, _Bp>,
56 typename _Trait = _VectorTraits<_Tp>>
57 _GLIBCXX_SIMD_INTRINSIC constexpr _Tp
58 __interleave128_lo(const _Ap& __av, const _Bp& __bv)
59 {
60 const _Tp __a(__av);
61 const _Tp __b(__bv);
62 if constexpr (sizeof(_Tp) == 16 && _Trait::_S_full_size == 2)
63 return _Tp{__a[0], __b[0]};
64 else if constexpr (sizeof(_Tp) == 16 && _Trait::_S_full_size == 4)
65 return _Tp{__a[0], __b[0], __a[1], __b[1]};
66 else if constexpr (sizeof(_Tp) == 16 && _Trait::_S_full_size == 8)
67 return _Tp{__a[0], __b[0], __a[1], __b[1],
68 __a[2], __b[2], __a[3], __b[3]};
69 else if constexpr (sizeof(_Tp) == 16 && _Trait::_S_full_size == 16)
70 return _Tp{__a[0], __b[0], __a[1], __b[1], __a[2], __b[2],
71 __a[3], __b[3], __a[4], __b[4], __a[5], __b[5],
72 __a[6], __b[6], __a[7], __b[7]};
73 else if constexpr (sizeof(_Tp) == 32 && _Trait::_S_full_size == 4)
74 return _Tp{__a[0], __b[0], __a[2], __b[2]};
75 else if constexpr (sizeof(_Tp) == 32 && _Trait::_S_full_size == 8)
76 return _Tp{__a[0], __b[0], __a[1], __b[1],
77 __a[4], __b[4], __a[5], __b[5]};
78 else if constexpr (sizeof(_Tp) == 32 && _Trait::_S_full_size == 16)
79 return _Tp{__a[0], __b[0], __a[1], __b[1], __a[2], __b[2],
80 __a[3], __b[3], __a[8], __b[8], __a[9], __b[9],
81 __a[10], __b[10], __a[11], __b[11]};
82 else if constexpr (sizeof(_Tp) == 32 && _Trait::_S_full_size == 32)
83 return _Tp{__a[0], __b[0], __a[1], __b[1], __a[2], __b[2], __a[3],
84 __b[3], __a[4], __b[4], __a[5], __b[5], __a[6], __b[6],
85 __a[7], __b[7], __a[16], __b[16], __a[17], __b[17], __a[18],
86 __b[18], __a[19], __b[19], __a[20], __b[20], __a[21], __b[21],
87 __a[22], __b[22], __a[23], __b[23]};
88 else if constexpr (sizeof(_Tp) == 64 && _Trait::_S_full_size == 8)
89 return _Tp{__a[0], __b[0], __a[2], __b[2],
90 __a[4], __b[4], __a[6], __b[6]};
91 else if constexpr (sizeof(_Tp) == 64 && _Trait::_S_full_size == 16)
92 return _Tp{__a[0], __b[0], __a[1], __b[1], __a[4], __b[4],
93 __a[5], __b[5], __a[8], __b[8], __a[9], __b[9],
94 __a[12], __b[12], __a[13], __b[13]};
95 else if constexpr (sizeof(_Tp) == 64 && _Trait::_S_full_size == 32)
96 return _Tp{__a[0], __b[0], __a[1], __b[1], __a[2], __b[2], __a[3],
97 __b[3], __a[8], __b[8], __a[9], __b[9], __a[10], __b[10],
98 __a[11], __b[11], __a[16], __b[16], __a[17], __b[17], __a[18],
99 __b[18], __a[19], __b[19], __a[24], __b[24], __a[25], __b[25],
100 __a[26], __b[26], __a[27], __b[27]};
101 else if constexpr (sizeof(_Tp) == 64 && _Trait::_S_full_size == 64)
102 return _Tp{__a[0], __b[0], __a[1], __b[1], __a[2], __b[2], __a[3],
103 __b[3], __a[4], __b[4], __a[5], __b[5], __a[6], __b[6],
104 __a[7], __b[7], __a[16], __b[16], __a[17], __b[17], __a[18],
105 __b[18], __a[19], __b[19], __a[20], __b[20], __a[21], __b[21],
106 __a[22], __b[22], __a[23], __b[23], __a[32], __b[32], __a[33],
107 __b[33], __a[34], __b[34], __a[35], __b[35], __a[36], __b[36],
108 __a[37], __b[37], __a[38], __b[38], __a[39], __b[39], __a[48],
109 __b[48], __a[49], __b[49], __a[50], __b[50], __a[51], __b[51],
110 __a[52], __b[52], __a[53], __b[53], __a[54], __b[54], __a[55],
111 __b[55]};
112 else
113 __assert_unreachable<_Tp>();
114 }
115
116// }}}
117// __is_zero{{{
118template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
119 _GLIBCXX_SIMD_INTRINSIC constexpr bool
120 __is_zero(_Tp __a)
121 {
122 if (!__builtin_is_constant_evaluated())
123 {
124 if constexpr (__have_avx)
125 {
126 if constexpr (_TVT::template _S_is<float, 8>)
127 return _mm256_testz_ps(__a, __a);
128 else if constexpr (_TVT::template _S_is<double, 4>)
129 return _mm256_testz_pd(__a, __a);
130 else if constexpr (sizeof(_Tp) == 32)
131 return _mm256_testz_si256(__to_intrin(__a), __to_intrin(__a));
132 else if constexpr (_TVT::template _S_is<float>)
133 return _mm_testz_ps(__to_intrin(__a), __to_intrin(__a));
134 else if constexpr (_TVT::template _S_is<double, 2>)
135 return _mm_testz_pd(__a, __a);
136 else
137 return _mm_testz_si128(__to_intrin(__a), __to_intrin(__a));
138 }
139 else if constexpr (__have_sse4_1)
140 return _mm_testz_si128(__intrin_bitcast<__m128i>(__a),
141 __intrin_bitcast<__m128i>(__a));
142 }
143 else if constexpr (sizeof(_Tp) <= 8)
144 return reinterpret_cast<__int_for_sizeof_t<_Tp>>(__a) == 0;
145 else
146 {
147 const auto __b = __vector_bitcast<_LLong>(__a);
148 if constexpr (sizeof(__b) == 16)
149 return (__b[0] | __b[1]) == 0;
150 else if constexpr (sizeof(__b) == 32)
151 return __is_zero(__lo128(__b) | __hi128(__b));
152 else if constexpr (sizeof(__b) == 64)
153 return __is_zero(__lo256(__b) | __hi256(__b));
154 else
155 __assert_unreachable<_Tp>();
156 }
157 }
158
159// }}}
160// __movemask{{{
161template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
162 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_CONST int
163 __movemask(_Tp __a)
164 {
165 if constexpr (sizeof(_Tp) == 32)
166 {
167 if constexpr (_TVT::template _S_is<float>)
168 return _mm256_movemask_ps(__to_intrin(__a));
169 else if constexpr (_TVT::template _S_is<double>)
170 return _mm256_movemask_pd(__to_intrin(__a));
171 else
172 return _mm256_movemask_epi8(__to_intrin(__a));
173 }
174 else if constexpr (_TVT::template _S_is<float>)
175 return _mm_movemask_ps(__to_intrin(__a));
176 else if constexpr (_TVT::template _S_is<double>)
177 return _mm_movemask_pd(__to_intrin(__a));
178 else
179 return _mm_movemask_epi8(__to_intrin(__a));
180 }
181
182// }}}
183// __testz{{{
184template <typename _TI, typename _TVT = _VectorTraits<_TI>>
185 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_CONST constexpr int
186 __testz(_TI __a, _TI __b)
187 {
188 static_assert(is_same_v<_TI, __intrinsic_type_t<typename _TVT::value_type,
189 _TVT::_S_full_size>>);
190 if (!__builtin_is_constant_evaluated())
191 {
192 if constexpr (sizeof(_TI) == 32)
193 {
194 if constexpr (_TVT::template _S_is<float>)
195 return _mm256_testz_ps(__to_intrin(__a), __to_intrin(__b));
196 else if constexpr (_TVT::template _S_is<double>)
197 return _mm256_testz_pd(__to_intrin(__a), __to_intrin(__b));
198 else
199 return _mm256_testz_si256(__to_intrin(__a), __to_intrin(__b));
200 }
201 else if constexpr (_TVT::template _S_is<float> && __have_avx)
202 return _mm_testz_ps(__to_intrin(__a), __to_intrin(__b));
203 else if constexpr (_TVT::template _S_is<double> && __have_avx)
204 return _mm_testz_pd(__to_intrin(__a), __to_intrin(__b));
205 else if constexpr (__have_sse4_1)
206 return _mm_testz_si128(__intrin_bitcast<__m128i>(__to_intrin(__a)),
207 __intrin_bitcast<__m128i>(__to_intrin(__b)));
208 else
209 return __movemask(0 == __and(__a, __b)) != 0;
210 }
211 else
212 return __is_zero(__and(__a, __b));
213 }
214
215// }}}
216// __testc{{{
217// requires SSE4.1 or above
218template <typename _TI, typename _TVT = _VectorTraits<_TI>>
219 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_CONST constexpr int
220 __testc(_TI __a, _TI __b)
221 {
222 static_assert(is_same_v<_TI, __intrinsic_type_t<typename _TVT::value_type,
223 _TVT::_S_full_size>>);
224 if (__builtin_is_constant_evaluated())
225 return __is_zero(__andnot(__a, __b));
226
227 if constexpr (sizeof(_TI) == 32)
228 {
229 if constexpr (_TVT::template _S_is<float>)
230 return _mm256_testc_ps(__a, __b);
231 else if constexpr (_TVT::template _S_is<double>)
232 return _mm256_testc_pd(__a, __b);
233 else
234 return _mm256_testc_si256(__to_intrin(__a), __to_intrin(__b));
235 }
236 else if constexpr (_TVT::template _S_is<float> && __have_avx)
237 return _mm_testc_ps(__to_intrin(__a), __to_intrin(__b));
238 else if constexpr (_TVT::template _S_is<double> && __have_avx)
239 return _mm_testc_pd(__to_intrin(__a), __to_intrin(__b));
240 else
241 {
242 static_assert(is_same_v<_TI, _TI> && __have_sse4_1);
243 return _mm_testc_si128(__intrin_bitcast<__m128i>(__to_intrin(__a)),
244 __intrin_bitcast<__m128i>(__to_intrin(__b)));
245 }
246 }
247
248// }}}
249// __testnzc{{{
250template <typename _TI, typename _TVT = _VectorTraits<_TI>>
251 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_CONST constexpr int
252 __testnzc(_TI __a, _TI __b)
253 {
254 static_assert(is_same_v<_TI, __intrinsic_type_t<typename _TVT::value_type,
255 _TVT::_S_full_size>>);
256 if (!__builtin_is_constant_evaluated())
257 {
258 if constexpr (sizeof(_TI) == 32)
259 {
260 if constexpr (_TVT::template _S_is<float>)
261 return _mm256_testnzc_ps(__a, __b);
262 else if constexpr (_TVT::template _S_is<double>)
263 return _mm256_testnzc_pd(__a, __b);
264 else
265 return _mm256_testnzc_si256(__to_intrin(__a), __to_intrin(__b));
266 }
267 else if constexpr (_TVT::template _S_is<float> && __have_avx)
268 return _mm_testnzc_ps(__to_intrin(__a), __to_intrin(__b));
269 else if constexpr (_TVT::template _S_is<double> && __have_avx)
270 return _mm_testnzc_pd(__to_intrin(__a), __to_intrin(__b));
271 else if constexpr (__have_sse4_1)
272 return _mm_testnzc_si128(__intrin_bitcast<__m128i>(__to_intrin(__a)),
273 __intrin_bitcast<__m128i>(__to_intrin(__b)));
274 else
275 return __movemask(0 == __and(__a, __b)) == 0
276 && __movemask(0 == __andnot(__a, __b)) == 0;
277 }
278 else
279 return !(__is_zero(__and(__a, __b)) || __is_zero(__andnot(__a, __b)));
280 }
281
282// }}}
283// __xzyw{{{
284// shuffles the complete vector, swapping the inner two quarters. Often useful
285// for AVX for fixing up a shuffle result.
286template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
287 _GLIBCXX_SIMD_INTRINSIC _Tp
288 __xzyw(_Tp __a)
289 {
290 if constexpr (sizeof(_Tp) == 16)
291 {
292 const auto __x = __vector_bitcast<conditional_t<
293 is_floating_point_v<typename _TVT::value_type>, float, int>>(__a);
294 return reinterpret_cast<_Tp>(
295 decltype(__x){__x[0], __x[2], __x[1], __x[3]});
296 }
297 else if constexpr (sizeof(_Tp) == 32)
298 {
299 const auto __x = __vector_bitcast<conditional_t<
300 is_floating_point_v<typename _TVT::value_type>, double, _LLong>>(__a);
301 return reinterpret_cast<_Tp>(
302 decltype(__x){__x[0], __x[2], __x[1], __x[3]});
303 }
304 else if constexpr (sizeof(_Tp) == 64)
305 {
306 const auto __x = __vector_bitcast<conditional_t<
307 is_floating_point_v<typename _TVT::value_type>, double, _LLong>>(__a);
308 return reinterpret_cast<_Tp>(decltype(__x){__x[0], __x[1], __x[4],
309 __x[5], __x[2], __x[3],
310 __x[6], __x[7]});
311 }
312 else
313 __assert_unreachable<_Tp>();
314 }
315
316// }}}
317// __maskload_epi32{{{
318template <typename _Tp>
319 _GLIBCXX_SIMD_INTRINSIC auto
320 __maskload_epi32(const int* __ptr, _Tp __k)
321 {
322 if constexpr (sizeof(__k) == 16)
323 return _mm_maskload_epi32(__ptr, __k);
324 else
325 return _mm256_maskload_epi32(__ptr, __k);
326 }
327
328// }}}
329// __maskload_epi64{{{
330template <typename _Tp>
331 _GLIBCXX_SIMD_INTRINSIC auto
332 __maskload_epi64(const _LLong* __ptr, _Tp __k)
333 {
334 if constexpr (sizeof(__k) == 16)
335 return _mm_maskload_epi64(__ptr, __k);
336 else
337 return _mm256_maskload_epi64(__ptr, __k);
338 }
339
340// }}}
341// __maskload_ps{{{
342template <typename _Tp>
343 _GLIBCXX_SIMD_INTRINSIC auto
344 __maskload_ps(const float* __ptr, _Tp __k)
345 {
346 if constexpr (sizeof(__k) == 16)
347 return _mm_maskload_ps(__ptr, __k);
348 else
349 return _mm256_maskload_ps(__ptr, __k);
350 }
351
352// }}}
353// __maskload_pd{{{
354template <typename _Tp>
355 _GLIBCXX_SIMD_INTRINSIC auto
356 __maskload_pd(const double* __ptr, _Tp __k)
357 {
358 if constexpr (sizeof(__k) == 16)
359 return _mm_maskload_pd(__ptr, __k);
360 else
361 return _mm256_maskload_pd(__ptr, __k);
362 }
363
364// }}}
365
366#ifdef __clang__
367template <size_t _Np, typename _Tp, typename _Kp>
368 _GLIBCXX_SIMD_INTRINSIC constexpr auto
369 __movm(_Kp __k) noexcept
370 {
371 static_assert(is_unsigned_v<_Kp>);
372 if constexpr (sizeof(_Tp) == 1 && __have_avx512bw)
373 {
374 if constexpr (_Np <= 16 && __have_avx512vl)
375 return __builtin_ia32_cvtmask2b128(__k);
376 else if constexpr (_Np <= 32 && __have_avx512vl)
377 return __builtin_ia32_cvtmask2b256(__k);
378 else
379 return __builtin_ia32_cvtmask2b512(__k);
380 }
381 else if constexpr (sizeof(_Tp) == 2 && __have_avx512bw)
382 {
383 if constexpr (_Np <= 8 && __have_avx512vl)
384 return __builtin_ia32_cvtmask2w128(__k);
385 else if constexpr (_Np <= 16 && __have_avx512vl)
386 return __builtin_ia32_cvtmask2w256(__k);
387 else
388 return __builtin_ia32_cvtmask2w512(__k);
389 }
390 else if constexpr (sizeof(_Tp) == 4 && __have_avx512dq)
391 {
392 if constexpr (_Np <= 4 && __have_avx512vl)
393 return __builtin_ia32_cvtmask2d128(__k);
394 else if constexpr (_Np <= 8 && __have_avx512vl)
395 return __builtin_ia32_cvtmask2d256(__k);
396 else
397 return __builtin_ia32_cvtmask2d512(__k);
398 }
399 else if constexpr (sizeof(_Tp) == 8 && __have_avx512dq)
400 {
401 if constexpr (_Np <= 2 && __have_avx512vl)
402 return __builtin_ia32_cvtmask2q128(__k);
403 else if constexpr (_Np <= 4 && __have_avx512vl)
404 return __builtin_ia32_cvtmask2q256(__k);
405 else
406 return __builtin_ia32_cvtmask2q512(__k);
407 }
408 else
409 __assert_unreachable<_Tp>();
410 }
411#endif // __clang__
412
413#ifdef _GLIBCXX_SIMD_WORKAROUND_PR85048
414#include "simd_x86_conversions.h"
415#endif
416
417// ISA & type detection {{{
418template <typename _Tp, size_t _Np>
419 constexpr bool
420 __is_sse_ps()
421 {
422 return __have_sse
423 && is_same_v<_Tp,
424 float> && sizeof(__intrinsic_type_t<_Tp, _Np>) == 16;
425 }
426
427template <typename _Tp, size_t _Np>
428 constexpr bool
429 __is_sse_pd()
430 {
431 return __have_sse2
432 && is_same_v<_Tp,
433 double> && sizeof(__intrinsic_type_t<_Tp, _Np>) == 16;
434 }
435
436template <typename _Tp, size_t _Np>
437 constexpr bool
438 __is_avx_ps()
439 {
440 return __have_avx
441 && is_same_v<_Tp,
442 float> && sizeof(__intrinsic_type_t<_Tp, _Np>) == 32;
443 }
444
445template <typename _Tp, size_t _Np>
446 constexpr bool
447 __is_avx_pd()
448 {
449 return __have_avx
450 && is_same_v<_Tp,
451 double> && sizeof(__intrinsic_type_t<_Tp, _Np>) == 32;
452 }
453
454template <typename _Tp, size_t _Np>
455 constexpr bool
456 __is_avx512_ps()
457 {
458 return __have_avx512f
459 && is_same_v<_Tp,
460 float> && sizeof(__intrinsic_type_t<_Tp, _Np>) == 64;
461 }
462
463template <typename _Tp, size_t _Np>
464 constexpr bool
465 __is_avx512_pd()
466 {
467 return __have_avx512f
468 && is_same_v<_Tp,
469 double> && sizeof(__intrinsic_type_t<_Tp, _Np>) == 64;
470 }
471
472// }}}
473struct _MaskImplX86Mixin;
474
475// _CommonImplX86 {{{
476struct _CommonImplX86 : _CommonImplBuiltin
477{
478#ifdef _GLIBCXX_SIMD_WORKAROUND_PR85048
479 // _S_converts_via_decomposition {{{
480 template <typename _From, typename _To, size_t _ToSize>
481 static constexpr bool
482 _S_converts_via_decomposition()
483 {
484 if constexpr (is_integral_v<
485 _From> && is_integral_v<_To> && sizeof(_From) == 8
486 && _ToSize == 16)
487 return (sizeof(_To) == 2 && !__have_ssse3)
488 || (sizeof(_To) == 1 && !__have_avx512f);
489 else if constexpr (is_floating_point_v<_From> && is_integral_v<_To>)
490 return ((sizeof(_From) == 4 || sizeof(_From) == 8) && sizeof(_To) == 8
491 && !__have_avx512dq)
492 || (sizeof(_From) == 8 && sizeof(_To) == 4 && !__have_sse4_1
493 && _ToSize == 16);
494 else if constexpr (
495 is_integral_v<_From> && is_floating_point_v<_To> && sizeof(_From) == 8
496 && !__have_avx512dq)
497 return (sizeof(_To) == 4 && _ToSize == 16)
498 || (sizeof(_To) == 8 && _ToSize < 64);
499 else
500 return false;
501 }
502
503 template <typename _From, typename _To, size_t _ToSize>
504 static inline constexpr bool __converts_via_decomposition_v
505 = _S_converts_via_decomposition<_From, _To, _ToSize>();
506
507 // }}}
508#endif
509 // _S_store {{{
510 using _CommonImplBuiltin::_S_store;
511
512 template <typename _Tp, size_t _Np>
513 _GLIBCXX_SIMD_INTRINSIC static constexpr void
514 _S_store(_SimdWrapper<_Tp, _Np> __x, void* __addr)
515 {
516 constexpr size_t _Bytes = _Np * sizeof(_Tp);
517
518 if (__builtin_is_constant_evaluated())
519 _CommonImplBuiltin::_S_store(__x, __addr);
520 else if constexpr ((_Bytes & (_Bytes - 1)) != 0 && __have_avx512bw_vl)
521 {
522 const auto __v = __to_intrin(__x);
523
524 if constexpr (_Bytes & 1)
525 {
526 if constexpr (_Bytes < 16)
527 _mm_mask_storeu_epi8(__addr, 0xffffu >> (16 - _Bytes),
528 __intrin_bitcast<__m128i>(__v));
529 else if constexpr (_Bytes < 32)
530 _mm256_mask_storeu_epi8(__addr, 0xffffffffu >> (32 - _Bytes),
531 __intrin_bitcast<__m256i>(__v));
532 else
533 _mm512_mask_storeu_epi8(__addr,
534 0xffffffffffffffffull >> (64 - _Bytes),
535 __intrin_bitcast<__m512i>(__v));
536 }
537 else if constexpr (_Bytes & 2)
538 {
539 if constexpr (_Bytes < 16)
540 _mm_mask_storeu_epi16(__addr, 0xffu >> (8 - _Bytes / 2),
541 __intrin_bitcast<__m128i>(__v));
542 else if constexpr (_Bytes < 32)
543 _mm256_mask_storeu_epi16(__addr, 0xffffu >> (16 - _Bytes / 2),
544 __intrin_bitcast<__m256i>(__v));
545 else
546 _mm512_mask_storeu_epi16(__addr,
547 0xffffffffull >> (32 - _Bytes / 2),
548 __intrin_bitcast<__m512i>(__v));
549 }
550 else if constexpr (_Bytes & 4)
551 {
552 if constexpr (_Bytes < 16)
553 _mm_mask_storeu_epi32(__addr, 0xfu >> (4 - _Bytes / 4),
554 __intrin_bitcast<__m128i>(__v));
555 else if constexpr (_Bytes < 32)
556 _mm256_mask_storeu_epi32(__addr, 0xffu >> (8 - _Bytes / 4),
557 __intrin_bitcast<__m256i>(__v));
558 else
559 _mm512_mask_storeu_epi32(__addr, 0xffffull >> (16 - _Bytes / 4),
560 __intrin_bitcast<__m512i>(__v));
561 }
562 else
563 {
564 static_assert(
565 _Bytes > 16,
566 "_Bytes < 16 && (_Bytes & 7) == 0 && (_Bytes & (_Bytes "
567 "- 1)) != 0 is impossible");
568 if constexpr (_Bytes < 32)
569 _mm256_mask_storeu_epi64(__addr, 0xfu >> (4 - _Bytes / 8),
570 __intrin_bitcast<__m256i>(__v));
571 else
572 _mm512_mask_storeu_epi64(__addr, 0xffull >> (8 - _Bytes / 8),
573 __intrin_bitcast<__m512i>(__v));
574 }
575 }
576 else
577 _CommonImplBuiltin::_S_store(__x, __addr);
578 }
579
580 // }}}
581 // _S_store_bool_array(_BitMask) {{{
582 template <size_t _Np, bool _Sanitized>
583 _GLIBCXX_SIMD_INTRINSIC static constexpr void
584 _S_store_bool_array(const _BitMask<_Np, _Sanitized> __x, bool* __mem)
585 {
586 if (__builtin_is_constant_evaluated())
587 _CommonImplBuiltin::_S_store_bool_array(__x, __mem);
588 else if constexpr (__have_avx512bw_vl) // don't care for BW w/o VL
589 _S_store<_Np>(1 & __vector_bitcast<_UChar, _Np>(
590 [=]() constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
591 if constexpr (_Np <= 16)
592 return _mm_movm_epi8(__x._M_to_bits());
593 else if constexpr (_Np <= 32)
594 return _mm256_movm_epi8(__x._M_to_bits());
595 else if constexpr (_Np <= 64)
596 return _mm512_movm_epi8(__x._M_to_bits());
597 else
598 __assert_unreachable<_SizeConstant<_Np>>();
599 }()),
600 __mem);
601 else if constexpr (__have_bmi2)
602 {
603 if constexpr (_Np <= 4)
604 _S_store<_Np>(_pdep_u32(__x._M_to_bits(), 0x01010101U), __mem);
605 else
606 __execute_n_times<__div_roundup(_Np, sizeof(size_t))>(
607 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
608 constexpr size_t __offset = __i * sizeof(size_t);
609 constexpr int __todo = std::min(sizeof(size_t), _Np - __offset);
610 if constexpr (__todo == 1)
611 __mem[__offset] = __x[__offset];
612 else
613 {
614 const auto __bools =
615#ifdef __x86_64__
616 _pdep_u64(__x.template _M_extract<__offset>().to_ullong(),
617 0x0101010101010101ULL);
618#else // __x86_64__
619 _pdep_u32(
620 __x.template _M_extract<__offset>()._M_to_bits(),
621 0x01010101U);
622#endif // __x86_64__
623 _S_store<__todo>(__bools, __mem + __offset);
624 }
625 });
626 }
627 else if constexpr (__have_sse2 && _Np > 7)
628 __execute_n_times<__div_roundup(_Np, 16)>([&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
629 constexpr int __offset = __i * 16;
630 constexpr int __todo = std::min(16, int(_Np) - __offset);
631 const int __bits = __x.template _M_extract<__offset>()._M_to_bits();
632 __vector_type16_t<_UChar> __bools;
633 if constexpr (__have_avx512f)
634 {
635 auto __as32bits
636 = _mm512_maskz_mov_epi32(__bits, __to_intrin(
637 __vector_broadcast<16>(1)));
638 auto __as16bits
639 = __xzyw(_mm256_packs_epi32(__lo256(__as32bits),
640 __todo > 8 ? __hi256(__as32bits)
641 : __m256i()));
642 __bools = __vector_bitcast<_UChar>(
643 _mm_packs_epi16(__lo128(__as16bits), __hi128(__as16bits)));
644 }
645 else
646 {
647 using _V = __vector_type_t<_UChar, 16>;
648 auto __tmp = _mm_cvtsi32_si128(__bits);
649 __tmp = _mm_unpacklo_epi8(__tmp, __tmp);
650 __tmp = _mm_unpacklo_epi16(__tmp, __tmp);
651 __tmp = _mm_unpacklo_epi32(__tmp, __tmp);
652 _V __tmp2 = reinterpret_cast<_V>(__tmp);
653 __tmp2 &= _V{1, 2, 4, 8, 16, 32, 64, 128,
654 1, 2, 4, 8, 16, 32, 64, 128}; // mask bit index
655 __bools = (__tmp2 == 0) + 1; // 0xff -> 0x00 | 0x00 -> 0x01
656 }
657 _S_store<__todo>(__bools, __mem + __offset);
658 });
659 else
660 _CommonImplBuiltin::_S_store_bool_array(__x, __mem);
661 }
662
663 // }}}
664 // _S_blend_avx512 {{{
665 // Returns: __k ? __b : __a
666 // TODO: reverse __a and __b to match COND_EXPR
667 // Requires: _TV to be a __vector_type_t matching valuetype for the bitmask
668 // __k
669 template <typename _Kp, typename _TV>
670 _GLIBCXX_SIMD_INTRINSIC static _TV
671 _S_blend_avx512(const _Kp __k, const _TV __a, const _TV __b) noexcept
672 {
673 static_assert(__is_vector_type_v<_TV>);
674 using _Tp = typename _VectorTraits<_TV>::value_type;
675 static_assert(sizeof(_TV) >= 16);
676 static_assert(sizeof(_Tp) <= 8);
677#ifdef __clang__
678 return __movm<_VectorTraits<_TV>::_S_full_size, _Tp>(__k) ? __b : __a;
679#else
680 using _IntT
681 = conditional_t<(sizeof(_Tp) > 2),
682 conditional_t<sizeof(_Tp) == 4, int, long long>,
683 conditional_t<sizeof(_Tp) == 1, char, short>>;
684 [[maybe_unused]] const auto __aa = __vector_bitcast<_IntT>(__a);
685 [[maybe_unused]] const auto __bb = __vector_bitcast<_IntT>(__b);
686 if constexpr (sizeof(_TV) == 64)
687 {
688 if constexpr (sizeof(_Tp) == 1)
689 return reinterpret_cast<_TV>(
690 __builtin_ia32_blendmb_512_mask(__aa, __bb, __k));
691 else if constexpr (sizeof(_Tp) == 2)
692 return reinterpret_cast<_TV>(
693 __builtin_ia32_blendmw_512_mask(__aa, __bb, __k));
694 else if constexpr (sizeof(_Tp) == 4 && is_floating_point_v<_Tp>)
695 return __builtin_ia32_blendmps_512_mask(__a, __b, __k);
696 else if constexpr (sizeof(_Tp) == 4)
697 return reinterpret_cast<_TV>(
698 __builtin_ia32_blendmd_512_mask(__aa, __bb, __k));
699 else if constexpr (sizeof(_Tp) == 8 && is_floating_point_v<_Tp>)
700 return __builtin_ia32_blendmpd_512_mask(__a, __b, __k);
701 else if constexpr (sizeof(_Tp) == 8)
702 return reinterpret_cast<_TV>(
703 __builtin_ia32_blendmq_512_mask(__aa, __bb, __k));
704 }
705 else if constexpr (sizeof(_TV) == 32)
706 {
707 if constexpr (sizeof(_Tp) == 1)
708 return reinterpret_cast<_TV>(
709 __builtin_ia32_blendmb_256_mask(__aa, __bb, __k));
710 else if constexpr (sizeof(_Tp) == 2)
711 return reinterpret_cast<_TV>(
712 __builtin_ia32_blendmw_256_mask(__aa, __bb, __k));
713 else if constexpr (sizeof(_Tp) == 4 && is_floating_point_v<_Tp>)
714 return __builtin_ia32_blendmps_256_mask(__a, __b, __k);
715 else if constexpr (sizeof(_Tp) == 4)
716 return reinterpret_cast<_TV>(
717 __builtin_ia32_blendmd_256_mask(__aa, __bb, __k));
718 else if constexpr (sizeof(_Tp) == 8 && is_floating_point_v<_Tp>)
719 return __builtin_ia32_blendmpd_256_mask(__a, __b, __k);
720 else if constexpr (sizeof(_Tp) == 8)
721 return reinterpret_cast<_TV>(
722 __builtin_ia32_blendmq_256_mask(__aa, __bb, __k));
723 }
724 else if constexpr (sizeof(_TV) == 16)
725 {
726 if constexpr (sizeof(_Tp) == 1)
727 return reinterpret_cast<_TV>(
728 __builtin_ia32_blendmb_128_mask(__aa, __bb, __k));
729 else if constexpr (sizeof(_Tp) == 2)
730 return reinterpret_cast<_TV>(
731 __builtin_ia32_blendmw_128_mask(__aa, __bb, __k));
732 else if constexpr (sizeof(_Tp) == 4 && is_floating_point_v<_Tp>)
733 return __builtin_ia32_blendmps_128_mask(__a, __b, __k);
734 else if constexpr (sizeof(_Tp) == 4)
735 return reinterpret_cast<_TV>(
736 __builtin_ia32_blendmd_128_mask(__aa, __bb, __k));
737 else if constexpr (sizeof(_Tp) == 8 && is_floating_point_v<_Tp>)
738 return __builtin_ia32_blendmpd_128_mask(__a, __b, __k);
739 else if constexpr (sizeof(_Tp) == 8)
740 return reinterpret_cast<_TV>(
741 __builtin_ia32_blendmq_128_mask(__aa, __bb, __k));
742 }
743#endif
744 }
745
746 // }}}
747 // _S_blend_intrin {{{
748 // Returns: __k ? __b : __a
749 // TODO: reverse __a and __b to match COND_EXPR
750 // Requires: _Tp to be an intrinsic type (integers blend per byte) and 16/32
751 // Bytes wide
752 template <typename _Tp>
753 _GLIBCXX_SIMD_INTRINSIC static _Tp
754 _S_blend_intrin(_Tp __k, _Tp __a, _Tp __b) noexcept
755 {
756 static_assert(is_same_v<decltype(__to_intrin(__a)), _Tp>);
757 constexpr struct
758 {
759 _GLIBCXX_SIMD_INTRINSIC __m128 operator()(__m128 __a, __m128 __b,
760 __m128 __k) const noexcept
761 {
762 return __builtin_ia32_blendvps(__a, __b, __k);
763 }
764 _GLIBCXX_SIMD_INTRINSIC __m128d operator()(__m128d __a, __m128d __b,
765 __m128d __k) const noexcept
766 {
767 return __builtin_ia32_blendvpd(__a, __b, __k);
768 }
769 _GLIBCXX_SIMD_INTRINSIC __m128i operator()(__m128i __a, __m128i __b,
770 __m128i __k) const noexcept
771 {
772 return reinterpret_cast<__m128i>(
773 __builtin_ia32_pblendvb128(reinterpret_cast<__v16qi>(__a),
774 reinterpret_cast<__v16qi>(__b),
775 reinterpret_cast<__v16qi>(__k)));
776 }
777 _GLIBCXX_SIMD_INTRINSIC __m256 operator()(__m256 __a, __m256 __b,
778 __m256 __k) const noexcept
779 {
780 return __builtin_ia32_blendvps256(__a, __b, __k);
781 }
782 _GLIBCXX_SIMD_INTRINSIC __m256d operator()(__m256d __a, __m256d __b,
783 __m256d __k) const noexcept
784 {
785 return __builtin_ia32_blendvpd256(__a, __b, __k);
786 }
787 _GLIBCXX_SIMD_INTRINSIC __m256i operator()(__m256i __a, __m256i __b,
788 __m256i __k) const noexcept
789 {
790 if constexpr (__have_avx2)
791 return reinterpret_cast<__m256i>(
792 __builtin_ia32_pblendvb256(reinterpret_cast<__v32qi>(__a),
793 reinterpret_cast<__v32qi>(__b),
794 reinterpret_cast<__v32qi>(__k)));
795 else
796 return reinterpret_cast<__m256i>(
797 __builtin_ia32_blendvps256(reinterpret_cast<__v8sf>(__a),
798 reinterpret_cast<__v8sf>(__b),
799 reinterpret_cast<__v8sf>(__k)));
800 }
801 } __eval;
802 return __eval(__a, __b, __k);
803 }
804
805 // }}}
806 // _S_blend {{{
807 // Returns: __k ? __at1 : __at0
808 // TODO: reverse __at0 and __at1 to match COND_EXPR
809 template <typename _Tp, size_t _Np>
810 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
811 _S_blend(_SimdWrapper<bool, _Np> __k, _SimdWrapper<_Tp, _Np> __at0,
812 _SimdWrapper<_Tp, _Np> __at1)
813 {
814 static_assert(is_same_v<_Tp, _Tp> && __have_avx512f);
815 if (__k._M_is_constprop() && __at0._M_is_constprop()
816 && __at1._M_is_constprop())
817 return __generate_from_n_evaluations<_Np, __vector_type_t<_Tp, _Np>>(
818 [&](auto __i) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
819 return __k[__i] ? __at1[__i] : __at0[__i];
820 });
821 else if constexpr (sizeof(__at0) == 64
822 || (__have_avx512vl && sizeof(__at0) >= 16))
823 return _S_blend_avx512(__k._M_data, __at0._M_data, __at1._M_data);
824 else
825 {
826 static_assert((__have_avx512vl && sizeof(__at0) < 16)
827 || !__have_avx512vl);
828 constexpr size_t __size = (__have_avx512vl ? 16 : 64) / sizeof(_Tp);
829 return __vector_bitcast<_Tp, _Np>(
830 _S_blend_avx512(__k._M_data, __vector_bitcast<_Tp, __size>(__at0),
831 __vector_bitcast<_Tp, __size>(__at1)));
832 }
833 }
834
835 template <typename _Tp, size_t _Np>
836 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
837 _S_blend(_SimdWrapper<__int_for_sizeof_t<_Tp>, _Np> __k,
838 _SimdWrapper<_Tp, _Np> __at0, _SimdWrapper<_Tp, _Np> __at1)
839 {
840 const auto __kk = __wrapper_bitcast<_Tp>(__k);
841 if (__builtin_is_constant_evaluated()
842 || (__kk._M_is_constprop() && __at0._M_is_constprop()
843 && __at1._M_is_constprop()))
844 {
845 auto __r = __or(__andnot(__kk, __at0), __and(__kk, __at1));
846 if (__r._M_is_constprop())
847 return __r;
848 }
849 if constexpr (((__have_avx512f && sizeof(__at0) == 64) || __have_avx512vl)
850 && (sizeof(_Tp) >= 4 || __have_avx512bw))
851 // convert to bitmask and call overload above
852 return _S_blend(
853 _SimdWrapper<bool, _Np>(
854 __make_dependent_t<_Tp, _MaskImplX86Mixin>::_S_to_bits(__k)
855 ._M_to_bits()),
856 __at0, __at1);
857 else
858 {
859 // Since GCC does not assume __k to be a mask, using the builtin
860 // conditional operator introduces an extra compare against 0 before
861 // blending. So we rather call the intrinsic here.
