25#ifndef _GLIBCXX_EXPERIMENTAL_SIMD_ABIS_H_
26#define _GLIBCXX_EXPERIMENTAL_SIMD_ABIS_H_
28#if __cplusplus >= 201703L
34_GLIBCXX_SIMD_BEGIN_NAMESPACE
37 static inline _GLIBCXX_SIMD_USE_CONSTEXPR _V _S_allbits
38 =
reinterpret_cast<_V
>(~__vector_type_t<char,
sizeof(_V) /
sizeof(
char)>());
42template <
typename _V,
typename = _VectorTraits<_V>>
43 static inline _GLIBCXX_SIMD_USE_CONSTEXPR _V _S_signmask
44 = __xor(_V() + 1, _V() - 1);
46template <
typename _V,
typename = _VectorTraits<_V>>
47 static inline _GLIBCXX_SIMD_USE_CONSTEXPR _V _S_absmask
48 = __andnot(_S_signmask<_V>, _S_allbits<_V>);
53template <
int... _Indices,
typename _Tp,
typename _TVT = _VectorTraits<_Tp>,
54 typename = __detail::__odr_helper>
56 __vector_permute(_Tp __x)
58 static_assert(
sizeof...(_Indices) == _TVT::_S_full_size);
59 return __make_vector<typename _TVT::value_type>(
60 (_Indices == -1 ? 0 : __x[_Indices == -1 ? 0 : _Indices])...);
66template <
int... _Indices,
typename _Tp,
typename _TVT = _VectorTraits<_Tp>,
67 typename = __detail::__odr_helper>
69 __vector_shuffle(_Tp __x, _Tp __y)
71 return _Tp{(_Indices == -1 ? 0
72 : _Indices < _TVT::_S_full_size
74 : __y[_Indices - _TVT::_S_full_size])...};
79template <
typename _Tp,
typename... _Args>
80 _GLIBCXX_SIMD_INTRINSIC
constexpr _SimdWrapper<_Tp,
sizeof...(_Args)>
81 __make_wrapper(
const _Args&... __args)
82 {
return __make_vector<_Tp>(__args...); }
86template <
typename _Tp,
size_t _ToN = 0,
typename _Up,
size_t _M,
87 size_t _Np = _ToN != 0 ? _ToN :
sizeof(_Up) * _M /
sizeof(_Tp)>
88 _GLIBCXX_SIMD_INTRINSIC
constexpr _SimdWrapper<_Tp, _Np>
89 __wrapper_bitcast(_SimdWrapper<_Up, _M> __x)
91 static_assert(_Np > 1);
92 return __intrin_bitcast<__vector_type_t<_Tp, _Np>>(__x._M_data);
98template <
unsigned __shift,
typename _Tp,
typename _TVT = _VectorTraits<_Tp>>
99 _GLIBCXX_SIMD_INTRINSIC _Tp
100 __shift_elements_right(_Tp __v)
102 [[maybe_unused]]
const auto __iv = __to_intrin(__v);
103 static_assert(__shift <=
sizeof(_Tp));
104 if constexpr (__shift == 0)
106 else if constexpr (__shift ==
sizeof(_Tp))
108#if _GLIBCXX_SIMD_X86INTRIN
109 else if constexpr (__have_sse && __shift == 8
110 && _TVT::template _S_is<float, 4>)
111 return _mm_movehl_ps(__iv, __iv);
112 else if constexpr (__have_sse2 && __shift == 8
113 && _TVT::template _S_is<double, 2>)
114 return _mm_unpackhi_pd(__iv, __iv);
115 else if constexpr (__have_sse2 &&
sizeof(_Tp) == 16)
116 return reinterpret_cast<typename _TVT::type
>(
117 _mm_srli_si128(
reinterpret_cast<__m128i
>(__iv), __shift));
118 else if constexpr (__shift == 16 &&
sizeof(_Tp) == 32)
128 return __zero_extend(__hi128(__v));
130 else if constexpr (__have_avx2 &&
sizeof(_Tp) == 32 && __shift < 16)
132 const auto __vll = __vector_bitcast<_LLong>(__v);
133 return reinterpret_cast<typename _TVT::type
>(
134 _mm256_alignr_epi8(_mm256_permute2x128_si256(__vll, __vll, 0x81),
137 else if constexpr (__have_avx &&
sizeof(_Tp) == 32 && __shift < 16)
139 const auto __vll = __vector_bitcast<_LLong>(__v);
140 return reinterpret_cast<typename _TVT::type
>(
141 __concat(_mm_alignr_epi8(__hi128(__vll), __lo128(__vll), __shift),
142 _mm_srli_si128(__hi128(__vll), __shift)));
144 else if constexpr (
sizeof(_Tp) == 32 && __shift > 16)
145 return __zero_extend(__shift_elements_right<__shift - 16>(__hi128(__v)));
146 else if constexpr (
sizeof(_Tp) == 64 && __shift == 32)
147 return __zero_extend(__hi256(__v));
148 else if constexpr (__have_avx512f &&
sizeof(_Tp) == 64)
150 if constexpr (__shift >= 48)
151 return __zero_extend(
152 __shift_elements_right<__shift - 48>(__extract<3, 4>(__v)));
153 else if constexpr (__shift >= 32)
154 return __zero_extend(
155 __shift_elements_right<__shift - 32>(__hi256(__v)));
156 else if constexpr (__shift % 8 == 0)
157 return reinterpret_cast<typename _TVT::type
>(
158 _mm512_alignr_epi64(__m512i(), __intrin_bitcast<__m512i>(__v),
160 else if constexpr (__shift % 4 == 0)
161 return reinterpret_cast<typename _TVT::type
>(
162 _mm512_alignr_epi32(__m512i(), __intrin_bitcast<__m512i>(__v),
164 else if constexpr (__have_avx512bw && __shift < 16)
166 const auto __vll = __vector_bitcast<_LLong>(__v);
167 return reinterpret_cast<typename _TVT::type
>(
168 _mm512_alignr_epi8(_mm512_shuffle_i32x4(__vll, __vll, 0xf9),
171 else if constexpr (__have_avx512bw && __shift < 32)
173 const auto __vll = __vector_bitcast<_LLong>(__v);
174 return reinterpret_cast<typename _TVT::type
>(
175 _mm512_alignr_epi8(_mm512_shuffle_i32x4(__vll, __m512i(), 0xee),
176 _mm512_shuffle_i32x4(__vll, __vll, 0xf9),
180 __assert_unreachable<_Tp>();
189 constexpr int __chunksize = __shift % 8 == 0 ? 8
190 : __shift % 4 == 0 ? 4
191 : __shift % 2 == 0 ? 2
193 auto __w = __vector_bitcast<__int_with_sizeof_t<__chunksize>>(__v);
194 using _Up =
decltype(__w);
195 return __intrin_bitcast<_Tp>(
196 __call_with_n_evaluations<(
sizeof(_Tp) - __shift) / __chunksize>(
197 [](
auto... __chunks) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
198 return _Up{__chunks...};
199 }, [&](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
200 return __w[__shift / __chunksize + __i];
207template <
int _Index,
int _Total,
int _Combine,
typename _Tp,
size_t _Np>
208 _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_CONST
constexpr
209 _SimdWrapper<_Tp, _Np / _Total * _Combine>
210 __extract_part(
const _SimdWrapper<_Tp, _Np> __x)
212 if constexpr (_Index % 2 == 0 && _Total % 2 == 0 && _Combine % 2 == 0)
213 return __extract_part<_Index / 2, _Total / 2, _Combine / 2>(__x);
216 constexpr size_t __values_per_part = _Np / _Total;
217 constexpr size_t __values_to_skip = _Index * __values_per_part;
218 constexpr size_t __return_size = __values_per_part * _Combine;
219 using _R = __vector_type_t<_Tp, __return_size>;
220 static_assert((_Index + _Combine) * __values_per_part *
sizeof(_Tp)
222 "out of bounds __extract_part");
229 if (__x._M_is_constprop())
230 return __generate_from_n_evaluations<__return_size, _R>(
231 [&](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
232 return __x[__values_to_skip + __i];
234 if constexpr (_Index == 0 && _Total == 1)
236 else if constexpr (_Index == 0)
237 return __intrin_bitcast<_R>(__as_vector(__x));
238#if _GLIBCXX_SIMD_X86INTRIN
239 else if constexpr (
sizeof(__x) == 32
240 && __return_size *
sizeof(_Tp) <= 16)
242 constexpr size_t __bytes_to_skip = __values_to_skip *
sizeof(_Tp);
243 if constexpr (__bytes_to_skip == 16)
244 return __vector_bitcast<_Tp, __return_size>(
245 __hi128(__as_vector(__x)));
247 return __vector_bitcast<_Tp, __return_size>(
248 _mm_alignr_epi8(__hi128(__vector_bitcast<_LLong>(__x)),
249 __lo128(__vector_bitcast<_LLong>(__x)),
253 else if constexpr (_Index > 0
254 && (__values_to_skip % __return_size != 0
256 && (__values_to_skip + __return_size) *
sizeof(_Tp)
258 &&
sizeof(__x) >= 16)
259 return __intrin_bitcast<_R>(
260 __shift_elements_right<__values_to_skip *
sizeof(_Tp)>(
265 __builtin_memcpy(&__r,
266 reinterpret_cast<const char*
>(&__x)
267 +
sizeof(_Tp) * __values_to_skip,
268 __return_size *
sizeof(_Tp));
276template <
int _Index,
int _Total,
int _Combine = 1,
size_t _Np>
277 _GLIBCXX_SIMD_INTRINSIC
constexpr _SimdWrapper<bool, _Np / _Total * _Combine>
278 __extract_part(
const _SimdWrapper<bool, _Np> __x)
280 static_assert(_Combine == 1,
"_Combine != 1 not implemented");
281 static_assert(__have_avx512f && _Total >= 2 && _Index + _Combine <= _Total && _Index >= 0);
282 return __x._M_data >> (_Index * _Np / _Total);
289template <
typename _To,
typename _From,
size_t... _I>
290 _GLIBCXX_SIMD_INTRINSIC
constexpr _To
291 __vector_convert(_From __a, index_sequence<_I...>)
293 using _Tp =
typename _VectorTraits<_To>::value_type;
294 return _To{
static_cast<_Tp
>(__a[_I])...};
297template <
typename _To,
typename _From,
size_t... _I>
298 _GLIBCXX_SIMD_INTRINSIC
constexpr _To
299 __vector_convert(_From __a, _From __b, index_sequence<_I...>)
301 using _Tp =
typename _VectorTraits<_To>::value_type;
302 return _To{
static_cast<_Tp
>(__a[_I])...,
static_cast<_Tp
>(__b[_I])...};
305template <
typename _To,
typename _From,
size_t... _I>
306 _GLIBCXX_SIMD_INTRINSIC
constexpr _To
307 __vector_convert(_From __a, _From __b, _From __c, index_sequence<_I...>)
309 using _Tp =
typename _VectorTraits<_To>::value_type;
310 return _To{
static_cast<_Tp
>(__a[_I])...,
static_cast<_Tp
>(__b[_I])...,
311 static_cast<_Tp
>(__c[_I])...};
314template <
typename _To,
typename _From,
size_t... _I>
315 _GLIBCXX_SIMD_INTRINSIC
constexpr _To
316 __vector_convert(_From __a, _From __b, _From __c, _From __d,
317 index_sequence<_I...>)
319 using _Tp =
typename _VectorTraits<_To>::value_type;
320 return _To{
static_cast<_Tp
>(__a[_I])...,
static_cast<_Tp
>(__b[_I])...,
321 static_cast<_Tp
>(__c[_I])...,
static_cast<_Tp
>(__d[_I])...};
324template <
typename _To,
typename _From,
size_t... _I>
325 _GLIBCXX_SIMD_INTRINSIC
constexpr _To
326 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
327 index_sequence<_I...>)
329 using _Tp =
typename _VectorTraits<_To>::value_type;
330 return _To{
static_cast<_Tp
>(__a[_I])...,
static_cast<_Tp
>(__b[_I])...,
331 static_cast<_Tp
>(__c[_I])...,
static_cast<_Tp
>(__d[_I])...,
332 static_cast<_Tp
>(__e[_I])...};
335template <
typename _To,
typename _From,
size_t... _I>
336 _GLIBCXX_SIMD_INTRINSIC
constexpr _To
337 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
338 _From __f, index_sequence<_I...>)
340 using _Tp =
typename _VectorTraits<_To>::value_type;
341 return _To{
static_cast<_Tp
>(__a[_I])...,
static_cast<_Tp
>(__b[_I])...,
342 static_cast<_Tp
>(__c[_I])...,
static_cast<_Tp
>(__d[_I])...,
343 static_cast<_Tp
>(__e[_I])...,
static_cast<_Tp
>(__f[_I])...};
346template <
typename _To,
typename _From,
size_t... _I>
347 _GLIBCXX_SIMD_INTRINSIC
constexpr _To
348 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
349 _From __f, _From __g, index_sequence<_I...>)
351 using _Tp =
typename _VectorTraits<_To>::value_type;
352 return _To{
static_cast<_Tp
>(__a[_I])...,
static_cast<_Tp
>(__b[_I])...,
353 static_cast<_Tp
>(__c[_I])...,
static_cast<_Tp
>(__d[_I])...,
354 static_cast<_Tp
>(__e[_I])...,
static_cast<_Tp
>(__f[_I])...,
355 static_cast<_Tp
>(__g[_I])...};
358template <
typename _To,
typename _From,
size_t... _I>
359 _GLIBCXX_SIMD_INTRINSIC
constexpr _To
360 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
361 _From __f, _From __g, _From __h, index_sequence<_I...>)
363 using _Tp =
typename _VectorTraits<_To>::value_type;
364 return _To{
static_cast<_Tp
>(__a[_I])...,
static_cast<_Tp
>(__b[_I])...,
365 static_cast<_Tp
>(__c[_I])...,
static_cast<_Tp
>(__d[_I])...,
366 static_cast<_Tp
>(__e[_I])...,
static_cast<_Tp
>(__f[_I])...,
367 static_cast<_Tp
>(__g[_I])...,
static_cast<_Tp
>(__h[_I])...};
370template <
typename _To,
typename _From,
size_t... _I>
371 _GLIBCXX_SIMD_INTRINSIC
constexpr _To
372 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
373 _From __f, _From __g, _From __h, _From __i,
374 index_sequence<_I...>)
376 using _Tp =
typename _VectorTraits<_To>::value_type;
377 return _To{
static_cast<_Tp
>(__a[_I])...,
static_cast<_Tp
>(__b[_I])...,
378 static_cast<_Tp
>(__c[_I])...,
static_cast<_Tp
>(__d[_I])...,
379 static_cast<_Tp
>(__e[_I])...,
static_cast<_Tp
>(__f[_I])...,
380 static_cast<_Tp
>(__g[_I])...,
static_cast<_Tp
>(__h[_I])...,
381 static_cast<_Tp
>(__i[_I])...};
384template <
typename _To,
typename _From,
size_t... _I>
385 _GLIBCXX_SIMD_INTRINSIC
constexpr _To
386 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
387 _From __f, _From __g, _From __h, _From __i, _From __j,
388 index_sequence<_I...>)
390 using _Tp =
typename _VectorTraits<_To>::value_type;
391 return _To{
static_cast<_Tp
>(__a[_I])...,
static_cast<_Tp
>(__b[_I])...,
392 static_cast<_Tp
>(__c[_I])...,
static_cast<_Tp
>(__d[_I])...,
393 static_cast<_Tp
>(__e[_I])...,
static_cast<_Tp
>(__f[_I])...,
394 static_cast<_Tp
>(__g[_I])...,
static_cast<_Tp
>(__h[_I])...,
395 static_cast<_Tp
