libstdc++
simd.h
1// Definition of the public simd interfaces -*- C++ -*-
2
3// Copyright (C) 2020-2021 Free Software Foundation, Inc.
4//
5// This file is part of the GNU ISO C++ Library. This library is free
6// software; you can redistribute it and/or modify it under the
7// terms of the GNU General Public License as published by the
8// Free Software Foundation; either version 3, or (at your option)
9// any later version.
10
11// This library is distributed in the hope that it will be useful,
12// but WITHOUT ANY WARRANTY; without even the implied warranty of
13// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14// GNU General Public License for more details.
15
16// Under Section 7 of GPL version 3, you are granted additional
17// permissions described in the GCC Runtime Library Exception, version
18// 3.1, as published by the Free Software Foundation.
19
20// You should have received a copy of the GNU General Public License and
21// a copy of the GCC Runtime Library Exception along with this program;
22// see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
23// <http://www.gnu.org/licenses/>.
24
25#ifndef _GLIBCXX_EXPERIMENTAL_SIMD_H
26#define _GLIBCXX_EXPERIMENTAL_SIMD_H
27
28#if __cplusplus >= 201703L
29
30#include "simd_detail.h"
31#include "numeric_traits.h"
32#include <bit>
33#include <bitset>
34#ifdef _GLIBCXX_DEBUG_UB
35#include <cstdio> // for stderr
36#endif
37#include <cstring>
38#include <functional>
39#include <iosfwd>
40#include <utility>
41
42#if _GLIBCXX_SIMD_X86INTRIN
43#include <x86intrin.h>
44#elif _GLIBCXX_SIMD_HAVE_NEON
45#include <arm_neon.h>
46#endif
47
48/** @ingroup ts_simd
49 * @{
50 */
51/* There are several closely related types, with the following naming
52 * convention:
53 * _Tp: vectorizable (arithmetic) type (or any type)
54 * _TV: __vector_type_t<_Tp, _Np>
55 * _TW: _SimdWrapper<_Tp, _Np>
56 * _TI: __intrinsic_type_t<_Tp, _Np>
57 * _TVT: _VectorTraits<_TV> or _VectorTraits<_TW>
58 * If one additional type is needed use _U instead of _T.
59 * Otherwise use _T\d, _TV\d, _TW\d, TI\d, _TVT\d.
60 *
61 * More naming conventions:
62 * _Ap or _Abi: An ABI tag from the simd_abi namespace
63 * _Ip: often used for integer types with sizeof(_Ip) == sizeof(_Tp),
64 * _IV, _IW as for _TV, _TW
65 * _Np: number of elements (not bytes)
66 * _Bytes: number of bytes
67 *
68 * Variable names:
69 * __k: mask object (vector- or bitmask)
70 */
71_GLIBCXX_SIMD_BEGIN_NAMESPACE
72
73#if !_GLIBCXX_SIMD_X86INTRIN
74using __m128 [[__gnu__::__vector_size__(16)]] = float;
75using __m128d [[__gnu__::__vector_size__(16)]] = double;
76using __m128i [[__gnu__::__vector_size__(16)]] = long long;
77using __m256 [[__gnu__::__vector_size__(32)]] = float;
78using __m256d [[__gnu__::__vector_size__(32)]] = double;
79using __m256i [[__gnu__::__vector_size__(32)]] = long long;
80using __m512 [[__gnu__::__vector_size__(64)]] = float;
81using __m512d [[__gnu__::__vector_size__(64)]] = double;
82using __m512i [[__gnu__::__vector_size__(64)]] = long long;
83#endif
84
85namespace simd_abi {
86// simd_abi forward declarations {{{
87// implementation details:
88struct _Scalar;
89
90template <int _Np>
91 struct _Fixed;
92
93// There are two major ABIs that appear on different architectures.
94// Both have non-boolean values packed into an N Byte register
95// -> #elements = N / sizeof(T)
96// Masks differ:
97// 1. Use value vector registers for masks (all 0 or all 1)
98// 2. Use bitmasks (mask registers) with one bit per value in the corresponding
99// value vector
100//
101// Both can be partially used, masking off the rest when doing horizontal
102// operations or operations that can trap (e.g. FP_INVALID or integer division
103// by 0). This is encoded as the number of used bytes.
104template <int _UsedBytes>
105 struct _VecBuiltin;
106
107template <int _UsedBytes>
108 struct _VecBltnBtmsk;
109
110template <typename _Tp, int _Np>
111 using _VecN = _VecBuiltin<sizeof(_Tp) * _Np>;
112
113template <int _UsedBytes = 16>
114 using _Sse = _VecBuiltin<_UsedBytes>;
115
116template <int _UsedBytes = 32>
117 using _Avx = _VecBuiltin<_UsedBytes>;
118
119template <int _UsedBytes = 64>
120 using _Avx512 = _VecBltnBtmsk<_UsedBytes>;
121
122template <int _UsedBytes = 16>
123 using _Neon = _VecBuiltin<_UsedBytes>;
124
125// implementation-defined:
126using __sse = _Sse<>;
127using __avx = _Avx<>;
128using __avx512 = _Avx512<>;
129using __neon = _Neon<>;
130using __neon128 = _Neon<16>;
131using __neon64 = _Neon<8>;
132
133// standard:
134template <typename _Tp, size_t _Np, typename...>
135 struct deduce;
136
137template <int _Np>
138 using fixed_size = _Fixed<_Np>;
139
140using scalar = _Scalar;
141
142// }}}
143} // namespace simd_abi
144// forward declarations is_simd(_mask), simd(_mask), simd_size {{{
145template <typename _Tp>
146 struct is_simd;
147
148template <typename _Tp>
149 struct is_simd_mask;
150
151template <typename _Tp, typename _Abi>
152 class simd;
153
154template <typename _Tp, typename _Abi>
155 class simd_mask;
156
157template <typename _Tp, typename _Abi>
158 struct simd_size;
159
160// }}}
161// load/store flags {{{
162struct element_aligned_tag
163{
164 template <typename _Tp, typename _Up = typename _Tp::value_type>
165 static constexpr size_t _S_alignment = alignof(_Up);
166
167 template <typename _Tp, typename _Up>
168 _GLIBCXX_SIMD_INTRINSIC static constexpr _Up*
169 _S_apply(_Up* __ptr)
170 { return __ptr; }
171};
172
173struct vector_aligned_tag
174{
175 template <typename _Tp, typename _Up = typename _Tp::value_type>
176 static constexpr size_t _S_alignment
177 = std::__bit_ceil(sizeof(_Up) * _Tp::size());
178
179 template <typename _Tp, typename _Up>
180 _GLIBCXX_SIMD_INTRINSIC static constexpr _Up*
181 _S_apply(_Up* __ptr)
182 {
183 return static_cast<_Up*>(
184 __builtin_assume_aligned(__ptr, _S_alignment<_Tp, _Up>));
185 }
186};
187
188template <size_t _Np> struct overaligned_tag
189{
190 template <typename _Tp, typename _Up = typename _Tp::value_type>
191 static constexpr size_t _S_alignment = _Np;
192
193 template <typename _Tp, typename _Up>
194 _GLIBCXX_SIMD_INTRINSIC static constexpr _Up*
195 _S_apply(_Up* __ptr)
196 { return static_cast<_Up*>(__builtin_assume_aligned(__ptr, _Np)); }
197};
198
199inline constexpr element_aligned_tag element_aligned = {};
200
201inline constexpr vector_aligned_tag vector_aligned = {};
202
203template <size_t _Np>
204 inline constexpr overaligned_tag<_Np> overaligned = {};
205
206// }}}
207template <size_t _Xp>
208 using _SizeConstant = integral_constant<size_t, _Xp>;
209
210namespace __detail
211{
212 struct _Minimum
213 {
214 template <typename _Tp>
215 _GLIBCXX_SIMD_INTRINSIC constexpr
216 _Tp
217 operator()(_Tp __a, _Tp __b) const
218 {
219 using std::min;
220 return min(__a, __b);
221 }
222 };
223
224 struct _Maximum
225 {
226 template <typename _Tp>
227 _GLIBCXX_SIMD_INTRINSIC constexpr
228 _Tp
229 operator()(_Tp __a, _Tp __b) const
230 {
231 using std::max;
232 return max(__a, __b);
233 }
234 };
235} // namespace __detail
236
237// unrolled/pack execution helpers
238// __execute_n_times{{{
239template <typename _Fp, size_t... _I>
240 _GLIBCXX_SIMD_INTRINSIC constexpr void
241 __execute_on_index_sequence(_Fp&& __f, index_sequence<_I...>)
242 { ((void)__f(_SizeConstant<_I>()), ...); }
243
244template <typename _Fp>
245 _GLIBCXX_SIMD_INTRINSIC constexpr void
246 __execute_on_index_sequence(_Fp&&, index_sequence<>)
247 { }
248
249template <size_t _Np, typename _Fp>
250 _GLIBCXX_SIMD_INTRINSIC constexpr void
251 __execute_n_times(_Fp&& __f)
252 {
253 __execute_on_index_sequence(static_cast<_Fp&&>(__f),
254 make_index_sequence<_Np>{});
255 }
256
257// }}}
258// __generate_from_n_evaluations{{{
259template <typename _R, typename _Fp, size_t... _I>
260 _GLIBCXX_SIMD_INTRINSIC constexpr _R
261 __execute_on_index_sequence_with_return(_Fp&& __f, index_sequence<_I...>)
262 { return _R{__f(_SizeConstant<_I>())...}; }
263
264template <size_t _Np, typename _R, typename _Fp>
265 _GLIBCXX_SIMD_INTRINSIC constexpr _R
266 __generate_from_n_evaluations(_Fp&& __f)
267 {
268 return __execute_on_index_sequence_with_return<_R>(
269 static_cast<_Fp&&>(__f), make_index_sequence<_Np>{});
270 }
271
272// }}}
273// __call_with_n_evaluations{{{
274template <size_t... _I, typename _F0, typename _FArgs>
275 _GLIBCXX_SIMD_INTRINSIC constexpr auto
276 __call_with_n_evaluations(index_sequence<_I...>, _F0&& __f0, _FArgs&& __fargs)
277 { return __f0(__fargs(_SizeConstant<_I>())...); }
278
279template <size_t _Np, typename _F0, typename _FArgs>
280 _GLIBCXX_SIMD_INTRINSIC constexpr auto
281 __call_with_n_evaluations(_F0&& __f0, _FArgs&& __fargs)
282 {
283 return __call_with_n_evaluations(make_index_sequence<_Np>{},
284 static_cast<_F0&&>(__f0),
285 static_cast<_FArgs&&>(__fargs));
286 }
287
288// }}}
289// __call_with_subscripts{{{
290template <size_t _First = 0, size_t... _It, typename _Tp, typename _Fp>
291 _GLIBCXX_SIMD_INTRINSIC constexpr auto
292 __call_with_subscripts(_Tp&& __x, index_sequence<_It...>, _Fp&& __fun)
293 { return __fun(__x[_First + _It]...); }
294
295template <size_t _Np, size_t _First = 0, typename _Tp, typename _Fp>
296 _GLIBCXX_SIMD_INTRINSIC constexpr auto
297 __call_with_subscripts(_Tp&& __x, _Fp&& __fun)
298 {
299 return __call_with_subscripts<_First>(static_cast<_Tp&&>(__x),
300 make_index_sequence<_Np>(),
301 static_cast<_Fp&&>(__fun));
302 }
303
304// }}}
305
306// vvv ---- type traits ---- vvv
307// integer type aliases{{{
308using _UChar = unsigned char;
309using _SChar = signed char;
310using _UShort = unsigned short;
311using _UInt = unsigned int;
312using _ULong = unsigned long;
313using _ULLong = unsigned long long;
314using _LLong = long long;
315
316//}}}
317// __first_of_pack{{{
318template <typename _T0, typename...>
319 struct __first_of_pack
320 { using type = _T0; };
321
322template <typename... _Ts>
323 using __first_of_pack_t = typename __first_of_pack<_Ts...>::type;
324
325//}}}
326// __value_type_or_identity_t {{{
327template <typename _Tp>
328 typename _Tp::value_type
329 __value_type_or_identity_impl(int);
330
331template <typename _Tp>
332 _Tp
333 __value_type_or_identity_impl(float);
334
335template <typename _Tp>
336 using __value_type_or_identity_t
337 = decltype(__value_type_or_identity_impl<_Tp>(int()));
338
339// }}}
340// __is_vectorizable {{{
341template <typename _Tp>
342 struct __is_vectorizable : public is_arithmetic<_Tp> {};
343
344template <>
345 struct __is_vectorizable<bool> : public false_type {};
346
347template <typename _Tp>
348 inline constexpr bool __is_vectorizable_v = __is_vectorizable<_Tp>::value;
349
350// Deduces to a vectorizable type
351template <typename _Tp, typename = enable_if_t<__is_vectorizable_v<_Tp>>>
352 using _Vectorizable = _Tp;
353
354// }}}
355// _LoadStorePtr / __is_possible_loadstore_conversion {{{
356template <typename _Ptr, typename _ValueType>
357 struct __is_possible_loadstore_conversion
358 : conjunction<__is_vectorizable<_Ptr>, __is_vectorizable<_ValueType>> {};
359
360template <>
361 struct __is_possible_loadstore_conversion<bool, bool> : true_type {};
362
363// Deduces to a type allowed for load/store with the given value type.
364template <typename _Ptr, typename _ValueType,
365 typename = enable_if_t<
366 __is_possible_loadstore_conversion<_Ptr, _ValueType>::value>>
367 using _LoadStorePtr = _Ptr;
368
369// }}}
370// __is_bitmask{{{
371template <typename _Tp, typename = void_t<>>
372 struct __is_bitmask : false_type {};
373
374template <typename _Tp>
375 inline constexpr bool __is_bitmask_v = __is_bitmask<_Tp>::value;
376
377// the __mmaskXX case:
378template <typename _Tp>
379 struct __is_bitmask<_Tp,
380 void_t<decltype(declval<unsigned&>() = declval<_Tp>() & 1u)>>
381 : true_type {};
382
383// }}}
384// __int_for_sizeof{{{
385#pragma GCC diagnostic push
386#pragma GCC diagnostic ignored "-Wpedantic"
387template <size_t _Bytes>
388 constexpr auto
389 __int_for_sizeof()
390 {
391 if constexpr (_Bytes == sizeof(int))
392 return int();
393 #ifdef __clang__
394 else if constexpr (_Bytes == sizeof(char))
395 return char();
396 #else
397 else if constexpr (_Bytes == sizeof(_SChar))
398 return _SChar();
399 #endif
400 else if constexpr (_Bytes == sizeof(short))
401 return short();
402 #ifndef __clang__
403 else if constexpr (_Bytes == sizeof(long))
404 return long();
405 #endif
406 else if constexpr (_Bytes == sizeof(_LLong))
407 return _LLong();
408 #ifdef __SIZEOF_INT128__
409 else if constexpr (_Bytes == sizeof(__int128))
410 return __int128();
411 #endif // __SIZEOF_INT128__
412 else if constexpr (_Bytes % sizeof(int) == 0)
413 {
414 constexpr size_t _Np = _Bytes / sizeof(int);
415 struct _Ip
416 {
417 int _M_data[_Np];
418
419 _GLIBCXX_SIMD_INTRINSIC constexpr _Ip
420 operator&(_Ip __rhs) const
421 {
422 return __generate_from_n_evaluations<_Np, _Ip>(
423 [&](auto __i) { return __rhs._M_data[__i] & _M_data[__i]; });
424 }
425
426 _GLIBCXX_SIMD_INTRINSIC constexpr _Ip
427 operator|(_Ip __rhs) const
428 {
429 return __generate_from_n_evaluations<_Np, _Ip>(
430 [&](auto __i) { return __rhs._M_data[__i] | _M_data[__i]; });
431 }
432
433 _GLIBCXX_SIMD_INTRINSIC constexpr _Ip
434 operator^(_Ip __rhs) const
435 {
436 return __generate_from_n_evaluations<_Np, _Ip>(
437 [&](auto __i) { return __rhs._M_data[__i] ^ _M_data[__i]; });
438 }
439
440 _GLIBCXX_SIMD_INTRINSIC constexpr _Ip
441 operator~() const
442 {
443 return __generate_from_n_evaluations<_Np, _Ip>(
444 [&](auto __i) { return ~_M_data[__i]; });
445 }
446 };
447 return _Ip{};
448 }
449 else
450 static_assert(_Bytes != _Bytes, "this should be unreachable");
451 }
452#pragma GCC diagnostic pop
453
454template <typename _Tp>
455 using __int_for_sizeof_t = decltype(__int_for_sizeof<sizeof(_Tp)>());
456
457template <size_t _Np>
458 using __int_with_sizeof_t = decltype(__int_for_sizeof<_Np>());
459
460// }}}
461// __is_fixed_size_abi{{{
462template <typename _Tp>
463 struct __is_fixed_size_abi : false_type {};
464
465template <int _Np>
466 struct __is_fixed_size_abi<simd_abi::fixed_size<_Np>> : true_type {};
467
468template <typename _Tp>
469 inline constexpr bool __is_fixed_size_abi_v = __is_fixed_size_abi<_Tp>::value;
470
471// }}}
472// constexpr feature detection{{{
473constexpr inline bool __have_mmx = _GLIBCXX_SIMD_HAVE_MMX;
474constexpr inline bool __have_sse = _GLIBCXX_SIMD_HAVE_SSE;
475constexpr inline bool __have_sse2 = _GLIBCXX_SIMD_HAVE_SSE2;
476constexpr inline bool __have_sse3 = _GLIBCXX_SIMD_HAVE_SSE3;
477constexpr inline bool __have_ssse3 = _GLIBCXX_SIMD_HAVE_SSSE3;
478constexpr inline bool __have_sse4_1 = _GLIBCXX_SIMD_HAVE_SSE4_1;
479constexpr inline bool __have_sse4_2 = _GLIBCXX_SIMD_HAVE_SSE4_2;
480constexpr inline bool __have_xop = _GLIBCXX_SIMD_HAVE_XOP;
481constexpr inline bool __have_avx = _GLIBCXX_SIMD_HAVE_AVX;
482constexpr inline bool __have_avx2 = _GLIBCXX_SIMD_HAVE_AVX2;
483constexpr inline bool __have_bmi = _GLIBCXX_SIMD_HAVE_BMI1;
484constexpr inline bool __have_bmi2 = _GLIBCXX_SIMD_HAVE_BMI2;
485constexpr inline bool __have_lzcnt = _GLIBCXX_SIMD_HAVE_LZCNT;
486constexpr inline bool __have_sse4a = _GLIBCXX_SIMD_HAVE_SSE4A;
487constexpr inline bool __have_fma = _GLIBCXX_SIMD_HAVE_FMA;
488constexpr inline bool __have_fma4 = _GLIBCXX_SIMD_HAVE_FMA4;
489constexpr inline bool __have_f16c = _GLIBCXX_SIMD_HAVE_F16C;
490constexpr inline bool __have_popcnt = _GLIBCXX_SIMD_HAVE_POPCNT;
491constexpr inline bool __have_avx512f = _GLIBCXX_SIMD_HAVE_AVX512F;
492constexpr inline bool __have_avx512dq = _GLIBCXX_SIMD_HAVE_AVX512DQ;
493constexpr inline bool __have_avx512vl = _GLIBCXX_SIMD_HAVE_AVX512VL;
494constexpr inline bool __have_avx512bw = _GLIBCXX_SIMD_HAVE_AVX512BW;
495constexpr inline bool __have_avx512dq_vl = __have_avx512dq && __have_avx512vl;
496constexpr inline bool __have_avx512bw_vl = __have_avx512bw && __have_avx512vl;
497
498constexpr inline bool __have_neon = _GLIBCXX_SIMD_HAVE_NEON;
499constexpr inline bool __have_neon_a32 = _GLIBCXX_SIMD_HAVE_NEON_A32;
500constexpr inline bool __have_neon_a64 = _GLIBCXX_SIMD_HAVE_NEON_A64;
501constexpr inline bool __support_neon_float =
502#if defined __GCC_IEC_559
503 __GCC_IEC_559 == 0;
504#elif defined __FAST_MATH__
505 true;
506#else
507 false;
508#endif
509
510#ifdef _ARCH_PWR10
511constexpr inline bool __have_power10vec = true;
512#else
513constexpr inline bool __have_power10vec = false;
514#endif
515#ifdef __POWER9_VECTOR__
516constexpr inline bool __have_power9vec = true;
517#else
518constexpr inline bool __have_power9vec = false;
519#endif
520#if defined __POWER8_VECTOR__
521constexpr inline bool __have_power8vec = true;
522#else
523constexpr inline bool __have_power8vec = __have_power9vec;
524#endif
525#if defined __VSX__
526constexpr inline bool __have_power_vsx = true;
527#else
528constexpr inline bool __have_power_vsx = __have_power8vec;
529#endif
530#if defined __ALTIVEC__
531constexpr inline bool __have_power_vmx = true;
532#else
533constexpr inline bool __have_power_vmx = __have_power_vsx;
534#endif
535
536// }}}
537// __is_scalar_abi {{{
538template <typename _Abi>
539 constexpr bool
540 __is_scalar_abi()
541 { return is_same_v<simd_abi::scalar, _Abi>; }
542
543// }}}
544// __abi_bytes_v {{{
545template <template <int> class _Abi, int _Bytes>
546 constexpr int
547 __abi_bytes_impl(_Abi<_Bytes>*)
548 { return _Bytes; }
549
550template <typename _Tp>
551 constexpr int
552 __abi_bytes_impl(_Tp*)
553 { return -1; }
554
555template <typename _Abi>
556 inline constexpr int __abi_bytes_v
557 = __abi_bytes_impl(static_cast<_Abi*>(nullptr));
558
559// }}}
560// __is_builtin_bitmask_abi {{{
561template <typename _Abi>
562 constexpr bool
563 __is_builtin_bitmask_abi()
564 { return is_same_v<simd_abi::_VecBltnBtmsk<__abi_bytes_v<_Abi>>, _Abi>; }
565
566// }}}
567// __is_sse_abi {{{
568template <typename _Abi>
569 constexpr bool
570 __is_sse_abi()
571 {
572 constexpr auto _Bytes = __abi_bytes_v<_Abi>;
573 return _Bytes <= 16 && is_same_v<simd_abi::_VecBuiltin<_Bytes>, _Abi>;
574 }
575
576// }}}
577// __is_avx_abi {{{
578template <typename _Abi>
579 constexpr bool
580 __is_avx_abi()
581 {
582 constexpr auto _Bytes = __abi_bytes_v<_Abi>;
583 return _Bytes > 16 && _Bytes <= 32
584 && is_same_v<simd_abi::_VecBuiltin<_Bytes>, _Abi>;
585 }
586
587// }}}
588// __is_avx512_abi {{{
589template <typename _Abi>
590 constexpr bool
591 __is_avx512_abi()
592 {
593 constexpr auto _Bytes = __abi_bytes_v<_Abi>;
594 return _Bytes <= 64 && is_same_v<simd_abi::_Avx512<_Bytes>, _Abi>;
595 }
596
597// }}}
598// __is_neon_abi {{{
599template <typename _Abi>
600 constexpr bool
601 __is_neon_abi()
602 {
603 constexpr auto _Bytes = __abi_bytes_v<_Abi>;
604 return _Bytes <= 16 && is_same_v<simd_abi::_VecBuiltin<_Bytes>, _Abi>;
605 }
606
607// }}}
608// __make_dependent_t {{{
609template <typename, typename _Up>
610 struct __make_dependent
611 { using type = _Up; };
612
613template <typename _Tp, typename _Up>
614 using __make_dependent_t = typename __make_dependent<_Tp, _Up>::type;
615
616// }}}
617// ^^^ ---- type traits ---- ^^^
618
619// __invoke_ub{{{
620template <typename... _Args>
621 [[noreturn]] _GLIBCXX_SIMD_ALWAYS_INLINE void
622 __invoke_ub([[maybe_unused]] const char* __msg,
623 [[maybe_unused]] const _Args&... __args)
624 {
625#ifdef _GLIBCXX_DEBUG_UB
626 __builtin_fprintf(stderr, __msg, __args...);
627 __builtin_trap();
628#else
629 __builtin_unreachable();
630#endif
631 }
632
633// }}}
634// __assert_unreachable{{{
635template <typename _Tp>
636 struct __assert_unreachable
637 { static_assert(!is_same_v<_Tp, _Tp>, "this should be unreachable"); };
638
639// }}}
640// __size_or_zero_v {{{
641template <typename _Tp, typename _Ap, size_t _Np = simd_size<_Tp, _Ap>::value>
642 constexpr size_t
643 __size_or_zero_dispatch(int)
644 { return _Np; }
645
646template <typename _Tp, typename _Ap>
647 constexpr size_t
648 __size_or_zero_dispatch(float)
649 { return 0; }
650
651template <typename _Tp, typename _Ap>
652 inline constexpr size_t __size_or_zero_v
653 = __size_or_zero_dispatch<_Tp, _Ap>(0);
654
655// }}}
656// __div_roundup {{{
657inline constexpr size_t
658__div_roundup(size_t __a, size_t __b)
659{ return (__a + __b - 1) / __b; }
660
661// }}}
662// _ExactBool{{{
663class _ExactBool
664{
665 const bool _M_data;
666
667public:
668 _GLIBCXX_SIMD_INTRINSIC constexpr _ExactBool(bool __b) : _M_data(__b) {}
669
670 _ExactBool(int) = delete;
671
672 _GLIBCXX_SIMD_INTRINSIC constexpr operator bool() const { return _M_data; }
673};
674
675// }}}
676// __may_alias{{{
677/**@internal
678 * Helper __may_alias<_Tp> that turns _Tp into the type to be used for an
679 * aliasing pointer. This adds the __may_alias attribute to _Tp (with compilers
680 * that support it).
681 */
682template <typename _Tp>
683 using __may_alias [[__gnu__::__may_alias__]] = _Tp;
684
685// }}}
686// _UnsupportedBase {{{
687// simd and simd_mask base for unsupported <_Tp, _Abi>
688struct _UnsupportedBase
689{
690 _UnsupportedBase() = delete;
691 _UnsupportedBase(const _UnsupportedBase&) = delete;
692 _UnsupportedBase& operator=(const _UnsupportedBase&) = delete;
693 ~_UnsupportedBase() = delete;
694};
695
696// }}}
697// _InvalidTraits {{{
698/**
699 * @internal
700 * Defines the implementation of __a given <_Tp, _Abi>.
701 *
702 * Implementations must ensure that only valid <_Tp, _Abi> instantiations are
703 * possible. Static assertions in the type definition do not suffice. It is
704 * important that SFINAE works.
705 */
706struct _InvalidTraits
707{
708 using _IsValid = false_type;
709 using _SimdBase = _UnsupportedBase;
710 using _MaskBase = _UnsupportedBase;
711
712 static constexpr size_t _S_full_size = 0;
713 static constexpr bool _S_is_partial = false;
714
715 static constexpr size_t _S_simd_align = 1;
716 struct _SimdImpl;
717 struct _SimdMember {};
718 struct _SimdCastType;
719
720 static constexpr size_t _S_mask_align = 1;
721 struct _MaskImpl;
722 struct _MaskMember {};
723 struct _MaskCastType;
724};
725
726// }}}
727// _SimdTraits {{{
728template <typename _Tp, typename _Abi, typename = void_t<>>
729 struct _SimdTraits : _InvalidTraits {};
730
731// }}}
732// __private_init, __bitset_init{{{
733/**
734 * @internal
735 * Tag used for private init constructor of simd and simd_mask
736 */
737inline constexpr struct _PrivateInit {} __private_init = {};
738
739inline constexpr struct _BitsetInit {} __bitset_init = {};
740
741// }}}
742// __is_narrowing_conversion<_From, _To>{{{
743template <typename _From, typename _To, bool = is_arithmetic_v<_From>,
744 bool = is_arithmetic_v<_To>>
745 struct __is_narrowing_conversion;
746
747// ignore "signed/unsigned mismatch" in the following trait.
748// The implicit conversions will do the right thing here.
