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