libstdc++
simd_fixed_size.h
1// Simd fixed_size ABI specific implementations -*- C++ -*-
2
3// Copyright (C) 2020-2023 Free Software Foundation, Inc.
4//
5// This file is part of the GNU ISO C++ Library. This library is free
6// software; you can redistribute it and/or modify it under the
7// terms of the GNU General Public License as published by the
8// Free Software Foundation; either version 3, or (at your option)
9// any later version.
10
11// This library is distributed in the hope that it will be useful,
12// but WITHOUT ANY WARRANTY; without even the implied warranty of
13// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14// GNU General Public License for more details.
15
16// Under Section 7 of GPL version 3, you are granted additional
17// permissions described in the GCC Runtime Library Exception, version
18// 3.1, as published by the Free Software Foundation.
19
20// You should have received a copy of the GNU General Public License and
21// a copy of the GCC Runtime Library Exception along with this program;
22// see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
23// <http://www.gnu.org/licenses/>.
24
25/*
26 * The fixed_size ABI gives the following guarantees:
27 * - simd objects are passed via the stack
28 * - memory layout of `simd<_Tp, _Np>` is equivalent to `array<_Tp, _Np>`
29 * - alignment of `simd<_Tp, _Np>` is `_Np * sizeof(_Tp)` if _Np is __a
30 * power-of-2 value, otherwise `std::__bit_ceil(_Np * sizeof(_Tp))` (Note:
31 * if the alignment were to exceed the system/compiler maximum, it is bounded
32 * to that maximum)
33 * - simd_mask objects are passed like bitset<_Np>
34 * - memory layout of `simd_mask<_Tp, _Np>` is equivalent to `bitset<_Np>`
35 * - alignment of `simd_mask<_Tp, _Np>` is equal to the alignment of
36 * `bitset<_Np>`
37 */
38
39#ifndef _GLIBCXX_EXPERIMENTAL_SIMD_FIXED_SIZE_H_
40#define _GLIBCXX_EXPERIMENTAL_SIMD_FIXED_SIZE_H_
41
42#if __cplusplus >= 201703L
43
44#include <array>
45
46_GLIBCXX_SIMD_BEGIN_NAMESPACE
47
48// __simd_tuple_element {{{
49template <size_t _I, typename _Tp>
50 struct __simd_tuple_element;
51
52template <typename _Tp, typename _A0, typename... _As>
53 struct __simd_tuple_element<0, _SimdTuple<_Tp, _A0, _As...>>
54 { using type = simd<_Tp, _A0>; };
55
56template <size_t _I, typename _Tp, typename _A0, typename... _As>
57 struct __simd_tuple_element<_I, _SimdTuple<_Tp, _A0, _As...>>
58 { using type = typename __simd_tuple_element<_I - 1, _SimdTuple<_Tp, _As...>>::type; };
59
60template <size_t _I, typename _Tp>
61 using __simd_tuple_element_t = typename __simd_tuple_element<_I, _Tp>::type;
62
63// }}}
64// __simd_tuple_concat {{{
65
66template <typename _Tp, typename... _A0s, typename... _A1s>
67 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdTuple<_Tp, _A0s..., _A1s...>
68 __simd_tuple_concat(const _SimdTuple<_Tp, _A0s...>& __left,
69 const _SimdTuple<_Tp, _A1s...>& __right)
70 {
71 if constexpr (sizeof...(_A0s) == 0)
72 return __right;
73 else if constexpr (sizeof...(_A1s) == 0)
74 return __left;
75 else
76 return {__left.first, __simd_tuple_concat(__left.second, __right)};
77 }
78
79template <typename _Tp, typename _A10, typename... _A1s>
80 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdTuple<_Tp, simd_abi::scalar, _A10, _A1s...>
81 __simd_tuple_concat(const _Tp& __left, const _SimdTuple<_Tp, _A10, _A1s...>& __right)
82 { return {__left, __right}; }
83
84// }}}
85// __simd_tuple_pop_front {{{
86// Returns the next _SimdTuple in __x that has _Np elements less.
87// Precondition: _Np must match the number of elements in __first (recursively)
88template <size_t _Np, typename _Tp>
89 _GLIBCXX_SIMD_INTRINSIC constexpr decltype(auto)
90 __simd_tuple_pop_front(_Tp&& __x)
91 {
92 if constexpr (_Np == 0)
93 return static_cast<_Tp&&>(__x);
94 else
95 {
96 using _Up = __remove_cvref_t<_Tp>;
97 static_assert(_Np >= _Up::_S_first_size);
98 return __simd_tuple_pop_front<_Np - _Up::_S_first_size>(__x.second);
99 }
100 }
101
102// }}}
103// __get_simd_at<_Np> {{{1
104struct __as_simd {};
105
106struct __as_simd_tuple {};
107
108template <typename _Tp, typename _A0, typename... _Abis>
109 _GLIBCXX_SIMD_INTRINSIC constexpr simd<_Tp, _A0>
110 __simd_tuple_get_impl(__as_simd, const _SimdTuple<_Tp, _A0, _Abis...>& __t, _SizeConstant<0>)
111 { return {__private_init, __t.first}; }
112
113template <typename _Tp, typename _A0, typename... _Abis>
114 _GLIBCXX_SIMD_INTRINSIC constexpr const auto&
115 __simd_tuple_get_impl(__as_simd_tuple, const _SimdTuple<_Tp, _A0, _Abis...>& __t,
116 _SizeConstant<0>)
117 { return __t.first; }
118
119template <typename _Tp, typename _A0, typename... _Abis>
120 _GLIBCXX_SIMD_INTRINSIC constexpr auto&
121 __simd_tuple_get_impl(__as_simd_tuple, _SimdTuple<_Tp, _A0, _Abis...>& __t, _SizeConstant<0>)
122 { return __t.first; }
123
124template <typename _R, size_t _Np, typename _Tp, typename... _Abis>
125 _GLIBCXX_SIMD_INTRINSIC constexpr auto
126 __simd_tuple_get_impl(_R, const _SimdTuple<_Tp, _Abis...>& __t, _SizeConstant<_Np>)
127 { return __simd_tuple_get_impl(_R(), __t.second, _SizeConstant<_Np - 1>()); }
128
129template <size_t _Np, typename _Tp, typename... _Abis>
130 _GLIBCXX_SIMD_INTRINSIC constexpr auto&
131 __simd_tuple_get_impl(__as_simd_tuple, _SimdTuple<_Tp, _Abis...>& __t, _SizeConstant<_Np>)
132 { return __simd_tuple_get_impl(__as_simd_tuple(), __t.second, _SizeConstant<_Np - 1>()); }
133
134template <size_t _Np, typename _Tp, typename... _Abis>
135 _GLIBCXX_SIMD_INTRINSIC constexpr auto
136 __get_simd_at(const _SimdTuple<_Tp, _Abis...>& __t)
137 { return __simd_tuple_get_impl(__as_simd(), __t, _SizeConstant<_Np>()); }
138
139// }}}
140// __get_tuple_at<_Np> {{{
141template <size_t _Np, typename _Tp, typename... _Abis>
142 _GLIBCXX_SIMD_INTRINSIC constexpr auto
143 __get_tuple_at(const _SimdTuple<_Tp, _Abis...>& __t)
144 { return __simd_tuple_get_impl(__as_simd_tuple(), __t, _SizeConstant<_Np>()); }
145
146template <size_t _Np, typename _Tp, typename... _Abis>
147 _GLIBCXX_SIMD_INTRINSIC constexpr auto&
148 __get_tuple_at(_SimdTuple<_Tp, _Abis...>& __t)
149 { return __simd_tuple_get_impl(__as_simd_tuple(), __t, _SizeConstant<_Np>()); }
150
151// __tuple_element_meta {{{1
152template <typename _Tp, typename _Abi, size_t _Offset>
153 struct __tuple_element_meta : public _Abi::_SimdImpl
154 {
155 static_assert(is_same_v<typename _Abi::_SimdImpl::abi_type,
156 _Abi>); // this fails e.g. when _SimdImpl is an
157 // alias for _SimdImplBuiltin<_DifferentAbi>
158 using value_type = _Tp;
159 using abi_type = _Abi;
160 using _Traits = _SimdTraits<_Tp, _Abi>;
161 using _MaskImpl = typename _Abi::_MaskImpl;
162 using _MaskMember = typename _Traits::_MaskMember;
163 using simd_type = simd<_Tp, _Abi>;
164 static constexpr size_t _S_offset = _Offset;
165 static constexpr size_t _S_size() { return simd_size<_Tp, _Abi>::value; }
166 static constexpr _MaskImpl _S_mask_impl = {};
167
168 template <size_t _Np, bool _Sanitized>
169 _GLIBCXX_SIMD_INTRINSIC static auto
170 _S_submask(_BitMask<_Np, _Sanitized> __bits)
171 { return __bits.template _M_extract<_Offset, _S_size()>(); }
172
173 template <size_t _Np, bool _Sanitized>
174 _GLIBCXX_SIMD_INTRINSIC static _MaskMember
175 _S_make_mask(_BitMask<_Np, _Sanitized> __bits)
176 {
177 return _MaskImpl::template _S_convert<_Tp>(
178 __bits.template _M_extract<_Offset, _S_size()>()._M_sanitized());
179 }
180
181 _GLIBCXX_SIMD_INTRINSIC static _ULLong
182 _S_mask_to_shifted_ullong(_MaskMember __k)
183 { return _MaskImpl::_S_to_bits(__k).to_ullong() << _Offset; }
184 };
185
186template <size_t _Offset, typename _Tp, typename _Abi, typename... _As>
187 _GLIBCXX_SIMD_INTRINSIC
188 __tuple_element_meta<_Tp, _Abi, _Offset>
189 __make_meta(const _SimdTuple<_Tp, _Abi, _As...>&)
190 { return {}; }
191
192// }}}1
193// _WithOffset wrapper class {{{
194template <size_t _Offset, typename _Base>
195 struct _WithOffset : public _Base
196 {
197 static inline constexpr size_t _S_offset = _Offset;
198
199 _GLIBCXX_SIMD_INTRINSIC char*
200 _M_as_charptr()
201 { return reinterpret_cast<char*>(this) + _S_offset * sizeof(typename _Base::value_type); }
202
203 _GLIBCXX_SIMD_INTRINSIC const char*
204 _M_as_charptr() const
205 { return reinterpret_cast<const char*>(this) + _S_offset * sizeof(typename _Base::value_type); }
206 };
207
208// make _WithOffset<_WithOffset> ill-formed to use:
209template <size_t _O0, size_t _O1, typename _Base>
210 struct _WithOffset<_O0, _WithOffset<_O1, _Base>> {};
211
212template <size_t _Offset, typename _Tp>
213 _GLIBCXX_SIMD_INTRINSIC
214 decltype(auto)
215 __add_offset(_Tp& __base)
216 { return static_cast<_WithOffset<_Offset, __remove_cvref_t<_Tp>>&>(__base); }
217
218template <size_t _Offset, typename _Tp>
219 _GLIBCXX_SIMD_INTRINSIC
220 decltype(auto)
221 __add_offset(const _Tp& __base)
