xref: /freebsd/contrib/llvm-project/clang/lib/Headers/avxintrin.h (revision 5ffd83dbcc34f10e07f6d3e968ae6365869615f4)
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 /* Define the default attributes for the functions in this file. */
43 #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx"), __min_vector_width__(256)))
44 #define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx"), __min_vector_width__(128)))
45 
46 /* Arithmetic */
47 /// Adds two 256-bit vectors of [4 x double].
48 ///
49 /// \headerfile <x86intrin.h>
50 ///
51 /// This intrinsic corresponds to the <c> VADDPD </c> instruction.
52 ///
53 /// \param __a
54 ///    A 256-bit vector of [4 x double] containing one of the source operands.
55 /// \param __b
56 ///    A 256-bit vector of [4 x double] containing one of the source operands.
57 /// \returns A 256-bit vector of [4 x double] containing the sums of both
58 ///    operands.
59 static __inline __m256d __DEFAULT_FN_ATTRS
60 _mm256_add_pd(__m256d __a, __m256d __b)
61 {
62   return (__m256d)((__v4df)__a+(__v4df)__b);
63 }
64 
65 /// Adds two 256-bit vectors of [8 x float].
66 ///
67 /// \headerfile <x86intrin.h>
68 ///
69 /// This intrinsic corresponds to the <c> VADDPS </c> instruction.
70 ///
71 /// \param __a
72 ///    A 256-bit vector of [8 x float] containing one of the source operands.
73 /// \param __b
74 ///    A 256-bit vector of [8 x float] containing one of the source operands.
75 /// \returns A 256-bit vector of [8 x float] containing the sums of both
76 ///    operands.
77 static __inline __m256 __DEFAULT_FN_ATTRS
78 _mm256_add_ps(__m256 __a, __m256 __b)
79 {
80   return (__m256)((__v8sf)__a+(__v8sf)__b);
81 }
82 
83 /// Subtracts two 256-bit vectors of [4 x double].
84 ///
85 /// \headerfile <x86intrin.h>
86 ///
87 /// This intrinsic corresponds to the <c> VSUBPD </c> instruction.
88 ///
89 /// \param __a
90 ///    A 256-bit vector of [4 x double] containing the minuend.
91 /// \param __b
92 ///    A 256-bit vector of [4 x double] containing the subtrahend.
93 /// \returns A 256-bit vector of [4 x double] containing the differences between
94 ///    both operands.
95 static __inline __m256d __DEFAULT_FN_ATTRS
96 _mm256_sub_pd(__m256d __a, __m256d __b)
97 {
98   return (__m256d)((__v4df)__a-(__v4df)__b);
99 }
100 
101 /// Subtracts two 256-bit vectors of [8 x float].
102 ///
103 /// \headerfile <x86intrin.h>
104 ///
105 /// This intrinsic corresponds to the <c> VSUBPS </c> instruction.
106 ///
107 /// \param __a
108 ///    A 256-bit vector of [8 x float] containing the minuend.
109 /// \param __b
110 ///    A 256-bit vector of [8 x float] containing the subtrahend.
111 /// \returns A 256-bit vector of [8 x float] containing the differences between
112 ///    both operands.
113 static __inline __m256 __DEFAULT_FN_ATTRS
114 _mm256_sub_ps(__m256 __a, __m256 __b)
115 {
116   return (__m256)((__v8sf)__a-(__v8sf)__b);
117 }
118 
119 /// Adds the even-indexed values and subtracts the odd-indexed values of
120 ///    two 256-bit vectors of [4 x double].
121 ///
122 /// \headerfile <x86intrin.h>
123 ///
124 /// This intrinsic corresponds to the <c> VADDSUBPD </c> instruction.
125 ///
126 /// \param __a
127 ///    A 256-bit vector of [4 x double] containing the left source operand.
128 /// \param __b
129 ///    A 256-bit vector of [4 x double] containing the right source operand.
130 /// \returns A 256-bit vector of [4 x double] containing the alternating sums
131 ///    and differences between both operands.
132 static __inline __m256d __DEFAULT_FN_ATTRS
133 _mm256_addsub_pd(__m256d __a, __m256d __b)
134 {
135   return (__m256d)__builtin_ia32_addsubpd256((__v4df)__a, (__v4df)__b);
136 }
137 
138 /// Adds the even-indexed values and subtracts the odd-indexed values of
139 ///    two 256-bit vectors of [8 x float].
140 ///
141 /// \headerfile <x86intrin.h>
142 ///
143 /// This intrinsic corresponds to the <c> VADDSUBPS </c> instruction.
144 ///
145 /// \param __a
146 ///    A 256-bit vector of [8 x float] containing the left source operand.
147 /// \param __b
148 ///    A 256-bit vector of [8 x float] containing the right source operand.
149 /// \returns A 256-bit vector of [8 x float] containing the alternating sums and
150 ///    differences between both operands.
151 static __inline __m256 __DEFAULT_FN_ATTRS
152 _mm256_addsub_ps(__m256 __a, __m256 __b)
153 {
154   return (__m256)__builtin_ia32_addsubps256((__v8sf)__a, (__v8sf)__b);
155 }
156 
157 /// Divides two 256-bit vectors of [4 x double].
158 ///
159 /// \headerfile <x86intrin.h>
160 ///
161 /// This intrinsic corresponds to the <c> VDIVPD </c> instruction.
162 ///
163 /// \param __a
164 ///    A 256-bit vector of [4 x double] containing the dividend.
165 /// \param __b
166 ///    A 256-bit vector of [4 x double] containing the divisor.
167 /// \returns A 256-bit vector of [4 x double] containing the quotients of both
168 ///    operands.
169 static __inline __m256d __DEFAULT_FN_ATTRS
170 _mm256_div_pd(__m256d __a, __m256d __b)
171 {
172   return (__m256d)((__v4df)__a/(__v4df)__b);
173 }
174 
175 /// Divides two 256-bit vectors of [8 x float].
176 ///
177 /// \headerfile <x86intrin.h>
178 ///
179 /// This intrinsic corresponds to the <c> VDIVPS </c> instruction.
180 ///
181 /// \param __a
182 ///    A 256-bit vector of [8 x float] containing the dividend.
183 /// \param __b
184 ///    A 256-bit vector of [8 x float] containing the divisor.
185 /// \returns A 256-bit vector of [8 x float] containing the quotients of both
186 ///    operands.
187 static __inline __m256 __DEFAULT_FN_ATTRS
188 _mm256_div_ps(__m256 __a, __m256 __b)
189 {
190   return (__m256)((__v8sf)__a/(__v8sf)__b);
191 }
192 
193 /// Compares two 256-bit vectors of [4 x double] and returns the greater
194 ///    of each pair of values.
195 ///
196 /// \headerfile <x86intrin.h>
197 ///
198 /// This intrinsic corresponds to the <c> VMAXPD </c> instruction.
199 ///
200 /// \param __a
201 ///    A 256-bit vector of [4 x double] containing one of the operands.
202 /// \param __b
203 ///    A 256-bit vector of [4 x double] containing one of the operands.
204 /// \returns A 256-bit vector of [4 x double] containing the maximum values
205 ///    between both operands.
206 static __inline __m256d __DEFAULT_FN_ATTRS
207 _mm256_max_pd(__m256d __a, __m256d __b)
208 {
209   return (__m256d)__builtin_ia32_maxpd256((__v4df)__a, (__v4df)__b);
210 }
211 
212 /// Compares two 256-bit vectors of [8 x float] and returns the greater
213 ///    of each pair of values.
214 ///
215 /// \headerfile <x86intrin.h>
216 ///
217 /// This intrinsic corresponds to the <c> VMAXPS </c> instruction.
218 ///
219 /// \param __a
220 ///    A 256-bit vector of [8 x float] containing one of the operands.
221 /// \param __b
222 ///    A 256-bit vector of [8 x float] containing one of the operands.
223 /// \returns A 256-bit vector of [8 x float] containing the maximum values
224 ///    between both operands.
225 static __inline __m256 __DEFAULT_FN_ATTRS
226 _mm256_max_ps(__m256 __a, __m256 __b)
227 {
228   return (__m256)__builtin_ia32_maxps256((__v8sf)__a, (__v8sf)__b);
229 }
230 
231 /// Compares two 256-bit vectors of [4 x double] and returns the lesser
232 ///    of each pair of values.
233 ///
234 /// \headerfile <x86intrin.h>
235 ///
236 /// This intrinsic corresponds to the <c> VMINPD </c> instruction.
237 ///
238 /// \param __a
239 ///    A 256-bit vector of [4 x double] containing one of the operands.
240 /// \param __b
241 ///    A 256-bit vector of [4 x double] containing one of the operands.
242 /// \returns A 256-bit vector of [4 x double] containing the minimum values
243 ///    between both operands.
244 static __inline __m256d __DEFAULT_FN_ATTRS
245 _mm256_min_pd(__m256d __a, __m256d __b)
246 {
247   return (__m256d)__builtin_ia32_minpd256((__v4df)__a, (__v4df)__b);
248 }
249 
250 /// Compares two 256-bit vectors of [8 x float] and returns the lesser
251 ///    of each pair of values.
252 ///
253 /// \headerfile <x86intrin.h>
254 ///
255 /// This intrinsic corresponds to the <c> VMINPS </c> instruction.
256 ///
257 /// \param __a
258 ///    A 256-bit vector of [8 x float] containing one of the operands.
259 /// \param __b
260 ///    A 256-bit vector of [8 x float] containing one of the operands.
261 /// \returns A 256-bit vector of [8 x float] containing the minimum values
262 ///    between both operands.
263 static __inline __m256 __DEFAULT_FN_ATTRS
264 _mm256_min_ps(__m256 __a, __m256 __b)
265 {
266   return (__m256)__builtin_ia32_minps256((__v8sf)__a, (__v8sf)__b);
267 }
268 
269 /// Multiplies two 256-bit vectors of [4 x double].
270 ///
271 /// \headerfile <x86intrin.h>
272 ///
273 /// This intrinsic corresponds to the <c> VMULPD </c> instruction.
274 ///
275 /// \param __a
276 ///    A 256-bit vector of [4 x double] containing one of the operands.
277 /// \param __b
278 ///    A 256-bit vector of [4 x double] containing one of the operands.
279 /// \returns A 256-bit vector of [4 x double] containing the products of both
280 ///    operands.
281 static __inline __m256d __DEFAULT_FN_ATTRS
282 _mm256_mul_pd(__m256d __a, __m256d __b)
283 {
284   return (__m256d)((__v4df)__a * (__v4df)__b);
285 }
286 
287 /// Multiplies two 256-bit vectors of [8 x float].
288 ///
289 /// \headerfile <x86intrin.h>
290 ///
291 /// This intrinsic corresponds to the <c> VMULPS </c> instruction.
292 ///
293 /// \param __a
294 ///    A 256-bit vector of [8 x float] containing one of the operands.
295 /// \param __b
296 ///    A 256-bit vector of [8 x float] containing one of the operands.
297 /// \returns A 256-bit vector of [8 x float] containing the products of both
298 ///    operands.
299 static __inline __m256 __DEFAULT_FN_ATTRS
300 _mm256_mul_ps(__m256 __a, __m256 __b)
301 {
302   return (__m256)((__v8sf)__a * (__v8sf)__b);
303 }
304 
305 /// Calculates the square roots of the values in a 256-bit vector of
306 ///    [4 x double].
307 ///
308 /// \headerfile <x86intrin.h>
309 ///
310 /// This intrinsic corresponds to the <c> VSQRTPD </c> instruction.
311 ///
312 /// \param __a
313 ///    A 256-bit vector of [4 x double].
314 /// \returns A 256-bit vector of [4 x double] containing the square roots of the
315 ///    values in the operand.
316 static __inline __m256d __DEFAULT_FN_ATTRS
317 _mm256_sqrt_pd(__m256d __a)
318 {
319   return (__m256d)__builtin_ia32_sqrtpd256((__v4df)__a);
320 }
321 
322 /// Calculates the square roots of the values in a 256-bit vector of
323 ///    [8 x float].
324 ///
325 /// \headerfile <x86intrin.h>
326 ///
327 /// This intrinsic corresponds to the <c> VSQRTPS </c> instruction.
328 ///
329 /// \param __a
330 ///    A 256-bit vector of [8 x float].
331 /// \returns A 256-bit vector of [8 x float] containing the square roots of the
332 ///    values in the operand.
333 static __inline __m256 __DEFAULT_FN_ATTRS
334 _mm256_sqrt_ps(__m256 __a)
335 {
336   return (__m256)__builtin_ia32_sqrtps256((__v8sf)__a);
337 }
338 
339 /// Calculates the reciprocal square roots of the values in a 256-bit
340 ///    vector of [8 x float].
341 ///
342 /// \headerfile <x86intrin.h>
343 ///
344 /// This intrinsic corresponds to the <c> VRSQRTPS </c> instruction.
345 ///
346 /// \param __a
347 ///    A 256-bit vector of [8 x float].
348 /// \returns A 256-bit vector of [8 x float] containing the reciprocal square
349 ///    roots of the values in the operand.
350 static __inline __m256 __DEFAULT_FN_ATTRS
351 _mm256_rsqrt_ps(__m256 __a)
352 {
353   return (__m256)__builtin_ia32_rsqrtps256((__v8sf)__a);
354 }
355 
356 /// Calculates the reciprocals of the values in a 256-bit vector of
357 ///    [8 x float].
358 ///
359 /// \headerfile <x86intrin.h>
360 ///
361 /// This intrinsic corresponds to the <c> VRCPPS </c> instruction.
362 ///
363 /// \param __a
364 ///    A 256-bit vector of [8 x float].
365 /// \returns A 256-bit vector of [8 x float] containing the reciprocals of the
366 ///    values in the operand.
367 static __inline __m256 __DEFAULT_FN_ATTRS
368 _mm256_rcp_ps(__m256 __a)
369 {
370   return (__m256)__builtin_ia32_rcpps256((__v8sf)__a);
371 }
372 
373 /// Rounds the values in a 256-bit vector of [4 x double] as specified
374 ///    by the byte operand. The source values are rounded to integer values and
375 ///    returned as 64-bit double-precision floating-point values.
376 ///
377 /// \headerfile <x86intrin.h>
378 ///
379 /// \code
380 /// __m256d _mm256_round_pd(__m256d V, const int M);
381 /// \endcode
382 ///
383 /// This intrinsic corresponds to the <c> VROUNDPD </c> instruction.
384 ///
385 /// \param V
386 ///    A 256-bit vector of [4 x double].
387 /// \param M
388 ///    An integer value that specifies the rounding operation. \n
389 ///    Bits [7:4] are reserved. \n
390 ///    Bit [3] is a precision exception value: \n
391 ///      0: A normal PE exception is used. \n
392 ///      1: The PE field is not updated. \n
393 ///    Bit [2] is the rounding control source: \n
394 ///      0: Use bits [1:0] of \a M. \n
395 ///      1: Use the current MXCSR setting. \n
396 ///    Bits [1:0] contain the rounding control definition: \n
397 ///      00: Nearest. \n
398 ///      01: Downward (toward negative infinity). \n
399 ///      10: Upward (toward positive infinity). \n
400 ///      11: Truncated.
401 /// \returns A 256-bit vector of [4 x double] containing the rounded values.
402 #define _mm256_round_pd(V, M) \
403     (__m256d)__builtin_ia32_roundpd256((__v4df)(__m256d)(V), (M))
404 
405 /// Rounds the values stored in a 256-bit vector of [8 x float] as
406 ///    specified by the byte operand. The source values are rounded to integer
407 ///    values and returned as floating-point values.
408 ///
409 /// \headerfile <x86intrin.h>
410 ///
411 /// \code
412 /// __m256 _mm256_round_ps(__m256 V, const int M);
413 /// \endcode
414 ///
415 /// This intrinsic corresponds to the <c> VROUNDPS </c> instruction.
416 ///
417 /// \param V
418 ///    A 256-bit vector of [8 x float].
419 /// \param M
420 ///    An integer value that specifies the rounding operation. \n
421 ///    Bits [7:4] are reserved. \n
422 ///    Bit [3] is a precision exception value: \n
423 ///      0: A normal PE exception is used. \n
424 ///      1: The PE field is not updated. \n
425 ///    Bit [2] is the rounding control source: \n
426 ///      0: Use bits [1:0] of \a M. \n
427 ///      1: Use the current MXCSR setting. \n
428 ///    Bits [1:0] contain the rounding control definition: \n
429 ///      00: Nearest. \n
430 ///      01: Downward (toward negative infinity). \n
431 ///      10: Upward (toward positive infinity). \n
432 ///      11: Truncated.
433 /// \returns A 256-bit vector of [8 x float] containing the rounded values.
434 #define _mm256_round_ps(V, M) \
435   (__m256)__builtin_ia32_roundps256((__v8sf)(__m256)(V), (M))
436 
437 /// Rounds up the values stored in a 256-bit vector of [4 x double]. The
438 ///    source values are rounded up to integer values and returned as 64-bit
439 ///    double-precision floating-point values.
440 ///
441 /// \headerfile <x86intrin.h>
442 ///
443 /// \code
444 /// __m256d _mm256_ceil_pd(__m256d V);
445 /// \endcode
446 ///
447 /// This intrinsic corresponds to the <c> VROUNDPD </c> instruction.
448 ///
449 /// \param V
450 ///    A 256-bit vector of [4 x double].
451 /// \returns A 256-bit vector of [4 x double] containing the rounded up values.
452 #define _mm256_ceil_pd(V)  _mm256_round_pd((V), _MM_FROUND_CEIL)
453 
454 /// Rounds down the values stored in a 256-bit vector of [4 x double].
455 ///    The source values are rounded down to integer values and returned as
456 ///    64-bit double-precision floating-point values.
457 ///
458 /// \headerfile <x86intrin.h>
459 ///
460 /// \code
461 /// __m256d _mm256_floor_pd(__m256d V);
462 /// \endcode
463 ///
464 /// This intrinsic corresponds to the <c> VROUNDPD </c> instruction.
465 ///
466 /// \param V
467 ///    A 256-bit vector of [4 x double].
468 /// \returns A 256-bit vector of [4 x double] containing the rounded down
469 ///    values.
470 #define _mm256_floor_pd(V) _mm256_round_pd((V), _MM_FROUND_FLOOR)
471 
472 /// Rounds up the values stored in a 256-bit vector of [8 x float]. The
473 ///    source values are rounded up to integer values and returned as
474 ///    floating-point values.
475 ///
476 /// \headerfile <x86intrin.h>
477 ///
478 /// \code
479 /// __m256 _mm256_ceil_ps(__m256 V);
480 /// \endcode
481 ///
482 /// This intrinsic corresponds to the <c> VROUNDPS </c> instruction.
483 ///
484 /// \param V
485 ///    A 256-bit vector of [8 x float].
486 /// \returns A 256-bit vector of [8 x float] containing the rounded up values.
487 #define _mm256_ceil_ps(V)  _mm256_round_ps((V), _MM_FROUND_CEIL)
488 
489 /// Rounds down the values stored in a 256-bit vector of [8 x float]. The
490 ///    source values are rounded down to integer values and returned as
491 ///    floating-point values.
492 ///
493 /// \headerfile <x86intrin.h>
494 ///
495 /// \code
496 /// __m256 _mm256_floor_ps(__m256 V);
497 /// \endcode
498 ///
499 /// This intrinsic corresponds to the <c> VROUNDPS </c> instruction.
500 ///
501 /// \param V
502 ///    A 256-bit vector of [8 x float].
503 /// \returns A 256-bit vector of [8 x float] containing the rounded down values.
504 #define _mm256_floor_ps(V) _mm256_round_ps((V), _MM_FROUND_FLOOR)
505 
506 /* Logical */
507 /// Performs a bitwise AND of two 256-bit vectors of [4 x double].
508 ///
509 /// \headerfile <x86intrin.h>
510 ///
511 /// This intrinsic corresponds to the <c> VANDPD </c> instruction.
512 ///
513 /// \param __a
514 ///    A 256-bit vector of [4 x double] containing one of the source operands.
515 /// \param __b
516 ///    A 256-bit vector of [4 x double] containing one of the source operands.
517 /// \returns A 256-bit vector of [4 x double] containing the bitwise AND of the
518 ///    values between both operands.
519 static __inline __m256d __DEFAULT_FN_ATTRS
520 _mm256_and_pd(__m256d __a, __m256d __b)
521 {
522   return (__m256d)((__v4du)__a & (__v4du)__b);
523 }
524 
525 /// Performs a bitwise AND of two 256-bit vectors of [8 x float].
526 ///
527 /// \headerfile <x86intrin.h>
528 ///
529 /// This intrinsic corresponds to the <c> VANDPS </c> instruction.
530 ///
531 /// \param __a
532 ///    A 256-bit vector of [8 x float] containing one of the source operands.
533 /// \param __b
534 ///    A 256-bit vector of [8 x float] containing one of the source operands.
535 /// \returns A 256-bit vector of [8 x float] containing the bitwise AND of the
536 ///    values between both operands.
537 static __inline __m256 __DEFAULT_FN_ATTRS
538 _mm256_and_ps(__m256 __a, __m256 __b)
539 {
540   return (__m256)((__v8su)__a & (__v8su)__b);
541 }
542 
543 /// Performs a bitwise AND of two 256-bit vectors of [4 x double], using
544 ///    the one's complement of the values contained in the first source operand.
545 ///
546 /// \headerfile <x86intrin.h>
547 ///
548 /// This intrinsic corresponds to the <c> VANDNPD </c> instruction.
549 ///
550 /// \param __a
551 ///    A 256-bit vector of [4 x double] containing the left source operand. The
552 ///    one's complement of this value is used in the bitwise AND.
553 /// \param __b
554 ///    A 256-bit vector of [4 x double] containing the right source operand.
555 /// \returns A 256-bit vector of [4 x double] containing the bitwise AND of the
556 ///    values of the second operand and the one's complement of the first
557 ///    operand.
558 static __inline __m256d __DEFAULT_FN_ATTRS
559 _mm256_andnot_pd(__m256d __a, __m256d __b)
560 {
561   return (__m256d)(~(__v4du)__a & (__v4du)__b);
562 }
563 
564 /// Performs a bitwise AND of two 256-bit vectors of [8 x float], using
565 ///    the one's complement of the values contained in the first source operand.
566 ///
567 /// \headerfile <x86intrin.h>
568 ///
569 /// This intrinsic corresponds to the <c> VANDNPS </c> instruction.
570 ///
571 /// \param __a
572 ///    A 256-bit vector of [8 x float] containing the left source operand. The
573 ///    one's complement of this value is used in the bitwise AND.
574 /// \param __b
575 ///    A 256-bit vector of [8 x float] containing the right source operand.
576 /// \returns A 256-bit vector of [8 x float] containing the bitwise AND of the
577 ///    values of the second operand and the one's complement of the first
578 ///    operand.
579 static __inline __m256 __DEFAULT_FN_ATTRS
580 _mm256_andnot_ps(__m256 __a, __m256 __b)
581 {
582   return (__m256)(~(__v8su)__a & (__v8su)__b);
583 }
584 
585 /// Performs a bitwise OR of two 256-bit vectors of [4 x double].
586 ///
587 /// \headerfile <x86intrin.h>
588 ///
589 /// This intrinsic corresponds to the <c> VORPD </c> instruction.
590 ///
591 /// \param __a
592 ///    A 256-bit vector of [4 x double] containing one of the source operands.
593 /// \param __b
594 ///    A 256-bit vector of [4 x double] containing one of the source operands.
595 /// \returns A 256-bit vector of [4 x double] containing the bitwise OR of the
596 ///    values between both operands.
597 static __inline __m256d __DEFAULT_FN_ATTRS
598 _mm256_or_pd(__m256d __a, __m256d __b)
599 {
600   return (__m256d)((__v4du)__a | (__v4du)__b);
601 }
602 
603 /// Performs a bitwise OR of two 256-bit vectors of [8 x float].
604 ///
605 /// \headerfile <x86intrin.h>
606 ///
607 /// This intrinsic corresponds to the <c> VORPS </c> instruction.
608 ///
609 /// \param __a
610 ///    A 256-bit vector of [8 x float] containing one of the source operands.
611 /// \param __b
612 ///    A 256-bit vector of [8 x float] containing one of the source operands.
613 /// \returns A 256-bit vector of [8 x float] containing the bitwise OR of the
614 ///    values between both operands.
615 static __inline __m256 __DEFAULT_FN_ATTRS
616 _mm256_or_ps(__m256 __a, __m256 __b)
617 {
618   return (__m256)((__v8su)__a | (__v8su)__b);
619 }
620 
621 /// Performs a bitwise XOR of two 256-bit vectors of [4 x double].
622 ///
623 /// \headerfile <x86intrin.h>
624 ///
625 /// This intrinsic corresponds to the <c> VXORPD </c> instruction.
626 ///
627 /// \param __a
628 ///    A 256-bit vector of [4 x double] containing one of the source operands.
629 /// \param __b
630 ///    A 256-bit vector of [4 x double] containing one of the source operands.
631 /// \returns A 256-bit vector of [4 x double] containing the bitwise XOR of the
632 ///    values between both operands.
633 static __inline __m256d __DEFAULT_FN_ATTRS
634 _mm256_xor_pd(__m256d __a, __m256d __b)
635 {
636   return (__m256d)((__v4du)__a ^ (__v4du)__b);
637 }
638 
639 /// Performs a bitwise XOR of two 256-bit vectors of [8 x float].
640 ///
641 /// \headerfile <x86intrin.h>
642 ///
643 /// This intrinsic corresponds to the <c> VXORPS </c> instruction.
644 ///
645 /// \param __a
646 ///    A 256-bit vector of [8 x float] containing one of the source operands.
647 /// \param __b
648 ///    A 256-bit vector of [8 x float] containing one of the source operands.
649 /// \returns A 256-bit vector of [8 x float] containing the bitwise XOR of the
650 ///    values between both operands.
651 static __inline __m256 __DEFAULT_FN_ATTRS
652 _mm256_xor_ps(__m256 __a, __m256 __b)
653 {
654   return (__m256)((__v8su)__a ^ (__v8su)__b);
655 }
656 
657 /* Horizontal arithmetic */
658 /// Horizontally adds the adjacent pairs of values contained in two
659 ///    256-bit vectors of [4 x double].
660 ///
661 /// \headerfile <x86intrin.h>
662 ///
663 /// This intrinsic corresponds to the <c> VHADDPD </c> instruction.
664 ///
665 /// \param __a
666 ///    A 256-bit vector of [4 x double] containing one of the source operands.
667 ///    The horizontal sums of the values are returned in the even-indexed
668 ///    elements of a vector of [4 x double].
669 /// \param __b
670 ///    A 256-bit vector of [4 x double] containing one of the source operands.
671 ///    The horizontal sums of the values are returned in the odd-indexed
672 ///    elements of a vector of [4 x double].
673 /// \returns A 256-bit vector of [4 x double] containing the horizontal sums of
674 ///    both operands.
675 static __inline __m256d __DEFAULT_FN_ATTRS
676 _mm256_hadd_pd(__m256d __a, __m256d __b)
677 {
678   return (__m256d)__builtin_ia32_haddpd256((__v4df)__a, (__v4df)__b);
679 }
680 
681 /// Horizontally adds the adjacent pairs of values contained in two
682 ///    256-bit vectors of [8 x float].
683 ///
684 /// \headerfile <x86intrin.h>
685 ///
686 /// This intrinsic corresponds to the <c> VHADDPS </c> instruction.
687 ///
688 /// \param __a
689 ///    A 256-bit vector of [8 x float] containing one of the source operands.
690 ///    The horizontal sums of the values are returned in the elements with
691 ///    index 0, 1, 4, 5 of a vector of [8 x float].
692 /// \param __b
693 ///    A 256-bit vector of [8 x float] containing one of the source operands.
694 ///    The horizontal sums of the values are returned in the elements with
695 ///    index 2, 3, 6, 7 of a vector of [8 x float].
696 /// \returns A 256-bit vector of [8 x float] containing the horizontal sums of
697 ///    both operands.
698 static __inline __m256 __DEFAULT_FN_ATTRS
699 _mm256_hadd_ps(__m256 __a, __m256 __b)
700 {
701   return (__m256)__builtin_ia32_haddps256((__v8sf)__a, (__v8sf)__b);
702 }
703 
704 /// Horizontally subtracts the adjacent pairs of values contained in two
705 ///    256-bit vectors of [4 x double].
706 ///
707 /// \headerfile <x86intrin.h>
708 ///
709 /// This intrinsic corresponds to the <c> VHSUBPD </c> instruction.
710 ///
711 /// \param __a
712 ///    A 256-bit vector of [4 x double] containing one of the source operands.
713 ///    The horizontal differences between the values are returned in the
714 ///    even-indexed elements of a vector of [4 x double].
715 /// \param __b
716 ///    A 256-bit vector of [4 x double] containing one of the source operands.
717 ///    The horizontal differences between the values are returned in the
718 ///    odd-indexed elements of a vector of [4 x double].
719 /// \returns A 256-bit vector of [4 x double] containing the horizontal
720 ///    differences of both operands.
721 static __inline __m256d __DEFAULT_FN_ATTRS
722 _mm256_hsub_pd(__m256d __a, __m256d __b)
723 {
724   return (__m256d)__builtin_ia32_hsubpd256((__v4df)__a, (__v4df)__b);
725 }
726 
727 /// Horizontally subtracts the adjacent pairs of values contained in two
728 ///    256-bit vectors of [8 x float].
729 ///
730 /// \headerfile <x86intrin.h>
731 ///
732 /// This intrinsic corresponds to the <c> VHSUBPS </c> instruction.
733 ///
734 /// \param __a
735 ///    A 256-bit vector of [8 x float] containing one of the source operands.
736 ///    The horizontal differences between the values are returned in the
737 ///    elements with index 0, 1, 4, 5 of a vector of [8 x float].
738 /// \param __b
739 ///    A 256-bit vector of [8 x float] containing one of the source operands.