>(__i[_I])...,
static_cast<_Tp
>(__j[_I])...};
398template <
typename _To,
typename _From,
size_t... _I>
399 _GLIBCXX_SIMD_INTRINSIC
constexpr _To
400 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
401 _From __f, _From __g, _From __h, _From __i, _From __j,
402 _From __k, index_sequence<_I...>)
404 using _Tp =
typename _VectorTraits<_To>::value_type;
405 return _To{
static_cast<_Tp
>(__a[_I])...,
static_cast<_Tp
>(__b[_I])...,
406 static_cast<_Tp
>(__c[_I])...,
static_cast<_Tp
>(__d[_I])...,
407 static_cast<_Tp
>(__e[_I])...,
static_cast<_Tp
>(__f[_I])...,
408 static_cast<_Tp
>(__g[_I])...,
static_cast<_Tp
>(__h[_I])...,
409 static_cast<_Tp
>(__i[_I])...,
static_cast<_Tp
>(__j[_I])...,
410 static_cast<_Tp
>(__k[_I])...};
413template <
typename _To,
typename _From,
size_t... _I>
414 _GLIBCXX_SIMD_INTRINSIC
constexpr _To
415 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
416 _From __f, _From __g, _From __h, _From __i, _From __j,
417 _From __k, _From __l, index_sequence<_I...>)
419 using _Tp =
typename _VectorTraits<_To>::value_type;
420 return _To{
static_cast<_Tp
>(__a[_I])...,
static_cast<_Tp
>(__b[_I])...,
421 static_cast<_Tp
>(__c[_I])...,
static_cast<_Tp
>(__d[_I])...,
422 static_cast<_Tp
>(__e[_I])...,
static_cast<_Tp
>(__f[_I])...,
423 static_cast<_Tp
>(__g[_I])...,
static_cast<_Tp
>(__h[_I])...,
424 static_cast<_Tp
>(__i[_I])...,
static_cast<_Tp
>(__j[_I])...,
425 static_cast<_Tp
>(__k[_I])...,
static_cast<_Tp
>(__l[_I])...};
428template <
typename _To,
typename _From,
size_t... _I>
429 _GLIBCXX_SIMD_INTRINSIC
constexpr _To
430 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
431 _From __f, _From __g, _From __h, _From __i, _From __j,
432 _From __k, _From __l, _From __m, index_sequence<_I...>)
434 using _Tp =
typename _VectorTraits<_To>::value_type;
435 return _To{
static_cast<_Tp
>(__a[_I])...,
static_cast<_Tp
>(__b[_I])...,
436 static_cast<_Tp
>(__c[_I])...,
static_cast<_Tp
>(__d[_I])...,
437 static_cast<_Tp
>(__e[_I])...,
static_cast<_Tp
>(__f[_I])...,
438 static_cast<_Tp
>(__g[_I])...,
static_cast<_Tp
>(__h[_I])...,
439 static_cast<_Tp
>(__i[_I])...,
static_cast<_Tp
>(__j[_I])...,
440 static_cast<_Tp
>(__k[_I])...,
static_cast<_Tp
>(__l[_I])...,
441 static_cast<_Tp
>(__m[_I])...};
444template <
typename _To,
typename _From,
size_t... _I>
445 _GLIBCXX_SIMD_INTRINSIC
constexpr _To
446 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
447 _From __f, _From __g, _From __h, _From __i, _From __j,
448 _From __k, _From __l, _From __m, _From __n,
449 index_sequence<_I...>)
451 using _Tp =
typename _VectorTraits<_To>::value_type;
452 return _To{
static_cast<_Tp
>(__a[_I])...,
static_cast<_Tp
>(__b[_I])...,
453 static_cast<_Tp
>(__c[_I])...,
static_cast<_Tp
>(__d[_I])...,
454 static_cast<_Tp
>(__e[_I])...,
static_cast<_Tp
>(__f[_I])...,
455 static_cast<_Tp
>(__g[_I])...,
static_cast<_Tp
>(__h[_I])...,
456 static_cast<_Tp
>(__i[_I])...,
static_cast<_Tp
>(__j[_I])...,
457 static_cast<_Tp
>(__k[_I])...,
static_cast<_Tp
>(__l[_I])...,
458 static_cast<_Tp
>(__m[_I])...,
static_cast<_Tp
>(__n[_I])...};
461template <
typename _To,
typename _From,
size_t... _I>
462 _GLIBCXX_SIMD_INTRINSIC
constexpr _To
463 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
464 _From __f, _From __g, _From __h, _From __i, _From __j,
465 _From __k, _From __l, _From __m, _From __n, _From __o,
466 index_sequence<_I...>)
468 using _Tp =
typename _VectorTraits<_To>::value_type;
469 return _To{
static_cast<_Tp
>(__a[_I])...,
static_cast<_Tp
>(__b[_I])...,
470 static_cast<_Tp
>(__c[_I])...,
static_cast<_Tp
>(__d[_I])...,
471 static_cast<_Tp
>(__e[_I])...,
static_cast<_Tp
>(__f[_I])...,
472 static_cast<_Tp
>(__g[_I])...,
static_cast<_Tp
>(__h[_I])...,
473 static_cast<_Tp
>(__i[_I])...,
static_cast<_Tp
>(__j[_I])...,
474 static_cast<_Tp
>(__k[_I])...,
static_cast<_Tp
>(__l[_I])...,
475 static_cast<_Tp
>(__m[_I])...,
static_cast<_Tp
>(__n[_I])...,
476 static_cast<_Tp
>(__o[_I])...};
479template <
typename _To,
typename _From,
size_t... _I>
480 _GLIBCXX_SIMD_INTRINSIC
constexpr _To
481 __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
482 _From __f, _From __g, _From __h, _From __i, _From __j,
483 _From __k, _From __l, _From __m, _From __n, _From __o,
484 _From __p, index_sequence<_I...>)
486 using _Tp =
typename _VectorTraits<_To>::value_type;
487 return _To{
static_cast<_Tp
>(__a[_I])...,
static_cast<_Tp
>(__b[_I])...,
488 static_cast<_Tp
>(__c[_I])...,
static_cast<_Tp
>(__d[_I])...,
489 static_cast<_Tp
>(__e[_I])...,
static_cast<_Tp
>(__f[_I])...,
490 static_cast<_Tp
>(__g[_I])...,
static_cast<_Tp
>(__h[_I])...,
491 static_cast<_Tp
>(__i[_I])...,
static_cast<_Tp
>(__j[_I])...,
492 static_cast<_Tp
>(__k[_I])...,
static_cast<_Tp
>(__l[_I])...,
493 static_cast<_Tp
>(__m[_I])...,
static_cast<_Tp
>(__n[_I])...,
494 static_cast<_Tp
>(__o[_I])...,
static_cast<_Tp
>(__p[_I])...};
500template <
typename _To,
typename... _From,
size_t _FromSize>
501 _GLIBCXX_SIMD_INTRINSIC
constexpr _To
502 __vector_convert(_SimdWrapper<_From, _FromSize>... __xs)
504#ifdef _GLIBCXX_SIMD_WORKAROUND_PR85048
505 using _From0 = __first_of_pack_t<_From...>;
506 using _FW = _SimdWrapper<_From0, _FromSize>;
507 if (!_FW::_S_is_partial && !(... && __xs._M_is_constprop()))
509 if constexpr ((
sizeof...(_From) & (
sizeof...(_From) - 1))
511 return __convert_x86<_To>(__as_vector(__xs)...);
513 return __vector_convert<_To>(__xs..., _FW{});
517 return __vector_convert<_To>(
518 __as_vector(__xs)...,
519 make_index_sequence<(
sizeof...(__xs) == 1 ?
std::min(
520 _VectorTraits<_To>::_S_full_size,
int(_FromSize))
526template <
typename _To,
typename _From,
typename... _More>
527 _GLIBCXX_SIMD_INTRINSIC
constexpr auto
528 __convert(_From __v0, _More... __vs)
530 static_assert((
true && ... && is_same_v<_From, _More>) );
531 if constexpr (__is_vectorizable_v<_From>)
533 using _V =
typename _VectorTraits<_To>::type;
534 using _Tp =
typename _VectorTraits<_To>::value_type;
535 return _V{
static_cast<_Tp
>(__v0),
static_cast<_Tp
>(__vs)...};
537 else if constexpr (__is_vector_type_v<_From>)
538 return __convert<_To>(__as_wrapper(__v0), __as_wrapper(__vs)...);
541 constexpr size_t __input_size = _From::_S_size * (1 +
sizeof...(_More));
542 if constexpr (__is_vectorizable_v<_To>)
543 return __convert<__vector_type_t<_To, __input_size>>(__v0, __vs...);
544 else if constexpr (!__is_vector_type_v<_To>)
545 return _To(__convert<typename _To::_BuiltinType>(__v0, __vs...));
549 sizeof...(_More) == 0
550 || _VectorTraits<_To>::_S_full_size >= __input_size,
551 "__convert(...) requires the input to fit into the output");
552 return __vector_convert<_To>(__v0, __vs...);
562template <
typename _To,
567 typename _From,
typename _FromVT = _VectorTraits<_From>>
568 _GLIBCXX_SIMD_INTRINSIC
auto
569 __convert_all(_From __v)
571 if constexpr (is_arithmetic_v<_To> && _NParts != 1)
573 static_assert(_Offset < _FromVT::_S_full_size);
575 = _NParts == 0 ? _FromVT::_S_partial_width - _Offset : _NParts;
576 return __generate_from_n_evaluations<_Np, array<_To, _Np>>(
577 [&](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
578 return static_cast<_To
>(__v[__i + _Offset]);
583 static_assert(__is_vector_type_v<_To>);
584 using _ToVT = _VectorTraits<_To>;
585 if constexpr (__is_vector_type_v<_From>)
586 return __convert_all<_To, _NParts>(__as_wrapper(__v));
587 else if constexpr (_NParts == 1)
589 static_assert(_Offset % _ToVT::_S_full_size == 0);
590 return array<_To, 1>{__vector_convert<_To>(
591 __extract_part<_Offset / _ToVT::_S_full_size,
592 __div_roundup(_FromVT::_S_partial_width,
593 _ToVT::_S_full_size)>(__v))};
595#if _GLIBCXX_SIMD_X86INTRIN
596 else if constexpr (!__have_sse4_1 && _Offset == 0
597 && is_integral_v<typename _FromVT::value_type>
598 &&
sizeof(
typename _FromVT::value_type)
599 <
sizeof(
typename _ToVT::value_type)
600 && !(
sizeof(
typename _FromVT::value_type) == 4
601 && is_same_v<typename _ToVT::value_type, double>))
603 using _ToT =
typename _ToVT::value_type;
604 using _FromT =
typename _FromVT::value_type;
608 : (_FromVT::_S_partial_width / _ToVT::_S_full_size);
609 using _R = array<_To, _Np>;
614 [[maybe_unused]]
auto __adjust
616 auto __vv) -> _SimdWrapper<_FromT,
decltype(__n)::value> {
617 return __vector_bitcast<_FromT, decltype(__n)::value>(__vv);
619 [[maybe_unused]]
const auto __vi = __to_intrin(__v);
621 = [](
auto __x0, [[maybe_unused]]
auto __x1) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
622 if constexpr (_Np == 1)
623 return _R{__intrin_bitcast<_To>(__x0)};
625 return _R{__intrin_bitcast<_To>(__x0),
626 __intrin_bitcast<_To>(__x1)};
629 if constexpr (_Np == 0)
631 else if constexpr (
sizeof(_FromT) == 1 &&
sizeof(_ToT) == 2)
633 static_assert(is_integral_v<_FromT>);
634 static_assert(is_integral_v<_ToT>);
635 if constexpr (is_unsigned_v<_FromT>)
636 return __make_array(_mm_unpacklo_epi8(__vi, __m128i()),
637 _mm_unpackhi_epi8(__vi, __m128i()));
640 _mm_srai_epi16(_mm_unpacklo_epi8(__vi, __vi), 8),
641 _mm_srai_epi16(_mm_unpackhi_epi8(__vi, __vi), 8));
643 else if constexpr (
sizeof(_FromT) == 2 &&
sizeof(_ToT) == 4)
645 static_assert(is_integral_v<_FromT>);
646 if constexpr (is_floating_point_v<_ToT>)
649 = __convert_all<__vector_type16_t<int>, _Np>(
650 __adjust(_SizeConstant<_Np * 4>(), __v));
651 return __generate_from_n_evaluations<_Np, _R>(
652 [&](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
653 return __vector_convert<_To>(__as_wrapper(__ints[__i]));
656 else if constexpr (is_unsigned_v<_FromT>)
657 return __make_array(_mm_unpacklo_epi16(__vi, __m128i()),
658 _mm_unpackhi_epi16(__vi, __m128i()));
661 _mm_srai_epi32(_mm_unpacklo_epi16(__vi, __vi), 16),
662 _mm_srai_epi32(_mm_unpackhi_epi16(__vi, __vi), 16));
664 else if constexpr (
sizeof(_FromT) == 4 &&
sizeof(_ToT) == 8
665 && is_integral_v<_FromT> && is_integral_v<_ToT>)
667 if constexpr (is_unsigned_v<_FromT>)
668 return __make_array(_mm_unpacklo_epi32(__vi, __m128i()),
669 _mm_unpackhi_epi32(__vi, __m128i()));
672 _mm_unpacklo_epi32(__vi, _mm_srai_epi32(__vi, 31)),
673 _mm_unpackhi_epi32(__vi, _mm_srai_epi32(__vi, 31)));
675 else if constexpr (
sizeof(_FromT) == 4 &&
sizeof(_ToT) == 8
676 && is_integral_v<_FromT> && is_integral_v<_ToT>)
678 if constexpr (is_unsigned_v<_FromT>)
679 return __make_array(_mm_unpacklo_epi32(__vi, __m128i()),
680 _mm_unpackhi_epi32(__vi, __m128i()));
683 _mm_unpacklo_epi32(__vi, _mm_srai_epi32(__vi, 31)),
684 _mm_unpackhi_epi32(__vi, _mm_srai_epi32(__vi, 31)));
686 else if constexpr (
sizeof(_FromT) == 1 &&
sizeof(_ToT) >= 4
687 && is_signed_v<_FromT>)
689 const __m128i __vv[2] = {_mm_unpacklo_epi8(__vi, __vi),
690 _mm_unpackhi_epi8(__vi, __vi)};
691 const __vector_type_t<int, 4> __vvvv[4] = {
692 __vector_bitcast<int>(_mm_unpacklo_epi16(__vv[0], __vv[0])),
693 __vector_bitcast<int>(_mm_unpackhi_epi16(__vv[0], __vv[0])),
694 __vector_bitcast<int>(_mm_unpacklo_epi16(__vv[1], __vv[1])),
695 __vector_bitcast<int>(_mm_unpackhi_epi16(__vv[1], __vv[1]))};
696 if constexpr (
sizeof(_ToT) == 4)
697 return __generate_from_n_evaluations<_Np, _R>(
698 [&](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
699 return __vector_convert<_To>(
700 _SimdWrapper<int, 4>(__vvvv[__i] >> 24));
702 else if constexpr (is_integral_v<_ToT>)
703 return __generate_from_n_evaluations<_Np, _R>(
704 [&](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
705 const auto __signbits = __to_intrin(__vvvv[__i / 2] >> 31);
706 const auto __sx32 = __to_intrin(__vvvv[__i / 2] >> 24);
707 return __vector_bitcast<_ToT>(
708 __i % 2 == 0 ? _mm_unpacklo_epi32(__sx32, __signbits)
709 : _mm_unpackhi_epi32(__sx32, __signbits));
712 return __generate_from_n_evaluations<_Np, _R>(
713 [&](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
714 const _SimdWrapper<int, 4> __int4 = __vvvv[__i / 2] >> 24;
715 return __vector_convert<_To>(
716 __i % 2 == 0 ? __int4
717 : _SimdWrapper<int, 4>(
718 _mm_unpackhi_epi64(__to_intrin(__int4),