749template <typename _From, typename _To>
750 struct __is_narrowing_conversion<_From, _To, true, true>
751 : public __bool_constant<(
752 __digits_v<_From> > __digits_v<_To>
753 || __finite_max_v<_From> > __finite_max_v<_To>
754 || __finite_min_v<_From> < __finite_min_v<_To>
755 || (is_signed_v<_From> && is_unsigned_v<_To>))> {};
756
757template <typename _Tp>
758 struct __is_narrowing_conversion<_Tp, bool, true, true>
759 : public true_type {};
760
761template <>
762 struct __is_narrowing_conversion<bool, bool, true, true>
763 : public false_type {};
764
765template <typename _Tp>
766 struct __is_narrowing_conversion<_Tp, _Tp, true, true>
767 : public false_type {};
768
769template <typename _From, typename _To>
770 struct __is_narrowing_conversion<_From, _To, false, true>
771 : public negation<is_convertible<_From, _To>> {};
772
773// }}}
774// __converts_to_higher_integer_rank{{{
775template <typename _From, typename _To, bool = (sizeof(_From) < sizeof(_To))>
776 struct __converts_to_higher_integer_rank : public true_type {};
777
778// this may fail for char -> short if sizeof(char) == sizeof(short)
779template <typename _From, typename _To>
780 struct __converts_to_higher_integer_rank<_From, _To, false>
781 : public is_same<decltype(declval<_From>() + declval<_To>()), _To> {};
782
783// }}}
784// __data(simd/simd_mask) {{{
785template <typename _Tp, typename _Ap>
786 _GLIBCXX_SIMD_INTRINSIC constexpr const auto&
787 __data(const simd<_Tp, _Ap>& __x);
788
789template <typename _Tp, typename _Ap>
790 _GLIBCXX_SIMD_INTRINSIC constexpr auto&
791 __data(simd<_Tp, _Ap>& __x);
792
793template <typename _Tp, typename _Ap>
794 _GLIBCXX_SIMD_INTRINSIC constexpr const auto&
795 __data(const simd_mask<_Tp, _Ap>& __x);
796
797template <typename _Tp, typename _Ap>
798 _GLIBCXX_SIMD_INTRINSIC constexpr auto&
799 __data(simd_mask<_Tp, _Ap>& __x);
800
801// }}}
802// _SimdConverter {{{
803template <typename _FromT, typename _FromA, typename _ToT, typename _ToA,
804 typename = void>
805 struct _SimdConverter;
806
807template <typename _Tp, typename _Ap>
808 struct _SimdConverter<_Tp, _Ap, _Tp, _Ap, void>
809 {
810 template <typename _Up>
811 _GLIBCXX_SIMD_INTRINSIC const _Up&
812 operator()(const _Up& __x)
813 { return __x; }
814 };
815
816// }}}
817// __to_value_type_or_member_type {{{
818template <typename _V>
819 _GLIBCXX_SIMD_INTRINSIC constexpr auto
820 __to_value_type_or_member_type(const _V& __x) -> decltype(__data(__x))
821 { return __data(__x); }
822
823template <typename _V>
824 _GLIBCXX_SIMD_INTRINSIC constexpr const typename _V::value_type&
825 __to_value_type_or_member_type(const typename _V::value_type& __x)
826 { return __x; }
827
828// }}}
829// __bool_storage_member_type{{{
830template <size_t _Size>
831 struct __bool_storage_member_type;
832
833template <size_t _Size>
834 using __bool_storage_member_type_t =
835 typename __bool_storage_member_type<_Size>::type;
836
837// }}}
838// _SimdTuple {{{
839// why not tuple?
840// 1. tuple gives no guarantee about the storage order, but I require
841// storage
842// equivalent to array<_Tp, _Np>
843// 2. direct access to the element type (first template argument)
844// 3. enforces equal element type, only different _Abi types are allowed
845template <typename _Tp, typename... _Abis>
846 struct _SimdTuple;
847
848//}}}
849// __fixed_size_storage_t {{{
850template <typename _Tp, int _Np>
851 struct __fixed_size_storage;
852
853template <typename _Tp, int _Np>
854 using __fixed_size_storage_t = typename __fixed_size_storage<_Tp, _Np>::type;
855
856// }}}
857// _SimdWrapper fwd decl{{{
858template <typename _Tp, size_t _Size, typename = void_t<>>
859 struct _SimdWrapper;
860
861template <typename _Tp>
862 using _SimdWrapper8 = _SimdWrapper<_Tp, 8 / sizeof(_Tp)>;
863template <typename _Tp>
864 using _SimdWrapper16 = _SimdWrapper<_Tp, 16 / sizeof(_Tp)>;
865template <typename _Tp>
866 using _SimdWrapper32 = _SimdWrapper<_Tp, 32 / sizeof(_Tp)>;
867template <typename _Tp>
868 using _SimdWrapper64 = _SimdWrapper<_Tp, 64 / sizeof(_Tp)>;
869
870// }}}
871// __is_simd_wrapper {{{
872template <typename _Tp>
873 struct __is_simd_wrapper : false_type {};
874
875template <typename _Tp, size_t _Np>
876 struct __is_simd_wrapper<_SimdWrapper<_Tp, _Np>> : true_type {};
877
878template <typename _Tp>
879 inline constexpr bool __is_simd_wrapper_v = __is_simd_wrapper<_Tp>::value;
880
881// }}}
882// _BitOps {{{
883struct _BitOps
884{
885 // _S_bit_iteration {{{
886 template <typename _Tp, typename _Fp>
887 static void
888 _S_bit_iteration(_Tp __mask, _Fp&& __f)
889 {
890 static_assert(sizeof(_ULLong) >= sizeof(_Tp));
891 conditional_t<sizeof(_Tp) <= sizeof(_UInt), _UInt, _ULLong> __k;
892 if constexpr (is_convertible_v<_Tp, decltype(__k)>)
893 __k = __mask;
894 else
895 __k = __mask.to_ullong();
896 while(__k)
897 {
898 __f(std::__countr_zero(__k));
899 __k &= (__k - 1);
900 }
901 }
902
903 //}}}
904};
905
906//}}}
907// __increment, __decrement {{{
908template <typename _Tp = void>
909 struct __increment
910 { constexpr _Tp operator()(_Tp __a) const { return ++__a; } };
911
912template <>
913 struct __increment<void>
914 {
915 template <typename _Tp>
916 constexpr _Tp
917 operator()(_Tp __a) const
918 { return ++__a; }
919 };
920
921template <typename _Tp = void>
922 struct __decrement
923 { constexpr _Tp operator()(_Tp __a) const { return --__a; } };
924
925template <>
926 struct __decrement<void>
927 {
928 template <typename _Tp>
929 constexpr _Tp
930 operator()(_Tp __a) const
931 { return --__a; }
932 };
933
934// }}}
935// _ValuePreserving(OrInt) {{{
936template <typename _From, typename _To,
937 typename = enable_if_t<negation<
938 __is_narrowing_conversion<__remove_cvref_t<_From>, _To>>::value>>
939 using _ValuePreserving = _From;
940
941template <typename _From, typename _To,
942 typename _DecayedFrom = __remove_cvref_t<_From>,
943 typename = enable_if_t<conjunction<
944 is_convertible<_From, _To>,
945 disjunction<
946 is_same<_DecayedFrom, _To>, is_same<_DecayedFrom, int>,
947 conjunction<is_same<_DecayedFrom, _UInt>, is_unsigned<_To>>,
948 negation<__is_narrowing_conversion<_DecayedFrom, _To>>>>::value>>
949 using _ValuePreservingOrInt = _From;
950
951// }}}
952// __intrinsic_type {{{
953template <typename _Tp, size_t _Bytes, typename = void_t<>>
954 struct __intrinsic_type;
955
956template <typename _Tp, size_t _Size>
957 using __intrinsic_type_t =
958 typename __intrinsic_type<_Tp, _Size * sizeof(_Tp)>::type;
959
960template <typename _Tp>
961 using __intrinsic_type2_t = typename __intrinsic_type<_Tp, 2>::type;
962template <typename _Tp>
963 using __intrinsic_type4_t = typename __intrinsic_type<_Tp, 4>::type;
964template <typename _Tp>
965 using __intrinsic_type8_t = typename __intrinsic_type<_Tp, 8>::type;
966template <typename _Tp>
967 using __intrinsic_type16_t = typename __intrinsic_type<_Tp, 16>::type;
968template <typename _Tp>
969 using __intrinsic_type32_t = typename __intrinsic_type<_Tp, 32>::type;
970template <typename _Tp>
971 using __intrinsic_type64_t = typename __intrinsic_type<_Tp, 64>::type;
972
973// }}}
974// _BitMask {{{
975template <size_t _Np, bool _Sanitized = false>
976 struct _BitMask;
977
978template <size_t _Np, bool _Sanitized>
979 struct __is_bitmask<_BitMask<_Np, _Sanitized>, void> : true_type {};
980
981template <size_t _Np>
982 using _SanitizedBitMask = _BitMask<_Np, true>;
983
984template <size_t _Np, bool _Sanitized>
985 struct _BitMask
986 {
987 static_assert(_Np > 0);
988
989 static constexpr size_t _NBytes = __div_roundup(_Np, __CHAR_BIT__);
990
991 using _Tp = conditional_t<_Np == 1, bool,
992 make_unsigned_t<__int_with_sizeof_t<std::min(
993 sizeof(_ULLong), std::__bit_ceil(_NBytes))>>>;
994
995 static constexpr int _S_array_size = __div_roundup(_NBytes, sizeof(_Tp));
996
997 _Tp _M_bits[_S_array_size];
998
999 static constexpr int _S_unused_bits
1000 = _Np == 1 ? 0 : _S_array_size * sizeof(_Tp) * __CHAR_BIT__ - _Np;
1001
1002 static constexpr _Tp _S_bitmask = +_Tp(~_Tp()) >> _S_unused_bits;
1003
1004 constexpr _BitMask() noexcept = default;
1005
1006 constexpr _BitMask(unsigned long long __x) noexcept
1007 : _M_bits{static_cast<_Tp>(__x)} {}
1008
1009 _BitMask(bitset<_Np> __x) noexcept : _BitMask(__x.to_ullong()) {}
1010
1011 constexpr _BitMask(const _BitMask&) noexcept = default;
1012
1013 template <bool _RhsSanitized, typename = enable_if_t<_RhsSanitized == false
1014 && _Sanitized == true>>
1015 constexpr _BitMask(const _BitMask<_Np, _RhsSanitized>& __rhs) noexcept
1016 : _BitMask(__rhs._M_sanitized()) {}
1017
1018 constexpr operator _SimdWrapper<bool, _Np>() const noexcept
1019 {
1020 static_assert(_S_array_size == 1);
1021 return _M_bits[0];
1022 }
1023
1024 // precondition: is sanitized
1025 constexpr _Tp
1026 _M_to_bits() const noexcept
1027 {
1028 static_assert(_S_array_size == 1);
1029 return _M_bits[0];
1030 }
1031
1032 // precondition: is sanitized
1033 constexpr unsigned long long
1034 to_ullong() const noexcept
1035 {
1036 static_assert(_S_array_size == 1);
1037 return _M_bits[0];
1038 }
1039
1040 // precondition: is sanitized
1041 constexpr unsigned long
1042 to_ulong() const noexcept
1043 {
1044 static_assert(_S_array_size == 1);
1045 return _M_bits[0];
1046 }
1047
1048 constexpr bitset<_Np>
1049 _M_to_bitset() const noexcept
1050 {
1051 static_assert(_S_array_size == 1);
1052 return _M_bits[0];
1053 }
1054
1055 constexpr decltype(auto)
1056 _M_sanitized() const noexcept
1057 {
1058 if constexpr (_Sanitized)
1059 return *this;
1060 else if constexpr (_Np == 1)
1061 return _SanitizedBitMask<_Np>(_M_bits[0]);
1062 else
1063 {
1064 _SanitizedBitMask<_Np> __r = {};
1065 for (int __i = 0; __i < _S_array_size; ++__i)
1066 __r._M_bits[__i] = _M_bits[__i];
1067 if constexpr (_S_unused_bits > 0)
1068 __r._M_bits[_S_array_size - 1] &= _S_bitmask;
1069 return __r;
1070 }
1071 }
1072
1073 template <size_t _Mp, bool _LSanitized>
1074 constexpr _BitMask<_Np + _Mp, _Sanitized>
1075 _M_prepend(_BitMask<_Mp, _LSanitized> __lsb) const noexcept
1076 {
1077 constexpr size_t _RN = _Np + _Mp;
1078 using _Rp = _BitMask<_RN, _Sanitized>;
1079 if constexpr (_Rp::_S_array_size == 1)
1080 {
1081 _Rp __r{{_M_bits[0]}};
1082 __r._M_bits[0] <<= _Mp;
1083 __r._M_bits[0] |= __lsb._M_sanitized()._M_bits[0];
1084 return __r;
1085 }
1086 else
1087 __assert_unreachable<_Rp>();
1088 }
1089
1090 // Return a new _BitMask with size _NewSize while dropping _DropLsb least
1091 // significant bits. If the operation implicitly produces a sanitized bitmask,
1092 // the result type will have _Sanitized set.
1093 template <size_t _DropLsb, size_t _NewSize = _Np - _DropLsb>
1094 constexpr auto
1095 _M_extract() const noexcept
1096 {
1097 static_assert(_Np > _DropLsb);
1098 static_assert(_DropLsb + _NewSize <= sizeof(_ULLong) * __CHAR_BIT__,
1099 "not implemented for bitmasks larger than one ullong");
1100 if constexpr (_NewSize == 1)
1101 // must sanitize because the return _Tp is bool
1102 return _SanitizedBitMask<1>(_M_bits[0] & (_Tp(1) << _DropLsb));
1103 else
1104 return _BitMask<_NewSize,
1105 ((_NewSize + _DropLsb == sizeof(_Tp) * __CHAR_BIT__
1106 && _NewSize + _DropLsb <= _Np)
1107 || ((_Sanitized || _Np == sizeof(_Tp) * __CHAR_BIT__)
1108 && _NewSize + _DropLsb >= _Np))>(_M_bits[0]
1109 >> _DropLsb);
1110 }
1111
1112 // True if all bits are set. Implicitly sanitizes if _Sanitized == false.
1113 constexpr bool
1114 all() const noexcept
1115 {
1116 if constexpr (_Np == 1)
1117 return _M_bits[0];
1118 else if constexpr (!_Sanitized)
1119 return _M_sanitized().all();
1120 else
1121 {
1122 constexpr _Tp __allbits = ~_Tp();
1123 for (int __i = 0; __i < _S_array_size - 1; ++__i)
1124 if (_M_bits[__i] != __allbits)
1125 return false;
1126 return _M_bits[_S_array_size - 1] == _S_bitmask;
1127 }
1128 }
1129
1130 // True if at least one bit is set. Implicitly sanitizes if _Sanitized ==
1131 // false.
1132 constexpr bool
1133 any() const noexcept
1134 {
1135 if constexpr (_Np == 1)
1136 return _M_bits[0];
1137 else if constexpr (!_Sanitized)
1138 return _M_sanitized().any();
1139 else
1140 {
1141 for (int __i = 0; __i < _S_array_size - 1; ++__i)
1142 if (_M_bits[__i] != 0)
1143 return true;
1144 return _M_bits[_S_array_size - 1] != 0;
1145 }
1146 }
1147
1148 // True if no bit is set. Implicitly sanitizes if _Sanitized == false.
1149 constexpr bool
1150 none() const noexcept
1151 {
1152 if constexpr (_Np == 1)
1153 return !_M_bits[0];
1154 else if constexpr (!_Sanitized)
1155 return _M_sanitized().none();
1156 else
1157 {
1158 for (int __i = 0; __i < _S_array_size - 1; ++__i)
1159 if (_M_bits[__i] != 0)
1160 return false;
1161 return _M_bits[_S_array_size - 1] == 0;
1162 }
1163 }
1164
1165 // Returns the number of set bits. Implicitly sanitizes if _Sanitized ==
1166 // false.
1167 constexpr int
1168 count() const noexcept
1169 {
1170 if constexpr (_Np == 1)
1171 return _M_bits[0];
1172 else if constexpr (!_Sanitized)
1173 return _M_sanitized().none();
1174 else
1175 {
1176 int __result = __builtin_popcountll(_M_bits[0]);
1177 for (int __i = 1; __i < _S_array_size; ++__i)
1178 __result += __builtin_popcountll(_M_bits[__i]);
1179 return __result;
1180 }
1181 }
1182
1183 // Returns the bit at offset __i as bool.
1184 constexpr bool
1185 operator[](size_t __i) const noexcept
1186 {
1187 if constexpr (_Np == 1)
1188 return _M_bits[0];
1189 else if constexpr (_S_array_size == 1)
1190 return (_M_bits[0] >> __i) & 1;
1191 else
1192 {
1193 const size_t __j = __i / (sizeof(_Tp) * __CHAR_BIT__);
1194 const size_t __shift = __i % (sizeof(_Tp) * __CHAR_BIT__);
1195 return (_M_bits[__j] >> __shift) & 1;
1196 }
1197 }
1198
1199 template <size_t __i>
1200 constexpr bool
1201 operator[](_SizeConstant<__i>) const noexcept
1202 {
1203 static_assert(__i < _Np);
1204 constexpr size_t __j = __i / (sizeof(_Tp) * __CHAR_BIT__);
1205 constexpr size_t __shift = __i % (sizeof(_Tp) * __CHAR_BIT__);
1206 return static_cast<bool>(_M_bits[__j] & (_Tp(1) << __shift));
1207 }
1208
1209 // Set the bit at offset __i to __x.
1210 constexpr void
1211 set(size_t __i, bool __x) noexcept
1212 {
1213 if constexpr (_Np == 1)
1214 _M_bits[0] = __x;
1215 else if constexpr (_S_array_size == 1)
1216 {
1217 _M_bits[0] &= ~_Tp(_Tp(1) << __i);
1218 _M_bits[0] |= _Tp(_Tp(__x) << __i);
1219 }
1220 else
1221 {
1222 const size_t __j = __i / (sizeof(_Tp) * __CHAR_BIT__);
1223 const size_t __shift = __i % (sizeof(_Tp) * __CHAR_BIT__);
1224 _M_bits[__j] &= ~_Tp(_Tp(1) << __shift);
1225 _M_bits[__j] |= _Tp(_Tp(__x) << __shift);
1226 }
1227 }
1228
1229 template <size_t __i>
1230 constexpr void
1231 set(_SizeConstant<__i>, bool __x) noexcept
1232 {
1233 static_assert(__i < _Np);
1234 if constexpr (_Np == 1)
1235 _M_bits[0] = __x;
1236 else
1237 {
1238 constexpr size_t __j = __i / (sizeof(_Tp) * __CHAR_BIT__);
1239 constexpr size_t __shift = __i % (sizeof(_Tp) * __CHAR_BIT__);
1240 constexpr _Tp __mask = ~_Tp(_Tp(1) << __shift);
1241 _M_bits[__j] &= __mask;
1242 _M_bits[__j] |= _Tp(_Tp(__x) << __shift);
1243 }
1244 }
1245
1246 // Inverts all bits. Sanitized input leads to sanitized output.
1247 constexpr _BitMask
1248 operator~() const noexcept
1249 {
1250 if constexpr (_Np == 1)
1251 return !_M_bits[0];
1252 else
1253 {
1254 _BitMask __result{};
1255 for (int __i = 0; __i < _S_array_size - 1; ++__i)
1256 __result._M_bits[__i] = ~_M_bits[__i];
1257 if constexpr (_Sanitized)
1258 __result._M_bits[_S_array_size - 1]
1259 = _M_bits[_S_array_size - 1] ^ _S_bitmask;
1260 else
1261 __result._M_bits[_S_array_size - 1] = ~_M_bits[_S_array_size - 1];
1262 return __result;
1263 }
1264 }
1265
1266 constexpr _BitMask&
1267 operator^=(const _BitMask& __b) & noexcept
1268 {
1269 __execute_n_times<_S_array_size>(
1270 [&](auto __i) { _M_bits[__i] ^= __b._M_bits[__i]; });
1271 return *this;
1272 }
1273
1274 constexpr _BitMask&
1275 operator|=(const _BitMask& __b) & noexcept
1276 {
1277 __execute_n_times<_S_array_size>(
1278 [&](auto __i) { _M_bits[__i] |= __b._M_bits[__i]; });
1279 return *this;
1280 }
1281
1282 constexpr _BitMask&
1283 operator&=(const _BitMask& __b) & noexcept
1284 {
1285 __execute_n_times<_S_array_size>(
1286 [&](auto __i) { _M_bits[__i] &= __b._M_bits[__i]; });
1287 return *this;
1288 }
1289
1290 friend constexpr _BitMask
1291 operator^(const _BitMask& __a, const _BitMask& __b) noexcept
1292 {
1293 _BitMask __r = __a;
1294 __r ^= __b;
1295 return __r;
1296 }
1297
1298 friend constexpr _BitMask
1299 operator|(const _BitMask& __a, const _BitMask& __b) noexcept
1300 {
1301 _BitMask __r = __a;
1302 __r |= __b;
1303 return __r;
1304 }
1305
1306 friend constexpr _BitMask
1307 operator&(const _BitMask& __a, const _BitMask& __b) noexcept
1308 {
1309 _BitMask __r = __a;
1310 __r &= __b;
1311 return __r;
1312 }
1313
1314 _GLIBCXX_SIMD_INTRINSIC
1315 constexpr bool
1316 _M_is_constprop() const
1317 {
1318 if constexpr (_S_array_size == 0)
1319 return __builtin_constant_p(_M_bits[0]);
1320 else
1321 {
1322 for (int __i = 0; __i < _S_array_size; ++__i)
1323 if (!__builtin_constant_p(_M_bits[__i]))
1324 return false;
1325 return true;
1326 }
1327 }
1328 };
1329
1330// }}}
1331
1332// vvv ---- builtin vector types [[gnu::vector_size(N)]] and operations ---- vvv
1333// __min_vector_size {{{
1334template <typename _Tp = void>
1335 static inline constexpr int __min_vector_size = 2 * sizeof(_Tp);
1336
1337#if _GLIBCXX_SIMD_HAVE_NEON
1338template <>
1339 inline constexpr int __min_vector_size<void> = 8;
1340#else
1341template <>
1342 inline constexpr int __min_vector_size<void> = 16;
1343#endif
1344
1345// }}}
1346// __vector_type {{{
1347template <typename _Tp, size_t _Np, typename = void>
1348 struct __vector_type_n {};
1349
1350// substition failure for 0-element case
1351template <typename _Tp>
1352 struct __vector_type_n<_Tp, 0, void> {};
1353
1354// special case 1-element to be _Tp itself
1355template <typename _Tp>
1356 struct __vector_type_n<_Tp, 1, enable_if_t<__is_vectorizable_v<_Tp>>>
1357 { using type = _Tp; };
1358
1359// else, use GNU-style builtin vector types
1360template <typename _Tp, size_t _Np>
1361 struct __vector_type_n<_Tp, _Np,
1362 enable_if_t<__is_vectorizable_v<_Tp> && _Np >= 2>>
1363 {
1364 static constexpr size_t _S_Np2 = std::__bit_ceil(_Np * sizeof(_Tp));
1365
1366 static constexpr size_t _S_Bytes =
1367#ifdef __i386__
1368 // Using [[gnu::vector_size(8)]] would wreak havoc on the FPU because
1369 // those objects are passed via MMX registers and nothing ever calls EMMS.
