source: Daodan/MSYS2/mingw32/include/c++/11.2.0/experimental/bits/simd_x86.h@ 1177

Last change on this file since 1177 was 1166, checked in by rossy, 3 years ago

Daodan: Replace MinGW build env with an up-to-date MSYS2 env
File size: 188.5 KB
Line 
1// Simd x86 specific implementations -*- C++ -*-
2
3// Copyright (C) 2020-2021 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 {
44 return reinterpret_cast<__vector_type_t<__int_for_sizeof_t<_Tp>, _Np>>(
45 __x._M_data);
46 }
47
48template <typename _TV,
49 typename _TVT
50 = enable_if_t<__is_vector_type_v<_TV>, _VectorTraits<_TV>>,
51 typename _Up = __int_for_sizeof_t<typename _TVT::value_type>>
52 _GLIBCXX_SIMD_INTRINSIC constexpr __vector_type_t<_Up, _TVT::_S_full_size>
53 __to_masktype(_TV __x)
54 { return reinterpret_cast<__vector_type_t<_Up, _TVT::_S_full_size>>(__x); }
55
56// }}}
57// __interleave128_lo {{{
58template <typename _Ap, typename _Bp, typename _Tp = common_type_t<_Ap, _Bp>,
59 typename _Trait = _VectorTraits<_Tp>>
60 _GLIBCXX_SIMD_INTRINSIC constexpr _Tp
61 __interleave128_lo(const _Ap& __av, const _Bp& __bv)
62 {
63 const _Tp __a(__av);
64 const _Tp __b(__bv);
65 if constexpr (sizeof(_Tp) == 16 && _Trait::_S_full_size == 2)
66 return _Tp{__a[0], __b[0]};
67 else if constexpr (sizeof(_Tp) == 16 && _Trait::_S_full_size == 4)
68 return _Tp{__a[0], __b[0], __a[1], __b[1]};
69 else if constexpr (sizeof(_Tp) == 16 && _Trait::_S_full_size == 8)
70 return _Tp{__a[0], __b[0], __a[1], __b[1],
71 __a[2], __b[2], __a[3], __b[3]};
72 else if constexpr (sizeof(_Tp) == 16 && _Trait::_S_full_size == 16)
73 return _Tp{__a[0], __b[0], __a[1], __b[1], __a[2], __b[2],
74 __a[3], __b[3], __a[4], __b[4], __a[5], __b[5],
75 __a[6], __b[6], __a[7], __b[7]};
76 else if constexpr (sizeof(_Tp) == 32 && _Trait::_S_full_size == 4)
77 return _Tp{__a[0], __b[0], __a[2], __b[2]};
78 else if constexpr (sizeof(_Tp) == 32 && _Trait::_S_full_size == 8)
79 return _Tp{__a[0], __b[0], __a[1], __b[1],
80 __a[4], __b[4], __a[5], __b[5]};
81 else if constexpr (sizeof(_Tp) == 32 && _Trait::_S_full_size == 16)
82 return _Tp{__a[0], __b[0], __a[1], __b[1], __a[2], __b[2],
83 __a[3], __b[3], __a[8], __b[8], __a[9], __b[9],
84 __a[10], __b[10], __a[11], __b[11]};
85 else if constexpr (sizeof(_Tp) == 32 && _Trait::_S_full_size == 32)
86 return _Tp{__a[0], __b[0], __a[1], __b[1], __a[2], __b[2], __a[3],
87 __b[3], __a[4], __b[4], __a[5], __b[5], __a[6], __b[6],
88 __a[7], __b[7], __a[16], __b[16], __a[17], __b[17], __a[18],
89 __b[18], __a[19], __b[19], __a[20], __b[20], __a[21], __b[21],
90 __a[22], __b[22], __a[23], __b[23]};
91 else if constexpr (sizeof(_Tp) == 64 && _Trait::_S_full_size == 8)
92 return _Tp{__a[0], __b[0], __a[2], __b[2],
93 __a[4], __b[4], __a[6], __b[6]};
94 else if constexpr (sizeof(_Tp) == 64 && _Trait::_S_full_size == 16)
95 return _Tp{__a[0], __b[0], __a[1], __b[1], __a[4], __b[4],
96 __a[5], __b[5], __a[8], __b[8], __a[9], __b[9],
97 __a[12], __b[12], __a[13], __b[13]};
98 else if constexpr (sizeof(_Tp) == 64 && _Trait::_S_full_size == 32)
99 return _Tp{__a[0], __b[0], __a[1], __b[1], __a[2], __b[2], __a[3],
100 __b[3], __a[8], __b[8], __a[9], __b[9], __a[10], __b[10],
101 __a[11], __b[11], __a[16], __b[16], __a[17], __b[17], __a[18],
102 __b[18], __a[19], __b[19], __a[24], __b[24], __a[25], __b[25],
103 __a[26], __b[26], __a[27], __b[27]};
104 else if constexpr (sizeof(_Tp) == 64 && _Trait::_S_full_size == 64)
105 return _Tp{__a[0], __b[0], __a[1], __b[1], __a[2], __b[2], __a[3],
106 __b[3], __a[4], __b[4], __a[5], __b[5], __a[6], __b[6],
107 __a[7], __b[7], __a[16], __b[16], __a[17], __b[17], __a[18],
108 __b[18], __a[19], __b[19], __a[20], __b[20], __a[21], __b[21],
109 __a[22], __b[22], __a[23], __b[23], __a[32], __b[32], __a[33],
110 __b[33], __a[34], __b[34], __a[35], __b[35], __a[36], __b[36],
111 __a[37], __b[37], __a[38], __b[38], __a[39], __b[39], __a[48],
112 __b[48], __a[49], __b[49], __a[50], __b[50], __a[51], __b[51],
113 __a[52], __b[52], __a[53], __b[53], __a[54], __b[54], __a[55],
114 __b[55]};
115 else
116 __assert_unreachable<_Tp>();
117 }
118
119// }}}
120// __is_zero{{{
121template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
122 _GLIBCXX_SIMD_INTRINSIC constexpr bool
123 __is_zero(_Tp __a)
124 {
125 if (!__builtin_is_constant_evaluated())
126 {
127 if constexpr (__have_avx)
128 {
129 if constexpr (_TVT::template _S_is<float, 8>)
130 return _mm256_testz_ps(__a, __a);
131 else if constexpr (_TVT::template _S_is<double, 4>)
132 return _mm256_testz_pd(__a, __a);
133 else if constexpr (sizeof(_Tp) == 32)
134 return _mm256_testz_si256(__to_intrin(__a), __to_intrin(__a));
135 else if constexpr (_TVT::template _S_is<float>)
136 return _mm_testz_ps(__to_intrin(__a), __to_intrin(__a));
137 else if constexpr (_TVT::template _S_is<double, 2>)
138 return _mm_testz_pd(__a, __a);
139 else
140 return _mm_testz_si128(__to_intrin(__a), __to_intrin(__a));
141 }
142 else if constexpr (__have_sse4_1)
143 return _mm_testz_si128(__intrin_bitcast<__m128i>(__a),
144 __intrin_bitcast<__m128i>(__a));
145 }
146 else if constexpr (sizeof(_Tp) <= 8)
147 return reinterpret_cast<__int_for_sizeof_t<_Tp>>(__a) == 0;
148 else
149 {
150 const auto __b = __vector_bitcast<_LLong>(__a);
151 if constexpr (sizeof(__b) == 16)
152 return (__b[0] | __b[1]) == 0;
153 else if constexpr (sizeof(__b) == 32)
154 return __is_zero(__lo128(__b) | __hi128(__b));
155 else if constexpr (sizeof(__b) == 64)
156 return __is_zero(__lo256(__b) | __hi256(__b));
157 else
158 __assert_unreachable<_Tp>();
159 }
160 }
161
162// }}}
163// __movemask{{{
164template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
165 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_CONST int
166 __movemask(_Tp __a)
167 {
168 if constexpr (sizeof(_Tp) == 32)
169 {
170 if constexpr (_TVT::template _S_is<float>)
171 return _mm256_movemask_ps(__to_intrin(__a));
172 else if constexpr (_TVT::template _S_is<double>)
173 return _mm256_movemask_pd(__to_intrin(__a));
174 else
175 return _mm256_movemask_epi8(__to_intrin(__a));
176 }
177 else if constexpr (_TVT::template _S_is<float>)
178 return _mm_movemask_ps(__to_intrin(__a));
179 else if constexpr (_TVT::template _S_is<double>)
180 return _mm_movemask_pd(__to_intrin(__a));
181 else
182 return _mm_movemask_epi8(__to_intrin(__a));
183 }
184
185// }}}
186// __testz{{{
187template <typename _TI, typename _TVT = _VectorTraits<_TI>>
188 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_CONST constexpr int
189 __testz(_TI __a, _TI __b)
190 {
191 static_assert(is_same_v<_TI, __intrinsic_type_t<typename _TVT::value_type,
192 _TVT::_S_full_size>>);
193 if (!__builtin_is_constant_evaluated())
194 {
195 if constexpr (sizeof(_TI) == 32)
196 {
197 if constexpr (_TVT::template _S_is<float>)
198 return _mm256_testz_ps(__to_intrin(__a), __to_intrin(__b));
199 else if constexpr (_TVT::template _S_is<double>)
200 return _mm256_testz_pd(__to_intrin(__a), __to_intrin(__b));
201 else
202 return _mm256_testz_si256(__to_intrin(__a), __to_intrin(__b));
203 }
204 else if constexpr (_TVT::template _S_is<float> && __have_avx)
205 return _mm_testz_ps(__to_intrin(__a), __to_intrin(__b));
206 else if constexpr (_TVT::template _S_is<double> && __have_avx)
207 return _mm_testz_pd(__to_intrin(__a), __to_intrin(__b));
208 else if constexpr (__have_sse4_1)
209 return _mm_testz_si128(__intrin_bitcast<__m128i>(__to_intrin(__a)),
210 __intrin_bitcast<__m128i>(__to_intrin(__b)));
211 else
212 return __movemask(0 == __and(__a, __b)) != 0;
213 }
214 else
215 return __is_zero(__and(__a, __b));
216 }
217
218// }}}
219// __testc{{{
220// requires SSE4.1 or above
221template <typename _TI, typename _TVT = _VectorTraits<_TI>>
222 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_CONST constexpr int
223 __testc(_TI __a, _TI __b)
224 {
225 static_assert(is_same_v<_TI, __intrinsic_type_t<typename _TVT::value_type,
226 _TVT::_S_full_size>>);
227 if (__builtin_is_constant_evaluated())
228 return __is_zero(__andnot(__a, __b));
229
230 if constexpr (sizeof(_TI) == 32)
231 {
232 if constexpr (_TVT::template _S_is<float>)
233 return _mm256_testc_ps(__a, __b);
234 else if constexpr (_TVT::template _S_is<double>)
235 return _mm256_testc_pd(__a, __b);
236 else
237 return _mm256_testc_si256(__to_intrin(__a), __to_intrin(__b));
238 }
239 else if constexpr (_TVT::template _S_is<float> && __have_avx)
240 return _mm_testc_ps(__to_intrin(__a), __to_intrin(__b));
241 else if constexpr (_TVT::template _S_is<double> && __have_avx)
242 return _mm_testc_pd(__to_intrin(__a), __to_intrin(__b));
243 else
244 {
245 static_assert(is_same_v<_TI, _TI> && __have_sse4_1);
246 return _mm_testc_si128(__intrin_bitcast<__m128i>(__to_intrin(__a)),
247 __intrin_bitcast<__m128i>(__to_intrin(__b)));
248 }
249 }
250
251// }}}
252// __testnzc{{{
253template <typename _TI, typename _TVT = _VectorTraits<_TI>>
254 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_CONST constexpr int
255 __testnzc(_TI __a, _TI __b)
256 {
257 static_assert(is_same_v<_TI, __intrinsic_type_t<typename _TVT::value_type,
258 _TVT::_S_full_size>>);
259 if (!__builtin_is_constant_evaluated())
260 {
261 if constexpr (sizeof(_TI) == 32)
262 {
263 if constexpr (_TVT::template _S_is<float>)
264 return _mm256_testnzc_ps(__a, __b);
265 else if constexpr (_TVT::template _S_is<double>)
266 return _mm256_testnzc_pd(__a, __b);
267 else
268 return _mm256_testnzc_si256(__to_intrin(__a), __to_intrin(__b));
269 }
270 else if constexpr (_TVT::template _S_is<float> && __have_avx)
271 return _mm_testnzc_ps(__to_intrin(__a), __to_intrin(__b));
272 else if constexpr (_TVT::template _S_is<double> && __have_avx)
273 return _mm_testnzc_pd(__to_intrin(__a), __to_intrin(__b));
274 else if constexpr (__have_sse4_1)
275 return _mm_testnzc_si128(__intrin_bitcast<__m128i>(__to_intrin(__a)),
276 __intrin_bitcast<__m128i>(__to_intrin(__b)));
277 else
278 return __movemask(0 == __and(__a, __b)) == 0
279 && __movemask(0 == __andnot(__a, __b)) == 0;
280 }
281 else
282 return !(__is_zero(__and(__a, __b)) || __is_zero(__andnot(__a, __b)));
283 }
284
285// }}}
286// __xzyw{{{
287// shuffles the complete vector, swapping the inner two quarters. Often useful
288// for AVX for fixing up a shuffle result.
289template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
290 _GLIBCXX_SIMD_INTRINSIC _Tp
291 __xzyw(_Tp __a)
292 {
293 if constexpr (sizeof(_Tp) == 16)
294 {
295 const auto __x = __vector_bitcast<conditional_t<
296 is_floating_point_v<typename _TVT::value_type>, float, int>>(__a);
297 return reinterpret_cast<_Tp>(
298 decltype(__x){__x[0], __x[2], __x[1], __x[3]});
299 }
300 else if constexpr (sizeof(_Tp) == 32)
301 {
302 const auto __x = __vector_bitcast<conditional_t<
303 is_floating_point_v<typename _TVT::value_type>, double, _LLong>>(__a);
304 return reinterpret_cast<_Tp>(
305 decltype(__x){__x[0], __x[2], __x[1], __x[3]});
306 }
307 else if constexpr (sizeof(_Tp) == 64)
308 {
309 const auto __x = __vector_bitcast<conditional_t<
310 is_floating_point_v<typename _TVT::value_type>, double, _LLong>>(__a);
311 return reinterpret_cast<_Tp>(decltype(__x){__x[0], __x[1], __x[4],
312 __x[5], __x[2], __x[3],
313 __x[6], __x[7]});
314 }
315 else
316 __assert_unreachable<_Tp>();
317 }
318
319// }}}
320// __maskload_epi32{{{
321template <typename _Tp>
322 _GLIBCXX_SIMD_INTRINSIC auto
323 __maskload_epi32(const int* __ptr, _Tp __k)
324 {
325 if constexpr (sizeof(__k) == 16)
326 return _mm_maskload_epi32(__ptr, __k);
327 else
328 return _mm256_maskload_epi32(__ptr, __k);
329 }
330
331// }}}
332// __maskload_epi64{{{
333template <typename _Tp>
334 _GLIBCXX_SIMD_INTRINSIC auto
335 __maskload_epi64(const _LLong* __ptr, _Tp __k)
336 {
337 if constexpr (sizeof(__k) == 16)
338 return _mm_maskload_epi64(__ptr, __k);
339 else
340 return _mm256_maskload_epi64(__ptr, __k);
341 }
342
343// }}}
344// __maskload_ps{{{
345template <typename _Tp>
346 _GLIBCXX_SIMD_INTRINSIC auto
347 __maskload_ps(const float* __ptr, _Tp __k)
348 {
349 if constexpr (sizeof(__k) == 16)
350 return _mm_maskload_ps(__ptr, __k);
351 else
352 return _mm256_maskload_ps(__ptr, __k);
353 }
354
355// }}}
356// __maskload_pd{{{
357template <typename _Tp>
358 _GLIBCXX_SIMD_INTRINSIC auto
359 __maskload_pd(const double* __ptr, _Tp __k)
360 {
361 if constexpr (sizeof(__k) == 16)
362 return _mm_maskload_pd(__ptr, __k);
363 else
364 return _mm256_maskload_pd(__ptr, __k);
365 }
366
367// }}}
368
369#ifdef _GLIBCXX_SIMD_WORKAROUND_PR85048
370#include "simd_x86_conversions.h"
371#endif
372
373// ISA & type detection {{{
374template <typename _Tp, size_t _Np>
375 constexpr bool
376 __is_sse_ps()
377 {
378 return __have_sse
379 && is_same_v<_Tp,
380 float> && sizeof(__intrinsic_type_t<_Tp, _Np>) == 16;
381 }
382
383template <typename _Tp, size_t _Np>
384 constexpr bool
385 __is_sse_pd()
386 {
387 return __have_sse2
388 && is_same_v<_Tp,
389 double> && sizeof(__intrinsic_type_t<_Tp, _Np>) == 16;
390 }
391
392template <typename _Tp, size_t _Np>
393 constexpr bool
394 __is_avx_ps()
395 {
396 return __have_avx
397 && is_same_v<_Tp,
398 float> && sizeof(__intrinsic_type_t<_Tp, _Np>) == 32;
399 }
400
401template <typename _Tp, size_t _Np>
402 constexpr bool
403 __is_avx_pd()
404 {
405 return __have_avx
406 && is_same_v<_Tp,
407 double> && sizeof(__intrinsic_type_t<_Tp, _Np>) == 32;
408 }
409
410template <typename _Tp, size_t _Np>
411 constexpr bool
412 __is_avx512_ps()
413 {
414 return __have_avx512f
415 && is_same_v<_Tp,
416 float> && sizeof(__intrinsic_type_t<_Tp, _Np>) == 64;
417 }
418
419template <typename _Tp, size_t _Np>
420 constexpr bool
421 __is_avx512_pd()
422 {
423 return __have_avx512f
424 && is_same_v<_Tp,
425 double> && sizeof(__intrinsic_type_t<_Tp, _Np>) == 64;
426 }
427
428// }}}
429struct _MaskImplX86Mixin;
430
431// _CommonImplX86 {{{
432struct _CommonImplX86 : _CommonImplBuiltin
433{
434#ifdef _GLIBCXX_SIMD_WORKAROUND_PR85048
435 // _S_converts_via_decomposition {{{
436 template <typename _From, typename _To, size_t _ToSize>
437 static constexpr bool _S_converts_via_decomposition()
438 {
439 if constexpr (is_integral_v<
440 _From> && is_integral_v<_To> && sizeof(_From) == 8
441 && _ToSize == 16)
442 return (sizeof(_To) == 2 && !__have_ssse3)
443 || (sizeof(_To) == 1 && !__have_avx512f);
444 else if constexpr (is_floating_point_v<_From> && is_integral_v<_To>)
445 return ((sizeof(_From) == 4 || sizeof(_From) == 8) && sizeof(_To) == 8
446 && !__have_avx512dq)
447 || (sizeof(_From) == 8 && sizeof(_To) == 4 && !__have_sse4_1
448 && _ToSize == 16);
449 else if constexpr (
450 is_integral_v<_From> && is_floating_point_v<_To> && sizeof(_From) == 8
451 && !__have_avx512dq)
452 return (sizeof(_To) == 4 && _ToSize == 16)
453 || (sizeof(_To) == 8 && _ToSize < 64);
454 else
455 return false;
456 }
457
458 template <typename _From, typename _To, size_t _ToSize>
459 static inline constexpr bool __converts_via_decomposition_v
460 = _S_converts_via_decomposition<_From, _To, _ToSize>();
461
462 // }}}
463#endif
464 // _S_store {{{
465 using _CommonImplBuiltin::_S_store;
466
467 template <typename _Tp, size_t _Np>
468 _GLIBCXX_SIMD_INTRINSIC static void _S_store(_SimdWrapper<_Tp, _Np> __x,
469 void* __addr)
470 {
471 constexpr size_t _Bytes = _Np * sizeof(_Tp);
472
473 if constexpr ((_Bytes & (_Bytes - 1)) != 0 && __have_avx512bw_vl)
474 {
475 const auto __v = __to_intrin(__x);
476
477 if constexpr (_Bytes & 1)
478 {
479 if constexpr (_Bytes < 16)
480 _mm_mask_storeu_epi8(__addr, 0xffffu >> (16 - _Bytes),
481 __intrin_bitcast<__m128i>(__v));
482 else if constexpr (_Bytes < 32)
483 _mm256_mask_storeu_epi8(__addr, 0xffffffffu >> (32 - _Bytes),
484 __intrin_bitcast<__m256i>(__v));
485 else
486 _mm512_mask_storeu_epi8(__addr,
487 0xffffffffffffffffull >> (64 - _Bytes),
488 __intrin_bitcast<__m512i>(__v));
489 }
490 else if constexpr (_Bytes & 2)
491 {
492 if constexpr (_Bytes < 16)
493 _mm_mask_storeu_epi16(__addr, 0xffu >> (8 - _Bytes / 2),
494 __intrin_bitcast<__m128i>(__v));
495 else if constexpr (_Bytes < 32)
496 _mm256_mask_storeu_epi16(__addr, 0xffffu >> (16 - _Bytes / 2),
497 __intrin_bitcast<__m256i>(__v));
498 else
499 _mm512_mask_storeu_epi16(__addr,
500 0xffffffffull >> (32 - _Bytes / 2),
501 __intrin_bitcast<__m512i>(__v));
502 }
503 else if constexpr (_Bytes & 4)
504 {
505 if constexpr (_Bytes < 16)
506 _mm_mask_storeu_epi32(__addr, 0xfu >> (4 - _Bytes / 4),
507 __intrin_bitcast<__m128i>(__v));
508 else if constexpr (_Bytes < 32)
509 _mm256_mask_storeu_epi32(__addr, 0xffu >> (8 - _Bytes / 4),
510 __intrin_bitcast<__m256i>(__v));
511 else
512 _mm512_mask_storeu_epi32(__addr, 0xffffull >> (16 - _Bytes / 4),
513 __intrin_bitcast<__m512i>(__v));
514 }
515 else
516 {
517 static_assert(
518 _Bytes > 16,
519 "_Bytes < 16 && (_Bytes & 7) == 0 && (_Bytes & (_Bytes "
520 "- 1)) != 0 is impossible");
521 if constexpr (_Bytes < 32)
522 _mm256_mask_storeu_epi64(__addr, 0xfu >> (4 - _Bytes / 8),
523 __intrin_bitcast<__m256i>(__v));
524 else
525 _mm512_mask_storeu_epi64(__addr, 0xffull >> (8 - _Bytes / 8),
526 __intrin_bitcast<__m512i>(__v));
527 }
528 }
529 else
530 _CommonImplBuiltin::_S_store(__x, __addr);
531 }
532
533 // }}}
534 // _S_store_bool_array(_BitMask) {{{
535 template <size_t _Np, bool _Sanitized>
536 _GLIBCXX_SIMD_INTRINSIC static constexpr void
537 _S_store_bool_array(const _BitMask<_Np, _Sanitized> __x, bool* __mem)
538 {
539 if constexpr (__have_avx512bw_vl) // don't care for BW w/o VL
540 _S_store<_Np>(1 & __vector_bitcast<_UChar, _Np>([=]() constexpr {
541 if constexpr (_Np <= 16)
542 return _mm_movm_epi8(__x._M_to_bits());
543 else if constexpr (_Np <= 32)
544 return _mm256_movm_epi8(__x._M_to_bits());
545 else if constexpr (_Np <= 64)
546 return _mm512_movm_epi8(__x._M_to_bits());
547 else
548 __assert_unreachable<_SizeConstant<_Np>>();
549 }()),
550 __mem);
551 else if constexpr (__have_bmi2)
552 {
553 if constexpr (_Np <= 4)
554 _S_store<_Np>(_pdep_u32(__x._M_to_bits(), 0x01010101U), __mem);
555 else
556 __execute_n_times<__div_roundup(_Np, sizeof(size_t))>(
557 [&](auto __i) {
558 constexpr size_t __offset = __i * sizeof(size_t);
559 constexpr int __todo = std::min(sizeof(size_t), _Np - __offset);
560 if constexpr (__todo == 1)
561 __mem[__offset] = __x[__offset];
562 else
563 {
564 const auto __bools =
565#ifdef __x86_64__
566 _pdep_u64(__x.template _M_extract<__offset>().to_ullong(),
567 0x0101010101010101ULL);
568#else // __x86_64__
569 _pdep_u32(
570 __x.template _M_extract<__offset>()._M_to_bits(),
571 0x01010101U);
572#endif // __x86_64__
573 _S_store<__todo>(__bools, __mem + __offset);
574 }
575 });
576 }
577 else if constexpr (__have_sse2 && _Np > 7)
578 __execute_n_times<__div_roundup(_Np, 16)>([&](auto __i) {
579 constexpr int __offset = __i * 16;
580 constexpr int __todo = std::min(16, int(_Np) - __offset);
581 const int __bits = __x.template _M_extract<__offset>()._M_to_bits();
582 __vector_type16_t<_UChar> __bools;
583 if constexpr (__have_avx512f)
584 {
585 auto __as32bits
586 = _mm512_maskz_mov_epi32(__bits, __to_intrin(
587 __vector_broadcast<16>(1)));
588 auto __as16bits
589 = __xzyw(_mm256_packs_epi32(__lo256(__as32bits),
590 __todo > 8 ? __hi256(__as32bits)
591 : __m256i()));
592 __bools = __vector_bitcast<_UChar>(
593 _mm_packs_epi16(__lo128(__as16bits), __hi128(__as16bits)));
594 }
595 else
596 {
597 using _V = __vector_type_t<_UChar, 16>;
598 auto __tmp = _mm_cvtsi32_si128(__bits);
599 __tmp = _mm_unpacklo_epi8(__tmp, __tmp);
600 __tmp = _mm_unpacklo_epi16(__tmp, __tmp);
601 __tmp = _mm_unpacklo_epi32(__tmp, __tmp);
602 _V __tmp2 = reinterpret_cast<_V>(__tmp);
603 __tmp2 &= _V{1, 2, 4, 8, 16, 32, 64, 128,
604 1, 2, 4, 8, 16, 32, 64, 128}; // mask bit index
605 __bools = (__tmp2 == 0) + 1; // 0xff -> 0x00 | 0x00 -> 0x01
606 }
607 _S_store<__todo>(__bools, __mem + __offset);
608 });
609 else
610 _CommonImplBuiltin::_S_store_bool_array(__x, __mem);
611 }
612
613 // }}}
614 // _S_blend_avx512 {{{
615 // Returns: __k ? __b : __a
616 // TODO: reverse __a and __b to match COND_EXPR
617 // Requires: _TV to be a __vector_type_t matching valuetype for the bitmask
618 // __k
619 template <typename _Kp, typename _TV>
620 _GLIBCXX_SIMD_INTRINSIC static _TV
621 _S_blend_avx512(const _Kp __k, const _TV __a, const _TV __b) noexcept
622 {
623#ifdef __clang__
624 // FIXME: this does a boolean choice, not a blend
625 return __k ? __a : __b;
626#else
627 static_assert(__is_vector_type_v<_TV>);
628 using _Tp = typename _VectorTraits<_TV>::value_type;
629 static_assert(sizeof(_TV) >= 16);
630 static_assert(sizeof(_Tp) <= 8);
631 using _IntT
632 = conditional_t<(sizeof(_Tp) > 2),
633 conditional_t<sizeof(_Tp) == 4, int, long long>,
634 conditional_t<sizeof(_Tp) == 1, char, short>>;
635 [[maybe_unused]] const auto __aa = __vector_bitcast<_IntT>(__a);
636 [[maybe_unused]] const auto __bb = __vector_bitcast<_IntT>(__b);
637 if constexpr (sizeof(_TV) == 64)
638 {
639 if constexpr (sizeof(_Tp) == 1)
640 return reinterpret_cast<_TV>(
641 __builtin_ia32_blendmb_512_mask(__aa, __bb, __k));
642 else if constexpr (sizeof(_Tp) == 2)
643 return reinterpret_cast<_TV>(
644 __builtin_ia32_blendmw_512_mask(__aa, __bb, __k));
645 else if constexpr (sizeof(_Tp) == 4 && is_floating_point_v<_Tp>)
646 return __builtin_ia32_blendmps_512_mask(__a, __b, __k);
647 else if constexpr (sizeof(_Tp) == 4)
648 return reinterpret_cast<_TV>(
649 __builtin_ia32_blendmd_512_mask(__aa, __bb, __k));
650 else if constexpr (sizeof(_Tp) == 8 && is_floating_point_v<_Tp>)
651 return __builtin_ia32_blendmpd_512_mask(__a, __b, __k);
652 else if constexpr (sizeof(_Tp) == 8)
653 return reinterpret_cast<_TV>(
654 __builtin_ia32_blendmq_512_mask(__aa, __bb, __k));
655 }
656 else if constexpr (sizeof(_TV) == 32)
657 {
658 if constexpr (sizeof(_Tp) == 1)
659 return reinterpret_cast<_TV>(
660 __builtin_ia32_blendmb_256_mask(__aa, __bb, __k));
661 else if constexpr (sizeof(_Tp) == 2)
662 return reinterpret_cast<_TV>(
663 __builtin_ia32_blendmw_256_mask(__aa, __bb, __k));
664 else if constexpr (sizeof(_Tp) == 4 && is_floating_point_v<_Tp>)
665 return __builtin_ia32_blendmps_256_mask(__a, __b, __k);
666 else if constexpr (sizeof(_Tp) == 4)
667 return reinterpret_cast<_TV>(
668 __builtin_ia32_blendmd_256_mask(__aa, __bb, __k));
669 else if constexpr (sizeof(_Tp) == 8 && is_floating_point_v<_Tp>)
670 return __builtin_ia32_blendmpd_256_mask(__a, __b, __k);
671 else if constexpr (sizeof(_Tp) == 8)
672 return reinterpret_cast<_TV>(
673 __builtin_ia32_blendmq_256_mask(__aa, __bb, __k));
674 }
675 else if constexpr (sizeof(_TV) == 16)
676 {
677 if constexpr (sizeof(_Tp) == 1)
678 return reinterpret_cast<_TV>(
679 __builtin_ia32_blendmb_128_mask(__aa, __bb, __k));
680 else if constexpr (sizeof(_Tp) == 2)
681 return reinterpret_cast<_TV>(
682 __builtin_ia32_blendmw_128_mask(__aa, __bb, __k));
683 else if constexpr (sizeof(_Tp) == 4 && is_floating_point_v<_Tp>)
684 return __builtin_ia32_blendmps_128_mask(__a, __b, __k);
685 else if constexpr (sizeof(_Tp) == 4)
686 return reinterpret_cast<_TV>(
687 __builtin_ia32_blendmd_128_mask(__aa, __bb, __k));
688 else if constexpr (sizeof(_Tp) == 8 && is_floating_point_v<_Tp>)
689 return __builtin_ia32_blendmpd_128_mask(__a, __b, __k);
690 else if constexpr (sizeof(_Tp) == 8)
691 return reinterpret_cast<_TV>(
692 __builtin_ia32_blendmq_128_mask(__aa, __bb, __k));
693 }
694#endif
695 }
696
697 // }}}
698 // _S_blend_intrin {{{
699 // Returns: __k ? __b : __a
700 // TODO: reverse __a and __b to match COND_EXPR
701 // Requires: _Tp to be an intrinsic type (integers blend per byte) and 16/32
702 // Bytes wide
703 template <typename _Tp>
704 _GLIBCXX_SIMD_INTRINSIC static _Tp _S_blend_intrin(_Tp __k, _Tp __a,
705 _Tp __b) noexcept
706 {
707 static_assert(is_same_v<decltype(__to_intrin(__a)), _Tp>);
708 constexpr struct
709 {
710 _GLIBCXX_SIMD_INTRINSIC __m128 operator()(__m128 __a, __m128 __b,
711 __m128 __k) const noexcept
712 {
713 return __builtin_ia32_blendvps(__a, __b, __k);
714 }
715 _GLIBCXX_SIMD_INTRINSIC __m128d operator()(__m128d __a, __m128d __b,
716 __m128d __k) const noexcept
717 {
718 return __builtin_ia32_blendvpd(__a, __b, __k);
719 }
720 _GLIBCXX_SIMD_INTRINSIC __m128i operator()(__m128i __a, __m128i __b,
721 __m128i __k) const noexcept
722 {
723 return reinterpret_cast<__m128i>(
724 __builtin_ia32_pblendvb128(reinterpret_cast<__v16qi>(__a),
725 reinterpret_cast<__v16qi>(__b),
726 reinterpret_cast<__v16qi>(__k)));
727 }
728 _GLIBCXX_SIMD_INTRINSIC __m256 operator()(__m256 __a, __m256 __b,
729 __m256 __k) const noexcept
730 {
731 return __builtin_ia32_blendvps256(__a, __b, __k);
732 }
733 _GLIBCXX_SIMD_INTRINSIC __m256d operator()(__m256d __a, __m256d __b,
734 __m256d __k) const noexcept
735 {
736 return __builtin_ia32_blendvpd256(__a, __b, __k);
737 }
738 _GLIBCXX_SIMD_INTRINSIC __m256i operator()(__m256i __a, __m256i __b,
739 __m256i __k) const noexcept
740 {
741 if constexpr (__have_avx2)
742 return reinterpret_cast<__m256i>(
743 __builtin_ia32_pblendvb256(reinterpret_cast<__v32qi>(__a),
744 reinterpret_cast<__v32qi>(__b),
745 reinterpret_cast<__v32qi>(__k)));
746 else
747 return reinterpret_cast<__m256i>(
748 __builtin_ia32_blendvps256(reinterpret_cast<__v8sf>(__a),
749 reinterpret_cast<__v8sf>(__b),
750 reinterpret_cast<__v8sf>(__k)));
751 }
752 } __eval;
753 return __eval(__a, __b, __k);
754 }
755
756 // }}}
757 // _S_blend {{{
758 // Returns: __k ? __at1 : __at0
759 // TODO: reverse __at0 and __at1 to match COND_EXPR
760 template <typename _Tp, size_t _Np>
761 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
762 _S_blend(_SimdWrapper<bool, _Np> __k, _SimdWrapper<_Tp, _Np> __at0,
763 _SimdWrapper<_Tp, _Np> __at1)
764 {
765 static_assert(is_same_v<_Tp, _Tp> && __have_avx512f);
766 if (__k._M_is_constprop() && __at0._M_is_constprop()
767 && __at1._M_is_constprop())
768 return __generate_from_n_evaluations<_Np,
769 __vector_type_t<_Tp, _Np>>([&](
770 auto __i) constexpr { return __k[__i] ? __at1[__i] : __at0[__i]; });
771 else if constexpr (sizeof(__at0) == 64
772 || (__have_avx512vl && sizeof(__at0) >= 16))
773 return _S_blend_avx512(__k._M_data, __at0._M_data, __at1._M_data);
774 else
775 {
776 static_assert((__have_avx512vl && sizeof(__at0) < 16)
777 || !__have_avx512vl);
778 constexpr size_t __size = (__have_avx512vl ? 16 : 64) / sizeof(_Tp);
779 return __vector_bitcast<_Tp, _Np>(
780 _S_blend_avx512(__k._M_data, __vector_bitcast<_Tp, __size>(__at0),
781 __vector_bitcast<_Tp, __size>(__at1)));
782 }
783 }
784
785 template <typename _Tp, size_t _Np>
786 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
787 _S_blend(_SimdWrapper<__int_for_sizeof_t<_Tp>, _Np> __k,
788 _SimdWrapper<_Tp, _Np> __at0, _SimdWrapper<_Tp, _Np> __at1)
789 {
790 const auto __kk = __wrapper_bitcast<_Tp>(__k);
791 if (__builtin_is_constant_evaluated()
792 || (__kk._M_is_constprop() && __at0._M_is_constprop()
793 && __at1._M_is_constprop()))
794 {
795 auto __r = __or(__andnot(__kk, __at0), __and(__kk, __at1));
796 if (__r._M_is_constprop())
797 return __r;
798 }
799 if constexpr (((__have_avx512f && sizeof(__at0) == 64) || __have_avx512vl)
800 && (sizeof(_Tp) >= 4 || __have_avx512bw))
801 // convert to bitmask and call overload above
802 return _S_blend(
803 _SimdWrapper<bool, _Np>(
804 __make_dependent_t<_Tp, _MaskImplX86Mixin>::_S_to_bits(__k)
805 ._M_to_bits()),
806 __at0, __at1);
807 else
808 {
809 // Since GCC does not assume __k to be a mask, using the builtin
810 // conditional operator introduces an extra compare against 0 before
811 // blending. So we rather call the intrinsic here.