719 __to_intrin(__int4))));
722 else if constexpr (
sizeof(_FromT) == 1 &&
sizeof(_ToT) == 4)
724 const auto __shorts = __convert_all<__vector_type16_t<
725 conditional_t<is_signed_v<_FromT>, short,
unsigned short>>>(
726 __adjust(_SizeConstant<(_Np + 1) / 2 * 8>(), __v));
727 return __generate_from_n_evaluations<_Np, _R>(
728 [&](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
729 return __convert_all<_To>(__shorts[__i / 2])[__i % 2];
732 else if constexpr (
sizeof(_FromT) == 2 &&
sizeof(_ToT) == 8
733 && is_signed_v<_FromT> && is_integral_v<_ToT>)
735 const __m128i __vv[2] = {_mm_unpacklo_epi16(__vi, __vi),
736 _mm_unpackhi_epi16(__vi, __vi)};
737 const __vector_type16_t<int> __vvvv[4]
738 = {__vector_bitcast<int>(
739 _mm_unpacklo_epi32(_mm_srai_epi32(__vv[0], 16),
740 _mm_srai_epi32(__vv[0], 31))),
741 __vector_bitcast<int>(
742 _mm_unpackhi_epi32(_mm_srai_epi32(__vv[0], 16),
743 _mm_srai_epi32(__vv[0], 31))),
744 __vector_bitcast<int>(
745 _mm_unpacklo_epi32(_mm_srai_epi32(__vv[1], 16),
746 _mm_srai_epi32(__vv[1], 31))),
747 __vector_bitcast<int>(
748 _mm_unpackhi_epi32(_mm_srai_epi32(__vv[1], 16),
749 _mm_srai_epi32(__vv[1], 31)))};
750 return __generate_from_n_evaluations<_Np, _R>(
751 [&](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
752 return __vector_bitcast<_ToT>(__vvvv[__i]);
755 else if constexpr (
sizeof(_FromT) <= 2 &&
sizeof(_ToT) == 8)
759 is_signed_v<_FromT> || is_floating_point_v<_ToT>, int,
761 __adjust(_SizeConstant<(_Np + 1) / 2 * 4>(), __v));
762 return __generate_from_n_evaluations<_Np, _R>(
763 [&](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
764 return __convert_all<_To>(__ints[__i / 2])[__i % 2];
768 __assert_unreachable<_To>();
771 else if constexpr ((_FromVT::_S_partial_width - _Offset)
772 > _ToVT::_S_full_size)
781 constexpr size_t _NTotal
782 = (_FromVT::_S_partial_width - _Offset) / _ToVT::_S_full_size;
783 constexpr size_t _Np = _NParts == 0 ? _NTotal : _NParts;
786 || (_Np == _NTotal + 1
787 && (_FromVT::_S_partial_width - _Offset) % _ToVT::_S_full_size
789 using _R = array<_To, _Np>;
790 if constexpr (_Np == 1)
791 return _R{__vector_convert<_To>(
792 __extract_part<_Offset, _FromVT::_S_partial_width,
793 _ToVT::_S_full_size>(__v))};
795 return __generate_from_n_evaluations<_Np, _R>(
796 [&](
auto __i)
constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
798 = __extract_part<__i * _ToVT::_S_full_size + _Offset,
799 _FromVT::_S_partial_width,
800 _ToVT::_S_full_size>(__v);
801 return __vector_convert<_To>(__part);
804 else if constexpr (_Offset == 0)
805 return array<_To, 1>{__vector_convert<_To>(__v)};
807 return array<_To, 1>{__vector_convert<_To>(
808 __extract_part<_Offset, _FromVT::_S_partial_width,
809 _FromVT::_S_partial_width - _Offset>(__v))};
816template <
typename _Tp,
typename _Mp,
typename _Abi,
size_t _Np>
820 using _SimdImpl =
typename _Abi::_SimdImpl;
821 using _MaskImpl =
typename _Abi::_MaskImpl;
824 using _SimdMember = _SimdWrapper<_Tp, _Np>;
825 using _MaskMember = _SimdWrapper<_Mp, _Np>;
826 static constexpr size_t _S_simd_align =
alignof(_SimdMember);
827 static constexpr size_t _S_mask_align =
alignof(_MaskMember);
831 static constexpr size_t _S_full_size = _SimdMember::_S_full_size;
832 static constexpr bool _S_is_partial = _SimdMember::_S_is_partial;
838 _GLIBCXX_SIMD_ALWAYS_INLINE
explicit
839 operator __intrinsic_type_t<_Tp, _Np>()
const
840 {
return __to_intrin(
static_cast<const simd<_Tp, _Abi>*
>(
this)->_M_data); }
842 _GLIBCXX_SIMD_ALWAYS_INLINE
explicit
843 operator __vector_type_t<_Tp, _Np>()
const
844 {
return __data(*
static_cast<const simd<_Tp, _Abi>*
>(
this)); }
849 _GLIBCXX_SIMD_ALWAYS_INLINE
explicit
850 operator __intrinsic_type_t<_Tp, _Np>()
const
851 {
return __data(*
static_cast<const simd<_Tp, _Abi>*
>(
this)); }
855 is_same<__intrinsic_type_t<_Tp, _Np>, __vector_type_t<_Tp, _Np>>::value,
856 _SimdBase1, _SimdBase2>;
862 _GLIBCXX_SIMD_ALWAYS_INLINE
explicit
863 operator __intrinsic_type_t<_Tp, _Np>()
const
864 {
return static_cast<const simd_mask<_Tp, _Abi>*
>(
this) ->_M_data.__intrin(); }
866 _GLIBCXX_SIMD_ALWAYS_INLINE
explicit
867 operator __vector_type_t<_Tp, _Np>()
const
868 {
return static_cast<const simd_mask<_Tp, _Abi>*
>(
this)->_M_data._M_data; }
873 _GLIBCXX_SIMD_ALWAYS_INLINE
explicit
874 operator __intrinsic_type_t<_Tp, _Np>()
const
875 {
return __data(*
static_cast<const simd_mask<_Tp, _Abi>*
>(
this)); }
879 is_same<__intrinsic_type_t<_Tp, _Np>, __vector_type_t<_Tp, _Np>>::value,
880 _MaskBase1, _MaskBase2>;
887 using _Up = __intrinsic_type_t<_Tp, _Np>;
891 _GLIBCXX_SIMD_ALWAYS_INLINE
892 _MaskCastType(_Up __x) : _M_data(__x) {}
894 _GLIBCXX_SIMD_ALWAYS_INLINE
895 operator _MaskMember()
const {
return _M_data; }
903 using _Ap = __intrinsic_type_t<_Tp, _Np>;
907 _GLIBCXX_SIMD_ALWAYS_INLINE
constexpr
908 _SimdCastType1(_Ap __a) : _M_data(__vector_bitcast<_Tp>(__a)) {}
910 _GLIBCXX_SIMD_ALWAYS_INLINE
constexpr
911 operator _SimdMember()
const {
return _M_data; }
916 using _Ap = __intrinsic_type_t<_Tp, _Np>;
917 using _Bp = __vector_type_t<_Tp, _Np>;
921 _GLIBCXX_SIMD_ALWAYS_INLINE
constexpr
922 _SimdCastType2(_Ap __a) : _M_data(__vector_bitcast<_Tp>(__a)) {}
924 _GLIBCXX_SIMD_ALWAYS_INLINE
constexpr
925 _SimdCastType2(_Bp __b) : _M_data(__b) {}
927 _GLIBCXX_SIMD_ALWAYS_INLINE
constexpr
928 operator _SimdMember()
const {
return _M_data; }
932 is_same<__intrinsic_type_t<_Tp, _Np>, __vector_type_t<_Tp, _Np>>::value,
933 _SimdCastType1, _SimdCastType2>;
938struct _CommonImplX86;
939struct _CommonImplNeon;
940struct _CommonImplBuiltin;
941template <
typename _Abi,
typename = __detail::__odr_helper>
struct _SimdImplBuiltin;
942template <
typename _Abi,
typename = __detail::__odr_helper>
struct _MaskImplBuiltin;
943template <
typename _Abi,
typename = __detail::__odr_helper>
struct _SimdImplX86;
944template <
typename _Abi,
typename = __detail::__odr_helper>
struct _MaskImplX86;
945template <
typename _Abi,
typename = __detail::__odr_helper>
struct _SimdImplNeon;
946template <
typename _Abi,
typename = __detail::__odr_helper>
struct _MaskImplNeon;
947template <
typename _Abi,
typename = __detail::__odr_helper>
struct _SimdImplPpc;
948template <
typename _Abi,
typename = __detail::__odr_helper>
struct _MaskImplPpc;
951template <
int _UsedBytes>
952 struct simd_abi::_VecBuiltin
954 template <
typename _Tp>
955 static constexpr size_t _S_size = _UsedBytes /
sizeof(_Tp);
958 struct _IsValidAbiTag : __bool_constant<(_UsedBytes > 1)> {};
960 template <
typename _Tp>
961 struct _IsValidSizeFor
962 : __bool_constant<(_UsedBytes / sizeof(_Tp) > 1
963 && _UsedBytes % sizeof(_Tp) == 0
964 && _UsedBytes <= __vectorized_sizeof<_Tp>()
965 && (!__have_avx512f || _UsedBytes <= 32))> {};
967 template <typename _Tp>
968 struct _IsValid : conjunction<_IsValidAbiTag, __is_vectorizable<_Tp>,
969 _IsValidSizeFor<_Tp>> {};
971 template <typename _Tp>
972 static constexpr bool _S_is_valid_v = _IsValid<_Tp>::value;
976#if _GLIBCXX_SIMD_X86INTRIN
977 using _CommonImpl = _CommonImplX86;
978 using _SimdImpl = _SimdImplX86<_VecBuiltin<_UsedBytes>>;
979 using _MaskImpl = _MaskImplX86<_VecBuiltin<_UsedBytes>>;
980#elif _GLIBCXX_SIMD_HAVE_NEON
981 using _CommonImpl = _CommonImplNeon;
982 using _SimdImpl = _SimdImplNeon<_VecBuiltin<_UsedBytes>>;
983 using _MaskImpl = _MaskImplNeon<_VecBuiltin<_UsedBytes>>;
985 using _CommonImpl = _CommonImplBuiltin;
987 using _SimdImpl = _SimdImplPpc<_VecBuiltin<_UsedBytes>>;
988 using _MaskImpl = _MaskImplPpc<_VecBuiltin<_UsedBytes>>;
990 using _SimdImpl = _SimdImplBuiltin<_VecBuiltin<_UsedBytes>>;
991 using _MaskImpl = _MaskImplBuiltin<_VecBuiltin<_UsedBytes>>;
997 template <typename _Tp>
998 using _MaskValueType = __int_for_sizeof_t<_Tp>;
1000 template <typename _Tp>
1002 = conditional_t<_S_is_valid_v<_Tp>,
1003 _GnuTraits<_Tp, _MaskValueType<_Tp>,
1004 _VecBuiltin<_UsedBytes>, _S_size<_Tp>>,
1009 template <typename _Tp>
1010 static constexpr size_t _S_full_size = __traits<_Tp>::_S_full_size;
1012 template <typename _Tp>
1013 static constexpr bool _S_is_partial = __traits<_Tp>::_S_is_partial;
1017 template <typename _Tp>
1018 using _MaskMember = _SimdWrapper<_MaskValueType<_Tp>, _S_size<_Tp>>;
1020 template <typename _Tp>
1021 _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
1024 using _UV = typename _MaskMember<_Tp>::_BuiltinType;
1025 if constexpr (!_MaskMember<_Tp>::_S_is_partial)
1029 constexpr auto __size = _S_size<_Tp>;
1030 _GLIBCXX_SIMD_USE_CONSTEXPR auto __r
1031 = __generate_vector<_UV>([](auto __i) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA
1032 { return __i < __size ? -1 : 0; });
1037 template <typename _Tp>
1038 _GLIBCXX_SIMD_INTRINSIC static constexpr __intrinsic_type_t<_Tp, _S_size<_Tp>>
1039 _S_implicit_mask_intrin()
1040 { return __to_intrin(__vector_bitcast<_Tp>(_S_implicit_mask<_Tp>()._M_data)); }
1042 template <typename _TW, typename _TVT = _VectorTraits<_TW>>
1043 _GLIBCXX_SIMD_INTRINSIC static constexpr _TW
1046 using _Tp = typename _TVT::value_type;
1047 if constexpr (!_MaskMember<_Tp>::_S_is_partial)
1050 return __and(__as_vector(__x),
1051 __vector_bitcast<_Tp>(_S_implicit_mask<_Tp>()));
1054 template <typename _TW, typename _TVT = _VectorTraits<_TW>>
1055 _GLIBCXX_SIMD_INTRINSIC static constexpr auto
1056 __make_padding_nonzero(_TW __x)
1058 using _Tp = typename _TVT::value_type;
1059 if constexpr (!_S_is_partial<_Tp>)
1063 _GLIBCXX_SIMD_USE_CONSTEXPR auto __implicit_mask
1064 = __vector_bitcast<_Tp>(_S_implicit_mask<_Tp>());
1065 if constexpr (is_integral_v<_Tp>)
1066 return __or(__x, ~__implicit_mask);
1069 _GLIBCXX_SIMD_USE_CONSTEXPR auto __one
1070 = __andnot(__implicit_mask,
1071 __vector_broadcast<_S_full_size<_Tp>>(_Tp(1)));
1075 return __or(__and(__x, __implicit_mask), __one);
1084template <int _UsedBytes>
1085 struct simd_abi::_VecBltnBtmsk
1087 template <typename _Tp>
1088 static constexpr size_t _S_size = _UsedBytes / sizeof(_Tp);
1091 struct _IsValidAbiTag : __bool_constant<(_UsedBytes > 1)> {};
1093 template <typename _Tp>
1094 struct _IsValidSizeFor
1095 : __bool_constant<(_UsedBytes / sizeof(_Tp) > 1
1096 && _UsedBytes % sizeof(_Tp) == 0 && _UsedBytes <= 64
1097 && (_UsedBytes > 32 || __have_avx512vl))> {};
1101 template <
typename _Tp>
1104 _IsValidAbiTag, __bool_constant<__have_avx512f>,
1105 __bool_constant<__have_avx512bw || (sizeof(_Tp) >= 4)>,
1106 __bool_constant<(__vectorized_sizeof<_Tp>() > sizeof(_Tp))>,
1107 _IsValidSizeFor<_Tp>> {};
1109 template <
typename _Tp>
1110 static constexpr bool _S_is_valid_v = _IsValid<_Tp>::value;
1114 #if _GLIBCXX_SIMD_X86INTRIN
1115 using _CommonImpl = _CommonImplX86;
1116 using _SimdImpl = _SimdImplX86<_VecBltnBtmsk<_UsedBytes>>;
1117 using _MaskImpl = _MaskImplX86<_VecBltnBtmsk<_UsedBytes>>;
1120 struct _MissingImpl;
1122 using _CommonImpl = _MissingImpl<_UsedBytes>;
1123 using _SimdImpl = _MissingImpl<_UsedBytes>;
1124 using _MaskImpl = _MissingImpl<_UsedBytes>;
1129 template <
typename _Tp>
1130 using _MaskMember = _SimdWrapper<bool, _S_size<_Tp>>;
1132 template <
typename _Tp>
1135 _GnuTraits<_Tp, bool, _VecBltnBtmsk<_UsedBytes>, _S_size<_Tp>>,
1140 template <
typename _Tp>
1141 static constexpr size_t _S_full_size = __traits<_Tp>::_S_full_size;
1142 template <
typename _Tp>
1143 static constexpr bool _S_is_partial = __traits<_Tp>::_S_is_partial;
1148 template <
typename _Tp>
1149 using _ImplicitMask = _SimdWrapper<bool, _S_size<_Tp>>;
1152 template <
size_t _Np>
1153 _GLIBCXX_SIMD_INTRINSIC
static constexpr __bool_storage_member_type_t<_Np>
1156 using _Tp = __bool_storage_member_type_t<_Np>;
1157 return _Np <
sizeof(_Tp) * __CHAR_BIT__ ? _Tp((1ULL << _Np) - 1) : ~_Tp();
1160 template <
typename _Tp>
1161 _GLIBCXX_SIMD_INTRINSIC
static constexpr _ImplicitMask<_Tp>
1163 {
return __implicit_mask_n<_S_size<_Tp>>(); }
1165 template <
typename _Tp>
1166 _GLIBCXX_SIMD_INTRINSIC
static constexpr __bool_storage_member_type_t<_S_size<_Tp>>
1167 _S_implicit_mask_intrin()
