ETISS 0.8.0
Extendable Translating Instruction Set Simulator (version 0.8.0)
•All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Macros Modules Pages
immintrin.h
Go to the documentation of this file.
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(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
14 defined(__MMX__)
15#include <mmintrin.h>
16#endif
17
18#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
19 defined(__SSE__)
20#include <xmmintrin.h>
21#endif
22
23#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
24 defined(__SSE2__)
25#include <emmintrin.h>
26#endif
27
28#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
29 defined(__SSE3__)
30#include <pmmintrin.h>
31#endif
32
33#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
34 defined(__SSSE3__)
35#include <tmmintrin.h>
36#endif
37
38#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
39 (defined(__SSE4_2__) || defined(__SSE4_1__))
40#include <smmintrin.h>
41#endif
42
43#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
44 (defined(__AES__) || defined(__PCLMUL__))
45#include <wmmintrin.h>
46#endif
47
48#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
49 defined(__CLFLUSHOPT__)
50#include <clflushoptintrin.h>
51#endif
52
53#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
54 defined(__CLWB__)
55#include <clwbintrin.h>
56#endif
57
58#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
59 defined(__AVX__)
60#include <avxintrin.h>
61#endif
62
63#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
64 defined(__AVX2__)
65#include <avx2intrin.h>
66#endif
67
68#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
69 defined(__F16C__)
70#include <f16cintrin.h>
71#endif
72
73#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
74 defined(__VPCLMULQDQ__)
75#include <vpclmulqdqintrin.h>
76#endif
77
78/* No feature check desired due to internal checks */
79#include <bmiintrin.h>
80
81#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
82 defined(__BMI2__)
83#include <bmi2intrin.h>
84#endif
85
86#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
87 defined(__LZCNT__)
88#include <lzcntintrin.h>
89#endif
90
91#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
92 defined(__POPCNT__)
93#include <popcntintrin.h>
94#endif
95
96#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
97 defined(__FMA__)
98#include <fmaintrin.h>
99#endif
100
101#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
102 defined(__AVX512F__)
103#include <avx512fintrin.h>
104#endif
105
106#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
107 defined(__AVX512VL__)
108#include <avx512vlintrin.h>
109#endif
110
111#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
112 defined(__AVX512BW__)
113#include <avx512bwintrin.h>
114#endif
115
116#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
117 defined(__AVX512BITALG__)
118#include <avx512bitalgintrin.h>
119#endif
120
121#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
122 defined(__AVX512CD__)
123#include <avx512cdintrin.h>
124#endif
125
126#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
127 defined(__AVX512VPOPCNTDQ__)
128#include <avx512vpopcntdqintrin.h>
129#endif
130
131#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
132 (defined(__AVX512VL__) && defined(__AVX512VPOPCNTDQ__))
133#include <avx512vpopcntdqvlintrin.h>
134#endif
135
136#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
137 defined(__AVX512VNNI__)
138#include <avx512vnniintrin.h>
139#endif
140
141#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
142 (defined(__AVX512VL__) && defined(__AVX512VNNI__))
143#include <avx512vlvnniintrin.h>
144#endif
145
146#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
147 defined(__AVX512DQ__)
148#include <avx512dqintrin.h>
149#endif
150
151#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
152 (defined(__AVX512VL__) && defined(__AVX512BITALG__))
153#include <avx512vlbitalgintrin.h>
154#endif
155
156#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
157 (defined(__AVX512VL__) && defined(__AVX512BW__))
158#include <avx512vlbwintrin.h>
159#endif
160
161#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
162 (defined(__AVX512VL__) && defined(__AVX512CD__))
163#include <avx512vlcdintrin.h>
164#endif
165
166#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
167 (defined(__AVX512VL__) && defined(__AVX512DQ__))
168#include <avx512vldqintrin.h>
169#endif
170
171#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
172 defined(__AVX512ER__)
