xref: /freebsd/contrib/llvm-project/clang/lib/Headers/avxintrin.h (revision 0ad011ececb978e22a9bff2acf76633b094f1ff6)
1 /*===---- avxintrin.h - AVX intrinsics -------------------------------------===
2  *
3  * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4  * See https://llvm.org/LICENSE.txt for license information.
5  * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6  *
7  *===-----------------------------------------------------------------------===
8  */
9 
10 #ifndef __IMMINTRIN_H
11 #error "Never use <avxintrin.h> directly; include <immintrin.h> instead."
12 #endif
13 
14 #ifndef __AVXINTRIN_H
15 #define __AVXINTRIN_H
16 
17 typedef double __v4df __attribute__ ((__vector_size__ (32)));
18 typedef float __v8sf __attribute__ ((__vector_size__ (32)));
19 typedef long long __v4di __attribute__ ((__vector_size__ (32)));
20 typedef int __v8si __attribute__ ((__vector_size__ (32)));
21 typedef short __v16hi __attribute__ ((__vector_size__ (32)));
22 typedef char __v32qi __attribute__ ((__vector_size__ (32)));
23 
24 /* Unsigned types */
25 typedef unsigned long long __v4du __attribute__ ((__vector_size__ (32)));
26 typedef unsigned int __v8su __attribute__ ((__vector_size__ (32)));
27 typedef unsigned short __v16hu __attribute__ ((__vector_size__ (32)));
28 typedef unsigned char __v32qu __attribute__ ((__vector_size__ (32)));
29 
30 /* We need an explicitly signed variant for char. Note that this shouldn't
31  * appear in the interface though. */
32 typedef signed char __v32qs __attribute__((__vector_size__(32)));
33 
34 typedef float __m256 __attribute__ ((__vector_size__ (32), __aligned__(32)));
35 typedef double __m256d __attribute__((__vector_size__(32), __aligned__(32)));
36 typedef long long __m256i __attribute__((__vector_size__(32), __aligned__(32)));
37 
38 typedef float __m256_u __attribute__ ((__vector_size__ (32), __aligned__(1)));
39 typedef double __m256d_u __attribute__((__vector_size__(32), __aligned__(1)));
40 typedef long long __m256i_u __attribute__((__vector_size__(32), __aligned__(1)));
41 
42 #ifdef __SSE2__
43 /* Both _Float16 and __bf16 require SSE2 being enabled. */
44 typedef _Float16 __v16hf __attribute__((__vector_size__(32), __aligned__(32)));
45 typedef _Float16 __m256h __attribute__((__vector_size__(32), __aligned__(32)));
46 typedef _Float16 __m256h_u __attribute__((__vector_size__(32), __aligned__(1)));
47 
48 typedef __bf16 __v16bf __attribute__((__vector_size__(32), __aligned__(32)));
49 typedef __bf16 __m256bh __attribute__((__vector_size__(32), __aligned__(32)));
50 #endif
51 
52 /* Define the default attributes for the functions in this file. */
53 #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx"), __min_vector_width__(256)))
54 #define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx"), __min_vector_width__(128)))
55 
56 /* Arithmetic */
57 /// Adds two 256-bit vectors of [4 x double].
58 ///
59 /// \headerfile <x86intrin.h>
60 ///
61 /// This intrinsic corresponds to the <c> VADDPD </c> instruction.
62 ///
63 /// \param __a
64 ///    A 256-bit vector of [4 x double] containing one of the source operands.
65 /// \param __b
66 ///    A 256-bit vector of [4 x double] containing one of the source operands.
67 /// \returns A 256-bit vector of [4 x double] containing the sums of both
68 ///    operands.
69 static __inline __m256d __DEFAULT_FN_ATTRS
70 _mm256_add_pd(__m256d __a, __m256d __b)
71 {
72   return (__m256d)((__v4df)__a+(__v4df)__b);
73 }
74 
75 /// Adds two 256-bit vectors of [8 x float].
76 ///
77 /// \headerfile <x86intrin.h>
78 ///
79 /// This intrinsic corresponds to the <c> VADDPS </c> instruction.
80 ///
81 /// \param __a
82 ///    A 256-bit vector of [8 x float] containing one of the source operands.
83 /// \param __b
84 ///    A 256-bit vector of [8 x float] containing one of the source operands.
85 /// \returns A 256-bit vector of [8 x float] containing the sums of both
86 ///    operands.
87 static __inline __m256 __DEFAULT_FN_ATTRS
88 _mm256_add_ps(__m256 __a, __m256 __b)
89 {
90   return (__m256)((__v8sf)__a+(__v8sf)__b);
91 }
92 
93 /// Subtracts two 256-bit vectors of [4 x double].
94 ///
95 /// \headerfile <x86intrin.h>
96 ///
97 /// This intrinsic corresponds to the <c> VSUBPD </c> instruction.
98 ///
99 /// \param __a
100 ///    A 256-bit vector of [4 x double] containing the minuend.
101 /// \param __b
102 ///    A 256-bit vector of [4 x double] containing the subtrahend.
103 /// \returns A 256-bit vector of [4 x double] containing the differences between
104 ///    both operands.
105 static __inline __m256d __DEFAULT_FN_ATTRS
106 _mm256_sub_pd(__m256d __a, __m256d __b)
107 {
108   return (__m256d)((__v4df)__a-(__v4df)__b);
109 }
110 
111 /// Subtracts two 256-bit vectors of [8 x float].
112 ///
113 /// \headerfile <x86intrin.h>
114 ///
115 /// This intrinsic corresponds to the <c> VSUBPS </c> instruction.
116 ///
117 /// \param __a
118 ///    A 256-bit vector of [8 x float] containing the minuend.
119 /// \param __b
120 ///    A 256-bit vector of [8 x float] containing the subtrahend.
121 /// \returns A 256-bit vector of [8 x float] containing the differences between
122 ///    both operands.
123 static __inline __m256 __DEFAULT_FN_ATTRS
124 _mm256_sub_ps(__m256 __a, __m256 __b)
125 {
126   return (__m256)((__v8sf)__a-(__v8sf)__b);
127 }
128 
129 /// Adds the even-indexed values and subtracts the odd-indexed values of
130 ///    two 256-bit vectors of [4 x double].
131 ///
132 /// \headerfile <x86intrin.h>
133 ///
134 /// This intrinsic corresponds to the <c> VADDSUBPD </c> instruction.
135 ///
136 /// \param __a
137 ///    A 256-bit vector of [4 x double] containing the left source operand.
138 /// \param __b
139 ///    A 256-bit vector of [4 x double] containing the right source operand.
140 /// \returns A 256-bit vector of [4 x double] containing the alternating sums
141 ///    and differences between both operands.
142 static __inline __m256d __DEFAULT_FN_ATTRS
143 _mm256_addsub_pd(__m256d __a, __m256d __b)
144 {
145   return (__m256d)__builtin_ia32_addsubpd256((__v4df)__a, (__v4df)__b);
146 }
147 
148 /// Adds the even-indexed values and subtracts the odd-indexed values of
149 ///    two 256-bit vectors of [8 x float].
150 ///
151 /// \headerfile <x86intrin.h>
152 ///
153 /// This intrinsic corresponds to the <c> VADDSUBPS </c> instruction.
154 ///
155 /// \param __a
156 ///    A 256-bit vector of [8 x float] containing the left source operand.
157 /// \param __b
158 ///    A 256-bit vector of [8 x float] containing the right source operand.
159 /// \returns A 256-bit vector of [8 x float] containing the alternating sums and
160 ///    differences between both operands.
161 static __inline __m256 __DEFAULT_FN_ATTRS
162 _mm256_addsub_ps(__m256 __a, __m256 __b)
163 {
164   return (__m256)__builtin_ia32_addsubps256((__v8sf)__a, (__v8sf)__b);
165 }
166 
167 /// Divides two 256-bit vectors of [4 x double].
168 ///
169 /// \headerfile <x86intrin.h>
170 ///
171 /// This intrinsic corresponds to the <c> VDIVPD </c> instruction.
172 ///
173 /// \param __a
174 ///    A 256-bit vector of [4 x double] containing the dividend.
175 /// \param __b
176 ///    A 256-bit vector of [4 x double] containing the divisor.
177 /// \returns A 256-bit vector of [4 x double] containing the quotients of both
178 ///    operands.
179 static __inline __m256d __DEFAULT_FN_ATTRS
180 _mm256_div_pd(__m256d __a, __m256d __b)
181 {
182   return (__m256d)((__v4df)__a/(__v4df)__b);
183 }
184 
185 /// Divides two 256-bit vectors of [8 x float].
186 ///
187 /// \headerfile <x86intrin.h>
188 ///
189 /// This intrinsic corresponds to the <c> VDIVPS </c> instruction.
190 ///
191 /// \param __a
192 ///    A 256-bit vector of [8 x float] containing the dividend.
193 /// \param __b
194 ///    A 256-bit vector of [8 x float] containing the divisor.
195 /// \returns A 256-bit vector of [8 x float] containing the quotients of both
196 ///    operands.
197 static __inline __m256 __DEFAULT_FN_ATTRS
198 _mm256_div_ps(__m256 __a, __m256 __b)
199 {
200   return (__m256)((__v8sf)__a/(__v8sf)__b);
201 }
202 
203 /// Compares two 256-bit vectors of [4 x double] and returns the greater
204 ///    of each pair of values.
205 ///
206 /// \headerfile <x86intrin.h>
207 ///
208 /// This intrinsic corresponds to the <c> VMAXPD </c> instruction.
209 ///
210 /// \param __a
211 ///    A 256-bit vector of [4 x double] containing one of the operands.
212 /// \param __b
213 ///    A 256-bit vector of [4 x double] containing one of the operands.
214 /// \returns A 256-bit vector of [4 x double] containing the maximum values
215 ///    between both operands.
216 static __inline __m256d __DEFAULT_FN_ATTRS
217 _mm256_max_pd(__m256d __a, __m256d __b)
218 {
219   return (__m256d)__builtin_ia32_maxpd256((__v4df)__a, (__v4df)__b);
220 }
221 
222 /// Compares two 256-bit vectors of [8 x float] and returns the greater
223 ///    of each pair of values.
224 ///
225 /// \headerfile <x86intrin.h>
226 ///
227 /// This intrinsic corresponds to the <c> VMAXPS </c> instruction.
228 ///
229 /// \param __a
230 ///    A 256-bit vector of [8 x float] containing one of the operands.
231 /// \param __b
232 ///    A 256-bit vector of [8 x float] containing one of the operands.
233 /// \returns A 256-bit vector of [8 x float] containing the maximum values
234 ///    between both operands.
235 static __inline __m256 __DEFAULT_FN_ATTRS
236 _mm256_max_ps(__m256 __a, __m256 __b)
237 {
238   return (__m256)__builtin_ia32_maxps256((__v8sf)__a, (__v8sf)__b);
239 }
240 
241 /// Compares two 256-bit vectors of [4 x double] and returns the lesser
242 ///    of each pair of values.
243 ///
244 /// \headerfile <x86intrin.h>
245 ///
246 /// This intrinsic corresponds to the <c> VMINPD </c> instruction.
247 ///
248 /// \param __a
249 ///    A 256-bit vector of [4 x double] containing one of the operands.
250 /// \param __b
251 ///    A 256-bit vector of [4 x double] containing one of the operands.
252 /// \returns A 256-bit vector of [4 x double] containing the minimum values
253 ///    between both operands.
254 static __inline __m256d __DEFAULT_FN_ATTRS
255 _mm256_min_pd(__m256d __a, __m256d __b)
256 {
257   return (__m256d)__builtin_ia32_minpd256((__v4df)__a, (__v4df)__b);
258 }
259 
260 /// Compares two 256-bit vectors of [8 x float] and returns the lesser
261 ///    of each pair of values.
262 ///
263 /// \headerfile <x86intrin.h>
264 ///
265 /// This intrinsic corresponds to the <c> VMINPS </c> instruction.
266 ///
267 /// \param __a
268 ///    A 256-bit vector of [8 x float] containing one of the operands.
269 /// \param __b
270 ///    A 256-bit vector of [8 x float] containing one of the operands.
271 /// \returns A 256-bit vector of [8 x float] containing the minimum values
272 ///    between both operands.
273 static __inline __m256 __DEFAULT_FN_ATTRS
274 _mm256_min_ps(__m256 __a, __m256 __b)
275 {
276   return (__m256)__builtin_ia32_minps256((__v8sf)__a, (__v8sf)__b);
277 }
278 
279 /// Multiplies two 256-bit vectors of [4 x double].
280 ///
281 /// \headerfile <x86intrin.h>
282 ///
283 /// This intrinsic corresponds to the <c> VMULPD </c> instruction.
284 ///
285 /// \param __a
286 ///    A 256-bit vector of [4 x double] containing one of the operands.
287 /// \param __b
288 ///    A 256-bit vector of [4 x double] containing one of the operands.
289 /// \returns A 256-bit vector of [4 x double] containing the products of both
290 ///    operands.
291 static __inline __m256d __DEFAULT_FN_ATTRS
292 _mm256_mul_pd(__m256d __a, __m256d __b)
293 {
294   return (__m256d)((__v4df)__a * (__v4df)__b);
295 }
296 
297 /// Multiplies two 256-bit vectors of [8 x float].
298 ///
299 /// \headerfile <x86intrin.h>
300 ///
301 /// This intrinsic corresponds to the <c> VMULPS </c> instruction.
302 ///
303 /// \param __a
304 ///    A 256-bit vector of [8 x float] containing one of the operands.
305 /// \param __b
306 ///    A 256-bit vector of [8 x float] containing one of the operands.
307 /// \returns A 256-bit vector of [8 x float] containing the products of both
308 ///    operands.
309 static __inline __m256 __DEFAULT_FN_ATTRS
310 _mm256_mul_ps(__m256 __a, __m256 __b)
311 {
312   return (__m256)((__v8sf)__a * (__v8sf)__b);
313 }
314 
315 /// Calculates the square roots of the values in a 256-bit vector of
316 ///    [4 x double].
317 ///
318 /// \headerfile <x86intrin.h>
319 ///
320 /// This intrinsic corresponds to the <c> VSQRTPD </c> instruction.
321 ///
322 /// \param __a
323 ///    A 256-bit vector of [4 x double].
324 /// \returns A 256-bit vector of [4 x double] containing the square roots of the
325 ///    values in the operand.
326 static __inline __m256d __DEFAULT_FN_ATTRS
327 _mm256_sqrt_pd(__m256d __a)
328 {
329   return (__m256d)__builtin_ia32_sqrtpd256((__v4df)__a);
330 }
331 
332 /// Calculates the square roots of the values in a 256-bit vector of
333 ///    [8 x float].
334 ///
335 /// \headerfile <x86intrin.h>
336 ///
337 /// This intrinsic corresponds to the <c> VSQRTPS </c> instruction.
338 ///
339 /// \param __a
340 ///    A 256-bit vector of [8 x float].
341 /// \returns A 256-bit vector of [8 x float] containing the square roots of the
342 ///    values in the operand.
343 static __inline __m256 __DEFAULT_FN_ATTRS
344 _mm256_sqrt_ps(__m256 __a)
345 {
346   return (__m256)__builtin_ia32_sqrtps256((__v8sf)__a);
347 }
348 
349 /// Calculates the reciprocal square roots of the values in a 256-bit
350 ///    vector of [8 x float].
351 ///
352 /// \headerfile <x86intrin.h>
353 ///
354 /// This intrinsic corresponds to the <c> VRSQRTPS </c> instruction.
355 ///
356 /// \param __a
357 ///    A 256-bit vector of [8 x float].
358 /// \returns A 256-bit vector of [8 x float] containing the reciprocal square
359 ///    roots of the values in the operand.
360 static __inline __m256 __DEFAULT_FN_ATTRS
361 _mm256_rsqrt_ps(__m256 __a)
362 {
363   return (__m256)__builtin_ia32_rsqrtps256((__v8sf)__a);
364 }
365 
366 /// Calculates the reciprocals of the values in a 256-bit vector of
367 ///    [8 x float].
368 ///
369 /// \headerfile <x86intrin.h>
370 ///
371 /// This intrinsic corresponds to the <c> VRCPPS </c> instruction.
372 ///
373 /// \param __a
374 ///    A 256-bit vector of [8 x float].
375 /// \returns A 256-bit vector of [8 x float] containing the reciprocals of the
376 ///    values in the operand.
377 static __inline __m256 __DEFAULT_FN_ATTRS
378 _mm256_rcp_ps(__m256 __a)
379 {
380   return (__m256)__builtin_ia32_rcpps256((__v8sf)__a);
381 }
382 
383 /// Rounds the values in a 256-bit vector of [4 x double] as specified
384 ///    by the byte operand. The source values are rounded to integer values and
385 ///    returned as 64-bit double-precision floating-point values.
386 ///
387 /// \headerfile <x86intrin.h>
388 ///
389 /// \code
390 /// __m256d _mm256_round_pd(__m256d V, const int M);
391 /// \endcode
392 ///
393 /// This intrinsic corresponds to the <c> VROUNDPD </c> instruction.
394 ///
395 /// \param V
396 ///    A 256-bit vector of [4 x double].
397 /// \param M
398 ///    An integer value that specifies the rounding operation. \n
399 ///    Bits [7:4] are reserved. \n
400 ///    Bit [3] is a precision exception value: \n
401 ///      0: A normal PE exception is used. \n
402 ///      1: The PE field is not updated. \n
403 ///    Bit [2] is the rounding control source: \n
404 ///      0: Use bits [1:0] of \a M. \n
405 ///      1: Use the current MXCSR setting. \n
406 ///    Bits [1:0] contain the rounding control definition: \n
407 ///      00: Nearest. \n
408 ///      01: Downward (toward negative infinity). \n
409 ///      10: Upward (toward positive infinity). \n
410 ///      11: Truncated.
411 /// \returns A 256-bit vector of [4 x double] containing the rounded values.
412 #define _mm256_round_pd(V, M) \
413   ((__m256d)__builtin_ia32_roundpd256((__v4df)(__m256d)(V), (M)))
414 
415 /// Rounds the values stored in a 256-bit vector of [8 x float] as
416 ///    specified by the byte operand. The source values are rounded to integer
417 ///    values and returned as floating-point values.
418 ///
419 /// \headerfile <x86intrin.h>
420 ///
421 /// \code
422 /// __m256 _mm256_round_ps(__m256 V, const int M);
423 /// \endcode
424 ///
425 /// This intrinsic corresponds to the <c> VROUNDPS </c> instruction.
426 ///
427 /// \param V
428 ///    A 256-bit vector of [8 x float].
429 /// \param M
430 ///    An integer value that specifies the rounding operation. \n
431 ///    Bits [7:4] are reserved. \n
432 ///    Bit [3] is a precision exception value: \n
433 ///      0: A normal PE exception is used. \n
434 ///      1: The PE field is not updated. \n
435 ///    Bit [2] is the rounding control source: \n
436 ///      0: Use bits [1:0] of \a M. \n
437 ///      1: Use the current MXCSR setting. \n
438 ///    Bits [1:0] contain the rounding control definition: \n
439 ///      00: Nearest. \n
440 ///      01: Downward (toward negative infinity). \n
441 ///      10: Upward (toward positive infinity). \n
442 ///      11: Truncated.
443 /// \returns A 256-bit vector of [8 x float] containing the rounded values.
444 #define _mm256_round_ps(V, M) \
445   ((__m256)__builtin_ia32_roundps256((__v8sf)(__m256)(V), (M)))
446 
447 /// Rounds up the values stored in a 256-bit vector of [4 x double]. The
448 ///    source values are rounded up to integer values and returned as 64-bit
449 ///    double-precision floating-point values.
450 ///
451 /// \headerfile <x86intrin.h>
452 ///
453 /// \code
454 /// __m256d _mm256_ceil_pd(__m256d V);
455 /// \endcode
456 ///
457 /// This intrinsic corresponds to the <c> VROUNDPD </c> instruction.
458 ///
459 /// \param V
460 ///    A 256-bit vector of [4 x double].
461 /// \returns A 256-bit vector of [4 x double] containing the rounded up values.
462 #define _mm256_ceil_pd(V)  _mm256_round_pd((V), _MM_FROUND_CEIL)
463 
464 /// Rounds down the values stored in a 256-bit vector of [4 x double].
465 ///    The source values are rounded down to integer values and returned as
466 ///    64-bit double-precision floating-point values.
467 ///
468 /// \headerfile <x86intrin.h>
469 ///
470 /// \code
471 /// __m256d _mm256_floor_pd(__m256d V);
472 /// \endcode
473 ///
474 /// This intrinsic corresponds to the <c> VROUNDPD </c> instruction.
475 ///
476 /// \param V
477 ///    A 256-bit vector of [4 x double].
478 /// \returns A 256-bit vector of [4 x double] containing the rounded down
479 ///    values.
480 #define _mm256_floor_pd(V) _mm256_round_pd((V), _MM_FROUND_FLOOR)
481 
482 /// Rounds up the values stored in a 256-bit vector of [8 x float]. The
483 ///    source values are rounded up to integer values and returned as
484 ///    floating-point values.
485 ///
486 /// \headerfile <x86intrin.h>
487 ///
488 /// \code
489 /// __m256 _mm256_ceil_ps(__m256 V);
490 /// \endcode
491 ///
492 /// This intrinsic corresponds to the <c> VROUNDPS </c> instruction.
493 ///
494 /// \param V
495 ///    A 256-bit vector of [8 x float].
496 /// \returns A 256-bit vector of [8 x float] containing the rounded up values.
497 #define _mm256_ceil_ps(V)  _mm256_round_ps((V), _MM_FROUND_CEIL)
498 
499 /// Rounds down the values stored in a 256-bit vector of [8 x float]. The
500 ///    source values are rounded down to integer values and returned as
501 ///    floating-point values.
502 ///
503 /// \headerfile <x86intrin.h>
504 ///
505 /// \code
506 /// __m256 _mm256_floor_ps(__m256 V);
507 /// \endcode
508 ///
509 /// This intrinsic corresponds to the <c> VROUNDPS </c> instruction.
510 ///
511 /// \param V
512 ///    A 256-bit vector of [8 x float].
513 /// \returns A 256-bit vector of [8 x float] containing the rounded down values.
514 #define _mm256_floor_ps(V) _mm256_round_ps((V), _MM_FROUND_FLOOR)
515 
516 /* Logical */
517 /// Performs a bitwise AND of two 256-bit vectors of [4 x double].
518 ///
519 /// \headerfile <x86intrin.h>
520 ///
521 /// This intrinsic corresponds to the <c> VANDPD </c> instruction.
522 ///
523 /// \param __a
524 ///    A 256-bit vector of [4 x double] containing one of the source operands.
525 /// \param __b
526 ///    A 256-bit vector of [4 x double] containing one of the source operands.
527 /// \returns A 256-bit vector of [4 x double] containing the bitwise AND of the
528 ///    values between both operands.
529 static __inline __m256d __DEFAULT_FN_ATTRS
530 _mm256_and_pd(__m256d __a, __m256d __b)
531 {
532   return (__m256d)((__v4du)__a & (__v4du)__b);
533 }
534 
535 /// Performs a bitwise AND of two 256-bit vectors of [8 x float].
536 ///
537 /// \headerfile <x86intrin.h>
538 ///
539 /// This intrinsic corresponds to the <c> VANDPS </c> instruction.
540 ///
541 /// \param __a
542 ///    A 256-bit vector of [8 x float] containing one of the source operands.
543 /// \param __b
544 ///    A 256-bit vector of [8 x float] containing one of the source operands.
545 /// \returns A 256-bit vector of [8 x float] containing the bitwise AND of the
546 ///    values between both operands.
547 static __inline __m256 __DEFAULT_FN_ATTRS
548 _mm256_and_ps(__m256 __a, __m256 __b)
549 {
550   return (__m256)((__v8su)__a & (__v8su)__b);
551 }
552 
553 /// Performs a bitwise AND of two 256-bit vectors of [4 x double], using
554 ///    the one's complement of the values contained in the first source operand.
555 ///
556 /// \headerfile <x86intrin.h>
557 ///
558 /// This intrinsic corresponds to the <c> VANDNPD </c> instruction.
559 ///
560 /// \param __a
561 ///    A 256-bit vector of [4 x double] containing the left source operand. The
562 ///    one's complement of this value is used in the bitwise AND.
563 /// \param __b
564 ///    A 256-bit vector of [4 x double] containing the right source operand.
565 /// \returns A 256-bit vector of [4 x double] containing the bitwise AND of the
566 ///    values of the second operand and the one's complement of the first
567 ///    operand.
568 static __inline __m256d __DEFAULT_FN_ATTRS
569 _mm256_andnot_pd(__m256d __a, __m256d __b)
570 {
571   return (__m256d)(~(__v4du)__a & (__v4du)__b);
572 }
573 
574 /// Performs a bitwise AND of two 256-bit vectors of [8 x float], using
575 ///    the one's complement of the values contained in the first source operand.
576 ///
577 /// \headerfile <x86intrin.h>
578 ///
579 /// This intrinsic corresponds to the <c> VANDNPS </c> instruction.
580 ///
581 /// \param __a
582 ///    A 256-bit vector of [8 x float] containing the left source operand. The
583 ///    one's complement of this value is used in the bitwise AND.
584 /// \param __b
585 ///    A 256-bit vector of [8 x float] containing the right source operand.
586 /// \returns A 256-bit vector of [8 x float] containing the bitwise AND of the
587 ///    values of the second operand and the one's complement of the first
588 ///    operand.
589 static __inline __m256 __DEFAULT_FN_ATTRS
590 _mm256_andnot_ps(__m256 __a, __m256 __b)
591 {
592   return (__m256)(~(__v8su)__a & (__v8su)__b);
593 }
594 
595 /// Performs a bitwise OR of two 256-bit vectors of [4 x double].
596 ///
597 /// \headerfile <x86intrin.h>
598 ///
599 /// This intrinsic corresponds to the <c> VORPD </c> instruction.
600 ///
601 /// \param __a
602 ///    A 256-bit vector of [4 x double] containing one of the source operands.
603 /// \param __b
604 ///    A 256-bit vector of [4 x double] containing one of the source operands.
605 /// \returns A 256-bit vector of [4 x double] containing the bitwise OR of the
606 ///    values between both operands.
607 static __inline __m256d __DEFAULT_FN_ATTRS
608 _mm256_or_pd(__m256d __a, __m256d __b)
609 {
610   return (__m256d)((__v4du)__a | (__v4du)__b);
611 }
612 
613 /// Performs a bitwise OR of two 256-bit vectors of [8 x float].
614 ///
615 /// \headerfile <x86intrin.h>
616 ///
617 /// This intrinsic corresponds to the <c> VORPS </c> instruction.
618 ///
619 /// \param __a
620 ///    A 256-bit vector of [8 x float] containing one of the source operands.
621 /// \param __b
622 ///    A 256-bit vector of [8 x float] containing one of the source operands.
623 /// \returns A 256-bit vector of [8 x float] containing the bitwise OR of the
624 ///    values between both operands.
625 static __inline __m256 __DEFAULT_FN_ATTRS
626 _mm256_or_ps(__m256 __a, __m256 __b)
627 {
628   return (__m256)((__v8su)__a | (__v8su)__b);
629 }
630 
631 /// Performs a bitwise XOR of two 256-bit vectors of [4 x double].
632 ///
633 /// \headerfile <x86intrin.h>
634 ///
635 /// This intrinsic corresponds to the <c> VXORPD </c> instruction.
636 ///
637 /// \param __a
638 ///    A 256-bit vector of [4 x double] containing one of the source operands.
639 /// \param __b
640 ///    A 256-bit vector of [4 x double] containing one of the source operands.
641 /// \returns A 256-bit vector of [4 x double] containing the bitwise XOR of the
642 ///    values between both operands.
643 static __inline __m256d __DEFAULT_FN_ATTRS
644 _mm256_xor_pd(__m256d __a, __m256d __b)
645 {
646   return (__m256d)((__v4du)__a ^ (__v4du)__b);
647 }
648 
649 /// Performs a bitwise XOR of two 256-bit vectors of [8 x float].
650 ///
651 /// \headerfile <x86intrin.h>
652 ///
653 /// This intrinsic corresponds to the <c> VXORPS </c> instruction.
654 ///
655 /// \param __a
656 ///    A 256-bit vector of [8 x float] containing one of the source operands.
657 /// \param __b
658 ///    A 256-bit vector of [8 x float] containing one of the source operands.
659 /// \returns A 256-bit vector of [8 x float] containing the bitwise XOR of the
660 ///    values between both operands.
661 static __inline __m256 __DEFAULT_FN_ATTRS
662 _mm256_xor_ps(__m256 __a, __m256 __b)
663 {
664   return (__m256)((__v8su)__a ^ (__v8su)__b);
665 }
666 
667 /* Horizontal arithmetic */
668 /// Horizontally adds the adjacent pairs of values contained in two
669 ///    256-bit vectors of [4 x double].
670 ///
671 /// \headerfile <x86intrin.h>
672 ///
673 /// This intrinsic corresponds to the <c> VHADDPD </c> instruction.
674 ///
675 /// \param __a
676 ///    A 256-bit vector of [4 x double] containing one of the source operands.
677 ///    The horizontal sums of the values are returned in the even-indexed
678 ///    elements of a vector of [4 x double].
679 /// \param __b
680 ///    A 256-bit vector of [4 x double] containing one of the source operands.
681 ///    The horizontal sums of the values are returned in the odd-indexed
682 ///    elements of a vector of [4 x double].
683 /// \returns A 256-bit vector of [4 x double] containing the horizontal sums of
684 ///    both operands.
685 static __inline __m256d __DEFAULT_FN_ATTRS
686 _mm256_hadd_pd(__m256d __a, __m256d __b)
687 {
688   return (__m256d)__builtin_ia32_haddpd256((__v4df)__a, (__v4df)__b);
689 }
690 
691 /// Horizontally adds the adjacent pairs of values contained in two
692 ///    256-bit vectors of [8 x float].
693 ///
694 /// \headerfile <x86intrin.h>
695 ///
696 /// This intrinsic corresponds to the <c> VHADDPS </c> instruction.
697 ///
698 /// \param __a
699 ///    A 256-bit vector of [8 x float] containing one of the source operands.
700 ///    The horizontal sums of the values are returned in the elements with
701 ///    index 0, 1, 4, 5 of a vector of [8 x float].
702 /// \param __b
703 ///    A 256-bit vector of [8 x float] containing one of the source operands.