222 { return static_cast<const _WithOffset<_Offset, __remove_cvref_t<_Tp>>&>(__base); }
223
224template <size_t _Offset, size_t _ExistingOffset, typename _Tp>
225 _GLIBCXX_SIMD_INTRINSIC
226 decltype(auto)
227 __add_offset(_WithOffset<_ExistingOffset, _Tp>& __base)
228 { return static_cast<_WithOffset<_Offset + _ExistingOffset, _Tp>&>(static_cast<_Tp&>(__base)); }
229
230template <size_t _Offset, size_t _ExistingOffset, typename _Tp>
231 _GLIBCXX_SIMD_INTRINSIC
232 decltype(auto)
233 __add_offset(const _WithOffset<_ExistingOffset, _Tp>& __base)
234 {
235 return static_cast<const _WithOffset<_Offset + _ExistingOffset, _Tp>&>(
236 static_cast<const _Tp&>(__base));
237 }
238
239template <typename _Tp>
240 constexpr inline size_t __offset = 0;
241
242template <size_t _Offset, typename _Tp>
243 constexpr inline size_t __offset<_WithOffset<_Offset, _Tp>>
244 = _WithOffset<_Offset, _Tp>::_S_offset;
245
246template <typename _Tp>
247 constexpr inline size_t __offset<const _Tp> = __offset<_Tp>;
248
249template <typename _Tp>
250 constexpr inline size_t __offset<_Tp&> = __offset<_Tp>;
251
252template <typename _Tp>
253 constexpr inline size_t __offset<_Tp&&> = __offset<_Tp>;
254
255// }}}
256// _SimdTuple specializations {{{1
257// empty {{{2
258template <typename _Tp>
259 struct _SimdTuple<_Tp>
260 {
261 using value_type = _Tp;
262 static constexpr size_t _S_tuple_size = 0;
263 static constexpr size_t _S_size() { return 0; }
264 };
265
266// _SimdTupleData {{{2
267template <typename _FirstType, typename _SecondType>
268 struct _SimdTupleData
269 {
270 _FirstType first;
271 _SecondType second;
272
273 _GLIBCXX_SIMD_INTRINSIC
274 constexpr bool
275 _M_is_constprop() const
276 {
277 if constexpr (is_class_v<_FirstType>)
278 return first._M_is_constprop() && second._M_is_constprop();
279 else
280 return __builtin_constant_p(first) && second._M_is_constprop();
281 }
282 };
283
284template <typename _FirstType, typename _Tp>
285 struct _SimdTupleData<_FirstType, _SimdTuple<_Tp>>
286 {
287 _FirstType first;
288 static constexpr _SimdTuple<_Tp> second = {};
289
290 _GLIBCXX_SIMD_INTRINSIC
291 constexpr bool
292 _M_is_constprop() const
293 {
294 if constexpr (is_class_v<_FirstType>)
295 return first._M_is_constprop();
296 else
297 return __builtin_constant_p(first);
298 }
299 };
300
301// 1 or more {{{2
302template <typename _Tp, typename _Abi0, typename... _Abis>
303 struct _SimdTuple<_Tp, _Abi0, _Abis...>
304 : _SimdTupleData<typename _SimdTraits<_Tp, _Abi0>::_SimdMember,
305 _SimdTuple<_Tp, _Abis...>>
306 {
307 static_assert(!__is_fixed_size_abi_v<_Abi0>);
308 using value_type = _Tp;
309 using _FirstType = typename _SimdTraits<_Tp, _Abi0>::_SimdMember;
310 using _FirstAbi = _Abi0;
311 using _SecondType = _SimdTuple<_Tp, _Abis...>;
312 static constexpr size_t _S_tuple_size = sizeof...(_Abis) + 1;
313
314 static constexpr size_t _S_size()
315 { return simd_size_v<_Tp, _Abi0> + _SecondType::_S_size(); }
316
317 static constexpr size_t _S_first_size = simd_size_v<_Tp, _Abi0>;
318 static constexpr bool _S_is_homogeneous = (is_same_v<_Abi0, _Abis> && ...);
319
320 using _Base = _SimdTupleData<typename _SimdTraits<_Tp, _Abi0>::_SimdMember,
321 _SimdTuple<_Tp, _Abis...>>;
322 using _Base::first;
323 using _Base::second;
324
325 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdTuple() = default;
326 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdTuple(const _SimdTuple&) = default;
327 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdTuple& operator=(const _SimdTuple&)
328 = default;
329
330 template <typename _Up>
331 _GLIBCXX_SIMD_INTRINSIC constexpr
332 _SimdTuple(_Up&& __x)
333 : _Base{static_cast<_Up&&>(__x)} {}
334
335 template <typename _Up, typename _Up2>
336 _GLIBCXX_SIMD_INTRINSIC constexpr
337 _SimdTuple(_Up&& __x, _Up2&& __y)
338 : _Base{static_cast<_Up&&>(__x), static_cast<_Up2&&>(__y)} {}
339
340 template <typename _Up>
341 _GLIBCXX_SIMD_INTRINSIC constexpr
342 _SimdTuple(_Up&& __x, _SimdTuple<_Tp>)
343 : _Base{static_cast<_Up&&>(__x)} {}
344
345 _GLIBCXX_SIMD_INTRINSIC char*
346 _M_as_charptr()
347 { return reinterpret_cast<char*>(this); }
348
349 _GLIBCXX_SIMD_INTRINSIC const char*
350 _M_as_charptr() const
351 { return reinterpret_cast<const char*>(this); }
352
353 template <size_t _Np>
354 _GLIBCXX_SIMD_INTRINSIC constexpr auto&
355 _M_at()
356 {
357 if constexpr (_Np == 0)
358 return first;
359 else
360 return second.template _M_at<_Np - 1>();
361 }
362
363 template <size_t _Np>
364 _GLIBCXX_SIMD_INTRINSIC constexpr const auto&
365 _M_at() const
366 {
367 if constexpr (_Np == 0)
368 return first;
369 else
370 return second.template _M_at<_Np - 1>();
371 }
372
373 template <size_t _Np>
374 _GLIBCXX_SIMD_INTRINSIC constexpr auto
375 _M_simd_at() const
376 {
377 if constexpr (_Np == 0)
378 return simd<_Tp, _Abi0>(__private_init, first);
379 else
380 return second.template _M_simd_at<_Np - 1>();
381 }
382
383 template <size_t _Offset = 0, typename _Fp>
384 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdTuple
385 _S_generate(_Fp&& __gen, _SizeConstant<_Offset> = {})
386 {
387 auto&& __first = __gen(__tuple_element_meta<_Tp, _Abi0, _Offset>());
388 if constexpr (_S_tuple_size == 1)
389 return {__first};
390 else
391 return {__first,
392 _SecondType::_S_generate(
393 static_cast<_Fp&&>(__gen),
394 _SizeConstant<_Offset + simd_size_v<_Tp, _Abi0>>())};
395 }
396
397 template <size_t _Offset = 0, typename _Fp, typename... _More>
398 _GLIBCXX_SIMD_INTRINSIC _SimdTuple
399 _M_apply_wrapped(_Fp&& __fun, const _More&... __more) const
400 {
401 auto&& __first
402 = __fun(__make_meta<_Offset>(*this), first, __more.first...);
403 if constexpr (_S_tuple_size == 1)
404 return {__first};
405 else
406 return {
407 __first,
408 second.template _M_apply_wrapped<_Offset + simd_size_v<_Tp, _Abi0>>(
409 static_cast<_Fp&&>(__fun), __more.second...)};
410 }
411
412 template <typename _Tup>
413 _GLIBCXX_SIMD_INTRINSIC constexpr decltype(auto)
414 _M_extract_argument(_Tup&& __tup) const
415 {
416 using _TupT = typename __remove_cvref_t<_Tup>::value_type;
417 if constexpr (is_same_v<_SimdTuple, __remove_cvref_t<_Tup>>)
418 return __tup.first;
419 else if (__builtin_is_constant_evaluated())
420 return __fixed_size_storage_t<_TupT, _S_first_size>::_S_generate(
421 [&](auto __meta) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
422 return __meta._S_generator(
423 [&](auto __i) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
424 return __tup[__i];
425 }, static_cast<_TupT*>(nullptr));
426 });
427 else
428 return [&]() { // not always_inline; allow the compiler to decide
429 __fixed_size_storage_t<_TupT, _S_first_size> __r;
430 __builtin_memcpy(__r._M_as_charptr(), __tup._M_as_charptr(),
431 sizeof(__r));
432 return __r;
433 }();
434 }
435
436 template <typename _Tup>
437 _GLIBCXX_SIMD_INTRINSIC constexpr auto&
438 _M_skip_argument(_Tup&& __tup) const
439 {
440 static_assert(_S_tuple_size > 1);
441 using _Up = __remove_cvref_t<_Tup>;
442 constexpr size_t __off = __offset<_Up>;
443 if constexpr (_S_first_size == _Up::_S_first_size && __off == 0)
444 return __tup.second;
445 else if constexpr (_S_first_size > _Up::_S_first_size
446 && _S_first_size % _Up::_S_first_size == 0
447 && __off == 0)
448 return __simd_tuple_pop_front<_S_first_size>(__tup);
449 else if constexpr (_S_first_size + __off < _Up::_S_first_size)
450 return __add_offset<_S_first_size>(__tup);
451 else if constexpr (_S_first_size + __off == _Up::_S_first_size)
452 return __tup.second;
453 else
454 __assert_unreachable<_Tup>();
455 }
456
457 template <size_t _Offset, typename... _More>
458 _GLIBCXX_SIMD_INTRINSIC constexpr void
459 _M_assign_front(const _SimdTuple<_Tp, _Abi0, _More...>& __x) &
460 {
461 static_assert(_Offset == 0);
462 first = __x.first;
463 if constexpr (sizeof...(_More) > 0)
464 {
465 static_assert(sizeof...(_Abis) >= sizeof...(_More));
466 second.template _M_assign_front<0>(__x.second);
467 }
468 }
469
470 template <size_t _Offset>
471 _GLIBCXX_SIMD_INTRINSIC constexpr void
472 _M_assign_front(const _FirstType& __x) &
473 {
474 static_assert(_Offset == 0);
475 first = __x;
476 }
477
478 template <size_t _Offset, typename... _As>
479 _GLIBCXX_SIMD_INTRINSIC constexpr void
480 _M_assign_front(const _SimdTuple<_Tp, _As...>& __x) &
481 {
482 __builtin_memcpy(_M_as_charptr() + _Offset * sizeof(value_type),
483 __x._M_as_charptr(),
484 sizeof(_Tp) * _SimdTuple<_Tp, _As...>::_S_size());
485 }
486
487 /*
488 * Iterate over the first objects in this _SimdTuple and call __fun for each
489 * of them. If additional arguments are passed via __more, chunk them into
490 * _SimdTuple or __vector_type_t objects of the same number of values.
491 */
492 template <typename _Fp, typename... _More>
493 _GLIBCXX_SIMD_INTRINSIC constexpr _SimdTuple
494 _M_apply_per_chunk(_Fp&& __fun, _More&&... __more) const
495 {
496 if constexpr ((...