740 ///    The horizontal differences between the values are returned in the
741 ///    elements with index 2, 3, 6, 7 of a vector of [8 x float].
742 /// \returns A 256-bit vector of [8 x float] containing the horizontal
743 ///    differences of both operands.
744 static __inline __m256 __DEFAULT_FN_ATTRS
745 _mm256_hsub_ps(__m256 __a, __m256 __b)
746 {
747   return (__m256)__builtin_ia32_hsubps256((__v8sf)__a, (__v8sf)__b);
748 }
749 
750 /* Vector permutations */
751 /// Copies the values in a 128-bit vector of [2 x double] as specified
752 ///    by the 128-bit integer vector operand.
753 ///
754 /// \headerfile <x86intrin.h>
755 ///
756 /// This intrinsic corresponds to the <c> VPERMILPD </c> instruction.
757 ///
758 /// \param __a
759 ///    A 128-bit vector of [2 x double].
760 /// \param __c
761 ///    A 128-bit integer vector operand specifying how the values are to be
762 ///    copied. \n
763 ///    Bit [1]: \n
764 ///      0: Bits [63:0] of the source are copied to bits [63:0] of the returned
765 ///         vector. \n
766 ///      1: Bits [127:64] of the source are copied to bits [63:0] of the
767 ///         returned vector. \n
768 ///    Bit [65]: \n
769 ///      0: Bits [63:0] of the source are copied to bits [127:64] of the
770 ///         returned vector. \n
771 ///      1: Bits [127:64] of the source are copied to bits [127:64] of the
772 ///         returned vector.
773 /// \returns A 128-bit vector of [2 x double] containing the copied values.
774 static __inline __m128d __DEFAULT_FN_ATTRS128
775 _mm_permutevar_pd(__m128d __a, __m128i __c)
776 {
777   return (__m128d)__builtin_ia32_vpermilvarpd((__v2df)__a, (__v2di)__c);
778 }
779 
780 /// Copies the values in a 256-bit vector of [4 x double] as specified
781 ///    by the 256-bit integer vector operand.
782 ///
783 /// \headerfile <x86intrin.h>
784 ///
785 /// This intrinsic corresponds to the <c> VPERMILPD </c> instruction.
786 ///
787 /// \param __a
788 ///    A 256-bit vector of [4 x double].
789 /// \param __c
790 ///    A 256-bit integer vector operand specifying how the values are to be
791 ///    copied. \n
792 ///    Bit [1]: \n
793 ///      0: Bits [63:0] of the source are copied to bits [63:0] of the returned
794 ///         vector. \n
795 ///      1: Bits [127:64] of the source are copied to bits [63:0] of the
796 ///         returned vector. \n
797 ///    Bit [65]: \n
798 ///      0: Bits [63:0] of the source are copied to bits [127:64] of the
799 ///         returned vector. \n
800 ///      1: Bits [127:64] of the source are copied to bits [127:64] of the
801 ///         returned vector. \n
802 ///    Bit [129]: \n
803 ///      0: Bits [191:128] of the source are copied to bits [191:128] of the
804 ///         returned vector. \n
805 ///      1: Bits [255:192] of the source are copied to bits [191:128] of the
806 ///         returned vector. \n
807 ///    Bit [193]: \n
808 ///      0: Bits [191:128] of the source are copied to bits [255:192] of the
809 ///         returned vector. \n
810 ///      1: Bits [255:192] of the source are copied to bits [255:192] of the
811 ///    returned vector.
812 /// \returns A 256-bit vector of [4 x double] containing the copied values.
813 static __inline __m256d __DEFAULT_FN_ATTRS
814 _mm256_permutevar_pd(__m256d __a, __m256i __c)
815 {
816   return (__m256d)__builtin_ia32_vpermilvarpd256((__v4df)__a, (__v4di)__c);
817 }
818 
819 /// Copies the values stored in a 128-bit vector of [4 x float] as
820 ///    specified by the 128-bit integer vector operand.
821 /// \headerfile <x86intrin.h>
822 ///
823 /// This intrinsic corresponds to the <c> VPERMILPS </c> instruction.
824 ///
825 /// \param __a
826 ///    A 128-bit vector of [4 x float].
827 /// \param __c
828 ///    A 128-bit integer vector operand specifying how the values are to be
829 ///    copied. \n
830 ///    Bits [1:0]: \n
831 ///      00: Bits [31:0] of the source are copied to bits [31:0] of the
832 ///          returned vector. \n
833 ///      01: Bits [63:32] of the source are copied to bits [31:0] of the
834 ///          returned vector. \n
835 ///      10: Bits [95:64] of the source are copied to bits [31:0] of the
836 ///          returned vector. \n
837 ///      11: Bits [127:96] of the source are copied to bits [31:0] of the
838 ///          returned vector. \n
839 ///    Bits [33:32]: \n
840 ///      00: Bits [31:0] of the source are copied to bits [63:32] of the
841 ///          returned vector. \n
842 ///      01: Bits [63:32] of the source are copied to bits [63:32] of the
843 ///          returned vector. \n
844 ///      10: Bits [95:64] of the source are copied to bits [63:32] of the
845 ///          returned vector. \n
846 ///      11: Bits [127:96] of the source are copied to bits [63:32] of the
847 ///          returned vector. \n
848 ///    Bits [65:64]: \n
849 ///      00: Bits [31:0] of the source are copied to bits [95:64] of the
850 ///          returned vector. \n
851 ///      01: Bits [63:32] of the source are copied to bits [95:64] of the
852 ///          returned vector. \n
853 ///      10: Bits [95:64] of the source are copied to bits [95:64] of the
854 ///          returned vector. \n
855 ///      11: Bits [127:96] of the source are copied to bits [95:64] of the
856 ///          returned vector. \n
857 ///    Bits [97:96]: \n
858 ///      00: Bits [31:0] of the source are copied to bits [127:96] of the
859 ///          returned vector. \n
860 ///      01: Bits [63:32] of the source are copied to bits [127:96] of the
861 ///          returned vector. \n
862 ///      10: Bits [95:64] of the source are copied to bits [127:96] of the
863 ///          returned vector. \n
864 ///      11: Bits [127:96] of the source are copied to bits [127:96] of the
865 ///          returned vector.
866 /// \returns A 128-bit vector of [4 x float] containing the copied values.
867 static __inline __m128 __DEFAULT_FN_ATTRS128
868 _mm_permutevar_ps(__m128 __a, __m128i __c)
869 {
870   return (__m128)__builtin_ia32_vpermilvarps((__v4sf)__a, (__v4si)__c);
871 }
872 
873 /// Copies the values stored in a 256-bit vector of [8 x float] as
874 ///    specified by the 256-bit integer vector operand.
875 ///
876 /// \headerfile <x86intrin.h>
877 ///
878 /// This intrinsic corresponds to the <c> VPERMILPS </c> instruction.
879 ///
880 /// \param __a
881 ///    A 256-bit vector of [8 x float].
882 /// \param __c
883 ///    A 256-bit integer vector operand specifying how the values are to be
884 ///    copied. \n
885 ///    Bits [1:0]: \n
886 ///      00: Bits [31:0] of the source are copied to bits [31:0] of the
887 ///          returned vector. \n
888 ///      01: Bits [63:32] of the source are copied to bits [31:0] of the
889 ///          returned vector. \n
890 ///      10: Bits [95:64] of the source are copied to bits [31:0] of the
891 ///          returned vector. \n
892 ///      11: Bits [127:96] of the source are copied to bits [31:0] of the
893 ///          returned vector. \n
894 ///    Bits [33:32]: \n
895 ///      00: Bits [31:0] of the source are copied to bits [63:32] of the
896 ///          returned vector. \n
897 ///      01: Bits [63:32] of the source are copied to bits [63:32] of the
898 ///          returned vector. \n
899 ///      10: Bits [95:64] of the source are copied to bits [63:32] of the
900 ///          returned vector. \n
901 ///      11: Bits [127:96] of the source are copied to bits [63:32] of the
902 ///          returned vector. \n
903 ///    Bits [65:64]: \n
904 ///      00: Bits [31:0] of the source are copied to bits [95:64] of the
905 ///          returned vector. \n
906 ///      01: Bits [63:32] of the source are copied to bits [95:64] of the
907 ///          returned vector. \n
908 ///      10: Bits [95:64] of the source are copied to bits [95:64] of the
909 ///          returned vector. \n
910 ///      11: Bits [127:96] of the source are copied to bits [95:64] of the
911 ///          returned vector. \n
912 ///    Bits [97:96]: \n
913 ///      00: Bits [31:0] of the source are copied to bits [127:96] of the
914 ///          returned vector. \n
915 ///      01: Bits [63:32] of the source are copied to bits [127:96] of the
916 ///          returned vector. \n
917 ///      10: Bits [95:64] of the source are copied to bits [127:96] of the
918 ///          returned vector. \n
919 ///      11: Bits [127:96] of the source are copied to bits [127:96] of the
920 ///          returned vector. \n
921 ///    Bits [129:128]: \n
922 ///      00: Bits [159:128] of the source are copied to bits [159:128] of the
923 ///          returned vector. \n
924 ///      01: Bits [191:160] of the source are copied to bits [159:128] of the
925 ///          returned vector. \n
926 ///      10: Bits [223:192] of the source are copied to bits [159:128] of the
927 ///          returned vector. \n
928 ///      11: Bits [255:224] of the source are copied to bits [159:128] of the
929 ///          returned vector. \n
930 ///    Bits [161:160]: \n
931 ///      00: Bits [159:128] of the source are copied to bits [191:160] of the
932 ///          returned vector. \n
933 ///      01: Bits [191:160] of the source are copied to bits [191:160] of the
934 ///          returned vector. \n
935 ///      10: Bits [223:192] of the source are copied to bits [191:160] of the
936 ///          returned vector. \n
937 ///      11: Bits [255:224] of the source are copied to bits [191:160] of the
938 ///          returned vector. \n
939 ///    Bits [193:192]: \n
940 ///      00: Bits [159:128] of the source are copied to bits [223:192] of the
941 ///          returned vector. \n
942 ///      01: Bits [191:160] of the source are copied to bits [223:192] of the
943 ///          returned vector. \n
944 ///      10: Bits [223:192] of the source are copied to bits [223:192] of the
945 ///          returned vector. \n
946 ///      11: Bits [255:224] of the source are copied to bits [223:192] of the
947 ///          returned vector. \n
948 ///    Bits [225:224]: \n
949 ///      00: Bits [159:128] of the source are copied to bits [255:224] of the
950 ///          returned vector. \n
951 ///      01: Bits [191:160] of the source are copied to bits [255:224] of the
952 ///          returned vector. \n
953 ///      10: Bits [223:192] of the source are copied to bits [255:224] of the
954 ///          returned vector. \n
955 ///      11: Bits [255:224] of the source are copied to bits [255:224] of the
956 ///          returned vector.
957 /// \returns A 256-bit vector of [8 x float] containing the copied values.
958 static __inline __m256 __DEFAULT_FN_ATTRS
959 _mm256_permutevar_ps(__m256 __a, __m256i __c)
960 {
961   return (__m256)__builtin_ia32_vpermilvarps256((__v8sf)__a, (__v8si)__c);
962 }
963 
964 /// Copies the values in a 128-bit vector of [2 x double] as specified
965 ///    by the immediate integer operand.
966 ///
967 /// \headerfile <x86intrin.h>
968 ///
969 /// \code
970 /// __m128d _mm_permute_pd(__m128d A, const int C);
971 /// \endcode
972 ///
973 /// This intrinsic corresponds to the <c> VPERMILPD </c> instruction.
974 ///
975 /// \param A
976 ///    A 128-bit vector of [2 x double].
977 /// \param C
978 ///    An immediate integer operand specifying how the values are to be
979 ///    copied. \n
980 ///    Bit [0]: \n
981 ///      0: Bits [63:0] of the source are copied to bits [63:0] of the returned
982 ///         vector. \n
983 ///      1: Bits [127:64] of the source are copied to bits [63:0] of the
984 ///         returned vector. \n
985 ///    Bit [1]: \n
986 ///      0: Bits [63:0] of the source are copied to bits [127:64] of the
987 ///         returned vector. \n
988 ///      1: Bits [127:64] of the source are copied to bits [127:64] of the
989 ///         returned vector.
990 /// \returns A 128-bit vector of [2 x double] containing the copied values.
991 #define _mm_permute_pd(A, C) \
992   (__m128d)__builtin_ia32_vpermilpd((__v2df)(__m128d)(A), (int)(C))
993 
994 /// Copies the values in a 256-bit vector of [4 x double] as specified by
995 ///    the immediate integer operand.
996 ///
997 /// \headerfile <x86intrin.h>
998 ///
999 /// \code
1000 /// __m256d _mm256_permute_pd(__m256d A, const int C);
1001 /// \endcode
1002 ///
1003 /// This intrinsic corresponds to the <c> VPERMILPD </c> instruction.
1004 ///
1005 /// \param A
1006 ///    A 256-bit vector of [4 x double].
1007 /// \param C
1008 ///    An immediate integer operand specifying how the values are to be
1009 ///    copied. \n
1010 ///    Bit [0]: \n
1011 ///      0: Bits [63:0] of the source are copied to bits [63:0] of the returned
1012 ///         vector. \n
1013 ///      1: Bits [127:64] of the source are copied to bits [63:0] of the
1014 ///         returned vector. \n
1015 ///    Bit [1]: \n
1016 ///      0: Bits [63:0] of the source are copied to bits [127:64] of the
1017 ///         returned vector. \n
1018 ///      1: Bits [127:64] of the source are copied to bits [127:64] of the
1019 ///         returned vector. \n
1020 ///    Bit [2]: \n
1021 ///      0: Bits [191:128] of the source are copied to bits [191:128] of the
1022 ///         returned vector. \n
1023 ///      1: Bits [255:192] of the source are copied to bits [191:128] of the
1024 ///         returned vector. \n
1025 ///    Bit [3]: \n
1026 ///      0: Bits [191:128] of the source are copied to bits [255:192] of the
1027 ///         returned vector. \n
1028 ///      1: Bits [255:192] of the source are copied to bits [255:192] of the
1029 ///         returned vector.
1030 /// \returns A 256-bit vector of [4 x double] containing the copied values.
1031 #define _mm256_permute_pd(A, C) \
1032   (__m256d)__builtin_ia32_vpermilpd256((__v4df)(__m256d)(A), (int)(C))
1033 
1034 /// Copies the values in a 128-bit vector of [4 x float] as specified by
1035 ///    the immediate integer operand.
1036 ///
1037 /// \headerfile <x86intrin.h>
1038 ///
1039 /// \code
1040 /// __m128 _mm_permute_ps(__m128 A, const int C);
1041 /// \endcode
1042 ///
1043 /// This intrinsic corresponds to the <c> VPERMILPS </c> instruction.
1044 ///
1045 /// \param A
1046 ///    A 128-bit vector of [4 x float].
1047 /// \param C
1048 ///    An immediate integer operand specifying how the values are to be
1049 ///    copied. \n
1050 ///    Bits [1:0]: \n
1051 ///      00: Bits [31:0] of the source are copied to bits [31:0] of the
1052 ///          returned vector. \n
1053 ///      01: Bits [63:32] of the source are copied to bits [31:0] of the
1054 ///          returned vector. \n
1055 ///      10: Bits [95:64] of the source are copied to bits [31:0] of the
1056 ///          returned vector. \n
1057 ///      11: Bits [127:96] of the source are copied to bits [31:0] of the
1058 ///          returned vector. \n
1059 ///    Bits [3:2]: \n
1060 ///      00: Bits [31:0] of the source are copied to bits [63:32] of the
1061 ///          returned vector. \n
1062 ///      01: Bits [63:32] of the source are copied to bits [63:32] of the
1063 ///          returned vector. \n
1064 ///      10: Bits [95:64] of the source are copied to bits [63:32] of the
1065 ///          returned vector. \n
1066 ///      11: Bits [127:96] of the source are copied to bits [63:32] of the
1067 ///          returned vector. \n
1068 ///    Bits [5:4]: \n
1069 ///      00: Bits [31:0] of the source are copied to bits [95:64] of the
1070 ///          returned vector. \n
1071 ///      01: Bits [63:32] of the source are copied to bits [95:64] of the
1072 ///          returned vector. \n
1073 ///      10: Bits [95:64] of the source are copied to bits [95:64] of the
1074 ///          returned vector. \n
1075 ///      11: Bits [127:96] of the source are copied to bits [95:64] of the
1076 ///          returned vector. \n
1077 ///    Bits [7:6]: \n
1078 ///      00: Bits [31:0] of the source are copied to bits [127:96] of the
1079 ///          returned vector. \n
1080 ///      01: Bits [63:32] of the source are copied to bits [127:96] of the
1081 ///          returned vector. \n
1082 ///      10: Bits [95:64] of the source are copied to bits [127:96] of the
1083 ///          returned vector. \n
1084 ///      11: Bits [127:96] of the source are copied to bits [127:96] of the
1085 ///          returned vector.
1086 /// \returns A 128-bit vector of [4 x float] containing the copied values.
1087 #define _mm_permute_ps(A, C) \
1088   (__m128)__builtin_ia32_vpermilps((__v4sf)(__m128)(A), (int)(C))
1089 
1090 /// Copies the values in a 256-bit vector of [8 x float] as specified by
1091 ///    the immediate integer operand.
1092 ///
1093 /// \headerfile <x86intrin.h>
1094 ///
1095 /// \code
1096 /// __m256 _mm256_permute_ps(__m256 A, const int C);
1097 /// \endcode
1098 ///
1099 /// This intrinsic corresponds to the <c> VPERMILPS </c> instruction.
1100 ///
1101 /// \param A
1102 ///    A 256-bit vector of [8 x float].
1103 /// \param C
1104 ///    An immediate integer operand specifying how the values are to be
1105 ///    copied. \n
1106 ///    Bits [1:0]: \n
1107 ///      00: Bits [31:0] of the source are copied to bits [31:0] of the
1108 ///          returned vector. \n
1109 ///      01: Bits [63:32] of the source are copied to bits [31:0] of the
1110 ///          returned vector. \n
1111 ///      10: Bits [95:64] of the source are copied to bits [31:0] of the
1112 ///          returned vector. \n
1113 ///      11: Bits [127:96] of the source are copied to bits [31:0] of the
1114 ///          returned vector. \n
1115 ///    Bits [3:2]: \n
1116 ///      00: Bits [31:0] of the source are copied to bits [63:32] of the
1117 ///          returned vector. \n
1118 ///      01: Bits [63:32] of the source are copied to bits [63:32] of the
1119 ///          returned vector. \n
1120 ///      10: Bits [95:64] of the source are copied to bits [63:32] of the
1121 ///          returned vector. \n
1122 ///      11: Bits [127:96] of the source are copied to bits [63:32] of the
1123 ///          returned vector. \n
1124 ///    Bits [5:4]: \n
1125 ///      00: Bits [31:0] of the source are copied to bits [95:64] of the
1126 ///          returned vector. \n
1127 ///      01: Bits [63:32] of the source are copied to bits [95:64] of the
1128 ///          returned vector. \n
1129 ///      10: Bits [95:64] of the source are copied to bits [95:64] of the
1130 ///          returned vector. \n
1131 ///      11: Bits [127:96] of the source are copied to bits [95:64] of the
1132 ///          returned vector. \n
1133 ///    Bits [7:6]: \n
1134 ///      00: Bits [31:0] of the source are copied to bits [127:96] of the
1135 ///          returned vector. \n
1136 ///      01: Bits [63:32] of the source are copied to bits [127:96] of the
1137 ///          returned vector. \n
1138 ///      10: Bits [95:64] of the source are copied to bits [127:96] of the
1139 ///          returned vector. \n
1140 ///      11: Bits [127:96] of the source are copied to bits [127:96] of the
1141 ///          returned vector. \n
1142 ///    Bits [1:0]: \n
1143 ///      00: Bits [159:128] of the source are copied to bits [159:128] of the
1144 ///          returned vector. \n
1145 ///      01: Bits [191:160] of the source are copied to bits [159:128] of the
1146 ///          returned vector. \n
1147 ///      10: Bits [223:192] of the source are copied to bits [159:128] of the
1148 ///          returned vector. \n
1149 ///      11: Bits [255:224] of the source are copied to bits [159:128] of the
1150 ///          returned vector. \n
1151 ///    Bits [3:2]: \n
1152 ///      00: Bits [159:128] of the source are copied to bits [191:160] of the
1153 ///          returned vector. \n
1154 ///      01: Bits [191:160] of the source are copied to bits [191:160] of the
1155 ///          returned vector. \n
1156 ///      10: Bits [223:192] of the source are copied to bits [191:160] of the
1157 ///          returned vector. \n
1158 ///      11: Bits [255:224] of the source are copied to bits [191:160] of the
1159 ///          returned vector. \n
1160 ///    Bits [5:4]: \n
1161 ///      00: Bits [159:128] of the source are copied to bits [223:192] of the
1162 ///          returned vector. \n
1163 ///      01: Bits [191:160] of the source are copied to bits [223:192] of the
1164 ///          returned vector. \n
1165 ///      10: Bits [223:192] of the source are copied to bits [223:192] of the
1166 ///          returned vector. \n
1167 ///      11: Bits [255:224] of the source are copied to bits [223:192] of the
1168 ///          returned vector. \n
1169 ///    Bits [7:6]: \n
1170 ///      00: Bits [159:128] of the source are copied to bits [255:224] of the
1171 ///          returned vector. \n
1172 ///      01: Bits [191:160] of the source are copied to bits [255:224] of the
1173 ///          returned vector. \n
1174 ///      10: Bits [223:192] of the source are copied to bits [255:224] of the
1175 ///          returned vector. \n
1176 ///      11: Bits [255:224] of the source are copied to bits [255:224] of the
1177 ///          returned vector.
1178 /// \returns A 256-bit vector of [8 x float] containing the copied values.
1179 #define _mm256_permute_ps(A, C) \
1180   (__m256)__builtin_ia32_vpermilps256((__v8sf)(__m256)(A), (int)(C))
1181 
1182 /// Permutes 128-bit data values stored in two 256-bit vectors of
1183 ///    [4 x double], as specified by the immediate integer operand.
1184 ///
1185 /// \headerfile <x86intrin.h>
1186 ///
1187 /// \code
1188 /// __m256d _mm256_permute2f128_pd(__m256d V1, __m256d V2, const int M);
1189 /// \endcode
1190 ///
1191 /// This intrinsic corresponds to the <c> VPERM2F128 </c> instruction.
1192 ///
1193 /// \param V1
1194 ///    A 256-bit vector of [4 x double].
1195 /// \param V2
1196 ///    A 256-bit vector of [4 x double.
1197 /// \param M
1198 ///    An immediate integer operand specifying how the values are to be
1199 ///    permuted. \n
1200 ///    Bits [1:0]: \n
1201 ///      00: Bits [127:0] of operand \a V1 are copied to bits [127:0] of the
1202 ///          destination. \n
1203 ///      01: Bits [255:128] of operand \a V1 are copied to bits [127:0] of the
1204 ///          destination. \n
1205 ///      10: Bits [127:0] of operand \a V2 are copied to bits [127:0] of the
1206 ///          destination. \n
1207 ///      11: Bits [255:128] of operand \a V2 are copied to bits [127:0] of the
1208 ///          destination. \n
1209 ///    Bits [5:4]: \n
1210 ///      00: Bits [127:0] of operand \a V1 are copied to bits [255:128] of the
1211 ///          destination. \n
1212 ///      01: Bits [255:128] of operand \a V1 are copied to bits [255:128] of the
1213 ///          destination. \n
1214 ///      10: Bits [127:0] of operand \a V2 are copied to bits [255:128] of the
1215 ///          destination. \n
1216 ///      11: Bits [255:128] of operand \a V2 are copied to bits [255:128] of the
1217 ///          destination.
1218 /// \returns A 256-bit vector of [4 x double] containing the copied values.
1219 #define _mm256_permute2f128_pd(V1, V2, M) \
1220   (__m256d)__builtin_ia32_vperm2f128_pd256((__v4df)(__m256d)(V1), \
1221                                            (__v4df)(__m256d)(V2), (int)(M))
1222 
1223 /// Permutes 128-bit data values stored in two 256-bit vectors of
1224 ///    [8 x float], as specified by the immediate integer operand.
1225 ///
1226 /// \headerfile <x86intrin.h>
1227 ///
1228 /// \code
1229 /// __m256 _mm256_permute2f128_ps(__m256 V1, __m256 V2, const int M);
1230 /// \endcode
1231 ///
1232 /// This intrinsic corresponds to the <c> VPERM2F128 </c> instruction.
1233 ///
1234 /// \param V1
1235 ///    A 256-bit vector of [8 x float].
1236 /// \param V2
1237 ///    A 256-bit vector of [8 x float].
1238 /// \param M
1239 ///    An immediate integer operand specifying how the values are to be
1240 ///    permuted. \n
1241 ///    Bits [1:0]: \n
1242 ///    00: Bits [127:0] of operand \a V1 are copied to bits [127:0] of the
1243 ///    destination. \n
1244 ///    01: Bits [255:128] of operand \a V1 are copied to bits [127:0] of the
1245 ///    destination. \n
1246 ///    10: Bits [127:0] of operand \a V2 are copied to bits [127:0] of the
1247 ///    destination. \n
1248 ///    11: Bits [255:128] of operand \a V2 are copied to bits [127:0] of the
1249 ///    destination. \n
1250 ///    Bits [5:4]: \n
1251 ///    00: Bits [127:0] of operand \a V1 are copied to bits [255:128] of the
1252 ///    destination. \n
1253 ///    01: Bits [255:128] of operand \a V1 are copied to bits [255:128] of the
1254 ///    destination. \n
1255 ///    10: Bits [127:0] of operand \a V2 are copied to bits [255:128] of the
1256 ///    destination. \n
1257 ///    11: Bits [255:128] of operand \a V2 are copied to bits [255:128] of the
1258 ///    destination.
1259 /// \returns A 256-bit vector of [8 x float] containing the copied values.
1260 #define _mm256_permute2f128_ps(V1, V2, M) \
1261   (__m256)__builtin_ia32_vperm2f128_ps256((__v8sf)(__m256)(V1), \
1262                                           (__v8sf)(__m256)(V2), (int)(M))
1263 
1264 /// Permutes 128-bit data values stored in two 256-bit integer vectors,
1265 ///    as specified by the immediate integer operand.
1266 ///
1267 /// \headerfile <x86intrin.h>
1268 ///
1269 /// \code
1270 /// __m256i _mm256_permute2f128_si256(__m256i V1, __m256i V2, const int M);
1271 /// \endcode
1272 ///
1273 /// This intrinsic corresponds to the <c> VPERM2F128 </c> instruction.
1274 ///
1275 /// \param V1
1276 ///    A 256-bit integer vector.
1277 /// \param V2
1278 ///    A 256-bit integer vector.
1279 /// \param M
1280 ///    An immediate integer operand specifying how the values are to be copied.
1281 ///    Bits [1:0]: \n
1282 ///    00: Bits [127:0] of operand \a V1 are copied to bits [127:0] of the
1283 ///    destination. \n
1284 ///    01: Bits [255:128] of operand \a V1 are copied to bits [127:0] of the
1285 ///    destination. \n
1286 ///    10: Bits [127:0] of operand \a V2 are copied to bits [127:0] of the
1287 ///    destination. \n
1288 ///    11: Bits [255:128] of operand \a V2 are copied to bits [127:0] of the
1289 ///    destination. \n
1290 ///    Bits [5:4]: \n
1291 ///    00: Bits [127:0] of operand \a V1 are copied to bits [255:128] of the
1292 ///    destination. \n
1293 ///    01: Bits [255:128] of operand \a V1 are copied to bits [255:128] of the
1294 ///    destination. \n
1295 ///    10: Bits [127:0] of operand \a V2 are copied to bits [255:128] of the
1296 ///    destination. \n
1297 ///    11: Bits [255:128] of operand \a V2 are copied to bits [255:128] of the
1298 ///    destination.
1299 /// \returns A 256-bit integer vector containing the copied values.
1300 #define _mm256_permute2f128_si256(V1, V2, M) \
1301   (__m256i)__builtin_ia32_vperm2f128_si256((__v8si)(__m256i)(V1), \
1302                                            (__v8si)(__m256i)(V2), (int)(M))
1303 
1304 /* Vector Blend */
1305 /// Merges 64-bit double-precision data values stored in either of the
1306 ///    two 256-bit vectors of [4 x double], as specified by the immediate
1307 ///    integer operand.
1308 ///
1309 /// \headerfile <x86intrin.h>
1310 ///
1311 /// \code
1312 /// __m256d _mm256_blend_pd(__m256d V1, __m256d V2, const int M);
1313 /// \endcode
1314 ///
1315 /// This intrinsic corresponds to the <c> VBLENDPD </c> instruction.
1316 ///
1317 /// \param V1
1318 ///    A 256-bit vector of [4 x double].
1319 /// \param V2
1320 ///    A 256-bit vector of [4 x double].
1321 /// \param M
1322 ///    An immediate integer operand, with mask bits [3:0] specifying how the
1323 ///    values are to be copied. The position of the mask bit corresponds to the
1324 ///    index of a copied value. When a mask bit is 0, the corresponding 64-bit
1325 ///    element in operand \a V1 is copied to the same position in the
1326 ///    destination. When a mask bit is 1, the corresponding 64-bit element in
1327 ///    operand \a V2 is copied to the same position in the destination.
1328 /// \returns A 256-bit vector of [4 x double] containing the copied values.
1329 #define _mm256_blend_pd(V1, V2, M) \
1330   (__m256d)__builtin_ia32_blendpd256((__v4df)(__m256d)(V1), \
1331                                      (__v4df)(__m256d)(V2), (int)(M))
1332 
1333 /// Merges 32-bit single-precision data values stored in either of the
1334 ///    two 256-bit vectors of [8 x float], as specified by the immediate
1335 ///    integer operand.
