1 /* ===-------- Intrin.h ---------------------------------------------------===
3 * Permission is hereby granted, free of charge, to any person obtaining a copy
4 * of this software and associated documentation files (the "Software"), to deal
5 * in the Software without restriction, including without limitation the rights
6 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
7 * copies of the Software, and to permit persons to whom the Software is
8 * furnished to do so, subject to the following conditions:
10 * The above copyright notice and this permission notice shall be included in
11 * all copies or substantial portions of the Software.
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
16 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
17 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
18 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
21 *===-----------------------------------------------------------------------===
24 /* Only include this if we're compiling for the windows platform. */
26 #include_next <Intrin.h>
32 /* First include the standard intrinsics. */
33 #if defined(__i386__) || defined(__x86_64__)
34 #include <x86intrin.h>
37 /* For the definition of jmp_buf. */
42 /* Define the default attributes for the functions in this file. */
43 #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__))
50 /* And the random ones that aren't in those files. */
51 __m64 _m_from_float(float);
52 __m64 _m_from_int(int _l);
53 void _m_prefetch(void *);
54 float _m_to_float(__m64);
55 int _m_to_int(__m64 _M);
58 /* Other assorted instruction intrinsics. */
59 void __addfsbyte(unsigned long, unsigned char);
60 void __addfsdword(unsigned long, unsigned long);
61 void __addfsword(unsigned long, unsigned short);
62 void __code_seg(const char *);
64 void __cpuid(int[4], int);
66 void __cpuidex(int[4], int, int);
67 void __debugbreak(void);
68 __int64 __emul(int, int);
69 unsigned __int64 __emulu(unsigned int, unsigned int);
70 void __cdecl __fastfail(unsigned int);
71 unsigned int __getcallerseflags(void);
74 unsigned char __inbyte(unsigned short);
75 void __inbytestring(unsigned short, unsigned char *, unsigned long);
76 void __incfsbyte(unsigned long);
77 void __incfsdword(unsigned long);
78 void __incfsword(unsigned long);
79 unsigned long __indword(unsigned short);
80 void __indwordstring(unsigned short, unsigned long *, unsigned long);
82 void __invlpg(void *);
83 unsigned short __inword(unsigned short);
84 void __inwordstring(unsigned short, unsigned short *, unsigned long);
86 unsigned __int64 __ll_lshift(unsigned __int64, int);
87 __int64 __ll_rshift(__int64, int);
88 void __llwpcb(void *);
89 unsigned char __lwpins32(unsigned int, unsigned int, unsigned int);
90 void __lwpval32(unsigned int, unsigned int, unsigned int);
91 unsigned int __lzcnt(unsigned int);
92 unsigned short __lzcnt16(unsigned short);
94 void __movsb(unsigned char *, unsigned char const *, size_t);
96 void __movsd(unsigned long *, unsigned long const *, size_t);
98 void __movsw(unsigned short *, unsigned short const *, size_t);
100 void __nvreg_restore_fence(void);
101 void __nvreg_save_fence(void);
102 void __outbyte(unsigned short, unsigned char);
103 void __outbytestring(unsigned short, unsigned char *, unsigned long);
104 void __outdword(unsigned short, unsigned long);
105 void __outdwordstring(unsigned short, unsigned long *, unsigned long);
106 void __outword(unsigned short, unsigned short);
107 void __outwordstring(unsigned short, unsigned short *, unsigned long);
109 unsigned int __popcnt(unsigned int);
111 unsigned short __popcnt16(unsigned short);
112 unsigned long __readcr0(void);
113 unsigned long __readcr2(void);
115 unsigned long __readcr3(void);
116 unsigned long __readcr4(void);
117 unsigned long __readcr8(void);
118 unsigned int __readdr(unsigned int);
121 unsigned char __readfsbyte(unsigned long);
123 unsigned long __readfsdword(unsigned long);
125 unsigned __int64 __readfsqword(unsigned long);
127 unsigned short __readfsword(unsigned long);
130 unsigned __int64 __readmsr(unsigned long);
131 unsigned __int64 __readpmc(unsigned long);
132 unsigned long __segmentlimit(unsigned long);
134 void *__slwpcb(void);
136 void __stosb(unsigned char *, unsigned char, size_t);
138 void __stosd(unsigned long *, unsigned long, size_t);
140 void __stosw(unsigned short *, unsigned short, size_t);
141 void __svm_clgi(void);
142 void __svm_invlpga(void *, int);