704 ///    The horizontal sums of the values are returned in the elements with
705 ///    index 2, 3, 6, 7 of a vector of [8 x float].
706 /// \returns A 256-bit vector of [8 x float] containing the horizontal sums of
707 ///    both operands.
708 static __inline __m256 __DEFAULT_FN_ATTRS
709 _mm256_hadd_ps(__m256 __a, __m256 __b)
710 {
711   return (__m256)__builtin_ia32_haddps256((__v8sf)__a, (__v8sf)__b);
712 }
713 
714 /// Horizontally subtracts the adjacent pairs of values contained in two
715 ///    256-bit vectors of [4 x double].
716 ///
717 /// \headerfile <x86intrin.h>
718 ///
719 /// This intrinsic corresponds to the <c> VHSUBPD </c> instruction.
720 ///
721 /// \param __a
722 ///    A 256-bit vector of [4 x double] containing one of the source operands.
723 ///    The horizontal differences between the values are returned in the
724 ///    even-indexed elements of a vector of [4 x double].
725 /// \param __b
726 ///    A 256-bit vector of [4 x double] containing one of the source operands.
727 ///    The horizontal differences between the values are returned in the
728 ///    odd-indexed elements of a vector of [4 x double].
729 /// \returns A 256-bit vector of [4 x double] containing the horizontal
730 ///    differences of both operands.
731 static __inline __m256d __DEFAULT_FN_ATTRS
732 _mm256_hsub_pd(__m256d __a, __m256d __b)
733 {
734   return (__m256d)__builtin_ia32_hsubpd256((__v4df)__a, (__v4df)__b);
735 }
736 
737 /// Horizontally subtracts the adjacent pairs of values contained in two
738 ///    256-bit vectors of [8 x float].
739 ///
740 /// \headerfile <x86intrin.h>
741 ///
742 /// This intrinsic corresponds to the <c> VHSUBPS </c> instruction.
743 ///
744 /// \param __a
745 ///    A 256-bit vector of [8 x float] containing one of the source operands.
746 ///    The horizontal differences between the values are returned in the
747 ///    elements with index 0, 1, 4, 5 of a vector of [8 x float].
748 /// \param __b
749 ///    A 256-bit vector of [8 x float] containing one of the source operands.
750 ///    The horizontal differences between the values are returned in the
751 ///    elements with index 2, 3, 6, 7 of a vector of [8 x float].
752 /// \returns A 256-bit vector of [8 x float] containing the horizontal
753 ///    differences of both operands.
754 static __inline __m256 __DEFAULT_FN_ATTRS
755 _mm256_hsub_ps(__m256 __a, __m256 __b)
756 {
757   return (__m256)__builtin_ia32_hsubps256((__v8sf)__a, (__v8sf)__b);
758 }
759 
760 /* Vector permutations */
761 /// Copies the values in a 128-bit vector of [2 x double] as specified
762 ///    by the 128-bit integer vector operand.
763 ///
764 /// \headerfile <x86intrin.h>
765 ///
766 /// This intrinsic corresponds to the <c> VPERMILPD </c> instruction.
767 ///
768 /// \param __a
769 ///    A 128-bit vector of [2 x double].
770 /// \param __c
771 ///    A 128-bit integer vector operand specifying how the values are to be
772 ///    copied. \n
773 ///    Bit [1]: \n
774 ///      0: Bits [63:0] of the source are copied to bits [63:0] of the returned
775 ///         vector. \n
776 ///      1: Bits [127:64] of the source are copied to bits [63:0] of the
777 ///         returned vector. \n
778 ///    Bit [65]: \n
779 ///      0: Bits [63:0] of the source are copied to bits [127:64] of the
780 ///         returned vector. \n
781 ///      1: Bits [127:64] of the source are copied to bits [127:64] of the
782 ///         returned vector.
783 /// \returns A 128-bit vector of [2 x double] containing the copied values.
784 static __inline __m128d __DEFAULT_FN_ATTRS128
785 _mm_permutevar_pd(__m128d __a, __m128i __c)
786 {
787   return (__m128d)__builtin_ia32_vpermilvarpd((__v2df)__a, (__v2di)__c);
788 }
789 
790 /// Copies the values in a 256-bit vector of [4 x double] as specified
791 ///    by the 256-bit integer vector operand.
792 ///
793 /// \headerfile <x86intrin.h>
794 ///
795 /// This intrinsic corresponds to the <c> VPERMILPD </c> instruction.
796 ///
797 /// \param __a
798 ///    A 256-bit vector of [4 x double].
799 /// \param __c
800 ///    A 256-bit integer vector operand specifying how the values are to be
801 ///    copied. \n
802 ///    Bit [1]: \n
803 ///      0: Bits [63:0] of the source are copied to bits [63:0] of the returned
804 ///         vector. \n
805 ///      1: Bits [127:64] of the source are copied to bits [63:0] of the
806 ///         returned vector. \n
807 ///    Bit [65]: \n
808 ///      0: Bits [63:0] of the source are copied to bits [127:64] of the
809 ///         returned vector. \n
810 ///      1: Bits [127:64] of the source are copied to bits [127:64] of the
811 ///         returned vector. \n
812 ///    Bit [129]: \n
813 ///      0: Bits [191:128] of the source are copied to bits [191:128] of the
814 ///         returned vector. \n
815 ///      1: Bits [255:192] of the source are copied to bits [191:128] of the
816 ///         returned vector. \n
817 ///    Bit [193]: \n
818 ///      0: Bits [191:128] of the source are copied to bits [255:192] of the
819 ///         returned vector. \n
820 ///      1: Bits [255:192] of the source are copied to bits [255:192] of the
821 ///    returned vector.
822 /// \returns A 256-bit vector of [4 x double] containing the copied values.
823 static __inline __m256d __DEFAULT_FN_ATTRS
824 _mm256_permutevar_pd(__m256d __a, __m256i __c)
825 {
826   return (__m256d)__builtin_ia32_vpermilvarpd256((__v4df)__a, (__v4di)__c);
827 }
828 
829 /// Copies the values stored in a 128-bit vector of [4 x float] as
830 ///    specified by the 128-bit integer vector operand.
831 /// \headerfile <x86intrin.h>
832 ///
833 /// This intrinsic corresponds to the <c> VPERMILPS </c> instruction.
834 ///
835 /// \param __a
836 ///    A 128-bit vector of [4 x float].
837 /// \param __c
838 ///    A 128-bit integer vector operand specifying how the values are to be
839 ///    copied. \n
840 ///    Bits [1:0]: \n
841 ///      00: Bits [31:0] of the source are copied to bits [31:0] of the
842 ///          returned vector. \n
843 ///      01: Bits [63:32] of the source are copied to bits [31:0] of the
844 ///          returned vector. \n
845 ///      10: Bits [95:64] of the source are copied to bits [31:0] of the
846 ///          returned vector. \n
847 ///      11: Bits [127:96] of the source are copied to bits [31:0] of the
848 ///          returned vector. \n
849 ///    Bits [33:32]: \n
850 ///      00: Bits [31:0] of the source are copied to bits [63:32] of the
851 ///          returned vector. \n
852 ///      01: Bits [63:32] of the source are copied to bits [63:32] of the
853 ///          returned vector. \n
854 ///      10: Bits [95:64] of the source are copied to bits [63:32] of the
855 ///          returned vector. \n
856 ///      11: Bits [127:96] of the source are copied to bits [63:32] of the
857 ///          returned vector. \n
858 ///    Bits [65:64]: \n
859 ///      00: Bits [31:0] of the source are copied to bits [95:64] of the
860 ///          returned vector. \n
861 ///      01: Bits [63:32] of the source are copied to bits [95:64] of the
862 ///          returned vector. \n
863 ///      10: Bits [95:64] of the source are copied to bits [95:64] of the
864 ///          returned vector. \n
865 ///      11: Bits [127:96] of the source are copied to bits [95:64] of the
866 ///          returned vector. \n
867 ///    Bits [97:96]: \n
868 ///      00: Bits [31:0] of the source are copied to bits [127:96] of the
869 ///          returned vector. \n
870 ///      01: Bits [63:32] of the source are copied to bits [127:96] of the
871 ///          returned vector. \n
872 ///      10: Bits [95:64] of the source are copied to bits [127:96] of the
873 ///          returned vector. \n
874 ///      11: Bits [127:96] of the source are copied to bits [127:96] of the
875 ///          returned vector.
876 /// \returns A 128-bit vector of [4 x float] containing the copied values.
877 static __inline __m128 __DEFAULT_FN_ATTRS128
878 _mm_permutevar_ps(__m128 __a, __m128i __c)
879 {
880   return (__m128)__builtin_ia32_vpermilvarps((__v4sf)__a, (__v4si)__c);
881 }
882 
883 /// Copies the values stored in a 256-bit vector of [8 x float] as
884 ///    specified by the 256-bit integer vector operand.
885 ///
886 /// \headerfile <x86intrin.h>
887 ///
888 /// This intrinsic corresponds to the <c> VPERMILPS </c> instruction.
889 ///
890 /// \param __a
891 ///    A 256-bit vector of [8 x float].
892 /// \param __c
893 ///    A 256-bit integer vector operand specifying how the values are to be
894 ///    copied. \n
895 ///    Bits [1:0]: \n
896 ///      00: Bits [31:0] of the source are copied to bits [31:0] of the
897 ///          returned vector. \n
898 ///      01: Bits [63:32] of the source are copied to bits [31:0] of the
899 ///          returned vector. \n
900 ///      10: Bits [95:64] of the source are copied to bits [31:0] of the
901 ///          returned vector. \n
902 ///      11: Bits [127:96] of the source are copied to bits [31:0] of the
903 ///          returned vector. \n
904 ///    Bits [33:32]: \n
905 ///      00: Bits [31:0] of the source are copied to bits [63:32] of the
906 ///          returned vector. \n
907 ///      01: Bits [63:32] of the source are copied to bits [63:32] of the
908 ///          returned vector. \n
909 ///      10: Bits [95:64] of the source are copied to bits [63:32] of the
910 ///          returned vector. \n
911 ///      11: Bits [127:96] of the source are copied to bits [63:32] of the
912 ///          returned vector. \n
913 ///    Bits [65:64]: \n
914 ///      00: Bits [31:0] of the source are copied to bits [95:64] of the
915 ///          returned vector. \n
916 ///      01: Bits [63:32] of the source are copied to bits [95:64] of the
917 ///          returned vector. \n
918 ///      10: Bits [95:64] of the source are copied to bits [95:64] of the
919 ///          returned vector. \n
920 ///      11: Bits [127:96] of the source are copied to bits [95:64] of the
921 ///          returned vector. \n
922 ///    Bits [97:96]: \n
923 ///      00: Bits [31:0] of the source are copied to bits [127:96] of the
924 ///          returned vector. \n
925 ///      01: Bits [63:32] of the source are copied to bits [127:96] of the
926 ///          returned vector. \n
927 ///      10: Bits [95:64] of the source are copied to bits [127:96] of the
928 ///          returned vector. \n
929 ///      11: Bits [127:96] of the source are copied to bits [127:96] of the
930 ///          returned vector. \n
931 ///    Bits [129:128]: \n
932 ///      00: Bits [159:128] of the source are copied to bits [159:128] of the
933 ///          returned vector. \n
934 ///      01: Bits [191:160] of the source are copied to bits [159:128] of the
935 ///          returned vector. \n
936 ///      10: Bits [223:192] of the source are copied to bits [159:128] of the
937 ///          returned vector. \n
938 ///      11: Bits [255:224] of the source are copied to bits [159:128] of the
939 ///          returned vector. \n
940 ///    Bits [161:160]: \n
941 ///      00: Bits [159:128] of the source are copied to bits [191:160] of the
942 ///          returned vector. \n
943 ///      01: Bits [191:160] of the source are copied to bits [191:160] of the
944 ///          returned vector. \n
945 ///      10: Bits [223:192] of the source are copied to bits [191:160] of the
946 ///          returned vector. \n
947 ///      11: Bits [255:224] of the source are copied to bits [191:160] of the
948 ///          returned vector. \n
949 ///    Bits [193:192]: \n
950 ///      00: Bits [159:128] of the source are copied to bits [223:192] of the
951 ///          returned vector. \n
952 ///      01: Bits [191:160] of the source are copied to bits [223:192] of the
953 ///          returned vector. \n
954 ///      10: Bits [223:192] of the source are copied to bits [223:192] of the
955 ///          returned vector. \n
956 ///      11: Bits [255:224] of the source are copied to bits [223:192] of the
957 ///          returned vector. \n
958 ///    Bits [225:224]: \n
959 ///      00: Bits [159:128] of the source are copied to bits [255:224] of the
960 ///          returned vector. \n
961 ///      01: Bits [191:160] of the source are copied to bits [255:224] of the
962 ///          returned vector. \n
963 ///      10: Bits [223:192] of the source are copied to bits [255:224] of the
964 ///          returned vector. \n
965 ///      11: Bits [255:224] of the source are copied to bits [255:224] of the
966 ///          returned vector.
967 /// \returns A 256-bit vector of [8 x float] containing the copied values.
968 static __inline __m256 __DEFAULT_FN_ATTRS
969 _mm256_permutevar_ps(__m256 __a, __m256i __c)
970 {
971   return (__m256)__builtin_ia32_vpermilvarps256((__v8sf)__a, (__v8si)__c);
972 }
973 
974 /// Copies the values in a 128-bit vector of [2 x double] as specified
975 ///    by the immediate integer operand.
976 ///
977 /// \headerfile <x86intrin.h>
978 ///
979 /// \code
980 /// __m128d _mm_permute_pd(__m128d A, const int C);
981 /// \endcode
982 ///
983 /// This intrinsic corresponds to the <c> VPERMILPD </c> instruction.
984 ///
985 /// \param A
986 ///    A 128-bit vector of [2 x double].
987 /// \param C
988 ///    An immediate integer operand specifying how the values are to be
989 ///    copied. \n
990 ///    Bit [0]: \n
991 ///      0: Bits [63:0] of the source are copied to bits [63:0] of the returned
992 ///         vector. \n
993 ///      1: Bits [127:64] of the source are copied to bits [63:0] of the
994 ///         returned vector. \n
995 ///    Bit [1]: \n
996 ///      0: Bits [63:0] of the source are copied to bits [127:64] of the
997 ///         returned vector. \n
998 ///      1: Bits [127:64] of the source are copied to bits [127:64] of the
999 ///         returned vector.
1000 /// \returns A 128-bit vector of [2 x double] containing the copied values.
1001 #define _mm_permute_pd(A, C) \
1002   ((__m128d)__builtin_ia32_vpermilpd((__v2df)(__m128d)(A), (int)(C)))
1003 
1004 /// Copies the values in a 256-bit vector of [4 x double] as specified by
1005 ///    the immediate integer operand.
1006 ///
1007 /// \headerfile <x86intrin.h>
1008 ///
1009 /// \code
1010 /// __m256d _mm256_permute_pd(__m256d A, const int C);
1011 /// \endcode
1012 ///
1013 /// This intrinsic corresponds to the <c> VPERMILPD </c> instruction.
1014 ///
1015 /// \param A
1016 ///    A 256-bit vector of [4 x double].
1017 /// \param C
1018 ///    An immediate integer operand specifying how the values are to be
1019 ///    copied. \n
1020 ///    Bit [0]: \n
1021 ///      0: Bits [63:0] of the source are copied to bits [63:0] of the returned
1022 ///         vector. \n
1023 ///      1: Bits [127:64] of the source are copied to bits [63:0] of the
1024 ///         returned vector. \n
1025 ///    Bit [1]: \n
1026 ///      0: Bits [63:0] of the source are copied to bits [127:64] of the
1027 ///         returned vector. \n
1028 ///      1: Bits [127:64] of the source are copied to bits [127:64] of the
1029 ///         returned vector. \n
1030 ///    Bit [2]: \n
1031 ///      0: Bits [191:128] of the source are copied to bits [191:128] of the
1032 ///         returned vector. \n
1033 ///      1: Bits [255:192] of the source are copied to bits [191:128] of the
1034 ///         returned vector. \n
1035 ///    Bit [3]: \n
1036 ///      0: Bits [191:128] of the source are copied to bits [255:192] of the
1037 ///         returned vector. \n
1038 ///      1: Bits [255:192] of the source are copied to bits [255:192] of the
1039 ///         returned vector.
1040 /// \returns A 256-bit vector of [4 x double] containing the copied values.
1041 #define _mm256_permute_pd(A, C) \
1042   ((__m256d)__builtin_ia32_vpermilpd256((__v4df)(__m256d)(A), (int)(C)))
1043 
1044 /// Copies the values in a 128-bit vector of [4 x float] as specified by
1045 ///    the immediate integer operand.
1046 ///
1047 /// \headerfile <x86intrin.h>
1048 ///
1049 /// \code
1050 /// __m128 _mm_permute_ps(__m128 A, const int C);
1051 /// \endcode
1052 ///
1053 /// This intrinsic corresponds to the <c> VPERMILPS </c> instruction.
1054 ///
1055 /// \param A
1056 ///    A 128-bit vector of [4 x float].
1057 /// \param C
1058 ///    An immediate integer operand specifying how the values are to be
1059 ///    copied. \n
1060 ///    Bits [1:0]: \n
1061 ///      00: Bits [31:0] of the source are copied to bits [31:0] of the
1062 ///          returned vector. \n
1063 ///      01: Bits [63:32] of the source are copied to bits [31:0] of the
1064 ///          returned vector. \n
1065 ///      10: Bits [95:64] of the source are copied to bits [31:0] of the
1066 ///          returned vector. \n
1067 ///      11: Bits [127:96] of the source are copied to bits [31:0] of the
1068 ///          returned vector. \n
1069 ///    Bits [3:2]: \n
1070 ///      00: Bits [31:0] of the source are copied to bits [63:32] of the
1071 ///          returned vector. \n
1072 ///      01: Bits [63:32] of the source are copied to bits [63:32] of the
1073 ///          returned vector. \n
1074 ///      10: Bits [95:64] of the source are copied to bits [63:32] of the
1075 ///          returned vector. \n
1076 ///      11: Bits [127:96] of the source are copied to bits [63:32] of the
1077 ///          returned vector. \n
1078 ///    Bits [5:4]: \n
1079 ///      00: Bits [31:0] of the source are copied to bits [95:64] of the
1080 ///          returned vector. \n
1081 ///      01: Bits [63:32] of the source are copied to bits [95:64] of the
1082 ///          returned vector. \n
1083 ///      10: Bits [95:64] of the source are copied to bits [95:64] of the
1084 ///          returned vector. \n
1085 ///      11: Bits [127:96] of the source are copied to bits [95:64] of the
1086 ///          returned vector. \n
1087 ///    Bits [7:6]: \n
1088 ///      00: Bits [31:0] of the source are copied to bits [127:96] of the
1089 ///          returned vector. \n
1090 ///      01: Bits [63:32] of the source are copied to bits [127:96] of the
1091 ///          returned vector. \n
1092 ///      10: Bits [95:64] of the source are copied to bits [127:96] of the
1093 ///          returned vector. \n
1094 ///      11: Bits [127:96] of the source are copied to bits [127:96] of the
1095 ///          returned vector.
1096 /// \returns A 128-bit vector of [4 x float] containing the copied values.
1097 #define _mm_permute_ps(A, C) \
1098   ((__m128)__builtin_ia32_vpermilps((__v4sf)(__m128)(A), (int)(C)))
1099 
1100 /// Copies the values in a 256-bit vector of [8 x float] as specified by
1101 ///    the immediate integer operand.
1102 ///
1103 /// \headerfile <x86intrin.h>
1104 ///
1105 /// \code
1106 /// __m256 _mm256_permute_ps(__m256 A, const int C);
1107 /// \endcode
1108 ///
1109 /// This intrinsic corresponds to the <c> VPERMILPS </c> instruction.
1110 ///
1111 /// \param A
1112 ///    A 256-bit vector of [8 x float].
1113 /// \param C
1114 ///    An immediate integer operand specifying how the values are to be
1115 ///    copied. \n
1116 ///    Bits [1:0]: \n
1117 ///      00: Bits [31:0] of the source are copied to bits [31:0] of the
1118 ///          returned vector. \n
1119 ///      01: Bits [63:32] of the source are copied to bits [31:0] of the
1120 ///          returned vector. \n
1121 ///      10: Bits [95:64] of the source are copied to bits [31:0] of the
1122 ///          returned vector. \n
1123 ///      11: Bits [127:96] of the source are copied to bits [31:0] of the
1124 ///          returned vector. \n
1125 ///    Bits [3:2]: \n
1126 ///      00: Bits [31:0] of the source are copied to bits [63:32] of the
1127 ///          returned vector. \n
1128 ///      01: Bits [63:32] of the source are copied to bits [63:32] of the
1129 ///          returned vector. \n
1130 ///      10: Bits [95:64] of the source are copied to bits [63:32] of the
1131 ///          returned vector. \n
1132 ///      11: Bits [127:96] of the source are copied to bits [63:32] of the
1133 ///          returned vector. \n
1134 ///    Bits [5:4]: \n
1135 ///      00: Bits [31:0] of the source are copied to bits [95:64] of the
1136 ///          returned vector. \n
1137 ///      01: Bits [63:32] of the source are copied to bits [95:64] of the
1138 ///          returned vector. \n
1139 ///      10: Bits [95:64] of the source are copied to bits [95:64] of the
1140 ///          returned vector. \n
1141 ///      11: Bits [127:96] of the source are copied to bits [95:64] of the
1142 ///          returned vector. \n
1143 ///    Bits [7:6]: \n
1144 ///      00: Bits [31:0] of the source are copied to bits [127:96] of the
1145 ///          returned vector. \n
1146 ///      01: Bits [63:32] of the source are copied to bits [127:96] of the
1147 ///          returned vector. \n
1148 ///      10: Bits [95:64] of the source are copied to bits [127:96] of the
1149 ///          returned vector. \n
1150 ///      11: Bits [127:96] of the source are copied to bits [127:96] of the
1151 ///          returned vector. \n
1152 ///    Bits [1:0]: \n
1153 ///      00: Bits [159:128] of the source are copied to bits [159:128] of the
1154 ///          returned vector. \n
1155 ///      01: Bits [191:160] of the source are copied to bits [159:128] of the
1156 ///          returned vector. \n
1157 ///      10: Bits [223:192] of the source are copied to bits [159:128] of the
1158 ///          returned vector. \n
1159 ///      11: Bits [255:224] of the source are copied to bits [159:128] of the
1160 ///          returned vector. \n
1161 ///    Bits [3:2]: \n
1162 ///      00: Bits [159:128] of the source are copied to bits [191:160] of the
1163 ///          returned vector. \n
1164 ///      01: Bits [191:160] of the source are copied to bits [191:160] of the
1165 ///          returned vector. \n
1166 ///      10: Bits [223:192] of the source are copied to bits [191:160] of the
1167 ///          returned vector. \n
1168 ///      11: Bits [255:224] of the source are copied to bits [191:160] of the
1169 ///          returned vector. \n
1170 ///    Bits [5:4]: \n
1171 ///      00: Bits [159:128] of the source are copied to bits [223:192] of the
1172 ///          returned vector. \n
1173 ///      01: Bits [191:160] of the source are copied to bits [223:192] of the
1174 ///          returned vector. \n
1175 ///      10: Bits [223:192] of the source are copied to bits [223:192] of the
1176 ///          returned vector. \n
1177 ///      11: Bits [255:224] of the source are copied to bits [223:192] of the
1178 ///          returned vector. \n
1179 ///    Bits [7:6]: \n
1180 ///      00: Bits [159:128] of the source are copied to bits [255:224] of the
1181 ///          returned vector. \n
1182 ///      01: Bits [191:160] of the source are copied to bits [255:224] of the
1183 ///          returned vector. \n
1184 ///      10: Bits [223:192] of the source are copied to bits [255:224] of the
1185 ///          returned vector. \n
1186 ///      11: Bits [255:224] of the source are copied to bits [255:224] of the
1187 ///          returned vector.
1188 /// \returns A 256-bit vector of [8 x float] containing the copied values.
1189 #define _mm256_permute_ps(A, C) \
1190   ((__m256)__builtin_ia32_vpermilps256((__v8sf)(__m256)(A), (int)(C)))
1191 
1192 /// Permutes 128-bit data values stored in two 256-bit vectors of
1193 ///    [4 x double], as specified by the immediate integer operand.
1194 ///
1195 /// \headerfile <x86intrin.h>
1196 ///
1197 /// \code
1198 /// __m256d _mm256_permute2f128_pd(__m256d V1, __m256d V2, const int M);
1199 /// \endcode
1200 ///
1201 /// This intrinsic corresponds to the <c> VPERM2F128 </c> instruction.
1202 ///
1203 /// \param V1
1204 ///    A 256-bit vector of [4 x double].
1205 /// \param V2
1206 ///    A 256-bit vector of [4 x double.
1207 /// \param M
1208 ///    An immediate integer operand specifying how the values are to be
1209 ///    permuted. \n
1210 ///    Bits [1:0]: \n
1211 ///      00: Bits [127:0] of operand \a V1 are copied to bits [127:0] of the
1212 ///          destination. \n
1213 ///      01: Bits [255:128] of operand \a V1 are copied to bits [127:0] of the
1214 ///          destination. \n
1215 ///      10: Bits [127:0] of operand \a V2 are copied to bits [127:0] of the
1216 ///          destination. \n
1217 ///      11: Bits [255:128] of operand \a V2 are copied to bits [127:0] of the
1218 ///          destination. \n
1219 ///    Bits [5:4]: \n
1220 ///      00: Bits [127:0] of operand \a V1 are copied to bits [255:128] of the
1221 ///          destination. \n
1222 ///      01: Bits [255:128] of operand \a V1 are copied to bits [255:128] of the
1223 ///          destination. \n
1224 ///      10: Bits [127:0] of operand \a V2 are copied to bits [255:128] of the
1225 ///          destination. \n
1226 ///      11: Bits [255:128] of operand \a V2 are copied to bits [255:128] of the
1227 ///          destination.
1228 /// \returns A 256-bit vector of [4 x double] containing the copied values.
1229 #define _mm256_permute2f128_pd(V1, V2, M) \
1230   ((__m256d)__builtin_ia32_vperm2f128_pd256((__v4df)(__m256d)(V1), \
1231                                             (__v4df)(__m256d)(V2), (int)(M)))
1232 
1233 /// Permutes 128-bit data values stored in two 256-bit vectors of
1234 ///    [8 x float], as specified by the immediate integer operand.
1235 ///
1236 /// \headerfile <x86intrin.h>
1237 ///
1238 /// \code
1239 /// __m256 _mm256_permute2f128_ps(__m256 V1, __m256 V2, const int M);
1240 /// \endcode
1241 ///
1242 /// This intrinsic corresponds to the <c> VPERM2F128 </c> instruction.
1243 ///
1244 /// \param V1
1245 ///    A 256-bit vector of [8 x float].
1246 /// \param V2
1247 ///    A 256-bit vector of [8 x float].
1248 /// \param M
1249 ///    An immediate integer operand specifying how the values are to be
1250 ///    permuted. \n
1251 ///    Bits [1:0]: \n
1252 ///    00: Bits [127:0] of operand \a V1 are copied to bits [127:0] of the
1253 ///    destination. \n
1254 ///    01: Bits [255:128] of operand \a V1 are copied to bits [127:0] of the
1255 ///    destination. \n
1256 ///    10: Bits [127:0] of operand \a V2 are copied to bits [127:0] of the
1257 ///    destination. \n
1258 ///    11: Bits [255:128] of operand \a V2 are copied to bits [127:0] of the
1259 ///    destination. \n
1260 ///    Bits [5:4]: \n
1261 ///    00: Bits [127:0] of operand \a V1 are copied to bits [255:128] of the
1262 ///    destination. \n
1263 ///    01: Bits [255:128] of operand \a V1 are copied to bits [255:128] of the
1264 ///    destination. \n
1265 ///    10: Bits [127:0] of operand \a V2 are copied to bits [255:128] of the
1266 ///    destination. \n
1267 ///    11: Bits [255:128] of operand \a V2 are copied to bits [255:128] of the
1268 ///    destination.
1269 /// \returns A 256-bit vector of [8 x float] containing the copied values.
1270 #define _mm256_permute2f128_ps(V1, V2, M) \
1271   ((__m256)__builtin_ia32_vperm2f128_ps256((__v8sf)(__m256)(V1), \
1272                                            (__v8sf)(__m256)(V2), (int)(M)))
1273 
1274 /// Permutes 128-bit data values stored in two 256-bit integer vectors,
1275 ///    as specified by the immediate integer operand.
1276 ///
1277 /// \headerfile <x86intrin.h>
1278 ///
1279 /// \code
1280 /// __m256i _mm256_permute2f128_si256(__m256i V1, __m256i V2, const int M);
1281 /// \endcode
1282 ///
1283 /// This intrinsic corresponds to the <c> VPERM2F128 </c> instruction.
1284 ///
1285 /// \param V1
1286 ///    A 256-bit integer vector.
1287 /// \param V2
1288 ///    A 256-bit integer vector.
1289 /// \param M
1290 ///    An immediate integer operand specifying how the values are to be copied.
1291 ///    Bits [1:0]: \n
1292 ///    00: Bits [127:0] of operand \a V1 are copied to bits [127:0] of the
1293 ///    destination. \n
1294 ///    01: Bits [255:128] of operand \a V1 are copied to bits [127:0] of the
1295 ///    destination. \n
1296 ///    10: Bits [127:0] of operand \a V2 are copied to bits [127:0] of the
1297 ///    destination. \n
1298 ///    11: Bits [255:128] of operand \a V2 are copied to bits [127:0] of the
1299 ///    destination. \n
1300 ///    Bits [5:4]: \n
1301 ///    00: Bits [127:0] of operand \a V1 are copied to bits [255:128] of the
1302 ///    destination. \n
1303 ///    01: Bits [255:128] of operand \a V1 are copied to bits [255:128] of the
1304 ///    destination. \n
1305 ///    10: Bits [127:0] of operand \a V2 are copied to bits [255:128] of the
1306 ///    destination. \n
1307 ///    11: Bits [255:128] of operand \a V2 are copied to bits [255:128] of the
1308 ///    destination.
1309 /// \returns A 256-bit integer vector containing the copied values.