497 || conjunction_v<
498 is_lvalue_reference<_More>,
499 negation<is_const<remove_reference_t<_More>>>>) )
500 {
501 // need to write back at least one of __more after calling __fun
502 auto&& __first = [&](auto... __args) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
503 auto __r = __fun(__tuple_element_meta<_Tp, _Abi0, 0>(), first,
504 __args...);
505 [[maybe_unused]] auto&& __ignore_me = {(
506 [](auto&& __dst, const auto& __src) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
507 if constexpr (is_assignable_v<decltype(__dst),
508 decltype(__dst)>)
509 {
510 __dst.template _M_assign_front<__offset<decltype(__dst)>>(
511 __src);
512 }
513 }(static_cast<_More&&>(__more), __args),
514 0)...};
515 return __r;
516 }(_M_extract_argument(__more)...);
517 if constexpr (_S_tuple_size == 1)
518 return {__first};
519 else
520 return {__first,
521 second._M_apply_per_chunk(static_cast<_Fp&&>(__fun),
522 _M_skip_argument(__more)...)};
523 }
524 else
525 {
526 auto&& __first = __fun(__tuple_element_meta<_Tp, _Abi0, 0>(), first,
527 _M_extract_argument(__more)...);
528 if constexpr (_S_tuple_size == 1)
529 return {__first};
530 else
531 return {__first,
532 second._M_apply_per_chunk(static_cast<_Fp&&>(__fun),
533 _M_skip_argument(__more)...)};
534 }
535 }
536
537 template <typename _R = _Tp, typename _Fp, typename... _More>
538 _GLIBCXX_SIMD_INTRINSIC auto
539 _M_apply_r(_Fp&& __fun, const _More&... __more) const
540 {
541 auto&& __first = __fun(__tuple_element_meta<_Tp, _Abi0, 0>(), first,
542 __more.first...);
543 if constexpr (_S_tuple_size == 1)
544 return __first;
545 else
546 return __simd_tuple_concat<_R>(
547 __first, second.template _M_apply_r<_R>(static_cast<_Fp&&>(__fun),
548 __more.second...));
549 }
550
551 template <typename _Fp, typename... _More>
552 _GLIBCXX_SIMD_INTRINSIC constexpr friend _SanitizedBitMask<_S_size()>
553 _M_test(const _Fp& __fun, const _SimdTuple& __x, const _More&... __more)
554 {
555 const _SanitizedBitMask<_S_first_size> __first
556 = _Abi0::_MaskImpl::_S_to_bits(
557 __fun(__tuple_element_meta<_Tp, _Abi0, 0>(), __x.first,
558 __more.first...));
559 if constexpr (_S_tuple_size == 1)
560 return __first;
561 else
562 return _M_test(__fun, __x.second, __more.second...)
563 ._M_prepend(__first);
564 }
565
566 template <typename _Up, _Up _I>
567 _GLIBCXX_SIMD_INTRINSIC constexpr _Tp
568 operator[](integral_constant<_Up, _I>) const noexcept
569 {
570 if constexpr (_I < simd_size_v<_Tp, _Abi0>)
571 return _M_subscript_read(_I);
572 else
573 return second[integral_constant<_Up, _I - simd_size_v<_Tp, _Abi0>>()];
574 }
575
576 _GLIBCXX_SIMD_INTRINSIC _Tp
577 operator[](size_t __i) const noexcept
578 {
579 if constexpr (_S_tuple_size == 1)
580 return _M_subscript_read(__i);
581 else
582 {
583#ifdef _GLIBCXX_SIMD_USE_ALIASING_LOADS
584 return reinterpret_cast<const __may_alias<_Tp>*>(this)[__i];
585#else
586 if constexpr (__is_scalar_abi<_Abi0>())
587 {
588 const _Tp* ptr = &first;
589 return ptr[__i];
590 }
591 else
592 return __i < simd_size_v<_Tp, _Abi0>
593 ? _M_subscript_read(__i)
594 : second[__i - simd_size_v<_Tp, _Abi0>];
595#endif
596 }
597 }
598
599 _GLIBCXX_SIMD_INTRINSIC void
600 _M_set(size_t __i, _Tp __val) noexcept
601 {
602 if constexpr (_S_tuple_size == 1)
603 return _M_subscript_write(__i, __val);
604 else
605 {
606#ifdef _GLIBCXX_SIMD_USE_ALIASING_LOADS
607 reinterpret_cast<__may_alias<_Tp>*>(this)[__i] = __val;
608#else
609 if (__i < simd_size_v<_Tp, _Abi0>)
610 _M_subscript_write(__i, __val);
611 else
612 second._M_set(__i - simd_size_v<_Tp, _Abi0>, __val);
613#endif
614 }
615 }
616
617 private:
618 // _M_subscript_read/_write {{{
619 _GLIBCXX_SIMD_INTRINSIC _Tp
620 _M_subscript_read([[maybe_unused]] size_t __i) const noexcept
621 {
622 if constexpr (__is_vectorizable_v<_FirstType>)
623 return first;
624 else
625 return first[__i];
626 }
627
628 _GLIBCXX_SIMD_INTRINSIC void
629 _M_subscript_write([[maybe_unused]] size_t __i, _Tp __y) noexcept
630 {
631 if constexpr (__is_vectorizable_v<_FirstType>)
632 first = __y;
633 else
634 first._M_set(__i, __y);
635 }
636
637 // }}}
638 };
639
640// __make_simd_tuple {{{1
641template <typename _Tp, typename _A0>
642 _GLIBCXX_SIMD_INTRINSIC _SimdTuple<_Tp, _A0>
643 __make_simd_tuple(simd<_Tp, _A0> __x0)
644 { return {__data(__x0)}; }
645
646template <typename _Tp, typename _A0, typename... _As>
647 _GLIBCXX_SIMD_INTRINSIC _SimdTuple<_Tp, _A0, _As...>
648 __make_simd_tuple(const simd<_Tp, _A0>& __x0, const simd<_Tp, _As>&... __xs)
649 { return {__data(__x0), __make_simd_tuple(__xs...)}; }
650
651template <typename _Tp, typename _A0>
652 _GLIBCXX_SIMD_INTRINSIC _SimdTuple<_Tp, _A0>
653 __make_simd_tuple(const typename _SimdTraits<_Tp, _A0>::_SimdMember& __arg0)
654 { return {__arg0}; }
655
656template <typename _Tp, typename _A0, typename _A1, typename... _Abis>
657 _GLIBCXX_SIMD_INTRINSIC _SimdTuple<_Tp, _A0, _A1, _Abis...>
658 __make_simd_tuple(
659 const typename _SimdTraits<_Tp, _A0>::_SimdMember& __arg0,
660 const typename _SimdTraits<_Tp, _A1>::_SimdMember& __arg1,
661 const typename _SimdTraits<_Tp, _Abis>::_SimdMember&... __args)
662 { return {__arg0, __make_simd_tuple<_Tp, _A1, _Abis...>(__arg1, __args...)}; }
663
664// __to_simd_tuple {{{1
665template <typename _Tp, size_t _Np, typename _V, size_t _NV, typename... _VX>
666 _GLIBCXX_SIMD_INTRINSIC constexpr __fixed_size_storage_t<_Tp, _Np>
667 __to_simd_tuple(const array<_V, _NV>& __from, const _VX... __fromX);
668
669template <typename _Tp, size_t _Np,
670 size_t _Offset = 0, // skip this many elements in __from0
671 typename _R = __fixed_size_storage_t<_Tp, _Np>, typename _V0,
672 typename _V0VT = _VectorTraits<_V0>, typename... _VX>
673 _GLIBCXX_SIMD_INTRINSIC _R constexpr __to_simd_tuple(const _V0 __from0, const _VX... __fromX)
674 {
675 static_assert(is_same_v<typename _V0VT::value_type, _Tp>);
676 static_assert(_Offset < _V0VT::_S_full_size);
677 using _R0 = __vector_type_t<_Tp, _R::_S_first_size>;
678 if constexpr (_R::_S_tuple_size == 1)
679 {
680 if constexpr (_Np == 1)
681 return _R{__from0[_Offset]};
682 else if constexpr (_Offset == 0 && _V0VT::_S_full_size >= _Np)
683 return _R{__intrin_bitcast<_R0>(__from0)};
684 else if constexpr (_Offset * 2 == _V0VT::_S_full_size
685 && _V0VT::_S_full_size / 2 >= _Np)
686 return _R{__intrin_bitcast<_R0>(__extract_part<1, 2>(__from0))};
687 else if constexpr (_Offset * 4 == _V0VT::_S_full_size
688 && _V0VT::_S_full_size / 4 >= _Np)
689 return _R{__intrin_bitcast<_R0>(__extract_part<1, 4>(__from0))};
690 else
691 __assert_unreachable<_Tp>();
692 }
693 else
694 {
695 if constexpr (1 == _R::_S_first_size)
696 { // extract one scalar and recurse
697 if constexpr (_Offset + 1 < _V0VT::_S_full_size)
698 return _R{__from0[_Offset],
699 __to_simd_tuple<_Tp, _Np - 1, _Offset + 1>(__from0,
700 __fromX...)};
701 else
702 return _R{__from0[_Offset],
703 __to_simd_tuple<_Tp, _Np - 1, 0>(__fromX...)};
704 }
705
706 // place __from0 into _R::first and recurse for __fromX -> _R::second
707 else if constexpr (_V0VT::_S_full_size == _R::_S_first_size
708 && _Offset == 0)
709 return _R{__from0,
710 __to_simd_tuple<_Tp, _Np - _R::_S_first_size>(__fromX...)};
711
712 // place lower part of __from0 into _R::first and recurse with _Offset
713 else if constexpr (_V0VT::_S_full_size > _R::_S_first_size
714 && _Offset == 0)
715 return _R{__intrin_bitcast<_R0>(__from0),
716 __to_simd_tuple<_Tp, _Np - _R::_S_first_size,
717 _R::_S_first_size>(__from0, __fromX...)};
718
719 // place lower part of second quarter of __from0 into _R::first and
720 // recurse with _Offset
721 else if constexpr (_Offset * 4 == _V0VT::_S_full_size
722 && _V0VT::_S_full_size >= 4 * _R::_S_first_size)
723 return _R{__intrin_bitcast<_R0>(__extract_part<2, 4>(__from0)),
724 __to_simd_tuple<_Tp, _Np - _R::_S_first_size,
725 _Offset + _R::_S_first_size>(__from0,
726 __fromX...)};
727
728 // place lower half of high half of __from0 into _R::first and recurse
729 // with _Offset
730 else if constexpr (_Offset * 2 == _V0VT::_S_full_size