143 void __svm_skinit(int);
144 void __svm_stgi(void);
145 void __svm_vmload(size_t);
146 void __svm_vmrun(size_t);
147 void __svm_vmsave(size_t);
149 unsigned __int64 __ull_rshift(unsigned __int64, int);
150 void __vmx_off(void);
151 void __vmx_vmptrst(unsigned __int64 *);
153 void __writecr0(unsigned int);
155 void __writecr3(unsigned int);
156 void __writecr4(unsigned int);
157 void __writecr8(unsigned int);
158 void __writedr(unsigned int, unsigned int);
159 void __writefsbyte(unsigned long, unsigned char);
160 void __writefsdword(unsigned long, unsigned long);
161 void __writefsqword(unsigned long, unsigned __int64);
162 void __writefsword(unsigned long, unsigned short);
163 void __writemsr(unsigned long, unsigned __int64);
165 void *_AddressOfReturnAddress(void);
167 unsigned char _BitScanForward(unsigned long *_Index, unsigned long _Mask);
169 unsigned char _BitScanReverse(unsigned long *_Index, unsigned long _Mask);
171 unsigned char _bittest(long const *, long);
173 unsigned char _bittestandcomplement(long *, long);
175 unsigned char _bittestandreset(long *, long);
177 unsigned char _bittestandset(long *, long);
178 unsigned __int64 __cdecl _byteswap_uint64(unsigned __int64);
179 unsigned long __cdecl _byteswap_ulong(unsigned long);
180 unsigned short __cdecl _byteswap_ushort(unsigned short);
181 void __cdecl _disable(void);
182 void __cdecl _enable(void);
183 long _InterlockedAddLargeStatistic(__int64 volatile *_Addend, long _Value);
185 long _InterlockedAnd(long volatile *_Value, long _Mask);
187 short _InterlockedAnd16(short volatile *_Value, short _Mask);
189 char _InterlockedAnd8(char volatile *_Value, char _Mask);
190 unsigned char _interlockedbittestandreset(long volatile *, long);
192 unsigned char _interlockedbittestandset(long volatile *, long);
194 long __cdecl _InterlockedCompareExchange(long volatile *_Destination,
195 long _Exchange, long _Comparand);
196 long _InterlockedCompareExchange_HLEAcquire(long volatile *, long, long);
197 long _InterlockedCompareExchange_HLERelease(long volatile *, long, long);
199 short _InterlockedCompareExchange16(short volatile *_Destination,
200 short _Exchange, short _Comparand);
202 __int64 _InterlockedCompareExchange64(__int64 volatile *_Destination,
203 __int64 _Exchange, __int64 _Comparand);
204 __int64 _InterlockedcompareExchange64_HLEAcquire(__int64 volatile *, __int64,
206 __int64 _InterlockedCompareExchange64_HLERelease(__int64 volatile *, __int64,
209 char _InterlockedCompareExchange8(char volatile *_Destination, char _Exchange,
211 void *_InterlockedCompareExchangePointer_HLEAcquire(void *volatile *, void *,
213 void *_InterlockedCompareExchangePointer_HLERelease(void *volatile *, void *,
216 long __cdecl _InterlockedDecrement(long volatile *_Addend);
218 short _InterlockedDecrement16(short volatile *_Addend);
219 long _InterlockedExchange(long volatile *_Target, long _Value);
221 short _InterlockedExchange16(short volatile *_Target, short _Value);
223 char _InterlockedExchange8(char volatile *_Target, char _Value);
225 long __cdecl _InterlockedExchangeAdd(long volatile *_Addend, long _Value);
226 long _InterlockedExchangeAdd_HLEAcquire(long volatile *, long);
227 long _InterlockedExchangeAdd_HLERelease(long volatile *, long);
229 short _InterlockedExchangeAdd16(short volatile *_Addend, short _Value);
230 __int64 _InterlockedExchangeAdd64_HLEAcquire(__int64 volatile *, __int64);
231 __int64 _InterlockedExchangeAdd64_HLERelease(__int64 volatile *, __int64);
233 char _InterlockedExchangeAdd8(char volatile *_Addend, char _Value);
235 long __cdecl _InterlockedIncrement(long volatile *_Addend);
237 short _InterlockedIncrement16(short volatile *_Addend);
239 long _InterlockedOr(long volatile *_Value, long _Mask);
241 short _InterlockedOr16(short volatile *_Value, short _Mask);
243 char _InterlockedOr8(char volatile *_Value, char _Mask);
245 long _InterlockedXor(long volatile *_Value, long _Mask);
247 short _InterlockedXor16(short volatile *_Value, short _Mask);
249 char _InterlockedXor8(char volatile *_Value, char _Mask);
250 void __cdecl _invpcid(unsigned int, void *);
252 unsigned long __cdecl _lrotl(unsigned long, int);
254 unsigned long __cdecl _lrotr(unsigned long, int);
257 void _ReadBarrier(void);
259 void _ReadWriteBarrier(void);
261 void *_ReturnAddress(void);
262 unsigned int _rorx_u32(unsigned int, const unsigned int);
264 unsigned int __cdecl _rotl(unsigned int _Value, int _Shift);