1310 #define _mm256_permute2f128_si256(V1, V2, M) \
1311   ((__m256i)__builtin_ia32_vperm2f128_si256((__v8si)(__m256i)(V1), \
1312                                             (__v8si)(__m256i)(V2), (int)(M)))
1313 
1314 /* Vector Blend */
1315 /// Merges 64-bit double-precision data values stored in either of the
1316 ///    two 256-bit vectors of [4 x double], as specified by the immediate
1317 ///    integer operand.
1318 ///
1319 /// \headerfile <x86intrin.h>
1320 ///
1321 /// \code
1322 /// __m256d _mm256_blend_pd(__m256d V1, __m256d V2, const int M);
1323 /// \endcode
1324 ///
1325 /// This intrinsic corresponds to the <c> VBLENDPD </c> instruction.
1326 ///
1327 /// \param V1
1328 ///    A 256-bit vector of [4 x double].
1329 /// \param V2
1330 ///    A 256-bit vector of [4 x double].
1331 /// \param M
1332 ///    An immediate integer operand, with mask bits [3:0] specifying how the
1333 ///    values are to be copied. The position of the mask bit corresponds to the
1334 ///    index of a copied value. When a mask bit is 0, the corresponding 64-bit
1335 ///    element in operand \a V1 is copied to the same position in the
1336 ///    destination. When a mask bit is 1, the corresponding 64-bit element in
1337 ///    operand \a V2 is copied to the same position in the destination.
1338 /// \returns A 256-bit vector of [4 x double] containing the copied values.
1339 #define _mm256_blend_pd(V1, V2, M) \
1340   ((__m256d)__builtin_ia32_blendpd256((__v4df)(__m256d)(V1), \
1341                                       (__v4df)(__m256d)(V2), (int)(M)))
1342 
1343 /// Merges 32-bit single-precision data values stored in either of the
1344 ///    two 256-bit vectors of [8 x float], as specified by the immediate
1345 ///    integer operand.
1346 ///
1347 /// \headerfile <x86intrin.h>
1348 ///
1349 /// \code
1350 /// __m256 _mm256_blend_ps(__m256 V1, __m256 V2, const int M);
1351 /// \endcode
1352 ///
1353 /// This intrinsic corresponds to the <c> VBLENDPS </c> instruction.
1354 ///
1355 /// \param V1
1356 ///    A 256-bit vector of [8 x float].
1357 /// \param V2
1358 ///    A 256-bit vector of [8 x float].
1359 /// \param M
1360 ///    An immediate integer operand, with mask bits [7:0] specifying how the
1361 ///    values are to be copied. The position of the mask bit corresponds to the
1362 ///    index of a copied value. When a mask bit is 0, the corresponding 32-bit
1363 ///    element in operand \a V1 is copied to the same position in the
1364 ///    destination. When a mask bit is 1, the corresponding 32-bit element in
1365 ///    operand \a V2 is copied to the same position in the destination.
1366 /// \returns A 256-bit vector of [8 x float] containing the copied values.
1367 #define _mm256_blend_ps(V1, V2, M) \
1368   ((__m256)__builtin_ia32_blendps256((__v8sf)(__m256)(V1), \
1369                                      (__v8sf)(__m256)(V2), (int)(M)))
1370 
1371 /// Merges 64-bit double-precision data values stored in either of the
1372 ///    two 256-bit vectors of [4 x double], as specified by the 256-bit vector
1373 ///    operand.
1374 ///
1375 /// \headerfile <x86intrin.h>
1376 ///
1377 /// This intrinsic corresponds to the <c> VBLENDVPD </c> instruction.
1378 ///
1379 /// \param __a
1380 ///    A 256-bit vector of [4 x double].
1381 /// \param __b
1382 ///    A 256-bit vector of [4 x double].
1383 /// \param __c
1384 ///    A 256-bit vector operand, with mask bits 255, 191, 127, and 63 specifying
1385 ///    how the values are to be copied. The position of the mask bit corresponds
1386 ///    to the most significant bit of a copied value. When a mask bit is 0, the
1387 ///    corresponding 64-bit element in operand \a __a is copied to the same
1388 ///    position in the destination. When a mask bit is 1, the corresponding
1389 ///    64-bit element in operand \a __b is copied to the same position in the
1390 ///    destination.
1391 /// \returns A 256-bit vector of [4 x double] containing the copied values.
1392 static __inline __m256d __DEFAULT_FN_ATTRS
1393 _mm256_blendv_pd(__m256d __a, __m256d __b, __m256d __c)
1394 {
1395   return (__m256d)__builtin_ia32_blendvpd256(
1396     (__v4df)__a, (__v4df)__b, (__v4df)__c);
1397 }
1398 
1399 /// Merges 32-bit single-precision data values stored in either of the
1400 ///    two 256-bit vectors of [8 x float], as specified by the 256-bit vector
1401 ///    operand.
1402 ///
1403 /// \headerfile <x86intrin.h>
1404 ///
1405 /// This intrinsic corresponds to the <c> VBLENDVPS </c> instruction.
1406 ///
1407 /// \param __a
1408 ///    A 256-bit vector of [8 x float].
1409 /// \param __b
1410 ///    A 256-bit vector of [8 x float].
1411 /// \param __c
1412 ///    A 256-bit vector operand, with mask bits 255, 223, 191, 159, 127, 95, 63,
1413 ///    and 31 specifying how the values are to be copied. The position of the
1414 ///    mask bit corresponds to the most significant bit of a copied value. When
1415 ///    a mask bit is 0, the corresponding 32-bit element in operand \a __a is
1416 ///    copied to the same position in the destination. When a mask bit is 1, the
1417 ///    corresponding 32-bit element in operand \a __b is copied to the same
1418 ///    position in the destination.
1419 /// \returns A 256-bit vector of [8 x float] containing the copied values.
1420 static __inline __m256 __DEFAULT_FN_ATTRS
1421 _mm256_blendv_ps(__m256 __a, __m256 __b, __m256 __c)
1422 {
1423   return (__m256)__builtin_ia32_blendvps256(
1424     (__v8sf)__a, (__v8sf)__b, (__v8sf)__c);
1425 }
1426 
1427 /* Vector Dot Product */
1428 /// Computes two dot products in parallel, using the lower and upper
1429 ///    halves of two [8 x float] vectors as input to the two computations, and
1430 ///    returning the two dot products in the lower and upper halves of the
1431 ///    [8 x float] result.
1432 ///
1433 ///    The immediate integer operand controls which input elements will
1434 ///    contribute to the dot product, and where the final results are returned.
1435 ///    In general, for each dot product, the four corresponding elements of the
1436 ///    input vectors are multiplied; the first two and second two products are
1437 ///    summed, then the two sums are added to form the final result.
1438 ///
1439 /// \headerfile <x86intrin.h>
1440 ///
1441 /// \code
1442 /// __m256 _mm256_dp_ps(__m256 V1, __m256 V2, const int M);
1443 /// \endcode
1444 ///
1445 /// This intrinsic corresponds to the <c> VDPPS </c> instruction.
1446 ///
1447 /// \param V1
1448 ///    A vector of [8 x float] values, treated as two [4 x float] vectors.
1449 /// \param V2
1450 ///    A vector of [8 x float] values, treated as two [4 x float] vectors.
1451 /// \param M
1452 ///    An immediate integer argument. Bits [7:4] determine which elements of
1453 ///    the input vectors are used, with bit [4] corresponding to the lowest
1454 ///    element and bit [7] corresponding to the highest element of each [4 x
1455 ///    float] subvector. If a bit is set, the corresponding elements from the
1456 ///    two input vectors are used as an input for dot product; otherwise that
1457 ///    input is treated as zero. Bits [3:0] determine which elements of the
1458 ///    result will receive a copy of the final dot product, with bit [0]
1459 ///    corresponding to the lowest element and bit [3] corresponding to the
1460 ///    highest element of each [4 x float] subvector. If a bit is set, the dot
1461 ///    product is returned in the corresponding element; otherwise that element
1462 ///    is set to zero. The bitmask is applied in the same way to each of the
1463 ///    two parallel dot product computations.
1464 /// \returns A 256-bit vector of [8 x float] containing the two dot products.
1465 #define _mm256_dp_ps(V1, V2, M) \
1466   ((__m256)__builtin_ia32_dpps256((__v8sf)(__m256)(V1), \
1467                                   (__v8sf)(__m256)(V2), (M)))
1468 
1469 /* Vector shuffle */
1470 /// Selects 8 float values from the 256-bit operands of [8 x float], as
1471 ///    specified by the immediate value operand.
1472 ///
1473 ///    The four selected elements in each operand are copied to the destination
1474 ///    according to the bits specified in the immediate operand. The selected
1475 ///    elements from the first 256-bit operand are copied to bits [63:0] and
1476 ///    bits [191:128] of the destination, and the selected elements from the
1477 ///    second 256-bit operand are copied to bits [127:64] and bits [255:192] of
1478 ///    the destination. For example, if bits [7:0] of the immediate operand
1479 ///    contain a value of 0xFF, the 256-bit destination vector would contain the
1480 ///    following values: b[7], b[7], a[7], a[7], b[3], b[3], a[3], a[3].
1481 ///
1482 /// \headerfile <x86intrin.h>
1483 ///
1484 /// \code
1485 /// __m256 _mm256_shuffle_ps(__m256 a, __m256 b, const int mask);
1486 /// \endcode
1487 ///
1488 /// This intrinsic corresponds to the <c> VSHUFPS </c> instruction.
1489 ///
1490 /// \param a
1491 ///    A 256-bit vector of [8 x float]. The four selected elements in this
1492 ///    operand are copied to bits [63:0] and bits [191:128] in the destination,
1493 ///    according to the bits specified in the immediate operand.
1494 /// \param b
1495 ///    A 256-bit vector of [8 x float]. The four selected elements in this
1496 ///    operand are copied to bits [127:64] and bits [255:192] in the
1497 ///    destination, according to the bits specified in the immediate operand.
1498 /// \param mask
1499 ///    An immediate value containing an 8-bit value specifying which elements to
1500 ///    copy from \a a and \a b \n.
1501 ///    Bits [3:0] specify the values copied from operand \a a. \n
1502 ///    Bits [7:4] specify the values copied from operand \a b. \n
1503 ///    The destinations within the 256-bit destination are assigned values as
1504 ///    follows, according to the bit value assignments described below: \n
1505 ///    Bits [1:0] are used to assign values to bits [31:0] and [159:128] in the
1506 ///    destination. \n
1507 ///    Bits [3:2] are used to assign values to bits [63:32] and [191:160] in the
1508 ///    destination. \n
1509 ///    Bits [5:4] are used to assign values to bits [95:64] and [223:192] in the
1510 ///    destination. \n
1511 ///    Bits [7:6] are used to assign values to bits [127:96] and [255:224] in
1512 ///    the destination. \n
1513 ///    Bit value assignments: \n
1514 ///    00: Bits [31:0] and [159:128] are copied from the selected operand. \n
1515 ///    01: Bits [63:32] and [191:160] are copied from the selected operand. \n
1516 ///    10: Bits [95:64] and [223:192] are copied from the selected operand. \n
1517 ///    11: Bits [127:96] and [255:224] are copied from the selected operand. \n
1518 ///    Note: To generate a mask, you can use the \c _MM_SHUFFLE macro.
1519 ///    <c>_MM_SHUFFLE(b6, b4, b2, b0)</c> can create an 8-bit mask of the form
1520 ///    <c>[b6, b4, b2, b0]</c>.
1521 /// \returns A 256-bit vector of [8 x float] containing the shuffled values.
1522 #define _mm256_shuffle_ps(a, b, mask) \
1523   ((__m256)__builtin_ia32_shufps256((__v8sf)(__m256)(a), \
1524                                     (__v8sf)(__m256)(b), (int)(mask)))
1525 
1526 /// Selects four double-precision values from the 256-bit operands of
1527 ///    [4 x double], as specified by the immediate value operand.
1528 ///
1529 ///    The selected elements from the first 256-bit operand are copied to bits
1530 ///    [63:0] and bits [191:128] in the destination, and the selected elements
1531 ///    from the second 256-bit operand are copied to bits [127:64] and bits
1532 ///    [255:192] in the destination. For example, if bits [3:0] of the immediate
1533 ///    operand contain a value of 0xF, the 256-bit destination vector would
1534 ///    contain the following values: b[3], a[3], b[1], a[1].
1535 ///
1536 /// \headerfile <x86intrin.h>
1537 ///
1538 /// \code
1539 /// __m256d _mm256_shuffle_pd(__m256d a, __m256d b, const int mask);
1540 /// \endcode
1541 ///
1542 /// This intrinsic corresponds to the <c> VSHUFPD </c> instruction.
1543 ///
1544 /// \param a
1545 ///    A 256-bit vector of [4 x double].
1546 /// \param b
1547 ///    A 256-bit vector of [4 x double].
1548 /// \param mask
1549 ///    An immediate value containing 8-bit values specifying which elements to
1550 ///    copy from \a a and \a b: \n
1551 ///    Bit [0]=0: Bits [63:0] are copied from \a a to bits [63:0] of the
1552 ///    destination. \n
1553 ///    Bit [0]=1: Bits [127:64] are copied from \a a to bits [63:0] of the
1554 ///    destination. \n
1555 ///    Bit [1]=0: Bits [63:0] are copied from \a b to bits [127:64] of the
1556 ///    destination. \n
1557 ///    Bit [1]=1: Bits [127:64] are copied from \a b to bits [127:64] of the
1558 ///    destination. \n
1559 ///    Bit [2]=0: Bits [191:128] are copied from \a a to bits [191:128] of the
1560 ///    destination. \n
1561 ///    Bit [2]=1: Bits [255:192] are copied from \a a to bits [191:128] of the
1562 ///    destination. \n
1563 ///    Bit [3]=0: Bits [191:128] are copied from \a b to bits [255:192] of the
1564 ///    destination. \n
1565 ///    Bit [3]=1: Bits [255:192] are copied from \a b to bits [255:192] of the
1566 ///    destination.
1567 /// \returns A 256-bit vector of [4 x double] containing the shuffled values.
1568 #define _mm256_shuffle_pd(a, b, mask) \
1569   ((__m256d)__builtin_ia32_shufpd256((__v4df)(__m256d)(a), \
1570                                      (__v4df)(__m256d)(b), (int)(mask)))
1571 
1572 /* Compare */
1573 #define _CMP_EQ_OQ    0x00 /* Equal (ordered, non-signaling)  */
1574 #define _CMP_LT_OS    0x01 /* Less-than (ordered, signaling)  */
1575 #define _CMP_LE_OS    0x02 /* Less-than-or-equal (ordered, signaling)  */
1576 #define _CMP_UNORD_Q  0x03 /* Unordered (non-signaling)  */
1577 #define _CMP_NEQ_UQ   0x04 /* Not-equal (unordered, non-signaling)  */
1578 #define _CMP_NLT_US   0x05 /* Not-less-than (unordered, signaling)  */
1579 #define _CMP_NLE_US   0x06 /* Not-less-than-or-equal (unordered, signaling)  */
1580 #define _CMP_ORD_Q    0x07 /* Ordered (non-signaling)   */
1581 #define _CMP_EQ_UQ    0x08 /* Equal (unordered, non-signaling)  */
1582 #define _CMP_NGE_US   0x09 /* Not-greater-than-or-equal (unordered, signaling)  */
1583 #define _CMP_NGT_US   0x0a /* Not-greater-than (unordered, signaling)  */
1584 #define _CMP_FALSE_OQ 0x0b /* False (ordered, non-signaling)  */
1585 #define _CMP_NEQ_OQ   0x0c /* Not-equal (ordered, non-signaling)  */
1586 #define _CMP_GE_OS    0x0d /* Greater-than-or-equal (ordered, signaling)  */
1587 #define _CMP_GT_OS    0x0e /* Greater-than (ordered, signaling)  */
1588 #define _CMP_TRUE_UQ  0x0f /* True (unordered, non-signaling)  */
1589 #define _CMP_EQ_OS    0x10 /* Equal (ordered, signaling)  */
1590 #define _CMP_LT_OQ    0x11 /* Less-than (ordered, non-signaling)  */
1591 #define _CMP_LE_OQ    0x12 /* Less-than-or-equal (ordered, non-signaling)  */
1592 #define _CMP_UNORD_S  0x13 /* Unordered (signaling)  */
1593 #define _CMP_NEQ_US   0x14 /* Not-equal (unordered, signaling)  */
1594 #define _CMP_NLT_UQ   0x15 /* Not-less-than (unordered, non-signaling)  */
1595 #define _CMP_NLE_UQ   0x16 /* Not-less-than-or-equal (unordered, non-signaling)  */
1596 #define _CMP_ORD_S    0x17 /* Ordered (signaling)  */
1597 #define _CMP_EQ_US    0x18 /* Equal (unordered, signaling)  */
1598 #define _CMP_NGE_UQ   0x19 /* Not-greater-than-or-equal (unordered, non-signaling)  */
1599 #define _CMP_NGT_UQ   0x1a /* Not-greater-than (unordered, non-signaling)  */
1600 #define _CMP_FALSE_OS 0x1b /* False (ordered, signaling)  */
1601 #define _CMP_NEQ_OS   0x1c /* Not-equal (ordered, signaling)  */
1602 #define _CMP_GE_OQ    0x1d /* Greater-than-or-equal (ordered, non-signaling)  */
1603 #define _CMP_GT_OQ    0x1e /* Greater-than (ordered, non-signaling)  */
1604 #define _CMP_TRUE_US  0x1f /* True (unordered, signaling)  */
1605 
1606 /// Compares each of the corresponding double-precision values of two
1607 ///    128-bit vectors of [2 x double], using the operation specified by the
1608 ///    immediate integer operand.
1609 ///
1610 ///    Returns a [2 x double] vector consisting of two doubles corresponding to
1611 ///    the two comparison results: zero if the comparison is false, and all 1's
1612 ///    if the comparison is true.
1613 ///
1614 /// \headerfile <x86intrin.h>
1615 ///
1616 /// \code
1617 /// __m128d _mm_cmp_pd(__m128d a, __m128d b, const int c);
1618 /// \endcode
1619 ///
1620 /// This intrinsic corresponds to the <c> VCMPPD </c> instruction.
1621 ///
1622 /// \param a
1623 ///    A 128-bit vector of [2 x double].
1624 /// \param b
1625 ///    A 128-bit vector of [2 x double].
1626 /// \param c
1627 ///    An immediate integer operand, with bits [4:0] specifying which comparison
1628 ///    operation to use: \n
1629 ///    0x00: Equal (ordered, non-signaling) \n
1630 ///    0x01: Less-than (ordered, signaling) \n
1631 ///    0x02: Less-than-or-equal (ordered, signaling) \n
1632 ///    0x03: Unordered (non-signaling) \n
1633 ///    0x04: Not-equal (unordered, non-signaling) \n
1634 ///    0x05: Not-less-than (unordered, signaling) \n
1635 ///    0x06: Not-less-than-or-equal (unordered, signaling) \n
1636 ///    0x07: Ordered (non-signaling) \n
1637 ///    0x08: Equal (unordered, non-signaling) \n
1638 ///    0x09: Not-greater-than-or-equal (unordered, signaling) \n
1639 ///    0x0A: Not-greater-than (unordered, signaling) \n
1640 ///    0x0B: False (ordered, non-signaling) \n
1641 ///    0x0C: Not-equal (ordered, non-signaling) \n
1642 ///    0x0D: Greater-than-or-equal (ordered, signaling) \n
1643 ///    0x0E: Greater-than (ordered, signaling) \n
1644 ///    0x0F: True (unordered, non-signaling) \n
1645 ///    0x10: Equal (ordered, signaling) \n
1646 ///    0x11: Less-than (ordered, non-signaling) \n
1647 ///    0x12: Less-than-or-equal (ordered, non-signaling) \n
1648 ///    0x13: Unordered (signaling) \n
1649 ///    0x14: Not-equal (unordered, signaling) \n
1650 ///    0x15: Not-less-than (unordered, non-signaling) \n
1651 ///    0x16: Not-less-than-or-equal (unordered, non-signaling) \n
1652 ///    0x17: Ordered (signaling) \n
1653 ///    0x18: Equal (unordered, signaling) \n
1654 ///    0x19: Not-greater-than-or-equal (unordered, non-signaling) \n
1655 ///    0x1A: Not-greater-than (unordered, non-signaling) \n
1656 ///    0x1B: False (ordered, signaling) \n
1657 ///    0x1C: Not-equal (ordered, signaling) \n
1658 ///    0x1D: Greater-than-or-equal (ordered, non-signaling) \n
1659 ///    0x1E: Greater-than (ordered, non-signaling) \n
1660 ///    0x1F: True (unordered, signaling)
1661 /// \returns A 128-bit vector of [2 x double] containing the comparison results.
1662 #define _mm_cmp_pd(a, b, c) \
1663   ((__m128d)__builtin_ia32_cmppd((__v2df)(__m128d)(a), \
1664                                  (__v2df)(__m128d)(b), (c)))
1665 
1666 /// Compares each of the corresponding values of two 128-bit vectors of
1667 ///    [4 x float], using the operation specified by the immediate integer
1668 ///    operand.
1669 ///
1670 ///    Returns a [4 x float] vector consisting of four floats corresponding to
1671 ///    the four comparison results: zero if the comparison is false, and all 1's
1672 ///    if the comparison is true.
1673 ///
1674 /// \headerfile <x86intrin.h>
1675 ///
1676 /// \code
1677 /// __m128 _mm_cmp_ps(__m128 a, __m128 b, const int c);
1678 /// \endcode
1679 ///
1680 /// This intrinsic corresponds to the <c> VCMPPS </c> instruction.
1681 ///
1682 /// \param a
1683 ///    A 128-bit vector of [4 x float].
1684 /// \param b
1685 ///    A 128-bit vector of [4 x float].
1686 /// \param c
1687 ///    An immediate integer operand, with bits [4:0] specifying which comparison
1688 ///    operation to use: \n
1689 ///    0x00: Equal (ordered, non-signaling) \n
1690 ///    0x01: Less-than (ordered, signaling) \n
1691 ///    0x02: Less-than-or-equal (ordered, signaling) \n
1692 ///    0x03: Unordered (non-signaling) \n
1693 ///    0x04: Not-equal (unordered, non-signaling) \n
1694 ///    0x05: Not-less-than (unordered, signaling) \n
1695 ///    0x06: Not-less-than-or-equal (unordered, signaling) \n
1696 ///    0x07: Ordered (non-signaling) \n
1697 ///    0x08: Equal (unordered, non-signaling) \n
1698 ///    0x09: Not-greater-than-or-equal (unordered, signaling) \n
1699 ///    0x0A: Not-greater-than (unordered, signaling) \n
1700 ///    0x0B: False (ordered, non-signaling) \n
1701 ///    0x0C: Not-equal (ordered, non-signaling) \n
1702 ///    0x0D: Greater-than-or-equal (ordered, signaling) \n
1703 ///    0x0E: Greater-than (ordered, signaling) \n
1704 ///    0x0F: True (unordered, non-signaling) \n
1705 ///    0x10: Equal (ordered, signaling) \n
1706 ///    0x11: Less-than (ordered, non-signaling) \n
1707 ///    0x12: Less-than-or-equal (ordered, non-signaling) \n
1708 ///    0x13: Unordered (signaling) \n
1709 ///    0x14: Not-equal (unordered, signaling) \n
1710 ///    0x15: Not-less-than (unordered, non-signaling) \n
1711 ///    0x16: Not-less-than-or-equal (unordered, non-signaling) \n
1712 ///    0x17: Ordered (signaling) \n
1713 ///    0x18: Equal (unordered, signaling) \n
1714 ///    0x19: Not-greater-than-or-equal (unordered, non-signaling) \n
1715 ///    0x1A: Not-greater-than (unordered, non-signaling) \n
1716 ///    0x1B: False (ordered, signaling) \n
1717 ///    0x1C: Not-equal (ordered, signaling) \n
1718 ///    0x1D: Greater-than-or-equal (ordered, non-signaling) \n
1719 ///    0x1E: Greater-than (ordered, non-signaling) \n
1720 ///    0x1F: True (unordered, signaling)
1721 /// \returns A 128-bit vector of [4 x float] containing the comparison results.
1722 #define _mm_cmp_ps(a, b, c) \
1723   ((__m128)__builtin_ia32_cmpps((__v4sf)(__m128)(a), \
1724                                 (__v4sf)(__m128)(b), (c)))
1725 
1726 /// Compares each of the corresponding double-precision values of two
1727 ///    256-bit vectors of [4 x double], using the operation specified by the
1728 ///    immediate integer operand.
1729 ///
1730 ///    Returns a [4 x double] vector consisting of four doubles corresponding to
1731 ///    the four comparison results: zero if the comparison is false, and all 1's
1732 ///    if the comparison is true.
1733 ///
1734 /// \headerfile <x86intrin.h>
1735 ///
1736 /// \code
1737 /// __m256d _mm256_cmp_pd(__m256d a, __m256d b, const int c);
1738 /// \endcode
1739 ///
1740 /// This intrinsic corresponds to the <c> VCMPPD </c> instruction.
1741 ///
1742 /// \param a
1743 ///    A 256-bit vector of [4 x double].
1744 /// \param b
1745 ///    A 256-bit vector of [4 x double].
1746 /// \param c
1747 ///    An immediate integer operand, with bits [4:0] specifying which comparison
1748 ///    operation to use: \n
1749 ///    0x00: Equal (ordered, non-signaling) \n
1750 ///    0x01: Less-than (ordered, signaling) \n
1751 ///    0x02: Less-than-or-equal (ordered, signaling) \n
1752 ///    0x03: Unordered (non-signaling) \n
1753 ///    0x04: Not-equal (unordered, non-signaling) \n
1754 ///    0x05: Not-less-than (unordered, signaling) \n
1755 ///    0x06: Not-less-than-or-equal (unordered, signaling) \n
1756 ///    0x07: Ordered (non-signaling) \n
1757 ///    0x08: Equal (unordered, non-signaling) \n
1758 ///    0x09: Not-greater-than-or-equal (unordered, signaling) \n
1759 ///    0x0A: Not-greater-than (unordered, signaling) \n
1760 ///    0x0B: False (ordered, non-signaling) \n
1761 ///    0x0C: Not-equal (ordered, non-signaling) \n
1762 ///    0x0D: Greater-than-or-equal (ordered, signaling) \n
1763 ///    0x0E: Greater-than (ordered, signaling) \n
1764 ///    0x0F: True (unordered, non-signaling) \n
1765 ///    0x10: Equal (ordered, signaling) \n
1766 ///    0x11: Less-than (ordered, non-signaling) \n
1767 ///    0x12: Less-than-or-equal (ordered, non-signaling) \n
1768 ///    0x13: Unordered (signaling) \n
1769 ///    0x14: Not-equal (unordered, signaling) \n
1770 ///    0x15: Not-less-than (unordered, non-signaling) \n
1771 ///    0x16: Not-less-than-or-equal (unordered, non-signaling) \n
1772 ///    0x17: Ordered (signaling) \n
1773 ///    0x18: Equal (unordered, signaling) \n
1774 ///    0x19: Not-greater-than-or-equal (unordered, non-signaling) \n
1775 ///    0x1A: Not-greater-than (unordered, non-signaling) \n
1776 ///    0x1B: False (ordered, signaling) \n
1777 ///    0x1C: Not-equal (ordered, signaling) \n
1778 ///    0x1D: Greater-than-or-equal (ordered, non-signaling) \n
1779 ///    0x1E: Greater-than (ordered, non-signaling) \n
1780 ///    0x1F: True (unordered, signaling)
1781 /// \returns A 256-bit vector of [4 x double] containing the comparison results.
1782 #define _mm256_cmp_pd(a, b, c) \
1783   ((__m256d)__builtin_ia32_cmppd256((__v4df)(__m256d)(a), \
1784                                     (__v4df)(__m256d)(b), (c)))
1785 
1786 /// Compares each of the corresponding values of two 256-bit vectors of
1787 ///    [8 x float], using the operation specified by the immediate integer
1788 ///    operand.
1789 ///
1790 ///    Returns a [8 x float] vector consisting of eight floats corresponding to
1791 ///    the eight comparison results: zero if the comparison is false, and all
1792 ///    1's if the comparison is true.
1793 ///
1794 /// \headerfile <x86intrin.h>
1795 ///
1796 /// \code
1797 /// __m256 _mm256_cmp_ps(__m256 a, __m256 b, const int c);
1798 /// \endcode
1799 ///
1800 /// This intrinsic corresponds to the <c> VCMPPS </c> instruction.
1801 ///
1802 /// \param a
1803 ///    A 256-bit vector of [8 x float].
1804 /// \param b
1805 ///    A 256-bit vector of [8 x float].
1806 /// \param c
1807 ///    An immediate integer operand, with bits [4:0] specifying which comparison
1808 ///    operation to use: \n
1809 ///    0x00: Equal (ordered, non-signaling) \n
1810 ///    0x01: Less-than (ordered, signaling) \n
1811 ///    0x02: Less-than-or-equal (ordered, signaling) \n
1812 ///    0x03: Unordered (non-signaling) \n
1813 ///    0x04: Not-equal (unordered, non-signaling) \n
1814 ///    0x05: Not-less-than (unordered, signaling) \n
1815 ///    0x06: Not-less-than-or-equal (unordered, signaling) \n
1816 ///    0x07: Ordered (non-signaling) \n
1817 ///    0x08: Equal (unordered, non-signaling) \n
1818 ///    0x09: Not-greater-than-or-equal (unordered, signaling) \n
1819 ///    0x0A: Not-greater-than (unordered, signaling) \n
1820 ///    0x0B: False (ordered, non-signaling) \n
1821 ///    0x0C: Not-equal (ordered, non-signaling) \n
1822 ///    0x0D: Greater-than-or-equal (ordered, signaling) \n
1823 ///    0x0E: Greater-than (ordered, signaling) \n
1824 ///    0x0F: True (unordered, non-signaling) \n
1825 ///    0x10: Equal (ordered, signaling) \n
1826 ///    0x11: Less-than (ordered, non-signaling) \n
1827 ///    0x12: Less-than-or-equal (ordered, non-signaling) \n
1828 ///    0x13: Unordered (signaling) \n
1829 ///    0x14: Not-equal (unordered, signaling) \n
1830 ///    0x15: Not-less-than (unordered, non-signaling) \n
1831 ///    0x16: Not-less-than-or-equal (unordered, non-signaling) \n
1832 ///    0x17: Ordered (signaling) \n
1833 ///    0x18: Equal (unordered, signaling) \n
1834 ///    0x19: Not-greater-than-or-equal (unordered, non-signaling) \n
1835 ///    0x1A: Not-greater-than (unordered, non-signaling) \n
1836 ///    0x1B: False (ordered, signaling) \n
1837 ///    0x1C: Not-equal (ordered, signaling) \n
1838 ///    0x1D: Greater-than-or-equal (ordered, non-signaling) \n
1839 ///    0x1E: Greater-than (ordered, non-signaling) \n
1840 ///    0x1F: True (unordered, signaling)
1841 /// \returns A 256-bit vector of [8 x float] containing the comparison results.
