1 /*===---- immintrin.h - Intel 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 #define __IMMINTRIN_H
12 
13 #if !defined(__i386__) && !defined(__x86_64__)
14 #error "This header is only meant to be used on x86 and x64 architecture"
15 #endif
16 
17 #include <x86gprintrin.h>
18 
19 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
20     defined(__MMX__)
21 #include <mmintrin.h>
22 #endif
23 
24 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
25     defined(__SSE__)
26 #include <xmmintrin.h>
27 #endif
28 
29 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
30     defined(__SSE2__)
31 #include <emmintrin.h>
32 #endif
33 
34 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
35     defined(__SSE3__)
36 #include <pmmintrin.h>
37 #endif
38 
39 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
40     defined(__SSSE3__)
41 #include <tmmintrin.h>
42 #endif
43 
44 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
45     (defined(__SSE4_2__) || defined(__SSE4_1__))
46 #include <smmintrin.h>
47 #endif
48 
49 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
50     (defined(__AES__) || defined(__PCLMUL__))
51 #include <wmmintrin.h>
52 #endif
53 
54 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
55     defined(__CLFLUSHOPT__)
56 #include <clflushoptintrin.h>
57 #endif
58 
59 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
60     defined(__CLWB__)
61 #include <clwbintrin.h>
62 #endif
63 
64 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
65     defined(__AVX__)
66 #include <avxintrin.h>
67 #endif
68 
69 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
70     defined(__AVX2__)
71 #include <avx2intrin.h>
72 #endif
73 
74 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
75     defined(__F16C__)
76 #include <f16cintrin.h>
77 #endif
78 
79 /* No feature check desired due to internal checks */
80 #include <bmiintrin.h>
81 
82 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
83     defined(__BMI2__)
84 #include <bmi2intrin.h>
85 #endif
86 
87 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
88     defined(__LZCNT__)
89 #include <lzcntintrin.h>
90 #endif
91 
92 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
93     defined(__POPCNT__)
94 #include <popcntintrin.h>
95 #endif
96 
97 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
98     defined(__FMA__)
99 #include <fmaintrin.h>
100 #endif
101 
102 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
103     defined(__AVX512F__)
104 #include <avx512fintrin.h>
105 #endif
106 
107 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
108     defined(__AVX512VL__)
109 #include <avx512vlintrin.h>
110 #endif
111 
112 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
113     defined(__AVX512BW__)
114 #include <avx512bwintrin.h>
115 #endif
116 
117 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
118     defined(__AVX512BITALG__)
119 #include <avx512bitalgintrin.h>
120 #endif
121 
122 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
123     defined(__AVX512CD__)
124 #include <avx512cdintrin.h>
125 #endif
126 
127 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
128     defined(__AVX512VPOPCNTDQ__)
129 #include <avx512vpopcntdqintrin.h>
130 #endif
131 
132 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
133     (defined(__AVX512VL__) && defined(__AVX512VPOPCNTDQ__))
134 #include <avx512vpopcntdqvlintrin.h>
135 #endif
136 
137 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
138     defined(__AVX512VNNI__)
139 #include <avx512vnniintrin.h>
140 #endif
141 
142 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
143     (defined(__AVX512VL__) && defined(__AVX512VNNI__))
144 #include <avx512vlvnniintrin.h>