1168 {
return __implicit_mask_n<_S_size<_Tp>>(); }
1170 template <
typename _Tp,
size_t _Np>
1171 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Tp, _Np>
1172 _S_masked(_SimdWrapper<_Tp, _Np> __x)
1174 if constexpr (is_same_v<_Tp, bool>)
1175 if constexpr (_Np < 8 || (_Np & (_Np - 1)) != 0)
1176 return _MaskImpl::_S_bit_and(
1177 __x, _SimdWrapper<_Tp, _Np>(
1178 __bool_storage_member_type_t<_Np>((1ULL << _Np) - 1)));
1182 return _S_masked(__x._M_data);
1185 template <
typename _TV>
1186 _GLIBCXX_SIMD_INTRINSIC
static constexpr _TV
1189 using _Tp =
typename _VectorTraits<_TV>::value_type;
1191 !__is_bitmask_v<_TV>,
1192 "_VecBltnBtmsk::_S_masked cannot work on bitmasks, since it doesn't "
1193 "know the number of elements. Use _SimdWrapper<bool, N> instead.");
1194 if constexpr (_S_is_partial<_Tp>)
1196 constexpr size_t _Np = _S_size<_Tp>;
1197 return __make_dependent_t<_TV, _CommonImpl>::_S_blend(
1198 _S_implicit_mask<_Tp>(), _SimdWrapper<_Tp, _Np>(),
1199 _SimdWrapper<_Tp, _Np>(__x));
1205 template <
typename _TV,
typename _TVT = _VectorTraits<_TV>>
1206 _GLIBCXX_SIMD_INTRINSIC
static constexpr auto
1207 __make_padding_nonzero(_TV __x)
1209 using _Tp =
typename _TVT::value_type;
1210 if constexpr (!_S_is_partial<_Tp>)
1214 constexpr size_t _Np = _S_size<_Tp>;
1215 if constexpr (is_integral_v<typename _TVT::value_type>)
1217 | __generate_vector<_Tp, _S_full_size<_Tp>>(
1218 [](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA -> _Tp {
1225 return __make_dependent_t<_TV, _CommonImpl>::_S_blend(
1226 _S_implicit_mask<_Tp>(),
1227 _SimdWrapper<_Tp, _Np>(
1228 __vector_broadcast<_S_full_size<_Tp>>(_Tp(1))),
1229 _SimdWrapper<_Tp, _Np>(__x))
1239struct _CommonImplBuiltin
1246 template <
typename _From,
typename _To,
size_t _ToSize>
1247 static inline constexpr bool __converts_via_decomposition_v
1248 =
sizeof(_From) !=
sizeof(_To);
1252 template <
typename _Tp,
size_t _Np,
size_t _Bytes = _Np * sizeof(_Tp)>
1253 _GLIBCXX_SIMD_INTRINSIC
static __vector_type_t<_Tp, _Np>
1254 _S_load(
const void* __p)
1256 static_assert(_Np > 1);
1257 static_assert(_Bytes %
sizeof(_Tp) == 0);
1258 using _Rp = __vector_type_t<_Tp, _Np>;
1259 if constexpr (
sizeof(_Rp) == _Bytes)
1262 __builtin_memcpy(&__r, __p, _Bytes);
1267#ifdef _GLIBCXX_SIMD_WORKAROUND_PR90424
1271 conditional_t<_Bytes % 8 == 0, long long, int>,
1272 conditional_t<_Bytes % 2 == 0, short, signed char>>,
1273 conditional_t<(_Bytes < 8 || _Np % 2 == 1 || _Np == 2), _Tp,
1275 using _V = __vector_type_t<_Up, _Np *
sizeof(_Tp) /
sizeof(_Up)>;
1276 if constexpr (
sizeof(_V) !=
sizeof(_Rp))
1279 __builtin_memcpy(&__r, __p, _Bytes);
1288 static_assert(_Bytes <=
sizeof(_V));
1289 __builtin_memcpy(&__r, __p, _Bytes);
1290 return reinterpret_cast<_Rp
>(__r);
1297 template <
size_t _Bytes>
1298 _GLIBCXX_SIMD_INTRINSIC
static void
1299 _S_memcpy(
char* __dst,
const char* __src)
1301 if constexpr (_Bytes > 0)
1303 constexpr size_t _Ns = std::__bit_floor(_Bytes);
1304 __builtin_memcpy(__dst, __src, _Ns);
1305 _S_memcpy<_Bytes - _Ns>(__dst + _Ns, __src + _Ns);
1309 template <
size_t _ReqBytes = 0,
typename _TV>
1310 _GLIBCXX_SIMD_INTRINSIC
static void
1311 _S_store(_TV __x,
void* __addr)
1313 constexpr size_t _Bytes = _ReqBytes == 0 ?
sizeof(__x) : _ReqBytes;
1314 static_assert(
sizeof(__x) >= _Bytes);
1316#if !defined __clang__ && _GLIBCXX_SIMD_WORKAROUND_PR90424
1317 if constexpr (__is_vector_type_v<_TV>)
1318 _S_memcpy<_Bytes>(
reinterpret_cast<char*
>(__addr),
reinterpret_cast<const char*
>(&__x));
1321 __builtin_memcpy(__addr, &__x, _Bytes);
1324 template <
typename _Tp,
size_t _Np>
1325 _GLIBCXX_SIMD_INTRINSIC
static void
1326 _S_store(_SimdWrapper<_Tp, _Np> __x,
void* __addr)
1327 { _S_store<_Np * sizeof(_Tp)>(__x._M_data, __addr); }
1331 template <
size_t _Np,
bool _Sanitized>
1332 _GLIBCXX_SIMD_INTRINSIC
static constexpr void
1333 _S_store_bool_array(_BitMask<_Np, _Sanitized> __x,
bool* __mem)
1335 if constexpr (_Np == 1)
1337 else if (__builtin_is_constant_evaluated())
1339 for (
size_t __i = 0; __i < _Np; ++__i)
1340 __mem[__i] = __x[__i];
1342 else if constexpr (_Np == 2)
1344 short __bool2 = (__x._M_to_bits() * 0x81) & 0x0101;
1345 _S_store<_Np>(__bool2, __mem);
1347 else if constexpr (_Np == 3)
1349 int __bool3 = (__x._M_to_bits() * 0x4081) & 0x010101;
1350 _S_store<_Np>(__bool3, __mem);
1354 __execute_n_times<__div_roundup(_Np, 4)>(
1355 [&](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1356 constexpr int __offset = __i * 4;
1357 constexpr int __remaining = _Np - __offset;
1358 if constexpr (__remaining > 4 && __remaining <= 7)
1360 const _ULLong __bool7
1361 = (__x.template _M_extract<__offset>()._M_to_bits()
1363 & 0x0101010101010101ULL;
1364 _S_store<__remaining>(__bool7, __mem + __offset);
1366 else if constexpr (__remaining >= 4)
1368 int __bits = __x.template _M_extract<__offset>()._M_to_bits();
1369 if constexpr (__remaining > 7)
1371 const int __bool4 = (__bits * 0x204081) & 0x01010101;
1372 _S_store<4>(__bool4, __mem + __offset);
1380 template <
typename _Tp,
size_t _Np>
1381 _GLIBCXX_SIMD_INTRINSIC
static constexpr auto
1382 _S_blend(_SimdWrapper<__int_for_sizeof_t<_Tp>, _Np> __k,
1383 _SimdWrapper<_Tp, _Np> __at0, _SimdWrapper<_Tp, _Np> __at1)
1384 {
return __k._M_data ? __at1._M_data : __at0._M_data; }
1391template <
typename _Abi,
typename>
1392 struct _SimdImplBuiltin
1395 template <
typename _Tp>
1396 static constexpr size_t _S_max_store_size = 16;
1398 using abi_type = _Abi;
1400 template <
typename _Tp>
1401 using _TypeTag = _Tp*;
1403 template <
typename _Tp>
1404 using _SimdMember =
typename _Abi::template __traits<_Tp>::_SimdMember;
1406 template <
typename _Tp>
1407 using _MaskMember =
typename _Abi::template _MaskMember<_Tp>;
1409 template <
typename _Tp>
1410 static constexpr size_t _S_size = _Abi::template _S_size<_Tp>;
1412 template <
typename _Tp>
1413 static constexpr size_t _S_full_size = _Abi::template _S_full_size<_Tp>;
1415 using _CommonImpl =
typename _Abi::_CommonImpl;
1416 using _SuperImpl =
typename _Abi::_SimdImpl;
1417 using _MaskImpl =
typename _Abi::_MaskImpl;
1420 template <
typename _Tp,
size_t _Np>
1421 _GLIBCXX_SIMD_INTRINSIC
static constexpr simd<_Tp, _Abi>
1422 _M_make_simd(_SimdWrapper<_Tp, _Np> __x)
1423 {
return {__private_init, __x}; }
1425 template <
typename _Tp,
size_t _Np>
1426 _GLIBCXX_SIMD_INTRINSIC
static constexpr simd<_Tp, _Abi>
1427 _M_make_simd(__intrinsic_type_t<_Tp, _Np> __x)
1428 {
return {__private_init, __vector_bitcast<_Tp>(__x)}; }
1431 template <
typename _Tp>
1432 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdMember<_Tp>
1433 _S_broadcast(_Tp __x)
noexcept
1434 {
return __vector_broadcast<_S_full_size<_Tp>>(__x); }
1437 template <
typename _Fp,
typename _Tp>
1438 inline static constexpr _SimdMember<_Tp>
1439 _S_generator(_Fp&& __gen, _TypeTag<_Tp>)
1441 return __generate_vector<_Tp, _S_full_size<_Tp>>(
1442 [&](
auto __i)
constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1443 if constexpr (__i < _S_size<_Tp>)
1451 template <
typename _Tp,
typename _Up>
1452 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdMember<_Tp>
1453 _S_load(
const _Up* __mem, _TypeTag<_Tp>)
noexcept
1455 constexpr size_t _Np = _S_size<_Tp>;
1456 constexpr size_t __max_load_size
1457 = (
sizeof(_Up) >= 4 && __have_avx512f) || __have_avx512bw ? 64
1458 : (is_floating_point_v<_Up> && __have_avx) || __have_avx2 ? 32
1460 constexpr size_t __bytes_to_load =
sizeof(_Up) * _Np;
1461 if (__builtin_is_constant_evaluated())
1462 return __generate_vector<_Tp, _S_full_size<_Tp>>(
1463 [&](
auto __i)
constexpr {
1464 return static_cast<_Tp
>(__i < _Np ? __mem[__i] : 0);
1466 else if constexpr (
sizeof(_Up) > 8 or __vectorized_sizeof<_Up>() <=
sizeof(_Up))
1467 return __generate_vector<_Tp, _SimdMember<_Tp>::_S_full_size>(
1468 [&](
auto __i)
constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1469 return static_cast<_Tp
>(__i < _Np ? __mem[__i] : 0);
1471 else if constexpr (is_same_v<_Up, _Tp>)
1472 return _CommonImpl::template _S_load<_Tp, _S_full_size<_Tp>,
1473 _Np *
sizeof(_Tp)>(__mem);
1474 else if constexpr (__bytes_to_load <= __max_load_size)
1475 return __convert<_SimdMember<_Tp>>(
1476 _CommonImpl::template _S_load<_Up, _Np>(__mem));
1477 else if constexpr (__bytes_to_load % __max_load_size == 0)
1479 constexpr size_t __n_loads = __bytes_to_load / __max_load_size;
1480 constexpr size_t __elements_per_load = _Np / __n_loads;
1481 return __call_with_n_evaluations<__n_loads>(
1482 [](
auto... __uncvted) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1483 return __convert<_SimdMember<_Tp>>(__uncvted...);
1484 }, [&](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1485 return _CommonImpl::template _S_load<_Up, __elements_per_load>(
1486 __mem + __i * __elements_per_load);
1489 else if constexpr (__bytes_to_load % (__max_load_size / 2) == 0
1490 && __max_load_size > 16)
1492 constexpr size_t __n_loads
1493 = __bytes_to_load / (__max_load_size / 2);
1494 constexpr size_t __elements_per_load = _Np / __n_loads;
1495 return __call_with_n_evaluations<__n_loads>(
1496 [](
auto... __uncvted) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1497 return __convert<_SimdMember<_Tp>>(__uncvted...);
1498 }, [&](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1499 return _CommonImpl::template _S_load<_Up, __elements_per_load>(
1500 __mem + __i * __elements_per_load);
1504 return __call_with_subscripts(
1505 __mem, make_index_sequence<_Np>(),
1506 [](
auto... __args) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1507 return __vector_type_t<_Tp, _S_full_size<_Tp>>{
static_cast<_Tp
>(__args)...};
1512 template <
typename _Tp,
size_t _Np,
typename _Up>
1513 static constexpr inline _SimdWrapper<_Tp, _Np>
1514 _S_masked_load(_SimdWrapper<_Tp, _Np> __merge, _MaskMember<_Tp> __k,
1515 const _Up* __mem)
noexcept
1517 _BitOps::_S_bit_iteration(_MaskImpl::_S_to_bits(__k),
1518 [&](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1519 __merge._M_set(__i,
static_cast<_Tp
>(__mem[__i]));
1525 template <
typename _Tp,
typename _Up>
1526 _GLIBCXX_SIMD_INTRINSIC
static constexpr void
1527 _S_store(_SimdMember<_Tp> __v, _Up* __mem, _TypeTag<_Tp>)
noexcept
1530 constexpr size_t _Np = _S_size<_Tp>;
1531 constexpr size_t __max_store_size
1532 = _SuperImpl::template _S_max_store_size<_Up>;
1533 if (__builtin_is_constant_evaluated())
1535 for (
size_t __i = 0; __i < _Np; ++__i)
1536 __mem[__i] = __v[__i];
1538 else if constexpr (
sizeof(_Up) > 8 or __vectorized_sizeof<_Up>() <=
sizeof(_Up))
1539 __execute_n_times<_Np>([&](
auto __i)
constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1540 __mem[__i] = __v[__i];
1542 else if constexpr (is_same_v<_Up, _Tp>)
1543 _CommonImpl::_S_store(__v, __mem);
1544 else if constexpr (
sizeof(_Up) * _Np <= __max_store_size)
1545 _CommonImpl::_S_store(_SimdWrapper<_Up, _Np>(__convert<_Up>(__v)),
1549 constexpr size_t __vsize = __max_store_size /
sizeof(_Up);
1551 constexpr size_t __stores = __div_roundup(_Np, __vsize);
1552 constexpr size_t __full_stores = _Np / __vsize;
1553 using _V = __vector_type_t<_Up, __vsize>;
1554 const array<_V, __stores> __converted
1555 = __convert_all<_V, __stores>(__v);
1556 __execute_n_times<__full_stores>(
1557 [&](
auto __i)
constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1558 _CommonImpl::_S_store(__converted[__i], __mem + __i * __vsize);
1560 if constexpr (__full_stores < __stores)
1561 _CommonImpl::template _S_store<(_Np - __full_stores * __vsize)
1563 __converted[__full_stores], __mem + __full_stores * __vsize);
1568 template <
typename _Tp,
size_t _Np>
1569 _GLIBCXX_SIMD_INTRINSIC
static constexpr void
1570 _S_masked_store_nocvt(_SimdWrapper<_Tp, _Np> __v, _Tp* __mem, _MaskMember<_Tp> __k)
1572 _BitOps::_S_bit_iteration(
1573 _MaskImpl::_S_to_bits(__k),
1574 [&](
auto __i)
constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1575 __mem[__i] = __v[__i];