862 if constexpr (__have_sse4_1)
863 return _S_blend_intrin(__to_intrin(__kk), __to_intrin(__at0),
864 __to_intrin(__at1));
865 else
866 return __or(__andnot(__kk, __at0), __and(__kk, __at1));
867 }
868 }
869
870 // }}}
871};
872
873// }}}
874// _SimdImplX86 {{{
875template <typename _Abi, typename>
876 struct _SimdImplX86 : _SimdImplBuiltin<_Abi>
877 {
878 using _Base = _SimdImplBuiltin<_Abi>;
879
880 template <typename _Tp>
881 using _MaskMember = typename _Base::template _MaskMember<_Tp>;
882
883 template <typename _Tp>
884 static constexpr size_t _S_full_size = _Abi::template _S_full_size<_Tp>;
885
886 template <typename _Tp>
887 static constexpr size_t _S_size = _Abi::template _S_size<_Tp>;
888
889 template <typename _Tp>
890 static constexpr size_t _S_max_store_size
891 = (sizeof(_Tp) >= 4 && __have_avx512f) || __have_avx512bw ? 64
892 : (is_floating_point_v<_Tp>&& __have_avx) || __have_avx2 ? 32
893 : 16;
894
895 using _MaskImpl = typename _Abi::_MaskImpl;
896
897 // _S_masked_load {{{
898 template <typename _Tp, size_t _Np, typename _Up>
899 static inline _SimdWrapper<_Tp, _Np>
900 _S_masked_load(_SimdWrapper<_Tp, _Np> __merge, _MaskMember<_Tp> __k,
901 const _Up* __mem) noexcept
902 {
903 static_assert(_Np == _S_size<_Tp>);
904 if constexpr (is_same_v<_Tp, _Up> || // no conversion
905 (sizeof(_Tp) == sizeof(_Up)
906 && is_integral_v<
907 _Tp> == is_integral_v<_Up>) // conversion via bit
908 // reinterpretation
909 )
910 {
911 [[maybe_unused]] const auto __intrin = __to_intrin(__merge);
912 if constexpr ((__is_avx512_abi<_Abi>() || __have_avx512bw_vl)
913 && sizeof(_Tp) == 1)
914 {
915 const auto __kk = _MaskImpl::_S_to_bits(__k)._M_to_bits();
916 if constexpr (sizeof(__intrin) == 16)
917 __merge = __vector_bitcast<_Tp, _Np>(
918 _mm_mask_loadu_epi8(__intrin, __kk, __mem));
919 else if constexpr (sizeof(__merge) == 32)
920 __merge = __vector_bitcast<_Tp, _Np>(
921 _mm256_mask_loadu_epi8(__intrin, __kk, __mem));
922 else if constexpr (sizeof(__merge) == 64)
923 __merge = __vector_bitcast<_Tp, _Np>(
924 _mm512_mask_loadu_epi8(__intrin, __kk, __mem));
925 else
926 __assert_unreachable<_Tp>();
927 }
928 else if constexpr ((__is_avx512_abi<_Abi>() || __have_avx512bw_vl)
929 && sizeof(_Tp) == 2)
930 {
931 const auto __kk = _MaskImpl::_S_to_bits(__k)._M_to_bits();
932 if constexpr (sizeof(__intrin) == 16)
933 __merge = __vector_bitcast<_Tp, _Np>(
934 _mm_mask_loadu_epi16(__intrin, __kk, __mem));
935 else if constexpr (sizeof(__intrin) == 32)
936 __merge = __vector_bitcast<_Tp, _Np>(
937 _mm256_mask_loadu_epi16(__intrin, __kk, __mem));
938 else if constexpr (sizeof(__intrin) == 64)
939 __merge = __vector_bitcast<_Tp, _Np>(
940 _mm512_mask_loadu_epi16(__intrin, __kk, __mem));
941 else
942 __assert_unreachable<_Tp>();
943 }
944 else if constexpr ((__is_avx512_abi<_Abi>() || __have_avx512vl)
945 && sizeof(_Tp) == 4 && is_integral_v<_Up>)
946 {
947 const auto __kk = _MaskImpl::_S_to_bits(__k)._M_to_bits();
948 if constexpr (sizeof(__intrin) == 16)
949 __merge = __vector_bitcast<_Tp, _Np>(
950 _mm_mask_loadu_epi32(__intrin, __kk, __mem));
951 else if constexpr (sizeof(__intrin) == 32)
952 __merge = __vector_bitcast<_Tp, _Np>(
953 _mm256_mask_loadu_epi32(__intrin, __kk, __mem));
954 else if constexpr (sizeof(__intrin) == 64)
955 __merge = __vector_bitcast<_Tp, _Np>(
956 _mm512_mask_loadu_epi32(__intrin, __kk, __mem));
957 else
958 __assert_unreachable<_Tp>();
959 }
960 else if constexpr ((__is_avx512_abi<_Abi>() || __have_avx512vl)
961 && sizeof(_Tp) == 4 && is_floating_point_v<_Up>)
962 {
963 const auto __kk = _MaskImpl::_S_to_bits(__k)._M_to_bits();
964 if constexpr (sizeof(__intrin) == 16)
965 __merge = __vector_bitcast<_Tp, _Np>(
966 _mm_mask_loadu_ps(__intrin, __kk, __mem));
967 else if constexpr (sizeof(__intrin) == 32)
968 __merge = __vector_bitcast<_Tp, _Np>(
969 _mm256_mask_loadu_ps(__intrin, __kk, __mem));
970 else if constexpr (sizeof(__intrin) == 64)
971 __merge = __vector_bitcast<_Tp, _Np>(
972 _mm512_mask_loadu_ps(__intrin, __kk, __mem));
973 else
974 __assert_unreachable<_Tp>();
975 }
976 else if constexpr (__have_avx2 && sizeof(_Tp) == 4
977 && is_integral_v<_Up>)
978 {
979 static_assert(sizeof(__intrin) == 16 || sizeof(__intrin) == 32);
980 __merge
981 = __or(__andnot(__vector_bitcast<_Tp>(__k), __merge._M_data),
982 __vector_bitcast<_Tp, _Np>(
983 __maskload_epi32(reinterpret_cast<const int*>(__mem),
984 __to_intrin(__k))));
985 }
986 else if constexpr (__have_avx && sizeof(_Tp) == 4)
987 {
988 static_assert(sizeof(__intrin) == 16 || sizeof(__intrin) == 32);
989 __merge
990 = __or(__andnot(__vector_bitcast<_Tp>(__k), __merge._M_data),
991 __vector_bitcast<_Tp, _Np>(
992 __maskload_ps(reinterpret_cast<const float*>(__mem),
993 __to_intrin(__k))));
994 }
995 else if constexpr ((__is_avx512_abi<_Abi>() || __have_avx512vl)
996 && sizeof(_Tp) == 8 && is_integral_v<_Up>)
997 {
998 const auto __kk = _MaskImpl::_S_to_bits(__k)._M_to_bits();
999 if constexpr (sizeof(__intrin) == 16)
1000 __merge = __vector_bitcast<_Tp, _Np>(
1001 _mm_mask_loadu_epi64(__intrin, __kk, __mem));
1002 else if constexpr (sizeof(__intrin) == 32)
1003 __merge = __vector_bitcast<_Tp, _Np>(
1004 _mm256_mask_loadu_epi64(__intrin, __kk, __mem));
1005 else if constexpr (sizeof(__intrin) == 64)
1006 __merge = __vector_bitcast<_Tp, _Np>(
1007 _mm512_mask_loadu_epi64(__intrin, __kk, __mem));
1008 else
1009 __assert_unreachable<_Tp>();
1010 }
1011 else if constexpr ((__is_avx512_abi<_Abi>() || __have_avx512vl)
1012 && sizeof(_Tp) == 8 && is_floating_point_v<_Up>)
1013 {
1014 const auto __kk = _MaskImpl::_S_to_bits(__k)._M_to_bits();
1015 if constexpr (sizeof(__intrin) == 16)
1016 __merge = __vector_bitcast<_Tp, _Np>(
1017 _mm_mask_loadu_pd(__intrin, __kk, __mem));
1018 else if constexpr (sizeof(__intrin) == 32)
1019 __merge = __vector_bitcast<_Tp, _Np>(
1020 _mm256_mask_loadu_pd(__intrin, __kk, __mem));
1021 else if constexpr (sizeof(__intrin) == 64)
1022 __merge = __vector_bitcast<_Tp, _Np>(
1023 _mm512_mask_loadu_pd(__intrin, __kk, __mem));
1024 else
1025 __assert_unreachable<_Tp>();
1026 }
1027 else if constexpr (__have_avx2 && sizeof(_Tp) == 8
1028 && is_integral_v<_Up>)
1029 {
1030 static_assert(sizeof(__intrin) == 16 || sizeof(__intrin) == 32);
1031 __merge
1032 = __or(__andnot(__vector_bitcast<_Tp>(__k), __merge._M_data),
1033 __vector_bitcast<_Tp, _Np>(__maskload_epi64(
1034 reinterpret_cast<const _LLong*>(__mem),
1035 __to_intrin(__k))));
1036 }
1037 else if constexpr (__have_avx && sizeof(_Tp) == 8)
1038 {
1039 static_assert(sizeof(__intrin) == 16 || sizeof(__intrin) == 32);
1040 __merge
1041 = __or(__andnot(__vector_bitcast<_Tp>(__k), __merge._M_data),
1042 __vector_bitcast<_Tp, _Np>(
1043 __maskload_pd(reinterpret_cast<const double*>(__mem),
1044 __to_intrin(__k))));
1045 }
1046 else
1047 _BitOps::_S_bit_iteration(_MaskImpl::_S_to_bits(__k),
1048 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1049 __merge._M_set(__i, static_cast<_Tp>(__mem[__i]));
1050 });
1051 }
1052 /* Very uncertain, that the following improves anything. Needs
1053 benchmarking
1054 * before it's activated.
1055 else if constexpr (sizeof(_Up) <= 8 && // no long double
1056 !__converts_via_decomposition_v<
1057 _Up, _Tp,
1058 sizeof(__merge)> // conversion via decomposition
1059 // is better handled via the
1060 // bit_iteration fallback below
1061 )
1062 {
1063 // TODO: copy pattern from _S_masked_store, which doesn't resort to
1064 // fixed_size
1065 using _Ap = simd_abi::deduce_t<_Up, _Np>;
1066 using _ATraits = _SimdTraits<_Up, _Ap>;
1067 using _AImpl = typename _ATraits::_SimdImpl;
1068 typename _ATraits::_SimdMember __uncvted{};
1069 typename _ATraits::_MaskMember __kk = _Ap::_MaskImpl::template
1070 _S_convert<_Up>(__k);
1071 __uncvted = _AImpl::_S_masked_load(__uncvted, __kk, __mem);
1072 _SimdConverter<_Up, _Ap, _Tp, _Abi> __converter;
1073 _Base::_S_masked_assign(__k, __merge, __converter(__uncvted));
1074 }
1075 */
1076 else
1077 __merge = _Base::_S_masked_load(__merge, __k, __mem);
1078 return __merge;
1079 }
1080
1081 // }}}
1082 // _S_masked_store_nocvt {{{
1083 template <typename _Tp, size_t _Np>
1084 _GLIBCXX_SIMD_INTRINSIC static void
1085 _S_masked_store_nocvt(_SimdWrapper<_Tp, _Np> __v, _Tp* __mem, _SimdWrapper<bool, _Np> __k)
1086 {
1087 [[maybe_unused]] const auto __vi = __to_intrin(__v);
1088 if constexpr (sizeof(__vi) == 64)
1089 {
1090 static_assert(sizeof(__v) == 64 && __have_avx512f);
1091 if constexpr (__have_avx512bw && sizeof(_Tp) == 1)
1092 _mm512_mask_storeu_epi8(__mem, __k, __vi);
1093 else if constexpr (__have_avx512bw && sizeof(_Tp) == 2)
1094 _mm512_mask_storeu_epi16(__mem, __k, __vi);
1095 else if constexpr (__have_avx512f && sizeof(_Tp) == 4)
1096 {
1097 if constexpr (is_integral_v<_Tp>)
1098 _mm512_mask_storeu_epi32(__mem, __k, __vi);
1099 else
1100 _mm512_mask_storeu_ps(__mem, __k, __vi);
1101 }
1102 else if constexpr (__have_avx512f && sizeof(_Tp) == 8)
1103 {
1104 if constexpr (is_integral_v<_Tp>)
1105 _mm512_mask_storeu_epi64(__mem, __k, __vi);
1106 else
1107 _mm512_mask_storeu_pd(__mem, __k, __vi);
1108 }
1109 else
1110 __assert_unreachable<_Tp>();
1111 }
1112 else if constexpr (sizeof(__vi) == 32)
1113 {
1114 if constexpr (__have_avx512bw_vl && sizeof(_Tp) == 1)
1115 _mm256_mask_storeu_epi8(__mem, __k, __vi);
1116 else if constexpr (__have_avx512bw_vl && sizeof(_Tp) == 2)
1117 _mm256_mask_storeu_epi16(__mem, __k, __vi);
1118 else if constexpr (__have_avx512vl && sizeof(_Tp) == 4)
1119 {
1120 if constexpr (is_integral_v<_Tp>)
1121 _mm256_mask_storeu_epi32(__mem, __k, __vi);
1122 else
1123 _mm256_mask_storeu_ps(__mem, __k, __vi);
1124 }
1125 else if constexpr (__have_avx512vl && sizeof(_Tp) == 8)
1126 {
1127 if constexpr (is_integral_v<_Tp>)
1128 _mm256_mask_storeu_epi64(__mem, __k, __vi);
1129 else
1130 _mm256_mask_storeu_pd(__mem, __k, __vi);
1131 }
1132 else if constexpr (__have_avx512f
1133 && (sizeof(_Tp) >= 4 || __have_avx512bw))
1134 {
1135 // use a 512-bit maskstore, using zero-extension of the bitmask
1136 _S_masked_store_nocvt(
1137 _SimdWrapper64<_Tp>(
1138 __intrin_bitcast<__vector_type64_t<_Tp>>(__v._M_data)),
1139 __mem, _SimdWrapper<bool, 64 / sizeof(_Tp)>(__k._M_data));
1140 }
1141 else
1142 _S_masked_store_nocvt(__v, __mem,
1143 _MaskImpl::template _S_to_maskvector<
1144 __int_for_sizeof_t<_Tp>, _Np>(__k));
1145 }
1146 else if constexpr (sizeof(__vi) == 16)
1147 {
1148 if constexpr (__have_avx512bw_vl && sizeof(_Tp) == 1)
1149 _mm_mask_storeu_epi8(__mem, __k, __vi);
1150 else if constexpr (__have_avx512bw_vl && sizeof(_Tp) == 2)
1151 _mm_mask_storeu_epi16(__mem, __k, __vi);
1152 else if constexpr (__have_avx512vl && sizeof(_Tp) == 4)
1153 {
1154 if constexpr (is_integral_v<_Tp>)
1155 _mm_mask_storeu_epi32(__mem, __k, __vi);
1156 else
1157 _mm_mask_storeu_ps(__mem, __k, __vi);
1158 }
1159 else if constexpr (__have_avx512vl && sizeof(_Tp) == 8)
1160 {
1161 if constexpr (is_integral_v<_Tp>)
1162 _mm_mask_storeu_epi64(__mem, __k, __vi);
1163 else
1164 _mm_mask_storeu_pd(__mem, __k, __vi);
1165 }
1166 else if constexpr (__have_avx512f
1167 && (sizeof(_Tp) >= 4 || __have_avx512bw))
1168 {
1169 // use a 512-bit maskstore, using zero-extension of the bitmask
1170 _S_masked_store_nocvt(
1171 _SimdWrapper64<_Tp>(
1172 __intrin_bitcast<__intrinsic_type64_t<_Tp>>(__v._M_data)),
1173 __mem, _SimdWrapper<bool, 64 / sizeof(_Tp)>(__k._M_data));
1174 }
1175 else
1176 _S_masked_store_nocvt(__v, __mem,
1177 _MaskImpl::template _S_to_maskvector<
1178 __int_for_sizeof_t<_Tp>, _Np>(__k));
1179 }
1180 else
1181 __assert_unreachable<_Tp>();
1182 }
1183
1184 template <typename _Tp, size_t _Np>
1185 _GLIBCXX_SIMD_INTRINSIC static void
1186 _S_masked_store_nocvt(_SimdWrapper<_Tp, _Np> __v, _Tp* __mem,
1187 _SimdWrapper<__int_for_sizeof_t<_Tp>, _Np> __k)
1188 {
1189 if constexpr (sizeof(__v) <= 16)
1190 {
1191 [[maybe_unused]] const auto __vi
1192 = __intrin_bitcast<__m128i>(__as_vector(__v));
1193 [[maybe_unused]] const auto __ki
1194 = __intrin_bitcast<__m128i>(__as_vector(__k));
1195 if constexpr (__have_avx512bw_vl && sizeof(_Tp) == 1)
1196 _mm_mask_storeu_epi8(__mem, _mm_movepi8_mask(__ki), __vi);
1197 else if constexpr (__have_avx512bw_vl && sizeof(_Tp) == 2)
1198 _mm_mask_storeu_epi16(__mem, _mm_movepi16_mask(__ki), __vi);
1199 else if constexpr (__have_avx2 && sizeof(_Tp) == 4
1200 && is_integral_v<_Tp>)
1201 _mm_maskstore_epi32(reinterpret_cast<int*>(__mem), __ki, __vi);
1202 else if constexpr (__have_avx && sizeof(_Tp) == 4)
1203 _mm_maskstore_ps(reinterpret_cast<float*>(__mem), __ki,
1204 __vector_bitcast<float>(__vi));
1205 else if constexpr (__have_avx2 && sizeof(_Tp) == 8
1206 && is_integral_v<_Tp>)
1207 _mm_maskstore_epi64(reinterpret_cast<_LLong*>(__mem), __ki, __vi);
1208 else if constexpr (__have_avx && sizeof(_Tp) == 8)
1209 _mm_maskstore_pd(reinterpret_cast<double*>(__mem), __ki,
1210 __vector_bitcast<double>(__vi));
1211 else
1212 _Base::_S_masked_store_nocvt(__v, __mem, __k);
1213 }
1214 else if constexpr (sizeof(__v) == 32)
1215 {
1216 [[maybe_unused]] const auto __vi
1217 = __intrin_bitcast<__m256i>(__as_vector(__v));
1218 [[maybe_unused]] const auto __ki
1219 = __intrin_bitcast<__m256i>(__as_vector(__k));
1220 if constexpr (__have_avx512bw_vl && sizeof(_Tp) == 1)
1221 _mm256_mask_storeu_epi8(__mem, _mm256_movepi8_mask(__ki), __vi);
1222 else if constexpr (__have_avx512bw_vl && sizeof(_Tp) == 2)
1223 _mm256_mask_storeu_epi16(__mem, _mm256_movepi16_mask(__ki), __vi);
1224 else if constexpr (__have_avx2 && sizeof(_Tp) == 4
1225 && is_integral_v<_Tp>)
1226 _mm256_maskstore_epi32(reinterpret_cast<int*>(__mem), __ki, __vi);
1227 else if constexpr (sizeof(_Tp) == 4)
1228 _mm256_maskstore_ps(reinterpret_cast<float*>(__mem), __ki,
1229 __vector_bitcast<float>(__v));
1230 else if constexpr (__have_avx2 && sizeof(_Tp) == 8
1231 && is_integral_v<_Tp>)
1232 _mm256_maskstore_epi64(reinterpret_cast<_LLong*>(__mem), __ki,
1233 __vi);
1234 else if constexpr (__have_avx && sizeof(_Tp) == 8)
1235 _mm256_maskstore_pd(reinterpret_cast<double*>(__mem), __ki,
1236 __vector_bitcast<double>(__v));
1237 else
1238 _Base::_S_masked_store_nocvt(__v, __mem, __k);
1239 }
1240 else
1241 __assert_unreachable<_Tp>();
1242 }
1243
1244 // }}}
1245 // _S_masked_store {{{
1246 template <typename _Tp, size_t _Np, typename _Up>
1247 _GLIBCXX_SIMD_INTRINSIC static void
1248 _S_masked_store(const _SimdWrapper<_Tp, _Np> __v, _Up* __mem,
1249 const _MaskMember<_Tp> __k) noexcept
1250 {
1251 if constexpr (is_integral_v<
1252 _Tp> && is_integral_v<_Up> && sizeof(_Tp) > sizeof(_Up)
1253 && __have_avx512f && (sizeof(_Tp) >= 4 || __have_avx512bw)
1254 && (sizeof(__v) == 64 || __have_avx512vl))
1255 { // truncating store
1256 const auto __vi = __to_intrin(__v);
1257 const auto __kk = _MaskImpl::_S_to_bits(__k)._M_to_bits();
1258 if constexpr (sizeof(_Tp) == 8 && sizeof(_Up) == 4
1259 && sizeof(__vi) == 64)
1260 _mm512_mask_cvtepi64_storeu_epi32(__mem, __kk, __vi);
1261 else if constexpr (sizeof(_Tp) == 8 && sizeof(_Up) == 4
1262 && sizeof(__vi) == 32)
1263 _mm256_mask_cvtepi64_storeu_epi32(__mem, __kk, __vi);
1264 else if constexpr (sizeof(_Tp) == 8 && sizeof(_Up) == 4
1265 && sizeof(__vi) == 16)
1266 _mm_mask_cvtepi64_storeu_epi32(__mem, __kk, __vi);
1267 else if constexpr (sizeof(_Tp) == 8 && sizeof(_Up) == 2
1268 && sizeof(__vi) == 64)
1269 _mm512_mask_cvtepi64_storeu_epi16(__mem, __kk, __vi);
1270 else if constexpr (sizeof(_Tp) == 8 && sizeof(_Up) == 2
1271 && sizeof(__vi) == 32)
1272 _mm256_mask_cvtepi64_storeu_epi16(__mem, __kk, __vi);
1273 else if constexpr (sizeof(_Tp) == 8 && sizeof(_Up) == 2
1274 && sizeof(__vi) == 16)
1275 _mm_mask_cvtepi64_storeu_epi16(__mem, __kk, __vi);
1276 else if constexpr (sizeof(_Tp) == 8 && sizeof(_Up) == 1
1277 && sizeof(__vi) == 64)
1278 _mm512_mask_cvtepi64_storeu_epi8(__mem, __kk, __vi);
1279 else if constexpr (sizeof(_Tp) == 8 && sizeof(_Up) == 1
1280 && sizeof(__vi) == 32)
1281 _mm256_mask_cvtepi64_storeu_epi8(__mem, __kk, __vi);
1282 else if constexpr (sizeof(_Tp) == 8 && sizeof(_Up) == 1
1283 && sizeof(__vi) == 16)
1284 _mm_mask_cvtepi64_storeu_epi8(__mem, __kk, __vi);
1285 else if constexpr (sizeof(_Tp) == 4 && sizeof(_Up) == 2
1286 && sizeof(__vi) == 64)
1287 _mm512_mask_cvtepi32_storeu_epi16(__mem, __kk, __vi);
1288 else if constexpr (sizeof(_Tp) == 4 && sizeof(_Up) == 2
1289 && sizeof(__vi) == 32)
1290 _mm256_mask_cvtepi32_storeu_epi16(__mem, __kk, __vi);
1291 else if constexpr (sizeof(_Tp) == 4 && sizeof(_Up) == 2
1292 && sizeof(__vi) == 16)
1293 _mm_mask_cvtepi32_storeu_epi16(__mem, __kk, __vi);
1294 else if constexpr (sizeof(_Tp) == 4 && sizeof(_Up) == 1
1295 && sizeof(__vi) == 64)
1296 _mm512_mask_cvtepi32_storeu_epi8(__mem, __kk, __vi);
1297 else if constexpr (sizeof(_Tp) == 4 && sizeof(_Up) == 1
1298 && sizeof(__vi) == 32)
1299 _mm256_mask_cvtepi32_storeu_epi8(__mem, __kk, __vi);
1300 else if constexpr (sizeof(_Tp) == 4 && sizeof(_Up) == 1
1301 && sizeof(__vi) == 16)
1302 _mm_mask_cvtepi32_storeu_epi8(__mem, __kk, __vi);
1303 else if constexpr (sizeof(_Tp) == 2 && sizeof(_Up) == 1
1304 && sizeof(__vi) == 64)
1305 _mm512_mask_cvtepi16_storeu_epi8(__mem, __kk, __vi);
1306 else if constexpr (sizeof(_Tp) == 2 && sizeof(_Up) == 1
1307 && sizeof(__vi) == 32)
1308 _mm256_mask_cvtepi16_storeu_epi8(__mem, __kk, __vi);
1309 else if constexpr (sizeof(_Tp) == 2 && sizeof(_Up) == 1
1310 && sizeof(__vi) == 16)
1311 _mm_mask_cvtepi16_storeu_epi8(__mem, __kk, __vi);
1312 else
1313 __assert_unreachable<_Tp>();
1314 }
1315 else
1316 _Base::_S_masked_store(__v, __mem, __k);
1317 }
1318
1319 // }}}
1320 // _S_multiplies {{{
1321 template <typename _V, typename _VVT = _VectorTraits<_V>>
1322 _GLIBCXX_SIMD_INTRINSIC static constexpr _V
1323 _S_multiplies(_V __x, _V __y)
1324 {
1325 using _Tp = typename _VVT::value_type;
1326 if (__builtin_is_constant_evaluated() || __x._M_is_constprop()
1327 || __y._M_is_constprop())
1328 return __as_vector(__x) * __as_vector(__y);
1329 else if constexpr (sizeof(_Tp) == 1)
1330 {
1331 if constexpr (sizeof(_V) == 2)
1332 {
1333 const auto __xs = reinterpret_cast<short>(__x._M_data);
1334 const auto __ys = reinterpret_cast<short>(__y._M_data);
1335 return reinterpret_cast<__vector_type_t<_Tp, 2>>(short(
1336 ((__xs * __ys) & 0xff) | ((__xs >> 8) * (__ys & 0xff00))));
1337 }
1338 else if constexpr (sizeof(_V) == 4 && _VVT::_S_partial_width == 3)
1339 {
1340 const auto __xi = reinterpret_cast<int>(__x._M_data);
1341 const auto __yi = reinterpret_cast<int>(__y._M_data);
1342 return reinterpret_cast<__vector_type_t<_Tp, 3>>(
1343 ((__xi * __yi) & 0xff)
1344 | (((__xi >> 8) * (__yi & 0xff00)) & 0xff00)
1345 | ((__xi >> 16) * (__yi & 0xff0000)));
1346 }
1347 else if constexpr (sizeof(_V) == 4)
1348 {
1349 const auto __xi = reinterpret_cast<int>(__x._M_data);
1350 const auto __yi = reinterpret_cast<int>(__y._M_data);
1351 return reinterpret_cast<__vector_type_t<_Tp, 4>>(
1352 ((__xi * __yi) & 0xff)
1353 | (((__xi >> 8) * (__yi & 0xff00)) & 0xff00)
1354 | (((__xi >> 16) * (__yi & 0xff0000)) & 0xff0000)
1355 | ((__xi >> 24) * (__yi & 0xff000000u)));
1356 }
1357 else if constexpr (sizeof(_V) == 8 && __have_avx2
1358 && is_signed_v<_Tp>)
1359 return __convert<typename _VVT::type>(
1360 __vector_bitcast<short>(_mm_cvtepi8_epi16(__to_intrin(__x)))
1361 * __vector_bitcast<short>(_mm_cvtepi8_epi16(__to_intrin(__y))));
1362 else if constexpr (sizeof(_V) == 8 && __have_avx2
1363 && is_unsigned_v<_Tp>)
1364 return __convert<typename _VVT::type>(
1365 __vector_bitcast<short>(_mm_cvtepu8_epi16(__to_intrin(__x)))
1366 * __vector_bitcast<short>(_mm_cvtepu8_epi16(__to_intrin(__y))));
1367 else
1368 {
1369 // codegen of `x*y` is suboptimal (as of GCC 9.0.1)
1370 constexpr size_t __full_size = _VVT::_S_full_size;
1371 constexpr int _Np = sizeof(_V) >= 16 ? __full_size / 2 : 8;
1372 using _ShortW = _SimdWrapper<short, _Np>;
1373 const _ShortW __even = __vector_bitcast<short, _Np>(__x)
1374 * __vector_bitcast<short, _Np>(__y);
1375 _ShortW __high_byte = _ShortW()._M_data - 256;
1376 //[&]() { asm("" : "+x"(__high_byte._M_data)); }();
1377 const _ShortW __odd
1378 = (__vector_bitcast<short, _Np>(__x) >> 8)
1379 * (__vector_bitcast<short, _Np>(__y) & __high_byte._M_data);
1380 if constexpr (__have_avx512bw && sizeof(_V) > 2)
1381 return _CommonImplX86::_S_blend_avx512(
1382 0xaaaa'aaaa'aaaa'aaaaLL, __vector_bitcast<_Tp>(__even),
1383 __vector_bitcast<_Tp>(__odd));
1384 else if constexpr (__have_sse4_1 && sizeof(_V) > 2)
1385 return _CommonImplX86::_S_blend_intrin(__to_intrin(
1386 __high_byte),
1387 __to_intrin(__even),
1388 __to_intrin(__odd));
1389 else
1390 return __to_intrin(
1391 __or(__andnot(__high_byte, __even), __odd));
1392 }
1393 }
1394 else
1395 return _Base::_S_multiplies(__x, __y);
1396 }
1397
1398 // }}}
1399 // _S_divides {{{
1400#ifdef _GLIBCXX_SIMD_WORKAROUND_PR90993
1401 template <typename _Tp, size_t _Np>
1402 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
1403 _S_divides(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1404 {
1405 if (!__builtin_is_constant_evaluated()
1406 && !__builtin_constant_p(__y._M_data))
1407 if constexpr (is_integral_v<_Tp> && sizeof(_Tp) <= 4)
1408 { // use divps - codegen of `x/y` is suboptimal (as of GCC 9.0.1)
1409 // Note that using floating-point division is likely to raise the
1410 // *Inexact* exception flag and thus appears like an invalid
1411 // "as-if" transformation. However, C++ doesn't specify how the
1412 // fpenv can be observed and points to C. C says that function
1413 // calls are assumed to potentially raise fp exceptions, unless
1414 // documented otherwise. Consequently, operator/, which is a
1415 // function call, may raise fp exceptions.