1370 _S_Np2 == 8 ? 16 :
1371#endif
1372 _S_Np2 < __min_vector_size<_Tp> ? __min_vector_size<_Tp>
1373 : _S_Np2;
1374
1375 using type [[__gnu__::__vector_size__(_S_Bytes)]] = _Tp;
1376 };
1377
1378template <typename _Tp, size_t _Bytes, size_t = _Bytes % sizeof(_Tp)>
1379 struct __vector_type;
1380
1381template <typename _Tp, size_t _Bytes>
1382 struct __vector_type<_Tp, _Bytes, 0>
1383 : __vector_type_n<_Tp, _Bytes / sizeof(_Tp)> {};
1384
1385template <typename _Tp, size_t _Size>
1386 using __vector_type_t = typename __vector_type_n<_Tp, _Size>::type;
1387
1388template <typename _Tp>
1389 using __vector_type2_t = typename __vector_type<_Tp, 2>::type;
1390template <typename _Tp>
1391 using __vector_type4_t = typename __vector_type<_Tp, 4>::type;
1392template <typename _Tp>
1393 using __vector_type8_t = typename __vector_type<_Tp, 8>::type;
1394template <typename _Tp>
1395 using __vector_type16_t = typename __vector_type<_Tp, 16>::type;
1396template <typename _Tp>
1397 using __vector_type32_t = typename __vector_type<_Tp, 32>::type;
1398template <typename _Tp>
1399 using __vector_type64_t = typename __vector_type<_Tp, 64>::type;
1400
1401// }}}
1402// __is_vector_type {{{
1403template <typename _Tp, typename = void_t<>>
1404 struct __is_vector_type : false_type {};
1405
1406template <typename _Tp>
1407 struct __is_vector_type<
1408 _Tp, void_t<typename __vector_type<
1409 remove_reference_t<decltype(declval<_Tp>()[0])>, sizeof(_Tp)>::type>>
1410 : is_same<_Tp, typename __vector_type<
1411 remove_reference_t<decltype(declval<_Tp>()[0])>,
1412 sizeof(_Tp)>::type> {};
1413
1414template <typename _Tp>
1415 inline constexpr bool __is_vector_type_v = __is_vector_type<_Tp>::value;
1416
1417// }}}
1418// __is_intrinsic_type {{{
1419#if _GLIBCXX_SIMD_HAVE_SSE_ABI
1420template <typename _Tp>
1421 using __is_intrinsic_type = __is_vector_type<_Tp>;
1422#else // not SSE (x86)
1423template <typename _Tp, typename = void_t<>>
1424 struct __is_intrinsic_type : false_type {};
1425
1426template <typename _Tp>
1427 struct __is_intrinsic_type<
1428 _Tp, void_t<typename __intrinsic_type<
1429 remove_reference_t<decltype(declval<_Tp>()[0])>, sizeof(_Tp)>::type>>
1430 : is_same<_Tp, typename __intrinsic_type<
1431 remove_reference_t<decltype(declval<_Tp>()[0])>,
1432 sizeof(_Tp)>::type> {};
1433#endif
1434
1435template <typename _Tp>
1436 inline constexpr bool __is_intrinsic_type_v = __is_intrinsic_type<_Tp>::value;
1437
1438// }}}
1439// _VectorTraits{{{
1440template <typename _Tp, typename = void_t<>>
1441 struct _VectorTraitsImpl;
1442
1443template <typename _Tp>
1444 struct _VectorTraitsImpl<_Tp, enable_if_t<__is_vector_type_v<_Tp>
1445 || __is_intrinsic_type_v<_Tp>>>
1446 {
1447 using type = _Tp;
1448 using value_type = remove_reference_t<decltype(declval<_Tp>()[0])>;
1449 static constexpr int _S_full_size = sizeof(_Tp) / sizeof(value_type);
1450 using _Wrapper = _SimdWrapper<value_type, _S_full_size>;
1451 template <typename _Up, int _W = _S_full_size>
1452 static constexpr bool _S_is
1453 = is_same_v<value_type, _Up> && _W == _S_full_size;
1454 };
1455
1456template <typename _Tp, size_t _Np>
1457 struct _VectorTraitsImpl<_SimdWrapper<_Tp, _Np>,
1458 void_t<__vector_type_t<_Tp, _Np>>>
1459 {
1460 using type = __vector_type_t<_Tp, _Np>;
1461 using value_type = _Tp;
1462 static constexpr int _S_full_size = sizeof(type) / sizeof(value_type);
1463 using _Wrapper = _SimdWrapper<_Tp, _Np>;
1464 static constexpr bool _S_is_partial = (_Np == _S_full_size);
1465 static constexpr int _S_partial_width = _Np;
1466 template <typename _Up, int _W = _S_full_size>
1467 static constexpr bool _S_is
1468 = is_same_v<value_type, _Up>&& _W == _S_full_size;
1469 };
1470
1471template <typename _Tp, typename = typename _VectorTraitsImpl<_Tp>::type>
1472 using _VectorTraits = _VectorTraitsImpl<_Tp>;
1473
1474// }}}
1475// __as_vector{{{
1476template <typename _V>
1477 _GLIBCXX_SIMD_INTRINSIC constexpr auto
1478 __as_vector(_V __x)
1479 {
1480 if constexpr (__is_vector_type_v<_V>)
1481 return __x;
1482 else if constexpr (is_simd<_V>::value || is_simd_mask<_V>::value)
1483 return __data(__x)._M_data;
1484 else if constexpr (__is_vectorizable_v<_V>)
1485 return __vector_type_t<_V, 2>{__x};
1486 else
1487 return __x._M_data;
1488 }
1489
1490// }}}
1491// __as_wrapper{{{
1492template <size_t _Np = 0, typename _V>
1493 _GLIBCXX_SIMD_INTRINSIC constexpr auto
1494 __as_wrapper(_V __x)
1495 {
1496 if constexpr (__is_vector_type_v<_V>)
1497 return _SimdWrapper<typename _VectorTraits<_V>::value_type,
1498 (_Np > 0 ? _Np : _VectorTraits<_V>::_S_full_size)>(__x);
1499 else if constexpr (is_simd<_V>::value || is_simd_mask<_V>::value)
1500 {
1501 static_assert(_V::size() == _Np);
1502 return __data(__x);
1503 }
1504 else
1505 {
1506 static_assert(_V::_S_size == _Np);
1507 return __x;
1508 }
1509 }
1510
1511// }}}
1512// __intrin_bitcast{{{
1513template <typename _To, typename _From>
1514 _GLIBCXX_SIMD_INTRINSIC constexpr _To
1515 __intrin_bitcast(_From __v)
1516 {
1517 static_assert((__is_vector_type_v<_From> || __is_intrinsic_type_v<_From>)
1518 && (__is_vector_type_v<_To> || __is_intrinsic_type_v<_To>));
1519 if constexpr (sizeof(_To) == sizeof(_From))
1520 return reinterpret_cast<_To>(__v);
1521 else if constexpr (sizeof(_From) > sizeof(_To))
1522 if constexpr (sizeof(_To) >= 16)
1523 return reinterpret_cast<const __may_alias<_To>&>(__v);
1524 else
1525 {
1526 _To __r;
1527 __builtin_memcpy(&__r, &__v, sizeof(_To));
1528 return __r;
1529 }
1530#if _GLIBCXX_SIMD_X86INTRIN && !defined __clang__
1531 else if constexpr (__have_avx && sizeof(_From) == 16 && sizeof(_To) == 32)
1532 return reinterpret_cast<_To>(__builtin_ia32_ps256_ps(
1533 reinterpret_cast<__vector_type_t<float, 4>>(__v)));
1534 else if constexpr (__have_avx512f && sizeof(_From) == 16
1535 && sizeof(_To) == 64)
1536 return reinterpret_cast<_To>(__builtin_ia32_ps512_ps(
1537 reinterpret_cast<__vector_type_t<float, 4>>(__v)));
1538 else if constexpr (__have_avx512f && sizeof(_From) == 32
1539 && sizeof(_To) == 64)
1540 return reinterpret_cast<_To>(__builtin_ia32_ps512_256ps(
1541 reinterpret_cast<__vector_type_t<float, 8>>(__v)));
1542#endif // _GLIBCXX_SIMD_X86INTRIN
1543 else if constexpr (sizeof(__v) <= 8)
1544 return reinterpret_cast<_To>(
1545 __vector_type_t<__int_for_sizeof_t<_From>, sizeof(_To) / sizeof(_From)>{
1546 reinterpret_cast<__int_for_sizeof_t<_From>>(__v)});
1547 else
1548 {
1549 static_assert(sizeof(_To) > sizeof(_From));
1550 _To __r = {};
1551 __builtin_memcpy(&__r, &__v, sizeof(_From));
1552 return __r;
1553 }
1554 }
1555
1556// }}}
1557// __vector_bitcast{{{
1558template <typename _To, size_t _NN = 0, typename _From,
1559 typename _FromVT = _VectorTraits<_From>,
1560 size_t _Np = _NN == 0 ? sizeof(_From) / sizeof(_To) : _NN>
1561 _GLIBCXX_SIMD_INTRINSIC constexpr __vector_type_t<_To, _Np>
1562 __vector_bitcast(_From __x)
1563 {
1564 using _R = __vector_type_t<_To, _Np>;
1565 return __intrin_bitcast<_R>(__x);
1566 }
1567
1568template <typename _To, size_t _NN = 0, typename _Tp, size_t _Nx,
1569 size_t _Np
1570 = _NN == 0 ? sizeof(_SimdWrapper<_Tp, _Nx>) / sizeof(_To) : _NN>
1571 _GLIBCXX_SIMD_INTRINSIC constexpr __vector_type_t<_To, _Np>
1572 __vector_bitcast(const _SimdWrapper<_Tp, _Nx>& __x)
1573 {
1574 static_assert(_Np > 1);
1575 return __intrin_bitcast<__vector_type_t<_To, _Np>>(__x._M_data);
1576 }
1577
1578// }}}
1579// __convert_x86 declarations {{{
1580#ifdef _GLIBCXX_SIMD_WORKAROUND_PR85048
1581template <typename _To, typename _Tp, typename _TVT = _VectorTraits<_Tp>>
1582 _To __convert_x86(_Tp);
1583
1584template <typename _To, typename _Tp, typename _TVT = _VectorTraits<_Tp>>
1585 _To __convert_x86(_Tp, _Tp);
1586
1587template <typename _To, typename _Tp, typename _TVT = _VectorTraits<_Tp>>
1588 _To __convert_x86(_Tp, _Tp, _Tp, _Tp);
1589
1590template <typename _To, typename _Tp, typename _TVT = _VectorTraits<_Tp>>
1591 _To __convert_x86(_Tp, _Tp, _Tp, _Tp, _Tp, _Tp, _Tp, _Tp);
1592
1593template <typename _To, typename _Tp, typename _TVT = _VectorTraits<_Tp>>
1594 _To __convert_x86(_Tp, _Tp, _Tp, _Tp, _Tp, _Tp, _Tp, _Tp, _Tp, _Tp, _Tp, _Tp,
1595 _Tp, _Tp, _Tp, _Tp);
1596#endif // _GLIBCXX_SIMD_WORKAROUND_PR85048
1597
1598//}}}
1599// __bit_cast {{{
1600template <typename _To, typename _From>
1601 _GLIBCXX_SIMD_INTRINSIC constexpr _To
1602 __bit_cast(const _From __x)
1603 {
1604 // TODO: implement with / replace by __builtin_bit_cast ASAP
1605 static_assert(sizeof(_To) == sizeof(_From));
1606 constexpr bool __to_is_vectorizable
1607 = is_arithmetic_v<_To> || is_enum_v<_To>;
1608 constexpr bool __from_is_vectorizable
1609 = is_arithmetic_v<_From> || is_enum_v<_From>;
1610 if constexpr (__is_vector_type_v<_To> && __is_vector_type_v<_From>)
1611 return reinterpret_cast<_To>(__x);
1612 else if constexpr (__is_vector_type_v<_To> && __from_is_vectorizable)
1613 {
1614 using _FV [[gnu::vector_size(sizeof(_From))]] = _From;
1615 return reinterpret_cast<_To>(_FV{__x});
1616 }
1617 else if constexpr (__to_is_vectorizable && __from_is_vectorizable)
1618 {
1619 using _TV [[gnu::vector_size(sizeof(_To))]] = _To;
1620 using _FV [[gnu::vector_size(sizeof(_From))]] = _From;
1621 return reinterpret_cast<_TV>(_FV{__x})[0];
1622 }
1623 else if constexpr (__to_is_vectorizable && __is_vector_type_v<_From>)
1624 {
1625 using _TV [[gnu::vector_size(sizeof(_To))]] = _To;
1626 return reinterpret_cast<_TV>(__x)[0];
1627 }
1628 else
1629 {
1630 _To __r;
1631 __builtin_memcpy(reinterpret_cast<char*>(&__r),
1632 reinterpret_cast<const char*>(&__x), sizeof(_To));
1633 return __r;
1634 }
1635 }
1636
1637// }}}
1638// __to_intrin {{{
1639template <typename _Tp, typename _TVT = _VectorTraits<_Tp>,
1640 typename _R
1641 = __intrinsic_type_t<typename _TVT::value_type, _TVT::_S_full_size>>
1642 _GLIBCXX_SIMD_INTRINSIC constexpr _R
1643 __to_intrin(_Tp __x)
1644 {
1645 static_assert(sizeof(__x) <= sizeof(_R),
1646 "__to_intrin may never drop values off the end");
1647 if constexpr (sizeof(__x) == sizeof(_R))
1648 return reinterpret_cast<_R>(__as_vector(__x));
1649 else
1650 {
1651 using _Up = __int_for_sizeof_t<_Tp>;
1652 return reinterpret_cast<_R>(
1653 __vector_type_t<_Up, sizeof(_R) / sizeof(_Up)>{__bit_cast<_Up>(__x)});
1654 }
1655 }
1656
1657// }}}
1658// __make_vector{{{
1659template <typename _Tp, typename... _Args>
1660 _GLIBCXX_SIMD_INTRINSIC constexpr __vector_type_t<_Tp, sizeof...(_Args)>
1661 __make_vector(const _Args&... __args)
1662 {
1663 return __vector_type_t<_Tp, sizeof...(_Args)>{static_cast<_Tp>(__args)...};
1664 }
1665
1666// }}}
1667// __vector_broadcast{{{
1668template <size_t _Np, typename _Tp, size_t... _I>
1669 _GLIBCXX_SIMD_INTRINSIC constexpr __vector_type_t<_Tp, _Np>
1670 __vector_broadcast_impl(_Tp __x, index_sequence<_I...>)
1671 { return __vector_type_t<_Tp, _Np>{((void)_I, __x)...}; }
1672
1673template <size_t _Np, typename _Tp>
1674 _GLIBCXX_SIMD_INTRINSIC constexpr __vector_type_t<_Tp, _Np>
1675 __vector_broadcast(_Tp __x)
1676 { return __vector_broadcast_impl<_Np, _Tp>(__x, make_index_sequence<_Np>()); }
1677
1678// }}}
1679// __generate_vector{{{
1680 template <typename _Tp, size_t _Np, typename _Gp, size_t... _I>
1681 _GLIBCXX_SIMD_INTRINSIC constexpr __vector_type_t<_Tp, _Np>
1682 __generate_vector_impl(_Gp&& __gen, index_sequence<_I...>)
1683 {
1684 return __vector_type_t<_Tp, _Np>{
1685 static_cast<_Tp>(__gen(_SizeConstant<_I>()))...};
1686 }
1687
1688template <typename _V, typename _VVT = _VectorTraits<_V>, typename _Gp>
1689 _GLIBCXX_SIMD_INTRINSIC constexpr _V
1690 __generate_vector(_Gp&& __gen)
1691 {
1692 if constexpr (__is_vector_type_v<_V>)
1693 return __generate_vector_impl<typename _VVT::value_type,
1694 _VVT::_S_full_size>(
1695 static_cast<_Gp&&>(__gen), make_index_sequence<_VVT::_S_full_size>());
1696 else
1697 return __generate_vector_impl<typename _VVT::value_type,
1698 _VVT::_S_partial_width>(
1699 static_cast<_Gp&&>(__gen),
1700 make_index_sequence<_VVT::_S_partial_width>());
1701 }
1702
1703template <typename _Tp, size_t _Np, typename _Gp>
1704 _GLIBCXX_SIMD_INTRINSIC constexpr __vector_type_t<_Tp, _Np>
1705 __generate_vector(_Gp&& __gen)
1706 {
1707 return __generate_vector_impl<_Tp, _Np>(static_cast<_Gp&&>(__gen),
1708 make_index_sequence<_Np>());
1709 }
1710
1711// }}}
1712// __xor{{{
1713template <typename _TW>
1714 _GLIBCXX_SIMD_INTRINSIC constexpr _TW
1715 __xor(_TW __a, _TW __b) noexcept
1716 {
1717 if constexpr (__is_vector_type_v<_TW> || __is_simd_wrapper_v<_TW>)
1718 {
1719 using _Tp = typename conditional_t<__is_simd_wrapper_v<_TW>, _TW,
1720 _VectorTraitsImpl<_TW>>::value_type;
1721 if constexpr (is_floating_point_v<_Tp>)
1722 {
1723 using _Ip = make_unsigned_t<__int_for_sizeof_t<_Tp>>;
1724 return __vector_bitcast<_Tp>(__vector_bitcast<_Ip>(__a)
1725 ^ __vector_bitcast<_Ip>(__b));
1726 }
1727 else if constexpr (__is_vector_type_v<_TW>)
1728 return __a ^ __b;
1729 else
1730 return __a._M_data ^ __b._M_data;
1731 }
1732 else
1733 return __a ^ __b;
1734 }
1735
1736// }}}
1737// __or{{{
1738template <typename _TW>
1739 _GLIBCXX_SIMD_INTRINSIC constexpr _TW
1740 __or(_TW __a, _TW __b) noexcept
1741 {
1742 if constexpr (__is_vector_type_v<_TW> || __is_simd_wrapper_v<_TW>)
1743 {
1744 using _Tp = typename conditional_t<__is_simd_wrapper_v<_TW>, _TW,
1745 _VectorTraitsImpl<_TW>>::value_type;
1746 if constexpr (is_floating_point_v<_Tp>)
1747 {
1748 using _Ip = make_unsigned_t<__int_for_sizeof_t<_Tp>>;
1749 return __vector_bitcast<_Tp>(__vector_bitcast<_Ip>(__a)
1750 | __vector_bitcast<_Ip>(__b));
1751 }
1752 else if constexpr (__is_vector_type_v<_TW>)
1753 return __a | __b;
1754 else
1755 return __a._M_data | __b._M_data;
1756 }
1757 else
1758 return __a | __b;
1759 }
1760
1761// }}}
1762// __and{{{
1763template <typename _TW>
1764 _GLIBCXX_SIMD_INTRINSIC constexpr _TW
1765 __and(_TW __a, _TW __b) noexcept
1766 {
1767 if constexpr (__is_vector_type_v<_TW> || __is_simd_wrapper_v<_TW>)
1768 {
1769 using _Tp = typename conditional_t<__is_simd_wrapper_v<_TW>, _TW,
1770 _VectorTraitsImpl<_TW>>::value_type;
1771 if constexpr (is_floating_point_v<_Tp>)
1772 {
1773 using _Ip = make_unsigned_t<__int_for_sizeof_t<_Tp>>;
1774 return __vector_bitcast<_Tp>(__vector_bitcast<_Ip>(__a)
1775 & __vector_bitcast<_Ip>(__b));
1776 }
1777 else if constexpr (__is_vector_type_v<_TW>)
1778 return __a & __b;
1779 else
1780 return __a._M_data & __b._M_data;
1781 }
1782 else
1783 return __a & __b;
1784 }
1785
1786// }}}
1787// __andnot{{{
1788#if _GLIBCXX_SIMD_X86INTRIN && !defined __clang__
1789static constexpr struct
1790{
1791 _GLIBCXX_SIMD_INTRINSIC __v4sf
1792 operator()(__v4sf __a, __v4sf __b) const noexcept
1793 { return __builtin_ia32_andnps(__a, __b); }
1794
1795 _GLIBCXX_SIMD_INTRINSIC __v2df
1796 operator()(__v2df __a, __v2df __b) const noexcept
1797 { return __builtin_ia32_andnpd(__a, __b); }
1798
1799 _GLIBCXX_SIMD_INTRINSIC __v2di
1800 operator()(__v2di __a, __v2di __b) const noexcept
1801 { return __builtin_ia32_pandn128(__a, __b); }
1802
1803 _GLIBCXX_SIMD_INTRINSIC __v8sf
1804 operator()(__v8sf __a, __v8sf __b) const noexcept
1805 { return __builtin_ia32_andnps256(__a, __b); }
1806
1807 _GLIBCXX_SIMD_INTRINSIC __v4df
1808 operator()(__v4df __a, __v4df __b) const noexcept
1809 { return __builtin_ia32_andnpd256(__a, __b); }
1810
1811 _GLIBCXX_SIMD_INTRINSIC __v4di
1812 operator()(__v4di __a, __v4di __b) const noexcept
1813 {
1814 if constexpr (__have_avx2)
1815 return __builtin_ia32_andnotsi256(__a, __b);
1816 else
1817 return reinterpret_cast<__v4di>(
1818 __builtin_ia32_andnpd256(reinterpret_cast<__v4df>(__a),
1819 reinterpret_cast<__v4df>(__b)));
1820 }
1821
1822 _GLIBCXX_SIMD_INTRINSIC __v16sf
1823 operator()(__v16sf __a, __v16sf __b) const noexcept
1824 {
1825 if constexpr (__have_avx512dq)
1826 return _mm512_andnot_ps(__a, __b);
1827 else
1828 return reinterpret_cast<__v16sf>(
1829 _mm512_andnot_si512(reinterpret_cast<__v8di>(__a),
1830 reinterpret_cast<__v8di>(__b)));
1831 }
1832
1833 _GLIBCXX_SIMD_INTRINSIC __v8df
1834 operator()(__v8df __a, __v8df __b) const noexcept
1835 {
1836 if constexpr (__have_avx512dq)
1837 return _mm512_andnot_pd(__a, __b);
1838 else
1839 return reinterpret_cast<__v8df>(
1840 _mm512_andnot_si512(reinterpret_cast<__v8di>(__a),
1841 reinterpret_cast<__v8di>(__b)));
1842 }
1843
1844 _GLIBCXX_SIMD_INTRINSIC __v8di
1845 operator()(__v8di __a, __v8di __b) const noexcept
1846 { return _mm512_andnot_si512(__a, __b); }
1847} _S_x86_andnot;
1848#endif // _GLIBCXX_SIMD_X86INTRIN && !__clang__
1849
1850template <typename _TW>
1851 _GLIBCXX_SIMD_INTRINSIC constexpr _TW
1852 __andnot(_TW __a, _TW __b) noexcept
1853 {
1854 if constexpr (__is_vector_type_v<_TW> || __is_simd_wrapper_v<_TW>)
1855 {
1856 using _TVT = conditional_t<__is_simd_wrapper_v<_TW>, _TW,
1857 _VectorTraitsImpl<_TW>>;
1858 using _Tp = typename _TVT::value_type;
1859#if _GLIBCXX_SIMD_X86INTRIN && !defined __clang__
1860 if constexpr (sizeof(_TW) >= 16)
1861 {
1862 const auto __ai = __to_intrin(__a);
1863 const auto __bi = __to_intrin(__b);
1864 if (!__builtin_is_constant_evaluated()
1865 && !(__builtin_constant_p(__ai) && __builtin_constant_p(__bi)))
1866 {
1867 const auto __r = _S_x86_andnot(__ai, __bi);
1868 if constexpr (is_convertible_v<decltype(__r), _TW>)
1869 return __r;
1870 else
1871 return reinterpret_cast<typename _TVT::type>(__r);
1872 }
1873 }
1874#endif // _GLIBCXX_SIMD_X86INTRIN
1875 using _Ip = make_unsigned_t<__int_for_sizeof_t<_Tp>>;
1876 return __vector_bitcast<_Tp>(~__vector_bitcast<_Ip>(__a)
1877 & __vector_bitcast<_Ip>(__b));
1878 }
1879 else
1880 return ~__a & __b;
1881 }
1882
1883// }}}
1884// __not{{{
1885template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
1886 _GLIBCXX_SIMD_INTRINSIC constexpr _Tp
1887 __not(_Tp __a) noexcept
1888 {
1889 if constexpr (is_floating_point_v<typename _TVT::value_type>)
1890 return reinterpret_cast<typename _TVT::type>(
1891 ~__vector_bitcast<unsigned>(__a));
1892 else
1893 return ~__a;
1894 }
1895
1896// }}}
1897// __concat{{{
1898template <typename _Tp, typename _TVT = _VectorTraits<_Tp>,
1899 typename _R = __vector_type_t<typename _TVT::value_type,
1900 _TVT::_S_full_size * 2>>
1901 constexpr _R
1902 __concat(_Tp a_, _Tp b_)
1903 {
1904#ifdef _GLIBCXX_SIMD_WORKAROUND_XXX_1
1905 using _W
1906 = conditional_t<is_floating_point_v<typename _TVT::value_type>, double,
1907 conditional_t<(sizeof(_Tp) >= 2 * sizeof(long long)),
1908 long long, typename _TVT::value_type>>;
1909 constexpr int input_width = sizeof(_Tp) / sizeof(_W);
1910 const auto __a = __vector_bitcast<_W>(a_);
1911 const auto __b = __vector_bitcast<_W>(b_);
1912 using _Up = __vector_type_t<_W, sizeof(_R) / sizeof(_W)>;
1913#else
1914 constexpr int input_width = _TVT::_S_full_size;
1915 const _Tp& __a = a_;
1916 const _Tp& __b = b_;
1917 using _Up = _R;
1918#endif
1919 if constexpr (input_width == 2)
1920 return reinterpret_cast<_R>(_Up{__a[0], __a[1], __b[0], __b[1]});
1921 else if constexpr (input_width == 4)
1922 return reinterpret_cast<_R>(
1923 _Up{__a[0], __a[1], __a[2], __a[3], __b[0], __b[1], __b[2], __b[3]});
1924 else if constexpr (input_width == 8)
1925 return reinterpret_cast<_R>(
1926 _Up{__a[0], __a[1], __a[2], __a[3], __a[4], __a[5], __a[6], __a[7],
1927 __b[0], __b[1], __b[2], __b[3], __b[4], __b[5], __b[6], __b[7]});
1928 else if constexpr (input_width == 16)
1929 return reinterpret_cast<_R>(
1930 _Up{__a[0], __a[1], __a[2], __a[3], __a[4], __a[5], __a[6],
1931 __a[7], __a[8], __a[9], __a[10], __a[11], __a[12], __a[13],
1932 __a[14], __a[15], __b[0], __b[1], __b[2], __b[3], __b[4],
1933 __b[5], __b[6], __b[7], __b[8], __b[9], __b[10], __b[11],
1934 __b[12], __b[13], __b[14], __b[15]});
1935 else if constexpr (input_width == 32)
1936 return reinterpret_cast<_R>(
1937 _Up{__a[0], __a[1], __a[2], __a[3], __a[4], __a[5], __a[6],
1938 __a[7], __a[8], __a[9], __a[10], __a[11], __a[12], __a[13],
1939 __a[14], __a[15], __a[16], __a[17], __a[18], __a[19], __a[20],
1940 __a[21], __a[22], __a[23], __a[24], __a[25], __a[26], __a[27],
1941 __a[28], __a[29], __a[30], __a[31], __b[0], __b[1], __b[2],
1942 __b[3], __b[4], __b[5], __b[6], __b[7], __b[8], __b[9],
1943 __b[10], __b[11], __b[12], __b[13], __b[14], __b[15], __b[16],
1944 __b[17], __b[18], __b[19], __b[20], __b[21], __b[22], __b[23],
1945 __b[24], __b[25], __b[26], __b[27], __b[28], __b[29], __b[30],
1946 __b[31]});
1947 }
1948
1949// }}}
1950// __zero_extend {{{
1951template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
1952 struct _ZeroExtendProxy
1953 {
1954 using value_type = typename _TVT::value_type;
1955 static constexpr size_t _Np = _TVT::_S_full_size;
1956 const _Tp __x;
1957
1958 template <typename _To, typename _ToVT = _VectorTraits<_To>,
1959 typename
1960 = enable_if_t<is_same_v<typename _ToVT::value_type, value_type>>>
1961 _GLIBCXX_SIMD_INTRINSIC operator _To() const
1962 {
1963 constexpr size_t _ToN = _ToVT::_S_full_size;
1964 if constexpr (_ToN == _Np)
1965 return __x;
1966 else if constexpr (_ToN == 2 * _Np)
1967 {
1968#ifdef _GLIBCXX_SIMD_WORKAROUND_XXX_3
1969 if constexpr (__have_avx && _TVT::template _S_is<float, 4>)
1970 return __vector_bitcast<value_type>(
1971 _mm256_insertf128_ps(__m256(), __x, 0));
1972 else if constexpr (__have_avx && _TVT::template _S_is<double, 2>)
1973 return __vector_bitcast<value_type>(
1974 _mm256_insertf128_pd(__m256d(), __x, 0));
1975 else if constexpr (__have_avx2 && _Np * sizeof(value_type) == 16)
1976 return __vector_bitcast<value_type>(
1977 _mm256_insertf128_si256(__m256i(), __to_intrin(__x), 0));
1978 else if constexpr (__have_avx512f && _TVT::template _S_is<float, 8>)
1979 {
1980 if constexpr (__have_avx512dq)
1981 return __vector_bitcast<value_type>(
1982 _mm512_insertf32x8(__m512(), __x, 0));
1983 else
1984 return reinterpret_cast<__m512>(
1985 _mm512_insertf64x4(__m512d(),
1986 reinterpret_cast<__m256d>(__x), 0));
1987 }
1988 else if constexpr (__have_avx512f
1989 && _TVT::template _S_is<double, 4>)
1990 return __vector_bitcast<value_type>(
1991 _mm512_insertf64x4(__m512d(), __x, 0));
1992 else if constexpr (__have_avx512f && _Np * sizeof(value_type) == 32)
1993 return __vector_bitcast<value_type>(
1994 _mm512_inserti64x4(__m512i(), __to_intrin(__x), 0));
1995#endif
1996 return __concat(__x, _Tp());
1997 }
1998 else if constexpr (_ToN == 4 * _Np)
1999 {
2000#ifdef _GLIBCXX_SIMD_WORKAROUND_XXX_3
2001 if constexpr (__have_avx512dq && _TVT::template _S_is<double, 2>)
2002 {
2003 return __vector_bitcast<value_type>(
2004 _mm512_insertf64x2(__m512d(), __x, 0));
2005 }
2006 else if constexpr (__have_avx512f
2007 && is_floating_point_v<value_type>)
2008 {
2009 return __vector_bitcast<value_type>(
2010 _mm512_insertf32x4(__m512(), reinterpret_cast<__m128>(__x),
2011 0));
2012 }
2013 else if constexpr (__have_avx512f && _Np * sizeof(value_type) == 16)
2014 {
2015 return __vector_bitcast<value_type>(
2016 _mm512_inserti32x4(__m512i(), __to_intrin(__x), 0));
2017 }
2018#endif
2019 return __concat(__concat(__x, _Tp()),
2020 __vector_type_t<value_type, _Np * 2>());
2021 }
2022 else if constexpr (_ToN == 8 * _Np)
2023 return __concat(operator __vector_type_t<value_type, _Np * 4>(),
2024 __vector_type_t<value_type, _Np * 4>());
2025 else if constexpr (_ToN == 16 * _Np)
2026 return __concat(operator __vector_type_t<value_type, _Np * 8>(),
2027 __vector_type_t<value_type, _Np * 8>());
2028 else
2029 __assert_unreachable<_Tp>();
2030 }
2031 };
2032