1580 template <
typename _TW,
typename _TVT = _VectorTraits<_TW>,
1581 typename _Tp =
typename _TVT::value_type,
typename _Up>
1582 static constexpr inline void
1583 _S_masked_store(
const _TW __v, _Up* __mem,
const _MaskMember<_Tp> __k)
noexcept
1585 constexpr size_t _TV_size = _S_size<_Tp>;
1586 [[maybe_unused]]
const auto __vi = __to_intrin(__v);
1587 constexpr size_t __max_store_size
1588 = _SuperImpl::template _S_max_store_size<_Up>;
1592 _Up> || (is_integral_v<_Tp> && is_integral_v<_Up> &&
sizeof(_Tp) ==
sizeof(_Up)))
1595 const _MaskMember<_Up> __kk = [&]() _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1596 if constexpr (__is_bitmask_v<
decltype(__k)>)
1597 return _MaskMember<_Up>(__k._M_data);
1599 return __wrapper_bitcast<__int_for_sizeof_t<_Up>>(__k);
1601 _SuperImpl::_S_masked_store_nocvt(__wrapper_bitcast<_Up>(__v),
1604 else if constexpr (__vectorized_sizeof<_Up>() >
sizeof(_Up)
1606 template __converts_via_decomposition_v<
1607 _Tp, _Up, __max_store_size>)
1611 constexpr size_t _UW_size
1612 =
std::min(_TV_size, __max_store_size /
sizeof(_Up));
1613 static_assert(_UW_size <= _TV_size);
1614 using _UW = _SimdWrapper<_Up, _UW_size>;
1615 using _UV = __vector_type_t<_Up, _UW_size>;
1616 using _UAbi = simd_abi::__no_sve_deduce_t<_Up, _UW_size>;
1617 if constexpr (_UW_size == _TV_size)
1619 const _UW __converted = __convert<_UW>(__v);
1620 _UAbi::_SimdImpl::_S_masked_store_nocvt(
1622 _UAbi::_MaskImpl::template _S_convert<
1623 __int_for_sizeof_t<_Up>>(__k));
1627 static_assert(_UW_size *
sizeof(_Up) == __max_store_size);
1628 constexpr size_t _NFullStores = _TV_size / _UW_size;
1629 constexpr size_t _NAllStores
1630 = __div_roundup(_TV_size, _UW_size);
1631 constexpr size_t _NParts = _S_full_size<_Tp> / _UW_size;
1632 const array<_UV, _NAllStores> __converted
1633 = __convert_all<_UV, _NAllStores>(__v);
1634 __execute_n_times<_NFullStores>([&](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1635 _UAbi::_SimdImpl::_S_masked_store_nocvt(
1636 _UW(__converted[__i]), __mem + __i * _UW_size,
1637 _UAbi::_MaskImpl::template _S_convert<
1638 __int_for_sizeof_t<_Up>>(
1639 __extract_part<__i, _NParts>(__k.__as_full_vector())));
1641 if constexpr (_NAllStores
1643 _UAbi::_SimdImpl::_S_masked_store_nocvt(
1644 _UW(__converted[_NFullStores]),
1645 __mem + _NFullStores * _UW_size,
1646 _UAbi::_MaskImpl::template _S_convert<
1647 __int_for_sizeof_t<_Up>>(
1648 __extract_part<_NFullStores, _NParts>(
1649 __k.__as_full_vector())));
1653 _BitOps::_S_bit_iteration(_MaskImpl::_S_to_bits(__k),
1654 [&](
auto __i)
constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
1655 __mem[__i] =
static_cast<_Up
>(__v[__i]);
1660 template <
typename _Tp,
size_t _Np>
1661 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Tp, _Np>
1662 _S_complement(_SimdWrapper<_Tp, _Np> __x)
noexcept
1664 if constexpr (is_floating_point_v<_Tp>)
1665 return __vector_bitcast<_Tp>(~__vector_bitcast<__int_for_sizeof_t<_Tp>>(__x));
1667 return ~__x._M_data;
1671 template <
typename _Tp,
size_t _Np>
1672 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Tp, _Np>
1673 _S_unary_minus(_SimdWrapper<_Tp, _Np> __x)
noexcept
1677 return -__x._M_data;
1681 template <
typename _Tp,
size_t _Np>
1682 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Tp, _Np>
1683 _S_plus(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1684 {
return __x._M_data + __y._M_data; }
1686 template <
typename _Tp,
size_t _Np>
1687 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Tp, _Np>
1688 _S_minus(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1689 {
return __x._M_data - __y._M_data; }
1691 template <
typename _Tp,
size_t _Np>
1692 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Tp, _Np>
1693 _S_multiplies(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1694 {
return __x._M_data * __y._M_data; }
1696 template <
typename _Tp,
size_t _Np>
1697 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Tp, _Np>
1698 _S_divides(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1702 if constexpr (!_Abi::template _S_is_partial<_Tp>)
1703 return __x._M_data / __y._M_data;
1705 return __x._M_data / _Abi::__make_padding_nonzero(__y._M_data);
1708 template <
typename _Tp,
size_t _Np>
1709 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Tp, _Np>
1710 _S_modulus(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1712 if constexpr (!_Abi::template _S_is_partial<_Tp>)
1713 return __x._M_data % __y._M_data;
1715 return __as_vector(__x)
1716 % _Abi::__make_padding_nonzero(__as_vector(__y));
1719 template <
typename _Tp,
size_t _Np>
1720 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Tp, _Np>
1721 _S_bit_and(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1722 {
return __and(__x, __y); }
1724 template <
typename _Tp,
size_t _Np>
1725 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Tp, _Np>
1726 _S_bit_or(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1727 {
return __or(__x, __y); }
1729 template <
typename _Tp,
size_t _Np>
1730 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Tp, _Np>
1731 _S_bit_xor(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1732 {
return __xor(__x, __y); }
1734 template <
typename _Tp,
size_t _Np>
1735 _GLIBCXX_SIMD_INTRINSIC
static _SimdWrapper<_Tp, _Np>
1736 _S_bit_shift_left(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1737 {
return __x._M_data << __y._M_data; }
1739 template <
typename _Tp,
size_t _Np>
1740 _GLIBCXX_SIMD_INTRINSIC
static _SimdWrapper<_Tp, _Np>
1741 _S_bit_shift_right(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1742 {
return __x._M_data >> __y._M_data; }
1744 template <
typename _Tp,
size_t _Np>
1745 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Tp, _Np>
1746 _S_bit_shift_left(_SimdWrapper<_Tp, _Np> __x,
int __y)
1747 {
return __x._M_data << __y; }
1749 template <
typename _Tp,
size_t _Np>
1750 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Tp, _Np>
1751 _S_bit_shift_right(_SimdWrapper<_Tp, _Np> __x,
int __y)
1752 {
return __x._M_data >> __y; }
1756 template <
typename _Tp,
size_t _Np>
1757 _GLIBCXX_SIMD_INTRINSIC
static constexpr _MaskMember<_Tp>
1758 _S_equal_to(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1759 {
return __x._M_data == __y._M_data; }
1762 template <
typename _Tp,
size_t _Np>
1763 _GLIBCXX_SIMD_INTRINSIC
static constexpr _MaskMember<_Tp>
1764 _S_not_equal_to(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1765 {
return __x._M_data != __y._M_data; }
1768 template <
typename _Tp,
size_t _Np>
1769 _GLIBCXX_SIMD_INTRINSIC
static constexpr _MaskMember<_Tp>
1770 _S_less(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1771 {
return __x._M_data < __y._M_data; }
1774 template <
typename _Tp,
size_t _Np>
1775 _GLIBCXX_SIMD_INTRINSIC
static constexpr _MaskMember<_Tp>
1776 _S_less_equal(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
1777 {
return __x._M_data <= __y._M_data; }
1780 template <
typename _Tp,
size_t _Np>
1781 _GLIBCXX_SIMD_INTRINSIC
static constexpr _MaskMember<_Tp>
1782 _S_negate(_SimdWrapper<_Tp, _Np> __x)
noexcept
1783 {
return !__x._M_data; }
1786 template <
typename _Tp,
size_t _Np>
1787 _GLIBCXX_SIMD_NORMAL_MATH _GLIBCXX_SIMD_INTRINSIC
static constexpr
1788 _SimdWrapper<_Tp, _Np>
1789 _S_min(_SimdWrapper<_Tp, _Np> __a, _SimdWrapper<_Tp, _Np> __b)
1790 {
return __a._M_data < __b._M_data ? __a._M_data : __b._M_data; }
1792 template <
typename _Tp,
size_t _Np>
1793 _GLIBCXX_SIMD_NORMAL_MATH _GLIBCXX_SIMD_INTRINSIC
static constexpr
1794 _SimdWrapper<_Tp, _Np>
1795 _S_max(_SimdWrapper<_Tp, _Np> __a, _SimdWrapper<_Tp, _Np> __b)
1796 {
return __a._M_data > __b._M_data ? __a._M_data : __b._M_data; }
1798 template <
typename _Tp,
size_t _Np>
1799 _GLIBCXX_SIMD_NORMAL_MATH _GLIBCXX_SIMD_INTRINSIC
static constexpr
1800 pair<_SimdWrapper<_Tp, _Np>, _SimdWrapper<_Tp, _Np>>
1801 _S_minmax(_SimdWrapper<_Tp, _Np> __a, _SimdWrapper<_Tp, _Np> __b)
1803 return {__a._M_data < __b._M_data ? __a._M_data : __b._M_data,
1804 __a._M_data < __b._M_data ? __b._M_data : __a._M_data};
1808 template <
size_t _Np,
size_t... _Is,
size_t... _Zeros,
typename _Tp,
1809 typename _BinaryOperation>
1810 _GLIBCXX_SIMD_INTRINSIC
static constexpr _Tp
1811 _S_reduce_partial(index_sequence<_Is...>, index_sequence<_Zeros...>,
1812 simd<_Tp, _Abi> __x, _BinaryOperation&& __binary_op)
1814 using _V = __vector_type_t<_Tp, _Np / 2>;
1815 static_assert(
sizeof(_V) <=
sizeof(__x));
1818 using _FullSimd = __deduced_simd<_Tp, _VectorTraits<_V>::_S_full_size>;
1819 using _HalfSimd = __deduced_simd<_Tp, _Np / 2>;
1820 const auto __xx = __as_vector(__x);
1821 return _HalfSimd::abi_type::_SimdImpl::_S_reduce(
1822 static_cast<_HalfSimd
>(__as_vector(__binary_op(
1823 static_cast<_FullSimd
>(__intrin_bitcast<_V>(__xx)),
1824 static_cast<_FullSimd
>(__intrin_bitcast<_V>(
1825 __vector_permute<(_Np / 2 + _Is)..., (
int(_Zeros * 0) - 1)...>(
1830 template <
typename _Tp,
typename _BinaryOperation>
1831 _GLIBCXX_SIMD_INTRINSIC
static constexpr _Tp
1832 _S_reduce(simd<_Tp, _Abi> __x, _BinaryOperation&& __binary_op)
1834 constexpr size_t _Np = simd_size_v<_Tp, _Abi>;
1835 if constexpr (_Np == 1)
1837 else if constexpr (_Np == 2)
1838 return __binary_op(simd<_Tp, simd_abi::scalar>(__x[0]),
1839 simd<_Tp, simd_abi::scalar>(__x[1]))[0];
1840 else if (__builtin_is_constant_evaluated())
1842 simd<_Tp, simd_abi::scalar> __acc = __x[0];
1843 for (
size_t __i = 1; __i < _Np; ++__i)
1844 __acc = __binary_op(__acc, simd<_Tp, simd_abi::scalar>(__x[__i]));
1847 else if constexpr (_Abi::template _S_is_partial<_Tp>)
1849 [[maybe_unused]]
constexpr auto __full_size
1850 = _Abi::template _S_full_size<_Tp>;
1851 if constexpr (_Np == 3)
1853 __binary_op(simd<_Tp, simd_abi::scalar>(__x[0]),
1854 simd<_Tp, simd_abi::scalar>(__x[1])),
1855 simd<_Tp, simd_abi::scalar>(__x[2]))[0];
1856 else if constexpr (is_same_v<__remove_cvref_t<_BinaryOperation>,
1859 using _Ap = simd_abi::__no_sve_deduce_t<_Tp, __full_size>;
1860 return _Ap::_SimdImpl::_S_reduce(
1861 simd<_Tp, _Ap>(__private_init,
1862 _Abi::_S_masked(__as_vector(__x))),
1865 else if constexpr (is_same_v<__remove_cvref_t<_BinaryOperation>,
1868 using _Ap = simd_abi::__no_sve_deduce_t<_Tp, __full_size>;
1869 using _TW = _SimdWrapper<_Tp, __full_size>;
1870 _GLIBCXX_SIMD_USE_CONSTEXPR
auto __implicit_mask_full
1871 = _Abi::template _S_implicit_mask<_Tp>().__as_full_vector();
1872 _GLIBCXX_SIMD_USE_CONSTEXPR _TW __one
1873 = __vector_broadcast<__full_size>(_Tp(1));
1874 const _TW __x_full = __data(__x).__as_full_vector();
1875 const _TW __x_padded_with_ones
1876 = _Ap::_CommonImpl::_S_blend(__implicit_mask_full, __one,
1878 return _Ap::_SimdImpl::_S_reduce(
1879 simd<_Tp, _Ap>(__private_init, __x_padded_with_ones),
1882 else if constexpr (_Np & 1)
1884 using _Ap = simd_abi::__no_sve_deduce_t<_Tp, _Np - 1>;
1886 simd<_Tp, simd_abi::scalar>(_Ap::_SimdImpl::_S_reduce(
1888 __intrin_bitcast<__vector_type_t<_Tp, _Np - 1>>(
1891 simd<_Tp, simd_abi::scalar>(__x[_Np - 1]))[0];
1894 return _S_reduce_partial<_Np>(
1895 make_index_sequence<_Np / 2>(),
1896 make_index_sequence<__full_size - _Np / 2>(), __x, __binary_op);
1898 else if constexpr (
sizeof(__x) == 16)
1900 if constexpr (_Np == 16)
1902 const auto __y = __data(__x);
1904 _M_make_simd<_Tp, _Np>(
1905 __vector_permute<0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6,
1907 _M_make_simd<_Tp, _Np>(
1908 __vector_permute<8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13,
1909 14, 14, 15, 15>(__y)));
1911 if constexpr (_Np >= 8)
1913 const auto __y = __vector_bitcast<short>(__data(__x));
1915 _M_make_simd<_Tp, _Np>(__vector_bitcast<_Tp>(
1916 __vector_permute<0, 0, 1, 1, 2, 2, 3, 3>(__y))),
1917 _M_make_simd<_Tp, _Np>(__vector_bitcast<_Tp>(
1918 __vector_permute<4, 4, 5, 5, 6, 6, 7, 7>(__y))));
1920 if constexpr (_Np >= 4)
1922 using _Up = conditional_t<is_floating_point_v<_Tp>, float,
int>;
1923 const auto __y = __vector_bitcast<_Up>(__data(__x));
