libstdc++
simd_builtin.h
1// Simd Abi specific implementations -*- C++ -*-
2
3// Copyright (C) 2020-2023 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_ABIS_H_
26#define _GLIBCXX_EXPERIMENTAL_SIMD_ABIS_H_
27
28#if __cplusplus >= 201703L
29
30#include <array>
31#include <cmath>
32#include <cstdlib>
33
34_GLIBCXX_SIMD_BEGIN_NAMESPACE
35// _S_allbits{{{
36template <typename _V>
37 static inline _GLIBCXX_SIMD_USE_CONSTEXPR _V _S_allbits
38 = reinterpret_cast<_V>(~__vector_type_t<char, sizeof(_V) / sizeof(char)>());
39
40// }}}
41// _S_signmask, _S_absmask{{{
42template <typename _V, typename = _VectorTraits<_V>>
43 static inline _GLIBCXX_SIMD_USE_CONSTEXPR _V _S_signmask
44 = __xor(_V() + 1, _V() - 1);
45
46template <typename _V, typename = _VectorTraits<_V>>
47 static inline _GLIBCXX_SIMD_USE_CONSTEXPR _V _S_absmask
48 = __andnot(_S_signmask<_V>, _S_allbits<_V>);
49
50//}}}
51// __vector_permute<Indices...>{{{
52// Index == -1 requests zeroing of the output element
53template <int... _Indices, typename _Tp, typename _TVT = _VectorTraits<_Tp>,
54 typename = __detail::__odr_helper>
55 _Tp
56 __vector_permute(_Tp __x)
57 {
58 static_assert(sizeof...(_Indices) == _TVT::_S_full_size);
59 return __make_vector<typename _TVT::value_type>(
60 (_Indices == -1 ? 0 : __x[_Indices == -1 ? 0 : _Indices])...);
61 }
62
63// }}}
64// __vector_shuffle<Indices...>{{{
65// Index == -1 requests zeroing of the output element
66template <int... _Indices, typename _Tp, typename _TVT = _VectorTraits<_Tp>,
67 typename = __detail::__odr_helper>
68 _Tp
69 __vector_shuffle(_Tp __x, _Tp __y)
70 {
71 return _Tp{(_Indices == -1 ? 0
72 : _Indices < _TVT::_S_full_size
73 ? __x[_Indices]
74 : __y[_Indices - _TVT::_S_full_size])...};
75 }
76
77// }}}
78// __make_wrapper{{{
79template <typename _Tp, typename... _Args>
80 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper<_Tp, sizeof...(_Args)>
81 __make_wrapper(const _Args&... __args)
82 { return __make_vector<_Tp>(__args...); }
83
84// }}}
85// __wrapper_bitcast{{{
86template <typename _Tp, size_t _ToN = 0, typename _Up, size_t _M,
87 size_t _Np = _ToN != 0 ? _ToN : sizeof(_Up) * _M / sizeof(_Tp)>
88 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper<_Tp, _Np>
89 __wrapper_bitcast(_SimdWrapper<_Up, _M> __x)
90 {
91 static_assert(_Np > 1);
92 return __intrin_bitcast<__vector_type_t<_Tp, _Np>>(__x._M_data);
93 }
94
95// }}}
96// __shift_elements_right{{{
97// if (__shift % 2ⁿ == 0) => the low n Bytes are correct
98template <unsigned __shift, typename _Tp, typename _TVT = _VectorTraits<_Tp>>
99 _GLIBCXX_SIMD_INTRINSIC _Tp
100 __shift_elements_right(_Tp __v)
101 {
102 [[maybe_unused]] const auto __iv = __to_intrin(__v);
103 static_assert(__shift <= sizeof(_Tp));
104 if constexpr (__shift == 0)
105 return __v;
106 else if constexpr (__shift == sizeof(_Tp))
107 return _Tp();
108#if _GLIBCXX_SIMD_X86INTRIN // {{{
109 else if constexpr (__have_sse && __shift == 8
110 && _TVT::template _S_is<float, 4>)
111 return _mm_movehl_ps(__iv, __iv);
112 else if constexpr (__have_sse2 && __shift == 8
113 && _TVT::template _S_is<double, 2>)
114 return _mm_unpackhi_pd(__iv, __iv);
115 else if constexpr (__have_sse2 && sizeof(_Tp) == 16)
116 return reinterpret_cast<typename _TVT::type>(
117 _mm_srli_si128(reinterpret_cast<__m128i>(__iv), __shift));
118 else if constexpr (__shift == 16 && sizeof(_Tp) == 32)
119 {
120 /*if constexpr (__have_avx && _TVT::template _S_is<double, 4>)
121 return _mm256_permute2f128_pd(__iv, __iv, 0x81);
122 else if constexpr (__have_avx && _TVT::template _S_is<float, 8>)
123 return _mm256_permute2f128_ps(__iv, __iv, 0x81);
124 else if constexpr (__have_avx)
125 return reinterpret_cast<typename _TVT::type>(
126 _mm256_permute2f128_si256(__iv, __iv, 0x81));
127 else*/
128 return __zero_extend(__hi128(__v));
129 }
130 else if constexpr (__have_avx2 && sizeof(_Tp) == 32 && __shift < 16)
131 {
132 const auto __vll = __vector_bitcast<_LLong>(__v);
133 return reinterpret_cast<typename _TVT::type>(
134 _mm256_alignr_epi8(_mm256_permute2x128_si256(__vll, __vll, 0x81),
135 __vll, __shift));
136 }
137 else if constexpr (__have_avx && sizeof(_Tp) == 32 && __shift < 16)
138 {
139 const auto __vll = __vector_bitcast<_LLong>(__v);
140 return reinterpret_cast<typename _TVT::type>(
141 __concat(_mm_alignr_epi8(__hi128(__vll), __lo128(__vll), __shift),
142 _mm_srli_si128(__hi128(__vll), __shift)));
143 }
144 else if constexpr (sizeof(_Tp) == 32 && __shift > 16)
145 return __zero_extend(__shift_elements_right<__shift - 16>(__hi128(__v)));
146 else if constexpr (sizeof(_Tp) == 64 && __shift == 32)
147 return __zero_extend(__hi256(__v));
148 else if constexpr (__have_avx512f && sizeof(_Tp) == 64)
149 {
150 if constexpr (__shift >= 48)
151 return __zero_extend(
152 __shift_elements_right<__shift - 48>(__extract<3, 4>(__v)));
153 else if constexpr (__shift >= 32)
154 return __zero_extend(
155 __shift_elements_right<__shift - 32>(__hi256(__v)));
156 else if constexpr (__shift % 8 == 0)
157 return reinterpret_cast<typename _TVT::type>(
158 _mm512_alignr_epi64(__m512i(), __intrin_bitcast<__m512i>(__v),
159 __shift / 8));
160 else if constexpr (__shift % 4 == 0)
161 return reinterpret_cast<typename _TVT::type>(
162 _mm512_alignr_epi32(__m512i(), __intrin_bitcast<__m512i>(__v),
163 __shift / 4));
164 else if constexpr (__have_avx512bw && __shift < 16)
165 {
166 const auto __vll = __vector_bitcast<_LLong>(__v);
167 return reinterpret_cast<typename _TVT::type>(
168 _mm512_alignr_epi8(_mm512_shuffle_i32x4(__vll, __vll, 0xf9),
169 __vll, __shift));
170 }
171 else if constexpr (__have_avx512bw && __shift < 32)
172 {
173 const auto __vll = __vector_bitcast<_LLong>(__v);
174 return reinterpret_cast<typename _TVT::type>(
175 _mm512_alignr_epi8(_mm512_shuffle_i32x4(__vll, __m512i(), 0xee),
176 _mm512_shuffle_i32x4(__vll, __vll, 0xf9),
177 __shift - 16));
178 }
179 else
180 __assert_unreachable<_Tp>();
181 }
182 /*
183 } else if constexpr (__shift % 16 == 0 && sizeof(_Tp) == 64)
184 return __auto_bitcast(__extract<__shift / 16, 4>(__v));
185 */
186#endif // _GLIBCXX_SIMD_X86INTRIN }}}
187 else
188 {
189 constexpr int __chunksize = __shift % 8 == 0 ? 8
190 : __shift % 4 == 0 ? 4
191 : __shift % 2 == 0 ? 2
192 : 1;
193 auto __w = __vector_bitcast<__int_with_sizeof_t<__chunksize>>(__v);
194 using _Up = decltype(__w);
195 return __intrin_bitcast<_Tp>(
196 __call_with_n_evaluations<(sizeof(_Tp) - __shift) / __chunksize>(
197 [](auto... __chunks) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
198 return _Up{__chunks...};
199 }, [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
200 return __w[__shift / __chunksize + __i];
201 }));
202 }
203 }
204
205// }}}
206// __extract_part(_SimdWrapper<_Tp, _Np>) {{{
207template <int _Index, int _Total, int _Combine, typename _Tp, size_t _Np>
208 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_CONST
209 _SimdWrapper<_Tp, _Np / _Total * _Combine>
210 __extract_part(const _SimdWrapper<_Tp, _Np> __x)
211 {
212 if constexpr (_Index % 2 == 0 && _Total % 2 == 0 && _Combine % 2 == 0)
213 return __extract_part<_Index / 2, _Total / 2, _Combine / 2>(__x);
214 else
215 {
216 constexpr size_t __values_per_part = _Np / _Total;
217 constexpr size_t __values_to_skip = _Index * __values_per_part;
218 constexpr size_t __return_size = __values_per_part * _Combine;
219 using _R = __vector_type_t<_Tp, __return_size>;
220 static_assert((_Index + _Combine) * __values_per_part * sizeof(_Tp)
221 <= sizeof(__x),
222 "out of bounds __extract_part");
223 // the following assertion would ensure no "padding" to be read
224 // static_assert(_Total >= _Index + _Combine, "_Total must be greater
225 // than _Index");
226
227 // static_assert(__return_size * _Total == _Np, "_Np must be divisible
228 // by _Total");
229 if (__x._M_is_constprop())
230 return __generate_from_n_evaluations<__return_size, _R>(
231 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
232 return __x[__values_to_skip + __i];
233 });
234 if constexpr (_Index == 0 && _Total == 1)
235 return __x;
236 else if constexpr (_Index == 0)
237 return __intrin_bitcast<_R>(__as_vector(__x));
238#if _GLIBCXX_SIMD_X86INTRIN // {{{
239 else if constexpr (sizeof(__x) == 32
240 && __return_size * sizeof(_Tp) <= 16)
241 {
242 constexpr size_t __bytes_to_skip = __values_to_skip * sizeof(_Tp);
243 if constexpr (__bytes_to_skip == 16)
244 return __vector_bitcast<_Tp, __return_size>(
245 __hi128(__as_vector(__x)));
246 else
247 return __vector_bitcast<_Tp, __return_size>(
248 _mm_alignr_epi8(__hi128(__vector_bitcast<_LLong>(__x)),
249 __lo128(__vector_bitcast<_LLong>(__x)),
250 __bytes_to_skip));
251 }
252#endif // _GLIBCXX_SIMD_X86INTRIN }}}
253 else if constexpr (_Index > 0
254 && (__values_to_skip % __return_size != 0
255 || sizeof(_R) >= 8)
256 && (__values_to_skip + __return_size) * sizeof(_Tp)
257 <= 64
258 && sizeof(__x) >= 16)
259 return __intrin_bitcast<_R>(
260 __shift_elements_right<__values_to_skip * sizeof(_Tp)>(
261 __as_vector(__x)));
262 else
263 {
264 _R __r = {};
265 __builtin_memcpy(&__r,
266 reinterpret_cast<const char*>(&__x)
267 + sizeof(_Tp) * __values_to_skip,
268 __return_size * sizeof(_Tp));
269 return __r;
270 }
271 }
272 }
273
274// }}}
275// __extract_part(_SimdWrapper<bool, _Np>) {{{
276template <int _Index, int _Total, int _Combine = 1, size_t _Np>
277 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper<bool, _Np / _Total * _Combine>
278 __extract_part(const _SimdWrapper<bool, _Np> __x)
279 {
280 static_assert(_Combine == 1, "_Combine != 1 not implemented");
281 static_assert(__have_avx512f && _Np == _Np);
282 static_assert(_Total >= 2 && _Index + _Combine <= _Total && _Index >= 0);
283 return __x._M_data >> (_Index * _Np / _Total);
284 }
285
286// }}}
287
288// __vector_convert {{{
289// implementation requires an index sequence
290template <typename _To, typename _From, size_t... _I>
291 _GLIBCXX_SIMD_INTRINSIC constexpr _To
292 __vector_convert(_From __a, index_sequence<_I...>)
293 {
294 using _Tp = typename _VectorTraits<_To>::value_type;
295 return _To{static_cast<_Tp>(__a[_I])...};
296 }
297
298template <typename _To, typename _From, size_t... _I>
299 _GLIBCXX_SIMD_INTRINSIC constexpr _To
300 __vector_convert(_From __a, _From __b, index_sequence<_I...>)
301 {
302 using _Tp = typename _VectorTraits<_To>::value_type;
303 return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...};
304 }
305
306template <typename _To, typename _From, size_t... _I>
307 _GLIBCXX_SIMD_INTRINSIC constexpr _To
308 __vector_convert(_From __a, _From __b, _From __c, index_sequence<_I...>)
309 {
310 using _Tp = typename _VectorTraits<_To>::value_type;
311 return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
312 static_cast<_Tp>(__c[_I])...};
313 }
314
315template <typename _To, typename _From, size_t... _I>
316 _GLIBCXX_SIMD_INTRINSIC constexpr _To
317 __vector_convert(_From __a, _From __b, _From __c, _From __d,
318 index_sequence<_I...>)
319 {
320 using _Tp = typename _VectorTraits<_To>::value_type;
321 return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
322 static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...};
323 }
324
325template <typename _To, typename _From, size_t... _I>
326 _GLIBCXX_SIMD_INTRINSIC constexpr _To
327 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
328 index_sequence<_I...>)
329 {
330 using _Tp = typename _VectorTraits<_To>::value_type;
331 return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
332 static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
333 static_cast<_Tp>(__e[_I])...};
334 }
335
336template <typename _To, typename _From, size_t... _I>
337 _GLIBCXX_SIMD_INTRINSIC constexpr _To
338 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
339 _From __f, index_sequence<_I...>)
340 {
341 using _Tp = typename _VectorTraits<_To>::value_type;
342 return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
343 static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
344 static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...};
345 }
346
347template <typename _To, typename _From, size_t... _I>
348 _GLIBCXX_SIMD_INTRINSIC constexpr _To
349 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
350 _From __f, _From __g, index_sequence<_I...>)
351 {
352 using _Tp = typename _VectorTraits<_To>::value_type;
353 return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
354 static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
355 static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...,
356 static_cast<_Tp>(__g[_I])...};
357 }
358
359template <typename _To, typename _From, size_t... _I>
360 _GLIBCXX_SIMD_INTRINSIC constexpr _To
361 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
362 _From __f, _From __g, _From __h, index_sequence<_I...>)
363 {
364 using _Tp = typename _VectorTraits<_To>::value_type;
365 return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
366 static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
367 static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...,
368 static_cast<_Tp>(__g[_I])..., static_cast<_Tp>(__h[_I])...};
369 }
370
371template <typename _To, typename _From, size_t... _I>
372 _GLIBCXX_SIMD_INTRINSIC constexpr _To
373 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
374 _From __f, _From __g, _From __h, _From __i,
375 index_sequence<_I...>)
376 {
377 using _Tp = typename _VectorTraits<_To>::value_type;
378 return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
379 static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
380 static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...,
381 static_cast<_Tp>(__g[_I])..., static_cast<_Tp>(__h[_I])...,
382 static_cast<_Tp>(__i[_I])...};
383 }
384
385template <typename _To, typename _From, size_t... _I>
386 _GLIBCXX_SIMD_INTRINSIC constexpr _To
387 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
388 _From __f, _From __g, _From __h, _From __i, _From __j,
389 index_sequence<_I...>)
390 {
391 using _Tp = typename _VectorTraits<_To>::value_type;
392 return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
393 static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
394 static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...,
395 static_cast<_Tp>(__g[_I])..., static_cast<_Tp>(__h[_I])...,
396 static_cast<_Tp>(__i[_I])..., static_cast<_Tp>(__j[_I])...};
397 }
398
399template <typename _To, typename _From, size_t... _I>
400 _GLIBCXX_SIMD_INTRINSIC constexpr _To
401 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
402 _From __f, _From __g, _From __h, _From __i, _From __j,
403 _From __k, index_sequence<_I...>)
404 {
405 using _Tp = typename _VectorTraits<_To>::value_type;
406 return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
407 static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
408 static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...,
409 static_cast<_Tp>(__g[_I])..., static_cast<_Tp>(__h[_I])...,
410 static_cast<_Tp>(__i[_I])..., static_cast<_Tp>(__j[_I])...,
411 static_cast<_Tp>(__k[_I])...};
412 }
413
414template <typename _To, typename _From, size_t... _I>
415 _GLIBCXX_SIMD_INTRINSIC constexpr _To
416 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
417 _From __f, _From __g, _From __h, _From __i, _From __j,
418 _From __k, _From __l, index_sequence<_I...>)
419 {
420 using _Tp = typename _VectorTraits<_To>::value_type;
421 return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
422 static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
423 static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...,
424 static_cast<_Tp>(__g[_I])..., static_cast<_Tp>(__h[_I])...,
425 static_cast<_Tp>(__i[_I])..., static_cast<_Tp>(__j[_I])...,
426 static_cast<_Tp>(__k[_I])..., static_cast<_Tp>(__l[_I])...};
427 }
428
429template <typename _To, typename _From, size_t... _I>
430 _GLIBCXX_SIMD_INTRINSIC constexpr _To
431 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
432 _From __f, _From __g, _From __h, _From __i, _From __j,
433 _From __k, _From __l, _From __m, index_sequence<_I...>)
434 {
435 using _Tp = typename _VectorTraits<_To>::value_type;
436 return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
437 static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
438 static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...,
439 static_cast<_Tp>(__g[_I])..., static_cast<_Tp>(__h[_I])...,
440 static_cast<_Tp>(__i[_I])..., static_cast<_Tp>(__j[_I])...,
441 static_cast<_Tp>(__k[_I])..., static_cast<_Tp>(__l[_I])...,
442 static_cast<_Tp>(__m[_I])...};
443 }
444
445template <typename _To, typename _From, size_t... _I>
446 _GLIBCXX_SIMD_INTRINSIC constexpr _To
447 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
448 _From __f, _From __g, _From __h, _From __i, _From __j,
449 _From __k, _From __l, _From __m, _From __n,
450 index_sequence<_I...>)
451 {
452 using _Tp = typename _VectorTraits<_To>::value_type;
453 return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
454 static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
455 static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...,
456 static_cast<_Tp>(__g[_I])..., static_cast<_Tp>(__h[_I])...,
457 static_cast<_Tp>(__i[_I])..., static_cast<_Tp>(__j[_I])...,
458 static_cast<_Tp>(__k[_I])..., static_cast<_Tp>(__l[_I])...,
459 static_cast<_Tp>(__m[_I])..., static_cast<_Tp>(__n[_I])...};
460 }
461
462template <typename _To, typename _From, size_t... _I>
463 _GLIBCXX_SIMD_INTRINSIC constexpr _To
464 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
465 _From __f, _From __g, _From __h, _From __i, _From __j,
466 _From __k, _From __l, _From __m, _From __n, _From __o,
467 index_sequence<_I...>)
468 {
469 using _Tp = typename _VectorTraits<_To>::value_type;
470 return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
471 static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
472 static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...,
473 static_cast<_Tp>(__g[_I])..., static_cast<_Tp>(__h[_I])...,
474 static_cast<_Tp>(__i[_I])..., static_cast<_Tp>(__j[_I])...,
475 static_cast<_Tp>(__k[_I])..., static_cast<_Tp>(__l[_I])...,
476 static_cast<_Tp>(__m[_I])..., static_cast<_Tp>(__n[_I])...,
477 static_cast<_Tp>(__o[_I])...};
478 }
479
480template <typename _To, typename _From, size_t... _I>
481 _GLIBCXX_SIMD_INTRINSIC constexpr _To
482 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
483 _From __f, _From __g, _From __h, _From __i, _From __j,
484 _From __k, _From __l, _From __m, _From __n, _From __o,
485 _From __p, index_sequence<_I...>)
486 {
487 using _Tp = typename _VectorTraits<_To>::value_type;
488 return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
489 static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
490 static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...,
491 static_cast<_Tp>(__g[_I])..., static_cast<_Tp>(__h[_I])...,
492 static_cast<_Tp>(__i[_I])..., static_cast<_Tp>(__j[_I])...,
493 static_cast<_Tp>(__k[_I])..., static_cast<_Tp>(__l[_I])...,
494 static_cast<_Tp>(__m[_I])..., static_cast<_Tp>(__n[_I])...,
495 static_cast<_Tp>(__o[_I])..., static_cast<_Tp>(__p[_I])...};
496 }
497
498// Defer actual conversion to the overload that takes an index sequence. Note
499// that this function adds zeros or drops values off the end if you don't ensure
500// matching width.