731 && _V0VT::_S_full_size >= 4 * _R::_S_first_size)
732 return _R{__intrin_bitcast<_R0>(__extract_part<2, 4>(__from0)),
733 __to_simd_tuple<_Tp, _Np - _R::_S_first_size,
734 _Offset + _R::_S_first_size>(__from0,
735 __fromX...)};
736
737 // place high half of __from0 into _R::first and recurse with __fromX
738 else if constexpr (_Offset * 2 == _V0VT::_S_full_size
739 && _V0VT::_S_full_size / 2 >= _R::_S_first_size)
740 return _R{__intrin_bitcast<_R0>(__extract_part<1, 2>(__from0)),
741 __to_simd_tuple<_Tp, _Np - _R::_S_first_size, 0>(
742 __fromX...)};
743
744 // ill-formed if some unforseen pattern is needed
745 else
746 __assert_unreachable<_Tp>();
747 }
748 }
749
750template <typename _Tp, size_t _Np, typename _V, size_t _NV, typename... _VX>
751 _GLIBCXX_SIMD_INTRINSIC constexpr __fixed_size_storage_t<_Tp, _Np>
752 __to_simd_tuple(const array<_V, _NV>& __from, const _VX... __fromX)
753 {
754 if constexpr (is_same_v<_Tp, _V>)
755 {
756 static_assert(
757 sizeof...(_VX) == 0,
758 "An array of scalars must be the last argument to __to_simd_tuple");
759 return __call_with_subscripts(
760 __from, make_index_sequence<_NV>(),
761 [&](const auto... __args) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
762 return __simd_tuple_concat(
763 _SimdTuple<_Tp, simd_abi::scalar>{__args}..., _SimdTuple<_Tp>());
764 });
765 }
766 else
767 return __call_with_subscripts(
768 __from, make_index_sequence<_NV>(),
769 [&](const auto... __args) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
770 return __to_simd_tuple<_Tp, _Np>(__args..., __fromX...);
771 });
772 }
773
774template <size_t, typename _Tp>
775 using __to_tuple_helper = _Tp;
776
777template <typename _Tp, typename _A0, size_t _NOut, size_t _Np,
778 size_t... _Indexes>
779 _GLIBCXX_SIMD_INTRINSIC __fixed_size_storage_t<_Tp, _NOut>
780 __to_simd_tuple_impl(index_sequence<_Indexes...>,
781 const array<__vector_type_t<_Tp, simd_size_v<_Tp, _A0>>, _Np>& __args)
782 {
783 return __make_simd_tuple<_Tp, __to_tuple_helper<_Indexes, _A0>...>(
784 __args[_Indexes]...);
785 }
786
787template <typename _Tp, typename _A0, size_t _NOut, size_t _Np,
788 typename _R = __fixed_size_storage_t<_Tp, _NOut>>
789 _GLIBCXX_SIMD_INTRINSIC _R
790 __to_simd_tuple_sized(
791 const array<__vector_type_t<_Tp, simd_size_v<_Tp, _A0>>, _Np>& __args)
792 {
793 static_assert(_Np * simd_size_v<_Tp, _A0> >= _NOut);
794 return __to_simd_tuple_impl<_Tp, _A0, _NOut>(
795 make_index_sequence<_R::_S_tuple_size>(), __args);
796 }
797
798// __optimize_simd_tuple {{{1
799template <typename _Tp>
800 _GLIBCXX_SIMD_INTRINSIC _SimdTuple<_Tp>
801 __optimize_simd_tuple(const _SimdTuple<_Tp>)
802 { return {}; }
803
804template <typename _Tp, typename _Ap>
805 _GLIBCXX_SIMD_INTRINSIC const _SimdTuple<_Tp, _Ap>&
806 __optimize_simd_tuple(const _SimdTuple<_Tp, _Ap>& __x)
807 { return __x; }
808
809template <typename _Tp, typename _A0, typename _A1, typename... _Abis,
810 typename _R = __fixed_size_storage_t<
811 _Tp, _SimdTuple<_Tp, _A0, _A1, _Abis...>::_S_size()>>
812 _GLIBCXX_SIMD_INTRINSIC _R
813 __optimize_simd_tuple(const _SimdTuple<_Tp, _A0, _A1, _Abis...>& __x)
814 {
815 using _Tup = _SimdTuple<_Tp, _A0, _A1, _Abis...>;
816 if constexpr (is_same_v<_R, _Tup>)
817 return __x;
818 else if constexpr (is_same_v<typename _R::_FirstType,
819 typename _Tup::_FirstType>)
820 return {__x.first, __optimize_simd_tuple(__x.second)};
821 else if constexpr (__is_scalar_abi<_A0>()
822 || _A0::template _S_is_partial<_Tp>)
823 return {__generate_from_n_evaluations<_R::_S_first_size,
824 typename _R::_FirstType>(
825 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA { return __x[__i]; }),
826 __optimize_simd_tuple(
827 __simd_tuple_pop_front<_R::_S_first_size>(__x))};
828 else if constexpr (is_same_v<_A0, _A1>
829 && _R::_S_first_size == simd_size_v<_Tp, _A0> + simd_size_v<_Tp, _A1>)
830 return {__concat(__x.template _M_at<0>(), __x.template _M_at<1>()),
831 __optimize_simd_tuple(__x.second.second)};
832 else if constexpr (sizeof...(_Abis) >= 2
833 && _R::_S_first_size == (4 * simd_size_v<_Tp, _A0>)
834 && simd_size_v<_Tp, _A0> == __simd_tuple_element_t<
835 (sizeof...(_Abis) >= 2 ? 3 : 0), _Tup>::size())
836 return {
837 __concat(__concat(__x.template _M_at<0>(), __x.template _M_at<1>()),
838 __concat(__x.template _M_at<2>(), __x.template _M_at<3>())),
839 __optimize_simd_tuple(__x.second.second.second.second)};
840 else
841 {
842 static_assert(sizeof(_R) == sizeof(__x));
843 _R __r;
844 __builtin_memcpy(__r._M_as_charptr(), __x._M_as_charptr(),
845 sizeof(_Tp) * _R::_S_size());
846 return __r;
847 }
848 }
849
850// __for_each(const _SimdTuple &, Fun) {{{1
851template <size_t _Offset = 0, typename _Tp, typename _A0, typename _Fp>
852 _GLIBCXX_SIMD_INTRINSIC constexpr void
853 __for_each(const _SimdTuple<_Tp, _A0>& __t, _Fp&& __fun)
854 { static_cast<_Fp&&>(__fun)(__make_meta<_Offset>(__t), __t.first); }
855
856template <size_t _Offset = 0, typename _Tp, typename _A0, typename _A1,
857 typename... _As, typename _Fp>
858 _GLIBCXX_SIMD_INTRINSIC constexpr void
859 __for_each(const _SimdTuple<_Tp, _A0, _A1, _As...>& __t, _Fp&& __fun)
860 {
861 __fun(__make_meta<_Offset>(__t), __t.first);
862 __for_each<_Offset + simd_size<_Tp, _A0>::value>(__t.second,
863 static_cast<_Fp&&>(__fun));
864 }
865
866// __for_each(_SimdTuple &, Fun) {{{1
867template <size_t _Offset = 0, typename _Tp, typename _A0, typename _Fp>
868 _GLIBCXX_SIMD_INTRINSIC constexpr void
869 __for_each(_SimdTuple<_Tp, _A0>& __t, _Fp&& __fun)
870 { static_cast<_Fp&&>(__fun)(__make_meta<_Offset>(__t), __t.first); }
871
872template <size_t _Offset = 0, typename _Tp, typename _A0, typename _A1,
873 typename... _As, typename _Fp>
874 _GLIBCXX_SIMD_INTRINSIC constexpr void
875 __for_each(_SimdTuple<_Tp, _A0, _A1, _As...>& __t, _Fp&& __fun)
876 {
877 __fun(__make_meta<_Offset>(__t), __t.first);
878 __for_each<_Offset + simd_size<_Tp, _A0>::value>(__t.second,
879 static_cast<_Fp&&>(__fun));
880 }
881
882// __for_each(_SimdTuple &, const _SimdTuple &, Fun) {{{1
883template <size_t _Offset = 0, typename _Tp, typename _A0, typename _Fp>
884 _GLIBCXX_SIMD_INTRINSIC constexpr void
885 __for_each(_SimdTuple<_Tp, _A0>& __a, const _SimdTuple<_Tp, _A0>& __b, _Fp&& __fun)
886 { static_cast<_Fp&&>(__fun)(__make_meta<_Offset>(__a), __a.first, __b.first); }
887
888template <size_t _Offset = 0, typename _Tp, typename _A0, typename _A1,
889 typename... _As, typename _Fp>
890 _GLIBCXX_SIMD_INTRINSIC constexpr void
891 __for_each(_SimdTuple<_Tp, _A0, _A1, _As...>& __a,
892 const _SimdTuple<_Tp, _A0, _A1, _As...>& __b, _Fp&& __fun)
893 {
894 __fun(__make_meta<_Offset>(__a), __a.first, __b.first);
895 __for_each<_Offset + simd_size<_Tp, _A0>::value>(__a.second, __b.second,
896 static_cast<_Fp&&>(__fun));
897 }
898
899// __for_each(const _SimdTuple &, const _SimdTuple &, Fun) {{{1
900template <size_t _Offset = 0, typename _Tp, typename _A0, typename _Fp>
901 _GLIBCXX_SIMD_INTRINSIC constexpr void
902 __for_each(const _SimdTuple<_Tp, _A0>& __a, const _SimdTuple<_Tp, _A0>& __b, _Fp&& __fun)
903 { static_cast<_Fp&&>(__fun)(__make_meta<_Offset>(__a), __a.first, __b.first); }
904
905template <size_t _Offset = 0, typename _Tp, typename _A0, typename _A1,
906 typename... _As, typename _Fp>
907 _GLIBCXX_SIMD_INTRINSIC constexpr void
908 __for_each(const _SimdTuple<_Tp, _A0, _A1, _As...>& __a,
909 const _SimdTuple<_Tp, _A0, _A1, _As...>& __b, _Fp&& __fun)
910 {
911 __fun(__make_meta<_Offset>(__a), __a.first, __b.first);
912 __for_each<_Offset + simd_size<_Tp, _A0>::value>(__a.second, __b.second,
913 static_cast<_Fp&&>(__fun));
914 }
915
916// }}}1
917// __extract_part(_SimdTuple) {{{
918template <int _Index, int _Total, int _Combine, typename _Tp, typename _A0, typename... _As>
919 _GLIBCXX_SIMD_INTRINSIC auto // __vector_type_t or _SimdTuple
920 __extract_part(const _SimdTuple<_Tp, _A0, _As...>& __x)
921 {
922 // worst cases:
923 // (a) 4, 4, 4 => 3, 3, 3, 3 (_Total = 4)
924 // (b) 2, 2, 2 => 3, 3 (_Total = 2)
925 // (c) 4, 2 => 2, 2, 2 (_Total = 3)
926 using _Tuple = _SimdTuple<_Tp, _A0, _As...>;