2033template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
2034 _GLIBCXX_SIMD_INTRINSIC _ZeroExtendProxy<_Tp, _TVT>
2035 __zero_extend(_Tp __x)
2036 { return {__x}; }
2037
2038// }}}
2039// __extract<_Np, By>{{{
2040template <int _Offset,
2041 int _SplitBy,
2042 typename _Tp,
2043 typename _TVT = _VectorTraits<_Tp>,
2044 typename _R = __vector_type_t<typename _TVT::value_type,
2045 _TVT::_S_full_size / _SplitBy>>
2046 _GLIBCXX_SIMD_INTRINSIC constexpr _R
2047 __extract(_Tp __in)
2048 {
2049 using value_type = typename _TVT::value_type;
2050#if _GLIBCXX_SIMD_X86INTRIN // {{{
2051 if constexpr (sizeof(_Tp) == 64 && _SplitBy == 4 && _Offset > 0)
2052 {
2053 if constexpr (__have_avx512dq && is_same_v<double, value_type>)
2054 return _mm512_extractf64x2_pd(__to_intrin(__in), _Offset);
2055 else if constexpr (is_floating_point_v<value_type>)
2056 return __vector_bitcast<value_type>(
2057 _mm512_extractf32x4_ps(__intrin_bitcast<__m512>(__in), _Offset));
2058 else
2059 return reinterpret_cast<_R>(
2060 _mm512_extracti32x4_epi32(__intrin_bitcast<__m512i>(__in),
2061 _Offset));
2062 }
2063 else
2064#endif // _GLIBCXX_SIMD_X86INTRIN }}}
2065 {
2066#ifdef _GLIBCXX_SIMD_WORKAROUND_XXX_1
2067 using _W = conditional_t<
2068 is_floating_point_v<value_type>, double,
2069 conditional_t<(sizeof(_R) >= 16), long long, value_type>>;
2070 static_assert(sizeof(_R) % sizeof(_W) == 0);
2071 constexpr int __return_width = sizeof(_R) / sizeof(_W);
2072 using _Up = __vector_type_t<_W, __return_width>;
2073 const auto __x = __vector_bitcast<_W>(__in);
2074#else
2075 constexpr int __return_width = _TVT::_S_full_size / _SplitBy;
2076 using _Up = _R;
2077 const __vector_type_t<value_type, _TVT::_S_full_size>& __x
2078 = __in; // only needed for _Tp = _SimdWrapper<value_type, _Np>
2079#endif
2080 constexpr int _O = _Offset * __return_width;
2081 return __call_with_subscripts<__return_width, _O>(
2082 __x, [](auto... __entries) {
2083 return reinterpret_cast<_R>(_Up{__entries...});
2084 });
2085 }
2086 }
2087
2088// }}}
2089// __lo/__hi64[z]{{{
2090template <typename _Tp,
2091 typename _R
2092 = __vector_type8_t<typename _VectorTraits<_Tp>::value_type>>
2093 _GLIBCXX_SIMD_INTRINSIC constexpr _R
2094 __lo64(_Tp __x)
2095 {
2096 _R __r{};
2097 __builtin_memcpy(&__r, &__x, 8);
2098 return __r;
2099 }
2100
2101template <typename _Tp,
2102 typename _R
2103 = __vector_type8_t<typename _VectorTraits<_Tp>::value_type>>
2104 _GLIBCXX_SIMD_INTRINSIC constexpr _R
2105 __hi64(_Tp __x)
2106 {
2107 static_assert(sizeof(_Tp) == 16, "use __hi64z if you meant it");
2108 _R __r{};
2109 __builtin_memcpy(&__r, reinterpret_cast<const char*>(&__x) + 8, 8);
2110 return __r;
2111 }
2112
2113template <typename _Tp,
2114 typename _R
2115 = __vector_type8_t<typename _VectorTraits<_Tp>::value_type>>
2116 _GLIBCXX_SIMD_INTRINSIC constexpr _R
2117 __hi64z([[maybe_unused]] _Tp __x)
2118 {
2119 _R __r{};
2120 if constexpr (sizeof(_Tp) == 16)
2121 __builtin_memcpy(&__r, reinterpret_cast<const char*>(&__x) + 8, 8);
2122 return __r;
2123 }
2124
2125// }}}
2126// __lo/__hi128{{{
2127template <typename _Tp>
2128 _GLIBCXX_SIMD_INTRINSIC constexpr auto
2129 __lo128(_Tp __x)
2130 { return __extract<0, sizeof(_Tp) / 16>(__x); }
2131
2132template <typename _Tp>
2133 _GLIBCXX_SIMD_INTRINSIC constexpr auto
2134 __hi128(_Tp __x)
2135 {
2136 static_assert(sizeof(__x) == 32);
2137 return __extract<1, 2>(__x);
2138 }
2139
2140// }}}
2141// __lo/__hi256{{{
2142template <typename _Tp>
2143 _GLIBCXX_SIMD_INTRINSIC constexpr auto
2144 __lo256(_Tp __x)
2145 {
2146 static_assert(sizeof(__x) == 64);
2147 return __extract<0, 2>(__x);
2148 }
2149
2150template <typename _Tp>
2151 _GLIBCXX_SIMD_INTRINSIC constexpr auto
2152 __hi256(_Tp __x)
2153 {
2154 static_assert(sizeof(__x) == 64);
2155 return __extract<1, 2>(__x);
2156 }
2157
2158// }}}
2159// __auto_bitcast{{{
2160template <typename _Tp>
2161 struct _AutoCast
2162 {
2163 static_assert(__is_vector_type_v<_Tp>);
2164
2165 const _Tp __x;
2166
2167 template <typename _Up, typename _UVT = _VectorTraits<_Up>>
2168 _GLIBCXX_SIMD_INTRINSIC constexpr operator _Up() const
2169 { return __intrin_bitcast<typename _UVT::type>(__x); }
2170 };
2171
2172template <typename _Tp>
2173 _GLIBCXX_SIMD_INTRINSIC constexpr _AutoCast<_Tp>
2174 __auto_bitcast(const _Tp& __x)
2175 { return {__x}; }
2176
2177template <typename _Tp, size_t _Np>
2178 _GLIBCXX_SIMD_INTRINSIC constexpr
2179 _AutoCast<typename _SimdWrapper<_Tp, _Np>::_BuiltinType>
2180 __auto_bitcast(const _SimdWrapper<_Tp, _Np>& __x)
2181 { return {__x._M_data}; }
2182
2183// }}}
2184// ^^^ ---- builtin vector types [[gnu::vector_size(N)]] and operations ---- ^^^
2185
2186#if _GLIBCXX_SIMD_HAVE_SSE_ABI
2187// __bool_storage_member_type{{{
2188#if _GLIBCXX_SIMD_HAVE_AVX512F && _GLIBCXX_SIMD_X86INTRIN
2189template <size_t _Size>
2190 struct __bool_storage_member_type
2191 {
2192 static_assert((_Size & (_Size - 1)) != 0,
2193 "This trait may only be used for non-power-of-2 sizes. "
2194 "Power-of-2 sizes must be specialized.");
2195 using type =
2196 typename __bool_storage_member_type<std::__bit_ceil(_Size)>::type;
2197 };
2198
2199template <>
2200 struct __bool_storage_member_type<1> { using type = bool; };
2201
2202template <>
2203 struct __bool_storage_member_type<2> { using type = __mmask8; };
2204
2205template <>
2206 struct __bool_storage_member_type<4> { using type = __mmask8; };
2207
2208template <>
2209 struct __bool_storage_member_type<8> { using type = __mmask8; };
2210
2211template <>
2212 struct __bool_storage_member_type<16> { using type = __mmask16; };
2213
2214template <>
2215 struct __bool_storage_member_type<32> { using type = __mmask32; };
2216
2217template <>
2218 struct __bool_storage_member_type<64> { using type = __mmask64; };
2219#endif // _GLIBCXX_SIMD_HAVE_AVX512F
2220
2221// }}}
2222// __intrinsic_type (x86){{{
2223// the following excludes bool via __is_vectorizable
2224#if _GLIBCXX_SIMD_HAVE_SSE
2225template <typename _Tp, size_t _Bytes>
2226 struct __intrinsic_type<_Tp, _Bytes,
2227 enable_if_t<__is_vectorizable_v<_Tp> && _Bytes <= 64>>
2228 {
2229 static_assert(!is_same_v<_Tp, long double>,
2230 "no __intrinsic_type support for long double on x86");
2231
2232 static constexpr size_t _S_VBytes = _Bytes <= 16 ? 16
2233 : _Bytes <= 32 ? 32
2234 : 64;
2235
2236 using type [[__gnu__::__vector_size__(_S_VBytes)]]
2237 = conditional_t<is_integral_v<_Tp>, long long int, _Tp>;
2238 };
2239#endif // _GLIBCXX_SIMD_HAVE_SSE
2240
2241// }}}
2242#endif // _GLIBCXX_SIMD_HAVE_SSE_ABI
2243// __intrinsic_type (ARM){{{
2244#if _GLIBCXX_SIMD_HAVE_NEON
2245template <>
2246 struct __intrinsic_type<float, 8, void>
2247 { using type = float32x2_t; };
2248
2249template <>
2250 struct __intrinsic_type<float, 16, void>
2251 { using type = float32x4_t; };
2252
2253#if _GLIBCXX_SIMD_HAVE_NEON_A64
2254template <>
2255 struct __intrinsic_type<double, 8, void>
2256 { using type = float64x1_t; };
2257
2258template <>
2259 struct __intrinsic_type<double, 16, void>
2260 { using type = float64x2_t; };
2261#endif
2262
2263#define _GLIBCXX_SIMD_ARM_INTRIN(_Bits, _Np) \
2264template <> \
2265 struct __intrinsic_type<__int_with_sizeof_t<_Bits / 8>, \
2266 _Np * _Bits / 8, void> \
2267 { using type = int##_Bits##x##_Np##_t; }; \
2268template <> \
2269 struct __intrinsic_type<make_unsigned_t<__int_with_sizeof_t<_Bits / 8>>, \
2270 _Np * _Bits / 8, void> \
2271 { using type = uint##_Bits##x##_Np##_t; }
2272_GLIBCXX_SIMD_ARM_INTRIN(8, 8);
2273_GLIBCXX_SIMD_ARM_INTRIN(8, 16);
2274_GLIBCXX_SIMD_ARM_INTRIN(16, 4);
2275_GLIBCXX_SIMD_ARM_INTRIN(16, 8);
2276_GLIBCXX_SIMD_ARM_INTRIN(32, 2);
2277_GLIBCXX_SIMD_ARM_INTRIN(32, 4);
2278_GLIBCXX_SIMD_ARM_INTRIN(64, 1);
2279_GLIBCXX_SIMD_ARM_INTRIN(64, 2);
2280#undef _GLIBCXX_SIMD_ARM_INTRIN
2281
2282template <typename _Tp, size_t _Bytes>
2283 struct __intrinsic_type<_Tp, _Bytes,
2284 enable_if_t<__is_vectorizable_v<_Tp> && _Bytes <= 16>>
2285 {
2286 static constexpr int _SVecBytes = _Bytes <= 8 ? 8 : 16;
2287 using _Ip = __int_for_sizeof_t<_Tp>;
2288 using _Up = conditional_t<
2289 is_floating_point_v<_Tp>, _Tp,
2290 conditional_t<is_unsigned_v<_Tp>, make_unsigned_t<_Ip>, _Ip>>;
2291 static_assert(!is_same_v<_Tp, _Up> || _SVecBytes != _Bytes,
2292 "should use explicit specialization above");
2293 using type = typename __intrinsic_type<_Up, _SVecBytes>::type;
2294 };
2295#endif // _GLIBCXX_SIMD_HAVE_NEON
2296
2297// }}}
2298// __intrinsic_type (PPC){{{
2299#ifdef __ALTIVEC__
2300template <typename _Tp>
2301 struct __intrinsic_type_impl;
2302
2303#define _GLIBCXX_SIMD_PPC_INTRIN(_Tp) \
2304 template <> \
2305 struct __intrinsic_type_impl<_Tp> { using type = __vector _Tp; }
2306_GLIBCXX_SIMD_PPC_INTRIN(float);
2307#ifdef __VSX__
2308_GLIBCXX_SIMD_PPC_INTRIN(double);
2309#endif
2310_GLIBCXX_SIMD_PPC_INTRIN(signed char);
2311_GLIBCXX_SIMD_PPC_INTRIN(unsigned char);
2312_GLIBCXX_SIMD_PPC_INTRIN(signed short);
2313_GLIBCXX_SIMD_PPC_INTRIN(unsigned short);
2314_GLIBCXX_SIMD_PPC_INTRIN(signed int);
2315_GLIBCXX_SIMD_PPC_INTRIN(unsigned int);
2316#if defined __VSX__ || __SIZEOF_LONG__ == 4
2317_GLIBCXX_SIMD_PPC_INTRIN(signed long);
2318_GLIBCXX_SIMD_PPC_INTRIN(unsigned long);
2319#endif
2320#ifdef __VSX__
2321_GLIBCXX_SIMD_PPC_INTRIN(signed long long);
2322_GLIBCXX_SIMD_PPC_INTRIN(unsigned long long);
2323#endif
2324#undef _GLIBCXX_SIMD_PPC_INTRIN
2325
2326template <typename _Tp, size_t _Bytes>
2327 struct __intrinsic_type<_Tp, _Bytes,
2328 enable_if_t<__is_vectorizable_v<_Tp> && _Bytes <= 16>>
2329 {
2330 static constexpr bool _S_is_ldouble = is_same_v<_Tp, long double>;
2331 // allow _Tp == long double with -mlong-double-64
2332 static_assert(!(_S_is_ldouble && sizeof(long double) > sizeof(double)),
2333 "no __intrinsic_type support for 128-bit floating point on PowerPC");
2334#ifndef __VSX__
2335 static_assert(!(is_same_v<_Tp, double>
2336 || (_S_is_ldouble && sizeof(long double) == sizeof(double))),
2337 "no __intrinsic_type support for 64-bit floating point on PowerPC w/o VSX");
2338#endif
2339 using type =
2340 typename __intrinsic_type_impl<
2341 conditional_t<is_floating_point_v<_Tp>,
2342 conditional_t<_S_is_ldouble, double, _Tp>,
2343 __int_for_sizeof_t<_Tp>>>::type;
2344 };
2345#endif // __ALTIVEC__
2346
2347// }}}
2348// _SimdWrapper<bool>{{{1
2349template <size_t _Width>
2350 struct _SimdWrapper<bool, _Width,
2351 void_t<typename __bool_storage_member_type<_Width>::type>>
2352 {
2353 using _BuiltinType = typename __bool_storage_member_type<_Width>::type;
2354 using value_type = bool;
2355
2356 static constexpr size_t _S_full_size = sizeof(_BuiltinType) * __CHAR_BIT__;
2357
2358 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper<bool, _S_full_size>
2359 __as_full_vector() const { return _M_data; }
2360
2361 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper() = default;
2362 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper(_BuiltinType __k)
2363 : _M_data(__k) {};
2364
2365 _GLIBCXX_SIMD_INTRINSIC operator const _BuiltinType&() const
2366 { return _M_data; }
2367
2368 _GLIBCXX_SIMD_INTRINSIC operator _BuiltinType&()
2369 { return _M_data; }
2370
2371 _GLIBCXX_SIMD_INTRINSIC _BuiltinType __intrin() const
2372 { return _M_data; }
2373
2374 _GLIBCXX_SIMD_INTRINSIC constexpr value_type operator[](size_t __i) const
2375 { return _M_data & (_BuiltinType(1) << __i); }
2376
2377 template <size_t __i>
2378 _GLIBCXX_SIMD_INTRINSIC constexpr value_type
2379 operator[](_SizeConstant<__i>) const
2380 { return _M_data & (_BuiltinType(1) << __i); }
2381
2382 _GLIBCXX_SIMD_INTRINSIC constexpr void _M_set(size_t __i, value_type __x)
2383 {
2384 if (__x)
2385 _M_data |= (_BuiltinType(1) << __i);
2386 else
2387 _M_data &= ~(_BuiltinType(1) << __i);
2388 }
2389
2390 _GLIBCXX_SIMD_INTRINSIC
2391 constexpr bool _M_is_constprop() const
2392 { return __builtin_constant_p(_M_data); }
2393
2394 _GLIBCXX_SIMD_INTRINSIC constexpr bool _M_is_constprop_none_of() const
2395 {
2396 if (__builtin_constant_p(_M_data))
2397 {
2398 constexpr int __nbits = sizeof(_BuiltinType) * __CHAR_BIT__;
2399 constexpr _BuiltinType __active_mask
2400 = ~_BuiltinType() >> (__nbits - _Width);
2401 return (_M_data & __active_mask) == 0;
2402 }
2403 return false;
2404 }
2405
2406 _GLIBCXX_SIMD_INTRINSIC constexpr bool _M_is_constprop_all_of() const
2407 {
2408 if (__builtin_constant_p(_M_data))
2409 {
2410 constexpr int __nbits = sizeof(_BuiltinType) * __CHAR_BIT__;
2411 constexpr _BuiltinType __active_mask
2412 = ~_BuiltinType() >> (__nbits - _Width);
2413 return (_M_data & __active_mask) == __active_mask;
2414 }
2415 return false;
2416 }
2417
2418 _BuiltinType _M_data;
2419 };
2420
2421// _SimdWrapperBase{{{1
2422template <bool _MustZeroInitPadding, typename _BuiltinType>
2423 struct _SimdWrapperBase;
2424
2425template <typename _BuiltinType>
2426 struct _SimdWrapperBase<false, _BuiltinType> // no padding or no SNaNs
2427 {
2428 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapperBase() = default;
2429 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapperBase(_BuiltinType __init)
2430 : _M_data(__init)
2431 {}
2432
2433 _BuiltinType _M_data;
2434 };
2435
2436template <typename _BuiltinType>
2437 struct _SimdWrapperBase<true, _BuiltinType> // with padding that needs to
2438 // never become SNaN
2439 {
2440 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapperBase() : _M_data() {}
2441 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapperBase(_BuiltinType __init)
2442 : _M_data(__init)
2443 {}
2444
2445 _BuiltinType _M_data;
2446 };
2447
2448// }}}
2449// _SimdWrapper{{{
2450template <typename _Tp, size_t _Width>
2451 struct _SimdWrapper<
2452 _Tp, _Width,
2453 void_t<__vector_type_t<_Tp, _Width>, __intrinsic_type_t<_Tp, _Width>>>
2454 : _SimdWrapperBase<__has_iec559_behavior<__signaling_NaN, _Tp>::value
2455 && sizeof(_Tp) * _Width
2456 == sizeof(__vector_type_t<_Tp, _Width>),
2457 __vector_type_t<_Tp, _Width>>
2458 {
2459 using _Base
2460 = _SimdWrapperBase<__has_iec559_behavior<__signaling_NaN, _Tp>::value
2461 && sizeof(_Tp) * _Width
2462 == sizeof(__vector_type_t<_Tp, _Width>),
2463 __vector_type_t<_Tp, _Width>>;
2464
2465 static_assert(__is_vectorizable_v<_Tp>);
2466 static_assert(_Width >= 2); // 1 doesn't make sense, use _Tp directly then
2467
2468 using _BuiltinType = __vector_type_t<_Tp, _Width>;
2469 using value_type = _Tp;
2470
2471 static inline constexpr size_t _S_full_size
2472 = sizeof(_BuiltinType) / sizeof(value_type);
2473 static inline constexpr int _S_size = _Width;
2474 static inline constexpr bool _S_is_partial = _S_full_size != _S_size;
2475
2476 using _Base::_M_data;
2477
2478 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper<_Tp, _S_full_size>
2479 __as_full_vector() const
2480 { return _M_data; }
2481
2482 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper(initializer_list<_Tp> __init)
2483 : _Base(__generate_from_n_evaluations<_Width, _BuiltinType>(
2484 [&](auto __i) { return __init.begin()[__i.value]; })) {}
2485
2486 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper() = default;
2487 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper(const _SimdWrapper&)
2488 = default;
2489 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper(_SimdWrapper&&) = default;
2490
2491 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper&
2492 operator=(const _SimdWrapper&) = default;
2493 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper&
2494 operator=(_SimdWrapper&&) = default;
2495
2496 template <typename _V, typename = enable_if_t<disjunction_v<
2497 is_same<_V, __vector_type_t<_Tp, _Width>>,
2498 is_same<_V, __intrinsic_type_t<_Tp, _Width>>>>>
2499 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper(_V __x)
2500 // __vector_bitcast can convert e.g. __m128 to __vector(2) float
2501 : _Base(__vector_bitcast<_Tp, _Width>(__x)) {}
2502
2503 template <typename... _As,
2504 typename = enable_if_t<((is_same_v<simd_abi::scalar, _As> && ...)
2505 && sizeof...(_As) <= _Width)>>
2506 _GLIBCXX_SIMD_INTRINSIC constexpr
2507 operator _SimdTuple<_Tp, _As...>() const
2508 {
2509 const auto& dd = _M_data; // workaround for GCC7 ICE
2510 return __generate_from_n_evaluations<sizeof...(_As),
2511 _SimdTuple<_Tp, _As...>>([&](
2512 auto __i) constexpr { return dd[int(__i)]; });
2513 }
2514
2515 _GLIBCXX_SIMD_INTRINSIC constexpr operator const _BuiltinType&() const
2516 { return _M_data; }
2517
2518 _GLIBCXX_SIMD_INTRINSIC constexpr operator _BuiltinType&()
2519 { return _M_data; }
2520
2521 _GLIBCXX_SIMD_INTRINSIC constexpr _Tp operator[](size_t __i) const
2522 { return _M_data[__i]; }
2523
2524 template <size_t __i>
2525 _GLIBCXX_SIMD_INTRINSIC constexpr _Tp operator[](_SizeConstant<__i>) const
2526 { return _M_data[__i]; }
2527
2528 _GLIBCXX_SIMD_INTRINSIC constexpr void _M_set(size_t __i, _Tp __x)
2529 { _M_data[__i] = __x; }
2530
2531 _GLIBCXX_SIMD_INTRINSIC
2532 constexpr bool _M_is_constprop() const
2533 { return __builtin_constant_p(_M_data); }
2534
2535 _GLIBCXX_SIMD_INTRINSIC constexpr bool _M_is_constprop_none_of() const
2536 {
2537 if (__builtin_constant_p(_M_data))
2538 {
2539 bool __r = true;
2540 if constexpr (is_floating_point_v<_Tp>)
2541 {
2542 using _Ip = __int_for_sizeof_t<_Tp>;
2543 const auto __intdata = __vector_bitcast<_Ip>(_M_data);
2544 __execute_n_times<_Width>(
2545 [&](auto __i) { __r &= __intdata[__i.value] == _Ip(); });
2546 }
2547 else
2548 __execute_n_times<_Width>(
2549 [&](auto __i) { __r &= _M_data[__i.value] == _Tp(); });
2550 if (__builtin_constant_p(__r))
2551 return __r;
2552 }
2553 return false;
2554 }
2555
2556 _GLIBCXX_SIMD_INTRINSIC constexpr bool _M_is_constprop_all_of() const
2557 {
2558 if (__builtin_constant_p(_M_data))
2559 {
2560 bool __r = true;
2561 if constexpr (is_floating_point_v<_Tp>)
2562 {
2563 using _Ip = __int_for_sizeof_t<_Tp>;
2564 const auto __intdata = __vector_bitcast<_Ip>(_M_data);
2565 __execute_n_times<_Width>(
2566 [&](auto __i) { __r &= __intdata[__i.value] == ~_Ip(); });
2567 }
2568 else
2569 __execute_n_times<_Width>(
2570 [&](auto __i) { __r &= _M_data[__i.value] == ~_Tp(); });
2571 if (__builtin_constant_p(__r))
2572 return __r;
2573 }
2574 return false;
2575 }
2576 };
2577
2578// }}}
2579
2580// __vectorized_sizeof {{{
2581template <typename _Tp>
2582 constexpr size_t
2583 __vectorized_sizeof()
2584 {
2585 if constexpr (!__is_vectorizable_v<_Tp>)
2586 return 0;
2587
2588 if constexpr (sizeof(_Tp) <= 8)
2589 {
2590 // X86:
2591 if constexpr (__have_avx512bw)
2592 return 64;
2593 if constexpr (__have_avx512f && sizeof(_Tp) >= 4)
2594 return 64;
2595 if constexpr (__have_avx2)
2596 return 32;
2597 if constexpr (__have_avx && is_floating_point_v<_Tp>)
2598 return 32;
2599 if constexpr (__have_sse2)
2600 return 16;
2601 if constexpr (__have_sse && is_same_v<_Tp, float>)
2602 return 16;
2603 /* The following is too much trouble because of mixed MMX and x87 code.
2604 * While nothing here explicitly calls MMX instructions of registers,
2605 * they are still emitted but no EMMS cleanup is done.
2606 if constexpr (__have_mmx && sizeof(_Tp) <= 4 && is_integral_v<_Tp>)
2607 return 8;
2608 */
2609
2610 // PowerPC:
2611 if constexpr (__have_power8vec
2612 || (__have_power_vmx && (sizeof(_Tp) < 8))
2613 || (__have_power_vsx && is_floating_point_v<_Tp>) )
2614 return 16;
2615
2616 // ARM:
2617 if constexpr (__have_neon_a64
2618 || (__have_neon_a32 && !is_same_v<_Tp, double>) )
2619 return 16;
2620 if constexpr (__have_neon
2621 && sizeof(_Tp) < 8
2622 // Only allow fp if the user allows non-ICE559 fp (e.g.
2623 // via -ffast-math). ARMv7 NEON fp is not conforming to
2624 // IEC559.
2625 && (__support_neon_float || !is_floating_point_v<_Tp>))
2626 return 16;
2627 }
2628
2629 return sizeof(_Tp);
2630 }
2631
2632// }}}
2633namespace simd_abi {
2634// most of simd_abi is defined in simd_detail.h
2635template <typename _Tp>
2636 inline constexpr int max_fixed_size
2637 = (__have_avx512bw && sizeof(_Tp) == 1) ? 64 : 32;
2638
2639// compatible {{{
2640#if defined __x86_64__ || defined __aarch64__
2641template <typename _Tp>
2642 using compatible = conditional_t<(sizeof(_Tp) <= 8), _VecBuiltin<16>, scalar>;
2643#elif defined __ARM_NEON
2644// FIXME: not sure, probably needs to be scalar (or dependent on the hard-float
2645// ABI?)
2646template <typename _Tp>
2647 using compatible
2648 = conditional_t<(sizeof(_Tp) < 8
2649 && (__support_neon_float || !is_floating_point_v<_Tp>)),
2650 _VecBuiltin<16>, scalar>;
2651#else
2652template <typename>
2653 using compatible = scalar;
2654#endif
2655
2656// }}}
2657// native {{{
2658template <typename _Tp>
2659 constexpr auto
2660 __determine_native_abi()
2661 {
2662 constexpr size_t __bytes = __vectorized_sizeof<_Tp>();
2663 if constexpr (__bytes == sizeof(_Tp))
2664 return static_cast<scalar*>(nullptr);
2665 else if constexpr (__have_avx512vl || (__have_avx512f && __bytes == 64))
2666 return static_cast<_VecBltnBtmsk<__bytes>*>(nullptr);
2667 else
2668 return static_cast<_VecBuiltin<__bytes>*>(nullptr);
2669 }
2670
2671template <typename _Tp, typename = enable_if_t<__is_vectorizable_v<_Tp>>>
2672 using native = remove_pointer_t<decltype(__determine_native_abi<_Tp>())>;
2673
2674// }}}
2675// __default_abi {{{
2676#if defined _GLIBCXX_SIMD_DEFAULT_ABI
2677template <typename _Tp>
2678 using __default_abi = _GLIBCXX_SIMD_DEFAULT_ABI<_Tp>;
2679#else
2680template <typename _Tp>
2681 using __default_abi = compatible<_Tp>;
2682#endif
2683
2684// }}}
2685} // namespace simd_abi
2686
2687// traits {{{1
2688// is_abi_tag {{{2
2689template <typename _Tp, typename = void_t<>>
2690 struct is_abi_tag : false_type {};
2691
2692template <typename _Tp>
2693 struct is_abi_tag<_Tp, void_t<typename _Tp::_IsValidAbiTag>>
2694 : public _Tp::_IsValidAbiTag {};
2695
2696template <typename _Tp>
2697 inline constexpr bool is_abi_tag_v = is_abi_tag<_Tp>::value;
2698
2699// is_simd(_mask) {{{2
2700template <typename _Tp>
2701 struct is_simd : public false_type {};
2702
2703template <typename _Tp>
2704 inline constexpr bool is_simd_v = is_simd<_Tp>::value;
2705
2706template <typename _Tp>
2707 struct is_simd_mask : public false_type {};
2708
2709template <typename _Tp>
2710inline constexpr bool is_simd_mask_v = is_simd_mask<_Tp>::value;
2711
2712// simd_size {{{2
2713template <typename _Tp, typename _Abi, typename = void>
2714 struct __simd_size_impl {};
2715
2716template <typename _Tp, typename _Abi>
2717 struct __simd_size_impl<
2718 _Tp, _Abi,
2719 enable_if_t<conjunction_v<__is_vectorizable<_Tp>, is_abi_tag<_Abi>>>>
2720 : _SizeConstant<_Abi::template _S_size<_Tp>> {};
2721
2722template <typename _Tp, typename _Abi = simd_abi::__default_abi<_Tp>>
2723 struct simd_size : __simd_size_impl<_Tp, _Abi> {};
2724
2725template <typename _Tp, typename _Abi = simd_abi::__default_abi<_Tp>>
2726 inline constexpr size_t simd_size_v = simd_size<_Tp, _Abi>::value;
2727
2728// simd_abi::deduce {{{2
2729template <typename _Tp, size_t _Np, typename = void>
2730 struct __deduce_impl;
2731
2732namespace simd_abi {
2733/**
2734 * @tparam _Tp The requested `value_type` for the elements.
2735 * @tparam _Np The requested number of elements.
2736 * @tparam _Abis This parameter is ignored, since this implementation cannot
2737 * make any use of it. Either __a good native ABI is matched and used as `type`
2738 * alias, or the `fixed_size<_Np>` ABI is used, which internally is built from
2739 * the best matching native ABIs.
