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
46extern "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);
52float _m_to_float(__m64);
53#endif
54
55/* Other assorted instruction intrinsics. */
56void __addfsbyte(unsigned long, unsigned char);
57void __addfsdword(unsigned long, unsigned long);
58void __addfsword(unsigned long, unsigned short);
59void __code_seg(const char *);
60static __inline__
61void __cpuid(int[4], int);
62static __inline__
63void __cpuidex(int[4], int, int);
64static __inline__
65__int64 __emul(int, int);
66static __inline__
67unsigned __int64 __emulu(unsigned int, unsigned int);
68unsigned int __getcallerseflags(void);
69static __inline__
70void __halt(void);
71unsigned char __inbyte(unsigned short);
72void __inbytestring(unsigned short, unsigned char *, unsigned long);
73void __incfsbyte(unsigned long);
74void __incfsdword(unsigned long);
75void __incfsword(unsigned long);
76unsigned long __indword(unsigned short);
77void __indwordstring(unsigned short, unsigned long *, unsigned long);
78void __int2c(void);
79void __invlpg(void *);
80unsigned short __inword(unsigned short);
81void __inwordstring(unsigned short, unsigned short *, unsigned long);
82void __lidt(void *);
83unsigned __int64 __ll_lshift(unsigned __int64, int);
84__int64 __ll_rshift(__int64, int);
85static __inline__
86void __movsb(unsigned char *, unsigned char const *, size_t);
87static __inline__
88void __movsd(unsigned long *, unsigned long const *, size_t);
89static __inline__
90void __movsw(unsigned short *, unsigned short const *, size_t);
91static __inline__
92void __nop(void);
93void __nvreg_restore_fence(void);
94void __nvreg_save_fence(void);
95void __outbyte(unsigned short, unsigned char);
96void __outbytestring(unsigned short, unsigned char *, unsigned long);
97void __outdword(unsigned short, unsigned long);
98void __outdwordstring(unsigned short, unsigned long *, unsigned long);
99void __outword(unsigned short, unsigned short);
100void __outwordstring(unsigned short, unsigned short *, unsigned long);
101unsigned long __readcr0(void);
102unsigned long __readcr2(void);
103unsigned __LPTRINT_TYPE__ __readcr3(void);
104unsigned long __readcr4(void);
105unsigned long __readcr8(void);
106unsigned int __readdr(unsigned int);
107#ifdef __i386__
108static __inline__
109unsigned char __readfsbyte(unsigned long);
110static __inline__
111unsigned __int64 __readfsqword(unsigned long);
112static __inline__
113unsigned short __readfsword(unsigned long);
114#endif
115static __inline__
116unsigned __int64 __readmsr(unsigned long);
117unsigned __int64 __readpmc(unsigned long);
118unsigned long __segmentlimit(unsigned long);
119void __sidt(void *);
120static __inline__
121void __stosb(unsigned char *, unsigned char, size_t);
122static __inline__
123void __stosd(unsigned long *, unsigned long, size_t);
124static __inline__
125void __stosw(unsigned short *, unsigned short, size_t);
126void __svm_clgi(void);
127void __svm_invlpga(void *, int);
128void __svm_skinit(int);
129void __svm_stgi(void);
130void __svm_vmload(size_t);
131void __svm_vmrun(size_t);
132void __svm_vmsave(size_t);
133void __ud2(void);
134unsigned __int64 __ull_rshift(unsigned __int64, int);
135void __vmx_off(void);
136void __vmx_vmptrst(unsigned __int64 *);
137void __wbinvd(void);
138void __writecr0(unsigned int);
139static __inline__
140void __writecr3(unsigned __INTPTR_TYPE__);
141void __writecr4(unsigned int);
142void __writecr8(unsigned int);
143void __writedr(unsigned int, unsigned int);
144void __writefsbyte(unsigned long, unsigned char);
145void __writefsdword(unsigned long, unsigned long);
146void __writefsqword(unsigned long, unsigned __int64);
147void __writefsword(unsigned long, unsigned short);
148void __writemsr(unsigned long, unsigned __int64);
149static __inline__
150void *_AddressOfReturnAddress(void);
151static __inline__
152unsigned char _BitScanForward(unsigned long *_Index, unsigned long _Mask);
153static __inline__
154unsigned char _BitScanReverse(unsigned long *_Index, unsigned long _Mask);
155unsigned char _bittest(long const *, long);
156unsigned char _bittestandcomplement(long *, long);
157unsigned char _bittestandreset(long *, long);
158unsigned char _bittestandset(long *, long);
159void __cdecl _disable(void);
160void __cdecl _enable(void);
161long _InterlockedAddLargeStatistic(__int64 volatile *_Addend, long _Value);
162unsigned char _interlockedbittestandreset(long volatile *, long);
163unsigned char _interlockedbittestandset(long volatile *, long);
164void *_InterlockedCompareExchangePointer_HLEAcquire(void *volatile *, void *,
165                                                    void *);
166void *_InterlockedCompareExchangePointer_HLERelease(void *volatile *, void *,
167                                                    void *);