1416 /*const struct _CsrGuard
1417 {
1418 const unsigned _M_data = _mm_getcsr();
1419 _CsrGuard()
1420 {
1421 _mm_setcsr(0x9f80); // turn off FP exceptions and
1422 flush-to-zero
1423 }
1424 ~_CsrGuard() { _mm_setcsr(_M_data); }
1425 } __csr;*/
1426 using _Float = conditional_t<sizeof(_Tp) == 4, double, float>;
1427 constexpr size_t __n_intermediate
1428 = std::min(_Np, (__have_avx512f ? 64
1429 : __have_avx ? 32
1430 : 16)
1431 / sizeof(_Float));
1432 using _FloatV = __vector_type_t<_Float, __n_intermediate>;
1433 constexpr size_t __n_floatv
1434 = __div_roundup(_Np, __n_intermediate);
1435 using _R = __vector_type_t<_Tp, _Np>;
1436 const auto __xf = __convert_all<_FloatV, __n_floatv>(__x);
1437 const auto __yf = __convert_all<_FloatV, __n_floatv>(
1438 _Abi::__make_padding_nonzero(__as_vector(__y)));
1439 return __call_with_n_evaluations<__n_floatv>(
1440 [](auto... __quotients) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1441 return __vector_convert<_R>(__quotients...);
1442 },
1443 [&__xf, &__yf](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA
1444 -> _SimdWrapper<_Float, __n_intermediate>
1445 {
1446#if __RECIPROCAL_MATH__
1447 // If -freciprocal-math is active, using the `/` operator is
1448 // incorrect because it may be translated to an imprecise
1449 // multiplication with reciprocal. We need to use inline
1450 // assembly to force a real division.
1451 _FloatV __r;
1452 if constexpr (__have_avx) // -mno-sse2avx is irrelevant
1453 // because once -mavx is given, GCC
1454 // emits VEX encoded vdivp[sd]
1455 {
1456 if constexpr (sizeof(_Tp) == 4)
1457 asm("vdivpd\t{%2, %1, %0|%0, %1, %2}"
1458 : "=x"(__r)
1459 : "x"(__xf[__i]), "x"(__yf[__i]));
1460 else
1461 asm("vdivps\t{%2, %1, %0|%0, %1, %2}"
1462 : "=x"(__r)
1463 : "x"(__xf[__i]), "x"(__yf[__i]));
1464 }
1465 else
1466 {
1467 __r = __xf[__i];
1468 if constexpr (sizeof(_Tp) == 4)
1469 asm("divpd\t{%1, %0|%0, %1}"
1470 : "=x"(__r)
1471 : "x"(__yf[__i]));
1472 else
1473 asm("divps\t{%1, %0|%0, %1}"
1474 : "=x"(__r)
1475 : "x"(__yf[__i]));
1476 }
1477 return __r;
1478#else
1479 return __xf[__i] / __yf[__i];
1480#endif
1481 });
1482 }
1483 /* 64-bit int division is potentially optimizable via double division if
1484 * the value in __x is small enough and the conversion between
1485 * int<->double is efficient enough:
1486 else if constexpr (is_integral_v<_Tp> && is_unsigned_v<_Tp> &&
1487 sizeof(_Tp) == 8)
1488 {
1489 if constexpr (__have_sse4_1 && sizeof(__x) == 16)
1490 {
1491 if (_mm_test_all_zeros(__x, __m128i{0xffe0'0000'0000'0000ull,
1492 0xffe0'0000'0000'0000ull}))
1493 {
1494 __x._M_data | 0x __vector_convert<__m128d>(__x._M_data)
1495 }
1496 }
1497 }
1498 */
1499 return _Base::_S_divides(__x, __y);
1500 }
1501#else
1502 using _Base::_S_divides;
1503#endif // _GLIBCXX_SIMD_WORKAROUND_PR90993
1504
1505 // }}}
1506 // _S_modulus {{{
1507 template <typename _Tp, size_t _Np>
1508 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
1509 _S_modulus(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1510 {
1511 if (__builtin_is_constant_evaluated()
1512 || __builtin_constant_p(__y._M_data) || sizeof(_Tp) >= 8)
1513 return _Base::_S_modulus(__x, __y);
1514 else
1515 return _Base::_S_minus(__x, _S_multiplies(__y, _S_divides(__x, __y)));
1516 }
1517
1518 // }}}
1519 // _S_bit_shift_left {{{
1520 // Notes on UB. C++2a [expr.shift] says:
1521 // -1- [...] The operands shall be of integral or unscoped enumeration type
1522 // and integral promotions are performed. The type of the result is that
1523 // of the promoted left operand. The behavior is undefined if the right
1524 // operand is negative, or greater than or equal to the width of the
1525 // promoted left operand.
1526 // -2- The value of E1 << E2 is the unique value congruent to E1×2^E2 modulo
1527 // 2^N, where N is the width of the type of the result.
1528 //
1529 // C++17 [expr.shift] says:
1530 // -2- The value of E1 << E2 is E1 left-shifted E2 bit positions; vacated
1531 // bits are zero-filled. If E1 has an unsigned type, the value of the
1532 // result is E1 × 2^E2 , reduced modulo one more than the maximum value
1533 // representable in the result type. Otherwise, if E1 has a signed type
1534 // and non-negative value, and E1 × 2^E2 is representable in the
1535 // corresponding unsigned type of the result type, then that value,
1536 // converted to the result type, is the resulting value; otherwise, the
1537 // behavior is undefined.
1538 //
1539 // Consequences:
1540 // With C++2a signed and unsigned types have the same UB
1541 // characteristics:
1542 // - left shift is not UB for 0 <= RHS < max(32, #bits(T))
1543 //
1544 // With C++17 there's little room for optimizations because the standard
1545 // requires all shifts to happen on promoted integrals (i.e. int). Thus,
1546 // short and char shifts must assume shifts affect bits of neighboring
1547 // values.
1548 #ifndef _GLIBCXX_SIMD_NO_SHIFT_OPT
1549 template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
1550 constexpr inline _GLIBCXX_CONST static typename _TVT::type
1551 _S_bit_shift_left(_Tp __xx, int __y)
1552 {
1553 using _V = typename _TVT::type;
1554 using _Up = typename _TVT::value_type;
1555 _V __x = __xx;
1556 [[maybe_unused]] const auto __ix = __to_intrin(__x);
1557 if (__builtin_is_constant_evaluated())
1558 return __x << __y;
1559#if __cplusplus > 201703
1560 // after C++17, signed shifts have no UB, and behave just like unsigned
1561 // shifts
1562 else if constexpr (sizeof(_Up) == 1 && is_signed_v<_Up>)
1563 return __vector_bitcast<_Up>(
1564 _S_bit_shift_left(__vector_bitcast<make_unsigned_t<_Up>>(__x),
1565 __y));
1566#endif
1567 else if constexpr (sizeof(_Up) == 1)
1568 {
1569 // (cf. https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83894)
1570 if (__builtin_constant_p(__y))
1571 {
1572 if (__y == 0)
1573 return __x;
1574 else if (__y == 1)
1575 return __x + __x;
1576 else if (__y == 2)
1577 {
1578 __x = __x + __x;
1579 return __x + __x;
1580 }
1581 else if (__y > 2 && __y < 8)
1582 {
1583 if constexpr (sizeof(__x) > sizeof(unsigned))
1584 {
1585 const _UChar __mask = 0xff << __y; // precomputed vector
1586 return __vector_bitcast<_Up>(
1587 __vector_bitcast<_UChar>(
1588 __vector_bitcast<unsigned>(__x) << __y)
1589 & __mask);
1590 }
1591 else
1592 {
1593 const unsigned __mask
1594 = (0xff & (0xff << __y)) * 0x01010101u;
1595 return reinterpret_cast<_V>(
1596 static_cast<__int_for_sizeof_t<_V>>(
1597 unsigned(
1598 reinterpret_cast<__int_for_sizeof_t<_V>>(__x)
1599 << __y)
1600 & __mask));
1601 }
1602 }
1603 else if (__y >= 8 && __y < 32)
1604 return _V();
1605 else
1606 __builtin_unreachable();
1607 }
1608 // general strategy in the following: use an sllv instead of sll
1609 // instruction, because it's 2 to 4 times faster:
1610 else if constexpr (__have_avx512bw_vl && sizeof(__x) == 16)
1611 return __vector_bitcast<_Up>(_mm256_cvtepi16_epi8(
1612 _mm256_sllv_epi16(_mm256_cvtepi8_epi16(__ix),
1613 _mm256_set1_epi16(__y))));
1614 else if constexpr (__have_avx512bw && sizeof(__x) == 32)
1615 return __vector_bitcast<_Up>(_mm512_cvtepi16_epi8(
1616 _mm512_sllv_epi16(_mm512_cvtepi8_epi16(__ix),
1617 _mm512_set1_epi16(__y))));
1618 else if constexpr (__have_avx512bw && sizeof(__x) == 64)
1619 {
1620 const auto __shift = _mm512_set1_epi16(__y);
1621 return __vector_bitcast<_Up>(
1622 __concat(_mm512_cvtepi16_epi8(_mm512_sllv_epi16(
1623 _mm512_cvtepi8_epi16(__lo256(__ix)), __shift)),
1624 _mm512_cvtepi16_epi8(_mm512_sllv_epi16(
1625 _mm512_cvtepi8_epi16(__hi256(__ix)), __shift))));
1626 }
1627 else if constexpr (__have_avx2 && sizeof(__x) == 32)
1628 {
1629#if 1
1630 const auto __shift = _mm_cvtsi32_si128(__y);
1631 auto __k
1632 = _mm256_sll_epi16(_mm256_slli_epi16(~__m256i(), 8), __shift);
1633 __k |= _mm256_srli_epi16(__k, 8);
1634 return __vector_bitcast<_Up>(_mm256_sll_epi32(__ix, __shift)
1635 & __k);
1636#else
1637 const _Up __k = 0xff << __y;
1638 return __vector_bitcast<_Up>(__vector_bitcast<int>(__x) << __y)
1639 & __k;
1640#endif
1641 }
1642 else
1643 {
1644 const auto __shift = _mm_cvtsi32_si128(__y);
1645 auto __k
1646 = _mm_sll_epi16(_mm_slli_epi16(~__m128i(), 8), __shift);
1647 __k |= _mm_srli_epi16(__k, 8);
1648 return __intrin_bitcast<_V>(_mm_sll_epi16(__ix, __shift) & __k);
1649 }
1650 }
1651 return __x << __y;
1652 }
1653
1654 template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
1655 constexpr inline _GLIBCXX_CONST static typename _TVT::type
1656 _S_bit_shift_left(_Tp __xx, typename _TVT::type __y)
1657 {
1658 using _V = typename _TVT::type;
1659 using _Up = typename _TVT::value_type;
1660 _V __x = __xx;
1661 [[maybe_unused]] const auto __ix = __to_intrin(__x);
1662 [[maybe_unused]] const auto __iy = __to_intrin(__y);
1663 if (__builtin_is_constant_evaluated())
1664 return __x << __y;
1665#if __cplusplus > 201703
1666 // after C++17, signed shifts have no UB, and behave just like unsigned
1667 // shifts
1668 else if constexpr (is_signed_v<_Up>)
1669 return __vector_bitcast<_Up>(
1670 _S_bit_shift_left(__vector_bitcast<make_unsigned_t<_Up>>(__x),
1671 __vector_bitcast<make_unsigned_t<_Up>>(__y)));
1672#endif
1673 else if constexpr (sizeof(_Up) == 1)
1674 {
1675 if constexpr (sizeof __ix == 64 && __have_avx512bw)
1676 return __vector_bitcast<_Up>(__concat(
1677 _mm512_cvtepi16_epi8(
1678 _mm512_sllv_epi16(_mm512_cvtepu8_epi16(__lo256(__ix)),
1679 _mm512_cvtepu8_epi16(__lo256(__iy)))),
1680 _mm512_cvtepi16_epi8(
1681 _mm512_sllv_epi16(_mm512_cvtepu8_epi16(__hi256(__ix)),
1682 _mm512_cvtepu8_epi16(__hi256(__iy))))));
1683 else if constexpr (sizeof __ix == 32 && __have_avx512bw)
1684 return __vector_bitcast<_Up>(_mm512_cvtepi16_epi8(
1685 _mm512_sllv_epi16(_mm512_cvtepu8_epi16(__ix),
1686 _mm512_cvtepu8_epi16(__iy))));
1687 else if constexpr (sizeof __x <= 8 && __have_avx512bw_vl)
1688 return __intrin_bitcast<_V>(
1689 _mm_cvtepi16_epi8(_mm_sllv_epi16(_mm_cvtepu8_epi16(__ix),
1690 _mm_cvtepu8_epi16(__iy))));
1691 else if constexpr (sizeof __ix == 16 && __have_avx512bw_vl)
1692 return __intrin_bitcast<_V>(_mm256_cvtepi16_epi8(
1693 _mm256_sllv_epi16(_mm256_cvtepu8_epi16(__ix),
1694 _mm256_cvtepu8_epi16(__iy))));
1695 else if constexpr (sizeof __ix == 16 && __have_avx512bw)
1696 return __intrin_bitcast<_V>(
1697 __lo128(_mm512_cvtepi16_epi8(_mm512_sllv_epi16(
1698 _mm512_cvtepu8_epi16(_mm256_castsi128_si256(__ix)),
1699 _mm512_cvtepu8_epi16(_mm256_castsi128_si256(__iy))))));
1700 else if constexpr (__have_sse4_1 && sizeof(__x) == 16)
1701 {
1702 auto __mask
1703 = __vector_bitcast<_Up>(__vector_bitcast<short>(__y) << 5);
1704 auto __x4
1705 = __vector_bitcast<_Up>(__vector_bitcast<short>(__x) << 4);
1706 __x4 &= char(0xf0);
1707 __x = reinterpret_cast<_V>(_CommonImplX86::_S_blend_intrin(
1708 __to_intrin(__mask), __to_intrin(__x), __to_intrin(__x4)));
1709 __mask += __mask;
1710 auto __x2
1711 = __vector_bitcast<_Up>(__vector_bitcast<short>(__x) << 2);
1712 __x2 &= char(0xfc);
1713 __x = reinterpret_cast<_V>(_CommonImplX86::_S_blend_intrin(
1714 __to_intrin(__mask), __to_intrin(__x), __to_intrin(__x2)));
1715 __mask += __mask;
1716 auto __x1 = __x + __x;
1717 __x = reinterpret_cast<_V>(_CommonImplX86::_S_blend_intrin(
1718 __to_intrin(__mask), __to_intrin(__x), __to_intrin(__x1)));
1719 return __x
1720 & ((__y & char(0xf8)) == 0); // y > 7 nulls the result
1721 }
1722 else if constexpr (sizeof(__x) == 16)
1723 {
1724 auto __mask
1725 = __vector_bitcast<_UChar>(__vector_bitcast<short>(__y) << 5);
1726 auto __x4
1727 = __vector_bitcast<_Up>(__vector_bitcast<short>(__x) << 4);
1728 __x4 &= char(0xf0);
1729 __x = __vector_bitcast<_SChar>(__mask) < 0 ? __x4 : __x;
1730 __mask += __mask;
1731 auto __x2
1732 = __vector_bitcast<_Up>(__vector_bitcast<short>(__x) << 2);
1733 __x2 &= char(0xfc);
1734 __x = __vector_bitcast<_SChar>(__mask) < 0 ? __x2 : __x;
1735 __mask += __mask;
1736 auto __x1 = __x + __x;
1737 __x = __vector_bitcast<_SChar>(__mask) < 0 ? __x1 : __x;
1738 return __x
1739 & ((__y & char(0xf8)) == 0); // y > 7 nulls the result
1740 }
1741 else
1742 return __x << __y;
1743 }
1744 else if constexpr (sizeof(_Up) == 2)
1745 {
1746 if constexpr (sizeof __ix == 64 && __have_avx512bw)
1747 return __vector_bitcast<_Up>(_mm512_sllv_epi16(__ix, __iy));
1748 else if constexpr (sizeof __ix == 32 && __have_avx512bw_vl)
1749 return __vector_bitcast<_Up>(_mm256_sllv_epi16(__ix, __iy));
1750 else if constexpr (sizeof __ix == 32 && __have_avx512bw)
1751 return __vector_bitcast<_Up>(
1752 __lo256(_mm512_sllv_epi16(_mm512_castsi256_si512(__ix),
1753 _mm512_castsi256_si512(__iy))));
1754 else if constexpr (sizeof __ix == 32 && __have_avx2)
1755 {
1756 const auto __ux = __vector_bitcast<unsigned>(__x);
1757 const auto __uy = __vector_bitcast<unsigned>(__y);
1758 return __vector_bitcast<_Up>(_mm256_blend_epi16(
1759 __auto_bitcast(__ux << (__uy & 0x0000ffffu)),
1760 __auto_bitcast((__ux & 0xffff0000u) << (__uy >> 16)), 0xaa));
1761 }
1762 else if constexpr (sizeof __ix == 16 && __have_avx512bw_vl)
1763 return __intrin_bitcast<_V>(_mm_sllv_epi16(__ix, __iy));
1764 else if constexpr (sizeof __ix == 16 && __have_avx512bw)
1765 return __intrin_bitcast<_V>(
1766 __lo128(_mm512_sllv_epi16(_mm512_castsi128_si512(__ix),
1767 _mm512_castsi128_si512(__iy))));
1768 else if constexpr (sizeof __ix == 16 && __have_avx2)
1769 {
1770 const auto __ux = __vector_bitcast<unsigned>(__ix);
1771 const auto __uy = __vector_bitcast<unsigned>(__iy);
1772 return __intrin_bitcast<_V>(_mm_blend_epi16(
1773 __auto_bitcast(__ux << (__uy & 0x0000ffffu)),
1774 __auto_bitcast((__ux & 0xffff0000u) << (__uy >> 16)), 0xaa));
1775 }
1776 else if constexpr (sizeof __ix == 16)
1777 {
1778 using _Float4 = __vector_type_t<float, 4>;
1779 using _Int4 = __vector_type_t<int, 4>;
1780 using _UInt4 = __vector_type_t<unsigned, 4>;
1781 const _UInt4 __yu
1782 = reinterpret_cast<_UInt4>(__to_intrin(__y + (0x3f8 >> 3)));
1783 return __x
1784 * __intrin_bitcast<_V>(
1785 __vector_convert<_Int4>(_SimdWrapper<float, 4>(
1786 reinterpret_cast<_Float4>(__yu << 23)))
1787 | (__vector_convert<_Int4>(_SimdWrapper<float, 4>(
1788 reinterpret_cast<_Float4>((__yu >> 16) << 23)))
1789 << 16));
1790 }
1791 else
1792 __assert_unreachable<_Tp>();
1793 }
1794 else if constexpr (sizeof(_Up) == 4 && sizeof __ix == 16
1795 && !__have_avx2)
1796 // latency is suboptimal, but throughput is at full speedup
1797 return __intrin_bitcast<_V>(
1798 __vector_bitcast<unsigned>(__ix)
1799 * __vector_convert<__vector_type16_t<int>>(
1800 _SimdWrapper<float, 4>(__vector_bitcast<float>(
1801 (__vector_bitcast<unsigned, 4>(__y) << 23) + 0x3f80'0000))));
1802 else if constexpr (sizeof(_Up) == 8 && sizeof __ix == 16
1803 && !__have_avx2)
1804 {
1805 const auto __lo = _mm_sll_epi64(__ix, __iy);
1806 const auto __hi
1807 = _mm_sll_epi64(__ix, _mm_unpackhi_epi64(__iy, __iy));
1808 if constexpr (__have_sse4_1)
1809 return __vector_bitcast<_Up>(_mm_blend_epi16(__lo, __hi, 0xf0));
1810 else
1811 return __vector_bitcast<_Up>(
1812 _mm_move_sd(__vector_bitcast<double>(__hi),
1813 __vector_bitcast<double>(__lo)));
1814 }
1815 else
1816 return __x << __y;
1817 }
1818#endif // _GLIBCXX_SIMD_NO_SHIFT_OPT
1819
1820 // }}}
1821 // _S_bit_shift_right {{{
1822#ifndef _GLIBCXX_SIMD_NO_SHIFT_OPT
1823 template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
1824 constexpr inline _GLIBCXX_CONST static typename _TVT::type
1825 _S_bit_shift_right(_Tp __xx, int __y)
1826 {
1827 using _V = typename _TVT::type;
1828 using _Up = typename _TVT::value_type;
1829 _V __x = __xx;
1830 [[maybe_unused]] const auto __ix = __to_intrin(__x);
1831 if (__builtin_is_constant_evaluated())
1832 return __x >> __y;
1833 else if (__builtin_constant_p(__y)
1834 && is_unsigned_v<
1835 _Up> && __y >= int(sizeof(_Up) * __CHAR_BIT__))
1836 return _V();
1837 else if constexpr (sizeof(_Up) == 1 && is_unsigned_v<_Up>) //{{{
1838 return __intrin_bitcast<_V>(__vector_bitcast<_UShort>(__ix) >> __y)
1839 & _Up(0xff >> __y);
1840 //}}}
1841 else if constexpr (sizeof(_Up) == 1 && is_signed_v<_Up>) //{{{
1842 return __intrin_bitcast<_V>(
1843 (__vector_bitcast<_UShort>(__vector_bitcast<short>(__ix)
1844 >> (__y + 8))
1845 << 8)
1846 | (__vector_bitcast<_UShort>(
1847 __vector_bitcast<short>(__vector_bitcast<_UShort>(__ix) << 8)
1848 >> __y)
1849 >> 8));
1850 //}}}
1851 // GCC optimizes sizeof == 2, 4, and unsigned 8 as expected
1852 else if constexpr (sizeof(_Up) == 8 && is_signed_v<_Up>) //{{{
1853 {
1854 if (__y > 32)
1855 return (__intrin_bitcast<_V>(__vector_bitcast<int>(__ix) >> 32)
1856 & _Up(0xffff'ffff'0000'0000ull))
1857 | __vector_bitcast<_Up>(
1858 __vector_bitcast<int>(__vector_bitcast<_ULLong>(__ix)
1859 >> 32)
1860 >> (__y - 32));
1861 else
1862 return __intrin_bitcast<_V>(__vector_bitcast<_ULLong>(__ix)
1863 >> __y)
1864 | __vector_bitcast<_Up>(
1865 __vector_bitcast<int>(__ix & -0x8000'0000'0000'0000ll)
1866 >> __y);
1867 }
1868 //}}}
1869 else
1870 return __x >> __y;
1871 }
1872
1873 template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
1874 constexpr inline _GLIBCXX_CONST static typename _TVT::type
1875 _S_bit_shift_right(_Tp __xx, typename _TVT::type __y)
1876 {
1877 using _V = typename _TVT::type;
1878 using _Up = typename _TVT::value_type;
1879 _V __x = __xx;
1880 [[maybe_unused]] const auto __ix = __to_intrin(__x);
1881 [[maybe_unused]] const auto __iy = __to_intrin(__y);
1882 if (__builtin_is_constant_evaluated()
1883 || (__builtin_constant_p(__x) && __builtin_constant_p(__y)))
1884 return __x >> __y;
1885 else if constexpr (sizeof(_Up) == 1) //{{{
1886 {
1887 if constexpr (sizeof(__x) <= 8 && __have_avx512bw_vl)
1888 return __intrin_bitcast<_V>(_mm_cvtepi16_epi8(
1889 is_signed_v<_Up> ? _mm_srav_epi16(_mm_cvtepi8_epi16(__ix),
1890 _mm_cvtepi8_epi16(__iy))
1891 : _mm_srlv_epi16(_mm_cvtepu8_epi16(__ix),
1892 _mm_cvtepu8_epi16(__iy))));
1893 if constexpr (sizeof(__x) == 16 && __have_avx512bw_vl)
1894 return __intrin_bitcast<_V>(_mm256_cvtepi16_epi8(
1895 is_signed_v<_Up>
1896 ? _mm256_srav_epi16(_mm256_cvtepi8_epi16(__ix),
1897 _mm256_cvtepi8_epi16(__iy))
1898 : _mm256_srlv_epi16(_mm256_cvtepu8_epi16(__ix),
1899 _mm256_cvtepu8_epi16(__iy))));
1900 else if constexpr (sizeof(__x) == 32 && __have_avx512bw)
1901 return __vector_bitcast<_Up>(_mm512_cvtepi16_epi8(
1902 is_signed_v<_Up>
1903 ? _mm512_srav_epi16(_mm512_cvtepi8_epi16(__ix),
1904 _mm512_cvtepi8_epi16(__iy))
1905 : _mm512_srlv_epi16(_mm512_cvtepu8_epi16(__ix),
1906 _mm512_cvtepu8_epi16(__iy))));
1907 else if constexpr (sizeof(__x) == 64 && is_signed_v<_Up>)
1908 return __vector_bitcast<_Up>(_mm512_mask_mov_epi8(
1909 _mm512_srav_epi16(__ix, _mm512_srli_epi16(__iy, 8)),
1910 0x5555'5555'5555'5555ull,
1911 _mm512_srav_epi16(
1912 _mm512_slli_epi16(__ix, 8),
1913 _mm512_maskz_add_epi8(0x5555'5555'5555'5555ull, __iy,
1914 _mm512_set1_epi16(8)))));
1915 else if constexpr (sizeof(__x) == 64 && is_unsigned_v<_Up>)
1916 return __vector_bitcast<_Up>(_mm512_mask_mov_epi8(
1917 _mm512_srlv_epi16(__ix, _mm512_srli_epi16(__iy, 8)),
1918 0x5555'5555'5555'5555ull,
1919 _mm512_srlv_epi16(
1920 _mm512_maskz_mov_epi8(0x5555'5555'5555'5555ull, __ix),
1921 _mm512_maskz_mov_epi8(0x5555'5555'5555'5555ull, __iy))));
1922 /* This has better throughput but higher latency than the impl below
1923 else if constexpr (__have_avx2 && sizeof(__x) == 16 &&
1924 is_unsigned_v<_Up>)
1925 {
1926 const auto __shorts = __to_intrin(_S_bit_shift_right(
1927 __vector_bitcast<_UShort>(_mm256_cvtepu8_epi16(__ix)),
1928 __vector_bitcast<_UShort>(_mm256_cvtepu8_epi16(__iy))));
1929 return __vector_bitcast<_Up>(
1930 _mm_packus_epi16(__lo128(__shorts), __hi128(__shorts)));
1931 }
1932 */
1933 else if constexpr (__have_avx2 && sizeof(__x) > 8)
1934 // the following uses vpsr[al]vd, which requires AVX2
1935 if constexpr (is_signed_v<_Up>)
1936 {
1937 const auto r3 = __vector_bitcast<_UInt>(
1938 (__vector_bitcast<int>(__x)
1939 >> (__vector_bitcast<_UInt>(__y) >> 24)))
1940 & 0xff000000u;
1941 const auto r2
1942 = __vector_bitcast<_UInt>(
1943 ((__vector_bitcast<int>(__x) << 8)
1944 >> ((__vector_bitcast<_UInt>(__y) << 8) >> 24)))
1945 & 0xff000000u;
1946 const auto r1
1947 = __vector_bitcast<_UInt>(
1948 ((__vector_bitcast<int>(__x) << 16)
1949 >> ((__vector_bitcast<_UInt>(__y) << 16) >> 24)))
1950 & 0xff000000u;
1951 const auto r0 = __vector_bitcast<_UInt>(
1952 (__vector_bitcast<int>(__x) << 24)
1953 >> ((__vector_bitcast<_UInt>(__y) << 24) >> 24));
1954 return __vector_bitcast<_Up>(r3 | (r2 >> 8) | (r1 >> 16)
1955 | (r0 >> 24));
1956 }
1957 else
1958 {
1959 const auto r3 = (__vector_bitcast<_UInt>(__x)
1960 >> (__vector_bitcast<_UInt>(__y) >> 24))
1961 & 0xff000000u;
1962 const auto r2
1963 = ((__vector_bitcast<_UInt>(__x) << 8)
1964 >> ((__vector_bitcast<_UInt>(__y) << 8) >> 24))
1965 & 0xff000000u;
1966 const auto r1
1967 = ((__vector_bitcast<_UInt>(__x) << 16)
1968 >> ((__vector_bitcast<_UInt>(__y) << 16) >> 24))
1969 & 0xff000000u;
1970 const auto r0
1971 = (__vector_bitcast<_UInt>(__x) << 24)
1972 >> ((__vector_bitcast<_UInt>(__y) << 24) >> 24);
1973 return __vector_bitcast<_Up>(r3 | (r2 >> 8) | (r1 >> 16)
1974 | (r0 >> 24));
1975 }
1976 else if constexpr (__have_sse4_1
1977 && is_unsigned_v<_Up> && sizeof(__x) > 2)
1978 {
1979 auto __x128 = __vector_bitcast<_Up>(__ix);
1980 auto __mask
1981 = __vector_bitcast<_Up>(__vector_bitcast<_UShort>(__iy) << 5);
1982 auto __x4 = __vector_bitcast<_Up>(
1983 (__vector_bitcast<_UShort>(__x128) >> 4) & _UShort(0xff0f));
1984 __x128 = __vector_bitcast<_Up>(_CommonImplX86::_S_blend_intrin(
1985 __to_intrin(__mask), __to_intrin(__x128), __to_intrin(__x4)));
1986 __mask += __mask;
1987 auto __x2 = __vector_bitcast<_Up>(
1988 (__vector_bitcast<_UShort>(__x128) >> 2) & _UShort(0xff3f));
1989 __x128 = __vector_bitcast<_Up>(_CommonImplX86::_S_blend_intrin(
1990 __to_intrin(__mask), __to_intrin(__x128), __to_intrin(__x2)));
1991 __mask += __mask;
1992 auto __x1 = __vector_bitcast<_Up>(
1993 (__vector_bitcast<_UShort>(__x128) >> 1) & _UShort(0xff7f));
1994 __x128 = __vector_bitcast<_Up>(_CommonImplX86::_S_blend_intrin(
1995 __to_intrin(__mask), __to_intrin(__x128), __to_intrin(__x1)));
1996 return __intrin_bitcast<_V>(
1997 __x128
1998 & ((__vector_bitcast<_Up>(__iy) & char(0xf8))
1999 == 0)); // y > 7 nulls the result
2000 }
2001 else if constexpr (__have_sse4_1
2002 && is_signed_v<_Up> && sizeof(__x) > 2)
2003 {
2004 auto __mask = __vector_bitcast<_UChar>(
2005 __vector_bitcast<_UShort>(__iy) << 5);
2006 auto __maskl = [&]() _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2007 return __to_intrin(__vector_bitcast<_UShort>(__mask) << 8);
2008 };
2009 auto __xh = __vector_bitcast<short>(__ix);
2010 auto __xl = __vector_bitcast<short>(__ix) << 8;
2011 auto __xh4 = __xh >> 4;
2012 auto __xl4 = __xl >> 4;
2013 __xh = __vector_bitcast<short>(_CommonImplX86::_S_blend_intrin(
2014 __to_intrin(__mask), __to_intrin(__xh), __to_intrin(__xh4)));
2015 __xl = __vector_bitcast<short>(
2016 _CommonImplX86::_S_blend_intrin(__maskl(), __to_intrin(__xl),
2017 __to_intrin(__xl4)));
2018 __mask += __mask;
2019 auto __xh2 = __xh >> 2;
2020 auto __xl2 = __xl >> 2;
2021 __xh = __vector_bitcast<short>(_CommonImplX86::_S_blend_intrin(
2022 __to_intrin(__mask), __to_intrin(__xh), __to_intrin(__xh2)));
2023 __xl = __vector_bitcast<short>(
2024 _CommonImplX86::_S_blend_intrin(__maskl(), __to_intrin(__xl),
2025 __to_intrin(__xl2)));
2026 __mask += __mask;
2027 auto __xh1 = __xh >> 1;
2028 auto __xl1 = __xl >> 1;
2029 __xh = __vector_bitcast<short>(_CommonImplX86::_S_blend_intrin(
2030 __to_intrin(__mask), __to_intrin(__xh), __to_intrin(__xh1)));
2031 __xl = __vector_bitcast<short>(
2032 _CommonImplX86::_S_blend_intrin(__maskl(), __to_intrin(__xl),
2033 __to_intrin(__xl1)));
2034 return __intrin_bitcast<_V>(
2035 (__vector_bitcast<_Up>((__xh & short(0xff00)))
2036 | __vector_bitcast<_Up>(__vector_bitcast<_UShort>(__xl)
2037 >> 8))
2038 & ((__vector_bitcast<_Up>(__iy) & char(0xf8))
2039 == 0)); // y > 7 nulls the result
2040 }
2041 else if constexpr (is_unsigned_v<_Up> && sizeof(__x) > 2) // SSE2
2042 {
2043 auto __mask
2044 = __vector_bitcast<_Up>(__vector_bitcast<_UShort>(__y) << 5);
2045 auto __x4 = __vector_bitcast<_Up>(
2046 (__vector_bitcast<_UShort>(__x) >> 4) & _UShort(0xff0f));
2047 __x = __mask > 0x7f ? __x4 : __x;
2048 __mask += __mask;
2049 auto __x2 = __vector_bitcast<_Up>(
2050 (__vector_bitcast<_UShort>(__x) >> 2) & _UShort(0xff3f));
2051 __x = __mask > 0x7f ? __x2 : __x;
2052 __mask += __mask;
2053 auto __x1 = __vector_bitcast<_Up>(
2054 (__vector_bitcast<_UShort>(__x) >> 1) & _UShort(0xff7f));
2055 __x = __mask > 0x7f ? __x1 : __x;
2056 return __x
2057 & ((__y & char(0xf8)) == 0); // y > 7 nulls the result
2058 }
2059 else if constexpr (sizeof(__x) > 2) // signed SSE2
2060 {
2061 static_assert(is_signed_v<_Up>);
2062 auto __maskh = __vector_bitcast<_UShort>(__y) << 5;
2063 auto __maskl = __vector_bitcast<_UShort>(__y) << (5 + 8);
2064 auto __xh = __vector_bitcast<short>(__x);
2065 auto __xl = __vector_bitcast<short>(__x) << 8;
2066 auto __xh4 = __xh >> 4;
2067 auto __xl4 = __xl >> 4;
2068 __xh = __maskh > 0x7fff ? __xh4 : __xh;
2069 __xl = __maskl > 0x7fff ? __xl4 : __xl;
2070 __maskh += __maskh;
2071 __maskl += __maskl;
2072 auto __xh2 = __xh >> 2;
2073 auto __xl2 = __xl >> 2;
2074 __xh = __maskh > 0x7fff ? __xh2 : __xh;
2075 __xl = __maskl > 0x7fff ? __xl2 : __xl;
2076 __maskh += __maskh;
2077 __maskl += __maskl;
2078 auto __xh1 = __xh >> 1;
2079 auto __xl1 = __xl >> 1;
2080 __xh = __maskh > 0x7fff ? __xh1 : __xh;
2081 __xl = __maskl > 0x7fff ? __xl1 : __xl;
2082 __x = __vector_bitcast<_Up>((__xh & short(0xff00)))
2083 | __vector_bitcast<_Up>(__vector_bitcast<_UShort>(__xl)
2084 >> 8);
2085 return __x
2086 & ((__y & char(0xf8)) == 0); // y > 7 nulls the result
2087 }
2088 else
2089 return __x >> __y;
2090 } //}}}
2091 else if constexpr (sizeof(_Up) == 2 && sizeof(__x) >= 4) //{{{
2092 {
2093 [[maybe_unused]] auto __blend_0xaa
2094 = [](auto __a, auto __b) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2095 if constexpr (sizeof(__a) == 16)
2096 return _mm_blend_epi16(__to_intrin(__a), __to_intrin(__b),
2097 0xaa);
2098 else if constexpr (sizeof(__a) == 32)
2099 return _mm256_blend_epi16(__to_intrin(__a), __to_intrin(__b),
2100 0xaa);
2101 else if constexpr (sizeof(__a) == 64)
2102 return _mm512_mask_blend_epi16(0xaaaa'aaaaU, __to_intrin(__a),
2103 __to_intrin(__b));
2104 else
2105 __assert_unreachable<decltype(__a)>();