2740 */
2741template <typename _Tp, size_t _Np, typename...>
2742 struct deduce : __deduce_impl<_Tp, _Np> {};
2743
2744template <typename _Tp, size_t _Np, typename... _Abis>
2745 using deduce_t = typename deduce<_Tp, _Np, _Abis...>::type;
2746} // namespace simd_abi
2747
2748// }}}2
2749// rebind_simd {{{2
2750template <typename _Tp, typename _V, typename = void>
2751 struct rebind_simd;
2752
2753template <typename _Tp, typename _Up, typename _Abi>
2754 struct rebind_simd<
2755 _Tp, simd<_Up, _Abi>,
2756 void_t<simd_abi::deduce_t<_Tp, simd_size_v<_Up, _Abi>, _Abi>>>
2757 {
2758 using type
2759 = simd<_Tp, simd_abi::deduce_t<_Tp, simd_size_v<_Up, _Abi>, _Abi>>;
2760 };
2761
2762template <typename _Tp, typename _Up, typename _Abi>
2763 struct rebind_simd<
2764 _Tp, simd_mask<_Up, _Abi>,
2765 void_t<simd_abi::deduce_t<_Tp, simd_size_v<_Up, _Abi>, _Abi>>>
2766 {
2767 using type
2768 = simd_mask<_Tp, simd_abi::deduce_t<_Tp, simd_size_v<_Up, _Abi>, _Abi>>;
2769 };
2770
2771template <typename _Tp, typename _V>
2772 using rebind_simd_t = typename rebind_simd<_Tp, _V>::type;
2773
2774// resize_simd {{{2
2775template <int _Np, typename _V, typename = void>
2776 struct resize_simd;
2777
2778template <int _Np, typename _Tp, typename _Abi>
2779 struct resize_simd<_Np, simd<_Tp, _Abi>,
2780 void_t<simd_abi::deduce_t<_Tp, _Np, _Abi>>>
2781 { using type = simd<_Tp, simd_abi::deduce_t<_Tp, _Np, _Abi>>; };
2782
2783template <int _Np, typename _Tp, typename _Abi>
2784 struct resize_simd<_Np, simd_mask<_Tp, _Abi>,
2785 void_t<simd_abi::deduce_t<_Tp, _Np, _Abi>>>
2786 { using type = simd_mask<_Tp, simd_abi::deduce_t<_Tp, _Np, _Abi>>; };
2787
2788template <int _Np, typename _V>
2789 using resize_simd_t = typename resize_simd<_Np, _V>::type;
2790
2791// }}}2
2792// memory_alignment {{{2
2793template <typename _Tp, typename _Up = typename _Tp::value_type>
2794 struct memory_alignment
2795 : public _SizeConstant<vector_aligned_tag::_S_alignment<_Tp, _Up>> {};
2796
2797template <typename _Tp, typename _Up = typename _Tp::value_type>
2798 inline constexpr size_t memory_alignment_v = memory_alignment<_Tp, _Up>::value;
2799
2800// class template simd [simd] {{{1
2801template <typename _Tp, typename _Abi = simd_abi::__default_abi<_Tp>>
2802 class simd;
2803
2804template <typename _Tp, typename _Abi>
2805 struct is_simd<simd<_Tp, _Abi>> : public true_type {};
2806
2807template <typename _Tp>
2808 using native_simd = simd<_Tp, simd_abi::native<_Tp>>;
2809
2810template <typename _Tp, int _Np>
2811 using fixed_size_simd = simd<_Tp, simd_abi::fixed_size<_Np>>;
2812
2813template <typename _Tp, size_t _Np>
2814 using __deduced_simd = simd<_Tp, simd_abi::deduce_t<_Tp, _Np>>;
2815
2816// class template simd_mask [simd_mask] {{{1
2817template <typename _Tp, typename _Abi = simd_abi::__default_abi<_Tp>>
2818 class simd_mask;
2819
2820template <typename _Tp, typename _Abi>
2821 struct is_simd_mask<simd_mask<_Tp, _Abi>> : public true_type {};
2822
2823template <typename _Tp>
2824 using native_simd_mask = simd_mask<_Tp, simd_abi::native<_Tp>>;
2825
2826template <typename _Tp, int _Np>
2827 using fixed_size_simd_mask = simd_mask<_Tp, simd_abi::fixed_size<_Np>>;
2828
2829template <typename _Tp, size_t _Np>
2830 using __deduced_simd_mask = simd_mask<_Tp, simd_abi::deduce_t<_Tp, _Np>>;
2831
2832// casts [simd.casts] {{{1
2833// static_simd_cast {{{2
2834template <typename _Tp, typename _Up, typename _Ap, bool = is_simd_v<_Tp>,
2835 typename = void>
2836 struct __static_simd_cast_return_type;
2837
2838template <typename _Tp, typename _A0, typename _Up, typename _Ap>
2839 struct __static_simd_cast_return_type<simd_mask<_Tp, _A0>, _Up, _Ap, false,
2840 void>
2841 : __static_simd_cast_return_type<simd<_Tp, _A0>, _Up, _Ap> {};
2842
2843template <typename _Tp, typename _Up, typename _Ap>
2844 struct __static_simd_cast_return_type<
2845 _Tp, _Up, _Ap, true, enable_if_t<_Tp::size() == simd_size_v<_Up, _Ap>>>
2846 { using type = _Tp; };
2847
2848template <typename _Tp, typename _Ap>
2849 struct __static_simd_cast_return_type<_Tp, _Tp, _Ap, false,
2850#ifdef _GLIBCXX_SIMD_FIX_P2TS_ISSUE66
2851 enable_if_t<__is_vectorizable_v<_Tp>>
2852#else
2853 void
2854#endif
2855 >
2856 { using type = simd<_Tp, _Ap>; };
2857
2858template <typename _Tp, typename = void>
2859 struct __safe_make_signed { using type = _Tp;};
2860
2861template <typename _Tp>
2862 struct __safe_make_signed<_Tp, enable_if_t<is_integral_v<_Tp>>>
2863 {
2864 // the extra make_unsigned_t is because of PR85951
2865 using type = make_signed_t<make_unsigned_t<_Tp>>;
2866 };
2867
2868template <typename _Tp>
2869 using safe_make_signed_t = typename __safe_make_signed<_Tp>::type;
2870
2871template <typename _Tp, typename _Up, typename _Ap>
2872 struct __static_simd_cast_return_type<_Tp, _Up, _Ap, false,
2873#ifdef _GLIBCXX_SIMD_FIX_P2TS_ISSUE66
2874 enable_if_t<__is_vectorizable_v<_Tp>>
2875#else
2876 void
2877#endif
2878 >
2879 {
2880 using type = conditional_t<
2881 (is_integral_v<_Up> && is_integral_v<_Tp> &&
2882#ifndef _GLIBCXX_SIMD_FIX_P2TS_ISSUE65
2883 is_signed_v<_Up> != is_signed_v<_Tp> &&
2884#endif
2885 is_same_v<safe_make_signed_t<_Up>, safe_make_signed_t<_Tp>>),
2886 simd<_Tp, _Ap>, fixed_size_simd<_Tp, simd_size_v<_Up, _Ap>>>;
2887 };
2888
2889template <typename _Tp, typename _Up, typename _Ap,
2890 typename _R
2891 = typename __static_simd_cast_return_type<_Tp, _Up, _Ap>::type>
2892 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR _R
2893 static_simd_cast(const simd<_Up, _Ap>& __x)
2894 {
2895 if constexpr (is_same<_R, simd<_Up, _Ap>>::value)
2896 return __x;
2897 else
2898 {
2899 _SimdConverter<_Up, _Ap, typename _R::value_type, typename _R::abi_type>
2900 __c;
2901 return _R(__private_init, __c(__data(__x)));
2902 }
2903 }
2904
2905namespace __proposed {
2906template <typename _Tp, typename _Up, typename _Ap,
2907 typename _R
2908 = typename __static_simd_cast_return_type<_Tp, _Up, _Ap>::type>
2909 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR typename _R::mask_type
2910 static_simd_cast(const simd_mask<_Up, _Ap>& __x)
2911 {
2912 using _RM = typename _R::mask_type;
2913 return {__private_init, _RM::abi_type::_MaskImpl::template _S_convert<
2914 typename _RM::simd_type::value_type>(__x)};
2915 }
2916} // namespace __proposed
2917
2918// simd_cast {{{2
2919template <typename _Tp, typename _Up, typename _Ap,
2920 typename _To = __value_type_or_identity_t<_Tp>>
2921 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR auto
2922 simd_cast(const simd<_ValuePreserving<_Up, _To>, _Ap>& __x)
2923 -> decltype(static_simd_cast<_Tp>(__x))
2924 { return static_simd_cast<_Tp>(__x); }
2925
2926namespace __proposed {
2927template <typename _Tp, typename _Up, typename _Ap,
2928 typename _To = __value_type_or_identity_t<_Tp>>
2929 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR auto
2930 simd_cast(const simd_mask<_ValuePreserving<_Up, _To>, _Ap>& __x)
2931 -> decltype(static_simd_cast<_Tp>(__x))
2932 { return static_simd_cast<_Tp>(__x); }
2933} // namespace __proposed
2934
2935// }}}2
2936// resizing_simd_cast {{{
2937namespace __proposed {
2938/* Proposed spec:
2939
2940template <class T, class U, class Abi>
2941T resizing_simd_cast(const simd<U, Abi>& x)
2942
2943p1 Constraints:
2944 - is_simd_v<T> is true and
2945 - T::value_type is the same type as U
2946
2947p2 Returns:
2948 A simd object with the i^th element initialized to x[i] for all i in the
2949 range of [0, min(T::size(), simd_size_v<U, Abi>)). If T::size() is larger
2950 than simd_size_v<U, Abi>, the remaining elements are value-initialized.
2951
2952template <class T, class U, class Abi>
2953T resizing_simd_cast(const simd_mask<U, Abi>& x)
2954
2955p1 Constraints: is_simd_mask_v<T> is true
2956
2957p2 Returns:
2958 A simd_mask object with the i^th element initialized to x[i] for all i in
2959the range of [0, min(T::size(), simd_size_v<U, Abi>)). If T::size() is larger
2960 than simd_size_v<U, Abi>, the remaining elements are initialized to false.
2961
2962 */
2963
2964template <typename _Tp, typename _Up, typename _Ap>
2965 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR enable_if_t<
2966 conjunction_v<is_simd<_Tp>, is_same<typename _Tp::value_type, _Up>>, _Tp>
2967 resizing_simd_cast(const simd<_Up, _Ap>& __x)
2968 {
2969 if constexpr (is_same_v<typename _Tp::abi_type, _Ap>)
2970 return __x;
2971 else if constexpr (simd_size_v<_Up, _Ap> == 1)
2972 {
2973 _Tp __r{};
2974 __r[0] = __x[0];
2975 return __r;
2976 }
2977 else if constexpr (_Tp::size() == 1)
2978 return __x[0];
2979 else if constexpr (sizeof(_Tp) == sizeof(__x)
2980 && !__is_fixed_size_abi_v<_Ap>)
2981 return {__private_init,
2982 __vector_bitcast<typename _Tp::value_type, _Tp::size()>(
2983 _Ap::_S_masked(__data(__x))._M_data)};
2984 else
2985 {
2986 _Tp __r{};
2987 __builtin_memcpy(&__data(__r), &__data(__x),
2988 sizeof(_Up)
2989 * std::min(_Tp::size(), simd_size_v<_Up, _Ap>));
2990 return __r;
2991 }
2992 }
2993
2994template <typename _Tp, typename _Up, typename _Ap>
2995 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR
2996 enable_if_t<is_simd_mask_v<_Tp>, _Tp>
2997 resizing_simd_cast(const simd_mask<_Up, _Ap>& __x)
2998 {
2999 return {__private_init, _Tp::abi_type::_MaskImpl::template _S_convert<
3000 typename _Tp::simd_type::value_type>(__x)};
3001 }
3002} // namespace __proposed
3003
3004// }}}
3005// to_fixed_size {{{2
3006template <typename _Tp, int _Np>
3007 _GLIBCXX_SIMD_INTRINSIC fixed_size_simd<_Tp, _Np>
3008 to_fixed_size(const fixed_size_simd<_Tp, _Np>& __x)
3009 { return __x; }
3010
3011template <typename _Tp, int _Np>
3012 _GLIBCXX_SIMD_INTRINSIC fixed_size_simd_mask<_Tp, _Np>
3013 to_fixed_size(const fixed_size_simd_mask<_Tp, _Np>& __x)
3014 { return __x; }
3015
3016template <typename _Tp, typename _Ap>
3017 _GLIBCXX_SIMD_INTRINSIC auto
3018 to_fixed_size(const simd<_Tp, _Ap>& __x)
3019 {
3020 return simd<_Tp, simd_abi::fixed_size<simd_size_v<_Tp, _Ap>>>([&__x](
3021 auto __i) constexpr { return __x[__i]; });
3022 }
3023
3024template <typename _Tp, typename _Ap>
3025 _GLIBCXX_SIMD_INTRINSIC auto
3026 to_fixed_size(const simd_mask<_Tp, _Ap>& __x)
3027 {
3028 constexpr int _Np = simd_mask<_Tp, _Ap>::size();
3029 fixed_size_simd_mask<_Tp, _Np> __r;
3030 __execute_n_times<_Np>([&](auto __i) constexpr { __r[__i] = __x[__i]; });
3031 return __r;
3032 }
3033
3034// to_native {{{2
3035template <typename _Tp, int _Np>
3036 _GLIBCXX_SIMD_INTRINSIC
3037 enable_if_t<(_Np == native_simd<_Tp>::size()), native_simd<_Tp>>
3038 to_native(const fixed_size_simd<_Tp, _Np>& __x)
3039 {
3040 alignas(memory_alignment_v<native_simd<_Tp>>) _Tp __mem[_Np];
3041 __x.copy_to(__mem, vector_aligned);
3042 return {__mem, vector_aligned};
3043 }
3044
3045template <typename _Tp, size_t _Np>
3046 _GLIBCXX_SIMD_INTRINSIC
3047 enable_if_t<(_Np == native_simd_mask<_Tp>::size()), native_simd_mask<_Tp>>
3048 to_native(const fixed_size_simd_mask<_Tp, _Np>& __x)
3049 {
3050 return native_simd_mask<_Tp>([&](auto __i) constexpr { return __x[__i]; });
3051 }
3052
3053// to_compatible {{{2
3054template <typename _Tp, size_t _Np>
3055 _GLIBCXX_SIMD_INTRINSIC enable_if_t<(_Np == simd<_Tp>::size()), simd<_Tp>>
3056 to_compatible(const simd<_Tp, simd_abi::fixed_size<_Np>>& __x)
3057 {
3058 alignas(memory_alignment_v<simd<_Tp>>) _Tp __mem[_Np];
3059 __x.copy_to(__mem, vector_aligned);
3060 return {__mem, vector_aligned};
3061 }
3062
3063template <typename _Tp, size_t _Np>
3064 _GLIBCXX_SIMD_INTRINSIC
3065 enable_if_t<(_Np == simd_mask<_Tp>::size()), simd_mask<_Tp>>
3066 to_compatible(const simd_mask<_Tp, simd_abi::fixed_size<_Np>>& __x)
3067 { return simd_mask<_Tp>([&](auto __i) constexpr { return __x[__i]; }); }
3068
3069// masked assignment [simd_mask.where] {{{1
3070
3071// where_expression {{{1
3072// const_where_expression<M, T> {{{2
3073template <typename _M, typename _Tp>
3074 class const_where_expression
3075 {
3076 using _V = _Tp;
3077 static_assert(is_same_v<_V, __remove_cvref_t<_Tp>>);
3078
3079 struct _Wrapper { using value_type = _V; };
3080
3081 protected:
3082 using _Impl = typename _V::_Impl;
3083
3084 using value_type =
3085 typename conditional_t<is_arithmetic_v<_V>, _Wrapper, _V>::value_type;
3086
3087 _GLIBCXX_SIMD_INTRINSIC friend const _M&
3088 __get_mask(const const_where_expression& __x)
3089 { return __x._M_k; }
3090
3091 _GLIBCXX_SIMD_INTRINSIC friend const _Tp&
3092 __get_lvalue(const const_where_expression& __x)
3093 { return __x._M_value; }
3094
3095 const _M& _M_k;
3096 _Tp& _M_value;
3097
3098 public:
3099 const_where_expression(const const_where_expression&) = delete;
3100 const_where_expression& operator=(const const_where_expression&) = delete;
3101
3102 _GLIBCXX_SIMD_INTRINSIC const_where_expression(const _M& __kk, const _Tp& dd)
3103 : _M_k(__kk), _M_value(const_cast<_Tp&>(dd)) {}
3104
3105 _GLIBCXX_SIMD_INTRINSIC _V
3106 operator-() const&&
3107 {
3108 return {__private_init,
3109 _Impl::template _S_masked_unary<negate>(__data(_M_k),
3110 __data(_M_value))};
3111 }
3112
3113 template <typename _Up, typename _Flags>
3114 [[nodiscard]] _GLIBCXX_SIMD_INTRINSIC _V
3115 copy_from(const _LoadStorePtr<_Up, value_type>* __mem, _Flags) const&&
3116 {
3117 return {__private_init,
3118 _Impl::_S_masked_load(__data(_M_value), __data(_M_k),
3119 _Flags::template _S_apply<_V>(__mem))};
3120 }
3121
3122 template <typename _Up, typename _Flags>
3123 _GLIBCXX_SIMD_INTRINSIC void
3124 copy_to(_LoadStorePtr<_Up, value_type>* __mem, _Flags) const&&
3125 {
3126 _Impl::_S_masked_store(__data(_M_value),
3127 _Flags::template _S_apply<_V>(__mem),
3128 __data(_M_k));
3129 }
3130 };
3131
3132// const_where_expression<bool, T> {{{2
3133template <typename _Tp>
3134 class const_where_expression<bool, _Tp>
3135 {
3136 using _M = bool;
3137 using _V = _Tp;
3138
3139 static_assert(is_same_v<_V, __remove_cvref_t<_Tp>>);
3140
3141 struct _Wrapper { using value_type = _V; };
3142
3143 protected:
3144 using value_type =
3145 typename conditional_t<is_arithmetic_v<_V>, _Wrapper, _V>::value_type;
3146
3147 _GLIBCXX_SIMD_INTRINSIC friend const _M&
3148 __get_mask(const const_where_expression& __x)
3149 { return __x._M_k; }
3150
3151 _GLIBCXX_SIMD_INTRINSIC friend const _Tp&
3152 __get_lvalue(const const_where_expression& __x)
3153 { return __x._M_value; }
3154
3155 const bool _M_k;
3156 _Tp& _M_value;
3157
3158 public:
3159 const_where_expression(const const_where_expression&) = delete;
3160 const_where_expression& operator=(const const_where_expression&) = delete;
3161
3162 _GLIBCXX_SIMD_INTRINSIC const_where_expression(const bool __kk, const _Tp& dd)
3163 : _M_k(__kk), _M_value(const_cast<_Tp&>(dd)) {}
3164
3165 _GLIBCXX_SIMD_INTRINSIC _V operator-() const&&
3166 { return _M_k ? -_M_value : _M_value; }
3167
3168 template <typename _Up, typename _Flags>
3169 [[nodiscard]] _GLIBCXX_SIMD_INTRINSIC _V
3170 copy_from(const _LoadStorePtr<_Up, value_type>* __mem, _Flags) const&&
3171 { return _M_k ? static_cast<_V>(__mem[0]) : _M_value; }
3172
3173 template <typename _Up, typename _Flags>
3174 _GLIBCXX_SIMD_INTRINSIC void
3175 copy_to(_LoadStorePtr<_Up, value_type>* __mem, _Flags) const&&
3176 {
3177 if (_M_k)
3178 __mem[0] = _M_value;
3179 }
3180 };
3181
3182// where_expression<M, T> {{{2
3183template <typename _M, typename _Tp>
3184 class where_expression : public const_where_expression<_M, _Tp>
3185 {
3186 using _Impl = typename const_where_expression<_M, _Tp>::_Impl;
3187
3188 static_assert(!is_const<_Tp>::value,
3189 "where_expression may only be instantiated with __a non-const "
3190 "_Tp parameter");
3191
3192 using typename const_where_expression<_M, _Tp>::value_type;
3193 using const_where_expression<_M, _Tp>::_M_k;
3194 using const_where_expression<_M, _Tp>::_M_value;
3195
3196 static_assert(
3197 is_same<typename _M::abi_type, typename _Tp::abi_type>::value, "");
3198 static_assert(_M::size() == _Tp::size(), "");
3199
3200 _GLIBCXX_SIMD_INTRINSIC friend _Tp& __get_lvalue(where_expression& __x)
3201 { return __x._M_value; }
3202
3203 public:
3204 where_expression(const where_expression&) = delete;
3205 where_expression& operator=(const where_expression&) = delete;
3206
3207 _GLIBCXX_SIMD_INTRINSIC where_expression(const _M& __kk, _Tp& dd)
3208 : const_where_expression<_M, _Tp>(__kk, dd) {}
3209
3210 template <typename _Up>
3211 _GLIBCXX_SIMD_INTRINSIC void operator=(_Up&& __x) &&
3212 {
3213 _Impl::_S_masked_assign(__data(_M_k), __data(_M_value),
3214 __to_value_type_or_member_type<_Tp>(
3215 static_cast<_Up&&>(__x)));
3216 }
3217
3218#define _GLIBCXX_SIMD_OP_(__op, __name) \
3219 template <typename _Up> \
3220 _GLIBCXX_SIMD_INTRINSIC void operator __op##=(_Up&& __x)&& \
3221 { \
3222 _Impl::template _S_masked_cassign( \
3223 __data(_M_k), __data(_M_value), \
3224 __to_value_type_or_member_type<_Tp>(static_cast<_Up&&>(__x)), \
3225 [](auto __impl, auto __lhs, auto __rhs) constexpr { \
3226 return __impl.__name(__lhs, __rhs); \
3227 }); \
3228 } \
3229 static_assert(true)
3230 _GLIBCXX_SIMD_OP_(+, _S_plus);
3231 _GLIBCXX_SIMD_OP_(-, _S_minus);
3232 _GLIBCXX_SIMD_OP_(*, _S_multiplies);
3233 _GLIBCXX_SIMD_OP_(/, _S_divides);
3234 _GLIBCXX_SIMD_OP_(%, _S_modulus);
3235 _GLIBCXX_SIMD_OP_(&, _S_bit_and);
3236 _GLIBCXX_SIMD_OP_(|, _S_bit_or);
3237 _GLIBCXX_SIMD_OP_(^, _S_bit_xor);
3238 _GLIBCXX_SIMD_OP_(<<, _S_shift_left);
3239 _GLIBCXX_SIMD_OP_(>>, _S_shift_right);
3240#undef _GLIBCXX_SIMD_OP_
3241
3242 _GLIBCXX_SIMD_INTRINSIC void operator++() &&
3243 {
3244 __data(_M_value)
3245 = _Impl::template _S_masked_unary<__increment>(__data(_M_k),
3246 __data(_M_value));
3247 }
3248
3249 _GLIBCXX_SIMD_INTRINSIC void operator++(int) &&
3250 {
3251 __data(_M_value)
3252 = _Impl::template _S_masked_unary<__increment>(__data(_M_k),
3253 __data(_M_value));
3254 }
3255
3256 _GLIBCXX_SIMD_INTRINSIC void operator--() &&
3257 {
3258 __data(_M_value)
3259 = _Impl::template _S_masked_unary<__decrement>(__data(_M_k),
3260 __data(_M_value));
3261 }
3262
3263 _GLIBCXX_SIMD_INTRINSIC void operator--(int) &&
3264 {
3265 __data(_M_value)
3266 = _Impl::template _S_masked_unary<__decrement>(__data(_M_k),
3267 __data(_M_value));
3268 }
3269
3270 // intentionally hides const_where_expression::copy_from
3271 template <typename _Up, typename _Flags>
3272 _GLIBCXX_SIMD_INTRINSIC void
3273 copy_from(const _LoadStorePtr<_Up, value_type>* __mem, _Flags) &&
3274 {
3275 __data(_M_value)
3276 = _Impl::_S_masked_load(__data(_M_value), __data(_M_k),
3277 _Flags::template _S_apply<_Tp>(__mem));
3278 }
3279 };
3280
3281// where_expression<bool, T> {{{2
3282template <typename _Tp>
3283 class where_expression<bool, _Tp> : public const_where_expression<bool, _Tp>
3284 {
3285 using _M = bool;
3286 using typename const_where_expression<_M, _Tp>::value_type;
3287 using const_where_expression<_M, _Tp>::_M_k;
3288 using const_where_expression<_M, _Tp>::_M_value;
3289
3290 public:
3291 where_expression(const where_expression&) = delete;
3292 where_expression& operator=(const where_expression&) = delete;
3293
3294 _GLIBCXX_SIMD_INTRINSIC where_expression(const _M& __kk, _Tp& dd)
3295 : const_where_expression<_M, _Tp>(__kk, dd) {}
3296
3297#define _GLIBCXX_SIMD_OP_(__op) \
3298 template <typename _Up> \
3299 _GLIBCXX_SIMD_INTRINSIC void operator __op(_Up&& __x)&& \
3300 { if (_M_k) _M_value __op static_cast<_Up&&>(__x); }
3301
3302 _GLIBCXX_SIMD_OP_(=)
3303 _GLIBCXX_SIMD_OP_(+=)
3304 _GLIBCXX_SIMD_OP_(-=)
3305 _GLIBCXX_SIMD_OP_(*=)
3306 _GLIBCXX_SIMD_OP_(/=)
3307 _GLIBCXX_SIMD_OP_(%=)
3308 _GLIBCXX_SIMD_OP_(&=)
3309 _GLIBCXX_SIMD_OP_(|=)
3310 _GLIBCXX_SIMD_OP_(^=)
3311 _GLIBCXX_SIMD_OP_(<<=)
3312 _GLIBCXX_SIMD_OP_(>>=)
3313 #undef _GLIBCXX_SIMD_OP_
3314
3315 _GLIBCXX_SIMD_INTRINSIC void operator++() &&
3316 { if (_M_k) ++_M_value; }
3317
3318 _GLIBCXX_SIMD_INTRINSIC void operator++(int) &&
3319 { if (_M_k) ++_M_value; }
3320
3321 _GLIBCXX_SIMD_INTRINSIC void operator--() &&
3322 { if (_M_k) --_M_value; }
3323
3324 _GLIBCXX_SIMD_INTRINSIC void operator--(int) &&
3325 { if (_M_k) --_M_value; }
3326
3327 // intentionally hides const_where_expression::copy_from
3328 template <typename _Up, typename _Flags>
3329 _GLIBCXX_SIMD_INTRINSIC void
3330 copy_from(const _LoadStorePtr<_Up, value_type>* __mem, _Flags) &&
3331 { if (_M_k) _M_value = __mem[0]; }
3332 };
3333
3334// where {{{1