145 #endif
146 
147 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
148     defined(__AVXVNNI__)
149 #include <avxvnniintrin.h>
150 #endif
151 
152 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
153     defined(__AVX512DQ__)
154 #include <avx512dqintrin.h>
155 #endif
156 
157 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
158     (defined(__AVX512VL__) && defined(__AVX512BITALG__))
159 #include <avx512vlbitalgintrin.h>
160 #endif
161 
162 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
163     (defined(__AVX512VL__) && defined(__AVX512BW__))
164 #include <avx512vlbwintrin.h>
165 #endif
166 
167 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
168     (defined(__AVX512VL__) && defined(__AVX512CD__))
169 #include <avx512vlcdintrin.h>
170 #endif
171 
172 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
173     (defined(__AVX512VL__) && defined(__AVX512DQ__))
174 #include <avx512vldqintrin.h>
175 #endif
176 
177 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
178     defined(__AVX512ER__)
179 #include <avx512erintrin.h>
180 #endif
181 
182 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
183     defined(__AVX512IFMA__)
184 #include <avx512ifmaintrin.h>
185 #endif
186 
187 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
188     (defined(__AVX512IFMA__) && defined(__AVX512VL__))
189 #include <avx512ifmavlintrin.h>
190 #endif
191 
192 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
193     defined(__AVX512VBMI__)
194 #include <avx512vbmiintrin.h>
195 #endif
196 
197 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
198     (defined(__AVX512VBMI__) && defined(__AVX512VL__))
199 #include <avx512vbmivlintrin.h>
200 #endif
201 
202 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
203     defined(__AVX512VBMI2__)
204 #include <avx512vbmi2intrin.h>
205 #endif
206 
207 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
208     (defined(__AVX512VBMI2__) && defined(__AVX512VL__))
209 #include <avx512vlvbmi2intrin.h>
210 #endif
211 
212 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
213     defined(__AVX512PF__)
214 #include <avx512pfintrin.h>
215 #endif
216 
217 /*
218  * FIXME: _Float16 type is legal only when HW support float16 operation.
219  * We use __AVX512FP16__ to identify if float16 is supported or not, so
220  * when float16 is not supported, the related header is not included.
221  *
222  */
223 #if defined(__AVX512FP16__)
224 #include <avx512fp16intrin.h>
225 #endif
226 
227 #if defined(__AVX512FP16__) && defined(__AVX512VL__)
228 #include <avx512vlfp16intrin.h>
229 #endif
230 
231 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
232     defined(__AVX512BF16__)
233 #include <avx512bf16intrin.h>
234 #endif
235 
236 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
237     (defined(__AVX512VL__) && defined(__AVX512BF16__))
238 #include <avx512vlbf16intrin.h>
239 #endif
240 
241 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
242     defined(__PKU__)
243 #include <pkuintrin.h>
244 #endif
245 
246 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
247     defined(__VPCLMULQDQ__)
248 #include <vpclmulqdqintrin.h>
249 #endif
250 
251 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
252     defined(__VAES__)
253 #include <vaesintrin.h>
254 #endif
255 
256 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
257     defined(__GFNI__)
258 #include <gfniintrin.h>
259 #endif
260 
261 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
262     defined(__RDPID__)
263 /// Returns the value of the IA32_TSC_AUX MSR (0xc0000103).
264 ///
265 /// \headerfile <immintrin.h>
266 ///
267 /// This intrinsic corresponds to the <c> RDPID </c> instruction.