501template <typename _To, typename... _From, size_t _FromSize>
502 _GLIBCXX_SIMD_INTRINSIC constexpr _To
503 __vector_convert(_SimdWrapper<_From, _FromSize>... __xs)
504 {
505#ifdef _GLIBCXX_SIMD_WORKAROUND_PR85048
506 using _From0 = __first_of_pack_t<_From...>;
507 using _FW = _SimdWrapper<_From0, _FromSize>;
508 if (!_FW::_S_is_partial && !(... && __xs._M_is_constprop()))
509 {
510 if constexpr ((sizeof...(_From) & (sizeof...(_From) - 1))
511 == 0) // power-of-two number of arguments
512 return __convert_x86<_To>(__as_vector(__xs)...);
513 else // append zeros and recurse until the above branch is taken
514 return __vector_convert<_To>(__xs..., _FW{});
515 }
516 else
517#endif
518 return __vector_convert<_To>(
519 __as_vector(__xs)...,
520 make_index_sequence<(sizeof...(__xs) == 1 ? std::min(
521 _VectorTraits<_To>::_S_full_size, int(_FromSize))
522 : _FromSize)>());
523 }
524
525// }}}
526// __convert function{{{
527template <typename _To, typename _From, typename... _More>
528 _GLIBCXX_SIMD_INTRINSIC constexpr auto
529 __convert(_From __v0, _More... __vs)
530 {
531 static_assert((true && ... && is_same_v<_From, _More>) );
532 if constexpr (__is_vectorizable_v<_From>)
533 {
534 using _V = typename _VectorTraits<_To>::type;
535 using _Tp = typename _VectorTraits<_To>::value_type;
536 return _V{static_cast<_Tp>(__v0), static_cast<_Tp>(__vs)...};
537 }
538 else if constexpr (__is_vector_type_v<_From>)
539 return __convert<_To>(__as_wrapper(__v0), __as_wrapper(__vs)...);
540 else // _SimdWrapper arguments
541 {
542 constexpr size_t __input_size = _From::_S_size * (1 + sizeof...(_More));
543 if constexpr (__is_vectorizable_v<_To>)
544 return __convert<__vector_type_t<_To, __input_size>>(__v0, __vs...);
545 else if constexpr (!__is_vector_type_v<_To>)
546 return _To(__convert<typename _To::_BuiltinType>(__v0, __vs...));
547 else
548 {
549 static_assert(
550 sizeof...(_More) == 0
551 || _VectorTraits<_To>::_S_full_size >= __input_size,
552 "__convert(...) requires the input to fit into the output");
553 return __vector_convert<_To>(__v0, __vs...);
554 }
555 }
556 }
557
558// }}}
559// __convert_all{{{
560// Converts __v into array<_To, N>, where N is _NParts if non-zero or
561// otherwise deduced from _To such that N * #elements(_To) <= #elements(__v).
562// Note: this function may return less than all converted elements
563template <typename _To,
564 size_t _NParts = 0, // allows to convert fewer or more (only last
565 // _To, to be partially filled) than all
566 size_t _Offset = 0, // where to start, # of elements (not Bytes or
567 // Parts)
568 typename _From, typename _FromVT = _VectorTraits<_From>>
569 _GLIBCXX_SIMD_INTRINSIC auto
570 __convert_all(_From __v)
571 {
572 if constexpr (is_arithmetic_v<_To> && _NParts != 1)
573 {
574 static_assert(_Offset < _FromVT::_S_full_size);
575 constexpr auto _Np
576 = _NParts == 0 ? _FromVT::_S_partial_width - _Offset : _NParts;
577 return __generate_from_n_evaluations<_Np, array<_To, _Np>>(
578 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
579 return static_cast<_To>(__v[__i + _Offset]);
580 });
581 }
582 else
583 {
584 static_assert(__is_vector_type_v<_To>);
585 using _ToVT = _VectorTraits<_To>;
586 if constexpr (__is_vector_type_v<_From>)
587 return __convert_all<_To, _NParts>(__as_wrapper(__v));
588 else if constexpr (_NParts == 1)
589 {
590 static_assert(_Offset % _ToVT::_S_full_size == 0);
591 return array<_To, 1>{__vector_convert<_To>(
592 __extract_part<_Offset / _ToVT::_S_full_size,
593 __div_roundup(_FromVT::_S_partial_width,
594 _ToVT::_S_full_size)>(__v))};
595 }
596#if _GLIBCXX_SIMD_X86INTRIN // {{{
597 else if constexpr (!__have_sse4_1 && _Offset == 0
598 && is_integral_v<typename _FromVT::value_type>
599 && sizeof(typename _FromVT::value_type)
600 < sizeof(typename _ToVT::value_type)
601 && !(sizeof(typename _FromVT::value_type) == 4
602 && is_same_v<typename _ToVT::value_type, double>))
603 {
604 using _ToT = typename _ToVT::value_type;
605 using _FromT = typename _FromVT::value_type;
606 constexpr size_t _Np
607 = _NParts != 0
608 ? _NParts
609 : (_FromVT::_S_partial_width / _ToVT::_S_full_size);
610 using _R = array<_To, _Np>;
611 // __adjust modifies its input to have _Np (use _SizeConstant)
612 // entries so that no unnecessary intermediate conversions are
613 // requested and, more importantly, no intermediate conversions are
614 // missing
615 [[maybe_unused]] auto __adjust
616 = [](auto __n,
617 auto __vv) -> _SimdWrapper<_FromT, decltype(__n)::value> {
618 return __vector_bitcast<_FromT, decltype(__n)::value>(__vv);
619 };
620 [[maybe_unused]] const auto __vi = __to_intrin(__v);
621 auto&& __make_array
622 = [](auto __x0, [[maybe_unused]] auto __x1) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
623 if constexpr (_Np == 1)
624 return _R{__intrin_bitcast<_To>(__x0)};
625 else
626 return _R{__intrin_bitcast<_To>(__x0),
627 __intrin_bitcast<_To>(__x1)};
628 };
629
630 if constexpr (_Np == 0)
631 return _R{};
632 else if constexpr (sizeof(_FromT) == 1 && sizeof(_ToT) == 2)
633 {
634 static_assert(is_integral_v<_FromT>);
635 static_assert(is_integral_v<_ToT>);
636 if constexpr (is_unsigned_v<_FromT>)
637 return __make_array(_mm_unpacklo_epi8(__vi, __m128i()),
638 _mm_unpackhi_epi8(__vi, __m128i()));
639 else
640 return __make_array(
641 _mm_srai_epi16(_mm_unpacklo_epi8(__vi, __vi), 8),
642 _mm_srai_epi16(_mm_unpackhi_epi8(__vi, __vi), 8));
643 }
644 else if constexpr (sizeof(_FromT) == 2 && sizeof(_ToT) == 4)
645 {
646 static_assert(is_integral_v<_FromT>);
647 if constexpr (is_floating_point_v<_ToT>)
648 {
649 const auto __ints
650 = __convert_all<__vector_type16_t<int>, _Np>(
651 __adjust(_SizeConstant<_Np * 4>(), __v));
652 return __generate_from_n_evaluations<_Np, _R>(
653 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
654 return __vector_convert<_To>(__as_wrapper(__ints[__i]));
655 });
656 }
657 else if constexpr (is_unsigned_v<_FromT>)
658 return __make_array(_mm_unpacklo_epi16(__vi, __m128i()),
659 _mm_unpackhi_epi16(__vi, __m128i()));
660 else
661 return __make_array(
662 _mm_srai_epi32(_mm_unpacklo_epi16(__vi, __vi), 16),
663 _mm_srai_epi32(_mm_unpackhi_epi16(__vi, __vi), 16));
664 }
665 else if constexpr (sizeof(_FromT) == 4 && sizeof(_ToT) == 8
666 && is_integral_v<_FromT> && is_integral_v<_ToT>)
667 {
668 if constexpr (is_unsigned_v<_FromT>)
669 return __make_array(_mm_unpacklo_epi32(__vi, __m128i()),
670 _mm_unpackhi_epi32(__vi, __m128i()));
671 else
672 return __make_array(
673 _mm_unpacklo_epi32(__vi, _mm_srai_epi32(__vi, 31)),
674 _mm_unpackhi_epi32(__vi, _mm_srai_epi32(__vi, 31)));
675 }
676 else if constexpr (sizeof(_FromT) == 4 && sizeof(_ToT) == 8
677 && is_integral_v<_FromT> && is_integral_v<_ToT>)
678 {
679 if constexpr (is_unsigned_v<_FromT>)
680 return __make_array(_mm_unpacklo_epi32(__vi, __m128i()),
681 _mm_unpackhi_epi32(__vi, __m128i()));
682 else
683 return __make_array(
684 _mm_unpacklo_epi32(__vi, _mm_srai_epi32(__vi, 31)),
685 _mm_unpackhi_epi32(__vi, _mm_srai_epi32(__vi, 31)));
686 }
687 else if constexpr (sizeof(_FromT) == 1 && sizeof(_ToT) >= 4
688 && is_signed_v<_FromT>)
689 {
690 const __m128i __vv[2] = {_mm_unpacklo_epi8(__vi, __vi),
691 _mm_unpackhi_epi8(__vi, __vi)};
692 const __vector_type_t<int, 4> __vvvv[4] = {
693 __vector_bitcast<int>(_mm_unpacklo_epi16(__vv[0], __vv[0])),
694 __vector_bitcast<int>(_mm_unpackhi_epi16(__vv[0], __vv[0])),
695 __vector_bitcast<int>(_mm_unpacklo_epi16(__vv[1], __vv[1])),
696 __vector_bitcast<int>(_mm_unpackhi_epi16(__vv[1], __vv[1]))};
697 if constexpr (sizeof(_ToT) == 4)
698 return __generate_from_n_evaluations<_Np, _R>(
699 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
700 return __vector_convert<_To>(
701 _SimdWrapper<int, 4>(__vvvv[__i] >> 24));
702 });
703 else if constexpr (is_integral_v<_ToT>)
704 return __generate_from_n_evaluations<_Np, _R>(
705 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
706 const auto __signbits = __to_intrin(__vvvv[__i / 2] >> 31);
707 const auto __sx32 = __to_intrin(__vvvv[__i / 2] >> 24);
708 return __vector_bitcast<_ToT>(
709 __i % 2 == 0 ? _mm_unpacklo_epi32(__sx32, __signbits)
710 : _mm_unpackhi_epi32(__sx32, __signbits));
711 });
712 else
713 return __generate_from_n_evaluations<_Np, _R>(
714 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
715 const _SimdWrapper<int, 4> __int4 = __vvvv[__i / 2] >> 24;
716 return __vector_convert<_To>(
717 __i % 2 == 0 ? __int4
718 : _SimdWrapper<int, 4>(
719 _mm_unpackhi_epi64(__to_intrin(__int4),
720 __to_intrin(__int4))));
721 });
722 }
723 else if constexpr (sizeof(_FromT) == 1 && sizeof(_ToT) == 4)
724 {
725 const auto __shorts = __convert_all<__vector_type16_t<
726 conditional_t<is_signed_v<_FromT>, short, unsigned short>>>(
727 __adjust(_SizeConstant<(_Np + 1) / 2 * 8>(), __v));
728 return __generate_from_n_evaluations<_Np, _R>(
729 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
730 return __convert_all<_To>(__shorts[__i / 2])[__i % 2];
731 });
732 }
733 else if constexpr (sizeof(_FromT) == 2 && sizeof(_ToT) == 8
734 && is_signed_v<_FromT> && is_integral_v<_ToT>)
735 {
736 const __m128i __vv[2] = {_mm_unpacklo_epi16(__vi, __vi),
737 _mm_unpackhi_epi16(__vi, __vi)};
738 const __vector_type16_t<int> __vvvv[4]
739 = {__vector_bitcast<int>(
740 _mm_unpacklo_epi32(_mm_srai_epi32(__vv[0], 16),
741 _mm_srai_epi32(__vv[0], 31))),
742 __vector_bitcast<int>(
743 _mm_unpackhi_epi32(_mm_srai_epi32(__vv[0], 16),
744 _mm_srai_epi32(__vv[0], 31))),
745 __vector_bitcast<int>(
746 _mm_unpacklo_epi32(_mm_srai_epi32(__vv[1], 16),
747 _mm_srai_epi32(__vv[1], 31))),
748 __vector_bitcast<int>(
749 _mm_unpackhi_epi32(_mm_srai_epi32(__vv[1], 16),
750 _mm_srai_epi32(__vv[1], 31)))};
751 return __generate_from_n_evaluations<_Np, _R>(
752 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
753 return __vector_bitcast<_ToT>(__vvvv[__i]);
754 });
755 }
756 else if constexpr (sizeof(_FromT) <= 2 && sizeof(_ToT) == 8)
757 {
758 const auto __ints
759 = __convert_all<__vector_type16_t<conditional_t<
760 is_signed_v<_FromT> || is_floating_point_v<_ToT>, int,
761 unsigned int>>>(
762 __adjust(_SizeConstant<(_Np + 1) / 2 * 4>(), __v));
763 return __generate_from_n_evaluations<_Np, _R>(
764 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
765 return __convert_all<_To>(__ints[__i / 2])[__i % 2];
766 });
767 }
768 else
769 __assert_unreachable<_To>();
770 }
771#endif // _GLIBCXX_SIMD_X86INTRIN }}}
772 else if constexpr ((_FromVT::_S_partial_width - _Offset)
773 > _ToVT::_S_full_size)
774 {
775 /*
776 static_assert(
777 (_FromVT::_S_partial_width & (_FromVT::_S_partial_width - 1)) ==
778 0,
779 "__convert_all only supports power-of-2 number of elements.