266 unsigned short _rotl16(unsigned short _Value, unsigned char _Shift);
268 unsigned __int64 __cdecl _rotl64(unsigned __int64 _Value, int _Shift);
270 unsigned char _rotl8(unsigned char _Value, unsigned char _Shift);
272 unsigned int __cdecl _rotr(unsigned int _Value, int _Shift);
274 unsigned short _rotr16(unsigned short _Value, unsigned char _Shift);
276 unsigned __int64 __cdecl _rotr64(unsigned __int64 _Value, int _Shift);
278 unsigned char _rotr8(unsigned char _Value, unsigned char _Shift);
279 int _sarx_i32(int, unsigned int);
281 int __cdecl _setjmp(jmp_buf);
283 unsigned int _shlx_u32(unsigned int, unsigned int);
284 unsigned int _shrx_u32(unsigned int, unsigned int);
285 void _Store_HLERelease(long volatile *, long);
286 void _Store64_HLERelease(__int64 volatile *, __int64);
287 void _StorePointer_HLERelease(void *volatile *, void *);
289 void _WriteBarrier(void);
290 unsigned __int32 xbegin(void);
293 #define _XCR_XFEATURE_ENABLED_MASK 0
294 unsigned __int64 __cdecl _xgetbv(unsigned int);
295 void __cdecl _xrstor(void const *, unsigned __int64);
296 void __cdecl _xsave(void *, unsigned __int64);
297 void __cdecl _xsaveopt(void *, unsigned __int64);
298 void __cdecl _xsetbv(unsigned int, unsigned __int64);
300 /* These additional intrinsics are turned on in x64/amd64/x86_64 mode. */
302 void __addgsbyte(unsigned long, unsigned char);
303 void __addgsdword(unsigned long, unsigned long);
304 void __addgsqword(unsigned long, unsigned __int64);
305 void __addgsword(unsigned long, unsigned short);
307 void __faststorefence(void);
308 void __incgsbyte(unsigned long);
309 void __incgsdword(unsigned long);
310 void __incgsqword(unsigned long);
311 void __incgsword(unsigned long);
312 unsigned char __lwpins64(unsigned __int64, unsigned int, unsigned int);
313 void __lwpval64(unsigned __int64, unsigned int, unsigned int);
314 unsigned __int64 __lzcnt64(unsigned __int64);
316 void __movsq(unsigned long long *, unsigned long long const *, size_t);
317 __int64 __mulh(__int64, __int64);
319 unsigned __int64 __popcnt64(unsigned __int64);
321 unsigned char __readgsbyte(unsigned long);
323 unsigned long __readgsdword(unsigned long);
325 unsigned __int64 __readgsqword(unsigned long);
326 unsigned short __readgsword(unsigned long);
327 unsigned __int64 __shiftleft128(unsigned __int64 _LowPart,
328 unsigned __int64 _HighPart,
329 unsigned char _Shift);
330 unsigned __int64 __shiftright128(unsigned __int64 _LowPart,
331 unsigned __int64 _HighPart,
332 unsigned char _Shift);
334 void __stosq(unsigned __int64 *, unsigned __int64, size_t);
335 unsigned char __vmx_on(unsigned __int64 *);
336 unsigned char __vmx_vmclear(unsigned __int64 *);
337 unsigned char __vmx_vmlaunch(void);
338 unsigned char __vmx_vmptrld(unsigned __int64 *);
339 unsigned char __vmx_vmread(size_t, size_t *);
340 unsigned char __vmx_vmresume(void);
341 unsigned char __vmx_vmwrite(size_t, size_t);
342 void __writegsbyte(unsigned long, unsigned char);
343 void __writegsdword(unsigned long, unsigned long);
344 void __writegsqword(unsigned long, unsigned __int64);
345 void __writegsword(unsigned long, unsigned short);
347 unsigned char _BitScanForward64(unsigned long *_Index, unsigned __int64 _Mask);
349 unsigned char _BitScanReverse64(unsigned long *_Index, unsigned __int64 _Mask);
351 unsigned char _bittest64(__int64 const *, __int64);
353 unsigned char _bittestandcomplement64(__int64 *, __int64);
355 unsigned char _bittestandreset64(__int64 *, __int64);
357 unsigned char _bittestandset64(__int64 *, __int64);
358 unsigned __int64 __cdecl _byteswap_uint64(unsigned __int64);
359 long _InterlockedAnd_np(long volatile *_Value, long _Mask);
360 short _InterlockedAnd16_np(short volatile *_Value, short _Mask);
361 __int64 _InterlockedAnd64_np(__int64 volatile *_Value, __int64 _Mask);
362 char _InterlockedAnd8_np(char volatile *_Value, char _Mask);
363 unsigned char _interlockedbittestandreset64(__int64 volatile *, __int64);
365 unsigned char _interlockedbittestandset64(__int64 volatile *, __int64);
366 long _InterlockedCompareExchange_np(long volatile *_Destination, long _Exchange,
368 unsigned char _InterlockedCompareExchange128(__int64 volatile *_Destination,
369 __int64 _ExchangeHigh,
370 __int64 _ExchangeLow,
371 __int64 *_CompareandResult);
372 unsigned char _InterlockedCompareExchange128_np(__int64 volatile *_Destination,
373 __int64 _ExchangeHigh,
374 __int64 _ExchangeLow,
375 __int64 *_ComparandResult);