812 if constexpr (__have_sse4_1)
813 return _S_blend_intrin(__to_intrin(__kk), __to_intrin(__at0),
814 __to_intrin(__at1));
815 else
816 return __or(__andnot(__kk, __at0), __and(__kk, __at1));
817 }
818 }
819
820 // }}}
821};
822
823// }}}
824// _SimdImplX86 {{{
825template <typename _Abi>
826 struct _SimdImplX86 : _SimdImplBuiltin<_Abi>
827 {
828 using _Base = _SimdImplBuiltin<_Abi>;
829
830 template <typename _Tp>
831 using _MaskMember = typename _Base::template _MaskMember<_Tp>;
832
833 template <typename _Tp>
834 static constexpr size_t _S_full_size = _Abi::template _S_full_size<_Tp>;
835
836 template <typename _Tp>
837 static constexpr size_t _S_size = _Abi::template _S_size<_Tp>;
838
839 template <typename _Tp>
840 static constexpr size_t _S_max_store_size
841 = (sizeof(_Tp) >= 4 && __have_avx512f) || __have_avx512bw ? 64
842 : (is_floating_point_v<_Tp>&& __have_avx) || __have_avx2 ? 32
843 : 16;
844 using _MaskImpl = typename _Abi::_MaskImpl;
845
846 // _S_masked_load {{{
847 template <typename _Tp, size_t _Np, typename _Up>
848 static inline _SimdWrapper<_Tp, _Np>
849 _S_masked_load(_SimdWrapper<_Tp, _Np> __merge, _MaskMember<_Tp> __k,
850 const _Up* __mem) noexcept
851 {
852 static_assert(_Np == _S_size<_Tp>);
853 if constexpr (is_same_v<_Tp, _Up> || // no conversion
854 (sizeof(_Tp) == sizeof(_Up)
855 && is_integral_v<
856 _Tp> == is_integral_v<_Up>) // conversion via bit
857 // reinterpretation
858 )
859 {
860 [[maybe_unused]] const auto __intrin = __to_intrin(__merge);
861 if constexpr ((__is_avx512_abi<_Abi>() || __have_avx512bw_vl)
862 && sizeof(_Tp) == 1)
863 {
864 const auto __kk = _MaskImpl::_S_to_bits(__k)._M_to_bits();
865 if constexpr (sizeof(__intrin) == 16)
866 __merge = __vector_bitcast<_Tp, _Np>(
867 _mm_mask_loadu_epi8(__intrin, __kk, __mem));
868 else if constexpr (sizeof(__merge) == 32)
869 __merge = __vector_bitcast<_Tp, _Np>(
870 _mm256_mask_loadu_epi8(__intrin, __kk, __mem));
871 else if constexpr (sizeof(__merge) == 64)
872 __merge = __vector_bitcast<_Tp, _Np>(
873 _mm512_mask_loadu_epi8(__intrin, __kk, __mem));
874 else
875 __assert_unreachable<_Tp>();
876 }
877 else if constexpr ((__is_avx512_abi<_Abi>() || __have_avx512bw_vl)
878 && sizeof(_Tp) == 2)
879 {
880 const auto __kk = _MaskImpl::_S_to_bits(__k)._M_to_bits();
881 if constexpr (sizeof(__intrin) == 16)
882 __merge = __vector_bitcast<_Tp, _Np>(
883 _mm_mask_loadu_epi16(__intrin, __kk, __mem));
884 else if constexpr (sizeof(__intrin) == 32)
885 __merge = __vector_bitcast<_Tp, _Np>(
886 _mm256_mask_loadu_epi16(__intrin, __kk, __mem));
887 else if constexpr (sizeof(__intrin) == 64)
888 __merge = __vector_bitcast<_Tp, _Np>(
889 _mm512_mask_loadu_epi16(__intrin, __kk, __mem));
890 else
891 __assert_unreachable<_Tp>();
892 }
893 else if constexpr ((__is_avx512_abi<_Abi>() || __have_avx512vl)
894 && sizeof(_Tp) == 4 && is_integral_v<_Up>)
895 {
896 const auto __kk = _MaskImpl::_S_to_bits(__k)._M_to_bits();
897 if constexpr (sizeof(__intrin) == 16)
898 __merge = __vector_bitcast<_Tp, _Np>(
899 _mm_mask_loadu_epi32(__intrin, __kk, __mem));
900 else if constexpr (sizeof(__intrin) == 32)
901 __merge = __vector_bitcast<_Tp, _Np>(
902 _mm256_mask_loadu_epi32(__intrin, __kk, __mem));
903 else if constexpr (sizeof(__intrin) == 64)
904 __merge = __vector_bitcast<_Tp, _Np>(
905 _mm512_mask_loadu_epi32(__intrin, __kk, __mem));
906 else
907 __assert_unreachable<_Tp>();
908 }
909 else if constexpr ((__is_avx512_abi<_Abi>() || __have_avx512vl)
910 && sizeof(_Tp) == 4 && is_floating_point_v<_Up>)
911 {
912 const auto __kk = _MaskImpl::_S_to_bits(__k)._M_to_bits();
913 if constexpr (sizeof(__intrin) == 16)
914 __merge = __vector_bitcast<_Tp, _Np>(
915 _mm_mask_loadu_ps(__intrin, __kk, __mem));
916 else if constexpr (sizeof(__intrin) == 32)
917 __merge = __vector_bitcast<_Tp, _Np>(
918 _mm256_mask_loadu_ps(__intrin, __kk, __mem));
919 else if constexpr (sizeof(__intrin) == 64)
920 __merge = __vector_bitcast<_Tp, _Np>(
921 _mm512_mask_loadu_ps(__intrin, __kk, __mem));
922 else
923 __assert_unreachable<_Tp>();
924 }
925 else if constexpr (__have_avx2 && sizeof(_Tp) == 4
926 && is_integral_v<_Up>)
927 {
928 static_assert(sizeof(__intrin) == 16 || sizeof(__intrin) == 32);
929 __merge
930 = __or(__andnot(__vector_bitcast<_Tp>(__k), __merge._M_data),
931 __vector_bitcast<_Tp, _Np>(
932 __maskload_epi32(reinterpret_cast<const int*>(__mem),
933 __to_intrin(__k))));
934 }
935 else if constexpr (__have_avx && sizeof(_Tp) == 4)
936 {
937 static_assert(sizeof(__intrin) == 16 || sizeof(__intrin) == 32);
938 __merge
939 = __or(__andnot(__vector_bitcast<_Tp>(__k), __merge._M_data),
940 __vector_bitcast<_Tp, _Np>(
941 __maskload_ps(reinterpret_cast<const float*>(__mem),
942 __to_intrin(__k))));
943 }
944 else if constexpr ((__is_avx512_abi<_Abi>() || __have_avx512vl)
945 && sizeof(_Tp) == 8 && 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_epi64(__intrin, __kk, __mem));
951 else if constexpr (sizeof(__intrin) == 32)
952 __merge = __vector_bitcast<_Tp, _Np>(
953 _mm256_mask_loadu_epi64(__intrin, __kk, __mem));
954 else if constexpr (sizeof(__intrin) == 64)
955 __merge = __vector_bitcast<_Tp, _Np>(
956 _mm512_mask_loadu_epi64(__intrin, __kk, __mem));
957 else
958 __assert_unreachable<_Tp>();
959 }
960 else if constexpr ((__is_avx512_abi<_Abi>() || __have_avx512vl)
961 && sizeof(_Tp) == 8 && 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_pd(__intrin, __kk, __mem));
967 else if constexpr (sizeof(__intrin) == 32)
968 __merge = __vector_bitcast<_Tp, _Np>(
969 _mm256_mask_loadu_pd(__intrin, __kk, __mem));
970 else if constexpr (sizeof(__intrin) == 64)
971 __merge = __vector_bitcast<_Tp, _Np>(
972 _mm512_mask_loadu_pd(__intrin, __kk, __mem));
973 else
974 __assert_unreachable<_Tp>();
975 }
976 else if constexpr (__have_avx2 && sizeof(_Tp) == 8
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>(__maskload_epi64(
983 reinterpret_cast<const _LLong*>(__mem),
984 __to_intrin(__k))));
985 }
986 else if constexpr (__have_avx && sizeof(_Tp) == 8)
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_pd(reinterpret_cast<const double*>(__mem),
993 __to_intrin(__k))));
994 }
995 else
996 _BitOps::_S_bit_iteration(_MaskImpl::_S_to_bits(__k),
997 [&](auto __i) {
998 __merge._M_set(__i, static_cast<_Tp>(
999 __mem[__i]));
1000 });
1001 }
1002 /* Very uncertain, that the following improves anything. Needs
1003 benchmarking
1004 * before it's activated.
1005 else if constexpr (sizeof(_Up) <= 8 && // no long double
1006 !__converts_via_decomposition_v<
1007 _Up, _Tp,
1008 sizeof(__merge)> // conversion via decomposition
1009 // is better handled via the
1010 // bit_iteration fallback below
1011 )
1012 {
1013 // TODO: copy pattern from _S_masked_store, which doesn't resort to
1014 // fixed_size
1015 using _Ap = simd_abi::deduce_t<_Up, _Np>;
1016 using _ATraits = _SimdTraits<_Up, _Ap>;
1017 using _AImpl = typename _ATraits::_SimdImpl;
1018 typename _ATraits::_SimdMember __uncvted{};
1019 typename _ATraits::_MaskMember __kk = _Ap::_MaskImpl::template
1020 _S_convert<_Up>(__k);
1021 __uncvted = _AImpl::_S_masked_load(__uncvted, __kk, __mem);
1022 _SimdConverter<_Up, _Ap, _Tp, _Abi> __converter;
1023 _Base::_S_masked_assign(__k, __merge, __converter(__uncvted));
1024 }
1025 */
1026 else
1027 __merge = _Base::_S_masked_load(__merge, __k, __mem);
1028 return __merge;
1029 }
1030
1031 // }}}
1032 // _S_masked_store_nocvt {{{
1033 template <typename _Tp, size_t _Np>
1034 _GLIBCXX_SIMD_INTRINSIC static void
1035 _S_masked_store_nocvt(_SimdWrapper<_Tp, _Np> __v, _Tp* __mem,
1036 _SimdWrapper<bool, _Np> __k)
1037 {
1038 [[maybe_unused]] const auto __vi = __to_intrin(__v);
1039 if constexpr (sizeof(__vi) == 64)
1040 {
1041 static_assert(sizeof(__v) == 64 && __have_avx512f);
1042 if constexpr (__have_avx512bw && sizeof(_Tp) == 1)
1043 _mm512_mask_storeu_epi8(__mem, __k, __vi);
1044 else if constexpr (__have_avx512bw && sizeof(_Tp) == 2)
1045 _mm512_mask_storeu_epi16(__mem, __k, __vi);
1046 else if constexpr (__have_avx512f && sizeof(_Tp) == 4)
1047 {
1048 if constexpr (is_integral_v<_Tp>)
1049 _mm512_mask_storeu_epi32(__mem, __k, __vi);
1050 else
1051 _mm512_mask_storeu_ps(__mem, __k, __vi);
1052 }
1053 else if constexpr (__have_avx512f && sizeof(_Tp) == 8)
1054 {
1055 if constexpr (is_integral_v<_Tp>)
1056 _mm512_mask_storeu_epi64(__mem, __k, __vi);
1057 else
1058 _mm512_mask_storeu_pd(__mem, __k, __vi);
1059 }
1060#if 0 // with KNL either sizeof(_Tp) >= 4 or sizeof(_vi) <= 32
1061 // with Skylake-AVX512, __have_avx512bw is true
1062 else if constexpr (__have_sse2)
1063 {
1064 using _M = __vector_type_t<_Tp, _Np>;
1065 using _MVT = _VectorTraits<_M>;
1066 _mm_maskmoveu_si128(__auto_bitcast(__extract<0, 4>(__v._M_data)),
1067 __auto_bitcast(_MaskImpl::template _S_convert<_Tp, _Np>(__k._M_data)),
1068 reinterpret_cast<char*>(__mem));
1069 _mm_maskmoveu_si128(__auto_bitcast(__extract<1, 4>(__v._M_data)),
1070 __auto_bitcast(_MaskImpl::template _S_convert<_Tp, _Np>(
1071 __k._M_data >> 1 * _MVT::_S_full_size)),
1072 reinterpret_cast<char*>(__mem) + 1 * 16);
1073 _mm_maskmoveu_si128(__auto_bitcast(__extract<2, 4>(__v._M_data)),
1074 __auto_bitcast(_MaskImpl::template _S_convert<_Tp, _Np>(
1075 __k._M_data >> 2 * _MVT::_S_full_size)),
1076 reinterpret_cast<char*>(__mem) + 2 * 16);
1077 if constexpr (_Np > 48 / sizeof(_Tp))
1078 _mm_maskmoveu_si128(
1079 __auto_bitcast(__extract<3, 4>(__v._M_data)),
1080 __auto_bitcast(_MaskImpl::template _S_convert<_Tp, _Np>(
1081 __k._M_data >> 3 * _MVT::_S_full_size)),
1082 reinterpret_cast<char*>(__mem) + 3 * 16);
1083 }
1084#endif
1085 else
1086 __assert_unreachable<_Tp>();
1087 }
1088 else if constexpr (sizeof(__vi) == 32)
1089 {
1090 if constexpr (__have_avx512bw_vl && sizeof(_Tp) == 1)
1091 _mm256_mask_storeu_epi8(__mem, __k, __vi);
1092 else if constexpr (__have_avx512bw_vl && sizeof(_Tp) == 2)
1093 _mm256_mask_storeu_epi16(__mem, __k, __vi);
1094 else if constexpr (__have_avx512vl && sizeof(_Tp) == 4)
1095 {
1096 if constexpr (is_integral_v<_Tp>)
1097 _mm256_mask_storeu_epi32(__mem, __k, __vi);
1098 else
1099 _mm256_mask_storeu_ps(__mem, __k, __vi);
1100 }
1101 else if constexpr (__have_avx512vl && sizeof(_Tp) == 8)
1102 {
1103 if constexpr (is_integral_v<_Tp>)
1104 _mm256_mask_storeu_epi64(__mem, __k, __vi);
1105 else
1106 _mm256_mask_storeu_pd(__mem, __k, __vi);
1107 }
1108 else if constexpr (__have_avx512f
1109 && (sizeof(_Tp) >= 4 || __have_avx512bw))
1110 {
1111 // use a 512-bit maskstore, using zero-extension of the bitmask
1112 _S_masked_store_nocvt(
1113 _SimdWrapper64<_Tp>(
1114 __intrin_bitcast<__vector_type64_t<_Tp>>(__v._M_data)),
1115 __mem, _SimdWrapper<bool, 64 / sizeof(_Tp)>(__k._M_data));
1116 }
1117 else
1118 _S_masked_store_nocvt(__v, __mem,
1119 _MaskImpl::template _S_to_maskvector<
1120 __int_for_sizeof_t<_Tp>, _Np>(__k));
1121 }
1122 else if constexpr (sizeof(__vi) == 16)
1123 {
1124 if constexpr (__have_avx512bw_vl && sizeof(_Tp) == 1)
1125 _mm_mask_storeu_epi8(__mem, __k, __vi);
1126 else if constexpr (__have_avx512bw_vl && sizeof(_Tp) == 2)
1127 _mm_mask_storeu_epi16(__mem, __k, __vi);
1128 else if constexpr (__have_avx512vl && sizeof(_Tp) == 4)
1129 {
1130 if constexpr (is_integral_v<_Tp>)
1131 _mm_mask_storeu_epi32(__mem, __k, __vi);
1132 else
1133 _mm_mask_storeu_ps(__mem, __k, __vi);
1134 }
1135 else if constexpr (__have_avx512vl && sizeof(_Tp) == 8)
1136 {
1137 if constexpr (is_integral_v<_Tp>)
1138 _mm_mask_storeu_epi64(__mem, __k, __vi);
1139 else
1140 _mm_mask_storeu_pd(__mem, __k, __vi);
1141 }
1142 else if constexpr (__have_avx512f
1143 && (sizeof(_Tp) >= 4 || __have_avx512bw))
1144 {
1145 // use a 512-bit maskstore, using zero-extension of the bitmask
1146 _S_masked_store_nocvt(
1147 _SimdWrapper64<_Tp>(
1148 __intrin_bitcast<__intrinsic_type64_t<_Tp>>(__v._M_data)),
1149 __mem, _SimdWrapper<bool, 64 / sizeof(_Tp)>(__k._M_data));
1150 }
1151 else
1152 _S_masked_store_nocvt(__v, __mem,
1153 _MaskImpl::template _S_to_maskvector<
1154 __int_for_sizeof_t<_Tp>, _Np>(__k));
1155 }
1156 else
1157 __assert_unreachable<_Tp>();
1158 }
1159
1160 template <typename _Tp, size_t _Np>
1161 _GLIBCXX_SIMD_INTRINSIC static void
1162 _S_masked_store_nocvt(_SimdWrapper<_Tp, _Np> __v, _Tp* __mem,
1163 _SimdWrapper<__int_for_sizeof_t<_Tp>, _Np> __k)
1164 {
1165 if constexpr (sizeof(__v) <= 16)
1166 {
1167 [[maybe_unused]] const auto __vi
1168 = __intrin_bitcast<__m128i>(__as_vector(__v));
1169 [[maybe_unused]] const auto __ki
1170 = __intrin_bitcast<__m128i>(__as_vector(__k));
1171 if constexpr (__have_avx512bw_vl && sizeof(_Tp) == 1)
1172 _mm_mask_storeu_epi8(__mem, _mm_movepi8_mask(__ki), __vi);
1173 else if constexpr (__have_avx512bw_vl && sizeof(_Tp) == 2)
1174 _mm_mask_storeu_epi16(__mem, _mm_movepi16_mask(__ki), __vi);
1175 else if constexpr (__have_avx2 && sizeof(_Tp) == 4
1176 && is_integral_v<_Tp>)
1177 _mm_maskstore_epi32(reinterpret_cast<int*>(__mem), __ki, __vi);
1178 else if constexpr (__have_avx && sizeof(_Tp) == 4)
1179 _mm_maskstore_ps(reinterpret_cast<float*>(__mem), __ki,
1180 __vector_bitcast<float>(__vi));
1181 else if constexpr (__have_avx2 && sizeof(_Tp) == 8
1182 && is_integral_v<_Tp>)
1183 _mm_maskstore_epi64(reinterpret_cast<_LLong*>(__mem), __ki, __vi);
1184 else if constexpr (__have_avx && sizeof(_Tp) == 8)
1185 _mm_maskstore_pd(reinterpret_cast<double*>(__mem), __ki,
1186 __vector_bitcast<double>(__vi));
1187 else if constexpr (__have_sse2)
1188 _mm_maskmoveu_si128(__vi, __ki, reinterpret_cast<char*>(__mem));
1189 }
1190 else if constexpr (sizeof(__v) == 32)
1191 {
1192 [[maybe_unused]] const auto __vi
1193 = __intrin_bitcast<__m256i>(__as_vector(__v));
1194 [[maybe_unused]] const auto __ki
1195 = __intrin_bitcast<__m256i>(__as_vector(__k));
1196 if constexpr (__have_avx512bw_vl && sizeof(_Tp) == 1)
1197 _mm256_mask_storeu_epi8(__mem, _mm256_movepi8_mask(__ki), __vi);
1198 else if constexpr (__have_avx512bw_vl && sizeof(_Tp) == 2)
1199 _mm256_mask_storeu_epi16(__mem, _mm256_movepi16_mask(__ki), __vi);
1200 else if constexpr (__have_avx2 && sizeof(_Tp) == 4
1201 && is_integral_v<_Tp>)
1202 _mm256_maskstore_epi32(reinterpret_cast<int*>(__mem), __ki, __vi);
1203 else if constexpr (sizeof(_Tp) == 4)
1204 _mm256_maskstore_ps(reinterpret_cast<float*>(__mem), __ki,
1205 __vector_bitcast<float>(__v));
1206 else if constexpr (__have_avx2 && sizeof(_Tp) == 8
1207 && is_integral_v<_Tp>)
1208 _mm256_maskstore_epi64(reinterpret_cast<_LLong*>(__mem), __ki,
1209 __vi);
1210 else if constexpr (__have_avx && sizeof(_Tp) == 8)
1211 _mm256_maskstore_pd(reinterpret_cast<double*>(__mem), __ki,
1212 __vector_bitcast<double>(__v));
1213 else if constexpr (__have_sse2)
1214 {
1215 _mm_maskmoveu_si128(__lo128(__vi), __lo128(__ki),
1216 reinterpret_cast<char*>(__mem));
1217 _mm_maskmoveu_si128(__hi128(__vi), __hi128(__ki),
1218 reinterpret_cast<char*>(__mem) + 16);
1219 }
1220 }
1221 else
1222 __assert_unreachable<_Tp>();
1223 }
1224
1225 // }}}
1226 // _S_masked_store {{{
1227 template <typename _Tp, size_t _Np, typename _Up>
1228 _GLIBCXX_SIMD_INTRINSIC static void
1229 _S_masked_store(const _SimdWrapper<_Tp, _Np> __v, _Up* __mem,
1230 const _MaskMember<_Tp> __k) noexcept
1231 {
1232 if constexpr (is_integral_v<
1233 _Tp> && is_integral_v<_Up> && sizeof(_Tp) > sizeof(_Up)
1234 && __have_avx512f && (sizeof(_Tp) >= 4 || __have_avx512bw)
1235 && (sizeof(__v) == 64 || __have_avx512vl))
1236 { // truncating store
1237 const auto __vi = __to_intrin(__v);
1238 const auto __kk = _MaskImpl::_S_to_bits(__k)._M_to_bits();
1239 if constexpr (sizeof(_Tp) == 8 && sizeof(_Up) == 4
1240 && sizeof(__vi) == 64)
1241 _mm512_mask_cvtepi64_storeu_epi32(__mem, __kk, __vi);
1242 else if constexpr (sizeof(_Tp) == 8 && sizeof(_Up) == 4
1243 && sizeof(__vi) == 32)
1244 _mm256_mask_cvtepi64_storeu_epi32(__mem, __kk, __vi);
1245 else if constexpr (sizeof(_Tp) == 8 && sizeof(_Up) == 4
1246 && sizeof(__vi) == 16)
1247 _mm_mask_cvtepi64_storeu_epi32(__mem, __kk, __vi);
1248 else if constexpr (sizeof(_Tp) == 8 && sizeof(_Up) == 2
1249 && sizeof(__vi) == 64)
1250 _mm512_mask_cvtepi64_storeu_epi16(__mem, __kk, __vi);
1251 else if constexpr (sizeof(_Tp) == 8 && sizeof(_Up) == 2
1252 && sizeof(__vi) == 32)
1253 _mm256_mask_cvtepi64_storeu_epi16(__mem, __kk, __vi);
1254 else if constexpr (sizeof(_Tp) == 8 && sizeof(_Up) == 2
1255 && sizeof(__vi) == 16)
1256 _mm_mask_cvtepi64_storeu_epi16(__mem, __kk, __vi);
1257 else if constexpr (sizeof(_Tp) == 8 && sizeof(_Up) == 1
1258 && sizeof(__vi) == 64)
1259 _mm512_mask_cvtepi64_storeu_epi8(__mem, __kk, __vi);
1260 else if constexpr (sizeof(_Tp) == 8 && sizeof(_Up) == 1
1261 && sizeof(__vi) == 32)
1262 _mm256_mask_cvtepi64_storeu_epi8(__mem, __kk, __vi);
1263 else if constexpr (sizeof(_Tp) == 8 && sizeof(_Up) == 1
1264 && sizeof(__vi) == 16)
1265 _mm_mask_cvtepi64_storeu_epi8(__mem, __kk, __vi);
1266 else if constexpr (sizeof(_Tp) == 4 && sizeof(_Up) == 2
1267 && sizeof(__vi) == 64)
1268 _mm512_mask_cvtepi32_storeu_epi16(__mem, __kk, __vi);
1269 else if constexpr (sizeof(_Tp) == 4 && sizeof(_Up) == 2
1270 && sizeof(__vi) == 32)
1271 _mm256_mask_cvtepi32_storeu_epi16(__mem, __kk, __vi);
1272 else if constexpr (sizeof(_Tp) == 4 && sizeof(_Up) == 2
1273 && sizeof(__vi) == 16)
1274 _mm_mask_cvtepi32_storeu_epi16(__mem, __kk, __vi);
1275 else if constexpr (sizeof(_Tp) == 4 && sizeof(_Up) == 1
1276 && sizeof(__vi) == 64)
1277 _mm512_mask_cvtepi32_storeu_epi8(__mem, __kk, __vi);
1278 else if constexpr (sizeof(_Tp) == 4 && sizeof(_Up) == 1
1279 && sizeof(__vi) == 32)
1280 _mm256_mask_cvtepi32_storeu_epi8(__mem, __kk, __vi);
1281 else if constexpr (sizeof(_Tp) == 4 && sizeof(_Up) == 1
1282 && sizeof(__vi) == 16)
1283 _mm_mask_cvtepi32_storeu_epi8(__mem, __kk, __vi);
1284 else if constexpr (sizeof(_Tp) == 2 && sizeof(_Up) == 1
1285 && sizeof(__vi) == 64)
1286 _mm512_mask_cvtepi16_storeu_epi8(__mem, __kk, __vi);
1287 else if constexpr (sizeof(_Tp) == 2 && sizeof(_Up) == 1
1288 && sizeof(__vi) == 32)
1289 _mm256_mask_cvtepi16_storeu_epi8(__mem, __kk, __vi);
1290 else if constexpr (sizeof(_Tp) == 2 && sizeof(_Up) == 1
1291 && sizeof(__vi) == 16)
1292 _mm_mask_cvtepi16_storeu_epi8(__mem, __kk, __vi);
1293 else
1294 __assert_unreachable<_Tp>();
1295 }
1296 else
1297 _Base::_S_masked_store(__v, __mem, __k);
1298 }
1299
1300 // }}}
1301 // _S_multiplies {{{
1302 template <typename _V, typename _VVT = _VectorTraits<_V>>
1303 _GLIBCXX_SIMD_INTRINSIC static constexpr _V _S_multiplies(_V __x, _V __y)
1304 {
1305 using _Tp = typename _VVT::value_type;
1306 if (__builtin_is_constant_evaluated() || __x._M_is_constprop()
1307 || __y._M_is_constprop())
1308 return __as_vector(__x) * __as_vector(__y);
1309 else if constexpr (sizeof(_Tp) == 1)
1310 {
1311 if constexpr (sizeof(_V) == 2)
1312 {
1313 const auto __xs = reinterpret_cast<short>(__x._M_data);
1314 const auto __ys = reinterpret_cast<short>(__y._M_data);
1315 return reinterpret_cast<__vector_type_t<_Tp, 2>>(short(
1316 ((__xs * __ys) & 0xff) | ((__xs >> 8) * (__ys & 0xff00))));
1317 }
1318 else if constexpr (sizeof(_V) == 4 && _VVT::_S_partial_width == 3)
1319 {
1320 const auto __xi = reinterpret_cast<int>(__x._M_data);
1321 const auto __yi = reinterpret_cast<int>(__y._M_data);
1322 return reinterpret_cast<__vector_type_t<_Tp, 3>>(
1323 ((__xi * __yi) & 0xff)
1324 | (((__xi >> 8) * (__yi & 0xff00)) & 0xff00)
1325 | ((__xi >> 16) * (__yi & 0xff0000)));
1326 }
1327 else if constexpr (sizeof(_V) == 4)
1328 {
1329 const auto __xi = reinterpret_cast<int>(__x._M_data);
1330 const auto __yi = reinterpret_cast<int>(__y._M_data);
1331 return reinterpret_cast<__vector_type_t<_Tp, 4>>(
1332 ((__xi * __yi) & 0xff)
1333 | (((__xi >> 8) * (__yi & 0xff00)) & 0xff00)
1334 | (((__xi >> 16) * (__yi & 0xff0000)) & 0xff0000)
1335 | ((__xi >> 24) * (__yi & 0xff000000u)));
1336 }
1337 else if constexpr (sizeof(_V) == 8 && __have_avx2
1338 && is_signed_v<_Tp>)
1339 return __convert<typename _VVT::type>(
1340 __vector_bitcast<short>(_mm_cvtepi8_epi16(__to_intrin(__x)))
1341 * __vector_bitcast<short>(_mm_cvtepi8_epi16(__to_intrin(__y))));
1342 else if constexpr (sizeof(_V) == 8 && __have_avx2
1343 && is_unsigned_v<_Tp>)
1344 return __convert<typename _VVT::type>(
1345 __vector_bitcast<short>(_mm_cvtepu8_epi16(__to_intrin(__x)))
1346 * __vector_bitcast<short>(_mm_cvtepu8_epi16(__to_intrin(__y))));
1347 else
1348 {
1349 // codegen of `x*y` is suboptimal (as of GCC 9.0.1)
1350 constexpr size_t __full_size = _VVT::_S_full_size;
1351 constexpr int _Np = sizeof(_V) >= 16 ? __full_size / 2 : 8;
1352 using _ShortW = _SimdWrapper<short, _Np>;
1353 const _ShortW __even = __vector_bitcast<short, _Np>(__x)
1354 * __vector_bitcast<short, _Np>(__y);
1355 _ShortW __high_byte = _ShortW()._M_data - 256;
1356 //[&]() { asm("" : "+x"(__high_byte._M_data)); }();
1357 const _ShortW __odd
1358 = (__vector_bitcast<short, _Np>(__x) >> 8)
1359 * (__vector_bitcast<short, _Np>(__y) & __high_byte._M_data);
1360 if constexpr (__have_avx512bw && sizeof(_V) > 2)
1361 return _CommonImplX86::_S_blend_avx512(
1362 0xaaaa'aaaa'aaaa'aaaaLL, __vector_bitcast<_Tp>(__even),
1363 __vector_bitcast<_Tp>(__odd));
1364 else if constexpr (__have_sse4_1 && sizeof(_V) > 2)
1365 return _CommonImplX86::_S_blend_intrin(__to_intrin(
1366 __high_byte),
1367 __to_intrin(__even),
1368 __to_intrin(__odd));
1369 else
1370 return __to_intrin(
1371 __or(__andnot(__high_byte, __even), __odd));
1372 }
1373 }
1374 else
1375 return _Base::_S_multiplies(__x, __y);
1376 }
1377
1378 // }}}
1379 // _S_divides {{{
1380#ifdef _GLIBCXX_SIMD_WORKAROUND_PR90993
1381 template <typename _Tp, size_t _Np>
1382 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
1383 _S_divides(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1384 {
1385 if (!__builtin_is_constant_evaluated()
1386 && !__builtin_constant_p(__y._M_data))
1387 if constexpr (is_integral_v<_Tp> && sizeof(_Tp) <= 4)
1388 { // use divps - codegen of `x/y` is suboptimal (as of GCC 9.0.1)
1389 // Note that using floating-point division is likely to raise the
1390 // *Inexact* exception flag and thus appears like an invalid
1391 // "as-if" transformation. However, C++ doesn't specify how the
1392 // fpenv can be observed and points to C. C says that function
1393 // calls are assumed to potentially raise fp exceptions, unless
1394 // documented otherwise. Consequently, operator/, which is a
1395 // function call, may raise fp exceptions.