780 Otherwise " "the return type cannot be array<_To, N>.");
781 */
782 constexpr size_t _NTotal
783 = (_FromVT::_S_partial_width - _Offset) / _ToVT::_S_full_size;
784 constexpr size_t _Np = _NParts == 0 ? _NTotal : _NParts;
785 static_assert(
786 _Np <= _NTotal
787 || (_Np == _NTotal + 1
788 && (_FromVT::_S_partial_width - _Offset) % _ToVT::_S_full_size
789 > 0));
790 using _R = array<_To, _Np>;
791 if constexpr (_Np == 1)
792 return _R{__vector_convert<_To>(
793 __extract_part<_Offset, _FromVT::_S_partial_width,
794 _ToVT::_S_full_size>(__v))};
795 else
796 return __generate_from_n_evaluations<_Np, _R>(
797 [&](auto __i) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
798 auto __part
799 = __extract_part<__i * _ToVT::_S_full_size + _Offset,
800 _FromVT::_S_partial_width,
801 _ToVT::_S_full_size>(__v);
802 return __vector_convert<_To>(__part);
803 });
804 }
805 else if constexpr (_Offset == 0)
806 return array<_To, 1>{__vector_convert<_To>(__v)};
807 else
808 return array<_To, 1>{__vector_convert<_To>(
809 __extract_part<_Offset, _FromVT::_S_partial_width,
810 _FromVT::_S_partial_width - _Offset>(__v))};
811 }
812 }
813
814// }}}
815
816// _GnuTraits {{{
817template <typename _Tp, typename _Mp, typename _Abi, size_t _Np>
818 struct _GnuTraits
819 {
820 using _IsValid = true_type;
821 using _SimdImpl = typename _Abi::_SimdImpl;
822 using _MaskImpl = typename _Abi::_MaskImpl;
823
824 // simd and simd_mask member types {{{
825 using _SimdMember = _SimdWrapper<_Tp, _Np>;
826 using _MaskMember = _SimdWrapper<_Mp, _Np>;
827 static constexpr size_t _S_simd_align = alignof(_SimdMember);
828 static constexpr size_t _S_mask_align = alignof(_MaskMember);
829
830 // }}}
831 // size metadata {{{
832 static constexpr size_t _S_full_size = _SimdMember::_S_full_size;
833 static constexpr bool _S_is_partial = _SimdMember::_S_is_partial;
834
835 // }}}
836 // _SimdBase / base class for simd, providing extra conversions {{{
837 struct _SimdBase2
838 {
839 _GLIBCXX_SIMD_ALWAYS_INLINE
840 explicit operator __intrinsic_type_t<_Tp, _Np>() const
841 {
842 return __to_intrin(static_cast<const simd<_Tp, _Abi>*>(this)->_M_data);
843 }
844 _GLIBCXX_SIMD_ALWAYS_INLINE
845 explicit operator __vector_type_t<_Tp, _Np>() const
846 {
847 return static_cast<const simd<_Tp, _Abi>*>(this)->_M_data.__builtin();
848 }
849 };
850
851 struct _SimdBase1
852 {
853 _GLIBCXX_SIMD_ALWAYS_INLINE
854 explicit operator __intrinsic_type_t<_Tp, _Np>() const
855 { return __data(*static_cast<const simd<_Tp, _Abi>*>(this)); }
856 };
857
858 using _SimdBase = conditional_t<
859 is_same<__intrinsic_type_t<_Tp, _Np>, __vector_type_t<_Tp, _Np>>::value,
860 _SimdBase1, _SimdBase2>;
861
862 // }}}
863 // _MaskBase {{{
864 struct _MaskBase2
865 {
866 _GLIBCXX_SIMD_ALWAYS_INLINE
867 explicit operator __intrinsic_type_t<_Tp, _Np>() const
868 {
869 return static_cast<const simd_mask<_Tp, _Abi>*>(this)
870 ->_M_data.__intrin();
871 }
872 _GLIBCXX_SIMD_ALWAYS_INLINE
873 explicit operator __vector_type_t<_Tp, _Np>() const
874 {
875 return static_cast<const simd_mask<_Tp, _Abi>*>(this)->_M_data._M_data;
876 }
877 };
878
879 struct _MaskBase1
880 {
881 _GLIBCXX_SIMD_ALWAYS_INLINE
882 explicit operator __intrinsic_type_t<_Tp, _Np>() const
883 { return __data(*static_cast<const simd_mask<_Tp, _Abi>*>(this)); }
884 };
885
886 using _MaskBase = conditional_t<
887 is_same<__intrinsic_type_t<_Tp, _Np>, __vector_type_t<_Tp, _Np>>::value,
888 _MaskBase1, _MaskBase2>;
889
890 // }}}
891 // _MaskCastType {{{
892 // parameter type of one explicit simd_mask constructor
893 class _MaskCastType
894 {
895 using _Up = __intrinsic_type_t<_Tp, _Np>;
896 _Up _M_data;
897
898 public:
899 _GLIBCXX_SIMD_ALWAYS_INLINE
900 _MaskCastType(_Up __x) : _M_data(__x) {}
901 _GLIBCXX_SIMD_ALWAYS_INLINE
902 operator _MaskMember() const { return _M_data; }
903 };
904
905 // }}}
906 // _SimdCastType {{{
907 // parameter type of one explicit simd constructor
908 class _SimdCastType1
909 {
910 using _Ap = __intrinsic_type_t<_Tp, _Np>;
911 _SimdMember _M_data;
912
913 public:
914 _GLIBCXX_SIMD_ALWAYS_INLINE
915 _SimdCastType1(_Ap __a) : _M_data(__vector_bitcast<_Tp>(__a)) {}
916 _GLIBCXX_SIMD_ALWAYS_INLINE
917 operator _SimdMember() const { return _M_data; }
918 };
919
920 class _SimdCastType2
921 {
922 using _Ap = __intrinsic_type_t<_Tp, _Np>;
923 using _Bp = __vector_type_t<_Tp, _Np>;
924 _SimdMember _M_data;
925
926 public:
927 _GLIBCXX_SIMD_ALWAYS_INLINE
928 _SimdCastType2(_Ap __a) : _M_data(__vector_bitcast<_Tp>(__a)) {}
929 _GLIBCXX_SIMD_ALWAYS_INLINE
930 _SimdCastType2(_Bp __b) : _M_data(__b) {}
931 _GLIBCXX_SIMD_ALWAYS_INLINE
932 operator _SimdMember() const { return _M_data; }
933 };
934
935 using _SimdCastType = conditional_t<
936 is_same<__intrinsic_type_t<_Tp, _Np>, __vector_type_t<_Tp, _Np>>::value,
937 _SimdCastType1, _SimdCastType2>;
938 //}}}
939 };
940
941// }}}
942struct _CommonImplX86;
943struct _CommonImplNeon;
944struct _CommonImplBuiltin;
945template <typename _Abi, typename = __detail::__odr_helper> struct _SimdImplBuiltin;
946template <typename _Abi, typename = __detail::__odr_helper> struct _MaskImplBuiltin;
947template <typename _Abi, typename = __detail::__odr_helper> struct _SimdImplX86;
948template <typename _Abi, typename = __detail::__odr_helper> struct _MaskImplX86;
949template <typename _Abi, typename = __detail::__odr_helper> struct _SimdImplNeon;
950template <typename _Abi, typename = __detail::__odr_helper> struct _MaskImplNeon;
951template <typename _Abi, typename = __detail::__odr_helper> struct _SimdImplPpc;
952template <typename _Abi, typename = __detail::__odr_helper> struct _MaskImplPpc;
953
954// simd_abi::_VecBuiltin {{{
955template <int _UsedBytes>
956 struct simd_abi::_VecBuiltin
957 {
958 template <typename _Tp>
959 static constexpr size_t _S_size = _UsedBytes / sizeof(_Tp);
960
961 // validity traits {{{
962 struct _IsValidAbiTag : __bool_constant<(_UsedBytes > 1)> {};
963
964 template <typename _Tp>
965 struct _IsValidSizeFor
966 : __bool_constant<(_UsedBytes / sizeof(_Tp) > 1
967 && _UsedBytes % sizeof(_Tp) == 0
968 && _UsedBytes <= __vectorized_sizeof<_Tp>()
969 && (!__have_avx512f || _UsedBytes <= 32))> {};
970
971 template <typename _Tp>
972 struct _IsValid : conjunction<_IsValidAbiTag, __is_vectorizable<_Tp>,
973 _IsValidSizeFor<_Tp>> {};
974
975 template <typename _Tp>
976 static constexpr bool _S_is_valid_v = _IsValid<_Tp>::value;
977
978 // }}}
979 // _SimdImpl/_MaskImpl {{{
980#if _GLIBCXX_SIMD_X86INTRIN
981 using _CommonImpl = _CommonImplX86;
982 using _SimdImpl = _SimdImplX86<_VecBuiltin<_UsedBytes>>;
983 using _MaskImpl = _MaskImplX86<_VecBuiltin<_UsedBytes>>;
984#elif _GLIBCXX_SIMD_HAVE_NEON
985 using _CommonImpl = _CommonImplNeon;
986 using _SimdImpl = _SimdImplNeon<_VecBuiltin<_UsedBytes>>;
987 using _MaskImpl = _MaskImplNeon<_VecBuiltin<_UsedBytes>>;
988#else
989 using _CommonImpl = _CommonImplBuiltin;
990#ifdef __ALTIVEC__
991 using _SimdImpl = _SimdImplPpc<_VecBuiltin<_UsedBytes>>;
992 using _MaskImpl = _MaskImplPpc<_VecBuiltin<_UsedBytes>>;
993#else
994 using _SimdImpl = _SimdImplBuiltin<_VecBuiltin<_UsedBytes>>;
995 using _MaskImpl = _MaskImplBuiltin<_VecBuiltin<_UsedBytes>>;
996#endif
997#endif
998
999 // }}}
1000 // __traits {{{
1001 template <typename _Tp>
1002 using _MaskValueType = __int_for_sizeof_t<_Tp>;
1003
1004 template <typename _Tp>
1005 using __traits
1006 = conditional_t<_S_is_valid_v<_Tp>,
1007 _GnuTraits<_Tp, _MaskValueType<_Tp>,
1008 _VecBuiltin<_UsedBytes>, _S_size<_Tp>>,
1009 _InvalidTraits>;
1010
1011 //}}}
1012 // size metadata {{{
1013 template <typename _Tp>
1014 static constexpr size_t _S_full_size = __traits<_Tp>::_S_full_size;
1015
1016 template <typename _Tp>
1017 static constexpr bool _S_is_partial = __traits<_Tp>::_S_is_partial;
1018
1019 // }}}
1020 // implicit masks {{{
1021 template <typename _Tp>
1022 using _MaskMember = _SimdWrapper<_MaskValueType<_Tp>, _S_size<_Tp>>;
1023
1024 template <typename _Tp>
1025 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
1026 _S_implicit_mask()
1027 {
1028 using _UV = typename _MaskMember<_Tp>::_BuiltinType;
1029 if constexpr (!_MaskMember<_Tp>::_S_is_partial)
1030 return ~_UV();
1031 else
1032 {
1033 constexpr auto __size = _S_size<_Tp>;
1034 _GLIBCXX_SIMD_USE_CONSTEXPR auto __r
1035 = __generate_vector<_UV>([](auto __i) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA
1036 { return __i < __size ? -1 : 0; });
1037 return __r;
1038 }
1039 }
1040
1041 template <typename _Tp>
1042 _GLIBCXX_SIMD_INTRINSIC static constexpr __intrinsic_type_t<_Tp,
1043 _S_size<_Tp>>
1044 _S_implicit_mask_intrin()
1045 {
1046 return __to_intrin(
1047 __vector_bitcast<_Tp>(_S_implicit_mask<_Tp>()._M_data));
1048 }
1049
1050 template <typename _TW, typename _TVT = _VectorTraits<_TW>>
1051 _GLIBCXX_SIMD_INTRINSIC static constexpr _TW _S_masked(_TW __x)
1052 {
1053 using _Tp = typename _TVT::value_type;
1054 if constexpr (!_MaskMember<_Tp>::_S_is_partial)
1055 return __x;
1056 else
1057 return __and(__as_vector(__x),
1058 __vector_bitcast<_Tp>(_S_implicit_mask<_Tp>()));
1059 }
1060
1061 template <typename _TW, typename _TVT = _VectorTraits<_TW>>
1062 _GLIBCXX_SIMD_INTRINSIC static constexpr auto
1063 __make_padding_nonzero(_TW __x)
1064 {
1065 using _Tp = typename _TVT::value_type;
1066 if constexpr (!_S_is_partial<_Tp>)
1067 return __x;
1068 else
1069 {
1070 _GLIBCXX_SIMD_USE_CONSTEXPR auto __implicit_mask
1071 = __vector_bitcast<_Tp>(_S_implicit_mask<_Tp>());
1072 if constexpr (is_integral_v<_Tp>)
1073 return __or(__x, ~__implicit_mask);
1074 else
1075 {
1076 _GLIBCXX_SIMD_USE_CONSTEXPR auto __one
1077 = __andnot(__implicit_mask,
1078 __vector_broadcast<_S_full_size<_Tp>>(_Tp(1)));
1079 // it's not enough to return `x | 1_in_padding` because the
1080 // padding in x might be inf or nan (independent of
1081 // __FINITE_MATH_ONLY__, because it's about padding bits)
1082 return __or(__and(__x, __implicit_mask), __one);
1083 }
1084 }
1085 }
1086 // }}}
1087 };
1088
1089// }}}
1090// simd_abi::_VecBltnBtmsk {{{
1091template <int _UsedBytes>
1092 struct simd_abi::_VecBltnBtmsk
1093 {
1094 template <typename _Tp>
1095 static constexpr size_t _S_size = _UsedBytes / sizeof(_Tp);
1096
1097 // validity traits {{{
1098 struct _IsValidAbiTag : __bool_constant<(_UsedBytes > 1)> {};
1099
1100 template <typename _Tp>
1101 struct _IsValidSizeFor
1102 : __bool_constant<(_UsedBytes / sizeof(_Tp) > 1
1103 && _UsedBytes % sizeof(_Tp) == 0 && _UsedBytes <= 64
1104 && (_UsedBytes > 32 || __have_avx512vl))> {};
1105
1106 // Bitmasks require at least AVX512F. If sizeof(_Tp) < 4 the AVX512BW is also
1107 // required.