376 short _InterlockedCompareExchange16_np(short volatile *_Destination,
377 short _Exchange, short _Comparand);
378 __int64 _InterlockedCompareExchange64_HLEAcquire(__int64 volatile *, __int64,
380 __int64 _InterlockedCompareExchange64_HLERelease(__int64 volatile *, __int64,
382 __int64 _InterlockedCompareExchange64_np(__int64 volatile *_Destination,
383 __int64 _Exchange, __int64 _Comparand);
384 void *_InterlockedCompareExchangePointer(void *volatile *_Destination,
385 void *_Exchange, void *_Comparand);
386 void *_InterlockedCompareExchangePointer_np(void *volatile *_Destination,
387 void *_Exchange, void *_Comparand);
389 __int64 _InterlockedDecrement64(__int64 volatile *_Addend);
391 __int64 _InterlockedExchange64(__int64 volatile *_Target, __int64 _Value);
393 __int64 _InterlockedExchangeAdd64(__int64 volatile *_Addend, __int64 _Value);
394 void *_InterlockedExchangePointer(void *volatile *_Target, void *_Value);
396 __int64 _InterlockedIncrement64(__int64 volatile *_Addend);
397 long _InterlockedOr_np(long volatile *_Value, long _Mask);
398 short _InterlockedOr16_np(short volatile *_Value, short _Mask);
400 __int64 _InterlockedOr64(__int64 volatile *_Value, __int64 _Mask);
401 __int64 _InterlockedOr64_np(__int64 volatile *_Value, __int64 _Mask);
402 char _InterlockedOr8_np(char volatile *_Value, char _Mask);
403 long _InterlockedXor_np(long volatile *_Value, long _Mask);
404 short _InterlockedXor16_np(short volatile *_Value, short _Mask);
406 __int64 _InterlockedXor64(__int64 volatile *_Value, __int64 _Mask);
407 __int64 _InterlockedXor64_np(__int64 volatile *_Value, __int64 _Mask);
408 char _InterlockedXor8_np(char volatile *_Value, char _Mask);
410 __int64 _mul128(__int64 _Multiplier, __int64 _Multiplicand,
411 __int64 *_HighProduct);
412 unsigned __int64 _rorx_u64(unsigned __int64, const unsigned int);
413 __int64 _sarx_i64(__int64, unsigned int);
415 int __cdecl _setjmpex(jmp_buf);
417 unsigned __int64 _shlx_u64(unsigned __int64, unsigned int);
418 unsigned __int64 _shrx_u64(unsigned __int64, unsigned int);
420 * Multiply two 64-bit integers and obtain a 64-bit result.
421 * The low-half is returned directly and the high half is in an out parameter.
423 static __inline__ unsigned __int64 __DEFAULT_FN_ATTRS
424 _umul128(unsigned __int64 _Multiplier, unsigned __int64 _Multiplicand,
425 unsigned __int64 *_HighProduct) {
426 unsigned __int128 _FullProduct =
427 (unsigned __int128)_Multiplier * (unsigned __int128)_Multiplicand;
428 *_HighProduct = _FullProduct >> 64;
431 static __inline__ unsigned __int64 __DEFAULT_FN_ATTRS
432 __umulh(unsigned __int64 _Multiplier, unsigned __int64 _Multiplicand) {
433 unsigned __int128 _FullProduct =
434 (unsigned __int128)_Multiplier * (unsigned __int128)_Multiplicand;
435 return _FullProduct >> 64;
437 void __cdecl _xrstor64(void const *, unsigned __int64);
438 void __cdecl _xsave64(void *, unsigned __int64);
439 void __cdecl _xsaveopt64(void *, unsigned __int64);
441 #endif /* __x86_64__ */
443 /*----------------------------------------------------------------------------*\
445 \*----------------------------------------------------------------------------*/
446 static __inline__ unsigned char __DEFAULT_FN_ATTRS
447 _rotl8(unsigned char _Value, unsigned char _Shift) {
449 return _Shift ? (_Value << _Shift) | (_Value >> (8 - _Shift)) : _Value;
451 static __inline__ unsigned char __DEFAULT_FN_ATTRS
452 _rotr8(unsigned char _Value, unsigned char _Shift) {
454 return _Shift ? (_Value >> _Shift) | (_Value << (8 - _Shift)) : _Value;
456 static __inline__ unsigned short __DEFAULT_FN_ATTRS
457 _rotl16(unsigned short _Value, unsigned char _Shift) {
459 return _Shift ? (_Value << _Shift) | (_Value >> (16 - _Shift)) : _Value;
461 static __inline__ unsigned short __DEFAULT_FN_ATTRS
462 _rotr16(unsigned short _Value, unsigned char _Shift) {
464 return _Shift ? (_Value >> _Shift) | (_Value << (16 - _Shift)) : _Value;
466 static __inline__ unsigned int __DEFAULT_FN_ATTRS
467 _rotl(unsigned int _Value, int _Shift) {
469 return _Shift ? (_Value << _Shift) | (_Value >> (32 - _Shift)) : _Value;
471 static __inline__ unsigned int __DEFAULT_FN_ATTRS
472 _rotr(unsigned int _Value, int _Shift) {
474 return _Shift ? (_Value >> _Shift) | (_Value << (32 - _Shift)) : _Value;
476 static __inline__ unsigned long __DEFAULT_FN_ATTRS
477 _lrotl(unsigned long _Value, int _Shift) {
479 return _Shift ? (_Value << _Shift) | (_Value >> (32 - _Shift)) : _Value;