1396 /*const struct _CsrGuard
1397 {
1398 const unsigned _M_data = _mm_getcsr();
1399 _CsrGuard()
1400 {
1401 _mm_setcsr(0x9f80); // turn off FP exceptions and
1402 flush-to-zero
1403 }
1404 ~_CsrGuard() { _mm_setcsr(_M_data); }
1405 } __csr;*/
1406 using _Float = conditional_t<sizeof(_Tp) == 4, double, float>;
1407 constexpr size_t __n_intermediate
1408 = std::min(_Np, (__have_avx512f ? 64
1409 : __have_avx ? 32
1410 : 16)
1411 / sizeof(_Float));
1412 using _FloatV = __vector_type_t<_Float, __n_intermediate>;
1413 constexpr size_t __n_floatv
1414 = __div_roundup(_Np, __n_intermediate);
1415 using _R = __vector_type_t<_Tp, _Np>;
1416 const auto __xf = __convert_all<_FloatV, __n_floatv>(__x);
1417 const auto __yf = __convert_all<_FloatV, __n_floatv>(
1418 _Abi::__make_padding_nonzero(__as_vector(__y)));
1419 return __call_with_n_evaluations<__n_floatv>(
1420 [](auto... __quotients) {
1421 return __vector_convert<_R>(__quotients...);
1422 },
1423 [&__xf,
1424 &__yf](auto __i) -> _SimdWrapper<_Float, __n_intermediate> {
1425#if !defined __clang__ && __GCC_IEC_559 == 0
1426 // If -freciprocal-math is active, using the `/` operator is
1427 // incorrect because it may be translated to an imprecise
1428 // multiplication with reciprocal. We need to use inline
1429 // assembly to force a real division.
1430 _FloatV __r;
1431 if constexpr (__have_avx) // -mno-sse2avx is irrelevant
1432 // because once -mavx is given, GCC
1433 // emits VEX encoded vdivp[sd]
1434 {
1435 if constexpr (sizeof(_Tp) == 4)
1436 asm("vdivpd\t{%2, %1, %0|%0, %1, %2}"
1437 : "=x"(__r)
1438 : "x"(__xf[__i]), "x"(__yf[__i]));
1439 else
1440 asm("vdivps\t{%2, %1, %0|%0, %1, %2}"
1441 : "=x"(__r)
1442 : "x"(__xf[__i]), "x"(__yf[__i]));
1443 }
1444 else
1445 {
1446 __r = __xf[__i];
1447 if constexpr (sizeof(_Tp) == 4)
1448 asm("divpd\t{%1, %0|%0, %1}"
1449 : "=x"(__r)
1450 : "x"(__yf[__i]));
1451 else
1452 asm("divps\t{%1, %0|%0, %1}"
1453 : "=x"(__r)
1454 : "x"(__yf[__i]));
1455 }
1456 return __r;
1457#else
1458 return __xf[__i] / __yf[__i];
1459#endif
1460 });
1461 }
1462 /* 64-bit int division is potentially optimizable via double division if
1463 * the value in __x is small enough and the conversion between
1464 * int<->double is efficient enough:
1465 else if constexpr (is_integral_v<_Tp> && is_unsigned_v<_Tp> &&
1466 sizeof(_Tp) == 8)
1467 {
1468 if constexpr (__have_sse4_1 && sizeof(__x) == 16)
1469 {
1470 if (_mm_test_all_zeros(__x, __m128i{0xffe0'0000'0000'0000ull,
1471 0xffe0'0000'0000'0000ull}))
1472 {
1473 __x._M_data | 0x __vector_convert<__m128d>(__x._M_data)
1474 }
1475 }
1476 }
1477 */
1478 return _Base::_S_divides(__x, __y);
1479 }
1480 #endif // _GLIBCXX_SIMD_WORKAROUND_PR90993
1481
1482 // }}}
1483 // _S_modulus {{{
1484 template <typename _Tp, size_t _Np>
1485 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
1486 _S_modulus(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1487 {
1488 if (__builtin_is_constant_evaluated()
1489 || __builtin_constant_p(__y._M_data) || sizeof(_Tp) >= 8)
1490 return _Base::_S_modulus(__x, __y);
1491 else
1492 return _Base::_S_minus(__x, _S_multiplies(__y, _S_divides(__x, __y)));
1493 }
1494
1495 // }}}
1496 // _S_bit_shift_left {{{
1497 // Notes on UB. C++2a [expr.shift] says:
1498 // -1- [...] The operands shall be of integral or unscoped enumeration type
1499 // and integral promotions are performed. The type of the result is that
1500 // of the promoted left operand. The behavior is undefined if the right
1501 // operand is negative, or greater than or equal to the width of the
1502 // promoted left operand.
1503 // -2- The value of E1 << E2 is the unique value congruent to E1×2^E2 modulo
1504 // 2^N, where N is the width of the type of the result.
1505 //
1506 // C++17 [expr.shift] says:
1507 // -2- The value of E1 << E2 is E1 left-shifted E2 bit positions; vacated
1508 // bits are zero-filled. If E1 has an unsigned type, the value of the
1509 // result is E1 × 2^E2 , reduced modulo one more than the maximum value
1510 // representable in the result type. Otherwise, if E1 has a signed type
1511 // and non-negative value, and E1 × 2^E2 is representable in the
1512 // corresponding unsigned type of the result type, then that value,
1513 // converted to the result type, is the resulting value; otherwise, the
1514 // behavior is undefined.
1515 //
1516 // Consequences:
1517 // With C++2a signed and unsigned types have the same UB
1518 // characteristics:
1519 // - left shift is not UB for 0 <= RHS < max(32, #bits(T))
1520 //
1521 // With C++17 there's little room for optimizations because the standard
1522 // requires all shifts to happen on promoted integrals (i.e. int). Thus,
1523 // short and char shifts must assume shifts affect bits of neighboring
1524 // values.
1525 #ifndef _GLIBCXX_SIMD_NO_SHIFT_OPT
1526 template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
1527 inline _GLIBCXX_CONST static typename _TVT::type
1528 _S_bit_shift_left(_Tp __xx, int __y)
1529 {
1530 using _V = typename _TVT::type;
1531 using _Up = typename _TVT::value_type;
1532 _V __x = __xx;
1533 [[maybe_unused]] const auto __ix = __to_intrin(__x);
1534 if (__builtin_is_constant_evaluated())
1535 return __x << __y;
1536#if __cplusplus > 201703
1537 // after C++17, signed shifts have no UB, and behave just like unsigned
1538 // shifts
1539 else if constexpr (sizeof(_Up) == 1 && is_signed_v<_Up>)
1540 return __vector_bitcast<_Up>(
1541 _S_bit_shift_left(__vector_bitcast<make_unsigned_t<_Up>>(__x),
1542 __y));
1543#endif
1544 else if constexpr (sizeof(_Up) == 1)
1545 {
1546 // (cf. https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83894)
1547 if (__builtin_constant_p(__y))
1548 {
1549 if (__y == 0)
1550 return __x;
1551 else if (__y == 1)
1552 return __x + __x;
1553 else if (__y == 2)
1554 {
1555 __x = __x + __x;
1556 return __x + __x;
1557 }
1558 else if (__y > 2 && __y < 8)
1559 {
1560 if constexpr (sizeof(__x) > sizeof(unsigned))
1561 {
1562 const _UChar __mask = 0xff << __y; // precomputed vector
1563 return __vector_bitcast<_Up>(
1564 __vector_bitcast<_UChar>(
1565 __vector_bitcast<unsigned>(__x) << __y)
1566 & __mask);
1567 }
1568 else
1569 {
1570 const unsigned __mask
1571 = (0xff & (0xff << __y)) * 0x01010101u;
1572 return reinterpret_cast<_V>(
1573 static_cast<__int_for_sizeof_t<_V>>(
1574 unsigned(
1575 reinterpret_cast<__int_for_sizeof_t<_V>>(__x)
1576 << __y)
1577 & __mask));
1578 }
1579 }
1580 else if (__y >= 8 && __y < 32)
1581 return _V();
1582 else
1583 __builtin_unreachable();
1584 }
1585 // general strategy in the following: use an sllv instead of sll
1586 // instruction, because it's 2 to 4 times faster:
1587 else if constexpr (__have_avx512bw_vl && sizeof(__x) == 16)
1588 return __vector_bitcast<_Up>(_mm256_cvtepi16_epi8(
1589 _mm256_sllv_epi16(_mm256_cvtepi8_epi16(__ix),
1590 _mm256_set1_epi16(__y))));
1591 else if constexpr (__have_avx512bw && sizeof(__x) == 32)
1592 return __vector_bitcast<_Up>(_mm512_cvtepi16_epi8(
1593 _mm512_sllv_epi16(_mm512_cvtepi8_epi16(__ix),
1594 _mm512_set1_epi16(__y))));
1595 else if constexpr (__have_avx512bw && sizeof(__x) == 64)
1596 {
1597 const auto __shift = _mm512_set1_epi16(__y);
1598 return __vector_bitcast<_Up>(
1599 __concat(_mm512_cvtepi16_epi8(_mm512_sllv_epi16(
1600 _mm512_cvtepi8_epi16(__lo256(__ix)), __shift)),
1601 _mm512_cvtepi16_epi8(_mm512_sllv_epi16(
1602 _mm512_cvtepi8_epi16(__hi256(__ix)), __shift))));
1603 }
1604 else if constexpr (__have_avx2 && sizeof(__x) == 32)
1605 {
1606#if 1
1607 const auto __shift = _mm_cvtsi32_si128(__y);
1608 auto __k
1609 = _mm256_sll_epi16(_mm256_slli_epi16(~__m256i(), 8), __shift);
1610 __k |= _mm256_srli_epi16(__k, 8);
1611 return __vector_bitcast<_Up>(_mm256_sll_epi32(__ix, __shift)
1612 & __k);
1613#else
1614 const _Up __k = 0xff << __y;
1615 return __vector_bitcast<_Up>(__vector_bitcast<int>(__x) << __y)
1616 & __k;
1617#endif
1618 }
1619 else
1620 {
1621 const auto __shift = _mm_cvtsi32_si128(__y);
1622 auto __k
1623 = _mm_sll_epi16(_mm_slli_epi16(~__m128i(), 8), __shift);
1624 __k |= _mm_srli_epi16(__k, 8);
1625 return __intrin_bitcast<_V>(_mm_sll_epi16(__ix, __shift) & __k);
1626 }
1627 }
1628 return __x << __y;
1629 }
1630
1631 template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
1632 inline _GLIBCXX_CONST static typename _TVT::type
1633 _S_bit_shift_left(_Tp __xx, typename _TVT::type __y)
1634 {
1635 using _V = typename _TVT::type;
1636 using _Up = typename _TVT::value_type;
1637 _V __x = __xx;
1638 [[maybe_unused]] const auto __ix = __to_intrin(__x);
1639 [[maybe_unused]] const auto __iy = __to_intrin(__y);
1640 if (__builtin_is_constant_evaluated())
1641 return __x << __y;
1642#if __cplusplus > 201703
1643 // after C++17, signed shifts have no UB, and behave just like unsigned
1644 // shifts
1645 else if constexpr (is_signed_v<_Up>)
1646 return __vector_bitcast<_Up>(
1647 _S_bit_shift_left(__vector_bitcast<make_unsigned_t<_Up>>(__x),
1648 __vector_bitcast<make_unsigned_t<_Up>>(__y)));
1649#endif
1650 else if constexpr (sizeof(_Up) == 1)
1651 {
1652 if constexpr (sizeof __ix == 64 && __have_avx512bw)
1653 return __vector_bitcast<_Up>(__concat(
1654 _mm512_cvtepi16_epi8(
1655 _mm512_sllv_epi16(_mm512_cvtepu8_epi16(__lo256(__ix)),
1656 _mm512_cvtepu8_epi16(__lo256(__iy)))),
1657 _mm512_cvtepi16_epi8(
1658 _mm512_sllv_epi16(_mm512_cvtepu8_epi16(__hi256(__ix)),
1659 _mm512_cvtepu8_epi16(__hi256(__iy))))));
1660 else if constexpr (sizeof __ix == 32 && __have_avx512bw)
1661 return __vector_bitcast<_Up>(_mm512_cvtepi16_epi8(
1662 _mm512_sllv_epi16(_mm512_cvtepu8_epi16(__ix),
1663 _mm512_cvtepu8_epi16(__iy))));
1664 else if constexpr (sizeof __x <= 8 && __have_avx512bw_vl)
1665 return __intrin_bitcast<_V>(
1666 _mm_cvtepi16_epi8(_mm_sllv_epi16(_mm_cvtepu8_epi16(__ix),
1667 _mm_cvtepu8_epi16(__iy))));
1668 else if constexpr (sizeof __ix == 16 && __have_avx512bw_vl)
1669 return __intrin_bitcast<_V>(_mm256_cvtepi16_epi8(
1670 _mm256_sllv_epi16(_mm256_cvtepu8_epi16(__ix),
1671 _mm256_cvtepu8_epi16(__iy))));
1672 else if constexpr (sizeof __ix == 16 && __have_avx512bw)
1673 return __intrin_bitcast<_V>(
1674 __lo128(_mm512_cvtepi16_epi8(_mm512_sllv_epi16(
1675 _mm512_cvtepu8_epi16(_mm256_castsi128_si256(__ix)),
1676 _mm512_cvtepu8_epi16(_mm256_castsi128_si256(__iy))))));
1677 else if constexpr (__have_sse4_1 && sizeof(__x) == 16)
1678 {
1679 auto __mask
1680 = __vector_bitcast<_Up>(__vector_bitcast<short>(__y) << 5);
1681 auto __x4
1682 = __vector_bitcast<_Up>(__vector_bitcast<short>(__x) << 4);
1683 __x4 &= char(0xf0);
1684 __x = reinterpret_cast<_V>(_CommonImplX86::_S_blend_intrin(
1685 __to_intrin(__mask), __to_intrin(__x), __to_intrin(__x4)));
1686 __mask += __mask;
1687 auto __x2
1688 = __vector_bitcast<_Up>(__vector_bitcast<short>(__x) << 2);
1689 __x2 &= char(0xfc);
1690 __x = reinterpret_cast<_V>(_CommonImplX86::_S_blend_intrin(
1691 __to_intrin(__mask), __to_intrin(__x), __to_intrin(__x2)));
1692 __mask += __mask;
1693 auto __x1 = __x + __x;
1694 __x = reinterpret_cast<_V>(_CommonImplX86::_S_blend_intrin(
1695 __to_intrin(__mask), __to_intrin(__x), __to_intrin(__x1)));
1696 return __x
1697 & ((__y & char(0xf8)) == 0); // y > 7 nulls the result
1698 }
1699 else if constexpr (sizeof(__x) == 16)
1700 {
1701 auto __mask
1702 = __vector_bitcast<_UChar>(__vector_bitcast<short>(__y) << 5);
1703 auto __x4
1704 = __vector_bitcast<_Up>(__vector_bitcast<short>(__x) << 4);
1705 __x4 &= char(0xf0);
1706 __x = __vector_bitcast<_SChar>(__mask) < 0 ? __x4 : __x;
1707 __mask += __mask;
1708 auto __x2
1709 = __vector_bitcast<_Up>(__vector_bitcast<short>(__x) << 2);
1710 __x2 &= char(0xfc);
1711 __x = __vector_bitcast<_SChar>(__mask) < 0 ? __x2 : __x;
1712 __mask += __mask;
1713 auto __x1 = __x + __x;
1714 __x = __vector_bitcast<_SChar>(__mask) < 0 ? __x1 : __x;
1715 return __x
1716 & ((__y & char(0xf8)) == 0); // y > 7 nulls the result
1717 }
1718 else
1719 return __x << __y;
1720 }
1721 else if constexpr (sizeof(_Up) == 2)
1722 {
1723 if constexpr (sizeof __ix == 64 && __have_avx512bw)
1724 return __vector_bitcast<_Up>(_mm512_sllv_epi16(__ix, __iy));
1725 else if constexpr (sizeof __ix == 32 && __have_avx512bw_vl)
1726 return __vector_bitcast<_Up>(_mm256_sllv_epi16(__ix, __iy));
1727 else if constexpr (sizeof __ix == 32 && __have_avx512bw)
1728 return __vector_bitcast<_Up>(
1729 __lo256(_mm512_sllv_epi16(_mm512_castsi256_si512(__ix),
1730 _mm512_castsi256_si512(__iy))));
1731 else if constexpr (sizeof __ix == 32 && __have_avx2)
1732 {
1733 const auto __ux = __vector_bitcast<unsigned>(__x);
1734 const auto __uy = __vector_bitcast<unsigned>(__y);
1735 return __vector_bitcast<_Up>(_mm256_blend_epi16(
1736 __auto_bitcast(__ux << (__uy & 0x0000ffffu)),
1737 __auto_bitcast((__ux & 0xffff0000u) << (__uy >> 16)), 0xaa));
1738 }
1739 else if constexpr (sizeof __ix == 16 && __have_avx512bw_vl)
1740 return __intrin_bitcast<_V>(_mm_sllv_epi16(__ix, __iy));
1741 else if constexpr (sizeof __ix == 16 && __have_avx512bw)
1742 return __intrin_bitcast<_V>(
1743 __lo128(_mm512_sllv_epi16(_mm512_castsi128_si512(__ix),
1744 _mm512_castsi128_si512(__iy))));
1745 else if constexpr (sizeof __ix == 16 && __have_avx2)
1746 {
1747 const auto __ux = __vector_bitcast<unsigned>(__ix);
1748 const auto __uy = __vector_bitcast<unsigned>(__iy);
1749 return __intrin_bitcast<_V>(_mm_blend_epi16(
1750 __auto_bitcast(__ux << (__uy & 0x0000ffffu)),
1751 __auto_bitcast((__ux & 0xffff0000u) << (__uy >> 16)), 0xaa));
1752 }
1753 else if constexpr (sizeof __ix == 16)
1754 {
1755 using _Float4 = __vector_type_t<float, 4>;
1756 using _Int4 = __vector_type_t<int, 4>;
1757 using _UInt4 = __vector_type_t<unsigned, 4>;
1758 const _UInt4 __yu
1759 = reinterpret_cast<_UInt4>(__to_intrin(__y + (0x3f8 >> 3)));
1760 return __x
1761 * __intrin_bitcast<_V>(
1762 __vector_convert<_Int4>(_SimdWrapper<float, 4>(
1763 reinterpret_cast<_Float4>(__yu << 23)))
1764 | (__vector_convert<_Int4>(_SimdWrapper<float, 4>(
1765 reinterpret_cast<_Float4>((__yu >> 16) << 23)))
1766 << 16));
1767 }
1768 else
1769 __assert_unreachable<_Tp>();
1770 }
1771 else if constexpr (sizeof(_Up) == 4 && sizeof __ix == 16
1772 && !__have_avx2)
1773 // latency is suboptimal, but throughput is at full speedup
1774 return __intrin_bitcast<_V>(
1775 __vector_bitcast<unsigned>(__ix)
1776 * __vector_convert<__vector_type16_t<int>>(
1777 _SimdWrapper<float, 4>(__vector_bitcast<float>(
1778 (__vector_bitcast<unsigned, 4>(__y) << 23) + 0x3f80'0000))));
1779 else if constexpr (sizeof(_Up) == 8 && sizeof __ix == 16
1780 && !__have_avx2)
1781 {
1782 const auto __lo = _mm_sll_epi64(__ix, __iy);
1783 const auto __hi
1784 = _mm_sll_epi64(__ix, _mm_unpackhi_epi64(__iy, __iy));
1785 if constexpr (__have_sse4_1)
1786 return __vector_bitcast<_Up>(_mm_blend_epi16(__lo, __hi, 0xf0));
1787 else
1788 return __vector_bitcast<_Up>(
1789 _mm_move_sd(__vector_bitcast<double>(__hi),
1790 __vector_bitcast<double>(__lo)));
1791 }
1792 else
1793 return __x << __y;
1794 }
1795#endif // _GLIBCXX_SIMD_NO_SHIFT_OPT
1796
1797 // }}}
1798 // _S_bit_shift_right {{{
1799#ifndef _GLIBCXX_SIMD_NO_SHIFT_OPT
1800 template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
1801 inline _GLIBCXX_CONST static typename _TVT::type
1802 _S_bit_shift_right(_Tp __xx, int __y)
1803 {
1804 using _V = typename _TVT::type;
1805 using _Up = typename _TVT::value_type;
1806 _V __x = __xx;
1807 [[maybe_unused]] const auto __ix = __to_intrin(__x);
1808 if (__builtin_is_constant_evaluated())
1809 return __x >> __y;
1810 else if (__builtin_constant_p(__y)
1811 && is_unsigned_v<
1812 _Up> && __y >= int(sizeof(_Up) * __CHAR_BIT__))
1813 return _V();
1814 else if constexpr (sizeof(_Up) == 1 && is_unsigned_v<_Up>) //{{{
1815 return __intrin_bitcast<_V>(__vector_bitcast<_UShort>(__ix) >> __y)
1816 & _Up(0xff >> __y);
1817 //}}}
1818 else if constexpr (sizeof(_Up) == 1 && is_signed_v<_Up>) //{{{
1819 return __intrin_bitcast<_V>(
1820 (__vector_bitcast<_UShort>(__vector_bitcast<short>(__ix)
1821 >> (__y + 8))
1822 << 8)
1823 | (__vector_bitcast<_UShort>(
1824 __vector_bitcast<short>(__vector_bitcast<_UShort>(__ix) << 8)
1825 >> __y)
1826 >> 8));
1827 //}}}
1828 // GCC optimizes sizeof == 2, 4, and unsigned 8 as expected
1829 else if constexpr (sizeof(_Up) == 8 && is_signed_v<_Up>) //{{{
1830 {
1831 if (__y > 32)
1832 return (__intrin_bitcast<_V>(__vector_bitcast<int>(__ix) >> 32)
1833 & _Up(0xffff'ffff'0000'0000ull))
1834 | __vector_bitcast<_Up>(
1835 __vector_bitcast<int>(__vector_bitcast<_ULLong>(__ix)
1836 >> 32)
1837 >> (__y - 32));
1838 else
1839 return __intrin_bitcast<_V>(__vector_bitcast<_ULLong>(__ix)
1840 >> __y)
1841 | __vector_bitcast<_Up>(
1842 __vector_bitcast<int>(__ix & -0x8000'0000'0000'0000ll)
1843 >> __y);
1844 }
1845 //}}}
1846 else
1847 return __x >> __y;
1848 }
1849
1850 template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
1851 inline _GLIBCXX_CONST static typename _TVT::type
1852 _S_bit_shift_right(_Tp __xx, typename _TVT::type __y)
1853 {
1854 using _V = typename _TVT::type;
1855 using _Up = typename _TVT::value_type;
1856 _V __x = __xx;
1857 [[maybe_unused]] const auto __ix = __to_intrin(__x);
1858 [[maybe_unused]] const auto __iy = __to_intrin(__y);
1859 if (__builtin_is_constant_evaluated()
1860 || (__builtin_constant_p(__x) && __builtin_constant_p(__y)))
1861 return __x >> __y;
1862 else if constexpr (sizeof(_Up) == 1) //{{{
1863 {
1864 if constexpr (sizeof(__x) <= 8 && __have_avx512bw_vl)
1865 return __intrin_bitcast<_V>(_mm_cvtepi16_epi8(
1866 is_signed_v<_Up> ? _mm_srav_epi16(_mm_cvtepi8_epi16(__ix),
1867 _mm_cvtepi8_epi16(__iy))
1868 : _mm_srlv_epi16(_mm_cvtepu8_epi16(__ix),
1869 _mm_cvtepu8_epi16(__iy))));
1870 if constexpr (sizeof(__x) == 16 && __have_avx512bw_vl)
1871 return __intrin_bitcast<_V>(_mm256_cvtepi16_epi8(
1872 is_signed_v<_Up>
1873 ? _mm256_srav_epi16(_mm256_cvtepi8_epi16(__ix),
1874 _mm256_cvtepi8_epi16(__iy))
1875 : _mm256_srlv_epi16(_mm256_cvtepu8_epi16(__ix),
1876 _mm256_cvtepu8_epi16(__iy))));
1877 else if constexpr (sizeof(__x) == 32 && __have_avx512bw)
1878 return __vector_bitcast<_Up>(_mm512_cvtepi16_epi8(
1879 is_signed_v<_Up>
1880 ? _mm512_srav_epi16(_mm512_cvtepi8_epi16(__ix),
1881 _mm512_cvtepi8_epi16(__iy))
1882 : _mm512_srlv_epi16(_mm512_cvtepu8_epi16(__ix),
1883 _mm512_cvtepu8_epi16(__iy))));
1884 else if constexpr (sizeof(__x) == 64 && is_signed_v<_Up>)
1885 return __vector_bitcast<_Up>(_mm512_mask_mov_epi8(
1886 _mm512_srav_epi16(__ix, _mm512_srli_epi16(__iy, 8)),
1887 0x5555'5555'5555'5555ull,
1888 _mm512_srav_epi16(
1889 _mm512_slli_epi16(__ix, 8),
1890 _mm512_maskz_add_epi8(0x5555'5555'5555'5555ull, __iy,
1891 _mm512_set1_epi16(8)))));
1892 else if constexpr (sizeof(__x) == 64 && is_unsigned_v<_Up>)
1893 return __vector_bitcast<_Up>(_mm512_mask_mov_epi8(
1894 _mm512_srlv_epi16(__ix, _mm512_srli_epi16(__iy, 8)),
1895 0x5555'5555'5555'5555ull,
1896 _mm512_srlv_epi16(
1897 _mm512_maskz_mov_epi8(0x5555'5555'5555'5555ull, __ix),
1898 _mm512_maskz_mov_epi8(0x5555'5555'5555'5555ull, __iy))));
1899 /* This has better throughput but higher latency than the impl below
1900 else if constexpr (__have_avx2 && sizeof(__x) == 16 &&
1901 is_unsigned_v<_Up>)
1902 {
1903 const auto __shorts = __to_intrin(_S_bit_shift_right(
1904 __vector_bitcast<_UShort>(_mm256_cvtepu8_epi16(__ix)),
1905 __vector_bitcast<_UShort>(_mm256_cvtepu8_epi16(__iy))));
1906 return __vector_bitcast<_Up>(
1907 _mm_packus_epi16(__lo128(__shorts), __hi128(__shorts)));
1908 }
1909 */
1910 else if constexpr (__have_avx2 && sizeof(__x) > 8)
1911 // the following uses vpsr[al]vd, which requires AVX2
1912 if constexpr (is_signed_v<_Up>)
1913 {
1914 const auto r3 = __vector_bitcast<_UInt>(
1915 (__vector_bitcast<int>(__x)
1916 >> (__vector_bitcast<_UInt>(__y) >> 24)))
1917 & 0xff000000u;
1918 const auto r2
1919 = __vector_bitcast<_UInt>(
1920 ((__vector_bitcast<int>(__x) << 8)
1921 >> ((__vector_bitcast<_UInt>(__y) << 8) >> 24)))
1922 & 0xff000000u;
1923 const auto r1
1924 = __vector_bitcast<_UInt>(
1925 ((__vector_bitcast<int>(__x) << 16)
1926 >> ((__vector_bitcast<_UInt>(__y) << 16) >> 24)))
1927 & 0xff000000u;
1928 const auto r0 = __vector_bitcast<_UInt>(
1929 (__vector_bitcast<int>(__x) << 24)
1930 >> ((__vector_bitcast<_UInt>(__y) << 24) >> 24));
1931 return __vector_bitcast<_Up>(r3 | (r2 >> 8) | (r1 >> 16)
1932 | (r0 >> 24));
1933 }
1934 else
1935 {
1936 const auto r3 = (__vector_bitcast<_UInt>(__x)
1937 >> (__vector_bitcast<_UInt>(__y) >> 24))
1938 & 0xff000000u;
1939 const auto r2
1940 = ((__vector_bitcast<_UInt>(__x) << 8)
1941 >> ((__vector_bitcast<_UInt>(__y) << 8) >> 24))
1942 & 0xff000000u;
1943 const auto r1
1944 = ((__vector_bitcast<_UInt>(__x) << 16)
1945 >> ((__vector_bitcast<_UInt>(__y) << 16) >> 24))
1946 & 0xff000000u;
1947 const auto r0
1948 = (__vector_bitcast<_UInt>(__x) << 24)
1949 >> ((__vector_bitcast<_UInt>(__y) << 24) >> 24);
1950 return __vector_bitcast<_Up>(r3 | (r2 >> 8) | (r1 >> 16)
1951 | (r0 >> 24));
1952 }
1953 else if constexpr (__have_sse4_1
1954 && is_unsigned_v<_Up> && sizeof(__x) > 2)
1955 {
1956 auto __x128 = __vector_bitcast<_Up>(__ix);
1957 auto __mask
1958 = __vector_bitcast<_Up>(__vector_bitcast<_UShort>(__iy) << 5);
1959 auto __x4 = __vector_bitcast<_Up>(
1960 (__vector_bitcast<_UShort>(__x128) >> 4) & _UShort(0xff0f));
1961 __x128 = __vector_bitcast<_Up>(_CommonImplX86::_S_blend_intrin(
1962 __to_intrin(__mask), __to_intrin(__x128), __to_intrin(__x4)));
1963 __mask += __mask;
1964 auto __x2 = __vector_bitcast<_Up>(
1965 (__vector_bitcast<_UShort>(__x128) >> 2) & _UShort(0xff3f));
1966 __x128 = __vector_bitcast<_Up>(_CommonImplX86::_S_blend_intrin(
1967 __to_intrin(__mask), __to_intrin(__x128), __to_intrin(__x2)));
1968 __mask += __mask;
1969 auto __x1 = __vector_bitcast<_Up>(
1970 (__vector_bitcast<_UShort>(__x128) >> 1) & _UShort(0xff7f));
1971 __x128 = __vector_bitcast<_Up>(_CommonImplX86::_S_blend_intrin(
1972 __to_intrin(__mask), __to_intrin(__x128), __to_intrin(__x1)));
1973 return __intrin_bitcast<_V>(
1974 __x128
1975 & ((__vector_bitcast<_Up>(__iy) & char(0xf8))
1976 == 0)); // y > 7 nulls the result
1977 }
1978 else if constexpr (__have_sse4_1
1979 && is_signed_v<_Up> && sizeof(__x) > 2)
1980 {
1981 auto __mask = __vector_bitcast<_UChar>(
1982 __vector_bitcast<_UShort>(__iy) << 5);
1983 auto __maskl = [&]() {
1984 return __to_intrin(__vector_bitcast<_UShort>(__mask) << 8);
1985 };
1986 auto __xh = __vector_bitcast<short>(__ix);
1987 auto __xl = __vector_bitcast<short>(__ix) << 8;
1988 auto __xh4 = __xh >> 4;
1989 auto __xl4 = __xl >> 4;
1990 __xh = __vector_bitcast<short>(_CommonImplX86::_S_blend_intrin(
1991 __to_intrin(__mask), __to_intrin(__xh), __to_intrin(__xh4)));
1992 __xl = __vector_bitcast<short>(
1993 _CommonImplX86::_S_blend_intrin(__maskl(), __to_intrin(__xl),
1994 __to_intrin(__xl4)));
1995 __mask += __mask;
1996 auto __xh2 = __xh >> 2;
1997 auto __xl2 = __xl >> 2;
1998 __xh = __vector_bitcast<short>(_CommonImplX86::_S_blend_intrin(
1999 __to_intrin(__mask), __to_intrin(__xh), __to_intrin(__xh2)));
2000 __xl = __vector_bitcast<short>(
2001 _CommonImplX86::_S_blend_intrin(__maskl(), __to_intrin(__xl),
2002 __to_intrin(__xl2)));
2003 __mask += __mask;
2004 auto __xh1 = __xh >> 1;
2005 auto __xl1 = __xl >> 1;
2006 __xh = __vector_bitcast<short>(_CommonImplX86::_S_blend_intrin(
2007 __to_intrin(__mask), __to_intrin(__xh), __to_intrin(__xh1)));
2008 __xl = __vector_bitcast<short>(
2009 _CommonImplX86::_S_blend_intrin(__maskl(), __to_intrin(__xl),
2010 __to_intrin(__xl1)));
2011 return __intrin_bitcast<_V>(
2012 (__vector_bitcast<_Up>((__xh & short(0xff00)))
2013 | __vector_bitcast<_Up>(__vector_bitcast<_UShort>(__xl)
2014 >> 8))
2015 & ((__vector_bitcast<_Up>(__iy) & char(0xf8))
2016 == 0)); // y > 7 nulls the result
2017 }
2018 else if constexpr (is_unsigned_v<_Up> && sizeof(__x) > 2) // SSE2
2019 {
2020 auto __mask
2021 = __vector_bitcast<_Up>(__vector_bitcast<_UShort>(__y) << 5);
2022 auto __x4 = __vector_bitcast<_Up>(
2023 (__vector_bitcast<_UShort>(__x) >> 4) & _UShort(0xff0f));
2024 __x = __mask > 0x7f ? __x4 : __x;
2025 __mask += __mask;
2026 auto __x2 = __vector_bitcast<_Up>(
2027 (__vector_bitcast<_UShort>(__x) >> 2) & _UShort(0xff3f));
2028 __x = __mask > 0x7f ? __x2 : __x;
2029 __mask += __mask;
2030 auto __x1 = __vector_bitcast<_Up>(
2031 (__vector_bitcast<_UShort>(__x) >> 1) & _UShort(0xff7f));
2032 __x = __mask > 0x7f ? __x1 : __x;
2033 return __x
2034 & ((__y & char(0xf8)) == 0); // y > 7 nulls the result
2035 }
2036 else if constexpr (sizeof(__x) > 2) // signed SSE2
2037 {
2038 static_assert(is_signed_v<_Up>);
2039 auto __maskh = __vector_bitcast<_UShort>(__y) << 5;
2040 auto __maskl = __vector_bitcast<_UShort>(__y) << (5 + 8);
2041 auto __xh = __vector_bitcast<short>(__x);
2042 auto __xl = __vector_bitcast<short>(__x) << 8;
2043 auto __xh4 = __xh >> 4;
2044 auto __xl4 = __xl >> 4;
2045 __xh = __maskh > 0x7fff ? __xh4 : __xh;
2046 __xl = __maskl > 0x7fff ? __xl4 : __xl;
2047 __maskh += __maskh;
2048 __maskl += __maskl;
2049 auto __xh2 = __xh >> 2;
2050 auto __xl2 = __xl >> 2;
2051 __xh = __maskh > 0x7fff ? __xh2 : __xh;
2052 __xl = __maskl > 0x7fff ? __xl2 : __xl;
2053 __maskh += __maskh;
2054 __maskl += __maskl;
2055 auto __xh1 = __xh >> 1;
2056 auto __xl1 = __xl >> 1;
2057 __xh = __maskh > 0x7fff ? __xh1 : __xh;
2058 __xl = __maskl > 0x7fff ? __xl1 : __xl;
2059 __x = __vector_bitcast<_Up>((__xh & short(0xff00)))
2060 | __vector_bitcast<_Up>(__vector_bitcast<_UShort>(__xl)
2061 >> 8);
2062 return __x
2063 & ((__y & char(0xf8)) == 0); // y > 7 nulls the result
2064 }
2065 else
2066 return __x >> __y;
2067 } //}}}
2068 else if constexpr (sizeof(_Up) == 2 && sizeof(__x) >= 4) //{{{
2069 {
2070 [[maybe_unused]] auto __blend_0xaa = [](auto __a, auto __b) {
2071 if constexpr (sizeof(__a) == 16)
2072 return _mm_blend_epi16(__to_intrin(__a), __to_intrin(__b),
2073 0xaa);
2074 else if constexpr (sizeof(__a) == 32)
2075 return _mm256_blend_epi16(__to_intrin(__a), __to_intrin(__b),
2076 0xaa);
2077 else if constexpr (sizeof(__a) == 64)
2078 return _mm512_mask_blend_epi16(0xaaaa'aaaaU, __to_intrin(__a),
2079 __to_intrin(__b));
2080 else
2081 __assert_unreachable<decltype(__a)>();