1336 ///
1337 /// \headerfile <x86intrin.h>
1338 ///
1339 /// \code
1340 /// __m256 _mm256_blend_ps(__m256 V1, __m256 V2, const int M);
1341 /// \endcode
1342 ///
1343 /// This intrinsic corresponds to the <c> VBLENDPS </c> instruction.
1344 ///
1345 /// \param V1
1346 ///    A 256-bit vector of [8 x float].
1347 /// \param V2
1348 ///    A 256-bit vector of [8 x float].
1349 /// \param M
1350 ///    An immediate integer operand, with mask bits [7:0] specifying how the
1351 ///    values are to be copied. The position of the mask bit corresponds to the
1352 ///    index of a copied value. When a mask bit is 0, the corresponding 32-bit
1353 ///    element in operand \a V1 is copied to the same position in the
1354 ///    destination. When a mask bit is 1, the corresponding 32-bit element in
1355 ///    operand \a V2 is copied to the same position in the destination.
1356 /// \returns A 256-bit vector of [8 x float] containing the copied values.
1357 #define _mm256_blend_ps(V1, V2, M) \
1358   (__m256)__builtin_ia32_blendps256((__v8sf)(__m256)(V1), \
1359                                     (__v8sf)(__m256)(V2), (int)(M))
1360 
1361 /// Merges 64-bit double-precision data values stored in either of the
1362 ///    two 256-bit vectors of [4 x double], as specified by the 256-bit vector
1363 ///    operand.
1364 ///
1365 /// \headerfile <x86intrin.h>
1366 ///
1367 /// This intrinsic corresponds to the <c> VBLENDVPD </c> instruction.
1368 ///
1369 /// \param __a
1370 ///    A 256-bit vector of [4 x double].
1371 /// \param __b
1372 ///    A 256-bit vector of [4 x double].
1373 /// \param __c
1374 ///    A 256-bit vector operand, with mask bits 255, 191, 127, and 63 specifying
1375 ///    how the values are to be copied. The position of the mask bit corresponds
1376 ///    to the most significant bit of a copied value. When a mask bit is 0, the
1377 ///    corresponding 64-bit element in operand \a __a is copied to the same
1378 ///    position in the destination. When a mask bit is 1, the corresponding
1379 ///    64-bit element in operand \a __b is copied to the same position in the
1380 ///    destination.
1381 /// \returns A 256-bit vector of [4 x double] containing the copied values.
1382 static __inline __m256d __DEFAULT_FN_ATTRS
1383 _mm256_blendv_pd(__m256d __a, __m256d __b, __m256d __c)
1384 {
1385   return (__m256d)__builtin_ia32_blendvpd256(
1386     (__v4df)__a, (__v4df)__b, (__v4df)__c);
1387 }
1388 
1389 /// Merges 32-bit single-precision data values stored in either of the
1390 ///    two 256-bit vectors of [8 x float], as specified by the 256-bit vector
1391 ///    operand.
1392 ///
1393 /// \headerfile <x86intrin.h>
1394 ///
1395 /// This intrinsic corresponds to the <c> VBLENDVPS </c> instruction.
1396 ///
1397 /// \param __a
1398 ///    A 256-bit vector of [8 x float].
1399 /// \param __b
1400 ///    A 256-bit vector of [8 x float].
1401 /// \param __c
1402 ///    A 256-bit vector operand, with mask bits 255, 223, 191, 159, 127, 95, 63,
1403 ///    and 31 specifying how the values are to be copied. The position of the
1404 ///    mask bit corresponds to the most significant bit of a copied value. When
1405 ///    a mask bit is 0, the corresponding 32-bit element in operand \a __a is
1406 ///    copied to the same position in the destination. When a mask bit is 1, the
1407 ///    corresponding 32-bit element in operand \a __b is copied to the same
1408 ///    position in the destination.
1409 /// \returns A 256-bit vector of [8 x float] containing the copied values.
1410 static __inline __m256 __DEFAULT_FN_ATTRS
1411 _mm256_blendv_ps(__m256 __a, __m256 __b, __m256 __c)
1412 {
1413   return (__m256)__builtin_ia32_blendvps256(
1414     (__v8sf)__a, (__v8sf)__b, (__v8sf)__c);
1415 }
1416 
1417 /* Vector Dot Product */
1418 /// Computes two dot products in parallel, using the lower and upper
1419 ///    halves of two [8 x float] vectors as input to the two computations, and
1420 ///    returning the two dot products in the lower and upper halves of the
1421 ///    [8 x float] result.
1422 ///
1423 ///    The immediate integer operand controls which input elements will
1424 ///    contribute to the dot product, and where the final results are returned.
1425 ///    In general, for each dot product, the four corresponding elements of the
1426 ///    input vectors are multiplied; the first two and second two products are
1427 ///    summed, then the two sums are added to form the final result.
1428 ///
1429 /// \headerfile <x86intrin.h>
1430 ///
1431 /// \code
1432 /// __m256 _mm256_dp_ps(__m256 V1, __m256 V2, const int M);
1433 /// \endcode
1434 ///
1435 /// This intrinsic corresponds to the <c> VDPPS </c> instruction.
1436 ///
1437 /// \param V1
1438 ///    A vector of [8 x float] values, treated as two [4 x float] vectors.
1439 /// \param V2
1440 ///    A vector of [8 x float] values, treated as two [4 x float] vectors.
1441 /// \param M
1442 ///    An immediate integer argument. Bits [7:4] determine which elements of
1443 ///    the input vectors are used, with bit [4] corresponding to the lowest
1444 ///    element and bit [7] corresponding to the highest element of each [4 x
1445 ///    float] subvector. If a bit is set, the corresponding elements from the
1446 ///    two input vectors are used as an input for dot product; otherwise that
1447 ///    input is treated as zero. Bits [3:0] determine which elements of the
1448 ///    result will receive a copy of the final dot product, with bit [0]
1449 ///    corresponding to the lowest element and bit [3] corresponding to the
1450 ///    highest element of each [4 x float] subvector. If a bit is set, the dot
1451 ///    product is returned in the corresponding element; otherwise that element
1452 ///    is set to zero. The bitmask is applied in the same way to each of the
1453 ///    two parallel dot product computations.
1454 /// \returns A 256-bit vector of [8 x float] containing the two dot products.
1455 #define _mm256_dp_ps(V1, V2, M) \
1456   (__m256)__builtin_ia32_dpps256((__v8sf)(__m256)(V1), \
1457                                  (__v8sf)(__m256)(V2), (M))
1458 
1459 /* Vector shuffle */
1460 /// Selects 8 float values from the 256-bit operands of [8 x float], as
1461 ///    specified by the immediate value operand.
1462 ///
1463 ///    The four selected elements in each operand are copied to the destination
1464 ///    according to the bits specified in the immediate operand. The selected
1465 ///    elements from the first 256-bit operand are copied to bits [63:0] and
1466 ///    bits [191:128] of the destination, and the selected elements from the
1467 ///    second 256-bit operand are copied to bits [127:64] and bits [255:192] of
1468 ///    the destination. For example, if bits [7:0] of the immediate operand
1469 ///    contain a value of 0xFF, the 256-bit destination vector would contain the
1470 ///    following values: b[7], b[7], a[7], a[7], b[3], b[3], a[3], a[3].
1471 ///
1472 /// \headerfile <x86intrin.h>
1473 ///
1474 /// \code
1475 /// __m256 _mm256_shuffle_ps(__m256 a, __m256 b, const int mask);
1476 /// \endcode
1477 ///
1478 /// This intrinsic corresponds to the <c> VSHUFPS </c> instruction.
1479 ///
1480 /// \param a
1481 ///    A 256-bit vector of [8 x float]. The four selected elements in this
1482 ///    operand are copied to bits [63:0] and bits [191:128] in the destination,
1483 ///    according to the bits specified in the immediate operand.
1484 /// \param b
1485 ///    A 256-bit vector of [8 x float]. The four selected elements in this
1486 ///    operand are copied to bits [127:64] and bits [255:192] in the
1487 ///    destination, according to the bits specified in the immediate operand.
1488 /// \param mask
1489 ///    An immediate value containing an 8-bit value specifying which elements to
1490 ///    copy from \a a and \a b \n.
1491 ///    Bits [3:0] specify the values copied from operand \a a. \n
1492 ///    Bits [7:4] specify the values copied from operand \a b. \n
1493 ///    The destinations within the 256-bit destination are assigned values as
1494 ///    follows, according to the bit value assignments described below: \n
1495 ///    Bits [1:0] are used to assign values to bits [31:0] and [159:128] in the
1496 ///    destination. \n
1497 ///    Bits [3:2] are used to assign values to bits [63:32] and [191:160] in the
1498 ///    destination. \n
1499 ///    Bits [5:4] are used to assign values to bits [95:64] and [223:192] in the
1500 ///    destination. \n
1501 ///    Bits [7:6] are used to assign values to bits [127:96] and [255:224] in
1502 ///    the destination. \n
1503 ///    Bit value assignments: \n
1504 ///    00: Bits [31:0] and [159:128] are copied from the selected operand. \n
1505 ///    01: Bits [63:32] and [191:160] are copied from the selected operand. \n
1506 ///    10: Bits [95:64] and [223:192] are copied from the selected operand. \n
1507 ///    11: Bits [127:96] and [255:224] are copied from the selected operand.
1508 /// \returns A 256-bit vector of [8 x float] containing the shuffled values.
1509 #define _mm256_shuffle_ps(a, b, mask) \
1510   (__m256)__builtin_ia32_shufps256((__v8sf)(__m256)(a), \
1511                                    (__v8sf)(__m256)(b), (int)(mask))
1512 
1513 /// Selects four double-precision values from the 256-bit operands of
1514 ///    [4 x double], as specified by the immediate value operand.
1515 ///
1516 ///    The selected elements from the first 256-bit operand are copied to bits
1517 ///    [63:0] and bits [191:128] in the destination, and the selected elements
1518 ///    from the second 256-bit operand are copied to bits [127:64] and bits
1519 ///    [255:192] in the destination. For example, if bits [3:0] of the immediate
1520 ///    operand contain a value of 0xF, the 256-bit destination vector would
1521 ///    contain the following values: b[3], a[3], b[1], a[1].
1522 ///
1523 /// \headerfile <x86intrin.h>
1524 ///
1525 /// \code
1526 /// __m256d _mm256_shuffle_pd(__m256d a, __m256d b, const int mask);
1527 /// \endcode
1528 ///
1529 /// This intrinsic corresponds to the <c> VSHUFPD </c> instruction.
1530 ///
1531 /// \param a
1532 ///    A 256-bit vector of [4 x double].
1533 /// \param b
1534 ///    A 256-bit vector of [4 x double].
1535 /// \param mask
1536 ///    An immediate value containing 8-bit values specifying which elements to
1537 ///    copy from \a a and \a b: \n
1538 ///    Bit [0]=0: Bits [63:0] are copied from \a a to bits [63:0] of the
1539 ///    destination. \n
1540 ///    Bit [0]=1: Bits [127:64] are copied from \a a to bits [63:0] of the
1541 ///    destination. \n
1542 ///    Bit [1]=0: Bits [63:0] are copied from \a b to bits [127:64] of the
1543 ///    destination. \n
1544 ///    Bit [1]=1: Bits [127:64] are copied from \a b to bits [127:64] of the
1545 ///    destination. \n
1546 ///    Bit [2]=0: Bits [191:128] are copied from \a a to bits [191:128] of the
1547 ///    destination. \n
1548 ///    Bit [2]=1: Bits [255:192] are copied from \a a to bits [191:128] of the
1549 ///    destination. \n
1550 ///    Bit [3]=0: Bits [191:128] are copied from \a b to bits [255:192] of the
1551 ///    destination. \n
1552 ///    Bit [3]=1: Bits [255:192] are copied from \a b to bits [255:192] of the
1553 ///    destination.
1554 /// \returns A 256-bit vector of [4 x double] containing the shuffled values.
1555 #define _mm256_shuffle_pd(a, b, mask) \
1556   (__m256d)__builtin_ia32_shufpd256((__v4df)(__m256d)(a), \
1557                                     (__v4df)(__m256d)(b), (int)(mask))
1558 
1559 /* Compare */
1560 #define _CMP_EQ_OQ    0x00 /* Equal (ordered, non-signaling)  */
1561 #define _CMP_LT_OS    0x01 /* Less-than (ordered, signaling)  */
1562 #define _CMP_LE_OS    0x02 /* Less-than-or-equal (ordered, signaling)  */
1563 #define _CMP_UNORD_Q  0x03 /* Unordered (non-signaling)  */
1564 #define _CMP_NEQ_UQ   0x04 /* Not-equal (unordered, non-signaling)  */
1565 #define _CMP_NLT_US   0x05 /* Not-less-than (unordered, signaling)  */
1566 #define _CMP_NLE_US   0x06 /* Not-less-than-or-equal (unordered, signaling)  */
1567 #define _CMP_ORD_Q    0x07 /* Ordered (non-signaling)   */
1568 #define _CMP_EQ_UQ    0x08 /* Equal (unordered, non-signaling)  */
1569 #define _CMP_NGE_US   0x09 /* Not-greater-than-or-equal (unordered, signaling)  */
1570 #define _CMP_NGT_US   0x0a /* Not-greater-than (unordered, signaling)  */
1571 #define _CMP_FALSE_OQ 0x0b /* False (ordered, non-signaling)  */
1572 #define _CMP_NEQ_OQ   0x0c /* Not-equal (ordered, non-signaling)  */
1573 #define _CMP_GE_OS    0x0d /* Greater-than-or-equal (ordered, signaling)  */
1574 #define _CMP_GT_OS    0x0e /* Greater-than (ordered, signaling)  */
1575 #define _CMP_TRUE_UQ  0x0f /* True (unordered, non-signaling)  */
1576 #define _CMP_EQ_OS    0x10 /* Equal (ordered, signaling)  */
1577 #define _CMP_LT_OQ    0x11 /* Less-than (ordered, non-signaling)  */
1578 #define _CMP_LE_OQ    0x12 /* Less-than-or-equal (ordered, non-signaling)  */
1579 #define _CMP_UNORD_S  0x13 /* Unordered (signaling)  */
1580 #define _CMP_NEQ_US   0x14 /* Not-equal (unordered, signaling)  */
1581 #define _CMP_NLT_UQ   0x15 /* Not-less-than (unordered, non-signaling)  */
1582 #define _CMP_NLE_UQ   0x16 /* Not-less-than-or-equal (unordered, non-signaling)  */
1583 #define _CMP_ORD_S    0x17 /* Ordered (signaling)  */
1584 #define _CMP_EQ_US    0x18 /* Equal (unordered, signaling)  */
1585 #define _CMP_NGE_UQ   0x19 /* Not-greater-than-or-equal (unordered, non-signaling)  */
1586 #define _CMP_NGT_UQ   0x1a /* Not-greater-than (unordered, non-signaling)  */
1587 #define _CMP_FALSE_OS 0x1b /* False (ordered, signaling)  */
1588 #define _CMP_NEQ_OS   0x1c /* Not-equal (ordered, signaling)  */
1589 #define _CMP_GE_OQ    0x1d /* Greater-than-or-equal (ordered, non-signaling)  */
1590 #define _CMP_GT_OQ    0x1e /* Greater-than (ordered, non-signaling)  */
1591 #define _CMP_TRUE_US  0x1f /* True (unordered, signaling)  */
1592 
1593 /// Compares each of the corresponding double-precision values of two
1594 ///    128-bit vectors of [2 x double], using the operation specified by the
1595 ///    immediate integer operand.
1596 ///
1597 ///    Returns a [2 x double] vector consisting of two doubles corresponding to
1598 ///    the two comparison results: zero if the comparison is false, and all 1's
1599 ///    if the comparison is true.
1600 ///
1601 /// \headerfile <x86intrin.h>
1602 ///
1603 /// \code
1604 /// __m128d _mm_cmp_pd(__m128d a, __m128d b, const int c);
1605 /// \endcode
1606 ///
1607 /// This intrinsic corresponds to the <c> VCMPPD </c> instruction.
1608 ///
1609 /// \param a
1610 ///    A 128-bit vector of [2 x double].
1611 /// \param b
1612 ///    A 128-bit vector of [2 x double].
1613 /// \param c
1614 ///    An immediate integer operand, with bits [4:0] specifying which comparison
1615 ///    operation to use: \n
1616 ///    0x00: Equal (ordered, non-signaling) \n
1617 ///    0x01: Less-than (ordered, signaling) \n
1618 ///    0x02: Less-than-or-equal (ordered, signaling) \n
1619 ///    0x03: Unordered (non-signaling) \n
1620 ///    0x04: Not-equal (unordered, non-signaling) \n
1621 ///    0x05: Not-less-than (unordered, signaling) \n
1622 ///    0x06: Not-less-than-or-equal (unordered, signaling) \n
1623 ///    0x07: Ordered (non-signaling) \n
1624 ///    0x08: Equal (unordered, non-signaling) \n
1625 ///    0x09: Not-greater-than-or-equal (unordered, signaling) \n
1626 ///    0x0A: Not-greater-than (unordered, signaling) \n
1627 ///    0x0B: False (ordered, non-signaling) \n
1628 ///    0x0C: Not-equal (ordered, non-signaling) \n
1629 ///    0x0D: Greater-than-or-equal (ordered, signaling) \n
1630 ///    0x0E: Greater-than (ordered, signaling) \n
1631 ///    0x0F: True (unordered, non-signaling) \n
1632 ///    0x10: Equal (ordered, signaling) \n
1633 ///    0x11: Less-than (ordered, non-signaling) \n
1634 ///    0x12: Less-than-or-equal (ordered, non-signaling) \n
1635 ///    0x13: Unordered (signaling) \n
1636 ///    0x14: Not-equal (unordered, signaling) \n
1637 ///    0x15: Not-less-than (unordered, non-signaling) \n
1638 ///    0x16: Not-less-than-or-equal (unordered, non-signaling) \n
1639 ///    0x17: Ordered (signaling) \n
1640 ///    0x18: Equal (unordered, signaling) \n
1641 ///    0x19: Not-greater-than-or-equal (unordered, non-signaling) \n
1642 ///    0x1A: Not-greater-than (unordered, non-signaling) \n
1643 ///    0x1B: False (ordered, signaling) \n
1644 ///    0x1C: Not-equal (ordered, signaling) \n
1645 ///    0x1D: Greater-than-or-equal (ordered, non-signaling) \n
1646 ///    0x1E: Greater-than (ordered, non-signaling) \n
1647 ///    0x1F: True (unordered, signaling)
1648 /// \returns A 128-bit vector of [2 x double] containing the comparison results.
1649 #define _mm_cmp_pd(a, b, c) \
1650   (__m128d)__builtin_ia32_cmppd((__v2df)(__m128d)(a), \
1651                                 (__v2df)(__m128d)(b), (c))
1652 
1653 /// Compares each of the corresponding values of two 128-bit vectors of
1654 ///    [4 x float], using the operation specified by the immediate integer
1655 ///    operand.
1656 ///
1657 ///    Returns a [4 x float] vector consisting of four floats corresponding to
1658 ///    the four comparison results: zero if the comparison is false, and all 1's
1659 ///    if the comparison is true.
1660 ///
1661 /// \headerfile <x86intrin.h>
1662 ///
1663 /// \code
1664 /// __m128 _mm_cmp_ps(__m128 a, __m128 b, const int c);
1665 /// \endcode
1666 ///
1667 /// This intrinsic corresponds to the <c> VCMPPS </c> instruction.
1668 ///
1669 /// \param a
1670 ///    A 128-bit vector of [4 x float].
1671 /// \param b
1672 ///    A 128-bit vector of [4 x float].
1673 /// \param c
1674 ///    An immediate integer operand, with bits [4:0] specifying which comparison
1675 ///    operation to use: \n
1676 ///    0x00: Equal (ordered, non-signaling) \n
1677 ///    0x01: Less-than (ordered, signaling) \n
1678 ///    0x02: Less-than-or-equal (ordered, signaling) \n
1679 ///    0x03: Unordered (non-signaling) \n
1680 ///    0x04: Not-equal (unordered, non-signaling) \n
1681 ///    0x05: Not-less-than (unordered, signaling) \n
1682 ///    0x06: Not-less-than-or-equal (unordered, signaling) \n
1683 ///    0x07: Ordered (non-signaling) \n
1684 ///    0x08: Equal (unordered, non-signaling) \n
1685 ///    0x09: Not-greater-than-or-equal (unordered, signaling) \n
1686 ///    0x0A: Not-greater-than (unordered, signaling) \n
1687 ///    0x0B: False (ordered, non-signaling) \n
1688 ///    0x0C: Not-equal (ordered, non-signaling) \n
1689 ///    0x0D: Greater-than-or-equal (ordered, signaling) \n
1690 ///    0x0E: Greater-than (ordered, signaling) \n
1691 ///    0x0F: True (unordered, non-signaling) \n
1692 ///    0x10: Equal (ordered, signaling) \n
1693 ///    0x11: Less-than (ordered, non-signaling) \n
1694 ///    0x12: Less-than-or-equal (ordered, non-signaling) \n
1695 ///    0x13: Unordered (signaling) \n
1696 ///    0x14: Not-equal (unordered, signaling) \n
1697 ///    0x15: Not-less-than (unordered, non-signaling) \n
1698 ///    0x16: Not-less-than-or-equal (unordered, non-signaling) \n
1699 ///    0x17: Ordered (signaling) \n
1700 ///    0x18: Equal (unordered, signaling) \n
1701 ///    0x19: Not-greater-than-or-equal (unordered, non-signaling) \n
1702 ///    0x1A: Not-greater-than (unordered, non-signaling) \n
1703 ///    0x1B: False (ordered, signaling) \n
1704 ///    0x1C: Not-equal (ordered, signaling) \n
1705 ///    0x1D: Greater-than-or-equal (ordered, non-signaling) \n
1706 ///    0x1E: Greater-than (ordered, non-signaling) \n
1707 ///    0x1F: True (unordered, signaling)
1708 /// \returns A 128-bit vector of [4 x float] containing the comparison results.
1709 #define _mm_cmp_ps(a, b, c) \
1710   (__m128)__builtin_ia32_cmpps((__v4sf)(__m128)(a), \
1711                                (__v4sf)(__m128)(b), (c))
1712 
1713 /// Compares each of the corresponding double-precision values of two
1714 ///    256-bit vectors of [4 x double], using the operation specified by the
1715 ///    immediate integer operand.
1716 ///
1717 ///    Returns a [4 x double] vector consisting of four doubles corresponding to
1718 ///    the four comparison results: zero if the comparison is false, and all 1's
1719 ///    if the comparison is true.
1720 ///
1721 /// \headerfile <x86intrin.h>
1722 ///
1723 /// \code
1724 /// __m256d _mm256_cmp_pd(__m256d a, __m256d b, const int c);
1725 /// \endcode
1726 ///
1727 /// This intrinsic corresponds to the <c> VCMPPD </c> instruction.
1728 ///
1729 /// \param a
1730 ///    A 256-bit vector of [4 x double].
1731 /// \param b
1732 ///    A 256-bit vector of [4 x double].
1733 /// \param c
1734 ///    An immediate integer operand, with bits [4:0] specifying which comparison
1735 ///    operation to use: \n
1736 ///    0x00: Equal (ordered, non-signaling) \n
1737 ///    0x01: Less-than (ordered, signaling) \n
1738 ///    0x02: Less-than-or-equal (ordered, signaling) \n
1739 ///    0x03: Unordered (non-signaling) \n
1740 ///    0x04: Not-equal (unordered, non-signaling) \n
1741 ///    0x05: Not-less-than (unordered, signaling) \n
1742 ///    0x06: Not-less-than-or-equal (unordered, signaling) \n
1743 ///    0x07: Ordered (non-signaling) \n
1744 ///    0x08: Equal (unordered, non-signaling) \n
1745 ///    0x09: Not-greater-than-or-equal (unordered, signaling) \n
1746 ///    0x0A: Not-greater-than (unordered, signaling) \n
1747 ///    0x0B: False (ordered, non-signaling) \n
1748 ///    0x0C: Not-equal (ordered, non-signaling) \n
1749 ///    0x0D: Greater-than-or-equal (ordered, signaling) \n
1750 ///    0x0E: Greater-than (ordered, signaling) \n
1751 ///    0x0F: True (unordered, non-signaling) \n
1752 ///    0x10: Equal (ordered, signaling) \n
1753 ///    0x11: Less-than (ordered, non-signaling) \n
1754 ///    0x12: Less-than-or-equal (ordered, non-signaling) \n
1755 ///    0x13: Unordered (signaling) \n
1756 ///    0x14: Not-equal (unordered, signaling) \n
1757 ///    0x15: Not-less-than (unordered, non-signaling) \n
1758 ///    0x16: Not-less-than-or-equal (unordered, non-signaling) \n
1759 ///    0x17: Ordered (signaling) \n
1760 ///    0x18: Equal (unordered, signaling) \n
1761 ///    0x19: Not-greater-than-or-equal (unordered, non-signaling) \n
1762 ///    0x1A: Not-greater-than (unordered, non-signaling) \n
1763 ///    0x1B: False (ordered, signaling) \n
1764 ///    0x1C: Not-equal (ordered, signaling) \n
1765 ///    0x1D: Greater-than-or-equal (ordered, non-signaling) \n
1766 ///    0x1E: Greater-than (ordered, non-signaling) \n
1767 ///    0x1F: True (unordered, signaling)
1768 /// \returns A 256-bit vector of [4 x double] containing the comparison results.
1769 #define _mm256_cmp_pd(a, b, c) \
1770   (__m256d)__builtin_ia32_cmppd256((__v4df)(__m256d)(a), \
1771                                    (__v4df)(__m256d)(b), (c))
1772 
1773 /// Compares each of the corresponding values of two 256-bit vectors of
1774 ///    [8 x float], using the operation specified by the immediate integer
1775 ///    operand.
1776 ///
1777 ///    Returns a [8 x float] vector consisting of eight floats corresponding to
1778 ///    the eight comparison results: zero if the comparison is false, and all
1779 ///    1's if the comparison is true.
1780 ///
1781 /// \headerfile <x86intrin.h>
1782 ///
1783 /// \code
1784 /// __m256 _mm256_cmp_ps(__m256 a, __m256 b, const int c);
1785 /// \endcode
1786 ///
1787 /// This intrinsic corresponds to the <c> VCMPPS </c> instruction.
1788 ///
1789 /// \param a
1790 ///    A 256-bit vector of [8 x float].
1791 /// \param b
1792 ///    A 256-bit vector of [8 x float].
1793 /// \param c
1794 ///    An immediate integer operand, with bits [4:0] specifying which comparison
1795 ///    operation to use: \n
1796 ///    0x00: Equal (ordered, non-signaling) \n
1797 ///    0x01: Less-than (ordered, signaling) \n
1798 ///    0x02: Less-than-or-equal (ordered, signaling) \n
1799 ///    0x03: Unordered (non-signaling) \n
1800 ///    0x04: Not-equal (unordered, non-signaling) \n
1801 ///    0x05: Not-less-than (unordered, signaling) \n
1802 ///    0x06: Not-less-than-or-equal (unordered, signaling) \n
1803 ///    0x07: Ordered (non-signaling) \n
1804 ///    0x08: Equal (unordered, non-signaling) \n
1805 ///    0x09: Not-greater-than-or-equal (unordered, signaling) \n
1806 ///    0x0A: Not-greater-than (unordered, signaling) \n
1807 ///    0x0B: False (ordered, non-signaling) \n
1808 ///    0x0C: Not-equal (ordered, non-signaling) \n
1809 ///    0x0D: Greater-than-or-equal (ordered, signaling) \n
1810 ///    0x0E: Greater-than (ordered, signaling) \n
1811 ///    0x0F: True (unordered, non-signaling) \n
1812 ///    0x10: Equal (ordered, signaling) \n
1813 ///    0x11: Less-than (ordered, non-signaling) \n
1814 ///    0x12: Less-than-or-equal (ordered, non-signaling) \n
1815 ///    0x13: Unordered (signaling) \n
1816 ///    0x14: Not-equal (unordered, signaling) \n
1817 ///    0x15: Not-less-than (unordered, non-signaling) \n
1818 ///    0x16: Not-less-than-or-equal (unordered, non-signaling) \n
1819 ///    0x17: Ordered (signaling) \n
1820 ///    0x18: Equal (unordered, signaling) \n
1821 ///    0x19: Not-greater-than-or-equal (unordered, non-signaling) \n
1822 ///    0x1A: Not-greater-than (unordered, non-signaling) \n
1823 ///    0x1B: False (ordered, signaling) \n
1824 ///    0x1C: Not-equal (ordered, signaling) \n
1825 ///    0x1D: Greater-than-or-equal (ordered, non-signaling) \n
1826 ///    0x1E: Greater-than (ordered, non-signaling) \n
1827 ///    0x1F: True (unordered, signaling)
1828 /// \returns A 256-bit vector of [8 x float] containing the comparison results.
1829 #define _mm256_cmp_ps(a, b, c) \
1830   (__m256)__builtin_ia32_cmpps256((__v8sf)(__m256)(a), \
1831                                   (__v8sf)(__m256)(b), (c))
1832 
1833 /// Compares each of the corresponding scalar double-precision values of
1834 ///    two 128-bit vectors of [2 x double], using the operation specified by the
1835 ///    immediate integer operand.
1836 ///
1837 ///    If the result is true, all 64 bits of the destination vector are set;
1838 ///    otherwise they are cleared.
1839 ///
1840 /// \headerfile <x86intrin.h>
1841 ///
1842 /// \code
1843 /// __m128d _mm_cmp_sd(__m128d a, __m128d b, const int c);
1844 /// \endcode
1845 ///
1846 /// This intrinsic corresponds to the <c> VCMPSD </c> instruction.
1847 ///
1848 /// \param a
1849 ///    A 128-bit vector of [2 x double].