481 static __inline__ unsigned long __DEFAULT_FN_ATTRS
482 _lrotr(unsigned long _Value, int _Shift) {
484 return _Shift ? (_Value >> _Shift) | (_Value << (32 - _Shift)) : _Value;
487 __inline__ unsigned __int64 __DEFAULT_FN_ATTRS
488 _rotl64(unsigned __int64 _Value, int _Shift) {
490 return _Shift ? (_Value << _Shift) | (_Value >> (64 - _Shift)) : _Value;
493 __inline__ unsigned __int64 __DEFAULT_FN_ATTRS
494 _rotr64(unsigned __int64 _Value, int _Shift) {
496 return _Shift ? (_Value >> _Shift) | (_Value << (64 - _Shift)) : _Value;
498 /*----------------------------------------------------------------------------*\
499 |* Bit Counting and Testing
500 \*----------------------------------------------------------------------------*/
501 static __inline__ unsigned char __DEFAULT_FN_ATTRS
502 _BitScanForward(unsigned long *_Index, unsigned long _Mask) {
505 *_Index = __builtin_ctzl(_Mask);
508 static __inline__ unsigned char __DEFAULT_FN_ATTRS
509 _BitScanReverse(unsigned long *_Index, unsigned long _Mask) {
512 *_Index = 31 - __builtin_clzl(_Mask);
515 static __inline__ unsigned short __DEFAULT_FN_ATTRS
516 __popcnt16(unsigned short _Value) {
517 return __builtin_popcount((int)_Value);
519 static __inline__ unsigned int __DEFAULT_FN_ATTRS
520 __popcnt(unsigned int _Value) {
521 return __builtin_popcount(_Value);
523 static __inline__ unsigned char __DEFAULT_FN_ATTRS
524 _bittest(long const *_BitBase, long _BitPos) {
525 return (*_BitBase >> _BitPos) & 1;
527 static __inline__ unsigned char __DEFAULT_FN_ATTRS
528 _bittestandcomplement(long *_BitBase, long _BitPos) {
529 unsigned char _Res = (*_BitBase >> _BitPos) & 1;
530 *_BitBase = *_BitBase ^ (1 << _BitPos);
533 static __inline__ unsigned char __DEFAULT_FN_ATTRS
534 _bittestandreset(long *_BitBase, long _BitPos) {
535 unsigned char _Res = (*_BitBase >> _BitPos) & 1;
536 *_BitBase = *_BitBase & ~(1 << _BitPos);
539 static __inline__ unsigned char __DEFAULT_FN_ATTRS
540 _bittestandset(long *_BitBase, long _BitPos) {
541 unsigned char _Res = (*_BitBase >> _BitPos) & 1;
542 *_BitBase = *_BitBase | (1 << _BitPos);
545 static __inline__ unsigned char __DEFAULT_FN_ATTRS
546 _interlockedbittestandset(long volatile *_BitBase, long _BitPos) {
547 long _PrevVal = __atomic_fetch_or(_BitBase, 1l << _BitPos, __ATOMIC_SEQ_CST);
548 return (_PrevVal >> _BitPos) & 1;
551 static __inline__ unsigned char __DEFAULT_FN_ATTRS
552 _BitScanForward64(unsigned long *_Index, unsigned __int64 _Mask) {
555 *_Index = __builtin_ctzll(_Mask);
558 static __inline__ unsigned char __DEFAULT_FN_ATTRS
559 _BitScanReverse64(unsigned long *_Index, unsigned __int64 _Mask) {
562 *_Index = 63 - __builtin_clzll(_Mask);
566 unsigned __int64 __DEFAULT_FN_ATTRS
567 __popcnt64(unsigned __int64 _Value) {
568 return __builtin_popcountll(_Value);
570 static __inline__ unsigned char __DEFAULT_FN_ATTRS
571 _bittest64(__int64 const *_BitBase, __int64 _BitPos) {
572 return (*_BitBase >> _BitPos) & 1;
574 static __inline__ unsigned char __DEFAULT_FN_ATTRS
575 _bittestandcomplement64(__int64 *_BitBase, __int64 _BitPos) {
576 unsigned char _Res = (*_BitBase >> _BitPos) & 1;
577 *_BitBase = *_BitBase ^ (1ll << _BitPos);
580 static __inline__ unsigned char __DEFAULT_FN_ATTRS
581 _bittestandreset64(__int64 *_BitBase, __int64 _BitPos) {
582 unsigned char _Res = (*_BitBase >> _BitPos) & 1;
583 *_BitBase = *_BitBase & ~(1ll << _BitPos);
586 static __inline__ unsigned char __DEFAULT_FN_ATTRS
587 _bittestandset64(__int64 *_BitBase, __int64 _BitPos) {
588 unsigned char _Res = (*_BitBase >> _BitPos) & 1;
589 *_BitBase = *_BitBase | (1ll << _BitPos);
592 static __inline__ unsigned char __DEFAULT_FN_ATTRS
593 _interlockedbittestandset64(__int64 volatile *_BitBase, __int64 _BitPos) {
595 __atomic_fetch_or(_BitBase, 1ll << _BitPos, __ATOMIC_SEQ_CST);
596 return (_PrevVal >> _BitPos) & 1;
599 /*----------------------------------------------------------------------------*\
600 |* Interlocked Exchange Add
601 \*----------------------------------------------------------------------------*/
602 static __inline__ char __DEFAULT_FN_ATTRS
603 _InterlockedExchangeAdd8(char volatile *_Addend, char _Value) {
604 return __atomic_fetch_add(_Addend, _Value, __ATOMIC_SEQ_CST);
606 static __inline__ short __DEFAULT_FN_ATTRS
607 _InterlockedExchangeAdd16(short volatile *_Addend, short _Value) {
608 return __atomic_fetch_add(_Addend, _Value, __ATOMIC_SEQ_CST);
611 static __inline__ __int64 __DEFAULT_FN_ATTRS