2106 };
2107 if constexpr (__have_avx512bw_vl && sizeof(_Tp) <= 16)
2108 return __intrin_bitcast<_V>(is_signed_v<_Up>
2109 ? _mm_srav_epi16(__ix, __iy)
2110 : _mm_srlv_epi16(__ix, __iy));
2111 else if constexpr (__have_avx512bw_vl && sizeof(_Tp) == 32)
2112 return __vector_bitcast<_Up>(is_signed_v<_Up>
2113 ? _mm256_srav_epi16(__ix, __iy)
2114 : _mm256_srlv_epi16(__ix, __iy));
2115 else if constexpr (__have_avx512bw && sizeof(_Tp) == 64)
2116 return __vector_bitcast<_Up>(is_signed_v<_Up>
2117 ? _mm512_srav_epi16(__ix, __iy)
2118 : _mm512_srlv_epi16(__ix, __iy));
2119 else if constexpr (__have_avx2 && is_signed_v<_Up>)
2120 return __intrin_bitcast<_V>(
2121 __blend_0xaa(((__vector_bitcast<int>(__ix) << 16)
2122 >> (__vector_bitcast<int>(__iy) & 0xffffu))
2123 >> 16,
2124 __vector_bitcast<int>(__ix)
2125 >> (__vector_bitcast<int>(__iy) >> 16)));
2126 else if constexpr (__have_avx2 && is_unsigned_v<_Up>)
2127 return __intrin_bitcast<_V>(
2128 __blend_0xaa((__vector_bitcast<_UInt>(__ix) & 0xffffu)
2129 >> (__vector_bitcast<_UInt>(__iy) & 0xffffu),
2130 __vector_bitcast<_UInt>(__ix)
2131 >> (__vector_bitcast<_UInt>(__iy) >> 16)));
2132 else if constexpr (__have_sse4_1)
2133 {
2134 auto __mask = __vector_bitcast<_UShort>(__iy);
2135 auto __x128 = __vector_bitcast<_Up>(__ix);
2136 //__mask *= 0x0808;
2137 __mask = (__mask << 3) | (__mask << 11);
2138 // do __x128 = 0 where __y[4] is set
2139 __x128 = __vector_bitcast<_Up>(
2140 _mm_blendv_epi8(__to_intrin(__x128), __m128i(),
2141 __to_intrin(__mask)));
2142 // do __x128 =>> 8 where __y[3] is set
2143 __x128 = __vector_bitcast<_Up>(
2144 _mm_blendv_epi8(__to_intrin(__x128), __to_intrin(__x128 >> 8),
2145 __to_intrin(__mask += __mask)));
2146 // do __x128 =>> 4 where __y[2] is set
2147 __x128 = __vector_bitcast<_Up>(
2148 _mm_blendv_epi8(__to_intrin(__x128), __to_intrin(__x128 >> 4),
2149 __to_intrin(__mask += __mask)));
2150 // do __x128 =>> 2 where __y[1] is set
2151 __x128 = __vector_bitcast<_Up>(
2152 _mm_blendv_epi8(__to_intrin(__x128), __to_intrin(__x128 >> 2),
2153 __to_intrin(__mask += __mask)));
2154 // do __x128 =>> 1 where __y[0] is set
2155 return __intrin_bitcast<_V>(
2156 _mm_blendv_epi8(__to_intrin(__x128), __to_intrin(__x128 >> 1),
2157 __to_intrin(__mask + __mask)));
2158 }
2159 else
2160 {
2161 auto __k = __vector_bitcast<_UShort>(__iy) << 11;
2162 auto __x128 = __vector_bitcast<_Up>(__ix);
2163 auto __mask
2164 = [](__vector_type16_t<_UShort> __kk) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2165 return __vector_bitcast<short>(__kk) < 0;
2166 };
2167 // do __x128 = 0 where __y[4] is set
2168 __x128 = __mask(__k) ? decltype(__x128)() : __x128;
2169 // do __x128 =>> 8 where __y[3] is set
2170 __x128 = __mask(__k += __k) ? __x128 >> 8 : __x128;
2171 // do __x128 =>> 4 where __y[2] is set
2172 __x128 = __mask(__k += __k) ? __x128 >> 4 : __x128;
2173 // do __x128 =>> 2 where __y[1] is set
2174 __x128 = __mask(__k += __k) ? __x128 >> 2 : __x128;
2175 // do __x128 =>> 1 where __y[0] is set
2176 return __intrin_bitcast<_V>(__mask(__k + __k) ? __x128 >> 1
2177 : __x128);
2178 }
2179 } //}}}
2180 else if constexpr (sizeof(_Up) == 4 && !__have_avx2) //{{{
2181 {
2182 if constexpr (is_unsigned_v<_Up>)
2183 {
2184 // x >> y == x * 2^-y == (x * 2^(31-y)) >> 31
2185 const __m128 __factor_f = reinterpret_cast<__m128>(
2186 0x4f00'0000u - (__vector_bitcast<unsigned, 4>(__y) << 23));
2187 const __m128i __factor
2188 = __builtin_constant_p(__factor_f)
2189 ? __to_intrin(
2190 __make_vector<unsigned>(__factor_f[0], __factor_f[1],
2191 __factor_f[2], __factor_f[3]))
2192 : _mm_cvttps_epi32(__factor_f);
2193 const auto __r02
2194 = _mm_srli_epi64(_mm_mul_epu32(__ix, __factor), 31);
2195 const auto __r13 = _mm_mul_epu32(_mm_srli_si128(__ix, 4),
2196 _mm_srli_si128(__factor, 4));
2197 if constexpr (__have_sse4_1)
2198 return __intrin_bitcast<_V>(
2199 _mm_blend_epi16(_mm_slli_epi64(__r13, 1), __r02, 0x33));
2200 else
2201 return __intrin_bitcast<_V>(
2202 __r02 | _mm_slli_si128(_mm_srli_epi64(__r13, 31), 4));
2203 }
2204 else
2205 {
2206 auto __shift = [](auto __a, auto __b) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2207 if constexpr (is_signed_v<_Up>)
2208 return _mm_sra_epi32(__a, __b);
2209 else
2210 return _mm_srl_epi32(__a, __b);
2211 };
2212 const auto __r0
2213 = __shift(__ix, _mm_unpacklo_epi32(__iy, __m128i()));
2214 const auto __r1 = __shift(__ix, _mm_srli_epi64(__iy, 32));
2215 const auto __r2
2216 = __shift(__ix, _mm_unpackhi_epi32(__iy, __m128i()));
2217 const auto __r3 = __shift(__ix, _mm_srli_si128(__iy, 12));
2218 if constexpr (__have_sse4_1)
2219 return __intrin_bitcast<_V>(
2220 _mm_blend_epi16(_mm_blend_epi16(__r1, __r0, 0x3),
2221 _mm_blend_epi16(__r3, __r2, 0x30), 0xf0));
2222 else
2223 return __intrin_bitcast<_V>(_mm_unpacklo_epi64(
2224 _mm_unpacklo_epi32(__r0, _mm_srli_si128(__r1, 4)),
2225 _mm_unpackhi_epi32(__r2, _mm_srli_si128(__r3, 4))));
2226 }
2227 } //}}}
2228 else
2229 return __x >> __y;
2230 }
2231#endif // _GLIBCXX_SIMD_NO_SHIFT_OPT
2232
2233 // }}}
2234 // compares {{{
2235 // _S_equal_to {{{
2236 template <typename _Tp, size_t _Np>
2237 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
2238 _S_equal_to(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
2239 {
2240 if constexpr (__is_avx512_abi<_Abi>()) // {{{
2241 {
2242 if (__builtin_is_constant_evaluated()
2243 || (__x._M_is_constprop() && __y._M_is_constprop()))
2244 return _MaskImpl::_S_to_bits(
2245 __as_wrapper<_Np>(__x._M_data == __y._M_data));
2246
2247 constexpr auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
2248 [[maybe_unused]] const auto __xi = __to_intrin(__x);
2249 [[maybe_unused]] const auto __yi = __to_intrin(__y);
2250 if constexpr (is_floating_point_v<_Tp>)
2251 {
2252 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
2253 return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_EQ_OQ);
2254 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
2255 return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_EQ_OQ);
2256 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
2257 return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_EQ_OQ);
2258 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
2259 return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_EQ_OQ);
2260 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
2261 return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_EQ_OQ);
2262 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
2263 return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_EQ_OQ);
2264 else
2265 __assert_unreachable<_Tp>();
2266 }
2267 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
2268 return _mm512_mask_cmpeq_epi64_mask(__k1, __xi, __yi);
2269 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
2270 return _mm512_mask_cmpeq_epi32_mask(__k1, __xi, __yi);
2271 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 2)
2272 return _mm512_mask_cmpeq_epi16_mask(__k1, __xi, __yi);
2273 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 1)
2274 return _mm512_mask_cmpeq_epi8_mask(__k1, __xi, __yi);
2275 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
2276 return _mm256_mask_cmpeq_epi64_mask(__k1, __xi, __yi);
2277 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
2278 return _mm256_mask_cmpeq_epi32_mask(__k1, __xi, __yi);
2279 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 2)
2280 return _mm256_mask_cmpeq_epi16_mask(__k1, __xi, __yi);
2281 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 1)
2282 return _mm256_mask_cmpeq_epi8_mask(__k1, __xi, __yi);
2283 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
2284 return _mm_mask_cmpeq_epi64_mask(__k1, __xi, __yi);
2285 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
2286 return _mm_mask_cmpeq_epi32_mask(__k1, __xi, __yi);
2287 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 2)
2288 return _mm_mask_cmpeq_epi16_mask(__k1, __xi, __yi);
2289 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 1)
2290 return _mm_mask_cmpeq_epi8_mask(__k1, __xi, __yi);
2291 else
2292 __assert_unreachable<_Tp>();
2293 } // }}}
2294 else if (__builtin_is_constant_evaluated())
2295 return _Base::_S_equal_to(__x, __y);
2296 else if constexpr (sizeof(__x) == 8)
2297 {
2298 const auto __r128 = __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__x)
2299 == __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__y);
2300 _MaskMember<_Tp> __r64{};
2301 __builtin_memcpy(&__r64._M_data, &__r128, sizeof(__r64));
2302 return __r64;
2303 }
2304 else
2305 return _Base::_S_equal_to(__x, __y);
2306 }
2307
2308 // }}}
2309 // _S_not_equal_to {{{
2310 template <typename _Tp, size_t _Np>
2311 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
2312 _S_not_equal_to(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
2313 {
2314 if constexpr (__is_avx512_abi<_Abi>()) // {{{
2315 {
2316 if (__builtin_is_constant_evaluated()
2317 || (__x._M_is_constprop() && __y._M_is_constprop()))
2318 return _MaskImpl::_S_to_bits(
2319 __as_wrapper<_Np>(__x._M_data != __y._M_data));
2320
2321 constexpr auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
2322 [[maybe_unused]] const auto __xi = __to_intrin(__x);
2323 [[maybe_unused]] const auto __yi = __to_intrin(__y);
2324 if constexpr (is_floating_point_v<_Tp>)
2325 {
2326 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
2327 return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_NEQ_UQ);
2328 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
2329 return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_NEQ_UQ);
2330 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
2331 return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_NEQ_UQ);
2332 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
2333 return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_NEQ_UQ);
2334 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
2335 return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_NEQ_UQ);
2336 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
2337 return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_NEQ_UQ);
2338 else
2339 __assert_unreachable<_Tp>();
2340 }
2341 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
2342 return _mm512_mask_cmpneq_epi64_mask(__k1, __xi, __yi);
2343 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
2344 return _mm512_mask_cmpneq_epi32_mask(__k1, __xi, __yi);
2345 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 2)
2346 return _mm512_mask_cmpneq_epi16_mask(__k1, __xi, __yi);
2347 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 1)
2348 return _mm512_mask_cmpneq_epi8_mask(__k1, __xi, __yi);
2349 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
2350 return _mm256_mask_cmpneq_epi64_mask(__k1, __xi, __yi);
2351 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
2352 return _mm256_mask_cmpneq_epi32_mask(__k1, __xi, __yi);
2353 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 2)
2354 return _mm256_mask_cmpneq_epi16_mask(__k1, __xi, __yi);
2355 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 1)
2356 return _mm256_mask_cmpneq_epi8_mask(__k1, __xi, __yi);
2357 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
2358 return _mm_mask_cmpneq_epi64_mask(__k1, __xi, __yi);
2359 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
2360 return _mm_mask_cmpneq_epi32_mask(__k1, __xi, __yi);
2361 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 2)
2362 return _mm_mask_cmpneq_epi16_mask(__k1, __xi, __yi);
2363 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 1)
2364 return _mm_mask_cmpneq_epi8_mask(__k1, __xi, __yi);
2365 else
2366 __assert_unreachable<_Tp>();
2367 } // }}}
2368 else if (__builtin_is_constant_evaluated())
2369 return _Base::_S_not_equal_to(__x, __y);
2370 else if constexpr (sizeof(__x) == 8)
2371 {
2372 const auto __r128 = __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__x)
2373 != __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__y);
2374 _MaskMember<_Tp> __r64{};
2375 __builtin_memcpy(&__r64._M_data, &__r128, sizeof(__r64));
2376 return __r64;
2377 }
2378 else
2379 return _Base::_S_not_equal_to(__x, __y);
2380 }
2381
2382 // }}}
2383 // _S_less {{{
2384 template <typename _Tp, size_t _Np>
2385 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
2386 _S_less(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
2387 {
2388 if constexpr (__is_avx512_abi<_Abi>()) // {{{
2389 {
2390 if (__builtin_is_constant_evaluated()
2391 || (__x._M_is_constprop() && __y._M_is_constprop()))
2392 return _MaskImpl::_S_to_bits(
2393 __as_wrapper<_Np>(__x._M_data < __y._M_data));
2394
2395 constexpr auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
2396 [[maybe_unused]] const auto __xi = __to_intrin(__x);
2397 [[maybe_unused]] const auto __yi = __to_intrin(__y);
2398 if constexpr (sizeof(__xi) == 64)
2399 {
2400 if constexpr (is_same_v<_Tp, float>)
2401 return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OS);
2402 else if constexpr (is_same_v<_Tp, double>)
2403 return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OS);
2404 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 1)
2405 return _mm512_mask_cmplt_epi8_mask(__k1, __xi, __yi);
2406 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 2)
2407 return _mm512_mask_cmplt_epi16_mask(__k1, __xi, __yi);
2408 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 4)
2409 return _mm512_mask_cmplt_epi32_mask(__k1, __xi, __yi);
2410 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 8)
2411 return _mm512_mask_cmplt_epi64_mask(__k1, __xi, __yi);
2412 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 1)
2413 return _mm512_mask_cmplt_epu8_mask(__k1, __xi, __yi);
2414 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 2)
2415 return _mm512_mask_cmplt_epu16_mask(__k1, __xi, __yi);
2416 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 4)
2417 return _mm512_mask_cmplt_epu32_mask(__k1, __xi, __yi);
2418 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 8)
2419 return _mm512_mask_cmplt_epu64_mask(__k1, __xi, __yi);
2420 else
2421 __assert_unreachable<_Tp>();
2422 }
2423 else if constexpr (sizeof(__xi) == 32)
2424 {
2425 if constexpr (is_same_v<_Tp, float>)
2426 return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OS);
2427 else if constexpr (is_same_v<_Tp, double>)
2428 return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OS);
2429 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 1)
2430 return _mm256_mask_cmplt_epi8_mask(__k1, __xi, __yi);
2431 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 2)
2432 return _mm256_mask_cmplt_epi16_mask(__k1, __xi, __yi);
2433 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 4)
2434 return _mm256_mask_cmplt_epi32_mask(__k1, __xi, __yi);
2435 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 8)
2436 return _mm256_mask_cmplt_epi64_mask(__k1, __xi, __yi);
2437 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 1)
2438 return _mm256_mask_cmplt_epu8_mask(__k1, __xi, __yi);
2439 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 2)
2440 return _mm256_mask_cmplt_epu16_mask(__k1, __xi, __yi);
2441 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 4)
2442 return _mm256_mask_cmplt_epu32_mask(__k1, __xi, __yi);
2443 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 8)
2444 return _mm256_mask_cmplt_epu64_mask(__k1, __xi, __yi);
2445 else
2446 __assert_unreachable<_Tp>();
2447 }
2448 else if constexpr (sizeof(__xi) == 16)
2449 {
2450 if constexpr (is_same_v<_Tp, float>)
2451 return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OS);
2452 else if constexpr (is_same_v<_Tp, double>)
2453 return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OS);
2454 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 1)
2455 return _mm_mask_cmplt_epi8_mask(__k1, __xi, __yi);
2456 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 2)
2457 return _mm_mask_cmplt_epi16_mask(__k1, __xi, __yi);
2458 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 4)
2459 return _mm_mask_cmplt_epi32_mask(__k1, __xi, __yi);
2460 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 8)
2461 return _mm_mask_cmplt_epi64_mask(__k1, __xi, __yi);
2462 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 1)
2463 return _mm_mask_cmplt_epu8_mask(__k1, __xi, __yi);
2464 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 2)
2465 return _mm_mask_cmplt_epu16_mask(__k1, __xi, __yi);
2466 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 4)
2467 return _mm_mask_cmplt_epu32_mask(__k1, __xi, __yi);
2468 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 8)
2469 return _mm_mask_cmplt_epu64_mask(__k1, __xi, __yi);
2470 else
2471 __assert_unreachable<_Tp>();
2472 }
2473 else
2474 __assert_unreachable<_Tp>();
2475 } // }}}
2476 else if (__builtin_is_constant_evaluated())
2477 return _Base::_S_less(__x, __y);
2478 else if constexpr (sizeof(__x) == 8)
2479 {
2480 const auto __r128 = __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__x)
2481 < __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__y);
2482 _MaskMember<_Tp> __r64{};
2483 __builtin_memcpy(&__r64._M_data, &__r128, sizeof(__r64));
2484 return __r64;
2485 }
2486 else
2487 return _Base::_S_less(__x, __y);
2488 }
2489
2490 // }}}
2491 // _S_less_equal {{{
2492 template <typename _Tp, size_t _Np>
2493 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
2494 _S_less_equal(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
2495 {
2496 if constexpr (__is_avx512_abi<_Abi>()) // {{{
2497 {
2498 if (__builtin_is_constant_evaluated()
2499 || (__x._M_is_constprop() && __y._M_is_constprop()))
2500 return _MaskImpl::_S_to_bits(
2501 __as_wrapper<_Np>(__x._M_data <= __y._M_data));
2502
2503 constexpr auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
2504 [[maybe_unused]] const auto __xi = __to_intrin(__x);
2505 [[maybe_unused]] const auto __yi = __to_intrin(__y);
2506 if constexpr (sizeof(__xi) == 64)
2507 {
2508 if constexpr (is_same_v<_Tp, float>)
2509 return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OS);
2510 else if constexpr (is_same_v<_Tp, double>)
2511 return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OS);
2512 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 1)
2513 return _mm512_mask_cmple_epi8_mask(__k1, __xi, __yi);
2514 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 2)
2515 return _mm512_mask_cmple_epi16_mask(__k1, __xi, __yi);
2516 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 4)
2517 return _mm512_mask_cmple_epi32_mask(__k1, __xi, __yi);
2518 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 8)
2519 return _mm512_mask_cmple_epi64_mask(__k1, __xi, __yi);
2520 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 1)
2521 return _mm512_mask_cmple_epu8_mask(__k1, __xi, __yi);
2522 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 2)
2523 return _mm512_mask_cmple_epu16_mask(__k1, __xi, __yi);
2524 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 4)
2525 return _mm512_mask_cmple_epu32_mask(__k1, __xi, __yi);
2526 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 8)
2527 return _mm512_mask_cmple_epu64_mask(__k1, __xi, __yi);
2528 else
2529 __assert_unreachable<_Tp>();
2530 }
2531 else if constexpr (sizeof(__xi) == 32)
2532 {
2533 if constexpr (is_same_v<_Tp, float>)
2534 return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OS);
2535 else if constexpr (is_same_v<_Tp, double>)
2536 return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OS);
2537 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 1)
2538 return _mm256_mask_cmple_epi8_mask(__k1, __xi, __yi);
2539 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 2)
2540 return _mm256_mask_cmple_epi16_mask(__k1, __xi, __yi);
2541 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 4)
2542 return _mm256_mask_cmple_epi32_mask(__k1, __xi, __yi);
2543 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 8)
2544 return _mm256_mask_cmple_epi64_mask(__k1, __xi, __yi);
2545 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 1)
2546 return _mm256_mask_cmple_epu8_mask(__k1, __xi, __yi);
2547 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 2)
2548 return _mm256_mask_cmple_epu16_mask(__k1, __xi, __yi);
2549 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 4)
2550 return _mm256_mask_cmple_epu32_mask(__k1, __xi, __yi);
2551 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 8)
2552 return _mm256_mask_cmple_epu64_mask(__k1, __xi, __yi);
2553 else
2554 __assert_unreachable<_Tp>();
2555 }
2556 else if constexpr (sizeof(__xi) == 16)
2557 {
2558 if constexpr (is_same_v<_Tp, float>)
2559 return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OS);
2560 else if constexpr (is_same_v<_Tp, double>)
2561 return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OS);
2562 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 1)
2563 return _mm_mask_cmple_epi8_mask(__k1, __xi, __yi);
2564 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 2)
2565 return _mm_mask_cmple_epi16_mask(__k1, __xi, __yi);
2566 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 4)
2567 return _mm_mask_cmple_epi32_mask(__k1, __xi, __yi);
2568 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 8)
2569 return _mm_mask_cmple_epi64_mask(__k1, __xi, __yi);
2570 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 1)
2571 return _mm_mask_cmple_epu8_mask(__k1, __xi, __yi);
2572 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 2)
2573 return _mm_mask_cmple_epu16_mask(__k1, __xi, __yi);
2574 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 4)
2575 return _mm_mask_cmple_epu32_mask(__k1, __xi, __yi);
2576 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 8)
2577 return _mm_mask_cmple_epu64_mask(__k1, __xi, __yi);
2578 else
2579 __assert_unreachable<_Tp>();
2580 }
2581 else
2582 __assert_unreachable<_Tp>();
2583 } // }}}
2584 else if (__builtin_is_constant_evaluated())
2585 return _Base::_S_less_equal(__x, __y);
2586 else if constexpr (sizeof(__x) == 8)
2587 {
2588 const auto __r128 = __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__x)
2589 <= __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__y);
2590 _MaskMember<_Tp> __r64{};
2591 __builtin_memcpy(&__r64._M_data, &__r128, sizeof(__r64));
2592 return __r64;
2593 }
2594 else
2595 return _Base::_S_less_equal(__x, __y);
2596 }
2597
2598 // }}} }}}
2599 // negation {{{
2600 template <typename _Tp, size_t _Np>
2601 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
2602 _S_negate(_SimdWrapper<_Tp, _Np> __x) noexcept
2603 {
2604 if constexpr (__is_avx512_abi<_Abi>())
2605 return _S_equal_to(__x, _SimdWrapper<_Tp, _Np>());
2606 else
2607 return _Base::_S_negate(__x);
2608 }
2609
2610 // }}}
2611 // math {{{
2612 using _Base::_S_abs;
2613
2614 // _S_sqrt {{{
2615 template <typename _Tp, size_t _Np>
2616 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
2617 _S_sqrt(_SimdWrapper<_Tp, _Np> __x)
2618 {
2619 if constexpr (__is_sse_ps<_Tp, _Np>())
2620 return __auto_bitcast(_mm_sqrt_ps(__to_intrin(__x)));
2621 else if constexpr (__is_sse_pd<_Tp, _Np>())
2622 return _mm_sqrt_pd(__x);
2623 else if constexpr (__is_avx_ps<_Tp, _Np>())
2624 return _mm256_sqrt_ps(__x);
2625 else if constexpr (__is_avx_pd<_Tp, _Np>())
2626 return _mm256_sqrt_pd(__x);
2627 else if constexpr (__is_avx512_ps<_Tp, _Np>())
2628 return _mm512_sqrt_ps(__x);
2629 else if constexpr (__is_avx512_pd<_Tp, _Np>())
2630 return _mm512_sqrt_pd(__x);
2631 else
2632 __assert_unreachable<_Tp>();
2633 }
2634
2635 // }}}
2636 // _S_ldexp {{{
2637 template <typename _Tp, size_t _Np>
2638 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
2639 _S_ldexp(_SimdWrapper<_Tp, _Np> __x,
2640 __fixed_size_storage_t<int, _Np> __exp)
2641 {
2642 if constexpr (sizeof(__x) == 64 || __have_avx512vl)
2643 {
2644 const auto __xi = __to_intrin(__x);
2645 constexpr _SimdConverter<int, simd_abi::fixed_size<_Np>, _Tp, _Abi>
2646 __cvt;
2647 const auto __expi = __to_intrin(__cvt(__exp));
2648 using _Up = __bool_storage_member_type_t<_Np>;
2649 constexpr _Up __k1 = _Np < sizeof(_Up) * __CHAR_BIT__ ? _Up((1ULL << _Np) - 1) : ~_Up();
2650 if constexpr (sizeof(__xi) == 16)
2651 {
2652 if constexpr (sizeof(_Tp) == 8)
2653 return _mm_maskz_scalef_pd(__k1, __xi, __expi);
2654 else
2655 return _mm_maskz_scalef_ps(__k1, __xi, __expi);
2656 }
2657 else if constexpr (sizeof(__xi) == 32)
2658 {
2659 if constexpr (sizeof(_Tp) == 8)
2660 return _mm256_maskz_scalef_pd(__k1, __xi, __expi);
2661 else
2662 return _mm256_maskz_scalef_ps(__k1, __xi, __expi);
2663 }
2664 else
2665 {
2666 static_assert(sizeof(__xi) == 64);
2667 if constexpr (sizeof(_Tp) == 8)
2668 return _mm512_maskz_scalef_pd(__k1, __xi, __expi);
2669 else
2670 return _mm512_maskz_scalef_ps(__k1, __xi, __expi);
2671 }
2672 }
2673 else
2674 return _Base::_S_ldexp(__x, __exp);
2675 }
2676
2677 // }}}
2678 // _S_trunc {{{
2679 template <typename _Tp, size_t _Np>
2680 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
2681 _S_trunc(_SimdWrapper<_Tp, _Np> __x)
2682 {
2683 if constexpr (__is_avx512_ps<_Tp, _Np>())
2684 return _mm512_roundscale_ps(__x, 0x0b);
2685 else if constexpr (__is_avx512_pd<_Tp, _Np>())
2686 return _mm512_roundscale_pd(__x, 0x0b);
2687 else if constexpr (__is_avx_ps<_Tp, _Np>())
2688 return _mm256_round_ps(__x, 0xb);
2689 else if constexpr (__is_avx_pd<_Tp, _Np>())
2690 return _mm256_round_pd(__x, 0xb);
2691 else if constexpr (__have_sse4_1 && __is_sse_ps<_Tp, _Np>())
2692 return __auto_bitcast(_mm_round_ps(__to_intrin(__x), 0xb));
2693 else if constexpr (__have_sse4_1 && __is_sse_pd<_Tp, _Np>())
2694 return _mm_round_pd(__x, 0xb);
2695 else if constexpr (__is_sse_ps<_Tp, _Np>())
2696 {
2697 auto __truncated
2698 = _mm_cvtepi32_ps(_mm_cvttps_epi32(__to_intrin(__x)));
2699 const auto __no_fractional_values
2700 = __vector_bitcast<int>(__vector_bitcast<_UInt>(__to_intrin(__x))
2701 & 0x7f800000u)
2702 < 0x4b000000; // the exponent is so large that no mantissa bits
2703 // signify fractional values (0x3f8 + 23*8 =
2704 // 0x4b0)
2705 return __no_fractional_values ? __truncated : __to_intrin(__x);
2706 }
2707 else
2708 return _Base::_S_trunc(__x);
2709 }
2710
2711 // }}}
2712 // _S_round {{{
2713 template <typename _Tp, size_t _Np>
2714 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
2715 _S_round(_SimdWrapper<_Tp, _Np> __x)
2716 {
2717 // Note that _MM_FROUND_TO_NEAREST_INT rounds ties to even, not away
2718 // from zero as required by std::round. Therefore this function is more
2719 // complicated.