268 static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__, __target__("rdpid")))
269 _rdpid_u32(void) {
270   return __builtin_ia32_rdpid();
271 }
272 #endif // __RDPID__
273 
274 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
275     defined(__RDRND__)
276 static __inline__ int __attribute__((__always_inline__, __nodebug__, __target__("rdrnd")))
277 _rdrand16_step(unsigned short *__p)
278 {
279   return (int)__builtin_ia32_rdrand16_step(__p);
280 }
281 
282 static __inline__ int __attribute__((__always_inline__, __nodebug__, __target__("rdrnd")))
283 _rdrand32_step(unsigned int *__p)
284 {
285   return (int)__builtin_ia32_rdrand32_step(__p);
286 }
287 
288 #ifdef __x86_64__
289 static __inline__ int __attribute__((__always_inline__, __nodebug__, __target__("rdrnd")))
290 _rdrand64_step(unsigned long long *__p)
291 {
292   return (int)__builtin_ia32_rdrand64_step(__p);
293 }
294 #endif
295 #endif /* __RDRND__ */
296 
297 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
298     defined(__FSGSBASE__)
299 #ifdef __x86_64__
300 static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase")))
301 _readfsbase_u32(void)
302 {
303   return __builtin_ia32_rdfsbase32();
304 }
305 
306 static __inline__ unsigned long long __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase")))
307 _readfsbase_u64(void)
308 {
309   return __builtin_ia32_rdfsbase64();
310 }
311 
312 static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase")))
313 _readgsbase_u32(void)
314 {
315   return __builtin_ia32_rdgsbase32();
316 }
317 
318 static __inline__ unsigned long long __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase")))
319 _readgsbase_u64(void)
320 {
321   return __builtin_ia32_rdgsbase64();
322 }
323 
324 static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase")))
325 _writefsbase_u32(unsigned int __V)
326 {
327   __builtin_ia32_wrfsbase32(__V);
328 }
329 
330 static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase")))
331 _writefsbase_u64(unsigned long long __V)
332 {
333   __builtin_ia32_wrfsbase64(__V);
334 }
335 
336 static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase")))
337 _writegsbase_u32(unsigned int __V)
338 {
339   __builtin_ia32_wrgsbase32(__V);
340 }
341 
342 static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase")))
343 _writegsbase_u64(unsigned long long __V)
344 {
345   __builtin_ia32_wrgsbase64(__V);
346 }
347 
348 #endif
349 #endif /* __FSGSBASE__ */
350 
351 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
352     defined(__MOVBE__)
353 
354 /* The structs used below are to force the load/store to be unaligned. This
355  * is accomplished with the __packed__ attribute. The __may_alias__ prevents
356  * tbaa metadata from being generated based on the struct and the type of the
357  * field inside of it.
358  */
359 
360 static __inline__ short __attribute__((__always_inline__, __nodebug__, __target__("movbe")))
361 _loadbe_i16(void const * __P) {
362   struct __loadu_i16 {
363     unsigned short __v;
364   } __attribute__((__packed__, __may_alias__));
365   return (short)__builtin_bswap16(((const struct __loadu_i16*)__P)->__v);
366 }
367 
368 static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("movbe")))
369 _storebe_i16(void * __P, short __D) {
370   struct __storeu_i16 {
371     unsigned short __v;
372   } __attribute__((__packed__, __may_alias__));
373   ((struct __storeu_i16*)__P)->__v = __builtin_bswap16((unsigned short)__D);
374 }
375 
376 static __inline__ int __attribute__((__always_inline__, __nodebug__, __target__("movbe")))
377 _loadbe_i32(void const * __P) {
378   struct __loadu_i32 {
379     unsigned int __v;
380   } __attribute__((__packed__, __may_alias__));
381   return (int)__builtin_bswap32(((const struct __loadu_i32*)__P)->__v);
382 }
383 
384 static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("movbe")))
385 _storebe_i32(void * __P, int __D) {
386   struct __storeu_i32 {
387     unsigned int __v;
388   } __attribute__((__packed__, __may_alias__));
389   ((struct __storeu_i32*)__P)->__v = __builtin_bswap32((unsigned int)__D);