927 static_assert(_Index + _Combine <= _Total && _Index >= 0 && _Total >= 1);
928 constexpr size_t _Np = _Tuple::_S_size();
929 static_assert(_Np >= _Total && _Np % _Total == 0);
930 constexpr size_t __values_per_part = _Np / _Total;
931 [[maybe_unused]] constexpr size_t __values_to_skip
932 = _Index * __values_per_part;
933 constexpr size_t __return_size = __values_per_part * _Combine;
934 using _RetAbi = simd_abi::deduce_t<_Tp, __return_size>;
935
936 // handle (optimize) the simple cases
937 if constexpr (_Index == 0 && _Tuple::_S_first_size == __return_size)
938 return __x.first._M_data;
939 else if constexpr (_Index == 0 && _Total == _Combine)
940 return __x;
941 else if constexpr (_Index == 0 && _Tuple::_S_first_size >= __return_size)
942 return __intrin_bitcast<__vector_type_t<_Tp, __return_size>>(
943 __as_vector(__x.first));
944
945 // recurse to skip unused data members at the beginning of _SimdTuple
946 else if constexpr (__values_to_skip >= _Tuple::_S_first_size)
947 { // recurse
948 if constexpr (_Tuple::_S_first_size % __values_per_part == 0)
949 {
950 constexpr int __parts_in_first
951 = _Tuple::_S_first_size / __values_per_part;
952 return __extract_part<_Index - __parts_in_first,
953 _Total - __parts_in_first, _Combine>(
954 __x.second);
955 }
956 else
957 return __extract_part<__values_to_skip - _Tuple::_S_first_size,
958 _Np - _Tuple::_S_first_size, __return_size>(
959 __x.second);
960 }
961
962 // extract from multiple _SimdTuple data members
963 else if constexpr (__return_size > _Tuple::_S_first_size - __values_to_skip)
964 {
965#ifdef _GLIBCXX_SIMD_USE_ALIASING_LOADS
966 const __may_alias<_Tp>* const element_ptr
967 = reinterpret_cast<const __may_alias<_Tp>*>(&__x) + __values_to_skip;
968 return __as_vector(simd<_Tp, _RetAbi>(element_ptr, element_aligned));
969#else
970 [[maybe_unused]] constexpr size_t __offset = __values_to_skip;
971 return __as_vector(simd<_Tp, _RetAbi>(
972 [&](auto __i) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
973 constexpr _SizeConstant<__i + __offset> __k;
974 return __x[__k];
975 }));
976#endif
977 }
978
979 // all of the return values are in __x.first
980 else if constexpr (_Tuple::_S_first_size % __values_per_part == 0)
981 return __extract_part<_Index, _Tuple::_S_first_size / __values_per_part,
982 _Combine>(__x.first);
983 else
984 return __extract_part<__values_to_skip, _Tuple::_S_first_size,
985 _Combine * __values_per_part>(__x.first);
986 }
987
988// }}}
989// __fixed_size_storage_t<_Tp, _Np>{{{
990template <typename _Tp, int _Np, typename _Tuple,
991 typename _Next = simd<_Tp, _AllNativeAbis::_BestAbi<_Tp, _Np>>,
992 int _Remain = _Np - int(_Next::size())>
993 struct __fixed_size_storage_builder;
994
995template <typename _Tp, int _Np>
996 struct __fixed_size_storage
997 : public __fixed_size_storage_builder<_Tp, _Np, _SimdTuple<_Tp>> {};
998
999template <typename _Tp, int _Np, typename... _As, typename _Next>
1000 struct __fixed_size_storage_builder<_Tp, _Np, _SimdTuple<_Tp, _As...>, _Next,
1001 0>
1002 { using type = _SimdTuple<_Tp, _As..., typename _Next::abi_type>; };
1003
1004template <typename _Tp, int _Np, typename... _As, typename _Next, int _Remain>
1005 struct __fixed_size_storage_builder<_Tp, _Np, _SimdTuple<_Tp, _As...>, _Next,
1006 _Remain>
1007 {
1008 using type = typename __fixed_size_storage_builder<
1009 _Tp, _Remain, _SimdTuple<_Tp, _As..., typename _Next::abi_type>>::type;
1010 };
1011
1012// }}}
1013// __autocvt_to_simd {{{
1014template <typename _Tp, bool = is_arithmetic_v<__remove_cvref_t<_Tp>>>
1015 struct __autocvt_to_simd
1016 {
1017 _Tp _M_data;
1018 using _TT = __remove_cvref_t<_Tp>;
1019
1020 _GLIBCXX_SIMD_INTRINSIC
1021 operator _TT()
1022 { return _M_data; }
1023
1024 _GLIBCXX_SIMD_INTRINSIC
1025 operator _TT&()
1026 {
1027 static_assert(is_lvalue_reference<_Tp>::value, "");
1028 static_assert(!is_const<_Tp>::value, "");
1029 return _M_data;
1030 }
1031
1032 _GLIBCXX_SIMD_INTRINSIC
1033 operator _TT*()
1034 {
1035 static_assert(is_lvalue_reference<_Tp>::value, "");
1036 static_assert(!is_const<_Tp>::value, "");
1037 return &_M_data;
1038 }
1039
1040 _GLIBCXX_SIMD_INTRINSIC constexpr
1041 __autocvt_to_simd(_Tp dd) : _M_data(dd) {}
1042
1043 template <typename _Abi>
1044 _GLIBCXX_SIMD_INTRINSIC
1045 operator simd<typename _TT::value_type, _Abi>()
1046 { return {__private_init, _M_data}; }
1047
1048 template <typename _Abi>
1049 _GLIBCXX_SIMD_INTRINSIC
1050 operator simd<typename _TT::value_type, _Abi>&()
1051 { return *reinterpret_cast<simd<typename _TT::value_type, _Abi>*>(&_M_data); }
1052
1053 template <typename _Abi>
1054 _GLIBCXX_SIMD_INTRINSIC
1055 operator simd<typename _TT::value_type, _Abi>*()
1056 { return reinterpret_cast<simd<typename _TT::value_type, _Abi>*>(&_M_data); }
1057 };
1058
1059template <typename _Tp>
1060 __autocvt_to_simd(_Tp &&) -> __autocvt_to_simd<_Tp>;
1061
1062template <typename _Tp>
1063 struct __autocvt_to_simd<_Tp, true>
1064 {
1065 using _TT = __remove_cvref_t<_Tp>;
1066 _Tp _M_data;
1067 fixed_size_simd<_TT, 1> _M_fd;
1068
1069 _GLIBCXX_SIMD_INTRINSIC
1070 constexpr __autocvt_to_simd(_Tp dd) : _M_data(dd), _M_fd(_M_data) {}
1071
1072 _GLIBCXX_SIMD_INTRINSIC
1073 ~__autocvt_to_simd()
1074 { _M_data = __data(_M_fd).first; }
1075
1076 _GLIBCXX_SIMD_INTRINSIC
1077 operator fixed_size_simd<_TT, 1>()
1078 { return _M_fd; }
1079
1080 _GLIBCXX_SIMD_INTRINSIC
1081 operator fixed_size_simd<_TT, 1> &()
1082 {
1083 static_assert(is_lvalue_reference<_Tp>::value, "");
1084 static_assert(!is_const<_Tp>::value, "");
1085 return _M_fd;
1086 }
1087
1088 _GLIBCXX_SIMD_INTRINSIC
1089 operator fixed_size_simd<_TT, 1> *()
1090 {
1091 static_assert(is_lvalue_reference<_Tp>::value, "");
1092 static_assert(!is_const<_Tp>::value, "");
1093 return &_M_fd;
1094 }
1095 };
1096
1097// }}}
1098
1099struct _CommonImplFixedSize;
1100template <int _Np, typename = __detail::__odr_helper> struct _SimdImplFixedSize;
1101template <int _Np, typename = __detail::__odr_helper> struct _MaskImplFixedSize;
1102// simd_abi::_Fixed {{{
1103template <int _Np>
1104 struct simd_abi::_Fixed
1105 {
1106 template <typename _Tp> static constexpr size_t _S_size = _Np;
1107 template <typename _Tp> static constexpr size_t _S_full_size = _Np;
1108 // validity traits {{{
1109 struct _IsValidAbiTag : public __bool_constant<(_Np > 0)> {};
1110
1111 template <typename _Tp>
1112 struct _IsValidSizeFor
1113 : __bool_constant<(_Np <= simd_abi::max_fixed_size<_Tp>)> {};
1114
1115 template <typename _Tp>
1116 struct _IsValid : conjunction<_IsValidAbiTag, __is_vectorizable<_Tp>,
1117 _IsValidSizeFor<_Tp>> {};
1118
1119 template <typename _Tp>
1120 static constexpr bool _S_is_valid_v = _IsValid<_Tp>::value;
1121
1122 // }}}
1123 // _S_masked {{{
1124 _GLIBCXX_SIMD_INTRINSIC static constexpr _SanitizedBitMask<_Np>
1125 _S_masked(_BitMask<_Np> __x)
1126 { return __x._M_sanitized(); }
1127
1128 _GLIBCXX_SIMD_INTRINSIC static constexpr _SanitizedBitMask<_Np>
1129 _S_masked(_SanitizedBitMask<_Np> __x)
1130 { return __x; }
1131
1132 // }}}
1133 // _*Impl {{{
1134 using _CommonImpl = _CommonImplFixedSize;
1135 using _SimdImpl = _SimdImplFixedSize<_Np>;
1136 using _MaskImpl = _MaskImplFixedSize<_Np>;
1137
1138 // }}}
1139 // __traits {{{
1140 template <typename _Tp, bool = _S_is_valid_v<_Tp>>
1141 struct __traits : _InvalidTraits {};
1142
1143 template <typename _Tp>
1144 struct __traits<_Tp, true>
1145 {
1146 using _IsValid = true_type;
1147 using _SimdImpl = _SimdImplFixedSize<_Np>;
1148 using _MaskImpl = _MaskImplFixedSize<_Np>;
1149
1150 // simd and simd_mask member types {{{
1151 using _SimdMember = __fixed_size_storage_t<_Tp, _Np>;
1152 using _MaskMember = _SanitizedBitMask<_Np>;
1153
1154 static constexpr size_t _S_simd_align
1155 = std::__bit_ceil(_Np * sizeof(_Tp));
1156
1157 static constexpr size_t _S_mask_align = alignof(_MaskMember);
1158
1159 // }}}
1160 // _SimdBase / base class for simd, providing extra conversions {{{
1161 struct _SimdBase
1162 {
1163 // The following ensures, function arguments are passed via the stack.