2720 using _V = __vector_type_t<_Tp, _Np>;
2721 _V __truncated;
2722 if constexpr (__is_avx512_ps<_Tp, _Np>())
2723 __truncated = _mm512_roundscale_ps(__x._M_data, 0x0b);
2724 else if constexpr (__is_avx512_pd<_Tp, _Np>())
2725 __truncated = _mm512_roundscale_pd(__x._M_data, 0x0b);
2726 else if constexpr (__is_avx_ps<_Tp, _Np>())
2727 __truncated = _mm256_round_ps(__x._M_data,
2728 _MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC);
2729 else if constexpr (__is_avx_pd<_Tp, _Np>())
2730 __truncated = _mm256_round_pd(__x._M_data,
2731 _MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC);
2732 else if constexpr (__have_sse4_1 && __is_sse_ps<_Tp, _Np>())
2733 __truncated = __auto_bitcast(
2734 _mm_round_ps(__to_intrin(__x),
2735 _MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC));
2736 else if constexpr (__have_sse4_1 && __is_sse_pd<_Tp, _Np>())
2737 __truncated
2738 = _mm_round_pd(__x._M_data, _MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC);
2739 else if constexpr (__is_sse_ps<_Tp, _Np>())
2740 __truncated = __auto_bitcast(
2741 _mm_cvtepi32_ps(_mm_cvttps_epi32(__to_intrin(__x))));
2742 else
2743 return _Base::_S_round(__x);
2744
2745 // x < 0 => truncated <= 0 && truncated >= x => x - truncated <= 0
2746 // x > 0 => truncated >= 0 && truncated <= x => x - truncated >= 0
2747
2748 const _V __rounded
2749 = __truncated
2750 + (__and(_S_absmask<_V>, __x._M_data - __truncated) >= _Tp(.5)
2751 ? __or(__and(_S_signmask<_V>, __x._M_data), _V() + 1)
2752 : _V());
2753 if constexpr (__have_sse4_1)
2754 return __rounded;
2755 else // adjust for missing range in cvttps_epi32
2756 return __and(_S_absmask<_V>, __x._M_data) < 0x1p23f ? __rounded
2757 : __x._M_data;
2758 }
2759
2760 // }}}
2761 // _S_nearbyint {{{
2762 template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
2763 _GLIBCXX_SIMD_INTRINSIC static _Tp
2764 _S_nearbyint(_Tp __x) noexcept
2765 {
2766 if constexpr (_TVT::template _S_is<float, 16>)
2767 return _mm512_roundscale_ps(__x, 0x0c);
2768 else if constexpr (_TVT::template _S_is<double, 8>)
2769 return _mm512_roundscale_pd(__x, 0x0c);
2770 else if constexpr (_TVT::template _S_is<float, 8>)
2771 return _mm256_round_ps(__x,
2772 _MM_FROUND_CUR_DIRECTION | _MM_FROUND_NO_EXC);
2773 else if constexpr (_TVT::template _S_is<double, 4>)
2774 return _mm256_round_pd(__x,
2775 _MM_FROUND_CUR_DIRECTION | _MM_FROUND_NO_EXC);
2776 else if constexpr (__have_sse4_1 && _TVT::template _S_is<float, 4>)
2777 return _mm_round_ps(__x,
2778 _MM_FROUND_CUR_DIRECTION | _MM_FROUND_NO_EXC);
2779 else if constexpr (__have_sse4_1 && _TVT::template _S_is<double, 2>)
2780 return _mm_round_pd(__x,
2781 _MM_FROUND_CUR_DIRECTION | _MM_FROUND_NO_EXC);
2782 else
2783 return _Base::_S_nearbyint(__x);
2784 }
2785
2786 // }}}
2787 // _S_rint {{{
2788 template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
2789 _GLIBCXX_SIMD_INTRINSIC static _Tp
2790 _S_rint(_Tp __x) noexcept
2791 {
2792 if constexpr (_TVT::template _S_is<float, 16>)
2793 return _mm512_roundscale_ps(__x, 0x04);
2794 else if constexpr (_TVT::template _S_is<double, 8>)
2795 return _mm512_roundscale_pd(__x, 0x04);
2796 else if constexpr (_TVT::template _S_is<float, 8>)
2797 return _mm256_round_ps(__x, _MM_FROUND_CUR_DIRECTION);
2798 else if constexpr (_TVT::template _S_is<double, 4>)
2799 return _mm256_round_pd(__x, _MM_FROUND_CUR_DIRECTION);
2800 else if constexpr (__have_sse4_1 && _TVT::template _S_is<float, 4>)
2801 return _mm_round_ps(__x, _MM_FROUND_CUR_DIRECTION);
2802 else if constexpr (__have_sse4_1 && _TVT::template _S_is<double, 2>)
2803 return _mm_round_pd(__x, _MM_FROUND_CUR_DIRECTION);
2804 else
2805 return _Base::_S_rint(__x);
2806 }
2807
2808 // }}}
2809 // _S_floor {{{
2810 template <typename _Tp, size_t _Np>
2811 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
2812 _S_floor(_SimdWrapper<_Tp, _Np> __x)
2813 {
2814 if constexpr (__is_avx512_ps<_Tp, _Np>())
2815 return _mm512_roundscale_ps(__x, 0x09);
2816 else if constexpr (__is_avx512_pd<_Tp, _Np>())
2817 return _mm512_roundscale_pd(__x, 0x09);
2818 else if constexpr (__is_avx_ps<_Tp, _Np>())
2819 return _mm256_round_ps(__x, 0x9);
2820 else if constexpr (__is_avx_pd<_Tp, _Np>())
2821 return _mm256_round_pd(__x, 0x9);
2822 else if constexpr (__have_sse4_1 && __is_sse_ps<_Tp, _Np>())
2823 return __auto_bitcast(_mm_round_ps(__to_intrin(__x), 0x9));
2824 else if constexpr (__have_sse4_1 && __is_sse_pd<_Tp, _Np>())
2825 return _mm_round_pd(__x, 0x9);
2826 else
2827 return _Base::_S_floor(__x);
2828 }
2829
2830 // }}}
2831 // _S_ceil {{{
2832 template <typename _Tp, size_t _Np>
2833 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
2834 _S_ceil(_SimdWrapper<_Tp, _Np> __x)
2835 {
2836 if constexpr (__is_avx512_ps<_Tp, _Np>())
2837 return _mm512_roundscale_ps(__x, 0x0a);
2838 else if constexpr (__is_avx512_pd<_Tp, _Np>())
2839 return _mm512_roundscale_pd(__x, 0x0a);
2840 else if constexpr (__is_avx_ps<_Tp, _Np>())
2841 return _mm256_round_ps(__x, 0xa);
2842 else if constexpr (__is_avx_pd<_Tp, _Np>())
2843 return _mm256_round_pd(__x, 0xa);
2844 else if constexpr (__have_sse4_1 && __is_sse_ps<_Tp, _Np>())
2845 return __auto_bitcast(_mm_round_ps(__to_intrin(__x), 0xa));
2846 else if constexpr (__have_sse4_1 && __is_sse_pd<_Tp, _Np>())
2847 return _mm_round_pd(__x, 0xa);
2848 else
2849 return _Base::_S_ceil(__x);
2850 }
2851
2852 // }}}
2853 // _S_signbit {{{
2854 template <typename _Tp, size_t _Np>
2855 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
2856 _S_signbit(_SimdWrapper<_Tp, _Np> __x)
2857 {
2858 if constexpr (__is_avx512_abi<_Abi>() && __have_avx512dq)
2859 {
2860 if constexpr (sizeof(__x) == 64 && sizeof(_Tp) == 4)
2861 return _mm512_movepi32_mask(
2862 __intrin_bitcast<__m512i>(__x._M_data));
2863 else if constexpr (sizeof(__x) == 64 && sizeof(_Tp) == 8)
2864 return _mm512_movepi64_mask(
2865 __intrin_bitcast<__m512i>(__x._M_data));
2866 else if constexpr (sizeof(__x) == 32 && sizeof(_Tp) == 4)
2867 return _mm256_movepi32_mask(
2868 __intrin_bitcast<__m256i>(__x._M_data));
2869 else if constexpr (sizeof(__x) == 32 && sizeof(_Tp) == 8)
2870 return _mm256_movepi64_mask(
2871 __intrin_bitcast<__m256i>(__x._M_data));
2872 else if constexpr (sizeof(__x) <= 16 && sizeof(_Tp) == 4)
2873 return _mm_movepi32_mask(__intrin_bitcast<__m128i>(__x._M_data));
2874 else if constexpr (sizeof(__x) <= 16 && sizeof(_Tp) == 8)
2875 return _mm_movepi64_mask(__intrin_bitcast<__m128i>(__x._M_data));
2876 }
2877 else if constexpr (__is_avx512_abi<_Abi>())
2878 {
2879 const auto __xi = __to_intrin(__x);
2880 [[maybe_unused]] constexpr auto __k1
2881 = _Abi::template _S_implicit_mask_intrin<_Tp>();
2882 if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
2883 return _mm_movemask_ps(__xi);
2884 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
2885 return _mm_movemask_pd(__xi);
2886 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
2887 return _mm256_movemask_ps(__xi);
2888 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
2889 return _mm256_movemask_pd(__xi);
2890 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
2891 return _mm512_mask_cmplt_epi32_mask(
2892 __k1, __intrin_bitcast<__m512i>(__xi), __m512i());
2893 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
2894 return _mm512_mask_cmplt_epi64_mask(
2895 __k1, __intrin_bitcast<__m512i>(__xi), __m512i());
2896 else
2897 __assert_unreachable<_Tp>();
2898 }
2899 else
2900 return _Base::_S_signbit(__x);
2901 /*{
2902 using _I = __int_for_sizeof_t<_Tp>;
2903 if constexpr (sizeof(__x) == 64)
2904 return _S_less(__vector_bitcast<_I>(__x), _I());
2905 else
2906 {
2907 const auto __xx = __vector_bitcast<_I>(__x._M_data);
2908 [[maybe_unused]] constexpr _I __signmask = __finite_min_v<_I>;
2909 if constexpr ((sizeof(_Tp) == 4 &&
2910 (__have_avx2 || sizeof(__x) == 16)) ||
2911 __have_avx512vl)
2912 {
2913 return __vector_bitcast<_Tp>(__xx >> __digits_v<_I>);
2914 }
2915 else if constexpr ((__have_avx2 ||
2916 (__have_ssse3 && sizeof(__x) == 16)))
2917 {
2918 return __vector_bitcast<_Tp>((__xx & __signmask) ==
2919 __signmask);
2920 }
2921 else
2922 { // SSE2/3 or AVX (w/o AVX2)
2923 constexpr auto __one = __vector_broadcast<_Np, _Tp>(1);
2924 return __vector_bitcast<_Tp>(
2925 __vector_bitcast<_Tp>(
2926 (__xx & __signmask) |
2927 __vector_bitcast<_I>(__one)) // -1 or 1
2928 != __one);
2929 }
2930 }
2931 }*/
2932 }
2933
2934 // }}}
2935 // _S_isnonzerovalue_mask {{{
2936 // (isnormal | is subnormal == !isinf & !isnan & !is zero)
2937 template <typename _Tp>
2938 _GLIBCXX_SIMD_INTRINSIC static auto
2939 _S_isnonzerovalue_mask(_Tp __x)
2940 {
2941 using _Traits = _VectorTraits<_Tp>;
2942 if constexpr (__have_avx512dq_vl)
2943 {
2944 if constexpr (_Traits::template _S_is<
2945 float, 2> || _Traits::template _S_is<float, 4>)
2946 return _knot_mask8(_mm_fpclass_ps_mask(__to_intrin(__x), 0x9f));
2947 else if constexpr (_Traits::template _S_is<float, 8>)
2948 return _knot_mask8(_mm256_fpclass_ps_mask(__x, 0x9f));
2949 else if constexpr (_Traits::template _S_is<float, 16>)
2950 return _knot_mask16(_mm512_fpclass_ps_mask(__x, 0x9f));
2951 else if constexpr (_Traits::template _S_is<double, 2>)
2952 return _knot_mask8(_mm_fpclass_pd_mask(__x, 0x9f));
2953 else if constexpr (_Traits::template _S_is<double, 4>)
2954 return _knot_mask8(_mm256_fpclass_pd_mask(__x, 0x9f));
2955 else if constexpr (_Traits::template _S_is<double, 8>)
2956 return _knot_mask8(_mm512_fpclass_pd_mask(__x, 0x9f));
2957 else
2958 __assert_unreachable<_Tp>();
2959 }
2960 else
2961 {
2962 using _Up = typename _Traits::value_type;
2963 constexpr size_t _Np = _Traits::_S_full_size;
2964 const auto __a = __x * __infinity_v<_Up>; // NaN if __x == 0
2965 const auto __b = __x * _Up(); // NaN if __x == inf
2966 if constexpr (__have_avx512vl && __is_sse_ps<_Up, _Np>())
2967 return _mm_cmp_ps_mask(__to_intrin(__a), __to_intrin(__b),
2968 _CMP_ORD_Q);
2969 else if constexpr (__have_avx512f && __is_sse_ps<_Up, _Np>())
2970 return __mmask8(0xf
2971 & _mm512_cmp_ps_mask(__auto_bitcast(__a),
2972 __auto_bitcast(__b),
2973 _CMP_ORD_Q));
2974 else if constexpr (__have_avx512vl && __is_sse_pd<_Up, _Np>())
2975 return _mm_cmp_pd_mask(__a, __b, _CMP_ORD_Q);
2976 else if constexpr (__have_avx512f && __is_sse_pd<_Up, _Np>())
2977 return __mmask8(0x3
2978 & _mm512_cmp_pd_mask(__auto_bitcast(__a),
2979 __auto_bitcast(__b),
2980 _CMP_ORD_Q));
2981 else if constexpr (__have_avx512vl && __is_avx_ps<_Up, _Np>())
2982 return _mm256_cmp_ps_mask(__a, __b, _CMP_ORD_Q);
2983 else if constexpr (__have_avx512f && __is_avx_ps<_Up, _Np>())
2984 return __mmask8(_mm512_cmp_ps_mask(__auto_bitcast(__a),
2985 __auto_bitcast(__b),
2986 _CMP_ORD_Q));
2987 else if constexpr (__have_avx512vl && __is_avx_pd<_Up, _Np>())
2988 return _mm256_cmp_pd_mask(__a, __b, _CMP_ORD_Q);
2989 else if constexpr (__have_avx512f && __is_avx_pd<_Up, _Np>())
2990 return __mmask8(0xf
2991 & _mm512_cmp_pd_mask(__auto_bitcast(__a),
2992 __auto_bitcast(__b),
2993 _CMP_ORD_Q));
2994 else if constexpr (__is_avx512_ps<_Up, _Np>())
2995 return _mm512_cmp_ps_mask(__a, __b, _CMP_ORD_Q);
2996 else if constexpr (__is_avx512_pd<_Up, _Np>())
2997 return _mm512_cmp_pd_mask(__a, __b, _CMP_ORD_Q);
2998 else
2999 __assert_unreachable<_Tp>();
3000 }
3001 }
3002
3003 // }}}
3004 // _S_isfinite {{{
3005 template <typename _Tp, size_t _Np>
3006 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
3007 _S_isfinite(_SimdWrapper<_Tp, _Np> __x)
3008 {
3009 static_assert(is_floating_point_v<_Tp>);
3010#if !__FINITE_MATH_ONLY__
3011 if constexpr (__is_avx512_abi<_Abi>() && __have_avx512dq)
3012 {
3013 const auto __xi = __to_intrin(__x);
3014 constexpr auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
3015 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
3016 return __k1 ^ _mm512_mask_fpclass_ps_mask(__k1, __xi, 0x99);
3017 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
3018 return __k1 ^ _mm512_mask_fpclass_pd_mask(__k1, __xi, 0x99);
3019 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3020 return __k1 ^ _mm256_mask_fpclass_ps_mask(__k1, __xi, 0x99);
3021 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3022 return __k1 ^ _mm256_mask_fpclass_pd_mask(__k1, __xi, 0x99);
3023 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3024 return __k1 ^ _mm_mask_fpclass_ps_mask(__k1, __xi, 0x99);
3025 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3026 return __k1 ^ _mm_mask_fpclass_pd_mask(__k1, __xi, 0x99);
3027 }
3028 else if constexpr (__is_avx512_abi<_Abi>())
3029 {
3030 // if all exponent bits are set, __x is either inf or NaN
3031 using _I = __int_for_sizeof_t<_Tp>;
3032 const auto __inf = __vector_bitcast<_I>(
3033 __vector_broadcast<_Np>(__infinity_v<_Tp>));
3034 return _S_less<_I, _Np>(__vector_bitcast<_I>(__x) & __inf, __inf);
3035 }
3036 else
3037#endif
3038 return _Base::_S_isfinite(__x);
3039 }
3040
3041 // }}}
3042 // _S_isinf {{{
3043 template <typename _Tp, size_t _Np>
3044 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
3045 _S_isinf(_SimdWrapper<_Tp, _Np> __x)
3046 {
3047#if !__FINITE_MATH_ONLY__
3048 if constexpr (__is_avx512_abi<_Abi>() && __have_avx512dq)
3049 {
3050 const auto __xi = __to_intrin(__x);
3051 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
3052 return _mm512_fpclass_ps_mask(__xi, 0x18);
3053 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
3054 return _mm512_fpclass_pd_mask(__xi, 0x18);
3055 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3056 return _mm256_fpclass_ps_mask(__xi, 0x18);
3057 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3058 return _mm256_fpclass_pd_mask(__xi, 0x18);
3059 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3060 return _mm_fpclass_ps_mask(__xi, 0x18);
3061 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3062 return _mm_fpclass_pd_mask(__xi, 0x18);
3063 else
3064 __assert_unreachable<_Tp>();
3065 }
3066 else if constexpr (__have_avx512dq_vl)
3067 {
3068 if constexpr (__is_sse_pd<_Tp, _Np>())
3069 return _mm_movm_epi64(_mm_fpclass_pd_mask(__x, 0x18));
3070 else if constexpr (__is_avx_pd<_Tp, _Np>())
3071 return _mm256_movm_epi64(_mm256_fpclass_pd_mask(__x, 0x18));
3072 else if constexpr (__is_sse_ps<_Tp, _Np>())
3073 return _mm_movm_epi32(
3074 _mm_fpclass_ps_mask(__to_intrin(__x), 0x18));
3075 else if constexpr (__is_avx_ps<_Tp, _Np>())
3076 return _mm256_movm_epi32(_mm256_fpclass_ps_mask(__x, 0x18));
3077 else
3078 __assert_unreachable<_Tp>();
3079 }
3080 else
3081#endif
3082 return _Base::_S_isinf(__x);
3083 }
3084
3085 // }}}
3086 // _S_isnormal {{{
3087 template <typename _Tp, size_t _Np>
3088 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
3089 _S_isnormal(_SimdWrapper<_Tp, _Np> __x)
3090 {
3091#if __FINITE_MATH_ONLY__
3092 [[maybe_unused]] constexpr int __mode = 0x26;
3093#else
3094 [[maybe_unused]] constexpr int __mode = 0xbf;
3095#endif
3096 if constexpr (__is_avx512_abi<_Abi>() && __have_avx512dq)
3097 {
3098 const auto __xi = __to_intrin(__x);
3099 const auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
3100 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
3101 return __k1 ^ _mm512_mask_fpclass_ps_mask(__k1, __xi, __mode);
3102 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
3103 return __k1 ^ _mm512_mask_fpclass_pd_mask(__k1, __xi, __mode);
3104 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3105 return __k1 ^ _mm256_mask_fpclass_ps_mask(__k1, __xi, __mode);
3106 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3107 return __k1 ^ _mm256_mask_fpclass_pd_mask(__k1, __xi, __mode);
3108 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3109 return __k1 ^ _mm_mask_fpclass_ps_mask(__k1, __xi, __mode);
3110 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3111 return __k1 ^ _mm_mask_fpclass_pd_mask(__k1, __xi, __mode);
3112 else
3113 __assert_unreachable<_Tp>();
3114 }
3115 else if constexpr (__have_avx512dq)
3116 {
3117 if constexpr (__have_avx512vl && __is_sse_ps<_Tp, _Np>())
3118 return _mm_movm_epi32(
3119 _knot_mask8(_mm_fpclass_ps_mask(__to_intrin(__x), __mode)));
3120 else if constexpr (__have_avx512vl && __is_avx_ps<_Tp, _Np>())
3121 return _mm256_movm_epi32(
3122 _knot_mask8(_mm256_fpclass_ps_mask(__x, __mode)));
3123 else if constexpr (__is_avx512_ps<_Tp, _Np>())
3124 return _knot_mask16(_mm512_fpclass_ps_mask(__x, __mode));
3125 else if constexpr (__have_avx512vl && __is_sse_pd<_Tp, _Np>())
3126 return _mm_movm_epi64(
3127 _knot_mask8(_mm_fpclass_pd_mask(__x, __mode)));
3128 else if constexpr (__have_avx512vl && __is_avx_pd<_Tp, _Np>())
3129 return _mm256_movm_epi64(
3130 _knot_mask8(_mm256_fpclass_pd_mask(__x, __mode)));
3131 else if constexpr (__is_avx512_pd<_Tp, _Np>())
3132 return _knot_mask8(_mm512_fpclass_pd_mask(__x, __mode));
3133 else
3134 __assert_unreachable<_Tp>();
3135 }
3136 else if constexpr (__is_avx512_abi<_Abi>())
3137 {
3138 using _I = __int_for_sizeof_t<_Tp>;
3139 const auto absn = __vector_bitcast<_I>(_S_abs(__x));
3140 const auto minn = __vector_bitcast<_I>(
3141 __vector_broadcast<_Np>(__norm_min_v<_Tp>));
3142#if __FINITE_MATH_ONLY__
3143 return _S_less_equal<_I, _Np>(minn, absn);
3144#else
3145 const auto infn
3146 = __vector_bitcast<_I>(__vector_broadcast<_Np>(__infinity_v<_Tp>));
3147 return __and(_S_less_equal<_I, _Np>(minn, absn),
3148 _S_less<_I, _Np>(absn, infn));
3149#endif
3150 }
3151 else
3152 return _Base::_S_isnormal(__x);
3153 }
3154
3155 // }}}
3156 // _S_isnan {{{
3157 template <typename _Tp, size_t _Np>
3158 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
3159 _S_isnan(_SimdWrapper<_Tp, _Np> __x)
3160 { return _S_isunordered(__x, __x); }
3161
3162 // }}}
3163 // _S_isunordered {{{
3164 template <typename _Tp, size_t _Np>
3165 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
3166 _S_isunordered([[maybe_unused]] _SimdWrapper<_Tp, _Np> __x,
3167 [[maybe_unused]] _SimdWrapper<_Tp, _Np> __y)
3168 {
3169#if __FINITE_MATH_ONLY__
3170 return {}; // false
3171#else
3172 const auto __xi = __to_intrin(__x);
3173 const auto __yi = __to_intrin(__y);
3174 if constexpr (__is_avx512_abi<_Abi>())
3175 {
3176 constexpr auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
3177 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
3178 return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_UNORD_Q);
3179 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
3180 return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_UNORD_Q);
3181 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3182 return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_UNORD_Q);
3183 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3184 return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_UNORD_Q);
3185 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3186 return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_UNORD_Q);
3187 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3188 return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_UNORD_Q);
3189 }
3190 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3191 return __to_masktype(_mm256_cmp_ps(__xi, __yi, _CMP_UNORD_Q));
3192 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3193 return __to_masktype(_mm256_cmp_pd(__xi, __yi, _CMP_UNORD_Q));
3194 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3195 return __auto_bitcast(_mm_cmpunord_ps(__xi, __yi));
3196 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3197 return __to_masktype(_mm_cmpunord_pd(__xi, __yi));
3198 else
3199 __assert_unreachable<_Tp>();
3200#endif
3201 }
3202
3203 // }}}
3204 // _S_isgreater {{{
3205 template <typename _Tp, size_t _Np>
3206 static constexpr _MaskMember<_Tp>
3207 _S_isgreater(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
3208 {
3209 const auto __xi = __to_intrin(__x);
3210 const auto __yi = __to_intrin(__y);
3211 if constexpr (__is_avx512_abi<_Abi>())
3212 {
3213 const auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
3214 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
3215 return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_GT_OQ);
3216 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
3217 return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_GT_OQ);
3218 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3219 return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_GT_OQ);
3220 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3221 return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_GT_OQ);
3222 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3223 return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_GT_OQ);
3224 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3225 return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_GT_OQ);
3226 else
3227 __assert_unreachable<_Tp>();
3228 }
3229 else if constexpr (__have_avx)
3230 {
3231 if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3232 return __to_masktype(_mm256_cmp_ps(__xi, __yi, _CMP_GT_OQ));
3233 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3234 return __to_masktype(_mm256_cmp_pd(__xi, __yi, _CMP_GT_OQ));
3235 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3236 return __auto_bitcast(_mm_cmp_ps(__xi, __yi, _CMP_GT_OQ));
3237 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3238 return __to_masktype(_mm_cmp_pd(__xi, __yi, _CMP_GT_OQ));
3239 else
3240 __assert_unreachable<_Tp>();
3241 }
3242 else if constexpr (__have_sse2 && sizeof(__xi) == 16
3243 && sizeof(_Tp) == 4)
3244 {
3245 const auto __xn = __vector_bitcast<int>(__xi);
3246 const auto __yn = __vector_bitcast<int>(__yi);
3247 const auto __xp = __xn < 0 ? -(__xn & 0x7fff'ffff) : __xn;
3248 const auto __yp = __yn < 0 ? -(__yn & 0x7fff'ffff) : __yn;
3249 return __auto_bitcast(
3250 __and(__to_masktype(_mm_cmpord_ps(__xi, __yi)), __xp > __yp));
3251 }
3252 else if constexpr (__have_sse2 && sizeof(__xi) == 16
3253 && sizeof(_Tp) == 8)
3254 return __vector_type_t<__int_with_sizeof_t<8>, 2>{
3255 -_mm_ucomigt_sd(__xi, __yi),
3256 -_mm_ucomigt_sd(_mm_unpackhi_pd(__xi, __xi),
3257 _mm_unpackhi_pd(__yi, __yi))};
3258 else
3259 return _Base::_S_isgreater(__x, __y);
3260 }
3261
3262 // }}}
3263 // _S_isgreaterequal {{{
3264 template <typename _Tp, size_t _Np>
3265 static constexpr _MaskMember<_Tp>
3266 _S_isgreaterequal(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
3267 {
3268 const auto __xi = __to_intrin(__x);
3269 const auto __yi = __to_intrin(__y);
3270 if constexpr (__is_avx512_abi<_Abi>())
3271 {
3272 const auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
3273 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
3274 return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_GE_OQ);
3275 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
3276 return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_GE_OQ);
3277 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3278 return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_GE_OQ);
3279 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3280 return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_GE_OQ);
3281 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3282 return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_GE_OQ);
3283 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3284 return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_GE_OQ);
3285 else
3286 __assert_unreachable<_Tp>();
3287 }
3288 else if constexpr (__have_avx)
3289 {
3290 if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3291 return __to_masktype(_mm256_cmp_ps(__xi, __yi, _CMP_GE_OQ));
3292 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3293 return __to_masktype(_mm256_cmp_pd(__xi, __yi, _CMP_GE_OQ));
3294 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3295 return __auto_bitcast(_mm_cmp_ps(__xi, __yi, _CMP_GE_OQ));
3296 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3297 return __to_masktype(_mm_cmp_pd(__xi, __yi, _CMP_GE_OQ));
3298 else
3299 __assert_unreachable<_Tp>();
3300 }
3301 else if constexpr (__have_sse2 && sizeof(__xi) == 16
3302 && sizeof(_Tp) == 4)
3303 {
3304 const auto __xn = __vector_bitcast<int>(__xi);
3305 const auto __yn = __vector_bitcast<int>(__yi);
3306 const auto __xp = __xn < 0 ? -(__xn & 0x7fff'ffff) : __xn;