1924 __x = __binary_op(__x,
1925 _M_make_simd<_Tp, _Np>(__vector_bitcast<_Tp>(
1926 __vector_permute<3, 2, 1, 0>(__y))));
1928 using _Up = conditional_t<is_floating_point_v<_Tp>, double, _LLong>;
1929 const auto __y = __vector_bitcast<_Up>(__data(__x));
1930 __x = __binary_op(__x, _M_make_simd<_Tp, _Np>(__vector_bitcast<_Tp>(
1931 __vector_permute<1, 1>(__y))));
1936 static_assert(
sizeof(__x) > __min_vector_size<_Tp>);
1937 static_assert((_Np & (_Np - 1)) == 0);
1938 using _Ap = simd_abi::__no_sve_deduce_t<_Tp, _Np / 2>;
1939 using _V = simd<_Tp, _Ap>;
1940 return _Ap::_SimdImpl::_S_reduce(
1941 __binary_op(_V(__private_init, __extract<0, 2>(__as_vector(__x))),
1943 __extract<1, 2>(__as_vector(__x)))),
1944 static_cast<_BinaryOperation&&
>(__binary_op));
1950#define _GLIBCXX_SIMD_MATH_FALLBACK(__name) \
1951 template <typename _Tp, typename... _More> \
1953 _S_##__name(const _Tp& __x, const _More&... __more) \
1955 return __generate_vector<_Tp>( \
1956 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA { \
1957 return __name(__x[__i], __more[__i]...); \
1961#define _GLIBCXX_SIMD_MATH_FALLBACK_MASKRET(__name) \
1962 template <typename _Tp, typename... _More> \
1963 static typename _Tp::mask_type \
1964 _S_##__name(const _Tp& __x, const _More&... __more) \
1966 return __generate_vector<_Tp>( \
1967 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA { \
1968 return __name(__x[__i], __more[__i]...); \
1972#define _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET(_RetTp, __name) \
1973 template <typename _Tp, typename... _More> \
1975 _S_##__name(const _Tp& __x, const _More&... __more) \
1977 return __fixed_size_storage_t<_RetTp, \
1978 _VectorTraits<_Tp>::_S_partial_width>:: \
1979 _S_generate([&](auto __meta) constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA { \
1980 return __meta._S_generator( \
1981 [&](auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA { \
1982 return __name(__x[__meta._S_offset + __i], \
1983 __more[__meta._S_offset + __i]...); \
1985 static_cast<_RetTp*>(nullptr)); \
1989 _GLIBCXX_SIMD_MATH_FALLBACK(acos)
1990 _GLIBCXX_SIMD_MATH_FALLBACK(asin)
1991 _GLIBCXX_SIMD_MATH_FALLBACK(atan)
1992 _GLIBCXX_SIMD_MATH_FALLBACK(atan2)
1993 _GLIBCXX_SIMD_MATH_FALLBACK(cos)
1994 _GLIBCXX_SIMD_MATH_FALLBACK(sin)
1995 _GLIBCXX_SIMD_MATH_FALLBACK(tan)
1996 _GLIBCXX_SIMD_MATH_FALLBACK(acosh)
1997 _GLIBCXX_SIMD_MATH_FALLBACK(asinh)
1998 _GLIBCXX_SIMD_MATH_FALLBACK(atanh)
1999 _GLIBCXX_SIMD_MATH_FALLBACK(cosh)
2000 _GLIBCXX_SIMD_MATH_FALLBACK(sinh)
2001 _GLIBCXX_SIMD_MATH_FALLBACK(tanh)
2002 _GLIBCXX_SIMD_MATH_FALLBACK(exp)
2003 _GLIBCXX_SIMD_MATH_FALLBACK(exp2)
2004 _GLIBCXX_SIMD_MATH_FALLBACK(expm1)
2005 _GLIBCXX_SIMD_MATH_FALLBACK(ldexp)
2006 _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET(
int, ilogb)
2007 _GLIBCXX_SIMD_MATH_FALLBACK(log)
2008 _GLIBCXX_SIMD_MATH_FALLBACK(log10)
2009 _GLIBCXX_SIMD_MATH_FALLBACK(log1p)
2010 _GLIBCXX_SIMD_MATH_FALLBACK(log2)
2011 _GLIBCXX_SIMD_MATH_FALLBACK(logb)
2014 _GLIBCXX_SIMD_MATH_FALLBACK(scalbn)
2015 _GLIBCXX_SIMD_MATH_FALLBACK(scalbln)
2016 _GLIBCXX_SIMD_MATH_FALLBACK(cbrt)
2017 _GLIBCXX_SIMD_MATH_FALLBACK(fabs)
2018 _GLIBCXX_SIMD_MATH_FALLBACK(pow)
2019 _GLIBCXX_SIMD_MATH_FALLBACK(sqrt)
2020 _GLIBCXX_SIMD_MATH_FALLBACK(erf)
2021 _GLIBCXX_SIMD_MATH_FALLBACK(erfc)
2022 _GLIBCXX_SIMD_MATH_FALLBACK(lgamma)
2023 _GLIBCXX_SIMD_MATH_FALLBACK(tgamma)
2025 _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET(
long, lrint)
2026 _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET(
long long, llrint)
2028 _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET(
long, lround)
2029 _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET(
long long, llround)
2031 _GLIBCXX_SIMD_MATH_FALLBACK(fmod)
2032 _GLIBCXX_SIMD_MATH_FALLBACK(remainder)
2034 template <
typename _Tp,
typename _TVT = _VectorTraits<_Tp>>
2036 _S_remquo(
const _Tp __x,
const _Tp __y,
2037 __fixed_size_storage_t<int, _TVT::_S_partial_width>* __z)
2039 return __generate_vector<_Tp>([&](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2041 auto __r = remquo(__x[__i], __y[__i], &__tmp);
2042 __z->_M_set(__i, __tmp);
2048 _GLIBCXX_SIMD_MATH_FALLBACK(nextafter)
2049 _GLIBCXX_SIMD_MATH_FALLBACK(fdim)
2050 _GLIBCXX_SIMD_MATH_FALLBACK(fmax)
2051 _GLIBCXX_SIMD_MATH_FALLBACK(fmin)
2052 _GLIBCXX_SIMD_MATH_FALLBACK(fma)
2054 template <
typename _Tp,
size_t _Np>
2055 static constexpr _MaskMember<_Tp>
2056 _S_isgreater(_SimdWrapper<_Tp, _Np> __x,
2057 _SimdWrapper<_Tp, _Np> __y)
noexcept
2059 using _Ip = __int_for_sizeof_t<_Tp>;
2060 const auto __xn = __vector_bitcast<_Ip>(__x);
2061 const auto __yn = __vector_bitcast<_Ip>(__y);
2062 const auto __xp = __xn < 0 ? -(__xn & __finite_max_v<_Ip>) : __xn;
2063 const auto __yp = __yn < 0 ? -(__yn & __finite_max_v<_Ip>) : __yn;
2064 return __andnot(_SuperImpl::_S_isunordered(__x, __y)._M_data,
2068 template <
typename _Tp,
size_t _Np>
2069 static constexpr _MaskMember<_Tp>
2070 _S_isgreaterequal(_SimdWrapper<_Tp, _Np> __x,
2071 _SimdWrapper<_Tp, _Np> __y)
noexcept
2073 using _Ip = __int_for_sizeof_t<_Tp>;
2074 const auto __xn = __vector_bitcast<_Ip>(__x);
2075 const auto __yn = __vector_bitcast<_Ip>(__y);
2076 const auto __xp = __xn < 0 ? -(__xn & __finite_max_v<_Ip>) : __xn;
2077 const auto __yp = __yn < 0 ? -(__yn & __finite_max_v<_Ip>) : __yn;
2078 return __andnot(_SuperImpl::_S_isunordered(__x, __y)._M_data,
2082 template <
typename _Tp,
size_t _Np>
2083 static constexpr _MaskMember<_Tp>
2084 _S_isless(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
noexcept
2086 using _Ip = __int_for_sizeof_t<_Tp>;
2087 const auto __xn = __vector_bitcast<_Ip>(__x);
2088 const auto __yn = __vector_bitcast<_Ip>(__y);
2089 const auto __xp = __xn < 0 ? -(__xn & __finite_max_v<_Ip>) : __xn;
2090 const auto __yp = __yn < 0 ? -(__yn & __finite_max_v<_Ip>) : __yn;
2091 return __andnot(_SuperImpl::_S_isunordered(__x, __y)._M_data,
2095 template <
typename _Tp,
size_t _Np>
2096 static constexpr _MaskMember<_Tp>
2097 _S_islessequal(_SimdWrapper<_Tp, _Np> __x,
2098 _SimdWrapper<_Tp, _Np> __y)
noexcept
2100 using _Ip = __int_for_sizeof_t<_Tp>;
2101 const auto __xn = __vector_bitcast<_Ip>(__x);
2102 const auto __yn = __vector_bitcast<_Ip>(__y);
2103 const auto __xp = __xn < 0 ? -(__xn & __finite_max_v<_Ip>) : __xn;
2104 const auto __yp = __yn < 0 ? -(__yn & __finite_max_v<_Ip>) : __yn;
2105 return __andnot(_SuperImpl::_S_isunordered(__x, __y)._M_data,
2109 template <
typename _Tp,
size_t _Np>
2110 static constexpr _MaskMember<_Tp>
2111 _S_islessgreater(_SimdWrapper<_Tp, _Np> __x,
2112 _SimdWrapper<_Tp, _Np> __y)
noexcept
2114 return __andnot(_SuperImpl::_S_isunordered(__x, __y),
2115 _SuperImpl::_S_not_equal_to(__x, __y));
2118#undef _GLIBCXX_SIMD_MATH_FALLBACK
2119#undef _GLIBCXX_SIMD_MATH_FALLBACK_MASKRET
2120#undef _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET
2122 template <
typename _Tp,
size_t _Np>
2123 _GLIBCXX_SIMD_INTRINSIC
static _SimdWrapper<_Tp, _Np>
2124 _S_abs(_SimdWrapper<_Tp, _Np> __x)
noexcept
2130 if constexpr (is_floating_point_v<_Tp>)
2135 return __and(_S_absmask<__vector_type_t<_Tp, _Np>>, __x._M_data);
2137 return __x._M_data < 0 ? -__x._M_data : __x._M_data;
2145 template <
typename _TV,
typename _UV>
2146 _GLIBCXX_SIMD_INTRINSIC
static constexpr _TV
2147 _S_plus_minus(_TV __x, _UV __y)
noexcept
2149#if defined __i386__ && !defined __SSE_MATH__
2150 if constexpr (
sizeof(__x) == 8)
2152 static_assert(is_same_v<_TV, __vector_type_t<float, 2>>);
2153 const auto __x4 = __vector_bitcast<float, 4>(__x);
2154 if constexpr (is_same_v<_TV, _UV>)
2155 return __vector_bitcast<float, 2>(
2156 _S_plus_minus(__x4, __vector_bitcast<float, 4>(__y)));
2158 return __vector_bitcast<float, 2>(_S_plus_minus(__x4, __y));
2161#if !defined __clang__ && __GCC_IEC_559 == 0
2162 if (__builtin_is_constant_evaluated()
2163 || (__builtin_constant_p(__x) && __builtin_constant_p(__y)))
2164 return (__x + __y) - __y;
2168 if constexpr(__have_sse)
2170 if constexpr (
sizeof(__x) >= 16)
2171 asm(
"" :
"+x"(__x));
2172 else if constexpr (is_same_v<__vector_type_t<float, 2>, _TV>)
2173 asm(
"" :
"+x"(__x[0]),
"+x"(__x[1]));
2175 __assert_unreachable<_TV>();
2177 else if constexpr(__have_neon)
2178 asm(
"" :
"+w"(__x));
2179 else if constexpr (__have_power_vmx)
2181 if constexpr (is_same_v<__vector_type_t<float, 2>, _TV>)
2182 asm(
"" :
"+fgr"(__x[0]),
"+fgr"(__x[1]));
2184 asm(
"" :
"+v"(__x));
2187 asm(
"" :
"+g"(__x));
2191 return (__x + __y) - __y;
2197 template <
typename _Tp,
typename _TVT = _VectorTraits<_Tp>>
2198 _GLIBCXX_SIMD_INTRINSIC
static _Tp
2199 _S_nearbyint(_Tp __x_)
noexcept
2201 using value_type =
typename _TVT::value_type;
2202 using _V =
typename _TVT::type;
2203 const _V __x = __x_;
2204 const _V __absx = __and(__x, _S_absmask<_V>);
2205 static_assert(__CHAR_BIT__ *
sizeof(1ull) >= __digits_v<value_type>);
2206 _GLIBCXX_SIMD_USE_CONSTEXPR _V __shifter_abs
2207 = _V() + (1ull << (__digits_v<value_type> - 1));
2208 const _V __shifter = __or(__and(_S_signmask<_V>, __x), __shifter_abs);
2209 const _V __shifted = _S_plus_minus(__x, __shifter);
2210 return __absx < __shifter_abs ? __shifted : __x;
2214 template <
typename _Tp,
typename _TVT = _VectorTraits<_Tp>>
2215 _GLIBCXX_SIMD_INTRINSIC
static _Tp
2216 _S_rint(_Tp __x)
noexcept
2217 {
return _SuperImpl::_S_nearbyint(__x); }
2220 template <
typename _Tp,
size_t _Np>
2221 _GLIBCXX_SIMD_INTRINSIC
static _SimdWrapper<_Tp, _Np>
2222 _S_trunc(_SimdWrapper<_Tp, _Np> __x)
2224 using _V = __vector_type_t<_Tp, _Np>;
2225 const _V __absx = __and(__x._M_data, _S_absmask<_V>);
2226 static_assert(__CHAR_BIT__ *
sizeof(1ull) >= __digits_v<_Tp>);
2227 constexpr _Tp __shifter = 1ull << (__digits_v<_Tp> - 1);
2228 _V __truncated = _S_plus_minus(__absx, __shifter);
2229 __truncated -= __truncated > __absx ? _V() + 1 : _V();
2230 return __absx < __shifter ? __or(__xor(__absx, __x._M_data), __truncated)
2235 template <
typename _Tp,
size_t _Np>
2236 _GLIBCXX_SIMD_INTRINSIC
static _SimdWrapper<_Tp, _Np>
2237 _S_round(_SimdWrapper<_Tp, _Np> __x)
2239 const auto __abs_x = _SuperImpl::_S_abs(__x);
2240 const auto __t_abs = _SuperImpl::_S_trunc(__abs_x)._M_data;
2242 = __t_abs + (__abs_x._M_data - __t_abs >= _Tp(.5) ? _Tp(1) : 0);
2243 return __or(__xor(__abs_x._M_data, __x._M_data), __r_abs);
2247 template <
typename _Tp,
size_t _Np>
2248 _GLIBCXX_SIMD_INTRINSIC
static _SimdWrapper<_Tp, _Np>
2249 _S_floor(_SimdWrapper<_Tp, _Np> __x)
2251 const auto __y = _SuperImpl::_S_trunc(__x)._M_data;
2252 const auto __negative_input
2253 = __vector_bitcast<_Tp>(__x._M_data < __vector_broadcast<_Np, _Tp>(0));
2255 = __andnot(__vector_bitcast<_Tp>(__y == __x._M_data), __negative_input);
2256 return __or(__andnot(__mask, __y),
2257 __and(__mask, __y - __vector_broadcast<_Np, _Tp>(1)));
2261 template <
typename _Tp,
size_t _Np>
2262 _GLIBCXX_SIMD_INTRINSIC
static _SimdWrapper<_Tp, _Np>
2263 _S_ceil(_SimdWrapper<_Tp, _Np> __x)
2265 const auto __y = _SuperImpl::_S_trunc(__x)._M_data;
2266 const auto __negative_input
2267 = __vector_bitcast<_Tp>(__x._M_data < __vector_broadcast<_Np, _Tp>(0));
2268 const auto __inv_mask
2269 = __or(__vector_bitcast<_Tp>(__y == __x._M_data), __negative_input);
2270 return __or(__and(__inv_mask, __y),
2271 __andnot(__inv_mask, __y + __vector_broadcast<_Np, _Tp>(1)));
2275 template <
typename _Tp,
size_t _Np>
2276 _GLIBCXX_SIMD_INTRINSIC
static _MaskMember<_Tp>
2277 _S_isnan([[maybe_unused]] _SimdWrapper<_Tp, _Np> __x)
2279#if __FINITE_MATH_ONLY__
2281#elif !defined __SUPPORT_SNAN__
2282 return ~(__x._M_data == __x._M_data);
2283#elif defined __STDC_IEC_559__
2284 using _Ip = __int_for_sizeof_t<_Tp>;
2285 const auto __absn = __vector_bitcast<_Ip>(_SuperImpl::_S_abs(__x));
2287 = __vector_bitcast<_Ip>(__vector_broadcast<_Np>(__infinity_v<_Tp>));
2288 return __infn < __absn;
2290#error "Not implemented: how to support SNaN but non-IEC559 floating-point?"