3335template <typename _Tp, typename _Ap>
3336 _GLIBCXX_SIMD_INTRINSIC where_expression<simd_mask<_Tp, _Ap>, simd<_Tp, _Ap>>
3337 where(const typename simd<_Tp, _Ap>::mask_type& __k, simd<_Tp, _Ap>& __value)
3338 { return {__k, __value}; }
3339
3340template <typename _Tp, typename _Ap>
3341 _GLIBCXX_SIMD_INTRINSIC
3342 const_where_expression<simd_mask<_Tp, _Ap>, simd<_Tp, _Ap>>
3343 where(const typename simd<_Tp, _Ap>::mask_type& __k,
3344 const simd<_Tp, _Ap>& __value)
3345 { return {__k, __value}; }
3346
3347template <typename _Tp, typename _Ap>
3348 _GLIBCXX_SIMD_INTRINSIC
3349 where_expression<simd_mask<_Tp, _Ap>, simd_mask<_Tp, _Ap>>
3350 where(const remove_const_t<simd_mask<_Tp, _Ap>>& __k,
3351 simd_mask<_Tp, _Ap>& __value)
3352 { return {__k, __value}; }
3353
3354template <typename _Tp, typename _Ap>
3355 _GLIBCXX_SIMD_INTRINSIC
3356 const_where_expression<simd_mask<_Tp, _Ap>, simd_mask<_Tp, _Ap>>
3357 where(const remove_const_t<simd_mask<_Tp, _Ap>>& __k,
3358 const simd_mask<_Tp, _Ap>& __value)
3359 { return {__k, __value}; }
3360
3361template <typename _Tp>
3362 _GLIBCXX_SIMD_INTRINSIC where_expression<bool, _Tp>
3363 where(_ExactBool __k, _Tp& __value)
3364 { return {__k, __value}; }
3365
3366template <typename _Tp>
3367 _GLIBCXX_SIMD_INTRINSIC const_where_expression<bool, _Tp>
3368 where(_ExactBool __k, const _Tp& __value)
3369 { return {__k, __value}; }
3370
3371 template <typename _Tp, typename _Ap>
3372 void where(bool __k, simd<_Tp, _Ap>& __value) = delete;
3373
3374 template <typename _Tp, typename _Ap>
3375 void where(bool __k, const simd<_Tp, _Ap>& __value) = delete;
3376
3377// proposed mask iterations {{{1
3378namespace __proposed {
3379template <size_t _Np>
3380 class where_range
3381 {
3382 const bitset<_Np> __bits;
3383
3384 public:
3385 where_range(bitset<_Np> __b) : __bits(__b) {}
3386
3387 class iterator
3388 {
3389 size_t __mask;
3390 size_t __bit;
3391
3392 _GLIBCXX_SIMD_INTRINSIC void __next_bit()
3393 { __bit = __builtin_ctzl(__mask); }
3394
3395 _GLIBCXX_SIMD_INTRINSIC void __reset_lsb()
3396 {
3397 // 01100100 - 1 = 01100011
3398 __mask &= (__mask - 1);
3399 // __asm__("btr %1,%0" : "+r"(__mask) : "r"(__bit));
3400 }
3401
3402 public:
3403 iterator(decltype(__mask) __m) : __mask(__m) { __next_bit(); }
3404 iterator(const iterator&) = default;
3405 iterator(iterator&&) = default;
3406
3407 _GLIBCXX_SIMD_ALWAYS_INLINE size_t operator->() const
3408 { return __bit; }
3409
3410 _GLIBCXX_SIMD_ALWAYS_INLINE size_t operator*() const
3411 { return __bit; }
3412
3413 _GLIBCXX_SIMD_ALWAYS_INLINE iterator& operator++()
3414 {
3415 __reset_lsb();
3416 __next_bit();
3417 return *this;
3418 }
3419
3420 _GLIBCXX_SIMD_ALWAYS_INLINE iterator operator++(int)
3421 {
3422 iterator __tmp = *this;
3423 __reset_lsb();
3424 __next_bit();
3425 return __tmp;
3426 }
3427
3428 _GLIBCXX_SIMD_ALWAYS_INLINE bool operator==(const iterator& __rhs) const
3429 { return __mask == __rhs.__mask; }
3430
3431 _GLIBCXX_SIMD_ALWAYS_INLINE bool operator!=(const iterator& __rhs) const
3432 { return __mask != __rhs.__mask; }
3433 };
3434
3435 iterator begin() const
3436 { return __bits.to_ullong(); }
3437
3438 iterator end() const
3439 { return 0; }
3440 };
3441
3442template <typename _Tp, typename _Ap>
3443 where_range<simd_size_v<_Tp, _Ap>>
3444 where(const simd_mask<_Tp, _Ap>& __k)
3445 { return __k.__to_bitset(); }
3446
3447} // namespace __proposed
3448
3449// }}}1
3450// reductions [simd.reductions] {{{1
3451template <typename _Tp, typename _Abi, typename _BinaryOperation = plus<>>
3452 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR _Tp
3453 reduce(const simd<_Tp, _Abi>& __v,
3454 _BinaryOperation __binary_op = _BinaryOperation())
3455 { return _Abi::_SimdImpl::_S_reduce(__v, __binary_op); }
3456
3457template <typename _M, typename _V, typename _BinaryOperation = plus<>>
3458 _GLIBCXX_SIMD_INTRINSIC typename _V::value_type
3459 reduce(const const_where_expression<_M, _V>& __x,
3460 typename _V::value_type __identity_element,
3461 _BinaryOperation __binary_op)
3462 {
3463 if (__builtin_expect(none_of(__get_mask(__x)), false))
3464 return __identity_element;
3465
3466 _V __tmp = __identity_element;
3467 _V::_Impl::_S_masked_assign(__data(__get_mask(__x)), __data(__tmp),
3468 __data(__get_lvalue(__x)));
3469 return reduce(__tmp, __binary_op);
3470 }
3471
3472template <typename _M, typename _V>
3473 _GLIBCXX_SIMD_INTRINSIC typename _V::value_type
3474 reduce(const const_where_expression<_M, _V>& __x, plus<> __binary_op = {})
3475 { return reduce(__x, 0, __binary_op); }
3476
3477template <typename _M, typename _V>
3478 _GLIBCXX_SIMD_INTRINSIC typename _V::value_type
3479 reduce(const const_where_expression<_M, _V>& __x, multiplies<> __binary_op)
3480 { return reduce(__x, 1, __binary_op); }
3481
3482template <typename _M, typename _V>
3483 _GLIBCXX_SIMD_INTRINSIC typename _V::value_type
3484 reduce(const const_where_expression<_M, _V>& __x, bit_and<> __binary_op)
3485 { return reduce(__x, ~typename _V::value_type(), __binary_op); }
3486
3487template <typename _M, typename _V>
3488 _GLIBCXX_SIMD_INTRINSIC typename _V::value_type
3489 reduce(const const_where_expression<_M, _V>& __x, bit_or<> __binary_op)
3490 { return reduce(__x, 0, __binary_op); }
3491
3492template <typename _M, typename _V>
3493 _GLIBCXX_SIMD_INTRINSIC typename _V::value_type
3494 reduce(const const_where_expression<_M, _V>& __x, bit_xor<> __binary_op)
3495 { return reduce(__x, 0, __binary_op); }
3496
3497template <typename _Tp, typename _Abi>
3498 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR _Tp
3499 hmin(const simd<_Tp, _Abi>& __v) noexcept
3500 {
3501 return _Abi::_SimdImpl::_S_reduce(__v, __detail::_Minimum());
3502 }
3503
3504template <typename _Tp, typename _Abi>
3505 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR _Tp
3506 hmax(const simd<_Tp, _Abi>& __v) noexcept
3507 {
3508 return _Abi::_SimdImpl::_S_reduce(__v, __detail::_Maximum());
3509 }
3510
3511template <typename _M, typename _V>
3512 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR
3513 typename _V::value_type
3514 hmin(const const_where_expression<_M, _V>& __x) noexcept
3515 {
3516 using _Tp = typename _V::value_type;
3517 constexpr _Tp __id_elem =
3518#ifdef __FINITE_MATH_ONLY__
3519 __finite_max_v<_Tp>;
3520#else
3521 __value_or<__infinity, _Tp>(__finite_max_v<_Tp>);
3522#endif
3523 _V __tmp = __id_elem;
3524 _V::_Impl::_S_masked_assign(__data(__get_mask(__x)), __data(__tmp),
3525 __data(__get_lvalue(__x)));
3526 return _V::abi_type::_SimdImpl::_S_reduce(__tmp, __detail::_Minimum());
3527 }
3528
3529template <typename _M, typename _V>
3530 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR
3531 typename _V::value_type
3532 hmax(const const_where_expression<_M, _V>& __x) noexcept
3533 {
3534 using _Tp = typename _V::value_type;
3535 constexpr _Tp __id_elem =
3536#ifdef __FINITE_MATH_ONLY__
3537 __finite_min_v<_Tp>;
3538#else
3539 [] {
3540 if constexpr (__value_exists_v<__infinity, _Tp>)
3541 return -__infinity_v<_Tp>;
3542 else
3543 return __finite_min_v<_Tp>;
3544 }();
3545#endif
3546 _V __tmp = __id_elem;
3547 _V::_Impl::_S_masked_assign(__data(__get_mask(__x)), __data(__tmp),
3548 __data(__get_lvalue(__x)));
3549 return _V::abi_type::_SimdImpl::_S_reduce(__tmp, __detail::_Maximum());
3550 }
3551
3552// }}}1
3553// algorithms [simd.alg] {{{
3554template <typename _Tp, typename _Ap>
3555 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR simd<_Tp, _Ap>
3556 min(const simd<_Tp, _Ap>& __a, const simd<_Tp, _Ap>& __b)
3557 { return {__private_init, _Ap::_SimdImpl::_S_min(__data(__a), __data(__b))}; }
3558
3559template <typename _Tp, typename _Ap>
3560 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR simd<_Tp, _Ap>
3561 max(const simd<_Tp, _Ap>& __a, const simd<_Tp, _Ap>& __b)
3562 { return {__private_init, _Ap::_SimdImpl::_S_max(__data(__a), __data(__b))}; }
3563
3564template <typename _Tp, typename _Ap>
3565 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR
3566 pair<simd<_Tp, _Ap>, simd<_Tp, _Ap>>
3567 minmax(const simd<_Tp, _Ap>& __a, const simd<_Tp, _Ap>& __b)
3568 {
3569 const auto pair_of_members
3570 = _Ap::_SimdImpl::_S_minmax(__data(__a), __data(__b));
3571 return {simd<_Tp, _Ap>(__private_init, pair_of_members.first),
3572 simd<_Tp, _Ap>(__private_init, pair_of_members.second)};
3573 }
3574
3575template <typename _Tp, typename _Ap>
3576 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR simd<_Tp, _Ap>
3577 clamp(const simd<_Tp, _Ap>& __v, const simd<_Tp, _Ap>& __lo,
3578 const simd<_Tp, _Ap>& __hi)
3579 {
3580 using _Impl = typename _Ap::_SimdImpl;
3581 return {__private_init,
3582 _Impl::_S_min(__data(__hi),
3583 _Impl::_S_max(__data(__lo), __data(__v)))};
3584 }
3585
3586// }}}
3587
3588template <size_t... _Sizes, typename _Tp, typename _Ap,
3589 typename = enable_if_t<((_Sizes + ...) == simd<_Tp, _Ap>::size())>>
3590 inline tuple<simd<_Tp, simd_abi::deduce_t<_Tp, _Sizes>>...>
3591 split(const simd<_Tp, _Ap>&);
3592
3593// __extract_part {{{
3594template <int _Index, int _Total, int _Combine = 1, typename _Tp, size_t _Np>
3595 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_CONST
3596 _SimdWrapper<_Tp, _Np / _Total * _Combine>
3597 __extract_part(const _SimdWrapper<_Tp, _Np> __x);
3598
3599template <int _Index, int _Parts, int _Combine = 1, typename _Tp, typename _A0,
3600 typename... _As>
3601 _GLIBCXX_SIMD_INTRINSIC auto
3602 __extract_part(const _SimdTuple<_Tp, _A0, _As...>& __x);
3603
3604// }}}
3605// _SizeList {{{
3606template <size_t _V0, size_t... _Values>
3607 struct _SizeList
3608 {
3609 template <size_t _I>
3610 static constexpr size_t _S_at(_SizeConstant<_I> = {})
3611 {
3612 if constexpr (_I == 0)
3613 return _V0;
3614 else
3615 return _SizeList<_Values...>::template _S_at<_I - 1>();
3616 }
3617
3618 template <size_t _I>
3619 static constexpr auto _S_before(_SizeConstant<_I> = {})
3620 {
3621 if constexpr (_I == 0)
3622 return _SizeConstant<0>();
3623 else
3624 return _SizeConstant<
3625 _V0 + _SizeList<_Values...>::template _S_before<_I - 1>()>();
3626 }
3627
3628 template <size_t _Np>
3629 static constexpr auto _S_pop_front(_SizeConstant<_Np> = {})
3630 {
3631 if constexpr (_Np == 0)
3632 return _SizeList();
3633 else
3634 return _SizeList<_Values...>::template _S_pop_front<_Np - 1>();
3635 }
3636 };
3637
3638// }}}
3639// __extract_center {{{
3640template <typename _Tp, size_t _Np>
3641 _GLIBCXX_SIMD_INTRINSIC _SimdWrapper<_Tp, _Np / 2>
3642 __extract_center(_SimdWrapper<_Tp, _Np> __x)
3643 {
3644 static_assert(_Np >= 4);
3645 static_assert(_Np % 4 == 0); // x0 - x1 - x2 - x3 -> return {x1, x2}
3646#if _GLIBCXX_SIMD_X86INTRIN // {{{
3647 if constexpr (__have_avx512f && sizeof(_Tp) * _Np == 64)
3648 {
3649 const auto __intrin = __to_intrin(__x);
3650 if constexpr (is_integral_v<_Tp>)
3651 return __vector_bitcast<_Tp>(_mm512_castsi512_si256(
3652 _mm512_shuffle_i32x4(__intrin, __intrin,
3653 1 + 2 * 0x4 + 2 * 0x10 + 3 * 0x40)));
3654 else if constexpr (sizeof(_Tp) == 4)
3655 return __vector_bitcast<_Tp>(_mm512_castps512_ps256(
3656 _mm512_shuffle_f32x4(__intrin, __intrin,
3657 1 + 2 * 0x4 + 2 * 0x10 + 3 * 0x40)));
3658 else if constexpr (sizeof(_Tp) == 8)
3659 return __vector_bitcast<_Tp>(_mm512_castpd512_pd256(
3660 _mm512_shuffle_f64x2(__intrin, __intrin,
3661 1 + 2 * 0x4 + 2 * 0x10 + 3 * 0x40)));
3662 else
3663 __assert_unreachable<_Tp>();
3664 }
3665 else if constexpr (sizeof(_Tp) * _Np == 32 && is_floating_point_v<_Tp>)
3666 return __vector_bitcast<_Tp>(
3667 _mm_shuffle_pd(__lo128(__vector_bitcast<double>(__x)),
3668 __hi128(__vector_bitcast<double>(__x)), 1));
3669 else if constexpr (sizeof(__x) == 32 && sizeof(_Tp) * _Np <= 32)
3670 return __vector_bitcast<_Tp>(
3671 _mm_alignr_epi8(__hi128(__vector_bitcast<_LLong>(__x)),
3672 __lo128(__vector_bitcast<_LLong>(__x)),
3673 sizeof(_Tp) * _Np / 4));
3674 else
3675#endif // _GLIBCXX_SIMD_X86INTRIN }}}
3676 {
3677 __vector_type_t<_Tp, _Np / 2> __r;
3678 __builtin_memcpy(&__r,
3679 reinterpret_cast<const char*>(&__x)
3680 + sizeof(_Tp) * _Np / 4,
3681 sizeof(_Tp) * _Np / 2);
3682 return __r;
3683 }
3684 }
3685
3686template <typename _Tp, typename _A0, typename... _As>
3687 _GLIBCXX_SIMD_INTRINSIC
3688 _SimdWrapper<_Tp, _SimdTuple<_Tp, _A0, _As...>::_S_size() / 2>
3689 __extract_center(const _SimdTuple<_Tp, _A0, _As...>& __x)
3690 {
3691 if constexpr (sizeof...(_As) == 0)
3692 return __extract_center(__x.first);
3693 else
3694 return __extract_part<1, 4, 2>(__x);
3695 }
3696
3697// }}}
3698// __split_wrapper {{{
3699template <size_t... _Sizes, typename _Tp, typename... _As>
3700 auto
3701 __split_wrapper(_SizeList<_Sizes...>, const _SimdTuple<_Tp, _As...>& __x)
3702 {
3703 return split<_Sizes...>(
3704 fixed_size_simd<_Tp, _SimdTuple<_Tp, _As...>::_S_size()>(__private_init,
3705 __x));
3706 }
3707
3708// }}}
3709
3710// split<simd>(simd) {{{
3711template <typename _V, typename _Ap,
3712 size_t _Parts = simd_size_v<typename _V::value_type, _Ap> / _V::size()>
3713 enable_if_t<simd_size_v<typename _V::value_type, _Ap> == _Parts * _V::size()
3714 && is_simd_v<_V>, array<_V, _Parts>>
3715 split(const simd<typename _V::value_type, _Ap>& __x)
3716 {
3717 using _Tp = typename _V::value_type;
3718 if constexpr (_Parts == 1)
3719 {
3720 return {simd_cast<_V>(__x)};
3721 }
3722 else if (__x._M_is_constprop())
3723 {
3724 return __generate_from_n_evaluations<_Parts, array<_V, _Parts>>([&](
3725 auto __i) constexpr {
3726 return _V([&](auto __j) constexpr {
3727 return __x[__i * _V::size() + __j];
3728 });
3729 });
3730 }
3731 else if constexpr (
3732 __is_fixed_size_abi_v<_Ap>
3733 && (is_same_v<typename _V::abi_type, simd_abi::scalar>
3734 || (__is_fixed_size_abi_v<typename _V::abi_type>
3735 && sizeof(_V) == sizeof(_Tp) * _V::size() // _V doesn't have padding
3736 )))
3737 {
3738 // fixed_size -> fixed_size (w/o padding) or scalar
3739#ifdef _GLIBCXX_SIMD_USE_ALIASING_LOADS
3740 const __may_alias<_Tp>* const __element_ptr
3741 = reinterpret_cast<const __may_alias<_Tp>*>(&__data(__x));
3742 return __generate_from_n_evaluations<_Parts, array<_V, _Parts>>([&](
3743 auto __i) constexpr {
3744 return _V(__element_ptr + __i * _V::size(), vector_aligned);
3745 });
3746#else
3747 const auto& __xx = __data(__x);
3748 return __generate_from_n_evaluations<_Parts, array<_V, _Parts>>([&](
3749 auto __i) constexpr {
3750 [[maybe_unused]] constexpr size_t __offset
3751 = decltype(__i)::value * _V::size();
3752 return _V([&](auto __j) constexpr {
3753 constexpr _SizeConstant<__j + __offset> __k;
3754 return __xx[__k];
3755 });
3756 });
3757#endif
3758 }
3759 else if constexpr (is_same_v<typename _V::abi_type, simd_abi::scalar>)
3760 {
3761 // normally memcpy should work here as well
3762 return __generate_from_n_evaluations<_Parts, array<_V, _Parts>>([&](
3763 auto __i) constexpr { return __x[__i]; });
3764 }
3765 else
3766 {
3767 return __generate_from_n_evaluations<_Parts, array<_V, _Parts>>([&](
3768 auto __i) constexpr {
3769 if constexpr (__is_fixed_size_abi_v<typename _V::abi_type>)
3770 return _V([&](auto __j) constexpr {
3771 return __x[__i * _V::size() + __j];
3772 });
3773 else
3774 return _V(__private_init,
3775 __extract_part<decltype(__i)::value, _Parts>(__data(__x)));
3776 });
3777 }
3778 }
3779
3780// }}}
3781// split<simd_mask>(simd_mask) {{{
3782template <typename _V, typename _Ap,
3783 size_t _Parts
3784 = simd_size_v<typename _V::simd_type::value_type, _Ap> / _V::size()>
3785 enable_if_t<is_simd_mask_v<_V> && simd_size_v<typename
3786 _V::simd_type::value_type, _Ap> == _Parts * _V::size(), array<_V, _Parts>>
3787 split(const simd_mask<typename _V::simd_type::value_type, _Ap>& __x)
3788 {
3789 if constexpr (is_same_v<_Ap, typename _V::abi_type>)
3790 return {__x};
3791 else if constexpr (_Parts == 1)
3792 return {__proposed::static_simd_cast<_V>(__x)};
3793 else if constexpr (_Parts == 2 && __is_sse_abi<typename _V::abi_type>()
3794 && __is_avx_abi<_Ap>())
3795 return {_V(__private_init, __lo128(__data(__x))),
3796 _V(__private_init, __hi128(__data(__x)))};
3797 else if constexpr (_V::size() <= __CHAR_BIT__ * sizeof(_ULLong))
3798 {
3799 const bitset __bits = __x.__to_bitset();
3800 return __generate_from_n_evaluations<_Parts, array<_V, _Parts>>([&](
3801 auto __i) constexpr {
3802 constexpr size_t __offset = __i * _V::size();
3803 return _V(__bitset_init, (__bits >> __offset).to_ullong());
3804 });
3805 }
3806 else
3807 {
3808 return __generate_from_n_evaluations<_Parts, array<_V, _Parts>>([&](
3809 auto __i) constexpr {
3810 constexpr size_t __offset = __i * _V::size();
3811 return _V(
3812 __private_init, [&](auto __j) constexpr {
3813 return __x[__j + __offset];
3814 });
3815 });
3816 }
3817 }
3818
3819// }}}
3820// split<_Sizes...>(simd) {{{
3821template <size_t... _Sizes, typename _Tp, typename _Ap, typename>
3822 _GLIBCXX_SIMD_ALWAYS_INLINE
3823 tuple<simd<_Tp, simd_abi::deduce_t<_Tp, _Sizes>>...>
3824 split(const simd<_Tp, _Ap>& __x)
3825 {
3826 using _SL = _SizeList<_Sizes...>;
3827 using _Tuple = tuple<__deduced_simd<_Tp, _Sizes>...>;
3828 constexpr size_t _Np = simd_size_v<_Tp, _Ap>;
3829 constexpr size_t _N0 = _SL::template _S_at<0>();
3830 using _V = __deduced_simd<_Tp, _N0>;
3831
3832 if (__x._M_is_constprop())
3833 return __generate_from_n_evaluations<sizeof...(_Sizes), _Tuple>([&](
3834 auto __i) constexpr {
3835 using _Vi = __deduced_simd<_Tp, _SL::_S_at(__i)>;
3836 constexpr size_t __offset = _SL::_S_before(__i);
3837 return _Vi([&](auto __j) constexpr { return __x[__offset + __j]; });
3838 });
3839 else if constexpr (_Np == _N0)
3840 {
3841 static_assert(sizeof...(_Sizes) == 1);
3842 return {simd_cast<_V>(__x)};
3843 }
3844 else if constexpr // split from fixed_size, such that __x::first.size == _N0
3845 (__is_fixed_size_abi_v<
3846 _Ap> && __fixed_size_storage_t<_Tp, _Np>::_S_first_size == _N0)
3847 {
3848 static_assert(
3849 !__is_fixed_size_abi_v<typename _V::abi_type>,
3850 "How can <_Tp, _Np> be __a single _SimdTuple entry but __a "
3851 "fixed_size_simd "
3852 "when deduced?");
3853 // extract first and recurse (__split_wrapper is needed to deduce a new
3854 // _Sizes pack)
3855 return tuple_cat(make_tuple(_V(__private_init, __data(__x).first)),
3856 __split_wrapper(_SL::template _S_pop_front<1>(),
3857 __data(__x).second));
3858 }
3859 else if constexpr ((!is_same_v<simd_abi::scalar,
3860 simd_abi::deduce_t<_Tp, _Sizes>> && ...)
3861 && (!__is_fixed_size_abi_v<
3862 simd_abi::deduce_t<_Tp, _Sizes>> && ...))
3863 {
3864 if constexpr (((_Sizes * 2 == _Np) && ...))
3865 return {{__private_init, __extract_part<0, 2>(__data(__x))},
3866 {__private_init, __extract_part<1, 2>(__data(__x))}};
3867 else if constexpr (is_same_v<_SizeList<_Sizes...>,
3868 _SizeList<_Np / 3, _Np / 3, _Np / 3>>)
3869 return {{__private_init, __extract_part<0, 3>(__data(__x))},
3870 {__private_init, __extract_part<1, 3>(__data(__x))},
3871 {__private_init, __extract_part<2, 3>(__data(__x))}};
3872 else if constexpr (is_same_v<_SizeList<_Sizes...>,
3873 _SizeList<2 * _Np / 3, _Np / 3>>)
3874 return {{__private_init, __extract_part<0, 3, 2>(__data(__x))},
3875 {__private_init, __extract_part<2, 3>(__data(__x))}};
3876 else if constexpr (is_same_v<_SizeList<_Sizes...>,
3877 _SizeList<_Np / 3, 2 * _Np / 3>>)
3878 return {{__private_init, __extract_part<0, 3>(__data(__x))},
3879 {__private_init, __extract_part<1, 3, 2>(__data(__x))}};
3880 else if constexpr (is_same_v<_SizeList<_Sizes...>,
3881 _SizeList<_Np / 2, _Np / 4, _Np / 4>>)
3882 return {{__private_init, __extract_part<0, 2>(__data(__x))},
3883 {__private_init, __extract_part<2, 4>(__data(__x))},
3884 {__private_init, __extract_part<3, 4>(__data(__x))}};
3885 else if constexpr (is_same_v<_SizeList<_Sizes...>,
3886 _SizeList<_Np / 4, _Np / 4, _Np / 2>>)
3887 return {{__private_init, __extract_part<0, 4>(__data(__x))},
3888 {__private_init, __extract_part<1, 4>(__data(__x))},
3889 {__private_init, __extract_part<1, 2>(__data(__x))}};
3890 else if constexpr (is_same_v<_SizeList<_Sizes...>,
3891 _SizeList<_Np / 4, _Np / 2, _Np / 4>>)
3892 return {{__private_init, __extract_part<0, 4>(__data(__x))},
3893 {__private_init, __extract_center(__data(__x))},
3894 {__private_init, __extract_part<3, 4>(__data(__x))}};
3895 else if constexpr (((_Sizes * 4 == _Np) && ...))