1842 #define _mm256_cmp_ps(a, b, c) \
1843   ((__m256)__builtin_ia32_cmpps256((__v8sf)(__m256)(a), \
1844                                    (__v8sf)(__m256)(b), (c)))
1845 
1846 /// Compares each of the corresponding scalar double-precision values of
1847 ///    two 128-bit vectors of [2 x double], using the operation specified by the
1848 ///    immediate integer operand.
1849 ///
1850 ///    If the result is true, all 64 bits of the destination vector are set;
1851 ///    otherwise they are cleared.
1852 ///
1853 /// \headerfile <x86intrin.h>
1854 ///
1855 /// \code
1856 /// __m128d _mm_cmp_sd(__m128d a, __m128d b, const int c);
1857 /// \endcode
1858 ///
1859 /// This intrinsic corresponds to the <c> VCMPSD </c> instruction.
1860 ///
1861 /// \param a
1862 ///    A 128-bit vector of [2 x double].
1863 /// \param b
1864 ///    A 128-bit vector of [2 x double].
1865 /// \param c
1866 ///    An immediate integer operand, with bits [4:0] specifying which comparison
1867 ///    operation to use: \n
1868 ///    0x00: Equal (ordered, non-signaling) \n
1869 ///    0x01: Less-than (ordered, signaling) \n
1870 ///    0x02: Less-than-or-equal (ordered, signaling) \n
1871 ///    0x03: Unordered (non-signaling) \n
1872 ///    0x04: Not-equal (unordered, non-signaling) \n
1873 ///    0x05: Not-less-than (unordered, signaling) \n
1874 ///    0x06: Not-less-than-or-equal (unordered, signaling) \n
1875 ///    0x07: Ordered (non-signaling) \n
1876 ///    0x08: Equal (unordered, non-signaling) \n
1877 ///    0x09: Not-greater-than-or-equal (unordered, signaling) \n
1878 ///    0x0A: Not-greater-than (unordered, signaling) \n
1879 ///    0x0B: False (ordered, non-signaling) \n
1880 ///    0x0C: Not-equal (ordered, non-signaling) \n
1881 ///    0x0D: Greater-than-or-equal (ordered, signaling) \n
1882 ///    0x0E: Greater-than (ordered, signaling) \n
1883 ///    0x0F: True (unordered, non-signaling) \n
1884 ///    0x10: Equal (ordered, signaling) \n
1885 ///    0x11: Less-than (ordered, non-signaling) \n
1886 ///    0x12: Less-than-or-equal (ordered, non-signaling) \n
1887 ///    0x13: Unordered (signaling) \n
1888 ///    0x14: Not-equal (unordered, signaling) \n
1889 ///    0x15: Not-less-than (unordered, non-signaling) \n
1890 ///    0x16: Not-less-than-or-equal (unordered, non-signaling) \n
1891 ///    0x17: Ordered (signaling) \n
1892 ///    0x18: Equal (unordered, signaling) \n
1893 ///    0x19: Not-greater-than-or-equal (unordered, non-signaling) \n
1894 ///    0x1A: Not-greater-than (unordered, non-signaling) \n
1895 ///    0x1B: False (ordered, signaling) \n
1896 ///    0x1C: Not-equal (ordered, signaling) \n
1897 ///    0x1D: Greater-than-or-equal (ordered, non-signaling) \n
1898 ///    0x1E: Greater-than (ordered, non-signaling) \n
1899 ///    0x1F: True (unordered, signaling)
1900 /// \returns A 128-bit vector of [2 x double] containing the comparison results.
1901 #define _mm_cmp_sd(a, b, c) \
1902   ((__m128d)__builtin_ia32_cmpsd((__v2df)(__m128d)(a), \
1903                                  (__v2df)(__m128d)(b), (c)))
1904 
1905 /// Compares each of the corresponding scalar values of two 128-bit
1906 ///    vectors of [4 x float], using the operation specified by the immediate
1907 ///    integer operand.
1908 ///
1909 ///    If the result is true, all 32 bits of the destination vector are set;
1910 ///    otherwise they are cleared.
1911 ///
1912 /// \headerfile <x86intrin.h>
1913 ///
1914 /// \code
1915 /// __m128 _mm_cmp_ss(__m128 a, __m128 b, const int c);
1916 /// \endcode
1917 ///
1918 /// This intrinsic corresponds to the <c> VCMPSS </c> instruction.
1919 ///
1920 /// \param a
1921 ///    A 128-bit vector of [4 x float].
1922 /// \param b
1923 ///    A 128-bit vector of [4 x float].
1924 /// \param c
1925 ///    An immediate integer operand, with bits [4:0] specifying which comparison
1926 ///    operation to use: \n
1927 ///    0x00: Equal (ordered, non-signaling) \n
1928 ///    0x01: Less-than (ordered, signaling) \n
1929 ///    0x02: Less-than-or-equal (ordered, signaling) \n
1930 ///    0x03: Unordered (non-signaling) \n
1931 ///    0x04: Not-equal (unordered, non-signaling) \n
1932 ///    0x05: Not-less-than (unordered, signaling) \n
1933 ///    0x06: Not-less-than-or-equal (unordered, signaling) \n
1934 ///    0x07: Ordered (non-signaling) \n
1935 ///    0x08: Equal (unordered, non-signaling) \n
1936 ///    0x09: Not-greater-than-or-equal (unordered, signaling) \n
1937 ///    0x0A: Not-greater-than (unordered, signaling) \n
1938 ///    0x0B: False (ordered, non-signaling) \n
1939 ///    0x0C: Not-equal (ordered, non-signaling) \n
1940 ///    0x0D: Greater-than-or-equal (ordered, signaling) \n
1941 ///    0x0E: Greater-than (ordered, signaling) \n
1942 ///    0x0F: True (unordered, non-signaling) \n
1943 ///    0x10: Equal (ordered, signaling) \n
1944 ///    0x11: Less-than (ordered, non-signaling) \n
1945 ///    0x12: Less-than-or-equal (ordered, non-signaling) \n
1946 ///    0x13: Unordered (signaling) \n
1947 ///    0x14: Not-equal (unordered, signaling) \n
1948 ///    0x15: Not-less-than (unordered, non-signaling) \n
1949 ///    0x16: Not-less-than-or-equal (unordered, non-signaling) \n
1950 ///    0x17: Ordered (signaling) \n
1951 ///    0x18: Equal (unordered, signaling) \n
1952 ///    0x19: Not-greater-than-or-equal (unordered, non-signaling) \n
1953 ///    0x1A: Not-greater-than (unordered, non-signaling) \n
1954 ///    0x1B: False (ordered, signaling) \n
1955 ///    0x1C: Not-equal (ordered, signaling) \n
1956 ///    0x1D: Greater-than-or-equal (ordered, non-signaling) \n
1957 ///    0x1E: Greater-than (ordered, non-signaling) \n
1958 ///    0x1F: True (unordered, signaling)
1959 /// \returns A 128-bit vector of [4 x float] containing the comparison results.
1960 #define _mm_cmp_ss(a, b, c) \
1961   ((__m128)__builtin_ia32_cmpss((__v4sf)(__m128)(a), \
1962                                 (__v4sf)(__m128)(b), (c)))
1963 
1964 /// Takes a [8 x i32] vector and returns the vector element value
1965 ///    indexed by the immediate constant operand.
1966 ///
1967 /// \headerfile <x86intrin.h>
1968 ///
1969 /// \code
1970 /// int _mm256_extract_epi32(__m256i X, const int N);
1971 /// \endcode
1972 ///
1973 /// This intrinsic corresponds to the <c> VEXTRACTF128+COMPOSITE </c>
1974 ///   instruction.
1975 ///
1976 /// \param X
1977 ///    A 256-bit vector of [8 x i32].
1978 /// \param N
1979 ///    An immediate integer operand with bits [2:0] determining which vector
1980 ///    element is extracted and returned.
1981 /// \returns A 32-bit integer containing the extracted 32 bits of extended
1982 ///    packed data.
1983 #define _mm256_extract_epi32(X, N) \
1984   ((int)__builtin_ia32_vec_ext_v8si((__v8si)(__m256i)(X), (int)(N)))
1985 
1986 /// Takes a [16 x i16] vector and returns the vector element value
1987 ///    indexed by the immediate constant operand.
1988 ///
1989 /// \headerfile <x86intrin.h>
1990 ///
1991 /// \code
1992 /// int _mm256_extract_epi16(__m256i X, const int N);
1993 /// \endcode
1994 ///
1995 /// This intrinsic corresponds to the <c> VEXTRACTF128+COMPOSITE </c>
1996 ///   instruction.
1997 ///
1998 /// \param X
1999 ///    A 256-bit integer vector of [16 x i16].
2000 /// \param N
2001 ///    An immediate integer operand with bits [3:0] determining which vector
2002 ///    element is extracted and returned.
2003 /// \returns A 32-bit integer containing the extracted 16 bits of zero extended
2004 ///    packed data.
2005 #define _mm256_extract_epi16(X, N) \
2006   ((int)(unsigned short)__builtin_ia32_vec_ext_v16hi((__v16hi)(__m256i)(X), \
2007                                                      (int)(N)))
2008 
2009 /// Takes a [32 x i8] vector and returns the vector element value
2010 ///    indexed by the immediate constant operand.
2011 ///
2012 /// \headerfile <x86intrin.h>
2013 ///
2014 /// \code
2015 /// int _mm256_extract_epi8(__m256i X, const int N);
2016 /// \endcode
2017 ///
2018 /// This intrinsic corresponds to the <c> VEXTRACTF128+COMPOSITE </c>
2019 ///   instruction.
2020 ///
2021 /// \param X
2022 ///    A 256-bit integer vector of [32 x i8].
2023 /// \param N
2024 ///    An immediate integer operand with bits [4:0] determining which vector
2025 ///    element is extracted and returned.
2026 /// \returns A 32-bit integer containing the extracted 8 bits of zero extended
2027 ///    packed data.
2028 #define _mm256_extract_epi8(X, N) \
2029   ((int)(unsigned char)__builtin_ia32_vec_ext_v32qi((__v32qi)(__m256i)(X), \
2030                                                     (int)(N)))
2031 
2032 #ifdef __x86_64__
2033 /// Takes a [4 x i64] vector and returns the vector element value
2034 ///    indexed by the immediate constant operand.
2035 ///
2036 /// \headerfile <x86intrin.h>
2037 ///
2038 /// \code
2039 /// long long _mm256_extract_epi64(__m256i X, const int N);
2040 /// \endcode
2041 ///
2042 /// This intrinsic corresponds to the <c> VEXTRACTF128+COMPOSITE </c>
2043 ///   instruction.
2044 ///
2045 /// \param X
2046 ///    A 256-bit integer vector of [4 x i64].
2047 /// \param N
2048 ///    An immediate integer operand with bits [1:0] determining which vector
2049 ///    element is extracted and returned.
2050 /// \returns A 64-bit integer containing the extracted 64 bits of extended
2051 ///    packed data.
2052 #define _mm256_extract_epi64(X, N) \
2053   ((long long)__builtin_ia32_vec_ext_v4di((__v4di)(__m256i)(X), (int)(N)))
2054 #endif
2055 
2056 /// Takes a [8 x i32] vector and replaces the vector element value
2057 ///    indexed by the immediate constant operand by a new value. Returns the
2058 ///    modified vector.
2059 ///
2060 /// \headerfile <x86intrin.h>
2061 ///
2062 /// \code
2063 /// __m256i _mm256_insert_epi32(__m256i X, int I, const int N);
2064 /// \endcode
2065 ///
2066 /// This intrinsic corresponds to the <c> VINSERTF128+COMPOSITE </c>
2067 ///   instruction.
2068 ///
2069 /// \param X
2070 ///    A vector of [8 x i32] to be used by the insert operation.
2071 /// \param I
2072 ///    An integer value. The replacement value for the insert operation.
2073 /// \param N
2074 ///    An immediate integer specifying the index of the vector element to be
2075 ///    replaced.
2076 /// \returns A copy of vector \a X, after replacing its element indexed by
2077 ///    \a N with \a I.
2078 #define _mm256_insert_epi32(X, I, N) \
2079   ((__m256i)__builtin_ia32_vec_set_v8si((__v8si)(__m256i)(X), \
2080                                         (int)(I), (int)(N)))
2081 
2082 
2083 /// Takes a [16 x i16] vector and replaces the vector element value
2084 ///    indexed by the immediate constant operand with a new value. Returns the
2085 ///    modified vector.
2086 ///
2087 /// \headerfile <x86intrin.h>
2088 ///
2089 /// \code
2090 /// __m256i _mm256_insert_epi16(__m256i X, int I, const int N);
2091 /// \endcode
2092 ///
2093 /// This intrinsic corresponds to the <c> VINSERTF128+COMPOSITE </c>
2094 ///   instruction.
2095 ///
2096 /// \param X
2097 ///    A vector of [16 x i16] to be used by the insert operation.
2098 /// \param I
2099 ///    An i16 integer value. The replacement value for the insert operation.
2100 /// \param N
2101 ///    An immediate integer specifying the index of the vector element to be
2102 ///    replaced.
2103 /// \returns A copy of vector \a X, after replacing its element indexed by
2104 ///    \a N with \a I.
2105 #define _mm256_insert_epi16(X, I, N) \
2106   ((__m256i)__builtin_ia32_vec_set_v16hi((__v16hi)(__m256i)(X), \
2107                                          (int)(I), (int)(N)))
2108 
2109 /// Takes a [32 x i8] vector and replaces the vector element value
2110 ///    indexed by the immediate constant operand with a new value. Returns the
2111 ///    modified vector.
2112 ///
2113 /// \headerfile <x86intrin.h>
2114 ///
2115 /// \code
2116 /// __m256i _mm256_insert_epi8(__m256i X, int I, const int N);
2117 /// \endcode
2118 ///
2119 /// This intrinsic corresponds to the <c> VINSERTF128+COMPOSITE </c>
2120 ///   instruction.
2121 ///
2122 /// \param X
2123 ///    A vector of [32 x i8] to be used by the insert operation.
2124 /// \param I
2125 ///    An i8 integer value. The replacement value for the insert operation.
2126 /// \param N
2127 ///    An immediate integer specifying the index of the vector element to be
2128 ///    replaced.
2129 /// \returns A copy of vector \a X, after replacing its element indexed by
2130 ///    \a N with \a I.
2131 #define _mm256_insert_epi8(X, I, N) \
2132   ((__m256i)__builtin_ia32_vec_set_v32qi((__v32qi)(__m256i)(X), \
2133                                          (int)(I), (int)(N)))
2134 
2135 #ifdef __x86_64__
2136 /// Takes a [4 x i64] vector and replaces the vector element value
2137 ///    indexed by the immediate constant operand with a new value. Returns the
2138 ///    modified vector.
2139 ///
2140 /// \headerfile <x86intrin.h>
2141 ///
2142 /// \code
2143 /// __m256i _mm256_insert_epi64(__m256i X, int I, const int N);
2144 /// \endcode
2145 ///
2146 /// This intrinsic corresponds to the <c> VINSERTF128+COMPOSITE </c>
2147 ///   instruction.
2148 ///
2149 /// \param X
2150 ///    A vector of [4 x i64] to be used by the insert operation.
2151 /// \param I
2152 ///    A 64-bit integer value. The replacement value for the insert operation.
2153 /// \param N
2154 ///    An immediate integer specifying the index of the vector element to be
2155 ///    replaced.
2156 /// \returns A copy of vector \a X, after replacing its element indexed by
2157 ///     \a N with \a I.
2158 #define _mm256_insert_epi64(X, I, N) \
2159   ((__m256i)__builtin_ia32_vec_set_v4di((__v4di)(__m256i)(X), \
2160                                         (long long)(I), (int)(N)))
2161 #endif
2162 
2163 /* Conversion */
2164 /// Converts a vector of [4 x i32] into a vector of [4 x double].
2165 ///
2166 /// \headerfile <x86intrin.h>
2167 ///
2168 /// This intrinsic corresponds to the <c> VCVTDQ2PD </c> instruction.
2169 ///
2170 /// \param __a
2171 ///    A 128-bit integer vector of [4 x i32].
2172 /// \returns A 256-bit vector of [4 x double] containing the converted values.
2173 static __inline __m256d __DEFAULT_FN_ATTRS
2174 _mm256_cvtepi32_pd(__m128i __a)
2175 {
2176   return (__m256d)__builtin_convertvector((__v4si)__a, __v4df);
2177 }
2178 
2179 /// Converts a vector of [8 x i32] into a vector of [8 x float].
2180 ///
2181 /// \headerfile <x86intrin.h>
2182 ///
2183 /// This intrinsic corresponds to the <c> VCVTDQ2PS </c> instruction.
2184 ///
2185 /// \param __a
2186 ///    A 256-bit integer vector.
2187 /// \returns A 256-bit vector of [8 x float] containing the converted values.
2188 static __inline __m256 __DEFAULT_FN_ATTRS
2189 _mm256_cvtepi32_ps(__m256i __a)
2190 {
2191   return (__m256)__builtin_convertvector((__v8si)__a, __v8sf);
2192 }
2193 
2194 /// Converts a 256-bit vector of [4 x double] into a 128-bit vector of
2195 ///    [4 x float].
2196 ///
2197 /// \headerfile <x86intrin.h>
2198 ///
2199 /// This intrinsic corresponds to the <c> VCVTPD2PS </c> instruction.
2200 ///
2201 /// \param __a
2202 ///    A 256-bit vector of [4 x double].
2203 /// \returns A 128-bit vector of [4 x float] containing the converted values.
2204 static __inline __m128 __DEFAULT_FN_ATTRS
2205 _mm256_cvtpd_ps(__m256d __a)
2206 {
2207   return (__m128)__builtin_ia32_cvtpd2ps256((__v4df) __a);
2208 }
2209 
2210 /// Converts a vector of [8 x float] into a vector of [8 x i32].
2211 ///
2212 /// \headerfile <x86intrin.h>
2213 ///
2214 /// This intrinsic corresponds to the <c> VCVTPS2DQ </c> instruction.
2215 ///
2216 /// \param __a
2217 ///    A 256-bit vector of [8 x float].
2218 /// \returns A 256-bit integer vector containing the converted values.
2219 static __inline __m256i __DEFAULT_FN_ATTRS
2220 _mm256_cvtps_epi32(__m256 __a)
2221 {
2222   return (__m256i)__builtin_ia32_cvtps2dq256((__v8sf) __a);
2223 }
2224 
2225 /// Converts a 128-bit vector of [4 x float] into a 256-bit vector of [4
2226 ///    x double].
2227 ///
2228 /// \headerfile <x86intrin.h>
2229 ///
2230 /// This intrinsic corresponds to the <c> VCVTPS2PD </c> instruction.
2231 ///
2232 /// \param __a
2233 ///    A 128-bit vector of [4 x float].
2234 /// \returns A 256-bit vector of [4 x double] containing the converted values.
2235 static __inline __m256d __DEFAULT_FN_ATTRS
2236 _mm256_cvtps_pd(__m128 __a)
2237 {
2238   return (__m256d)__builtin_convertvector((__v4sf)__a, __v4df);
2239 }
2240 
2241 /// Converts a 256-bit vector of [4 x double] into a 128-bit vector of [4
2242 ///    x i32], truncating the result by rounding towards zero when it is
2243 ///    inexact.
2244 ///
2245 /// \headerfile <x86intrin.h>
2246 ///
2247 /// This intrinsic corresponds to the <c> VCVTTPD2DQ </c> instruction.
2248 ///
2249 /// \param __a
2250 ///    A 256-bit vector of [4 x double].
2251 /// \returns A 128-bit integer vector containing the converted values.
2252 static __inline __m128i __DEFAULT_FN_ATTRS
2253 _mm256_cvttpd_epi32(__m256d __a)
2254 {
2255   return (__m128i)__builtin_ia32_cvttpd2dq256((__v4df) __a);
2256 }
2257 
2258 /// Converts a 256-bit vector of [4 x double] into a 128-bit vector of [4
2259 ///    x i32]. When a conversion is inexact, the value returned is rounded
2260 ///    according to the rounding control bits in the MXCSR register.
2261 ///
2262 /// \headerfile <x86intrin.h>
2263 ///
2264 /// This intrinsic corresponds to the <c> VCVTPD2DQ </c> instruction.
2265 ///
2266 /// \param __a
2267 ///    A 256-bit vector of [4 x double].
2268 /// \returns A 128-bit integer vector containing the converted values.
2269 static __inline __m128i __DEFAULT_FN_ATTRS
2270 _mm256_cvtpd_epi32(__m256d __a)
2271 {
2272   return (__m128i)__builtin_ia32_cvtpd2dq256((__v4df) __a);
2273 }
2274 
2275 /// Converts a vector of [8 x float] into a vector of [8 x i32],
2276 ///    truncating the result by rounding towards zero when it is inexact.
2277 ///
2278 /// \headerfile <x86intrin.h>
2279 ///
2280 /// This intrinsic corresponds to the <c> VCVTTPS2DQ </c> instruction.
2281 ///
2282 /// \param __a
2283 ///    A 256-bit vector of [8 x float].
2284 /// \returns A 256-bit integer vector containing the converted values.
2285 static __inline __m256i __DEFAULT_FN_ATTRS
2286 _mm256_cvttps_epi32(__m256 __a)
2287 {
2288   return (__m256i)__builtin_ia32_cvttps2dq256((__v8sf) __a);
2289 }
2290 
2291 /// Returns the first element of the input vector of [4 x double].
2292 ///
2293 /// \headerfile <x86intrin.h>
2294 ///
2295 /// This intrinsic is a utility function and does not correspond to a specific
2296 ///    instruction.
2297 ///
2298 /// \param __a
2299 ///    A 256-bit vector of [4 x double].
2300 /// \returns A 64 bit double containing the first element of the input vector.
2301 static __inline double __DEFAULT_FN_ATTRS
2302 _mm256_cvtsd_f64(__m256d __a)
2303 {
2304  return __a[0];
2305 }
2306 
2307 /// Returns the first element of the input vector of [8 x i32].
2308 ///
2309 /// \headerfile <x86intrin.h>
2310 ///
2311 /// This intrinsic is a utility function and does not correspond to a specific
2312 ///    instruction.
2313 ///
2314 /// \param __a
2315 ///    A 256-bit vector of [8 x i32].
2316 /// \returns A 32 bit integer containing the first element of the input vector.
2317 static __inline int __DEFAULT_FN_ATTRS
2318 _mm256_cvtsi256_si32(__m256i __a)
2319 {
2320  __v8si __b = (__v8si)__a;
2321  return __b[0];
2322 }
2323 
2324 /// Returns the first element of the input vector of [8 x float].
2325 ///
2326 /// \headerfile <x86intrin.h>
2327 ///
2328 /// This intrinsic is a utility function and does not correspond to a specific
2329 ///    instruction.
2330 ///
2331 /// \param __a
2332 ///    A 256-bit vector of [8 x float].
2333 /// \returns A 32 bit float containing the first element of the input vector.
2334 static __inline float __DEFAULT_FN_ATTRS
2335 _mm256_cvtss_f32(__m256 __a)
2336 {
2337  return __a[0];
2338 }
2339 
2340 /* Vector replicate */
2341 /// Moves and duplicates odd-indexed values from a 256-bit vector of
2342 ///    [8 x float] to float values in a 256-bit vector of [8 x float].
2343 ///
2344 /// \headerfile <x86intrin.h>
2345 ///
2346 /// This intrinsic corresponds to the <c> VMOVSHDUP </c> instruction.
2347 ///
2348 /// \param __a
2349 ///    A 256-bit vector of [8 x float]. \n
2350 ///    Bits [255:224] of \a __a are written to bits [255:224] and [223:192] of
2351 ///    the return value. \n
2352 ///    Bits [191:160] of \a __a are written to bits [191:160] and [159:128] of
2353 ///    the return value. \n
2354 ///    Bits [127:96] of \a __a are written to bits [127:96] and [95:64] of the
2355 ///    return value. \n
2356 ///    Bits [63:32] of \a __a are written to bits [63:32] and [31:0] of the
2357 ///    return value.
2358 /// \returns A 256-bit vector of [8 x float] containing the moved and duplicated
2359 ///    values.
2360 static __inline __m256 __DEFAULT_FN_ATTRS
2361 _mm256_movehdup_ps(__m256 __a)
2362 {
2363   return __builtin_shufflevector((__v8sf)__a, (__v8sf)__a, 1, 1, 3, 3, 5, 5, 7, 7);
2364 }
2365 
2366 /// Moves and duplicates even-indexed values from a 256-bit vector of
2367 ///    [8 x float] to float values in a 256-bit vector of [8 x float].
2368 ///
2369 /// \headerfile <x86intrin.h>
2370 ///
2371 /// This intrinsic corresponds to the <c> VMOVSLDUP </c> instruction.
2372 ///
2373 /// \param __a
2374 ///    A 256-bit vector of [8 x float]. \n
2375 ///    Bits [223:192] of \a __a are written to bits [255:224] and [223:192] of
2376 ///    the return value. \n
2377 ///    Bits [159:128] of \a __a are written to bits [191:160] and [159:128] of
2378 ///    the return value. \n
2379 ///    Bits [95:64] of \a __a are written to bits [127:96] and [95:64] of the
2380 ///    return value. \n
2381 ///    Bits [31:0] of \a __a are written to bits [63:32] and [31:0] of the
2382 ///    return value.
2383 /// \returns A 256-bit vector of [8 x float] containing the moved and duplicated
2384 ///    values.
2385 static __inline __m256 __DEFAULT_FN_ATTRS
2386 _mm256_moveldup_ps(__m256 __a)
2387 {
2388   return __builtin_shufflevector((__v8sf)__a, (__v8sf)__a, 0, 0, 2, 2, 4, 4, 6, 6);
2389 }
2390 
2391 /// Moves and duplicates double-precision floating point values from a
2392 ///    256-bit vector of [4 x double] to double-precision values in a 256-bit
2393 ///    vector of [4 x double].
2394 ///
2395 /// \headerfile <x86intrin.h>
2396 ///
2397 /// This intrinsic corresponds to the <c> VMOVDDUP </c> instruction.
2398 ///
2399 /// \param __a
2400 ///    A 256-bit vector of [4 x double]. \n
2401 ///    Bits [63:0] of \a __a are written to bits [127:64] and [63:0] of the
2402 ///    return value. \n
2403 ///    Bits [191:128] of \a __a are written to bits [255:192] and [191:128] of
2404 ///    the return value.
2405 /// \returns A 256-bit vector of [4 x double] containing the moved and
2406 ///    duplicated values.
2407 static __inline __m256d __DEFAULT_FN_ATTRS
2408 _mm256_movedup_pd(__m256d __a)
2409 {
2410   return __builtin_shufflevector((__v4df)__a, (__v4df)__a, 0, 0, 2, 2);
2411 }
2412 
2413 /* Unpack and Interleave */
2414 /// Unpacks the odd-indexed vector elements from two 256-bit vectors of
2415 ///    [4 x double] and interleaves them into a 256-bit vector of [4 x double].
2416 ///
2417 /// \headerfile <x86intrin.h>
2418 ///
2419 /// This intrinsic corresponds to the <c> VUNPCKHPD </c> instruction.
2420 ///
2421 /// \param __a
2422 ///    A 256-bit floating-point vector of [4 x double]. \n
2423 ///    Bits [127:64] are written to bits [63:0] of the return value. \n
2424 ///    Bits [255:192] are written to bits [191:128] of the return value. \n
2425 /// \param __b
2426 ///    A 256-bit floating-point vector of [4 x double]. \n
2427 ///    Bits [127:64] are written to bits [127:64] of the return value. \n
2428 ///    Bits [255:192] are written to bits [255:192] of the return value. \n
2429 /// \returns A 256-bit vector of [4 x double] containing the interleaved values.
2430 static __inline __m256d __DEFAULT_FN_ATTRS
2431 _mm256_unpackhi_pd(__m256d __a, __m256d __b)
2432 {
2433   return __builtin_shufflevector((__v4df)__a, (__v4df)__b, 1, 5, 1+2, 5+2);
2434 }
2435 
2436 /// Unpacks the even-indexed vector elements from two 256-bit vectors of
2437 ///    [4 x double] and interleaves them into a 256-bit vector of [4 x double].
2438 ///
2439 /// \headerfile <x86intrin.h>
2440 ///
2441 /// This intrinsic corresponds to the <c> VUNPCKLPD </c> instruction.
2442 ///
2443 /// \param __a
2444 ///    A 256-bit floating-point vector of [4 x double]. \n
2445 ///    Bits [63:0] are written to bits [63:0] of the return value. \n
2446 ///    Bits [191:128] are written to bits [191:128] of the return value.
2447 /// \param __b
2448 ///    A 256-bit floating-point vector of [4 x double]. \n
2449 ///    Bits [63:0] are written to bits [127:64] of the return value. \n
2450 ///    Bits [191:128] are written to bits [255:192] of the return value. \n
2451 /// \returns A 256-bit vector of [4 x double] containing the interleaved values.
2452 static __inline __m256d __DEFAULT_FN_ATTRS
2453 _mm256_unpacklo_pd(__m256d __a, __m256d __b)
2454 {
2455   return __builtin_shufflevector((__v4df)__a, (__v4df)__b, 0, 4, 0+2, 4+2);
2456 }
2457 
2458 /// Unpacks the 32-bit vector elements 2, 3, 6 and 7 from each of the
2459 ///    two 256-bit vectors of [8 x float] and interleaves them into a 256-bit
2460 ///    vector of [8 x float].
2461 ///
2462 /// \headerfile <x86intrin.h>
2463 ///
2464 /// This intrinsic corresponds to the <c> VUNPCKHPS </c> instruction.
2465 ///
2466 /// \param __a
2467 ///    A 256-bit vector of [8 x float]. \n
2468 ///    Bits [95:64] are written to bits [31:0] of the return value. \n
2469 ///    Bits [127:96] are written to bits [95:64] of the return value. \n
2470 ///    Bits [223:192] are written to bits [159:128] of the return value. \n
2471 ///    Bits [255:224] are written to bits [223:192] of the return value.
