1 /* ===-------- intrin.h ---------------------------------------------------===
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 /* Only include this if we're compiling for the windows platform. */
11 #ifndef _MSC_VER
12 #include_next <intrin.h>
13 #else
14 
15 #ifndef __INTRIN_H
16 #define __INTRIN_H
17 
18 /* First include the standard intrinsics. */
19 #if defined(__i386__) || defined(__x86_64__)
20 #include <x86intrin.h>
21 #endif
22 
23 #if defined(__arm__)
24 #include <armintr.h>
25 #endif
26 
27 #if defined(__aarch64__)
28 #include <arm64intr.h>
29 #endif
30 
31 /* For the definition of jmp_buf. */
32 #if __STDC_HOSTED__
33 #include <setjmp.h>
34 #endif
35 
36 /* Define the default attributes for the functions in this file. */
37 #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__))
38 
39 #if __x86_64__
40 #define __LPTRINT_TYPE__ __int64
41 #else
42 #define __LPTRINT_TYPE__ long
43 #endif
44 
45 #ifdef __cplusplus
46 extern "C" {
47 #endif
48 
49 #if defined(__MMX__)
50 /* And the random ones that aren't in those files. */
51 __m64 _m_from_float(float);
52 float _m_to_float(__m64);
53 #endif
54 
55 /* Other assorted instruction intrinsics. */
56 void __addfsbyte(unsigned long, unsigned char);
57 void __addfsdword(unsigned long, unsigned long);
58 void __addfsword(unsigned long, unsigned short);
59 void __code_seg(const char *);
60 static __inline__
61 void __cpuid(int[4], int);
62 static __inline__
63 void __cpuidex(int[4], int, int);
64 static __inline__
65 __int64 __emul(int, int);
66 static __inline__
67 unsigned __int64 __emulu(unsigned int, unsigned int);
68 unsigned int __getcallerseflags(void);
69 static __inline__
70 void __halt(void);
71 unsigned char __inbyte(unsigned short);
72 void __inbytestring(unsigned short, unsigned char *, unsigned long);
73 void __incfsbyte(unsigned long);
74 void __incfsdword(unsigned long);
75 void __incfsword(unsigned long);
76 unsigned long __indword(unsigned short);
77 void __indwordstring(unsigned short, unsigned long *, unsigned long);
78 void __int2c(void);
79 void __invlpg(void *);
80 unsigned short __inword(unsigned short);
81 void __inwordstring(unsigned short, unsigned short *, unsigned long);
82 void __lidt(void *);
83 unsigned __int64 __ll_lshift(unsigned __int64, int);
84 __int64 __ll_rshift(__int64, int);
85 static __inline__
86 void __movsb(unsigned char *, unsigned char const *, size_t);
87 static __inline__
88 void __movsd(unsigned long *, unsigned long const *, size_t);
89 static __inline__
90 void __movsw(unsigned short *, unsigned short const *, size_t);
91 static __inline__
92 void __nop(void);
93 void __nvreg_restore_fence(void);
94 void __nvreg_save_fence(void);
95 void __outbyte(unsigned short, unsigned char);
96 void __outbytestring(unsigned short, unsigned char *, unsigned long);
97 void __outdword(unsigned short, unsigned long);
98 void __outdwordstring(unsigned short, unsigned long *, unsigned long);
99 void __outword(unsigned short, unsigned short);
100 void __outwordstring(unsigned short, unsigned short *, unsigned long);
101 unsigned long __readcr0(void);
102 unsigned long __readcr2(void);
103 unsigned __LPTRINT_TYPE__ __readcr3(void);
104 unsigned long __readcr4(void);
105 unsigned long __readcr8(void);
106 unsigned int __readdr(unsigned int);
107 #ifdef __i386__
108 static __inline__
109 unsigned char __readfsbyte(unsigned long);
110 static __inline__
111 unsigned __int64 __readfsqword(unsigned long);
112 static __inline__
113 unsigned short __readfsword(unsigned long);
114 #endif
115 static __inline__
116 unsigned __int64 __readmsr(unsigned long);