1850 /// \param b
1851 ///    A 128-bit vector of [2 x double].
1852 /// \param c
1853 ///    An immediate integer operand, with bits [4:0] specifying which comparison
1854 ///    operation to use: \n
1855 ///    0x00: Equal (ordered, non-signaling) \n
1856 ///    0x01: Less-than (ordered, signaling) \n
1857 ///    0x02: Less-than-or-equal (ordered, signaling) \n
1858 ///    0x03: Unordered (non-signaling) \n
1859 ///    0x04: Not-equal (unordered, non-signaling) \n
1860 ///    0x05: Not-less-than (unordered, signaling) \n
1861 ///    0x06: Not-less-than-or-equal (unordered, signaling) \n
1862 ///    0x07: Ordered (non-signaling) \n
1863 ///    0x08: Equal (unordered, non-signaling) \n
1864 ///    0x09: Not-greater-than-or-equal (unordered, signaling) \n
1865 ///    0x0A: Not-greater-than (unordered, signaling) \n
1866 ///    0x0B: False (ordered, non-signaling) \n
1867 ///    0x0C: Not-equal (ordered, non-signaling) \n
1868 ///    0x0D: Greater-than-or-equal (ordered, signaling) \n
1869 ///    0x0E: Greater-than (ordered, signaling) \n
1870 ///    0x0F: True (unordered, non-signaling) \n
1871 ///    0x10: Equal (ordered, signaling) \n
1872 ///    0x11: Less-than (ordered, non-signaling) \n
1873 ///    0x12: Less-than-or-equal (ordered, non-signaling) \n
1874 ///    0x13: Unordered (signaling) \n
1875 ///    0x14: Not-equal (unordered, signaling) \n
1876 ///    0x15: Not-less-than (unordered, non-signaling) \n
1877 ///    0x16: Not-less-than-or-equal (unordered, non-signaling) \n
1878 ///    0x17: Ordered (signaling) \n
1879 ///    0x18: Equal (unordered, signaling) \n
1880 ///    0x19: Not-greater-than-or-equal (unordered, non-signaling) \n
1881 ///    0x1A: Not-greater-than (unordered, non-signaling) \n
1882 ///    0x1B: False (ordered, signaling) \n
1883 ///    0x1C: Not-equal (ordered, signaling) \n
1884 ///    0x1D: Greater-than-or-equal (ordered, non-signaling) \n
1885 ///    0x1E: Greater-than (ordered, non-signaling) \n
1886 ///    0x1F: True (unordered, signaling)
1887 /// \returns A 128-bit vector of [2 x double] containing the comparison results.
1888 #define _mm_cmp_sd(a, b, c) \
1889   (__m128d)__builtin_ia32_cmpsd((__v2df)(__m128d)(a), \
1890                                 (__v2df)(__m128d)(b), (c))
1891 
1892 /// Compares each of the corresponding scalar values of two 128-bit
1893 ///    vectors of [4 x float], using the operation specified by the immediate
1894 ///    integer operand.
1895 ///
1896 ///    If the result is true, all 32 bits of the destination vector are set;
1897 ///    otherwise they are cleared.
1898 ///
1899 /// \headerfile <x86intrin.h>
1900 ///
1901 /// \code
1902 /// __m128 _mm_cmp_ss(__m128 a, __m128 b, const int c);
1903 /// \endcode
1904 ///
1905 /// This intrinsic corresponds to the <c> VCMPSS </c> instruction.
1906 ///
1907 /// \param a
1908 ///    A 128-bit vector of [4 x float].
1909 /// \param b
1910 ///    A 128-bit vector of [4 x float].
1911 /// \param c
1912 ///    An immediate integer operand, with bits [4:0] specifying which comparison
1913 ///    operation to use: \n
1914 ///    0x00: Equal (ordered, non-signaling) \n
1915 ///    0x01: Less-than (ordered, signaling) \n
1916 ///    0x02: Less-than-or-equal (ordered, signaling) \n
1917 ///    0x03: Unordered (non-signaling) \n
1918 ///    0x04: Not-equal (unordered, non-signaling) \n
1919 ///    0x05: Not-less-than (unordered, signaling) \n
1920 ///    0x06: Not-less-than-or-equal (unordered, signaling) \n
1921 ///    0x07: Ordered (non-signaling) \n
1922 ///    0x08: Equal (unordered, non-signaling) \n
1923 ///    0x09: Not-greater-than-or-equal (unordered, signaling) \n
1924 ///    0x0A: Not-greater-than (unordered, signaling) \n
1925 ///    0x0B: False (ordered, non-signaling) \n
1926 ///    0x0C: Not-equal (ordered, non-signaling) \n
1927 ///    0x0D: Greater-than-or-equal (ordered, signaling) \n
1928 ///    0x0E: Greater-than (ordered, signaling) \n
1929 ///    0x0F: True (unordered, non-signaling) \n
1930 ///    0x10: Equal (ordered, signaling) \n
1931 ///    0x11: Less-than (ordered, non-signaling) \n
1932 ///    0x12: Less-than-or-equal (ordered, non-signaling) \n
1933 ///    0x13: Unordered (signaling) \n
1934 ///    0x14: Not-equal (unordered, signaling) \n
1935 ///    0x15: Not-less-than (unordered, non-signaling) \n
1936 ///    0x16: Not-less-than-or-equal (unordered, non-signaling) \n
1937 ///    0x17: Ordered (signaling) \n
1938 ///    0x18: Equal (unordered, signaling) \n
1939 ///    0x19: Not-greater-than-or-equal (unordered, non-signaling) \n
1940 ///    0x1A: Not-greater-than (unordered, non-signaling) \n
1941 ///    0x1B: False (ordered, signaling) \n
1942 ///    0x1C: Not-equal (ordered, signaling) \n
1943 ///    0x1D: Greater-than-or-equal (ordered, non-signaling) \n
1944 ///    0x1E: Greater-than (ordered, non-signaling) \n
1945 ///    0x1F: True (unordered, signaling)
1946 /// \returns A 128-bit vector of [4 x float] containing the comparison results.
1947 #define _mm_cmp_ss(a, b, c) \
1948   (__m128)__builtin_ia32_cmpss((__v4sf)(__m128)(a), \
1949                                (__v4sf)(__m128)(b), (c))
1950 
1951 /// Takes a [8 x i32] vector and returns the vector element value
1952 ///    indexed by the immediate constant operand.
1953 ///
1954 /// \headerfile <x86intrin.h>
1955 ///
1956 /// This intrinsic corresponds to the <c> VEXTRACTF128+COMPOSITE </c>
1957 ///   instruction.
1958 ///
1959 /// \param __a
1960 ///    A 256-bit vector of [8 x i32].
1961 /// \param __imm
1962 ///    An immediate integer operand with bits [2:0] determining which vector
1963 ///    element is extracted and returned.
1964 /// \returns A 32-bit integer containing the extracted 32 bits of extended
1965 ///    packed data.
1966 #define _mm256_extract_epi32(X, N) \
1967   (int)__builtin_ia32_vec_ext_v8si((__v8si)(__m256i)(X), (int)(N))
1968 
1969 /// Takes a [16 x i16] vector and returns the vector element value
1970 ///    indexed by the immediate constant operand.
1971 ///
1972 /// \headerfile <x86intrin.h>
1973 ///
1974 /// This intrinsic corresponds to the <c> VEXTRACTF128+COMPOSITE </c>
1975 ///   instruction.
1976 ///
1977 /// \param __a
1978 ///    A 256-bit integer vector of [16 x i16].
1979 /// \param __imm
1980 ///    An immediate integer operand with bits [3:0] determining which vector
1981 ///    element is extracted and returned.
1982 /// \returns A 32-bit integer containing the extracted 16 bits of zero extended
1983 ///    packed data.
1984 #define _mm256_extract_epi16(X, N) \
1985   (int)(unsigned short)__builtin_ia32_vec_ext_v16hi((__v16hi)(__m256i)(X), \
1986                                                     (int)(N))
1987 
1988 /// Takes a [32 x i8] vector and returns the vector element value
1989 ///    indexed by the immediate constant operand.
1990 ///
1991 /// \headerfile <x86intrin.h>
1992 ///
1993 /// This intrinsic corresponds to the <c> VEXTRACTF128+COMPOSITE </c>
1994 ///   instruction.
1995 ///
1996 /// \param __a
1997 ///    A 256-bit integer vector of [32 x i8].
1998 /// \param __imm
1999 ///    An immediate integer operand with bits [4:0] determining which vector
2000 ///    element is extracted and returned.
2001 /// \returns A 32-bit integer containing the extracted 8 bits of zero extended
2002 ///    packed data.
2003 #define _mm256_extract_epi8(X, N) \
2004   (int)(unsigned char)__builtin_ia32_vec_ext_v32qi((__v32qi)(__m256i)(X), \
2005                                                    (int)(N))
2006 
2007 #ifdef __x86_64__
2008 /// Takes a [4 x i64] vector and returns the vector element value
2009 ///    indexed by the immediate constant operand.
2010 ///
2011 /// \headerfile <x86intrin.h>
2012 ///
2013 /// This intrinsic corresponds to the <c> VEXTRACTF128+COMPOSITE </c>
2014 ///   instruction.
2015 ///
2016 /// \param __a
2017 ///    A 256-bit integer vector of [4 x i64].
2018 /// \param __imm
2019 ///    An immediate integer operand with bits [1:0] determining which vector
2020 ///    element is extracted and returned.
2021 /// \returns A 64-bit integer containing the extracted 64 bits of extended
2022 ///    packed data.
2023 #define _mm256_extract_epi64(X, N) \
2024   (long long)__builtin_ia32_vec_ext_v4di((__v4di)(__m256i)(X), (int)(N))
2025 #endif
2026 
2027 /// Takes a [8 x i32] vector and replaces the vector element value
2028 ///    indexed by the immediate constant operand by a new value. Returns the
2029 ///    modified vector.
2030 ///
2031 /// \headerfile <x86intrin.h>
2032 ///
2033 /// This intrinsic corresponds to the <c> VINSERTF128+COMPOSITE </c>
2034 ///   instruction.
2035 ///
2036 /// \param __a
2037 ///    A vector of [8 x i32] to be used by the insert operation.
2038 /// \param __b
2039 ///    An integer value. The replacement value for the insert operation.
2040 /// \param __imm
2041 ///    An immediate integer specifying the index of the vector element to be
2042 ///    replaced.
2043 /// \returns A copy of vector \a __a, after replacing its element indexed by
2044 ///    \a __imm with \a __b.
2045 #define _mm256_insert_epi32(X, I, N) \
2046   (__m256i)__builtin_ia32_vec_set_v8si((__v8si)(__m256i)(X), \
2047                                        (int)(I), (int)(N))
2048 
2049 
2050 /// Takes a [16 x i16] vector and replaces the vector element value
2051 ///    indexed by the immediate constant operand with a new value. Returns the
2052 ///    modified vector.
2053 ///
2054 /// \headerfile <x86intrin.h>
2055 ///
2056 /// This intrinsic corresponds to the <c> VINSERTF128+COMPOSITE </c>
2057 ///   instruction.
2058 ///
2059 /// \param __a
2060 ///    A vector of [16 x i16] to be used by the insert operation.
2061 /// \param __b
2062 ///    An i16 integer value. The replacement value for the insert operation.
2063 /// \param __imm
2064 ///    An immediate integer specifying the index of the vector element to be
2065 ///    replaced.
2066 /// \returns A copy of vector \a __a, after replacing its element indexed by
2067 ///    \a __imm with \a __b.
2068 #define _mm256_insert_epi16(X, I, N) \
2069   (__m256i)__builtin_ia32_vec_set_v16hi((__v16hi)(__m256i)(X), \
2070                                         (int)(I), (int)(N))
2071 
2072 /// Takes a [32 x i8] vector and replaces the vector element value
2073 ///    indexed by the immediate constant operand with a new value. Returns the
2074 ///    modified vector.
2075 ///
2076 /// \headerfile <x86intrin.h>
2077 ///
2078 /// This intrinsic corresponds to the <c> VINSERTF128+COMPOSITE </c>
2079 ///   instruction.
2080 ///
2081 /// \param __a
2082 ///    A vector of [32 x i8] to be used by the insert operation.
2083 /// \param __b
2084 ///    An i8 integer value. The replacement value for the insert operation.
2085 /// \param __imm
2086 ///    An immediate integer specifying the index of the vector element to be
2087 ///    replaced.
2088 /// \returns A copy of vector \a __a, after replacing its element indexed by
2089 ///    \a __imm with \a __b.
2090 #define _mm256_insert_epi8(X, I, N) \
2091   (__m256i)__builtin_ia32_vec_set_v32qi((__v32qi)(__m256i)(X), \
2092                                         (int)(I), (int)(N))
2093 
2094 #ifdef __x86_64__
2095 /// Takes a [4 x i64] vector and replaces the vector element value
2096 ///    indexed by the immediate constant operand with a new value. Returns the
2097 ///    modified vector.
2098 ///
2099 /// \headerfile <x86intrin.h>
2100 ///
2101 /// This intrinsic corresponds to the <c> VINSERTF128+COMPOSITE </c>
2102 ///   instruction.
2103 ///
2104 /// \param __a
2105 ///    A vector of [4 x i64] to be used by the insert operation.
2106 /// \param __b
2107 ///    A 64-bit integer value. The replacement value for the insert operation.
2108 /// \param __imm
2109 ///    An immediate integer specifying the index of the vector element to be
2110 ///    replaced.
2111 /// \returns A copy of vector \a __a, after replacing its element indexed by
2112 ///     \a __imm with \a __b.
2113 #define _mm256_insert_epi64(X, I, N) \
2114   (__m256i)__builtin_ia32_vec_set_v4di((__v4di)(__m256i)(X), \
2115                                        (long long)(I), (int)(N))
2116 #endif
2117 
2118 /* Conversion */
2119 /// Converts a vector of [4 x i32] into a vector of [4 x double].
2120 ///
2121 /// \headerfile <x86intrin.h>
2122 ///
2123 /// This intrinsic corresponds to the <c> VCVTDQ2PD </c> instruction.
2124 ///
2125 /// \param __a
2126 ///    A 128-bit integer vector of [4 x i32].
2127 /// \returns A 256-bit vector of [4 x double] containing the converted values.
2128 static __inline __m256d __DEFAULT_FN_ATTRS
2129 _mm256_cvtepi32_pd(__m128i __a)
2130 {
2131   return (__m256d)__builtin_convertvector((__v4si)__a, __v4df);
2132 }
2133 
2134 /// Converts a vector of [8 x i32] into a vector of [8 x float].
2135 ///
2136 /// \headerfile <x86intrin.h>
2137 ///
2138 /// This intrinsic corresponds to the <c> VCVTDQ2PS </c> instruction.
2139 ///
2140 /// \param __a
2141 ///    A 256-bit integer vector.
2142 /// \returns A 256-bit vector of [8 x float] containing the converted values.
2143 static __inline __m256 __DEFAULT_FN_ATTRS
2144 _mm256_cvtepi32_ps(__m256i __a)
2145 {
2146   return (__m256)__builtin_convertvector((__v8si)__a, __v8sf);
2147 }
2148 
2149 /// Converts a 256-bit vector of [4 x double] into a 128-bit vector of
2150 ///    [4 x float].
2151 ///
2152 /// \headerfile <x86intrin.h>
2153 ///
2154 /// This intrinsic corresponds to the <c> VCVTPD2PS </c> instruction.
2155 ///
2156 /// \param __a
2157 ///    A 256-bit vector of [4 x double].
2158 /// \returns A 128-bit vector of [4 x float] containing the converted values.
2159 static __inline __m128 __DEFAULT_FN_ATTRS
2160 _mm256_cvtpd_ps(__m256d __a)
2161 {
2162   return (__m128)__builtin_ia32_cvtpd2ps256((__v4df) __a);
2163 }
2164 
2165 /// Converts a vector of [8 x float] into a vector of [8 x i32].
2166 ///
2167 /// \headerfile <x86intrin.h>
2168 ///
2169 /// This intrinsic corresponds to the <c> VCVTPS2DQ </c> instruction.
2170 ///
2171 /// \param __a
2172 ///    A 256-bit vector of [8 x float].
2173 /// \returns A 256-bit integer vector containing the converted values.
2174 static __inline __m256i __DEFAULT_FN_ATTRS
2175 _mm256_cvtps_epi32(__m256 __a)
2176 {
2177   return (__m256i)__builtin_ia32_cvtps2dq256((__v8sf) __a);
2178 }
2179 
2180 /// Converts a 128-bit vector of [4 x float] into a 256-bit vector of [4
2181 ///    x double].
2182 ///
2183 /// \headerfile <x86intrin.h>
2184 ///
2185 /// This intrinsic corresponds to the <c> VCVTPS2PD </c> instruction.
2186 ///
2187 /// \param __a
2188 ///    A 128-bit vector of [4 x float].
2189 /// \returns A 256-bit vector of [4 x double] containing the converted values.
2190 static __inline __m256d __DEFAULT_FN_ATTRS
2191 _mm256_cvtps_pd(__m128 __a)
2192 {
2193   return (__m256d)__builtin_convertvector((__v4sf)__a, __v4df);
2194 }
2195 
2196 /// Converts a 256-bit vector of [4 x double] into a 128-bit vector of [4
2197 ///    x i32], truncating the result by rounding towards zero when it is
2198 ///    inexact.
2199 ///
2200 /// \headerfile <x86intrin.h>
2201 ///
2202 /// This intrinsic corresponds to the <c> VCVTTPD2DQ </c> instruction.
2203 ///
2204 /// \param __a
2205 ///    A 256-bit vector of [4 x double].
2206 /// \returns A 128-bit integer vector containing the converted values.
2207 static __inline __m128i __DEFAULT_FN_ATTRS
2208 _mm256_cvttpd_epi32(__m256d __a)
2209 {
2210   return (__m128i)__builtin_ia32_cvttpd2dq256((__v4df) __a);
2211 }
2212 
2213 /// Converts a 256-bit vector of [4 x double] into a 128-bit vector of [4
2214 ///    x i32]. When a conversion is inexact, the value returned is rounded
2215 ///    according to the rounding control bits in the MXCSR register.
2216 ///
2217 /// \headerfile <x86intrin.h>
2218 ///
2219 /// This intrinsic corresponds to the <c> VCVTPD2DQ </c> instruction.
2220 ///
2221 /// \param __a
2222 ///    A 256-bit vector of [4 x double].
2223 /// \returns A 128-bit integer vector containing the converted values.
2224 static __inline __m128i __DEFAULT_FN_ATTRS
2225 _mm256_cvtpd_epi32(__m256d __a)
2226 {
2227   return (__m128i)__builtin_ia32_cvtpd2dq256((__v4df) __a);
2228 }
2229 
2230 /// Converts a vector of [8 x float] into a vector of [8 x i32],
2231 ///    truncating the result by rounding towards zero when it is inexact.
2232 ///
2233 /// \headerfile <x86intrin.h>
2234 ///
2235 /// This intrinsic corresponds to the <c> VCVTTPS2DQ </c> instruction.
2236 ///
2237 /// \param __a
2238 ///    A 256-bit vector of [8 x float].
2239 /// \returns A 256-bit integer vector containing the converted values.
2240 static __inline __m256i __DEFAULT_FN_ATTRS
2241 _mm256_cvttps_epi32(__m256 __a)
2242 {
2243   return (__m256i)__builtin_ia32_cvttps2dq256((__v8sf) __a);
2244 }
2245 
2246 /// Returns the first element of the input vector of [4 x double].
2247 ///
2248 /// \headerfile <avxintrin.h>
2249 ///
2250 /// This intrinsic is a utility function and does not correspond to a specific
2251 ///    instruction.
2252 ///
2253 /// \param __a
2254 ///    A 256-bit vector of [4 x double].
2255 /// \returns A 64 bit double containing the first element of the input vector.
2256 static __inline double __DEFAULT_FN_ATTRS
2257 _mm256_cvtsd_f64(__m256d __a)
2258 {
2259  return __a[0];
2260 }
2261 
2262 /// Returns the first element of the input vector of [8 x i32].
2263 ///
2264 /// \headerfile <avxintrin.h>
2265 ///
2266 /// This intrinsic is a utility function and does not correspond to a specific
2267 ///    instruction.
2268 ///
2269 /// \param __a
2270 ///    A 256-bit vector of [8 x i32].
2271 /// \returns A 32 bit integer containing the first element of the input vector.
2272 static __inline int __DEFAULT_FN_ATTRS
2273 _mm256_cvtsi256_si32(__m256i __a)
2274 {
2275  __v8si __b = (__v8si)__a;
2276  return __b[0];
2277 }
2278 
2279 /// Returns the first element of the input vector of [8 x float].
2280 ///
2281 /// \headerfile <avxintrin.h>
2282 ///
2283 /// This intrinsic is a utility function and does not correspond to a specific
2284 ///    instruction.
2285 ///
2286 /// \param __a
2287 ///    A 256-bit vector of [8 x float].
2288 /// \returns A 32 bit float containing the first element of the input vector.
2289 static __inline float __DEFAULT_FN_ATTRS
2290 _mm256_cvtss_f32(__m256 __a)
2291 {
2292  return __a[0];
2293 }
2294 
2295 /* Vector replicate */
2296 /// Moves and duplicates odd-indexed values from a 256-bit vector of
2297 ///    [8 x float] to float values in a 256-bit vector of [8 x float].
2298 ///
2299 /// \headerfile <x86intrin.h>
2300 ///
2301 /// This intrinsic corresponds to the <c> VMOVSHDUP </c> instruction.
2302 ///
2303 /// \param __a
2304 ///    A 256-bit vector of [8 x float]. \n
2305 ///    Bits [255:224] of \a __a are written to bits [255:224] and [223:192] of
2306 ///    the return value. \n
2307 ///    Bits [191:160] of \a __a are written to bits [191:160] and [159:128] of
2308 ///    the return value. \n
2309 ///    Bits [127:96] of \a __a are written to bits [127:96] and [95:64] of the
2310 ///    return value. \n
2311 ///    Bits [63:32] of \a __a are written to bits [63:32] and [31:0] of the
2312 ///    return value.
2313 /// \returns A 256-bit vector of [8 x float] containing the moved and duplicated
2314 ///    values.
2315 static __inline __m256 __DEFAULT_FN_ATTRS
2316 _mm256_movehdup_ps(__m256 __a)
2317 {
2318   return __builtin_shufflevector((__v8sf)__a, (__v8sf)__a, 1, 1, 3, 3, 5, 5, 7, 7);
2319 }
2320 
2321 /// Moves and duplicates even-indexed values from a 256-bit vector of
2322 ///    [8 x float] to float values in a 256-bit vector of [8 x float].
2323 ///
2324 /// \headerfile <x86intrin.h>
2325 ///
2326 /// This intrinsic corresponds to the <c> VMOVSLDUP </c> instruction.
2327 ///
2328 /// \param __a
2329 ///    A 256-bit vector of [8 x float]. \n
2330 ///    Bits [223:192] of \a __a are written to bits [255:224] and [223:192] of
2331 ///    the return value. \n
2332 ///    Bits [159:128] of \a __a are written to bits [191:160] and [159:128] of
2333 ///    the return value. \n
2334 ///    Bits [95:64] of \a __a are written to bits [127:96] and [95:64] of the
2335 ///    return value. \n
2336 ///    Bits [31:0] of \a __a are written to bits [63:32] and [31:0] of the
2337 ///    return value.
2338 /// \returns A 256-bit vector of [8 x float] containing the moved and duplicated
2339 ///    values.
2340 static __inline __m256 __DEFAULT_FN_ATTRS
2341 _mm256_moveldup_ps(__m256 __a)
2342 {
2343   return __builtin_shufflevector((__v8sf)__a, (__v8sf)__a, 0, 0, 2, 2, 4, 4, 6, 6);
2344 }
2345 
2346 /// Moves and duplicates double-precision floating point values from a
2347 ///    256-bit vector of [4 x double] to double-precision values in a 256-bit
2348 ///    vector of [4 x double].
2349 ///
2350 /// \headerfile <x86intrin.h>
2351 ///
2352 /// This intrinsic corresponds to the <c> VMOVDDUP </c> instruction.
2353 ///
2354 /// \param __a
2355 ///    A 256-bit vector of [4 x double]. \n
2356 ///    Bits [63:0] of \a __a are written to bits [127:64] and [63:0] of the
2357 ///    return value. \n
2358 ///    Bits [191:128] of \a __a are written to bits [255:192] and [191:128] of
2359 ///    the return value.
2360 /// \returns A 256-bit vector of [4 x double] containing the moved and
2361 ///    duplicated values.
2362 static __inline __m256d __DEFAULT_FN_ATTRS
2363 _mm256_movedup_pd(__m256d __a)
2364 {
2365   return __builtin_shufflevector((__v4df)__a, (__v4df)__a, 0, 0, 2, 2);
2366 }
2367 
2368 /* Unpack and Interleave */
2369 /// Unpacks the odd-indexed vector elements from two 256-bit vectors of
2370 ///    [4 x double] and interleaves them into a 256-bit vector of [4 x double].
2371 ///
2372 /// \headerfile <x86intrin.h>
2373 ///
2374 /// This intrinsic corresponds to the <c> VUNPCKHPD </c> instruction.
2375 ///
2376 /// \param __a
2377 ///    A 256-bit floating-point vector of [4 x double]. \n
2378 ///    Bits [127:64] are written to bits [63:0] of the return value. \n
2379 ///    Bits [255:192] are written to bits [191:128] of the return value. \n
2380 /// \param __b
2381 ///    A 256-bit floating-point vector of [4 x double]. \n
2382 ///    Bits [127:64] are written to bits [127:64] of the return value. \n
2383 ///    Bits [255:192] are written to bits [255:192] of the return value. \n
2384 /// \returns A 256-bit vector of [4 x double] containing the interleaved values.
2385 static __inline __m256d __DEFAULT_FN_ATTRS
2386 _mm256_unpackhi_pd(__m256d __a, __m256d __b)
2387 {
2388   return __builtin_shufflevector((__v4df)__a, (__v4df)__b, 1, 5, 1+2, 5+2);
2389 }
2390 
2391 /// Unpacks the even-indexed vector elements from two 256-bit vectors of
2392 ///    [4 x double] and interleaves them into a 256-bit vector of [4 x double].
2393 ///
2394 /// \headerfile <x86intrin.h>
2395 ///
2396 /// This intrinsic corresponds to the <c> VUNPCKLPD </c> instruction.
2397 ///
2398 /// \param __a
2399 ///    A 256-bit floating-point vector of [4 x double]. \n
2400 ///    Bits [63:0] are written to bits [63:0] of the return value. \n
2401 ///    Bits [191:128] are written to bits [191:128] of the return value.
2402 /// \param __b
2403 ///    A 256-bit floating-point vector of [4 x double]. \n
2404 ///    Bits [63:0] are written to bits [127:64] of the return value. \n
2405 ///    Bits [191:128] are written to bits [255:192] of the return value. \n
2406 /// \returns A 256-bit vector of [4 x double] containing the interleaved values.
2407 static __inline __m256d __DEFAULT_FN_ATTRS
2408 _mm256_unpacklo_pd(__m256d __a, __m256d __b)
2409 {
2410   return __builtin_shufflevector((__v4df)__a, (__v4df)__b, 0, 4, 0+2, 4+2);
2411 }
2412 
2413 /// Unpacks the 32-bit vector elements 2, 3, 6 and 7 from each of the
2414 ///    two 256-bit vectors of [8 x float] and interleaves them into a 256-bit
2415 ///    vector of [8 x float].
2416 ///
2417 /// \headerfile <x86intrin.h>
2418 ///
2419 /// This intrinsic corresponds to the <c> VUNPCKHPS </c> instruction.
2420 ///
2421 /// \param __a
2422 ///    A 256-bit vector of [8 x float]. \n
2423 ///    Bits [95:64] are written to bits [31:0] of the return value. \n
2424 ///    Bits [127:96] are written to bits [95:64] of the return value. \n
2425 ///    Bits [223:192] are written to bits [159:128] of the return value. \n
2426 ///    Bits [255:224] are written to bits [223:192] of the return value.
2427 /// \param __b
2428 ///    A 256-bit vector of [8 x float]. \n
2429 ///    Bits [95:64] are written to bits [63:32] of the return value. \n
2430 ///    Bits [127:96] are written to bits [127:96] of the return value. \n
2431 ///    Bits [223:192] are written to bits [191:160] of the return value. \n
2432 ///    Bits [255:224] are written to bits [255:224] of the return value.
2433 /// \returns A 256-bit vector of [8 x float] containing the interleaved values.
2434 static __inline __m256 __DEFAULT_FN_ATTRS
2435 _mm256_unpackhi_ps(__m256 __a, __m256 __b)
2436 {
2437   return __builtin_shufflevector((__v8sf)__a, (__v8sf)__b, 2, 10, 2+1, 10+1, 6, 14, 6+1, 14+1);
2438 }
2439 
2440 /// Unpacks the 32-bit vector elements 0, 1, 4 and 5 from each of the
2441 ///    two 256-bit vectors of [8 x float] and interleaves them into a 256-bit
2442 ///    vector of [8 x float].
2443 ///
2444 /// \headerfile <x86intrin.h>
2445 ///
2446 /// This intrinsic corresponds to the <c> VUNPCKLPS </c> instruction.
2447 ///
2448 /// \param __a
2449 ///    A 256-bit vector of [8 x float]. \n
2450 ///    Bits [31:0] are written to bits [31:0] of the return value. \n
2451 ///    Bits [63:32] are written to bits [95:64] of the return value. \n
2452 ///    Bits [159:128] are written to bits [159:128] of the return value. \n
2453 ///    Bits [191:160] are written to bits [223:192] of the return value.
2454 /// \param __b
2455 ///    A 256-bit vector of [8 x float]. \n
2456 ///    Bits [31:0] are written to bits [63:32] of the return value. \n
2457 ///    Bits [63:32] are written to bits [127:96] of the return value. \n
2458 ///    Bits [159:128] are written to bits [191:160] of the return value. \n
2459 ///    Bits [191:160] are written to bits [255:224] of the return value.