1108 template <typename _Tp>
1109 struct _IsValid
1110 : conjunction<
1111 _IsValidAbiTag, __bool_constant<__have_avx512f>,
1112 __bool_constant<__have_avx512bw || (sizeof(_Tp) >= 4)>,
1113 __bool_constant<(__vectorized_sizeof<_Tp>() > sizeof(_Tp))>,
1114 _IsValidSizeFor<_Tp>> {};
1115
1116 template <typename _Tp>
1117 static constexpr bool _S_is_valid_v = _IsValid<_Tp>::value;
1118
1119 // }}}
1120 // simd/_MaskImpl {{{
1121 #if _GLIBCXX_SIMD_X86INTRIN
1122 using _CommonImpl = _CommonImplX86;
1123 using _SimdImpl = _SimdImplX86<_VecBltnBtmsk<_UsedBytes>>;
1124 using _MaskImpl = _MaskImplX86<_VecBltnBtmsk<_UsedBytes>>;
1125 #else
1126 template <int>
1127 struct _MissingImpl;
1128
1129 using _CommonImpl = _MissingImpl<_UsedBytes>;
1130 using _SimdImpl = _MissingImpl<_UsedBytes>;
1131 using _MaskImpl = _MissingImpl<_UsedBytes>;
1132 #endif
1133
1134 // }}}
1135 // __traits {{{
1136 template <typename _Tp>
1137 using _MaskMember = _SimdWrapper<bool, _S_size<_Tp>>;
1138
1139 template <typename _Tp>
1140 using __traits = conditional_t<
1141 _S_is_valid_v<_Tp>,
1142 _GnuTraits<_Tp, bool, _VecBltnBtmsk<_UsedBytes>, _S_size<_Tp>>,
1143 _InvalidTraits>;
1144
1145 //}}}
1146 // size metadata {{{
1147 template <typename _Tp>
1148 static constexpr size_t _S_full_size = __traits<_Tp>::_S_full_size;
1149 template <typename _Tp>
1150 static constexpr bool _S_is_partial = __traits<_Tp>::_S_is_partial;
1151
1152 // }}}
1153 // implicit mask {{{
1154 private:
1155 template <typename _Tp>
1156 using _ImplicitMask = _SimdWrapper<bool, _S_size<_Tp>>;
1157
1158 public:
1159 template <size_t _Np>
1160 _GLIBCXX_SIMD_INTRINSIC static constexpr __bool_storage_member_type_t<_Np>
1161 __implicit_mask_n()
1162 {
1163 using _Tp = __bool_storage_member_type_t<_Np>;
1164 return _Np < sizeof(_Tp) * __CHAR_BIT__ ? _Tp((1ULL << _Np) - 1) : ~_Tp();
1165 }
1166
1167 template <typename _Tp>
1168 _GLIBCXX_SIMD_INTRINSIC static constexpr _ImplicitMask<_Tp>
1169 _S_implicit_mask()
1170 { return __implicit_mask_n<_S_size<_Tp>>(); }
1171
1172 template <typename _Tp>
1173 _GLIBCXX_SIMD_INTRINSIC static constexpr __bool_storage_member_type_t<
1174 _S_size<_Tp>>
1175 _S_implicit_mask_intrin()
1176 { return __implicit_mask_n<_S_size<_Tp>>(); }
1177
1178 template <typename _Tp, size_t _Np>
1179 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
1180 _S_masked(_SimdWrapper<_Tp, _Np> __x)
1181 {
1182 if constexpr (is_same_v<_Tp, bool>)
1183 if constexpr (_Np < 8 || (_Np & (_Np - 1)) != 0)
1184 return _MaskImpl::_S_bit_and(
1185 __x, _SimdWrapper<_Tp, _Np>(
1186 __bool_storage_member_type_t<_Np>((1ULL << _Np) - 1)));
1187 else
1188 return __x;
1189 else
1190 return _S_masked(__x._M_data);
1191 }
1192
1193 template <typename _TV>
1194 _GLIBCXX_SIMD_INTRINSIC static constexpr _TV
1195 _S_masked(_TV __x)
1196 {
1197 using _Tp = typename _VectorTraits<_TV>::value_type;
1198 static_assert(
1199 !__is_bitmask_v<_TV>,
1200 "_VecBltnBtmsk::_S_masked cannot work on bitmasks, since it doesn't "
1201 "know the number of elements. Use _SimdWrapper<bool, N> instead.");
1202 if constexpr (_S_is_partial<_Tp>)
1203 {
1204 constexpr size_t _Np = _S_size<_Tp>;
1205 return __make_dependent_t<_TV, _CommonImpl>::_S_blend(
1206 _S_implicit_mask<_Tp>(), _SimdWrapper<_Tp, _Np>(),
1207 _SimdWrapper<_Tp, _Np>(__x));
1208 }
1209 else
1210 return __x;
1211 }
1212
1213 template <typename _TV, typename _TVT = _VectorTraits<_TV>>
1214 _GLIBCXX_SIMD_INTRINSIC static constexpr auto
1215 __make_padding_nonzero(_TV __x)
1216 {
1217 using _Tp = typename _TVT::value_type;
1218 if constexpr (!_S_is_partial<_Tp>)
1219 return __x;
1220 else
1221 {
1222 constexpr size_t _Np = _S_size<_Tp>;
1223 if constexpr (is_integral_v<typename _TVT::value_type>)
1224 return __x
1225 | __generate_vector<_Tp, _S_full_size<_Tp>>(
1226 [](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA -> _Tp {
1227 if (__i < _Np)
1228 return 0;
1229 else
1230 return 1;
1231 });
1232 else
1233 return __make_dependent_t<_TV, _CommonImpl>::_S_blend(
1234 _S_implicit_mask<_Tp>(),
1235 _SimdWrapper<_Tp, _Np>(
1236 __vector_broadcast<_S_full_size<_Tp>>(_Tp(1))),
1237 _SimdWrapper<_Tp, _Np>(__x))
1238 ._M_data;
1239 }
1240 }
1241
1242 // }}}
1243 };
1244
1245//}}}
1246// _CommonImplBuiltin {{{
1247struct _CommonImplBuiltin
1248{
1249 // _S_converts_via_decomposition{{{
1250 // This lists all cases where a __vector_convert needs to fall back to
1251 // conversion of individual scalars (i.e. decompose the input vector into
1252 // scalars, convert, compose output vector). In those cases, _S_masked_load &
1253 // _S_masked_store prefer to use the _S_bit_iteration implementation.
1254 template <typename _From, typename _To, size_t _ToSize>
1255 static inline constexpr bool __converts_via_decomposition_v
1256 = sizeof(_From) != sizeof(_To);
1257
1258 // }}}
1259 // _S_load{{{
1260 template <typename _Tp, size_t _Np, size_t _Bytes = _Np * sizeof(_Tp)>
1261 _GLIBCXX_SIMD_INTRINSIC static __vector_type_t<_Tp, _Np>
1262 _S_load(const void* __p)
1263 {
1264 static_assert(_Np > 1);
1265 static_assert(_Bytes % sizeof(_Tp) == 0);
1266 using _Rp = __vector_type_t<_Tp, _Np>;
1267 if constexpr (sizeof(_Rp) == _Bytes)
1268 {
1269 _Rp __r;
1270 __builtin_memcpy(&__r, __p, _Bytes);
1271 return __r;
1272 }
1273 else
1274 {
1275#ifdef _GLIBCXX_SIMD_WORKAROUND_PR90424
1276 using _Up = conditional_t<
1277 is_integral_v<_Tp>,
1278 conditional_t<_Bytes % 4 == 0,
1279 conditional_t<_Bytes % 8 == 0, long long, int>,
1280 conditional_t<_Bytes % 2 == 0, short, signed char>>,
1281 conditional_t<(_Bytes < 8 || _Np % 2 == 1 || _Np == 2), _Tp,
1282 double>>;
1283 using _V = __vector_type_t<_Up, _Np * sizeof(_Tp) / sizeof(_Up)>;
1284 if constexpr (sizeof(_V) != sizeof(_Rp))
1285 { // on i386 with 4 < _Bytes <= 8
1286 _Rp __r{};
1287 __builtin_memcpy(&__r, __p, _Bytes);
1288 return __r;
1289 }
1290 else
1291#else // _GLIBCXX_SIMD_WORKAROUND_PR90424
1292 using _V = _Rp;
1293#endif // _GLIBCXX_SIMD_WORKAROUND_PR90424
1294 {
1295 _V __r{};
1296 static_assert(_Bytes <= sizeof(_V));
1297 __builtin_memcpy(&__r, __p, _Bytes);
1298 return reinterpret_cast<_Rp>(__r);
1299 }
1300 }
1301 }
1302
1303 // }}}
1304 // _S_store {{{
1305 template <size_t _ReqBytes = 0, typename _TV>
1306 _GLIBCXX_SIMD_INTRINSIC static void _S_store(_TV __x, void* __addr)
1307 {
1308 constexpr size_t _Bytes = _ReqBytes == 0 ? sizeof(__x) : _ReqBytes;
1309 static_assert(sizeof(__x) >= _Bytes);
1310
1311 if constexpr (__is_vector_type_v<_TV>)
1312 {
1313 using _Tp = typename _VectorTraits<_TV>::value_type;
1314 constexpr size_t _Np = _Bytes / sizeof(_Tp);
1315 static_assert(_Np * sizeof(_Tp) == _Bytes);
1316
1317#ifdef _GLIBCXX_SIMD_WORKAROUND_PR90424
1318 using _Up = conditional_t<
1319 (is_integral_v<_Tp> || _Bytes < 4),
1320 conditional_t<(sizeof(__x) > sizeof(long long)), long long, _Tp>,
1321 float>;
1322 const auto __v = __vector_bitcast<_Up>(__x);
1323#else // _GLIBCXX_SIMD_WORKAROUND_PR90424
1324 const __vector_type_t<_Tp, _Np> __v = __x;
1325#endif // _GLIBCXX_SIMD_WORKAROUND_PR90424
1326
1327 if constexpr ((_Bytes & (_Bytes - 1)) != 0)
1328 {
1329 constexpr size_t _MoreBytes = std::__bit_ceil(_Bytes);
1330 alignas(decltype(__v)) char __tmp[_MoreBytes];
1331 __builtin_memcpy(__tmp, &__v, _MoreBytes);
1332 __builtin_memcpy(__addr, __tmp, _Bytes);
1333 }
1334 else
1335 __builtin_memcpy(__addr, &__v, _Bytes);
1336 }
1337 else
1338 __builtin_memcpy(__addr, &__x, _Bytes);
1339 }
1340
1341 template <typename _Tp, size_t _Np>
1342 _GLIBCXX_SIMD_INTRINSIC static void _S_store(_SimdWrapper<_Tp, _Np> __x,
1343 void* __addr)
1344 { _S_store<_Np * sizeof(_Tp)>(__x._M_data, __addr); }
1345
1346 // }}}
1347 // _S_store_bool_array(_BitMask) {{{
1348 template <size_t _Np, bool _Sanitized>
1349 _GLIBCXX_SIMD_INTRINSIC static constexpr void
1350 _S_store_bool_array(_BitMask<_Np, _Sanitized> __x, bool* __mem)
1351 {
1352 if constexpr (_Np == 1)
1353 __mem[0] = __x[0];
1354 else if constexpr (_Np == 2)
1355 {
1356 short __bool2 = (__x._M_to_bits() * 0x81) & 0x0101;
1357 _S_store<_Np>(__bool2, __mem);
1358 }
1359 else if constexpr (_Np == 3)
1360 {
1361 int __bool3 = (__x._M_to_bits() * 0x4081) & 0x010101;
1362 _S_store<_Np>(__bool3, __mem);
1363 }
1364 else
1365 {
1366 __execute_n_times<__div_roundup(_Np, 4)>(
1367 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1368 constexpr int __offset = __i * 4;
1369 constexpr int __remaining = _Np - __offset;
1370 if constexpr (__remaining > 4 && __remaining <= 7)
1371 {
1372 const _ULLong __bool7
1373 = (__x.template _M_extract<__offset>()._M_to_bits()
1374 * 0x40810204081ULL)
1375 & 0x0101010101010101ULL;
1376 _S_store<__remaining>(__bool7, __mem + __offset);
1377 }
1378 else if constexpr (__remaining >= 4)
1379 {
1380 int __bits = __x.template _M_extract<__offset>()._M_to_bits();
1381 if constexpr (__remaining > 7)
1382 __bits &= 0xf;
1383 const int __bool4 = (__bits * 0x204081) & 0x01010101;
1384 _S_store<4>(__bool4, __mem + __offset);
1385 }
1386 });
1387 }
1388 }
1389
1390 // }}}
1391 // _S_blend{{{
1392 template <typename _Tp, size_t _Np>
1393 _GLIBCXX_SIMD_INTRINSIC static constexpr auto
1394 _S_blend(_SimdWrapper<__int_for_sizeof_t<_Tp>, _Np> __k,
1395 _SimdWrapper<_Tp, _Np> __at0, _SimdWrapper<_Tp, _Np> __at1)
1396 { return __k._M_data ? __at1._M_data : __at0._M_data; }
1397
1398 // }}}
1399};
1400
1401// }}}
1402// _SimdImplBuiltin {{{1
1403template <typename _Abi, typename>
1404 struct _SimdImplBuiltin
1405 {
1406 // member types {{{2
1407 template <typename _Tp>
1408 static constexpr size_t _S_max_store_size = 16;
1409
1410 using abi_type = _Abi;
1411
1412 template <typename _Tp>
1413 using _TypeTag = _Tp*;
1414
1415 template <typename _Tp>
1416 using _SimdMember = typename _Abi::template __traits<_Tp>::_SimdMember;
1417
1418 template <typename _Tp>
1419 using _MaskMember = typename _Abi::template _MaskMember<_Tp>;
1420
1421 template <typename _Tp>
1422 static constexpr size_t _S_size = _Abi::template _S_size<_Tp>;
1423
1424 template <typename _Tp>
1425 static constexpr size_t _S_full_size = _Abi::template _S_full_size<_Tp>;
1426
1427 using _CommonImpl = typename _Abi::_CommonImpl;
1428 using _SuperImpl = typename _Abi::_SimdImpl;
1429 using _MaskImpl = typename _Abi::_MaskImpl;
1430
1431 // _M_make_simd(_SimdWrapper/__intrinsic_type_t) {{{2
1432 template <typename _Tp, size_t _Np>
1433 _GLIBCXX_SIMD_INTRINSIC static simd<_Tp, _Abi>
1434 _M_make_simd(_SimdWrapper<_Tp, _Np> __x)
1435 { return {__private_init, __x}; }
1436
1437 template <typename _Tp, size_t _Np>
1438 _GLIBCXX_SIMD_INTRINSIC static simd<_Tp, _Abi>
1439 _M_make_simd(__intrinsic_type_t<_Tp, _Np> __x)
1440 { return {__private_init, __vector_bitcast<_Tp>(__x)}; }
1441
1442 // _S_broadcast {{{2
1443 template <typename _Tp>
1444 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdMember<_Tp>
1445 _S_broadcast(_Tp __x) noexcept
1446 { return __vector_broadcast<_S_full_size<_Tp>>(__x); }
1447
1448 // _S_generator {{{2
1449 template <typename _Fp, typename _Tp>
1450 inline static constexpr _SimdMember<_Tp> _S_generator(_Fp&& __gen,
1451 _TypeTag<_Tp>)
1452 {
1453 return __generate_vector<_Tp, _S_full_size<_Tp>>(
1454 [&](auto __i) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1455 if constexpr (__i < _S_size<_Tp>)
1456 return __gen(__i);
1457 else
1458 return 0;
1459 });
1460 }
1461
1462 // _S_load {{{2
1463 template <typename _Tp, typename _Up>
1464 _GLIBCXX_SIMD_INTRINSIC static _SimdMember<_Tp>
1465 _S_load(const _Up* __mem, _TypeTag<_Tp>) noexcept
1466 {
1467 constexpr size_t _Np = _S_size<_Tp>;
1468 constexpr size_t __max_load_size
1469 = (sizeof(_Up) >= 4 && __have_avx512f) || __have_avx512bw ? 64
1470 : (is_floating_point_v<_Up> && __have_avx) || __have_avx2 ? 32
1471 : 16;
1472 constexpr size_t __bytes_to_load = sizeof(_Up) * _Np;
1473 if constexpr (sizeof(_Up) > 8)
1474 return __generate_vector<_Tp, _SimdMember<_Tp>::_S_full_size>(
1475 [&](auto __i) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1476 return static_cast<_Tp>(__i < _Np ? __mem[__i] : 0);
1477 });
1478 else if constexpr (is_same_v<_Up, _Tp>)
1479 return _CommonImpl::template _S_load<_Tp, _S_full_size<_Tp>,
1480 _Np * sizeof(_Tp)>(__mem);
1481 else if constexpr (__bytes_to_load <= __max_load_size)
1482 return __convert<_SimdMember<_Tp>>(
1483 _CommonImpl::template _S_load<_Up, _Np>(__mem));
1484 else if constexpr (__bytes_to_load % __max_load_size == 0)
1485 {
1486 constexpr size_t __n_loads = __bytes_to_load / __max_load_size;
1487 constexpr size_t __elements_per_load = _Np / __n_loads;
1488 return __call_with_n_evaluations<__n_loads>(
1489 [](auto... __uncvted) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1490 return __convert<_SimdMember<_Tp>>(__uncvted...);
1491 }, [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1492 return _CommonImpl::template _S_load<_Up, __elements_per_load>(
1493 __mem + __i * __elements_per_load);
1494 });
1495 }
1496 else if constexpr (__bytes_to_load % (__max_load_size / 2) == 0
1497 && __max_load_size > 16)
1498 { // e.g. int[] -> <char, 12> with AVX2
1499 constexpr size_t __n_loads
1500 = __bytes_to_load / (__max_load_size / 2);
1501 constexpr size_t __elements_per_load = _Np / __n_loads;
1502 return __call_with_n_evaluations<__n_loads>(
1503 [](auto... __uncvted) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1504 return __convert<_SimdMember<_Tp>>(__uncvted...);
1505 }, [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1506 return _CommonImpl::template _S_load<_Up, __elements_per_load>(
1507 __mem + __i * __elements_per_load);
1508 });
1509 }
1510 else // e.g. int[] -> <char, 9>
1511 return __call_with_subscripts(
1512 __mem, make_index_sequence<_Np>(),
1513 [](auto... __args) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1514 return __vector_type_t<_Tp, _S_full_size<_Tp>>{static_cast<_Tp>(__args)...};
1515 });
1516 }
1517
1518 // _S_masked_load {{{2
1519 template <typename _Tp, size_t _Np, typename _Up>
1520 static inline _SimdWrapper<_Tp, _Np>
1521 _S_masked_load(_SimdWrapper<_Tp, _Np> __merge, _MaskMember<_Tp> __k,
1522 const _Up* __mem) noexcept
1523 {
1524 _BitOps::_S_bit_iteration(_MaskImpl::_S_to_bits(__k),
1525 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1526 __merge._M_set(__i, static_cast<_Tp>(__mem[__i]));
1527 });
1528 return __merge;
1529 }
1530
1531 // _S_store {{{2
1532 template <typename _Tp, typename _Up>
1533 _GLIBCXX_SIMD_INTRINSIC static void
1534 _S_store(_SimdMember<_Tp> __v, _Up* __mem, _TypeTag<_Tp>) noexcept