168long _InterlockedExchangeAdd_HLEAcquire(long volatile *, long);
169long _InterlockedExchangeAdd_HLERelease(long volatile *, long);
170__int64 _InterlockedExchangeAdd64_HLEAcquire(__int64 volatile *, __int64);
171__int64 _InterlockedExchangeAdd64_HLERelease(__int64 volatile *, __int64);
172static __inline__ void
173__attribute__((__deprecated__("use other intrinsics or C++11 atomics instead")))
174_ReadBarrier(void);
175static __inline__ void
176__attribute__((__deprecated__("use other intrinsics or C++11 atomics instead")))
177_ReadWriteBarrier(void);
178unsigned int _rorx_u32(unsigned int, const unsigned int);
179int _sarx_i32(int, unsigned int);
180#if __STDC_HOSTED__
181int __cdecl _setjmp(jmp_buf);
182#endif
183unsigned int _shlx_u32(unsigned int, unsigned int);
184unsigned int _shrx_u32(unsigned int, unsigned int);
185void _Store_HLERelease(long volatile *, long);
186void _Store64_HLERelease(__int64 volatile *, __int64);
187void _StorePointer_HLERelease(void *volatile *, void *);
188static __inline__ void
189__attribute__((__deprecated__("use other intrinsics or C++11 atomics instead")))
190_WriteBarrier(void);
191unsigned __int32 xbegin(void);
192void _xend(void);
193
194/* These additional intrinsics are turned on in x64/amd64/x86_64 mode. */
195#ifdef __x86_64__
196void __addgsbyte(unsigned long, unsigned char);
197void __addgsdword(unsigned long, unsigned long);
198void __addgsqword(unsigned long, unsigned __int64);
199void __addgsword(unsigned long, unsigned short);
200static __inline__
201void __faststorefence(void);
202void __incgsbyte(unsigned long);
203void __incgsdword(unsigned long);
204void __incgsqword(unsigned long);
205void __incgsword(unsigned long);
206static __inline__
207void __movsq(unsigned long long *, unsigned long long const *, size_t);
208static __inline__
209unsigned char __readgsbyte(unsigned long);
210static __inline__
211unsigned long __readgsdword(unsigned long);
212static __inline__
213unsigned __int64 __readgsqword(unsigned long);
214unsigned short __readgsword(unsigned long);
215unsigned __int64 __shiftleft128(unsigned __int64 _LowPart,
216                                unsigned __int64 _HighPart,
217                                unsigned char _Shift);
218unsigned __int64 __shiftright128(unsigned __int64 _LowPart,
219                                 unsigned __int64 _HighPart,
220                                 unsigned char _Shift);
221static __inline__
222void __stosq(unsigned __int64 *, unsigned __int64, size_t);
223unsigned char __vmx_on(unsigned __int64 *);
224unsigned char __vmx_vmclear(unsigned __int64 *);
225unsigned char __vmx_vmlaunch(void);
226unsigned char __vmx_vmptrld(unsigned __int64 *);
227unsigned char __vmx_vmread(size_t, size_t *);
228unsigned char __vmx_vmresume(void);
229unsigned char __vmx_vmwrite(size_t, size_t);
230void __writegsbyte(unsigned long, unsigned char);
231void __writegsdword(unsigned long, unsigned long);
232void __writegsqword(unsigned long, unsigned __int64);
233void __writegsword(unsigned long, unsigned short);
234unsigned char _bittest64(__int64 const *, __int64);
235unsigned char _bittestandcomplement64(__int64 *, __int64);
236unsigned char _bittestandreset64(__int64 *, __int64);
237unsigned char _bittestandset64(__int64 *, __int64);
238long _InterlockedAnd_np(long volatile *_Value, long _Mask);
239short _InterlockedAnd16_np(short volatile *_Value, short _Mask);
240__int64 _InterlockedAnd64_np(__int64 volatile *_Value, __int64 _Mask);
241char _InterlockedAnd8_np(char volatile *_Value, char _Mask);
242unsigned char _interlockedbittestandreset64(__int64 volatile *, __int64);
243unsigned char _interlockedbittestandset64(__int64 volatile *, __int64);
244long _InterlockedCompareExchange_np(long volatile *_Destination, long _Exchange,
245                                    long _Comparand);
246unsigned char _InterlockedCompareExchange128(__int64 volatile *_Destination,
247                                             __int64 _ExchangeHigh,
248                                             __int64 _ExchangeLow,
249                                             __int64 *_CompareandResult);
250unsigned char _InterlockedCompareExchange128_np(__int64 volatile *_Destination,
251                                                __int64 _ExchangeHigh,
252                                                __int64 _ExchangeLow,
253                                                __int64 *_ComparandResult);
254short _InterlockedCompareExchange16_np(short volatile *_Destination,
255                                       short _Exchange, short _Comparand);
256__int64 _InterlockedCompareExchange64_np(__int64 volatile *_Destination,
257                                         __int64 _Exchange, __int64 _Comparand);
258void *_InterlockedCompareExchangePointer_np(void *volatile *_Destination,
