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 float _m_to_float(__m64
);
55 /* Other assorted instruction intrinsics. */
56 void __addfsbyte(unsigned long, unsigned char);
57 void __addfsdword(unsigned long, unsigned long);
58 void __addfsword(unsigned long, unsigned short);
59 void __code_seg(const char *);
61 void __cpuid(int[4], int);
63 void __cpuidex(int[4], int, int);
64 void __debugbreak(void);
65 __int64
__emul(int, int);
66 unsigned __int64
__emulu(unsigned int, unsigned int);
67 void __cdecl
__fastfail(unsigned int);
68 unsigned int __getcallerseflags(void);
71 unsigned char __inbyte(unsigned short);
72 void __inbytestring(unsigned short, unsigned char *, unsigned long);
73 void __incfsbyte(unsigned long);
74 void __incfsdword(unsigned long);
75 void __incfsword(unsigned long);
76 unsigned long __indword(unsigned short);
77 void __indwordstring(unsigned short, unsigned long *, unsigned long);
79 void __invlpg(void *);
80 unsigned short __inword(unsigned short);
81 void __inwordstring(unsigned short, unsigned short *, unsigned long);
83 unsigned __int64
__ll_lshift(unsigned __int64
, int);
84 __int64
__ll_rshift(__int64
, int);
85 void __llwpcb(void *);
86 unsigned char __lwpins32(unsigned int, unsigned int, unsigned int);
87 void __lwpval32(unsigned int, unsigned int, unsigned int);
88 unsigned int __lzcnt(unsigned int);
89 unsigned short __lzcnt16(unsigned short);
91 void __movsb(unsigned char *, unsigned char const *, size_t);
93 void __movsd(unsigned long *, unsigned long const *, size_t);
95 void __movsw(unsigned short *, unsigned short const *, size_t);
97 void __nvreg_restore_fence(void);
98 void __nvreg_save_fence(void);
99 void __outbyte(unsigned short, unsigned char);
100 void __outbytestring(unsigned short, unsigned char *, unsigned long);
101 void __outdword(unsigned short, unsigned long);
102 void __outdwordstring(unsigned short, unsigned long *, unsigned long);
103 void __outword(unsigned short, unsigned short);
104 void __outwordstring(unsigned short, unsigned short *, unsigned long);
106 unsigned int __popcnt(unsigned int);
108 unsigned short __popcnt16(unsigned short);
109 unsigned long __readcr0(void);
110 unsigned long __readcr2(void);
112 unsigned long __readcr3(void);
113 unsigned long __readcr4(void);
114 unsigned long __readcr8(void);
115 unsigned int __readdr(unsigned int);
118 unsigned char __readfsbyte(unsigned long);
120 unsigned long __readfsdword(unsigned long);
122 unsigned __int64
__readfsqword(unsigned long);
124 unsigned short __readfsword(unsigned long);
127 unsigned __int64
__readmsr(unsigned long);
128 unsigned __int64
__readpmc(unsigned long);
129 unsigned long __segmentlimit(unsigned long);
131 void *__slwpcb(void);
133 void __stosb(unsigned char *, unsigned char, size_t);
135 void __stosd(unsigned long *, unsigned long, size_t);
137 void __stosw(unsigned short *, unsigned short, size_t);
138 void __svm_clgi(void);
139 void __svm_invlpga(void *, int);
140 void __svm_skinit(int);
141 void __svm_stgi(void);
142 void __svm_vmload(size_t);
143 void __svm_vmrun(size_t);
144 void __svm_vmsave(size_t);
146 unsigned __int64
__ull_rshift(unsigned __int64
, int);
147 void __vmx_off(void);
148 void __vmx_vmptrst(unsigned __int64
*);
150 void __writecr0(unsigned int);
152 void __writecr3(unsigned int);
153 void __writecr4(unsigned int);
154 void __writecr8(unsigned int);
155 void __writedr(unsigned int, unsigned int);
156 void __writefsbyte(unsigned long, unsigned char);
157 void __writefsdword(unsigned long, unsigned long);
158 void __writefsqword(unsigned long, unsigned __int64
);
159 void __writefsword(unsigned long, unsigned short);
160 void __writemsr(unsigned long, unsigned __int64
);
162 void *_AddressOfReturnAddress(void);
164 unsigned char _BitScanForward(unsigned long *_Index
, unsigned long _Mask
);
166 unsigned char _BitScanReverse(unsigned long *_Index
, unsigned long _Mask
);
168 unsigned char _bittest(long const *, long);
170 unsigned char _bittestandcomplement(long *, long);
172 unsigned char _bittestandreset(long *, long);
174 unsigned char _bittestandset(long *, long);
175 unsigned __int64 __cdecl
_byteswap_uint64(unsigned __int64
);
176 unsigned long __cdecl
_byteswap_ulong(unsigned long);
177 unsigned short __cdecl
_byteswap_ushort(unsigned short);
178 void __cdecl
_disable(void);
179 void __cdecl
_enable(void);
180 long _InterlockedAddLargeStatistic(__int64
volatile *_Addend
, long _Value
);
182 long _InterlockedAnd(long volatile *_Value
, long _Mask
);