117 unsigned __int64 __readpmc(unsigned long);
118 unsigned long __segmentlimit(unsigned long);
119 void __sidt(void *);
120 static __inline__
121 void __stosb(unsigned char *, unsigned char, size_t);
122 static __inline__
123 void __stosd(unsigned long *, unsigned long, size_t);
124 static __inline__
125 void __stosw(unsigned short *, unsigned short, size_t);
126 void __svm_clgi(void);
127 void __svm_invlpga(void *, int);
128 void __svm_skinit(int);
129 void __svm_stgi(void);
130 void __svm_vmload(size_t);
131 void __svm_vmrun(size_t);
132 void __svm_vmsave(size_t);
133 void __ud2(void);
134 unsigned __int64 __ull_rshift(unsigned __int64, int);
135 void __vmx_off(void);
136 void __vmx_vmptrst(unsigned __int64 *);
137 void __wbinvd(void);
138 void __writecr0(unsigned int);
139 static __inline__
140 void __writecr3(unsigned __INTPTR_TYPE__);
141 void __writecr4(unsigned int);
142 void __writecr8(unsigned int);
143 void __writedr(unsigned int, unsigned int);
144 void __writefsbyte(unsigned long, unsigned char);
145 void __writefsdword(unsigned long, unsigned long);
146 void __writefsqword(unsigned long, unsigned __int64);
147 void __writefsword(unsigned long, unsigned short);
148 void __writemsr(unsigned long, unsigned __int64);
149 static __inline__
150 void *_AddressOfReturnAddress(void);
151 static __inline__
152 unsigned char _BitScanForward(unsigned long *_Index, unsigned long _Mask);
153 static __inline__
154 unsigned char _BitScanReverse(unsigned long *_Index, unsigned long _Mask);
155 unsigned char _bittest(long const *, long);
156 unsigned char _bittestandcomplement(long *, long);
157 unsigned char _bittestandreset(long *, long);
158 unsigned char _bittestandset(long *, long);
159 void __cdecl _disable(void);
160 void __cdecl _enable(void);
161 long _InterlockedAddLargeStatistic(__int64 volatile *_Addend, long _Value);
162 unsigned char _interlockedbittestandreset(long volatile *, long);
163 unsigned char _interlockedbittestandset(long volatile *, long);
164 void *_InterlockedCompareExchangePointer_HLEAcquire(void *volatile *, void *,
165                                                     void *);
166 void *_InterlockedCompareExchangePointer_HLERelease(void *volatile *, void *,
167                                                     void *);
168 long _InterlockedExchangeAdd_HLEAcquire(long volatile *, long);
169 long _InterlockedExchangeAdd_HLERelease(long volatile *, long);
170 __int64 _InterlockedExchangeAdd64_HLEAcquire(__int64 volatile *, __int64);
171 __int64 _InterlockedExchangeAdd64_HLERelease(__int64 volatile *, __int64);
172 static __inline__ void
173 __attribute__((__deprecated__("use other intrinsics or C++11 atomics instead")))
174 _ReadBarrier(void);
175 static __inline__ void
176 __attribute__((__deprecated__("use other intrinsics or C++11 atomics instead")))
177 _ReadWriteBarrier(void);
178 unsigned int _rorx_u32(unsigned int, const unsigned int);
179 int _sarx_i32(int, unsigned int);
180 #if __STDC_HOSTED__
181 int __cdecl _setjmp(jmp_buf);
182 #endif
183 unsigned int _shlx_u32(unsigned int, unsigned int);
184 unsigned int _shrx_u32(unsigned int, unsigned int);
185 void _Store_HLERelease(long volatile *, long);
186 void _Store64_HLERelease(__int64 volatile *, __int64);
187 void _StorePointer_HLERelease(void *volatile *, void *);
188 static __inline__ void
189 __attribute__((__deprecated__("use other intrinsics or C++11 atomics instead")))
190 _WriteBarrier(void);
191 unsigned __int32 xbegin(void);
192 void _xend(void);
193 
194 /* These additional intrinsics are turned on in x64/amd64/x86_64 mode. */