612 _InterlockedExchangeAdd64(__int64 volatile *_Addend, __int64 _Value) {
613 return __atomic_fetch_add(_Addend, _Value, __ATOMIC_SEQ_CST);
616 /*----------------------------------------------------------------------------*\
617 |* Interlocked Exchange Sub
618 \*----------------------------------------------------------------------------*/
619 static __inline__ char __DEFAULT_FN_ATTRS
620 _InterlockedExchangeSub8(char volatile *_Subend, char _Value) {
621 return __atomic_fetch_sub(_Subend, _Value, __ATOMIC_SEQ_CST);
623 static __inline__ short __DEFAULT_FN_ATTRS
624 _InterlockedExchangeSub16(short volatile *_Subend, short _Value) {
625 return __atomic_fetch_sub(_Subend, _Value, __ATOMIC_SEQ_CST);
627 static __inline__ long __DEFAULT_FN_ATTRS
628 _InterlockedExchangeSub(long volatile *_Subend, long _Value) {
629 return __atomic_fetch_sub(_Subend, _Value, __ATOMIC_SEQ_CST);
632 static __inline__ __int64 __DEFAULT_FN_ATTRS
633 _InterlockedExchangeSub64(__int64 volatile *_Subend, __int64 _Value) {
634 return __atomic_fetch_sub(_Subend, _Value, __ATOMIC_SEQ_CST);
637 /*----------------------------------------------------------------------------*\
638 |* Interlocked Increment
639 \*----------------------------------------------------------------------------*/
640 static __inline__ short __DEFAULT_FN_ATTRS
641 _InterlockedIncrement16(short volatile *_Value) {
642 return __atomic_add_fetch(_Value, 1, __ATOMIC_SEQ_CST);
645 static __inline__ __int64 __DEFAULT_FN_ATTRS
646 _InterlockedIncrement64(__int64 volatile *_Value) {
647 return __atomic_add_fetch(_Value, 1, __ATOMIC_SEQ_CST);
650 /*----------------------------------------------------------------------------*\
651 |* Interlocked Decrement
652 \*----------------------------------------------------------------------------*/
653 static __inline__ short __DEFAULT_FN_ATTRS
654 _InterlockedDecrement16(short volatile *_Value) {
655 return __atomic_sub_fetch(_Value, 1, __ATOMIC_SEQ_CST);
658 static __inline__ __int64 __DEFAULT_FN_ATTRS
659 _InterlockedDecrement64(__int64 volatile *_Value) {
660 return __atomic_sub_fetch(_Value, 1, __ATOMIC_SEQ_CST);
663 /*----------------------------------------------------------------------------*\
665 \*----------------------------------------------------------------------------*/
666 static __inline__ char __DEFAULT_FN_ATTRS
667 _InterlockedAnd8(char volatile *_Value, char _Mask) {
668 return __atomic_and_fetch(_Value, _Mask, __ATOMIC_SEQ_CST);
670 static __inline__ short __DEFAULT_FN_ATTRS
671 _InterlockedAnd16(short volatile *_Value, short _Mask) {
672 return __atomic_and_fetch(_Value, _Mask, __ATOMIC_SEQ_CST);
674 static __inline__ long __DEFAULT_FN_ATTRS
675 _InterlockedAnd(long volatile *_Value, long _Mask) {
676 return __atomic_and_fetch(_Value, _Mask, __ATOMIC_SEQ_CST);
679 static __inline__ __int64 __DEFAULT_FN_ATTRS
680 _InterlockedAnd64(__int64 volatile *_Value, __int64 _Mask) {
681 return __atomic_and_fetch(_Value, _Mask, __ATOMIC_SEQ_CST);
684 /*----------------------------------------------------------------------------*\
686 \*----------------------------------------------------------------------------*/
687 static __inline__ char __DEFAULT_FN_ATTRS
688 _InterlockedOr8(char volatile *_Value, char _Mask) {
689 return __atomic_or_fetch(_Value, _Mask, __ATOMIC_SEQ_CST);
691 static __inline__ short __DEFAULT_FN_ATTRS
692 _InterlockedOr16(short volatile *_Value, short _Mask) {
693 return __atomic_or_fetch(_Value, _Mask, __ATOMIC_SEQ_CST);
695 static __inline__ long __DEFAULT_FN_ATTRS
696 _InterlockedOr(long volatile *_Value, long _Mask) {
697 return __atomic_or_fetch(_Value, _Mask, __ATOMIC_SEQ_CST);
700 static __inline__ __int64 __DEFAULT_FN_ATTRS
701 _InterlockedOr64(__int64 volatile *_Value, __int64 _Mask) {
702 return __atomic_or_fetch(_Value, _Mask, __ATOMIC_SEQ_CST);
705 /*----------------------------------------------------------------------------*\
707 \*----------------------------------------------------------------------------*/
708 static __inline__ char __DEFAULT_FN_ATTRS
709 _InterlockedXor8(char volatile *_Value, char _Mask) {
710 return __atomic_xor_fetch(_Value, _Mask, __ATOMIC_SEQ_CST);
712 static __inline__ short __DEFAULT_FN_ATTRS
713 _InterlockedXor16(short volatile *_Value, short _Mask) {
714 return __atomic_xor_fetch(_Value, _Mask, __ATOMIC_SEQ_CST);
716 static __inline__ long __DEFAULT_FN_ATTRS
717 _InterlockedXor(long volatile *_Value, long _Mask) {