1164 // This is important for ABI compatibility across TU boundaries
1165 _GLIBCXX_SIMD_ALWAYS_INLINE
1166 _SimdBase(const _SimdBase&) {}
1167 _SimdBase() = default;
1168
1169 _GLIBCXX_SIMD_ALWAYS_INLINE explicit
1170 operator const _SimdMember &() const
1171 { return static_cast<const simd<_Tp, _Fixed>*>(this)->_M_data; }
1172
1173 _GLIBCXX_SIMD_ALWAYS_INLINE explicit
1174 operator array<_Tp, _Np>() const
1175 {
1176 array<_Tp, _Np> __r;
1177 // _SimdMember can be larger because of higher alignment
1178 static_assert(sizeof(__r) <= sizeof(_SimdMember), "");
1179 __builtin_memcpy(__r.data(), &static_cast<const _SimdMember&>(*this),
1180 sizeof(__r));
1181 return __r;
1182 }
1183 };
1184
1185 // }}}
1186 // _MaskBase {{{
1187 // empty. The bitset interface suffices
1188 struct _MaskBase {};
1189
1190 // }}}
1191 // _SimdCastType {{{
1192 struct _SimdCastType
1193 {
1194 _GLIBCXX_SIMD_ALWAYS_INLINE
1195 _SimdCastType(const array<_Tp, _Np>&);
1196
1197 _GLIBCXX_SIMD_ALWAYS_INLINE
1198 _SimdCastType(const _SimdMember& dd) : _M_data(dd) {}
1199
1200 _GLIBCXX_SIMD_ALWAYS_INLINE explicit
1201 operator const _SimdMember &() const { return _M_data; }
1202
1203 private:
1204 const _SimdMember& _M_data;
1205 };
1206
1207 // }}}
1208 // _MaskCastType {{{
1209 class _MaskCastType
1210 {
1211 _MaskCastType() = delete;
1212 };
1213 // }}}
1214 };
1215 // }}}
1216 };
1217
1218// }}}
1219// _CommonImplFixedSize {{{
1220struct _CommonImplFixedSize
1221{
1222 // _S_store {{{
1223 template <typename _Tp, typename... _As>
1224 _GLIBCXX_SIMD_INTRINSIC static void
1225 _S_store(const _SimdTuple<_Tp, _As...>& __x, void* __addr)
1226 {
1227 constexpr size_t _Np = _SimdTuple<_Tp, _As...>::_S_size();
1228 __builtin_memcpy(__addr, &__x, _Np * sizeof(_Tp));
1229 }
1230
1231 // }}}
1232};
1233
1234// }}}
1235// _SimdImplFixedSize {{{1
1236// fixed_size should not inherit from _SimdMathFallback in order for
1237// specializations in the used _SimdTuple Abis to get used
1238template <int _Np, typename>
1239 struct _SimdImplFixedSize
1240 {
1241 // member types {{{2
1242 using _MaskMember = _SanitizedBitMask<_Np>;
1243
1244 template <typename _Tp>
1245 using _SimdMember = __fixed_size_storage_t<_Tp, _Np>;
1246
1247 template <typename _Tp>
1248 static constexpr size_t _S_tuple_size = _SimdMember<_Tp>::_S_tuple_size;
1249
1250 template <typename _Tp>
1251 using _Simd = simd<_Tp, simd_abi::fixed_size<_Np>>;
1252
1253 template <typename _Tp>
1254 using _TypeTag = _Tp*;
1255
1256 // broadcast {{{2
1257 template <typename _Tp>
1258 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdMember<_Tp>
1259 _S_broadcast(_Tp __x) noexcept
1260 {
1261 return _SimdMember<_Tp>::_S_generate(
1262 [&](auto __meta) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1263 return __meta._S_broadcast(__x);
1264 });
1265 }
1266
1267 // _S_generator {{{2
1268 template <typename _Fp, typename _Tp>
1269 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdMember<_Tp>
1270 _S_generator(_Fp&& __gen, _TypeTag<_Tp>)
1271 {
1272 return _SimdMember<_Tp>::_S_generate(
1273 [&__gen](auto __meta) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1274 return __meta._S_generator(
1275 [&](auto __i) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1276 return __i < _Np ? __gen(_SizeConstant<__meta._S_offset + __i>())
1277 : 0;
1278 },
1279 _TypeTag<_Tp>());
1280 });
1281 }
1282
1283 // _S_load {{{2
1284 template <typename _Tp, typename _Up>
1285 _GLIBCXX_SIMD_INTRINSIC static _SimdMember<_Tp>
1286 _S_load(const _Up* __mem, _TypeTag<_Tp>) noexcept
1287 {
1288 return _SimdMember<_Tp>::_S_generate(
1289 [&](auto __meta) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1290 return __meta._S_load(&__mem[__meta._S_offset], _TypeTag<_Tp>());
1291 });
1292 }
1293
1294 // _S_masked_load {{{2
1295 template <typename _Tp, typename... _As, typename _Up>
1296 _GLIBCXX_SIMD_INTRINSIC static _SimdTuple<_Tp, _As...>
1297 _S_masked_load(const _SimdTuple<_Tp, _As...>& __old,
1298 const _MaskMember __bits, const _Up* __mem) noexcept
1299 {
1300 auto __merge = __old;
1301 __for_each(__merge, [&](auto __meta, auto& __native) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1302 if (__meta._S_submask(__bits).any())
1303#pragma GCC diagnostic push
1304 // __mem + __mem._S_offset could be UB ([expr.add]/4.3, but it punts
1305 // the responsibility for avoiding UB to the caller of the masked load
1306 // via the mask. Consequently, the compiler may assume this branch is
1307 // unreachable, if the pointer arithmetic is UB.
1308#pragma GCC diagnostic ignored "-Warray-bounds"
1309 __native
1310 = __meta._S_masked_load(__native, __meta._S_make_mask(__bits),
1311 __mem + __meta._S_offset);
1312#pragma GCC diagnostic pop
1313 });
1314 return __merge;
1315 }
1316
1317 // _S_store {{{2
1318 template <typename _Tp, typename _Up>
1319 _GLIBCXX_SIMD_INTRINSIC static void
1320 _S_store(const _SimdMember<_Tp>& __v, _Up* __mem, _TypeTag<_Tp>) noexcept
1321 {
1322 __for_each(__v, [&](auto __meta, auto __native) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1323 __meta._S_store(__native, &__mem[__meta._S_offset], _TypeTag<_Tp>());
1324 });
1325 }
1326
1327 // _S_masked_store {{{2
1328 template <typename _Tp, typename... _As, typename _Up>
1329 _GLIBCXX_SIMD_INTRINSIC static void
1330 _S_masked_store(const _SimdTuple<_Tp, _As...>& __v, _Up* __mem,
1331 const _MaskMember __bits) noexcept
1332 {
1333 __for_each(__v, [&](auto __meta, auto __native) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1334 if (__meta._S_submask(__bits).any())
1335#pragma GCC diagnostic push
1336 // __mem + __mem._S_offset could be UB ([expr.add]/4.3, but it punts
1337 // the responsibility for avoiding UB to the caller of the masked
1338 // store via the mask. Consequently, the compiler may assume this
1339 // branch is unreachable, if the pointer arithmetic is UB.
1340#pragma GCC diagnostic ignored "-Warray-bounds"
1341 __meta._S_masked_store(__native, __mem + __meta._S_offset,
1342 __meta._S_make_mask(__bits));
1343#pragma GCC diagnostic pop
1344 });
1345 }
1346
1347 // negation {{{2
1348 template <typename _Tp, typename... _As>
1349 static inline _MaskMember
1350 _S_negate(const _SimdTuple<_Tp, _As...>& __x) noexcept
1351 {
1352 _MaskMember __bits = 0;
1353 __for_each(
1354 __x, [&__bits](auto __meta, auto __native) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1355 __bits
1356 |= __meta._S_mask_to_shifted_ullong(__meta._S_negate(__native));
1357 });
1358 return __bits;
1359 }
1360
1361 // reductions {{{2
1362 template <typename _Tp, typename _BinaryOperation>
1363 static constexpr inline _Tp _S_reduce(const _Simd<_Tp>& __x,
1364 const _BinaryOperation& __binary_op)
1365 {
1366 using _Tup = _SimdMember<_Tp>;
1367 const _Tup& __tup = __data(__x);
1368 if constexpr (_Tup::_S_tuple_size == 1)
1369 return _Tup::_FirstAbi::_SimdImpl::_S_reduce(
1370 __tup.template _M_simd_at<0>(), __binary_op);
1371 else if constexpr (_Tup::_S_tuple_size == 2 && _Tup::_S_size() > 2
1372 && _Tup::_SecondType::_S_size() == 1)
1373 {
1374 return __binary_op(simd<_Tp, simd_abi::scalar>(
1375 reduce(__tup.template _M_simd_at<0>(),
1376 __binary_op)),
1377 __tup.template _M_simd_at<1>())[0];
1378 }
1379 else if constexpr (_Tup::_S_tuple_size == 2 && _Tup::_S_size() > 4
1380 && _Tup::_SecondType::_S_size() == 2)
1381 {
1382 return __binary_op(
1383 simd<_Tp, simd_abi::scalar>(
1384 reduce(__tup.template _M_simd_at<0>(), __binary_op)),
1385 simd<_Tp, simd_abi::scalar>(
1386 reduce(__tup.template _M_simd_at<1>(), __binary_op)))[0];
1387 }
1388 else
1389 {
1390 const auto& __x2 = __call_with_n_evaluations<
1391 __div_roundup(_Tup::_S_tuple_size, 2)>(
1392 [](auto __first_simd, auto... __remaining) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1393 if constexpr (sizeof...(__remaining) == 0)
1394 return __first_simd;
1395 else
1396 {
1397 using _Tup2
1398 = _SimdTuple<_Tp,
1399 typename decltype(__first_simd)::abi_type,
1400 typename decltype(__remaining)::abi_type...>;
1401 return fixed_size_simd<_Tp, _Tup2::_S_size()>(
1402 __private_init,
1403 __make_simd_tuple(__first_simd, __remaining...));
1404 }
1405 },
1406 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1407 auto __left = __tup.template _M_simd_at<2 * __i>();
1408 if constexpr (2 * __i + 1 == _Tup::_S_tuple_size)
1409 return __left;
1410 else
1411 {
1412 auto __right = __tup.template _M_simd_at<2 * __i + 1>();
1413 using _LT = decltype(__left);
1414 using _RT = decltype(__right);
1415 if constexpr (_LT::size() == _RT::size())
1416 return __binary_op(__left, __right);
1417 else
1418 {
1419 _GLIBCXX_SIMD_USE_CONSTEXPR_API
1420 typename _LT::mask_type __k(
1421 __private_init,
1422 [](auto __j) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1423 return __j < _RT::size();
1424 });
1425 _LT __ext_right = __left;
1426 where(__k, __ext_right)
1427 = __proposed::resizing_simd_cast<_LT>(__right);
1428 where(__k, __left) = __binary_op(__left, __ext_right);
1429 return __left;
1430 }
1431 }
1432 });
1433 return reduce(__x2, __binary_op);
1434 }
1435 }
1436
1437 // _S_min, _S_max {{{2
1438 template <typename _Tp, typename... _As>
1439 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdTuple<_Tp, _As...>
1440 _S_min(const _SimdTuple<_Tp, _As...>& __a, const _SimdTuple<_Tp, _As...>& __b)
1441 {
1442 return __a._M_apply_per_chunk(
1443 [](auto __impl, auto __aa, auto __bb) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1444 return __impl._S_min(__aa, __bb);
1445 },
1446 __b);
1447 }
1448
1449 template <typename _Tp, typename... _As>
1450 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdTuple<_Tp, _As...>
1451 _S_max(const _SimdTuple<_Tp, _As...>& __a, const _SimdTuple<_Tp, _As...>& __b)
1452 {
1453 return __a._M_apply_per_chunk(
1454 [](auto __impl, auto __aa, auto __bb) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1455 return __impl._S_max(__aa, __bb);
1456 },
1457 __b);
1458 }
1459
1460 // _S_complement {{{2
1461 template <typename _Tp, typename... _As>
1462 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdTuple<_Tp, _As...>
1463 _S_complement(const _SimdTuple<_Tp, _As...>& __x) noexcept
1464 {
1465 return __x._M_apply_per_chunk(
1466 [](auto __impl, auto __xx) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1467 return __impl._S_complement(__xx);
1468 });
1469 }
1470
1471 // _S_unary_minus {{{2
1472 template <typename _Tp, typename... _As>
1473 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdTuple<_Tp, _As...>
1474 _S_unary_minus(const _SimdTuple<_Tp, _As...>& __x) noexcept
1475 {
1476 return __x._M_apply_per_chunk(
1477 [](auto __impl, auto __xx) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1478 return __impl._S_unary_minus(__xx);
1479 });
1480 }
1481
1482 // arithmetic operators {{{2
1483
1484#define _GLIBCXX_SIMD_FIXED_OP(name_, op_) \
1485 template <typename _Tp, typename... _As> \
1486 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdTuple<_Tp, _As...> name_( \
1487 const _SimdTuple<_Tp, _As...>& __x, const _SimdTuple<_Tp, _As...>& __y) \
1488 { \
1489 return __x._M_apply_per_chunk( \