195 #ifdef __x86_64__
196 void __addgsbyte(unsigned long, unsigned char);
197 void __addgsdword(unsigned long, unsigned long);
198 void __addgsqword(unsigned long, unsigned __int64);
199 void __addgsword(unsigned long, unsigned short);
200 static __inline__
201 void __faststorefence(void);
202 void __incgsbyte(unsigned long);
203 void __incgsdword(unsigned long);
204 void __incgsqword(unsigned long);
205 void __incgsword(unsigned long);
206 static __inline__
207 void __movsq(unsigned long long *, unsigned long long const *, size_t);
208 static __inline__
209 unsigned char __readgsbyte(unsigned long);
210 static __inline__
211 unsigned long __readgsdword(unsigned long);
212 static __inline__
213 unsigned __int64 __readgsqword(unsigned long);
214 unsigned short __readgsword(unsigned long);
215 unsigned __int64 __shiftleft128(unsigned __int64 _LowPart,
216                                 unsigned __int64 _HighPart,
217                                 unsigned char _Shift);
218 unsigned __int64 __shiftright128(unsigned __int64 _LowPart,
219                                  unsigned __int64 _HighPart,
220                                  unsigned char _Shift);
221 static __inline__
222 void __stosq(unsigned __int64 *, unsigned __int64, size_t);
223 unsigned char __vmx_on(unsigned __int64 *);
224 unsigned char __vmx_vmclear(unsigned __int64 *);
225 unsigned char __vmx_vmlaunch(void);
226 unsigned char __vmx_vmptrld(unsigned __int64 *);
227 unsigned char __vmx_vmread(size_t, size_t *);
228 unsigned char __vmx_vmresume(void);
229 unsigned char __vmx_vmwrite(size_t, size_t);
230 void __writegsbyte(unsigned long, unsigned char);
231 void __writegsdword(unsigned long, unsigned long);
232 void __writegsqword(unsigned long, unsigned __int64);
233 void __writegsword(unsigned long, unsigned short);
234 unsigned char _bittest64(__int64 const *, __int64);
235 unsigned char _bittestandcomplement64(__int64 *, __int64);
236 unsigned char _bittestandreset64(__int64 *, __int64);
237 unsigned char _bittestandset64(__int64 *, __int64);
238 long _InterlockedAnd_np(long volatile *_Value, long _Mask);
239 short _InterlockedAnd16_np(short volatile *_Value, short _Mask);
240 __int64 _InterlockedAnd64_np(__int64 volatile *_Value, __int64 _Mask);
241 char _InterlockedAnd8_np(char volatile *_Value, char _Mask);
242 unsigned char _interlockedbittestandreset64(__int64 volatile *, __int64);
243 unsigned char _interlockedbittestandset64(__int64 volatile *, __int64);
244 long _InterlockedCompareExchange_np(long volatile *_Destination, long _Exchange,
245                                     long _Comparand);
246 unsigned char _InterlockedCompareExchange128(__int64 volatile *_Destination,
247                                              __int64 _ExchangeHigh,
248                                              __int64 _ExchangeLow,
249                                              __int64 *_CompareandResult);
250 unsigned char _InterlockedCompareExchange128_np(__int64 volatile *_Destination,
251                                                 __int64 _ExchangeHigh,
252                                                 __int64 _ExchangeLow,
253                                                 __int64 *_ComparandResult);
254 short _InterlockedCompareExchange16_np(short volatile *_Destination,
255                                        short _Exchange, short _Comparand);
256 __int64 _InterlockedCompareExchange64_np(__int64 volatile *_Destination,
257                                          __int64 _Exchange, __int64 _Comparand);