3896 return {{__private_init, __extract_part<0, 4>(__data(__x))},
3897 {__private_init, __extract_part<1, 4>(__data(__x))},
3898 {__private_init, __extract_part<2, 4>(__data(__x))},
3899 {__private_init, __extract_part<3, 4>(__data(__x))}};
3900 // else fall through
3901 }
3902#ifdef _GLIBCXX_SIMD_USE_ALIASING_LOADS
3903 const __may_alias<_Tp>* const __element_ptr
3904 = reinterpret_cast<const __may_alias<_Tp>*>(&__x);
3905 return __generate_from_n_evaluations<sizeof...(_Sizes), _Tuple>([&](
3906 auto __i) constexpr {
3907 using _Vi = __deduced_simd<_Tp, _SL::_S_at(__i)>;
3908 constexpr size_t __offset = _SL::_S_before(__i);
3909 constexpr size_t __base_align = alignof(simd<_Tp, _Ap>);
3910 constexpr size_t __a
3911 = __base_align - ((__offset * sizeof(_Tp)) % __base_align);
3912 constexpr size_t __b = ((__a - 1) & __a) ^ __a;
3913 constexpr size_t __alignment = __b == 0 ? __a : __b;
3914 return _Vi(__element_ptr + __offset, overaligned<__alignment>);
3915 });
3916#else
3917 return __generate_from_n_evaluations<sizeof...(_Sizes), _Tuple>([&](
3918 auto __i) constexpr {
3919 using _Vi = __deduced_simd<_Tp, _SL::_S_at(__i)>;
3920 const auto& __xx = __data(__x);
3921 using _Offset = decltype(_SL::_S_before(__i));
3922 return _Vi([&](auto __j) constexpr {
3923 constexpr _SizeConstant<_Offset::value + __j> __k;
3924 return __xx[__k];
3925 });
3926 });
3927#endif
3928 }
3929
3930// }}}
3931
3932// __subscript_in_pack {{{
3933template <size_t _I, typename _Tp, typename _Ap, typename... _As>
3934 _GLIBCXX_SIMD_INTRINSIC constexpr _Tp
3935 __subscript_in_pack(const simd<_Tp, _Ap>& __x, const simd<_Tp, _As>&... __xs)
3936 {
3937 if constexpr (_I < simd_size_v<_Tp, _Ap>)
3938 return __x[_I];
3939 else
3940 return __subscript_in_pack<_I - simd_size_v<_Tp, _Ap>>(__xs...);
3941 }
3942
3943// }}}
3944// __store_pack_of_simd {{{
3945template <typename _Tp, typename _A0, typename... _As>
3946 _GLIBCXX_SIMD_INTRINSIC void
3947 __store_pack_of_simd(char* __mem, const simd<_Tp, _A0>& __x0,
3948 const simd<_Tp, _As>&... __xs)
3949 {
3950 constexpr size_t __n_bytes = sizeof(_Tp) * simd_size_v<_Tp, _A0>;
3951 __builtin_memcpy(__mem, &__data(__x0), __n_bytes);
3952 if constexpr (sizeof...(__xs) > 0)
3953 __store_pack_of_simd(__mem + __n_bytes, __xs...);
3954 }
3955
3956// }}}
3957// concat(simd...) {{{
3958template <typename _Tp, typename... _As>
3959 inline _GLIBCXX_SIMD_CONSTEXPR
3960 simd<_Tp, simd_abi::deduce_t<_Tp, (simd_size_v<_Tp, _As> + ...)>>
3961 concat(const simd<_Tp, _As>&... __xs)
3962 {
3963 using _Rp = __deduced_simd<_Tp, (simd_size_v<_Tp, _As> + ...)>;
3964 if constexpr (sizeof...(__xs) == 1)
3965 return simd_cast<_Rp>(__xs...);
3966 else if ((... && __xs._M_is_constprop()))
3967 return simd<_Tp,
3968 simd_abi::deduce_t<_Tp, (simd_size_v<_Tp, _As> + ...)>>([&](
3969 auto __i) constexpr { return __subscript_in_pack<__i>(__xs...); });
3970 else
3971 {
3972 _Rp __r{};
3973 __store_pack_of_simd(reinterpret_cast<char*>(&__data(__r)), __xs...);
3974 return __r;
3975 }
3976 }
3977
3978// }}}
3979// concat(array<simd>) {{{
3980template <typename _Tp, typename _Abi, size_t _Np>
3981 _GLIBCXX_SIMD_ALWAYS_INLINE
3982 _GLIBCXX_SIMD_CONSTEXPR __deduced_simd<_Tp, simd_size_v<_Tp, _Abi> * _Np>
3983 concat(const array<simd<_Tp, _Abi>, _Np>& __x)
3984 {
3985 return __call_with_subscripts<_Np>(__x, [](const auto&... __xs) {
3986 return concat(__xs...);
3987 });
3988 }
3989
3990// }}}
3991
3992/// @cond undocumented
3993// _SmartReference {{{
3994template <typename _Up, typename _Accessor = _Up,
3995 typename _ValueType = typename _Up::value_type>
3996 class _SmartReference
3997 {
3998 friend _Accessor;
3999 int _M_index;
4000 _Up& _M_obj;
4001
4002 _GLIBCXX_SIMD_INTRINSIC constexpr _ValueType _M_read() const noexcept
4003 {
4004 if constexpr (is_arithmetic_v<_Up>)
4005 return _M_obj;
4006 else
4007 return _M_obj[_M_index];
4008 }
4009
4010 template <typename _Tp>
4011 _GLIBCXX_SIMD_INTRINSIC constexpr void _M_write(_Tp&& __x) const
4012 { _Accessor::_S_set(_M_obj, _M_index, static_cast<_Tp&&>(__x)); }
4013
4014 public:
4015 _GLIBCXX_SIMD_INTRINSIC constexpr
4016 _SmartReference(_Up& __o, int __i) noexcept
4017 : _M_index(__i), _M_obj(__o) {}
4018
4019 using value_type = _ValueType;
4020
4021 _GLIBCXX_SIMD_INTRINSIC _SmartReference(const _SmartReference&) = delete;
4022
4023 _GLIBCXX_SIMD_INTRINSIC constexpr operator value_type() const noexcept
4024 { return _M_read(); }
4025
4026 template <typename _Tp,
4027 typename
4028 = _ValuePreservingOrInt<__remove_cvref_t<_Tp>, value_type>>
4029 _GLIBCXX_SIMD_INTRINSIC constexpr _SmartReference operator=(_Tp&& __x) &&
4030 {
4031 _M_write(static_cast<_Tp&&>(__x));
4032 return {_M_obj, _M_index};
4033 }
4034
4035#define _GLIBCXX_SIMD_OP_(__op) \
4036 template <typename _Tp, \
4037 typename _TT \
4038 = decltype(declval<value_type>() __op declval<_Tp>()), \
4039 typename = _ValuePreservingOrInt<__remove_cvref_t<_Tp>, _TT>, \
4040 typename = _ValuePreservingOrInt<_TT, value_type>> \
4041 _GLIBCXX_SIMD_INTRINSIC constexpr _SmartReference \
4042 operator __op##=(_Tp&& __x) && \
4043 { \
4044 const value_type& __lhs = _M_read(); \
4045 _M_write(__lhs __op __x); \
4046 return {_M_obj, _M_index}; \
4047 }
4048 _GLIBCXX_SIMD_ALL_ARITHMETICS(_GLIBCXX_SIMD_OP_);
4049 _GLIBCXX_SIMD_ALL_SHIFTS(_GLIBCXX_SIMD_OP_);
4050 _GLIBCXX_SIMD_ALL_BINARY(_GLIBCXX_SIMD_OP_);
4051#undef _GLIBCXX_SIMD_OP_
4052
4053 template <typename _Tp = void,
4054 typename
4055 = decltype(++declval<conditional_t<true, value_type, _Tp>&>())>
4056 _GLIBCXX_SIMD_INTRINSIC constexpr _SmartReference operator++() &&
4057 {
4058 value_type __x = _M_read();
4059 _M_write(++__x);
4060 return {_M_obj, _M_index};
4061 }
4062
4063 template <typename _Tp = void,
4064 typename
4065 = decltype(declval<conditional_t<true, value_type, _Tp>&>()++)>
4066 _GLIBCXX_SIMD_INTRINSIC constexpr value_type operator++(int) &&
4067 {
4068 const value_type __r = _M_read();
4069 value_type __x = __r;
4070 _M_write(++__x);
4071 return __r;
4072 }
4073
4074 template <typename _Tp = void,
4075 typename
4076 = decltype(--declval<conditional_t<true, value_type, _Tp>&>())>
4077 _GLIBCXX_SIMD_INTRINSIC constexpr _SmartReference operator--() &&
4078 {
4079 value_type __x = _M_read();
4080 _M_write(--__x);
4081 return {_M_obj, _M_index};
4082 }
4083
4084 template <typename _Tp = void,
4085 typename
4086 = decltype(declval<conditional_t<true, value_type, _Tp>&>()--)>
4087 _GLIBCXX_SIMD_INTRINSIC constexpr value_type operator--(int) &&
4088 {
4089 const value_type __r = _M_read();
4090 value_type __x = __r;
4091 _M_write(--__x);
4092 return __r;
4093 }
4094
4095 _GLIBCXX_SIMD_INTRINSIC friend void
4096 swap(_SmartReference&& __a, _SmartReference&& __b) noexcept(
4097 conjunction<
4098 is_nothrow_constructible<value_type, _SmartReference&&>,
4099 is_nothrow_assignable<_SmartReference&&, value_type&&>>::value)
4100 {
4101 value_type __tmp = static_cast<_SmartReference&&>(__a);
4102 static_cast<_SmartReference&&>(__a) = static_cast<value_type>(__b);
4103 static_cast<_SmartReference&&>(__b) = std::move(__tmp);
4104 }
4105
4106 _GLIBCXX_SIMD_INTRINSIC friend void
4107 swap(value_type& __a, _SmartReference&& __b) noexcept(
4108 conjunction<
4109 is_nothrow_constructible<value_type, value_type&&>,
4110 is_nothrow_assignable<value_type&, value_type&&>,
4111 is_nothrow_assignable<_SmartReference&&, value_type&&>>::value)
4112 {
4113 value_type __tmp(std::move(__a));
4114 __a = static_cast<value_type>(__b);
4115 static_cast<_SmartReference&&>(__b) = std::move(__tmp);
4116 }
4117
4118 _GLIBCXX_SIMD_INTRINSIC friend void
4119 swap(_SmartReference&& __a, value_type& __b) noexcept(
4120 conjunction<
4121 is_nothrow_constructible<value_type, _SmartReference&&>,
4122 is_nothrow_assignable<value_type&, value_type&&>,
4123 is_nothrow_assignable<_SmartReference&&, value_type&&>>::value)
4124 {
4125 value_type __tmp(__a);
4126 static_cast<_SmartReference&&>(__a) = std::move(__b);
4127 __b = std::move(__tmp);
4128 }
4129 };
4130
4131// }}}
4132// __scalar_abi_wrapper {{{
4133template <int _Bytes>
4134 struct __scalar_abi_wrapper
4135 {
4136 template <typename _Tp> static constexpr size_t _S_full_size = 1;
4137 template <typename _Tp> static constexpr size_t _S_size = 1;
4138 template <typename _Tp> static constexpr size_t _S_is_partial = false;
4139
4140 template <typename _Tp, typename _Abi = simd_abi::scalar>
4141 static constexpr bool _S_is_valid_v
4142 = _Abi::template _IsValid<_Tp>::value && sizeof(_Tp) == _Bytes;
4143 };
4144
4145// }}}
4146// __decay_abi metafunction {{{
4147template <typename _Tp>
4148 struct __decay_abi { using type = _Tp; };
4149
4150template <int _Bytes>
4151 struct __decay_abi<__scalar_abi_wrapper<_Bytes>>
4152 { using type = simd_abi::scalar; };
4153
4154// }}}
4155// __find_next_valid_abi metafunction {{{1
4156// Given an ABI tag A<N>, find an N2 < N such that A<N2>::_S_is_valid_v<_Tp> ==
4157// true, N2 is a power-of-2, and A<N2>::_S_is_partial<_Tp> is false. Break
4158// recursion at 2 elements in the resulting ABI tag. In this case
4159// type::_S_is_valid_v<_Tp> may be false.
4160template <template <int> class _Abi, int _Bytes, typename _Tp>
4161 struct __find_next_valid_abi
4162 {
4163 static constexpr auto _S_choose()
4164 {
4165 constexpr int _NextBytes = std::__bit_ceil(_Bytes) / 2;
4166 using _NextAbi = _Abi<_NextBytes>;
4167 if constexpr (_NextBytes < sizeof(_Tp) * 2) // break recursion
4168 return _Abi<_Bytes>();
4169 else if constexpr (_NextAbi::template _S_is_partial<_Tp> == false
4170 && _NextAbi::template _S_is_valid_v<_Tp>)
4171 return _NextAbi();
4172 else
4173 return __find_next_valid_abi<_Abi, _NextBytes, _Tp>::_S_choose();
4174 }
4175
4176 using type = decltype(_S_choose());
4177 };
4178
4179template <int _Bytes, typename _Tp>
4180 struct __find_next_valid_abi<__scalar_abi_wrapper, _Bytes, _Tp>
4181 { using type = simd_abi::scalar; };
4182
4183// _AbiList {{{1
4184template <template <int> class...>
4185 struct _AbiList
4186 {
4187 template <typename, int> static constexpr bool _S_has_valid_abi = false;
4188 template <typename, int> using _FirstValidAbi = void;
4189 template <typename, int> using _BestAbi = void;
4190 };
4191
4192template <template <int> class _A0, template <int> class... _Rest>
4193 struct _AbiList<_A0, _Rest...>
4194 {
4195 template <typename _Tp, int _Np>
4196 static constexpr bool _S_has_valid_abi
4197 = _A0<sizeof(_Tp) * _Np>::template _S_is_valid_v<
4198 _Tp> || _AbiList<_Rest...>::template _S_has_valid_abi<_Tp, _Np>;
4199
4200 template <typename _Tp, int _Np>
4201 using _FirstValidAbi = conditional_t<
4202 _A0<sizeof(_Tp) * _Np>::template _S_is_valid_v<_Tp>,
4203 typename __decay_abi<_A0<sizeof(_Tp) * _Np>>::type,
4204 typename _AbiList<_Rest...>::template _FirstValidAbi<_Tp, _Np>>;
4205
4206 template <typename _Tp, int _Np>
4207 static constexpr auto _S_determine_best_abi()
4208 {
4209 static_assert(_Np >= 1);
4210 constexpr int _Bytes = sizeof(_Tp) * _Np;
4211 if constexpr (_Np == 1)
4212 return __make_dependent_t<_Tp, simd_abi::scalar>{};
4213 else
4214 {
4215 constexpr int __fullsize = _A0<_Bytes>::template _S_full_size<_Tp>;
4216 // _A0<_Bytes> is good if:
4217 // 1. The ABI tag is valid for _Tp
4218 // 2. The storage overhead is no more than padding to fill the next
4219 // power-of-2 number of bytes
4220 if constexpr (_A0<_Bytes>::template _S_is_valid_v<
4221 _Tp> && __fullsize / 2 < _Np)
4222 return typename __decay_abi<_A0<_Bytes>>::type{};
4223 else
4224 {
4225 using _Bp =
4226 typename __find_next_valid_abi<_A0, _Bytes, _Tp>::type;
4227 if constexpr (_Bp::template _S_is_valid_v<
4228 _Tp> && _Bp::template _S_size<_Tp> <= _Np)
4229 return _Bp{};
4230 else
4231 return
4232 typename _AbiList<_Rest...>::template _BestAbi<_Tp, _Np>{};
4233 }
4234 }
4235 }
4236
4237 template <typename _Tp, int _Np>
4238 using _BestAbi = decltype(_S_determine_best_abi<_Tp, _Np>());
4239 };
4240
4241// }}}1
4242
4243// the following lists all native ABIs, which makes them accessible to
4244// simd_abi::deduce and select_best_vector_type_t (for fixed_size). Order
4245// matters: Whatever comes first has higher priority.
4246using _AllNativeAbis = _AbiList<simd_abi::_VecBltnBtmsk, simd_abi::_VecBuiltin,
4247 __scalar_abi_wrapper>;
4248
4249// valid _SimdTraits specialization {{{1
4250template <typename _Tp, typename _Abi>
4251 struct _SimdTraits<_Tp, _Abi, void_t<typename _Abi::template _IsValid<_Tp>>>
4252 : _Abi::template __traits<_Tp> {};
4253
4254// __deduce_impl specializations {{{1
4255// try all native ABIs (including scalar) first
4256template <typename _Tp, size_t _Np>
4257 struct __deduce_impl<
4258 _Tp, _Np, enable_if_t<_AllNativeAbis::template _S_has_valid_abi<_Tp, _Np>>>
4259 { using type = _AllNativeAbis::_FirstValidAbi<_Tp, _Np>; };
4260
4261// fall back to fixed_size only if scalar and native ABIs don't match
4262template <typename _Tp, size_t _Np, typename = void>
4263 struct __deduce_fixed_size_fallback {};
4264
4265template <typename _Tp, size_t _Np>
4266 struct __deduce_fixed_size_fallback<_Tp, _Np,
4267 enable_if_t<simd_abi::fixed_size<_Np>::template _S_is_valid_v<_Tp>>>
4268 { using type = simd_abi::fixed_size<_Np>; };
4269
4270template <typename _Tp, size_t _Np, typename>
4271 struct __deduce_impl : public __deduce_fixed_size_fallback<_Tp, _Np> {};
4272
4273//}}}1
4274/// @endcond
4275
4276// simd_mask {{{
4277template <typename _Tp, typename _Abi>
4278 class simd_mask : public _SimdTraits<_Tp, _Abi>::_MaskBase
4279 {
4280 // types, tags, and friends {{{
4281 using _Traits = _SimdTraits<_Tp, _Abi>;
4282 using _MemberType = typename _Traits::_MaskMember;
4283
4284 // We map all masks with equal element sizeof to a single integer type, the
4285 // one given by __int_for_sizeof_t<_Tp>. This is the approach
4286 // [[gnu::vector_size(N)]] types take as well and it reduces the number of
4287 // template specializations in the implementation classes.
4288 using _Ip = __int_for_sizeof_t<_Tp>;
4289 static constexpr _Ip* _S_type_tag = nullptr;
4290
4291 friend typename _Traits::_MaskBase;
4292 friend class simd<_Tp, _Abi>; // to construct masks on return
4293 friend typename _Traits::_SimdImpl; // to construct masks on return and
4294 // inspect data on masked operations
4295 public:
4296 using _Impl = typename _Traits::_MaskImpl;
4297 friend _Impl;
4298
4299 // }}}
4300 // member types {{{
4301 using value_type = bool;
4302 using reference = _SmartReference<_MemberType, _Impl, value_type>;
4303 using simd_type = simd<_Tp, _Abi>;
4304 using abi_type = _Abi;
4305
4306 // }}}
4307 static constexpr size_t size() // {{{
4308 { return __size_or_zero_v<_Tp, _Abi>; }
4309
4310 // }}}
4311 // constructors & assignment {{{
4312 simd_mask() = default;
4313 simd_mask(const simd_mask&) = default;
4314 simd_mask(simd_mask&&) = default;
4315 simd_mask& operator=(const simd_mask&) = default;
4316 simd_mask& operator=(simd_mask&&) = default;
4317
4318 // }}}
4319 // access to internal representation (optional feature) {{{
4320 _GLIBCXX_SIMD_ALWAYS_INLINE explicit
4321 simd_mask(typename _Traits::_MaskCastType __init)
4322 : _M_data{__init} {}
4323 // conversions to internal type is done in _MaskBase
4324
4325 // }}}
4326 // bitset interface (extension to be proposed) {{{
4327 // TS_FEEDBACK:
4328 // Conversion of simd_mask to and from bitset makes it much easier to
4329 // interface with other facilities. I suggest adding `static
4330 // simd_mask::from_bitset` and `simd_mask::to_bitset`.
4331 _GLIBCXX_SIMD_ALWAYS_INLINE static simd_mask
4332 __from_bitset(bitset<size()> bs)
4333 { return {__bitset_init, bs}; }
4334
4335 _GLIBCXX_SIMD_ALWAYS_INLINE bitset<size()>
4336 __to_bitset() const
4337 { return _Impl::_S_to_bits(_M_data)._M_to_bitset(); }
4338
4339 // }}}
4340 // explicit broadcast constructor {{{
4341 _GLIBCXX_SIMD_ALWAYS_INLINE explicit _GLIBCXX_SIMD_CONSTEXPR
4342 simd_mask(value_type __x)
4343 : _M_data(_Impl::template _S_broadcast<_Ip>(__x)) {}
4344
4345 // }}}
4346 // implicit type conversion constructor {{{
4347 #ifdef _GLIBCXX_SIMD_ENABLE_IMPLICIT_MASK_CAST
4348 // proposed improvement
4349 template <typename _Up, typename _A2,
4350 typename = enable_if_t<simd_size_v<_Up, _A2> == size()>>
4351 _GLIBCXX_SIMD_ALWAYS_INLINE explicit(sizeof(_MemberType)
4352 != sizeof(typename _SimdTraits<_Up, _A2>::_MaskMember))
4353 simd_mask(const simd_mask<_Up, _A2>& __x)
4354 : simd_mask(__proposed::static_simd_cast<simd_mask>(__x)) {}
4355 #else
4356 // conforming to ISO/IEC 19570:2018
4357 template <typename _Up, typename = enable_if_t<conjunction<
4358 is_same<abi_type, simd_abi::fixed_size<size()>>,
4359 is_same<_Up, _Up>>::value>>
4360 _GLIBCXX_SIMD_ALWAYS_INLINE
4361 simd_mask(const simd_mask<_Up, simd_abi::fixed_size<size()>>& __x)
4362 : _M_data(_Impl::_S_from_bitmask(__data(__x), _S_type_tag)) {}
4363 #endif
4364
4365 // }}}
4366 // load constructor {{{
4367 template <typename _Flags>
4368 _GLIBCXX_SIMD_ALWAYS_INLINE
4369 simd_mask(const value_type* __mem, _Flags)
4370 : _M_data(_Impl::template _S_load<_Ip>(
4371 _Flags::template _S_apply<simd_mask>(__mem))) {}
4372
4373 template <typename _Flags>
4374 _GLIBCXX_SIMD_ALWAYS_INLINE
4375 simd_mask(const value_type* __mem, simd_mask __k, _Flags)
4376 : _M_data{}
4377 {
4378 _M_data
4379 = _Impl::_S_masked_load(_M_data, __k._M_data,
4380 _Flags::template _S_apply<simd_mask>(__mem));
4381 }
4382
4383 // }}}
4384 // loads [simd_mask.load] {{{
4385 template <typename _Flags>
4386 _GLIBCXX_SIMD_ALWAYS_INLINE void
4387 copy_from(const value_type* __mem, _Flags)
4388 {
4389 _M_data = _Impl::template _S_load<_Ip>(
4390 _Flags::template _S_apply<simd_mask>(__mem));
4391 }
4392
4393 // }}}
4394 // stores [simd_mask.store] {{{
4395 template <typename _Flags>
4396 _GLIBCXX_SIMD_ALWAYS_INLINE void
4397 copy_to(value_type* __mem, _Flags) const
4398 { _Impl::_S_store(_M_data, _Flags::template _S_apply<simd_mask>(__mem)); }
4399
4400 // }}}
4401 // scalar access {{{
4402 _GLIBCXX_SIMD_ALWAYS_INLINE reference
4403 operator[](size_t __i)
4404 {
4405 if (__i >= size())
4406 __invoke_ub("Subscript %d is out of range [0, %d]", __i, size() - 1);
4407 return {_M_data, int(__i)};
4408 }
4409
4410 _GLIBCXX_SIMD_ALWAYS_INLINE value_type
4411 operator[](size_t __i) const
4412 {
4413 if (__i >= size())
4414 __invoke_ub("Subscript %d is out of range [0, %d]", __i, size() - 1);
4415 if constexpr (__is_scalar_abi<_Abi>())
4416 return _M_data;
4417 else
4418 return static_cast<bool>(_M_data[__i]);
4419 }
4420
4421 // }}}
4422 // negation {{{
4423 _GLIBCXX_SIMD_ALWAYS_INLINE simd_mask
4424 operator!() const
4425 { return {__private_init, _Impl::_S_bit_not(_M_data)}; }
4426
4427 // }}}
4428 // simd_mask binary operators [simd_mask.binary] {{{
4429 #ifdef _GLIBCXX_SIMD_ENABLE_IMPLICIT_MASK_CAST
4430 // simd_mask<int> && simd_mask<uint> needs disambiguation
4431 template <typename _Up, typename _A2,
4432 typename
4433 = enable_if_t<is_convertible_v<simd_mask<_Up, _A2>, simd_mask>>>
4434 _GLIBCXX_SIMD_ALWAYS_INLINE friend simd_mask
4435 operator&&(const simd_mask& __x, const simd_mask<_Up, _A2>& __y)
4436 {
4437 return {__private_init,
4438 _Impl::_S_logical_and(__x._M_data, simd_mask(__y)._M_data)};
4439 }
4440
4441 template <typename _Up, typename _A2,
4442 typename
4443 = enable_if_t<is_convertible_v<simd_mask<_Up, _A2>, simd_mask>>>
4444 _GLIBCXX_SIMD_ALWAYS_INLINE friend simd_mask
4445 operator||(const simd_mask& __x, const simd_mask<_Up, _A2>& __y)
4446 {
4447 return {__private_init,
4448 _Impl::_S_logical_or(__x._M_data, simd_mask(__y)._M_data)};
4449 }
4450 #endif // _GLIBCXX_SIMD_ENABLE_IMPLICIT_MASK_CAST
4451
4452 _GLIBCXX_SIMD_ALWAYS_INLINE friend simd_mask
4453 operator&&(const simd_mask& __x, const simd_mask& __y)
4454 {
4455 return {__private_init, _Impl::_S_logical_and(__x._M_data, __y._M_data)};
4456 }
4457
4458 _GLIBCXX_SIMD_ALWAYS_INLINE friend simd_mask
4459 operator||(const simd_mask& __x, const simd_mask& __y)
4460 {
4461 return {__private_init, _Impl::_S_logical_or(__x._M_data, __y._M_data)};
4462 }
4463
4464 _GLIBCXX_SIMD_ALWAYS_INLINE friend simd_mask
4465 operator&(const simd_mask& __x, const simd_mask& __y)
4466 { return {__private_init, _Impl::_S_bit_and(__x._M_data, __y._M_data)}; }
4467
4468 _GLIBCXX_SIMD_ALWAYS_INLINE friend simd_mask
4469 operator|(const simd_mask& __x, const simd_mask& __y)
4470 { return {__private_init, _Impl::_S_bit_or(__x._M_data, __y._M_data)}; }
4471
4472 _GLIBCXX_SIMD_ALWAYS_INLINE friend simd_mask
4473 operator^(const simd_mask& __x, const simd_mask& __y)
4474 { return {__private_init, _Impl::_S_bit_xor(__x._M_data, __y._M_data)}; }
4475
4476 _GLIBCXX_SIMD_ALWAYS_INLINE friend simd_mask&
4477 operator&=(simd_mask& __x, const simd_mask& __y)
4478 {
4479 __x._M_data = _Impl::_S_bit_and(__x._M_data, __y._M_data);
4480 return __x;
4481 }
4482
4483 _GLIBCXX_SIMD_ALWAYS_INLINE friend simd_mask&
4484 operator|=(simd_mask& __x, const simd_mask& __y)
4485 {
4486 __x._M_data = _Impl::_S_bit_or(__x._M_data, __y._M_data);
4487 return __x;
4488 }
4489
4490 _GLIBCXX_SIMD_ALWAYS_INLINE friend simd_mask&
4491 operator^=(simd_mask& __x, const simd_mask& __y)
4492 {
4493 __x._M_data = _Impl::_S_bit_xor(__x._M_data, __y._M_data);
4494 return __x;
4495 }
4496
4497 // }}}
4498 // simd_mask compares [simd_mask.comparison] {{{
4499 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd_mask
4500 operator==(const simd_mask& __x, const simd_mask& __y)
4501 { return !operator!=(__x, __y); }
4502
4503 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd_mask
4504 operator!=(const simd_mask& __x, const simd_mask& __y)
4505 { return {__private_init, _Impl::_S_bit_xor(__x._M_data, __y._M_data)}; }
4506
4507 // }}}
4508 // private_init ctor {{{
4509 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR
4510 simd_mask(_PrivateInit, typename _Traits::_MaskMember __init)
4511 : _M_data(__init) {}
4512
4513 // }}}
4514 // private_init generator ctor {{{
4515 template <typename _Fp, typename = decltype(bool(declval<_Fp>()(size_t())))>
4516 _GLIBCXX_SIMD_INTRINSIC constexpr
4517 simd_mask(_PrivateInit, _Fp&& __gen)
4518 : _M_data()
4519 {
4520 __execute_n_times<size()>([&](auto __i) constexpr {
4521 _Impl::_S_set(_M_data, __i, __gen(__i));
4522 });
4523 }
4524
4525 // }}}
4526 // bitset_init ctor {{{
4527 _GLIBCXX_SIMD_INTRINSIC simd_mask(_BitsetInit, bitset<size()> __init)
4528 : _M_data(
4529 _Impl::_S_from_bitmask(_SanitizedBitMask<size()>(__init), _S_type_tag))
4530 {}
4531
4532 // }}}
4533 // __cvt {{{
4534 // TS_FEEDBACK:
4535 // The conversion operator this implements should be a ctor on simd_mask.
4536 // Once you call .__cvt() on a simd_mask it converts conveniently.