1535 {
1536 // TODO: converting int -> "smaller int" can be optimized with AVX512
1537 constexpr size_t _Np = _S_size<_Tp>;
1538 constexpr size_t __max_store_size
1539 = _SuperImpl::template _S_max_store_size<_Up>;
1540 if constexpr (sizeof(_Up) > 8)
1541 __execute_n_times<_Np>([&](auto __i) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1542 __mem[__i] = __v[__i];
1543 });
1544 else if constexpr (is_same_v<_Up, _Tp>)
1545 _CommonImpl::_S_store(__v, __mem);
1546 else if constexpr (sizeof(_Up) * _Np <= __max_store_size)
1547 _CommonImpl::_S_store(_SimdWrapper<_Up, _Np>(__convert<_Up>(__v)),
1548 __mem);
1549 else
1550 {
1551 constexpr size_t __vsize = __max_store_size / sizeof(_Up);
1552 // round up to convert the last partial vector as well:
1553 constexpr size_t __stores = __div_roundup(_Np, __vsize);
1554 constexpr size_t __full_stores = _Np / __vsize;
1555 using _V = __vector_type_t<_Up, __vsize>;
1556 const array<_V, __stores> __converted
1557 = __convert_all<_V, __stores>(__v);
1558 __execute_n_times<__full_stores>(
1559 [&](auto __i) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1560 _CommonImpl::_S_store(__converted[__i], __mem + __i * __vsize);
1561 });
1562 if constexpr (__full_stores < __stores)
1563 _CommonImpl::template _S_store<(_Np - __full_stores * __vsize)
1564 * sizeof(_Up)>(
1565 __converted[__full_stores], __mem + __full_stores * __vsize);
1566 }
1567 }
1568
1569 // _S_masked_store_nocvt {{{2
1570 template <typename _Tp, size_t _Np>
1571 _GLIBCXX_SIMD_INTRINSIC static void
1572 _S_masked_store_nocvt(_SimdWrapper<_Tp, _Np> __v, _Tp* __mem,
1573 _MaskMember<_Tp> __k)
1574 {
1575 _BitOps::_S_bit_iteration(
1576 _MaskImpl::_S_to_bits(__k),
1577 [&](auto __i) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1578 __mem[__i] = __v[__i];
1579 });
1580 }
1581
1582 // _S_masked_store {{{2
1583 template <typename _TW, typename _TVT = _VectorTraits<_TW>,
1584 typename _Tp = typename _TVT::value_type, typename _Up>
1585 static inline void
1586 _S_masked_store(const _TW __v, _Up* __mem, const _MaskMember<_Tp> __k)
1587 noexcept
1588 {
1589 constexpr size_t _TV_size = _S_size<_Tp>;
1590 [[maybe_unused]] const auto __vi = __to_intrin(__v);
1591 constexpr size_t __max_store_size
1592 = _SuperImpl::template _S_max_store_size<_Up>;
1593 if constexpr (
1594 is_same_v<
1595 _Tp,
1596 _Up> || (is_integral_v<_Tp> && is_integral_v<_Up> && sizeof(_Tp) == sizeof(_Up)))
1597 {
1598 // bitwise or no conversion, reinterpret:
1599 const _MaskMember<_Up> __kk = [&]() _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1600 if constexpr (__is_bitmask_v<decltype(__k)>)
1601 return _MaskMember<_Up>(__k._M_data);
1602 else
1603 return __wrapper_bitcast<__int_for_sizeof_t<_Up>>(__k);
1604 }();
1605 _SuperImpl::_S_masked_store_nocvt(__wrapper_bitcast<_Up>(__v),
1606 __mem, __kk);
1607 }
1608 else if constexpr (__vectorized_sizeof<_Up>() > sizeof(_Up)
1609 && !_CommonImpl::
1610 template __converts_via_decomposition_v<
1611 _Tp, _Up, __max_store_size>)
1612 { // conversion via decomposition is better handled via the
1613 // bit_iteration
1614 // fallback below
1615 constexpr size_t _UW_size
1616 = std::min(_TV_size, __max_store_size / sizeof(_Up));
1617 static_assert(_UW_size <= _TV_size);
1618 using _UW = _SimdWrapper<_Up, _UW_size>;
1619 using _UV = __vector_type_t<_Up, _UW_size>;
1620 using _UAbi = simd_abi::deduce_t<_Up, _UW_size>;
1621 if constexpr (_UW_size == _TV_size) // one convert+store
1622 {
1623 const _UW __converted = __convert<_UW>(__v);
1624 _SuperImpl::_S_masked_store_nocvt(
1625 __converted, __mem,
1626 _UAbi::_MaskImpl::template _S_convert<
1627 __int_for_sizeof_t<_Up>>(__k));
1628 }
1629 else
1630 {
1631 static_assert(_UW_size * sizeof(_Up) == __max_store_size);
1632 constexpr size_t _NFullStores = _TV_size / _UW_size;
1633 constexpr size_t _NAllStores
1634 = __div_roundup(_TV_size, _UW_size);
1635 constexpr size_t _NParts = _S_full_size<_Tp> / _UW_size;
1636 const array<_UV, _NAllStores> __converted
1637 = __convert_all<_UV, _NAllStores>(__v);
1638 __execute_n_times<_NFullStores>([&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1639 _SuperImpl::_S_masked_store_nocvt(
1640 _UW(__converted[__i]), __mem + __i * _UW_size,
1641 _UAbi::_MaskImpl::template _S_convert<
1642 __int_for_sizeof_t<_Up>>(
1643 __extract_part<__i, _NParts>(__k.__as_full_vector())));
1644 });
1645 if constexpr (_NAllStores
1646 > _NFullStores) // one partial at the end
1647 _SuperImpl::_S_masked_store_nocvt(
1648 _UW(__converted[_NFullStores]),
1649 __mem + _NFullStores * _UW_size,
1650 _UAbi::_MaskImpl::template _S_convert<
1651 __int_for_sizeof_t<_Up>>(
1652 __extract_part<_NFullStores, _NParts>(
1653 __k.__as_full_vector())));
1654 }
1655 }
1656 else
1657 _BitOps::_S_bit_iteration(_MaskImpl::_S_to_bits(__k),
1658 [&](auto __i) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1659 __mem[__i] = static_cast<_Up>(__v[__i]);
1660 });
1661 }
1662
1663 // _S_complement {{{2
1664 template <typename _Tp, size_t _Np>
1665 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
1666 _S_complement(_SimdWrapper<_Tp, _Np> __x) noexcept
1667 {
1668 if constexpr (is_floating_point_v<_Tp>)
1669 return __vector_bitcast<_Tp>(~__vector_bitcast<__int_for_sizeof_t<_Tp>>(__x));
1670 else
1671 return ~__x._M_data;
1672 }
1673
1674 // _S_unary_minus {{{2
1675 template <typename _Tp, size_t _Np>
1676 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
1677 _S_unary_minus(_SimdWrapper<_Tp, _Np> __x) noexcept
1678 {
1679 // GCC doesn't use the psign instructions, but pxor & psub seem to be
1680 // just as good a choice as pcmpeqd & psign. So meh.
1681 return -__x._M_data;
1682 }
1683
1684 // arithmetic operators {{{2
1685 template <typename _Tp, size_t _Np>
1686 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
1687 _S_plus(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1688 { return __x._M_data + __y._M_data; }
1689
1690 template <typename _Tp, size_t _Np>
1691 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
1692 _S_minus(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1693 { return __x._M_data - __y._M_data; }
1694
1695 template <typename _Tp, size_t _Np>
1696 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
1697 _S_multiplies(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1698 { return __x._M_data * __y._M_data; }
1699
1700 template <typename _Tp, size_t _Np>
1701 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
1702 _S_divides(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1703 {
1704 // Note that division by 0 is always UB, so we must ensure we avoid the
1705 // case for partial registers
1706 if constexpr (!_Abi::template _S_is_partial<_Tp>)
1707 return __x._M_data / __y._M_data;
1708 else
1709 return __x._M_data / _Abi::__make_padding_nonzero(__y._M_data);
1710 }
1711
1712 template <typename _Tp, size_t _Np>
1713 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
1714 _S_modulus(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1715 {
1716 if constexpr (!_Abi::template _S_is_partial<_Tp>)
1717 return __x._M_data % __y._M_data;
1718 else
1719 return __as_vector(__x)
1720 % _Abi::__make_padding_nonzero(__as_vector(__y));
1721 }
1722
1723 template <typename _Tp, size_t _Np>
1724 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
1725 _S_bit_and(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1726 { return __and(__x, __y); }
1727
1728 template <typename _Tp, size_t _Np>
1729 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
1730 _S_bit_or(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1731 { return __or(__x, __y); }
1732
1733 template <typename _Tp, size_t _Np>
1734 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
1735 _S_bit_xor(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1736 { return __xor(__x, __y); }
1737
1738 template <typename _Tp, size_t _Np>
1739 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
1740 _S_bit_shift_left(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1741 { return __x._M_data << __y._M_data; }
1742
1743 template <typename _Tp, size_t _Np>
1744 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
1745 _S_bit_shift_right(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1746 { return __x._M_data >> __y._M_data; }
1747
1748 template <typename _Tp, size_t _Np>
1749 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
1750 _S_bit_shift_left(_SimdWrapper<_Tp, _Np> __x, int __y)
1751 { return __x._M_data << __y; }
1752
1753 template <typename _Tp, size_t _Np>
1754 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
1755 _S_bit_shift_right(_SimdWrapper<_Tp, _Np> __x, int __y)
1756 { return __x._M_data >> __y; }
1757
1758 // compares {{{2
1759 // _S_equal_to {{{3
1760 template <typename _Tp, size_t _Np>
1761 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
1762 _S_equal_to(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1763 { return __x._M_data == __y._M_data; }
1764
1765 // _S_not_equal_to {{{3
1766 template <typename _Tp, size_t _Np>
1767 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
1768 _S_not_equal_to(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1769 { return __x._M_data != __y._M_data; }
1770
1771 // _S_less {{{3
1772 template <typename _Tp, size_t _Np>
1773 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
1774 _S_less(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1775 { return __x._M_data < __y._M_data; }
1776
1777 // _S_less_equal {{{3
1778 template <typename _Tp, size_t _Np>
1779 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
1780 _S_less_equal(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1781 { return __x._M_data <= __y._M_data; }
1782
1783 // _S_negate {{{2
1784 template <typename _Tp, size_t _Np>
1785 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
1786 _S_negate(_SimdWrapper<_Tp, _Np> __x) noexcept
1787 { return !__x._M_data; }
1788
1789 // _S_min, _S_max, _S_minmax {{{2
1790 template <typename _Tp, size_t _Np>
1791 _GLIBCXX_SIMD_NORMAL_MATH _GLIBCXX_SIMD_INTRINSIC static constexpr
1792 _SimdWrapper<_Tp, _Np>
1793 _S_min(_SimdWrapper<_Tp, _Np> __a, _SimdWrapper<_Tp, _Np> __b)
1794 { return __a._M_data < __b._M_data ? __a._M_data : __b._M_data; }
1795
1796 template <typename _Tp, size_t _Np>
1797 _GLIBCXX_SIMD_NORMAL_MATH _GLIBCXX_SIMD_INTRINSIC static constexpr
1798 _SimdWrapper<_Tp, _Np>
1799 _S_max(_SimdWrapper<_Tp, _Np> __a, _SimdWrapper<_Tp, _Np> __b)
1800 { return __a._M_data > __b._M_data ? __a._M_data : __b._M_data; }
1801
1802 template <typename _Tp, size_t _Np>
1803 _GLIBCXX_SIMD_NORMAL_MATH _GLIBCXX_SIMD_INTRINSIC static constexpr
1804 pair<_SimdWrapper<_Tp, _Np>, _SimdWrapper<_Tp, _Np>>
1805 _S_minmax(_SimdWrapper<_Tp, _Np> __a, _SimdWrapper<_Tp, _Np> __b)
1806 {
1807 return {__a._M_data < __b._M_data ? __a._M_data : __b._M_data,
1808 __a._M_data < __b._M_data ? __b._M_data : __a._M_data};
1809 }
1810
1811 // reductions {{{2
1812 template <size_t _Np, size_t... _Is, size_t... _Zeros, typename _Tp,
1813 typename _BinaryOperation>
1814 _GLIBCXX_SIMD_INTRINSIC static _Tp
1815 _S_reduce_partial(index_sequence<_Is...>, index_sequence<_Zeros...>,
1816 simd<_Tp, _Abi> __x, _BinaryOperation&& __binary_op)
1817 {
1818 using _V = __vector_type_t<_Tp, _Np / 2>;
1819 static_assert(sizeof(_V) <= sizeof(__x));
1820 // _S_full_size is the size of the smallest native SIMD register that
1821 // can store _Np/2 elements:
1822 using _FullSimd = __deduced_simd<_Tp, _VectorTraits<_V>::_S_full_size>;
1823 using _HalfSimd = __deduced_simd<_Tp, _Np / 2>;
1824 const auto __xx = __as_vector(__x);
1825 return _HalfSimd::abi_type::_SimdImpl::_S_reduce(
1826 static_cast<_HalfSimd>(__as_vector(__binary_op(
1827 static_cast<_FullSimd>(__intrin_bitcast<_V>(__xx)),
1828 static_cast<_FullSimd>(__intrin_bitcast<_V>(
1829 __vector_permute<(_Np / 2 + _Is)..., (int(_Zeros * 0) - 1)...>(
1830 __xx)))))),
1831 __binary_op);
1832 }
1833
1834 template <typename _Tp, typename _BinaryOperation>
1835 _GLIBCXX_SIMD_INTRINSIC static constexpr _Tp
1836 _S_reduce(simd<_Tp, _Abi> __x, _BinaryOperation&& __binary_op)
1837 {
1838 constexpr size_t _Np = simd_size_v<_Tp, _Abi>;
1839 if constexpr (_Np == 1)
1840 return __x[0];
1841 else if constexpr (_Np == 2)
1842 return __binary_op(simd<_Tp, simd_abi::scalar>(__x[0]),
1843 simd<_Tp, simd_abi::scalar>(__x[1]))[0];
1844 else if constexpr (_Abi::template _S_is_partial<_Tp>) //{{{
1845 {
1846 [[maybe_unused]] constexpr auto __full_size
1847 = _Abi::template _S_full_size<_Tp>;
1848 if constexpr (_Np == 3)
1849 return __binary_op(
1850 __binary_op(simd<_Tp, simd_abi::scalar>(__x[0]),
1851 simd<_Tp, simd_abi::scalar>(__x[1])),
1852 simd<_Tp, simd_abi::scalar>(__x[2]))[0];
1853 else if constexpr (is_same_v<__remove_cvref_t<_BinaryOperation>,
1854 plus<>>)
1855 {
1856 using _Ap = simd_abi::deduce_t<_Tp, __full_size>;
1857 return _Ap::_SimdImpl::_S_reduce(
1858 simd<_Tp, _Ap>(__private_init,
1859 _Abi::_S_masked(__as_vector(__x))),
1860 __binary_op);
1861 }
1862 else if constexpr (is_same_v<__remove_cvref_t<_BinaryOperation>,
1863 multiplies<>>)
1864 {
1865 using _Ap = simd_abi::deduce_t<_Tp, __full_size>;
1866 using _TW = _SimdWrapper<_Tp, __full_size>;
1867 _GLIBCXX_SIMD_USE_CONSTEXPR auto __implicit_mask_full
1868 = _Abi::template _S_implicit_mask<_Tp>().__as_full_vector();
1869 _GLIBCXX_SIMD_USE_CONSTEXPR _TW __one
1870 = __vector_broadcast<__full_size>(_Tp(1));
1871 const _TW __x_full = __data(__x).__as_full_vector();
1872 const _TW __x_padded_with_ones
1873 = _Ap::_CommonImpl::_S_blend(__implicit_mask_full, __one,
1874 __x_full);
1875 return _Ap::_SimdImpl::_S_reduce(
1876 simd<_Tp, _Ap>(__private_init, __x_padded_with_ones),
1877 __binary_op);
1878 }
1879 else if constexpr (_Np & 1)
1880 {
1881 using _Ap = simd_abi::deduce_t<_Tp, _Np - 1>;
1882 return __binary_op(
1883 simd<_Tp, simd_abi::scalar>(_Ap::_SimdImpl::_S_reduce(
1884 simd<_Tp, _Ap>(
1885 __intrin_bitcast<__vector_type_t<_Tp, _Np - 1>>(
1886 __as_vector(__x))),
1887 __binary_op)),
1888 simd<_Tp, simd_abi::scalar>(__x[_Np - 1]))[0];
1889 }
1890 else
1891 return _S_reduce_partial<_Np>(
1892 make_index_sequence<_Np / 2>(),
1893 make_index_sequence<__full_size - _Np / 2>(), __x, __binary_op);
1894 } //}}}
1895 else if constexpr (sizeof(__x) == 16) //{{{
1896 {
1897 if constexpr (_Np == 16)
1898 {
1899 const auto __y = __data(__x);
1900 __x = __binary_op(
1901 _M_make_simd<_Tp, _Np>(
1902 __vector_permute<0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6,
1903 7, 7>(__y)),