258 void *_InterlockedCompareExchangePointer_np(void *volatile *_Destination,
259                                             void *_Exchange, void *_Comparand);
260 long _InterlockedOr_np(long volatile *_Value, long _Mask);
261 short _InterlockedOr16_np(short volatile *_Value, short _Mask);
262 __int64 _InterlockedOr64_np(__int64 volatile *_Value, __int64 _Mask);
263 char _InterlockedOr8_np(char volatile *_Value, char _Mask);
264 long _InterlockedXor_np(long volatile *_Value, long _Mask);
265 short _InterlockedXor16_np(short volatile *_Value, short _Mask);
266 __int64 _InterlockedXor64_np(__int64 volatile *_Value, __int64 _Mask);
267 char _InterlockedXor8_np(char volatile *_Value, char _Mask);
268 unsigned __int64 _rorx_u64(unsigned __int64, const unsigned int);
269 __int64 _sarx_i64(__int64, unsigned int);
270 unsigned __int64 _shlx_u64(unsigned __int64, unsigned int);
271 unsigned __int64 _shrx_u64(unsigned __int64, unsigned int);
272 static __inline__
273 __int64 __mulh(__int64, __int64);
274 static __inline__
275 unsigned __int64 __umulh(unsigned __int64, unsigned __int64);
276 static __inline__
277 __int64 _mul128(__int64, __int64, __int64*);
278 static __inline__
279 unsigned __int64 _umul128(unsigned __int64,
280                           unsigned __int64,
281                           unsigned __int64*);
282 
283 #endif /* __x86_64__ */
284 
285 #if defined(__x86_64__) || defined(__arm__) || defined(__aarch64__)
286 
287 static __inline__
288 unsigned char _BitScanForward64(unsigned long *_Index, unsigned __int64 _Mask);
289 static __inline__
290 unsigned char _BitScanReverse64(unsigned long *_Index, unsigned __int64 _Mask);
291 
292 static __inline__
293 __int64 _InterlockedDecrement64(__int64 volatile *_Addend);
294 static __inline__
295 __int64 _InterlockedExchange64(__int64 volatile *_Target, __int64 _Value);
296 static __inline__
297 __int64 _InterlockedExchangeAdd64(__int64 volatile *_Addend, __int64 _Value);
298 static __inline__
299 __int64 _InterlockedExchangeSub64(__int64 volatile *_Subend, __int64 _Value);
300 static __inline__
301 __int64 _InterlockedIncrement64(__int64 volatile *_Addend);
302 static __inline__
303 __int64 _InterlockedOr64(__int64 volatile *_Value, __int64 _Mask);
304 static __inline__
305 __int64 _InterlockedXor64(__int64 volatile *_Value, __int64 _Mask);
306 static __inline__
307 __int64 _InterlockedAnd64(__int64 volatile *_Value, __int64 _Mask);
308 
309 #endif
310 
311 /*----------------------------------------------------------------------------*\
312 |* Interlocked Exchange Add
313 \*----------------------------------------------------------------------------*/
314 #if defined(__arm__) || defined(__aarch64__)
315 char _InterlockedExchangeAdd8_acq(char volatile *_Addend, char _Value);
316 char _InterlockedExchangeAdd8_nf(char volatile *_Addend, char _Value);
317 char _InterlockedExchangeAdd8_rel(char volatile *_Addend, char _Value);
318 short _InterlockedExchangeAdd16_acq(short volatile *_Addend, short _Value);
319 short _InterlockedExchangeAdd16_nf(short volatile *_Addend, short _Value);
320 short _InterlockedExchangeAdd16_rel(short volatile *_Addend, short _Value);
321 long _InterlockedExchangeAdd_acq(long volatile *_Addend, long _Value);
322 long _InterlockedExchangeAdd_nf(long volatile *_Addend, long _Value);
323 long _InterlockedExchangeAdd_rel(long volatile *_Addend, long _Value);
324 __int64 _InterlockedExchangeAdd64_acq(__int64 volatile *_Addend, __int64 _Value);
325 __int64 _InterlockedExchangeAdd64_nf(__int64 volatile *_Addend, __int64 _Value);
326 __int64 _InterlockedExchangeAdd64_rel(__int64 volatile *_Addend, __int64 _Value);