2472 /// \param __b
2473 ///    A 256-bit vector of [8 x float]. \n
2474 ///    Bits [95:64] are written to bits [63:32] of the return value. \n
2475 ///    Bits [127:96] are written to bits [127:96] of the return value. \n
2476 ///    Bits [223:192] are written to bits [191:160] of the return value. \n
2477 ///    Bits [255:224] are written to bits [255:224] of the return value.
2478 /// \returns A 256-bit vector of [8 x float] containing the interleaved values.
2479 static __inline __m256 __DEFAULT_FN_ATTRS
2480 _mm256_unpackhi_ps(__m256 __a, __m256 __b)
2481 {
2482   return __builtin_shufflevector((__v8sf)__a, (__v8sf)__b, 2, 10, 2+1, 10+1, 6, 14, 6+1, 14+1);
2483 }
2484 
2485 /// Unpacks the 32-bit vector elements 0, 1, 4 and 5 from each of the
2486 ///    two 256-bit vectors of [8 x float] and interleaves them into a 256-bit
2487 ///    vector of [8 x float].
2488 ///
2489 /// \headerfile <x86intrin.h>
2490 ///
2491 /// This intrinsic corresponds to the <c> VUNPCKLPS </c> instruction.
2492 ///
2493 /// \param __a
2494 ///    A 256-bit vector of [8 x float]. \n
2495 ///    Bits [31:0] are written to bits [31:0] of the return value. \n
2496 ///    Bits [63:32] are written to bits [95:64] of the return value. \n
2497 ///    Bits [159:128] are written to bits [159:128] of the return value. \n
2498 ///    Bits [191:160] are written to bits [223:192] of the return value.
2499 /// \param __b
2500 ///    A 256-bit vector of [8 x float]. \n
2501 ///    Bits [31:0] are written to bits [63:32] of the return value. \n
2502 ///    Bits [63:32] are written to bits [127:96] of the return value. \n
2503 ///    Bits [159:128] are written to bits [191:160] of the return value. \n
2504 ///    Bits [191:160] are written to bits [255:224] of the return value.
2505 /// \returns A 256-bit vector of [8 x float] containing the interleaved values.
2506 static __inline __m256 __DEFAULT_FN_ATTRS
2507 _mm256_unpacklo_ps(__m256 __a, __m256 __b)
2508 {
2509   return __builtin_shufflevector((__v8sf)__a, (__v8sf)__b, 0, 8, 0+1, 8+1, 4, 12, 4+1, 12+1);
2510 }
2511 
2512 /* Bit Test */
2513 /// Given two 128-bit floating-point vectors of [2 x double], perform an
2514 ///    element-by-element comparison of the double-precision element in the
2515 ///    first source vector and the corresponding element in the second source
2516 ///    vector.
2517 ///
2518 ///    The EFLAGS register is updated as follows: \n
2519 ///    If there is at least one pair of double-precision elements where the
2520 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2521 ///    ZF flag is set to 1. \n
2522 ///    If there is at least one pair of double-precision elements where the
2523 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2524 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2525 ///    This intrinsic returns the value of the ZF flag.
2526 ///
2527 /// \headerfile <x86intrin.h>
2528 ///
2529 /// This intrinsic corresponds to the <c> VTESTPD </c> instruction.
2530 ///
2531 /// \param __a
2532 ///    A 128-bit vector of [2 x double].
2533 /// \param __b
2534 ///    A 128-bit vector of [2 x double].
2535 /// \returns the ZF flag in the EFLAGS register.
2536 static __inline int __DEFAULT_FN_ATTRS128
2537 _mm_testz_pd(__m128d __a, __m128d __b)
2538 {
2539   return __builtin_ia32_vtestzpd((__v2df)__a, (__v2df)__b);
2540 }
2541 
2542 /// Given two 128-bit floating-point vectors of [2 x double], perform an
2543 ///    element-by-element comparison of the double-precision element in the
2544 ///    first source vector and the corresponding element in the second source
2545 ///    vector.
2546 ///
2547 ///    The EFLAGS register is updated as follows: \n
2548 ///    If there is at least one pair of double-precision elements where the
2549 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2550 ///    ZF flag is set to 1. \n
2551 ///    If there is at least one pair of double-precision elements where the
2552 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2553 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2554 ///    This intrinsic returns the value of the CF flag.
2555 ///
2556 /// \headerfile <x86intrin.h>
2557 ///
2558 /// This intrinsic corresponds to the <c> VTESTPD </c> instruction.
2559 ///
2560 /// \param __a
2561 ///    A 128-bit vector of [2 x double].
2562 /// \param __b
2563 ///    A 128-bit vector of [2 x double].
2564 /// \returns the CF flag in the EFLAGS register.
2565 static __inline int __DEFAULT_FN_ATTRS128
2566 _mm_testc_pd(__m128d __a, __m128d __b)
2567 {
2568   return __builtin_ia32_vtestcpd((__v2df)__a, (__v2df)__b);
2569 }
2570 
2571 /// Given two 128-bit floating-point vectors of [2 x double], perform an
2572 ///    element-by-element comparison of the double-precision element in the
2573 ///    first source vector and the corresponding element in the second source
2574 ///    vector.
2575 ///
2576 ///    The EFLAGS register is updated as follows: \n
2577 ///    If there is at least one pair of double-precision elements where the
2578 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2579 ///    ZF flag is set to 1. \n
2580 ///    If there is at least one pair of double-precision elements where the
2581 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2582 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2583 ///    This intrinsic returns 1 if both the ZF and CF flags are set to 0,
2584 ///    otherwise it returns 0.
2585 ///
2586 /// \headerfile <x86intrin.h>
2587 ///
2588 /// This intrinsic corresponds to the <c> VTESTPD </c> instruction.
2589 ///
2590 /// \param __a
2591 ///    A 128-bit vector of [2 x double].
2592 /// \param __b
2593 ///    A 128-bit vector of [2 x double].
2594 /// \returns 1 if both the ZF and CF flags are set to 0, otherwise returns 0.
2595 static __inline int __DEFAULT_FN_ATTRS128
2596 _mm_testnzc_pd(__m128d __a, __m128d __b)
2597 {
2598   return __builtin_ia32_vtestnzcpd((__v2df)__a, (__v2df)__b);
2599 }
2600 
2601 /// Given two 128-bit floating-point vectors of [4 x float], perform an
2602 ///    element-by-element comparison of the single-precision element in the
2603 ///    first source vector and the corresponding element in the second source
2604 ///    vector.
2605 ///
2606 ///    The EFLAGS register is updated as follows: \n
2607 ///    If there is at least one pair of single-precision elements where the
2608 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2609 ///    ZF flag is set to 1. \n
2610 ///    If there is at least one pair of single-precision elements where the
2611 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2612 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2613 ///    This intrinsic returns the value of the ZF flag.
2614 ///
2615 /// \headerfile <x86intrin.h>
2616 ///
2617 /// This intrinsic corresponds to the <c> VTESTPS </c> instruction.
2618 ///
2619 /// \param __a
2620 ///    A 128-bit vector of [4 x float].
2621 /// \param __b
2622 ///    A 128-bit vector of [4 x float].
2623 /// \returns the ZF flag.
2624 static __inline int __DEFAULT_FN_ATTRS128
2625 _mm_testz_ps(__m128 __a, __m128 __b)
2626 {
2627   return __builtin_ia32_vtestzps((__v4sf)__a, (__v4sf)__b);
2628 }
2629 
2630 /// Given two 128-bit floating-point vectors of [4 x float], perform an
2631 ///    element-by-element comparison of the single-precision element in the
2632 ///    first source vector and the corresponding element in the second source
2633 ///    vector.
2634 ///
2635 ///    The EFLAGS register is updated as follows: \n
2636 ///    If there is at least one pair of single-precision elements where the
2637 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2638 ///    ZF flag is set to 1. \n
2639 ///    If there is at least one pair of single-precision elements where the
2640 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2641 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2642 ///    This intrinsic returns the value of the CF flag.
2643 ///
2644 /// \headerfile <x86intrin.h>
2645 ///
2646 /// This intrinsic corresponds to the <c> VTESTPS </c> instruction.
2647 ///
2648 /// \param __a
2649 ///    A 128-bit vector of [4 x float].
2650 /// \param __b
2651 ///    A 128-bit vector of [4 x float].
2652 /// \returns the CF flag.
2653 static __inline int __DEFAULT_FN_ATTRS128
2654 _mm_testc_ps(__m128 __a, __m128 __b)
2655 {
2656   return __builtin_ia32_vtestcps((__v4sf)__a, (__v4sf)__b);
2657 }
2658 
2659 /// Given two 128-bit floating-point vectors of [4 x float], perform an
2660 ///    element-by-element comparison of the single-precision element in the
2661 ///    first source vector and the corresponding element in the second source
2662 ///    vector.
2663 ///
2664 ///    The EFLAGS register is updated as follows: \n
2665 ///    If there is at least one pair of single-precision elements where the
2666 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2667 ///    ZF flag is set to 1. \n
2668 ///    If there is at least one pair of single-precision elements where the
2669 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2670 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2671 ///    This intrinsic returns 1 if both the ZF and CF flags are set to 0,
2672 ///    otherwise it returns 0.
2673 ///
2674 /// \headerfile <x86intrin.h>
2675 ///
2676 /// This intrinsic corresponds to the <c> VTESTPS </c> instruction.
2677 ///
2678 /// \param __a
2679 ///    A 128-bit vector of [4 x float].
2680 /// \param __b
2681 ///    A 128-bit vector of [4 x float].
2682 /// \returns 1 if both the ZF and CF flags are set to 0, otherwise returns 0.
2683 static __inline int __DEFAULT_FN_ATTRS128
2684 _mm_testnzc_ps(__m128 __a, __m128 __b)
2685 {
2686   return __builtin_ia32_vtestnzcps((__v4sf)__a, (__v4sf)__b);
2687 }
2688 
2689 /// Given two 256-bit floating-point vectors of [4 x double], perform an
2690 ///    element-by-element comparison of the double-precision elements in the
2691 ///    first source vector and the corresponding elements in the second source
2692 ///    vector.
2693 ///
2694 ///    The EFLAGS register is updated as follows: \n
2695 ///    If there is at least one pair of double-precision elements where the
2696 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2697 ///    ZF flag is set to 1. \n
2698 ///    If there is at least one pair of double-precision elements where the
2699 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2700 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2701 ///    This intrinsic returns the value of the ZF flag.
2702 ///
2703 /// \headerfile <x86intrin.h>
2704 ///
2705 /// This intrinsic corresponds to the <c> VTESTPD </c> instruction.
2706 ///
2707 /// \param __a
2708 ///    A 256-bit vector of [4 x double].
2709 /// \param __b
2710 ///    A 256-bit vector of [4 x double].
2711 /// \returns the ZF flag.
2712 static __inline int __DEFAULT_FN_ATTRS
2713 _mm256_testz_pd(__m256d __a, __m256d __b)
2714 {
2715   return __builtin_ia32_vtestzpd256((__v4df)__a, (__v4df)__b);
2716 }
2717 
2718 /// Given two 256-bit floating-point vectors of [4 x double], perform an
2719 ///    element-by-element comparison of the double-precision elements in the
2720 ///    first source vector and the corresponding elements in the second source
2721 ///    vector.
2722 ///
2723 ///    The EFLAGS register is updated as follows: \n
2724 ///    If there is at least one pair of double-precision elements where the
2725 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2726 ///    ZF flag is set to 1. \n
2727 ///    If there is at least one pair of double-precision elements where the
2728 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2729 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2730 ///    This intrinsic returns the value of the CF flag.
2731 ///
2732 /// \headerfile <x86intrin.h>
2733 ///
2734 /// This intrinsic corresponds to the <c> VTESTPD </c> instruction.
2735 ///
2736 /// \param __a
2737 ///    A 256-bit vector of [4 x double].
2738 /// \param __b
2739 ///    A 256-bit vector of [4 x double].
2740 /// \returns the CF flag.
2741 static __inline int __DEFAULT_FN_ATTRS
2742 _mm256_testc_pd(__m256d __a, __m256d __b)
2743 {
2744   return __builtin_ia32_vtestcpd256((__v4df)__a, (__v4df)__b);
2745 }
2746 
2747 /// Given two 256-bit floating-point vectors of [4 x double], perform an
2748 ///    element-by-element comparison of the double-precision elements in the
2749 ///    first source vector and the corresponding elements in the second source
2750 ///    vector.
2751 ///
2752 ///    The EFLAGS register is updated as follows: \n
2753 ///    If there is at least one pair of double-precision elements where the
2754 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2755 ///    ZF flag is set to 1. \n
2756 ///    If there is at least one pair of double-precision elements where the
2757 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2758 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2759 ///    This intrinsic returns 1 if both the ZF and CF flags are set to 0,
2760 ///    otherwise it returns 0.
2761 ///
2762 /// \headerfile <x86intrin.h>
2763 ///
2764 /// This intrinsic corresponds to the <c> VTESTPD </c> instruction.
2765 ///
2766 /// \param __a
2767 ///    A 256-bit vector of [4 x double].
2768 /// \param __b
2769 ///    A 256-bit vector of [4 x double].
2770 /// \returns 1 if both the ZF and CF flags are set to 0, otherwise returns 0.
2771 static __inline int __DEFAULT_FN_ATTRS
2772 _mm256_testnzc_pd(__m256d __a, __m256d __b)
2773 {
2774   return __builtin_ia32_vtestnzcpd256((__v4df)__a, (__v4df)__b);
2775 }
2776 
2777 /// Given two 256-bit floating-point vectors of [8 x float], perform an
2778 ///    element-by-element comparison of the single-precision element in the
2779 ///    first source vector and the corresponding element in the second source
2780 ///    vector.
2781 ///
2782 ///    The EFLAGS register is updated as follows: \n
2783 ///    If there is at least one pair of single-precision elements where the
2784 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2785 ///    ZF flag is set to 1. \n
2786 ///    If there is at least one pair of single-precision elements where the
2787 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2788 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2789 ///    This intrinsic returns the value of the ZF flag.
2790 ///
2791 /// \headerfile <x86intrin.h>
2792 ///
2793 /// This intrinsic corresponds to the <c> VTESTPS </c> instruction.
2794 ///
2795 /// \param __a
2796 ///    A 256-bit vector of [8 x float].
2797 /// \param __b
2798 ///    A 256-bit vector of [8 x float].
2799 /// \returns the ZF flag.
2800 static __inline int __DEFAULT_FN_ATTRS
2801 _mm256_testz_ps(__m256 __a, __m256 __b)
2802 {
2803   return __builtin_ia32_vtestzps256((__v8sf)__a, (__v8sf)__b);
2804 }
2805 
2806 /// Given two 256-bit floating-point vectors of [8 x float], perform an
2807 ///    element-by-element comparison of the single-precision element in the
2808 ///    first source vector and the corresponding element in the second source
2809 ///    vector.
2810 ///
2811 ///    The EFLAGS register is updated as follows: \n
2812 ///    If there is at least one pair of single-precision elements where the
2813 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2814 ///    ZF flag is set to 1. \n
2815 ///    If there is at least one pair of single-precision elements where the
2816 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2817 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2818 ///    This intrinsic returns the value of the CF flag.
2819 ///
2820 /// \headerfile <x86intrin.h>
2821 ///
2822 /// This intrinsic corresponds to the <c> VTESTPS </c> instruction.
2823 ///
2824 /// \param __a
2825 ///    A 256-bit vector of [8 x float].
2826 /// \param __b
2827 ///    A 256-bit vector of [8 x float].
2828 /// \returns the CF flag.
2829 static __inline int __DEFAULT_FN_ATTRS
2830 _mm256_testc_ps(__m256 __a, __m256 __b)
2831 {
2832   return __builtin_ia32_vtestcps256((__v8sf)__a, (__v8sf)__b);
2833 }
2834 
2835 /// Given two 256-bit floating-point vectors of [8 x float], perform an
2836 ///    element-by-element comparison of the single-precision elements in the
2837 ///    first source vector and the corresponding elements in the second source
2838 ///    vector.
2839 ///
2840 ///    The EFLAGS register is updated as follows: \n
2841 ///    If there is at least one pair of single-precision elements where the
2842 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2843 ///    ZF flag is set to 1. \n
2844 ///    If there is at least one pair of single-precision elements where the
2845 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2846 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2847 ///    This intrinsic returns 1 if both the ZF and CF flags are set to 0,
2848 ///    otherwise it returns 0.
2849 ///
2850 /// \headerfile <x86intrin.h>
2851 ///
2852 /// This intrinsic corresponds to the <c> VTESTPS </c> instruction.
2853 ///
2854 /// \param __a
2855 ///    A 256-bit vector of [8 x float].
2856 /// \param __b
2857 ///    A 256-bit vector of [8 x float].
2858 /// \returns 1 if both the ZF and CF flags are set to 0, otherwise returns 0.
2859 static __inline int __DEFAULT_FN_ATTRS
2860 _mm256_testnzc_ps(__m256 __a, __m256 __b)
2861 {
2862   return __builtin_ia32_vtestnzcps256((__v8sf)__a, (__v8sf)__b);
2863 }
2864 
2865 /// Given two 256-bit integer vectors, perform a bit-by-bit comparison
2866 ///    of the two source vectors.
2867 ///
2868 ///    The EFLAGS register is updated as follows: \n
2869 ///    If there is at least one pair of bits where both bits are 1, the ZF flag
2870 ///    is set to 0. Otherwise the ZF flag is set to 1. \n
2871 ///    If there is at least one pair of bits where the bit from the first source
2872 ///    vector is 0 and the bit from the second source vector is 1, the CF flag
2873 ///    is set to 0. Otherwise the CF flag is set to 1. \n
2874 ///    This intrinsic returns the value of the ZF flag.
2875 ///
2876 /// \headerfile <x86intrin.h>
2877 ///
2878 /// This intrinsic corresponds to the <c> VPTEST </c> instruction.
2879 ///
2880 /// \param __a
2881 ///    A 256-bit integer vector.
2882 /// \param __b
2883 ///    A 256-bit integer vector.
2884 /// \returns the ZF flag.
2885 static __inline int __DEFAULT_FN_ATTRS
2886 _mm256_testz_si256(__m256i __a, __m256i __b)
2887 {
2888   return __builtin_ia32_ptestz256((__v4di)__a, (__v4di)__b);
2889 }
2890 
2891 /// Given two 256-bit integer vectors, perform a bit-by-bit comparison
2892 ///    of the two source vectors.
2893 ///
2894 ///    The EFLAGS register is updated as follows: \n
2895 ///    If there is at least one pair of bits where both bits are 1, the ZF flag
2896 ///    is set to 0. Otherwise the ZF flag is set to 1. \n
2897 ///    If there is at least one pair of bits where the bit from the first source
2898 ///    vector is 0 and the bit from the second source vector is 1, the CF flag
2899 ///    is set to 0. Otherwise the CF flag is set to 1. \n
2900 ///    This intrinsic returns the value of the CF flag.
2901 ///
2902 /// \headerfile <x86intrin.h>
2903 ///
2904 /// This intrinsic corresponds to the <c> VPTEST </c> instruction.
2905 ///
2906 /// \param __a
2907 ///    A 256-bit integer vector.
2908 /// \param __b
2909 ///    A 256-bit integer vector.
2910 /// \returns the CF flag.
2911 static __inline int __DEFAULT_FN_ATTRS
2912 _mm256_testc_si256(__m256i __a, __m256i __b)
2913 {
2914   return __builtin_ia32_ptestc256((__v4di)__a, (__v4di)__b);
2915 }
2916 
2917 /// Given two 256-bit integer vectors, perform a bit-by-bit comparison
2918 ///    of the two source vectors.
2919 ///
2920 ///    The EFLAGS register is updated as follows: \n
2921 ///    If there is at least one pair of bits where both bits are 1, the ZF flag
2922 ///    is set to 0. Otherwise the ZF flag is set to 1. \n
2923 ///    If there is at least one pair of bits where the bit from the first source
2924 ///    vector is 0 and the bit from the second source vector is 1, the CF flag
2925 ///    is set to 0. Otherwise the CF flag is set to 1. \n
2926 ///    This intrinsic returns 1 if both the ZF and CF flags are set to 0,
2927 ///    otherwise it returns 0.
2928 ///
2929 /// \headerfile <x86intrin.h>
2930 ///
2931 /// This intrinsic corresponds to the <c> VPTEST </c> instruction.
2932 ///
2933 /// \param __a
2934 ///    A 256-bit integer vector.
2935 /// \param __b
2936 ///    A 256-bit integer vector.
2937 /// \returns 1 if both the ZF and CF flags are set to 0, otherwise returns 0.
2938 static __inline int __DEFAULT_FN_ATTRS
2939 _mm256_testnzc_si256(__m256i __a, __m256i __b)
2940 {
2941   return __builtin_ia32_ptestnzc256((__v4di)__a, (__v4di)__b);
2942 }
2943 
2944 /* Vector extract sign mask */
2945 /// Extracts the sign bits of double-precision floating point elements
2946 ///    in a 256-bit vector of [4 x double] and writes them to the lower order
2947 ///    bits of the return value.
2948 ///
2949 /// \headerfile <x86intrin.h>
2950 ///
2951 /// This intrinsic corresponds to the <c> VMOVMSKPD </c> instruction.
2952 ///
2953 /// \param __a
2954 ///    A 256-bit vector of [4 x double] containing the double-precision
2955 ///    floating point values with sign bits to be extracted.
2956 /// \returns The sign bits from the operand, written to bits [3:0].
2957 static __inline int __DEFAULT_FN_ATTRS
2958 _mm256_movemask_pd(__m256d __a)
2959 {
2960   return __builtin_ia32_movmskpd256((__v4df)__a);
2961 }
2962 
2963 /// Extracts the sign bits of single-precision floating point elements
2964 ///    in a 256-bit vector of [8 x float] and writes them to the lower order
2965 ///    bits of the return value.
2966 ///
2967 /// \headerfile <x86intrin.h>
2968 ///
2969 /// This intrinsic corresponds to the <c> VMOVMSKPS </c> instruction.
2970 ///
2971 /// \param __a
2972 ///    A 256-bit vector of [8 x float] containing the single-precision floating
2973 ///    point values with sign bits to be extracted.
2974 /// \returns The sign bits from the operand, written to bits [7:0].
2975 static __inline int __DEFAULT_FN_ATTRS
2976 _mm256_movemask_ps(__m256 __a)
2977 {
2978   return __builtin_ia32_movmskps256((__v8sf)__a);
2979 }
2980 
2981 /* Vector __zero */
2982 /// Zeroes the contents of all XMM or YMM registers.
2983 ///
2984 /// \headerfile <x86intrin.h>
2985 ///
2986 /// This intrinsic corresponds to the <c> VZEROALL </c> instruction.
2987 static __inline void __attribute__((__always_inline__, __nodebug__, __target__("avx")))
2988 _mm256_zeroall(void)
2989 {
2990   __builtin_ia32_vzeroall();
2991 }
2992 
2993 /// Zeroes the upper 128 bits (bits 255:128) of all YMM registers.
2994 ///
2995 /// \headerfile <x86intrin.h>
2996 ///
2997 /// This intrinsic corresponds to the <c> VZEROUPPER </c> instruction.
2998 static __inline void __attribute__((__always_inline__, __nodebug__, __target__("avx")))
2999 _mm256_zeroupper(void)
3000 {
3001   __builtin_ia32_vzeroupper();
3002 }
3003 
3004 /* Vector load with broadcast */
3005 /// Loads a scalar single-precision floating point value from the
3006 ///    specified address pointed to by \a __a and broadcasts it to the elements
3007 ///    of a [4 x float] vector.
3008 ///
3009 /// \headerfile <x86intrin.h>
3010 ///
3011 /// This intrinsic corresponds to the <c> VBROADCASTSS </c> instruction.
3012 ///
3013 /// \param __a
3014 ///    The single-precision floating point value to be broadcast.
3015 /// \returns A 128-bit vector of [4 x float] whose 32-bit elements are set
3016 ///    equal to the broadcast value.
3017 static __inline __m128 __DEFAULT_FN_ATTRS128
3018 _mm_broadcast_ss(float const *__a)
3019 {
3020   struct __mm_broadcast_ss_struct {
3021     float __f;
3022   } __attribute__((__packed__, __may_alias__));
3023   float __f = ((const struct __mm_broadcast_ss_struct*)__a)->__f;
3024   return __extension__ (__m128){ __f, __f, __f, __f };
3025 }
3026 
3027 /// Loads a scalar double-precision floating point value from the
3028 ///    specified address pointed to by \a __a and broadcasts it to the elements
3029 ///    of a [4 x double] vector.
3030 ///
3031 /// \headerfile <x86intrin.h>
3032 ///
3033 /// This intrinsic corresponds to the <c> VBROADCASTSD </c> instruction.
3034 ///
3035 /// \param __a
3036 ///    The double-precision floating point value to be broadcast.
3037 /// \returns A 256-bit vector of [4 x double] whose 64-bit elements are set
3038 ///    equal to the broadcast value.
3039 static __inline __m256d __DEFAULT_FN_ATTRS
3040 _mm256_broadcast_sd(double const *__a)
3041 {
3042   struct __mm256_broadcast_sd_struct {
3043     double __d;
3044   } __attribute__((__packed__, __may_alias__));
3045   double __d = ((const struct __mm256_broadcast_sd_struct*)__a)->__d;
3046   return __extension__ (__m256d)(__v4df){ __d, __d, __d, __d };
3047 }
3048 
3049 /// Loads a scalar single-precision floating point value from the
3050 ///    specified address pointed to by \a __a and broadcasts it to the elements
3051 ///    of a [8 x float] vector.
3052 ///
3053 /// \headerfile <x86intrin.h>
3054 ///
3055 /// This intrinsic corresponds to the <c> VBROADCASTSS </c> instruction.
3056 ///
3057 /// \param __a
3058 ///    The single-precision floating point value to be broadcast.
3059 /// \returns A 256-bit vector of [8 x float] whose 32-bit elements are set
3060 ///    equal to the broadcast value.
3061 static __inline __m256 __DEFAULT_FN_ATTRS
3062 _mm256_broadcast_ss(float const *__a)
3063 {
3064   struct __mm256_broadcast_ss_struct {
3065     float __f;
3066   } __attribute__((__packed__, __may_alias__));
3067   float __f = ((const struct __mm256_broadcast_ss_struct*)__a)->__f;
3068   return __extension__ (__m256)(__v8sf){ __f, __f, __f, __f, __f, __f, __f, __f };
3069 }
3070 
3071 /// Loads the data from a 128-bit vector of [2 x double] from the
3072 ///    specified address pointed to by \a __a and broadcasts it to 128-bit
3073 ///    elements in a 256-bit vector of [4 x double].
3074 ///
3075 /// \headerfile <x86intrin.h>
3076 ///
3077 /// This intrinsic corresponds to the <c> VBROADCASTF128 </c> instruction.
3078 ///
3079 /// \param __a
3080 ///    The 128-bit vector of [2 x double] to be broadcast.
3081 /// \returns A 256-bit vector of [4 x double] whose 128-bit elements are set
3082 ///    equal to the broadcast value.
3083 static __inline __m256d __DEFAULT_FN_ATTRS
3084 _mm256_broadcast_pd(__m128d const *__a)
3085 {
3086   __m128d __b = _mm_loadu_pd((const double *)__a);
3087   return (__m256d)__builtin_shufflevector((__v2df)__b, (__v2df)__b,
3088                                           0, 1, 0, 1);
3089 }
3090 
3091 /// Loads the data from a 128-bit vector of [4 x float] from the
3092 ///    specified address pointed to by \a __a and broadcasts it to 128-bit
3093 ///    elements in a 256-bit vector of [8 x float].
3094 ///
3095 /// \headerfile <x86intrin.h>
3096 ///
3097 /// This intrinsic corresponds to the <c> VBROADCASTF128 </c> instruction.
3098 ///
3099 /// \param __a
3100 ///    The 128-bit vector of [4 x float] to be broadcast.
3101 /// \returns A 256-bit vector of [8 x float] whose 128-bit elements are set
3102 ///    equal to the broadcast value.
3103 static __inline __m256 __DEFAULT_FN_ATTRS
3104 _mm256_broadcast_ps(__m128 const *__a)
3105 {
3106   __m128 __b = _mm_loadu_ps((const float *)__a);
3107   return (__m256)__builtin_shufflevector((__v4sf)__b, (__v4sf)__b,
3108                                          0, 1, 2, 3, 0, 1, 2, 3);
3109 }
3110 
3111 /* SIMD load ops */
3112 /// Loads 4 double-precision floating point values from a 32-byte aligned
3113 ///    memory location pointed to by \a __p into a vector of [4 x double].
3114 ///
3115 /// \headerfile <x86intrin.h>
3116 ///
3117 /// This intrinsic corresponds to the <c> VMOVAPD </c> instruction.
3118 ///
3119 /// \param __p
3120 ///    A 32-byte aligned pointer to a memory location containing
3121 ///    double-precision floating point values.
3122 /// \returns A 256-bit vector of [4 x double] containing the moved values.
3123 static __inline __m256d __DEFAULT_FN_ATTRS
3124 _mm256_load_pd(double const *__p)
3125 {
3126   return *(const __m256d *)__p;
3127 }
3128 
3129 /// Loads 8 single-precision floating point values from a 32-byte aligned
3130 ///    memory location pointed to by \a __p into a vector of [8 x float].
3131 ///
3132 /// \headerfile <x86intrin.h>
3133 ///
3134 /// This intrinsic corresponds to the <c> VMOVAPS </c> instruction.
3135 ///
3136 /// \param __p
3137 ///    A 32-byte aligned pointer to a memory location containing float values.
3138 /// \returns A 256-bit vector of [8 x float] containing the moved values.
3139 static __inline __m256 __DEFAULT_FN_ATTRS
3140 _mm256_load_ps(float const *__p)
3141 {
3142   return *(const __m256 *)__p;
3143 }
3144 
3145 /// Loads 4 double-precision floating point values from an unaligned
3146 ///    memory location pointed to by \a __p into a vector of [4 x double].
3147 ///
3148 /// \headerfile <x86intrin.h>
3149 ///
3150 /// This intrinsic corresponds to the <c> VMOVUPD </c> instruction.
3151 ///
3152 /// \param __p
3153 ///    A pointer to a memory location containing double-precision floating
3154 ///    point values.