2295 template <
typename _Tp,
size_t _Np>
2296 _GLIBCXX_SIMD_INTRINSIC
static _MaskMember<_Tp>
2297 _S_isfinite([[maybe_unused]] _SimdWrapper<_Tp, _Np> __x)
2299#if __FINITE_MATH_ONLY__
2300 using _UV =
typename _MaskMember<_Tp>::_BuiltinType;
2301 _GLIBCXX_SIMD_USE_CONSTEXPR _UV __alltrue = ~_UV();
2305 using _Ip = __int_for_sizeof_t<_Tp>;
2306 const auto __absn = __vector_bitcast<_Ip>(_SuperImpl::_S_abs(__x));
2308 = __vector_bitcast<_Ip>(__vector_broadcast<_Np>(__finite_max_v<_Tp>));
2309 return __absn <= __maxn;
2314 template <
typename _Tp,
size_t _Np>
2315 _GLIBCXX_SIMD_INTRINSIC
static _MaskMember<_Tp>
2316 _S_isunordered(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
2317 {
return __or(_S_isnan(__x), _S_isnan(__y)); }
2320 template <
typename _Tp,
size_t _Np>
2321 _GLIBCXX_SIMD_INTRINSIC
static _MaskMember<_Tp>
2322 _S_signbit(_SimdWrapper<_Tp, _Np> __x)
2324 using _Ip = __int_for_sizeof_t<_Tp>;
2325 return __vector_bitcast<_Ip>(__x) < 0;
2332 template <
typename _Tp,
size_t _Np>
2333 _GLIBCXX_SIMD_INTRINSIC
static _MaskMember<_Tp>
2334 _S_isinf([[maybe_unused]] _SimdWrapper<_Tp, _Np> __x)
2336#if __FINITE_MATH_ONLY__
2339 return _SuperImpl::template _S_equal_to<_Tp, _Np>(_SuperImpl::_S_abs(__x),
2340 __vector_broadcast<_Np>(
2341 __infinity_v<_Tp>));
2356 template <
typename _Tp,
size_t _Np>
2357 _GLIBCXX_SIMD_INTRINSIC
static _MaskMember<_Tp>
2358 _S_isnormal(_SimdWrapper<_Tp, _Np> __x)
2360 using _Ip = __int_for_sizeof_t<_Tp>;
2361 const auto __absn = __vector_bitcast<_Ip>(_SuperImpl::_S_abs(__x));
2363 = __vector_bitcast<_Ip>(__vector_broadcast<_Np>(__norm_min_v<_Tp>));
2364#if __FINITE_MATH_ONLY__
2365 return __absn >= __minn;
2368 = __vector_bitcast<_Ip>(__vector_broadcast<_Np>(__finite_max_v<_Tp>));
2369 return __minn <= __absn && __absn <= __maxn;
2374 template <
typename _Tp,
size_t _Np>
2375 _GLIBCXX_SIMD_INTRINSIC
static __fixed_size_storage_t<int, _Np>
2376 _S_fpclassify(_SimdWrapper<_Tp, _Np> __x)
2378 if constexpr(__have_sve)
2380 __fixed_size_storage_t<int, _Np> __r{};
2381 __execute_n_times<_Np>(
2382 [&](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2383 __r._M_set(__i, std::fpclassify(__x[__i]));
2389 using _I = __int_for_sizeof_t<_Tp>;
2391 = __vector_bitcast<_I>(__to_intrin(_SuperImpl::_S_abs(__x)));
2392 constexpr size_t _NI =
sizeof(__xn) /
sizeof(_I);
2393 _GLIBCXX_SIMD_USE_CONSTEXPR
auto __minn
2394 = __vector_bitcast<_I>(__vector_broadcast<_NI>(__norm_min_v<_Tp>));
2396 _GLIBCXX_SIMD_USE_CONSTEXPR
auto __fp_normal
2397 = __vector_broadcast<_NI, _I>(FP_NORMAL);
2398#if !__FINITE_MATH_ONLY__
2399 _GLIBCXX_SIMD_USE_CONSTEXPR
auto __infn
2400 = __vector_bitcast<_I>(__vector_broadcast<_NI>(__infinity_v<_Tp>));
2401 _GLIBCXX_SIMD_USE_CONSTEXPR
auto __fp_nan
2402 = __vector_broadcast<_NI, _I>(FP_NAN);
2403 _GLIBCXX_SIMD_USE_CONSTEXPR
auto __fp_infinite
2404 = __vector_broadcast<_NI, _I>(FP_INFINITE);
2406#ifndef __FAST_MATH__
2407 _GLIBCXX_SIMD_USE_CONSTEXPR
auto __fp_subnormal
2408 = __vector_broadcast<_NI, _I>(FP_SUBNORMAL);
2410 _GLIBCXX_SIMD_USE_CONSTEXPR
auto __fp_zero
2411 = __vector_broadcast<_NI, _I>(FP_ZERO);
2413 __vector_type_t<_I, _NI>
2414 __tmp = __xn < __minn
2418 ? (__xn == 0 ? __fp_zero : __fp_subnormal)
2420#
if __FINITE_MATH_ONLY__
2423 : (__xn < __infn ? __fp_normal
2424 : (__xn == __infn ? __fp_infinite : __fp_nan));
2427 if constexpr (
sizeof(_I) ==
sizeof(
int))
2429 using _FixedInt = __fixed_size_storage_t<int, _Np>;
2430 const auto __as_int = __vector_bitcast<int, _Np>(__tmp);
2431 if constexpr (_FixedInt::_S_tuple_size == 1)
2433 else if constexpr (_FixedInt::_S_tuple_size == 2
2435 typename _FixedInt::_SecondType::_FirstAbi,
2437 return {__extract<0, 2>(__as_int), __as_int[_Np - 1]};
2438 else if constexpr (_FixedInt::_S_tuple_size == 2)
2439 return {__extract<0, 2>(__as_int),
2440 __auto_bitcast(__extract<1, 2>(__as_int))};
2442 __assert_unreachable<_Tp>();
2444 else if constexpr (_Np == 2 &&
sizeof(_I) == 8
2445 && __fixed_size_storage_t<int, _Np>::_S_tuple_size == 2)
2447 const auto __aslong = __vector_bitcast<_LLong>(__tmp);
2448 return {int(__aslong[0]), {int(__aslong[1])}};
2450#if _GLIBCXX_SIMD_X86INTRIN
2451 else if constexpr (
sizeof(_Tp) == 8 &&
sizeof(__tmp) == 32
2452 && __fixed_size_storage_t<int, _Np>::_S_tuple_size == 1)
2453 return {_mm_packs_epi32(__to_intrin(__lo128(__tmp)),
2454 __to_intrin(__hi128(__tmp)))};
2455 else if constexpr (
sizeof(_Tp) == 8 &&
sizeof(__tmp) == 64
2456 && __fixed_size_storage_t<int, _Np>::_S_tuple_size == 1)
2457 return {_mm512_cvtepi64_epi32(__to_intrin(__tmp))};
2459 else if constexpr (__fixed_size_storage_t<int, _Np>::_S_tuple_size == 1)
2460 return {__call_with_subscripts<_Np>(__vector_bitcast<_LLong>(__tmp),
2461 [](
auto... __l) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2462 return __make_wrapper<int>(__l...);
2465 __assert_unreachable<_Tp>();
2470 template <
typename _Tp,
size_t _Np>
2471 _GLIBCXX_SIMD_INTRINSIC
static constexpr void
2472 _S_increment(_SimdWrapper<_Tp, _Np>& __x)
2473 { __x = __x._M_data + 1; }
2475 template <
typename _Tp,
size_t _Np>
2476 _GLIBCXX_SIMD_INTRINSIC
static constexpr void
2477 _S_decrement(_SimdWrapper<_Tp, _Np>& __x)
2478 { __x = __x._M_data - 1; }
2481 template <
typename _Tp,
size_t _Np,
typename _Up>
2482 _GLIBCXX_SIMD_INTRINSIC
static constexpr void
2483 _S_set(_SimdWrapper<_Tp, _Np>& __v,
int __i, _Up&& __x)
noexcept
2484 { __v._M_set(__i,
static_cast<_Up&&
>(__x)); }
2487 template <
typename _Tp,
typename _K,
size_t _Np>
2488 _GLIBCXX_SIMD_INTRINSIC
static constexpr void
2489 _S_masked_assign(_SimdWrapper<_K, _Np> __k, _SimdWrapper<_Tp, _Np>& __lhs,
2490 __type_identity_t<_SimdWrapper<_Tp, _Np>> __rhs)
2492 if (__k._M_is_constprop_none_of())
2494 else if (__k._M_is_constprop_all_of())
2497 __lhs = _CommonImpl::_S_blend(__k, __lhs, __rhs);
2500 template <
typename _Tp,
typename _K,
size_t _Np>
2501 _GLIBCXX_SIMD_INTRINSIC
static constexpr void
2502 _S_masked_assign(_SimdWrapper<_K, _Np> __k, _SimdWrapper<_Tp, _Np>& __lhs,
2503 __type_identity_t<_Tp> __rhs)
2505 if (__k._M_is_constprop_none_of())
2507 else if (__k._M_is_constprop_all_of())
2508 __lhs = __vector_broadcast<_Np>(__rhs);
2509 else if (__builtin_constant_p(__rhs) && __rhs == 0)
2511 if constexpr (!is_same_v<bool, _K>)
2515 = __andnot(__vector_bitcast<_Tp>(__k), __lhs._M_data);
2519 = _CommonImpl::_S_blend(__k, __lhs, _SimdWrapper<_Tp, _Np>());
2522 __lhs = _CommonImpl::_S_blend(__k, __lhs,
2523 _SimdWrapper<_Tp, _Np>(
2524 __vector_broadcast<_Np>(__rhs)));
2528 template <
typename _Op,
typename _Tp,
typename _K,
size_t _Np>
2529 _GLIBCXX_SIMD_INTRINSIC
static constexpr void
2530 _S_masked_cassign(
const _SimdWrapper<_K, _Np> __k,
2531 _SimdWrapper<_Tp, _Np>& __lhs,
2532 const __type_identity_t<_SimdWrapper<_Tp, _Np>> __rhs,
2535 if (__k._M_is_constprop_none_of())
2537 else if (__k._M_is_constprop_all_of())
2538 __lhs = __op(_SuperImpl{}, __lhs, __rhs);
2540 __lhs = _CommonImpl::_S_blend(__k, __lhs,
2541 __op(_SuperImpl{}, __lhs, __rhs));
2544 template <
typename _Op,
typename _Tp,
typename _K,
size_t _Np>
2545 _GLIBCXX_SIMD_INTRINSIC
static constexpr void
2546 _S_masked_cassign(
const _SimdWrapper<_K, _Np> __k,
2547 _SimdWrapper<_Tp, _Np>& __lhs,
2548 const __type_identity_t<_Tp> __rhs, _Op __op)
2549 { _S_masked_cassign(__k, __lhs, __vector_broadcast<_Np>(__rhs), __op); }
2552 template <
template <
typename>
class _Op,
typename _Tp,
typename _K,
2554 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Tp, _Np>
2555 _S_masked_unary(
const _SimdWrapper<_K, _Np> __k,
2556 const _SimdWrapper<_Tp, _Np> __v)
2558 if (__k._M_is_constprop_none_of())
2560 auto __vv = _M_make_simd(__v);
2561 _Op<
decltype(__vv)> __op;
2562 if (__k._M_is_constprop_all_of())
2563 return __data(__op(__vv));
2564 else if constexpr (is_same_v<_Op<void>, __increment<void>>)
2566 static_assert(not std::is_same_v<_K, bool>);
2567 if constexpr (is_integral_v<_Tp>)
2569 return __v._M_data - __vector_bitcast<_Tp>(__k._M_data);
2570 else if constexpr (not __have_avx2)
2572 + __vector_bitcast<_Tp>(__k._M_data & __builtin_bit_cast(
2576 else if constexpr (is_same_v<_Op<void>, __decrement<void>>)
2578 static_assert(not std::is_same_v<_K, bool>);
2579 if constexpr (is_integral_v<_Tp>)
2581 return __v._M_data + __vector_bitcast<_Tp>(__k._M_data);
2582 else if constexpr (not __have_avx2)
2584 - __vector_bitcast<_Tp>(__k._M_data & __builtin_bit_cast(
2588 return _CommonImpl::_S_blend(__k, __v, __data(__op(__vv)));
2595struct _MaskImplBuiltinMixin
2597 template <
typename _Tp>
2598 using _TypeTag = _Tp*;
2601 template <
typename _Up,
size_t _ToN = 1>
2602 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Up, _ToN>
2603 _S_to_maskvector(
bool __x)
2605 static_assert(is_same_v<_Up, __int_for_sizeof_t<_Up>>);
2606 return __x ? __vector_type_t<_Up, _ToN>{~_Up()}
2607 : __vector_type_t<_Up, _ToN>{};
2610 template <
typename _Up,
size_t _UpN = 0,
size_t _Np,
bool _Sanitized,
2611 size_t _ToN = _UpN == 0 ? _Np : _UpN>
2612 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Up, _ToN>
2613 _S_to_maskvector(_BitMask<_Np, _Sanitized> __x)
2615 static_assert(is_same_v<_Up, __int_for_sizeof_t<_Up>>);
2616 return __generate_vector<__vector_type_t<_Up, _ToN>>(
2617 [&](
auto __i)
constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2618 if constexpr (__i < _Np)
2619 return __x[__i] ? ~_Up() : _Up();
2625 template <
typename _Up,
size_t _UpN = 0,
typename _Tp,
size_t _Np,
2626 size_t _ToN = _UpN == 0 ? _Np : _UpN>
2627 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Up, _ToN>
2628 _S_to_maskvector(_SimdWrapper<_Tp, _Np> __x)
2630 static_assert(is_same_v<_Up, __int_for_sizeof_t<_Up>>);
2631 using _TW = _SimdWrapper<_Tp, _Np>;
2632 using _UW = _SimdWrapper<_Up, _ToN>;
2633 if constexpr (
sizeof(_Up) ==
sizeof(_Tp) &&
sizeof(_TW) ==
sizeof(_UW))
2634 return __wrapper_bitcast<_Up, _ToN>(__x);
2635 else if constexpr (is_same_v<_Tp, bool>)
2636 return _S_to_maskvector<_Up, _ToN>(_BitMask<_Np>(__x._M_data));
2663 return __generate_vector<__vector_type_t<_Up, _ToN>>(
2664 [&](
auto __i)
constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2665 if constexpr (__i < _Np)
2666 return _Up(__x[__i.value]);
2676 template <
typename _Tp,
size_t _Np>
2677 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SanitizedBitMask<_Np>
2678 _S_to_bits(_SimdWrapper<_Tp, _Np> __x)