2460 /// \returns A 256-bit vector of [8 x float] containing the interleaved values.
2461 static __inline __m256 __DEFAULT_FN_ATTRS
2462 _mm256_unpacklo_ps(__m256 __a, __m256 __b)
2463 {
2464   return __builtin_shufflevector((__v8sf)__a, (__v8sf)__b, 0, 8, 0+1, 8+1, 4, 12, 4+1, 12+1);
2465 }
2466 
2467 /* Bit Test */
2468 /// Given two 128-bit floating-point vectors of [2 x double], perform an
2469 ///    element-by-element comparison of the double-precision element in the
2470 ///    first source vector and the corresponding element in the second source
2471 ///    vector.
2472 ///
2473 ///    The EFLAGS register is updated as follows: \n
2474 ///    If there is at least one pair of double-precision elements where the
2475 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2476 ///    ZF flag is set to 1. \n
2477 ///    If there is at least one pair of double-precision elements where the
2478 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2479 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2480 ///    This intrinsic returns the value of the ZF flag.
2481 ///
2482 /// \headerfile <x86intrin.h>
2483 ///
2484 /// This intrinsic corresponds to the <c> VTESTPD </c> instruction.
2485 ///
2486 /// \param __a
2487 ///    A 128-bit vector of [2 x double].
2488 /// \param __b
2489 ///    A 128-bit vector of [2 x double].
2490 /// \returns the ZF flag in the EFLAGS register.
2491 static __inline int __DEFAULT_FN_ATTRS128
2492 _mm_testz_pd(__m128d __a, __m128d __b)
2493 {
2494   return __builtin_ia32_vtestzpd((__v2df)__a, (__v2df)__b);
2495 }
2496 
2497 /// Given two 128-bit floating-point vectors of [2 x double], perform an
2498 ///    element-by-element comparison of the double-precision element in the
2499 ///    first source vector and the corresponding element in the second source
2500 ///    vector.
2501 ///
2502 ///    The EFLAGS register is updated as follows: \n
2503 ///    If there is at least one pair of double-precision elements where the
2504 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2505 ///    ZF flag is set to 1. \n
2506 ///    If there is at least one pair of double-precision elements where the
2507 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2508 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2509 ///    This intrinsic returns the value of the CF flag.
2510 ///
2511 /// \headerfile <x86intrin.h>
2512 ///
2513 /// This intrinsic corresponds to the <c> VTESTPD </c> instruction.
2514 ///
2515 /// \param __a
2516 ///    A 128-bit vector of [2 x double].
2517 /// \param __b
2518 ///    A 128-bit vector of [2 x double].
2519 /// \returns the CF flag in the EFLAGS register.
2520 static __inline int __DEFAULT_FN_ATTRS128
2521 _mm_testc_pd(__m128d __a, __m128d __b)
2522 {
2523   return __builtin_ia32_vtestcpd((__v2df)__a, (__v2df)__b);
2524 }
2525 
2526 /// Given two 128-bit floating-point vectors of [2 x double], perform an
2527 ///    element-by-element comparison of the double-precision element in the
2528 ///    first source vector and the corresponding element in the second source
2529 ///    vector.
2530 ///
2531 ///    The EFLAGS register is updated as follows: \n
2532 ///    If there is at least one pair of double-precision elements where the
2533 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2534 ///    ZF flag is set to 1. \n
2535 ///    If there is at least one pair of double-precision elements where the
2536 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2537 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2538 ///    This intrinsic returns 1 if both the ZF and CF flags are set to 0,
2539 ///    otherwise it returns 0.
2540 ///
2541 /// \headerfile <x86intrin.h>
2542 ///
2543 /// This intrinsic corresponds to the <c> VTESTPD </c> instruction.
2544 ///
2545 /// \param __a
2546 ///    A 128-bit vector of [2 x double].
2547 /// \param __b
2548 ///    A 128-bit vector of [2 x double].
2549 /// \returns 1 if both the ZF and CF flags are set to 0, otherwise returns 0.
2550 static __inline int __DEFAULT_FN_ATTRS128
2551 _mm_testnzc_pd(__m128d __a, __m128d __b)
2552 {
2553   return __builtin_ia32_vtestnzcpd((__v2df)__a, (__v2df)__b);
2554 }
2555 
2556 /// Given two 128-bit floating-point vectors of [4 x float], perform an
2557 ///    element-by-element comparison of the single-precision element in the
2558 ///    first source vector and the corresponding element in the second source
2559 ///    vector.
2560 ///
2561 ///    The EFLAGS register is updated as follows: \n
2562 ///    If there is at least one pair of single-precision elements where the
2563 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2564 ///    ZF flag is set to 1. \n
2565 ///    If there is at least one pair of single-precision elements where the
2566 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2567 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2568 ///    This intrinsic returns the value of the ZF flag.
2569 ///
2570 /// \headerfile <x86intrin.h>
2571 ///
2572 /// This intrinsic corresponds to the <c> VTESTPS </c> instruction.
2573 ///
2574 /// \param __a
2575 ///    A 128-bit vector of [4 x float].
2576 /// \param __b
2577 ///    A 128-bit vector of [4 x float].
2578 /// \returns the ZF flag.
2579 static __inline int __DEFAULT_FN_ATTRS128
2580 _mm_testz_ps(__m128 __a, __m128 __b)
2581 {
2582   return __builtin_ia32_vtestzps((__v4sf)__a, (__v4sf)__b);
2583 }
2584 
2585 /// Given two 128-bit floating-point vectors of [4 x float], perform an
2586 ///    element-by-element comparison of the single-precision element in the
2587 ///    first source vector and the corresponding element in the second source
2588 ///    vector.
2589 ///
2590 ///    The EFLAGS register is updated as follows: \n
2591 ///    If there is at least one pair of single-precision elements where the
2592 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2593 ///    ZF flag is set to 1. \n
2594 ///    If there is at least one pair of single-precision elements where the
2595 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2596 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2597 ///    This intrinsic returns the value of the CF flag.
2598 ///
2599 /// \headerfile <x86intrin.h>
2600 ///
2601 /// This intrinsic corresponds to the <c> VTESTPS </c> instruction.
2602 ///
2603 /// \param __a
2604 ///    A 128-bit vector of [4 x float].
2605 /// \param __b
2606 ///    A 128-bit vector of [4 x float].
2607 /// \returns the CF flag.
2608 static __inline int __DEFAULT_FN_ATTRS128
2609 _mm_testc_ps(__m128 __a, __m128 __b)
2610 {
2611   return __builtin_ia32_vtestcps((__v4sf)__a, (__v4sf)__b);
2612 }
2613 
2614 /// Given two 128-bit floating-point vectors of [4 x float], perform an
2615 ///    element-by-element comparison of the single-precision element in the
2616 ///    first source vector and the corresponding element in the second source
2617 ///    vector.
2618 ///
2619 ///    The EFLAGS register is updated as follows: \n
2620 ///    If there is at least one pair of single-precision elements where the
2621 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2622 ///    ZF flag is set to 1. \n
2623 ///    If there is at least one pair of single-precision elements where the
2624 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2625 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2626 ///    This intrinsic returns 1 if both the ZF and CF flags are set to 0,
2627 ///    otherwise it returns 0.
2628 ///
2629 /// \headerfile <x86intrin.h>
2630 ///
2631 /// This intrinsic corresponds to the <c> VTESTPS </c> instruction.
2632 ///
2633 /// \param __a
2634 ///    A 128-bit vector of [4 x float].
2635 /// \param __b
2636 ///    A 128-bit vector of [4 x float].
2637 /// \returns 1 if both the ZF and CF flags are set to 0, otherwise returns 0.
2638 static __inline int __DEFAULT_FN_ATTRS128
2639 _mm_testnzc_ps(__m128 __a, __m128 __b)
2640 {
2641   return __builtin_ia32_vtestnzcps((__v4sf)__a, (__v4sf)__b);
2642 }
2643 
2644 /// Given two 256-bit floating-point vectors of [4 x double], perform an
2645 ///    element-by-element comparison of the double-precision elements in the
2646 ///    first source vector and the corresponding elements in the second source
2647 ///    vector.
2648 ///
2649 ///    The EFLAGS register is updated as follows: \n
2650 ///    If there is at least one pair of double-precision elements where the
2651 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2652 ///    ZF flag is set to 1. \n
2653 ///    If there is at least one pair of double-precision elements where the
2654 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2655 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2656 ///    This intrinsic returns the value of the ZF flag.
2657 ///
2658 /// \headerfile <x86intrin.h>
2659 ///
2660 /// This intrinsic corresponds to the <c> VTESTPD </c> instruction.
2661 ///
2662 /// \param __a
2663 ///    A 256-bit vector of [4 x double].
2664 /// \param __b
2665 ///    A 256-bit vector of [4 x double].
2666 /// \returns the ZF flag.
2667 static __inline int __DEFAULT_FN_ATTRS
2668 _mm256_testz_pd(__m256d __a, __m256d __b)
2669 {
2670   return __builtin_ia32_vtestzpd256((__v4df)__a, (__v4df)__b);
2671 }
2672 
2673 /// Given two 256-bit floating-point vectors of [4 x double], perform an
2674 ///    element-by-element comparison of the double-precision elements in the
2675 ///    first source vector and the corresponding elements in the second source
2676 ///    vector.
2677 ///
2678 ///    The EFLAGS register is updated as follows: \n
2679 ///    If there is at least one pair of double-precision elements where the
2680 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2681 ///    ZF flag is set to 1. \n
2682 ///    If there is at least one pair of double-precision elements where the
2683 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2684 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2685 ///    This intrinsic returns the value of the CF flag.
2686 ///
2687 /// \headerfile <x86intrin.h>
2688 ///
2689 /// This intrinsic corresponds to the <c> VTESTPD </c> instruction.
2690 ///
2691 /// \param __a
2692 ///    A 256-bit vector of [4 x double].
2693 /// \param __b
2694 ///    A 256-bit vector of [4 x double].
2695 /// \returns the CF flag.
2696 static __inline int __DEFAULT_FN_ATTRS
2697 _mm256_testc_pd(__m256d __a, __m256d __b)
2698 {
2699   return __builtin_ia32_vtestcpd256((__v4df)__a, (__v4df)__b);
2700 }
2701 
2702 /// Given two 256-bit floating-point vectors of [4 x double], perform an
2703 ///    element-by-element comparison of the double-precision elements in the
2704 ///    first source vector and the corresponding elements in the second source
2705 ///    vector.
2706 ///
2707 ///    The EFLAGS register is updated as follows: \n
2708 ///    If there is at least one pair of double-precision elements where the
2709 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2710 ///    ZF flag is set to 1. \n
2711 ///    If there is at least one pair of double-precision elements where the
2712 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2713 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2714 ///    This intrinsic returns 1 if both the ZF and CF flags are set to 0,
2715 ///    otherwise it returns 0.
2716 ///
2717 /// \headerfile <x86intrin.h>
2718 ///
2719 /// This intrinsic corresponds to the <c> VTESTPD </c> instruction.
2720 ///
2721 /// \param __a
2722 ///    A 256-bit vector of [4 x double].
2723 /// \param __b
2724 ///    A 256-bit vector of [4 x double].
2725 /// \returns 1 if both the ZF and CF flags are set to 0, otherwise returns 0.
2726 static __inline int __DEFAULT_FN_ATTRS
2727 _mm256_testnzc_pd(__m256d __a, __m256d __b)
2728 {
2729   return __builtin_ia32_vtestnzcpd256((__v4df)__a, (__v4df)__b);
2730 }
2731 
2732 /// Given two 256-bit floating-point vectors of [8 x float], perform an
2733 ///    element-by-element comparison of the single-precision element in the
2734 ///    first source vector and the corresponding element in the second source
2735 ///    vector.
2736 ///
2737 ///    The EFLAGS register is updated as follows: \n
2738 ///    If there is at least one pair of single-precision elements where the
2739 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2740 ///    ZF flag is set to 1. \n
2741 ///    If there is at least one pair of single-precision elements where the
2742 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2743 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2744 ///    This intrinsic returns the value of the ZF flag.
2745 ///
2746 /// \headerfile <x86intrin.h>
2747 ///
2748 /// This intrinsic corresponds to the <c> VTESTPS </c> instruction.
2749 ///
2750 /// \param __a
2751 ///    A 256-bit vector of [8 x float].
2752 /// \param __b
2753 ///    A 256-bit vector of [8 x float].
2754 /// \returns the ZF flag.
2755 static __inline int __DEFAULT_FN_ATTRS
2756 _mm256_testz_ps(__m256 __a, __m256 __b)
2757 {
2758   return __builtin_ia32_vtestzps256((__v8sf)__a, (__v8sf)__b);
2759 }
2760 
2761 /// Given two 256-bit floating-point vectors of [8 x float], perform an
2762 ///    element-by-element comparison of the single-precision element in the
2763 ///    first source vector and the corresponding element in the second source
2764 ///    vector.
2765 ///
2766 ///    The EFLAGS register is updated as follows: \n
2767 ///    If there is at least one pair of single-precision elements where the
2768 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2769 ///    ZF flag is set to 1. \n
2770 ///    If there is at least one pair of single-precision elements where the
2771 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2772 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2773 ///    This intrinsic returns the value of the CF flag.
2774 ///
2775 /// \headerfile <x86intrin.h>
2776 ///
2777 /// This intrinsic corresponds to the <c> VTESTPS </c> instruction.
2778 ///
2779 /// \param __a
2780 ///    A 256-bit vector of [8 x float].
2781 /// \param __b
2782 ///    A 256-bit vector of [8 x float].
2783 /// \returns the CF flag.
2784 static __inline int __DEFAULT_FN_ATTRS
2785 _mm256_testc_ps(__m256 __a, __m256 __b)
2786 {
2787   return __builtin_ia32_vtestcps256((__v8sf)__a, (__v8sf)__b);
2788 }
2789 
2790 /// Given two 256-bit floating-point vectors of [8 x float], perform an
2791 ///    element-by-element comparison of the single-precision elements in the
2792 ///    first source vector and the corresponding elements in the second source
2793 ///    vector.
2794 ///
2795 ///    The EFLAGS register is updated as follows: \n
2796 ///    If there is at least one pair of single-precision elements where the
2797 ///    sign-bits of both elements are 1, the ZF flag is set to 0. Otherwise the
2798 ///    ZF flag is set to 1. \n
2799 ///    If there is at least one pair of single-precision elements where the
2800 ///    sign-bit of the first element is 0 and the sign-bit of the second element
2801 ///    is 1, the CF flag is set to 0. Otherwise the CF flag is set to 1. \n
2802 ///    This intrinsic returns 1 if both the ZF and CF flags are set to 0,
2803 ///    otherwise it returns 0.
2804 ///
2805 /// \headerfile <x86intrin.h>
2806 ///
2807 /// This intrinsic corresponds to the <c> VTESTPS </c> instruction.
2808 ///
2809 /// \param __a
2810 ///    A 256-bit vector of [8 x float].
2811 /// \param __b
2812 ///    A 256-bit vector of [8 x float].
2813 /// \returns 1 if both the ZF and CF flags are set to 0, otherwise returns 0.
2814 static __inline int __DEFAULT_FN_ATTRS
2815 _mm256_testnzc_ps(__m256 __a, __m256 __b)
2816 {
2817   return __builtin_ia32_vtestnzcps256((__v8sf)__a, (__v8sf)__b);
2818 }
2819 
2820 /// Given two 256-bit integer vectors, perform a bit-by-bit comparison
2821 ///    of the two source vectors.
2822 ///
2823 ///    The EFLAGS register is updated as follows: \n
2824 ///    If there is at least one pair of bits where both bits are 1, the ZF flag
2825 ///    is set to 0. Otherwise the ZF flag is set to 1. \n
2826 ///    If there is at least one pair of bits where the bit from the first source
2827 ///    vector is 0 and the bit from the second source vector is 1, the CF flag
2828 ///    is set to 0. Otherwise the CF flag is set to 1. \n
2829 ///    This intrinsic returns the value of the ZF flag.
2830 ///
2831 /// \headerfile <x86intrin.h>
2832 ///
2833 /// This intrinsic corresponds to the <c> VPTEST </c> instruction.
2834 ///
2835 /// \param __a
2836 ///    A 256-bit integer vector.
2837 /// \param __b
2838 ///    A 256-bit integer vector.
2839 /// \returns the ZF flag.
2840 static __inline int __DEFAULT_FN_ATTRS
2841 _mm256_testz_si256(__m256i __a, __m256i __b)
2842 {
2843   return __builtin_ia32_ptestz256((__v4di)__a, (__v4di)__b);
2844 }
2845 
2846 /// Given two 256-bit integer vectors, perform a bit-by-bit comparison
2847 ///    of the two source vectors.
2848 ///
2849 ///    The EFLAGS register is updated as follows: \n
2850 ///    If there is at least one pair of bits where both bits are 1, the ZF flag
2851 ///    is set to 0. Otherwise the ZF flag is set to 1. \n
2852 ///    If there is at least one pair of bits where the bit from the first source
2853 ///    vector is 0 and the bit from the second source vector is 1, the CF flag
2854 ///    is set to 0. Otherwise the CF flag is set to 1. \n
2855 ///    This intrinsic returns the value of the CF flag.
2856 ///
2857 /// \headerfile <x86intrin.h>
2858 ///
2859 /// This intrinsic corresponds to the <c> VPTEST </c> instruction.
2860 ///
2861 /// \param __a
2862 ///    A 256-bit integer vector.
2863 /// \param __b
2864 ///    A 256-bit integer vector.
2865 /// \returns the CF flag.
2866 static __inline int __DEFAULT_FN_ATTRS
2867 _mm256_testc_si256(__m256i __a, __m256i __b)
2868 {
2869   return __builtin_ia32_ptestc256((__v4di)__a, (__v4di)__b);
2870 }
2871 
2872 /// Given two 256-bit integer vectors, perform a bit-by-bit comparison
2873 ///    of the two source vectors.
2874 ///
2875 ///    The EFLAGS register is updated as follows: \n
2876 ///    If there is at least one pair of bits where both bits are 1, the ZF flag
2877 ///    is set to 0. Otherwise the ZF flag is set to 1. \n
2878 ///    If there is at least one pair of bits where the bit from the first source
2879 ///    vector is 0 and the bit from the second source vector is 1, the CF flag
2880 ///    is set to 0. Otherwise the CF flag is set to 1. \n
2881 ///    This intrinsic returns 1 if both the ZF and CF flags are set to 0,
2882 ///    otherwise it returns 0.
2883 ///
2884 /// \headerfile <x86intrin.h>
2885 ///
2886 /// This intrinsic corresponds to the <c> VPTEST </c> instruction.
2887 ///
2888 /// \param __a
2889 ///    A 256-bit integer vector.
2890 /// \param __b
2891 ///    A 256-bit integer vector.
2892 /// \returns 1 if both the ZF and CF flags are set to 0, otherwise returns 0.
2893 static __inline int __DEFAULT_FN_ATTRS
2894 _mm256_testnzc_si256(__m256i __a, __m256i __b)
2895 {
2896   return __builtin_ia32_ptestnzc256((__v4di)__a, (__v4di)__b);
2897 }
2898 
2899 /* Vector extract sign mask */
2900 /// Extracts the sign bits of double-precision floating point elements
2901 ///    in a 256-bit vector of [4 x double] and writes them to the lower order
2902 ///    bits of the return value.
2903 ///
2904 /// \headerfile <x86intrin.h>
2905 ///
2906 /// This intrinsic corresponds to the <c> VMOVMSKPD </c> instruction.
2907 ///
2908 /// \param __a
2909 ///    A 256-bit vector of [4 x double] containing the double-precision
2910 ///    floating point values with sign bits to be extracted.
2911 /// \returns The sign bits from the operand, written to bits [3:0].
2912 static __inline int __DEFAULT_FN_ATTRS
2913 _mm256_movemask_pd(__m256d __a)
2914 {
2915   return __builtin_ia32_movmskpd256((__v4df)__a);
2916 }
2917 
2918 /// Extracts the sign bits of single-precision floating point elements
2919 ///    in a 256-bit vector of [8 x float] and writes them to the lower order
2920 ///    bits of the return value.
2921 ///
2922 /// \headerfile <x86intrin.h>
2923 ///
2924 /// This intrinsic corresponds to the <c> VMOVMSKPS </c> instruction.
2925 ///
2926 /// \param __a
2927 ///    A 256-bit vector of [8 x float] containing the single-precision floating
2928 ///    point values with sign bits to be extracted.
2929 /// \returns The sign bits from the operand, written to bits [7:0].
2930 static __inline int __DEFAULT_FN_ATTRS
2931 _mm256_movemask_ps(__m256 __a)
2932 {
2933   return __builtin_ia32_movmskps256((__v8sf)__a);
2934 }
2935 
2936 /* Vector __zero */
2937 /// Zeroes the contents of all XMM or YMM registers.
2938 ///
2939 /// \headerfile <x86intrin.h>
2940 ///
2941 /// This intrinsic corresponds to the <c> VZEROALL </c> instruction.
2942 static __inline void __attribute__((__always_inline__, __nodebug__, __target__("avx")))
2943 _mm256_zeroall(void)
2944 {
2945   __builtin_ia32_vzeroall();
2946 }
2947 
2948 /// Zeroes the upper 128 bits (bits 255:128) of all YMM registers.
2949 ///
2950 /// \headerfile <x86intrin.h>
2951 ///
2952 /// This intrinsic corresponds to the <c> VZEROUPPER </c> instruction.
2953 static __inline void __attribute__((__always_inline__, __nodebug__, __target__("avx")))
2954 _mm256_zeroupper(void)
2955 {
2956   __builtin_ia32_vzeroupper();
2957 }
2958 
2959 /* Vector load with broadcast */
2960 /// Loads a scalar single-precision floating point value from the
2961 ///    specified address pointed to by \a __a and broadcasts it to the elements
2962 ///    of a [4 x float] vector.
2963 ///
2964 /// \headerfile <x86intrin.h>
2965 ///
2966 /// This intrinsic corresponds to the <c> VBROADCASTSS </c> instruction.
2967 ///
2968 /// \param __a
2969 ///    The single-precision floating point value to be broadcast.
2970 /// \returns A 128-bit vector of [4 x float] whose 32-bit elements are set
2971 ///    equal to the broadcast value.
2972 static __inline __m128 __DEFAULT_FN_ATTRS128
2973 _mm_broadcast_ss(float const *__a)
2974 {
2975   float __f = *__a;
2976   return __extension__ (__m128)(__v4sf){ __f, __f, __f, __f };
2977 }
2978 
2979 /// Loads a scalar double-precision floating point value from the
2980 ///    specified address pointed to by \a __a and broadcasts it to the elements
2981 ///    of a [4 x double] vector.
2982 ///
2983 /// \headerfile <x86intrin.h>
2984 ///
2985 /// This intrinsic corresponds to the <c> VBROADCASTSD </c> instruction.
2986 ///
2987 /// \param __a
2988 ///    The double-precision floating point value to be broadcast.
2989 /// \returns A 256-bit vector of [4 x double] whose 64-bit elements are set
2990 ///    equal to the broadcast value.
2991 static __inline __m256d __DEFAULT_FN_ATTRS
2992 _mm256_broadcast_sd(double const *__a)
2993 {
2994   double __d = *__a;
2995   return __extension__ (__m256d)(__v4df){ __d, __d, __d, __d };
2996 }
2997 
2998 /// Loads a scalar single-precision floating point value from the
2999 ///    specified address pointed to by \a __a and broadcasts it to the elements
3000 ///    of a [8 x float] vector.
3001 ///
3002 /// \headerfile <x86intrin.h>
3003 ///
3004 /// This intrinsic corresponds to the <c> VBROADCASTSS </c> instruction.
3005 ///
3006 /// \param __a
3007 ///    The single-precision floating point value to be broadcast.
3008 /// \returns A 256-bit vector of [8 x float] whose 32-bit elements are set
3009 ///    equal to the broadcast value.
3010 static __inline __m256 __DEFAULT_FN_ATTRS
3011 _mm256_broadcast_ss(float const *__a)
3012 {
3013   float __f = *__a;
3014   return __extension__ (__m256)(__v8sf){ __f, __f, __f, __f, __f, __f, __f, __f };
3015 }
3016 
3017 /// Loads the data from a 128-bit vector of [2 x double] from the
3018 ///    specified address pointed to by \a __a and broadcasts it to 128-bit
3019 ///    elements in a 256-bit vector of [4 x double].
3020 ///
3021 /// \headerfile <x86intrin.h>
3022 ///
3023 /// This intrinsic corresponds to the <c> VBROADCASTF128 </c> instruction.
3024 ///
3025 /// \param __a
3026 ///    The 128-bit vector of [2 x double] to be broadcast.
3027 /// \returns A 256-bit vector of [4 x double] whose 128-bit elements are set
3028 ///    equal to the broadcast value.
3029 static __inline __m256d __DEFAULT_FN_ATTRS
3030 _mm256_broadcast_pd(__m128d const *__a)
3031 {
3032   __m128d __b = _mm_loadu_pd((const double *)__a);
3033   return (__m256d)__builtin_shufflevector((__v2df)__b, (__v2df)__b,
3034                                           0, 1, 0, 1);
3035 }
3036 
3037 /// Loads the data from a 128-bit vector of [4 x float] from the
3038 ///    specified address pointed to by \a __a and broadcasts it to 128-bit
3039 ///    elements in a 256-bit vector of [8 x float].
3040 ///
3041 /// \headerfile <x86intrin.h>
3042 ///
3043 /// This intrinsic corresponds to the <c> VBROADCASTF128 </c> instruction.
3044 ///
3045 /// \param __a
3046 ///    The 128-bit vector of [4 x float] to be broadcast.
3047 /// \returns A 256-bit vector of [8 x float] whose 128-bit elements are set
3048 ///    equal to the broadcast value.
3049 static __inline __m256 __DEFAULT_FN_ATTRS
3050 _mm256_broadcast_ps(__m128 const *__a)
3051 {
3052   __m128 __b = _mm_loadu_ps((const float *)__a);
3053   return (__m256)__builtin_shufflevector((__v4sf)__b, (__v4sf)__b,
3054                                          0, 1, 2, 3, 0, 1, 2, 3);
3055 }
3056 
3057 /* SIMD load ops */
3058 /// Loads 4 double-precision floating point values from a 32-byte aligned
3059 ///    memory location pointed to by \a __p into a vector of [4 x double].
3060 ///
3061 /// \headerfile <x86intrin.h>
3062 ///
3063 /// This intrinsic corresponds to the <c> VMOVAPD </c> instruction.
3064 ///
3065 /// \param __p
3066 ///    A 32-byte aligned pointer to a memory location containing
3067 ///    double-precision floating point values.
3068 /// \returns A 256-bit vector of [4 x double] containing the moved values.
3069 static __inline __m256d __DEFAULT_FN_ATTRS
3070 _mm256_load_pd(double const *__p)
3071 {
3072   return *(const __m256d *)__p;
3073 }
3074 
3075 /// Loads 8 single-precision floating point values from a 32-byte aligned
3076 ///    memory location pointed to by \a __p into a vector of [8 x float].
3077 ///
3078 /// \headerfile <x86intrin.h>
3079 ///
3080 /// This intrinsic corresponds to the <c> VMOVAPS </c> instruction.
3081 ///
3082 /// \param __p
3083 ///    A 32-byte aligned pointer to a memory location containing float values.
3084 /// \returns A 256-bit vector of [8 x float] containing the moved values.
3085 static __inline __m256 __DEFAULT_FN_ATTRS
3086 _mm256_load_ps(float const *__p)
3087 {
3088   return *(const __m256 *)__p;
3089 }
3090 
3091 /// Loads 4 double-precision floating point values from an unaligned
3092 ///    memory location pointed to by \a __p into a vector of [4 x double].
3093 ///
3094 /// \headerfile <x86intrin.h>
3095 ///
3096 /// This intrinsic corresponds to the <c> VMOVUPD </c> instruction.
3097 ///
3098 /// \param __p
3099 ///    A pointer to a memory location containing double-precision floating
3100 ///    point values.
3101 /// \returns A 256-bit vector of [4 x double] containing the moved values.
3102 static __inline __m256d __DEFAULT_FN_ATTRS
3103 _mm256_loadu_pd(double const *__p)
3104 {
3105   struct __loadu_pd {
3106     __m256d_u __v;
3107   } __attribute__((__packed__, __may_alias__));
3108   return ((const struct __loadu_pd*)__p)->__v;
3109 }
3110 
3111 /// Loads 8 single-precision floating point values from an unaligned
3112 ///    memory location pointed to by \a __p into a vector of [8 x float].
3113 ///
3114 /// \headerfile <x86intrin.h>
3115 ///
3116 /// This intrinsic corresponds to the <c> VMOVUPS </c> instruction.
3117 ///
3118 /// \param __p
3119 ///    A pointer to a memory location containing single-precision floating
3120 ///    point values.
3121 /// \returns A 256-bit vector of [8 x float] containing the moved values.
3122 static __inline __m256 __DEFAULT_FN_ATTRS
3123 _mm256_loadu_ps(float const *__p)
3124 {
3125   struct __loadu_ps {
3126     __m256_u __v;
3127   } __attribute__((__packed__, __may_alias__));
3128   return ((const struct __loadu_ps*)__p)->__v;
3129 }
3130 
3131 /// Loads 256 bits of integer data from a 32-byte aligned memory
3132 ///    location pointed to by \a __p into elements of a 256-bit integer vector.
3133 ///
3134 /// \headerfile <x86intrin.h>
3135 ///
3136 /// This intrinsic corresponds to the <c> VMOVDQA </c> instruction.
3137 ///
3138 /// \param __p
3139 ///    A 32-byte aligned pointer to a 256-bit integer vector containing integer
3140 ///    values.