390 }
391 
392 #ifdef __x86_64__
393 static __inline__ long long __attribute__((__always_inline__, __nodebug__, __target__("movbe")))
394 _loadbe_i64(void const * __P) {
395   struct __loadu_i64 {
396     unsigned long long __v;
397   } __attribute__((__packed__, __may_alias__));
398   return (long long)__builtin_bswap64(((const struct __loadu_i64*)__P)->__v);
399 }
400 
401 static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("movbe")))
402 _storebe_i64(void * __P, long long __D) {
403   struct __storeu_i64 {
404     unsigned long long __v;
405   } __attribute__((__packed__, __may_alias__));
406   ((struct __storeu_i64*)__P)->__v = __builtin_bswap64((unsigned long long)__D);
407 }
408 #endif
409 #endif /* __MOVBE */
410 
411 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
412     defined(__RTM__)
413 #include <rtmintrin.h>
414 #include <xtestintrin.h>
415 #endif
416 
417 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
418     defined(__SHA__)
419 #include <shaintrin.h>
420 #endif
421 
422 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
423     defined(__FXSR__)
424 #include <fxsrintrin.h>
425 #endif
426 
427 /* No feature check desired due to internal MSC_VER checks */
428 #include <xsaveintrin.h>
429 
430 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
431     defined(__XSAVEOPT__)
432 #include <xsaveoptintrin.h>
433 #endif
434 
435 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
436     defined(__XSAVEC__)
437 #include <xsavecintrin.h>
438 #endif
439 
440 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
441     defined(__XSAVES__)
442 #include <xsavesintrin.h>
443 #endif
444 
445 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
446     defined(__SHSTK__)
447 #include <cetintrin.h>
448 #endif
449 
450 /* Some intrinsics inside adxintrin.h are available only on processors with ADX,
451  * whereas others are also available at all times. */
452 #include <adxintrin.h>
453 
454 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
455     defined(__RDSEED__)
456 #include <rdseedintrin.h>
457 #endif
458 
459 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
460     defined(__WBNOINVD__)
461 #include <wbnoinvdintrin.h>
462 #endif
463 
464 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
465     defined(__CLDEMOTE__)
466 #include <cldemoteintrin.h>
467 #endif
468 
469 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
470     defined(__WAITPKG__)
471 #include <waitpkgintrin.h>
472 #endif
473 
474 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
475     defined(__MOVDIRI__) || defined(__MOVDIR64B__)
476 #include <movdirintrin.h>
477 #endif
478 
479 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
480     defined(__PCONFIG__)
481 #include <pconfigintrin.h>
482 #endif
483 
484 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
485     defined(__SGX__)
486 #include <sgxintrin.h>
487 #endif
488 
489 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
490     defined(__PTWRITE__)
491 #include <ptwriteintrin.h>
492 #endif
493 
494 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
495     defined(__INVPCID__)
496 #include <invpcidintrin.h>
497 #endif
498 
499 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
500     defined(__KL__) || defined(__WIDEKL__)
501 #include <keylockerintrin.h>
502 #endif
503 
504 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
505     defined(__AMXTILE__) || defined(__AMXINT8__) || defined(__AMXBF16__)
506 #include <amxintrin.h>
507 #endif
508 
509 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
510     defined(__AVX512VP2INTERSECT__)
511 #include <avx512vp2intersectintrin.h>
512 #endif
513 
514 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
515     (defined(__AVX512VL__) && defined(__AVX512VP2INTERSECT__))
516 #include <avx512vlvp2intersectintrin.h>
517 #endif
518 
519 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
520     defined(__ENQCMD__)
521 #include <enqcmdintrin.h>
522 #endif
523 
524 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
525     defined(__SERIALIZE__)