4537 // A useful variation: add `explicit(sizeof(_Tp) != sizeof(_Up))`
4538 struct _CvtProxy
4539 {
4540 template <typename _Up, typename _A2,
4541 typename
4542 = enable_if_t<simd_size_v<_Up, _A2> == simd_size_v<_Tp, _Abi>>>
4543 operator simd_mask<_Up, _A2>() &&
4544 {
4545 using namespace std::experimental::__proposed;
4546 return static_simd_cast<simd_mask<_Up, _A2>>(_M_data);
4547 }
4548
4549 const simd_mask<_Tp, _Abi>& _M_data;
4550 };
4551
4552 _GLIBCXX_SIMD_INTRINSIC _CvtProxy
4553 __cvt() const
4554 { return {*this}; }
4555
4556 // }}}
4557 // operator?: overloads (suggested extension) {{{
4558 #ifdef __GXX_CONDITIONAL_IS_OVERLOADABLE__
4559 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd_mask
4560 operator?:(const simd_mask& __k, const simd_mask& __where_true,
4561 const simd_mask& __where_false)
4562 {
4563 auto __ret = __where_false;
4564 _Impl::_S_masked_assign(__k._M_data, __ret._M_data, __where_true._M_data);
4565 return __ret;
4566 }
4567
4568 template <typename _U1, typename _U2,
4569 typename _Rp = simd<common_type_t<_U1, _U2>, _Abi>,
4570 typename = enable_if_t<conjunction_v<
4571 is_convertible<_U1, _Rp>, is_convertible<_U2, _Rp>,
4572 is_convertible<simd_mask, typename _Rp::mask_type>>>>
4573 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend _Rp
4574 operator?:(const simd_mask& __k, const _U1& __where_true,
4575 const _U2& __where_false)
4576 {
4577 _Rp __ret = __where_false;
4578 _Rp::_Impl::_S_masked_assign(
4579 __data(static_cast<typename _Rp::mask_type>(__k)), __data(__ret),
4580 __data(static_cast<_Rp>(__where_true)));
4581 return __ret;
4582 }
4583
4584 #ifdef _GLIBCXX_SIMD_ENABLE_IMPLICIT_MASK_CAST
4585 template <typename _Kp, typename _Ak, typename _Up, typename _Au,
4586 typename = enable_if_t<
4587 conjunction_v<is_convertible<simd_mask<_Kp, _Ak>, simd_mask>,
4588 is_convertible<simd_mask<_Up, _Au>, simd_mask>>>>
4589 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd_mask
4590 operator?:(const simd_mask<_Kp, _Ak>& __k, const simd_mask& __where_true,
4591 const simd_mask<_Up, _Au>& __where_false)
4592 {
4593 simd_mask __ret = __where_false;
4594 _Impl::_S_masked_assign(simd_mask(__k)._M_data, __ret._M_data,
4595 __where_true._M_data);
4596 return __ret;
4597 }
4598 #endif // _GLIBCXX_SIMD_ENABLE_IMPLICIT_MASK_CAST
4599 #endif // __GXX_CONDITIONAL_IS_OVERLOADABLE__
4600
4601 // }}}
4602 // _M_is_constprop {{{
4603 _GLIBCXX_SIMD_INTRINSIC constexpr bool
4604 _M_is_constprop() const
4605 {
4606 if constexpr (__is_scalar_abi<_Abi>())
4607 return __builtin_constant_p(_M_data);
4608 else
4609 return _M_data._M_is_constprop();
4610 }
4611
4612 // }}}
4613
4614 private:
4615 friend const auto& __data<_Tp, abi_type>(const simd_mask&);
4616 friend auto& __data<_Tp, abi_type>(simd_mask&);
4617 alignas(_Traits::_S_mask_align) _MemberType _M_data;
4618 };
4619
4620// }}}
4621
4622/// @cond undocumented
4623// __data(simd_mask) {{{
4624template <typename _Tp, typename _Ap>
4625 _GLIBCXX_SIMD_INTRINSIC constexpr const auto&
4626 __data(const simd_mask<_Tp, _Ap>& __x)
4627 { return __x._M_data; }
4628
4629template <typename _Tp, typename _Ap>
4630 _GLIBCXX_SIMD_INTRINSIC constexpr auto&
4631 __data(simd_mask<_Tp, _Ap>& __x)
4632 { return __x._M_data; }
4633
4634// }}}
4635/// @endcond
4636
4637// simd_mask reductions [simd_mask.reductions] {{{
4638template <typename _Tp, typename _Abi>
4639 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR bool
4640 all_of(const simd_mask<_Tp, _Abi>& __k) noexcept
4641 {
4642 if (__builtin_is_constant_evaluated() || __k._M_is_constprop())
4643 {
4644 for (size_t __i = 0; __i < simd_size_v<_Tp, _Abi>; ++__i)
4645 if (!__k[__i])
4646 return false;
4647 return true;
4648 }
4649 else
4650 return _Abi::_MaskImpl::_S_all_of(__k);
4651 }
4652
4653template <typename _Tp, typename _Abi>
4654 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR bool
4655 any_of(const simd_mask<_Tp, _Abi>& __k) noexcept
4656 {
4657 if (__builtin_is_constant_evaluated() || __k._M_is_constprop())
4658 {
4659 for (size_t __i = 0; __i < simd_size_v<_Tp, _Abi>; ++__i)
4660 if (__k[__i])
4661 return true;
4662 return false;
4663 }
4664 else
4665 return _Abi::_MaskImpl::_S_any_of(__k);
4666 }
4667
4668template <typename _Tp, typename _Abi>
4669 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR bool
4670 none_of(const simd_mask<_Tp, _Abi>& __k) noexcept
4671 {
4672 if (__builtin_is_constant_evaluated() || __k._M_is_constprop())
4673 {
4674 for (size_t __i = 0; __i < simd_size_v<_Tp, _Abi>; ++__i)
4675 if (__k[__i])
4676 return false;
4677 return true;
4678 }
4679 else
4680 return _Abi::_MaskImpl::_S_none_of(__k);
4681 }
4682
4683template <typename _Tp, typename _Abi>
4684 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR bool
4685 some_of(const simd_mask<_Tp, _Abi>& __k) noexcept
4686 {
4687 if (__builtin_is_constant_evaluated() || __k._M_is_constprop())
4688 {
4689 for (size_t __i = 1; __i < simd_size_v<_Tp, _Abi>; ++__i)
4690 if (__k[__i] != __k[__i - 1])
4691 return true;
4692 return false;
4693 }
4694 else
4695 return _Abi::_MaskImpl::_S_some_of(__k);
4696 }
4697
4698template <typename _Tp, typename _Abi>
4699 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR int
4700 popcount(const simd_mask<_Tp, _Abi>& __k) noexcept
4701 {
4702 if (__builtin_is_constant_evaluated() || __k._M_is_constprop())
4703 {
4704 const int __r = __call_with_subscripts<simd_size_v<_Tp, _Abi>>(
4705 __k, [](auto... __elements) { return ((__elements != 0) + ...); });
4706 if (__builtin_is_constant_evaluated() || __builtin_constant_p(__r))
4707 return __r;
4708 }
4709 return _Abi::_MaskImpl::_S_popcount(__k);
4710 }
4711
4712template <typename _Tp, typename _Abi>
4713 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR int
4714 find_first_set(const simd_mask<_Tp, _Abi>& __k)
4715 {
4716 if (__builtin_is_constant_evaluated() || __k._M_is_constprop())
4717 {
4718 constexpr size_t _Np = simd_size_v<_Tp, _Abi>;
4719 const size_t _Idx = __call_with_n_evaluations<_Np>(
4720 [](auto... __indexes) { return std::min({__indexes...}); },
4721 [&](auto __i) { return __k[__i] ? +__i : _Np; });
4722 if (_Idx >= _Np)
4723 __invoke_ub("find_first_set(empty mask) is UB");
4724 if (__builtin_constant_p(_Idx))
4725 return _Idx;
4726 }
4727 return _Abi::_MaskImpl::_S_find_first_set(__k);
4728 }
4729
4730template <typename _Tp, typename _Abi>
4731 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR int
4732 find_last_set(const simd_mask<_Tp, _Abi>& __k)
4733 {
4734 if (__builtin_is_constant_evaluated() || __k._M_is_constprop())
4735 {
4736 constexpr size_t _Np = simd_size_v<_Tp, _Abi>;
4737 const int _Idx = __call_with_n_evaluations<_Np>(
4738 [](auto... __indexes) { return std::max({__indexes...}); },
4739 [&](auto __i) { return __k[__i] ? int(__i) : -1; });
4740 if (_Idx < 0)
4741 __invoke_ub("find_first_set(empty mask) is UB");
4742 if (__builtin_constant_p(_Idx))
4743 return _Idx;
4744 }
4745 return _Abi::_MaskImpl::_S_find_last_set(__k);
4746 }
4747
4748_GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR bool
4749all_of(_ExactBool __x) noexcept
4750{ return __x; }
4751
4752_GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR bool
4753any_of(_ExactBool __x) noexcept
4754{ return __x; }
4755
4756_GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR bool
4757none_of(_ExactBool __x) noexcept
4758{ return !__x; }
4759
4760_GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR bool
4761some_of(_ExactBool) noexcept
4762{ return false; }
4763
4764_GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR int
4765popcount(_ExactBool __x) noexcept
4766{ return __x; }
4767
4768_GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR int
4769find_first_set(_ExactBool)
4770{ return 0; }
4771
4772_GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR int
4773find_last_set(_ExactBool)
4774{ return 0; }
4775
4776// }}}
4777
4778/// @cond undocumented
4779// _SimdIntOperators{{{1
4780template <typename _V, typename _Impl, bool>
4781 class _SimdIntOperators {};
4782
4783template <typename _V, typename _Impl>
4784 class _SimdIntOperators<_V, _Impl, true>
4785 {
4786 _GLIBCXX_SIMD_INTRINSIC const _V& __derived() const
4787 { return *static_cast<const _V*>(this); }
4788
4789 template <typename _Tp>
4790 _GLIBCXX_SIMD_INTRINSIC static _GLIBCXX_SIMD_CONSTEXPR _V
4791 _S_make_derived(_Tp&& __d)
4792 { return {__private_init, static_cast<_Tp&&>(__d)}; }
4793
4794 public:
4795 _GLIBCXX_SIMD_CONSTEXPR friend _V& operator%=(_V& __lhs, const _V& __x)
4796 { return __lhs = __lhs % __x; }
4797
4798 _GLIBCXX_SIMD_CONSTEXPR friend _V& operator&=(_V& __lhs, const _V& __x)
4799 { return __lhs = __lhs & __x; }
4800
4801 _GLIBCXX_SIMD_CONSTEXPR friend _V& operator|=(_V& __lhs, const _V& __x)
4802 { return __lhs = __lhs | __x; }
4803
4804 _GLIBCXX_SIMD_CONSTEXPR friend _V& operator^=(_V& __lhs, const _V& __x)
4805 { return __lhs = __lhs ^ __x; }
4806
4807 _GLIBCXX_SIMD_CONSTEXPR friend _V& operator<<=(_V& __lhs, const _V& __x)
4808 { return __lhs = __lhs << __x; }
4809
4810 _GLIBCXX_SIMD_CONSTEXPR friend _V& operator>>=(_V& __lhs, const _V& __x)
4811 { return __lhs = __lhs >> __x; }
4812
4813 _GLIBCXX_SIMD_CONSTEXPR friend _V& operator<<=(_V& __lhs, int __x)
4814 { return __lhs = __lhs << __x; }
4815
4816 _GLIBCXX_SIMD_CONSTEXPR friend _V& operator>>=(_V& __lhs, int __x)
4817 { return __lhs = __lhs >> __x; }
4818
4819 _GLIBCXX_SIMD_CONSTEXPR friend _V operator%(const _V& __x, const _V& __y)
4820 {
4821 return _SimdIntOperators::_S_make_derived(
4822 _Impl::_S_modulus(__data(__x), __data(__y)));
4823 }
4824
4825 _GLIBCXX_SIMD_CONSTEXPR friend _V operator&(const _V& __x, const _V& __y)
4826 {
4827 return _SimdIntOperators::_S_make_derived(
4828 _Impl::_S_bit_and(__data(__x), __data(__y)));
4829 }
4830
4831 _GLIBCXX_SIMD_CONSTEXPR friend _V operator|(const _V& __x, const _V& __y)
4832 {
4833 return _SimdIntOperators::_S_make_derived(
4834 _Impl::_S_bit_or(__data(__x), __data(__y)));
4835 }
4836
4837 _GLIBCXX_SIMD_CONSTEXPR friend _V operator^(const _V& __x, const _V& __y)
4838 {
4839 return _SimdIntOperators::_S_make_derived(
4840 _Impl::_S_bit_xor(__data(__x), __data(__y)));
4841 }
4842
4843 _GLIBCXX_SIMD_CONSTEXPR friend _V operator<<(const _V& __x, const _V& __y)
4844 {
4845 return _SimdIntOperators::_S_make_derived(
4846 _Impl::_S_bit_shift_left(__data(__x), __data(__y)));
4847 }
4848
4849 _GLIBCXX_SIMD_CONSTEXPR friend _V operator>>(const _V& __x, const _V& __y)
4850 {
4851 return _SimdIntOperators::_S_make_derived(
4852 _Impl::_S_bit_shift_right(__data(__x), __data(__y)));
4853 }
4854
4855 template <typename _VV = _V>
4856 _GLIBCXX_SIMD_CONSTEXPR friend _V operator<<(const _V& __x, int __y)
4857 {
4858 using _Tp = typename _VV::value_type;
4859 if (__y < 0)
4860 __invoke_ub("The behavior is undefined if the right operand of a "
4861 "shift operation is negative. [expr.shift]\nA shift by "
4862 "%d was requested",
4863 __y);
4864 if (size_t(__y) >= sizeof(declval<_Tp>() << __y) * __CHAR_BIT__)
4865 __invoke_ub(
4866 "The behavior is undefined if the right operand of a "
4867 "shift operation is greater than or equal to the width of the "
4868 "promoted left operand. [expr.shift]\nA shift by %d was requested",
4869 __y);
4870 return _SimdIntOperators::_S_make_derived(
4871 _Impl::_S_bit_shift_left(__data(__x), __y));
4872 }
4873
4874 template <typename _VV = _V>
4875 _GLIBCXX_SIMD_CONSTEXPR friend _V operator>>(const _V& __x, int __y)
4876 {
4877 using _Tp = typename _VV::value_type;
4878 if (__y < 0)
4879 __invoke_ub(
4880 "The behavior is undefined if the right operand of a shift "
4881 "operation is negative. [expr.shift]\nA shift by %d was requested",
4882 __y);
4883 if (size_t(__y) >= sizeof(declval<_Tp>() << __y) * __CHAR_BIT__)
4884 __invoke_ub(
4885 "The behavior is undefined if the right operand of a shift "
4886 "operation is greater than or equal to the width of the promoted "
4887 "left operand. [expr.shift]\nA shift by %d was requested",
4888 __y);
4889 return _SimdIntOperators::_S_make_derived(
4890 _Impl::_S_bit_shift_right(__data(__x), __y));
4891 }
4892
4893 // unary operators (for integral _Tp)
4894 _GLIBCXX_SIMD_CONSTEXPR _V operator~() const
4895 { return {__private_init, _Impl::_S_complement(__derived()._M_data)}; }
4896 };
4897
4898//}}}1
4899/// @endcond
4900
4901// simd {{{
4902template <typename _Tp, typename _Abi>
4903 class simd : public _SimdIntOperators<
4904 simd<_Tp, _Abi>, typename _SimdTraits<_Tp, _Abi>::_SimdImpl,
4905 conjunction<is_integral<_Tp>,
4906 typename _SimdTraits<_Tp, _Abi>::_IsValid>::value>,
4907 public _SimdTraits<_Tp, _Abi>::_SimdBase
4908 {
4909 using _Traits = _SimdTraits<_Tp, _Abi>;
4910 using _MemberType = typename _Traits::_SimdMember;
4911 using _CastType = typename _Traits::_SimdCastType;
4912 static constexpr _Tp* _S_type_tag = nullptr;
4913 friend typename _Traits::_SimdBase;
4914
4915 public:
4916 using _Impl = typename _Traits::_SimdImpl;
4917 friend _Impl;
4918 friend _SimdIntOperators<simd, _Impl, true>;
4919
4920 using value_type = _Tp;
4921 using reference = _SmartReference<_MemberType, _Impl, value_type>;
4922 using mask_type = simd_mask<_Tp, _Abi>;
4923 using abi_type = _Abi;
4924
4925 static constexpr size_t size()
4926 { return __size_or_zero_v<_Tp, _Abi>; }
4927
4928 _GLIBCXX_SIMD_CONSTEXPR simd() = default;
4929 _GLIBCXX_SIMD_CONSTEXPR simd(const simd&) = default;
4930 _GLIBCXX_SIMD_CONSTEXPR simd(simd&&) noexcept = default;
4931 _GLIBCXX_SIMD_CONSTEXPR simd& operator=(const simd&) = default;
4932 _GLIBCXX_SIMD_CONSTEXPR simd& operator=(simd&&) noexcept = default;
4933
4934 // implicit broadcast constructor
4935 template <typename _Up,
4936 typename = enable_if_t<!is_same_v<__remove_cvref_t<_Up>, bool>>>
4937 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR
4938 simd(_ValuePreservingOrInt<_Up, value_type>&& __x)
4939 : _M_data(
4940 _Impl::_S_broadcast(static_cast<value_type>(static_cast<_Up&&>(__x))))
4941 {}
4942
4943 // implicit type conversion constructor (convert from fixed_size to
4944 // fixed_size)
4945 template <typename _Up>
4946 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR
4947 simd(const simd<_Up, simd_abi::fixed_size<size()>>& __x,
4949 conjunction<
4950 is_same<simd_abi::fixed_size<size()>, abi_type>,
4951 negation<__is_narrowing_conversion<_Up, value_type>>,
4952 __converts_to_higher_integer_rank<_Up, value_type>>::value,
4953 void*> = nullptr)
4954 : simd{static_cast<array<_Up, size()>>(__x).data(), vector_aligned} {}
4955
4956 // explicit type conversion constructor
4957#ifdef _GLIBCXX_SIMD_ENABLE_STATIC_CAST
4958 template <typename _Up, typename _A2,
4959 typename = decltype(static_simd_cast<simd>(
4960 declval<const simd<_Up, _A2>&>()))>
4961 _GLIBCXX_SIMD_ALWAYS_INLINE explicit _GLIBCXX_SIMD_CONSTEXPR
4962 simd(const simd<_Up, _A2>& __x)
4963 : simd(static_simd_cast<simd>(__x)) {}
4964#endif // _GLIBCXX_SIMD_ENABLE_STATIC_CAST
4965
4966 // generator constructor
4967 template <typename _Fp>
4968 _GLIBCXX_SIMD_ALWAYS_INLINE explicit _GLIBCXX_SIMD_CONSTEXPR
4969 simd(_Fp&& __gen, _ValuePreservingOrInt<decltype(declval<_Fp>()(
4970 declval<_SizeConstant<0>&>())),
4971 value_type>* = nullptr)
4972 : _M_data(_Impl::_S_generator(static_cast<_Fp&&>(__gen), _S_type_tag)) {}
4973
4974 // load constructor
4975 template <typename _Up, typename _Flags>
4976 _GLIBCXX_SIMD_ALWAYS_INLINE
4977 simd(const _Up* __mem, _Flags)
4978 : _M_data(
4979 _Impl::_S_load(_Flags::template _S_apply<simd>(__mem), _S_type_tag))
4980 {}
4981
4982 // loads [simd.load]
4983 template <typename _Up, typename _Flags>
4984 _GLIBCXX_SIMD_ALWAYS_INLINE void
4985 copy_from(const _Vectorizable<_Up>* __mem, _Flags)
4986 {
4987 _M_data = static_cast<decltype(_M_data)>(
4988 _Impl::_S_load(_Flags::template _S_apply<simd>(__mem), _S_type_tag));
4989 }
4990
4991 // stores [simd.store]
4992 template <typename _Up, typename _Flags>
4993 _GLIBCXX_SIMD_ALWAYS_INLINE void
4994 copy_to(_Vectorizable<_Up>* __mem, _Flags) const
4995 {
4996 _Impl::_S_store(_M_data, _Flags::template _S_apply<simd>(__mem),
4997 _S_type_tag);
4998 }
4999
5000 // scalar access
5001 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR reference
5002 operator[](size_t __i)
5003 { return {_M_data, int(__i)}; }
5004
5005 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR value_type
5006 operator[]([[maybe_unused]] size_t __i) const
5007 {
5008 if constexpr (__is_scalar_abi<_Abi>())
5009 {
5010 _GLIBCXX_DEBUG_ASSERT(__i == 0);
5011 return _M_data;
5012 }
5013 else
5014 return _M_data[__i];
5015 }
5016
5017 // increment and decrement:
5018 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR simd&
5019 operator++()
5020 {
5021 _Impl::_S_increment(_M_data);
5022 return *this;
5023 }
5024
5025 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR simd
5026 operator++(int)
5027 {
5028 simd __r = *this;
5029 _Impl::_S_increment(_M_data);
5030 return __r;
5031 }
5032
5033 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR simd&
5034 operator--()
5035 {
5036 _Impl::_S_decrement(_M_data);
5037 return *this;
5038 }
5039
5040 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR simd
5041 operator--(int)
5042 {
5043 simd __r = *this;
5044 _Impl::_S_decrement(_M_data);
5045 return __r;
5046 }
5047
5048 // unary operators (for any _Tp)
5049 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR mask_type
5050 operator!() const
5051 { return {__private_init, _Impl::_S_negate(_M_data)}; }
5052
5053 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR simd
5054 operator+() const
5055 { return *this; }
5056
5057 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR simd
5058 operator-() const
5059 { return {__private_init, _Impl::_S_unary_minus(_M_data)}; }
5060
5061 // access to internal representation (suggested extension)
5062 _GLIBCXX_SIMD_ALWAYS_INLINE explicit _GLIBCXX_SIMD_CONSTEXPR
5063 simd(_CastType __init) : _M_data(__init) {}
5064
5065 // compound assignment [simd.cassign]
5066 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd&
5067 operator+=(simd& __lhs, const simd& __x)
5068 { return __lhs = __lhs + __x; }
5069
5070 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd&
5071 operator-=(simd& __lhs, const simd& __x)
5072 { return __lhs = __lhs - __x; }
5073
5074 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd&
5075 operator*=(simd& __lhs, const simd& __x)
5076 { return __lhs = __lhs * __x; }
5077
5078 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd&
5079 operator/=(simd& __lhs, const simd& __x)
5080 { return __lhs = __lhs / __x; }
5081
5082 // binary operators [simd.binary]
5083 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd
5084 operator+(const simd& __x, const simd& __y)
5085 { return {__private_init, _Impl::_S_plus(__x._M_data, __y._M_data)}; }
5086
5087 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd
5088 operator-(const simd& __x, const simd& __y)
5089 { return {__private_init, _Impl::_S_minus(__x._M_data, __y._M_data)}; }
5090
5091 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd
5092 operator*(const simd& __x, const simd& __y)
5093 { return {__private_init, _Impl::_S_multiplies(__x._M_data, __y._M_data)}; }
5094
5095 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd
5096 operator/(const simd& __x, const simd& __y)
5097 { return {__private_init, _Impl::_S_divides(__x._M_data, __y._M_data)}; }
5098
5099 // compares [simd.comparison]
5100 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend mask_type
5101 operator==(const simd& __x, const simd& __y)
5102 { return simd::_S_make_mask(_Impl::_S_equal_to(__x._M_data, __y._M_data)); }
5103
5104 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend mask_type
5105 operator!=(const simd& __x, const simd& __y)
5106 {
5107 return simd::_S_make_mask(
5108 _Impl::_S_not_equal_to(__x._M_data, __y._M_data));
5109 }
5110
5111 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend mask_type
5112 operator<(const simd& __x, const simd& __y)
5113 { return simd::_S_make_mask(_Impl::_S_less(__x._M_data, __y._M_data)); }
5114
5115 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend mask_type
5116 operator<=(const simd& __x, const simd& __y)
5117 {
5118 return simd::_S_make_mask(_Impl::_S_less_equal(__x._M_data, __y._M_data));
5119 }
5120
5121 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend mask_type
5122 operator>(const simd& __x, const simd& __y)
5123 { return simd::_S_make_mask(_Impl::_S_less(__y._M_data, __x._M_data)); }
5124
5125 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend mask_type
5126 operator>=(const simd& __x, const simd& __y)
5127 {
5128 return simd::_S_make_mask(_Impl::_S_less_equal(__y._M_data, __x._M_data));
5129 }
5130
5131 // operator?: overloads (suggested extension) {{{
5132#ifdef __GXX_CONDITIONAL_IS_OVERLOADABLE__
5133 _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd
5134 operator?:(const mask_type& __k, const simd& __where_true,
5135 const simd& __where_false)
5136 {
5137 auto __ret = __where_false;
5138 _Impl::_S_masked_assign(__data(__k), __data(__ret), __data(__where_true));
5139 return __ret;
5140 }
5141
5142#endif // __GXX_CONDITIONAL_IS_OVERLOADABLE__
5143 // }}}
5144
5145 // "private" because of the first arguments's namespace
5146 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR
5147 simd(_PrivateInit, const _MemberType& __init)
5148 : _M_data(__init) {}
5149
5150 // "private" because of the first arguments's namespace
5151 _GLIBCXX_SIMD_INTRINSIC
5152 simd(_BitsetInit, bitset<size()> __init) : _M_data()
5153 { where(mask_type(__bitset_init, __init), *this) = ~*this; }
5154
5155 _GLIBCXX_SIMD_INTRINSIC constexpr bool
5156 _M_is_constprop() const
5157 {
5158 if constexpr (__is_scalar_abi<_Abi>())
5159 return __builtin_constant_p(_M_data);
5160 else
5161 return _M_data._M_is_constprop();
5162 }
5163
5164 private:
5165 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR static mask_type
5166 _S_make_mask(typename mask_type::_MemberType __k)
5167 { return {__private_init, __k}; }
5168
5169 friend const auto& __data<value_type, abi_type>(const simd&);
5170 friend auto& __data<value_type, abi_type>(simd&);
5171 alignas(_Traits::_S_simd_align) _MemberType _M_data;
5172 };
5173
5174// }}}
5175/// @cond undocumented
5176// __data {{{
5177template <typename _Tp, typename _Ap>
5178 _GLIBCXX_SIMD_INTRINSIC constexpr const auto&
5179 __data(const simd<_Tp, _Ap>& __x)
5180 { return __x._M_data; }
5181
5182template <typename _Tp, typename _Ap>
5183 _GLIBCXX_SIMD_INTRINSIC constexpr auto&
5184 __data(simd<_Tp, _Ap>& __x)
5185 { return __x._M_data; }
5186
5187// }}}
5188namespace __float_bitwise_operators { //{{{
5189template <typename _Tp, typename _Ap>
5190 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR simd<_Tp, _Ap>
5191 operator^(const simd<_Tp, _Ap>& __a, const simd<_Tp, _Ap>& __b)
5192 {
5193 return {__private_init,
5194 _Ap::_SimdImpl::_S_bit_xor(__data(__a), __data(__b))};
5195 }
5196
5197template <typename _Tp, typename _Ap>
5198 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR simd<_Tp, _Ap>
5199 operator|(const simd<_Tp, _Ap>& __a, const simd<_Tp, _Ap>& __b)
5200 {
5201 return {__private_init,
5202 _Ap::_SimdImpl::_S_bit_or(__data(__a), __data(__b))};
5203 }
5204
5205template <typename _Tp, typename _Ap>
5206 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR simd<_Tp, _Ap>
5207 operator&(const simd<_Tp, _Ap>& __a, const simd<_Tp, _Ap>& __b)
5208 {
5209 return {__private_init,
5210 _Ap::_SimdImpl::_S_bit_and(__data(__a), __data(__b))};
5211 }
5212} // namespace __float_bitwise_operators }}}
5213/// @endcond
5214
5215/// @}
5216_GLIBCXX_SIMD_END_NAMESPACE
5217
5218#endif // __cplusplus >= 201703L
5219#endif // _GLIBCXX_EXPERIMENTAL_SIMD_H
5220
5221// vim: foldmethod=marker foldmarker={{{,}}}
constexpr duration< __common_rep_t< _Rep2, _Rep1 >, _Period > operator*(const _Rep1 &__s, const duration< _Rep2, _Period > &__d)
Definition: chrono:700
constexpr duration< __common_rep_t< _Rep1, __disable_if_is_duration< _Rep2 > >, _Period > operator%(const duration< _Rep1, _Period > &__d, const _Rep2 &__s)
Definition: chrono:729
constexpr time_point< _Clock, typename common_type< duration< _Rep1, _Period1 >, _Dur2 >::type > operator+(const duration< _Rep1, _Period1 > &__lhs, const time_point< _Clock, _Dur2 > &__rhs)
Adjust a time point forwards by the given duration.
Definition: chrono:1016
constexpr common_type< duration< _Rep1, _Period1 >, duration< _Rep2, _Period2 > >::type operator-(const duration< _Rep1, _Period1 > &__lhs, const duration< _Rep2, _Period2 > &__rhs)
The difference between two durations.
Definition: chrono:660
constexpr duration< __common_rep_t< _Rep1, __disable_if_is_duration< _Rep2 > >, _Period > operator/(const duration< _Rep1, _Period > &__d, const _Rep2 &__s)
Definition: chrono:706
typename remove_reference< _Tp >::type remove_reference_t
Alias template for remove_reference.
Definition: type_traits:1645
typename make_unsigned< _Tp >::type make_unsigned_t
Alias template for make_unsigned.
Definition: type_traits:1980
void void_t
A metafunction that always yields void, used for detecting valid types.
Definition: type_traits:2607
integral_constant< bool, true > true_type
The type used as a compile-time boolean with true value.
Definition: type_traits:83
typename conditional< _Cond, _Iftrue, _Iffalse >::type conditional_t
Alias template for conditional.
Definition: type_traits:2589
integral_constant< bool, false > false_type
The type used as a compile-time boolean with false value.
Definition: type_traits:86
typename remove_const< _Tp >::type remove_const_t
Alias template for remove_const.
Definition: type_traits:1576
typename enable_if< _Cond, _Tp >::type enable_if_t
Alias template for enable_if.
Definition: type_traits:2585
constexpr auto tuple_cat(_Tpls &&... __tpls) -> typename __tuple_cat_result< _Tpls... >::__type
tuple_cat
Definition: tuple:1732
auto declval() noexcept -> decltype(__declval< _Tp >(0))
Definition: type_traits:2364
constexpr std::remove_reference< _Tp >::type && move(_Tp &&__t) noexcept
Convert a value to an rvalue.
Definition: move.h:104
void swap(any &__x, any &__y) noexcept
Exchange the states of two any objects.
Definition: any:428
_Tp * end(valarray< _Tp > &__va) noexcept
Return an iterator pointing to one past the last element of the valarray.
Definition: valarray:1239
_Tp * begin(valarray< _Tp > &__va) noexcept
Return an iterator pointing to the first element of the valarray.
Definition: valarray:1217
constexpr const _Tp & clamp(const _Tp &, const _Tp &, const _Tp &)
Returns the value clamped between lo and hi.
Definition: stl_algo.h:3656
constexpr const _Tp & max(const _Tp &, const _Tp &)
This does what you think it does.
Definition: stl_algobase.h:254
constexpr pair< const _Tp &, const _Tp & > minmax(const _Tp &, const _Tp &)
Determines min and max at once as an ordered pair.
Definition: stl_algo.h:3301
constexpr const _Tp & min(const _Tp &, const _Tp &)
This does what you think it does.
Definition: stl_algobase.h:230
constexpr _Tp reduce(_InputIterator __first, _InputIterator __last, _Tp __init, _BinaryOperation __binary_op)
Calculate reduction of values in a range.
Definition: numeric:289
bitset< _Nb > operator&(const bitset< _Nb > &__x, const bitset< _Nb > &__y) noexcept
Global bitwise operations on bitsets.
Definition: bitset:1435
constexpr auto size(const _Container &__cont) noexcept(noexcept(__cont.size())) -> decltype(__cont.size())
Return the size of a container.
Definition: range_access.h:245
bitset< _Nb > operator|(const bitset< _Nb > &__x, const bitset< _Nb > &__y) noexcept
Global bitwise operations on bitsets.
Definition: bitset:1444
std::basic_istream< _CharT, _Traits > & operator>>(std::basic_istream< _CharT, _Traits > &__is, bitset< _Nb > &__x)
Global I/O operators for bitsets.
Definition: bitset:1472
std::basic_ostream< _CharT, _Traits > & operator<<(std::basic_ostream< _CharT, _Traits > &__os, const bitset< _Nb > &__x)
Global I/O operators for bitsets.
Definition: bitset:1540
constexpr auto data(_Container &__cont) noexcept(noexcept(__cont.data())) -> decltype(__cont.data())
Return the data pointer of a container.
Definition: range_access.h:290
bitset< _Nb > operator^(const bitset< _Nb > &__x, const bitset< _Nb > &__y) noexcept
Global bitwise operations on bitsets.
Definition: bitset:1453