3307 const auto __yp = __yn < 0 ? -(__yn & 0x7fff'ffff) : __yn;
3308 return __auto_bitcast(
3309 __and(__to_masktype(_mm_cmpord_ps(__xi, __yi)), __xp >= __yp));
3310 }
3311 else if constexpr (__have_sse2 && sizeof(__xi) == 16
3312 && sizeof(_Tp) == 8)
3313 return __vector_type_t<__int_with_sizeof_t<8>, 2>{
3314 -_mm_ucomige_sd(__xi, __yi),
3315 -_mm_ucomige_sd(_mm_unpackhi_pd(__xi, __xi),
3316 _mm_unpackhi_pd(__yi, __yi))};
3317 else
3318 return _Base::_S_isgreaterequal(__x, __y);
3319 }
3320
3321 // }}}
3322 // _S_isless {{{
3323 template <typename _Tp, size_t _Np>
3324 static constexpr _MaskMember<_Tp>
3325 _S_isless(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
3326 {
3327 const auto __xi = __to_intrin(__x);
3328 const auto __yi = __to_intrin(__y);
3329 if constexpr (__is_avx512_abi<_Abi>())
3330 {
3331 const auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
3332 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
3333 return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OQ);
3334 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
3335 return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OQ);
3336 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3337 return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OQ);
3338 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3339 return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OQ);
3340 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3341 return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OQ);
3342 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3343 return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OQ);
3344 else
3345 __assert_unreachable<_Tp>();
3346 }
3347 else if constexpr (__have_avx)
3348 {
3349 if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3350 return __to_masktype(_mm256_cmp_ps(__xi, __yi, _CMP_LT_OQ));
3351 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3352 return __to_masktype(_mm256_cmp_pd(__xi, __yi, _CMP_LT_OQ));
3353 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3354 return __auto_bitcast(_mm_cmp_ps(__xi, __yi, _CMP_LT_OQ));
3355 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3356 return __to_masktype(_mm_cmp_pd(__xi, __yi, _CMP_LT_OQ));
3357 else
3358 __assert_unreachable<_Tp>();
3359 }
3360 else if constexpr (__have_sse2 && sizeof(__xi) == 16
3361 && sizeof(_Tp) == 4)
3362 {
3363 const auto __xn = __vector_bitcast<int>(__xi);
3364 const auto __yn = __vector_bitcast<int>(__yi);
3365 const auto __xp = __xn < 0 ? -(__xn & 0x7fff'ffff) : __xn;
3366 const auto __yp = __yn < 0 ? -(__yn & 0x7fff'ffff) : __yn;
3367 return __auto_bitcast(
3368 __and(__to_masktype(_mm_cmpord_ps(__xi, __yi)), __xp < __yp));
3369 }
3370 else if constexpr (__have_sse2 && sizeof(__xi) == 16
3371 && sizeof(_Tp) == 8)
3372 return __vector_type_t<__int_with_sizeof_t<8>, 2>{
3373 -_mm_ucomigt_sd(__yi, __xi),
3374 -_mm_ucomigt_sd(_mm_unpackhi_pd(__yi, __yi),
3375 _mm_unpackhi_pd(__xi, __xi))};
3376 else
3377 return _Base::_S_isless(__x, __y);
3378 }
3379
3380 // }}}
3381 // _S_islessequal {{{
3382 template <typename _Tp, size_t _Np>
3383 static constexpr _MaskMember<_Tp>
3384 _S_islessequal(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
3385 {
3386 const auto __xi = __to_intrin(__x);
3387 const auto __yi = __to_intrin(__y);
3388 if constexpr (__is_avx512_abi<_Abi>())
3389 {
3390 const auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
3391 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
3392 return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OQ);
3393 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
3394 return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OQ);
3395 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3396 return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OQ);
3397 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3398 return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OQ);
3399 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3400 return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OQ);
3401 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3402 return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OQ);
3403 else
3404 __assert_unreachable<_Tp>();
3405 }
3406 else if constexpr (__have_avx)
3407 {
3408 if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3409 return __to_masktype(_mm256_cmp_ps(__xi, __yi, _CMP_LE_OQ));
3410 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3411 return __to_masktype(_mm256_cmp_pd(__xi, __yi, _CMP_LE_OQ));
3412 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3413 return __auto_bitcast(_mm_cmp_ps(__xi, __yi, _CMP_LE_OQ));
3414 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3415 return __to_masktype(_mm_cmp_pd(__xi, __yi, _CMP_LE_OQ));
3416 else
3417 __assert_unreachable<_Tp>();
3418 }
3419 else if constexpr (__have_sse2 && sizeof(__xi) == 16
3420 && sizeof(_Tp) == 4)
3421 {
3422 const auto __xn = __vector_bitcast<int>(__xi);
3423 const auto __yn = __vector_bitcast<int>(__yi);
3424 const auto __xp = __xn < 0 ? -(__xn & 0x7fff'ffff) : __xn;
3425 const auto __yp = __yn < 0 ? -(__yn & 0x7fff'ffff) : __yn;
3426 return __auto_bitcast(
3427 __and(__to_masktype(_mm_cmpord_ps(__xi, __yi)), __xp <= __yp));
3428 }
3429 else if constexpr (__have_sse2 && sizeof(__xi) == 16
3430 && sizeof(_Tp) == 8)
3431 return __vector_type_t<__int_with_sizeof_t<8>, 2>{
3432 -_mm_ucomige_sd(__yi, __xi),
3433 -_mm_ucomige_sd(_mm_unpackhi_pd(__yi, __yi),
3434 _mm_unpackhi_pd(__xi, __xi))};
3435 else
3436 return _Base::_S_islessequal(__x, __y);
3437 }
3438
3439 // }}}
3440 // _S_islessgreater {{{
3441 template <typename _Tp, size_t _Np>
3442 static constexpr _MaskMember<_Tp>
3443 _S_islessgreater(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
3444 {
3445 const auto __xi = __to_intrin(__x);
3446 const auto __yi = __to_intrin(__y);
3447 if constexpr (__is_avx512_abi<_Abi>())
3448 {
3449 const auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
3450 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
3451 return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_NEQ_OQ);
3452 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
3453 return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_NEQ_OQ);
3454 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3455 return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_NEQ_OQ);
3456 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3457 return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_NEQ_OQ);
3458 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3459 return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_NEQ_OQ);
3460 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3461 return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_NEQ_OQ);
3462 else
3463 __assert_unreachable<_Tp>();
3464 }
3465 else if constexpr (__have_avx)
3466 {
3467 if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3468 return __to_masktype(_mm256_cmp_ps(__xi, __yi, _CMP_NEQ_OQ));
3469 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3470 return __to_masktype(_mm256_cmp_pd(__xi, __yi, _CMP_NEQ_OQ));
3471 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3472 return __auto_bitcast(_mm_cmp_ps(__xi, __yi, _CMP_NEQ_OQ));
3473 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3474 return __to_masktype(_mm_cmp_pd(__xi, __yi, _CMP_NEQ_OQ));
3475 else
3476 __assert_unreachable<_Tp>();
3477 }
3478 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3479 return __auto_bitcast(
3480 __and(_mm_cmpord_ps(__xi, __yi), _mm_cmpneq_ps(__xi, __yi)));
3481 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3482 return __to_masktype(
3483 __and(_mm_cmpord_pd(__xi, __yi), _mm_cmpneq_pd(__xi, __yi)));
3484 else
3485 __assert_unreachable<_Tp>();
3486 }
3487
3488 //}}} }}}
3489 template <template <typename> class _Op, typename _Tp, typename _K, size_t _Np>
3490 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
3491 _S_masked_unary(const _SimdWrapper<_K, _Np> __k, const _SimdWrapper<_Tp, _Np> __v)
3492 {
3493 if (__k._M_is_constprop_none_of())
3494 return __v;
3495 else if (__k._M_is_constprop_all_of())
3496 {
3497 auto __vv = _Base::_M_make_simd(__v);
3498 _Op<decltype(__vv)> __op;
3499 return __data(__op(__vv));
3500 }
3501 else if constexpr (__is_bitmask_v<decltype(__k)>
3502 && (is_same_v<_Op<void>, __increment<void>>
3503 || is_same_v<_Op<void>, __decrement<void>>))
3504 {
3505 // optimize masked unary increment and decrement as masked sub +/-1
3506 constexpr int __pm_one
3507 = is_same_v<_Op<void>, __increment<void>> ? -1 : 1;
3508#ifdef __clang__
3509 return __movm<_Np, _Tp>(__k._M_data) ? __v._M_data - __pm_one : __v._M_data;
3510#else // __clang__
3511 using _TV = __vector_type_t<_Tp, _Np>;
3512 constexpr size_t __bytes = sizeof(__v) < 16 ? 16 : sizeof(__v);
3513 constexpr size_t __width = __bytes / sizeof(_Tp);
3514 if constexpr (is_integral_v<_Tp>)
3515 {
3516 constexpr bool __lp64 = sizeof(long) == sizeof(long long);
3517 using _Ip = std::make_signed_t<_Tp>;
3518 using _Up = std::conditional_t<
3519 std::is_same_v<_Ip, long>,
3522 std::is_same_v<_Ip, signed char>, char, _Ip>>;
3523 const auto __value = __intrin_bitcast<__vector_type_t<_Up, __width>>(__v._M_data);
3524#define _GLIBCXX_SIMD_MASK_SUB(_Sizeof, _Width, _Instr) \
3525 if constexpr (sizeof(_Tp) == _Sizeof && sizeof(__value) == _Width) \
3526 return __intrin_bitcast<_TV>(__builtin_ia32_##_Instr##_mask(__value, \
3527 __vector_broadcast<__width>(_Up(__pm_one)), __value, __k._M_data))
3528 _GLIBCXX_SIMD_MASK_SUB(1, 64, psubb512);
3529 _GLIBCXX_SIMD_MASK_SUB(1, 32, psubb256);
3530 _GLIBCXX_SIMD_MASK_SUB(1, 16, psubb128);
3531 _GLIBCXX_SIMD_MASK_SUB(2, 64, psubw512);
3532 _GLIBCXX_SIMD_MASK_SUB(2, 32, psubw256);
3533 _GLIBCXX_SIMD_MASK_SUB(2, 16, psubw128);
3534 _GLIBCXX_SIMD_MASK_SUB(4, 64, psubd512);
3535 _GLIBCXX_SIMD_MASK_SUB(4, 32, psubd256);
3536 _GLIBCXX_SIMD_MASK_SUB(4, 16, psubd128);
3537 _GLIBCXX_SIMD_MASK_SUB(8, 64, psubq512);
3538 _GLIBCXX_SIMD_MASK_SUB(8, 32, psubq256);
3539 _GLIBCXX_SIMD_MASK_SUB(8, 16, psubq128);
3540#undef _GLIBCXX_SIMD_MASK_SUB
3541 }
3542 else
3543 {
3544 const auto __value = __intrin_bitcast<__vector_type_t<_Tp, __width>>(__v._M_data);
3545#define _GLIBCXX_SIMD_MASK_SUB_512(_Sizeof, _Width, _Instr) \
3546 if constexpr (sizeof(_Tp) == _Sizeof && sizeof(__value) == _Width) \
3547 return __builtin_ia32_##_Instr##_mask( \
3548 __value, __vector_broadcast<__width>(_Tp(__pm_one)), __value, \
3549 __k._M_data, _MM_FROUND_CUR_DIRECTION)
3550#define _GLIBCXX_SIMD_MASK_SUB(_Sizeof, _Width, _Instr) \
3551 if constexpr (sizeof(_Tp) == _Sizeof && sizeof(__value) == _Width) \
3552 return __intrin_bitcast<_TV>(__builtin_ia32_##_Instr##_mask( \
3553 __value, __vector_broadcast<__width>(_Tp(__pm_one)), __value, \
3554 __k._M_data))
3555 _GLIBCXX_SIMD_MASK_SUB_512(4, 64, subps512);
3556 _GLIBCXX_SIMD_MASK_SUB(4, 32, subps256);
3557 _GLIBCXX_SIMD_MASK_SUB(4, 16, subps128);
3558 _GLIBCXX_SIMD_MASK_SUB_512(8, 64, subpd512);
3559 _GLIBCXX_SIMD_MASK_SUB(8, 32, subpd256);
3560 _GLIBCXX_SIMD_MASK_SUB(8, 16, subpd128);
3561#undef _GLIBCXX_SIMD_MASK_SUB_512
3562#undef _GLIBCXX_SIMD_MASK_SUB
3563 }
3564#endif // __clang__
3565 }
3566 else
3567 return _Base::template _S_masked_unary<_Op>(__k, __v);
3568 }
3569 };
3570
3571// }}}
3572// _MaskImplX86Mixin {{{
3573struct _MaskImplX86Mixin
3574{
3575 template <typename _Tp>
3576 using _TypeTag = _Tp*;
3577
3578 using _Base = _MaskImplBuiltinMixin;
3579
3580 // _S_to_maskvector(bool) {{{
3581 template <typename _Up, size_t _ToN = 1, typename _Tp>
3582 _GLIBCXX_SIMD_INTRINSIC static constexpr
3583 enable_if_t<is_same_v<_Tp, bool>, _SimdWrapper<_Up, _ToN>>
3584 _S_to_maskvector(_Tp __x)
3585 {
3586 static_assert(is_same_v<_Up, __int_for_sizeof_t<_Up>>);
3587 return __x ? __vector_type_t<_Up, _ToN>{~_Up()}
3588 : __vector_type_t<_Up, _ToN>();
3589 }
3590
3591 // }}}
3592 // _S_to_maskvector(_SanitizedBitMask) {{{
3593 template <typename _Up, size_t _UpN = 0, size_t _Np, size_t _ToN = _UpN == 0 ? _Np : _UpN>
3594 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Up, _ToN>
3595 _S_to_maskvector(_SanitizedBitMask<_Np> __x)
3596 {
3597 static_assert(is_same_v<_Up, __int_for_sizeof_t<_Up>>);
3598 using _UV = __vector_type_t<_Up, _ToN>;
3599 using _UI = __intrinsic_type_t<_Up, _ToN>;
3600 [[maybe_unused]] const auto __k = __x._M_to_bits();
3601 if constexpr (_Np == 1)
3602 return _S_to_maskvector<_Up, _ToN>(__k);
3603 else if (__x._M_is_constprop() || __builtin_is_constant_evaluated())
3604 return __generate_from_n_evaluations<std::min(_ToN, _Np), _UV>(
3605 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA -> _Up { return -__x[__i.value]; });
3606 else if constexpr (sizeof(_Up) == 1)
3607 {
3608 if constexpr (sizeof(_UI) == 16)
3609 {
3610 if constexpr (__have_avx512bw_vl)
3611 return __intrin_bitcast<_UV>(_mm_movm_epi8(__k));
3612 else if constexpr (__have_avx512bw)
3613 return __intrin_bitcast<_UV>(__lo128(_mm512_movm_epi8(__k)));
3614 else if constexpr (__have_avx512f)
3615 {
3616 auto __as32bits = _mm512_maskz_mov_epi32(__k, ~__m512i());
3617 auto __as16bits
3618 = __xzyw(_mm256_packs_epi32(__lo256(__as32bits),
3619 __hi256(__as32bits)));
3620 return __intrin_bitcast<_UV>(
3621 _mm_packs_epi16(__lo128(__as16bits), __hi128(__as16bits)));
3622 }
3623 else if constexpr (__have_ssse3)
3624 {
3625 const auto __bitmask = __to_intrin(
3626 __make_vector<_UChar>(1, 2, 4, 8, 16, 32, 64, 128, 1, 2, 4,
3627 8, 16, 32, 64, 128));
3628 return __intrin_bitcast<_UV>(
3629 __vector_bitcast<_Up>(
3630 _mm_shuffle_epi8(__to_intrin(
3631 __vector_type_t<_ULLong, 2>{__k}),
3632 _mm_setr_epi8(0, 0, 0, 0, 0, 0, 0, 0, 1,
3633 1, 1, 1, 1, 1, 1, 1))
3634 & __bitmask)
3635 != 0);
3636 }
3637 // else fall through
3638 }
3639 else if constexpr (sizeof(_UI) == 32)
3640 {
3641 if constexpr (__have_avx512bw_vl)
3642 return __vector_bitcast<_Up>(_mm256_movm_epi8(__k));
3643 else if constexpr (__have_avx512bw)
3644 return __vector_bitcast<_Up>(__lo256(_mm512_movm_epi8(__k)));
3645 else if constexpr (__have_avx512f)
3646 {
3647 auto __as16bits = // 0 16 1 17 ... 15 31
3648 _mm512_srli_epi32(_mm512_maskz_mov_epi32(__k, ~__m512i()),
3649 16)
3650 | _mm512_slli_epi32(_mm512_maskz_mov_epi32(__k >> 16,
3651 ~__m512i()),
3652 16);
3653 auto __0_16_1_17 = __xzyw(_mm256_packs_epi16(
3654 __lo256(__as16bits),
3655 __hi256(__as16bits)) // 0 16 1 17 2 18 3 19 8 24 9 25 ...
3656 );
3657 // deinterleave:
3658 return __vector_bitcast<_Up>(__xzyw(_mm256_shuffle_epi8(
3659 __0_16_1_17, // 0 16 1 17 2 ...
3660 _mm256_setr_epi8(0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9,
3661 11, 13, 15, 0, 2, 4, 6, 8, 10, 12, 14, 1,
3662 3, 5, 7, 9, 11, 13,
3663 15)))); // 0-7 16-23 8-15 24-31 -> xzyw
3664 // 0-3 8-11 16-19 24-27
3665 // 4-7 12-15 20-23 28-31
3666 }
3667 else if constexpr (__have_avx2)
3668 {
3669 const auto __bitmask
3670 = _mm256_broadcastsi128_si256(__to_intrin(
3671 __make_vector<_UChar>(1, 2, 4, 8, 16, 32, 64, 128, 1, 2,
3672 4, 8, 16, 32, 64, 128)));
3673 return __vector_bitcast<_Up>(
3674 __vector_bitcast<_Up>(
3675 _mm256_shuffle_epi8(
3676 _mm256_broadcastsi128_si256(
3677 __to_intrin(__vector_type_t<_ULLong, 2>{__k})),
3678 _mm256_setr_epi8(0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1,
3679 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2, 3, 3,
3680 3, 3, 3, 3, 3, 3))
3681 & __bitmask)
3682 != 0);
3683 }
3684 // else fall through
3685 }
3686 else if constexpr (sizeof(_UI) == 64)
3687 return reinterpret_cast<_UV>(_mm512_movm_epi8(__k));
3688 if constexpr (std::min(_ToN, _Np) <= 4)
3689 {
3690 if constexpr (_Np > 7) // avoid overflow
3691 __x &= _SanitizedBitMask<_Np>(0x0f);
3692 const _UInt __char_mask
3693 = ((_UInt(__x.to_ulong()) * 0x00204081U) & 0x01010101ULL)
3694 * 0xff;
3695 _UV __r = {};
3696 __builtin_memcpy(&__r, &__char_mask,
3697 std::min(sizeof(__r), sizeof(__char_mask)));
3698 return __r;
3699 }
3700 else if constexpr (std::min(_ToN, _Np) <= 7)
3701 {
3702 if constexpr (_Np > 7) // avoid overflow
3703 __x &= _SanitizedBitMask<_Np>(0x7f);
3704 const _ULLong __char_mask
3705 = ((__x.to_ulong() * 0x40810204081ULL) & 0x0101010101010101ULL)
3706 * 0xff;
3707 _UV __r = {};
3708 __builtin_memcpy(&__r, &__char_mask,
3709 std::min(sizeof(__r), sizeof(__char_mask)));
3710 return __r;
3711 }
3712 }
3713 else if constexpr (sizeof(_Up) == 2)
3714 {
3715 if constexpr (sizeof(_UI) == 16)
3716 {
3717 if constexpr (__have_avx512bw_vl)
3718 return __intrin_bitcast<_UV>(_mm_movm_epi16(__k));
3719 else if constexpr (__have_avx512bw)
3720 return __intrin_bitcast<_UV>(__lo128(_mm512_movm_epi16(__k)));
3721 else if constexpr (__have_avx512f)
3722 {
3723 __m256i __as32bits = {};
3724 if constexpr (__have_avx512vl)
3725 __as32bits = _mm256_maskz_mov_epi32(__k, ~__m256i());
3726 else
3727 __as32bits
3728 = __lo256(_mm512_maskz_mov_epi32(__k, ~__m512i()));
3729 return __intrin_bitcast<_UV>(
3730 _mm_packs_epi32(__lo128(__as32bits), __hi128(__as32bits)));
3731 }
3732 // else fall through
3733 }
3734 else if constexpr (sizeof(_UI) == 32)
3735 {
3736 if constexpr (__have_avx512bw_vl)
3737 return __vector_bitcast<_Up>(_mm256_movm_epi16(__k));
3738 else if constexpr (__have_avx512bw)
3739 return __vector_bitcast<_Up>(__lo256(_mm512_movm_epi16(__k)));
3740 else if constexpr (__have_avx512f)
3741 {
3742 auto __as32bits = _mm512_maskz_mov_epi32(__k, ~__m512i());
3743 return __vector_bitcast<_Up>(
3744 __xzyw(_mm256_packs_epi32(__lo256(__as32bits),
3745 __hi256(__as32bits))));
3746 }
3747 // else fall through
3748 }
3749 else if constexpr (sizeof(_UI) == 64)
3750 return __vector_bitcast<_Up>(_mm512_movm_epi16(__k));
3751 }
3752 else if constexpr (sizeof(_Up) == 4)
3753 {
3754 if constexpr (sizeof(_UI) == 16)
3755 {
3756 if constexpr (__have_avx512dq_vl)
3757 return __intrin_bitcast<_UV>(_mm_movm_epi32(__k));
3758 else if constexpr (__have_avx512dq)
3759 return __intrin_bitcast<_UV>(__lo128(_mm512_movm_epi32(__k)));
3760 else if constexpr (__have_avx512vl)
3761 return __intrin_bitcast<_UV>(
3762 _mm_maskz_mov_epi32(__k, ~__m128i()));
3763 else if constexpr (__have_avx512f)
3764 return __intrin_bitcast<_UV>(
3765 __lo128(_mm512_maskz_mov_epi32(__k, ~__m512i())));
3766 // else fall through
3767 }
3768 else if constexpr (sizeof(_UI) == 32)
3769 {
3770 if constexpr (__have_avx512dq_vl)
3771 return __vector_bitcast<_Up>(_mm256_movm_epi32(__k));
3772 else if constexpr (__have_avx512dq)
3773 return __vector_bitcast<_Up>(__lo256(_mm512_movm_epi32(__k)));
3774 else if constexpr (__have_avx512vl)
3775 return __vector_bitcast<_Up>(
3776 _mm256_maskz_mov_epi32(__k, ~__m256i()));
3777 else if constexpr (__have_avx512f)
3778 return __vector_bitcast<_Up>(
3779 __lo256(_mm512_maskz_mov_epi32(__k, ~__m512i())));
3780 // else fall through
3781 }
3782 else if constexpr (sizeof(_UI) == 64)
3783 return __vector_bitcast<_Up>(
3784 __have_avx512dq ? _mm512_movm_epi32(__k)
3785 : _mm512_maskz_mov_epi32(__k, ~__m512i()));
3786 }
3787 else if constexpr (sizeof(_Up) == 8)
3788 {
3789 if constexpr (sizeof(_UI) == 16)
3790 {
3791 if constexpr (__have_avx512dq_vl)
3792 return __vector_bitcast<_Up>(_mm_movm_epi64(__k));
3793 else if constexpr (__have_avx512dq)
3794 return __vector_bitcast<_Up>(__lo128(_mm512_movm_epi64(__k)));
3795 else if constexpr (__have_avx512vl)
3796 return __vector_bitcast<_Up>(
3797 _mm_maskz_mov_epi64(__k, ~__m128i()));
3798 else if constexpr (__have_avx512f)
3799 return __vector_bitcast<_Up>(
3800 __lo128(_mm512_maskz_mov_epi64(__k, ~__m512i())));
3801 // else fall through
3802 }
3803 else if constexpr (sizeof(_UI) == 32)
3804 {
3805 if constexpr (__have_avx512dq_vl)
3806 return __vector_bitcast<_Up>(_mm256_movm_epi64(__k));
3807 else if constexpr (__have_avx512dq)
3808 return __vector_bitcast<_Up>(__lo256(_mm512_movm_epi64(__k)));
3809 else if constexpr (__have_avx512vl)
3810 return __vector_bitcast<_Up>(
3811 _mm256_maskz_mov_epi64(__k, ~__m256i()));
3812 else if constexpr (__have_avx512f)
3813 return __vector_bitcast<_Up>(
3814 __lo256(_mm512_maskz_mov_epi64(__k, ~__m512i())));
3815 // else fall through
3816 }
3817 else if constexpr (sizeof(_UI) == 64)
3818 return __vector_bitcast<_Up>(
3819 __have_avx512dq ? _mm512_movm_epi64(__k)
3820 : _mm512_maskz_mov_epi64(__k, ~__m512i()));
3821 }
3822
3823 using _UpUInt = make_unsigned_t<_Up>;
3824 using _V = __vector_type_t<_UpUInt, _ToN>;
3825 constexpr size_t __bits_per_element = sizeof(_Up) * __CHAR_BIT__;
3826 if constexpr (_ToN == 2)
3827 {
3828 return __vector_bitcast<_Up>(_V{_UpUInt(-__x[0]), _UpUInt(-__x[1])});
3829 }
3830 else if constexpr (!__have_avx2 && __have_avx && sizeof(_V) == 32)
3831 {
3832 if constexpr (sizeof(_Up) == 4)
3833 return __vector_bitcast<_Up>(_mm256_cmp_ps(
3834 _mm256_and_ps(_mm256_castsi256_ps(_mm256_set1_epi32(__k)),
3835 _mm256_castsi256_ps(_mm256_setr_epi32(
3836 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80))),
3837 _mm256_setzero_ps(), _CMP_NEQ_UQ));
3838 else if constexpr (sizeof(_Up) == 8)
3839 return __vector_bitcast<_Up>(_mm256_cmp_pd(
3840 _mm256_and_pd(_mm256_castsi256_pd(_mm256_set1_epi64x(__k)),
3841 _mm256_castsi256_pd(
3842 _mm256_setr_epi64x(0x01, 0x02, 0x04, 0x08))),
3843 _mm256_setzero_pd(), _CMP_NEQ_UQ));
3844 else
3845 __assert_unreachable<_Up>();
3846 }
3847 else if constexpr (__bits_per_element >= _ToN)
3848 {
3849 constexpr auto __bitmask
3850 = __generate_vector<_V>([](auto __i)
3851 constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA -> _UpUInt
3852 { return __i < _ToN ? 1ull << __i : 0; });
3853 const auto __bits
3854 = __vector_broadcast<_ToN, _UpUInt>(__k) & __bitmask;
3855 if constexpr (__bits_per_element > _ToN)
3856 return __vector_bitcast<_Up>(__bits) > 0;
3857 else
3858 return __vector_bitcast<_Up>(__bits != 0);
3859 }
3860 else
3861 {
3862 const _V __tmp
3863 = __generate_vector<_V>([&](auto __i) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
3864 return static_cast<_UpUInt>(
3865 __k >> (__bits_per_element * (__i / __bits_per_element)));
3866 })
3867 & __generate_vector<_V>([](auto __i) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
3868 return static_cast<_UpUInt>(1ull
3869 << (__i % __bits_per_element));
3870 }); // mask bit index
3871 return __intrin_bitcast<_UV>(__tmp != _V());
3872 }
3873 }
3874
3875 // }}}
3876 // _S_to_maskvector(_SimdWrapper) {{{
3877 template <typename _Up, size_t _UpN = 0, typename _Tp, size_t _Np,
3878 size_t _ToN = _UpN == 0 ? _Np : _UpN>
3879 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Up, _ToN>
3880 _S_to_maskvector(_SimdWrapper<_Tp, _Np> __x)
3881 {
3882 static_assert(is_same_v<_Up, __int_for_sizeof_t<_Up>>);
3883 using _TW = _SimdWrapper<_Tp, _Np>;
3884 using _UW = _SimdWrapper<_Up, _ToN>;
3885 using _UI = __intrinsic_type_t<_Up, _ToN>;
3886 if constexpr (is_same_v<_Tp, bool>) // bits -> vector
3887 return _S_to_maskvector<_Up, _ToN>(
3888 _BitMask<_Np>(__x._M_data)._M_sanitized());
3889 // vector -> vector bitcast
3890 else if constexpr (sizeof(_Up) == sizeof(_Tp)
3891 && sizeof(_TW) == sizeof(_UW))
3892 return __wrapper_bitcast<_Up, _ToN>(
3893 _ToN <= _Np
3894 ? __x
3895 : simd_abi::_VecBuiltin<sizeof(_Tp) * _Np>::_S_masked(__x));
3896 else // vector -> vector {{{
3897 {
3898 if (__x._M_is_constprop() || __builtin_is_constant_evaluated())
3899 {
3900 const auto __y = __vector_bitcast<__int_for_sizeof_t<_Tp>>(__x);
3901 return __generate_from_n_evaluations<std::min(_ToN, _Np),
3902 __vector_type_t<_Up, _ToN>>(
3903 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA -> _Up { return __y[__i.value]; });
3904 }
3905 using _To = __vector_type_t<_Up, _ToN>;
3906 [[maybe_unused]] constexpr size_t _FromN = _Np;
3907 constexpr int _FromBytes = sizeof(_Tp);
3908 constexpr int _ToBytes = sizeof(_Up);
3909 const auto __k = __x._M_data;
3910
3911 if constexpr (_FromBytes == _ToBytes)
3912 return __intrin_bitcast<_To>(__k);
3913 else if constexpr (sizeof(_UI) == 16 && sizeof(__k) == 16)
3914 { // SSE -> SSE {{{
3915 if constexpr (_FromBytes == 4 && _ToBytes == 8)
3916 return __intrin_bitcast<_To>(__interleave128_lo(__k, __k));