2680 static_assert(!is_same_v<_Tp, bool>);
2681 static_assert(_Np <= __CHAR_BIT__ *
sizeof(_ULLong));
2682 using _Up = make_unsigned_t<__int_for_sizeof_t<_Tp>>;
2684 = __vector_bitcast<_Up>(__x) >> (
sizeof(_Up) * __CHAR_BIT__ - 1);
2686 __execute_n_times<_Np>(
2687 [&](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2688 __r |= _ULLong(__bools[__i.value]) << __i;
2697template <
typename _Abi,
typename>
2698 struct _MaskImplBuiltin : _MaskImplBuiltinMixin
2700 using _MaskImplBuiltinMixin::_S_to_bits;
2701 using _MaskImplBuiltinMixin::_S_to_maskvector;
2704 template <
typename _Tp>
2705 using _SimdMember =
typename _Abi::template __traits<_Tp>::_SimdMember;
2707 template <
typename _Tp>
2708 using _MaskMember =
typename _Abi::template _MaskMember<_Tp>;
2710 using _SuperImpl =
typename _Abi::_MaskImpl;
2711 using _CommonImpl =
typename _Abi::_CommonImpl;
2713 template <
typename _Tp>
2714 static constexpr size_t _S_size = simd_size_v<_Tp, _Abi>;
2718 template <
typename _Tp>
2719 _GLIBCXX_SIMD_INTRINSIC
static constexpr _MaskMember<_Tp>
2720 _S_broadcast(
bool __x)
2721 {
return __x ? _Abi::template _S_implicit_mask<_Tp>() : _MaskMember<_Tp>(); }
2725 template <
typename _Tp>
2726 _GLIBCXX_SIMD_INTRINSIC
static constexpr _MaskMember<_Tp>
2727 _S_load(
const bool* __mem)
2729 using _I = __int_for_sizeof_t<_Tp>;
2730 if (not __builtin_is_constant_evaluated())
2731 if constexpr (
sizeof(_Tp) ==
sizeof(
bool))
2734 = _CommonImpl::template _S_load<_I, _S_size<_Tp>>(__mem);
2738 return __generate_vector<_I, _S_size<_Tp>>(
2739 [&](
auto __i)
constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2740 return __mem[__i] ? ~_I() : _I();
2746 template <
typename _Tp,
size_t _Np,
bool _Sanitized>
2747 _GLIBCXX_SIMD_INTRINSIC
static constexpr auto
2748 _S_convert(_BitMask<_Np, _Sanitized> __x)
2750 if constexpr (__is_builtin_bitmask_abi<_Abi>())
2751 return _SimdWrapper<
bool, simd_size_v<_Tp, _Abi>>(__x._M_to_bits());
2753 return _SuperImpl::template _S_to_maskvector<__int_for_sizeof_t<_Tp>,
2755 __x._M_sanitized());
2758 template <
typename _Tp,
size_t _Np>
2759 _GLIBCXX_SIMD_INTRINSIC
static constexpr auto
2760 _S_convert(_SimdWrapper<bool, _Np> __x)
2762 if constexpr (__is_builtin_bitmask_abi<_Abi>())
2763 return _SimdWrapper<
bool, simd_size_v<_Tp, _Abi>>(__x._M_data);
2765 return _SuperImpl::template _S_to_maskvector<__int_for_sizeof_t<_Tp>,
2767 _BitMask<_Np>(__x._M_data)._M_sanitized());
2770 template <
typename _Tp,
typename _Up,
size_t _Np>
2771 _GLIBCXX_SIMD_INTRINSIC
static constexpr auto
2772 _S_convert(_SimdWrapper<_Up, _Np> __x)
2774 if constexpr (__is_builtin_bitmask_abi<_Abi>())
2775 return _SimdWrapper<
bool, simd_size_v<_Tp, _Abi>>(
2776 _SuperImpl::_S_to_bits(__x));
2778 return _SuperImpl::template _S_to_maskvector<__int_for_sizeof_t<_Tp>,
2782 template <
typename _Tp,
typename _Up,
typename _UAbi>
2783 _GLIBCXX_SIMD_INTRINSIC
static constexpr auto
2784 _S_convert(simd_mask<_Up, _UAbi> __x)
2786 if constexpr (__is_builtin_bitmask_abi<_Abi>())
2788 using _R = _SimdWrapper<bool, simd_size_v<_Tp, _Abi>>;
2789 if constexpr (__is_builtin_bitmask_abi<_UAbi>())
2790 return _R(__data(__x));
2791 else if constexpr (__is_scalar_abi<_UAbi>())
2792 return _R(__data(__x));
2793 else if constexpr (__is_fixed_size_abi_v<_UAbi>)
2794 return _R(__data(__x)._M_to_bits());
2796 return _R(_UAbi::_MaskImpl::_S_to_bits(__data(__x))._M_to_bits());
2800 if constexpr(__is_sve_abi<_UAbi>())
2802 simd_mask<_Tp> __r(
false);
2803 constexpr size_t __min_size =
std::min(__r.size(), __x.size());
2804 __execute_n_times<__min_size>(
2805 [&](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2806 __r[__i] = __x[__i];
2811 return _SuperImpl::template _S_to_maskvector<__int_for_sizeof_t<_Tp>,
2818 template <
typename _Tp,
size_t _Np>
2819 static inline _SimdWrapper<_Tp, _Np>
2820 _S_masked_load(_SimdWrapper<_Tp, _Np> __merge,
2821 _SimdWrapper<_Tp, _Np> __mask,
const bool* __mem)
noexcept
2824 auto __tmp = __wrapper_bitcast<__int_for_sizeof_t<_Tp>>(__merge);
2825 _BitOps::_S_bit_iteration(_SuperImpl::_S_to_bits(__mask),
2826 [&](
auto __i) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2827 __tmp._M_set(__i, -__mem[__i]);
2829 __merge = __wrapper_bitcast<_Tp>(__tmp);
2834 template <
typename _Tp,
size_t _Np>
2835 _GLIBCXX_SIMD_INTRINSIC
static constexpr void
2836 _S_store(_SimdWrapper<_Tp, _Np> __v,
bool* __mem)
noexcept
2838 __execute_n_times<_Np>([&](
auto __i)
constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2839 __mem[__i] = __v[__i];
2844 template <
typename _Tp,
size_t _Np>
2846 _S_masked_store(
const _SimdWrapper<_Tp, _Np> __v,
bool* __mem,
2847 const _SimdWrapper<_Tp, _Np> __k)
noexcept
2849 _BitOps::_S_bit_iteration(_SuperImpl::_S_to_bits(__k),
2850 [&](
auto __i)
constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2851 __mem[__i] = __v[__i];
2856 template <
size_t _Np,
typename _Tp>
2857 _GLIBCXX_SIMD_INTRINSIC
static _MaskMember<_Tp>
2858 _S_from_bitmask(_SanitizedBitMask<_Np> __bits, _TypeTag<_Tp>)
2859 {
return _SuperImpl::template _S_to_maskvector<_Tp, _S_size<_Tp>>(__bits); }
2862 template <
typename _Tp,
size_t _Np>
2863 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Tp, _Np>
2864 _S_logical_and(
const _SimdWrapper<_Tp, _Np>& __x,
const _SimdWrapper<_Tp, _Np>& __y)
2865 {
return __and(__x._M_data, __y._M_data); }
2867 template <
typename _Tp,
size_t _Np>
2868 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Tp, _Np>
2869 _S_logical_or(
const _SimdWrapper<_Tp, _Np>& __x,
const _SimdWrapper<_Tp, _Np>& __y)
2870 {
return __or(__x._M_data, __y._M_data); }
2872 template <
typename _Tp,
size_t _Np>
2873 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Tp, _Np>
2874 _S_bit_not(
const _SimdWrapper<_Tp, _Np>& __x)
2876 if constexpr (_Abi::template _S_is_partial<_Tp>)
2877 return __andnot(__x, __wrapper_bitcast<_Tp>(
2878 _Abi::template _S_implicit_mask<_Tp>()));
2880 return __not(__x._M_data);
2883 template <
typename _Tp,
size_t _Np>
2884 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Tp, _Np>
2885 _S_bit_and(
const _SimdWrapper<_Tp, _Np>& __x,
const _SimdWrapper<_Tp, _Np>& __y)
2886 {
return __and(__x._M_data, __y._M_data); }
2888 template <
typename _Tp,
size_t _Np>
2889 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Tp, _Np>
2890 _S_bit_or(
const _SimdWrapper<_Tp, _Np>& __x,
const _SimdWrapper<_Tp, _Np>& __y)
2891 {
return __or(__x._M_data, __y._M_data); }
2893 template <
typename _Tp,
size_t _Np>
2894 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SimdWrapper<_Tp, _Np>
2895 _S_bit_xor(
const _SimdWrapper<_Tp, _Np>& __x,
const _SimdWrapper<_Tp, _Np>& __y)
2896 {
return __xor(__x._M_data, __y._M_data); }
2899 template <
typename _Tp,
size_t _Np>
2900 static constexpr void
2901 _S_set(_SimdWrapper<_Tp, _Np>& __k,
int __i,
bool __x)
noexcept
2903 if constexpr (is_same_v<_Tp, bool>)
2904 __k._M_set(__i, __x);
2907 static_assert(is_same_v<_Tp, __int_for_sizeof_t<_Tp>>);
2908 if (__builtin_is_constant_evaluated())
2910 __k = __generate_from_n_evaluations<_Np,
2911 __vector_type_t<_Tp, _Np>>(
2912 [&](
auto __j) _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA {
2913 if (__i ==
static_cast<int>(__j))
2920 __k._M_data[__i] = -__x;
2925 template <
typename _Tp,
size_t _Np>
2926 _GLIBCXX_SIMD_INTRINSIC
static void
2927 _S_masked_assign(_SimdWrapper<_Tp, _Np> __k, _SimdWrapper<_Tp, _Np>& __lhs,
2928 __type_identity_t<_SimdWrapper<_Tp, _Np>> __rhs)
2929 { __lhs = _CommonImpl::_S_blend(__k, __lhs, __rhs); }
2931 template <
typename _Tp,
size_t _Np>
2932 _GLIBCXX_SIMD_INTRINSIC
static void
2933 _S_masked_assign(_SimdWrapper<_Tp, _Np> __k, _SimdWrapper<_Tp, _Np>& __lhs,
bool __rhs)
2935 if (__builtin_constant_p(__rhs))
2938 __lhs = __andnot(__k, __lhs);
2940 __lhs = __or(__k, __lhs);
2943 __lhs = _CommonImpl::_S_blend(__k, __lhs,
2944 __data(simd_mask<_Tp, _Abi>(__rhs)));
2949 template <
typename _Tp>
2950 _GLIBCXX_SIMD_INTRINSIC
static bool
2951 _S_all_of(simd_mask<_Tp, _Abi> __k)
2953 return __call_with_subscripts(
2954 __data(__k), make_index_sequence<_S_size<_Tp>>(),
2955 [](
const auto... __ent)
constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA
2956 {
return (... && !(__ent == 0)); });
2961 template <
typename _Tp>
2962 _GLIBCXX_SIMD_INTRINSIC
static bool
2963 _S_any_of(simd_mask<_Tp, _Abi> __k)
2965 return __call_with_subscripts(
2966 __data(__k), make_index_sequence<_S_size<_Tp>>(),
2967 [](
const auto... __ent)
constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA
2968 {
return (... || !(__ent == 0)); });
2973 template <
typename _Tp>
2974 _GLIBCXX_SIMD_INTRINSIC
static bool
2975 _S_none_of(simd_mask<_Tp, _Abi> __k)
2977 return __call_with_subscripts(
2978 __data(__k), make_index_sequence<_S_size<_Tp>>(),
2979 [](
const auto... __ent)
constexpr _GLIBCXX_SIMD_ALWAYS_INLINE_LAMBDA
2980 {
return (... && (__ent == 0)); });
2985 template <
typename _Tp>
2986 _GLIBCXX_SIMD_INTRINSIC
static bool
2987 _S_some_of(simd_mask<_Tp, _Abi> __k)
2989 const int __n_true = _SuperImpl::_S_popcount(__k);
2990 return __n_true > 0 && __n_true < int(_S_size<_Tp>);
2995 template <
typename _Tp>
2996 _GLIBCXX_SIMD_INTRINSIC
static int
2997 _S_popcount(simd_mask<_Tp, _Abi> __k)
2999 using _I = __int_for_sizeof_t<_Tp>;
3000 if constexpr (is_default_constructible_v<simd<_I, _Abi>>)
3002 simd<_I, _Abi>(__private_init, __wrapper_bitcast<_I>(__data(__k))));
3004 return -
reduce(__bit_cast<rebind_simd_t<_I, simd<_Tp, _Abi>>>(
3005 simd<_Tp, _Abi>(__private_init, __data(__k))));
3010 template <
typename _Tp>
3011 _GLIBCXX_SIMD_INTRINSIC
static int
3012 _S_find_first_set(simd_mask<_Tp, _Abi> __k)
3013 {
return std::__countr_zero(_SuperImpl::_S_to_bits(__data(__k))._M_to_bits()); }
3017 template <
typename _Tp>
3018 _GLIBCXX_SIMD_INTRINSIC
static int
3019 _S_find_last_set(simd_mask<_Tp, _Abi> __k)
3020 {
return std::__bit_width(_SuperImpl::_S_to_bits(__data(__k))._M_to_bits()) - 1; }
3026_GLIBCXX_SIMD_END_NAMESPACE
__bool_constant< true > true_type
The type used as a compile-time boolean with true value.
typename conditional< _Cond, _Iftrue, _Iffalse >::type conditional_t
Alias template for conditional.
constexpr const _Tp & min(const _Tp &, const _Tp &)
This does what you think it does.
constexpr _Tp reduce(_InputIterator __first, _InputIterator __last, _Tp __init, _BinaryOperation __binary_op)
Calculate reduction of values in a range.