1904 _M_make_simd<_Tp, _Np>(
1905 __vector_permute<8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13,
1906 14, 14, 15, 15>(__y)));
1907 }
1908 if constexpr (_Np >= 8)
1909 {
1910 const auto __y = __vector_bitcast<short>(__data(__x));
1911 __x = __binary_op(
1912 _M_make_simd<_Tp, _Np>(__vector_bitcast<_Tp>(
1913 __vector_permute<0, 0, 1, 1, 2, 2, 3, 3>(__y))),
1914 _M_make_simd<_Tp, _Np>(__vector_bitcast<_Tp>(
1915 __vector_permute<4, 4, 5, 5, 6, 6, 7, 7>(__y))));
1916 }
1917 if constexpr (_Np >= 4)
1918 {
1919 using _Up = conditional_t<is_floating_point_v<_Tp>, float, int>;
1920 const auto __y = __vector_bitcast<_Up>(__data(__x));
1921 __x = __binary_op(__x,
1922 _M_make_simd<_Tp, _Np>(__vector_bitcast<_Tp>(
1923 __vector_permute<3, 2, 1, 0>(__y))));
1924 }
1925 using _Up = conditional_t<is_floating_point_v<_Tp>, double, _LLong>;
1926 const auto __y = __vector_bitcast<_Up>(__data(__x));
1927 __x = __binary_op(__x, _M_make_simd<_Tp, _Np>(__vector_bitcast<_Tp>(
1928 __vector_permute<1, 1>(__y))));
1929 return __x[0];
1930 } //}}}
1931 else
1932 {
1933 static_assert(sizeof(__x) > __min_vector_size<_Tp>);
1934 static_assert((_Np & (_Np - 1)) == 0); // _Np must be a power of 2
1935 using _Ap = simd_abi::deduce_t<_Tp, _Np / 2>;
1936 using _V = simd<_Tp, _Ap>;
1937 return _Ap::_SimdImpl::_S_reduce(
1938 __binary_op(_V(__private_init, __extract<0, 2>(__as_vector(__x))),
1939 _V(__private_init,
1940 __extract<1, 2>(__as_vector(__x)))),
1941 static_cast<_BinaryOperation&&>(__binary_op));
1942 }
1943 }
1944
1945 // math {{{2
1946 // frexp, modf and copysign implemented in simd_math.h
1947#define _GLIBCXX_SIMD_MATH_FALLBACK(__name) \
1948 template <typename _Tp, typename... _More> \
1949 static _Tp _S_##__name(const _Tp& __x, const _More&... __more) \
1950 { \
1951 return __generate_vector<_Tp>( \
1952 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA { \
1953 return __name(__x[__i], __more[__i]...); \
1954 }); \
1955 }
1956
1957#define _GLIBCXX_SIMD_MATH_FALLBACK_MASKRET(__name) \
1958 template <typename _Tp, typename... _More> \
1959 static typename _Tp::mask_type _S_##__name(const _Tp& __x, \
1960 const _More&... __more) \
1961 { \
1962 return __generate_vector<_Tp>( \
1963 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA { \
1964 return __name(__x[__i], __more[__i]...); \
1965 }); \
1966 }
1967
1968#define _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET(_RetTp, __name) \
1969 template <typename _Tp, typename... _More> \
1970 static auto _S_##__name(const _Tp& __x, const _More&... __more) \
1971 { \
1972 return __fixed_size_storage_t<_RetTp, \
1973 _VectorTraits<_Tp>::_S_partial_width>:: \
1974 _S_generate([&](auto __meta) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA { \
1975 return __meta._S_generator( \
1976 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA { \
1977 return __name(__x[__meta._S_offset + __i], \
1978 __more[__meta._S_offset + __i]...); \
1979 }, \
1980 static_cast<_RetTp*>(nullptr)); \
1981 }); \
1982 }
1983
1984 _GLIBCXX_SIMD_MATH_FALLBACK(acos)
1985 _GLIBCXX_SIMD_MATH_FALLBACK(asin)
1986 _GLIBCXX_SIMD_MATH_FALLBACK(atan)
1987 _GLIBCXX_SIMD_MATH_FALLBACK(atan2)
1988 _GLIBCXX_SIMD_MATH_FALLBACK(cos)
1989 _GLIBCXX_SIMD_MATH_FALLBACK(sin)
1990 _GLIBCXX_SIMD_MATH_FALLBACK(tan)
1991 _GLIBCXX_SIMD_MATH_FALLBACK(acosh)
1992 _GLIBCXX_SIMD_MATH_FALLBACK(asinh)
1993 _GLIBCXX_SIMD_MATH_FALLBACK(atanh)
1994 _GLIBCXX_SIMD_MATH_FALLBACK(cosh)
1995 _GLIBCXX_SIMD_MATH_FALLBACK(sinh)
1996 _GLIBCXX_SIMD_MATH_FALLBACK(tanh)
1997 _GLIBCXX_SIMD_MATH_FALLBACK(exp)
1998 _GLIBCXX_SIMD_MATH_FALLBACK(exp2)
1999 _GLIBCXX_SIMD_MATH_FALLBACK(expm1)
2000 _GLIBCXX_SIMD_MATH_FALLBACK(ldexp)
2001 _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET(int, ilogb)
2002 _GLIBCXX_SIMD_MATH_FALLBACK(log)
2003 _GLIBCXX_SIMD_MATH_FALLBACK(log10)
2004 _GLIBCXX_SIMD_MATH_FALLBACK(log1p)
2005 _GLIBCXX_SIMD_MATH_FALLBACK(log2)
2006 _GLIBCXX_SIMD_MATH_FALLBACK(logb)
2007
2008 // modf implemented in simd_math.h
2009 _GLIBCXX_SIMD_MATH_FALLBACK(scalbn)
2010 _GLIBCXX_SIMD_MATH_FALLBACK(scalbln)
2011 _GLIBCXX_SIMD_MATH_FALLBACK(cbrt)
2012 _GLIBCXX_SIMD_MATH_FALLBACK(fabs)
2013 _GLIBCXX_SIMD_MATH_FALLBACK(pow)
2014 _GLIBCXX_SIMD_MATH_FALLBACK(sqrt)
2015 _GLIBCXX_SIMD_MATH_FALLBACK(erf)
2016 _GLIBCXX_SIMD_MATH_FALLBACK(erfc)
2017 _GLIBCXX_SIMD_MATH_FALLBACK(lgamma)
2018 _GLIBCXX_SIMD_MATH_FALLBACK(tgamma)
2019
2020 _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET(long, lrint)
2021 _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET(long long, llrint)
2022
2023 _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET(long, lround)
2024 _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET(long long, llround)
2025
2026 _GLIBCXX_SIMD_MATH_FALLBACK(fmod)
2027 _GLIBCXX_SIMD_MATH_FALLBACK(remainder)
2028
2029 template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
2030 static _Tp
2031 _S_remquo(const _Tp __x, const _Tp __y,
2032 __fixed_size_storage_t<int, _TVT::_S_partial_width>* __z)
2033 {
2034 return __generate_vector<_Tp>([&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2035 int __tmp;
2036 auto __r = remquo(__x[__i], __y[__i], &__tmp);
2037 __z->_M_set(__i, __tmp);
2038 return __r;
2039 });
2040 }
2041
2042 // copysign in simd_math.h
2043 _GLIBCXX_SIMD_MATH_FALLBACK(nextafter)
2044 _GLIBCXX_SIMD_MATH_FALLBACK(fdim)
2045 _GLIBCXX_SIMD_MATH_FALLBACK(fmax)
2046 _GLIBCXX_SIMD_MATH_FALLBACK(fmin)
2047 _GLIBCXX_SIMD_MATH_FALLBACK(fma)
2048
2049 template <typename _Tp, size_t _Np>
2050 static constexpr _MaskMember<_Tp>
2051 _S_isgreater(_SimdWrapper<_Tp, _Np> __x,
2052 _SimdWrapper<_Tp, _Np> __y) noexcept
2053 {
2054 using _Ip = __int_for_sizeof_t<_Tp>;
2055 const auto __xn = __vector_bitcast<_Ip>(__x);
2056 const auto __yn = __vector_bitcast<_Ip>(__y);
2057 const auto __xp = __xn < 0 ? -(__xn & __finite_max_v<_Ip>) : __xn;
2058 const auto __yp = __yn < 0 ? -(__yn & __finite_max_v<_Ip>) : __yn;
2059 return __andnot(_SuperImpl::_S_isunordered(__x, __y)._M_data,
2060 __xp > __yp);
2061 }
2062
2063 template <typename _Tp, size_t _Np>
2064 static constexpr _MaskMember<_Tp>
2065 _S_isgreaterequal(_SimdWrapper<_Tp, _Np> __x,
2066 _SimdWrapper<_Tp, _Np> __y) noexcept
2067 {
2068 using _Ip = __int_for_sizeof_t<_Tp>;
2069 const auto __xn = __vector_bitcast<_Ip>(__x);
2070 const auto __yn = __vector_bitcast<_Ip>(__y);
2071 const auto __xp = __xn < 0 ? -(__xn & __finite_max_v<_Ip>) : __xn;
2072 const auto __yp = __yn < 0 ? -(__yn & __finite_max_v<_Ip>) : __yn;
2073 return __andnot(_SuperImpl::_S_isunordered(__x, __y)._M_data,
2074 __xp >= __yp);
2075 }
2076
2077 template <typename _Tp, size_t _Np>
2078 static constexpr _MaskMember<_Tp>
2079 _S_isless(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y) noexcept
2080 {
2081 using _Ip = __int_for_sizeof_t<_Tp>;
2082 const auto __xn = __vector_bitcast<_Ip>(__x);
2083 const auto __yn = __vector_bitcast<_Ip>(__y);
2084 const auto __xp = __xn < 0 ? -(__xn & __finite_max_v<_Ip>) : __xn;
2085 const auto __yp = __yn < 0 ? -(__yn & __finite_max_v<_Ip>) : __yn;
2086 return __andnot(_SuperImpl::_S_isunordered(__x, __y)._M_data,
2087 __xp < __yp);
2088 }
2089
2090 template <typename _Tp, size_t _Np>
2091 static constexpr _MaskMember<_Tp>
2092 _S_islessequal(_SimdWrapper<_Tp, _Np> __x,
2093 _SimdWrapper<_Tp, _Np> __y) noexcept
2094 {
2095 using _Ip = __int_for_sizeof_t<_Tp>;
2096 const auto __xn = __vector_bitcast<_Ip>(__x);
2097 const auto __yn = __vector_bitcast<_Ip>(__y);
2098 const auto __xp = __xn < 0 ? -(__xn & __finite_max_v<_Ip>) : __xn;
2099 const auto __yp = __yn < 0 ? -(__yn & __finite_max_v<_Ip>) : __yn;
2100 return __andnot(_SuperImpl::_S_isunordered(__x, __y)._M_data,
2101 __xp <= __yp);
2102 }
2103
2104 template <typename _Tp, size_t _Np>
2105 static constexpr _MaskMember<_Tp>
2106 _S_islessgreater(_SimdWrapper<_Tp, _Np> __x,
2107 _SimdWrapper<_Tp, _Np> __y) noexcept
2108 {
2109 return __andnot(_SuperImpl::_S_isunordered(__x, __y),
2110 _SuperImpl::_S_not_equal_to(__x, __y));
2111 }
2112
2113#undef _GLIBCXX_SIMD_MATH_FALLBACK
2114#undef _GLIBCXX_SIMD_MATH_FALLBACK_MASKRET
2115#undef _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET
2116 // _S_abs {{{3
2117 template <typename _Tp, size_t _Np>
2118 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
2119 _S_abs(_SimdWrapper<_Tp, _Np> __x) noexcept
2120 {
2121 // if (__builtin_is_constant_evaluated())
2122 // {
2123 // return __x._M_data < 0 ? -__x._M_data : __x._M_data;
2124 // }
2125 if constexpr (is_floating_point_v<_Tp>)
2126 // `v < 0 ? -v : v` cannot compile to the efficient implementation of
2127 // masking the signbit off because it must consider v == -0
2128
2129 // ~(-0.) & v would be easy, but breaks with fno-signed-zeros
2130 return __and(_S_absmask<__vector_type_t<_Tp, _Np>>, __x._M_data);
2131 else
2132 return __x._M_data < 0 ? -__x._M_data : __x._M_data;
2133 }
2134
2135 // }}}3
2136 // _S_plus_minus {{{
2137 // Returns __x + __y - __y without -fassociative-math optimizing to __x.
2138 // - _TV must be __vector_type_t<floating-point type, N>.
2139 // - _UV must be _TV or floating-point type.
2140 template <typename _TV, typename _UV>
2141 _GLIBCXX_SIMD_INTRINSIC static constexpr _TV _S_plus_minus(_TV __x,
2142 _UV __y) noexcept
2143 {
2144 #if defined __i386__ && !defined __SSE_MATH__
2145 if constexpr (sizeof(__x) == 8)
2146 { // operations on __x would use the FPU
2147 static_assert(is_same_v<_TV, __vector_type_t<float, 2>>);
2148 const auto __x4 = __vector_bitcast<float, 4>(__x);
2149 if constexpr (is_same_v<_TV, _UV>)
2150 return __vector_bitcast<float, 2>(
2151 _S_plus_minus(__x4, __vector_bitcast<float, 4>(__y)));
2152 else
2153 return __vector_bitcast<float, 2>(_S_plus_minus(__x4, __y));
2154 }
2155 #endif
2156 #if !defined __clang__ && __GCC_IEC_559 == 0
2157 if (__builtin_is_constant_evaluated()
2158 || (__builtin_constant_p(__x) && __builtin_constant_p(__y)))
2159 return (__x + __y) - __y;
2160 else
2161 return [&] {
2162 __x += __y;
2163 if constexpr(__have_sse)
2164 {
2165 if constexpr (sizeof(__x) >= 16)
2166 asm("" : "+x"(__x));
2167 else if constexpr (is_same_v<__vector_type_t<float, 2>, _TV>)
2168 asm("" : "+x"(__x[0]), "+x"(__x[1]));
2169 else
2170 __assert_unreachable<_TV>();
2171 }
2172 else if constexpr(__have_neon)
2173 asm("" : "+w"(__x));
2174 else if constexpr (__have_power_vmx)
2175 {
2176 if constexpr (is_same_v<__vector_type_t<float, 2>, _TV>)
2177 asm("" : "+fgr"(__x[0]), "+fgr"(__x[1]));
2178 else
2179 asm("" : "+v"(__x));
2180 }
2181 else
2182 asm("" : "+g"(__x));
2183 return __x - __y;
2184 }();
2185 #else
2186 return (__x + __y) - __y;
2187 #endif
2188 }
2189
2190 // }}}
2191 // _S_nearbyint {{{3
2192 template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
2193 _GLIBCXX_SIMD_INTRINSIC static _Tp _S_nearbyint(_Tp __x_) noexcept
2194 {
2195 using value_type = typename _TVT::value_type;
2196 using _V = typename _TVT::type;
2197 const _V __x = __x_;
2198 const _V __absx = __and(__x, _S_absmask<_V>);
2199 static_assert(__CHAR_BIT__ * sizeof(1ull) >= __digits_v<value_type>);
2200 _GLIBCXX_SIMD_USE_CONSTEXPR _V __shifter_abs
2201 = _V() + (1ull << (__digits_v<value_type> - 1));
2202 const _V __shifter = __or(__and(_S_signmask<_V>, __x), __shifter_abs);
2203 const _V __shifted = _S_plus_minus(__x, __shifter);
2204 return __absx < __shifter_abs ? __shifted : __x;
2205 }
2206
2207 // _S_rint {{{3
2208 template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
2209 _GLIBCXX_SIMD_INTRINSIC static _Tp _S_rint(_Tp __x) noexcept
2210 {
2211 return _SuperImpl::_S_nearbyint(__x);
2212 }
2213
2214 // _S_trunc {{{3
2215 template <typename _Tp, size_t _Np>
2216 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
2217 _S_trunc(_SimdWrapper<_Tp, _Np> __x)
2218 {
2219 using _V = __vector_type_t<_Tp, _Np>;
2220 const _V __absx = __and(__x._M_data, _S_absmask<_V>);
2221 static_assert(__CHAR_BIT__ * sizeof(1ull) >= __digits_v<_Tp>);
2222 constexpr _Tp __shifter = 1ull << (__digits_v<_Tp> - 1);
2223 _V __truncated = _S_plus_minus(__absx, __shifter);
2224 __truncated -= __truncated > __absx ? _V() + 1 : _V();
2225 return __absx < __shifter ? __or(__xor(__absx, __x._M_data), __truncated)
2226 : __x._M_data;
2227 }
2228
2229 // _S_round {{{3
2230 template <typename _Tp, size_t _Np>
2231 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
2232 _S_round(_SimdWrapper<_Tp, _Np> __x)
2233 {
2234 const auto __abs_x = _SuperImpl::_S_abs(__x);
2235 const auto __t_abs = _SuperImpl::_S_trunc(__abs_x)._M_data;
2236 const auto __r_abs // round(abs(x)) =
2237 = __t_abs + (__abs_x._M_data - __t_abs >= _Tp(.5) ? _Tp(1) : 0);
2238 return __or(__xor(__abs_x._M_data, __x._M_data), __r_abs);
2239 }
2240
2241 // _S_floor {{{3
2242 template <typename _Tp, size_t _Np>
2243 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
2244 _S_floor(_SimdWrapper<_Tp, _Np> __x)
2245 {
2246 const auto __y = _SuperImpl::_S_trunc(__x)._M_data;
2247 const auto __negative_input
2248 = __vector_bitcast<_Tp>(__x._M_data < __vector_broadcast<_Np, _Tp>(0));
2249 const auto __mask
2250 = __andnot(__vector_bitcast<_Tp>(__y == __x._M_data), __negative_input);
2251 return __or(__andnot(__mask, __y),
2252 __and(__mask, __y - __vector_broadcast<_Np, _Tp>(1)));
2253 }
2254
2255 // _S_ceil {{{3
2256 template <typename _Tp, size_t _Np>
2257 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
2258 _S_ceil(_SimdWrapper<_Tp, _Np> __x)
2259 {
2260 const auto __y = _SuperImpl::_S_trunc(__x)._M_data;
2261 const auto __negative_input
2262 = __vector_bitcast<_Tp>(__x._M_data < __vector_broadcast<_Np, _Tp>(0));
2263 const auto __inv_mask
2264 = __or(__vector_bitcast<_Tp>(__y == __x._M_data), __negative_input);
2265 return __or(__and(__inv_mask, __y),
2266 __andnot(__inv_mask, __y + __vector_broadcast<_Np, _Tp>(1)));
2267 }
2268
2269 // _S_isnan {{{3
2270 template <typename _Tp, size_t _Np>
2271 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
2272 _S_isnan([[maybe_unused]] _SimdWrapper<_Tp, _Np> __x)
2273 {
2274 #if __FINITE_MATH_ONLY__
2275 return {}; // false
2276 #elif !defined __SUPPORT_SNAN__
2277 return ~(__x._M_data == __x._M_data);
2278 #elif defined __STDC_IEC_559__
2279 using _Ip = __int_for_sizeof_t<_Tp>;
2280 const auto __absn = __vector_bitcast<_Ip>(_SuperImpl::_S_abs(__x));
2281 const auto __infn
2282 = __vector_bitcast<_Ip>(__vector_broadcast<_Np>(__infinity_v<_Tp>));
2283 return __infn < __absn;
2284 #else
2285 #error "Not implemented: how to support SNaN but non-IEC559 floating-point?"