2082 };
2083 if constexpr (__have_avx512bw_vl && sizeof(_Tp) <= 16)
2084 return __intrin_bitcast<_V>(is_signed_v<_Up>
2085 ? _mm_srav_epi16(__ix, __iy)
2086 : _mm_srlv_epi16(__ix, __iy));
2087 else if constexpr (__have_avx512bw_vl && sizeof(_Tp) == 32)
2088 return __vector_bitcast<_Up>(is_signed_v<_Up>
2089 ? _mm256_srav_epi16(__ix, __iy)
2090 : _mm256_srlv_epi16(__ix, __iy));
2091 else if constexpr (__have_avx512bw && sizeof(_Tp) == 64)
2092 return __vector_bitcast<_Up>(is_signed_v<_Up>
2093 ? _mm512_srav_epi16(__ix, __iy)
2094 : _mm512_srlv_epi16(__ix, __iy));
2095 else if constexpr (__have_avx2 && is_signed_v<_Up>)
2096 return __intrin_bitcast<_V>(
2097 __blend_0xaa(((__vector_bitcast<int>(__ix) << 16)
2098 >> (__vector_bitcast<int>(__iy) & 0xffffu))
2099 >> 16,
2100 __vector_bitcast<int>(__ix)
2101 >> (__vector_bitcast<int>(__iy) >> 16)));
2102 else if constexpr (__have_avx2 && is_unsigned_v<_Up>)
2103 return __intrin_bitcast<_V>(
2104 __blend_0xaa((__vector_bitcast<_UInt>(__ix) & 0xffffu)
2105 >> (__vector_bitcast<_UInt>(__iy) & 0xffffu),
2106 __vector_bitcast<_UInt>(__ix)
2107 >> (__vector_bitcast<_UInt>(__iy) >> 16)));
2108 else if constexpr (__have_sse4_1)
2109 {
2110 auto __mask = __vector_bitcast<_UShort>(__iy);
2111 auto __x128 = __vector_bitcast<_Up>(__ix);
2112 //__mask *= 0x0808;
2113 __mask = (__mask << 3) | (__mask << 11);
2114 // do __x128 = 0 where __y[4] is set
2115 __x128 = __vector_bitcast<_Up>(
2116 _mm_blendv_epi8(__to_intrin(__x128), __m128i(),
2117 __to_intrin(__mask)));
2118 // do __x128 =>> 8 where __y[3] is set
2119 __x128 = __vector_bitcast<_Up>(
2120 _mm_blendv_epi8(__to_intrin(__x128), __to_intrin(__x128 >> 8),
2121 __to_intrin(__mask += __mask)));
2122 // do __x128 =>> 4 where __y[2] is set
2123 __x128 = __vector_bitcast<_Up>(
2124 _mm_blendv_epi8(__to_intrin(__x128), __to_intrin(__x128 >> 4),
2125 __to_intrin(__mask += __mask)));
2126 // do __x128 =>> 2 where __y[1] is set
2127 __x128 = __vector_bitcast<_Up>(
2128 _mm_blendv_epi8(__to_intrin(__x128), __to_intrin(__x128 >> 2),
2129 __to_intrin(__mask += __mask)));
2130 // do __x128 =>> 1 where __y[0] is set
2131 return __intrin_bitcast<_V>(
2132 _mm_blendv_epi8(__to_intrin(__x128), __to_intrin(__x128 >> 1),
2133 __to_intrin(__mask + __mask)));
2134 }
2135 else
2136 {
2137 auto __k = __vector_bitcast<_UShort>(__iy) << 11;
2138 auto __x128 = __vector_bitcast<_Up>(__ix);
2139 auto __mask = [](__vector_type16_t<_UShort> __kk) {
2140 return __vector_bitcast<short>(__kk) < 0;
2141 };
2142 // do __x128 = 0 where __y[4] is set
2143 __x128 = __mask(__k) ? decltype(__x128)() : __x128;
2144 // do __x128 =>> 8 where __y[3] is set
2145 __x128 = __mask(__k += __k) ? __x128 >> 8 : __x128;
2146 // do __x128 =>> 4 where __y[2] is set
2147 __x128 = __mask(__k += __k) ? __x128 >> 4 : __x128;
2148 // do __x128 =>> 2 where __y[1] is set
2149 __x128 = __mask(__k += __k) ? __x128 >> 2 : __x128;
2150 // do __x128 =>> 1 where __y[0] is set
2151 return __intrin_bitcast<_V>(__mask(__k + __k) ? __x128 >> 1
2152 : __x128);
2153 }
2154 } //}}}
2155 else if constexpr (sizeof(_Up) == 4 && !__have_avx2) //{{{
2156 {
2157 if constexpr (is_unsigned_v<_Up>)
2158 {
2159 // x >> y == x * 2^-y == (x * 2^(31-y)) >> 31
2160 const __m128 __factor_f = reinterpret_cast<__m128>(
2161 0x4f00'0000u - (__vector_bitcast<unsigned, 4>(__y) << 23));
2162 const __m128i __factor
2163 = __builtin_constant_p(__factor_f)
2164 ? __to_intrin(
2165 __make_vector<unsigned>(__factor_f[0], __factor_f[1],
2166 __factor_f[2], __factor_f[3]))
2167 : _mm_cvttps_epi32(__factor_f);
2168 const auto __r02
2169 = _mm_srli_epi64(_mm_mul_epu32(__ix, __factor), 31);
2170 const auto __r13 = _mm_mul_epu32(_mm_srli_si128(__ix, 4),
2171 _mm_srli_si128(__factor, 4));
2172 if constexpr (__have_sse4_1)
2173 return __intrin_bitcast<_V>(
2174 _mm_blend_epi16(_mm_slli_epi64(__r13, 1), __r02, 0x33));
2175 else
2176 return __intrin_bitcast<_V>(
2177 __r02 | _mm_slli_si128(_mm_srli_epi64(__r13, 31), 4));
2178 }
2179 else
2180 {
2181 auto __shift = [](auto __a, auto __b) {
2182 if constexpr (is_signed_v<_Up>)
2183 return _mm_sra_epi32(__a, __b);
2184 else
2185 return _mm_srl_epi32(__a, __b);
2186 };
2187 const auto __r0
2188 = __shift(__ix, _mm_unpacklo_epi32(__iy, __m128i()));
2189 const auto __r1 = __shift(__ix, _mm_srli_epi64(__iy, 32));
2190 const auto __r2
2191 = __shift(__ix, _mm_unpackhi_epi32(__iy, __m128i()));
2192 const auto __r3 = __shift(__ix, _mm_srli_si128(__iy, 12));
2193 if constexpr (__have_sse4_1)
2194 return __intrin_bitcast<_V>(
2195 _mm_blend_epi16(_mm_blend_epi16(__r1, __r0, 0x3),
2196 _mm_blend_epi16(__r3, __r2, 0x30), 0xf0));
2197 else
2198 return __intrin_bitcast<_V>(_mm_unpacklo_epi64(
2199 _mm_unpacklo_epi32(__r0, _mm_srli_si128(__r1, 4)),
2200 _mm_unpackhi_epi32(__r2, _mm_srli_si128(__r3, 4))));
2201 }
2202 } //}}}
2203 else
2204 return __x >> __y;
2205 }
2206#endif // _GLIBCXX_SIMD_NO_SHIFT_OPT
2207
2208 // }}}
2209 // compares {{{
2210 // _S_equal_to {{{
2211 template <typename _Tp, size_t _Np>
2212 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
2213 _S_equal_to(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
2214 {
2215 if constexpr (__is_avx512_abi<_Abi>()) // {{{
2216 {
2217 if (__builtin_is_constant_evaluated()
2218 || (__x._M_is_constprop() && __y._M_is_constprop()))
2219 return _MaskImpl::_S_to_bits(
2220 __as_wrapper<_Np>(__x._M_data == __y._M_data));
2221
2222 constexpr auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
2223 [[maybe_unused]] const auto __xi = __to_intrin(__x);
2224 [[maybe_unused]] const auto __yi = __to_intrin(__y);
2225 if constexpr (is_floating_point_v<_Tp>)
2226 {
2227 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
2228 return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_EQ_OQ);
2229 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
2230 return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_EQ_OQ);
2231 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
2232 return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_EQ_OQ);
2233 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
2234 return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_EQ_OQ);
2235 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
2236 return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_EQ_OQ);
2237 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
2238 return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_EQ_OQ);
2239 else
2240 __assert_unreachable<_Tp>();
2241 }
2242 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
2243 return _mm512_mask_cmpeq_epi64_mask(__k1, __xi, __yi);
2244 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
2245 return _mm512_mask_cmpeq_epi32_mask(__k1, __xi, __yi);
2246 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 2)
2247 return _mm512_mask_cmpeq_epi16_mask(__k1, __xi, __yi);
2248 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 1)
2249 return _mm512_mask_cmpeq_epi8_mask(__k1, __xi, __yi);
2250 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
2251 return _mm256_mask_cmpeq_epi64_mask(__k1, __xi, __yi);
2252 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
2253 return _mm256_mask_cmpeq_epi32_mask(__k1, __xi, __yi);
2254 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 2)
2255 return _mm256_mask_cmpeq_epi16_mask(__k1, __xi, __yi);
2256 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 1)
2257 return _mm256_mask_cmpeq_epi8_mask(__k1, __xi, __yi);
2258 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
2259 return _mm_mask_cmpeq_epi64_mask(__k1, __xi, __yi);
2260 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
2261 return _mm_mask_cmpeq_epi32_mask(__k1, __xi, __yi);
2262 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 2)
2263 return _mm_mask_cmpeq_epi16_mask(__k1, __xi, __yi);
2264 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 1)
2265 return _mm_mask_cmpeq_epi8_mask(__k1, __xi, __yi);
2266 else
2267 __assert_unreachable<_Tp>();
2268 } // }}}
2269 else if (__builtin_is_constant_evaluated())
2270 return _Base::_S_equal_to(__x, __y);
2271 else if constexpr (sizeof(__x) == 8) // {{{
2272 {
2273 const auto __r128 = __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__x)
2274 == __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__y);
2275 _MaskMember<_Tp> __r64;
2276 __builtin_memcpy(&__r64._M_data, &__r128, sizeof(__r64));
2277 return __r64;
2278 } // }}}
2279 else
2280 return _Base::_S_equal_to(__x, __y);
2281 }
2282
2283 // }}}
2284 // _S_not_equal_to {{{
2285 template <typename _Tp, size_t _Np>
2286 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
2287 _S_not_equal_to(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
2288 {
2289 if constexpr (__is_avx512_abi<_Abi>()) // {{{
2290 {
2291 if (__builtin_is_constant_evaluated()
2292 || (__x._M_is_constprop() && __y._M_is_constprop()))
2293 return _MaskImpl::_S_to_bits(
2294 __as_wrapper<_Np>(__x._M_data != __y._M_data));
2295
2296 constexpr auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
2297 [[maybe_unused]] const auto __xi = __to_intrin(__x);
2298 [[maybe_unused]] const auto __yi = __to_intrin(__y);
2299 if constexpr (is_floating_point_v<_Tp>)
2300 {
2301 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
2302 return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_NEQ_UQ);
2303 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
2304 return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_NEQ_UQ);
2305 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
2306 return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_NEQ_UQ);
2307 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
2308 return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_NEQ_UQ);
2309 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
2310 return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_NEQ_UQ);
2311 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
2312 return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_NEQ_UQ);
2313 else
2314 __assert_unreachable<_Tp>();
2315 }
2316 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
2317 return ~_mm512_mask_cmpeq_epi64_mask(__k1, __xi, __yi);
2318 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
2319 return ~_mm512_mask_cmpeq_epi32_mask(__k1, __xi, __yi);
2320 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 2)
2321 return ~_mm512_mask_cmpeq_epi16_mask(__k1, __xi, __yi);
2322 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 1)
2323 return ~_mm512_mask_cmpeq_epi8_mask(__k1, __xi, __yi);
2324 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
2325 return ~_mm256_mask_cmpeq_epi64_mask(__k1, __xi, __yi);
2326 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
2327 return ~_mm256_mask_cmpeq_epi32_mask(__k1, __xi, __yi);
2328 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 2)
2329 return ~_mm256_mask_cmpeq_epi16_mask(__k1, __xi, __yi);
2330 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 1)
2331 return ~_mm256_mask_cmpeq_epi8_mask(__k1, __xi, __yi);
2332 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
2333 return ~_mm_mask_cmpeq_epi64_mask(__k1, __xi, __yi);
2334 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
2335 return ~_mm_mask_cmpeq_epi32_mask(__k1, __xi, __yi);
2336 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 2)
2337 return ~_mm_mask_cmpeq_epi16_mask(__k1, __xi, __yi);
2338 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 1)
2339 return ~_mm_mask_cmpeq_epi8_mask(__k1, __xi, __yi);
2340 else
2341 __assert_unreachable<_Tp>();
2342 } // }}}
2343 else if constexpr (!__builtin_is_constant_evaluated() // {{{
2344 && sizeof(__x) == 8)
2345 {
2346 const auto __r128 = __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__x)
2347 != __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__y);
2348 _MaskMember<_Tp> __r64;
2349 __builtin_memcpy(&__r64._M_data, &__r128, sizeof(__r64));
2350 return __r64;
2351 } // }}}
2352 else
2353 return _Base::_S_not_equal_to(__x, __y);
2354 }
2355
2356 // }}}
2357 // _S_less {{{
2358 template <typename _Tp, size_t _Np>
2359 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
2360 _S_less(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
2361 {
2362 if constexpr (__is_avx512_abi<_Abi>()) // {{{
2363 {
2364 if (__builtin_is_constant_evaluated()
2365 || (__x._M_is_constprop() && __y._M_is_constprop()))
2366 return _MaskImpl::_S_to_bits(
2367 __as_wrapper<_Np>(__x._M_data < __y._M_data));
2368
2369 constexpr auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
2370 [[maybe_unused]] const auto __xi = __to_intrin(__x);
2371 [[maybe_unused]] const auto __yi = __to_intrin(__y);
2372 if constexpr (sizeof(__xi) == 64)
2373 {
2374 if constexpr (is_same_v<_Tp, float>)
2375 return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OS);
2376 else if constexpr (is_same_v<_Tp, double>)
2377 return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OS);
2378 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 1)
2379 return _mm512_mask_cmplt_epi8_mask(__k1, __xi, __yi);
2380 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 2)
2381 return _mm512_mask_cmplt_epi16_mask(__k1, __xi, __yi);
2382 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 4)
2383 return _mm512_mask_cmplt_epi32_mask(__k1, __xi, __yi);
2384 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 8)
2385 return _mm512_mask_cmplt_epi64_mask(__k1, __xi, __yi);
2386 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 1)
2387 return _mm512_mask_cmplt_epu8_mask(__k1, __xi, __yi);
2388 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 2)
2389 return _mm512_mask_cmplt_epu16_mask(__k1, __xi, __yi);
2390 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 4)
2391 return _mm512_mask_cmplt_epu32_mask(__k1, __xi, __yi);
2392 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 8)
2393 return _mm512_mask_cmplt_epu64_mask(__k1, __xi, __yi);
2394 else
2395 __assert_unreachable<_Tp>();
2396 }
2397 else if constexpr (sizeof(__xi) == 32)
2398 {
2399 if constexpr (is_same_v<_Tp, float>)
2400 return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OS);
2401 else if constexpr (is_same_v<_Tp, double>)
2402 return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OS);
2403 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 1)
2404 return _mm256_mask_cmplt_epi8_mask(__k1, __xi, __yi);
2405 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 2)
2406 return _mm256_mask_cmplt_epi16_mask(__k1, __xi, __yi);
2407 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 4)
2408 return _mm256_mask_cmplt_epi32_mask(__k1, __xi, __yi);
2409 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 8)
2410 return _mm256_mask_cmplt_epi64_mask(__k1, __xi, __yi);
2411 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 1)
2412 return _mm256_mask_cmplt_epu8_mask(__k1, __xi, __yi);
2413 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 2)
2414 return _mm256_mask_cmplt_epu16_mask(__k1, __xi, __yi);
2415 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 4)
2416 return _mm256_mask_cmplt_epu32_mask(__k1, __xi, __yi);
2417 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 8)
2418 return _mm256_mask_cmplt_epu64_mask(__k1, __xi, __yi);
2419 else
2420 __assert_unreachable<_Tp>();
2421 }
2422 else if constexpr (sizeof(__xi) == 16)
2423 {
2424 if constexpr (is_same_v<_Tp, float>)
2425 return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OS);
2426 else if constexpr (is_same_v<_Tp, double>)
2427 return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OS);
2428 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 1)
2429 return _mm_mask_cmplt_epi8_mask(__k1, __xi, __yi);
2430 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 2)
2431 return _mm_mask_cmplt_epi16_mask(__k1, __xi, __yi);
2432 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 4)
2433 return _mm_mask_cmplt_epi32_mask(__k1, __xi, __yi);
2434 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 8)
2435 return _mm_mask_cmplt_epi64_mask(__k1, __xi, __yi);
2436 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 1)
2437 return _mm_mask_cmplt_epu8_mask(__k1, __xi, __yi);
2438 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 2)
2439 return _mm_mask_cmplt_epu16_mask(__k1, __xi, __yi);
2440 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 4)
2441 return _mm_mask_cmplt_epu32_mask(__k1, __xi, __yi);
2442 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 8)
2443 return _mm_mask_cmplt_epu64_mask(__k1, __xi, __yi);
2444 else
2445 __assert_unreachable<_Tp>();
2446 }
2447 else
2448 __assert_unreachable<_Tp>();
2449 } // }}}
2450 else if constexpr (!__builtin_is_constant_evaluated() // {{{
2451 && sizeof(__x) == 8)
2452 {
2453 const auto __r128 = __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__x)
2454 < __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__y);
2455 _MaskMember<_Tp> __r64;
2456 __builtin_memcpy(&__r64._M_data, &__r128, sizeof(__r64));
2457 return __r64;
2458 } // }}}
2459 else
2460 return _Base::_S_less(__x, __y);
2461 }
2462
2463 // }}}
2464 // _S_less_equal {{{
2465 template <typename _Tp, size_t _Np>
2466 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
2467 _S_less_equal(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
2468 {
2469 if constexpr (__is_avx512_abi<_Abi>()) // {{{
2470 {
2471 if (__builtin_is_constant_evaluated()
2472 || (__x._M_is_constprop() && __y._M_is_constprop()))
2473 return _MaskImpl::_S_to_bits(
2474 __as_wrapper<_Np>(__x._M_data <= __y._M_data));
2475
2476 constexpr auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
2477 [[maybe_unused]] const auto __xi = __to_intrin(__x);
2478 [[maybe_unused]] const auto __yi = __to_intrin(__y);
2479 if constexpr (sizeof(__xi) == 64)
2480 {
2481 if constexpr (is_same_v<_Tp, float>)
2482 return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OS);
2483 else if constexpr (is_same_v<_Tp, double>)
2484 return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OS);
2485 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 1)
2486 return _mm512_mask_cmple_epi8_mask(__k1, __xi, __yi);
2487 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 2)
2488 return _mm512_mask_cmple_epi16_mask(__k1, __xi, __yi);
2489 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 4)
2490 return _mm512_mask_cmple_epi32_mask(__k1, __xi, __yi);
2491 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 8)
2492 return _mm512_mask_cmple_epi64_mask(__k1, __xi, __yi);
2493 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 1)
2494 return _mm512_mask_cmple_epu8_mask(__k1, __xi, __yi);
2495 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 2)
2496 return _mm512_mask_cmple_epu16_mask(__k1, __xi, __yi);
2497 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 4)
2498 return _mm512_mask_cmple_epu32_mask(__k1, __xi, __yi);
2499 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 8)
2500 return _mm512_mask_cmple_epu64_mask(__k1, __xi, __yi);
2501 else
2502 __assert_unreachable<_Tp>();
2503 }
2504 else if constexpr (sizeof(__xi) == 32)
2505 {
2506 if constexpr (is_same_v<_Tp, float>)
2507 return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OS);
2508 else if constexpr (is_same_v<_Tp, double>)
2509 return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OS);
2510 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 1)
2511 return _mm256_mask_cmple_epi8_mask(__k1, __xi, __yi);
2512 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 2)
2513 return _mm256_mask_cmple_epi16_mask(__k1, __xi, __yi);
2514 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 4)
2515 return _mm256_mask_cmple_epi32_mask(__k1, __xi, __yi);
2516 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 8)
2517 return _mm256_mask_cmple_epi64_mask(__k1, __xi, __yi);
2518 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 1)
2519 return _mm256_mask_cmple_epu8_mask(__k1, __xi, __yi);
2520 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 2)
2521 return _mm256_mask_cmple_epu16_mask(__k1, __xi, __yi);
2522 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 4)
2523 return _mm256_mask_cmple_epu32_mask(__k1, __xi, __yi);
2524 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 8)
2525 return _mm256_mask_cmple_epu64_mask(__k1, __xi, __yi);
2526 else
2527 __assert_unreachable<_Tp>();
2528 }
2529 else if constexpr (sizeof(__xi) == 16)
2530 {
2531 if constexpr (is_same_v<_Tp, float>)
2532 return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OS);
2533 else if constexpr (is_same_v<_Tp, double>)
2534 return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OS);
2535 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 1)
2536 return _mm_mask_cmple_epi8_mask(__k1, __xi, __yi);
2537 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 2)
2538 return _mm_mask_cmple_epi16_mask(__k1, __xi, __yi);
2539 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 4)
2540 return _mm_mask_cmple_epi32_mask(__k1, __xi, __yi);
2541 else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 8)
2542 return _mm_mask_cmple_epi64_mask(__k1, __xi, __yi);
2543 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 1)
2544 return _mm_mask_cmple_epu8_mask(__k1, __xi, __yi);
2545 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 2)
2546 return _mm_mask_cmple_epu16_mask(__k1, __xi, __yi);
2547 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 4)
2548 return _mm_mask_cmple_epu32_mask(__k1, __xi, __yi);
2549 else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) == 8)
2550 return _mm_mask_cmple_epu64_mask(__k1, __xi, __yi);
2551 else
2552 __assert_unreachable<_Tp>();
2553 }
2554 else
2555 __assert_unreachable<_Tp>();
2556 } // }}}
2557 else if constexpr (!__builtin_is_constant_evaluated() // {{{
2558 && sizeof(__x) == 8)
2559 {
2560 const auto __r128 = __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__x)
2561 <= __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__y);
2562 _MaskMember<_Tp> __r64;
2563 __builtin_memcpy(&__r64._M_data, &__r128, sizeof(__r64));
2564 return __r64;
2565 } // }}}
2566 else
2567 return _Base::_S_less_equal(__x, __y);
2568 }
2569
2570 // }}} }}}
2571 // negation {{{
2572 template <typename _Tp, size_t _Np>
2573 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
2574 _S_negate(_SimdWrapper<_Tp, _Np> __x) noexcept
2575 {
2576 if constexpr (__is_avx512_abi<_Abi>())
2577 return _S_equal_to(__x, _SimdWrapper<_Tp, _Np>());
2578 else
2579 return _Base::_S_negate(__x);
2580 }
2581
2582 // }}}
2583 // math {{{
2584 using _Base::_S_abs;
2585
2586 // _S_sqrt {{{
2587 template <typename _Tp, size_t _Np>
2588 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
2589 _S_sqrt(_SimdWrapper<_Tp, _Np> __x)
2590 {
2591 if constexpr (__is_sse_ps<_Tp, _Np>())
2592 return __auto_bitcast(_mm_sqrt_ps(__to_intrin(__x)));
2593 else if constexpr (__is_sse_pd<_Tp, _Np>())
2594 return _mm_sqrt_pd(__x);
2595 else if constexpr (__is_avx_ps<_Tp, _Np>())
2596 return _mm256_sqrt_ps(__x);
2597 else if constexpr (__is_avx_pd<_Tp, _Np>())
2598 return _mm256_sqrt_pd(__x);
2599 else if constexpr (__is_avx512_ps<_Tp, _Np>())
2600 return _mm512_sqrt_ps(__x);
2601 else if constexpr (__is_avx512_pd<_Tp, _Np>())
2602 return _mm512_sqrt_pd(__x);
2603 else
2604 __assert_unreachable<_Tp>();
2605 }
2606
2607 // }}}
2608 // _S_ldexp {{{
2609 template <typename _Tp, size_t _Np>
2610 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
2611 _S_ldexp(_SimdWrapper<_Tp, _Np> __x,
2612 __fixed_size_storage_t<int, _Np> __exp)
2613 {
2614 if constexpr (__is_avx512_abi<_Abi>())
2615 {
2616 const auto __xi = __to_intrin(__x);
2617 constexpr _SimdConverter<int, simd_abi::fixed_size<_Np>, _Tp, _Abi>
2618 __cvt;
2619 const auto __expi = __to_intrin(__cvt(__exp));
2620 constexpr auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
2621 if constexpr (sizeof(__xi) == 16)
2622 {
2623 if constexpr (sizeof(_Tp) == 8)
2624 return _mm_maskz_scalef_pd(__k1, __xi, __expi);
2625 else
2626 return _mm_maskz_scalef_ps(__k1, __xi, __expi);
2627 }
2628 else if constexpr (sizeof(__xi) == 32)
2629 {
2630 if constexpr (sizeof(_Tp) == 8)
2631 return _mm256_maskz_scalef_pd(__k1, __xi, __expi);
2632 else
2633 return _mm256_maskz_scalef_ps(__k1, __xi, __expi);
2634 }
2635 else
2636 {
2637 static_assert(sizeof(__xi) == 64);
2638 if constexpr (sizeof(_Tp) == 8)
2639 return _mm512_maskz_scalef_pd(__k1, __xi, __expi);
2640 else
2641 return _mm512_maskz_scalef_ps(__k1, __xi, __expi);
2642 }
2643 }
2644 else
2645 return _Base::_S_ldexp(__x, __exp);
2646 }
2647
2648 // }}}
2649 // _S_trunc {{{
2650 template <typename _Tp, size_t _Np>
2651 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
2652 _S_trunc(_SimdWrapper<_Tp, _Np> __x)
2653 {
2654 if constexpr (__is_avx512_ps<_Tp, _Np>())
2655 return _mm512_roundscale_ps(__x, 0x0b);
2656 else if constexpr (__is_avx512_pd<_Tp, _Np>())
2657 return _mm512_roundscale_pd(__x, 0x0b);
2658 else if constexpr (__is_avx_ps<_Tp, _Np>())
2659 return _mm256_round_ps(__x, 0x3);
2660 else if constexpr (__is_avx_pd<_Tp, _Np>())
2661 return _mm256_round_pd(__x, 0x3);
2662 else if constexpr (__have_sse4_1 && __is_sse_ps<_Tp, _Np>())
2663 return __auto_bitcast(_mm_round_ps(__to_intrin(__x), 0x3));
2664 else if constexpr (__have_sse4_1 && __is_sse_pd<_Tp, _Np>())
2665 return _mm_round_pd(__x, 0x3);
2666 else if constexpr (__is_sse_ps<_Tp, _Np>())
2667 {
2668 auto __truncated
2669 = _mm_cvtepi32_ps(_mm_cvttps_epi32(__to_intrin(__x)));
2670 const auto __no_fractional_values
2671 = __vector_bitcast<int>(__vector_bitcast<_UInt>(__to_intrin(__x))
2672 & 0x7f800000u)
2673 < 0x4b000000; // the exponent is so large that no mantissa bits
2674 // signify fractional values (0x3f8 + 23*8 =
2675 // 0x4b0)
2676 return __no_fractional_values ? __truncated : __to_intrin(__x);
2677 }
2678 else
2679 return _Base::_S_trunc(__x);
2680 }
2681
2682 // }}}
2683 // _S_round {{{
2684 template <typename _Tp, size_t _Np>
2685 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
2686 _S_round(_SimdWrapper<_Tp, _Np> __x)
2687 {
2688 // Note that _MM_FROUND_TO_NEAREST_INT rounds ties to even, not away
2689 // from zero as required by std::round. Therefore this function is more
2690 // complicated.