3155 /// \returns A 256-bit vector of [4 x double] containing the moved values.
3156 static __inline __m256d __DEFAULT_FN_ATTRS
3157 _mm256_loadu_pd(double const *__p)
3158 {
3159   struct __loadu_pd {
3160     __m256d_u __v;
3161   } __attribute__((__packed__, __may_alias__));
3162   return ((const struct __loadu_pd*)__p)->__v;
3163 }
3164 
3165 /// Loads 8 single-precision floating point values from an unaligned
3166 ///    memory location pointed to by \a __p into a vector of [8 x float].
3167 ///
3168 /// \headerfile <x86intrin.h>
3169 ///
3170 /// This intrinsic corresponds to the <c> VMOVUPS </c> instruction.
3171 ///
3172 /// \param __p
3173 ///    A pointer to a memory location containing single-precision floating
3174 ///    point values.
3175 /// \returns A 256-bit vector of [8 x float] containing the moved values.
3176 static __inline __m256 __DEFAULT_FN_ATTRS
3177 _mm256_loadu_ps(float const *__p)
3178 {
3179   struct __loadu_ps {
3180     __m256_u __v;
3181   } __attribute__((__packed__, __may_alias__));
3182   return ((const struct __loadu_ps*)__p)->__v;
3183 }
3184 
3185 /// Loads 256 bits of integer data from a 32-byte aligned memory
3186 ///    location pointed to by \a __p into elements of a 256-bit integer vector.
3187 ///
3188 /// \headerfile <x86intrin.h>
3189 ///
3190 /// This intrinsic corresponds to the <c> VMOVDQA </c> instruction.
3191 ///
3192 /// \param __p
3193 ///    A 32-byte aligned pointer to a 256-bit integer vector containing integer
3194 ///    values.
3195 /// \returns A 256-bit integer vector containing the moved values.
3196 static __inline __m256i __DEFAULT_FN_ATTRS
3197 _mm256_load_si256(__m256i const *__p)
3198 {
3199   return *__p;
3200 }
3201 
3202 /// Loads 256 bits of integer data from an unaligned memory location
3203 ///    pointed to by \a __p into a 256-bit integer vector.
3204 ///
3205 /// \headerfile <x86intrin.h>
3206 ///
3207 /// This intrinsic corresponds to the <c> VMOVDQU </c> instruction.
3208 ///
3209 /// \param __p
3210 ///    A pointer to a 256-bit integer vector containing integer values.
3211 /// \returns A 256-bit integer vector containing the moved values.
3212 static __inline __m256i __DEFAULT_FN_ATTRS
3213 _mm256_loadu_si256(__m256i_u const *__p)
3214 {
3215   struct __loadu_si256 {
3216     __m256i_u __v;
3217   } __attribute__((__packed__, __may_alias__));
3218   return ((const struct __loadu_si256*)__p)->__v;
3219 }
3220 
3221 /// Loads 256 bits of integer data from an unaligned memory location
3222 ///    pointed to by \a __p into a 256-bit integer vector. This intrinsic may
3223 ///    perform better than \c _mm256_loadu_si256 when the data crosses a cache
3224 ///    line boundary.
3225 ///
3226 /// \headerfile <x86intrin.h>
3227 ///
3228 /// This intrinsic corresponds to the <c> VLDDQU </c> instruction.
3229 ///
3230 /// \param __p
3231 ///    A pointer to a 256-bit integer vector containing integer values.
3232 /// \returns A 256-bit integer vector containing the moved values.
3233 static __inline __m256i __DEFAULT_FN_ATTRS
3234 _mm256_lddqu_si256(__m256i_u const *__p)
3235 {
3236   return (__m256i)__builtin_ia32_lddqu256((char const *)__p);
3237 }
3238 
3239 /* SIMD store ops */
3240 /// Stores double-precision floating point values from a 256-bit vector
3241 ///    of [4 x double] to a 32-byte aligned memory location pointed to by
3242 ///    \a __p.
3243 ///
3244 /// \headerfile <x86intrin.h>
3245 ///
3246 /// This intrinsic corresponds to the <c> VMOVAPD </c> instruction.
3247 ///
3248 /// \param __p
3249 ///    A 32-byte aligned pointer to a memory location that will receive the
3250 ///    double-precision floaing point values.
3251 /// \param __a
3252 ///    A 256-bit vector of [4 x double] containing the values to be moved.
3253 static __inline void __DEFAULT_FN_ATTRS
3254 _mm256_store_pd(double *__p, __m256d __a)
3255 {
3256   *(__m256d *)__p = __a;
3257 }
3258 
3259 /// Stores single-precision floating point values from a 256-bit vector
3260 ///    of [8 x float] to a 32-byte aligned memory location pointed to by \a __p.
3261 ///
3262 /// \headerfile <x86intrin.h>
3263 ///
3264 /// This intrinsic corresponds to the <c> VMOVAPS </c> instruction.
3265 ///
3266 /// \param __p
3267 ///    A 32-byte aligned pointer to a memory location that will receive the
3268 ///    float values.
3269 /// \param __a
3270 ///    A 256-bit vector of [8 x float] containing the values to be moved.
3271 static __inline void __DEFAULT_FN_ATTRS
3272 _mm256_store_ps(float *__p, __m256 __a)
3273 {
3274   *(__m256 *)__p = __a;
3275 }
3276 
3277 /// Stores double-precision floating point values from a 256-bit vector
3278 ///    of [4 x double] to an unaligned memory location pointed to by \a __p.
3279 ///
3280 /// \headerfile <x86intrin.h>
3281 ///
3282 /// This intrinsic corresponds to the <c> VMOVUPD </c> instruction.
3283 ///
3284 /// \param __p
3285 ///    A pointer to a memory location that will receive the double-precision
3286 ///    floating point values.
3287 /// \param __a
3288 ///    A 256-bit vector of [4 x double] containing the values to be moved.
3289 static __inline void __DEFAULT_FN_ATTRS
3290 _mm256_storeu_pd(double *__p, __m256d __a)
3291 {
3292   struct __storeu_pd {
3293     __m256d_u __v;
3294   } __attribute__((__packed__, __may_alias__));
3295   ((struct __storeu_pd*)__p)->__v = __a;
3296 }
3297 
3298 /// Stores single-precision floating point values from a 256-bit vector
3299 ///    of [8 x float] to an unaligned memory location pointed to by \a __p.
3300 ///
3301 /// \headerfile <x86intrin.h>
3302 ///
3303 /// This intrinsic corresponds to the <c> VMOVUPS </c> instruction.
3304 ///
3305 /// \param __p
3306 ///    A pointer to a memory location that will receive the float values.
3307 /// \param __a
3308 ///    A 256-bit vector of [8 x float] containing the values to be moved.
3309 static __inline void __DEFAULT_FN_ATTRS
3310 _mm256_storeu_ps(float *__p, __m256 __a)
3311 {
3312   struct __storeu_ps {
3313     __m256_u __v;
3314   } __attribute__((__packed__, __may_alias__));
3315   ((struct __storeu_ps*)__p)->__v = __a;
3316 }
3317 
3318 /// Stores integer values from a 256-bit integer vector to a 32-byte
3319 ///    aligned memory location pointed to by \a __p.
3320 ///
3321 /// \headerfile <x86intrin.h>
3322 ///
3323 /// This intrinsic corresponds to the <c> VMOVDQA </c> instruction.
3324 ///
3325 /// \param __p
3326 ///    A 32-byte aligned pointer to a memory location that will receive the
3327 ///    integer values.
3328 /// \param __a
3329 ///    A 256-bit integer vector containing the values to be moved.
3330 static __inline void __DEFAULT_FN_ATTRS
3331 _mm256_store_si256(__m256i *__p, __m256i __a)
3332 {
3333   *__p = __a;
3334 }
3335 
3336 /// Stores integer values from a 256-bit integer vector to an unaligned
3337 ///    memory location pointed to by \a __p.
3338 ///
3339 /// \headerfile <x86intrin.h>
3340 ///
3341 /// This intrinsic corresponds to the <c> VMOVDQU </c> instruction.
3342 ///
3343 /// \param __p
3344 ///    A pointer to a memory location that will receive the integer values.
3345 /// \param __a
3346 ///    A 256-bit integer vector containing the values to be moved.
3347 static __inline void __DEFAULT_FN_ATTRS
3348 _mm256_storeu_si256(__m256i_u *__p, __m256i __a)
3349 {
3350   struct __storeu_si256 {
3351     __m256i_u __v;
3352   } __attribute__((__packed__, __may_alias__));
3353   ((struct __storeu_si256*)__p)->__v = __a;
3354 }
3355 
3356 /* Conditional load ops */
3357 /// Conditionally loads double-precision floating point elements from a
3358 ///    memory location pointed to by \a __p into a 128-bit vector of
3359 ///    [2 x double], depending on the mask bits associated with each data
3360 ///    element.
3361 ///
3362 /// \headerfile <x86intrin.h>
3363 ///
3364 /// This intrinsic corresponds to the <c> VMASKMOVPD </c> instruction.
3365 ///
3366 /// \param __p
3367 ///    A pointer to a memory location that contains the double-precision
3368 ///    floating point values.
3369 /// \param __m
3370 ///    A 128-bit integer vector containing the mask. The most significant bit of
3371 ///    each data element represents the mask bits. If a mask bit is zero, the
3372 ///    corresponding value in the memory location is not loaded and the
3373 ///    corresponding field in the return value is set to zero.
3374 /// \returns A 128-bit vector of [2 x double] containing the loaded values.
3375 static __inline __m128d __DEFAULT_FN_ATTRS128
3376 _mm_maskload_pd(double const *__p, __m128i __m)
3377 {
3378   return (__m128d)__builtin_ia32_maskloadpd((const __v2df *)__p, (__v2di)__m);
3379 }
3380 
3381 /// Conditionally loads double-precision floating point elements from a
3382 ///    memory location pointed to by \a __p into a 256-bit vector of
3383 ///    [4 x double], depending on the mask bits associated with each data
3384 ///    element.
3385 ///
3386 /// \headerfile <x86intrin.h>
3387 ///
3388 /// This intrinsic corresponds to the <c> VMASKMOVPD </c> instruction.
3389 ///
3390 /// \param __p
3391 ///    A pointer to a memory location that contains the double-precision
3392 ///    floating point values.
3393 /// \param __m
3394 ///    A 256-bit integer vector of [4 x quadword] containing the mask. The most
3395 ///    significant bit of each quadword element represents the mask bits. If a
3396 ///    mask bit is zero, the corresponding value in the memory location is not
3397 ///    loaded and the corresponding field in the return value is set to zero.
3398 /// \returns A 256-bit vector of [4 x double] containing the loaded values.
3399 static __inline __m256d __DEFAULT_FN_ATTRS
3400 _mm256_maskload_pd(double const *__p, __m256i __m)
3401 {
3402   return (__m256d)__builtin_ia32_maskloadpd256((const __v4df *)__p,
3403                                                (__v4di)__m);
3404 }
3405 
3406 /// Conditionally loads single-precision floating point elements from a
3407 ///    memory location pointed to by \a __p into a 128-bit vector of
3408 ///    [4 x float], depending on the mask bits associated with each data
3409 ///    element.
3410 ///
3411 /// \headerfile <x86intrin.h>
3412 ///
3413 /// This intrinsic corresponds to the <c> VMASKMOVPS </c> instruction.
3414 ///
3415 /// \param __p
3416 ///    A pointer to a memory location that contains the single-precision
3417 ///    floating point values.
3418 /// \param __m
3419 ///    A 128-bit integer vector containing the mask. The most significant bit of
3420 ///    each data element represents the mask bits. If a mask bit is zero, the
3421 ///    corresponding value in the memory location is not loaded and the
3422 ///    corresponding field in the return value is set to zero.
3423 /// \returns A 128-bit vector of [4 x float] containing the loaded values.
3424 static __inline __m128 __DEFAULT_FN_ATTRS128
3425 _mm_maskload_ps(float const *__p, __m128i __m)
3426 {
3427   return (__m128)__builtin_ia32_maskloadps((const __v4sf *)__p, (__v4si)__m);
3428 }
3429 
3430 /// Conditionally loads single-precision floating point elements from a
3431 ///    memory location pointed to by \a __p into a 256-bit vector of
3432 ///    [8 x float], depending on the mask bits associated with each data
3433 ///    element.
3434 ///
3435 /// \headerfile <x86intrin.h>
3436 ///
3437 /// This intrinsic corresponds to the <c> VMASKMOVPS </c> instruction.
3438 ///
3439 /// \param __p
3440 ///    A pointer to a memory location that contains the single-precision
3441 ///    floating point values.
3442 /// \param __m
3443 ///    A 256-bit integer vector of [8 x dword] containing the mask. The most
3444 ///    significant bit of each dword element represents the mask bits. If a mask
3445 ///    bit is zero, the corresponding value in the memory location is not loaded
3446 ///    and the corresponding field in the return value is set to zero.
3447 /// \returns A 256-bit vector of [8 x float] containing the loaded values.
3448 static __inline __m256 __DEFAULT_FN_ATTRS
3449 _mm256_maskload_ps(float const *__p, __m256i __m)
3450 {
3451   return (__m256)__builtin_ia32_maskloadps256((const __v8sf *)__p, (__v8si)__m);
3452 }
3453 
3454 /* Conditional store ops */
3455 /// Moves single-precision floating point values from a 256-bit vector
3456 ///    of [8 x float] to a memory location pointed to by \a __p, according to
3457 ///    the specified mask.
3458 ///
3459 /// \headerfile <x86intrin.h>
3460 ///
3461 /// This intrinsic corresponds to the <c> VMASKMOVPS </c> instruction.
3462 ///
3463 /// \param __p
3464 ///    A pointer to a memory location that will receive the float values.
3465 /// \param __m
3466 ///    A 256-bit integer vector of [8 x dword] containing the mask. The most
3467 ///    significant bit of each dword element in the mask vector represents the
3468 ///    mask bits. If a mask bit is zero, the corresponding value from vector
3469 ///    \a __a is not stored and the corresponding field in the memory location
3470 ///    pointed to by \a __p is not changed.
3471 /// \param __a
3472 ///    A 256-bit vector of [8 x float] containing the values to be stored.
3473 static __inline void __DEFAULT_FN_ATTRS
3474 _mm256_maskstore_ps(float *__p, __m256i __m, __m256 __a)
3475 {
3476   __builtin_ia32_maskstoreps256((__v8sf *)__p, (__v8si)__m, (__v8sf)__a);
3477 }
3478 
3479 /// Moves double-precision values from a 128-bit vector of [2 x double]
3480 ///    to a memory location pointed to by \a __p, according to the specified
3481 ///    mask.
3482 ///
3483 /// \headerfile <x86intrin.h>
3484 ///
3485 /// This intrinsic corresponds to the <c> VMASKMOVPD </c> instruction.
3486 ///
3487 /// \param __p
3488 ///    A pointer to a memory location that will receive the float values.
3489 /// \param __m
3490 ///    A 128-bit integer vector containing the mask. The most significant bit of
3491 ///    each field in the mask vector represents the mask bits. If a mask bit is
3492 ///    zero, the corresponding value from vector \a __a is not stored and the
3493 ///    corresponding field in the memory location pointed to by \a __p is not
3494 ///    changed.
3495 /// \param __a
3496 ///    A 128-bit vector of [2 x double] containing the values to be stored.
3497 static __inline void __DEFAULT_FN_ATTRS128
3498 _mm_maskstore_pd(double *__p, __m128i __m, __m128d __a)
3499 {
3500   __builtin_ia32_maskstorepd((__v2df *)__p, (__v2di)__m, (__v2df)__a);
3501 }
3502 
3503 /// Moves double-precision values from a 256-bit vector of [4 x double]
3504 ///    to a memory location pointed to by \a __p, according to the specified
3505 ///    mask.
3506 ///
3507 /// \headerfile <x86intrin.h>
3508 ///
3509 /// This intrinsic corresponds to the <c> VMASKMOVPD </c> instruction.
3510 ///
3511 /// \param __p
3512 ///    A pointer to a memory location that will receive the float values.
3513 /// \param __m
3514 ///    A 256-bit integer vector of [4 x quadword] containing the mask. The most
3515 ///    significant bit of each quadword element in the mask vector represents
3516 ///    the mask bits. If a mask bit is zero, the corresponding value from vector
3517 ///    __a is not stored and the corresponding field in the memory location
3518 ///    pointed to by \a __p is not changed.
3519 /// \param __a
3520 ///    A 256-bit vector of [4 x double] containing the values to be stored.
3521 static __inline void __DEFAULT_FN_ATTRS
3522 _mm256_maskstore_pd(double *__p, __m256i __m, __m256d __a)
3523 {
3524   __builtin_ia32_maskstorepd256((__v4df *)__p, (__v4di)__m, (__v4df)__a);
3525 }
3526 
3527 /// Moves single-precision floating point values from a 128-bit vector
3528 ///    of [4 x float] to a memory location pointed to by \a __p, according to
3529 ///    the specified mask.
3530 ///
3531 /// \headerfile <x86intrin.h>
3532 ///
3533 /// This intrinsic corresponds to the <c> VMASKMOVPS </c> instruction.
3534 ///
3535 /// \param __p
3536 ///    A pointer to a memory location that will receive the float values.
3537 /// \param __m
3538 ///    A 128-bit integer vector containing the mask. The most significant bit of
3539 ///    each field in the mask vector represents the mask bits. If a mask bit is
3540 ///    zero, the corresponding value from vector __a is not stored and the
3541 ///    corresponding field in the memory location pointed to by \a __p is not
3542 ///    changed.
3543 /// \param __a
3544 ///    A 128-bit vector of [4 x float] containing the values to be stored.
3545 static __inline void __DEFAULT_FN_ATTRS128
3546 _mm_maskstore_ps(float *__p, __m128i __m, __m128 __a)
3547 {
3548   __builtin_ia32_maskstoreps((__v4sf *)__p, (__v4si)__m, (__v4sf)__a);
3549 }
3550 
3551 /* Cacheability support ops */
3552 /// Moves integer data from a 256-bit integer vector to a 32-byte
3553 ///    aligned memory location. To minimize caching, the data is flagged as
3554 ///    non-temporal (unlikely to be used again soon).
3555 ///
3556 /// \headerfile <x86intrin.h>
3557 ///
3558 /// This intrinsic corresponds to the <c> VMOVNTDQ </c> instruction.
3559 ///
3560 /// \param __a
3561 ///    A pointer to a 32-byte aligned memory location that will receive the
3562 ///    integer values.
3563 /// \param __b
3564 ///    A 256-bit integer vector containing the values to be moved.
3565 static __inline void __DEFAULT_FN_ATTRS
3566 _mm256_stream_si256(__m256i *__a, __m256i __b)
3567 {
3568   typedef __v4di __v4di_aligned __attribute__((aligned(32)));
3569   __builtin_nontemporal_store((__v4di_aligned)__b, (__v4di_aligned*)__a);
3570 }
3571 
3572 /// Moves double-precision values from a 256-bit vector of [4 x double]
3573 ///    to a 32-byte aligned memory location. To minimize caching, the data is
3574 ///    flagged as non-temporal (unlikely to be used again soon).
3575 ///
3576 /// \headerfile <x86intrin.h>
3577 ///
3578 /// This intrinsic corresponds to the <c> VMOVNTPD </c> instruction.
3579 ///
3580 /// \param __a
3581 ///    A pointer to a 32-byte aligned memory location that will receive the
3582 ///    double-precision floating-point values.
3583 /// \param __b
3584 ///    A 256-bit vector of [4 x double] containing the values to be moved.
3585 static __inline void __DEFAULT_FN_ATTRS
3586 _mm256_stream_pd(double *__a, __m256d __b)
3587 {
3588   typedef __v4df __v4df_aligned __attribute__((aligned(32)));
3589   __builtin_nontemporal_store((__v4df_aligned)__b, (__v4df_aligned*)__a);
3590 }
3591 
3592 /// Moves single-precision floating point values from a 256-bit vector
3593 ///    of [8 x float] to a 32-byte aligned memory location. To minimize
3594 ///    caching, the data is flagged as non-temporal (unlikely to be used again
3595 ///    soon).
3596 ///
3597 /// \headerfile <x86intrin.h>
3598 ///
3599 /// This intrinsic corresponds to the <c> VMOVNTPS </c> instruction.
3600 ///
3601 /// \param __p
3602 ///    A pointer to a 32-byte aligned memory location that will receive the
3603 ///    single-precision floating point values.
3604 /// \param __a
3605 ///    A 256-bit vector of [8 x float] containing the values to be moved.
3606 static __inline void __DEFAULT_FN_ATTRS
3607 _mm256_stream_ps(float *__p, __m256 __a)
3608 {
3609   typedef __v8sf __v8sf_aligned __attribute__((aligned(32)));
3610   __builtin_nontemporal_store((__v8sf_aligned)__a, (__v8sf_aligned*)__p);
3611 }
3612 
3613 /* Create vectors */
3614 /// Create a 256-bit vector of [4 x double] with undefined values.
3615 ///
3616 /// \headerfile <x86intrin.h>
3617 ///
3618 /// This intrinsic has no corresponding instruction.
3619 ///
3620 /// \returns A 256-bit vector of [4 x double] containing undefined values.
3621 static __inline__ __m256d __DEFAULT_FN_ATTRS
3622 _mm256_undefined_pd(void)
3623 {
3624   return (__m256d)__builtin_ia32_undef256();
3625 }
3626 
3627 /// Create a 256-bit vector of [8 x float] with undefined values.
3628 ///
3629 /// \headerfile <x86intrin.h>
3630 ///
3631 /// This intrinsic has no corresponding instruction.
3632 ///
3633 /// \returns A 256-bit vector of [8 x float] containing undefined values.
3634 static __inline__ __m256 __DEFAULT_FN_ATTRS
3635 _mm256_undefined_ps(void)
3636 {
3637   return (__m256)__builtin_ia32_undef256();
3638 }
3639 
3640 /// Create a 256-bit integer vector with undefined values.
3641 ///
3642 /// \headerfile <x86intrin.h>
3643 ///
3644 /// This intrinsic has no corresponding instruction.
3645 ///
3646 /// \returns A 256-bit integer vector containing undefined values.
3647 static __inline__ __m256i __DEFAULT_FN_ATTRS
3648 _mm256_undefined_si256(void)
3649 {
3650   return (__m256i)__builtin_ia32_undef256();
3651 }
3652 
3653 /// Constructs a 256-bit floating-point vector of [4 x double]
3654 ///    initialized with the specified double-precision floating-point values.
3655 ///
3656 /// \headerfile <x86intrin.h>
3657 ///
3658 /// This intrinsic corresponds to the <c> VUNPCKLPD+VINSERTF128 </c>
3659 ///   instruction.
3660 ///
3661 /// \param __a
3662 ///    A double-precision floating-point value used to initialize bits [255:192]
3663 ///    of the result.
3664 /// \param __b
3665 ///    A double-precision floating-point value used to initialize bits [191:128]
3666 ///    of the result.
3667 /// \param __c
3668 ///    A double-precision floating-point value used to initialize bits [127:64]
3669 ///    of the result.
3670 /// \param __d
3671 ///    A double-precision floating-point value used to initialize bits [63:0]
3672 ///    of the result.
3673 /// \returns An initialized 256-bit floating-point vector of [4 x double].
3674 static __inline __m256d __DEFAULT_FN_ATTRS
3675 _mm256_set_pd(double __a, double __b, double __c, double __d)
3676 {
3677   return __extension__ (__m256d){ __d, __c, __b, __a };
3678 }
3679 
3680 /// Constructs a 256-bit floating-point vector of [8 x float] initialized
3681 ///    with the specified single-precision floating-point values.
3682 ///
3683 /// \headerfile <x86intrin.h>
3684 ///
3685 /// This intrinsic is a utility function and does not correspond to a specific
3686 ///   instruction.
3687 ///
3688 /// \param __a
3689 ///    A single-precision floating-point value used to initialize bits [255:224]
3690 ///    of the result.
3691 /// \param __b
3692 ///    A single-precision floating-point value used to initialize bits [223:192]
3693 ///    of the result.
3694 /// \param __c
3695 ///    A single-precision floating-point value used to initialize bits [191:160]
3696 ///    of the result.
3697 /// \param __d
3698 ///    A single-precision floating-point value used to initialize bits [159:128]
3699 ///    of the result.
3700 /// \param __e
3701 ///    A single-precision floating-point value used to initialize bits [127:96]
3702 ///    of the result.
3703 /// \param __f
3704 ///    A single-precision floating-point value used to initialize bits [95:64]
3705 ///    of the result.
3706 /// \param __g
3707 ///    A single-precision floating-point value used to initialize bits [63:32]
3708 ///    of the result.
3709 /// \param __h
3710 ///    A single-precision floating-point value used to initialize bits [31:0]
3711 ///    of the result.
3712 /// \returns An initialized 256-bit floating-point vector of [8 x float].
3713 static __inline __m256 __DEFAULT_FN_ATTRS
3714 _mm256_set_ps(float __a, float __b, float __c, float __d,
3715               float __e, float __f, float __g, float __h)
3716 {
3717   return __extension__ (__m256){ __h, __g, __f, __e, __d, __c, __b, __a };
3718 }
3719 
3720 /// Constructs a 256-bit integer vector initialized with the specified
3721 ///    32-bit integral values.
3722 ///
3723 /// \headerfile <x86intrin.h>
3724 ///
3725 /// This intrinsic is a utility function and does not correspond to a specific
3726 ///   instruction.
3727 ///
3728 /// \param __i0
3729 ///    A 32-bit integral value used to initialize bits [255:224] of the result.
3730 /// \param __i1
3731 ///    A 32-bit integral value used to initialize bits [223:192] of the result.
3732 /// \param __i2
3733 ///    A 32-bit integral value used to initialize bits [191:160] of the result.
3734 /// \param __i3
3735 ///    A 32-bit integral value used to initialize bits [159:128] of the result.
3736 /// \param __i4
3737 ///    A 32-bit integral value used to initialize bits [127:96] of the result.
3738 /// \param __i5
3739 ///    A 32-bit integral value used to initialize bits [95:64] of the result.
3740 /// \param __i6
3741 ///    A 32-bit integral value used to initialize bits [63:32] of the result.
3742 /// \param __i7
3743 ///    A 32-bit integral value used to initialize bits [31:0] of the result.
3744 /// \returns An initialized 256-bit integer vector.
3745 static __inline __m256i __DEFAULT_FN_ATTRS
3746 _mm256_set_epi32(int __i0, int __i1, int __i2, int __i3,
3747                  int __i4, int __i5, int __i6, int __i7)
3748 {
3749   return __extension__ (__m256i)(__v8si){ __i7, __i6, __i5, __i4, __i3, __i2, __i1, __i0 };
3750 }
3751 
3752 /// Constructs a 256-bit integer vector initialized with the specified
3753 ///    16-bit integral values.
3754 ///
3755 /// \headerfile <x86intrin.h>
3756 ///
3757 /// This intrinsic is a utility function and does not correspond to a specific
3758 ///   instruction.
3759 ///
3760 /// \param __w15
3761 ///    A 16-bit integral value used to initialize bits [255:240] of the result.
3762 /// \param __w14
3763 ///    A 16-bit integral value used to initialize bits [239:224] of the result.
3764 /// \param __w13
3765 ///    A 16-bit integral value used to initialize bits [223:208] of the result.
3766 /// \param __w12
3767 ///    A 16-bit integral value used to initialize bits [207:192] of the result.
3768 /// \param __w11
3769 ///    A 16-bit integral value used to initialize bits [191:176] of the result.
3770 /// \param __w10
3771 ///    A 16-bit integral value used to initialize bits [175:160] of the result.
3772 /// \param __w09
3773 ///    A 16-bit integral value used to initialize bits [159:144] of the result.
3774 /// \param __w08
3775 ///    A 16-bit integral value used to initialize bits [143:128] of the result.
3776 /// \param __w07
3777 ///    A 16-bit integral value used to initialize bits [127:112] of the result.
3778 /// \param __w06
3779 ///    A 16-bit integral value used to initialize bits [111:96] of the result.
3780 /// \param __w05
3781 ///    A 16-bit integral value used to initialize bits [95:80] of the result.
3782 /// \param __w04
3783 ///    A 16-bit integral value used to initialize bits [79:64] of the result.
3784 /// \param __w03
3785 ///    A 16-bit integral value used to initialize bits [63:48] of the result.
3786 /// \param __w02
3787 ///    A 16-bit integral value used to initialize bits [47:32] of the result.
3788 /// \param __w01
3789 ///    A 16-bit integral value used to initialize bits [31:16] of the result.
3790 /// \param __w00
3791 ///    A 16-bit integral value used to initialize bits [15:0] of the result.
3792 /// \returns An initialized 256-bit integer vector.
3793 static __inline __m256i __DEFAULT_FN_ATTRS
3794 _mm256_set_epi16(short __w15, short __w14, short __w13, short __w12,
3795                  short __w11, short __w10, short __w09, short __w08,
3796                  short __w07, short __w06, short __w05, short __w04,
3797                  short __w03, short __w02, short __w01, short __w00)
3798 {
3799   return __extension__ (__m256i)(__v16hi){ __w00, __w01, __w02, __w03, __w04, __w05, __w06,
3800     __w07, __w08, __w09, __w10, __w11, __w12, __w13, __w14, __w15 };
3801 }
3802 
3803 /// Constructs a 256-bit integer vector initialized with the specified
3804 ///    8-bit integral values.
3805 ///
3806 /// \headerfile <x86intrin.h>
3807 ///
3808 /// This intrinsic is a utility function and does not correspond to a specific
3809 ///   instruction.
3810 ///
3811 /// \param __b31
3812 ///    An 8-bit integral value used to initialize bits [255:248] of the result.
3813 /// \param __b30
3814 ///    An 8-bit integral value used to initialize bits [247:240] of the result.
3815 /// \param __b29
3816 ///    An 8-bit integral value used to initialize bits [239:232] of the result.
3817 /// \param __b28
3818 ///    An 8-bit integral value used to initialize bits [231:224] of the result.
3819 /// \param __b27
3820 ///    An 8-bit integral value used to initialize bits [223:216] of the result.
3821 /// \param __b26
3822 ///    An 8-bit integral value used to initialize bits [215:208] of the result.
3823 /// \param __b25
3824 ///    An 8-bit integral value used to initialize bits [207:200] of the result.
3825 /// \param __b24
3826 ///    An 8-bit integral value used to initialize bits [199:192] of the result.