526 #include <serializeintrin.h>
527 #endif
528 
529 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
530     defined(__TSXLDTRK__)
531 #include <tsxldtrkintrin.h>
532 #endif
533 
534 #if defined(_MSC_VER) && __has_extension(gnu_asm)
535 /* Define the default attributes for these intrinsics */
536 #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__))
537 #ifdef __cplusplus
538 extern "C" {
539 #endif
540 /*----------------------------------------------------------------------------*\
541 |* Interlocked Exchange HLE
542 \*----------------------------------------------------------------------------*/
543 #if defined(__i386__) || defined(__x86_64__)
544 static __inline__ long __DEFAULT_FN_ATTRS
545 _InterlockedExchange_HLEAcquire(long volatile *_Target, long _Value) {
546   __asm__ __volatile__(".byte 0xf2 ; lock ; xchg {%0, %1|%1, %0}"
547                        : "+r" (_Value), "+m" (*_Target) :: "memory");
548   return _Value;
549 }
550 static __inline__ long __DEFAULT_FN_ATTRS
551 _InterlockedExchange_HLERelease(long volatile *_Target, long _Value) {
552   __asm__ __volatile__(".byte 0xf3 ; lock ; xchg {%0, %1|%1, %0}"
553                        : "+r" (_Value), "+m" (*_Target) :: "memory");
554   return _Value;
555 }
556 #endif
557 #if defined(__x86_64__)
558 static __inline__ __int64 __DEFAULT_FN_ATTRS
559 _InterlockedExchange64_HLEAcquire(__int64 volatile *_Target, __int64 _Value) {
560   __asm__ __volatile__(".byte 0xf2 ; lock ; xchg {%0, %1|%1, %0}"
561                        : "+r" (_Value), "+m" (*_Target) :: "memory");
562   return _Value;
563 }
564 static __inline__ __int64 __DEFAULT_FN_ATTRS
565 _InterlockedExchange64_HLERelease(__int64 volatile *_Target, __int64 _Value) {
566   __asm__ __volatile__(".byte 0xf3 ; lock ; xchg {%0, %1|%1, %0}"
567                        : "+r" (_Value), "+m" (*_Target) :: "memory");
568   return _Value;
569 }
570 #endif
571 /*----------------------------------------------------------------------------*\
572 |* Interlocked Compare Exchange HLE
573 \*----------------------------------------------------------------------------*/
574 #if defined(__i386__) || defined(__x86_64__)
575 static __inline__ long __DEFAULT_FN_ATTRS
576 _InterlockedCompareExchange_HLEAcquire(long volatile *_Destination,
577                               long _Exchange, long _Comparand) {
578   __asm__ __volatile__(".byte 0xf2 ; lock ; cmpxchg {%2, %1|%1, %2}"
579                        : "+a" (_Comparand), "+m" (*_Destination)
580                        : "r" (_Exchange) : "memory");
581   return _Comparand;
582 }
583 static __inline__ long __DEFAULT_FN_ATTRS
584 _InterlockedCompareExchange_HLERelease(long volatile *_Destination,
585                               long _Exchange, long _Comparand) {
586   __asm__ __volatile__(".byte 0xf3 ; lock ; cmpxchg {%2, %1|%1, %2}"
587                        : "+a" (_Comparand), "+m" (*_Destination)
588                        : "r" (_Exchange) : "memory");
589   return _Comparand;
590 }
591 #endif
592 #if defined(__x86_64__)
593 static __inline__ __int64 __DEFAULT_FN_ATTRS
594 _InterlockedCompareExchange64_HLEAcquire(__int64 volatile *_Destination,
595                               __int64 _Exchange, __int64 _Comparand) {
596   __asm__ __volatile__(".byte 0xf2 ; lock ; cmpxchg {%2, %1|%1, %2}"
597                        : "+a" (_Comparand), "+m" (*_Destination)
598                        : "r" (_Exchange) : "memory");
599   return _Comparand;
600 }
601 static __inline__ __int64 __DEFAULT_FN_ATTRS
602 _InterlockedCompareExchange64_HLERelease(__int64 volatile *_Destination,
603                               __int64 _Exchange, __int64 _Comparand) {
604   __asm__ __volatile__(".byte 0xf3 ; lock ; cmpxchg {%2, %1|%1, %2}"
605                        : "+a" (_Comparand), "+m" (*_Destination)
606                        : "r" (_Exchange) : "memory");
607   return _Comparand;
608 }
609 #endif
610 #ifdef __cplusplus
611 }
612 #endif
613 
614 #undef __DEFAULT_FN_ATTRS
615 
616 #endif /* defined(_MSC_VER) && __has_extension(gnu_asm) */
617 
618 #endif /* __IMMINTRIN_H */
619