259                                            void *_Exchange, void *_Comparand);
260long _InterlockedOr_np(long volatile *_Value, long _Mask);
261short _InterlockedOr16_np(short volatile *_Value, short _Mask);
262__int64 _InterlockedOr64_np(__int64 volatile *_Value, __int64 _Mask);
263char _InterlockedOr8_np(char volatile *_Value, char _Mask);
264long _InterlockedXor_np(long volatile *_Value, long _Mask);
265short _InterlockedXor16_np(short volatile *_Value, short _Mask);
266__int64 _InterlockedXor64_np(__int64 volatile *_Value, __int64 _Mask);
267char _InterlockedXor8_np(char volatile *_Value, char _Mask);
268unsigned __int64 _rorx_u64(unsigned __int64, const unsigned int);
269__int64 _sarx_i64(__int64, unsigned int);
270unsigned __int64 _shlx_u64(unsigned __int64, unsigned int);
271unsigned __int64 _shrx_u64(unsigned __int64, unsigned int);
272static __inline__
273__int64 __mulh(__int64, __int64);
274static __inline__
275unsigned __int64 __umulh(unsigned __int64, unsigned __int64);
276static __inline__
277__int64 _mul128(__int64, __int64, __int64*);
278static __inline__
279unsigned __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
287static __inline__
288unsigned char _BitScanForward64(unsigned long *_Index, unsigned __int64 _Mask);
289static __inline__
290unsigned char _BitScanReverse64(unsigned long *_Index, unsigned __int64 _Mask);
291
292static __inline__
293__int64 _InterlockedDecrement64(__int64 volatile *_Addend);
294static __inline__
295__int64 _InterlockedExchange64(__int64 volatile *_Target, __int64 _Value);
296static __inline__
297__int64 _InterlockedExchangeAdd64(__int64 volatile *_Addend, __int64 _Value);
298static __inline__
299__int64 _InterlockedExchangeSub64(__int64 volatile *_Subend, __int64 _Value);
300static __inline__
301__int64 _InterlockedIncrement64(__int64 volatile *_Addend);
302static __inline__
303__int64 _InterlockedOr64(__int64 volatile *_Value, __int64 _Mask);
304static __inline__
305__int64 _InterlockedXor64(__int64 volatile *_Value, __int64 _Mask);
306static __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__)
315char _InterlockedExchangeAdd8_acq(char volatile *_Addend, char _Value);
316char _InterlockedExchangeAdd8_nf(char volatile *_Addend, char _Value);
317char _InterlockedExchangeAdd8_rel(char volatile *_Addend, char _Value);
318short _InterlockedExchangeAdd16_acq(short volatile *_Addend, short _Value);
319short _InterlockedExchangeAdd16_nf(short volatile *_Addend, short _Value);
320short _InterlockedExchangeAdd16_rel(short volatile *_Addend, short _Value);
321long _InterlockedExchangeAdd_acq(long volatile *_Addend, long _Value);
322long _InterlockedExchangeAdd_nf(long volatile *_Addend, long _Value);
323long _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__)
332short _InterlockedIncrement16_acq(short volatile *_Value);
333short _InterlockedIncrement16_nf(short volatile *_Value);
334short _InterlockedIncrement16_rel(short volatile *_Value);
335long _InterlockedIncrement_acq(long volatile *_Value);
336long _InterlockedIncrement_nf(long volatile *_Value);
337long _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__)
346short _InterlockedDecrement16_acq(short volatile *_Value);
347short _InterlockedDecrement16_nf(short volatile *_Value);
348short _InterlockedDecrement16_rel(short volatile *_Value);
349long _InterlockedDecrement_acq(long volatile *_Value);
350long _InterlockedDecrement_nf(long volatile *_Value);
351long _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__)
360char _InterlockedAnd8_acq(char volatile *_Value, char _Mask);
361char _InterlockedAnd8_nf(char volatile *_Value, char _Mask);
362char _InterlockedAnd8_rel(char volatile *_Value, char _Mask);
363short _InterlockedAnd16_acq(short volatile *_Value, short _Mask);
364short _InterlockedAnd16_nf(short volatile *_Value, short _Mask);
365short _InterlockedAnd16_rel(short volatile *_Value, short _Mask);
366long _InterlockedAnd_acq(long volatile *_Value, long _Mask);
367long _InterlockedAnd_nf(long volatile *_Value, long _Mask);
368long _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__)
377unsigned char _interlockedbittestandset_acq(long volatile *_BitBase,
378                                            long _BitPos);
379unsigned char _interlockedbittestandset_nf(long volatile *_BitBase,
380                                           long _BitPos);
381unsigned char _interlockedbittestandset_rel(long volatile *_BitBase,
382                                            long _BitPos);
383unsigned char _interlockedbittestandreset_acq(long volatile *_BitBase,
384                                              long _BitPos);
385unsigned char _interlockedbittestandreset_nf(long volatile *_BitBase,
386                                             long _BitPos);
387unsigned char _interlockedbittestandreset_rel(long volatile *_BitBase,