173#include <avx512erintrin.h>
174#endif
175
176#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
177 defined(__AVX512IFMA__)
178#include <avx512ifmaintrin.h>
179#endif
180
181#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
182 (defined(__AVX512IFMA__) && defined(__AVX512VL__))
183#include <avx512ifmavlintrin.h>
184#endif
185
186#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
187 defined(__AVX512VBMI__)
188#include <avx512vbmiintrin.h>
189#endif
190
191#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
192 (defined(__AVX512VBMI__) && defined(__AVX512VL__))
193#include <avx512vbmivlintrin.h>
194#endif
195
196#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
197 defined(__AVX512VBMI2__)
198#include <avx512vbmi2intrin.h>
199#endif
200
201#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
202 (defined(__AVX512VBMI2__) && defined(__AVX512VL__))
203#include <avx512vlvbmi2intrin.h>
204#endif
205
206#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
207 defined(__AVX512PF__)
208#include <avx512pfintrin.h>
209#endif
210
211#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
212 defined(__AVX512BF16__)
213#include <avx512bf16intrin.h>
214#endif
215
216#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
217 (defined(__AVX512VL__) && defined(__AVX512BF16__))
218#include <avx512vlbf16intrin.h>
219#endif
220
221#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
222 defined(__PKU__)
223#include <pkuintrin.h>
224#endif
225
226#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
227 defined(__VAES__)
228#include <vaesintrin.h>
229#endif
230
231#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
232 defined(__GFNI__)
233#include <gfniintrin.h>
234#endif
235
236#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
237 defined(__RDPID__)
243static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__, __target__("rdpid")))
244_rdpid_u32(void) {
245 return __builtin_ia32_rdpid();
246}
247#endif // __RDPID__
248
249#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
250 defined(__RDRND__)
251static __inline__ int __attribute__((__always_inline__, __nodebug__, __target__("rdrnd")))
252_rdrand16_step(unsigned short *__p)
253{
254 return __builtin_ia32_rdrand16_step(__p);
255}
256
257static __inline__ int __attribute__((__always_inline__, __nodebug__, __target__("rdrnd")))
258_rdrand32_step(unsigned int *__p)
259{
260 return __builtin_ia32_rdrand32_step(__p);
261}
262
263#ifdef __x86_64__
264static __inline__ int __attribute__((__always_inline__, __nodebug__, __target__("rdrnd")))
265_rdrand64_step(unsigned long long *__p)
266{
267 return __builtin_ia32_rdrand64_step(__p);
268}
269#endif
270#endif /* __RDRND__ */
271
272#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
273 defined(__FSGSBASE__)
274#ifdef __x86_64__
275static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase")))
276_readfsbase_u32(void)
277{
278 return __builtin_ia32_rdfsbase32();
279}
280
281static __inline__ unsigned long long __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase")))
282_readfsbase_u64(void)
283{
284 return __builtin_ia32_rdfsbase64();
285}
286
287static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase")))
288_readgsbase_u32(void)
289{
290 return __builtin_ia32_rdgsbase32();
291}
292
293static __inline__ unsigned long long __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase")))
294_readgsbase_u64(void)
295{
296 return __builtin_ia32_rdgsbase64();
297}
298
299static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase")))
300_writefsbase_u32(unsigned int __V)
301{
302 __builtin_ia32_wrfsbase32(__V);
303}
304
305static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase")))
306_writefsbase_u64(unsigned long long __V)
307{
308 __builtin_ia32_wrfsbase64(__V);
309}
310
311static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase")))
312_writegsbase_u32(unsigned int __V)
313{
314 __builtin_ia32_wrgsbase32(__V);
315}
316
317static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase")))
318_writegsbase_u64(unsigned long long __V)
319{
320 __builtin_ia32_wrgsbase64(__V);
321}
322
323#endif
324#endif /* __FSGSBASE__ */
325
326#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
327 defined(__MOVBE__)
328
329/* The structs used below are to force the load/store to be unaligned. This
330 * is accomplished with the __packed__ attribute. The __may_alias__ prevents
331 * tbaa metadata from being generated based on the struct and the type of the
332 * field inside of it.