2691 using _V = __vector_type_t<_Tp, _Np>;
2692 _V __truncated;
2693 if constexpr (__is_avx512_ps<_Tp, _Np>())
2694 __truncated = _mm512_roundscale_ps(__x._M_data, 0x0b);
2695 else if constexpr (__is_avx512_pd<_Tp, _Np>())
2696 __truncated = _mm512_roundscale_pd(__x._M_data, 0x0b);
2697 else if constexpr (__is_avx_ps<_Tp, _Np>())
2698 __truncated = _mm256_round_ps(__x._M_data,
2699 _MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC);
2700 else if constexpr (__is_avx_pd<_Tp, _Np>())
2701 __truncated = _mm256_round_pd(__x._M_data,
2702 _MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC);
2703 else if constexpr (__have_sse4_1 && __is_sse_ps<_Tp, _Np>())
2704 __truncated = __auto_bitcast(
2705 _mm_round_ps(__to_intrin(__x),
2706 _MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC));
2707 else if constexpr (__have_sse4_1 && __is_sse_pd<_Tp, _Np>())
2708 __truncated
2709 = _mm_round_pd(__x._M_data, _MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC);
2710 else if constexpr (__is_sse_ps<_Tp, _Np>())
2711 __truncated = __auto_bitcast(
2712 _mm_cvtepi32_ps(_mm_cvttps_epi32(__to_intrin(__x))));
2713 else
2714 return _Base::_S_round(__x);
2715
2716 // x < 0 => truncated <= 0 && truncated >= x => x - truncated <= 0
2717 // x > 0 => truncated >= 0 && truncated <= x => x - truncated >= 0
2718
2719 const _V __rounded
2720 = __truncated
2721 + (__and(_S_absmask<_V>, __x._M_data - __truncated) >= _Tp(.5)
2722 ? __or(__and(_S_signmask<_V>, __x._M_data), _V() + 1)
2723 : _V());
2724 if constexpr (__have_sse4_1)
2725 return __rounded;
2726 else // adjust for missing range in cvttps_epi32
2727 return __and(_S_absmask<_V>, __x._M_data) < 0x1p23f ? __rounded
2728 : __x._M_data;
2729 }
2730
2731 // }}}
2732 // _S_nearbyint {{{
2733 template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
2734 _GLIBCXX_SIMD_INTRINSIC static _Tp _S_nearbyint(_Tp __x) noexcept
2735 {
2736 if constexpr (_TVT::template _S_is<float, 16>)
2737 return _mm512_roundscale_ps(__x, 0x0c);
2738 else if constexpr (_TVT::template _S_is<double, 8>)
2739 return _mm512_roundscale_pd(__x, 0x0c);
2740 else if constexpr (_TVT::template _S_is<float, 8>)
2741 return _mm256_round_ps(__x,
2742 _MM_FROUND_CUR_DIRECTION | _MM_FROUND_NO_EXC);
2743 else if constexpr (_TVT::template _S_is<double, 4>)
2744 return _mm256_round_pd(__x,
2745 _MM_FROUND_CUR_DIRECTION | _MM_FROUND_NO_EXC);
2746 else if constexpr (__have_sse4_1 && _TVT::template _S_is<float, 4>)
2747 return _mm_round_ps(__x,
2748 _MM_FROUND_CUR_DIRECTION | _MM_FROUND_NO_EXC);
2749 else if constexpr (__have_sse4_1 && _TVT::template _S_is<double, 2>)
2750 return _mm_round_pd(__x,
2751 _MM_FROUND_CUR_DIRECTION | _MM_FROUND_NO_EXC);
2752 else
2753 return _Base::_S_nearbyint(__x);
2754 }
2755
2756 // }}}
2757 // _S_rint {{{
2758 template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
2759 _GLIBCXX_SIMD_INTRINSIC static _Tp _S_rint(_Tp __x) noexcept
2760 {
2761 if constexpr (_TVT::template _S_is<float, 16>)
2762 return _mm512_roundscale_ps(__x, 0x04);
2763 else if constexpr (_TVT::template _S_is<double, 8>)
2764 return _mm512_roundscale_pd(__x, 0x04);
2765 else if constexpr (_TVT::template _S_is<float, 8>)
2766 return _mm256_round_ps(__x, _MM_FROUND_CUR_DIRECTION);
2767 else if constexpr (_TVT::template _S_is<double, 4>)
2768 return _mm256_round_pd(__x, _MM_FROUND_CUR_DIRECTION);
2769 else if constexpr (__have_sse4_1 && _TVT::template _S_is<float, 4>)
2770 return _mm_round_ps(__x, _MM_FROUND_CUR_DIRECTION);
2771 else if constexpr (__have_sse4_1 && _TVT::template _S_is<double, 2>)
2772 return _mm_round_pd(__x, _MM_FROUND_CUR_DIRECTION);
2773 else
2774 return _Base::_S_rint(__x);
2775 }
2776
2777 // }}}
2778 // _S_floor {{{
2779 template <typename _Tp, size_t _Np>
2780 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
2781 _S_floor(_SimdWrapper<_Tp, _Np> __x)
2782 {
2783 if constexpr (__is_avx512_ps<_Tp, _Np>())
2784 return _mm512_roundscale_ps(__x, 0x09);
2785 else if constexpr (__is_avx512_pd<_Tp, _Np>())
2786 return _mm512_roundscale_pd(__x, 0x09);
2787 else if constexpr (__is_avx_ps<_Tp, _Np>())
2788 return _mm256_round_ps(__x, 0x1);
2789 else if constexpr (__is_avx_pd<_Tp, _Np>())
2790 return _mm256_round_pd(__x, 0x1);
2791 else if constexpr (__have_sse4_1 && __is_sse_ps<_Tp, _Np>())
2792 return __auto_bitcast(_mm_floor_ps(__to_intrin(__x)));
2793 else if constexpr (__have_sse4_1 && __is_sse_pd<_Tp, _Np>())
2794 return _mm_floor_pd(__x);
2795 else
2796 return _Base::_S_floor(__x);
2797 }
2798
2799 // }}}
2800 // _S_ceil {{{
2801 template <typename _Tp, size_t _Np>
2802 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
2803 _S_ceil(_SimdWrapper<_Tp, _Np> __x)
2804 {
2805 if constexpr (__is_avx512_ps<_Tp, _Np>())
2806 return _mm512_roundscale_ps(__x, 0x0a);
2807 else if constexpr (__is_avx512_pd<_Tp, _Np>())
2808 return _mm512_roundscale_pd(__x, 0x0a);
2809 else if constexpr (__is_avx_ps<_Tp, _Np>())
2810 return _mm256_round_ps(__x, 0x2);
2811 else if constexpr (__is_avx_pd<_Tp, _Np>())
2812 return _mm256_round_pd(__x, 0x2);
2813 else if constexpr (__have_sse4_1 && __is_sse_ps<_Tp, _Np>())
2814 return __auto_bitcast(_mm_ceil_ps(__to_intrin(__x)));
2815 else if constexpr (__have_sse4_1 && __is_sse_pd<_Tp, _Np>())
2816 return _mm_ceil_pd(__x);
2817 else
2818 return _Base::_S_ceil(__x);
2819 }
2820
2821 // }}}
2822 // _S_signbit {{{
2823 template <typename _Tp, size_t _Np>
2824 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
2825 _S_signbit(_SimdWrapper<_Tp, _Np> __x)
2826 {
2827 if constexpr (__is_avx512_abi<_Abi>() && __have_avx512dq)
2828 {
2829 if constexpr (sizeof(__x) == 64 && sizeof(_Tp) == 4)
2830 return _mm512_movepi32_mask(
2831 __intrin_bitcast<__m512i>(__x._M_data));
2832 else if constexpr (sizeof(__x) == 64 && sizeof(_Tp) == 8)
2833 return _mm512_movepi64_mask(
2834 __intrin_bitcast<__m512i>(__x._M_data));
2835 else if constexpr (sizeof(__x) == 32 && sizeof(_Tp) == 4)
2836 return _mm256_movepi32_mask(
2837 __intrin_bitcast<__m256i>(__x._M_data));
2838 else if constexpr (sizeof(__x) == 32 && sizeof(_Tp) == 8)
2839 return _mm256_movepi64_mask(
2840 __intrin_bitcast<__m256i>(__x._M_data));
2841 else if constexpr (sizeof(__x) <= 16 && sizeof(_Tp) == 4)
2842 return _mm_movepi32_mask(__intrin_bitcast<__m128i>(__x._M_data));
2843 else if constexpr (sizeof(__x) <= 16 && sizeof(_Tp) == 8)
2844 return _mm_movepi64_mask(__intrin_bitcast<__m128i>(__x._M_data));
2845 }
2846 else if constexpr (__is_avx512_abi<_Abi>())
2847 {
2848 const auto __xi = __to_intrin(__x);
2849 [[maybe_unused]] constexpr auto __k1
2850 = _Abi::template _S_implicit_mask_intrin<_Tp>();
2851 if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
2852 return _mm_movemask_ps(__xi);
2853 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
2854 return _mm_movemask_pd(__xi);
2855 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
2856 return _mm256_movemask_ps(__xi);
2857 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
2858 return _mm256_movemask_pd(__xi);
2859 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
2860 return _mm512_mask_cmplt_epi32_mask(
2861 __k1, __intrin_bitcast<__m512i>(__xi), __m512i());
2862 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
2863 return _mm512_mask_cmplt_epi64_mask(
2864 __k1, __intrin_bitcast<__m512i>(__xi), __m512i());
2865 else
2866 __assert_unreachable<_Tp>();
2867 }
2868 else
2869 return _Base::_S_signbit(__x);
2870 /*{
2871 using _I = __int_for_sizeof_t<_Tp>;
2872 if constexpr (sizeof(__x) == 64)
2873 return _S_less(__vector_bitcast<_I>(__x), _I());
2874 else
2875 {
2876 const auto __xx = __vector_bitcast<_I>(__x._M_data);
2877 [[maybe_unused]] constexpr _I __signmask = __finite_min_v<_I>;
2878 if constexpr ((sizeof(_Tp) == 4 &&
2879 (__have_avx2 || sizeof(__x) == 16)) ||
2880 __have_avx512vl)
2881 {
2882 return __vector_bitcast<_Tp>(__xx >> __digits_v<_I>);
2883 }
2884 else if constexpr ((__have_avx2 ||
2885 (__have_ssse3 && sizeof(__x) == 16)))
2886 {
2887 return __vector_bitcast<_Tp>((__xx & __signmask) ==
2888 __signmask);
2889 }
2890 else
2891 { // SSE2/3 or AVX (w/o AVX2)
2892 constexpr auto __one = __vector_broadcast<_Np, _Tp>(1);
2893 return __vector_bitcast<_Tp>(
2894 __vector_bitcast<_Tp>(
2895 (__xx & __signmask) |
2896 __vector_bitcast<_I>(__one)) // -1 or 1
2897 != __one);
2898 }
2899 }
2900 }*/
2901 }
2902
2903 // }}}
2904 // _S_isnonzerovalue_mask {{{
2905 // (isnormal | is subnormal == !isinf & !isnan & !is zero)
2906 template <typename _Tp>
2907 _GLIBCXX_SIMD_INTRINSIC static auto _S_isnonzerovalue_mask(_Tp __x)
2908 {
2909 using _Traits = _VectorTraits<_Tp>;
2910 if constexpr (__have_avx512dq_vl)
2911 {
2912 if constexpr (_Traits::template _S_is<
2913 float, 2> || _Traits::template _S_is<float, 4>)
2914 return _knot_mask8(_mm_fpclass_ps_mask(__to_intrin(__x), 0x9f));
2915 else if constexpr (_Traits::template _S_is<float, 8>)
2916 return _knot_mask8(_mm256_fpclass_ps_mask(__x, 0x9f));
2917 else if constexpr (_Traits::template _S_is<float, 16>)
2918 return _knot_mask16(_mm512_fpclass_ps_mask(__x, 0x9f));
2919 else if constexpr (_Traits::template _S_is<double, 2>)
2920 return _knot_mask8(_mm_fpclass_pd_mask(__x, 0x9f));
2921 else if constexpr (_Traits::template _S_is<double, 4>)
2922 return _knot_mask8(_mm256_fpclass_pd_mask(__x, 0x9f));
2923 else if constexpr (_Traits::template _S_is<double, 8>)
2924 return _knot_mask8(_mm512_fpclass_pd_mask(__x, 0x9f));
2925 else
2926 __assert_unreachable<_Tp>();
2927 }
2928 else
2929 {
2930 using _Up = typename _Traits::value_type;
2931 constexpr size_t _Np = _Traits::_S_full_size;
2932 const auto __a = __x * __infinity_v<_Up>; // NaN if __x == 0
2933 const auto __b = __x * _Up(); // NaN if __x == inf
2934 if constexpr (__have_avx512vl && __is_sse_ps<_Up, _Np>())
2935 return _mm_cmp_ps_mask(__to_intrin(__a), __to_intrin(__b),
2936 _CMP_ORD_Q);
2937 else if constexpr (__have_avx512f && __is_sse_ps<_Up, _Np>())
2938 return __mmask8(0xf
2939 & _mm512_cmp_ps_mask(__auto_bitcast(__a),
2940 __auto_bitcast(__b),
2941 _CMP_ORD_Q));
2942 else if constexpr (__have_avx512vl && __is_sse_pd<_Up, _Np>())
2943 return _mm_cmp_pd_mask(__a, __b, _CMP_ORD_Q);
2944 else if constexpr (__have_avx512f && __is_sse_pd<_Up, _Np>())
2945 return __mmask8(0x3
2946 & _mm512_cmp_pd_mask(__auto_bitcast(__a),
2947 __auto_bitcast(__b),
2948 _CMP_ORD_Q));
2949 else if constexpr (__have_avx512vl && __is_avx_ps<_Up, _Np>())
2950 return _mm256_cmp_ps_mask(__a, __b, _CMP_ORD_Q);
2951 else if constexpr (__have_avx512f && __is_avx_ps<_Up, _Np>())
2952 return __mmask8(_mm512_cmp_ps_mask(__auto_bitcast(__a),
2953 __auto_bitcast(__b),
2954 _CMP_ORD_Q));
2955 else if constexpr (__have_avx512vl && __is_avx_pd<_Up, _Np>())
2956 return _mm256_cmp_pd_mask(__a, __b, _CMP_ORD_Q);
2957 else if constexpr (__have_avx512f && __is_avx_pd<_Up, _Np>())
2958 return __mmask8(0xf
2959 & _mm512_cmp_pd_mask(__auto_bitcast(__a),
2960 __auto_bitcast(__b),
2961 _CMP_ORD_Q));
2962 else if constexpr (__is_avx512_ps<_Up, _Np>())
2963 return _mm512_cmp_ps_mask(__a, __b, _CMP_ORD_Q);
2964 else if constexpr (__is_avx512_pd<_Up, _Np>())
2965 return _mm512_cmp_pd_mask(__a, __b, _CMP_ORD_Q);
2966 else
2967 __assert_unreachable<_Tp>();
2968 }
2969 }
2970
2971 // }}}
2972 // _S_isfinite {{{
2973 template <typename _Tp, size_t _Np>
2974 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
2975 _S_isfinite(_SimdWrapper<_Tp, _Np> __x)
2976 {
2977 static_assert(is_floating_point_v<_Tp>);
2978#if !__FINITE_MATH_ONLY__
2979 if constexpr (__is_avx512_abi<_Abi>() && __have_avx512dq)
2980 {
2981 const auto __xi = __to_intrin(__x);
2982 constexpr auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
2983 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
2984 return __k1 ^ _mm512_mask_fpclass_ps_mask(__k1, __xi, 0x99);
2985 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
2986 return __k1 ^ _mm512_mask_fpclass_pd_mask(__k1, __xi, 0x99);
2987 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
2988 return __k1 ^ _mm256_mask_fpclass_ps_mask(__k1, __xi, 0x99);
2989 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
2990 return __k1 ^ _mm256_mask_fpclass_pd_mask(__k1, __xi, 0x99);
2991 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
2992 return __k1 ^ _mm_mask_fpclass_ps_mask(__k1, __xi, 0x99);
2993 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
2994 return __k1 ^ _mm_mask_fpclass_pd_mask(__k1, __xi, 0x99);
2995 }
2996 else if constexpr (__is_avx512_abi<_Abi>())
2997 {
2998 // if all exponent bits are set, __x is either inf or NaN
2999 using _I = __int_for_sizeof_t<_Tp>;
3000 const auto __inf = __vector_bitcast<_I>(
3001 __vector_broadcast<_Np>(__infinity_v<_Tp>));
3002 return _S_less<_I, _Np>(__vector_bitcast<_I>(__x) & __inf, __inf);
3003 }
3004 else
3005#endif
3006 return _Base::_S_isfinite(__x);
3007 }
3008
3009 // }}}
3010 // _S_isinf {{{
3011 template <typename _Tp, size_t _Np>
3012 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
3013 _S_isinf(_SimdWrapper<_Tp, _Np> __x)
3014 {
3015#if !__FINITE_MATH_ONLY__
3016 if constexpr (__is_avx512_abi<_Abi>() && __have_avx512dq)
3017 {
3018 const auto __xi = __to_intrin(__x);
3019 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
3020 return _mm512_fpclass_ps_mask(__xi, 0x18);
3021 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
3022 return _mm512_fpclass_pd_mask(__xi, 0x18);
3023 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3024 return _mm256_fpclass_ps_mask(__xi, 0x18);
3025 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3026 return _mm256_fpclass_pd_mask(__xi, 0x18);
3027 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3028 return _mm_fpclass_ps_mask(__xi, 0x18);
3029 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3030 return _mm_fpclass_pd_mask(__xi, 0x18);
3031 else
3032 __assert_unreachable<_Tp>();
3033 }
3034 else if constexpr (__have_avx512dq_vl)
3035 {
3036 if constexpr (__is_sse_pd<_Tp, _Np>())
3037 return _mm_movm_epi64(_mm_fpclass_pd_mask(__x, 0x18));
3038 else if constexpr (__is_avx_pd<_Tp, _Np>())
3039 return _mm256_movm_epi64(_mm256_fpclass_pd_mask(__x, 0x18));
3040 else if constexpr (__is_sse_ps<_Tp, _Np>())
3041 return _mm_movm_epi32(
3042 _mm_fpclass_ps_mask(__to_intrin(__x), 0x18));
3043 else if constexpr (__is_avx_ps<_Tp, _Np>())
3044 return _mm256_movm_epi32(_mm256_fpclass_ps_mask(__x, 0x18));
3045 else
3046 __assert_unreachable<_Tp>();
3047 }
3048 else
3049#endif
3050 return _Base::_S_isinf(__x);
3051 }
3052
3053 // }}}
3054 // _S_isnormal {{{
3055 template <typename _Tp, size_t _Np>
3056 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
3057 _S_isnormal(_SimdWrapper<_Tp, _Np> __x)
3058 {
3059#if __FINITE_MATH_ONLY__
3060 [[maybe_unused]] constexpr int __mode = 0x26;
3061#else
3062 [[maybe_unused]] constexpr int __mode = 0xbf;
3063#endif
3064 if constexpr (__is_avx512_abi<_Abi>() && __have_avx512dq)
3065 {
3066 const auto __xi = __to_intrin(__x);
3067 const auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
3068 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
3069 return __k1 ^ _mm512_mask_fpclass_ps_mask(__k1, __xi, __mode);
3070 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
3071 return __k1 ^ _mm512_mask_fpclass_pd_mask(__k1, __xi, __mode);
3072 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3073 return __k1 ^ _mm256_mask_fpclass_ps_mask(__k1, __xi, __mode);
3074 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3075 return __k1 ^ _mm256_mask_fpclass_pd_mask(__k1, __xi, __mode);
3076 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3077 return __k1 ^ _mm_mask_fpclass_ps_mask(__k1, __xi, __mode);
3078 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3079 return __k1 ^ _mm_mask_fpclass_pd_mask(__k1, __xi, __mode);
3080 else
3081 __assert_unreachable<_Tp>();
3082 }
3083 else if constexpr (__have_avx512dq)
3084 {
3085 if constexpr (__have_avx512vl && __is_sse_ps<_Tp, _Np>())
3086 return _mm_movm_epi32(
3087 _knot_mask8(_mm_fpclass_ps_mask(__to_intrin(__x), __mode)));
3088 else if constexpr (__have_avx512vl && __is_avx_ps<_Tp, _Np>())
3089 return _mm256_movm_epi32(
3090 _knot_mask8(_mm256_fpclass_ps_mask(__x, __mode)));
3091 else if constexpr (__is_avx512_ps<_Tp, _Np>())
3092 return _knot_mask16(_mm512_fpclass_ps_mask(__x, __mode));
3093 else if constexpr (__have_avx512vl && __is_sse_pd<_Tp, _Np>())
3094 return _mm_movm_epi64(
3095 _knot_mask8(_mm_fpclass_pd_mask(__x, __mode)));
3096 else if constexpr (__have_avx512vl && __is_avx_pd<_Tp, _Np>())
3097 return _mm256_movm_epi64(
3098 _knot_mask8(_mm256_fpclass_pd_mask(__x, __mode)));
3099 else if constexpr (__is_avx512_pd<_Tp, _Np>())
3100 return _knot_mask8(_mm512_fpclass_pd_mask(__x, __mode));
3101 else
3102 __assert_unreachable<_Tp>();
3103 }
3104 else if constexpr (__is_avx512_abi<_Abi>())
3105 {
3106 using _I = __int_for_sizeof_t<_Tp>;
3107 const auto absn = __vector_bitcast<_I>(_S_abs(__x));
3108 const auto minn = __vector_bitcast<_I>(
3109 __vector_broadcast<_Np>(__norm_min_v<_Tp>));
3110#if __FINITE_MATH_ONLY__
3111 return _S_less_equal<_I, _Np>(minn, absn);
3112#else
3113 const auto infn
3114 = __vector_bitcast<_I>(__vector_broadcast<_Np>(__infinity_v<_Tp>));
3115 return __and(_S_less_equal<_I, _Np>(minn, absn),
3116 _S_less<_I, _Np>(absn, infn));
3117#endif
3118 }
3119 else
3120 return _Base::_S_isnormal(__x);
3121 }
3122
3123 // }}}
3124 // _S_isnan {{{
3125 template <typename _Tp, size_t _Np>
3126 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
3127 _S_isnan(_SimdWrapper<_Tp, _Np> __x)
3128 { return _S_isunordered(__x, __x); }
3129
3130 // }}}
3131 // _S_isunordered {{{
3132 template <typename _Tp, size_t _Np>
3133 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
3134 _S_isunordered([[maybe_unused]] _SimdWrapper<_Tp, _Np> __x,
3135 [[maybe_unused]] _SimdWrapper<_Tp, _Np> __y)
3136 {
3137#if __FINITE_MATH_ONLY__
3138 return {}; // false
3139#else
3140 const auto __xi = __to_intrin(__x);
3141 const auto __yi = __to_intrin(__y);
3142 if constexpr (__is_avx512_abi<_Abi>())
3143 {
3144 constexpr auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
3145 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
3146 return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_UNORD_Q);
3147 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
3148 return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_UNORD_Q);
3149 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3150 return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_UNORD_Q);
3151 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3152 return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_UNORD_Q);
3153 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3154 return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_UNORD_Q);
3155 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3156 return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_UNORD_Q);
3157 }
3158 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3159 return __to_masktype(_mm256_cmp_ps(__xi, __yi, _CMP_UNORD_Q));
3160 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3161 return __to_masktype(_mm256_cmp_pd(__xi, __yi, _CMP_UNORD_Q));
3162 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3163 return __auto_bitcast(_mm_cmpunord_ps(__xi, __yi));
3164 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3165 return __to_masktype(_mm_cmpunord_pd(__xi, __yi));
3166 else
3167 __assert_unreachable<_Tp>();
3168#endif
3169 }
3170
3171 // }}}
3172 // _S_isgreater {{{
3173 template <typename _Tp, size_t _Np>
3174 static constexpr _MaskMember<_Tp> _S_isgreater(_SimdWrapper<_Tp, _Np> __x,
3175 _SimdWrapper<_Tp, _Np> __y)
3176 {
3177 const auto __xi = __to_intrin(__x);
3178 const auto __yi = __to_intrin(__y);
3179 if constexpr (__is_avx512_abi<_Abi>())
3180 {
3181 const auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
3182 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
3183 return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_GT_OQ);
3184 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
3185 return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_GT_OQ);
3186 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3187 return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_GT_OQ);
3188 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3189 return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_GT_OQ);
3190 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3191 return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_GT_OQ);
3192 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3193 return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_GT_OQ);
3194 else
3195 __assert_unreachable<_Tp>();
3196 }
3197 else if constexpr (__have_avx)
3198 {
3199 if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3200 return __to_masktype(_mm256_cmp_ps(__xi, __yi, _CMP_GT_OQ));
3201 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3202 return __to_masktype(_mm256_cmp_pd(__xi, __yi, _CMP_GT_OQ));
3203 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3204 return __auto_bitcast(_mm_cmp_ps(__xi, __yi, _CMP_GT_OQ));
3205 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3206 return __to_masktype(_mm_cmp_pd(__xi, __yi, _CMP_GT_OQ));
3207 else
3208 __assert_unreachable<_Tp>();
3209 }
3210 else if constexpr (__have_sse2 && sizeof(__xi) == 16
3211 && sizeof(_Tp) == 4)
3212 {
3213 const auto __xn = __vector_bitcast<int>(__xi);
3214 const auto __yn = __vector_bitcast<int>(__yi);
3215 const auto __xp = __xn < 0 ? -(__xn & 0x7fff'ffff) : __xn;
3216 const auto __yp = __yn < 0 ? -(__yn & 0x7fff'ffff) : __yn;
3217 return __auto_bitcast(
3218 __and(__to_masktype(_mm_cmpord_ps(__xi, __yi)), __xp > __yp));
3219 }
3220 else if constexpr (__have_sse2 && sizeof(__xi) == 16
3221 && sizeof(_Tp) == 8)
3222 return __vector_type_t<__int_with_sizeof_t<8>, 2>{
3223 -_mm_ucomigt_sd(__xi, __yi),
3224 -_mm_ucomigt_sd(_mm_unpackhi_pd(__xi, __xi),
3225 _mm_unpackhi_pd(__yi, __yi))};
3226 else
3227 return _Base::_S_isgreater(__x, __y);
3228 }
3229
3230 // }}}
3231 // _S_isgreaterequal {{{
3232 template <typename _Tp, size_t _Np>
3233 static constexpr _MaskMember<_Tp>
3234 _S_isgreaterequal(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
3235 {
3236 const auto __xi = __to_intrin(__x);
3237 const auto __yi = __to_intrin(__y);
3238 if constexpr (__is_avx512_abi<_Abi>())
3239 {
3240 const auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
3241 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
3242 return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_GE_OQ);
3243 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
3244 return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_GE_OQ);
3245 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3246 return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_GE_OQ);