327 #endif
328 /*----------------------------------------------------------------------------*\
329 |* Interlocked Increment
330 \*----------------------------------------------------------------------------*/
331 #if defined(__arm__) || defined(__aarch64__)
332 short _InterlockedIncrement16_acq(short volatile *_Value);
333 short _InterlockedIncrement16_nf(short volatile *_Value);
334 short _InterlockedIncrement16_rel(short volatile *_Value);
335 long _InterlockedIncrement_acq(long volatile *_Value);
336 long _InterlockedIncrement_nf(long volatile *_Value);
337 long _InterlockedIncrement_rel(long volatile *_Value);
338 __int64 _InterlockedIncrement64_acq(__int64 volatile *_Value);
339 __int64 _InterlockedIncrement64_nf(__int64 volatile *_Value);
340 __int64 _InterlockedIncrement64_rel(__int64 volatile *_Value);
341 #endif
342 /*----------------------------------------------------------------------------*\
343 |* Interlocked Decrement
344 \*----------------------------------------------------------------------------*/
345 #if defined(__arm__) || defined(__aarch64__)
346 short _InterlockedDecrement16_acq(short volatile *_Value);
347 short _InterlockedDecrement16_nf(short volatile *_Value);
348 short _InterlockedDecrement16_rel(short volatile *_Value);
349 long _InterlockedDecrement_acq(long volatile *_Value);
350 long _InterlockedDecrement_nf(long volatile *_Value);
351 long _InterlockedDecrement_rel(long volatile *_Value);
352 __int64 _InterlockedDecrement64_acq(__int64 volatile *_Value);
353 __int64 _InterlockedDecrement64_nf(__int64 volatile *_Value);
354 __int64 _InterlockedDecrement64_rel(__int64 volatile *_Value);
355 #endif
356 /*----------------------------------------------------------------------------*\
357 |* Interlocked And
358 \*----------------------------------------------------------------------------*/
359 #if defined(__arm__) || defined(__aarch64__)
360 char _InterlockedAnd8_acq(char volatile *_Value, char _Mask);
361 char _InterlockedAnd8_nf(char volatile *_Value, char _Mask);
362 char _InterlockedAnd8_rel(char volatile *_Value, char _Mask);
363 short _InterlockedAnd16_acq(short volatile *_Value, short _Mask);
364 short _InterlockedAnd16_nf(short volatile *_Value, short _Mask);
365 short _InterlockedAnd16_rel(short volatile *_Value, short _Mask);
366 long _InterlockedAnd_acq(long volatile *_Value, long _Mask);
367 long _InterlockedAnd_nf(long volatile *_Value, long _Mask);
368 long _InterlockedAnd_rel(long volatile *_Value, long _Mask);
369 __int64 _InterlockedAnd64_acq(__int64 volatile *_Value, __int64 _Mask);
370 __int64 _InterlockedAnd64_nf(__int64 volatile *_Value, __int64 _Mask);
371 __int64 _InterlockedAnd64_rel(__int64 volatile *_Value, __int64 _Mask);
372 #endif
373 /*----------------------------------------------------------------------------*\
374 |* Bit Counting and Testing
375 \*----------------------------------------------------------------------------*/
376 #if defined(__arm__) || defined(__aarch64__)
377 unsigned char _interlockedbittestandset_acq(long volatile *_BitBase,
378                                             long _BitPos);
379 unsigned char _interlockedbittestandset_nf(long volatile *_BitBase,
380                                            long _BitPos);
381 unsigned char _interlockedbittestandset_rel(long volatile *_BitBase,
382                                             long _BitPos);
383 unsigned char _interlockedbittestandreset_acq(long volatile *_BitBase,
384                                               long _BitPos);