718 return __atomic_xor_fetch(_Value, _Mask, __ATOMIC_SEQ_CST);
721 static __inline__ __int64 __DEFAULT_FN_ATTRS
722 _InterlockedXor64(__int64 volatile *_Value, __int64 _Mask) {
723 return __atomic_xor_fetch(_Value, _Mask, __ATOMIC_SEQ_CST);
726 /*----------------------------------------------------------------------------*\
727 |* Interlocked Exchange
728 \*----------------------------------------------------------------------------*/
729 static __inline__ char __DEFAULT_FN_ATTRS
730 _InterlockedExchange8(char volatile *_Target, char _Value) {
731 __atomic_exchange(_Target, &_Value, &_Value, __ATOMIC_SEQ_CST);
734 static __inline__ short __DEFAULT_FN_ATTRS
735 _InterlockedExchange16(short volatile *_Target, short _Value) {
736 __atomic_exchange(_Target, &_Value, &_Value, __ATOMIC_SEQ_CST);
740 static __inline__ __int64 __DEFAULT_FN_ATTRS
741 _InterlockedExchange64(__int64 volatile *_Target, __int64 _Value) {
742 __atomic_exchange(_Target, &_Value, &_Value, __ATOMIC_SEQ_CST);
746 /*----------------------------------------------------------------------------*\
747 |* Interlocked Compare Exchange
748 \*----------------------------------------------------------------------------*/
749 static __inline__ char __DEFAULT_FN_ATTRS
750 _InterlockedCompareExchange8(char volatile *_Destination,
751 char _Exchange, char _Comparand) {
752 __atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0,
753 __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST);
756 static __inline__ short __DEFAULT_FN_ATTRS
757 _InterlockedCompareExchange16(short volatile *_Destination,
758 short _Exchange, short _Comparand) {
759 __atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0,
760 __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST);
763 static __inline__ __int64 __DEFAULT_FN_ATTRS
764 _InterlockedCompareExchange64(__int64 volatile *_Destination,
765 __int64 _Exchange, __int64 _Comparand) {
766 __atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0,
767 __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST);
770 /*----------------------------------------------------------------------------*\
772 \*----------------------------------------------------------------------------*/
773 #if defined(__i386__) || defined(__x86_64__)
774 static __inline__ void __DEFAULT_FN_ATTRS
775 __attribute__((__deprecated__("use other intrinsics or C++11 atomics instead")))
776 _ReadWriteBarrier(void) {
777 __asm__ volatile ("" : : : "memory");
779 static __inline__ void __DEFAULT_FN_ATTRS
780 __attribute__((__deprecated__("use other intrinsics or C++11 atomics instead")))
782 __asm__ volatile ("" : : : "memory");
784 static __inline__ void __DEFAULT_FN_ATTRS
785 __attribute__((__deprecated__("use other intrinsics or C++11 atomics instead")))
786 _WriteBarrier(void) {
787 __asm__ volatile ("" : : : "memory");
791 static __inline__ void __DEFAULT_FN_ATTRS
792 __faststorefence(void) {
793 __asm__ volatile("lock orq $0, (%%rsp)" : : : "memory");
796 /*----------------------------------------------------------------------------*\
798 |* (Pointers in address space #256 and #257 are relative to the GS and FS
799 |* segment registers, respectively.)
800 \*----------------------------------------------------------------------------*/
801 #define __ptr_to_addr_space(__addr_space_nbr, __type, __offset) \
802 ((volatile __type __attribute__((__address_space__(__addr_space_nbr)))*) \
806 static __inline__ unsigned char __DEFAULT_FN_ATTRS
807 __readfsbyte(unsigned long __offset) {
808 return *__ptr_to_addr_space(257, unsigned char, __offset);
810 static __inline__ unsigned __int64 __DEFAULT_FN_ATTRS
811 __readfsqword(unsigned long __offset) {
812 return *__ptr_to_addr_space(257, unsigned __int64, __offset);
814 static __inline__ unsigned short __DEFAULT_FN_ATTRS
815 __readfsword(unsigned long __offset) {
816 return *__ptr_to_addr_space(257, unsigned short, __offset);
820 static __inline__ unsigned char __DEFAULT_FN_ATTRS
821 __readgsbyte(unsigned long __offset) {
822 return *__ptr_to_addr_space(256, unsigned char, __offset);
824 static __inline__ unsigned long __DEFAULT_FN_ATTRS
825 __readgsdword(unsigned long __offset) {
826 return *__ptr_to_addr_space(256, unsigned long, __offset);
828 static __inline__ unsigned __int64 __DEFAULT_FN_ATTRS
829 __readgsqword(unsigned long __offset) {
830 return *__ptr_to_addr_space(256, unsigned __int64, __offset);
832 static __inline__ unsigned short __DEFAULT_FN_ATTRS