3247 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3248 return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_GE_OQ);
3249 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3250 return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_GE_OQ);
3251 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3252 return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_GE_OQ);
3253 else
3254 __assert_unreachable<_Tp>();
3255 }
3256 else if constexpr (__have_avx)
3257 {
3258 if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3259 return __to_masktype(_mm256_cmp_ps(__xi, __yi, _CMP_GE_OQ));
3260 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3261 return __to_masktype(_mm256_cmp_pd(__xi, __yi, _CMP_GE_OQ));
3262 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3263 return __auto_bitcast(_mm_cmp_ps(__xi, __yi, _CMP_GE_OQ));
3264 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3265 return __to_masktype(_mm_cmp_pd(__xi, __yi, _CMP_GE_OQ));
3266 else
3267 __assert_unreachable<_Tp>();
3268 }
3269 else if constexpr (__have_sse2 && sizeof(__xi) == 16
3270 && sizeof(_Tp) == 4)
3271 {
3272 const auto __xn = __vector_bitcast<int>(__xi);
3273 const auto __yn = __vector_bitcast<int>(__yi);
3274 const auto __xp = __xn < 0 ? -(__xn & 0x7fff'ffff) : __xn;
3275 const auto __yp = __yn < 0 ? -(__yn & 0x7fff'ffff) : __yn;
3276 return __auto_bitcast(
3277 __and(__to_masktype(_mm_cmpord_ps(__xi, __yi)), __xp >= __yp));
3278 }
3279 else if constexpr (__have_sse2 && sizeof(__xi) == 16
3280 && sizeof(_Tp) == 8)
3281 return __vector_type_t<__int_with_sizeof_t<8>, 2>{
3282 -_mm_ucomige_sd(__xi, __yi),
3283 -_mm_ucomige_sd(_mm_unpackhi_pd(__xi, __xi),
3284 _mm_unpackhi_pd(__yi, __yi))};
3285 else
3286 return _Base::_S_isgreaterequal(__x, __y);
3287 }
3288
3289 // }}}
3290 // _S_isless {{{
3291 template <typename _Tp, size_t _Np>
3292 static constexpr _MaskMember<_Tp> _S_isless(_SimdWrapper<_Tp, _Np> __x,
3293 _SimdWrapper<_Tp, _Np> __y)
3294 {
3295 const auto __xi = __to_intrin(__x);
3296 const auto __yi = __to_intrin(__y);
3297 if constexpr (__is_avx512_abi<_Abi>())
3298 {
3299 const auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
3300 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
3301 return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OQ);
3302 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
3303 return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OQ);
3304 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3305 return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OQ);
3306 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3307 return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OQ);
3308 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3309 return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OQ);
3310 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3311 return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OQ);
3312 else
3313 __assert_unreachable<_Tp>();
3314 }
3315 else if constexpr (__have_avx)
3316 {
3317 if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3318 return __to_masktype(_mm256_cmp_ps(__xi, __yi, _CMP_LT_OQ));
3319 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3320 return __to_masktype(_mm256_cmp_pd(__xi, __yi, _CMP_LT_OQ));
3321 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3322 return __auto_bitcast(_mm_cmp_ps(__xi, __yi, _CMP_LT_OQ));
3323 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3324 return __to_masktype(_mm_cmp_pd(__xi, __yi, _CMP_LT_OQ));
3325 else
3326 __assert_unreachable<_Tp>();
3327 }
3328 else if constexpr (__have_sse2 && sizeof(__xi) == 16
3329 && sizeof(_Tp) == 4)
3330 {
3331 const auto __xn = __vector_bitcast<int>(__xi);
3332 const auto __yn = __vector_bitcast<int>(__yi);
3333 const auto __xp = __xn < 0 ? -(__xn & 0x7fff'ffff) : __xn;
3334 const auto __yp = __yn < 0 ? -(__yn & 0x7fff'ffff) : __yn;
3335 return __auto_bitcast(
3336 __and(__to_masktype(_mm_cmpord_ps(__xi, __yi)), __xp < __yp));
3337 }
3338 else if constexpr (__have_sse2 && sizeof(__xi) == 16
3339 && sizeof(_Tp) == 8)
3340 return __vector_type_t<__int_with_sizeof_t<8>, 2>{
3341 -_mm_ucomigt_sd(__yi, __xi),
3342 -_mm_ucomigt_sd(_mm_unpackhi_pd(__yi, __yi),
3343 _mm_unpackhi_pd(__xi, __xi))};
3344 else
3345 return _Base::_S_isless(__x, __y);
3346 }
3347
3348 // }}}
3349 // _S_islessequal {{{
3350 template <typename _Tp, size_t _Np>
3351 static constexpr _MaskMember<_Tp>
3352 _S_islessequal(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
3353 {
3354 const auto __xi = __to_intrin(__x);
3355 const auto __yi = __to_intrin(__y);
3356 if constexpr (__is_avx512_abi<_Abi>())
3357 {
3358 const auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
3359 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
3360 return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OQ);
3361 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
3362 return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OQ);
3363 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3364 return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OQ);
3365 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3366 return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OQ);
3367 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3368 return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OQ);
3369 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3370 return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OQ);
3371 else
3372 __assert_unreachable<_Tp>();
3373 }
3374 else if constexpr (__have_avx)
3375 {
3376 if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3377 return __to_masktype(_mm256_cmp_ps(__xi, __yi, _CMP_LE_OQ));
3378 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3379 return __to_masktype(_mm256_cmp_pd(__xi, __yi, _CMP_LE_OQ));
3380 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3381 return __auto_bitcast(_mm_cmp_ps(__xi, __yi, _CMP_LE_OQ));
3382 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3383 return __to_masktype(_mm_cmp_pd(__xi, __yi, _CMP_LE_OQ));
3384 else
3385 __assert_unreachable<_Tp>();
3386 }
3387 else if constexpr (__have_sse2 && sizeof(__xi) == 16
3388 && sizeof(_Tp) == 4)
3389 {
3390 const auto __xn = __vector_bitcast<int>(__xi);
3391 const auto __yn = __vector_bitcast<int>(__yi);
3392 const auto __xp = __xn < 0 ? -(__xn & 0x7fff'ffff) : __xn;
3393 const auto __yp = __yn < 0 ? -(__yn & 0x7fff'ffff) : __yn;
3394 return __auto_bitcast(
3395 __and(__to_masktype(_mm_cmpord_ps(__xi, __yi)), __xp <= __yp));
3396 }
3397 else if constexpr (__have_sse2 && sizeof(__xi) == 16
3398 && sizeof(_Tp) == 8)
3399 return __vector_type_t<__int_with_sizeof_t<8>, 2>{
3400 -_mm_ucomige_sd(__yi, __xi),
3401 -_mm_ucomige_sd(_mm_unpackhi_pd(__yi, __yi),
3402 _mm_unpackhi_pd(__xi, __xi))};
3403 else
3404 return _Base::_S_islessequal(__x, __y);
3405 }
3406
3407 // }}}
3408 // _S_islessgreater {{{
3409 template <typename _Tp, size_t _Np>
3410 static constexpr _MaskMember<_Tp>
3411 _S_islessgreater(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
3412 {
3413 const auto __xi = __to_intrin(__x);
3414 const auto __yi = __to_intrin(__y);
3415 if constexpr (__is_avx512_abi<_Abi>())
3416 {
3417 const auto __k1 = _Abi::template _S_implicit_mask_intrin<_Tp>();
3418 if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 4)
3419 return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_NEQ_OQ);
3420 else if constexpr (sizeof(__xi) == 64 && sizeof(_Tp) == 8)
3421 return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_NEQ_OQ);
3422 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3423 return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_NEQ_OQ);
3424 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3425 return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_NEQ_OQ);
3426 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3427 return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_NEQ_OQ);
3428 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3429 return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_NEQ_OQ);
3430 else
3431 __assert_unreachable<_Tp>();
3432 }
3433 else if constexpr (__have_avx)
3434 {
3435 if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 4)
3436 return __to_masktype(_mm256_cmp_ps(__xi, __yi, _CMP_NEQ_OQ));
3437 else if constexpr (sizeof(__xi) == 32 && sizeof(_Tp) == 8)
3438 return __to_masktype(_mm256_cmp_pd(__xi, __yi, _CMP_NEQ_OQ));
3439 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3440 return __auto_bitcast(_mm_cmp_ps(__xi, __yi, _CMP_NEQ_OQ));
3441 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3442 return __to_masktype(_mm_cmp_pd(__xi, __yi, _CMP_NEQ_OQ));
3443 else
3444 __assert_unreachable<_Tp>();
3445 }
3446 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 4)
3447 return __auto_bitcast(
3448 __and(_mm_cmpord_ps(__xi, __yi), _mm_cmpneq_ps(__xi, __yi)));
3449 else if constexpr (sizeof(__xi) == 16 && sizeof(_Tp) == 8)
3450 return __to_masktype(
3451 __and(_mm_cmpord_pd(__xi, __yi), _mm_cmpneq_pd(__xi, __yi)));
3452 else
3453 __assert_unreachable<_Tp>();
3454 }
3455
3456 //}}} }}}
3457 };
3458
3459// }}}
3460// _MaskImplX86Mixin {{{
3461struct _MaskImplX86Mixin
3462{
3463 template <typename _Tp>
3464 using _TypeTag = _Tp*;
3465
3466 using _Base = _MaskImplBuiltinMixin;
3467
3468 // _S_to_maskvector(bool) {{{
3469 template <typename _Up, size_t _ToN = 1, typename _Tp>
3470 _GLIBCXX_SIMD_INTRINSIC static constexpr enable_if_t<
3471 is_same_v<_Tp, bool>, _SimdWrapper<_Up, _ToN>>
3472 _S_to_maskvector(_Tp __x)
3473 {
3474 static_assert(is_same_v<_Up, __int_for_sizeof_t<_Up>>);
3475 return __x ? __vector_type_t<_Up, _ToN>{~_Up()}
3476 : __vector_type_t<_Up, _ToN>();
3477 }
3478
3479 // }}}
3480 // _S_to_maskvector(_SanitizedBitMask) {{{
3481 template <typename _Up, size_t _UpN = 0, size_t _Np,
3482 size_t _ToN = _UpN == 0 ? _Np : _UpN>
3483 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Up, _ToN>
3484 _S_to_maskvector(_SanitizedBitMask<_Np> __x)
3485 {
3486 static_assert(is_same_v<_Up, __int_for_sizeof_t<_Up>>);
3487 using _UV = __vector_type_t<_Up, _ToN>;
3488 using _UI = __intrinsic_type_t<_Up, _ToN>;
3489 [[maybe_unused]] const auto __k = __x._M_to_bits();
3490 if constexpr (_Np == 1)
3491 return _S_to_maskvector<_Up, _ToN>(__k);
3492 else if (__x._M_is_constprop() || __builtin_is_constant_evaluated())
3493 return __generate_from_n_evaluations<std::min(_ToN, _Np), _UV>(
3494 [&](auto __i) -> _Up { return -__x[__i.value]; });
3495 else if constexpr (sizeof(_Up) == 1)
3496 {
3497 if constexpr (sizeof(_UI) == 16)
3498 {
3499 if constexpr (__have_avx512bw_vl)
3500 return __intrin_bitcast<_UV>(_mm_movm_epi8(__k));
3501 else if constexpr (__have_avx512bw)
3502 return __intrin_bitcast<_UV>(__lo128(_mm512_movm_epi8(__k)));
3503 else if constexpr (__have_avx512f)
3504 {
3505 auto __as32bits = _mm512_maskz_mov_epi32(__k, ~__m512i());
3506 auto __as16bits
3507 = __xzyw(_mm256_packs_epi32(__lo256(__as32bits),
3508 __hi256(__as32bits)));
3509 return __intrin_bitcast<_UV>(
3510 _mm_packs_epi16(__lo128(__as16bits), __hi128(__as16bits)));
3511 }
3512 else if constexpr (__have_ssse3)
3513 {
3514 const auto __bitmask = __to_intrin(
3515 __make_vector<_UChar>(1, 2, 4, 8, 16, 32, 64, 128, 1, 2, 4,
3516 8, 16, 32, 64, 128));
3517 return __intrin_bitcast<_UV>(
3518 __vector_bitcast<_Up>(
3519 _mm_shuffle_epi8(__to_intrin(
3520 __vector_type_t<_ULLong, 2>{__k}),
3521 _mm_setr_epi8(0, 0, 0, 0, 0, 0, 0, 0, 1,
3522 1, 1, 1, 1, 1, 1, 1))
3523 & __bitmask)
3524 != 0);
3525 }
3526 // else fall through
3527 }
3528 else if constexpr (sizeof(_UI) == 32)
3529 {
3530 if constexpr (__have_avx512bw_vl)
3531 return __vector_bitcast<_Up>(_mm256_movm_epi8(__k));
3532 else if constexpr (__have_avx512bw)
3533 return __vector_bitcast<_Up>(__lo256(_mm512_movm_epi8(__k)));
3534 else if constexpr (__have_avx512f)
3535 {
3536 auto __as16bits = // 0 16 1 17 ... 15 31
3537 _mm512_srli_epi32(_mm512_maskz_mov_epi32(__k, ~__m512i()),
3538 16)
3539 | _mm512_slli_epi32(_mm512_maskz_mov_epi32(__k >> 16,
3540 ~__m512i()),
3541 16);
3542 auto __0_16_1_17 = __xzyw(_mm256_packs_epi16(
3543 __lo256(__as16bits),
3544 __hi256(__as16bits)) // 0 16 1 17 2 18 3 19 8 24 9 25 ...
3545 );
3546 // deinterleave:
3547 return __vector_bitcast<_Up>(__xzyw(_mm256_shuffle_epi8(
3548 __0_16_1_17, // 0 16 1 17 2 ...
3549 _mm256_setr_epi8(0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9,
3550 11, 13, 15, 0, 2, 4, 6, 8, 10, 12, 14, 1,
3551 3, 5, 7, 9, 11, 13,
3552 15)))); // 0-7 16-23 8-15 24-31 -> xzyw
3553 // 0-3 8-11 16-19 24-27
3554 // 4-7 12-15 20-23 28-31
3555 }
3556 else if constexpr (__have_avx2)
3557 {
3558 const auto __bitmask
3559 = _mm256_broadcastsi128_si256(__to_intrin(
3560 __make_vector<_UChar>(1, 2, 4, 8, 16, 32, 64, 128, 1, 2,
3561 4, 8, 16, 32, 64, 128)));
3562 return __vector_bitcast<_Up>(
3563 __vector_bitcast<_Up>(
3564 _mm256_shuffle_epi8(
3565 _mm256_broadcastsi128_si256(
3566 __to_intrin(__vector_type_t<_ULLong, 2>{__k})),
3567 _mm256_setr_epi8(0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1,
3568 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2, 3, 3,
3569 3, 3, 3, 3, 3, 3))
3570 & __bitmask)
3571 != 0);
3572 }
3573 // else fall through
3574 }
3575 else if constexpr (sizeof(_UI) == 64)
3576 return reinterpret_cast<_UV>(_mm512_movm_epi8(__k));
3577 if constexpr (std::min(_ToN, _Np) <= 4)
3578 {
3579 if constexpr (_Np > 7) // avoid overflow
3580 __x &= _SanitizedBitMask<_Np>(0x0f);
3581 const _UInt __char_mask
3582 = ((_UInt(__x.to_ulong()) * 0x00204081U) & 0x01010101ULL)
3583 * 0xff;
3584 _UV __r = {};
3585 __builtin_memcpy(&__r, &__char_mask,
3586 std::min(sizeof(__r), sizeof(__char_mask)));
3587 return __r;
3588 }
3589 else if constexpr (std::min(_ToN, _Np) <= 7)
3590 {
3591 if constexpr (_Np > 7) // avoid overflow
3592 __x &= _SanitizedBitMask<_Np>(0x7f);
3593 const _ULLong __char_mask
3594 = ((__x.to_ulong() * 0x40810204081ULL) & 0x0101010101010101ULL)
3595 * 0xff;
3596 _UV __r = {};
3597 __builtin_memcpy(&__r, &__char_mask,
3598 std::min(sizeof(__r), sizeof(__char_mask)));
3599 return __r;
3600 }
3601 }
3602 else if constexpr (sizeof(_Up) == 2)
3603 {
3604 if constexpr (sizeof(_UI) == 16)
3605 {
3606 if constexpr (__have_avx512bw_vl)
3607 return __intrin_bitcast<_UV>(_mm_movm_epi16(__k));
3608 else if constexpr (__have_avx512bw)
3609 return __intrin_bitcast<_UV>(__lo128(_mm512_movm_epi16(__k)));
3610 else if constexpr (__have_avx512f)
3611 {
3612 __m256i __as32bits = {};
3613 if constexpr (__have_avx512vl)
3614 __as32bits = _mm256_maskz_mov_epi32(__k, ~__m256i());
3615 else
3616 __as32bits
3617 = __lo256(_mm512_maskz_mov_epi32(__k, ~__m512i()));
3618 return __intrin_bitcast<_UV>(
3619 _mm_packs_epi32(__lo128(__as32bits), __hi128(__as32bits)));
3620 }
3621 // else fall through
3622 }
3623 else if constexpr (sizeof(_UI) == 32)
3624 {
3625 if constexpr (__have_avx512bw_vl)
3626 return __vector_bitcast<_Up>(_mm256_movm_epi16(__k));
3627 else if constexpr (__have_avx512bw)
3628 return __vector_bitcast<_Up>(__lo256(_mm512_movm_epi16(__k)));
3629 else if constexpr (__have_avx512f)
3630 {
3631 auto __as32bits = _mm512_maskz_mov_epi32(__k, ~__m512i());
3632 return __vector_bitcast<_Up>(
3633 __xzyw(_mm256_packs_epi32(__lo256(__as32bits),
3634 __hi256(__as32bits))));
3635 }
3636 // else fall through
3637 }
3638 else if constexpr (sizeof(_UI) == 64)
3639 return __vector_bitcast<_Up>(_mm512_movm_epi16(__k));
3640 }
3641 else if constexpr (sizeof(_Up) == 4)
3642 {
3643 if constexpr (sizeof(_UI) == 16)
3644 {
3645 if constexpr (__have_avx512dq_vl)
3646 return __intrin_bitcast<_UV>(_mm_movm_epi32(__k));
3647 else if constexpr (__have_avx512dq)
3648 return __intrin_bitcast<_UV>(__lo128(_mm512_movm_epi32(__k)));
3649 else if constexpr (__have_avx512vl)
3650 return __intrin_bitcast<_UV>(
3651 _mm_maskz_mov_epi32(__k, ~__m128i()));
3652 else if constexpr (__have_avx512f)
3653 return __intrin_bitcast<_UV>(
3654 __lo128(_mm512_maskz_mov_epi32(__k, ~__m512i())));
3655 // else fall through
3656 }
3657 else if constexpr (sizeof(_UI) == 32)
3658 {
3659 if constexpr (__have_avx512dq_vl)
3660 return __vector_bitcast<_Up>(_mm256_movm_epi32(__k));
3661 else if constexpr (__have_avx512dq)
3662 return __vector_bitcast<_Up>(__lo256(_mm512_movm_epi32(__k)));
3663 else if constexpr (__have_avx512vl)
3664 return __vector_bitcast<_Up>(
3665 _mm256_maskz_mov_epi32(__k, ~__m256i()));
3666 else if constexpr (__have_avx512f)
3667 return __vector_bitcast<_Up>(
3668 __lo256(_mm512_maskz_mov_epi32(__k, ~__m512i())));
3669 // else fall through
3670 }
3671 else if constexpr (sizeof(_UI) == 64)
3672 return __vector_bitcast<_Up>(
3673 __have_avx512dq ? _mm512_movm_epi32(__k)
3674 : _mm512_maskz_mov_epi32(__k, ~__m512i()));
3675 }
3676 else if constexpr (sizeof(_Up) == 8)
3677 {
3678 if constexpr (sizeof(_UI) == 16)
3679 {
3680 if constexpr (__have_avx512dq_vl)
3681 return __vector_bitcast<_Up>(_mm_movm_epi64(__k));
3682 else if constexpr (__have_avx512dq)
3683 return __vector_bitcast<_Up>(__lo128(_mm512_movm_epi64(__k)));
3684 else if constexpr (__have_avx512vl)
3685 return __vector_bitcast<_Up>(
3686 _mm_maskz_mov_epi64(__k, ~__m128i()));
3687 else if constexpr (__have_avx512f)
3688 return __vector_bitcast<_Up>(
3689 __lo128(_mm512_maskz_mov_epi64(__k, ~__m512i())));
3690 // else fall through
3691 }
3692 else if constexpr (sizeof(_UI) == 32)
3693 {
3694 if constexpr (__have_avx512dq_vl)
3695 return __vector_bitcast<_Up>(_mm256_movm_epi64(__k));
3696 else if constexpr (__have_avx512dq)
3697 return __vector_bitcast<_Up>(__lo256(_mm512_movm_epi64(__k)));
3698 else if constexpr (__have_avx512vl)
3699 return __vector_bitcast<_Up>(
3700 _mm256_maskz_mov_epi64(__k, ~__m256i()));
3701 else if constexpr (__have_avx512f)
3702 return __vector_bitcast<_Up>(
3703 __lo256(_mm512_maskz_mov_epi64(__k, ~__m512i())));
3704 // else fall through
3705 }
3706 else if constexpr (sizeof(_UI) == 64)
3707 return __vector_bitcast<_Up>(
3708 __have_avx512dq ? _mm512_movm_epi64(__k)
3709 : _mm512_maskz_mov_epi64(__k, ~__m512i()));
3710 }
3711
3712 using _UpUInt = make_unsigned_t<_Up>;
3713 using _V = __vector_type_t<_UpUInt, _ToN>;
3714 constexpr size_t __bits_per_element = sizeof(_Up) * __CHAR_BIT__;
3715 if constexpr (_ToN == 2)
3716 {
3717 return __vector_bitcast<_Up>(_V{_UpUInt(-__x[0]), _UpUInt(-__x[1])});
3718 }
3719 else if constexpr (!__have_avx2 && __have_avx && sizeof(_V) == 32)
3720 {
3721 if constexpr (sizeof(_Up) == 4)
3722 return __vector_bitcast<_Up>(_mm256_cmp_ps(
3723 _mm256_and_ps(_mm256_castsi256_ps(_mm256_set1_epi32(__k)),
3724 _mm256_castsi256_ps(_mm256_setr_epi32(
3725 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80))),
3726 _mm256_setzero_ps(), _CMP_NEQ_UQ));
3727 else if constexpr (sizeof(_Up) == 8)
3728 return __vector_bitcast<_Up>(_mm256_cmp_pd(
3729 _mm256_and_pd(_mm256_castsi256_pd(_mm256_set1_epi64x(__k)),
3730 _mm256_castsi256_pd(
3731 _mm256_setr_epi64x(0x01, 0x02, 0x04, 0x08))),
3732 _mm256_setzero_pd(), _CMP_NEQ_UQ));
3733 else
3734 __assert_unreachable<_Up>();
3735 }
3736 else if constexpr (__bits_per_element >= _ToN)
3737 {
3738 constexpr auto __bitmask
3739 = __generate_vector<_V>([](auto __i) constexpr->_UpUInt {
3740 return __i < _ToN ? 1ull << __i : 0;
3741 });
3742 const auto __bits
3743 = __vector_broadcast<_ToN, _UpUInt>(__k) & __bitmask;
3744 if constexpr (__bits_per_element > _ToN)
3745 return __vector_bitcast<_Up>(__bits) > 0;
3746 else
3747 return __vector_bitcast<_Up>(__bits != 0);
3748 }
3749 else
3750 {
3751 const _V __tmp
3752 = __generate_vector<_V>([&](auto __i) constexpr {
3753 return static_cast<_UpUInt>(
3754 __k >> (__bits_per_element * (__i / __bits_per_element)));
3755 })
3756 & __generate_vector<_V>([](auto __i) constexpr {
3757 return static_cast<_UpUInt>(1ull
3758 << (__i % __bits_per_element));
3759 }); // mask bit index
3760 return __intrin_bitcast<_UV>(__tmp != _V());
3761 }
3762 }
3763
3764 // }}}
3765 // _S_to_maskvector(_SimdWrapper) {{{
3766 template <typename _Up, size_t _UpN = 0, typename _Tp, size_t _Np,
3767 size_t _ToN = _UpN == 0 ? _Np : _UpN>
3768 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Up, _ToN>
3769 _S_to_maskvector(_SimdWrapper<_Tp, _Np> __x)
3770 {
3771 static_assert(is_same_v<_Up, __int_for_sizeof_t<_Up>>);
3772 using _TW = _SimdWrapper<_Tp, _Np>;
3773 using _UW = _SimdWrapper<_Up, _ToN>;
3774 using _UI = __intrinsic_type_t<_Up, _ToN>;
3775 if constexpr (is_same_v<_Tp, bool>) // bits -> vector
3776 return _S_to_maskvector<_Up, _ToN>(
3777 _BitMask<_Np>(__x._M_data)._M_sanitized());
3778 // vector -> vector bitcast
3779 else if constexpr (sizeof(_Up) == sizeof(_Tp)
3780 && sizeof(_TW) == sizeof(_UW))
3781 return __wrapper_bitcast<_Up, _ToN>(
3782 _ToN <= _Np
3783 ? __x
3784 : simd_abi::_VecBuiltin<sizeof(_Tp) * _Np>::_S_masked(__x));
3785 else // vector -> vector {{{
3786 {
3787 if (__x._M_is_constprop() || __builtin_is_constant_evaluated())
3788 {
3789 const auto __y = __vector_bitcast<__int_for_sizeof_t<_Tp>>(__x);
3790 return __generate_from_n_evaluations<std::min(_ToN, _Np),
3791 __vector_type_t<_Up, _ToN>>(
3792 [&](auto __i) -> _Up { return __y[__i.value]; });
3793 }
3794 using _To = __vector_type_t<_Up, _ToN>;
3795 [[maybe_unused]] constexpr size_t _FromN = _Np;
3796 constexpr int _FromBytes = sizeof(_Tp);
3797 constexpr int _ToBytes = sizeof(_Up);
3798 const auto __k = __x._M_data;
3799
3800 if constexpr (_FromBytes == _ToBytes)
3801 return __intrin_bitcast<_To>(__k);
3802 else if constexpr (sizeof(_UI) == 16 && sizeof(__k) == 16)
3803 { // SSE -> SSE {{{
3804 if constexpr (_FromBytes == 4 && _ToBytes == 8)
3805 return __intrin_bitcast<_To>(__interleave128_lo(__k, __k));
3806 else if constexpr (_FromBytes == 2 && _ToBytes == 8)
3807 {
3808 const auto __y
3809 = __vector_bitcast<int>(__interleave128_lo(__k, __k));
3810 return __intrin_bitcast<_To>(__interleave128_lo(__y, __y));
3811 }
3812 else if constexpr (_FromBytes == 1 && _ToBytes == 8)
3813 {
3814 auto __y
3815 = __vector_bitcast<short>(__interleave128_lo(__k, __k));
3816 auto __z
3817 = __vector_bitcast<int>(__interleave128_lo(__y, __y));
3818 return __intrin_bitcast<_To>(__interleave128_lo(__z, __z));
3819 }
3820 else if constexpr (_FromBytes == 8 && _ToBytes == 4
3821 && __have_sse2)
3822 return __intrin_bitcast<_To>(
3823 _mm_packs_epi32(__vector_bitcast<_LLong>(__k), __m128i()));
3824 else if constexpr (_FromBytes == 8 && _ToBytes == 4)
3825 return __vector_shuffle<1, 3, 6, 7>(__vector_bitcast<_Up>(__k),
3826 _UI());
3827 else if constexpr (_FromBytes == 2 && _ToBytes == 4)
3828 return __intrin_bitcast<_To>(__interleave128_lo(__k, __k));
3829 else if constexpr (_FromBytes == 1 && _ToBytes == 4)
3830 {
3831 const auto __y
3832 = __vector_bitcast<short>(__interleave128_lo(__k, __k));