3141 /// \returns A 256-bit integer vector containing the moved values.
3142 static __inline __m256i __DEFAULT_FN_ATTRS
3143 _mm256_load_si256(__m256i const *__p)
3144 {
3145   return *__p;
3146 }
3147 
3148 /// Loads 256 bits of integer data from an unaligned memory location
3149 ///    pointed to by \a __p into a 256-bit integer vector.
3150 ///
3151 /// \headerfile <x86intrin.h>
3152 ///
3153 /// This intrinsic corresponds to the <c> VMOVDQU </c> instruction.
3154 ///
3155 /// \param __p
3156 ///    A pointer to a 256-bit integer vector containing integer values.
3157 /// \returns A 256-bit integer vector containing the moved values.
3158 static __inline __m256i __DEFAULT_FN_ATTRS
3159 _mm256_loadu_si256(__m256i_u const *__p)
3160 {
3161   struct __loadu_si256 {
3162     __m256i_u __v;
3163   } __attribute__((__packed__, __may_alias__));
3164   return ((const struct __loadu_si256*)__p)->__v;
3165 }
3166 
3167 /// Loads 256 bits of integer data from an unaligned memory location
3168 ///    pointed to by \a __p into a 256-bit integer vector. This intrinsic may
3169 ///    perform better than \c _mm256_loadu_si256 when the data crosses a cache
3170 ///    line boundary.
3171 ///
3172 /// \headerfile <x86intrin.h>
3173 ///
3174 /// This intrinsic corresponds to the <c> VLDDQU </c> instruction.
3175 ///
3176 /// \param __p
3177 ///    A pointer to a 256-bit integer vector containing integer values.
3178 /// \returns A 256-bit integer vector containing the moved values.
3179 static __inline __m256i __DEFAULT_FN_ATTRS
3180 _mm256_lddqu_si256(__m256i const *__p)
3181 {
3182   return (__m256i)__builtin_ia32_lddqu256((char const *)__p);
3183 }
3184 
3185 /* SIMD store ops */
3186 /// Stores double-precision floating point values from a 256-bit vector
3187 ///    of [4 x double] to a 32-byte aligned memory location pointed to by
3188 ///    \a __p.
3189 ///
3190 /// \headerfile <x86intrin.h>
3191 ///
3192 /// This intrinsic corresponds to the <c> VMOVAPD </c> instruction.
3193 ///
3194 /// \param __p
3195 ///    A 32-byte aligned pointer to a memory location that will receive the
3196 ///    double-precision floaing point values.
3197 /// \param __a
3198 ///    A 256-bit vector of [4 x double] containing the values to be moved.
3199 static __inline void __DEFAULT_FN_ATTRS
3200 _mm256_store_pd(double *__p, __m256d __a)
3201 {
3202   *(__m256d *)__p = __a;
3203 }
3204 
3205 /// Stores single-precision floating point values from a 256-bit vector
3206 ///    of [8 x float] to a 32-byte aligned memory location pointed to by \a __p.
3207 ///
3208 /// \headerfile <x86intrin.h>
3209 ///
3210 /// This intrinsic corresponds to the <c> VMOVAPS </c> instruction.
3211 ///
3212 /// \param __p
3213 ///    A 32-byte aligned pointer to a memory location that will receive the
3214 ///    float values.
3215 /// \param __a
3216 ///    A 256-bit vector of [8 x float] containing the values to be moved.
3217 static __inline void __DEFAULT_FN_ATTRS
3218 _mm256_store_ps(float *__p, __m256 __a)
3219 {
3220   *(__m256 *)__p = __a;
3221 }
3222 
3223 /// Stores double-precision floating point values from a 256-bit vector
3224 ///    of [4 x double] to an unaligned memory location pointed to by \a __p.
3225 ///
3226 /// \headerfile <x86intrin.h>
3227 ///
3228 /// This intrinsic corresponds to the <c> VMOVUPD </c> instruction.
3229 ///
3230 /// \param __p
3231 ///    A pointer to a memory location that will receive the double-precision
3232 ///    floating point values.
3233 /// \param __a
3234 ///    A 256-bit vector of [4 x double] containing the values to be moved.
3235 static __inline void __DEFAULT_FN_ATTRS
3236 _mm256_storeu_pd(double *__p, __m256d __a)
3237 {
3238   struct __storeu_pd {
3239     __m256d_u __v;
3240   } __attribute__((__packed__, __may_alias__));
3241   ((struct __storeu_pd*)__p)->__v = __a;
3242 }
3243 
3244 /// Stores single-precision floating point values from a 256-bit vector
3245 ///    of [8 x float] to an unaligned memory location pointed to by \a __p.
3246 ///
3247 /// \headerfile <x86intrin.h>
3248 ///
3249 /// This intrinsic corresponds to the <c> VMOVUPS </c> instruction.
3250 ///
3251 /// \param __p
3252 ///    A pointer to a memory location that will receive the float values.
3253 /// \param __a
3254 ///    A 256-bit vector of [8 x float] containing the values to be moved.
3255 static __inline void __DEFAULT_FN_ATTRS
3256 _mm256_storeu_ps(float *__p, __m256 __a)
3257 {
3258   struct __storeu_ps {
3259     __m256_u __v;
3260   } __attribute__((__packed__, __may_alias__));
3261   ((struct __storeu_ps*)__p)->__v = __a;
3262 }
3263 
3264 /// Stores integer values from a 256-bit integer vector to a 32-byte
3265 ///    aligned memory location pointed to by \a __p.
3266 ///
3267 /// \headerfile <x86intrin.h>
3268 ///
3269 /// This intrinsic corresponds to the <c> VMOVDQA </c> instruction.
3270 ///
3271 /// \param __p
3272 ///    A 32-byte aligned pointer to a memory location that will receive the
3273 ///    integer values.
3274 /// \param __a
3275 ///    A 256-bit integer vector containing the values to be moved.
3276 static __inline void __DEFAULT_FN_ATTRS
3277 _mm256_store_si256(__m256i *__p, __m256i __a)
3278 {
3279   *__p = __a;
3280 }
3281 
3282 /// Stores integer values from a 256-bit integer vector to an unaligned
3283 ///    memory location pointed to by \a __p.
3284 ///
3285 /// \headerfile <x86intrin.h>
3286 ///
3287 /// This intrinsic corresponds to the <c> VMOVDQU </c> instruction.
3288 ///
3289 /// \param __p
3290 ///    A pointer to a memory location that will receive the integer values.
3291 /// \param __a
3292 ///    A 256-bit integer vector containing the values to be moved.
3293 static __inline void __DEFAULT_FN_ATTRS
3294 _mm256_storeu_si256(__m256i_u *__p, __m256i __a)
3295 {
3296   struct __storeu_si256 {
3297     __m256i_u __v;
3298   } __attribute__((__packed__, __may_alias__));
3299   ((struct __storeu_si256*)__p)->__v = __a;
3300 }
3301 
3302 /* Conditional load ops */
3303 /// Conditionally loads double-precision floating point elements from a
3304 ///    memory location pointed to by \a __p into a 128-bit vector of
3305 ///    [2 x double], depending on the mask bits associated with each data
3306 ///    element.
3307 ///
3308 /// \headerfile <x86intrin.h>
3309 ///
3310 /// This intrinsic corresponds to the <c> VMASKMOVPD </c> instruction.
3311 ///
3312 /// \param __p
3313 ///    A pointer to a memory location that contains the double-precision
3314 ///    floating point values.
3315 /// \param __m
3316 ///    A 128-bit integer vector containing the mask. The most significant bit of
3317 ///    each data element represents the mask bits. If a mask bit is zero, the
3318 ///    corresponding value in the memory location is not loaded and the
3319 ///    corresponding field in the return value is set to zero.
3320 /// \returns A 128-bit vector of [2 x double] containing the loaded values.
3321 static __inline __m128d __DEFAULT_FN_ATTRS128
3322 _mm_maskload_pd(double const *__p, __m128i __m)
3323 {
3324   return (__m128d)__builtin_ia32_maskloadpd((const __v2df *)__p, (__v2di)__m);
3325 }
3326 
3327 /// Conditionally loads double-precision floating point elements from a
3328 ///    memory location pointed to by \a __p into a 256-bit vector of
3329 ///    [4 x double], depending on the mask bits associated with each data
3330 ///    element.
3331 ///
3332 /// \headerfile <x86intrin.h>
3333 ///
3334 /// This intrinsic corresponds to the <c> VMASKMOVPD </c> instruction.
3335 ///
3336 /// \param __p
3337 ///    A pointer to a memory location that contains the double-precision
3338 ///    floating point values.
3339 /// \param __m
3340 ///    A 256-bit integer vector of [4 x quadword] containing the mask. The most
3341 ///    significant bit of each quadword element represents the mask bits. If a
3342 ///    mask bit is zero, the corresponding value in the memory location is not
3343 ///    loaded and the corresponding field in the return value is set to zero.
3344 /// \returns A 256-bit vector of [4 x double] containing the loaded values.
3345 static __inline __m256d __DEFAULT_FN_ATTRS
3346 _mm256_maskload_pd(double const *__p, __m256i __m)
3347 {
3348   return (__m256d)__builtin_ia32_maskloadpd256((const __v4df *)__p,
3349                                                (__v4di)__m);
3350 }
3351 
3352 /// Conditionally loads single-precision floating point elements from a
3353 ///    memory location pointed to by \a __p into a 128-bit vector of
3354 ///    [4 x float], depending on the mask bits associated with each data
3355 ///    element.
3356 ///
3357 /// \headerfile <x86intrin.h>
3358 ///
3359 /// This intrinsic corresponds to the <c> VMASKMOVPS </c> instruction.
3360 ///
3361 /// \param __p
3362 ///    A pointer to a memory location that contains the single-precision
3363 ///    floating point values.
3364 /// \param __m
3365 ///    A 128-bit integer vector containing the mask. The most significant bit of
3366 ///    each data element represents the mask bits. If a mask bit is zero, the
3367 ///    corresponding value in the memory location is not loaded and the
3368 ///    corresponding field in the return value is set to zero.
3369 /// \returns A 128-bit vector of [4 x float] containing the loaded values.
3370 static __inline __m128 __DEFAULT_FN_ATTRS128
3371 _mm_maskload_ps(float const *__p, __m128i __m)
3372 {
3373   return (__m128)__builtin_ia32_maskloadps((const __v4sf *)__p, (__v4si)__m);
3374 }
3375 
3376 /// Conditionally loads single-precision floating point elements from a
3377 ///    memory location pointed to by \a __p into a 256-bit vector of
3378 ///    [8 x float], depending on the mask bits associated with each data
3379 ///    element.
3380 ///
3381 /// \headerfile <x86intrin.h>
3382 ///
3383 /// This intrinsic corresponds to the <c> VMASKMOVPS </c> instruction.
3384 ///
3385 /// \param __p
3386 ///    A pointer to a memory location that contains the single-precision
3387 ///    floating point values.
3388 /// \param __m
3389 ///    A 256-bit integer vector of [8 x dword] containing the mask. The most
3390 ///    significant bit of each dword element represents the mask bits. If a mask
3391 ///    bit is zero, the corresponding value in the memory location is not loaded
3392 ///    and the corresponding field in the return value is set to zero.
3393 /// \returns A 256-bit vector of [8 x float] containing the loaded values.
3394 static __inline __m256 __DEFAULT_FN_ATTRS
3395 _mm256_maskload_ps(float const *__p, __m256i __m)
3396 {
3397   return (__m256)__builtin_ia32_maskloadps256((const __v8sf *)__p, (__v8si)__m);
3398 }
3399 
3400 /* Conditional store ops */
3401 /// Moves single-precision floating point values from a 256-bit vector
3402 ///    of [8 x float] to a memory location pointed to by \a __p, according to
3403 ///    the specified mask.
3404 ///
3405 /// \headerfile <x86intrin.h>
3406 ///
3407 /// This intrinsic corresponds to the <c> VMASKMOVPS </c> instruction.
3408 ///
3409 /// \param __p
3410 ///    A pointer to a memory location that will receive the float values.
3411 /// \param __m
3412 ///    A 256-bit integer vector of [8 x dword] containing the mask. The most
3413 ///    significant bit of each dword element in the mask vector represents the
3414 ///    mask bits. If a mask bit is zero, the corresponding value from vector
3415 ///    \a __a is not stored and the corresponding field in the memory location
3416 ///    pointed to by \a __p is not changed.
3417 /// \param __a
3418 ///    A 256-bit vector of [8 x float] containing the values to be stored.
3419 static __inline void __DEFAULT_FN_ATTRS
3420 _mm256_maskstore_ps(float *__p, __m256i __m, __m256 __a)
3421 {
3422   __builtin_ia32_maskstoreps256((__v8sf *)__p, (__v8si)__m, (__v8sf)__a);
3423 }
3424 
3425 /// Moves double-precision values from a 128-bit vector of [2 x double]
3426 ///    to a memory location pointed to by \a __p, according to the specified
3427 ///    mask.
3428 ///
3429 /// \headerfile <x86intrin.h>
3430 ///
3431 /// This intrinsic corresponds to the <c> VMASKMOVPD </c> instruction.
3432 ///
3433 /// \param __p
3434 ///    A pointer to a memory location that will receive the float values.
3435 /// \param __m
3436 ///    A 128-bit integer vector containing the mask. The most significant bit of
3437 ///    each field in the mask vector represents the mask bits. If a mask bit is
3438 ///    zero, the corresponding value from vector \a __a is not stored and the
3439 ///    corresponding field in the memory location pointed to by \a __p is not
3440 ///    changed.
3441 /// \param __a
3442 ///    A 128-bit vector of [2 x double] containing the values to be stored.
3443 static __inline void __DEFAULT_FN_ATTRS128
3444 _mm_maskstore_pd(double *__p, __m128i __m, __m128d __a)
3445 {
3446   __builtin_ia32_maskstorepd((__v2df *)__p, (__v2di)__m, (__v2df)__a);
3447 }
3448 
3449 /// Moves double-precision values from a 256-bit vector of [4 x double]
3450 ///    to a memory location pointed to by \a __p, according to the specified
3451 ///    mask.
3452 ///
3453 /// \headerfile <x86intrin.h>
3454 ///
3455 /// This intrinsic corresponds to the <c> VMASKMOVPD </c> instruction.
3456 ///
3457 /// \param __p
3458 ///    A pointer to a memory location that will receive the float values.
3459 /// \param __m
3460 ///    A 256-bit integer vector of [4 x quadword] containing the mask. The most
3461 ///    significant bit of each quadword element in the mask vector represents
3462 ///    the mask bits. If a mask bit is zero, the corresponding value from vector
3463 ///    __a is not stored and the corresponding field in the memory location
3464 ///    pointed to by \a __p is not changed.
3465 /// \param __a
3466 ///    A 256-bit vector of [4 x double] containing the values to be stored.
3467 static __inline void __DEFAULT_FN_ATTRS
3468 _mm256_maskstore_pd(double *__p, __m256i __m, __m256d __a)
3469 {
3470   __builtin_ia32_maskstorepd256((__v4df *)__p, (__v4di)__m, (__v4df)__a);
3471 }
3472 
3473 /// Moves single-precision floating point values from a 128-bit vector
3474 ///    of [4 x float] to a memory location pointed to by \a __p, according to
3475 ///    the specified mask.
3476 ///
3477 /// \headerfile <x86intrin.h>
3478 ///
3479 /// This intrinsic corresponds to the <c> VMASKMOVPS </c> instruction.
3480 ///
3481 /// \param __p
3482 ///    A pointer to a memory location that will receive the float values.
3483 /// \param __m
3484 ///    A 128-bit integer vector containing the mask. The most significant bit of
3485 ///    each field in the mask vector represents the mask bits. If a mask bit is
3486 ///    zero, the corresponding value from vector __a is not stored and the
3487 ///    corresponding field in the memory location pointed to by \a __p is not
3488 ///    changed.
3489 /// \param __a
3490 ///    A 128-bit vector of [4 x float] containing the values to be stored.
3491 static __inline void __DEFAULT_FN_ATTRS128
3492 _mm_maskstore_ps(float *__p, __m128i __m, __m128 __a)
3493 {
3494   __builtin_ia32_maskstoreps((__v4sf *)__p, (__v4si)__m, (__v4sf)__a);
3495 }
3496 
3497 /* Cacheability support ops */
3498 /// Moves integer data from a 256-bit integer vector to a 32-byte
3499 ///    aligned memory location. To minimize caching, the data is flagged as
3500 ///    non-temporal (unlikely to be used again soon).
3501 ///
3502 /// \headerfile <x86intrin.h>
3503 ///
3504 /// This intrinsic corresponds to the <c> VMOVNTDQ </c> instruction.
3505 ///
3506 /// \param __a
3507 ///    A pointer to a 32-byte aligned memory location that will receive the
3508 ///    integer values.
3509 /// \param __b
3510 ///    A 256-bit integer vector containing the values to be moved.
3511 static __inline void __DEFAULT_FN_ATTRS
3512 _mm256_stream_si256(__m256i *__a, __m256i __b)
3513 {
3514   typedef __v4di __v4di_aligned __attribute__((aligned(32)));
3515   __builtin_nontemporal_store((__v4di_aligned)__b, (__v4di_aligned*)__a);
3516 }
3517 
3518 /// Moves double-precision values from a 256-bit vector of [4 x double]
3519 ///    to a 32-byte aligned memory location. To minimize caching, the data is
3520 ///    flagged as non-temporal (unlikely to be used again soon).
3521 ///
3522 /// \headerfile <x86intrin.h>
3523 ///
3524 /// This intrinsic corresponds to the <c> VMOVNTPD </c> instruction.
3525 ///
3526 /// \param __a
3527 ///    A pointer to a 32-byte aligned memory location that will receive the
3528 ///    double-precision floating-point values.
3529 /// \param __b
3530 ///    A 256-bit vector of [4 x double] containing the values to be moved.
3531 static __inline void __DEFAULT_FN_ATTRS
3532 _mm256_stream_pd(double *__a, __m256d __b)
3533 {
3534   typedef __v4df __v4df_aligned __attribute__((aligned(32)));
3535   __builtin_nontemporal_store((__v4df_aligned)__b, (__v4df_aligned*)__a);
3536 }
3537 
3538 /// Moves single-precision floating point values from a 256-bit vector
3539 ///    of [8 x float] to a 32-byte aligned memory location. To minimize
3540 ///    caching, the data is flagged as non-temporal (unlikely to be used again
3541 ///    soon).
3542 ///
3543 /// \headerfile <x86intrin.h>
3544 ///
3545 /// This intrinsic corresponds to the <c> VMOVNTPS </c> instruction.
3546 ///
3547 /// \param __p
3548 ///    A pointer to a 32-byte aligned memory location that will receive the
3549 ///    single-precision floating point values.
3550 /// \param __a
3551 ///    A 256-bit vector of [8 x float] containing the values to be moved.
3552 static __inline void __DEFAULT_FN_ATTRS
3553 _mm256_stream_ps(float *__p, __m256 __a)
3554 {
3555   typedef __v8sf __v8sf_aligned __attribute__((aligned(32)));
3556   __builtin_nontemporal_store((__v8sf_aligned)__a, (__v8sf_aligned*)__p);
3557 }
3558 
3559 /* Create vectors */
3560 /// Create a 256-bit vector of [4 x double] with undefined values.
3561 ///
3562 /// \headerfile <x86intrin.h>
3563 ///
3564 /// This intrinsic has no corresponding instruction.
3565 ///
3566 /// \returns A 256-bit vector of [4 x double] containing undefined values.
3567 static __inline__ __m256d __DEFAULT_FN_ATTRS
3568 _mm256_undefined_pd(void)
3569 {
3570   return (__m256d)__builtin_ia32_undef256();
3571 }
3572 
3573 /// Create a 256-bit vector of [8 x float] with undefined values.
3574 ///
3575 /// \headerfile <x86intrin.h>
3576 ///
3577 /// This intrinsic has no corresponding instruction.
3578 ///
3579 /// \returns A 256-bit vector of [8 x float] containing undefined values.
3580 static __inline__ __m256 __DEFAULT_FN_ATTRS
3581 _mm256_undefined_ps(void)
3582 {
3583   return (__m256)__builtin_ia32_undef256();
3584 }
3585 
3586 /// Create a 256-bit integer vector with undefined values.
3587 ///
3588 /// \headerfile <x86intrin.h>
3589 ///
3590 /// This intrinsic has no corresponding instruction.
3591 ///
3592 /// \returns A 256-bit integer vector containing undefined values.
3593 static __inline__ __m256i __DEFAULT_FN_ATTRS
3594 _mm256_undefined_si256(void)
3595 {
3596   return (__m256i)__builtin_ia32_undef256();
3597 }
3598 
3599 /// Constructs a 256-bit floating-point vector of [4 x double]
3600 ///    initialized with the specified double-precision floating-point values.
3601 ///
3602 /// \headerfile <x86intrin.h>
3603 ///
3604 /// This intrinsic corresponds to the <c> VUNPCKLPD+VINSERTF128 </c>
3605 ///   instruction.
3606 ///
3607 /// \param __a
3608 ///    A double-precision floating-point value used to initialize bits [255:192]
3609 ///    of the result.
3610 /// \param __b
3611 ///    A double-precision floating-point value used to initialize bits [191:128]
3612 ///    of the result.
3613 /// \param __c
3614 ///    A double-precision floating-point value used to initialize bits [127:64]
3615 ///    of the result.
3616 /// \param __d
3617 ///    A double-precision floating-point value used to initialize bits [63:0]
3618 ///    of the result.
3619 /// \returns An initialized 256-bit floating-point vector of [4 x double].
3620 static __inline __m256d __DEFAULT_FN_ATTRS
3621 _mm256_set_pd(double __a, double __b, double __c, double __d)
3622 {
3623   return __extension__ (__m256d){ __d, __c, __b, __a };
3624 }
3625 
3626 /// Constructs a 256-bit floating-point vector of [8 x float] initialized
3627 ///    with the specified single-precision floating-point values.
3628 ///
3629 /// \headerfile <x86intrin.h>
3630 ///
3631 /// This intrinsic is a utility function and does not correspond to a specific
3632 ///   instruction.
3633 ///
3634 /// \param __a
3635 ///    A single-precision floating-point value used to initialize bits [255:224]
3636 ///    of the result.
3637 /// \param __b
3638 ///    A single-precision floating-point value used to initialize bits [223:192]
3639 ///    of the result.
3640 /// \param __c
3641 ///    A single-precision floating-point value used to initialize bits [191:160]
3642 ///    of the result.
3643 /// \param __d
3644 ///    A single-precision floating-point value used to initialize bits [159:128]
3645 ///    of the result.
3646 /// \param __e
3647 ///    A single-precision floating-point value used to initialize bits [127:96]
3648 ///    of the result.
3649 /// \param __f
3650 ///    A single-precision floating-point value used to initialize bits [95:64]
3651 ///    of the result.
3652 /// \param __g
3653 ///    A single-precision floating-point value used to initialize bits [63:32]
3654 ///    of the result.
3655 /// \param __h
3656 ///    A single-precision floating-point value used to initialize bits [31:0]
3657 ///    of the result.
3658 /// \returns An initialized 256-bit floating-point vector of [8 x float].
3659 static __inline __m256 __DEFAULT_FN_ATTRS
3660 _mm256_set_ps(float __a, float __b, float __c, float __d,
3661               float __e, float __f, float __g, float __h)
3662 {
3663   return __extension__ (__m256){ __h, __g, __f, __e, __d, __c, __b, __a };
3664 }
3665 
3666 /// Constructs a 256-bit integer vector initialized with the specified
3667 ///    32-bit integral values.
3668 ///
3669 /// \headerfile <x86intrin.h>
3670 ///
3671 /// This intrinsic is a utility function and does not correspond to a specific
3672 ///   instruction.
3673 ///
3674 /// \param __i0
3675 ///    A 32-bit integral value used to initialize bits [255:224] of the result.
3676 /// \param __i1
3677 ///    A 32-bit integral value used to initialize bits [223:192] of the result.
3678 /// \param __i2
3679 ///    A 32-bit integral value used to initialize bits [191:160] of the result.
3680 /// \param __i3
3681 ///    A 32-bit integral value used to initialize bits [159:128] of the result.
3682 /// \param __i4
3683 ///    A 32-bit integral value used to initialize bits [127:96] of the result.
3684 /// \param __i5
3685 ///    A 32-bit integral value used to initialize bits [95:64] of the result.
3686 /// \param __i6
3687 ///    A 32-bit integral value used to initialize bits [63:32] of the result.
3688 /// \param __i7
3689 ///    A 32-bit integral value used to initialize bits [31:0] of the result.
3690 /// \returns An initialized 256-bit integer vector.
3691 static __inline __m256i __DEFAULT_FN_ATTRS
3692 _mm256_set_epi32(int __i0, int __i1, int __i2, int __i3,
3693                  int __i4, int __i5, int __i6, int __i7)
3694 {
3695   return __extension__ (__m256i)(__v8si){ __i7, __i6, __i5, __i4, __i3, __i2, __i1, __i0 };
3696 }
3697 
3698 /// Constructs a 256-bit integer vector initialized with the specified
3699 ///    16-bit integral values.
3700 ///
3701 /// \headerfile <x86intrin.h>
3702 ///
3703 /// This intrinsic is a utility function and does not correspond to a specific
3704 ///   instruction.
3705 ///
3706 /// \param __w15
3707 ///    A 16-bit integral value used to initialize bits [255:240] of the result.
3708 /// \param __w14
3709 ///    A 16-bit integral value used to initialize bits [239:224] of the result.
3710 /// \param __w13
3711 ///    A 16-bit integral value used to initialize bits [223:208] of the result.
3712 /// \param __w12
3713 ///    A 16-bit integral value used to initialize bits [207:192] of the result.
3714 /// \param __w11
3715 ///    A 16-bit integral value used to initialize bits [191:176] of the result.
3716 /// \param __w10
3717 ///    A 16-bit integral value used to initialize bits [175:160] of the result.
3718 /// \param __w09
3719 ///    A 16-bit integral value used to initialize bits [159:144] of the result.
3720 /// \param __w08
3721 ///    A 16-bit integral value used to initialize bits [143:128] of the result.
3722 /// \param __w07
3723 ///    A 16-bit integral value used to initialize bits [127:112] of the result.
3724 /// \param __w06
3725 ///    A 16-bit integral value used to initialize bits [111:96] of the result.
3726 /// \param __w05
3727 ///    A 16-bit integral value used to initialize bits [95:80] of the result.
3728 /// \param __w04
3729 ///    A 16-bit integral value used to initialize bits [79:64] of the result.
3730 /// \param __w03
3731 ///    A 16-bit integral value used to initialize bits [63:48] of the result.
3732 /// \param __w02
3733 ///    A 16-bit integral value used to initialize bits [47:32] of the result.
3734 /// \param __w01
3735 ///    A 16-bit integral value used to initialize bits [31:16] of the result.
3736 /// \param __w00
3737 ///    A 16-bit integral value used to initialize bits [15:0] of the result.
3738 /// \returns An initialized 256-bit integer vector.
3739 static __inline __m256i __DEFAULT_FN_ATTRS
3740 _mm256_set_epi16(short __w15, short __w14, short __w13, short __w12,
3741                  short __w11, short __w10, short __w09, short __w08,
3742                  short __w07, short __w06, short __w05, short __w04,
3743                  short __w03, short __w02, short __w01, short __w00)
3744 {
3745   return __extension__ (__m256i)(__v16hi){ __w00, __w01, __w02, __w03, __w04, __w05, __w06,
3746     __w07, __w08, __w09, __w10, __w11, __w12, __w13, __w14, __w15 };
3747 }
3748 
3749 /// Constructs a 256-bit integer vector initialized with the specified
3750 ///    8-bit integral values.
3751 ///
3752 /// \headerfile <x86intrin.h>
3753 ///
3754 /// This intrinsic is a utility function and does not correspond to a specific
3755 ///   instruction.
3756 ///
3757 /// \param __b31
3758 ///    An 8-bit integral value used to initialize bits [255:248] of the result.
3759 /// \param __b30
3760 ///    An 8-bit integral value used to initialize bits [247:240] of the result.
3761 /// \param __b29
3762 ///    An 8-bit integral value used to initialize bits [239:232] of the result.
3763 /// \param __b28
3764 ///    An 8-bit integral value used to initialize bits [231:224] of the result.
3765 /// \param __b27
3766 ///    An 8-bit integral value used to initialize bits [223:216] of the result.
3767 /// \param __b26
3768 ///    An 8-bit integral value used to initialize bits [215:208] of the result.
3769 /// \param __b25
3770 ///    An 8-bit integral value used to initialize bits [207:200] of the result.
3771 /// \param __b24
3772 ///    An 8-bit integral value used to initialize bits [199:192] of the result.
3773 /// \param __b23
3774 ///    An 8-bit integral value used to initialize bits [191:184] of the result.
3775 /// \param __b22
3776 ///    An 8-bit integral value used to initialize bits [183:176] of the result.
3777 /// \param __b21
3778 ///    An 8-bit integral value used to initialize bits [175:168] of the result.
3779 /// \param __b20
3780 ///    An 8-bit integral value used to initialize bits [167:160] of the result.
3781 /// \param __b19
3782 ///    An 8-bit integral value used to initialize bits [159:152] of the result.
3783 /// \param __b18
3784 ///    An 8-bit integral value used to initialize bits [151:144] of the result.
3785 /// \param __b17
3786 ///    An 8-bit integral value used to initialize bits [143:136] of the result.
3787 /// \param __b16
3788 ///    An 8-bit integral value used to initialize bits [135:128] of the result.
3789 /// \param __b15
3790 ///    An 8-bit integral value used to initialize bits [127:120] of the result.
3791 /// \param __b14
3792 ///    An 8-bit integral value used to initialize bits [119:112] of the result.
3793 /// \param __b13
3794 ///    An 8-bit integral value used to initialize bits [111:104] of the result.