2286 #endif
2287 }
2288
2289 // _S_isfinite {{{3
2290 template <typename _Tp, size_t _Np>
2291 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
2292 _S_isfinite([[maybe_unused]] _SimdWrapper<_Tp, _Np> __x)
2293 {
2294 #if __FINITE_MATH_ONLY__
2295 using _UV = typename _MaskMember<_Tp>::_BuiltinType;
2296 _GLIBCXX_SIMD_USE_CONSTEXPR _UV __alltrue = ~_UV();
2297 return __alltrue;
2298 #else
2299 // if all exponent bits are set, __x is either inf or NaN
2300 using _Ip = __int_for_sizeof_t<_Tp>;
2301 const auto __absn = __vector_bitcast<_Ip>(_SuperImpl::_S_abs(__x));
2302 const auto __maxn
2303 = __vector_bitcast<_Ip>(__vector_broadcast<_Np>(__finite_max_v<_Tp>));
2304 return __absn <= __maxn;
2305 #endif
2306 }
2307
2308 // _S_isunordered {{{3
2309 template <typename _Tp, size_t _Np>
2310 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
2311 _S_isunordered(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
2312 {
2313 return __or(_S_isnan(__x), _S_isnan(__y));
2314 }
2315
2316 // _S_signbit {{{3
2317 template <typename _Tp, size_t _Np>
2318 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
2319 _S_signbit(_SimdWrapper<_Tp, _Np> __x)
2320 {
2321 using _Ip = __int_for_sizeof_t<_Tp>;
2322 return __vector_bitcast<_Ip>(__x) < 0;
2323 // Arithmetic right shift (SRA) would also work (instead of compare), but
2324 // 64-bit SRA isn't available on x86 before AVX512. And in general,
2325 // compares are more likely to be efficient than SRA.
2326 }
2327
2328 // _S_isinf {{{3
2329 template <typename _Tp, size_t _Np>
2330 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
2331 _S_isinf([[maybe_unused]] _SimdWrapper<_Tp, _Np> __x)
2332 {
2333 #if __FINITE_MATH_ONLY__
2334 return {}; // false
2335 #else
2336 return _SuperImpl::template _S_equal_to<_Tp, _Np>(_SuperImpl::_S_abs(__x),
2337 __vector_broadcast<_Np>(
2338 __infinity_v<_Tp>));
2339 // alternative:
2340 // compare to inf using the corresponding integer type
2341 /*
2342 return
2343 __vector_bitcast<_Tp>(__vector_bitcast<__int_for_sizeof_t<_Tp>>(
2344 _S_abs(__x)._M_data)
2345 ==
2346 __vector_bitcast<__int_for_sizeof_t<_Tp>>(__vector_broadcast<_Np>(
2347 __infinity_v<_Tp>)));
2348 */
2349 #endif
2350 }
2351
2352 // _S_isnormal {{{3
2353 template <typename _Tp, size_t _Np>
2354 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
2355 _S_isnormal(_SimdWrapper<_Tp, _Np> __x)
2356 {
2357 using _Ip = __int_for_sizeof_t<_Tp>;
2358 const auto __absn = __vector_bitcast<_Ip>(_SuperImpl::_S_abs(__x));
2359 const auto __minn
2360 = __vector_bitcast<_Ip>(__vector_broadcast<_Np>(__norm_min_v<_Tp>));
2361 #if __FINITE_MATH_ONLY__
2362 return __absn >= __minn;
2363 #else
2364 const auto __maxn
2365 = __vector_bitcast<_Ip>(__vector_broadcast<_Np>(__finite_max_v<_Tp>));
2366 return __minn <= __absn && __absn <= __maxn;
2367 #endif
2368 }
2369
2370 // _S_fpclassify {{{3
2371 template <typename _Tp, size_t _Np>
2372 _GLIBCXX_SIMD_INTRINSIC static __fixed_size_storage_t<int, _Np>
2373 _S_fpclassify(_SimdWrapper<_Tp, _Np> __x)
2374 {
2375 using _I = __int_for_sizeof_t<_Tp>;
2376 const auto __xn
2377 = __vector_bitcast<_I>(__to_intrin(_SuperImpl::_S_abs(__x)));
2378 constexpr size_t _NI = sizeof(__xn) / sizeof(_I);
2379 _GLIBCXX_SIMD_USE_CONSTEXPR auto __minn
2380 = __vector_bitcast<_I>(__vector_broadcast<_NI>(__norm_min_v<_Tp>));
2381 _GLIBCXX_SIMD_USE_CONSTEXPR auto __infn
2382 = __vector_bitcast<_I>(__vector_broadcast<_NI>(__infinity_v<_Tp>));
2383
2384 _GLIBCXX_SIMD_USE_CONSTEXPR auto __fp_normal
2385 = __vector_broadcast<_NI, _I>(FP_NORMAL);
2386 #if !__FINITE_MATH_ONLY__
2387 _GLIBCXX_SIMD_USE_CONSTEXPR auto __fp_nan
2388 = __vector_broadcast<_NI, _I>(FP_NAN);
2389 _GLIBCXX_SIMD_USE_CONSTEXPR auto __fp_infinite
2390 = __vector_broadcast<_NI, _I>(FP_INFINITE);
2391 #endif
2392 #ifndef __FAST_MATH__
2393 _GLIBCXX_SIMD_USE_CONSTEXPR auto __fp_subnormal
2394 = __vector_broadcast<_NI, _I>(FP_SUBNORMAL);
2395 #endif
2396 _GLIBCXX_SIMD_USE_CONSTEXPR auto __fp_zero
2397 = __vector_broadcast<_NI, _I>(FP_ZERO);
2398
2399 __vector_type_t<_I, _NI>
2400 __tmp = __xn < __minn
2401 #ifdef __FAST_MATH__
2402 ? __fp_zero
2403 #else
2404 ? (__xn == 0 ? __fp_zero : __fp_subnormal)
2405 #endif
2406 #if __FINITE_MATH_ONLY__
2407 : __fp_normal;
2408 #else
2409 : (__xn < __infn ? __fp_normal
2410 : (__xn == __infn ? __fp_infinite : __fp_nan));
2411 #endif
2412
2413 if constexpr (sizeof(_I) == sizeof(int))
2414 {
2415 using _FixedInt = __fixed_size_storage_t<int, _Np>;
2416 const auto __as_int = __vector_bitcast<int, _Np>(__tmp);
2417 if constexpr (_FixedInt::_S_tuple_size == 1)
2418 return {__as_int};
2419 else if constexpr (_FixedInt::_S_tuple_size == 2
2420 && is_same_v<
2421 typename _FixedInt::_SecondType::_FirstAbi,
2422 simd_abi::scalar>)
2423 return {__extract<0, 2>(__as_int), __as_int[_Np - 1]};
2424 else if constexpr (_FixedInt::_S_tuple_size == 2)
2425 return {__extract<0, 2>(__as_int),
2426 __auto_bitcast(__extract<1, 2>(__as_int))};
2427 else
2428 __assert_unreachable<_Tp>();
2429 }
2430 else if constexpr (_Np == 2 && sizeof(_I) == 8
2431 && __fixed_size_storage_t<int, _Np>::_S_tuple_size == 2)
2432 {
2433 const auto __aslong = __vector_bitcast<_LLong>(__tmp);
2434 return {int(__aslong[0]), {int(__aslong[1])}};
2435 }
2436 #if _GLIBCXX_SIMD_X86INTRIN
2437 else if constexpr (sizeof(_Tp) == 8 && sizeof(__tmp) == 32
2438 && __fixed_size_storage_t<int, _Np>::_S_tuple_size == 1)
2439 return {_mm_packs_epi32(__to_intrin(__lo128(__tmp)),
2440 __to_intrin(__hi128(__tmp)))};
2441 else if constexpr (sizeof(_Tp) == 8 && sizeof(__tmp) == 64
2442 && __fixed_size_storage_t<int, _Np>::_S_tuple_size == 1)
2443 return {_mm512_cvtepi64_epi32(__to_intrin(__tmp))};
2444 #endif // _GLIBCXX_SIMD_X86INTRIN
2445 else if constexpr (__fixed_size_storage_t<int, _Np>::_S_tuple_size == 1)
2446 return {__call_with_subscripts<_Np>(__vector_bitcast<_LLong>(__tmp),
2447 [](auto... __l) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2448 return __make_wrapper<int>(__l...);
2449 })};
2450 else
2451 __assert_unreachable<_Tp>();
2452 }
2453
2454 // _S_increment & _S_decrement{{{2
2455 template <typename _Tp, size_t _Np>
2456 _GLIBCXX_SIMD_INTRINSIC static void
2457 _S_increment(_SimdWrapper<_Tp, _Np>& __x)
2458 { __x = __x._M_data + 1; }
2459
2460 template <typename _Tp, size_t _Np>
2461 _GLIBCXX_SIMD_INTRINSIC static void
2462 _S_decrement(_SimdWrapper<_Tp, _Np>& __x)
2463 { __x = __x._M_data - 1; }
2464
2465 // smart_reference access {{{2
2466 template <typename _Tp, size_t _Np, typename _Up>
2467 _GLIBCXX_SIMD_INTRINSIC constexpr static void
2468 _S_set(_SimdWrapper<_Tp, _Np>& __v, int __i, _Up&& __x) noexcept
2469 { __v._M_set(__i, static_cast<_Up&&>(__x)); }
2470
2471 // _S_masked_assign{{{2
2472 template <typename _Tp, typename _K, size_t _Np>
2473 _GLIBCXX_SIMD_INTRINSIC static void
2474 _S_masked_assign(_SimdWrapper<_K, _Np> __k, _SimdWrapper<_Tp, _Np>& __lhs,
2475 __type_identity_t<_SimdWrapper<_Tp, _Np>> __rhs)
2476 {
2477 if (__k._M_is_constprop_none_of())
2478 return;
2479 else if (__k._M_is_constprop_all_of())
2480 __lhs = __rhs;
2481 else
2482 __lhs = _CommonImpl::_S_blend(__k, __lhs, __rhs);
2483 }
2484
2485 template <typename _Tp, typename _K, size_t _Np>
2486 _GLIBCXX_SIMD_INTRINSIC static void
2487 _S_masked_assign(_SimdWrapper<_K, _Np> __k, _SimdWrapper<_Tp, _Np>& __lhs,
2488 __type_identity_t<_Tp> __rhs)
2489 {
2490 if (__k._M_is_constprop_none_of())
2491 return;
2492 else if (__k._M_is_constprop_all_of())
2493 __lhs = __vector_broadcast<_Np>(__rhs);
2494 else if (__builtin_constant_p(__rhs) && __rhs == 0)
2495 {
2496 if constexpr (!is_same_v<bool, _K>)
2497 // the __andnot optimization only makes sense if __k._M_data is a
2498 // vector register
2499 __lhs._M_data
2500 = __andnot(__vector_bitcast<_Tp>(__k), __lhs._M_data);
2501 else
2502 // for AVX512/__mmask, a _mm512_maskz_mov is best
2503 __lhs
2504 = _CommonImpl::_S_blend(__k, __lhs, _SimdWrapper<_Tp, _Np>());
2505 }
2506 else
2507 __lhs = _CommonImpl::_S_blend(__k, __lhs,
2508 _SimdWrapper<_Tp, _Np>(
2509 __vector_broadcast<_Np>(__rhs)));
2510 }
2511
2512 // _S_masked_cassign {{{2
2513 template <typename _Op, typename _Tp, typename _K, size_t _Np>
2514 _GLIBCXX_SIMD_INTRINSIC static void
2515 _S_masked_cassign(const _SimdWrapper<_K, _Np> __k,
2516 _SimdWrapper<_Tp, _Np>& __lhs,
2517 const __type_identity_t<_SimdWrapper<_Tp, _Np>> __rhs,
2518 _Op __op)
2519 {
2520 if (__k._M_is_constprop_none_of())
2521 return;
2522 else if (__k._M_is_constprop_all_of())
2523 __lhs = __op(_SuperImpl{}, __lhs, __rhs);
2524 else
2525 __lhs = _CommonImpl::_S_blend(__k, __lhs,
2526 __op(_SuperImpl{}, __lhs, __rhs));
2527 }
2528
2529 template <typename _Op, typename _Tp, typename _K, size_t _Np>
2530 _GLIBCXX_SIMD_INTRINSIC static void
2531 _S_masked_cassign(const _SimdWrapper<_K, _Np> __k,
2532 _SimdWrapper<_Tp, _Np>& __lhs,
2533 const __type_identity_t<_Tp> __rhs, _Op __op)
2534 { _S_masked_cassign(__k, __lhs, __vector_broadcast<_Np>(__rhs), __op); }
2535
2536 // _S_masked_unary {{{2
2537 template <template <typename> class _Op, typename _Tp, typename _K,
2538 size_t _Np>
2539 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
2540 _S_masked_unary(const _SimdWrapper<_K, _Np> __k,
2541 const _SimdWrapper<_Tp, _Np> __v)
2542 {
2543 if (__k._M_is_constprop_none_of())
2544 return __v;
2545 auto __vv = _M_make_simd(__v);
2546 _Op<decltype(__vv)> __op;
2547 if (__k._M_is_constprop_all_of())
2548 return __data(__op(__vv));
2549 else
2550 return _CommonImpl::_S_blend(__k, __v, __data(__op(__vv)));
2551 }
2552
2553 //}}}2
2554 };
2555
2556// _MaskImplBuiltinMixin {{{1
2557struct _MaskImplBuiltinMixin
2558{
2559 template <typename _Tp>
2560 using _TypeTag = _Tp*;
2561
2562 // _S_to_maskvector {{{
2563 template <typename _Up, size_t _ToN = 1>
2564 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Up, _ToN>
2565 _S_to_maskvector(bool __x)
2566 {
2567 static_assert(is_same_v<_Up, __int_for_sizeof_t<_Up>>);
2568 return __x ? __vector_type_t<_Up, _ToN>{~_Up()}
2569 : __vector_type_t<_Up, _ToN>{};
2570 }
2571
2572 template <typename _Up, size_t _UpN = 0, size_t _Np, bool _Sanitized,
2573 size_t _ToN = _UpN == 0 ? _Np : _UpN>
2574 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Up, _ToN>
2575 _S_to_maskvector(_BitMask<_Np, _Sanitized> __x)
2576 {
2577 static_assert(is_same_v<_Up, __int_for_sizeof_t<_Up>>);
2578 return __generate_vector<__vector_type_t<_Up, _ToN>>(
2579 [&](auto __i) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2580 if constexpr (__i < _Np)
2581 return __x[__i] ? ~_Up() : _Up();
2582 else
2583 return _Up();
2584 });
2585 }
2586
2587 template <typename _Up, size_t _UpN = 0, typename _Tp, size_t _Np,
2588 size_t _ToN = _UpN == 0 ? _Np : _UpN>
2589 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Up, _ToN>
2590 _S_to_maskvector(_SimdWrapper<_Tp, _Np> __x)
2591 {
2592 static_assert(is_same_v<_Up, __int_for_sizeof_t<_Up>>);
2593 using _TW = _SimdWrapper<_Tp, _Np>;
2594 using _UW = _SimdWrapper<_Up, _ToN>;
2595 if constexpr (sizeof(_Up) == sizeof(_Tp) && sizeof(_TW) == sizeof(_UW))
2596 return __wrapper_bitcast<_Up, _ToN>(__x);
2597 else if constexpr (is_same_v<_Tp, bool>) // bits -> vector
2598 return _S_to_maskvector<_Up, _ToN>(_BitMask<_Np>(__x._M_data));
2599 else
2600 { // vector -> vector
2601 /*
2602 [[maybe_unused]] const auto __y = __vector_bitcast<_Up>(__x._M_data);
2603 if constexpr (sizeof(_Tp) == 8 && sizeof(_Up) == 4 && sizeof(__y) ==
2604 16) return __vector_permute<1, 3, -1, -1>(__y); else if constexpr
2605 (sizeof(_Tp) == 4 && sizeof(_Up) == 2
2606 && sizeof(__y) == 16)
2607 return __vector_permute<1, 3, 5, 7, -1, -1, -1, -1>(__y);
2608 else if constexpr (sizeof(_Tp) == 8 && sizeof(_Up) == 2
2609 && sizeof(__y) == 16)
2610 return __vector_permute<3, 7, -1, -1, -1, -1, -1, -1>(__y);
2611 else if constexpr (sizeof(_Tp) == 2 && sizeof(_Up) == 1
2612 && sizeof(__y) == 16)
2613 return __vector_permute<1, 3, 5, 7, 9, 11, 13, 15, -1, -1, -1, -1,
2614 -1, -1, -1, -1>(__y); else if constexpr (sizeof(_Tp) == 4 &&
2615 sizeof(_Up) == 1
2616 && sizeof(__y) == 16)
2617 return __vector_permute<3, 7, 11, 15, -1, -1, -1, -1, -1, -1, -1,
2618 -1, -1, -1, -1, -1>(__y); else if constexpr (sizeof(_Tp) == 8 &&
2619 sizeof(_Up) == 1
2620 && sizeof(__y) == 16)
2621 return __vector_permute<7, 15, -1, -1, -1, -1, -1, -1, -1, -1, -1,
2622 -1, -1, -1, -1, -1>(__y); else
2623 */
2624 {
2625 return __generate_vector<__vector_type_t<_Up, _ToN>>(