833 __readgsword(unsigned long __offset) {
834 return *__ptr_to_addr_space(256, unsigned short, __offset);
837 #undef __ptr_to_addr_space
838 /*----------------------------------------------------------------------------*\
840 \*----------------------------------------------------------------------------*/
841 #if defined(__i386__) || defined(__x86_64__)
842 static __inline__ void __DEFAULT_FN_ATTRS
843 __movsb(unsigned char *__dst, unsigned char const *__src, size_t __n) {
844 __asm__("rep movsb" : : "D"(__dst), "S"(__src), "c"(__n)
845 : "%edi", "%esi", "%ecx");
847 static __inline__ void __DEFAULT_FN_ATTRS
848 __movsd(unsigned long *__dst, unsigned long const *__src, size_t __n) {
849 __asm__("rep movsl" : : "D"(__dst), "S"(__src), "c"(__n)
850 : "%edi", "%esi", "%ecx");
852 static __inline__ void __DEFAULT_FN_ATTRS
853 __movsw(unsigned short *__dst, unsigned short const *__src, size_t __n) {
854 __asm__("rep movsh" : : "D"(__dst), "S"(__src), "c"(__n)
855 : "%edi", "%esi", "%ecx");
857 static __inline__ void __DEFAULT_FN_ATTRS
858 __stosb(unsigned char *__dst, unsigned char __x, size_t __n) {
859 __asm__("rep stosb" : : "D"(__dst), "a"(__x), "c"(__n)
862 static __inline__ void __DEFAULT_FN_ATTRS
863 __stosd(unsigned long *__dst, unsigned long __x, size_t __n) {
864 __asm__("rep stosl" : : "D"(__dst), "a"(__x), "c"(__n)
867 static __inline__ void __DEFAULT_FN_ATTRS
868 __stosw(unsigned short *__dst, unsigned short __x, size_t __n) {
869 __asm__("rep stosh" : : "D"(__dst), "a"(__x), "c"(__n)
874 static __inline__ void __DEFAULT_FN_ATTRS
875 __movsq(unsigned long long *__dst, unsigned long long const *__src, size_t __n) {
876 __asm__("rep movsq" : : "D"(__dst), "S"(__src), "c"(__n)
877 : "%edi", "%esi", "%ecx");
879 static __inline__ void __DEFAULT_FN_ATTRS
880 __stosq(unsigned __int64 *__dst, unsigned __int64 __x, size_t __n) {
881 __asm__("rep stosq" : : "D"(__dst), "a"(__x), "c"(__n)
886 /*----------------------------------------------------------------------------*\
888 \*----------------------------------------------------------------------------*/
889 static __inline__ void * __DEFAULT_FN_ATTRS
890 _AddressOfReturnAddress(void) {
891 return (void*)((char*)__builtin_frame_address(0) + sizeof(void*));
893 static __inline__ void * __DEFAULT_FN_ATTRS
894 _ReturnAddress(void) {
895 return __builtin_return_address(0);
897 #if defined(__i386__) || defined(__x86_64__)
898 static __inline__ void __DEFAULT_FN_ATTRS
899 __cpuid(int __info[4], int __level) {
900 __asm__ ("cpuid" : "=a"(__info[0]), "=b" (__info[1]), "=c"(__info[2]), "=d"(__info[3])
903 static __inline__ void __DEFAULT_FN_ATTRS
904 __cpuidex(int __info[4], int __level, int __ecx) {
905 __asm__ ("cpuid" : "=a"(__info[0]), "=b" (__info[1]), "=c"(__info[2]), "=d"(__info[3])
906 : "a"(__level), "c"(__ecx));
908 static __inline__ unsigned __int64 __cdecl __DEFAULT_FN_ATTRS
909 _xgetbv(unsigned int __xcr_no) {
910 unsigned int __eax, __edx;
911 __asm__ ("xgetbv" : "=a" (__eax), "=d" (__edx) : "c" (__xcr_no));
912 return ((unsigned __int64)__edx << 32) | __eax;
914 static __inline__ void __DEFAULT_FN_ATTRS
916 __asm__ volatile ("hlt");
920 /*----------------------------------------------------------------------------*\
921 |* Privileged intrinsics
922 \*----------------------------------------------------------------------------*/
923 #if defined(__i386__) || defined(__x86_64__)
924 static __inline__ unsigned __int64 __DEFAULT_FN_ATTRS
925 __readmsr(unsigned long __register) {
926 // Loads the contents of a 64-bit model specific register (MSR) specified in
927 // the ECX register into registers EDX:EAX. The EDX register is loaded with
928 // the high-order 32 bits of the MSR and the EAX register is loaded with the
929 // low-order 32 bits. If less than 64 bits are implemented in the MSR being
930 // read, the values returned to EDX:EAX in unimplemented bit locations are
934 __asm__ ("rdmsr" : "=d"(__edx), "=a"(__eax) : "c"(__register));
935 return (((unsigned __int64)__edx) << 32) | (unsigned __int64)__eax;
938 static __inline__ unsigned long __DEFAULT_FN_ATTRS
940 unsigned long __cr3_val;
941 __asm__ __volatile__ ("mov %%cr3, %0" : "=q"(__cr3_val) : : "memory");
945 static __inline__ void __DEFAULT_FN_ATTRS
946 __writecr3(unsigned int __cr3_val) {
947 __asm__ ("mov %0, %%cr3" : : "q"(__cr3_val) : "memory");
955 #undef __DEFAULT_FN_ATTRS
957 #endif /* __INTRIN_H */
958 #endif /* _MSC_VER */