3833 return __intrin_bitcast<_To>(__interleave128_lo(__y, __y));
3834 }
3835 else if constexpr (_FromBytes == 8 && _ToBytes == 2)
3836 {
3837 if constexpr (__have_sse2 && !__have_ssse3)
3838 return __intrin_bitcast<_To>(_mm_packs_epi32(
3839 _mm_packs_epi32(__vector_bitcast<_LLong>(__k), __m128i()),
3840 __m128i()));
3841 else
3842 return __intrin_bitcast<_To>(
3843 __vector_permute<3, 7, -1, -1, -1, -1, -1, -1>(
3844 __vector_bitcast<_Up>(__k)));
3845 }
3846 else if constexpr (_FromBytes == 4 && _ToBytes == 2)
3847 return __intrin_bitcast<_To>(
3848 _mm_packs_epi32(__vector_bitcast<_LLong>(__k), __m128i()));
3849 else if constexpr (_FromBytes == 1 && _ToBytes == 2)
3850 return __intrin_bitcast<_To>(__interleave128_lo(__k, __k));
3851 else if constexpr (_FromBytes == 8 && _ToBytes == 1
3852 && __have_ssse3)
3853 return __intrin_bitcast<_To>(
3854 _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
3855 _mm_setr_epi8(7, 15, -1, -1, -1, -1, -1, -1,
3856 -1, -1, -1, -1, -1, -1, -1,
3857 -1)));
3858 else if constexpr (_FromBytes == 8 && _ToBytes == 1)
3859 {
3860 auto __y
3861 = _mm_packs_epi32(__vector_bitcast<_LLong>(__k), __m128i());
3862 __y = _mm_packs_epi32(__y, __m128i());
3863 return __intrin_bitcast<_To>(_mm_packs_epi16(__y, __m128i()));
3864 }
3865 else if constexpr (_FromBytes == 4 && _ToBytes == 1
3866 && __have_ssse3)
3867 return __intrin_bitcast<_To>(
3868 _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
3869 _mm_setr_epi8(3, 7, 11, 15, -1, -1, -1, -1,
3870 -1, -1, -1, -1, -1, -1, -1,
3871 -1)));
3872 else if constexpr (_FromBytes == 4 && _ToBytes == 1)
3873 {
3874 const auto __y
3875 = _mm_packs_epi32(__vector_bitcast<_LLong>(__k), __m128i());
3876 return __intrin_bitcast<_To>(_mm_packs_epi16(__y, __m128i()));
3877 }
3878 else if constexpr (_FromBytes == 2 && _ToBytes == 1)
3879 return __intrin_bitcast<_To>(
3880 _mm_packs_epi16(__vector_bitcast<_LLong>(__k), __m128i()));
3881 else
3882 __assert_unreachable<_Tp>();
3883 } // }}}
3884 else if constexpr (sizeof(_UI) == 32 && sizeof(__k) == 32)
3885 { // AVX -> AVX {{{
3886 if constexpr (_FromBytes == _ToBytes)
3887 __assert_unreachable<_Tp>();
3888 else if constexpr (_FromBytes == _ToBytes * 2)
3889 {
3890 const auto __y = __vector_bitcast<_LLong>(__k);
3891 return __intrin_bitcast<_To>(_mm256_castsi128_si256(
3892 _mm_packs_epi16(__lo128(__y), __hi128(__y))));
3893 }
3894 else if constexpr (_FromBytes == _ToBytes * 4)
3895 {
3896 const auto __y = __vector_bitcast<_LLong>(__k);
3897 return __intrin_bitcast<_To>(_mm256_castsi128_si256(
3898 _mm_packs_epi16(_mm_packs_epi16(__lo128(__y), __hi128(__y)),
3899 __m128i())));
3900 }
3901 else if constexpr (_FromBytes == _ToBytes * 8)
3902 {
3903 const auto __y = __vector_bitcast<_LLong>(__k);
3904 return __intrin_bitcast<_To>(
3905 _mm256_castsi128_si256(_mm_shuffle_epi8(
3906 _mm_packs_epi16(__lo128(__y), __hi128(__y)),
3907 _mm_setr_epi8(3, 7, 11, 15, -1, -1, -1, -1, -1, -1, -1,
3908 -1, -1, -1, -1, -1))));
3909 }
3910 else if constexpr (_FromBytes * 2 == _ToBytes)
3911 {
3912 auto __y = __xzyw(__to_intrin(__k));
3913 if constexpr (is_floating_point_v<
3914 _Tp> || (!__have_avx2 && _FromBytes == 4))
3915 {
3916 const auto __yy = __vector_bitcast<float>(__y);
3917 return __intrin_bitcast<_To>(
3918 _mm256_unpacklo_ps(__yy, __yy));
3919 }
3920 else
3921 return __intrin_bitcast<_To>(
3922 _mm256_unpacklo_epi8(__y, __y));
3923 }
3924 else if constexpr (_FromBytes * 4 == _ToBytes)
3925 {
3926 auto __y
3927 = _mm_unpacklo_epi8(__lo128(__vector_bitcast<_LLong>(__k)),
3928 __lo128(__vector_bitcast<_LLong>(
3929 __k))); // drops 3/4 of input
3930 return __intrin_bitcast<_To>(
3931 __concat(_mm_unpacklo_epi16(__y, __y),
3932 _mm_unpackhi_epi16(__y, __y)));
3933 }
3934 else if constexpr (_FromBytes == 1 && _ToBytes == 8)
3935 {
3936 auto __y
3937 = _mm_unpacklo_epi8(__lo128(__vector_bitcast<_LLong>(__k)),
3938 __lo128(__vector_bitcast<_LLong>(
3939 __k))); // drops 3/4 of input
3940 __y
3941 = _mm_unpacklo_epi16(__y,
3942 __y); // drops another 1/2 => 7/8 total
3943 return __intrin_bitcast<_To>(
3944 __concat(_mm_unpacklo_epi32(__y, __y),
3945 _mm_unpackhi_epi32(__y, __y)));
3946 }
3947 else
3948 __assert_unreachable<_Tp>();
3949 } // }}}
3950 else if constexpr (sizeof(_UI) == 32 && sizeof(__k) == 16)
3951 { // SSE -> AVX {{{
3952 if constexpr (_FromBytes == _ToBytes)
3953 return __intrin_bitcast<_To>(
3954 __intrinsic_type_t<_Tp, 32 / sizeof(_Tp)>(
3955 __zero_extend(__to_intrin(__k))));
3956 else if constexpr (_FromBytes * 2 == _ToBytes)
3957 { // keep all
3958 return __intrin_bitcast<_To>(
3959 __concat(_mm_unpacklo_epi8(__vector_bitcast<_LLong>(__k),
3960 __vector_bitcast<_LLong>(__k)),
3961 _mm_unpackhi_epi8(__vector_bitcast<_LLong>(__k),
3962 __vector_bitcast<_LLong>(__k))));
3963 }
3964 else if constexpr (_FromBytes * 4 == _ToBytes)
3965 {
3966 if constexpr (__have_avx2)
3967 {
3968 return __intrin_bitcast<_To>(_mm256_shuffle_epi8(
3969 __concat(__vector_bitcast<_LLong>(__k),
3970 __vector_bitcast<_LLong>(__k)),
3971 _mm256_setr_epi8(0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 3,
3972 3, 3, 3, 4, 4, 4, 4, 5, 5, 5, 5, 6, 6,
3973 6, 6, 7, 7, 7, 7)));
3974 }
3975 else
3976 {
3977 return __intrin_bitcast<_To>(__concat(
3978 _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
3979 _mm_setr_epi8(0, 0, 0, 0, 1, 1, 1, 1,
3980 2, 2, 2, 2, 3, 3, 3, 3)),
3981 _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
3982 _mm_setr_epi8(4, 4, 4, 4, 5, 5, 5, 5,
3983 6, 6, 6, 6, 7, 7, 7,
3984 7))));
3985 }
3986 }
3987 else if constexpr (_FromBytes * 8 == _ToBytes)
3988 {
3989 if constexpr (__have_avx2)
3990 {
3991 return __intrin_bitcast<_To>(_mm256_shuffle_epi8(
3992 __concat(__vector_bitcast<_LLong>(__k),
3993 __vector_bitcast<_LLong>(__k)),
3994 _mm256_setr_epi8(0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1,
3995 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2, 3, 3,
3996 3, 3, 3, 3, 3, 3)));
3997 }
3998 else
3999 {
4000 return __intrin_bitcast<_To>(__concat(
4001 _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
4002 _mm_setr_epi8(0, 0, 0, 0, 0, 0, 0, 0,
4003 1, 1, 1, 1, 1, 1, 1, 1)),
4004 _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
4005 _mm_setr_epi8(2, 2, 2, 2, 2, 2, 2, 2,
4006 3, 3, 3, 3, 3, 3, 3,
4007 3))));
4008 }
4009 }
4010 else if constexpr (_FromBytes == _ToBytes * 2)
4011 return __intrin_bitcast<_To>(__m256i(__zero_extend(
4012 _mm_packs_epi16(__vector_bitcast<_LLong>(__k), __m128i()))));
4013 else if constexpr (_FromBytes == 8 && _ToBytes == 2)
4014 {
4015 return __intrin_bitcast<_To>(__m256i(__zero_extend(
4016 _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
4017 _mm_setr_epi8(6, 7, 14, 15, -1, -1, -1, -1,
4018 -1, -1, -1, -1, -1, -1, -1,
4019 -1)))));
4020 }
4021 else if constexpr (_FromBytes == 4 && _ToBytes == 1)
4022 {
4023 return __intrin_bitcast<_To>(__m256i(__zero_extend(
4024 _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
4025 _mm_setr_epi8(3, 7, 11, 15, -1, -1, -1, -1,
4026 -1, -1, -1, -1, -1, -1, -1,
4027 -1)))));
4028 }
4029 else if constexpr (_FromBytes == 8 && _ToBytes == 1)
4030 {
4031 return __intrin_bitcast<_To>(__m256i(__zero_extend(
4032 _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
4033 _mm_setr_epi8(7, 15, -1, -1, -1, -1, -1,
4034 -1, -1, -1, -1, -1, -1, -1,
4035 -1, -1)))));
4036 }
4037 else
4038 static_assert(!is_same_v<_Tp, _Tp>, "should be unreachable");
4039 } // }}}
4040 else if constexpr (sizeof(_UI) == 16 && sizeof(__k) == 32)
4041 { // AVX -> SSE {{{
4042 if constexpr (_FromBytes == _ToBytes)
4043 { // keep low 1/2
4044 return __intrin_bitcast<_To>(__lo128(__k));
4045 }
4046 else if constexpr (_FromBytes == _ToBytes * 2)
4047 { // keep all
4048 auto __y = __vector_bitcast<_LLong>(__k);
4049 return __intrin_bitcast<_To>(
4050 _mm_packs_epi16(__lo128(__y), __hi128(__y)));
4051 }
4052 else if constexpr (_FromBytes == _ToBytes * 4)
4053 { // add 1/2 undef
4054 auto __y = __vector_bitcast<_LLong>(__k);
4055 return __intrin_bitcast<_To>(
4056 _mm_packs_epi16(_mm_packs_epi16(__lo128(__y), __hi128(__y)),
4057 __m128i()));
4058 }
4059 else if constexpr (_FromBytes == 8 && _ToBytes == 1)
4060 { // add 3/4 undef
4061 auto __y = __vector_bitcast<_LLong>(__k);
4062 return __intrin_bitcast<_To>(_mm_shuffle_epi8(
4063 _mm_packs_epi16(__lo128(__y), __hi128(__y)),
4064 _mm_setr_epi8(3, 7, 11, 15, -1, -1, -1, -1, -1, -1, -1, -1,
4065 -1, -1, -1, -1)));
4066 }
4067 else if constexpr (_FromBytes * 2 == _ToBytes)
4068 { // keep low 1/4
4069 auto __y = __lo128(__vector_bitcast<_LLong>(__k));
4070 return __intrin_bitcast<_To>(_mm_unpacklo_epi8(__y, __y));
4071 }
4072 else if constexpr (_FromBytes * 4 == _ToBytes)
4073 { // keep low 1/8
4074 auto __y = __lo128(__vector_bitcast<_LLong>(__k));
4075 __y = _mm_unpacklo_epi8(__y, __y);
4076 return __intrin_bitcast<_To>(_mm_unpacklo_epi8(__y, __y));
4077 }
4078 else if constexpr (_FromBytes * 8 == _ToBytes)
4079 { // keep low 1/16
4080 auto __y = __lo128(__vector_bitcast<_LLong>(__k));
4081 __y = _mm_unpacklo_epi8(__y, __y);
4082 __y = _mm_unpacklo_epi8(__y, __y);
4083 return __intrin_bitcast<_To>(_mm_unpacklo_epi8(__y, __y));
4084 }
4085 else
4086 static_assert(!is_same_v<_Tp, _Tp>, "should be unreachable");
4087 } // }}}
4088 else
4089 return _Base::template _S_to_maskvector<_Up, _ToN>(__x);
4090 /*
4091 if constexpr (_FromBytes > _ToBytes) {
4092 const _To __y = __vector_bitcast<_Up>(__k);
4093 return [&] <size_t... _Is> (index_sequence<_Is...>) {
4094 constexpr int _Stride = _FromBytes / _ToBytes;
4095 return _To{__y[(_Is + 1) * _Stride - 1]...};
4096 }(make_index_sequence<std::min(_ToN, _FromN)>());
4097 } else {
4098 // {0, 0, 1, 1} (_Dups = 2, _Is<4>)
4099 // {0, 0, 0, 0, 1, 1, 1, 1} (_Dups = 4, _Is<8>)
4100 // {0, 0, 1, 1, 2, 2, 3, 3} (_Dups = 2, _Is<8>)
4101 // ...
4102 return [&] <size_t... _Is> (index_sequence<_Is...>) {
4103 constexpr int __dup = _ToBytes / _FromBytes;
4104 return __intrin_bitcast<_To>(_From{__k[_Is / __dup]...});
4105 }(make_index_sequence<_FromN>());
4106 }
4107 */
4108 } // }}}
4109 }
4110
4111 // }}}
4112 // _S_to_bits {{{
4113 template <typename _Tp, size_t _Np>
4114 _GLIBCXX_SIMD_INTRINSIC static constexpr _SanitizedBitMask<_Np>
4115 _S_to_bits(_SimdWrapper<_Tp, _Np> __x)
4116 {
4117 if constexpr (is_same_v<_Tp, bool>)
4118 return _BitMask<_Np>(__x._M_data)._M_sanitized();
4119 else
4120 {
4121 static_assert(is_same_v<_Tp, __int_for_sizeof_t<_Tp>>);
4122 if (__builtin_is_constant_evaluated()
4123 || __builtin_constant_p(__x._M_data))
4124 {
4125 const auto __bools = -__x._M_data;
4126 const _ULLong __k = __call_with_n_evaluations<_Np>(
4127 [](auto... __bits) { return (__bits | ...); },
4128 [&](auto __i) { return _ULLong(__bools[+__i]) << __i; });
4129 if (__builtin_is_constant_evaluated()
4130 || __builtin_constant_p(__k))
4131 return __k;
4132 }
4133 const auto __xi = __to_intrin(__x);
4134 if constexpr (sizeof(_Tp) == 1)
4135 if constexpr (sizeof(__xi) == 16)
4136 if constexpr (__have_avx512bw_vl)
4137 return _BitMask<_Np>(_mm_movepi8_mask(__xi));
4138 else // implies SSE2
4139 return _BitMask<_Np>(_mm_movemask_epi8(__xi));
4140 else if constexpr (sizeof(__xi) == 32)
4141 if constexpr (__have_avx512bw_vl)
4142 return _BitMask<_Np>(_mm256_movepi8_mask(__xi));
4143 else // implies AVX2
4144 return _BitMask<_Np>(_mm256_movemask_epi8(__xi));
4145 else // implies AVX512BW
4146 return _BitMask<_Np>(_mm512_movepi8_mask(__xi));
4147
4148 else if constexpr (sizeof(_Tp) == 2)
4149 if constexpr (sizeof(__xi) == 16)
4150 if constexpr (__have_avx512bw_vl)
4151 return _BitMask<_Np>(_mm_movepi16_mask(__xi));
4152 else if constexpr (__have_avx512bw)
4153 return _BitMask<_Np>(_mm512_movepi16_mask(__zero_extend(__xi)));
4154 else // implies SSE2
4155 return _BitMask<_Np>(
4156 _mm_movemask_epi8(_mm_packs_epi16(__xi, __m128i())));
4157 else if constexpr (sizeof(__xi) == 32)
4158 if constexpr (__have_avx512bw_vl)
4159 return _BitMask<_Np>(_mm256_movepi16_mask(__xi));
4160 else if constexpr (__have_avx512bw)
4161 return _BitMask<_Np>(_mm512_movepi16_mask(__zero_extend(__xi)));
4162 else // implies SSE2
4163 return _BitMask<_Np>(_mm_movemask_epi8(
4164 _mm_packs_epi16(__lo128(__xi), __hi128(__xi))));
4165 else // implies AVX512BW
4166 return _BitMask<_Np>(_mm512_movepi16_mask(__xi));
4167
4168 else if constexpr (sizeof(_Tp) == 4)
4169 if constexpr (sizeof(__xi) == 16)
4170 if constexpr (__have_avx512dq_vl)
4171 return _BitMask<_Np>(_mm_movepi32_mask(__xi));
4172 else if constexpr (__have_avx512vl)
4173 return _BitMask<_Np>(_mm_cmplt_epi32_mask(__xi, __m128i()));
4174 else if constexpr (__have_avx512dq)
4175 return _BitMask<_Np>(_mm512_movepi32_mask(__zero_extend(__xi)));
4176 else if constexpr (__have_avx512f)
4177 return _BitMask<_Np>(
4178 _mm512_cmplt_epi32_mask(__zero_extend(__xi), __m512i()));
4179 else // implies SSE
4180 return _BitMask<_Np>(
4181 _mm_movemask_ps(reinterpret_cast<__m128>(__xi)));
4182 else if constexpr (sizeof(__xi) == 32)
4183 if constexpr (__have_avx512dq_vl)
4184 return _BitMask<_Np>(_mm256_movepi32_mask(__xi));
4185 else if constexpr (__have_avx512dq)
4186 return _BitMask<_Np>(_mm512_movepi32_mask(__zero_extend(__xi)));
4187 else if constexpr (__have_avx512vl)
4188 return _BitMask<_Np>(_mm256_cmplt_epi32_mask(__xi, __m256i()));
4189 else if constexpr (__have_avx512f)
4190 return _BitMask<_Np>(
4191 _mm512_cmplt_epi32_mask(__zero_extend(__xi), __m512i()));
4192 else // implies AVX
4193 return _BitMask<_Np>(
4194 _mm256_movemask_ps(reinterpret_cast<__m256>(__xi)));
4195 else // implies AVX512??
4196 if constexpr (__have_avx512dq)
4197 return _BitMask<_Np>(_mm512_movepi32_mask(__xi));
4198 else // implies AVX512F
4199 return _BitMask<_Np>(_mm512_cmplt_epi32_mask(__xi, __m512i()));
4200
4201 else if constexpr (sizeof(_Tp) == 8)
4202 if constexpr (sizeof(__xi) == 16)
4203 if constexpr (__have_avx512dq_vl)
4204 return _BitMask<_Np>(_mm_movepi64_mask(__xi));
4205 else if constexpr (__have_avx512dq)
4206 return _BitMask<_Np>(_mm512_movepi64_mask(__zero_extend(__xi)));
4207 else if constexpr (__have_avx512vl)
4208 return _BitMask<_Np>(_mm_cmplt_epi64_mask(__xi, __m128i()));
4209 else if constexpr (__have_avx512f)
4210 return _BitMask<_Np>(
4211 _mm512_cmplt_epi64_mask(__zero_extend(__xi), __m512i()));
4212 else // implies SSE2
4213 return _BitMask<_Np>(
4214 _mm_movemask_pd(reinterpret_cast<__m128d>(__xi)));
4215 else if constexpr (sizeof(__xi) == 32)
4216 if constexpr (__have_avx512dq_vl)
4217 return _BitMask<_Np>(_mm256_movepi64_mask(__xi));
4218 else if constexpr (__have_avx512dq)
4219 return _BitMask<_Np>(_mm512_movepi64_mask(__zero_extend(__xi)));
4220 else if constexpr (__have_avx512vl)
4221 return _BitMask<_Np>(_mm256_cmplt_epi64_mask(__xi, __m256i()));
4222 else if constexpr (__have_avx512f)
4223 return _BitMask<_Np>(
4224 _mm512_cmplt_epi64_mask(__zero_extend(__xi), __m512i()));
4225 else // implies AVX
4226 return _BitMask<_Np>(
4227 _mm256_movemask_pd(reinterpret_cast<__m256d>(__xi)));
4228 else // implies AVX512??
4229 if constexpr (__have_avx512dq)
4230 return _BitMask<_Np>(_mm512_movepi64_mask(__xi));
4231 else // implies AVX512F
4232 return _BitMask<_Np>(_mm512_cmplt_epi64_mask(__xi, __m512i()));
4233
4234 else
4235 __assert_unreachable<_Tp>();
4236 }
4237 }
4238 // }}}
4239};
4240
4241// }}}
4242// _MaskImplX86 {{{
4243template <typename _Abi>
4244 struct _MaskImplX86 : _MaskImplX86Mixin, _MaskImplBuiltin<_Abi>
4245 {
4246 using _MaskImplX86Mixin::_S_to_bits;
4247 using _MaskImplX86Mixin::_S_to_maskvector;
4248 using _MaskImplBuiltin<_Abi>::_S_convert;
4249
4250 // member types {{{
4251 template <typename _Tp>
4252 using _SimdMember = typename _Abi::template __traits<_Tp>::_SimdMember;
4253
4254 template <typename _Tp>
4255 using _MaskMember = typename _Abi::template _MaskMember<_Tp>;
4256
4257 template <typename _Tp>
4258 static constexpr size_t _S_size = simd_size_v<_Tp, _Abi>;
4259
4260 using _Base = _MaskImplBuiltin<_Abi>;
4261
4262 // }}}
4263 // _S_broadcast {{{
4264 template <typename _Tp>
4265 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
4266 _S_broadcast(bool __x)
4267 {
4268 if constexpr (__is_avx512_abi<_Abi>())
4269 return __x ? _Abi::_S_masked(_MaskMember<_Tp>(-1))
4270 : _MaskMember<_Tp>();
4271 else
4272 return _Base::template _S_broadcast<_Tp>(__x);
4273 }
4274
4275 // }}}
4276 // _S_load {{{
4277 template <typename _Tp>
4278 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
4279 _S_load(const bool* __mem)
4280 {
4281 static_assert(is_same_v<_Tp, __int_for_sizeof_t<_Tp>>);
4282 if constexpr (__have_avx512bw)
4283 {
4284 const auto __to_vec_or_bits = [](auto __bits) -> decltype(auto) {
4285 if constexpr (__is_avx512_abi<_Abi>())
4286 return __bits;
4287 else
4288 return _S_to_maskvector<_Tp>(
4289 _BitMask<_S_size<_Tp>>(__bits)._M_sanitized());
4290 };
4291
4292 if constexpr (_S_size<_Tp> <= 16 && __have_avx512vl)
4293 {
4294 __m128i __a = {};
4295 __builtin_memcpy(&__a, __mem, _S_size<_Tp>);
4296 return __to_vec_or_bits(_mm_test_epi8_mask(__a, __a));
4297 }
4298 else if constexpr (_S_size<_Tp> <= 32 && __have_avx512vl)
4299 {
4300 __m256i __a = {};
4301 __builtin_memcpy(&__a, __mem, _S_size<_Tp>);
4302 return __to_vec_or_bits(_mm256_test_epi8_mask(__a, __a));
4303 }
4304 else if constexpr (_S_size<_Tp> <= 64)
4305 {
4306 __m512i __a = {};
4307 __builtin_memcpy(&__a, __mem, _S_size<_Tp>);
4308 return __to_vec_or_bits(_mm512_test_epi8_mask(__a, __a));
4309 }
4310 }
4311 else if constexpr (__is_avx512_abi<_Abi>())
4312 {
4313 if constexpr (_S_size<_Tp> <= 8)
4314 {
4315 __m128i __a = {};
4316 __builtin_memcpy(&__a, __mem, _S_size<_Tp>);
4317 const auto __b = _mm512_cvtepi8_epi64(__a);
4318 return _mm512_test_epi64_mask(__b, __b);
4319 }
4320 else if constexpr (_S_size<_Tp> <= 16)
4321 {
4322 __m128i __a = {};
4323 __builtin_memcpy(&__a, __mem, _S_size<_Tp>);
4324 const auto __b = _mm512_cvtepi8_epi32(__a);
4325 return _mm512_test_epi32_mask(__b, __b);
4326 }
4327 else if constexpr (_S_size<_Tp> <= 32)
4328 {
4329 __m128i __a = {};
4330 __builtin_memcpy(&__a, __mem, 16);
4331 const auto __b = _mm512_cvtepi8_epi32(__a);
4332 __builtin_memcpy(&__a, __mem + 16, _S_size<_Tp> - 16);
4333 const auto __c = _mm512_cvtepi8_epi32(__a);
4334 return _mm512_test_epi32_mask(__b, __b)
4335 | (_mm512_test_epi32_mask(__c, __c) << 16);
4336 }
4337 else if constexpr (_S_size<_Tp> <= 64)
4338 {
4339 __m128i __a = {};
4340 __builtin_memcpy(&__a, __mem, 16);
4341 const auto __b = _mm512_cvtepi8_epi32(__a);
4342 __builtin_memcpy(&__a, __mem + 16, 16);
4343 const auto __c = _mm512_cvtepi8_epi32(__a);
4344 if constexpr (_S_size<_Tp> <= 48)
4345 {
4346 __builtin_memcpy(&__a, __mem + 32, _S_size<_Tp> - 32);
4347 const auto __d = _mm512_cvtepi8_epi32(__a);
4348 return _mm512_test_epi32_mask(__b, __b)
4349 | (_mm512_test_epi32_mask(__c, __c) << 16)
4350 | (_ULLong(_mm512_test_epi32_mask(__d, __d)) << 32);
4351 }
4352 else
4353 {
4354 __builtin_memcpy(&__a, __mem + 16, 16);
4355 const auto __d = _mm512_cvtepi8_epi32(__a);
4356 __builtin_memcpy(&__a, __mem + 32, _S_size<_Tp> - 48);
4357 const auto __e = _mm512_cvtepi8_epi32(__a);
4358 return _mm512_test_epi32_mask(__b, __b)
4359 | (_mm512_test_epi32_mask(__c, __c) << 16)
4360 | (_ULLong(_mm512_test_epi32_mask(__d, __d)) << 32)
4361 | (_ULLong(_mm512_test_epi32_mask(__e, __e)) << 48);
4362 }
4363 }
4364 else
4365 __assert_unreachable<_Tp>();
4366 }
4367 else if constexpr (sizeof(_Tp) == 8 && _S_size<_Tp> == 2)
4368 return __vector_bitcast<_Tp>(
4369 __vector_type16_t<int>{-int(__mem[0]), -int(__mem[0]),
4370 -int(__mem[1]), -int(__mem[1])});
4371 else if constexpr (sizeof(_Tp) == 8 && _S_size<_Tp> <= 4 && __have_avx)
4372 {
4373 int __bool4 = 0;
4374 __builtin_memcpy(&__bool4, __mem, _S_size<_Tp>);
4375 const auto __k = __to_intrin(
4376 (__vector_broadcast<4>(__bool4)
4377 & __make_vector<int>(0x1, 0x100, 0x10000,
4378 _S_size<_Tp> == 4 ? 0x1000000 : 0))
4379 != 0);
4380 return __vector_bitcast<_Tp>(
4381 __concat(_mm_unpacklo_epi32(__k, __k),
4382 _mm_unpackhi_epi32(__k, __k)));
4383 }
4384 else if constexpr (sizeof(_Tp) == 4 && _S_size<_Tp> <= 4)
4385 {
4386 int __bools = 0;
4387 __builtin_memcpy(&__bools, __mem, _S_size<_Tp>);
4388 if constexpr (__have_sse2)
4389 {
4390 __m128i __k = _mm_cvtsi32_si128(__bools);
4391 __k = _mm_cmpgt_epi16(_mm_unpacklo_epi8(__k, __k), __m128i());
4392 return __vector_bitcast<_Tp, _S_size<_Tp>>(
4393 _mm_unpacklo_epi16(__k, __k));
4394 }
4395 else
4396 {
4397 __m128 __k = _mm_cvtpi8_ps(_mm_cvtsi32_si64(__bools));
4398 _mm_empty();
4399 return __vector_bitcast<_Tp, _S_size<_Tp>>(
4400 _mm_cmpgt_ps(__k, __m128()));
4401 }
4402 }
4403 else if constexpr (sizeof(_Tp) == 4 && _S_size<_Tp> <= 8)
4404 {
4405 __m128i __k = {};
4406 __builtin_memcpy(&__k, __mem, _S_size<_Tp>);
4407 __k = _mm_cmpgt_epi16(_mm_unpacklo_epi8(__k, __k), __m128i());
4408 return __vector_bitcast<_Tp>(
4409 __concat(_mm_unpacklo_epi16(__k, __k),
4410 _mm_unpackhi_epi16(__k, __k)));
4411 }
4412 else if constexpr (sizeof(_Tp) == 2 && _S_size<_Tp> <= 16)
4413 {
4414 __m128i __k = {};
4415 __builtin_memcpy(&__k, __mem, _S_size<_Tp>);
4416 __k = _mm_cmpgt_epi8(__k, __m128i());
4417 if constexpr (_S_size<_Tp> <= 8)
4418 return __vector_bitcast<_Tp, _S_size<_Tp>>(
4419 _mm_unpacklo_epi8(__k, __k));
4420 else
4421 return __concat(_mm_unpacklo_epi8(__k, __k),
4422 _mm_unpackhi_epi8(__k, __k));
4423 }
4424 else
4425 return _Base::template _S_load<_Tp>(__mem);
4426 }
4427
4428 // }}}
4429 // _S_from_bitmask{{{
4430 template <size_t _Np, typename _Tp>
4431 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
4432 _S_from_bitmask(_SanitizedBitMask<_Np> __bits, _TypeTag<_Tp>)
4433 {
4434 static_assert(is_same_v<_Tp, __int_for_sizeof_t<_Tp>>);
4435 if constexpr (__is_avx512_abi<_Abi>())
4436 return __bits._M_to_bits();
4437 else
4438 return _S_to_maskvector<_Tp, _S_size<_Tp>>(__bits);
4439 }
4440
4441 // }}}
4442 // _S_masked_load {{{2
4443 template <typename _Tp, size_t _Np>
4444 static inline _SimdWrapper<_Tp, _Np>
4445 _S_masked_load(_SimdWrapper<_Tp, _Np> __merge,
4446 _SimdWrapper<_Tp, _Np> __mask, const bool* __mem) noexcept
4447 {
4448 if constexpr (__is_avx512_abi<_Abi>())
4449 {
4450 if constexpr (__have_avx512bw_vl)
4451 {
4452 if constexpr (_Np <= 16)
4453 {
4454 const auto __a
4455 = _mm_mask_loadu_epi8(__m128i(), __mask, __mem);
4456 return (__merge & ~__mask) | _mm_test_epi8_mask(__a, __a);
4457 }
4458 else if constexpr (_Np <= 32)
4459 {
4460 const auto __a
4461 = _mm256_mask_loadu_epi8(__m256i(), __mask, __mem);
4462 return (__merge & ~__mask)
4463 | _mm256_test_epi8_mask(__a, __a);
4464 }
4465 else if constexpr (_Np <= 64)
4466 {
4467 const auto __a
4468 = _mm512_mask_loadu_epi8(__m512i(), __mask, __mem);
4469 return (__merge & ~__mask)
4470 | _mm512_test_epi8_mask(__a, __a);
4471 }
4472 else
4473 __assert_unreachable<_Tp>();