385 unsigned char _interlockedbittestandreset_nf(long volatile *_BitBase,
386                                              long _BitPos);
387 unsigned char _interlockedbittestandreset_rel(long volatile *_BitBase,
388                                               long _BitPos);
389 #endif
390 /*----------------------------------------------------------------------------*\
391 |* Interlocked Or
392 \*----------------------------------------------------------------------------*/
393 #if defined(__arm__) || defined(__aarch64__)
394 char _InterlockedOr8_acq(char volatile *_Value, char _Mask);
395 char _InterlockedOr8_nf(char volatile *_Value, char _Mask);
396 char _InterlockedOr8_rel(char volatile *_Value, char _Mask);
397 short _InterlockedOr16_acq(short volatile *_Value, short _Mask);
398 short _InterlockedOr16_nf(short volatile *_Value, short _Mask);
399 short _InterlockedOr16_rel(short volatile *_Value, short _Mask);
400 long _InterlockedOr_acq(long volatile *_Value, long _Mask);
401 long _InterlockedOr_nf(long volatile *_Value, long _Mask);
402 long _InterlockedOr_rel(long volatile *_Value, long _Mask);
403 __int64 _InterlockedOr64_acq(__int64 volatile *_Value, __int64 _Mask);
404 __int64 _InterlockedOr64_nf(__int64 volatile *_Value, __int64 _Mask);
405 __int64 _InterlockedOr64_rel(__int64 volatile *_Value, __int64 _Mask);
406 #endif
407 /*----------------------------------------------------------------------------*\
408 |* Interlocked Xor
409 \*----------------------------------------------------------------------------*/
410 #if defined(__arm__) || defined(__aarch64__)
411 char _InterlockedXor8_acq(char volatile *_Value, char _Mask);
412 char _InterlockedXor8_nf(char volatile *_Value, char _Mask);
413 char _InterlockedXor8_rel(char volatile *_Value, char _Mask);
414 short _InterlockedXor16_acq(short volatile *_Value, short _Mask);
415 short _InterlockedXor16_nf(short volatile *_Value, short _Mask);
416 short _InterlockedXor16_rel(short volatile *_Value, short _Mask);
417 long _InterlockedXor_acq(long volatile *_Value, long _Mask);
418 long _InterlockedXor_nf(long volatile *_Value, long _Mask);
419 long _InterlockedXor_rel(long volatile *_Value, long _Mask);
420 __int64 _InterlockedXor64_acq(__int64 volatile *_Value, __int64 _Mask);
421 __int64 _InterlockedXor64_nf(__int64 volatile *_Value, __int64 _Mask);
422 __int64 _InterlockedXor64_rel(__int64 volatile *_Value, __int64 _Mask);
423 #endif
424 /*----------------------------------------------------------------------------*\
425 |* Interlocked Exchange
426 \*----------------------------------------------------------------------------*/
427 #if defined(__arm__) || defined(__aarch64__)
428 char _InterlockedExchange8_acq(char volatile *_Target, char _Value);
429 char _InterlockedExchange8_nf(char volatile *_Target, char _Value);
430 char _InterlockedExchange8_rel(char volatile *_Target, char _Value);
431 short _InterlockedExchange16_acq(short volatile *_Target, short _Value);
432 short _InterlockedExchange16_nf(short volatile *_Target, short _Value);
433 short _InterlockedExchange16_rel(short volatile *_Target, short _Value);
434 long _InterlockedExchange_acq(long volatile *_Target, long _Value);
435 long _InterlockedExchange_nf(long volatile *_Target, long _Value);
436 long _InterlockedExchange_rel(long volatile *_Target, long _Value);
437 __int64 _InterlockedExchange64_acq(__int64 volatile *_Target, __int64 _Value);
438 __int64 _InterlockedExchange64_nf(__int64 volatile *_Target, __int64 _Value);
439 __int64 _InterlockedExchange64_rel(__int64 volatile *_Target, __int64 _Value);
440 #endif