3795 /// \param __b12
3796 ///    An 8-bit integral value used to initialize bits [103:96] of the result.
3797 /// \param __b11
3798 ///    An 8-bit integral value used to initialize bits [95:88] of the result.
3799 /// \param __b10
3800 ///    An 8-bit integral value used to initialize bits [87:80] of the result.
3801 /// \param __b09
3802 ///    An 8-bit integral value used to initialize bits [79:72] of the result.
3803 /// \param __b08
3804 ///    An 8-bit integral value used to initialize bits [71:64] of the result.
3805 /// \param __b07
3806 ///    An 8-bit integral value used to initialize bits [63:56] of the result.
3807 /// \param __b06
3808 ///    An 8-bit integral value used to initialize bits [55:48] of the result.
3809 /// \param __b05
3810 ///    An 8-bit integral value used to initialize bits [47:40] of the result.
3811 /// \param __b04
3812 ///    An 8-bit integral value used to initialize bits [39:32] of the result.
3813 /// \param __b03
3814 ///    An 8-bit integral value used to initialize bits [31:24] of the result.
3815 /// \param __b02
3816 ///    An 8-bit integral value used to initialize bits [23:16] of the result.
3817 /// \param __b01
3818 ///    An 8-bit integral value used to initialize bits [15:8] of the result.
3819 /// \param __b00
3820 ///    An 8-bit integral value used to initialize bits [7:0] of the result.
3821 /// \returns An initialized 256-bit integer vector.
3822 static __inline __m256i __DEFAULT_FN_ATTRS
3823 _mm256_set_epi8(char __b31, char __b30, char __b29, char __b28,
3824                 char __b27, char __b26, char __b25, char __b24,
3825                 char __b23, char __b22, char __b21, char __b20,
3826                 char __b19, char __b18, char __b17, char __b16,
3827                 char __b15, char __b14, char __b13, char __b12,
3828                 char __b11, char __b10, char __b09, char __b08,
3829                 char __b07, char __b06, char __b05, char __b04,
3830                 char __b03, char __b02, char __b01, char __b00)
3831 {
3832   return __extension__ (__m256i)(__v32qi){
3833     __b00, __b01, __b02, __b03, __b04, __b05, __b06, __b07,
3834     __b08, __b09, __b10, __b11, __b12, __b13, __b14, __b15,
3835     __b16, __b17, __b18, __b19, __b20, __b21, __b22, __b23,
3836     __b24, __b25, __b26, __b27, __b28, __b29, __b30, __b31
3837   };
3838 }
3839 
3840 /// Constructs a 256-bit integer vector initialized with the specified
3841 ///    64-bit integral values.
3842 ///
3843 /// \headerfile <x86intrin.h>
3844 ///
3845 /// This intrinsic corresponds to the <c> VPUNPCKLQDQ+VINSERTF128 </c>
3846 ///   instruction.
3847 ///
3848 /// \param __a
3849 ///    A 64-bit integral value used to initialize bits [255:192] of the result.
3850 /// \param __b
3851 ///    A 64-bit integral value used to initialize bits [191:128] of the result.
3852 /// \param __c
3853 ///    A 64-bit integral value used to initialize bits [127:64] of the result.
3854 /// \param __d
3855 ///    A 64-bit integral value used to initialize bits [63:0] of the result.
3856 /// \returns An initialized 256-bit integer vector.
3857 static __inline __m256i __DEFAULT_FN_ATTRS
3858 _mm256_set_epi64x(long long __a, long long __b, long long __c, long long __d)
3859 {
3860   return __extension__ (__m256i)(__v4di){ __d, __c, __b, __a };
3861 }
3862 
3863 /* Create vectors with elements in reverse order */
3864 /// Constructs a 256-bit floating-point vector of [4 x double],
3865 ///    initialized in reverse order with the specified double-precision
3866 ///    floating-point values.
3867 ///
3868 /// \headerfile <x86intrin.h>
3869 ///
3870 /// This intrinsic corresponds to the <c> VUNPCKLPD+VINSERTF128 </c>
3871 ///   instruction.
3872 ///
3873 /// \param __a
3874 ///    A double-precision floating-point value used to initialize bits [63:0]
3875 ///    of the result.
3876 /// \param __b
3877 ///    A double-precision floating-point value used to initialize bits [127:64]
3878 ///    of the result.
3879 /// \param __c
3880 ///    A double-precision floating-point value used to initialize bits [191:128]
3881 ///    of the result.
3882 /// \param __d
3883 ///    A double-precision floating-point value used to initialize bits [255:192]
3884 ///    of the result.
3885 /// \returns An initialized 256-bit floating-point vector of [4 x double].
3886 static __inline __m256d __DEFAULT_FN_ATTRS
3887 _mm256_setr_pd(double __a, double __b, double __c, double __d)
3888 {
3889   return _mm256_set_pd(__d, __c, __b, __a);
3890 }
3891 
3892 /// Constructs a 256-bit floating-point vector of [8 x float],
3893 ///    initialized in reverse order with the specified single-precision
3894 ///    float-point values.
3895 ///
3896 /// \headerfile <x86intrin.h>
3897 ///
3898 /// This intrinsic is a utility function and does not correspond to a specific
3899 ///   instruction.
3900 ///
3901 /// \param __a
3902 ///    A single-precision floating-point value used to initialize bits [31:0]
3903 ///    of the result.
3904 /// \param __b
3905 ///    A single-precision floating-point value used to initialize bits [63:32]
3906 ///    of the result.
3907 /// \param __c
3908 ///    A single-precision floating-point value used to initialize bits [95:64]
3909 ///    of the result.
3910 /// \param __d
3911 ///    A single-precision floating-point value used to initialize bits [127:96]
3912 ///    of the result.
3913 /// \param __e
3914 ///    A single-precision floating-point value used to initialize bits [159:128]
3915 ///    of the result.
3916 /// \param __f
3917 ///    A single-precision floating-point value used to initialize bits [191:160]
3918 ///    of the result.
3919 /// \param __g
3920 ///    A single-precision floating-point value used to initialize bits [223:192]
3921 ///    of the result.
3922 /// \param __h
3923 ///    A single-precision floating-point value used to initialize bits [255:224]
3924 ///    of the result.
3925 /// \returns An initialized 256-bit floating-point vector of [8 x float].
3926 static __inline __m256 __DEFAULT_FN_ATTRS
3927 _mm256_setr_ps(float __a, float __b, float __c, float __d,
3928                float __e, float __f, float __g, float __h)
3929 {
3930   return _mm256_set_ps(__h, __g, __f, __e, __d, __c, __b, __a);
3931 }
3932 
3933 /// Constructs a 256-bit integer vector, initialized in reverse order
3934 ///    with the specified 32-bit integral values.
3935 ///
3936 /// \headerfile <x86intrin.h>
3937 ///
3938 /// This intrinsic is a utility function and does not correspond to a specific
3939 ///   instruction.
3940 ///
3941 /// \param __i0
3942 ///    A 32-bit integral value used to initialize bits [31:0] of the result.
3943 /// \param __i1
3944 ///    A 32-bit integral value used to initialize bits [63:32] of the result.
3945 /// \param __i2
3946 ///    A 32-bit integral value used to initialize bits [95:64] of the result.
3947 /// \param __i3
3948 ///    A 32-bit integral value used to initialize bits [127:96] of the result.
3949 /// \param __i4
3950 ///    A 32-bit integral value used to initialize bits [159:128] of the result.
3951 /// \param __i5
3952 ///    A 32-bit integral value used to initialize bits [191:160] of the result.
3953 /// \param __i6
3954 ///    A 32-bit integral value used to initialize bits [223:192] of the result.
3955 /// \param __i7
3956 ///    A 32-bit integral value used to initialize bits [255:224] of the result.
3957 /// \returns An initialized 256-bit integer vector.
3958 static __inline __m256i __DEFAULT_FN_ATTRS
3959 _mm256_setr_epi32(int __i0, int __i1, int __i2, int __i3,
3960                   int __i4, int __i5, int __i6, int __i7)
3961 {
3962   return _mm256_set_epi32(__i7, __i6, __i5, __i4, __i3, __i2, __i1, __i0);
3963 }
3964 
3965 /// Constructs a 256-bit integer vector, initialized in reverse order
3966 ///    with the specified 16-bit integral values.
3967 ///
3968 /// \headerfile <x86intrin.h>
3969 ///
3970 /// This intrinsic is a utility function and does not correspond to a specific
3971 ///   instruction.
3972 ///
3973 /// \param __w15
3974 ///    A 16-bit integral value used to initialize bits [15:0] of the result.
3975 /// \param __w14
3976 ///    A 16-bit integral value used to initialize bits [31:16] of the result.
3977 /// \param __w13
3978 ///    A 16-bit integral value used to initialize bits [47:32] of the result.
3979 /// \param __w12
3980 ///    A 16-bit integral value used to initialize bits [63:48] of the result.
3981 /// \param __w11
3982 ///    A 16-bit integral value used to initialize bits [79:64] of the result.
3983 /// \param __w10
3984 ///    A 16-bit integral value used to initialize bits [95:80] of the result.
3985 /// \param __w09
3986 ///    A 16-bit integral value used to initialize bits [111:96] of the result.
3987 /// \param __w08
3988 ///    A 16-bit integral value used to initialize bits [127:112] of the result.
3989 /// \param __w07
3990 ///    A 16-bit integral value used to initialize bits [143:128] of the result.
3991 /// \param __w06
3992 ///    A 16-bit integral value used to initialize bits [159:144] of the result.
3993 /// \param __w05
3994 ///    A 16-bit integral value used to initialize bits [175:160] of the result.
3995 /// \param __w04
3996 ///    A 16-bit integral value used to initialize bits [191:176] of the result.
3997 /// \param __w03
3998 ///    A 16-bit integral value used to initialize bits [207:192] of the result.
3999 /// \param __w02
4000 ///    A 16-bit integral value used to initialize bits [223:208] of the result.
4001 /// \param __w01
4002 ///    A 16-bit integral value used to initialize bits [239:224] of the result.
4003 /// \param __w00
4004 ///    A 16-bit integral value used to initialize bits [255:240] of the result.
4005 /// \returns An initialized 256-bit integer vector.
4006 static __inline __m256i __DEFAULT_FN_ATTRS
4007 _mm256_setr_epi16(short __w15, short __w14, short __w13, short __w12,
4008        short __w11, short __w10, short __w09, short __w08,
4009        short __w07, short __w06, short __w05, short __w04,
4010        short __w03, short __w02, short __w01, short __w00)
4011 {
4012   return _mm256_set_epi16(__w00, __w01, __w02, __w03,
4013                           __w04, __w05, __w06, __w07,
4014                           __w08, __w09, __w10, __w11,
4015                           __w12, __w13, __w14, __w15);
4016 }
4017 
4018 /// Constructs a 256-bit integer vector, initialized in reverse order
4019 ///    with the specified 8-bit integral values.
4020 ///
4021 /// \headerfile <x86intrin.h>
4022 ///
4023 /// This intrinsic is a utility function and does not correspond to a specific
4024 ///   instruction.
4025 ///
4026 /// \param __b31
4027 ///    An 8-bit integral value used to initialize bits [7:0] of the result.
4028 /// \param __b30
4029 ///    An 8-bit integral value used to initialize bits [15:8] of the result.
4030 /// \param __b29
4031 ///    An 8-bit integral value used to initialize bits [23:16] of the result.
4032 /// \param __b28
4033 ///    An 8-bit integral value used to initialize bits [31:24] of the result.
4034 /// \param __b27
4035 ///    An 8-bit integral value used to initialize bits [39:32] of the result.
4036 /// \param __b26
4037 ///    An 8-bit integral value used to initialize bits [47:40] of the result.
4038 /// \param __b25
4039 ///    An 8-bit integral value used to initialize bits [55:48] of the result.
4040 /// \param __b24
4041 ///    An 8-bit integral value used to initialize bits [63:56] of the result.
4042 /// \param __b23
4043 ///    An 8-bit integral value used to initialize bits [71:64] of the result.
4044 /// \param __b22
4045 ///    An 8-bit integral value used to initialize bits [79:72] of the result.
4046 /// \param __b21
4047 ///    An 8-bit integral value used to initialize bits [87:80] of the result.
4048 /// \param __b20
4049 ///    An 8-bit integral value used to initialize bits [95:88] of the result.
4050 /// \param __b19
4051 ///    An 8-bit integral value used to initialize bits [103:96] of the result.
4052 /// \param __b18
4053 ///    An 8-bit integral value used to initialize bits [111:104] of the result.
4054 /// \param __b17
4055 ///    An 8-bit integral value used to initialize bits [119:112] of the result.
4056 /// \param __b16
4057 ///    An 8-bit integral value used to initialize bits [127:120] of the result.
4058 /// \param __b15
4059 ///    An 8-bit integral value used to initialize bits [135:128] of the result.
4060 /// \param __b14
4061 ///    An 8-bit integral value used to initialize bits [143:136] of the result.
4062 /// \param __b13
4063 ///    An 8-bit integral value used to initialize bits [151:144] of the result.
4064 /// \param __b12
4065 ///    An 8-bit integral value used to initialize bits [159:152] of the result.
4066 /// \param __b11
4067 ///    An 8-bit integral value used to initialize bits [167:160] of the result.
4068 /// \param __b10
4069 ///    An 8-bit integral value used to initialize bits [175:168] of the result.
4070 /// \param __b09
4071 ///    An 8-bit integral value used to initialize bits [183:176] of the result.
4072 /// \param __b08
4073 ///    An 8-bit integral value used to initialize bits [191:184] of the result.
4074 /// \param __b07
4075 ///    An 8-bit integral value used to initialize bits [199:192] of the result.
4076 /// \param __b06
4077 ///    An 8-bit integral value used to initialize bits [207:200] of the result.
4078 /// \param __b05
4079 ///    An 8-bit integral value used to initialize bits [215:208] of the result.
4080 /// \param __b04
4081 ///    An 8-bit integral value used to initialize bits [223:216] of the result.
4082 /// \param __b03
4083 ///    An 8-bit integral value used to initialize bits [231:224] of the result.
4084 /// \param __b02
4085 ///    An 8-bit integral value used to initialize bits [239:232] of the result.
4086 /// \param __b01
4087 ///    An 8-bit integral value used to initialize bits [247:240] of the result.
4088 /// \param __b00
4089 ///    An 8-bit integral value used to initialize bits [255:248] of the result.
4090 /// \returns An initialized 256-bit integer vector.
4091 static __inline __m256i __DEFAULT_FN_ATTRS
4092 _mm256_setr_epi8(char __b31, char __b30, char __b29, char __b28,
4093                  char __b27, char __b26, char __b25, char __b24,
4094                  char __b23, char __b22, char __b21, char __b20,
4095                  char __b19, char __b18, char __b17, char __b16,
4096                  char __b15, char __b14, char __b13, char __b12,
4097                  char __b11, char __b10, char __b09, char __b08,
4098                  char __b07, char __b06, char __b05, char __b04,
4099                  char __b03, char __b02, char __b01, char __b00)
4100 {
4101   return _mm256_set_epi8(__b00, __b01, __b02, __b03, __b04, __b05, __b06, __b07,
4102                          __b08, __b09, __b10, __b11, __b12, __b13, __b14, __b15,
4103                          __b16, __b17, __b18, __b19, __b20, __b21, __b22, __b23,
4104                          __b24, __b25, __b26, __b27, __b28, __b29, __b30, __b31);
4105 }
4106 
4107 /// Constructs a 256-bit integer vector, initialized in reverse order
4108 ///    with the specified 64-bit integral values.
4109 ///
4110 /// \headerfile <x86intrin.h>
4111 ///
4112 /// This intrinsic corresponds to the <c> VPUNPCKLQDQ+VINSERTF128 </c>
4113 ///   instruction.
4114 ///
4115 /// \param __a
4116 ///    A 64-bit integral value used to initialize bits [63:0] of the result.
4117 /// \param __b
4118 ///    A 64-bit integral value used to initialize bits [127:64] of the result.
4119 /// \param __c
4120 ///    A 64-bit integral value used to initialize bits [191:128] of the result.
4121 /// \param __d
4122 ///    A 64-bit integral value used to initialize bits [255:192] of the result.
4123 /// \returns An initialized 256-bit integer vector.
4124 static __inline __m256i __DEFAULT_FN_ATTRS
4125 _mm256_setr_epi64x(long long __a, long long __b, long long __c, long long __d)
4126 {
4127   return _mm256_set_epi64x(__d, __c, __b, __a);
4128 }
4129 
4130 /* Create vectors with repeated elements */
4131 /// Constructs a 256-bit floating-point vector of [4 x double], with each
4132 ///    of the four double-precision floating-point vector elements set to the
4133 ///    specified double-precision floating-point value.
4134 ///
4135 /// \headerfile <x86intrin.h>
4136 ///
4137 /// This intrinsic corresponds to the <c> VMOVDDUP+VINSERTF128 </c> instruction.
4138 ///
4139 /// \param __w
4140 ///    A double-precision floating-point value used to initialize each vector
4141 ///    element of the result.
4142 /// \returns An initialized 256-bit floating-point vector of [4 x double].
4143 static __inline __m256d __DEFAULT_FN_ATTRS
4144 _mm256_set1_pd(double __w)
4145 {
4146   return _mm256_set_pd(__w, __w, __w, __w);
4147 }
4148 
4149 /// Constructs a 256-bit floating-point vector of [8 x float], with each
4150 ///    of the eight single-precision floating-point vector elements set to the
4151 ///    specified single-precision floating-point value.
4152 ///
4153 /// \headerfile <x86intrin.h>
4154 ///
4155 /// This intrinsic corresponds to the <c> VPERMILPS+VINSERTF128 </c>
4156 ///   instruction.
4157 ///
4158 /// \param __w
4159 ///    A single-precision floating-point value used to initialize each vector
4160 ///    element of the result.
4161 /// \returns An initialized 256-bit floating-point vector of [8 x float].
4162 static __inline __m256 __DEFAULT_FN_ATTRS
4163 _mm256_set1_ps(float __w)
4164 {
4165   return _mm256_set_ps(__w, __w, __w, __w, __w, __w, __w, __w);
4166 }
4167 
4168 /// Constructs a 256-bit integer vector of [8 x i32], with each of the
4169 ///    32-bit integral vector elements set to the specified 32-bit integral
4170 ///    value.
4171 ///
4172 /// \headerfile <x86intrin.h>
4173 ///
4174 /// This intrinsic corresponds to the <c> VPERMILPS+VINSERTF128 </c>
4175 ///   instruction.
4176 ///
4177 /// \param __i
4178 ///    A 32-bit integral value used to initialize each vector element of the
4179 ///    result.
4180 /// \returns An initialized 256-bit integer vector of [8 x i32].
4181 static __inline __m256i __DEFAULT_FN_ATTRS
4182 _mm256_set1_epi32(int __i)
4183 {
4184   return _mm256_set_epi32(__i, __i, __i, __i, __i, __i, __i, __i);
4185 }
4186 
4187 /// Constructs a 256-bit integer vector of [16 x i16], with each of the
4188 ///    16-bit integral vector elements set to the specified 16-bit integral
4189 ///    value.
4190 ///
4191 /// \headerfile <x86intrin.h>
4192 ///
4193 /// This intrinsic corresponds to the <c> VPSHUFB+VINSERTF128 </c> instruction.
4194 ///
4195 /// \param __w
4196 ///    A 16-bit integral value used to initialize each vector element of the
4197 ///    result.
4198 /// \returns An initialized 256-bit integer vector of [16 x i16].
4199 static __inline __m256i __DEFAULT_FN_ATTRS
4200 _mm256_set1_epi16(short __w)
4201 {
4202   return _mm256_set_epi16(__w, __w, __w, __w, __w, __w, __w, __w,
4203                           __w, __w, __w, __w, __w, __w, __w, __w);
4204 }
4205 
4206 /// Constructs a 256-bit integer vector of [32 x i8], with each of the
4207 ///    8-bit integral vector elements set to the specified 8-bit integral value.
4208 ///
4209 /// \headerfile <x86intrin.h>
4210 ///
4211 /// This intrinsic corresponds to the <c> VPSHUFB+VINSERTF128 </c> instruction.
4212 ///
4213 /// \param __b
4214 ///    An 8-bit integral value used to initialize each vector element of the
4215 ///    result.
4216 /// \returns An initialized 256-bit integer vector of [32 x i8].
4217 static __inline __m256i __DEFAULT_FN_ATTRS
4218 _mm256_set1_epi8(char __b)
4219 {
4220   return _mm256_set_epi8(__b, __b, __b, __b, __b, __b, __b, __b,
4221                          __b, __b, __b, __b, __b, __b, __b, __b,
4222                          __b, __b, __b, __b, __b, __b, __b, __b,
4223                          __b, __b, __b, __b, __b, __b, __b, __b);
4224 }
4225 
4226 /// Constructs a 256-bit integer vector of [4 x i64], with each of the
4227 ///    64-bit integral vector elements set to the specified 64-bit integral
4228 ///    value.
4229 ///
4230 /// \headerfile <x86intrin.h>
4231 ///
4232 /// This intrinsic corresponds to the <c> VMOVDDUP+VINSERTF128 </c> instruction.
4233 ///
4234 /// \param __q
4235 ///    A 64-bit integral value used to initialize each vector element of the
4236 ///    result.
4237 /// \returns An initialized 256-bit integer vector of [4 x i64].
4238 static __inline __m256i __DEFAULT_FN_ATTRS
4239 _mm256_set1_epi64x(long long __q)
4240 {
4241   return _mm256_set_epi64x(__q, __q, __q, __q);
4242 }
4243 
4244 /* Create __zeroed vectors */
4245 /// Constructs a 256-bit floating-point vector of [4 x double] with all
4246 ///    vector elements initialized to zero.
4247 ///
4248 /// \headerfile <x86intrin.h>
4249 ///
4250 /// This intrinsic corresponds to the <c> VXORPS </c> instruction.
4251 ///
4252 /// \returns A 256-bit vector of [4 x double] with all elements set to zero.
4253 static __inline __m256d __DEFAULT_FN_ATTRS
4254 _mm256_setzero_pd(void)
4255 {
4256   return __extension__ (__m256d){ 0, 0, 0, 0 };
4257 }
4258 
4259 /// Constructs a 256-bit floating-point vector of [8 x float] with all
4260 ///    vector elements initialized to zero.
4261 ///
4262 /// \headerfile <x86intrin.h>
4263 ///
4264 /// This intrinsic corresponds to the <c> VXORPS </c> instruction.
4265 ///
4266 /// \returns A 256-bit vector of [8 x float] with all elements set to zero.
4267 static __inline __m256 __DEFAULT_FN_ATTRS
4268 _mm256_setzero_ps(void)
4269 {
4270   return __extension__ (__m256){ 0, 0, 0, 0, 0, 0, 0, 0 };
4271 }
4272 
4273 /// Constructs a 256-bit integer vector initialized to zero.
4274 ///
4275 /// \headerfile <x86intrin.h>
4276 ///
4277 /// This intrinsic corresponds to the <c> VXORPS </c> instruction.
4278 ///
4279 /// \returns A 256-bit integer vector initialized to zero.
4280 static __inline __m256i __DEFAULT_FN_ATTRS
4281 _mm256_setzero_si256(void)
4282 {
4283   return __extension__ (__m256i)(__v4di){ 0, 0, 0, 0 };
4284 }
4285 
4286 /* Cast between vector types */
4287 /// Casts a 256-bit floating-point vector of [4 x double] into a 256-bit
4288 ///    floating-point vector of [8 x float].
4289 ///
4290 /// \headerfile <x86intrin.h>
4291 ///
4292 /// This intrinsic has no corresponding instruction.
4293 ///
4294 /// \param __a
4295 ///    A 256-bit floating-point vector of [4 x double].
4296 /// \returns A 256-bit floating-point vector of [8 x float] containing the same
4297 ///    bitwise pattern as the parameter.
4298 static __inline __m256 __DEFAULT_FN_ATTRS
4299 _mm256_castpd_ps(__m256d __a)
4300 {
4301   return (__m256)__a;
4302 }
4303 
4304 /// Casts a 256-bit floating-point vector of [4 x double] into a 256-bit
4305 ///    integer vector.
4306 ///
4307 /// \headerfile <x86intrin.h>
4308 ///
4309 /// This intrinsic has no corresponding instruction.
4310 ///
4311 /// \param __a
4312 ///    A 256-bit floating-point vector of [4 x double].
4313 /// \returns A 256-bit integer vector containing the same bitwise pattern as the
4314 ///    parameter.
4315 static __inline __m256i __DEFAULT_FN_ATTRS
4316 _mm256_castpd_si256(__m256d __a)
4317 {
4318   return (__m256i)__a;
4319 }
4320 
4321 /// Casts a 256-bit floating-point vector of [8 x float] into a 256-bit
4322 ///    floating-point vector of [4 x double].
4323 ///
4324 /// \headerfile <x86intrin.h>
4325 ///
4326 /// This intrinsic has no corresponding instruction.
4327 ///
4328 /// \param __a
4329 ///    A 256-bit floating-point vector of [8 x float].
4330 /// \returns A 256-bit floating-point vector of [4 x double] containing the same
4331 ///    bitwise pattern as the parameter.
4332 static __inline __m256d __DEFAULT_FN_ATTRS
4333 _mm256_castps_pd(__m256 __a)
4334 {
4335   return (__m256d)__a;
4336 }
4337 
4338 /// Casts a 256-bit floating-point vector of [8 x float] into a 256-bit
4339 ///    integer vector.
4340 ///
4341 /// \headerfile <x86intrin.h>
4342 ///
4343 /// This intrinsic has no corresponding instruction.
4344 ///
4345 /// \param __a
4346 ///    A 256-bit floating-point vector of [8 x float].
4347 /// \returns A 256-bit integer vector containing the same bitwise pattern as the
4348 ///    parameter.
4349 static __inline __m256i __DEFAULT_FN_ATTRS
4350 _mm256_castps_si256(__m256 __a)
4351 {
4352   return (__m256i)__a;
4353 }
4354 
4355 /// Casts a 256-bit integer vector into a 256-bit floating-point vector
4356 ///    of [8 x float].
4357 ///
4358 /// \headerfile <x86intrin.h>
4359 ///
4360 /// This intrinsic has no corresponding instruction.
4361 ///
4362 /// \param __a
4363 ///    A 256-bit integer vector.
4364 /// \returns A 256-bit floating-point vector of [8 x float] containing the same
4365 ///    bitwise pattern as the parameter.
4366 static __inline __m256 __DEFAULT_FN_ATTRS
4367 _mm256_castsi256_ps(__m256i __a)
4368 {
4369   return (__m256)__a;
4370 }
4371 
4372 /// Casts a 256-bit integer vector into a 256-bit floating-point vector
4373 ///    of [4 x double].
4374 ///
4375 /// \headerfile <x86intrin.h>
4376 ///
4377 /// This intrinsic has no corresponding instruction.
4378 ///
4379 /// \param __a
4380 ///    A 256-bit integer vector.
4381 /// \returns A 256-bit floating-point vector of [4 x double] containing the same
4382 ///    bitwise pattern as the parameter.
4383 static __inline __m256d __DEFAULT_FN_ATTRS
4384 _mm256_castsi256_pd(__m256i __a)
4385 {
4386   return (__m256d)__a;
4387 }
4388 
4389 /// Returns the lower 128 bits of a 256-bit floating-point vector of
4390 ///    [4 x double] as a 128-bit floating-point vector of [2 x double].
4391 ///
4392 /// \headerfile <x86intrin.h>
4393 ///
4394 /// This intrinsic has no corresponding instruction.
4395 ///
4396 /// \param __a
4397 ///    A 256-bit floating-point vector of [4 x double].
4398 /// \returns A 128-bit floating-point vector of [2 x double] containing the
4399 ///    lower 128 bits of the parameter.
4400 static __inline __m128d __DEFAULT_FN_ATTRS
4401 _mm256_castpd256_pd128(__m256d __a)
4402 {
4403   return __builtin_shufflevector((__v4df)__a, (__v4df)__a, 0, 1);
4404 }
4405 
4406 /// Returns the lower 128 bits of a 256-bit floating-point vector of
4407 ///    [8 x float] as a 128-bit floating-point vector of [4 x float].
4408 ///
4409 /// \headerfile <x86intrin.h>
4410 ///
4411 /// This intrinsic has no corresponding instruction.
4412 ///
4413 /// \param __a
4414 ///    A 256-bit floating-point vector of [8 x float].
4415 /// \returns A 128-bit floating-point vector of [4 x float] containing the
4416 ///    lower 128 bits of the parameter.
4417 static __inline __m128 __DEFAULT_FN_ATTRS
4418 _mm256_castps256_ps128(__m256 __a)
4419 {
4420   return __builtin_shufflevector((__v8sf)__a, (__v8sf)__a, 0, 1, 2, 3);
4421 }
4422 
4423 /// Truncates a 256-bit integer vector into a 128-bit integer vector.
4424 ///
4425 /// \headerfile <x86intrin.h>
4426 ///
4427 /// This intrinsic has no corresponding instruction.
4428 ///
4429 /// \param __a
4430 ///    A 256-bit integer vector.
4431 /// \returns A 128-bit integer vector containing the lower 128 bits of the
4432 ///    parameter.
4433 static __inline __m128i __DEFAULT_FN_ATTRS
4434 _mm256_castsi256_si128(__m256i __a)
4435 {
4436   return __builtin_shufflevector((__v4di)__a, (__v4di)__a, 0, 1);
4437 }
4438 
4439 /// Constructs a 256-bit floating-point vector of [4 x double] from a
4440 ///    128-bit floating-point vector of [2 x double].
4441 ///
4442 ///    The lower 128 bits contain the value of the source vector. The contents
4443 ///    of the upper 128 bits are undefined.
4444 ///
4445 /// \headerfile <x86intrin.h>
4446 ///
4447 /// This intrinsic has no corresponding instruction.