3917 else if constexpr (_FromBytes == 2 && _ToBytes == 8)
3918 {
3919 const auto __y
3920 = __vector_bitcast<int>(__interleave128_lo(__k, __k));
3921 return __intrin_bitcast<_To>(__interleave128_lo(__y, __y));
3922 }
3923 else if constexpr (_FromBytes == 1 && _ToBytes == 8)
3924 {
3925 auto __y
3926 = __vector_bitcast<short>(__interleave128_lo(__k, __k));
3927 auto __z
3928 = __vector_bitcast<int>(__interleave128_lo(__y, __y));
3929 return __intrin_bitcast<_To>(__interleave128_lo(__z, __z));
3930 }
3931 else if constexpr (_FromBytes == 8 && _ToBytes == 4
3932 && __have_sse2)
3933 return __intrin_bitcast<_To>(
3934 _mm_packs_epi32(__vector_bitcast<_LLong>(__k), __m128i()));
3935 else if constexpr (_FromBytes == 8 && _ToBytes == 4)
3936 return __vector_shuffle<1, 3, 6, 7>(__vector_bitcast<_Up>(__k),
3937 _UI());
3938 else if constexpr (_FromBytes == 2 && _ToBytes == 4)
3939 return __intrin_bitcast<_To>(__interleave128_lo(__k, __k));
3940 else if constexpr (_FromBytes == 1 && _ToBytes == 4)
3941 {
3942 const auto __y
3943 = __vector_bitcast<short>(__interleave128_lo(__k, __k));
3944 return __intrin_bitcast<_To>(__interleave128_lo(__y, __y));
3945 }
3946 else if constexpr (_FromBytes == 8 && _ToBytes == 2)
3947 {
3948 if constexpr (__have_sse2 && !__have_ssse3)
3949 return __intrin_bitcast<_To>(_mm_packs_epi32(
3950 _mm_packs_epi32(__vector_bitcast<_LLong>(__k), __m128i()),
3951 __m128i()));
3952 else
3953 return __intrin_bitcast<_To>(
3954 __vector_permute<3, 7, -1, -1, -1, -1, -1, -1>(
3955 __vector_bitcast<_Up>(__k)));
3956 }
3957 else if constexpr (_FromBytes == 4 && _ToBytes == 2)
3958 return __intrin_bitcast<_To>(
3959 _mm_packs_epi32(__vector_bitcast<_LLong>(__k), __m128i()));
3960 else if constexpr (_FromBytes == 1 && _ToBytes == 2)
3961 return __intrin_bitcast<_To>(__interleave128_lo(__k, __k));
3962 else if constexpr (_FromBytes == 8 && _ToBytes == 1
3963 && __have_ssse3)
3964 return __intrin_bitcast<_To>(
3965 _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
3966 _mm_setr_epi8(7, 15, -1, -1, -1, -1, -1, -1,
3967 -1, -1, -1, -1, -1, -1, -1,
3968 -1)));
3969 else if constexpr (_FromBytes == 8 && _ToBytes == 1)
3970 {
3971 auto __y
3972 = _mm_packs_epi32(__vector_bitcast<_LLong>(__k), __m128i());
3973 __y = _mm_packs_epi32(__y, __m128i());
3974 return __intrin_bitcast<_To>(_mm_packs_epi16(__y, __m128i()));
3975 }
3976 else if constexpr (_FromBytes == 4 && _ToBytes == 1
3977 && __have_ssse3)
3978 return __intrin_bitcast<_To>(
3979 _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
3980 _mm_setr_epi8(3, 7, 11, 15, -1, -1, -1, -1,
3981 -1, -1, -1, -1, -1, -1, -1,
3982 -1)));
3983 else if constexpr (_FromBytes == 4 && _ToBytes == 1)
3984 {
3985 const auto __y
3986 = _mm_packs_epi32(__vector_bitcast<_LLong>(__k), __m128i());
3987 return __intrin_bitcast<_To>(_mm_packs_epi16(__y, __m128i()));
3988 }
3989 else if constexpr (_FromBytes == 2 && _ToBytes == 1)
3990 return __intrin_bitcast<_To>(
3991 _mm_packs_epi16(__vector_bitcast<_LLong>(__k), __m128i()));
3992 else
3993 __assert_unreachable<_Tp>();
3994 } // }}}
3995 else if constexpr (sizeof(_UI) == 32 && sizeof(__k) == 32)
3996 { // AVX -> AVX {{{
3997 if constexpr (_FromBytes == _ToBytes)
3998 __assert_unreachable<_Tp>();
3999 else if constexpr (_FromBytes == _ToBytes * 2)
4000 {
4001 const auto __y = __vector_bitcast<_LLong>(__k);
4002 return __intrin_bitcast<_To>(_mm256_castsi128_si256(
4003 _mm_packs_epi16(__lo128(__y), __hi128(__y))));
4004 }
4005 else if constexpr (_FromBytes == _ToBytes * 4)
4006 {
4007 const auto __y = __vector_bitcast<_LLong>(__k);
4008 return __intrin_bitcast<_To>(_mm256_castsi128_si256(
4009 _mm_packs_epi16(_mm_packs_epi16(__lo128(__y), __hi128(__y)),
4010 __m128i())));
4011 }
4012 else if constexpr (_FromBytes == _ToBytes * 8)
4013 {
4014 const auto __y = __vector_bitcast<_LLong>(__k);
4015 return __intrin_bitcast<_To>(
4016 _mm256_castsi128_si256(_mm_shuffle_epi8(
4017 _mm_packs_epi16(__lo128(__y), __hi128(__y)),
4018 _mm_setr_epi8(3, 7, 11, 15, -1, -1, -1, -1, -1, -1, -1,
4019 -1, -1, -1, -1, -1))));
4020 }
4021 else if constexpr (_FromBytes * 2 == _ToBytes)
4022 {
4023 auto __y = __xzyw(__to_intrin(__k));
4024 if constexpr (is_floating_point_v<
4025 _Tp> || (!__have_avx2 && _FromBytes == 4))
4026 {
4027 const auto __yy = __vector_bitcast<float>(__y);
4028 return __intrin_bitcast<_To>(
4029 _mm256_unpacklo_ps(__yy, __yy));
4030 }
4031 else
4032 return __intrin_bitcast<_To>(
4033 _mm256_unpacklo_epi8(__y, __y));
4034 }
4035 else if constexpr (_FromBytes * 4 == _ToBytes)
4036 {
4037 auto __y
4038 = _mm_unpacklo_epi8(__lo128(__vector_bitcast<_LLong>(__k)),
4039 __lo128(__vector_bitcast<_LLong>(
4040 __k))); // drops 3/4 of input
4041 return __intrin_bitcast<_To>(
4042 __concat(_mm_unpacklo_epi16(__y, __y),
4043 _mm_unpackhi_epi16(__y, __y)));
4044 }
4045 else if constexpr (_FromBytes == 1 && _ToBytes == 8)
4046 {
4047 auto __y
4048 = _mm_unpacklo_epi8(__lo128(__vector_bitcast<_LLong>(__k)),
4049 __lo128(__vector_bitcast<_LLong>(
4050 __k))); // drops 3/4 of input
4051 __y
4052 = _mm_unpacklo_epi16(__y,
4053 __y); // drops another 1/2 => 7/8 total
4054 return __intrin_bitcast<_To>(
4055 __concat(_mm_unpacklo_epi32(__y, __y),
4056 _mm_unpackhi_epi32(__y, __y)));
4057 }
4058 else
4059 __assert_unreachable<_Tp>();
4060 } // }}}
4061 else if constexpr (sizeof(_UI) == 32 && sizeof(__k) == 16)
4062 { // SSE -> AVX {{{
4063 if constexpr (_FromBytes == _ToBytes)
4064 return __intrin_bitcast<_To>(
4065 __intrinsic_type_t<_Tp, 32 / sizeof(_Tp)>(
4066 __zero_extend(__to_intrin(__k))));
4067 else if constexpr (_FromBytes * 2 == _ToBytes)
4068 { // keep all
4069 return __intrin_bitcast<_To>(
4070 __concat(_mm_unpacklo_epi8(__vector_bitcast<_LLong>(__k),
4071 __vector_bitcast<_LLong>(__k)),
4072 _mm_unpackhi_epi8(__vector_bitcast<_LLong>(__k),
4073 __vector_bitcast<_LLong>(__k))));
4074 }
4075 else if constexpr (_FromBytes * 4 == _ToBytes)
4076 {
4077 if constexpr (__have_avx2)
4078 {
4079 return __intrin_bitcast<_To>(_mm256_shuffle_epi8(
4080 __concat(__vector_bitcast<_LLong>(__k),
4081 __vector_bitcast<_LLong>(__k)),
4082 _mm256_setr_epi8(0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 3,
4083 3, 3, 3, 4, 4, 4, 4, 5, 5, 5, 5, 6, 6,
4084 6, 6, 7, 7, 7, 7)));
4085 }
4086 else
4087 {
4088 return __intrin_bitcast<_To>(__concat(
4089 _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
4090 _mm_setr_epi8(0, 0, 0, 0, 1, 1, 1, 1,
4091 2, 2, 2, 2, 3, 3, 3, 3)),
4092 _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
4093 _mm_setr_epi8(4, 4, 4, 4, 5, 5, 5, 5,
4094 6, 6, 6, 6, 7, 7, 7,
4095 7))));
4096 }
4097 }
4098 else if constexpr (_FromBytes * 8 == _ToBytes)
4099 {
4100 if constexpr (__have_avx2)
4101 {
4102 return __intrin_bitcast<_To>(_mm256_shuffle_epi8(
4103 __concat(__vector_bitcast<_LLong>(__k),
4104 __vector_bitcast<_LLong>(__k)),
4105 _mm256_setr_epi8(0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1,
4106 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2, 3, 3,
4107 3, 3, 3, 3, 3, 3)));
4108 }
4109 else
4110 {
4111 return __intrin_bitcast<_To>(__concat(
4112 _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
4113 _mm_setr_epi8(0, 0, 0, 0, 0, 0, 0, 0,
4114 1, 1, 1, 1, 1, 1, 1, 1)),
4115 _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
4116 _mm_setr_epi8(2, 2, 2, 2, 2, 2, 2, 2,
4117 3, 3, 3, 3, 3, 3, 3,
4118 3))));
4119 }
4120 }
4121 else if constexpr (_FromBytes == _ToBytes * 2)
4122 return __intrin_bitcast<_To>(__m256i(__zero_extend(
4123 _mm_packs_epi16(__vector_bitcast<_LLong>(__k), __m128i()))));
4124 else if constexpr (_FromBytes == 8 && _ToBytes == 2)
4125 {
4126 return __intrin_bitcast<_To>(__m256i(__zero_extend(
4127 _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
4128 _mm_setr_epi8(6, 7, 14, 15, -1, -1, -1, -1,
4129 -1, -1, -1, -1, -1, -1, -1,
4130 -1)))));
4131 }
4132 else if constexpr (_FromBytes == 4 && _ToBytes == 1)
4133 {
4134 return __intrin_bitcast<_To>(__m256i(__zero_extend(
4135 _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
4136 _mm_setr_epi8(3, 7, 11, 15, -1, -1, -1, -1,
4137 -1, -1, -1, -1, -1, -1, -1,
4138 -1)))));
4139 }
4140 else if constexpr (_FromBytes == 8 && _ToBytes == 1)
4141 {
4142 return __intrin_bitcast<_To>(__m256i(__zero_extend(
4143 _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
4144 _mm_setr_epi8(7, 15, -1, -1, -1, -1, -1,
4145 -1, -1, -1, -1, -1, -1, -1,
4146 -1, -1)))));
4147 }
4148 else
4149 static_assert(!is_same_v<_Tp, _Tp>, "should be unreachable");
4150 } // }}}
4151 else if constexpr (sizeof(_UI) == 16 && sizeof(__k) == 32)
4152 { // AVX -> SSE {{{
4153 if constexpr (_FromBytes == _ToBytes)
4154 { // keep low 1/2
4155 return __intrin_bitcast<_To>(__lo128(__k));
4156 }
4157 else if constexpr (_FromBytes == _ToBytes * 2)
4158 { // keep all
4159 auto __y = __vector_bitcast<_LLong>(__k);
4160 return __intrin_bitcast<_To>(
4161 _mm_packs_epi16(__lo128(__y), __hi128(__y)));
4162 }
4163 else if constexpr (_FromBytes == _ToBytes * 4)
4164 { // add 1/2 undef
4165 auto __y = __vector_bitcast<_LLong>(__k);
4166 return __intrin_bitcast<_To>(
4167 _mm_packs_epi16(_mm_packs_epi16(__lo128(__y), __hi128(__y)),
4168 __m128i()));
4169 }
4170 else if constexpr (_FromBytes == 8 && _ToBytes == 1)
4171 { // add 3/4 undef
4172 auto __y = __vector_bitcast<_LLong>(__k);
4173 return __intrin_bitcast<_To>(_mm_shuffle_epi8(
4174 _mm_packs_epi16(__lo128(__y), __hi128(__y)),
4175 _mm_setr_epi8(3, 7, 11, 15, -1, -1, -1, -1, -1, -1, -1, -1,
4176 -1, -1, -1, -1)));
4177 }
4178 else if constexpr (_FromBytes * 2 == _ToBytes)
4179 { // keep low 1/4
4180 auto __y = __lo128(__vector_bitcast<_LLong>(__k));
4181 return __intrin_bitcast<_To>(_mm_unpacklo_epi8(__y, __y));
4182 }
4183 else if constexpr (_FromBytes * 4 == _ToBytes)
4184 { // keep low 1/8
4185 auto __y = __lo128(__vector_bitcast<_LLong>(__k));
4186 __y = _mm_unpacklo_epi8(__y, __y);
4187 return __intrin_bitcast<_To>(_mm_unpacklo_epi8(__y, __y));
4188 }
4189 else if constexpr (_FromBytes * 8 == _ToBytes)
4190 { // keep low 1/16
4191 auto __y = __lo128(__vector_bitcast<_LLong>(__k));
4192 __y = _mm_unpacklo_epi8(__y, __y);
4193 __y = _mm_unpacklo_epi8(__y, __y);
4194 return __intrin_bitcast<_To>(_mm_unpacklo_epi8(__y, __y));
4195 }
4196 else
4197 static_assert(!is_same_v<_Tp, _Tp>, "should be unreachable");
4198 } // }}}
4199 else
4200 return _Base::template _S_to_maskvector<_Up, _ToN>(__x);
4201 /*
4202 if constexpr (_FromBytes > _ToBytes) {
4203 const _To __y = __vector_bitcast<_Up>(__k);
4204 return [&] <size_t... _Is> (index_sequence<_Is...>) {
4205 constexpr int _Stride = _FromBytes / _ToBytes;
4206 return _To{__y[(_Is + 1) * _Stride - 1]...};
4207 }(make_index_sequence<std::min(_ToN, _FromN)>());
4208 } else {
4209 // {0, 0, 1, 1} (_Dups = 2, _Is<4>)
4210 // {0, 0, 0, 0, 1, 1, 1, 1} (_Dups = 4, _Is<8>)
4211 // {0, 0, 1, 1, 2, 2, 3, 3} (_Dups = 2, _Is<8>)
4212 // ...
4213 return [&] <size_t... _Is> (index_sequence<_Is...>) {
4214 constexpr int __dup = _ToBytes / _FromBytes;
4215 return __intrin_bitcast<_To>(_From{__k[_Is / __dup]...});
4216 }(make_index_sequence<_FromN>());
4217 }
4218 */
4219 } // }}}
4220 }
4221
4222 // }}}
4223 // _S_to_bits {{{
4224 template <typename _Tp, size_t _Np>
4225 _GLIBCXX_SIMD_INTRINSIC static constexpr _SanitizedBitMask<_Np>
4226 _S_to_bits(_SimdWrapper<_Tp, _Np> __x)
4227 {
4228 if constexpr (is_same_v<_Tp, bool>)
4229 return _BitMask<_Np>(__x._M_data)._M_sanitized();
4230 else
4231 {
4232 static_assert(is_same_v<_Tp, __int_for_sizeof_t<_Tp>>);
4233 if (__builtin_is_constant_evaluated()
4234 || __builtin_constant_p(__x._M_data))
4235 {
4236 const auto __bools = -__x._M_data;
4237 const _ULLong __k = __call_with_n_evaluations<_Np>(
4238 [](auto... __bits) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
4239 return (__bits | ...);
4240 }, [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
4241 return _ULLong(__bools[+__i]) << __i;
4242 });
4243 if (__builtin_is_constant_evaluated()
4244 || __builtin_constant_p(__k))
4245 return __k;
4246 }
4247 const auto __xi = __to_intrin(__x);
4248 if constexpr (sizeof(_Tp) == 1)
4249 if constexpr (sizeof(__xi) == 16)
4250 if constexpr (__have_avx512bw_vl)
4251 return _BitMask<_Np>(_mm_movepi8_mask(__xi));
4252 else // implies SSE2
4253 return _BitMask<_Np>(_mm_movemask_epi8(__xi));
4254 else if constexpr (sizeof(__xi) == 32)
4255 if constexpr (__have_avx512bw_vl)
4256 return _BitMask<_Np>(_mm256_movepi8_mask(__xi));
4257 else // implies AVX2
4258 return _BitMask<_Np>(_mm256_movemask_epi8(__xi));
4259 else // implies AVX512BW
4260 return _BitMask<_Np>(_mm512_movepi8_mask(__xi));
4261
4262 else if constexpr (sizeof(_Tp) == 2)
4263 if constexpr (sizeof(__xi) == 16)
4264 if constexpr (__have_avx512bw_vl)
4265 return _BitMask<_Np>(_mm_movepi16_mask(__xi));
4266 else if constexpr (__have_avx512bw)
4267 return _BitMask<_Np>(_mm512_movepi16_mask(__zero_extend(__xi)));
4268 else // implies SSE2
4269 return _BitMask<_Np>(
4270 _mm_movemask_epi8(_mm_packs_epi16(__xi, __m128i())));
4271 else if constexpr (sizeof(__xi) == 32)
4272 if constexpr (__have_avx512bw_vl)
4273 return _BitMask<_Np>(_mm256_movepi16_mask(__xi));
4274 else if constexpr (__have_avx512bw)
4275 return _BitMask<_Np>(_mm512_movepi16_mask(__zero_extend(__xi)));
4276 else // implies SSE2
4277 return _BitMask<_Np>(_mm_movemask_epi8(
4278 _mm_packs_epi16(__lo128(__xi), __hi128(__xi))));
4279 else // implies AVX512BW
4280 return _BitMask<_Np>(_mm512_movepi16_mask(__xi));
4281
4282 else if constexpr (sizeof(_Tp) == 4)
4283 if constexpr (sizeof(__xi) == 16)
4284 if constexpr (__have_avx512dq_vl)
4285 return _BitMask<_Np>(_mm_movepi32_mask(__xi));
4286 else if constexpr (__have_avx512vl)
4287 return _BitMask<_Np>(_mm_cmplt_epi32_mask(__xi, __m128i()));
4288 else if constexpr (__have_avx512dq)
4289 return _BitMask<_Np>(_mm512_movepi32_mask(__zero_extend(__xi)));
4290 else if constexpr (__have_avx512f)
4291 return _BitMask<_Np>(
4292 _mm512_cmplt_epi32_mask(__zero_extend(__xi), __m512i()));
4293 else // implies SSE
4294 return _BitMask<_Np>(
4295 _mm_movemask_ps(reinterpret_cast<__m128>(__xi)));
4296 else if constexpr (sizeof(__xi) == 32)
4297 if constexpr (__have_avx512dq_vl)
4298 return _BitMask<_Np>(_mm256_movepi32_mask(__xi));
4299 else if constexpr (__have_avx512dq)
4300 return _BitMask<_Np>(_mm512_movepi32_mask(__zero_extend(__xi)));
4301 else if constexpr (__have_avx512vl)
4302 return _BitMask<_Np>(_mm256_cmplt_epi32_mask(__xi, __m256i()));
4303 else if constexpr (__have_avx512f)
4304 return _BitMask<_Np>(
4305 _mm512_cmplt_epi32_mask(__zero_extend(__xi), __m512i()));
4306 else // implies AVX
4307 return _BitMask<_Np>(
4308 _mm256_movemask_ps(reinterpret_cast<__m256>(__xi)));
4309 else // implies AVX512??
4310 if constexpr (__have_avx512dq)
4311 return _BitMask<_Np>(_mm512_movepi32_mask(__xi));
4312 else // implies AVX512F
4313 return _BitMask<_Np>(_mm512_cmplt_epi32_mask(__xi, __m512i()));
4314
4315 else if constexpr (sizeof(_Tp) == 8)
4316 if constexpr (sizeof(__xi) == 16)
4317 if constexpr (__have_avx512dq_vl)
4318 return _BitMask<_Np>(_mm_movepi64_mask(__xi));
4319 else if constexpr (__have_avx512dq)
4320 return _BitMask<_Np>(_mm512_movepi64_mask(__zero_extend(__xi)));
4321 else if constexpr (__have_avx512vl)
4322 return _BitMask<_Np>(_mm_cmplt_epi64_mask(__xi, __m128i()));
4323 else if constexpr (__have_avx512f)
4324 return _BitMask<_Np>(
4325 _mm512_cmplt_epi64_mask(__zero_extend(__xi), __m512i()));
4326 else // implies SSE2
4327 return _BitMask<_Np>(
4328 _mm_movemask_pd(reinterpret_cast<__m128d>(__xi)));
4329 else if constexpr (sizeof(__xi) == 32)
4330 if constexpr (__have_avx512dq_vl)
4331 return _BitMask<_Np>(_mm256_movepi64_mask(__xi));
4332 else if constexpr (__have_avx512dq)
4333 return _BitMask<_Np>(_mm512_movepi64_mask(__zero_extend(__xi)));
4334 else if constexpr (__have_avx512vl)
4335 return _BitMask<_Np>(_mm256_cmplt_epi64_mask(__xi, __m256i()));
4336 else if constexpr (__have_avx512f)
4337 return _BitMask<_Np>(
4338 _mm512_cmplt_epi64_mask(__zero_extend(__xi), __m512i()));
4339 else // implies AVX
4340 return _BitMask<_Np>(
4341 _mm256_movemask_pd(reinterpret_cast<__m256d>(__xi)));
4342 else // implies AVX512??
4343 if constexpr (__have_avx512dq)
4344 return _BitMask<_Np>(_mm512_movepi64_mask(__xi));
4345 else // implies AVX512F
4346 return _BitMask<_Np>(_mm512_cmplt_epi64_mask(__xi, __m512i()));
4347
4348 else
4349 __assert_unreachable<_Tp>();
4350 }
4351 }
4352 // }}}
4353};
4354
4355// }}}
4356// _MaskImplX86 {{{
4357template <typename _Abi, typename>
4358 struct _MaskImplX86 : _MaskImplX86Mixin, _MaskImplBuiltin<_Abi>
4359 {
4360 using _MaskImplX86Mixin::_S_to_bits;
4361 using _MaskImplX86Mixin::_S_to_maskvector;
4362 using _MaskImplBuiltin<_Abi>::_S_convert;
4363
4364 // member types {{{
4365 template <typename _Tp>
4366 using _SimdMember = typename _Abi::template __traits<_Tp>::_SimdMember;
4367
4368 template <typename _Tp>
4369 using _MaskMember = typename _Abi::template _MaskMember<_Tp>;
4370
4371 template <typename _Tp>
4372 static constexpr size_t _S_size = simd_size_v<_Tp, _Abi>;
4373
4374 using _Base = _MaskImplBuiltin<_Abi>;
4375
4376 // }}}
4377 // _S_broadcast {{{
4378 template <typename _Tp>
4379 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
4380 _S_broadcast(bool __x)
4381 {
4382 if constexpr (__is_avx512_abi<_Abi>())
4383 return __x ? _Abi::_S_masked(_MaskMember<_Tp>(-1))
4384 : _MaskMember<_Tp>();
4385 else
4386 return _Base::template _S_broadcast<_Tp>(__x);
4387 }
4388
4389 // }}}
4390 // _S_load {{{
4391 template <typename _Tp>
4392 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
4393 _S_load(const bool* __mem)
4394 {
4395 static_assert(is_same_v<_Tp, __int_for_sizeof_t<_Tp>>);
4396 if (__builtin_is_constant_evaluated())
4397 {
4398 if constexpr (__is_avx512_abi<_Abi>())
4399 {
4400 _MaskMember<_Tp> __r{};
4401 for (size_t __i = 0; __i < _S_size<_Tp>; ++__i)
4402 __r._M_data |= _ULLong(__mem[__i]) << __i;
4403 return __r;
4404 }
4405 else
4406 return _Base::template _S_load<_Tp>(__mem);
4407 }
4408 else if constexpr (__have_avx512bw)
4409 {
4410 const auto __to_vec_or_bits
4411 = [](auto __bits) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA -> decltype(auto) {
4412 if constexpr (__is_avx512_abi<_Abi>())
4413 return __bits;
4414 else
4415 return _S_to_maskvector<_Tp>(
4416 _BitMask<_S_size<_Tp>>(__bits)._M_sanitized());
4417 };
4418
4419 if constexpr (_S_size<_Tp> <= 16 && __have_avx512vl)
4420 {
4421 __m128i __a = {};
4422 __builtin_memcpy(&__a, __mem, _S_size<_Tp>);
4423 return __to_vec_or_bits(_mm_test_epi8_mask(__a, __a));
4424 }
4425 else if constexpr (_S_size<_Tp> <= 32 && __have_avx512vl)
4426 {
4427 __m256i __a = {};
4428 __builtin_memcpy(&__a, __mem, _S_size<_Tp>);
4429 return __to_vec_or_bits(_mm256_test_epi8_mask(__a, __a));
4430 }
4431 else if constexpr (_S_size<_Tp> <= 64)
4432 {
4433 __m512i __a = {};
4434 __builtin_memcpy(&__a, __mem, _S_size<_Tp>);
4435 return __to_vec_or_bits(_mm512_test_epi8_mask(__a, __a));
4436 }
4437 }
4438 else if constexpr (__is_avx512_abi<_Abi>())
4439 {
4440 if constexpr (_S_size<_Tp> <= 8)
4441 {
4442 __m128i __a = {};
4443 __builtin_memcpy(&__a, __mem, _S_size<_Tp>);
4444 const auto __b = _mm512_cvtepi8_epi64(__a);
4445 return _mm512_test_epi64_mask(__b, __b);
4446 }
4447 else if constexpr (_S_size<_Tp> <= 16)
4448 {
4449 __m128i __a = {};
4450 __builtin_memcpy(&__a, __mem, _S_size<_Tp>);
4451 const auto __b = _mm512_cvtepi8_epi32(__a);
4452 return _mm512_test_epi32_mask(__b, __b);
4453 }
4454 else if constexpr (_S_size<_Tp> <= 32)
4455 {
4456 __m128i __a = {};
4457 __builtin_memcpy(&__a, __mem, 16);
4458 const auto __b = _mm512_cvtepi8_epi32(__a);
4459 __builtin_memcpy(&__a, __mem + 16, _S_size<_Tp> - 16);
4460 const auto __c = _mm512_cvtepi8_epi32(__a);
4461 return _mm512_test_epi32_mask(__b, __b)
4462 | (_mm512_test_epi32_mask(__c, __c) << 16);
4463 }
4464 else if constexpr (_S_size<_Tp> <= 64)
4465 {
4466 __m128i __a = {};
4467 __builtin_memcpy(&__a, __mem, 16);
4468 const auto __b = _mm512_cvtepi8_epi32(__a);
4469 __builtin_memcpy(&__a, __mem + 16, 16);
4470 const auto __c = _mm512_cvtepi8_epi32(__a);
4471 if constexpr (_S_size<_Tp> <= 48)
4472 {
4473 __builtin_memcpy(&__a, __mem + 32, _S_size<_Tp> - 32);
4474 const auto __d = _mm512_cvtepi8_epi32(__a);
4475 return _mm512_test_epi32_mask(__b, __b)
4476 | (_mm512_test_epi32_mask(__c, __c) << 16)
4477 | (_ULLong(_mm512_test_epi32_mask(__d, __d)) << 32);
4478 }
4479 else
4480 {
4481 __builtin_memcpy(&__a, __mem + 16, 16);
4482 const auto __d = _mm512_cvtepi8_epi32(__a);
4483 __builtin_memcpy(&__a, __mem + 32, _S_size<_Tp> - 48);
4484 const auto __e = _mm512_cvtepi8_epi32(__a);
4485 return _mm512_test_epi32_mask(__b, __b)
4486 | (_mm512_test_epi32_mask(__c, __c) << 16)
4487 | (_ULLong(_mm512_test_epi32_mask(__d, __d)) << 32)
4488 | (_ULLong(_mm512_test_epi32_mask(__e, __e)) << 48);
4489 }
4490 }
4491 else
4492 __assert_unreachable<_Tp>();
4493 }
4494 else if constexpr (sizeof(_Tp) == 8 && _S_size<_Tp> == 2)
4495 return __vector_bitcast<_Tp>(
4496 __vector_type16_t<int>{-int(__mem[0]), -int(__mem[0]),
4497 -int(__mem[1]), -int(__mem[1])});
4498 else if constexpr (sizeof(_Tp) == 8 && _S_size<_Tp> <= 4 && __have_avx)
4499 {
4500 int __bool4 = 0;
4501 __builtin_memcpy(&__bool4, __mem, _S_size<_Tp>);
4502 const auto __k = __to_intrin(
4503 (__vector_broadcast<4>(__bool4)
4504 & __make_vector<int>(0x1, 0x100, 0x10000,
4505 _S_size<_Tp> == 4 ? 0x1000000 : 0))
4506 != 0);
4507 return __vector_bitcast<_Tp>(
4508 __concat(_mm_unpacklo_epi32(__k, __k),
4509 _mm_unpackhi_epi32(__k, __k)));
4510 }
4511 else if constexpr (sizeof(_Tp) == 4 && _S_size<_Tp> <= 4)
4512 {
4513 int __bools = 0;
4514 __builtin_memcpy(&__bools, __mem, _S_size<_Tp>);
4515 if constexpr (__have_sse2)
4516 {
4517 __m128i __k = _mm_cvtsi32_si128(__bools);
4518 __k = _mm_cmpgt_epi16(_mm_unpacklo_epi8(__k, __k), __m128i());
4519 return __vector_bitcast<_Tp, _S_size<_Tp>>(
4520 _mm_unpacklo_epi16(__k, __k));
4521 }
4522 else
4523 {
4524 __m128 __k = _mm_cvtpi8_ps(_mm_cvtsi32_si64(__bools));
4525 _mm_empty();
4526 return __vector_bitcast<_Tp, _S_size<_Tp>>(
4527 _mm_cmpgt_ps(__k, __m128()));
4528 }
4529 }
4530 else if constexpr (sizeof(_Tp) == 4 && _S_size<_Tp> <= 8)
4531 {
4532 __m128i __k = {};
4533 __builtin_memcpy(&__k, __mem, _S_size<_Tp>);
4534 __k = _mm_cmpgt_epi16(_mm_unpacklo_epi8(__k, __k), __m128i());
4535 return __vector_bitcast<_Tp>(
4536 __concat(_mm_unpacklo_epi16(__k, __k),
4537 _mm_unpackhi_epi16(__k, __k)));
4538 }
4539 else if constexpr (sizeof(_Tp) == 2 && _S_size<_Tp> <= 16)
4540 {
4541 __m128i __k = {};
4542 __builtin_memcpy(&__k, __mem, _S_size<_Tp>);
4543 __k = _mm_cmpgt_epi8(__k, __m128i());
4544 if constexpr (_S_size<_Tp> <= 8)
4545 return __vector_bitcast<_Tp, _S_size<_Tp>>(
4546 _mm_unpacklo_epi8(__k, __k));
4547 else
4548 return __concat(_mm_unpacklo_epi8(__k, __k),
4549 _mm_unpackhi_epi8(__k, __k));
4550 }
4551 else
4552 return _Base::template _S_load<_Tp>(__mem);
4553 }
4554
4555 // }}}
4556 // _S_from_bitmask{{{
4557 template <size_t _Np, typename _Tp>
4558 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
4559 _S_from_bitmask(_SanitizedBitMask<_Np> __bits, _TypeTag<_Tp>)
4560 {
4561 static_assert(is_same_v<_Tp, __int_for_sizeof_t<_Tp>>);
4562 if constexpr (__is_avx512_abi<_Abi>())
4563 return __bits._M_to_bits();
4564 else
4565 return _S_to_maskvector<_Tp, _S_size<_Tp>>(__bits);
4566 }
4567
4568 // }}}
4569 // _S_masked_load {{{2
4570 template <typename _Tp, size_t _Np>
4571 static inline _SimdWrapper<_Tp, _Np>
4572 _S_masked_load(_SimdWrapper<_Tp, _Np> __merge,
4573 _SimdWrapper<_Tp, _Np> __mask, const bool* __mem) noexcept
4574 {
4575 if constexpr (__is_avx512_abi<_Abi>())
4576 {
4577 if constexpr (__have_avx512bw_vl)
4578 {
4579 if constexpr (_Np <= 16)
4580 {
4581 const auto __a
4582 = _mm_mask_loadu_epi8(__m128i(), __mask, __mem);
4583 return (__merge & ~__mask) | _mm_test_epi8_mask(__a, __a);
4584 }
4585 else if constexpr (_Np <= 32)
4586 {
4587 const auto __a
4588 = _mm256_mask_loadu_epi8(__m256i(), __mask, __mem);
4589 return (__merge & ~__mask)