1490 [](auto __impl, auto __xx, auto __yy) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA { \
1491 return __impl.name_(__xx, __yy); \
1492 }, \
1493 __y); \
1494 }
1495
1496 _GLIBCXX_SIMD_FIXED_OP(_S_plus, +)
1497 _GLIBCXX_SIMD_FIXED_OP(_S_minus, -)
1498 _GLIBCXX_SIMD_FIXED_OP(_S_multiplies, *)
1499 _GLIBCXX_SIMD_FIXED_OP(_S_divides, /)
1500 _GLIBCXX_SIMD_FIXED_OP(_S_modulus, %)
1501 _GLIBCXX_SIMD_FIXED_OP(_S_bit_and, &)
1502 _GLIBCXX_SIMD_FIXED_OP(_S_bit_or, |)
1503 _GLIBCXX_SIMD_FIXED_OP(_S_bit_xor, ^)
1504 _GLIBCXX_SIMD_FIXED_OP(_S_bit_shift_left, <<)
1505 _GLIBCXX_SIMD_FIXED_OP(_S_bit_shift_right, >>)
1506#undef _GLIBCXX_SIMD_FIXED_OP
1507
1508 template <typename _Tp, typename... _As>
1509 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdTuple<_Tp, _As...>
1510 _S_bit_shift_left(const _SimdTuple<_Tp, _As...>& __x, int __y)
1511 {
1512 return __x._M_apply_per_chunk(
1513 [__y](auto __impl, auto __xx) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1514 return __impl._S_bit_shift_left(__xx, __y);
1515 });
1516 }
1517
1518 template <typename _Tp, typename... _As>
1519 _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdTuple<_Tp, _As...>
1520 _S_bit_shift_right(const _SimdTuple<_Tp, _As...>& __x, int __y)
1521 {
1522 return __x._M_apply_per_chunk(
1523 [__y](auto __impl, auto __xx) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1524 return __impl._S_bit_shift_right(__xx, __y);
1525 });
1526 }
1527
1528 // math {{{2
1529#define _GLIBCXX_SIMD_APPLY_ON_TUPLE(_RetTp, __name) \
1530 template <typename _Tp, typename... _As, typename... _More> \
1531 static inline __fixed_size_storage_t<_RetTp, _Np> \
1532 _S_##__name(const _SimdTuple<_Tp, _As...>& __x, \
1533 const _More&... __more) \
1534 { \
1535 if constexpr (sizeof...(_More) == 0) \
1536 { \
1537 if constexpr (is_same_v<_Tp, _RetTp>) \
1538 return __x._M_apply_per_chunk( \
1539 [](auto __impl, auto __xx) \
1540 constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA \
1541 { \
1542 using _V = typename decltype(__impl)::simd_type; \
1543 return __data(__name(_V(__private_init, __xx))); \
1544 }); \
1545 else \
1546 return __optimize_simd_tuple( \
1547 __x.template _M_apply_r<_RetTp>( \
1548 [](auto __impl, auto __xx) \
1549 _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA \
1550 { return __impl._S_##__name(__xx); })); \
1551 } \
1552 else if constexpr ( \
1553 is_same_v< \
1554 _Tp, \
1555 _RetTp> && (... && is_same_v<_SimdTuple<_Tp, _As...>, _More>) ) \
1556 return __x._M_apply_per_chunk( \
1557 [](auto __impl, auto __xx, auto... __pack) \
1558 constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA \
1559 { \
1560 using _V = typename decltype(__impl)::simd_type; \
1561 return __data(__name(_V(__private_init, __xx), \
1562 _V(__private_init, __pack)...)); \
1563 }, __more...); \
1564 else if constexpr (is_same_v<_Tp, _RetTp>) \
1565 return __x._M_apply_per_chunk( \
1566 [](auto __impl, auto __xx, auto... __pack) \
1567 constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA \
1568 { \
1569 using _V = typename decltype(__impl)::simd_type; \
1570 return __data(__name(_V(__private_init, __xx), \
1571 __autocvt_to_simd(__pack)...)); \
1572 }, __more...); \
1573 else \
1574 __assert_unreachable<_Tp>(); \
1575 }
1576
1577 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, acos)
1578 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, asin)
1579 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, atan)
1580 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, atan2)
1581 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, cos)
1582 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, sin)
1583 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, tan)
1584 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, acosh)
1585 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, asinh)
1586 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, atanh)
1587 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, cosh)
1588 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, sinh)
1589 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, tanh)
1590 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, exp)
1591 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, exp2)
1592 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, expm1)
1593 _GLIBCXX_SIMD_APPLY_ON_TUPLE(int, ilogb)
1594 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, log)
1595 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, log10)
1596 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, log1p)
1597 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, log2)
1598 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, logb)
1599 // modf implemented in simd_math.h
1600 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp,
1601 scalbn) // double scalbn(double x, int exp);
1602 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, scalbln)
1603 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, cbrt)
1604 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, abs)
1605 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, fabs)
1606 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, pow)
1607 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, sqrt)
1608 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, erf)
1609 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, erfc)
1610 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, lgamma)
1611 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, tgamma)
1612 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, trunc)
1613 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, ceil)
1614 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, floor)
1615 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, nearbyint)
1616
1617 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, rint)
1618 _GLIBCXX_SIMD_APPLY_ON_TUPLE(long, lrint)
1619 _GLIBCXX_SIMD_APPLY_ON_TUPLE(long long, llrint)
1620
1621 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, round)
1622 _GLIBCXX_SIMD_APPLY_ON_TUPLE(long, lround)
1623 _GLIBCXX_SIMD_APPLY_ON_TUPLE(long long, llround)
1624
1625 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, ldexp)
1626 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, fmod)
1627 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, remainder)
1628 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, copysign)
1629 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, nextafter)
1630 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, fdim)
1631 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, fmax)
1632 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, fmin)
1633 _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, fma)
1634 _GLIBCXX_SIMD_APPLY_ON_TUPLE(int, fpclassify)
1635#undef _GLIBCXX_SIMD_APPLY_ON_TUPLE
1636
1637 template <typename _Tp, typename... _Abis>
1638 static inline _SimdTuple<_Tp, _Abis...>
1639 _S_remquo(const _SimdTuple<_Tp, _Abis...>& __x, const _SimdTuple<_Tp, _Abis...>& __y,
1640 __fixed_size_storage_t<int, _SimdTuple<_Tp, _Abis...>::_S_size()>* __z)
1641 {
1642 return __x._M_apply_per_chunk(
1643 [](auto __impl, const auto __xx, const auto __yy, auto& __zz)
1644 _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA
1645 { return __impl._S_remquo(__xx, __yy, &__zz); },
1646 __y, *__z);
1647 }
1648
1649 template <typename _Tp, typename... _As>
1650 static inline _SimdTuple<_Tp, _As...>
1651 _S_frexp(const _SimdTuple<_Tp, _As...>& __x,
1652 __fixed_size_storage_t<int, _Np>& __exp) noexcept
1653 {
1654 return __x._M_apply_per_chunk(
1655 [](auto __impl, const auto& __a, auto& __b) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1656 return __data(frexp(typename decltype(__impl)::simd_type(__private_init, __a),
1657 __autocvt_to_simd(__b)));
1658 }, __exp);
1659 }
1660
1661#define _GLIBCXX_SIMD_TEST_ON_TUPLE_(name_) \
1662 template <typename _Tp, typename... _As> \
1663 static inline _MaskMember \
1664 _S_##name_(const _SimdTuple<_Tp, _As...>& __x) noexcept \
1665 { \
1666 return _M_test([] (auto __impl, auto __xx) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA { \
1667 return __impl._S_##name_(__xx); \
1668 }, __x); \
1669 }
1670
1671 _GLIBCXX_SIMD_TEST_ON_TUPLE_(isinf)
1672 _GLIBCXX_SIMD_TEST_ON_TUPLE_(isfinite)
1673 _GLIBCXX_SIMD_TEST_ON_TUPLE_(isnan)
1674 _GLIBCXX_SIMD_TEST_ON_TUPLE_(isnormal)
1675 _GLIBCXX_SIMD_TEST_ON_TUPLE_(signbit)
1676#undef _GLIBCXX_SIMD_TEST_ON_TUPLE_
1677
1678 // _S_increment & _S_decrement{{{2
1679 template <typename... _Ts>
1680 _GLIBCXX_SIMD_INTRINSIC static constexpr void
1681 _S_increment(_SimdTuple<_Ts...>& __x)
1682 {
1683 __for_each(
1684 __x, [](auto __meta, auto& native) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1685 __meta._S_increment(native);
1686 });
1687 }
1688
1689 template <typename... _Ts>
1690 _GLIBCXX_SIMD_INTRINSIC static constexpr void
1691 _S_decrement(_SimdTuple<_Ts...>& __x)
1692 {
1693 __for_each(
1694 __x, [](auto __meta, auto& native) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1695 __meta._S_decrement(native);
1696 });
1697 }
1698
1699 // compares {{{2
1700#define _GLIBCXX_SIMD_CMP_OPERATIONS(__cmp) \
1701 template <typename _Tp, typename... _As> \
1702 _GLIBCXX_SIMD_INTRINSIC constexpr static _MaskMember \
1703 __cmp(const _SimdTuple<_Tp, _As...>& __x, \
1704 const _SimdTuple<_Tp, _As...>& __y) \
1705 { \
1706 return _M_test([](auto __impl, auto __xx, auto __yy) \
1707 constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA \
1708 { return __impl.__cmp(__xx, __yy); }, \
1709 __x, __y); \
1710 }
1711
1712 _GLIBCXX_SIMD_CMP_OPERATIONS(_S_equal_to)
1713 _GLIBCXX_SIMD_CMP_OPERATIONS(_S_not_equal_to)
1714 _GLIBCXX_SIMD_CMP_OPERATIONS(_S_less)
1715 _GLIBCXX_SIMD_CMP_OPERATIONS(_S_less_equal)
1716 _GLIBCXX_SIMD_CMP_OPERATIONS(_S_isless)
1717 _GLIBCXX_SIMD_CMP_OPERATIONS(_S_islessequal)
1718 _GLIBCXX_SIMD_CMP_OPERATIONS(_S_isgreater)
1719 _GLIBCXX_SIMD_CMP_OPERATIONS(_S_isgreaterequal)
1720 _GLIBCXX_SIMD_CMP_OPERATIONS(_S_islessgreater)
1721 _GLIBCXX_SIMD_CMP_OPERATIONS(_S_isunordered)
1722#undef _GLIBCXX_SIMD_CMP_OPERATIONS
1723
1724 // smart_reference access {{{2
1725 template <typename _Tp, typename... _As, typename _Up>
1726 _GLIBCXX_SIMD_INTRINSIC static void
1727 _S_set(_SimdTuple<_Tp, _As...>& __v, int __i, _Up&& __x) noexcept
1728 { __v._M_set(__i, static_cast<_Up&&>(__x)); }
1729
1730 // _S_masked_assign {{{2
1731 template <typename _Tp, typename... _As>
1732 _GLIBCXX_SIMD_INTRINSIC static void
1733 _S_masked_assign(const _MaskMember __bits, _SimdTuple<_Tp, _As...>& __lhs,
1734 const __type_identity_t<_SimdTuple<_Tp, _As...>>& __rhs)
1735 {
1736 __for_each(__lhs, __rhs,
1737 [&](auto __meta, auto& __native_lhs, auto __native_rhs)
1738 constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA
1739 {
1740 __meta._S_masked_assign(__meta._S_make_mask(__bits), __native_lhs,
1741 __native_rhs);
1742 });
1743 }
1744
1745 // Optimization for the case where the RHS is a scalar. No need to broadcast
1746 // the scalar to a simd first.