2626 [&](auto __i) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2627 if constexpr (__i < _Np)
2628 return _Up(__x[__i.value]);
2629 else
2630 return _Up();
2631 });
2632 }
2633 }
2634 }
2635
2636 // }}}
2637 // _S_to_bits {{{
2638 template <typename _Tp, size_t _Np>
2639 _GLIBCXX_SIMD_INTRINSIC static constexpr _SanitizedBitMask<_Np>
2640 _S_to_bits(_SimdWrapper<_Tp, _Np> __x)
2641 {
2642 static_assert(!is_same_v<_Tp, bool>);
2643 static_assert(_Np <= __CHAR_BIT__ * sizeof(_ULLong));
2644 using _Up = make_unsigned_t<__int_for_sizeof_t<_Tp>>;
2645 const auto __bools
2646 = __vector_bitcast<_Up>(__x) >> (sizeof(_Up) * __CHAR_BIT__ - 1);
2647 _ULLong __r = 0;
2648 __execute_n_times<_Np>(
2649 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2650 __r |= _ULLong(__bools[__i.value]) << __i;
2651 });
2652 return __r;
2653 }
2654
2655 // }}}
2656};
2657
2658// _MaskImplBuiltin {{{1
2659template <typename _Abi, typename>
2660 struct _MaskImplBuiltin : _MaskImplBuiltinMixin
2661 {
2662 using _MaskImplBuiltinMixin::_S_to_bits;
2663 using _MaskImplBuiltinMixin::_S_to_maskvector;
2664
2665 // member types {{{
2666 template <typename _Tp>
2667 using _SimdMember = typename _Abi::template __traits<_Tp>::_SimdMember;
2668
2669 template <typename _Tp>
2670 using _MaskMember = typename _Abi::template _MaskMember<_Tp>;
2671
2672 using _SuperImpl = typename _Abi::_MaskImpl;
2673 using _CommonImpl = typename _Abi::_CommonImpl;
2674
2675 template <typename _Tp>
2676 static constexpr size_t _S_size = simd_size_v<_Tp, _Abi>;
2677
2678 // }}}
2679 // _S_broadcast {{{
2680 template <typename _Tp>
2681 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
2682 _S_broadcast(bool __x)
2683 {
2684 return __x ? _Abi::template _S_implicit_mask<_Tp>()
2685 : _MaskMember<_Tp>();
2686 }
2687
2688 // }}}
2689 // _S_load {{{
2690 template <typename _Tp>
2691 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
2692 _S_load(const bool* __mem)
2693 {
2694 using _I = __int_for_sizeof_t<_Tp>;
2695 if constexpr (sizeof(_Tp) == sizeof(bool))
2696 {
2697 const auto __bools
2698 = _CommonImpl::template _S_load<_I, _S_size<_Tp>>(__mem);
2699 // bool is {0, 1}, everything else is UB
2700 return __bools > 0;
2701 }
2702 else
2703 return __generate_vector<_I, _S_size<_Tp>>(
2704 [&](auto __i) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2705 return __mem[__i] ? ~_I() : _I();
2706 });
2707 }
2708
2709 // }}}
2710 // _S_convert {{{
2711 template <typename _Tp, size_t _Np, bool _Sanitized>
2712 _GLIBCXX_SIMD_INTRINSIC static constexpr auto
2713 _S_convert(_BitMask<_Np, _Sanitized> __x)
2714 {
2715 if constexpr (__is_builtin_bitmask_abi<_Abi>())
2716 return _SimdWrapper<bool, simd_size_v<_Tp, _Abi>>(__x._M_to_bits());
2717 else
2718 return _SuperImpl::template _S_to_maskvector<__int_for_sizeof_t<_Tp>,
2719 _S_size<_Tp>>(
2720 __x._M_sanitized());
2721 }
2722
2723 template <typename _Tp, size_t _Np>
2724 _GLIBCXX_SIMD_INTRINSIC static constexpr auto
2725 _S_convert(_SimdWrapper<bool, _Np> __x)
2726 {
2727 if constexpr (__is_builtin_bitmask_abi<_Abi>())
2728 return _SimdWrapper<bool, simd_size_v<_Tp, _Abi>>(__x._M_data);
2729 else
2730 return _SuperImpl::template _S_to_maskvector<__int_for_sizeof_t<_Tp>,
2731 _S_size<_Tp>>(
2732 _BitMask<_Np>(__x._M_data)._M_sanitized());
2733 }
2734
2735 template <typename _Tp, typename _Up, size_t _Np>
2736 _GLIBCXX_SIMD_INTRINSIC static constexpr auto
2737 _S_convert(_SimdWrapper<_Up, _Np> __x)
2738 {
2739 if constexpr (__is_builtin_bitmask_abi<_Abi>())
2740 return _SimdWrapper<bool, simd_size_v<_Tp, _Abi>>(
2741 _SuperImpl::_S_to_bits(__x));
2742 else
2743 return _SuperImpl::template _S_to_maskvector<__int_for_sizeof_t<_Tp>,
2744 _S_size<_Tp>>(__x);
2745 }
2746
2747 template <typename _Tp, typename _Up, typename _UAbi>
2748 _GLIBCXX_SIMD_INTRINSIC static constexpr auto
2749 _S_convert(simd_mask<_Up, _UAbi> __x)
2750 {
2751 if constexpr (__is_builtin_bitmask_abi<_Abi>())
2752 {
2753 using _R = _SimdWrapper<bool, simd_size_v<_Tp, _Abi>>;
2754 if constexpr (__is_builtin_bitmask_abi<_UAbi>()) // bits -> bits
2755 return _R(__data(__x));
2756 else if constexpr (__is_scalar_abi<_UAbi>()) // bool -> bits
2757 return _R(__data(__x));
2758 else if constexpr (__is_fixed_size_abi_v<_UAbi>) // bitset -> bits
2759 return _R(__data(__x)._M_to_bits());
2760 else // vector -> bits
2761 return _R(_UAbi::_MaskImpl::_S_to_bits(__data(__x))._M_to_bits());
2762 }
2763 else
2764 return _SuperImpl::template _S_to_maskvector<__int_for_sizeof_t<_Tp>,
2765 _S_size<_Tp>>(
2766 __data(__x));
2767 }
2768
2769 // }}}
2770 // _S_masked_load {{{2
2771 template <typename _Tp, size_t _Np>
2772 static inline _SimdWrapper<_Tp, _Np>
2773 _S_masked_load(_SimdWrapper<_Tp, _Np> __merge,
2774 _SimdWrapper<_Tp, _Np> __mask, const bool* __mem) noexcept
2775 {
2776 // AVX(2) has 32/64 bit maskload, but nothing at 8 bit granularity
2777 auto __tmp = __wrapper_bitcast<__int_for_sizeof_t<_Tp>>(__merge);
2778 _BitOps::_S_bit_iteration(_SuperImpl::_S_to_bits(__mask),
2779 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2780 __tmp._M_set(__i, -__mem[__i]);
2781 });
2782 __merge = __wrapper_bitcast<_Tp>(__tmp);
2783 return __merge;
2784 }
2785
2786 // _S_store {{{2
2787 template <typename _Tp, size_t _Np>
2788 _GLIBCXX_SIMD_INTRINSIC static void _S_store(_SimdWrapper<_Tp, _Np> __v,
2789 bool* __mem) noexcept
2790 {
2791 __execute_n_times<_Np>([&](auto __i) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2792 __mem[__i] = __v[__i];
2793 });
2794 }
2795
2796 // _S_masked_store {{{2
2797 template <typename _Tp, size_t _Np>
2798 static inline void
2799 _S_masked_store(const _SimdWrapper<_Tp, _Np> __v, bool* __mem,
2800 const _SimdWrapper<_Tp, _Np> __k) noexcept
2801 {
2802 _BitOps::_S_bit_iteration(_SuperImpl::_S_to_bits(__k),
2803 [&](auto __i) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2804 __mem[__i] = __v[__i];
2805 });
2806 }
2807
2808 // _S_from_bitmask{{{2
2809 template <size_t _Np, typename _Tp>
2810 _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
2811 _S_from_bitmask(_SanitizedBitMask<_Np> __bits, _TypeTag<_Tp>)
2812 {
2813 return _SuperImpl::template _S_to_maskvector<_Tp, _S_size<_Tp>>(__bits);
2814 }
2815
2816 // logical and bitwise operators {{{2
2817 template <typename _Tp, size_t _Np>
2818 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
2819 _S_logical_and(const _SimdWrapper<_Tp, _Np>& __x,
2820 const _SimdWrapper<_Tp, _Np>& __y)
2821 { return __and(__x._M_data, __y._M_data); }
2822
2823 template <typename _Tp, size_t _Np>
2824 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
2825 _S_logical_or(const _SimdWrapper<_Tp, _Np>& __x,
2826 const _SimdWrapper<_Tp, _Np>& __y)
2827 { return __or(__x._M_data, __y._M_data); }
2828
2829 template <typename _Tp, size_t _Np>
2830 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
2831 _S_bit_not(const _SimdWrapper<_Tp, _Np>& __x)
2832 {
2833 if constexpr (_Abi::template _S_is_partial<_Tp>)
2834 return __andnot(__x, __wrapper_bitcast<_Tp>(
2835 _Abi::template _S_implicit_mask<_Tp>()));
2836 else
2837 return __not(__x._M_data);
2838 }
2839
2840 template <typename _Tp, size_t _Np>
2841 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
2842 _S_bit_and(const _SimdWrapper<_Tp, _Np>& __x,
2843 const _SimdWrapper<_Tp, _Np>& __y)
2844 { return __and(__x._M_data, __y._M_data); }
2845
2846 template <typename _Tp, size_t _Np>
2847 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
2848 _S_bit_or(const _SimdWrapper<_Tp, _Np>& __x,
2849 const _SimdWrapper<_Tp, _Np>& __y)
2850 { return __or(__x._M_data, __y._M_data); }
2851
2852 template <typename _Tp, size_t _Np>
2853 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
2854 _S_bit_xor(const _SimdWrapper<_Tp, _Np>& __x,
2855 const _SimdWrapper<_Tp, _Np>& __y)
2856 { return __xor(__x._M_data, __y._M_data); }
2857
2858 // smart_reference access {{{2
2859 template <typename _Tp, size_t _Np>
2860 static constexpr void _S_set(_SimdWrapper<_Tp, _Np>& __k, int __i,
2861 bool __x) noexcept
2862 {
2863 if constexpr (is_same_v<_Tp, bool>)
2864 __k._M_set(__i, __x);
2865 else
2866 {
2867 static_assert(is_same_v<_Tp, __int_for_sizeof_t<_Tp>>);
2868 if (__builtin_is_constant_evaluated())
2869 {
2870 __k = __generate_from_n_evaluations<_Np,
2871 __vector_type_t<_Tp, _Np>>(
2872 [&](auto __j) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2873 if (__i == __j)
2874 return _Tp(-__x);
2875 else
2876 return __k[+__j];
2877 });
2878 }
2879 else
2880 __k._M_data[__i] = -__x;
2881 }
2882 }
2883
2884 // _S_masked_assign{{{2
2885 template <typename _Tp, size_t _Np>
2886 _GLIBCXX_SIMD_INTRINSIC static void
2887 _S_masked_assign(_SimdWrapper<_Tp, _Np> __k,
2888 _SimdWrapper<_Tp, _Np>& __lhs,
2889 __type_identity_t<_SimdWrapper<_Tp, _Np>> __rhs)
2890 { __lhs = _CommonImpl::_S_blend(__k, __lhs, __rhs); }
2891
2892 template <typename _Tp, size_t _Np>
2893 _GLIBCXX_SIMD_INTRINSIC static void
2894 _S_masked_assign(_SimdWrapper<_Tp, _Np> __k,
2895 _SimdWrapper<_Tp, _Np>& __lhs, bool __rhs)
2896 {
2897 if (__builtin_constant_p(__rhs))
2898 {
2899 if (__rhs == false)
2900 __lhs = __andnot(__k, __lhs);
2901 else
2902 __lhs = __or(__k, __lhs);
2903 return;
2904 }
2905 __lhs = _CommonImpl::_S_blend(__k, __lhs,
2906 __data(simd_mask<_Tp, _Abi>(__rhs)));
2907 }
2908
2909 //}}}2
2910 // _S_all_of {{{
2911 template <typename _Tp>
2912 _GLIBCXX_SIMD_INTRINSIC static bool
2913 _S_all_of(simd_mask<_Tp, _Abi> __k)
2914 {
2915 return __call_with_subscripts(
2916 __data(__k), make_index_sequence<_S_size<_Tp>>(),
2917 [](const auto... __ent) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA
2918 { return (... && !(__ent == 0)); });
2919 }
2920
2921 // }}}
2922 // _S_any_of {{{
2923 template <typename _Tp>
2924 _GLIBCXX_SIMD_INTRINSIC static bool
2925 _S_any_of(simd_mask<_Tp, _Abi> __k)
2926 {
2927 return __call_with_subscripts(
2928 __data(__k), make_index_sequence<_S_size<_Tp>>(),
2929 [](const auto... __ent) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA
2930 { return (... || !(__ent == 0)); });
2931 }
2932
2933 // }}}
2934 // _S_none_of {{{
2935 template <typename _Tp>
2936 _GLIBCXX_SIMD_INTRINSIC static bool
2937 _S_none_of(simd_mask<_Tp, _Abi> __k)
2938 {
2939 return __call_with_subscripts(
2940 __data(__k), make_index_sequence<_S_size<_Tp>>(),
2941 [](const auto... __ent) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA
2942 { return (... && (__ent == 0)); });
2943 }
2944
2945 // }}}
2946 // _S_some_of {{{
2947 template <typename _Tp>
2948 _GLIBCXX_SIMD_INTRINSIC static bool
2949 _S_some_of(simd_mask<_Tp, _Abi> __k)
2950 {
2951 const int __n_true = _SuperImpl::_S_popcount(__k);
2952 return __n_true > 0 && __n_true < int(_S_size<_Tp>);
2953 }
2954
2955 // }}}
2956 // _S_popcount {{{
2957 template <typename _Tp>
2958 _GLIBCXX_SIMD_INTRINSIC static int
2959 _S_popcount(simd_mask<_Tp, _Abi> __k)
2960 {
2961 using _I = __int_for_sizeof_t<_Tp>;
2962 if constexpr (is_default_constructible_v<simd<_I, _Abi>>)
2963 return -reduce(
2964 simd<_I, _Abi>(__private_init, __wrapper_bitcast<_I>(__data(__k))));
2965 else
2966 return -reduce(__bit_cast<rebind_simd_t<_I, simd<_Tp, _Abi>>>(
2967 simd<_Tp, _Abi>(__private_init, __data(__k))));
2968 }
2969
2970 // }}}
2971 // _S_find_first_set {{{
2972 template <typename _Tp>
2973 _GLIBCXX_SIMD_INTRINSIC static int
2974 _S_find_first_set(simd_mask<_Tp, _Abi> __k)
2975 {
2976 return std::__countr_zero(
2977 _SuperImpl::_S_to_bits(__data(__k))._M_to_bits());
2978 }
2979
2980 // }}}
2981 // _S_find_last_set {{{
2982 template <typename _Tp>
2983 _GLIBCXX_SIMD_INTRINSIC static int
2984 _S_find_last_set(simd_mask<_Tp, _Abi> __k)
2985 {
2986 return std::__bit_width(
2987 _SuperImpl::_S_to_bits(__data(__k))._M_to_bits()) - 1;
2988 }
2989
2990 // }}}
2991 };
2992
2993//}}}1
2994_GLIBCXX_SIMD_END_NAMESPACE
2995#endif // __cplusplus >= 201703L
2996#endif // _GLIBCXX_EXPERIMENTAL_SIMD_ABIS_H_
2997
2998// vim: foldmethod=marker foldmarker={{{,}}} sw=2 noet ts=8 sts=2 tw=100
integral_constant< bool, true > true_type
The type used as a compile-time boolean with true value.
Definition: type_traits:82
typename conditional< _Cond, _Iftrue, _Iffalse >::type conditional_t
Alias template for conditional.
Definition: type_traits:2614
constexpr const _Tp & min(const _Tp &, const _Tp &)
This does what you think it does.
Definition: stl_algobase.h:233
constexpr _Tp reduce(_InputIterator __first, _InputIterator __last, _Tp __init, _BinaryOperation __binary_op)
Calculate reduction of values in a range.
Definition: numeric:287