4590 | _mm256_test_epi8_mask(__a, __a);
4591 }
4592 else if constexpr (_Np <= 64)
4593 {
4594 const auto __a
4595 = _mm512_mask_loadu_epi8(__m512i(), __mask, __mem);
4596 return (__merge & ~__mask)
4597 | _mm512_test_epi8_mask(__a, __a);
4598 }
4599 else
4600 __assert_unreachable<_Tp>();
4601 }
4602 else
4603 {
4604 _BitOps::_S_bit_iteration(__mask, [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
4605 __merge._M_set(__i, __mem[__i]);
4606 });
4607 return __merge;
4608 }
4609 }
4610 else if constexpr (__have_avx512bw_vl && _Np == 32 && sizeof(_Tp) == 1)
4611 {
4612 const auto __k = _S_to_bits(__mask)._M_to_bits();
4613 __merge = _mm256_mask_sub_epi8(__to_intrin(__merge), __k, __m256i(),
4614 _mm256_mask_loadu_epi8(__m256i(),
4615 __k, __mem));
4616 }
4617 else if constexpr (__have_avx512bw_vl && _Np == 16 && sizeof(_Tp) == 1)
4618 {
4619 const auto __k = _S_to_bits(__mask)._M_to_bits();
4620 __merge
4621 = _mm_mask_sub_epi8(__vector_bitcast<_LLong>(__merge), __k,
4622 __m128i(),
4623 _mm_mask_loadu_epi8(__m128i(), __k, __mem));
4624 }
4625 else if constexpr (__have_avx512bw_vl && _Np == 16 && sizeof(_Tp) == 2)
4626 {
4627 const auto __k = _S_to_bits(__mask)._M_to_bits();
4628 __merge = _mm256_mask_sub_epi16(
4629 __vector_bitcast<_LLong>(__merge), __k, __m256i(),
4630 _mm256_cvtepi8_epi16(_mm_mask_loadu_epi8(__m128i(), __k, __mem)));
4631 }
4632 else if constexpr (__have_avx512bw_vl && _Np == 8 && sizeof(_Tp) == 2)
4633 {
4634 const auto __k = _S_to_bits(__mask)._M_to_bits();
4635 __merge = _mm_mask_sub_epi16(
4636 __vector_bitcast<_LLong>(__merge), __k, __m128i(),
4637 _mm_cvtepi8_epi16(_mm_mask_loadu_epi8(__m128i(), __k, __mem)));
4638 }
4639 else if constexpr (__have_avx512bw_vl && _Np == 8 && sizeof(_Tp) == 4)
4640 {
4641 const auto __k = _S_to_bits(__mask)._M_to_bits();
4642 __merge = __vector_bitcast<_Tp>(_mm256_mask_sub_epi32(
4643 __vector_bitcast<_LLong>(__merge), __k, __m256i(),
4644 _mm256_cvtepi8_epi32(
4645 _mm_mask_loadu_epi8(__m128i(), __k, __mem))));
4646 }
4647 else if constexpr (__have_avx512bw_vl && _Np == 4 && sizeof(_Tp) == 4)
4648 {
4649 const auto __k = _S_to_bits(__mask)._M_to_bits();
4650 __merge = __vector_bitcast<_Tp>(_mm_mask_sub_epi32(
4651 __vector_bitcast<_LLong>(__merge), __k, __m128i(),
4652 _mm_cvtepi8_epi32(_mm_mask_loadu_epi8(__m128i(), __k, __mem))));
4653 }
4654 else if constexpr (__have_avx512bw_vl && _Np == 4 && sizeof(_Tp) == 8)
4655 {
4656 const auto __k = _S_to_bits(__mask)._M_to_bits();
4657 __merge = __vector_bitcast<_Tp>(_mm256_mask_sub_epi64(
4658 __vector_bitcast<_LLong>(__merge), __k, __m256i(),
4659 _mm256_cvtepi8_epi64(
4660 _mm_mask_loadu_epi8(__m128i(), __k, __mem))));
4661 }
4662 else if constexpr (__have_avx512bw_vl && _Np == 2 && sizeof(_Tp) == 8)
4663 {
4664 const auto __k = _S_to_bits(__mask)._M_to_bits();
4665 __merge = __vector_bitcast<_Tp>(_mm_mask_sub_epi64(
4666 __vector_bitcast<_LLong>(__merge), __k, __m128i(),
4667 _mm_cvtepi8_epi64(_mm_mask_loadu_epi8(__m128i(), __k, __mem))));
4668 }
4669 else
4670 return _Base::_S_masked_load(__merge, __mask, __mem);
4671 return __merge;
4672 }
4673
4674 // _S_store {{{2
4675 template <typename _Tp, size_t _Np>
4676 _GLIBCXX_SIMD_INTRINSIC static constexpr void
4677 _S_store(_SimdWrapper<_Tp, _Np> __v, bool* __mem) noexcept
4678 {
4679 if (__builtin_is_constant_evaluated())
4680 _Base::_S_store(__v, __mem);
4681 else if constexpr (__is_avx512_abi<_Abi>())
4682 {
4683 if constexpr (__have_avx512bw_vl)
4684 _CommonImplX86::_S_store<_Np>(
4685 __vector_bitcast<char>([](auto __data) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
4686 if constexpr (_Np <= 16)
4687 return _mm_maskz_set1_epi8(__data, 1);
4688 else if constexpr (_Np <= 32)
4689 return _mm256_maskz_set1_epi8(__data, 1);
4690 else
4691 return _mm512_maskz_set1_epi8(__data, 1);
4692 }(__v._M_data)),
4693 __mem);
4694 else if constexpr (_Np <= 8)
4695 _CommonImplX86::_S_store<_Np>(
4696 __vector_bitcast<char>(
4697#if defined __x86_64__
4698 __make_wrapper<_ULLong>(
4699 _pdep_u64(__v._M_data, 0x0101010101010101ULL), 0ull)
4700#else
4701 __make_wrapper<_UInt>(_pdep_u32(__v._M_data, 0x01010101U),
4702 _pdep_u32(__v._M_data >> 4,
4703 0x01010101U))
4704#endif
4705 ),
4706 __mem);
4707 else if constexpr (_Np <= 16)
4708 _mm512_mask_cvtepi32_storeu_epi8(
4709 __mem, 0xffffu >> (16 - _Np),
4710 _mm512_maskz_set1_epi32(__v._M_data, 1));
4711 else
4712 __assert_unreachable<_Tp>();
4713 }
4714 else if constexpr (__is_sse_abi<_Abi>()) //{{{
4715 {
4716 if constexpr (_Np == 2 && sizeof(_Tp) == 8)
4717 {
4718 const auto __k = __vector_bitcast<int>(__v);
4719 __mem[0] = -__k[1];
4720 __mem[1] = -__k[3];
4721 }
4722 else if constexpr (_Np <= 4 && sizeof(_Tp) == 4)
4723 {
4724 if constexpr (__have_sse2)
4725 {
4726 const unsigned __bool4
4727 = __vector_bitcast<_UInt>(_mm_packs_epi16(
4728 _mm_packs_epi32(__intrin_bitcast<__m128i>(
4729 __to_intrin(__v)),
4730 __m128i()),
4731 __m128i()))[0]
4732 & 0x01010101u;
4733 __builtin_memcpy(__mem, &__bool4, _Np);
4734 }
4735 else if constexpr (__have_mmx)
4736 {
4737 const __m64 __k = _mm_cvtps_pi8(
4738 __and(__to_intrin(__v), _mm_set1_ps(1.f)));
4739 __builtin_memcpy(__mem, &__k, _Np);
4740 _mm_empty();
4741 }
4742 else
4743 return _Base::_S_store(__v, __mem);
4744 }
4745 else if constexpr (_Np <= 8 && sizeof(_Tp) == 2)
4746 {
4747 _CommonImplX86::_S_store<_Np>(
4748 __vector_bitcast<char>(_mm_packs_epi16(
4749 __to_intrin(__vector_bitcast<_UShort>(__v) >> 15),
4750 __m128i())),
4751 __mem);
4752 }
4753 else if constexpr (_Np <= 16 && sizeof(_Tp) == 1)
4754 _CommonImplX86::_S_store<_Np>(__v._M_data & 1, __mem);
4755 else
4756 __assert_unreachable<_Tp>();
4757 } // }}}
4758 else if constexpr (__is_avx_abi<_Abi>()) // {{{
4759 {
4760 if constexpr (_Np <= 4 && sizeof(_Tp) == 8)
4761 {
4762 auto __k = __intrin_bitcast<__m256i>(__to_intrin(__v));
4763 int __bool4{};
4764 if constexpr (__have_avx2)
4765 __bool4 = _mm256_movemask_epi8(__k);
4766 else
4767 __bool4 = (_mm_movemask_epi8(__lo128(__k))
4768 | (_mm_movemask_epi8(__hi128(__k)) << 16));
4769 __bool4 &= 0x01010101;
4770 __builtin_memcpy(__mem, &__bool4, _Np);
4771 }
4772 else if constexpr (_Np <= 8 && sizeof(_Tp) == 4)
4773 {
4774 const auto __k = __intrin_bitcast<__m256i>(__to_intrin(__v));
4775 const auto __k2
4776 = _mm_srli_epi16(_mm_packs_epi16(__lo128(__k), __hi128(__k)),
4777 15);
4778 const auto __k3
4779 = __vector_bitcast<char>(_mm_packs_epi16(__k2, __m128i()));
4780 _CommonImplX86::_S_store<_Np>(__k3, __mem);
4781 }
4782 else if constexpr (_Np <= 16 && sizeof(_Tp) == 2)
4783 {
4784 if constexpr (__have_avx2)
4785 {
4786 const auto __x = _mm256_srli_epi16(__to_intrin(__v), 15);
4787 const auto __bools = __vector_bitcast<char>(
4788 _mm_packs_epi16(__lo128(__x), __hi128(__x)));
4789 _CommonImplX86::_S_store<_Np>(__bools, __mem);
4790 }
4791 else
4792 {
4793 const auto __bools
4794 = 1
4795 & __vector_bitcast<_UChar>(
4796 _mm_packs_epi16(__lo128(__to_intrin(__v)),
4797 __hi128(__to_intrin(__v))));
4798 _CommonImplX86::_S_store<_Np>(__bools, __mem);
4799 }
4800 }
4801 else if constexpr (_Np <= 32 && sizeof(_Tp) == 1)
4802 _CommonImplX86::_S_store<_Np>(1 & __v._M_data, __mem);
4803 else
4804 __assert_unreachable<_Tp>();
4805 } // }}}
4806 else
4807 __assert_unreachable<_Tp>();
4808 }
4809
4810 // _S_masked_store {{{2
4811 template <typename _Tp, size_t _Np>
4812 static inline void
4813 _S_masked_store(const _SimdWrapper<_Tp, _Np> __v, bool* __mem,
4814 const _SimdWrapper<_Tp, _Np> __k) noexcept
4815 {
4816 if constexpr (__is_avx512_abi<_Abi>())
4817 {
4818 static_assert(is_same_v<_Tp, bool>);
4819 if constexpr (_Np <= 16 && __have_avx512bw_vl)
4820 _mm_mask_storeu_epi8(__mem, __k, _mm_maskz_set1_epi8(__v, 1));
4821 else if constexpr (_Np <= 16)
4822 _mm512_mask_cvtepi32_storeu_epi8(__mem, __k,
4823 _mm512_maskz_set1_epi32(__v, 1));
4824 else if constexpr (_Np <= 32 && __have_avx512bw_vl)
4825 _mm256_mask_storeu_epi8(__mem, __k,
4826 _mm256_maskz_set1_epi8(__v, 1));
4827 else if constexpr (_Np <= 32 && __have_avx512bw)
4828 _mm256_mask_storeu_epi8(__mem, __k,
4829 __lo256(_mm512_maskz_set1_epi8(__v, 1)));
4830 else if constexpr (_Np <= 64 && __have_avx512bw)
4831 _mm512_mask_storeu_epi8(__mem, __k,
4832 _mm512_maskz_set1_epi8(__v, 1));
4833 else
4834 __assert_unreachable<_Tp>();
4835 }
4836 else
4837 _Base::_S_masked_store(__v, __mem, __k);
4838 }
4839
4840 // logical and bitwise operators {{{2
4841 template <typename _Tp, size_t _Np>
4842 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
4843 _S_logical_and(const _SimdWrapper<_Tp, _Np>& __x, const _SimdWrapper<_Tp, _Np>& __y)
4844 {
4845 if constexpr (is_same_v<_Tp, bool>)
4846 {
4847 if (__builtin_is_constant_evaluated())
4848 return __x._M_data & __y._M_data;
4849 else if constexpr (__have_avx512dq && _Np <= 8)
4850 return _kand_mask8(__x._M_data, __y._M_data);
4851 else if constexpr (_Np <= 16)
4852 return _kand_mask16(__x._M_data, __y._M_data);
4853 else if constexpr (__have_avx512bw && _Np <= 32)
4854 return _kand_mask32(__x._M_data, __y._M_data);
4855 else if constexpr (__have_avx512bw && _Np <= 64)
4856 return _kand_mask64(__x._M_data, __y._M_data);
4857 else
4858 __assert_unreachable<_Tp>();
4859 }
4860 else
4861 return _Base::_S_logical_and(__x, __y);
4862 }
4863
4864 template <typename _Tp, size_t _Np>
4865 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
4866 _S_logical_or(const _SimdWrapper<_Tp, _Np>& __x, const _SimdWrapper<_Tp, _Np>& __y)
4867 {
4868 if constexpr (is_same_v<_Tp, bool>)
4869 {
4870 if (__builtin_is_constant_evaluated())
4871 return __x._M_data | __y._M_data;
4872 else if constexpr (__have_avx512dq && _Np <= 8)
4873 return _kor_mask8(__x._M_data, __y._M_data);
4874 else if constexpr (_Np <= 16)
4875 return _kor_mask16(__x._M_data, __y._M_data);
4876 else if constexpr (__have_avx512bw && _Np <= 32)
4877 return _kor_mask32(__x._M_data, __y._M_data);
4878 else if constexpr (__have_avx512bw && _Np <= 64)
4879 return _kor_mask64(__x._M_data, __y._M_data);
4880 else
4881 __assert_unreachable<_Tp>();
4882 }
4883 else
4884 return _Base::_S_logical_or(__x, __y);
4885 }
4886
4887 template <typename _Tp, size_t _Np>
4888 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
4889 _S_bit_not(const _SimdWrapper<_Tp, _Np>& __x)
4890 {
4891 if constexpr (is_same_v<_Tp, bool>)
4892 {
4893 if (__builtin_is_constant_evaluated())
4894 return __x._M_data ^ _Abi::template __implicit_mask_n<_Np>();
4895 else if constexpr (__have_avx512dq && _Np <= 8)
4896 return _kandn_mask8(__x._M_data,
4897 _Abi::template __implicit_mask_n<_Np>());
4898 else if constexpr (_Np <= 16)
4899 return _kandn_mask16(__x._M_data,
4900 _Abi::template __implicit_mask_n<_Np>());
4901 else if constexpr (__have_avx512bw && _Np <= 32)
4902 return _kandn_mask32(__x._M_data,
4903 _Abi::template __implicit_mask_n<_Np>());
4904 else if constexpr (__have_avx512bw && _Np <= 64)
4905 return _kandn_mask64(__x._M_data,
4906 _Abi::template __implicit_mask_n<_Np>());
4907 else
4908 __assert_unreachable<_Tp>();
4909 }
4910 else
4911 return _Base::_S_bit_not(__x);
4912 }
4913
4914 template <typename _Tp, size_t _Np>
4915 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
4916 _S_bit_and(const _SimdWrapper<_Tp, _Np>& __x, const _SimdWrapper<_Tp, _Np>& __y)
4917 {
4918 if constexpr (is_same_v<_Tp, bool>)
4919 {
4920 if (__builtin_is_constant_evaluated())
4921 return __x._M_data & __y._M_data;
4922 else if constexpr (__have_avx512dq && _Np <= 8)
4923 return _kand_mask8(__x._M_data, __y._M_data);
4924 else if constexpr (_Np <= 16)
4925 return _kand_mask16(__x._M_data, __y._M_data);
4926 else if constexpr (__have_avx512bw && _Np <= 32)
4927 return _kand_mask32(__x._M_data, __y._M_data);
4928 else if constexpr (__have_avx512bw && _Np <= 64)
4929 return _kand_mask64(__x._M_data, __y._M_data);
4930 else
4931 __assert_unreachable<_Tp>();
4932 }
4933 else
4934 return _Base::_S_bit_and(__x, __y);
4935 }
4936
4937 template <typename _Tp, size_t _Np>
4938 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
4939 _S_bit_or(const _SimdWrapper<_Tp, _Np>& __x, const _SimdWrapper<_Tp, _Np>& __y)
4940 {
4941 if constexpr (is_same_v<_Tp, bool>)
4942 {
4943 if (__builtin_is_constant_evaluated())
4944 return __x._M_data | __y._M_data;
4945 else if constexpr (__have_avx512dq && _Np <= 8)
4946 return _kor_mask8(__x._M_data, __y._M_data);
4947 else if constexpr (_Np <= 16)
4948 return _kor_mask16(__x._M_data, __y._M_data);
4949 else if constexpr (__have_avx512bw && _Np <= 32)
4950 return _kor_mask32(__x._M_data, __y._M_data);
4951 else if constexpr (__have_avx512bw && _Np <= 64)
4952 return _kor_mask64(__x._M_data, __y._M_data);
4953 else
4954 __assert_unreachable<_Tp>();
4955 }
4956 else
4957 return _Base::_S_bit_or(__x, __y);
4958 }
4959
4960 template <typename _Tp, size_t _Np>
4961 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
4962 _S_bit_xor(const _SimdWrapper<_Tp, _Np>& __x, const _SimdWrapper<_Tp, _Np>& __y)
4963 {
4964 if constexpr (is_same_v<_Tp, bool>)
4965 {
4966 if (__builtin_is_constant_evaluated())
4967 return __x._M_data ^ __y._M_data;
4968 else if constexpr (__have_avx512dq && _Np <= 8)
4969 return _kxor_mask8(__x._M_data, __y._M_data);
4970 else if constexpr (_Np <= 16)
4971 return _kxor_mask16(__x._M_data, __y._M_data);
4972 else if constexpr (__have_avx512bw && _Np <= 32)
4973 return _kxor_mask32(__x._M_data, __y._M_data);
4974 else if constexpr (__have_avx512bw && _Np <= 64)
4975 return _kxor_mask64(__x._M_data, __y._M_data);
4976 else
4977 __assert_unreachable<_Tp>();
4978 }
4979 else
4980 return _Base::_S_bit_xor(__x, __y);
4981 }
4982
4983 //}}}2
4984 // _S_masked_assign{{{
4985 template <size_t _Np>
4986 _GLIBCXX_SIMD_INTRINSIC static void
4987 _S_masked_assign(_SimdWrapper<bool, _Np> __k,
4988 _SimdWrapper<bool, _Np>& __lhs, _SimdWrapper<bool, _Np> __rhs)
4989 {
4990 __lhs._M_data
4991 = (~__k._M_data & __lhs._M_data) | (__k._M_data & __rhs._M_data);
4992 }
4993
4994 template <size_t _Np>
4995 _GLIBCXX_SIMD_INTRINSIC static void
4996 _S_masked_assign(_SimdWrapper<bool, _Np> __k,
4997 _SimdWrapper<bool, _Np>& __lhs, bool __rhs)
4998 {
4999 if (__rhs)
5000 __lhs._M_data = __k._M_data | __lhs._M_data;
5001 else
5002 __lhs._M_data = ~__k._M_data & __lhs._M_data;
5003 }
5004
5005 using _MaskImplBuiltin<_Abi>::_S_masked_assign;
5006
5007 //}}}
5008 // _S_all_of {{{
5009 template <typename _Tp>
5010 _GLIBCXX_SIMD_INTRINSIC static bool
5011 _S_all_of(simd_mask<_Tp, _Abi> __k)
5012 {
5013 if constexpr (__is_sse_abi<_Abi>() || __is_avx_abi<_Abi>())
5014 {
5015 constexpr size_t _Np = simd_size_v<_Tp, _Abi>;
5016 using _TI = __intrinsic_type_t<_Tp, _Np>;
5017 const _TI __a = reinterpret_cast<_TI>(__to_intrin(__data(__k)));
5018 if constexpr (__have_sse4_1)
5019 {
5020 _GLIBCXX_SIMD_USE_CONSTEXPR _TI __b
5021 = _Abi::template _S_implicit_mask_intrin<_Tp>();
5022 return 0 != __testc(__a, __b);
5023 }
5024 else if constexpr (is_same_v<_Tp, float>)
5025 return (_mm_movemask_ps(__a) & ((1 << _Np) - 1))
5026 == (1 << _Np) - 1;
5027 else if constexpr (is_same_v<_Tp, double>)
5028 return (_mm_movemask_pd(__a) & ((1 << _Np) - 1))
5029 == (1 << _Np) - 1;
5030 else
5031 return (_mm_movemask_epi8(__a) & ((1 << (_Np * sizeof(_Tp))) - 1))
5032 == (1 << (_Np * sizeof(_Tp))) - 1;
5033 }
5034 else if constexpr (__is_avx512_abi<_Abi>())
5035 {
5036 constexpr auto _Mask = _Abi::template _S_implicit_mask<_Tp>();
5037 const auto __kk = __k._M_data._M_data;
5038 if constexpr (sizeof(__kk) == 1)
5039 {
5040 if constexpr (__have_avx512dq)
5041 return _kortestc_mask8_u8(__kk, _Mask == 0xff
5042 ? __kk
5043 : __mmask8(~_Mask));
5044 else
5045 return _kortestc_mask16_u8(__kk, __mmask16(~_Mask));
5046 }
5047 else if constexpr (sizeof(__kk) == 2)
5048 return _kortestc_mask16_u8(__kk, _Mask == 0xffff
5049 ? __kk
5050 : __mmask16(~_Mask));
5051 else if constexpr (sizeof(__kk) == 4 && __have_avx512bw)
5052 return _kortestc_mask32_u8(__kk, _Mask == 0xffffffffU
5053 ? __kk
5054 : __mmask32(~_Mask));
5055 else if constexpr (sizeof(__kk) == 8 && __have_avx512bw)
5056 return _kortestc_mask64_u8(__kk, _Mask == 0xffffffffffffffffULL
5057 ? __kk
5058 : __mmask64(~_Mask));
5059 else
5060 __assert_unreachable<_Tp>();
5061 }
5062 }
5063
5064 // }}}
5065 // _S_any_of {{{
5066 template <typename _Tp>
5067 _GLIBCXX_SIMD_INTRINSIC static bool
5068 _S_any_of(simd_mask<_Tp, _Abi> __k)
5069 {
5070 if constexpr (__is_sse_abi<_Abi>() || __is_avx_abi<_Abi>())
5071 {
5072 constexpr size_t _Np = simd_size_v<_Tp, _Abi>;
5073 using _TI = __intrinsic_type_t<_Tp, _Np>;
5074 const _TI __a = reinterpret_cast<_TI>(__to_intrin(__data(__k)));
5075 if constexpr (__have_sse4_1)
5076 {
5077 if constexpr (_Abi::template _S_is_partial<
5078 _Tp> || sizeof(__k) < 16)
5079 {
5080 _GLIBCXX_SIMD_USE_CONSTEXPR _TI __b
5081 = _Abi::template _S_implicit_mask_intrin<_Tp>();
5082 return 0 == __testz(__a, __b);
5083 }
5084 else
5085 return 0 == __testz(__a, __a);
5086 }
5087 else if constexpr (is_same_v<_Tp, float>)
5088 return (_mm_movemask_ps(__a) & ((1 << _Np) - 1)) != 0;
5089 else if constexpr (is_same_v<_Tp, double>)
5090 return (_mm_movemask_pd(__a) & ((1 << _Np) - 1)) != 0;
5091 else
5092 return (_mm_movemask_epi8(__a) & ((1 << (_Np * sizeof(_Tp))) - 1))
5093 != 0;
5094 }
5095 else if constexpr (__is_avx512_abi<_Abi>())
5096 return (__k._M_data._M_data & _Abi::template _S_implicit_mask<_Tp>())
5097 != 0;
5098 }
5099
5100 // }}}
5101 // _S_none_of {{{
5102 template <typename _Tp>
5103 _GLIBCXX_SIMD_INTRINSIC static bool
5104 _S_none_of(simd_mask<_Tp, _Abi> __k)
5105 {
5106 if constexpr (__is_sse_abi<_Abi>() || __is_avx_abi<_Abi>())
5107 {
5108 constexpr size_t _Np = simd_size_v<_Tp, _Abi>;
5109 using _TI = __intrinsic_type_t<_Tp, _Np>;
5110 const _TI __a = reinterpret_cast<_TI>(__to_intrin(__data(__k)));
5111 if constexpr (__have_sse4_1)
5112 {
5113 if constexpr (_Abi::template _S_is_partial<
5114 _Tp> || sizeof(__k) < 16)
5115 {
5116 _GLIBCXX_SIMD_USE_CONSTEXPR _TI __b
5117 = _Abi::template _S_implicit_mask_intrin<_Tp>();
5118 return 0 != __testz(__a, __b);
5119 }
5120 else
5121 return 0 != __testz(__a, __a);
5122 }
5123 else if constexpr (is_same_v<_Tp, float>)
5124 return (__movemask(__a) & ((1 << _Np) - 1)) == 0;
5125 else if constexpr (is_same_v<_Tp, double>)
5126 return (__movemask(__a) & ((1 << _Np) - 1)) == 0;
5127 else
5128 return (__movemask(__a) & int((1ull << (_Np * sizeof(_Tp))) - 1))
5129 == 0;
5130 }
5131 else if constexpr (__is_avx512_abi<_Abi>())
5132 return (__k._M_data._M_data & _Abi::template _S_implicit_mask<_Tp>())
5133 == 0;
5134 }
5135
5136 // }}}
5137 // _S_some_of {{{
5138 template <typename _Tp>
5139 _GLIBCXX_SIMD_INTRINSIC static bool
5140 _S_some_of(simd_mask<_Tp, _Abi> __k)
5141 {
5142 if constexpr (__is_sse_abi<_Abi>() || __is_avx_abi<_Abi>())
5143 {
5144 constexpr size_t _Np = simd_size_v<_Tp, _Abi>;
5145 using _TI = __intrinsic_type_t<_Tp, _Np>;
5146 const _TI __a = reinterpret_cast<_TI>(__to_intrin(__data(__k)));
5147 if constexpr (__have_sse4_1)
5148 {
5149 _GLIBCXX_SIMD_USE_CONSTEXPR _TI __b
5150 = _Abi::template _S_implicit_mask_intrin<_Tp>();
5151 return 0 != __testnzc(__a, __b);
5152 }
5153 else if constexpr (is_same_v<_Tp, float>)
5154 {
5155 constexpr int __allbits = (1 << _Np) - 1;
5156 const auto __tmp = _mm_movemask_ps(__a) & __allbits;
5157 return __tmp > 0 && __tmp < __allbits;
5158 }
5159 else if constexpr (is_same_v<_Tp, double>)
5160 {
5161 constexpr int __allbits = (1 << _Np) - 1;
5162 const auto __tmp = _mm_movemask_pd(__a) & __allbits;
5163 return __tmp > 0 && __tmp < __allbits;
5164 }
5165 else
5166 {
5167 constexpr int __allbits = (1 << (_Np * sizeof(_Tp))) - 1;
5168 const auto __tmp = _mm_movemask_epi8(__a) & __allbits;
5169 return __tmp > 0 && __tmp < __allbits;
5170 }
5171 }
5172 else if constexpr (__is_avx512_abi<_Abi>())
5173 return _S_any_of(__k) && !_S_all_of(__k);
5174 else
5175 __assert_unreachable<_Tp>();
5176 }
5177
5178 // }}}
5179 // _S_popcount {{{
5180 template <typename _Tp>
5181 _GLIBCXX_SIMD_INTRINSIC static int
5182 _S_popcount(simd_mask<_Tp, _Abi> __k)
5183 {
5184 constexpr size_t _Np = simd_size_v<_Tp, _Abi>;
5185 const auto __kk = _Abi::_S_masked(__k._M_data)._M_data;
5186 if constexpr (__is_avx512_abi<_Abi>())
5187 {
5188 if constexpr (_Np > 32)
5189 return __builtin_popcountll(__kk);
5190 else
5191 return __builtin_popcount(__kk);
5192 }
5193 else
5194 {
5195 if constexpr (__have_popcnt)
5196 {
5197 int __bits
5198 = __movemask(__to_intrin(__vector_bitcast<_Tp>(__kk)));
5199 const int __count = __builtin_popcount(__bits);
5200 return is_integral_v<_Tp> ? __count / sizeof(_Tp) : __count;
5201 }
5202 else if constexpr (_Np == 2 && sizeof(_Tp) == 8)
5203 {
5204 const int mask = _mm_movemask_pd(__auto_bitcast(__kk));
5205 return mask - (mask >> 1);
5206 }
5207 else if constexpr (_Np <= 4 && sizeof(_Tp) == 8)
5208 {
5209 auto __x = -(__lo128(__kk) + __hi128(__kk));
5210 return __x[0] + __x[1];
5211 }
5212 else if constexpr (_Np <= 4 && sizeof(_Tp) == 4)
5213 {
5214 if constexpr (__have_sse2)
5215 {
5216 __m128i __x = __intrin_bitcast<__m128i>(__to_intrin(__kk));
5217 __x = _mm_add_epi32(
5218 __x, _mm_shuffle_epi32(__x, _MM_SHUFFLE(0, 1, 2, 3)));
5219 __x = _mm_add_epi32(
5220 __x, _mm_shufflelo_epi16(__x, _MM_SHUFFLE(1, 0, 3, 2)));
5221 return -_mm_cvtsi128_si32(__x);
5222 }
5223 else
5224 return __builtin_popcount(
5225 _mm_movemask_ps(__auto_bitcast(__kk)));
5226 }
5227 else if constexpr (_Np <= 8 && sizeof(_Tp) == 2)
5228 {
5229 auto __x = __to_intrin(__kk);
5230 __x = _mm_add_epi16(__x,
5231 _mm_shuffle_epi32(__x,
5232 _MM_SHUFFLE(0, 1, 2, 3)));
5233 __x = _mm_add_epi16(
5234 __x, _mm_shufflelo_epi16(__x, _MM_SHUFFLE(0, 1, 2, 3)));
5235 __x = _mm_add_epi16(
5236 __x, _mm_shufflelo_epi16(__x, _MM_SHUFFLE(2, 3, 0, 1)));
5237 return -short(_mm_extract_epi16(__x, 0));
5238 }
5239 else if constexpr (_Np <= 16 && sizeof(_Tp) == 1)
5240 {
5241 auto __x = __to_intrin(__kk);
5242 __x = _mm_add_epi8(__x,
5243 _mm_shuffle_epi32(__x,
5244 _MM_SHUFFLE(0, 1, 2, 3)));
5245 __x = _mm_add_epi8(__x,
5246 _mm_shufflelo_epi16(__x, _MM_SHUFFLE(0, 1, 2,
5247 3)));
5248 __x = _mm_add_epi8(__x,
5249 _mm_shufflelo_epi16(__x, _MM_SHUFFLE(2, 3, 0,
5250 1)));
5251 auto __y = -__vector_bitcast<_UChar>(__x);
5252 if constexpr (__have_sse4_1)
5253 return __y[0] + __y[1];
5254 else
5255 {
5256 unsigned __z = _mm_extract_epi16(__to_intrin(__y), 0);
5257 return (__z & 0xff) + (__z >> 8);
5258 }
5259 }
5260 else if constexpr (sizeof(__kk) == 32)
5261 {
5262 // The following works only as long as the implementations above
5263 // use a summation
5264 using _I = __int_for_sizeof_t<_Tp>;
5265 const auto __as_int = __vector_bitcast<_I>(__kk);
5266 _MaskImplX86<simd_abi::__sse>::_S_popcount(
5267 simd_mask<_I, simd_abi::__sse>(__private_init,
5268 __lo128(__as_int)
5269 + __hi128(__as_int)));
5270 }
5271 else
5272 __assert_unreachable<_Tp>();
5273 }
5274 }
5275
5276 // }}}
5277 // _S_find_first_set {{{
5278 template <typename _Tp>
5279 _GLIBCXX_SIMD_INTRINSIC static int
5280 _S_find_first_set(simd_mask<_Tp, _Abi> __k)
5281 {
5282 if constexpr (__is_avx512_abi<_Abi>())
5283 return std::__countr_zero(__k._M_data._M_data);
5284 else
5285 return _Base::_S_find_first_set(__k);
5286 }
5287
5288 // }}}
5289 // _S_find_last_set {{{
5290 template <typename _Tp>
5291 _GLIBCXX_SIMD_INTRINSIC static int
5292 _S_find_last_set(simd_mask<_Tp, _Abi> __k)
5293 {
5294 if constexpr (__is_avx512_abi<_Abi>())
5295 return std::__bit_width(_Abi::_S_masked(__k._M_data)._M_data) - 1;
5296 else
5297 return _Base::_S_find_last_set(__k);
5298 }
5299
5300 // }}}
5301 };
5302
5303// }}}
5304
5305_GLIBCXX_SIMD_END_NAMESPACE
5306#endif // __cplusplus >= 201703L
5307#endif // _GLIBCXX_EXPERIMENTAL_SIMD_X86_H_
5308
5309// vim: foldmethod=marker sw=2 noet ts=8 sts=2 tw=80
typename conditional< _Cond, _Iftrue, _Iffalse >::type conditional_t
Alias template for conditional.
Definition type_traits:2713
typename make_signed< _Tp >::type make_signed_t
Alias template for make_signed.
Definition type_traits:2070
constexpr const _Tp & min(const _Tp &, const _Tp &)
This does what you think it does.