333 */
334
335static __inline__ short __attribute__((__always_inline__, __nodebug__, __target__("movbe")))
336_loadbe_i16(void const * __P) {
337 struct __loadu_i16 {
338 short __v;
339 } __attribute__((__packed__, __may_alias__));
340 return __builtin_bswap16(((const struct __loadu_i16*)__P)->__v);
341}
342
343static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("movbe")))
344_storebe_i16(void * __P, short __D) {
345 struct __storeu_i16 {
346 short __v;
347 } __attribute__((__packed__, __may_alias__));
348 ((struct __storeu_i16*)__P)->__v = __builtin_bswap16(__D);
349}
350
351static __inline__ int __attribute__((__always_inline__, __nodebug__, __target__("movbe")))
352_loadbe_i32(void const * __P) {
353 struct __loadu_i32 {
354 int __v;
355 } __attribute__((__packed__, __may_alias__));
356 return __builtin_bswap32(((const struct __loadu_i32*)__P)->__v);
357}
358
359static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("movbe")))
360_storebe_i32(void * __P, int __D) {
361 struct __storeu_i32 {
362 int __v;
363 } __attribute__((__packed__, __may_alias__));
364 ((struct __storeu_i32*)__P)->__v = __builtin_bswap32(__D);
365}
366
367#ifdef __x86_64__
368static __inline__ long long __attribute__((__always_inline__, __nodebug__, __target__("movbe")))
369_loadbe_i64(void const * __P) {
370 struct __loadu_i64 {
371 long long __v;
372 } __attribute__((__packed__, __may_alias__));
373 return __builtin_bswap64(((const struct __loadu_i64*)__P)->__v);
374}
375
376static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("movbe")))
377_storebe_i64(void * __P, long long __D) {
378 struct __storeu_i64 {
379 long long __v;
380 } __attribute__((__packed__, __may_alias__));
381 ((struct __storeu_i64*)__P)->__v = __builtin_bswap64(__D);
382}
383#endif
384#endif /* __MOVBE */
385
386#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
387 defined(__RTM__)
388#include <rtmintrin.h>
389#include <xtestintrin.h>
390#endif
391
392#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
393 defined(__SHA__)
394#include <shaintrin.h>
395#endif
396
397#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
398 defined(__FXSR__)
399#include <fxsrintrin.h>
400#endif
401
402/* No feature check desired due to internal MSC_VER checks */
403#include <xsaveintrin.h>
404
405#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
406 defined(__XSAVEOPT__)
407#include <xsaveoptintrin.h>
408#endif
409
410#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
411 defined(__XSAVEC__)
412#include <xsavecintrin.h>
413#endif
414
415#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
416 defined(__XSAVES__)
417#include <xsavesintrin.h>
418#endif
419
420#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
421 defined(__SHSTK__)
422#include <cetintrin.h>
423#endif
424
425/* Some intrinsics inside adxintrin.h are available only on processors with ADX,
426 * whereas others are also available at all times. */
427#include <adxintrin.h>
428
429#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
430 defined(__RDSEED__)
431#include <rdseedintrin.h>
432#endif
433
434#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
435 defined(__WBNOINVD__)
436#include <wbnoinvdintrin.h>
437#endif
438
439#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
440 defined(__CLDEMOTE__)
441#include <cldemoteintrin.h>
442#endif
443
444#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
445 defined(__WAITPKG__)
446#include <waitpkgintrin.h>
447#endif
448
449#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
450 defined(__MOVDIRI__) || defined(__MOVDIR64B__)
451#include <movdirintrin.h>
452#endif
453
454#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
455 defined(__PCONFIG__)
456#include <pconfigintrin.h>
457#endif
458
459#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
460 defined(__SGX__)
461#include <sgxintrin.h>
462#endif
463
464#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
465 defined(__PTWRITE__)
466#include <ptwriteintrin.h>
467#endif
468
469#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
470 defined(__INVPCID__)
471#include <invpcidintrin.h>
472#endif
473
474#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
475 defined(__AMXTILE__) || defined(__AMXINT8__) || defined(__AMXBF16__)
476#include <amxintrin.h>
477#endif
478
479#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
480 defined(__AVX512VP2INTERSECT__)
481#include <avx512vp2intersectintrin.h>
482#endif
483
484#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