3827 /// \param __b23
3828 ///    An 8-bit integral value used to initialize bits [191:184] of the result.
3829 /// \param __b22
3830 ///    An 8-bit integral value used to initialize bits [183:176] of the result.
3831 /// \param __b21
3832 ///    An 8-bit integral value used to initialize bits [175:168] of the result.
3833 /// \param __b20
3834 ///    An 8-bit integral value used to initialize bits [167:160] of the result.
3835 /// \param __b19
3836 ///    An 8-bit integral value used to initialize bits [159:152] of the result.
3837 /// \param __b18
3838 ///    An 8-bit integral value used to initialize bits [151:144] of the result.
3839 /// \param __b17
3840 ///    An 8-bit integral value used to initialize bits [143:136] of the result.
3841 /// \param __b16
3842 ///    An 8-bit integral value used to initialize bits [135:128] of the result.
3843 /// \param __b15
3844 ///    An 8-bit integral value used to initialize bits [127:120] of the result.
3845 /// \param __b14
3846 ///    An 8-bit integral value used to initialize bits [119:112] of the result.
3847 /// \param __b13
3848 ///    An 8-bit integral value used to initialize bits [111:104] of the result.
3849 /// \param __b12
3850 ///    An 8-bit integral value used to initialize bits [103:96] of the result.
3851 /// \param __b11
3852 ///    An 8-bit integral value used to initialize bits [95:88] of the result.
3853 /// \param __b10
3854 ///    An 8-bit integral value used to initialize bits [87:80] of the result.
3855 /// \param __b09
3856 ///    An 8-bit integral value used to initialize bits [79:72] of the result.
3857 /// \param __b08
3858 ///    An 8-bit integral value used to initialize bits [71:64] of the result.
3859 /// \param __b07
3860 ///    An 8-bit integral value used to initialize bits [63:56] of the result.
3861 /// \param __b06
3862 ///    An 8-bit integral value used to initialize bits [55:48] of the result.
3863 /// \param __b05
3864 ///    An 8-bit integral value used to initialize bits [47:40] of the result.
3865 /// \param __b04
3866 ///    An 8-bit integral value used to initialize bits [39:32] of the result.
3867 /// \param __b03
3868 ///    An 8-bit integral value used to initialize bits [31:24] of the result.
3869 /// \param __b02
3870 ///    An 8-bit integral value used to initialize bits [23:16] of the result.
3871 /// \param __b01
3872 ///    An 8-bit integral value used to initialize bits [15:8] of the result.
3873 /// \param __b00
3874 ///    An 8-bit integral value used to initialize bits [7:0] of the result.
3875 /// \returns An initialized 256-bit integer vector.
3876 static __inline __m256i __DEFAULT_FN_ATTRS
3877 _mm256_set_epi8(char __b31, char __b30, char __b29, char __b28,
3878                 char __b27, char __b26, char __b25, char __b24,
3879                 char __b23, char __b22, char __b21, char __b20,
3880                 char __b19, char __b18, char __b17, char __b16,
3881                 char __b15, char __b14, char __b13, char __b12,
3882                 char __b11, char __b10, char __b09, char __b08,
3883                 char __b07, char __b06, char __b05, char __b04,
3884                 char __b03, char __b02, char __b01, char __b00)
3885 {
3886   return __extension__ (__m256i)(__v32qi){
3887     __b00, __b01, __b02, __b03, __b04, __b05, __b06, __b07,
3888     __b08, __b09, __b10, __b11, __b12, __b13, __b14, __b15,
3889     __b16, __b17, __b18, __b19, __b20, __b21, __b22, __b23,
3890     __b24, __b25, __b26, __b27, __b28, __b29, __b30, __b31
3891   };
3892 }
3893 
3894 /// Constructs a 256-bit integer vector initialized with the specified
3895 ///    64-bit integral values.
3896 ///
3897 /// \headerfile <x86intrin.h>
3898 ///
3899 /// This intrinsic corresponds to the <c> VPUNPCKLQDQ+VINSERTF128 </c>
3900 ///   instruction.
3901 ///
3902 /// \param __a
3903 ///    A 64-bit integral value used to initialize bits [255:192] of the result.
3904 /// \param __b
3905 ///    A 64-bit integral value used to initialize bits [191:128] of the result.
3906 /// \param __c
3907 ///    A 64-bit integral value used to initialize bits [127:64] of the result.
3908 /// \param __d
3909 ///    A 64-bit integral value used to initialize bits [63:0] of the result.
3910 /// \returns An initialized 256-bit integer vector.
3911 static __inline __m256i __DEFAULT_FN_ATTRS
3912 _mm256_set_epi64x(long long __a, long long __b, long long __c, long long __d)
3913 {
3914   return __extension__ (__m256i)(__v4di){ __d, __c, __b, __a };
3915 }
3916 
3917 /* Create vectors with elements in reverse order */
3918 /// Constructs a 256-bit floating-point vector of [4 x double],
3919 ///    initialized in reverse order with the specified double-precision
3920 ///    floating-point values.
3921 ///
3922 /// \headerfile <x86intrin.h>
3923 ///
3924 /// This intrinsic corresponds to the <c> VUNPCKLPD+VINSERTF128 </c>
3925 ///   instruction.
3926 ///
3927 /// \param __a
3928 ///    A double-precision floating-point value used to initialize bits [63:0]
3929 ///    of the result.
3930 /// \param __b
3931 ///    A double-precision floating-point value used to initialize bits [127:64]
3932 ///    of the result.
3933 /// \param __c
3934 ///    A double-precision floating-point value used to initialize bits [191:128]
3935 ///    of the result.
3936 /// \param __d
3937 ///    A double-precision floating-point value used to initialize bits [255:192]
3938 ///    of the result.
3939 /// \returns An initialized 256-bit floating-point vector of [4 x double].
3940 static __inline __m256d __DEFAULT_FN_ATTRS
3941 _mm256_setr_pd(double __a, double __b, double __c, double __d)
3942 {
3943   return _mm256_set_pd(__d, __c, __b, __a);
3944 }
3945 
3946 /// Constructs a 256-bit floating-point vector of [8 x float],
3947 ///    initialized in reverse order with the specified single-precision
3948 ///    float-point values.
3949 ///
3950 /// \headerfile <x86intrin.h>
3951 ///
3952 /// This intrinsic is a utility function and does not correspond to a specific
3953 ///   instruction.
3954 ///
3955 /// \param __a
3956 ///    A single-precision floating-point value used to initialize bits [31:0]
3957 ///    of the result.
3958 /// \param __b
3959 ///    A single-precision floating-point value used to initialize bits [63:32]
3960 ///    of the result.
3961 /// \param __c
3962 ///    A single-precision floating-point value used to initialize bits [95:64]
3963 ///    of the result.
3964 /// \param __d
3965 ///    A single-precision floating-point value used to initialize bits [127:96]
3966 ///    of the result.
3967 /// \param __e
3968 ///    A single-precision floating-point value used to initialize bits [159:128]
3969 ///    of the result.
3970 /// \param __f
3971 ///    A single-precision floating-point value used to initialize bits [191:160]
3972 ///    of the result.
3973 /// \param __g
3974 ///    A single-precision floating-point value used to initialize bits [223:192]
3975 ///    of the result.
3976 /// \param __h
3977 ///    A single-precision floating-point value used to initialize bits [255:224]
3978 ///    of the result.
3979 /// \returns An initialized 256-bit floating-point vector of [8 x float].
3980 static __inline __m256 __DEFAULT_FN_ATTRS
3981 _mm256_setr_ps(float __a, float __b, float __c, float __d,
3982                float __e, float __f, float __g, float __h)
3983 {
3984   return _mm256_set_ps(__h, __g, __f, __e, __d, __c, __b, __a);
3985 }
3986 
3987 /// Constructs a 256-bit integer vector, initialized in reverse order
3988 ///    with the specified 32-bit integral values.
3989 ///
3990 /// \headerfile <x86intrin.h>
3991 ///
3992 /// This intrinsic is a utility function and does not correspond to a specific
3993 ///   instruction.
3994 ///
3995 /// \param __i0
3996 ///    A 32-bit integral value used to initialize bits [31:0] of the result.
3997 /// \param __i1
3998 ///    A 32-bit integral value used to initialize bits [63:32] of the result.
3999 /// \param __i2
4000 ///    A 32-bit integral value used to initialize bits [95:64] of the result.
4001 /// \param __i3
4002 ///    A 32-bit integral value used to initialize bits [127:96] of the result.
4003 /// \param __i4
4004 ///    A 32-bit integral value used to initialize bits [159:128] of the result.
4005 /// \param __i5
4006 ///    A 32-bit integral value used to initialize bits [191:160] of the result.
4007 /// \param __i6
4008 ///    A 32-bit integral value used to initialize bits [223:192] of the result.
4009 /// \param __i7
4010 ///    A 32-bit integral value used to initialize bits [255:224] of the result.
4011 /// \returns An initialized 256-bit integer vector.
4012 static __inline __m256i __DEFAULT_FN_ATTRS
4013 _mm256_setr_epi32(int __i0, int __i1, int __i2, int __i3,
4014                   int __i4, int __i5, int __i6, int __i7)
4015 {
4016   return _mm256_set_epi32(__i7, __i6, __i5, __i4, __i3, __i2, __i1, __i0);
4017 }
4018 
4019 /// Constructs a 256-bit integer vector, initialized in reverse order
4020 ///    with the specified 16-bit integral values.
4021 ///
4022 /// \headerfile <x86intrin.h>
4023 ///
4024 /// This intrinsic is a utility function and does not correspond to a specific
4025 ///   instruction.
4026 ///
4027 /// \param __w15
4028 ///    A 16-bit integral value used to initialize bits [15:0] of the result.
4029 /// \param __w14
4030 ///    A 16-bit integral value used to initialize bits [31:16] of the result.
4031 /// \param __w13
4032 ///    A 16-bit integral value used to initialize bits [47:32] of the result.
4033 /// \param __w12
4034 ///    A 16-bit integral value used to initialize bits [63:48] of the result.
4035 /// \param __w11
4036 ///    A 16-bit integral value used to initialize bits [79:64] of the result.
4037 /// \param __w10
4038 ///    A 16-bit integral value used to initialize bits [95:80] of the result.
4039 /// \param __w09
4040 ///    A 16-bit integral value used to initialize bits [111:96] of the result.
4041 /// \param __w08
4042 ///    A 16-bit integral value used to initialize bits [127:112] of the result.
4043 /// \param __w07
4044 ///    A 16-bit integral value used to initialize bits [143:128] of the result.
4045 /// \param __w06
4046 ///    A 16-bit integral value used to initialize bits [159:144] of the result.
4047 /// \param __w05
4048 ///    A 16-bit integral value used to initialize bits [175:160] of the result.
4049 /// \param __w04
4050 ///    A 16-bit integral value used to initialize bits [191:176] of the result.
4051 /// \param __w03
4052 ///    A 16-bit integral value used to initialize bits [207:192] of the result.
4053 /// \param __w02
4054 ///    A 16-bit integral value used to initialize bits [223:208] of the result.
4055 /// \param __w01
4056 ///    A 16-bit integral value used to initialize bits [239:224] of the result.
4057 /// \param __w00
4058 ///    A 16-bit integral value used to initialize bits [255:240] of the result.
4059 /// \returns An initialized 256-bit integer vector.
4060 static __inline __m256i __DEFAULT_FN_ATTRS
4061 _mm256_setr_epi16(short __w15, short __w14, short __w13, short __w12,
4062        short __w11, short __w10, short __w09, short __w08,
4063        short __w07, short __w06, short __w05, short __w04,
4064        short __w03, short __w02, short __w01, short __w00)
4065 {
4066   return _mm256_set_epi16(__w00, __w01, __w02, __w03,
4067                           __w04, __w05, __w06, __w07,
4068                           __w08, __w09, __w10, __w11,
4069                           __w12, __w13, __w14, __w15);
4070 }
4071 
4072 /// Constructs a 256-bit integer vector, initialized in reverse order
4073 ///    with the specified 8-bit integral values.
4074 ///
4075 /// \headerfile <x86intrin.h>
4076 ///
4077 /// This intrinsic is a utility function and does not correspond to a specific
4078 ///   instruction.
4079 ///
4080 /// \param __b31
4081 ///    An 8-bit integral value used to initialize bits [7:0] of the result.
4082 /// \param __b30
4083 ///    An 8-bit integral value used to initialize bits [15:8] of the result.
4084 /// \param __b29
4085 ///    An 8-bit integral value used to initialize bits [23:16] of the result.
4086 /// \param __b28
4087 ///    An 8-bit integral value used to initialize bits [31:24] of the result.
4088 /// \param __b27
4089 ///    An 8-bit integral value used to initialize bits [39:32] of the result.
4090 /// \param __b26
4091 ///    An 8-bit integral value used to initialize bits [47:40] of the result.
4092 /// \param __b25
4093 ///    An 8-bit integral value used to initialize bits [55:48] of the result.
4094 /// \param __b24
4095 ///    An 8-bit integral value used to initialize bits [63:56] of the result.
4096 /// \param __b23
4097 ///    An 8-bit integral value used to initialize bits [71:64] of the result.
4098 /// \param __b22
4099 ///    An 8-bit integral value used to initialize bits [79:72] of the result.
4100 /// \param __b21
4101 ///    An 8-bit integral value used to initialize bits [87:80] of the result.
4102 /// \param __b20
4103 ///    An 8-bit integral value used to initialize bits [95:88] of the result.
4104 /// \param __b19
4105 ///    An 8-bit integral value used to initialize bits [103:96] of the result.
4106 /// \param __b18
4107 ///    An 8-bit integral value used to initialize bits [111:104] of the result.
4108 /// \param __b17
4109 ///    An 8-bit integral value used to initialize bits [119:112] of the result.
4110 /// \param __b16
4111 ///    An 8-bit integral value used to initialize bits [127:120] of the result.
4112 /// \param __b15
4113 ///    An 8-bit integral value used to initialize bits [135:128] of the result.
4114 /// \param __b14
4115 ///    An 8-bit integral value used to initialize bits [143:136] of the result.
4116 /// \param __b13
4117 ///    An 8-bit integral value used to initialize bits [151:144] of the result.
4118 /// \param __b12
4119 ///    An 8-bit integral value used to initialize bits [159:152] of the result.
4120 /// \param __b11
4121 ///    An 8-bit integral value used to initialize bits [167:160] of the result.
4122 /// \param __b10
4123 ///    An 8-bit integral value used to initialize bits [175:168] of the result.
4124 /// \param __b09
4125 ///    An 8-bit integral value used to initialize bits [183:176] of the result.
4126 /// \param __b08
4127 ///    An 8-bit integral value used to initialize bits [191:184] of the result.
4128 /// \param __b07
4129 ///    An 8-bit integral value used to initialize bits [199:192] of the result.
4130 /// \param __b06
4131 ///    An 8-bit integral value used to initialize bits [207:200] of the result.
4132 /// \param __b05
4133 ///    An 8-bit integral value used to initialize bits [215:208] of the result.
4134 /// \param __b04
4135 ///    An 8-bit integral value used to initialize bits [223:216] of the result.
4136 /// \param __b03
4137 ///    An 8-bit integral value used to initialize bits [231:224] of the result.
4138 /// \param __b02
4139 ///    An 8-bit integral value used to initialize bits [239:232] of the result.
4140 /// \param __b01
4141 ///    An 8-bit integral value used to initialize bits [247:240] of the result.
4142 /// \param __b00
4143 ///    An 8-bit integral value used to initialize bits [255:248] of the result.
4144 /// \returns An initialized 256-bit integer vector.
4145 static __inline __m256i __DEFAULT_FN_ATTRS
4146 _mm256_setr_epi8(char __b31, char __b30, char __b29, char __b28,
4147                  char __b27, char __b26, char __b25, char __b24,
4148                  char __b23, char __b22, char __b21, char __b20,
4149                  char __b19, char __b18, char __b17, char __b16,
4150                  char __b15, char __b14, char __b13, char __b12,
4151                  char __b11, char __b10, char __b09, char __b08,
4152                  char __b07, char __b06, char __b05, char __b04,
4153                  char __b03, char __b02, char __b01, char __b00)
4154 {
4155   return _mm256_set_epi8(__b00, __b01, __b02, __b03, __b04, __b05, __b06, __b07,
4156                          __b08, __b09, __b10, __b11, __b12, __b13, __b14, __b15,
4157                          __b16, __b17, __b18, __b19, __b20, __b21, __b22, __b23,
4158                          __b24, __b25, __b26, __b27, __b28, __b29, __b30, __b31);
4159 }
4160 
4161 /// Constructs a 256-bit integer vector, initialized in reverse order
4162 ///    with the specified 64-bit integral values.
4163 ///
4164 /// \headerfile <x86intrin.h>
4165 ///
4166 /// This intrinsic corresponds to the <c> VPUNPCKLQDQ+VINSERTF128 </c>
4167 ///   instruction.
4168 ///
4169 /// \param __a
4170 ///    A 64-bit integral value used to initialize bits [63:0] of the result.
4171 /// \param __b
4172 ///    A 64-bit integral value used to initialize bits [127:64] of the result.
4173 /// \param __c
4174 ///    A 64-bit integral value used to initialize bits [191:128] of the result.
4175 /// \param __d
4176 ///    A 64-bit integral value used to initialize bits [255:192] of the result.
4177 /// \returns An initialized 256-bit integer vector.
4178 static __inline __m256i __DEFAULT_FN_ATTRS
4179 _mm256_setr_epi64x(long long __a, long long __b, long long __c, long long __d)
4180 {
4181   return _mm256_set_epi64x(__d, __c, __b, __a);
4182 }
4183 
4184 /* Create vectors with repeated elements */
4185 /// Constructs a 256-bit floating-point vector of [4 x double], with each
4186 ///    of the four double-precision floating-point vector elements set to the
4187 ///    specified double-precision floating-point value.
4188 ///
4189 /// \headerfile <x86intrin.h>
4190 ///
4191 /// This intrinsic corresponds to the <c> VMOVDDUP+VINSERTF128 </c> instruction.
4192 ///
4193 /// \param __w
4194 ///    A double-precision floating-point value used to initialize each vector
4195 ///    element of the result.
4196 /// \returns An initialized 256-bit floating-point vector of [4 x double].
4197 static __inline __m256d __DEFAULT_FN_ATTRS
4198 _mm256_set1_pd(double __w)
4199 {
4200   return _mm256_set_pd(__w, __w, __w, __w);
4201 }
4202 
4203 /// Constructs a 256-bit floating-point vector of [8 x float], with each
4204 ///    of the eight single-precision floating-point vector elements set to the
4205 ///    specified single-precision floating-point value.
4206 ///
4207 /// \headerfile <x86intrin.h>
4208 ///
4209 /// This intrinsic corresponds to the <c> VPERMILPS+VINSERTF128 </c>
4210 ///   instruction.
4211 ///
4212 /// \param __w
4213 ///    A single-precision floating-point value used to initialize each vector
4214 ///    element of the result.
4215 /// \returns An initialized 256-bit floating-point vector of [8 x float].
4216 static __inline __m256 __DEFAULT_FN_ATTRS
4217 _mm256_set1_ps(float __w)
4218 {
4219   return _mm256_set_ps(__w, __w, __w, __w, __w, __w, __w, __w);
4220 }
4221 
4222 /// Constructs a 256-bit integer vector of [8 x i32], with each of the
4223 ///    32-bit integral vector elements set to the specified 32-bit integral
4224 ///    value.
4225 ///
4226 /// \headerfile <x86intrin.h>
4227 ///
4228 /// This intrinsic corresponds to the <c> VPERMILPS+VINSERTF128 </c>
4229 ///   instruction.
4230 ///
4231 /// \param __i
4232 ///    A 32-bit integral value used to initialize each vector element of the
4233 ///    result.
4234 /// \returns An initialized 256-bit integer vector of [8 x i32].
4235 static __inline __m256i __DEFAULT_FN_ATTRS
4236 _mm256_set1_epi32(int __i)
4237 {
4238   return _mm256_set_epi32(__i, __i, __i, __i, __i, __i, __i, __i);
4239 }
4240 
4241 /// Constructs a 256-bit integer vector of [16 x i16], with each of the
4242 ///    16-bit integral vector elements set to the specified 16-bit integral
4243 ///    value.
4244 ///
4245 /// \headerfile <x86intrin.h>
4246 ///
4247 /// This intrinsic corresponds to the <c> VPSHUFB+VINSERTF128 </c> instruction.
4248 ///
4249 /// \param __w
4250 ///    A 16-bit integral value used to initialize each vector element of the
4251 ///    result.
4252 /// \returns An initialized 256-bit integer vector of [16 x i16].
4253 static __inline __m256i __DEFAULT_FN_ATTRS
4254 _mm256_set1_epi16(short __w)
4255 {
4256   return _mm256_set_epi16(__w, __w, __w, __w, __w, __w, __w, __w,
4257                           __w, __w, __w, __w, __w, __w, __w, __w);
4258 }
4259 
4260 /// Constructs a 256-bit integer vector of [32 x i8], with each of the
4261 ///    8-bit integral vector elements set to the specified 8-bit integral value.
4262 ///
4263 /// \headerfile <x86intrin.h>
4264 ///
4265 /// This intrinsic corresponds to the <c> VPSHUFB+VINSERTF128 </c> instruction.
4266 ///
4267 /// \param __b
4268 ///    An 8-bit integral value used to initialize each vector element of the
4269 ///    result.
4270 /// \returns An initialized 256-bit integer vector of [32 x i8].
4271 static __inline __m256i __DEFAULT_FN_ATTRS
4272 _mm256_set1_epi8(char __b)
4273 {
4274   return _mm256_set_epi8(__b, __b, __b, __b, __b, __b, __b, __b,
4275                          __b, __b, __b, __b, __b, __b, __b, __b,
4276                          __b, __b, __b, __b, __b, __b, __b, __b,
4277                          __b, __b, __b, __b, __b, __b, __b, __b);
4278 }
4279 
4280 /// Constructs a 256-bit integer vector of [4 x i64], with each of the
4281 ///    64-bit integral vector elements set to the specified 64-bit integral
4282 ///    value.
4283 ///
4284 /// \headerfile <x86intrin.h>
4285 ///
4286 /// This intrinsic corresponds to the <c> VMOVDDUP+VINSERTF128 </c> instruction.
4287 ///
4288 /// \param __q
4289 ///    A 64-bit integral value used to initialize each vector element of the
4290 ///    result.
4291 /// \returns An initialized 256-bit integer vector of [4 x i64].
4292 static __inline __m256i __DEFAULT_FN_ATTRS
4293 _mm256_set1_epi64x(long long __q)
4294 {
4295   return _mm256_set_epi64x(__q, __q, __q, __q);
4296 }
4297 
4298 /* Create __zeroed vectors */
4299 /// Constructs a 256-bit floating-point vector of [4 x double] with all
4300 ///    vector elements initialized to zero.
4301 ///
4302 /// \headerfile <x86intrin.h>
4303 ///
4304 /// This intrinsic corresponds to the <c> VXORPS </c> instruction.
4305 ///
4306 /// \returns A 256-bit vector of [4 x double] with all elements set to zero.
4307 static __inline __m256d __DEFAULT_FN_ATTRS
4308 _mm256_setzero_pd(void)
4309 {
4310   return __extension__ (__m256d){ 0.0, 0.0, 0.0, 0.0 };
4311 }
4312 
4313 /// Constructs a 256-bit floating-point vector of [8 x float] with all
4314 ///    vector elements initialized to zero.
4315 ///
4316 /// \headerfile <x86intrin.h>
4317 ///
4318 /// This intrinsic corresponds to the <c> VXORPS </c> instruction.
4319 ///
4320 /// \returns A 256-bit vector of [8 x float] with all elements set to zero.
4321 static __inline __m256 __DEFAULT_FN_ATTRS
4322 _mm256_setzero_ps(void)
4323 {
4324   return __extension__ (__m256){ 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f };
4325 }
4326 
4327 /// Constructs a 256-bit integer vector initialized to zero.
4328 ///
4329 /// \headerfile <x86intrin.h>
4330 ///
4331 /// This intrinsic corresponds to the <c> VXORPS </c> instruction.
4332 ///
4333 /// \returns A 256-bit integer vector initialized to zero.
4334 static __inline __m256i __DEFAULT_FN_ATTRS
4335 _mm256_setzero_si256(void)
4336 {
4337   return __extension__ (__m256i)(__v4di){ 0, 0, 0, 0 };
4338 }
4339 
4340 /* Cast between vector types */
4341 /// Casts a 256-bit floating-point vector of [4 x double] into a 256-bit
4342 ///    floating-point vector of [8 x float].
4343 ///
4344 /// \headerfile <x86intrin.h>
4345 ///
4346 /// This intrinsic has no corresponding instruction.
4347 ///
4348 /// \param __a
4349 ///    A 256-bit floating-point vector of [4 x double].
4350 /// \returns A 256-bit floating-point vector of [8 x float] containing the same
4351 ///    bitwise pattern as the parameter.
4352 static __inline __m256 __DEFAULT_FN_ATTRS
4353 _mm256_castpd_ps(__m256d __a)
4354 {
4355   return (__m256)__a;
4356 }
4357 
4358 /// Casts a 256-bit floating-point vector of [4 x double] into a 256-bit
4359 ///    integer vector.
4360 ///
4361 /// \headerfile <x86intrin.h>
4362 ///
4363 /// This intrinsic has no corresponding instruction.
4364 ///
4365 /// \param __a
4366 ///    A 256-bit floating-point vector of [4 x double].
4367 /// \returns A 256-bit integer vector containing the same bitwise pattern as the
4368 ///    parameter.
4369 static __inline __m256i __DEFAULT_FN_ATTRS
4370 _mm256_castpd_si256(__m256d __a)
4371 {
4372   return (__m256i)__a;
4373 }
4374 
4375 /// Casts a 256-bit floating-point vector of [8 x float] into a 256-bit
4376 ///    floating-point vector of [4 x double].
4377 ///
4378 /// \headerfile <x86intrin.h>
4379 ///
4380 /// This intrinsic has no corresponding instruction.
4381 ///
4382 /// \param __a
4383 ///    A 256-bit floating-point vector of [8 x float].
4384 /// \returns A 256-bit floating-point vector of [4 x double] containing the same
4385 ///    bitwise pattern as the parameter.
4386 static __inline __m256d __DEFAULT_FN_ATTRS
4387 _mm256_castps_pd(__m256 __a)
4388 {
4389   return (__m256d)__a;
4390 }
4391 
4392 /// Casts a 256-bit floating-point vector of [8 x float] into a 256-bit
4393 ///    integer vector.
4394 ///
4395 /// \headerfile <x86intrin.h>
4396 ///
4397 /// This intrinsic has no corresponding instruction.
4398 ///
4399 /// \param __a
4400 ///    A 256-bit floating-point vector of [8 x float].
4401 /// \returns A 256-bit integer vector containing the same bitwise pattern as the
4402 ///    parameter.
4403 static __inline __m256i __DEFAULT_FN_ATTRS
4404 _mm256_castps_si256(__m256 __a)
4405 {
4406   return (__m256i)__a;
4407 }
4408 
4409 /// Casts a 256-bit integer vector into a 256-bit floating-point vector
4410 ///    of [8 x float].
4411 ///
4412 /// \headerfile <x86intrin.h>
4413 ///
4414 /// This intrinsic has no corresponding instruction.
4415 ///
4416 /// \param __a
4417 ///    A 256-bit integer vector.
4418 /// \returns A 256-bit floating-point vector of [8 x float] containing the same
4419 ///    bitwise pattern as the parameter.
4420 static __inline __m256 __DEFAULT_FN_ATTRS
4421 _mm256_castsi256_ps(__m256i __a)
4422 {
4423   return (__m256)__a;
4424 }
4425 
4426 /// Casts a 256-bit integer vector into a 256-bit floating-point vector
4427 ///    of [4 x double].
4428 ///
4429 /// \headerfile <x86intrin.h>
4430 ///
4431 /// This intrinsic has no corresponding instruction.
4432 ///
4433 /// \param __a
4434 ///    A 256-bit integer vector.
4435 /// \returns A 256-bit floating-point vector of [4 x double] containing the same
4436 ///    bitwise pattern as the parameter.
4437 static __inline __m256d __DEFAULT_FN_ATTRS
4438 _mm256_castsi256_pd(__m256i __a)
4439 {
4440   return (__m256d)__a;
4441 }
4442 
4443 /// Returns the lower 128 bits of a 256-bit floating-point vector of
4444 ///    [4 x double] as a 128-bit floating-point vector of [2 x double].
4445 ///
4446 /// \headerfile <x86intrin.h>
4447 ///
4448 /// This intrinsic has no corresponding instruction.
4449 ///
4450 /// \param __a
4451 ///    A 256-bit floating-point vector of [4 x double].
4452 /// \returns A 128-bit floating-point vector of [2 x double] containing the
4453 ///    lower 128 bits of the parameter.
4454 static __inline __m128d __DEFAULT_FN_ATTRS
4455 _mm256_castpd256_pd128(__m256d __a)
4456 {
4457   return __builtin_shufflevector((__v4df)__a, (__v4df)__a, 0, 1);
4458 }
4459 
4460 /// Returns the lower 128 bits of a 256-bit floating-point vector of
4461 ///    [8 x float] as a 128-bit floating-point vector of [4 x float].
4462 ///
4463 /// \headerfile <x86intrin.h>
4464 ///
4465 /// This intrinsic has no corresponding instruction.
4466 ///
4467 /// \param __a
4468 ///    A 256-bit floating-point vector of [8 x float].
4469 /// \returns A 128-bit floating-point vector of [4 x float] containing the
4470 ///    lower 128 bits of the parameter.
4471 static __inline __m128 __DEFAULT_FN_ATTRS
4472 _mm256_castps256_ps128(__m256 __a)
4473 {
4474   return __builtin_shufflevector((__v8sf)__a, (__v8sf)__a, 0, 1, 2, 3);
4475 }
4476 
4477 /// Truncates a 256-bit integer vector into a 128-bit integer vector.
4478 ///
4479 /// \headerfile <x86intrin.h>
4480 ///
4481 /// This intrinsic has no corresponding instruction.
4482 ///
4483 /// \param __a
4484 ///    A 256-bit integer vector.
4485 /// \returns A 128-bit integer vector containing the lower 128 bits of the
4486 ///    parameter.