441 /*----------------------------------------------------------------------------*\
442 |* Interlocked Compare Exchange
443 \*----------------------------------------------------------------------------*/
444 #if defined(__arm__) || defined(__aarch64__)
445 char _InterlockedCompareExchange8_acq(char volatile *_Destination,
446                              char _Exchange, char _Comparand);
447 char _InterlockedCompareExchange8_nf(char volatile *_Destination,
448                              char _Exchange, char _Comparand);
449 char _InterlockedCompareExchange8_rel(char volatile *_Destination,
450                              char _Exchange, char _Comparand);
451 short _InterlockedCompareExchange16_acq(short volatile *_Destination,
452                               short _Exchange, short _Comparand);
453 short _InterlockedCompareExchange16_nf(short volatile *_Destination,
454                               short _Exchange, short _Comparand);
455 short _InterlockedCompareExchange16_rel(short volatile *_Destination,
456                               short _Exchange, short _Comparand);
457 long _InterlockedCompareExchange_acq(long volatile *_Destination,
458                               long _Exchange, long _Comparand);
459 long _InterlockedCompareExchange_nf(long volatile *_Destination,
460                               long _Exchange, long _Comparand);
461 long _InterlockedCompareExchange_rel(long volatile *_Destination,
462                               long _Exchange, long _Comparand);
463 __int64 _InterlockedCompareExchange64_acq(__int64 volatile *_Destination,
464                               __int64 _Exchange, __int64 _Comparand);
465 __int64 _InterlockedCompareExchange64_nf(__int64 volatile *_Destination,
466                               __int64 _Exchange, __int64 _Comparand);
467 __int64 _InterlockedCompareExchange64_rel(__int64 volatile *_Destination,
468                               __int64 _Exchange, __int64 _Comparand);
469 #endif
470 
471 /*----------------------------------------------------------------------------*\
472 |* movs, stos
473 \*----------------------------------------------------------------------------*/
474 #if defined(__i386__) || defined(__x86_64__)
475 static __inline__ void __DEFAULT_FN_ATTRS
476 __movsb(unsigned char *__dst, unsigned char const *__src, size_t __n) {
477   __asm__ __volatile__("rep movsb" : "+D"(__dst), "+S"(__src), "+c"(__n)
478                        : : "memory");
479 }
480 static __inline__ void __DEFAULT_FN_ATTRS
481 __movsd(unsigned long *__dst, unsigned long const *__src, size_t __n) {
482   __asm__ __volatile__("rep movsl" : "+D"(__dst), "+S"(__src), "+c"(__n)
483                        : : "memory");
484 }
485 static __inline__ void __DEFAULT_FN_ATTRS
486 __movsw(unsigned short *__dst, unsigned short const *__src, size_t __n) {
487   __asm__ __volatile__("rep movsw" : "+D"(__dst), "+S"(__src), "+c"(__n)
488                        : : "memory");
489 }
490 static __inline__ void __DEFAULT_FN_ATTRS
491 __stosd(unsigned long *__dst, unsigned long __x, size_t __n) {
492   __asm__ __volatile__("rep stosl" : "+D"(__dst), "+c"(__n) : "a"(__x)
493                        : "memory");
494 }
495 static __inline__ void __DEFAULT_FN_ATTRS
496 __stosw(unsigned short *__dst, unsigned short __x, size_t __n) {
497   __asm__ __volatile__("rep stosw" : "+D"(__dst), "+c"(__n) : "a"(__x)
498                        : "memory");
499 }
500 #endif
501 #ifdef __x86_64__
502 static __inline__ void __DEFAULT_FN_ATTRS
503 __movsq(unsigned long long *__dst, unsigned long long const *__src, size_t __n) {