1747 template <typename _Tp, typename... _As>
1748 _GLIBCXX_SIMD_INTRINSIC static void
1749 _S_masked_assign(const _MaskMember __bits, _SimdTuple<_Tp, _As...>& __lhs,
1750 const __type_identity_t<_Tp> __rhs)
1751 {
1752 __for_each(
1753 __lhs, [&](auto __meta, auto& __native_lhs) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1754 __meta._S_masked_assign(__meta._S_make_mask(__bits), __native_lhs,
1755 __rhs);
1756 });
1757 }
1758
1759 // _S_masked_cassign {{{2
1760 template <typename _Op, typename _Tp, typename... _As>
1761 static inline void
1762 _S_masked_cassign(const _MaskMember __bits, _SimdTuple<_Tp, _As...>& __lhs,
1763 const _SimdTuple<_Tp, _As...>& __rhs, _Op __op)
1764 {
1765 __for_each(__lhs, __rhs,
1766 [&](auto __meta, auto& __native_lhs, auto __native_rhs)
1767 constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA
1768 {
1769 __meta.template _S_masked_cassign(__meta._S_make_mask(__bits),
1770 __native_lhs, __native_rhs, __op);
1771 });
1772 }
1773
1774 // Optimization for the case where the RHS is a scalar. No need to broadcast
1775 // the scalar to a simd first.
1776 template <typename _Op, typename _Tp, typename... _As>
1777 static inline void
1778 _S_masked_cassign(const _MaskMember __bits, _SimdTuple<_Tp, _As...>& __lhs,
1779 const _Tp& __rhs, _Op __op)
1780 {
1781 __for_each(
1782 __lhs, [&](auto __meta, auto& __native_lhs) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1783 __meta.template _S_masked_cassign(__meta._S_make_mask(__bits),
1784 __native_lhs, __rhs, __op);
1785 });
1786 }
1787
1788 // _S_masked_unary {{{2
1789 template <template <typename> class _Op, typename _Tp, typename... _As>
1790 static inline _SimdTuple<_Tp, _As...>
1791 _S_masked_unary(const _MaskMember __bits, const _SimdTuple<_Tp, _As...>& __v)
1792 {
1793 return __v._M_apply_wrapped([&__bits](auto __meta,
1794 auto __native) constexpr {
1795 return __meta.template _S_masked_unary<_Op>(__meta._S_make_mask(
1796 __bits),
1797 __native);
1798 });
1799 }
1800
1801 // }}}2
1802 };
1803
1804// _MaskImplFixedSize {{{1
1805template <int _Np, typename>
1806 struct _MaskImplFixedSize
1807 {
1808 static_assert(
1809 sizeof(_ULLong) * __CHAR_BIT__ >= _Np,
1810 "The fixed_size implementation relies on one _ULLong being able to store "
1811 "all boolean elements."); // required in load & store
1812
1813 // member types {{{
1814 using _Abi = simd_abi::fixed_size<_Np>;
1815
1816 using _MaskMember = _SanitizedBitMask<_Np>;
1817
1818 template <typename _Tp>
1819 using _FirstAbi = typename __fixed_size_storage_t<_Tp, _Np>::_FirstAbi;
1820
1821 template <typename _Tp>
1822 using _TypeTag = _Tp*;
1823
1824 // }}}
1825 // _S_broadcast {{{
1826 template <typename>
1827 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember
1828 _S_broadcast(bool __x)
1829 { return __x ? ~_MaskMember() : _MaskMember(); }
1830
1831 // }}}
1832 // _S_load {{{
1833 template <typename>
1834 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember
1835 _S_load(const bool* __mem)
1836 {
1837 using _Ip = __int_for_sizeof_t<bool>;
1838 // the following load uses element_aligned and relies on __mem already
1839 // carrying alignment information from when this load function was
1840 // called.
1841 const simd<_Ip, _Abi> __bools(reinterpret_cast<const __may_alias<_Ip>*>(
1842 __mem),
1843 element_aligned);
1844 return __data(__bools != 0);
1845 }
1846
1847 // }}}
1848 // _S_to_bits {{{
1849 template <bool _Sanitized>
1850 _GLIBCXX_SIMD_INTRINSIC static constexpr _SanitizedBitMask<_Np>
1851 _S_to_bits(_BitMask<_Np, _Sanitized> __x)
1852 {
1853 if constexpr (_Sanitized)
1854 return __x;
1855 else
1856 return __x._M_sanitized();
1857 }
1858
1859 // }}}
1860 // _S_convert {{{
1861 template <typename _Tp, typename _Up, typename _UAbi>
1862 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember
1863 _S_convert(simd_mask<_Up, _UAbi> __x)
1864 {
1865 return _UAbi::_MaskImpl::_S_to_bits(__data(__x))
1866 .template _M_extract<0, _Np>();
1867 }
1868
1869 // }}}
1870 // _S_from_bitmask {{{2
1871 template <typename _Tp>
1872 _GLIBCXX_SIMD_INTRINSIC static _MaskMember
1873 _S_from_bitmask(_MaskMember __bits, _TypeTag<_Tp>) noexcept
1874 { return __bits; }
1875
1876 // _S_load {{{2
1877 static inline _MaskMember
1878 _S_load(const bool* __mem) noexcept
1879 {
1880 // TODO: _UChar is not necessarily the best type to use here. For smaller
1881 // _Np _UShort, _UInt, _ULLong, float, and double can be more efficient.
1882 _ULLong __r = 0;
1883 using _Vs = __fixed_size_storage_t<_UChar, _Np>;
1884 __for_each(_Vs{}, [&](auto __meta, auto) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1885 __r |= __meta._S_mask_to_shifted_ullong(
1886 __meta._S_mask_impl._S_load(&__mem[__meta._S_offset],
1887 _SizeConstant<__meta._S_size()>()));
1888 });
1889 return __r;
1890 }
1891
1892 // _S_masked_load {{{2
1893 static inline _MaskMember
1894 _S_masked_load(_MaskMember __merge, _MaskMember __mask, const bool* __mem) noexcept
1895 {
1896 _BitOps::_S_bit_iteration(__mask.to_ullong(),
1897 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1898 __merge.set(__i, __mem[__i]);
1899 });
1900 return __merge;
1901 }
1902
1903 // _S_store {{{2
1904 static inline void
1905 _S_store(const _MaskMember __bitmask, bool* __mem) noexcept
1906 {
1907 if constexpr (_Np == 1)
1908 __mem[0] = __bitmask[0];
1909 else
1910 _FirstAbi<_UChar>::_CommonImpl::_S_store_bool_array(__bitmask, __mem);
1911 }
1912
1913 // _S_masked_store {{{2
1914 static inline void
1915 _S_masked_store(const _MaskMember __v, bool* __mem, const _MaskMember __k) noexcept
1916 {
1917 _BitOps::_S_bit_iteration(
1918 __k, [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA { __mem[__i] = __v[__i]; });
1919 }
1920
1921 // logical and bitwise operators {{{2
1922 _GLIBCXX_SIMD_INTRINSIC static _MaskMember
1923 _S_logical_and(const _MaskMember& __x, const _MaskMember& __y) noexcept
1924 { return __x & __y; }
1925
1926 _GLIBCXX_SIMD_INTRINSIC static _MaskMember
1927 _S_logical_or(const _MaskMember& __x, const _MaskMember& __y) noexcept
1928 { return __x | __y; }
1929
1930 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember
1931 _S_bit_not(const _MaskMember& __x) noexcept
1932 { return ~__x; }
1933
1934 _GLIBCXX_SIMD_INTRINSIC static _MaskMember
1935 _S_bit_and(const _MaskMember& __x, const _MaskMember& __y) noexcept
1936 { return __x & __y; }
1937
1938 _GLIBCXX_SIMD_INTRINSIC static _MaskMember
1939 _S_bit_or(const _MaskMember& __x, const _MaskMember& __y) noexcept
1940 { return __x | __y; }
1941
1942 _GLIBCXX_SIMD_INTRINSIC static _MaskMember
1943 _S_bit_xor(const _MaskMember& __x, const _MaskMember& __y) noexcept
1944 { return __x ^ __y; }
1945
1946 // smart_reference access {{{2
1947 _GLIBCXX_SIMD_INTRINSIC static void
1948 _S_set(_MaskMember& __k, int __i, bool __x) noexcept
1949 { __k.set(__i, __x); }
1950
1951 // _S_masked_assign {{{2
1952 _GLIBCXX_SIMD_INTRINSIC static void
1953 _S_masked_assign(const _MaskMember __k, _MaskMember& __lhs, const _MaskMember __rhs)
1954 { __lhs = (__lhs & ~__k) | (__rhs & __k); }
1955
1956 // Optimization for the case where the RHS is a scalar.
1957 _GLIBCXX_SIMD_INTRINSIC static void
1958 _S_masked_assign(const _MaskMember __k, _MaskMember& __lhs, const bool __rhs)
1959 {
1960 if (__rhs)
1961 __lhs |= __k;
1962 else
1963 __lhs &= ~__k;
1964 }
1965
1966 // }}}2
1967 // _S_all_of {{{
1968 template <typename _Tp>
1969 _GLIBCXX_SIMD_INTRINSIC static bool
1970 _S_all_of(simd_mask<_Tp, _Abi> __k)
1971 { return __data(__k).all(); }
1972
1973 // }}}
1974 // _S_any_of {{{
1975 template <typename _Tp>
1976 _GLIBCXX_SIMD_INTRINSIC static bool
1977 _S_any_of(simd_mask<_Tp, _Abi> __k)
1978 { return __data(__k).any(); }
1979
1980 // }}}
1981 // _S_none_of {{{
1982 template <typename _Tp>
1983 _GLIBCXX_SIMD_INTRINSIC static bool
1984 _S_none_of(simd_mask<_Tp, _Abi> __k)
1985 { return __data(__k).none(); }
1986
1987 // }}}
1988 // _S_some_of {{{
1989 template <typename _Tp>
1990 _GLIBCXX_SIMD_INTRINSIC static bool
1991 _S_some_of([[maybe_unused]] simd_mask<_Tp, _Abi> __k)
1992 {
1993 if constexpr (_Np == 1)
1994 return false;
1995 else
1996 return __data(__k).any() && !__data(__k).all();
1997 }
1998
1999 // }}}
2000 // _S_popcount {{{
2001 template <typename _Tp>
2002 _GLIBCXX_SIMD_INTRINSIC static int
2003 _S_popcount(simd_mask<_Tp, _Abi> __k)
2004 { return __data(__k).count(); }
2005
2006 // }}}
2007 // _S_find_first_set {{{
2008 template <typename _Tp>
2009 _GLIBCXX_SIMD_INTRINSIC static int
2010 _S_find_first_set(simd_mask<_Tp, _Abi> __k)
2011 { return std::__countr_zero(__data(__k).to_ullong()); }
2012
2013 // }}}
2014 // _S_find_last_set {{{
2015 template <typename _Tp>
2016 _GLIBCXX_SIMD_INTRINSIC static int
2017 _S_find_last_set(simd_mask<_Tp, _Abi> __k)
2018 { return std::__bit_width(__data(__k).to_ullong()) - 1; }
2019
2020 // }}}
2021 };
2022// }}}1
2023
2024_GLIBCXX_SIMD_END_NAMESPACE
2025#endif // __cplusplus >= 201703L
2026#endif // _GLIBCXX_EXPERIMENTAL_SIMD_FIXED_SIZE_H_
2027
2028// vim: foldmethod=marker sw=2 noet ts=8 sts=2 tw=80
std::reduce
constexpr _Tp reduce(_InputIterator __first, _InputIterator __last, _Tp __init, _BinaryOperation __binary_op)
Calculate reduction of values in a range.
Definition: numeric:287
__gnu_debug::__base
constexpr _Iterator __base(_Iterator __it)
Definition: helper_functions.h:306