4487 static __inline __m128i __DEFAULT_FN_ATTRS
4488 _mm256_castsi256_si128(__m256i __a)
4489 {
4490   return __builtin_shufflevector((__v4di)__a, (__v4di)__a, 0, 1);
4491 }
4492 
4493 /// Constructs a 256-bit floating-point vector of [4 x double] from a
4494 ///    128-bit floating-point vector of [2 x double].
4495 ///
4496 ///    The lower 128 bits contain the value of the source vector. The contents
4497 ///    of the upper 128 bits are undefined.
4498 ///
4499 /// \headerfile <x86intrin.h>
4500 ///
4501 /// This intrinsic has no corresponding instruction.
4502 ///
4503 /// \param __a
4504 ///    A 128-bit vector of [2 x double].
4505 /// \returns A 256-bit floating-point vector of [4 x double]. The lower 128 bits
4506 ///    contain the value of the parameter. The contents of the upper 128 bits
4507 ///    are undefined.
4508 static __inline __m256d __DEFAULT_FN_ATTRS
4509 _mm256_castpd128_pd256(__m128d __a)
4510 {
4511   return __builtin_shufflevector(
4512       (__v2df)__a, (__v2df)__builtin_nondeterministic_value(__a), 0, 1, 2, 3);
4513 }
4514 
4515 /// Constructs a 256-bit floating-point vector of [8 x float] from a
4516 ///    128-bit floating-point vector of [4 x float].
4517 ///
4518 ///    The lower 128 bits contain the value of the source vector. The contents
4519 ///    of the upper 128 bits are undefined.
4520 ///
4521 /// \headerfile <x86intrin.h>
4522 ///
4523 /// This intrinsic has no corresponding instruction.
4524 ///
4525 /// \param __a
4526 ///    A 128-bit vector of [4 x float].
4527 /// \returns A 256-bit floating-point vector of [8 x float]. The lower 128 bits
4528 ///    contain the value of the parameter. The contents of the upper 128 bits
4529 ///    are undefined.
4530 static __inline __m256 __DEFAULT_FN_ATTRS
4531 _mm256_castps128_ps256(__m128 __a)
4532 {
4533   return __builtin_shufflevector((__v4sf)__a,
4534                                  (__v4sf)__builtin_nondeterministic_value(__a),
4535                                  0, 1, 2, 3, 4, 5, 6, 7);
4536 }
4537 
4538 /// Constructs a 256-bit integer vector from a 128-bit integer vector.
4539 ///
4540 ///    The lower 128 bits contain the value of the source vector. The contents
4541 ///    of the upper 128 bits are undefined.
4542 ///
4543 /// \headerfile <x86intrin.h>
4544 ///
4545 /// This intrinsic has no corresponding instruction.
4546 ///
4547 /// \param __a
4548 ///    A 128-bit integer vector.
4549 /// \returns A 256-bit integer vector. The lower 128 bits contain the value of
4550 ///    the parameter. The contents of the upper 128 bits are undefined.
4551 static __inline __m256i __DEFAULT_FN_ATTRS
4552 _mm256_castsi128_si256(__m128i __a)
4553 {
4554   return __builtin_shufflevector(
4555       (__v2di)__a, (__v2di)__builtin_nondeterministic_value(__a), 0, 1, 2, 3);
4556 }
4557 
4558 /// Constructs a 256-bit floating-point vector of [4 x double] from a
4559 ///    128-bit floating-point vector of [2 x double]. The lower 128 bits
4560 ///    contain the value of the source vector. The upper 128 bits are set
4561 ///    to zero.
4562 ///
4563 /// \headerfile <x86intrin.h>
4564 ///
4565 /// This intrinsic has no corresponding instruction.
4566 ///
4567 /// \param __a
4568 ///    A 128-bit vector of [2 x double].
4569 /// \returns A 256-bit floating-point vector of [4 x double]. The lower 128 bits
4570 ///    contain the value of the parameter. The upper 128 bits are set to zero.
4571 static __inline __m256d __DEFAULT_FN_ATTRS
4572 _mm256_zextpd128_pd256(__m128d __a)
4573 {
4574   return __builtin_shufflevector((__v2df)__a, (__v2df)_mm_setzero_pd(), 0, 1, 2, 3);
4575 }
4576 
4577 /// Constructs a 256-bit floating-point vector of [8 x float] from a
4578 ///    128-bit floating-point vector of [4 x float]. The lower 128 bits contain
4579 ///    the value of the source vector. The upper 128 bits are set to zero.
4580 ///
4581 /// \headerfile <x86intrin.h>
4582 ///
4583 /// This intrinsic has no corresponding instruction.
4584 ///
4585 /// \param __a
4586 ///    A 128-bit vector of [4 x float].
4587 /// \returns A 256-bit floating-point vector of [8 x float]. The lower 128 bits
4588 ///    contain the value of the parameter. The upper 128 bits are set to zero.
4589 static __inline __m256 __DEFAULT_FN_ATTRS
4590 _mm256_zextps128_ps256(__m128 __a)
4591 {
4592   return __builtin_shufflevector((__v4sf)__a, (__v4sf)_mm_setzero_ps(), 0, 1, 2, 3, 4, 5, 6, 7);
4593 }
4594 
4595 /// Constructs a 256-bit integer vector from a 128-bit integer vector.
4596 ///    The lower 128 bits contain the value of the source vector. The upper
4597 ///    128 bits are set to zero.
4598 ///
4599 /// \headerfile <x86intrin.h>
4600 ///
4601 /// This intrinsic has no corresponding instruction.
4602 ///
4603 /// \param __a
4604 ///    A 128-bit integer vector.
4605 /// \returns A 256-bit integer vector. The lower 128 bits contain the value of
4606 ///    the parameter. The upper 128 bits are set to zero.
4607 static __inline __m256i __DEFAULT_FN_ATTRS
4608 _mm256_zextsi128_si256(__m128i __a)
4609 {
4610   return __builtin_shufflevector((__v2di)__a, (__v2di)_mm_setzero_si128(), 0, 1, 2, 3);
4611 }
4612 
4613 /*
4614    Vector insert.
4615    We use macros rather than inlines because we only want to accept
4616    invocations where the immediate M is a constant expression.
4617 */
4618 /// Constructs a new 256-bit vector of [8 x float] by first duplicating
4619 ///    a 256-bit vector of [8 x float] given in the first parameter, and then
4620 ///    replacing either the upper or the lower 128 bits with the contents of a
4621 ///    128-bit vector of [4 x float] in the second parameter.
4622 ///
4623 ///    The immediate integer parameter determines between the upper or the lower
4624 ///    128 bits.
4625 ///
4626 /// \headerfile <x86intrin.h>
4627 ///
4628 /// \code
4629 /// __m256 _mm256_insertf128_ps(__m256 V1, __m128 V2, const int M);
4630 /// \endcode
4631 ///
4632 /// This intrinsic corresponds to the <c> VINSERTF128 </c> instruction.
4633 ///
4634 /// \param V1
4635 ///    A 256-bit vector of [8 x float]. This vector is copied to the result
4636 ///    first, and then either the upper or the lower 128 bits of the result will
4637 ///    be replaced by the contents of \a V2.
4638 /// \param V2
4639 ///    A 128-bit vector of [4 x float]. The contents of this parameter are
4640 ///    written to either the upper or the lower 128 bits of the result depending
4641 ///    on the value of parameter \a M.
4642 /// \param M
4643 ///    An immediate integer. The least significant bit determines how the values
4644 ///    from the two parameters are interleaved: \n
4645 ///    If bit [0] of \a M is 0, \a V2 are copied to bits [127:0] of the result,
4646 ///    and bits [255:128] of \a V1 are copied to bits [255:128] of the
4647 ///    result. \n
4648 ///    If bit [0] of \a M is 1, \a V2 are copied to bits [255:128] of the
4649 ///    result, and bits [127:0] of \a V1 are copied to bits [127:0] of the
4650 ///    result.
4651 /// \returns A 256-bit vector of [8 x float] containing the interleaved values.
4652 #define _mm256_insertf128_ps(V1, V2, M) \
4653   ((__m256)__builtin_ia32_vinsertf128_ps256((__v8sf)(__m256)(V1), \
4654                                             (__v4sf)(__m128)(V2), (int)(M)))
4655 
4656 /// Constructs a new 256-bit vector of [4 x double] by first duplicating
4657 ///    a 256-bit vector of [4 x double] given in the first parameter, and then
4658 ///    replacing either the upper or the lower 128 bits with the contents of a
4659 ///    128-bit vector of [2 x double] in the second parameter.
4660 ///
4661 ///    The immediate integer parameter determines between the upper or the lower
4662 ///    128 bits.
4663 ///
4664 /// \headerfile <x86intrin.h>
4665 ///
4666 /// \code
4667 /// __m256d _mm256_insertf128_pd(__m256d V1, __m128d V2, const int M);
4668 /// \endcode
4669 ///
4670 /// This intrinsic corresponds to the <c> VINSERTF128 </c> instruction.
4671 ///
4672 /// \param V1
4673 ///    A 256-bit vector of [4 x double]. This vector is copied to the result
4674 ///    first, and then either the upper or the lower 128 bits of the result will
4675 ///    be replaced by the contents of \a V2.
4676 /// \param V2
4677 ///    A 128-bit vector of [2 x double]. The contents of this parameter are
4678 ///    written to either the upper or the lower 128 bits of the result depending
4679 ///    on the value of parameter \a M.
4680 /// \param M
4681 ///    An immediate integer. The least significant bit determines how the values
4682 ///    from the two parameters are interleaved: \n
4683 ///    If bit [0] of \a M is 0, \a V2 are copied to bits [127:0] of the result,
4684 ///    and bits [255:128] of \a V1 are copied to bits [255:128] of the
4685 ///    result. \n
4686 ///    If bit [0] of \a M is 1, \a V2 are copied to bits [255:128] of the
4687 ///    result, and bits [127:0] of \a V1 are copied to bits [127:0] of the
4688 ///    result.
4689 /// \returns A 256-bit vector of [4 x double] containing the interleaved values.
4690 #define _mm256_insertf128_pd(V1, V2, M) \
4691   ((__m256d)__builtin_ia32_vinsertf128_pd256((__v4df)(__m256d)(V1), \
4692                                              (__v2df)(__m128d)(V2), (int)(M)))
4693 
4694 /// Constructs a new 256-bit integer vector by first duplicating a
4695 ///    256-bit integer vector given in the first parameter, and then replacing
4696 ///    either the upper or the lower 128 bits with the contents of a 128-bit
4697 ///    integer vector in the second parameter.
4698 ///
4699 ///    The immediate integer parameter determines between the upper or the lower
4700 ///    128 bits.
4701 ///
4702 /// \headerfile <x86intrin.h>
4703 ///
4704 /// \code
4705 /// __m256i _mm256_insertf128_si256(__m256i V1, __m128i V2, const int M);
4706 /// \endcode
4707 ///
4708 /// This intrinsic corresponds to the <c> VINSERTF128 </c> instruction.
4709 ///
4710 /// \param V1
4711 ///    A 256-bit integer vector. This vector is copied to the result first, and
4712 ///    then either the upper or the lower 128 bits of the result will be
4713 ///    replaced by the contents of \a V2.
4714 /// \param V2
4715 ///    A 128-bit integer vector. The contents of this parameter are written to
4716 ///    either the upper or the lower 128 bits of the result depending on the
4717 ///     value of parameter \a M.
4718 /// \param M
4719 ///    An immediate integer. The least significant bit determines how the values
4720 ///    from the two parameters are interleaved: \n
4721 ///    If bit [0] of \a M is 0, \a V2 are copied to bits [127:0] of the result,
4722 ///    and bits [255:128] of \a V1 are copied to bits [255:128] of the
4723 ///    result. \n
4724 ///    If bit [0] of \a M is 1, \a V2 are copied to bits [255:128] of the
4725 ///    result, and bits [127:0] of \a V1 are copied to bits [127:0] of the
4726 ///    result.
4727 /// \returns A 256-bit integer vector containing the interleaved values.
4728 #define _mm256_insertf128_si256(V1, V2, M) \
4729   ((__m256i)__builtin_ia32_vinsertf128_si256((__v8si)(__m256i)(V1), \
4730                                              (__v4si)(__m128i)(V2), (int)(M)))
4731 
4732 /*
4733    Vector extract.
4734    We use macros rather than inlines because we only want to accept
4735    invocations where the immediate M is a constant expression.
4736 */
4737 /// Extracts either the upper or the lower 128 bits from a 256-bit vector
4738 ///    of [8 x float], as determined by the immediate integer parameter, and
4739 ///    returns the extracted bits as a 128-bit vector of [4 x float].
4740 ///
4741 /// \headerfile <x86intrin.h>
4742 ///
4743 /// \code
4744 /// __m128 _mm256_extractf128_ps(__m256 V, const int M);
4745 /// \endcode
4746 ///
4747 /// This intrinsic corresponds to the <c> VEXTRACTF128 </c> instruction.
4748 ///
4749 /// \param V
4750 ///    A 256-bit vector of [8 x float].
4751 /// \param M
4752 ///    An immediate integer. The least significant bit determines which bits are
4753 ///    extracted from the first parameter: \n
4754 ///    If bit [0] of \a M is 0, bits [127:0] of \a V are copied to the
4755 ///    result. \n
4756 ///    If bit [0] of \a M is 1, bits [255:128] of \a V are copied to the result.
4757 /// \returns A 128-bit vector of [4 x float] containing the extracted bits.
4758 #define _mm256_extractf128_ps(V, M) \
4759   ((__m128)__builtin_ia32_vextractf128_ps256((__v8sf)(__m256)(V), (int)(M)))
4760 
4761 /// Extracts either the upper or the lower 128 bits from a 256-bit vector
4762 ///    of [4 x double], as determined by the immediate integer parameter, and
4763 ///    returns the extracted bits as a 128-bit vector of [2 x double].
4764 ///
4765 /// \headerfile <x86intrin.h>
4766 ///
4767 /// \code
4768 /// __m128d _mm256_extractf128_pd(__m256d V, const int M);
4769 /// \endcode
4770 ///
4771 /// This intrinsic corresponds to the <c> VEXTRACTF128 </c> instruction.
4772 ///
4773 /// \param V
4774 ///    A 256-bit vector of [4 x double].
4775 /// \param M
4776 ///    An immediate integer. The least significant bit determines which bits are
4777 ///    extracted from the first parameter: \n
4778 ///    If bit [0] of \a M is 0, bits [127:0] of \a V are copied to the
4779 ///    result. \n
4780 ///    If bit [0] of \a M is 1, bits [255:128] of \a V are copied to the result.
4781 /// \returns A 128-bit vector of [2 x double] containing the extracted bits.
4782 #define _mm256_extractf128_pd(V, M) \
4783   ((__m128d)__builtin_ia32_vextractf128_pd256((__v4df)(__m256d)(V), (int)(M)))
4784 
4785 /// Extracts either the upper or the lower 128 bits from a 256-bit
4786 ///    integer vector, as determined by the immediate integer parameter, and
4787 ///    returns the extracted bits as a 128-bit integer vector.
4788 ///
4789 /// \headerfile <x86intrin.h>
4790 ///
4791 /// \code
4792 /// __m128i _mm256_extractf128_si256(__m256i V, const int M);
4793 /// \endcode
4794 ///
4795 /// This intrinsic corresponds to the <c> VEXTRACTF128 </c> instruction.
4796 ///
4797 /// \param V
4798 ///    A 256-bit integer vector.
4799 /// \param M
4800 ///    An immediate integer. The least significant bit determines which bits are
4801 ///    extracted from the first parameter:  \n
4802 ///    If bit [0] of \a M is 0, bits [127:0] of \a V are copied to the
4803 ///    result. \n
4804 ///    If bit [0] of \a M is 1, bits [255:128] of \a V are copied to the result.
4805 /// \returns A 128-bit integer vector containing the extracted bits.
4806 #define _mm256_extractf128_si256(V, M) \
4807   ((__m128i)__builtin_ia32_vextractf128_si256((__v8si)(__m256i)(V), (int)(M)))
4808 
4809 /// Constructs a 256-bit floating-point vector of [8 x float] by
4810 ///    concatenating two 128-bit floating-point vectors of [4 x float].
4811 ///
4812 /// \headerfile <x86intrin.h>
4813 ///
4814 /// This intrinsic corresponds to the <c> VINSERTF128 </c> instruction.
4815 ///
4816 /// \param __hi
4817 ///    A 128-bit floating-point vector of [4 x float] to be copied to the upper
4818 ///    128 bits of the result.
4819 /// \param __lo
4820 ///    A 128-bit floating-point vector of [4 x float] to be copied to the lower
4821 ///    128 bits of the result.
4822 /// \returns A 256-bit floating-point vector of [8 x float] containing the
4823 ///    concatenated result.
4824 static __inline __m256 __DEFAULT_FN_ATTRS
4825 _mm256_set_m128 (__m128 __hi, __m128 __lo)
4826 {
4827   return (__m256) __builtin_shufflevector((__v4sf)__lo, (__v4sf)__hi, 0, 1, 2, 3, 4, 5, 6, 7);
4828 }
4829 
4830 /// Constructs a 256-bit floating-point vector of [4 x double] by
4831 ///    concatenating two 128-bit floating-point vectors of [2 x double].
4832 ///
4833 /// \headerfile <x86intrin.h>
4834 ///
4835 /// This intrinsic corresponds to the <c> VINSERTF128 </c> instruction.
4836 ///
4837 /// \param __hi
4838 ///    A 128-bit floating-point vector of [2 x double] to be copied to the upper
4839 ///    128 bits of the result.
4840 /// \param __lo
4841 ///    A 128-bit floating-point vector of [2 x double] to be copied to the lower
4842 ///    128 bits of the result.
4843 /// \returns A 256-bit floating-point vector of [4 x double] containing the
4844 ///    concatenated result.
4845 static __inline __m256d __DEFAULT_FN_ATTRS
4846 _mm256_set_m128d (__m128d __hi, __m128d __lo)
4847 {
4848   return (__m256d) __builtin_shufflevector((__v2df)__lo, (__v2df)__hi, 0, 1, 2, 3);
4849 }
4850 
4851 /// Constructs a 256-bit integer vector by concatenating two 128-bit
4852 ///    integer vectors.
4853 ///
4854 /// \headerfile <x86intrin.h>
4855 ///
4856 /// This intrinsic corresponds to the <c> VINSERTF128 </c> instruction.
4857 ///
4858 /// \param __hi
4859 ///    A 128-bit integer vector to be copied to the upper 128 bits of the
4860 ///    result.
4861 /// \param __lo
4862 ///    A 128-bit integer vector to be copied to the lower 128 bits of the
4863 ///    result.
4864 /// \returns A 256-bit integer vector containing the concatenated result.
4865 static __inline __m256i __DEFAULT_FN_ATTRS
4866 _mm256_set_m128i (__m128i __hi, __m128i __lo)
4867 {
4868   return (__m256i) __builtin_shufflevector((__v2di)__lo, (__v2di)__hi, 0, 1, 2, 3);
4869 }
4870 
4871 /// Constructs a 256-bit floating-point vector of [8 x float] by
4872 ///    concatenating two 128-bit floating-point vectors of [4 x float]. This is
4873 ///    similar to _mm256_set_m128, but the order of the input parameters is
4874 ///    swapped.
4875 ///
4876 /// \headerfile <x86intrin.h>
4877 ///
4878 /// This intrinsic corresponds to the <c> VINSERTF128 </c> instruction.
4879 ///
4880 /// \param __lo
4881 ///    A 128-bit floating-point vector of [4 x float] to be copied to the lower
4882 ///    128 bits of the result.
4883 /// \param __hi
4884 ///    A 128-bit floating-point vector of [4 x float] to be copied to the upper
4885 ///    128 bits of the result.
4886 /// \returns A 256-bit floating-point vector of [8 x float] containing the
4887 ///    concatenated result.
4888 static __inline __m256 __DEFAULT_FN_ATTRS
4889 _mm256_setr_m128 (__m128 __lo, __m128 __hi)
4890 {
4891   return _mm256_set_m128(__hi, __lo);
4892 }
4893 
4894 /// Constructs a 256-bit floating-point vector of [4 x double] by
4895 ///    concatenating two 128-bit floating-point vectors of [2 x double]. This is
4896 ///    similar to _mm256_set_m128d, but the order of the input parameters is
4897 ///    swapped.
4898 ///
4899 /// \headerfile <x86intrin.h>
4900 ///
4901 /// This intrinsic corresponds to the <c> VINSERTF128 </c> instruction.
4902 ///
4903 /// \param __lo
4904 ///    A 128-bit floating-point vector of [2 x double] to be copied to the lower
4905 ///    128 bits of the result.
4906 /// \param __hi
4907 ///    A 128-bit floating-point vector of [2 x double] to be copied to the upper
4908 ///    128 bits of the result.
4909 /// \returns A 256-bit floating-point vector of [4 x double] containing the
4910 ///    concatenated result.
4911 static __inline __m256d __DEFAULT_FN_ATTRS
4912 _mm256_setr_m128d (__m128d __lo, __m128d __hi)
4913 {
4914   return (__m256d)_mm256_set_m128d(__hi, __lo);
4915 }
4916 
4917 /// Constructs a 256-bit integer vector by concatenating two 128-bit
4918 ///    integer vectors. This is similar to _mm256_set_m128i, but the order of
4919 ///    the input parameters is swapped.
4920 ///
4921 /// \headerfile <x86intrin.h>
4922 ///
4923 /// This intrinsic corresponds to the <c> VINSERTF128 </c> instruction.
4924 ///
4925 /// \param __lo
4926 ///    A 128-bit integer vector to be copied to the lower 128 bits of the
4927 ///    result.
4928 /// \param __hi
4929 ///    A 128-bit integer vector to be copied to the upper 128 bits of the
4930 ///    result.
4931 /// \returns A 256-bit integer vector containing the concatenated result.
4932 static __inline __m256i __DEFAULT_FN_ATTRS
4933 _mm256_setr_m128i (__m128i __lo, __m128i __hi)
4934 {
4935   return (__m256i)_mm256_set_m128i(__hi, __lo);
4936 }
4937 
4938 /* SIMD load ops (unaligned) */
4939 /// Loads two 128-bit floating-point vectors of [4 x float] from
4940 ///    unaligned memory locations and constructs a 256-bit floating-point vector
4941 ///    of [8 x float] by concatenating the two 128-bit vectors.
4942 ///
4943 /// \headerfile <x86intrin.h>
4944 ///
4945 /// This intrinsic corresponds to load instructions followed by the
4946 ///   <c> VINSERTF128 </c> instruction.
4947 ///
4948 /// \param __addr_hi
4949 ///    A pointer to a 128-bit memory location containing 4 consecutive
4950 ///    single-precision floating-point values. These values are to be copied to
4951 ///    bits[255:128] of the result. The address of the memory location does not
4952 ///    have to be aligned.
4953 /// \param __addr_lo
4954 ///    A pointer to a 128-bit memory location containing 4 consecutive
4955 ///    single-precision floating-point values. These values are to be copied to
4956 ///    bits[127:0] of the result. The address of the memory location does not
4957 ///    have to be aligned.
4958 /// \returns A 256-bit floating-point vector of [8 x float] containing the
4959 ///    concatenated result.
4960 static __inline __m256 __DEFAULT_FN_ATTRS
4961 _mm256_loadu2_m128(float const *__addr_hi, float const *__addr_lo)
4962 {
4963   return _mm256_set_m128(_mm_loadu_ps(__addr_hi), _mm_loadu_ps(__addr_lo));
4964 }
4965 
4966 /// Loads two 128-bit floating-point vectors of [2 x double] from
4967 ///    unaligned memory locations and constructs a 256-bit floating-point vector
4968 ///    of [4 x double] by concatenating the two 128-bit vectors.
4969 ///
4970 /// \headerfile <x86intrin.h>
4971 ///
4972 /// This intrinsic corresponds to load instructions followed by the
4973 ///   <c> VINSERTF128 </c> instruction.
4974 ///
4975 /// \param __addr_hi
4976 ///    A pointer to a 128-bit memory location containing two consecutive
4977 ///    double-precision floating-point values. These values are to be copied to
4978 ///    bits[255:128] of the result. The address of the memory location does not
4979 ///    have to be aligned.
4980 /// \param __addr_lo
4981 ///    A pointer to a 128-bit memory location containing two consecutive
4982 ///    double-precision floating-point values. These values are to be copied to
4983 ///    bits[127:0] of the result. The address of the memory location does not
4984 ///    have to be aligned.
4985 /// \returns A 256-bit floating-point vector of [4 x double] containing the
4986 ///    concatenated result.
4987 static __inline __m256d __DEFAULT_FN_ATTRS
4988 _mm256_loadu2_m128d(double const *__addr_hi, double const *__addr_lo)
4989 {
4990   return _mm256_set_m128d(_mm_loadu_pd(__addr_hi), _mm_loadu_pd(__addr_lo));
4991 }
4992 
4993 /// Loads two 128-bit integer vectors from unaligned memory locations and
4994 ///    constructs a 256-bit integer vector by concatenating the two 128-bit
4995 ///    vectors.
4996 ///
4997 /// \headerfile <x86intrin.h>
4998 ///
4999 /// This intrinsic corresponds to load instructions followed by the
5000 ///   <c> VINSERTF128 </c> instruction.
5001 ///
5002 /// \param __addr_hi
5003 ///    A pointer to a 128-bit memory location containing a 128-bit integer
5004 ///    vector. This vector is to be copied to bits[255:128] of the result. The
5005 ///    address of the memory location does not have to be aligned.
5006 /// \param __addr_lo
5007 ///    A pointer to a 128-bit memory location containing a 128-bit integer
5008 ///    vector. This vector is to be copied to bits[127:0] of the result. The
5009 ///    address of the memory location does not have to be aligned.
5010 /// \returns A 256-bit integer vector containing the concatenated result.
5011 static __inline __m256i __DEFAULT_FN_ATTRS
5012 _mm256_loadu2_m128i(__m128i_u const *__addr_hi, __m128i_u const *__addr_lo)
5013 {
5014    return _mm256_set_m128i(_mm_loadu_si128(__addr_hi), _mm_loadu_si128(__addr_lo));
5015 }
5016 
5017 /* SIMD store ops (unaligned) */
5018 /// Stores the upper and lower 128 bits of a 256-bit floating-point
5019 ///    vector of [8 x float] into two different unaligned memory locations.
5020 ///
5021 /// \headerfile <x86intrin.h>
5022 ///
5023 /// This intrinsic corresponds to the <c> VEXTRACTF128 </c> instruction and the
5024 ///   store instructions.
5025 ///
5026 /// \param __addr_hi
5027 ///    A pointer to a 128-bit memory location. Bits[255:128] of \a __a are to be
5028 ///    copied to this memory location. The address of this memory location does
5029 ///    not have to be aligned.
5030 /// \param __addr_lo
5031 ///    A pointer to a 128-bit memory location. Bits[127:0] of \a __a are to be
5032 ///    copied to this memory location. The address of this memory location does
5033 ///    not have to be aligned.
5034 /// \param __a
5035 ///    A 256-bit floating-point vector of [8 x float].
5036 static __inline void __DEFAULT_FN_ATTRS
5037 _mm256_storeu2_m128(float *__addr_hi, float *__addr_lo, __m256 __a)
5038 {
5039   __m128 __v128;
5040 
5041   __v128 = _mm256_castps256_ps128(__a);
5042   _mm_storeu_ps(__addr_lo, __v128);
5043   __v128 = _mm256_extractf128_ps(__a, 1);
5044   _mm_storeu_ps(__addr_hi, __v128);
5045 }
5046 
5047 /// Stores the upper and lower 128 bits of a 256-bit floating-point
5048 ///    vector of [4 x double] into two different unaligned memory locations.
5049 ///
5050 /// \headerfile <x86intrin.h>
5051 ///
5052 /// This intrinsic corresponds to the <c> VEXTRACTF128 </c> instruction and the
5053 ///   store instructions.
5054 ///
5055 /// \param __addr_hi
5056 ///    A pointer to a 128-bit memory location. Bits[255:128] of \a __a are to be
5057 ///    copied to this memory location. The address of this memory location does
5058 ///    not have to be aligned.
5059 /// \param __addr_lo
5060 ///    A pointer to a 128-bit memory location. Bits[127:0] of \a __a are to be
5061 ///    copied to this memory location. The address of this memory location does
5062 ///    not have to be aligned.
5063 /// \param __a
5064 ///    A 256-bit floating-point vector of [4 x double].
5065 static __inline void __DEFAULT_FN_ATTRS
5066 _mm256_storeu2_m128d(double *__addr_hi, double *__addr_lo, __m256d __a)
5067 {
5068   __m128d __v128;
5069 
5070   __v128 = _mm256_castpd256_pd128(__a);
5071   _mm_storeu_pd(__addr_lo, __v128);
5072   __v128 = _mm256_extractf128_pd(__a, 1);
5073   _mm_storeu_pd(__addr_hi, __v128);
5074 }
5075 
5076 /// Stores the upper and lower 128 bits of a 256-bit integer vector into
5077 ///    two different unaligned memory locations.
5078 ///
5079 /// \headerfile <x86intrin.h>
5080 ///
5081 /// This intrinsic corresponds to the <c> VEXTRACTF128 </c> instruction and the
5082 ///   store instructions.
5083 ///
5084 /// \param __addr_hi
5085 ///    A pointer to a 128-bit memory location. Bits[255:128] of \a __a are to be
5086 ///    copied to this memory location. The address of this memory location does
5087 ///    not have to be aligned.
5088 /// \param __addr_lo
5089 ///    A pointer to a 128-bit memory location. Bits[127:0] of \a __a are to be
5090 ///    copied to this memory location. The address of this memory location does
5091 ///    not have to be aligned.
5092 /// \param __a
5093 ///    A 256-bit integer vector.
5094 static __inline void __DEFAULT_FN_ATTRS
5095 _mm256_storeu2_m128i(__m128i_u *__addr_hi, __m128i_u *__addr_lo, __m256i __a)
5096 {
5097   __m128i __v128;
5098 
5099   __v128 = _mm256_castsi256_si128(__a);
5100   _mm_storeu_si128(__addr_lo, __v128);
5101   __v128 = _mm256_extractf128_si256(__a, 1);
5102   _mm_storeu_si128(__addr_hi, __v128);
5103 }
5104 
5105 #undef __DEFAULT_FN_ATTRS
5106 #undef __DEFAULT_FN_ATTRS128
5107 
5108 #endif /* __AVXINTRIN_H */
5109