485 (defined(__AVX512VL__) && defined(__AVX512VP2INTERSECT__))
486#include <avx512vlvp2intersectintrin.h>
487#endif
488
489#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
490 defined(__ENQCMD__)
491#include <enqcmdintrin.h>
492#endif
493
494#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
495 defined(__SERIALIZE__)
496#include <serializeintrin.h>
497#endif
498
499#if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \
500 defined(__TSXLDTRK__)
501#include <tsxldtrkintrin.h>
502#endif
503
504#if defined(_MSC_VER) && __has_extension(gnu_asm)
505/* Define the default attributes for these intrinsics */
506#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__))
507#ifdef __cplusplus
508extern "C" {
509#endif
510/*----------------------------------------------------------------------------*\
511|* Interlocked Exchange HLE
512\*----------------------------------------------------------------------------*/
513#if defined(__i386__) || defined(__x86_64__)
514static __inline__ long __DEFAULT_FN_ATTRS
515_InterlockedExchange_HLEAcquire(long volatile *_Target, long _Value) {
516 __asm__ __volatile__(".byte 0xf2 ; lock ; xchg %0, %1"
517 : "+r" (_Value), "+m" (*_Target) :: "memory");
518 return _Value;
519}
520static __inline__ long __DEFAULT_FN_ATTRS
521_InterlockedExchange_HLERelease(long volatile *_Target, long _Value) {
522 __asm__ __volatile__(".byte 0xf3 ; lock ; xchg %0, %1"
523 : "+r" (_Value), "+m" (*_Target) :: "memory");
524 return _Value;
525}
526#endif
527#if defined(__x86_64__)
528static __inline__ __int64 __DEFAULT_FN_ATTRS
529_InterlockedExchange64_HLEAcquire(__int64 volatile *_Target, __int64 _Value) {
530 __asm__ __volatile__(".byte 0xf2 ; lock ; xchg %0, %1"
531 : "+r" (_Value), "+m" (*_Target) :: "memory");
532 return _Value;
533}
534static __inline__ __int64 __DEFAULT_FN_ATTRS
535_InterlockedExchange64_HLERelease(__int64 volatile *_Target, __int64 _Value) {
536 __asm__ __volatile__(".byte 0xf3 ; lock ; xchg %0, %1"
537 : "+r" (_Value), "+m" (*_Target) :: "memory");
538 return _Value;
539}
540#endif
541/*----------------------------------------------------------------------------*\
542|* Interlocked Compare Exchange HLE
543\*----------------------------------------------------------------------------*/
544#if defined(__i386__) || defined(__x86_64__)
545static __inline__ long __DEFAULT_FN_ATTRS
546_InterlockedCompareExchange_HLEAcquire(long volatile *_Destination,
547 long _Exchange, long _Comparand) {
548 __asm__ __volatile__(".byte 0xf2 ; lock ; cmpxchg %2, %1"
549 : "+a" (_Comparand), "+m" (*_Destination)
550 : "r" (_Exchange) : "memory");
551 return _Comparand;
552}
553static __inline__ long __DEFAULT_FN_ATTRS
554_InterlockedCompareExchange_HLERelease(long volatile *_Destination,
555 long _Exchange, long _Comparand) {
556 __asm__ __volatile__(".byte 0xf3 ; lock ; cmpxchg %2, %1"
557 : "+a" (_Comparand), "+m" (*_Destination)
558 : "r" (_Exchange) : "memory");
559 return _Comparand;
560}
561#endif
562#if defined(__x86_64__)
563static __inline__ __int64 __DEFAULT_FN_ATTRS
564_InterlockedCompareExchange64_HLEAcquire(__int64 volatile *_Destination,
565 __int64 _Exchange, __int64 _Comparand) {
566 __asm__ __volatile__(".byte 0xf2 ; lock ; cmpxchg %2, %1"
567 : "+a" (_Comparand), "+m" (*_Destination)
568 : "r" (_Exchange) : "memory");
569 return _Comparand;
570}
571static __inline__ __int64 __DEFAULT_FN_ATTRS
572_InterlockedCompareExchange64_HLERelease(__int64 volatile *_Destination,
573 __int64 _Exchange, __int64 _Comparand) {
574 __asm__ __volatile__(".byte 0xf3 ; lock ; cmpxchg %2, %1"
575 : "+a" (_Comparand), "+m" (*_Destination)
576 : "r" (_Exchange) : "memory");
577 return _Comparand;
578}
579#endif
580#ifdef __cplusplus
581}
582#endif
583
584#undef __DEFAULT_FN_ATTRS
585
586#endif /* defined(_MSC_VER) && __has_extension(gnu_asm) */
587
588#endif /* __IMMINTRIN_H */
__attribute__
_Float16 __2f16 __attribute__((ext_vector_type(2)))
Zeroes the upper 128 bits (bits 255:128) of all YMM registers.
Definition __clang_hip_libdevice_declares.h:294
__DEFAULT_FN_ATTRS
#define __DEFAULT_FN_ATTRS
Definition __wmmintrin_aes.h:18
__p
static __inline unsigned char unsigned int unsigned int unsigned int * __p
Definition adxintrin.h:24
__D
static __inline__ void short __D
Definition immintrin.h:344
__v
struct __storeu_i16 *__P __v
Definition immintrin.h:348
emmintrin.h
mmintrin.h
pmmintrin.h
Generated on Mon Feb 17 2025 13:12:28 for ETISS 0.8.0 by doxygen 1.9.8