504   __asm__ __volatile__("rep movsq" : "+D"(__dst), "+S"(__src), "+c"(__n)
505                        : : "memory");
506 }
507 static __inline__ void __DEFAULT_FN_ATTRS
508 __stosq(unsigned __int64 *__dst, unsigned __int64 __x, size_t __n) {
509   __asm__ __volatile__("rep stosq" : "+D"(__dst), "+c"(__n) : "a"(__x)
510                        : "memory");
511 }
512 #endif
513 
514 /*----------------------------------------------------------------------------*\
515 |* Misc
516 \*----------------------------------------------------------------------------*/
517 #if defined(__i386__) || defined(__x86_64__)
518 static __inline__ void __DEFAULT_FN_ATTRS
519 __cpuid(int __info[4], int __level) {
520   __asm__ ("cpuid" : "=a"(__info[0]), "=b" (__info[1]), "=c"(__info[2]), "=d"(__info[3])
521                    : "a"(__level), "c"(0));
522 }
523 static __inline__ void __DEFAULT_FN_ATTRS
524 __cpuidex(int __info[4], int __level, int __ecx) {
525   __asm__ ("cpuid" : "=a"(__info[0]), "=b" (__info[1]), "=c"(__info[2]), "=d"(__info[3])
526                    : "a"(__level), "c"(__ecx));
527 }
528 static __inline__ void __DEFAULT_FN_ATTRS
529 __halt(void) {
530   __asm__ volatile ("hlt");
531 }
532 #endif
533 
534 #if defined(__i386__) || defined(__x86_64__) || defined(__aarch64__)
535 static __inline__ void __DEFAULT_FN_ATTRS
536 __nop(void) {
537   __asm__ volatile ("nop");
538 }
539 #endif
540 
541 /*----------------------------------------------------------------------------*\
542 |* MS AArch64 specific
543 \*----------------------------------------------------------------------------*/
544 #if defined(__aarch64__)
545 unsigned __int64 __getReg(int);
546 long _InterlockedAdd(long volatile *Addend, long Value);
547 __int64 _ReadStatusReg(int);
548 void _WriteStatusReg(int, __int64);
549 
550 unsigned short __cdecl _byteswap_ushort(unsigned short val);
551 unsigned long __cdecl _byteswap_ulong (unsigned long val);
552 unsigned __int64 __cdecl _byteswap_uint64(unsigned __int64 val);
553 #endif
554 
555 /*----------------------------------------------------------------------------*\
556 |* Privileged intrinsics
557 \*----------------------------------------------------------------------------*/
558 #if defined(__i386__) || defined(__x86_64__)
559 static __inline__ unsigned __int64 __DEFAULT_FN_ATTRS
560 __readmsr(unsigned long __register) {
561   // Loads the contents of a 64-bit model specific register (MSR) specified in
562   // the ECX register into registers EDX:EAX. The EDX register is loaded with
563   // the high-order 32 bits of the MSR and the EAX register is loaded with the
564   // low-order 32 bits. If less than 64 bits are implemented in the MSR being
565   // read, the values returned to EDX:EAX in unimplemented bit locations are
566   // undefined.
567   unsigned long __edx;
568   unsigned long __eax;
569   __asm__ ("rdmsr" : "=d"(__edx), "=a"(__eax) : "c"(__register));
570   return (((unsigned __int64)__edx) << 32) | (unsigned __int64)__eax;
571 }
572 #endif
573 
574 static __inline__ unsigned __LPTRINT_TYPE__ __DEFAULT_FN_ATTRS
575 __readcr3(void) {
576   unsigned __LPTRINT_TYPE__ __cr3_val;
577   __asm__ __volatile__ ("mov %%cr3, %0" : "=r"(__cr3_val) : : "memory");
578   return __cr3_val;
579 }
580 
581 static __inline__ void __DEFAULT_FN_ATTRS
582 __writecr3(unsigned __INTPTR_TYPE__ __cr3_val) {
583   __asm__ ("mov %0, %%cr3" : : "r"(__cr3_val) : "memory");
584 }
585 
586 #ifdef __cplusplus
587 }
588 #endif
589 
590 #undef __LPTRINT_TYPE__
591 
592 #undef __DEFAULT_FN_ATTRS
593 
594 #endif /* __INTRIN_H */
595 #endif /* _MSC_VER */
596