388                                              long _BitPos);
389#endif
390/*----------------------------------------------------------------------------*\
391|* Interlocked Or
392\*----------------------------------------------------------------------------*/
393#if defined(__arm__) || defined(__aarch64__)
394char _InterlockedOr8_acq(char volatile *_Value, char _Mask);
395char _InterlockedOr8_nf(char volatile *_Value, char _Mask);
396char _InterlockedOr8_rel(char volatile *_Value, char _Mask);
397short _InterlockedOr16_acq(short volatile *_Value, short _Mask);
398short _InterlockedOr16_nf(short volatile *_Value, short _Mask);
399short _InterlockedOr16_rel(short volatile *_Value, short _Mask);
400long _InterlockedOr_acq(long volatile *_Value, long _Mask);
401long _InterlockedOr_nf(long volatile *_Value, long _Mask);
402long _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__)
411char _InterlockedXor8_acq(char volatile *_Value, char _Mask);
412char _InterlockedXor8_nf(char volatile *_Value, char _Mask);
413char _InterlockedXor8_rel(char volatile *_Value, char _Mask);
414short _InterlockedXor16_acq(short volatile *_Value, short _Mask);
415short _InterlockedXor16_nf(short volatile *_Value, short _Mask);
416short _InterlockedXor16_rel(short volatile *_Value, short _Mask);
417long _InterlockedXor_acq(long volatile *_Value, long _Mask);
418long _InterlockedXor_nf(long volatile *_Value, long _Mask);
419long _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__)
428char _InterlockedExchange8_acq(char volatile *_Target, char _Value);
429char _InterlockedExchange8_nf(char volatile *_Target, char _Value);
430char _InterlockedExchange8_rel(char volatile *_Target, char _Value);
431short _InterlockedExchange16_acq(short volatile *_Target, short _Value);
432short _InterlockedExchange16_nf(short volatile *_Target, short _Value);
433short _InterlockedExchange16_rel(short volatile *_Target, short _Value);
434long _InterlockedExchange_acq(long volatile *_Target, long _Value);
435long _InterlockedExchange_nf(long volatile *_Target, long _Value);
436long _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__)
445char _InterlockedCompareExchange8_acq(char volatile *_Destination,
446                             char _Exchange, char _Comparand);
447char _InterlockedCompareExchange8_nf(char volatile *_Destination,
448                             char _Exchange, char _Comparand);
449char _InterlockedCompareExchange8_rel(char volatile *_Destination,
450                             char _Exchange, char _Comparand);
451short _InterlockedCompareExchange16_acq(short volatile *_Destination,
452                              short _Exchange, short _Comparand);
453short _InterlockedCompareExchange16_nf(short volatile *_Destination,
454                              short _Exchange, short _Comparand);
455short _InterlockedCompareExchange16_rel(short volatile *_Destination,
456                              short _Exchange, short _Comparand);
457long _InterlockedCompareExchange_acq(long volatile *_Destination,
458                              long _Exchange, long _Comparand);
459long _InterlockedCompareExchange_nf(long volatile *_Destination,
460                              long _Exchange, long _Comparand);
461long _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__)
475static __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}
480static __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}
485static __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}
490static __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}
495static __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__
502static __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}
507static __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__)
518static __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}
523static __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}
528static __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__)
535static __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__)
545unsigned __int64 __getReg(int);
546long _InterlockedAdd(long volatile *Addend, long Value);
547__int64 _ReadStatusReg(int);
548void _WriteStatusReg(int, __int64);
549
550unsigned short __cdecl _byteswap_ushort(unsigned short val);
551unsigned long __cdecl _byteswap_ulong (unsigned long val);
552unsigned __int64 __cdecl _byteswap_uint64(unsigned __int64 val);
553#endif
554
555/*----------------------------------------------------------------------------*\
556|* Privileged intrinsics
557\*----------------------------------------------------------------------------*/
558#if defined(__i386__) || defined(__x86_64__)
559static __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
574static __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
581static __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