4448 ///
4449 /// \param __a
4450 ///    A 128-bit vector of [2 x double].
4451 /// \returns A 256-bit floating-point vector of [4 x double]. The lower 128 bits
4452 ///    contain the value of the parameter. The contents of the upper 128 bits
4453 ///    are undefined.
4454 static __inline __m256d __DEFAULT_FN_ATTRS
4455 _mm256_castpd128_pd256(__m128d __a)
4456 {
4457   return __builtin_shufflevector((__v2df)__a, (__v2df)__a, 0, 1, -1, -1);
4458 }
4459 
4460 /// Constructs a 256-bit floating-point vector of [8 x float] from a
4461 ///    128-bit floating-point vector of [4 x float].
4462 ///
4463 ///    The lower 128 bits contain the value of the source vector. The contents
4464 ///    of the upper 128 bits are undefined.
4465 ///
4466 /// \headerfile <x86intrin.h>
4467 ///
4468 /// This intrinsic has no corresponding instruction.
4469 ///
4470 /// \param __a
4471 ///    A 128-bit vector of [4 x float].
4472 /// \returns A 256-bit floating-point vector of [8 x float]. The lower 128 bits
4473 ///    contain the value of the parameter. The contents of the upper 128 bits
4474 ///    are undefined.
4475 static __inline __m256 __DEFAULT_FN_ATTRS
4476 _mm256_castps128_ps256(__m128 __a)
4477 {
4478   return __builtin_shufflevector((__v4sf)__a, (__v4sf)__a, 0, 1, 2, 3, -1, -1, -1, -1);
4479 }
4480 
4481 /// Constructs a 256-bit integer vector from a 128-bit integer vector.
4482 ///
4483 ///    The lower 128 bits contain the value of the source vector. The contents
4484 ///    of the upper 128 bits are undefined.
4485 ///
4486 /// \headerfile <x86intrin.h>
4487 ///
4488 /// This intrinsic has no corresponding instruction.
4489 ///
4490 /// \param __a
4491 ///    A 128-bit integer vector.
4492 /// \returns A 256-bit integer vector. The lower 128 bits contain the value of
4493 ///    the parameter. The contents of the upper 128 bits are undefined.
4494 static __inline __m256i __DEFAULT_FN_ATTRS
4495 _mm256_castsi128_si256(__m128i __a)
4496 {
4497   return __builtin_shufflevector((__v2di)__a, (__v2di)__a, 0, 1, -1, -1);
4498 }
4499 
4500 /// Constructs a 256-bit floating-point vector of [4 x double] from a
4501 ///    128-bit floating-point vector of [2 x double]. The lower 128 bits
4502 ///    contain the value of the source vector. The upper 128 bits are set
4503 ///    to zero.
4504 ///
4505 /// \headerfile <x86intrin.h>
4506 ///
4507 /// This intrinsic has no corresponding instruction.
4508 ///
4509 /// \param __a
4510 ///    A 128-bit vector of [2 x double].
4511 /// \returns A 256-bit floating-point vector of [4 x double]. The lower 128 bits
4512 ///    contain the value of the parameter. The upper 128 bits are set to zero.
4513 static __inline __m256d __DEFAULT_FN_ATTRS
4514 _mm256_zextpd128_pd256(__m128d __a)
4515 {
4516   return __builtin_shufflevector((__v2df)__a, (__v2df)_mm_setzero_pd(), 0, 1, 2, 3);
4517 }
4518 
4519 /// Constructs a 256-bit floating-point vector of [8 x float] from a
4520 ///    128-bit floating-point vector of [4 x float]. The lower 128 bits contain
4521 ///    the value of the source vector. The upper 128 bits are set to zero.
4522 ///
4523 /// \headerfile <x86intrin.h>
4524 ///
4525 /// This intrinsic has no corresponding instruction.
4526 ///
4527 /// \param __a
4528 ///    A 128-bit vector of [4 x float].
4529 /// \returns A 256-bit floating-point vector of [8 x float]. The lower 128 bits
4530 ///    contain the value of the parameter. The upper 128 bits are set to zero.
4531 static __inline __m256 __DEFAULT_FN_ATTRS
4532 _mm256_zextps128_ps256(__m128 __a)
4533 {
4534   return __builtin_shufflevector((__v4sf)__a, (__v4sf)_mm_setzero_ps(), 0, 1, 2, 3, 4, 5, 6, 7);
4535 }
4536 
4537 /// Constructs a 256-bit integer vector from a 128-bit integer vector.
4538 ///    The lower 128 bits contain the value of the source vector. The upper
4539 ///    128 bits are set to zero.
4540 ///
4541 /// \headerfile <x86intrin.h>
4542 ///
4543 /// This intrinsic has no corresponding instruction.
4544 ///
4545 /// \param __a
4546 ///    A 128-bit integer vector.
4547 /// \returns A 256-bit integer vector. The lower 128 bits contain the value of
4548 ///    the parameter. The upper 128 bits are set to zero.
4549 static __inline __m256i __DEFAULT_FN_ATTRS
4550 _mm256_zextsi128_si256(__m128i __a)
4551 {
4552   return __builtin_shufflevector((__v2di)__a, (__v2di)_mm_setzero_si128(), 0, 1, 2, 3);
4553 }
4554 
4555 /*
4556    Vector insert.
4557    We use macros rather than inlines because we only want to accept
4558    invocations where the immediate M is a constant expression.
4559 */
4560 /// Constructs a new 256-bit vector of [8 x float] by first duplicating
4561 ///    a 256-bit vector of [8 x float] given in the first parameter, and then
4562 ///    replacing either the upper or the lower 128 bits with the contents of a
4563 ///    128-bit vector of [4 x float] in the second parameter.
4564 ///
4565 ///    The immediate integer parameter determines between the upper or the lower
4566 ///    128 bits.
4567 ///
4568 /// \headerfile <x86intrin.h>
4569 ///
4570 /// \code
4571 /// __m256 _mm256_insertf128_ps(__m256 V1, __m128 V2, const int M);
4572 /// \endcode
4573 ///
4574 /// This intrinsic corresponds to the <c> VINSERTF128 </c> instruction.
4575 ///
4576 /// \param V1
4577 ///    A 256-bit vector of [8 x float]. This vector is copied to the result
4578 ///    first, and then either the upper or the lower 128 bits of the result will
4579 ///    be replaced by the contents of \a V2.
4580 /// \param V2
4581 ///    A 128-bit vector of [4 x float]. The contents of this parameter are
4582 ///    written to either the upper or the lower 128 bits of the result depending
4583 ///    on the value of parameter \a M.
4584 /// \param M
4585 ///    An immediate integer. The least significant bit determines how the values
4586 ///    from the two parameters are interleaved: \n
4587 ///    If bit [0] of \a M is 0, \a V2 are copied to bits [127:0] of the result,
4588 ///    and bits [255:128] of \a V1 are copied to bits [255:128] of the
4589 ///    result. \n
4590 ///    If bit [0] of \a M is 1, \a V2 are copied to bits [255:128] of the
4591 ///    result, and bits [127:0] of \a V1 are copied to bits [127:0] of the
4592 ///    result.
4593 /// \returns A 256-bit vector of [8 x float] containing the interleaved values.
4594 #define _mm256_insertf128_ps(V1, V2, M) \
4595   (__m256)__builtin_ia32_vinsertf128_ps256((__v8sf)(__m256)(V1), \
4596                                            (__v4sf)(__m128)(V2), (int)(M))
4597 
4598 /// Constructs a new 256-bit vector of [4 x double] by first duplicating
4599 ///    a 256-bit vector of [4 x double] given in the first parameter, and then
4600 ///    replacing either the upper or the lower 128 bits with the contents of a
4601 ///    128-bit vector of [2 x double] in the second parameter.
4602 ///
4603 ///    The immediate integer parameter determines between the upper or the lower
4604 ///    128 bits.
4605 ///
4606 /// \headerfile <x86intrin.h>
4607 ///
4608 /// \code
4609 /// __m256d _mm256_insertf128_pd(__m256d V1, __m128d V2, const int M);
4610 /// \endcode
4611 ///
4612 /// This intrinsic corresponds to the <c> VINSERTF128 </c> instruction.
4613 ///
4614 /// \param V1
4615 ///    A 256-bit vector of [4 x double]. This vector is copied to the result
4616 ///    first, and then either the upper or the lower 128 bits of the result will
4617 ///    be replaced by the contents of \a V2.
4618 /// \param V2
4619 ///    A 128-bit vector of [2 x double]. The contents of this parameter are
4620 ///    written to either the upper or the lower 128 bits of the result depending
4621 ///    on the value of parameter \a M.
4622 /// \param M
4623 ///    An immediate integer. The least significant bit determines how the values
4624 ///    from the two parameters are interleaved: \n
4625 ///    If bit [0] of \a M is 0, \a V2 are copied to bits [127:0] of the result,
4626 ///    and bits [255:128] of \a V1 are copied to bits [255:128] of the
4627 ///    result. \n
4628 ///    If bit [0] of \a M is 1, \a V2 are copied to bits [255:128] of the
4629 ///    result, and bits [127:0] of \a V1 are copied to bits [127:0] of the
4630 ///    result.
4631 /// \returns A 256-bit vector of [4 x double] containing the interleaved values.
4632 #define _mm256_insertf128_pd(V1, V2, M) \
4633   (__m256d)__builtin_ia32_vinsertf128_pd256((__v4df)(__m256d)(V1), \
4634                                             (__v2df)(__m128d)(V2), (int)(M))
4635 
4636 /// Constructs a new 256-bit integer vector by first duplicating a
4637 ///    256-bit integer vector given in the first parameter, and then replacing
4638 ///    either the upper or the lower 128 bits with the contents of a 128-bit
4639 ///    integer vector in the second parameter.
4640 ///
4641 ///    The immediate integer parameter determines between the upper or the lower
4642 ///    128 bits.
4643 ///
4644 /// \headerfile <x86intrin.h>
4645 ///
4646 /// \code
4647 /// __m256i _mm256_insertf128_si256(__m256i V1, __m128i V2, const int M);
4648 /// \endcode
4649 ///
4650 /// This intrinsic corresponds to the <c> VINSERTF128 </c> instruction.
4651 ///
4652 /// \param V1
4653 ///    A 256-bit integer vector. This vector is copied to the result first, and
4654 ///    then either the upper or the lower 128 bits of the result will be
4655 ///    replaced by the contents of \a V2.
4656 /// \param V2
4657 ///    A 128-bit integer vector. The contents of this parameter are written to
4658 ///    either the upper or the lower 128 bits of the result depending on the
4659 ///     value of parameter \a M.
4660 /// \param M
4661 ///    An immediate integer. The least significant bit determines how the values
4662 ///    from the two parameters are interleaved: \n
4663 ///    If bit [0] of \a M is 0, \a V2 are copied to bits [127:0] of the result,
4664 ///    and bits [255:128] of \a V1 are copied to bits [255:128] of the
4665 ///    result. \n
4666 ///    If bit [0] of \a M is 1, \a V2 are copied to bits [255:128] of the
4667 ///    result, and bits [127:0] of \a V1 are copied to bits [127:0] of the
4668 ///    result.
4669 /// \returns A 256-bit integer vector containing the interleaved values.
4670 #define _mm256_insertf128_si256(V1, V2, M) \
4671   (__m256i)__builtin_ia32_vinsertf128_si256((__v8si)(__m256i)(V1), \
4672                                             (__v4si)(__m128i)(V2), (int)(M))
4673 
4674 /*
4675    Vector extract.
4676    We use macros rather than inlines because we only want to accept
4677    invocations where the immediate M is a constant expression.
4678 */
4679 /// Extracts either the upper or the lower 128 bits from a 256-bit vector
4680 ///    of [8 x float], as determined by the immediate integer parameter, and
4681 ///    returns the extracted bits as a 128-bit vector of [4 x float].
4682 ///
4683 /// \headerfile <x86intrin.h>
4684 ///
4685 /// \code
4686 /// __m128 _mm256_extractf128_ps(__m256 V, const int M);
4687 /// \endcode
4688 ///
4689 /// This intrinsic corresponds to the <c> VEXTRACTF128 </c> instruction.
4690 ///
4691 /// \param V
4692 ///    A 256-bit vector of [8 x float].
4693 /// \param M
4694 ///    An immediate integer. The least significant bit determines which bits are
4695 ///    extracted from the first parameter: \n
4696 ///    If bit [0] of \a M is 0, bits [127:0] of \a V are copied to the
4697 ///    result. \n
4698 ///    If bit [0] of \a M is 1, bits [255:128] of \a V are copied to the result.
4699 /// \returns A 128-bit vector of [4 x float] containing the extracted bits.
4700 #define _mm256_extractf128_ps(V, M) \
4701   (__m128)__builtin_ia32_vextractf128_ps256((__v8sf)(__m256)(V), (int)(M))
4702 
4703 /// Extracts either the upper or the lower 128 bits from a 256-bit vector
4704 ///    of [4 x double], as determined by the immediate integer parameter, and
4705 ///    returns the extracted bits as a 128-bit vector of [2 x double].
4706 ///
4707 /// \headerfile <x86intrin.h>
4708 ///
4709 /// \code
4710 /// __m128d _mm256_extractf128_pd(__m256d V, const int M);
4711 /// \endcode
4712 ///
4713 /// This intrinsic corresponds to the <c> VEXTRACTF128 </c> instruction.
4714 ///
4715 /// \param V
4716 ///    A 256-bit vector of [4 x double].
4717 /// \param M
4718 ///    An immediate integer. The least significant bit determines which bits are
4719 ///    extracted from the first parameter: \n
4720 ///    If bit [0] of \a M is 0, bits [127:0] of \a V are copied to the
4721 ///    result. \n
4722 ///    If bit [0] of \a M is 1, bits [255:128] of \a V are copied to the result.
4723 /// \returns A 128-bit vector of [2 x double] containing the extracted bits.
4724 #define _mm256_extractf128_pd(V, M) \
4725   (__m128d)__builtin_ia32_vextractf128_pd256((__v4df)(__m256d)(V), (int)(M))
4726 
4727 /// Extracts either the upper or the lower 128 bits from a 256-bit
4728 ///    integer vector, as determined by the immediate integer parameter, and
4729 ///    returns the extracted bits as a 128-bit integer vector.
4730 ///
4731 /// \headerfile <x86intrin.h>
4732 ///
4733 /// \code
4734 /// __m128i _mm256_extractf128_si256(__m256i V, const int M);
4735 /// \endcode
4736 ///
4737 /// This intrinsic corresponds to the <c> VEXTRACTF128 </c> instruction.
4738 ///
4739 /// \param V
4740 ///    A 256-bit integer vector.
4741 /// \param M
4742 ///    An immediate integer. The least significant bit determines which bits are
4743 ///    extracted from the first parameter:  \n
4744 ///    If bit [0] of \a M is 0, bits [127:0] of \a V are copied to the
4745 ///    result. \n
4746 ///    If bit [0] of \a M is 1, bits [255:128] of \a V are copied to the result.
4747 /// \returns A 128-bit integer vector containing the extracted bits.
4748 #define _mm256_extractf128_si256(V, M) \
4749   (__m128i)__builtin_ia32_vextractf128_si256((__v8si)(__m256i)(V), (int)(M))
4750 
4751 /* SIMD load ops (unaligned) */
4752 /// Loads two 128-bit floating-point vectors of [4 x float] from
4753 ///    unaligned memory locations and constructs a 256-bit floating-point vector
4754 ///    of [8 x float] by concatenating the two 128-bit vectors.
4755 ///
4756 /// \headerfile <x86intrin.h>
4757 ///
4758 /// This intrinsic corresponds to load instructions followed by the
4759 ///   <c> VINSERTF128 </c> instruction.
4760 ///
4761 /// \param __addr_hi
4762 ///    A pointer to a 128-bit memory location containing 4 consecutive
4763 ///    single-precision floating-point values. These values are to be copied to
4764 ///    bits[255:128] of the result. The address of the memory location does not
4765 ///    have to be aligned.
4766 /// \param __addr_lo
4767 ///    A pointer to a 128-bit memory location containing 4 consecutive
4768 ///    single-precision floating-point values. These values are to be copied to
4769 ///    bits[127:0] of the result. The address of the memory location does not
4770 ///    have to be aligned.
4771 /// \returns A 256-bit floating-point vector of [8 x float] containing the
4772 ///    concatenated result.
4773 static __inline __m256 __DEFAULT_FN_ATTRS
4774 _mm256_loadu2_m128(float const *__addr_hi, float const *__addr_lo)
4775 {
4776   __m256 __v256 = _mm256_castps128_ps256(_mm_loadu_ps(__addr_lo));
4777   return _mm256_insertf128_ps(__v256, _mm_loadu_ps(__addr_hi), 1);
4778 }
4779 
4780 /// Loads two 128-bit floating-point vectors of [2 x double] from
4781 ///    unaligned memory locations and constructs a 256-bit floating-point vector
4782 ///    of [4 x double] by concatenating the two 128-bit vectors.
4783 ///
4784 /// \headerfile <x86intrin.h>
4785 ///
4786 /// This intrinsic corresponds to load instructions followed by the
4787 ///   <c> VINSERTF128 </c> instruction.
4788 ///
4789 /// \param __addr_hi
4790 ///    A pointer to a 128-bit memory location containing two consecutive
4791 ///    double-precision floating-point values. These values are to be copied to
4792 ///    bits[255:128] of the result. The address of the memory location does not
4793 ///    have to be aligned.
4794 /// \param __addr_lo
4795 ///    A pointer to a 128-bit memory location containing two consecutive
4796 ///    double-precision floating-point values. These values are to be copied to
4797 ///    bits[127:0] of the result. The address of the memory location does not
4798 ///    have to be aligned.
4799 /// \returns A 256-bit floating-point vector of [4 x double] containing the
4800 ///    concatenated result.
4801 static __inline __m256d __DEFAULT_FN_ATTRS
4802 _mm256_loadu2_m128d(double const *__addr_hi, double const *__addr_lo)
4803 {
4804   __m256d __v256 = _mm256_castpd128_pd256(_mm_loadu_pd(__addr_lo));
4805   return _mm256_insertf128_pd(__v256, _mm_loadu_pd(__addr_hi), 1);
4806 }
4807 
4808 /// Loads two 128-bit integer vectors from unaligned memory locations and
4809 ///    constructs a 256-bit integer vector by concatenating the two 128-bit
4810 ///    vectors.
4811 ///
4812 /// \headerfile <x86intrin.h>
4813 ///
4814 /// This intrinsic corresponds to load instructions followed by the
4815 ///   <c> VINSERTF128 </c> instruction.
4816 ///
4817 /// \param __addr_hi
4818 ///    A pointer to a 128-bit memory location containing a 128-bit integer
4819 ///    vector. This vector is to be copied to bits[255:128] of the result. The
4820 ///    address of the memory location does not have to be aligned.
4821 /// \param __addr_lo
4822 ///    A pointer to a 128-bit memory location containing a 128-bit integer
4823 ///    vector. This vector is to be copied to bits[127:0] of the result. The
4824 ///    address of the memory location does not have to be aligned.
4825 /// \returns A 256-bit integer vector containing the concatenated result.
4826 static __inline __m256i __DEFAULT_FN_ATTRS
4827 _mm256_loadu2_m128i(__m128i_u const *__addr_hi, __m128i_u const *__addr_lo)
4828 {
4829   __m256i __v256 = _mm256_castsi128_si256(_mm_loadu_si128(__addr_lo));
4830   return _mm256_insertf128_si256(__v256, _mm_loadu_si128(__addr_hi), 1);
4831 }
4832 
4833 /* SIMD store ops (unaligned) */
4834 /// Stores the upper and lower 128 bits of a 256-bit floating-point
4835 ///    vector of [8 x float] into two different unaligned memory locations.
4836 ///
4837 /// \headerfile <x86intrin.h>
4838 ///
4839 /// This intrinsic corresponds to the <c> VEXTRACTF128 </c> instruction and the
4840 ///   store instructions.
4841 ///
4842 /// \param __addr_hi
4843 ///    A pointer to a 128-bit memory location. Bits[255:128] of \a __a are to be
4844 ///    copied to this memory location. The address of this memory location does
4845 ///    not have to be aligned.
4846 /// \param __addr_lo
4847 ///    A pointer to a 128-bit memory location. Bits[127:0] of \a __a are to be
4848 ///    copied to this memory location. The address of this memory location does
4849 ///    not have to be aligned.
4850 /// \param __a
4851 ///    A 256-bit floating-point vector of [8 x float].
4852 static __inline void __DEFAULT_FN_ATTRS
4853 _mm256_storeu2_m128(float *__addr_hi, float *__addr_lo, __m256 __a)
4854 {
4855   __m128 __v128;
4856 
4857   __v128 = _mm256_castps256_ps128(__a);
4858   _mm_storeu_ps(__addr_lo, __v128);
4859   __v128 = _mm256_extractf128_ps(__a, 1);
4860   _mm_storeu_ps(__addr_hi, __v128);
4861 }
4862 
4863 /// Stores the upper and lower 128 bits of a 256-bit floating-point
4864 ///    vector of [4 x double] into two different unaligned memory locations.
4865 ///
4866 /// \headerfile <x86intrin.h>
4867 ///
4868 /// This intrinsic corresponds to the <c> VEXTRACTF128 </c> instruction and the
4869 ///   store instructions.
4870 ///
4871 /// \param __addr_hi
4872 ///    A pointer to a 128-bit memory location. Bits[255:128] of \a __a are to be
4873 ///    copied to this memory location. The address of this memory location does
4874 ///    not have to be aligned.
4875 /// \param __addr_lo
4876 ///    A pointer to a 128-bit memory location. Bits[127:0] of \a __a are to be
4877 ///    copied to this memory location. The address of this memory location does
4878 ///    not have to be aligned.
4879 /// \param __a
4880 ///    A 256-bit floating-point vector of [4 x double].
4881 static __inline void __DEFAULT_FN_ATTRS
4882 _mm256_storeu2_m128d(double *__addr_hi, double *__addr_lo, __m256d __a)
4883 {
4884   __m128d __v128;
4885 
4886   __v128 = _mm256_castpd256_pd128(__a);
4887   _mm_storeu_pd(__addr_lo, __v128);
4888   __v128 = _mm256_extractf128_pd(__a, 1);
4889   _mm_storeu_pd(__addr_hi, __v128);
4890 }
4891 
4892 /// Stores the upper and lower 128 bits of a 256-bit integer vector into
4893 ///    two different unaligned memory locations.
4894 ///
4895 /// \headerfile <x86intrin.h>
4896 ///
4897 /// This intrinsic corresponds to the <c> VEXTRACTF128 </c> instruction and the
4898 ///   store instructions.
4899 ///
4900 /// \param __addr_hi
4901 ///    A pointer to a 128-bit memory location. Bits[255:128] of \a __a are to be
4902 ///    copied to this memory location. The address of this memory location does
4903 ///    not have to be aligned.
4904 /// \param __addr_lo
4905 ///    A pointer to a 128-bit memory location. Bits[127:0] of \a __a are to be
4906 ///    copied to this memory location. The address of this memory location does
4907 ///    not have to be aligned.
4908 /// \param __a
4909 ///    A 256-bit integer vector.
4910 static __inline void __DEFAULT_FN_ATTRS
4911 _mm256_storeu2_m128i(__m128i_u *__addr_hi, __m128i_u *__addr_lo, __m256i __a)
4912 {
4913   __m128i __v128;
4914 
4915   __v128 = _mm256_castsi256_si128(__a);
4916   _mm_storeu_si128(__addr_lo, __v128);
4917   __v128 = _mm256_extractf128_si256(__a, 1);
4918   _mm_storeu_si128(__addr_hi, __v128);
4919 }
4920 
4921 /// Constructs a 256-bit floating-point vector of [8 x float] by
4922 ///    concatenating two 128-bit floating-point vectors of [4 x float].
4923 ///
4924 /// \headerfile <x86intrin.h>
4925 ///
4926 /// This intrinsic corresponds to the <c> VINSERTF128 </c> instruction.
4927 ///
4928 /// \param __hi
4929 ///    A 128-bit floating-point vector of [4 x float] to be copied to the upper
4930 ///    128 bits of the result.
4931 /// \param __lo
4932 ///    A 128-bit floating-point vector of [4 x float] to be copied to the lower
4933 ///    128 bits of the result.
4934 /// \returns A 256-bit floating-point vector of [8 x float] containing the
4935 ///    concatenated result.
4936 static __inline __m256 __DEFAULT_FN_ATTRS
4937 _mm256_set_m128 (__m128 __hi, __m128 __lo)
4938 {
4939   return (__m256) __builtin_shufflevector((__v4sf)__lo, (__v4sf)__hi, 0, 1, 2, 3, 4, 5, 6, 7);
4940 }
4941 
4942 /// Constructs a 256-bit floating-point vector of [4 x double] by
4943 ///    concatenating two 128-bit floating-point vectors of [2 x double].
4944 ///
4945 /// \headerfile <x86intrin.h>
4946 ///
4947 /// This intrinsic corresponds to the <c> VINSERTF128 </c> instruction.
4948 ///
4949 /// \param __hi
4950 ///    A 128-bit floating-point vector of [2 x double] to be copied to the upper
4951 ///    128 bits of the result.
4952 /// \param __lo
4953 ///    A 128-bit floating-point vector of [2 x double] to be copied to the lower
4954 ///    128 bits of the result.
4955 /// \returns A 256-bit floating-point vector of [4 x double] containing the
4956 ///    concatenated result.
4957 static __inline __m256d __DEFAULT_FN_ATTRS
4958 _mm256_set_m128d (__m128d __hi, __m128d __lo)
4959 {
4960   return (__m256d) __builtin_shufflevector((__v2df)__lo, (__v2df)__hi, 0, 1, 2, 3);
4961 }
4962 
4963 /// Constructs a 256-bit integer vector by concatenating two 128-bit
4964 ///    integer vectors.
4965 ///
4966 /// \headerfile <x86intrin.h>
4967 ///
4968 /// This intrinsic corresponds to the <c> VINSERTF128 </c> instruction.
4969 ///
4970 /// \param __hi
4971 ///    A 128-bit integer vector to be copied to the upper 128 bits of the
4972 ///    result.
4973 /// \param __lo
4974 ///    A 128-bit integer vector to be copied to the lower 128 bits of the
4975 ///    result.
4976 /// \returns A 256-bit integer vector containing the concatenated result.
4977 static __inline __m256i __DEFAULT_FN_ATTRS
4978 _mm256_set_m128i (__m128i __hi, __m128i __lo)
4979 {
4980   return (__m256i) __builtin_shufflevector((__v2di)__lo, (__v2di)__hi, 0, 1, 2, 3);
4981 }
4982 
4983 /// Constructs a 256-bit floating-point vector of [8 x float] by
4984 ///    concatenating two 128-bit floating-point vectors of [4 x float]. This is
4985 ///    similar to _mm256_set_m128, but the order of the input parameters is
4986 ///    swapped.
4987 ///
4988 /// \headerfile <x86intrin.h>
4989 ///
4990 /// This intrinsic corresponds to the <c> VINSERTF128 </c> instruction.
4991 ///
4992 /// \param __lo
4993 ///    A 128-bit floating-point vector of [4 x float] to be copied to the lower
4994 ///    128 bits of the result.
4995 /// \param __hi
4996 ///    A 128-bit floating-point vector of [4 x float] to be copied to the upper
4997 ///    128 bits of the result.
4998 /// \returns A 256-bit floating-point vector of [8 x float] containing the
4999 ///    concatenated result.
5000 static __inline __m256 __DEFAULT_FN_ATTRS
5001 _mm256_setr_m128 (__m128 __lo, __m128 __hi)
5002 {
5003   return _mm256_set_m128(__hi, __lo);
5004 }
5005 
5006 /// Constructs a 256-bit floating-point vector of [4 x double] by
5007 ///    concatenating two 128-bit floating-point vectors of [2 x double]. This is
5008 ///    similar to _mm256_set_m128d, but the order of the input parameters is
5009 ///    swapped.
5010 ///
5011 /// \headerfile <x86intrin.h>
5012 ///
5013 /// This intrinsic corresponds to the <c> VINSERTF128 </c> instruction.
5014 ///
5015 /// \param __lo
5016 ///    A 128-bit floating-point vector of [2 x double] to be copied to the lower
5017 ///    128 bits of the result.
5018 /// \param __hi
5019 ///    A 128-bit floating-point vector of [2 x double] to be copied to the upper
5020 ///    128 bits of the result.
5021 /// \returns A 256-bit floating-point vector of [4 x double] containing the
5022 ///    concatenated result.
5023 static __inline __m256d __DEFAULT_FN_ATTRS
5024 _mm256_setr_m128d (__m128d __lo, __m128d __hi)
5025 {
5026   return (__m256d)_mm256_set_m128d(__hi, __lo);
5027 }
5028 
5029 /// Constructs a 256-bit integer vector by concatenating two 128-bit
5030 ///    integer vectors. This is similar to _mm256_set_m128i, but the order of
5031 ///    the input parameters is swapped.
5032 ///
5033 /// \headerfile <x86intrin.h>
5034 ///
5035 /// This intrinsic corresponds to the <c> VINSERTF128 </c> instruction.
5036 ///
5037 /// \param __lo
5038 ///    A 128-bit integer vector to be copied to the lower 128 bits of the
5039 ///    result.
5040 /// \param __hi
5041 ///    A 128-bit integer vector to be copied to the upper 128 bits of the
5042 ///    result.
5043 /// \returns A 256-bit integer vector containing the concatenated result.
5044 static __inline __m256i __DEFAULT_FN_ATTRS
5045 _mm256_setr_m128i (__m128i __lo, __m128i __hi)
5046 {
5047   return (__m256i)_mm256_set_m128i(__hi, __lo);
5048 }
5049 
5050 #undef __DEFAULT_FN_ATTRS
5051 #undef __DEFAULT_FN_ATTRS128
5052 
5053 #endif /* __AVXINTRIN_H */
5054