4474 }
4475 else
4476 {
4477 _BitOps::_S_bit_iteration(__mask, [&](auto __i) {
4478 __merge._M_set(__i, __mem[__i]);
4479 });
4480 return __merge;
4481 }
4482 }
4483 else if constexpr (__have_avx512bw_vl && _Np == 32 && sizeof(_Tp) == 1)
4484 {
4485 const auto __k = _S_to_bits(__mask)._M_to_bits();
4486 __merge = _mm256_mask_sub_epi8(__to_intrin(__merge), __k, __m256i(),
4487 _mm256_mask_loadu_epi8(__m256i(),
4488 __k, __mem));
4489 }
4490 else if constexpr (__have_avx512bw_vl && _Np == 16 && sizeof(_Tp) == 1)
4491 {
4492 const auto __k = _S_to_bits(__mask)._M_to_bits();
4493 __merge
4494 = _mm_mask_sub_epi8(__vector_bitcast<_LLong>(__merge), __k,
4495 __m128i(),
4496 _mm_mask_loadu_epi8(__m128i(), __k, __mem));
4497 }
4498 else if constexpr (__have_avx512bw_vl && _Np == 16 && sizeof(_Tp) == 2)
4499 {
4500 const auto __k = _S_to_bits(__mask)._M_to_bits();
4501 __merge = _mm256_mask_sub_epi16(
4502 __vector_bitcast<_LLong>(__merge), __k, __m256i(),
4503 _mm256_cvtepi8_epi16(_mm_mask_loadu_epi8(__m128i(), __k, __mem)));
4504 }
4505 else if constexpr (__have_avx512bw_vl && _Np == 8 && sizeof(_Tp) == 2)
4506 {
4507 const auto __k = _S_to_bits(__mask)._M_to_bits();
4508 __merge = _mm_mask_sub_epi16(
4509 __vector_bitcast<_LLong>(__merge), __k, __m128i(),
4510 _mm_cvtepi8_epi16(_mm_mask_loadu_epi8(__m128i(), __k, __mem)));
4511 }
4512 else if constexpr (__have_avx512bw_vl && _Np == 8 && sizeof(_Tp) == 4)
4513 {
4514 const auto __k = _S_to_bits(__mask)._M_to_bits();
4515 __merge = __vector_bitcast<_Tp>(_mm256_mask_sub_epi32(
4516 __vector_bitcast<_LLong>(__merge), __k, __m256i(),
4517 _mm256_cvtepi8_epi32(
4518 _mm_mask_loadu_epi8(__m128i(), __k, __mem))));
4519 }
4520 else if constexpr (__have_avx512bw_vl && _Np == 4 && sizeof(_Tp) == 4)
4521 {
4522 const auto __k = _S_to_bits(__mask)._M_to_bits();
4523 __merge = __vector_bitcast<_Tp>(_mm_mask_sub_epi32(
4524 __vector_bitcast<_LLong>(__merge), __k, __m128i(),
4525 _mm_cvtepi8_epi32(_mm_mask_loadu_epi8(__m128i(), __k, __mem))));
4526 }
4527 else if constexpr (__have_avx512bw_vl && _Np == 4 && sizeof(_Tp) == 8)
4528 {
4529 const auto __k = _S_to_bits(__mask)._M_to_bits();
4530 __merge = __vector_bitcast<_Tp>(_mm256_mask_sub_epi64(
4531 __vector_bitcast<_LLong>(__merge), __k, __m256i(),
4532 _mm256_cvtepi8_epi64(
4533 _mm_mask_loadu_epi8(__m128i(), __k, __mem))));
4534 }
4535 else if constexpr (__have_avx512bw_vl && _Np == 2 && sizeof(_Tp) == 8)
4536 {
4537 const auto __k = _S_to_bits(__mask)._M_to_bits();
4538 __merge = __vector_bitcast<_Tp>(_mm_mask_sub_epi64(
4539 __vector_bitcast<_LLong>(__merge), __k, __m128i(),
4540 _mm_cvtepi8_epi64(_mm_mask_loadu_epi8(__m128i(), __k, __mem))));
4541 }
4542 else
4543 return _Base::_S_masked_load(__merge, __mask, __mem);
4544 return __merge;
4545 }
4546
4547 // _S_store {{{2
4548 template <typename _Tp, size_t _Np>
4549 _GLIBCXX_SIMD_INTRINSIC static void _S_store(_SimdWrapper<_Tp, _Np> __v,
4550 bool* __mem) noexcept
4551 {
4552 if constexpr (__is_avx512_abi<_Abi>())
4553 {
4554 if constexpr (__have_avx512bw_vl)
4555 _CommonImplX86::_S_store<_Np>(
4556 __vector_bitcast<char>([](auto __data) {
4557 if constexpr (_Np <= 16)
4558 return _mm_maskz_set1_epi8(__data, 1);
4559 else if constexpr (_Np <= 32)
4560 return _mm256_maskz_set1_epi8(__data, 1);
4561 else
4562 return _mm512_maskz_set1_epi8(__data, 1);
4563 }(__v._M_data)),
4564 __mem);
4565 else if constexpr (_Np <= 8)
4566 _CommonImplX86::_S_store<_Np>(
4567 __vector_bitcast<char>(
4568#if defined __x86_64__
4569 __make_wrapper<_ULLong>(
4570 _pdep_u64(__v._M_data, 0x0101010101010101ULL), 0ull)
4571#else
4572 __make_wrapper<_UInt>(_pdep_u32(__v._M_data, 0x01010101U),
4573 _pdep_u32(__v._M_data >> 4,
4574 0x01010101U))
4575#endif
4576 ),
4577 __mem);
4578 else if constexpr (_Np <= 16)
4579 _mm512_mask_cvtepi32_storeu_epi8(
4580 __mem, 0xffffu >> (16 - _Np),
4581 _mm512_maskz_set1_epi32(__v._M_data, 1));
4582 else
4583 __assert_unreachable<_Tp>();
4584 }
4585 else if constexpr (__is_sse_abi<_Abi>()) //{{{
4586 {
4587 if constexpr (_Np == 2 && sizeof(_Tp) == 8)
4588 {
4589 const auto __k = __vector_bitcast<int>(__v);
4590 __mem[0] = -__k[1];
4591 __mem[1] = -__k[3];
4592 }
4593 else if constexpr (_Np <= 4 && sizeof(_Tp) == 4)
4594 {
4595 if constexpr (__have_sse2)
4596 {
4597 const unsigned __bool4
4598 = __vector_bitcast<_UInt>(_mm_packs_epi16(
4599 _mm_packs_epi32(__intrin_bitcast<__m128i>(
4600 __to_intrin(__v)),
4601 __m128i()),
4602 __m128i()))[0]
4603 & 0x01010101u;
4604 __builtin_memcpy(__mem, &__bool4, _Np);
4605 }
4606 else if constexpr (__have_mmx)
4607 {
4608 const __m64 __k = _mm_cvtps_pi8(
4609 __and(__to_intrin(__v), _mm_set1_ps(1.f)));
4610 __builtin_memcpy(__mem, &__k, _Np);
4611 _mm_empty();
4612 }
4613 else
4614 return _Base::_S_store(__v, __mem);
4615 }
4616 else if constexpr (_Np <= 8 && sizeof(_Tp) == 2)
4617 {
4618 _CommonImplX86::_S_store<_Np>(
4619 __vector_bitcast<char>(_mm_packs_epi16(
4620 __to_intrin(__vector_bitcast<_UShort>(__v) >> 15),
4621 __m128i())),
4622 __mem);
4623 }
4624 else if constexpr (_Np <= 16 && sizeof(_Tp) == 1)
4625 _CommonImplX86::_S_store<_Np>(__v._M_data & 1, __mem);
4626 else
4627 __assert_unreachable<_Tp>();
4628 } // }}}
4629 else if constexpr (__is_avx_abi<_Abi>()) // {{{
4630 {
4631 if constexpr (_Np <= 4 && sizeof(_Tp) == 8)
4632 {
4633 auto __k = __intrin_bitcast<__m256i>(__to_intrin(__v));
4634 int __bool4;
4635 if constexpr (__have_avx2)
4636 __bool4 = _mm256_movemask_epi8(__k);
4637 else
4638 __bool4 = (_mm_movemask_epi8(__lo128(__k))
4639 | (_mm_movemask_epi8(__hi128(__k)) << 16));
4640 __bool4 &= 0x01010101;
4641 __builtin_memcpy(__mem, &__bool4, _Np);
4642 }
4643 else if constexpr (_Np <= 8 && sizeof(_Tp) == 4)
4644 {
4645 const auto __k = __intrin_bitcast<__m256i>(__to_intrin(__v));
4646 const auto __k2
4647 = _mm_srli_epi16(_mm_packs_epi16(__lo128(__k), __hi128(__k)),
4648 15);
4649 const auto __k3
4650 = __vector_bitcast<char>(_mm_packs_epi16(__k2, __m128i()));
4651 _CommonImplX86::_S_store<_Np>(__k3, __mem);
4652 }
4653 else if constexpr (_Np <= 16 && sizeof(_Tp) == 2)
4654 {
4655 if constexpr (__have_avx2)
4656 {
4657 const auto __x = _mm256_srli_epi16(__to_intrin(__v), 15);
4658 const auto __bools = __vector_bitcast<char>(
4659 _mm_packs_epi16(__lo128(__x), __hi128(__x)));
4660 _CommonImplX86::_S_store<_Np>(__bools, __mem);
4661 }
4662 else
4663 {
4664 const auto __bools
4665 = 1
4666 & __vector_bitcast<_UChar>(
4667 _mm_packs_epi16(__lo128(__to_intrin(__v)),
4668 __hi128(__to_intrin(__v))));
4669 _CommonImplX86::_S_store<_Np>(__bools, __mem);
4670 }
4671 }
4672 else if constexpr (_Np <= 32 && sizeof(_Tp) == 1)
4673 _CommonImplX86::_S_store<_Np>(1 & __v._M_data, __mem);
4674 else
4675 __assert_unreachable<_Tp>();
4676 } // }}}
4677 else
4678 __assert_unreachable<_Tp>();
4679 }
4680
4681 // _S_masked_store {{{2
4682 template <typename _Tp, size_t _Np>
4683 static inline void
4684 _S_masked_store(const _SimdWrapper<_Tp, _Np> __v, bool* __mem,
4685 const _SimdWrapper<_Tp, _Np> __k) noexcept
4686 {
4687 if constexpr (__is_avx512_abi<_Abi>())
4688 {
4689 static_assert(is_same_v<_Tp, bool>);
4690 if constexpr (_Np <= 16 && __have_avx512bw_vl)
4691 _mm_mask_storeu_epi8(__mem, __k, _mm_maskz_set1_epi8(__v, 1));
4692 else if constexpr (_Np <= 16)
4693 _mm512_mask_cvtepi32_storeu_epi8(__mem, __k,
4694 _mm512_maskz_set1_epi32(__v, 1));
4695 else if constexpr (_Np <= 32 && __have_avx512bw_vl)
4696 _mm256_mask_storeu_epi8(__mem, __k,
4697 _mm256_maskz_set1_epi8(__v, 1));
4698 else if constexpr (_Np <= 32 && __have_avx512bw)
4699 _mm256_mask_storeu_epi8(__mem, __k,
4700 __lo256(_mm512_maskz_set1_epi8(__v, 1)));
4701 else if constexpr (_Np <= 64 && __have_avx512bw)
4702 _mm512_mask_storeu_epi8(__mem, __k,
4703 _mm512_maskz_set1_epi8(__v, 1));
4704 else
4705 __assert_unreachable<_Tp>();
4706 }
4707 else
4708 _Base::_S_masked_store(__v, __mem, __k);
4709 }
4710
4711 // logical and bitwise operators {{{2
4712 template <typename _Tp, size_t _Np>
4713 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
4714 _S_logical_and(const _SimdWrapper<_Tp, _Np>& __x,
4715 const _SimdWrapper<_Tp, _Np>& __y)
4716 {
4717 if constexpr (is_same_v<_Tp, bool>)
4718 {
4719 if constexpr (__have_avx512dq && _Np <= 8)
4720 return _kand_mask8(__x._M_data, __y._M_data);
4721 else if constexpr (_Np <= 16)
4722 return _kand_mask16(__x._M_data, __y._M_data);
4723 else if constexpr (__have_avx512bw && _Np <= 32)
4724 return _kand_mask32(__x._M_data, __y._M_data);
4725 else if constexpr (__have_avx512bw && _Np <= 64)
4726 return _kand_mask64(__x._M_data, __y._M_data);
4727 else
4728 __assert_unreachable<_Tp>();
4729 }
4730 else
4731 return _Base::_S_logical_and(__x, __y);
4732 }
4733
4734 template <typename _Tp, size_t _Np>
4735 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
4736 _S_logical_or(const _SimdWrapper<_Tp, _Np>& __x,
4737 const _SimdWrapper<_Tp, _Np>& __y)
4738 {
4739 if constexpr (is_same_v<_Tp, bool>)
4740 {
4741 if constexpr (__have_avx512dq && _Np <= 8)
4742 return _kor_mask8(__x._M_data, __y._M_data);
4743 else if constexpr (_Np <= 16)
4744 return _kor_mask16(__x._M_data, __y._M_data);
4745 else if constexpr (__have_avx512bw && _Np <= 32)
4746 return _kor_mask32(__x._M_data, __y._M_data);
4747 else if constexpr (__have_avx512bw && _Np <= 64)
4748 return _kor_mask64(__x._M_data, __y._M_data);
4749 else
4750 __assert_unreachable<_Tp>();
4751 }
4752 else
4753 return _Base::_S_logical_or(__x, __y);
4754 }
4755
4756 template <typename _Tp, size_t _Np>
4757 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
4758 _S_bit_not(const _SimdWrapper<_Tp, _Np>& __x)
4759 {
4760 if constexpr (is_same_v<_Tp, bool>)
4761 {
4762 if constexpr (__have_avx512dq && _Np <= 8)
4763 return _kandn_mask8(__x._M_data,
4764 _Abi::template __implicit_mask_n<_Np>());
4765 else if constexpr (_Np <= 16)
4766 return _kandn_mask16(__x._M_data,
4767 _Abi::template __implicit_mask_n<_Np>());
4768 else if constexpr (__have_avx512bw && _Np <= 32)
4769 return _kandn_mask32(__x._M_data,
4770 _Abi::template __implicit_mask_n<_Np>());
4771 else if constexpr (__have_avx512bw && _Np <= 64)
4772 return _kandn_mask64(__x._M_data,
4773 _Abi::template __implicit_mask_n<_Np>());
4774 else
4775 __assert_unreachable<_Tp>();
4776 }
4777 else
4778 return _Base::_S_bit_not(__x);
4779 }
4780
4781 template <typename _Tp, size_t _Np>
4782 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
4783 _S_bit_and(const _SimdWrapper<_Tp, _Np>& __x,
4784 const _SimdWrapper<_Tp, _Np>& __y)
4785 {
4786 if constexpr (is_same_v<_Tp, bool>)
4787 {
4788 if constexpr (__have_avx512dq && _Np <= 8)
4789 return _kand_mask8(__x._M_data, __y._M_data);
4790 else if constexpr (_Np <= 16)
4791 return _kand_mask16(__x._M_data, __y._M_data);
4792 else if constexpr (__have_avx512bw && _Np <= 32)
4793 return _kand_mask32(__x._M_data, __y._M_data);
4794 else if constexpr (__have_avx512bw && _Np <= 64)
4795 return _kand_mask64(__x._M_data, __y._M_data);
4796 else
4797 __assert_unreachable<_Tp>();
4798 }
4799 else
4800 return _Base::_S_bit_and(__x, __y);
4801 }
4802
4803 template <typename _Tp, size_t _Np>
4804 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
4805 _S_bit_or(const _SimdWrapper<_Tp, _Np>& __x,
4806 const _SimdWrapper<_Tp, _Np>& __y)
4807 {
4808 if constexpr (is_same_v<_Tp, bool>)
4809 {
4810 if constexpr (__have_avx512dq && _Np <= 8)
4811 return _kor_mask8(__x._M_data, __y._M_data);
4812 else if constexpr (_Np <= 16)
4813 return _kor_mask16(__x._M_data, __y._M_data);
4814 else if constexpr (__have_avx512bw && _Np <= 32)
4815 return _kor_mask32(__x._M_data, __y._M_data);
4816 else if constexpr (__have_avx512bw && _Np <= 64)
4817 return _kor_mask64(__x._M_data, __y._M_data);
4818 else
4819 __assert_unreachable<_Tp>();
4820 }
4821 else
4822 return _Base::_S_bit_or(__x, __y);
4823 }
4824
4825 template <typename _Tp, size_t _Np>
4826 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
4827 _S_bit_xor(const _SimdWrapper<_Tp, _Np>& __x,
4828 const _SimdWrapper<_Tp, _Np>& __y)
4829 {
4830 if constexpr (is_same_v<_Tp, bool>)
4831 {
4832 if constexpr (__have_avx512dq && _Np <= 8)
4833 return _kxor_mask8(__x._M_data, __y._M_data);
4834 else if constexpr (_Np <= 16)
4835 return _kxor_mask16(__x._M_data, __y._M_data);
4836 else if constexpr (__have_avx512bw && _Np <= 32)
4837 return _kxor_mask32(__x._M_data, __y._M_data);
4838 else if constexpr (__have_avx512bw && _Np <= 64)
4839 return _kxor_mask64(__x._M_data, __y._M_data);
4840 else
4841 __assert_unreachable<_Tp>();
4842 }
4843 else
4844 return _Base::_S_bit_xor(__x, __y);
4845 }
4846
4847 //}}}2
4848 // _S_masked_assign{{{
4849 template <size_t _Np>
4850 _GLIBCXX_SIMD_INTRINSIC static void
4851 _S_masked_assign(_SimdWrapper<bool, _Np> __k,
4852 _SimdWrapper<bool, _Np>& __lhs,
4853 _SimdWrapper<bool, _Np> __rhs)
4854 {
4855 __lhs._M_data
4856 = (~__k._M_data & __lhs._M_data) | (__k._M_data & __rhs._M_data);
4857 }
4858
4859 template <size_t _Np>
4860 _GLIBCXX_SIMD_INTRINSIC static void
4861 _S_masked_assign(_SimdWrapper<bool, _Np> __k,
4862 _SimdWrapper<bool, _Np>& __lhs, bool __rhs)
4863 {
4864 if (__rhs)
4865 __lhs._M_data = __k._M_data | __lhs._M_data;
4866 else
4867 __lhs._M_data = ~__k._M_data & __lhs._M_data;
4868 }
4869
4870 using _MaskImplBuiltin<_Abi>::_S_masked_assign;
4871
4872 //}}}
4873 // _S_all_of {{{
4874 template <typename _Tp>
4875 _GLIBCXX_SIMD_INTRINSIC static bool _S_all_of(simd_mask<_Tp, _Abi> __k)
4876 {
4877 if constexpr (__is_sse_abi<_Abi>() || __is_avx_abi<_Abi>())
4878 {
4879 constexpr size_t _Np = simd_size_v<_Tp, _Abi>;
4880 using _TI = __intrinsic_type_t<_Tp, _Np>;
4881 const _TI __a = reinterpret_cast<_TI>(__to_intrin(__data(__k)));
4882 if constexpr (__have_sse4_1)
4883 {
4884 _GLIBCXX_SIMD_USE_CONSTEXPR _TI __b
4885 = _Abi::template _S_implicit_mask_intrin<_Tp>();
4886 return 0 != __testc(__a, __b);
4887 }
4888 else if constexpr (is_same_v<_Tp, float>)
4889 return (_mm_movemask_ps(__a) & ((1 << _Np) - 1))
4890 == (1 << _Np) - 1;
4891 else if constexpr (is_same_v<_Tp, double>)
4892 return (_mm_movemask_pd(__a) & ((1 << _Np) - 1))
4893 == (1 << _Np) - 1;
4894 else
4895 return (_mm_movemask_epi8(__a) & ((1 << (_Np * sizeof(_Tp))) - 1))
4896 == (1 << (_Np * sizeof(_Tp))) - 1;
4897 }
4898 else if constexpr (__is_avx512_abi<_Abi>())
4899 {
4900 constexpr auto _Mask = _Abi::template _S_implicit_mask<_Tp>();
4901 const auto __kk = __k._M_data._M_data;
4902 if constexpr (sizeof(__kk) == 1)
4903 {
4904 if constexpr (__have_avx512dq)
4905 return _kortestc_mask8_u8(__kk, _Mask == 0xff
4906 ? __kk
4907 : __mmask8(~_Mask));
4908 else
4909 return _kortestc_mask16_u8(__kk, __mmask16(~_Mask));
4910 }
4911 else if constexpr (sizeof(__kk) == 2)
4912 return _kortestc_mask16_u8(__kk, _Mask == 0xffff
4913 ? __kk
4914 : __mmask16(~_Mask));
4915 else if constexpr (sizeof(__kk) == 4 && __have_avx512bw)
4916 return _kortestc_mask32_u8(__kk, _Mask == 0xffffffffU
4917 ? __kk
4918 : __mmask32(~_Mask));
4919 else if constexpr (sizeof(__kk) == 8 && __have_avx512bw)
4920 return _kortestc_mask64_u8(__kk, _Mask == 0xffffffffffffffffULL
4921 ? __kk
4922 : __mmask64(~_Mask));
4923 else
4924 __assert_unreachable<_Tp>();
4925 }
4926 }
4927
4928 // }}}
4929 // _S_any_of {{{
4930 template <typename _Tp>
4931 _GLIBCXX_SIMD_INTRINSIC static bool _S_any_of(simd_mask<_Tp, _Abi> __k)
4932 {
4933 if constexpr (__is_sse_abi<_Abi>() || __is_avx_abi<_Abi>())
4934 {
4935 constexpr size_t _Np = simd_size_v<_Tp, _Abi>;
4936 using _TI = __intrinsic_type_t<_Tp, _Np>;
4937 const _TI __a = reinterpret_cast<_TI>(__to_intrin(__data(__k)));
4938 if constexpr (__have_sse4_1)
4939 {
4940 if constexpr (_Abi::template _S_is_partial<
4941 _Tp> || sizeof(__k) < 16)
4942 {
4943 _GLIBCXX_SIMD_USE_CONSTEXPR _TI __b
4944 = _Abi::template _S_implicit_mask_intrin<_Tp>();
4945 return 0 == __testz(__a, __b);
4946 }
4947 else
4948 return 0 == __testz(__a, __a);
4949 }
4950 else if constexpr (is_same_v<_Tp, float>)
4951 return (_mm_movemask_ps(__a) & ((1 << _Np) - 1)) != 0;
4952 else if constexpr (is_same_v<_Tp, double>)
4953 return (_mm_movemask_pd(__a) & ((1 << _Np) - 1)) != 0;
4954 else
4955 return (_mm_movemask_epi8(__a) & ((1 << (_Np * sizeof(_Tp))) - 1))
4956 != 0;
4957 }
4958 else if constexpr (__is_avx512_abi<_Abi>())
4959 return (__k._M_data._M_data & _Abi::template _S_implicit_mask<_Tp>())
4960 != 0;
4961 }
4962
4963 // }}}
4964 // _S_none_of {{{
4965 template <typename _Tp>
4966 _GLIBCXX_SIMD_INTRINSIC static bool _S_none_of(simd_mask<_Tp, _Abi> __k)
4967 {
4968 if constexpr (__is_sse_abi<_Abi>() || __is_avx_abi<_Abi>())
4969 {
4970 constexpr size_t _Np = simd_size_v<_Tp, _Abi>;
4971 using _TI = __intrinsic_type_t<_Tp, _Np>;
4972 const _TI __a = reinterpret_cast<_TI>(__to_intrin(__data(__k)));
4973 if constexpr (__have_sse4_1)
4974 {
4975 if constexpr (_Abi::template _S_is_partial<
4976 _Tp> || sizeof(__k) < 16)
4977 {
4978 _GLIBCXX_SIMD_USE_CONSTEXPR _TI __b
4979 = _Abi::template _S_implicit_mask_intrin<_Tp>();
4980 return 0 != __testz(__a, __b);
4981 }
4982 else
4983 return 0 != __testz(__a, __a);
4984 }
4985 else if constexpr (is_same_v<_Tp, float>)
4986 return (__movemask(__a) & ((1 << _Np) - 1)) == 0;
4987 else if constexpr (is_same_v<_Tp, double>)
4988 return (__movemask(__a) & ((1 << _Np) - 1)) == 0;
4989 else
4990 return (__movemask(__a) & int((1ull << (_Np * sizeof(_Tp))) - 1))
4991 == 0;
4992 }
4993 else if constexpr (__is_avx512_abi<_Abi>())
4994 return (__k._M_data._M_data & _Abi::template _S_implicit_mask<_Tp>())
4995 == 0;
4996 }
4997
4998 // }}}
4999 // _S_some_of {{{
5000 template <typename _Tp>
5001 _GLIBCXX_SIMD_INTRINSIC static bool _S_some_of(simd_mask<_Tp, _Abi> __k)
5002 {
5003 if constexpr (__is_sse_abi<_Abi>() || __is_avx_abi<_Abi>())
5004 {
5005 constexpr size_t _Np = simd_size_v<_Tp, _Abi>;
5006 using _TI = __intrinsic_type_t<_Tp, _Np>;
5007 const _TI __a = reinterpret_cast<_TI>(__to_intrin(__data(__k)));
5008 if constexpr (__have_sse4_1)
5009 {
5010 _GLIBCXX_SIMD_USE_CONSTEXPR _TI __b
5011 = _Abi::template _S_implicit_mask_intrin<_Tp>();
5012 return 0 != __testnzc(__a, __b);
5013 }
5014 else if constexpr (is_same_v<_Tp, float>)
5015 {
5016 constexpr int __allbits = (1 << _Np) - 1;
5017 const auto __tmp = _mm_movemask_ps(__a) & __allbits;
5018 return __tmp > 0 && __tmp < __allbits;
5019 }
5020 else if constexpr (is_same_v<_Tp, double>)
5021 {
5022 constexpr int __allbits = (1 << _Np) - 1;
5023 const auto __tmp = _mm_movemask_pd(__a) & __allbits;
5024 return __tmp > 0 && __tmp < __allbits;
5025 }
5026 else
5027 {
5028 constexpr int __allbits = (1 << (_Np * sizeof(_Tp))) - 1;
5029 const auto __tmp = _mm_movemask_epi8(__a) & __allbits;
5030 return __tmp > 0 && __tmp < __allbits;
5031 }
5032 }
5033 else if constexpr (__is_avx512_abi<_Abi>())
5034 return _S_any_of(__k) && !_S_all_of(__k);
5035 else
5036 __assert_unreachable<_Tp>();
5037 }
5038
5039 // }}}
5040 // _S_popcount {{{
5041 template <typename _Tp>
5042 _GLIBCXX_SIMD_INTRINSIC static int _S_popcount(simd_mask<_Tp, _Abi> __k)
5043 {
5044 constexpr size_t _Np = simd_size_v<_Tp, _Abi>;
5045 const auto __kk = _Abi::_S_masked(__k._M_data)._M_data;
5046 if constexpr (__is_avx512_abi<_Abi>())
5047 {
5048 if constexpr (_Np > 32)
5049 return __builtin_popcountll(__kk);
5050 else
5051 return __builtin_popcount(__kk);
5052 }
5053 else
5054 {
5055 if constexpr (__have_popcnt)
5056 {
5057 int __bits
5058 = __movemask(__to_intrin(__vector_bitcast<_Tp>(__kk)));
5059 const int __count = __builtin_popcount(__bits);
5060 return is_integral_v<_Tp> ? __count / sizeof(_Tp) : __count;
5061 }
5062 else if constexpr (_Np == 2 && sizeof(_Tp) == 8)
5063 {
5064 const int mask = _mm_movemask_pd(__auto_bitcast(__kk));
5065 return mask - (mask >> 1);
5066 }
5067 else if constexpr (_Np <= 4 && sizeof(_Tp) == 8)
5068 {
5069 auto __x = -(__lo128(__kk) + __hi128(__kk));
5070 return __x[0] + __x[1];
5071 }
5072 else if constexpr (_Np <= 4 && sizeof(_Tp) == 4)
5073 {
5074 if constexpr (__have_sse2)
5075 {
5076 __m128i __x = __intrin_bitcast<__m128i>(__to_intrin(__kk));
5077 __x = _mm_add_epi32(
5078 __x, _mm_shuffle_epi32(__x, _MM_SHUFFLE(0, 1, 2, 3)));
5079 __x = _mm_add_epi32(
5080 __x, _mm_shufflelo_epi16(__x, _MM_SHUFFLE(1, 0, 3, 2)));
5081 return -_mm_cvtsi128_si32(__x);
5082 }
5083 else
5084 return __builtin_popcount(
5085 _mm_movemask_ps(__auto_bitcast(__kk)));
5086 }
5087 else if constexpr (_Np <= 8 && sizeof(_Tp) == 2)
5088 {
5089 auto __x = __to_intrin(__kk);
5090 __x = _mm_add_epi16(__x,
5091 _mm_shuffle_epi32(__x,
5092 _MM_SHUFFLE(0, 1, 2, 3)));
5093 __x = _mm_add_epi16(
5094 __x, _mm_shufflelo_epi16(__x, _MM_SHUFFLE(0, 1, 2, 3)));
5095 __x = _mm_add_epi16(
5096 __x, _mm_shufflelo_epi16(__x, _MM_SHUFFLE(2, 3, 0, 1)));
5097 return -short(_mm_extract_epi16(__x, 0));
5098 }
5099 else if constexpr (_Np <= 16 && sizeof(_Tp) == 1)
5100 {
5101 auto __x = __to_intrin(__kk);
5102 __x = _mm_add_epi8(__x,
5103 _mm_shuffle_epi32(__x,
5104 _MM_SHUFFLE(0, 1, 2, 3)));
5105 __x = _mm_add_epi8(__x,
5106 _mm_shufflelo_epi16(__x, _MM_SHUFFLE(0, 1, 2,
5107 3)));
5108 __x = _mm_add_epi8(__x,
5109 _mm_shufflelo_epi16(__x, _MM_SHUFFLE(2, 3, 0,
5110 1)));
5111 auto __y = -__vector_bitcast<_UChar>(__x);
5112 if constexpr (__have_sse4_1)
5113 return __y[0] + __y[1];
5114 else
5115 {
5116 unsigned __z = _mm_extract_epi16(__to_intrin(__y), 0);
5117 return (__z & 0xff) + (__z >> 8);
5118 }
5119 }
5120 else if constexpr (sizeof(__kk) == 32)
5121 {
5122 // The following works only as long as the implementations above
5123 // use a summation
5124 using _I = __int_for_sizeof_t<_Tp>;
5125 const auto __as_int = __vector_bitcast<_I>(__kk);
5126 _MaskImplX86<simd_abi::__sse>::_S_popcount(
5127 simd_mask<_I, simd_abi::__sse>(__private_init,
5128 __lo128(__as_int)
5129 + __hi128(__as_int)));
5130 }
5131 else
5132 __assert_unreachable<_Tp>();
5133 }
5134 }
5135
5136 // }}}
5137 // _S_find_first_set {{{
5138 template <typename _Tp>
5139 _GLIBCXX_SIMD_INTRINSIC static int
5140 _S_find_first_set(simd_mask<_Tp, _Abi> __k)
5141 {
5142 if constexpr (__is_avx512_abi<_Abi>())
5143 return std::__countr_zero(__k._M_data._M_data);
5144 else
5145 return _Base::_S_find_first_set(__k);
5146 }
5147
5148 // }}}
5149 // _S_find_last_set {{{
5150 template <typename _Tp>
5151 _GLIBCXX_SIMD_INTRINSIC static int
5152 _S_find_last_set(simd_mask<_Tp, _Abi> __k)
5153 {
5154 if constexpr (__is_avx512_abi<_Abi>())
5155 return std::__bit_width(__k._M_data._M_data) - 1;
5156 else
5157 return _Base::_S_find_last_set(__k);
5158 }
5159
5160 // }}}
5161 };
5162
5163// }}}
5164
5165_GLIBCXX_SIMD_END_NAMESPACE
5166#endif // __cplusplus >= 201703L
5167#endif // _GLIBCXX_EXPERIMENTAL_SIMD_X86_H_
5168
5169// vim: foldmethod=marker sw=2 noet ts=8 sts=2 tw=80
Note: See TracBrowser for help on using the repository browser.