184 short _InterlockedAnd16(short volatile *_Value
, short _Mask
);
186 char _InterlockedAnd8(char volatile *_Value
, char _Mask
);
187 unsigned char _interlockedbittestandreset(long volatile *, long);
189 unsigned char _interlockedbittestandset(long volatile *, long);
191 long __cdecl
_InterlockedCompareExchange(long volatile *_Destination
,
192 long _Exchange
, long _Comparand
);
193 long _InterlockedCompareExchange_HLEAcquire(long volatile *, long, long);
194 long _InterlockedCompareExchange_HLERelease(long volatile *, long, long);
196 short _InterlockedCompareExchange16(short volatile *_Destination
,
197 short _Exchange
, short _Comparand
);
199 __int64
_InterlockedCompareExchange64(__int64
volatile *_Destination
,
200 __int64 _Exchange
, __int64 _Comparand
);
201 __int64
_InterlockedcompareExchange64_HLEAcquire(__int64
volatile *, __int64
,
203 __int64
_InterlockedCompareExchange64_HLERelease(__int64
volatile *, __int64
,
206 char _InterlockedCompareExchange8(char volatile *_Destination
, char _Exchange
,
208 void *_InterlockedCompareExchangePointer_HLEAcquire(void *volatile *, void *,
210 void *_InterlockedCompareExchangePointer_HLERelease(void *volatile *, void *,
213 long __cdecl
_InterlockedDecrement(long volatile *_Addend
);
215 short _InterlockedDecrement16(short volatile *_Addend
);
216 long _InterlockedExchange(long volatile *_Target
, long _Value
);
218 short _InterlockedExchange16(short volatile *_Target
, short _Value
);
220 char _InterlockedExchange8(char volatile *_Target
, char _Value
);
222 long __cdecl
_InterlockedExchangeAdd(long volatile *_Addend
, long _Value
);
223 long _InterlockedExchangeAdd_HLEAcquire(long volatile *, long);
224 long _InterlockedExchangeAdd_HLERelease(long volatile *, long);
226 short _InterlockedExchangeAdd16(short volatile *_Addend
, short _Value
);
227 __int64
_InterlockedExchangeAdd64_HLEAcquire(__int64
volatile *, __int64
);
228 __int64
_InterlockedExchangeAdd64_HLERelease(__int64
volatile *, __int64
);
230 char _InterlockedExchangeAdd8(char volatile *_Addend
, char _Value
);
232 long __cdecl
_InterlockedIncrement(long volatile *_Addend
);
234 short _InterlockedIncrement16(short volatile *_Addend
);
236 long _InterlockedOr(long volatile *_Value
, long _Mask
);
238 short _InterlockedOr16(short volatile *_Value
, short _Mask
);
240 char _InterlockedOr8(char volatile *_Value
, char _Mask
);
242 long _InterlockedXor(long volatile *_Value
, long _Mask
);
244 short _InterlockedXor16(short volatile *_Value
, short _Mask
);
246 char _InterlockedXor8(char volatile *_Value
, char _Mask
);
247 void __cdecl
_invpcid(unsigned int, void *);
249 unsigned long __cdecl
_lrotl(unsigned long, int);
251 unsigned long __cdecl
_lrotr(unsigned long, int);
254 void _ReadBarrier(void);
256 void _ReadWriteBarrier(void);
258 void *_ReturnAddress(void);
259 unsigned int _rorx_u32(unsigned int, const unsigned int);
261 unsigned int __cdecl
_rotl(unsigned int _Value
, int _Shift
);
263 unsigned short _rotl16(unsigned short _Value
, unsigned char _Shift
);
265 unsigned __int64 __cdecl
_rotl64(unsigned __int64 _Value
, int _Shift
);
267 unsigned char _rotl8(unsigned char _Value
, unsigned char _Shift
);
269 unsigned int __cdecl
_rotr(unsigned int _Value
, int _Shift
);
271 unsigned short _rotr16(unsigned short _Value
, unsigned char _Shift
);
273 unsigned __int64 __cdecl
_rotr64(unsigned __int64 _Value
, int _Shift
);
275 unsigned char _rotr8(unsigned char _Value
, unsigned char _Shift
);
276 int _sarx_i32(int, unsigned int);
278 int __cdecl
_setjmp(jmp_buf);
280 unsigned int _shlx_u32(unsigned int, unsigned int);
281 unsigned int _shrx_u32(unsigned int, unsigned int);
282 void _Store_HLERelease(long volatile *, long);
283 void _Store64_HLERelease(__int64
volatile *, __int64
);
284 void _StorePointer_HLERelease(void *volatile *, void *);
286 void _WriteBarrier(void);
287 unsigned __int32
xbegin(void);
290 #define _XCR_XFEATURE_ENABLED_MASK 0
291 unsigned __int64 __cdecl
_xgetbv(unsigned int);
292 void __cdecl
_xsetbv(unsigned int, unsigned __int64
);
294 /* These additional intrinsics are turned on in x64/amd64/x86_64 mode. */
296 void __addgsbyte(unsigned long, unsigned char);
297 void __addgsdword(unsigned long, unsigned long);
298 void __addgsqword(unsigned long, unsigned __int64
);
299 void __addgsword(unsigned long, unsigned short);
301 void __faststorefence(void);
302 void __incgsbyte(unsigned long);
303 void __incgsdword(unsigned long);
304 void __incgsqword(unsigned long);
305 void __incgsword(unsigned long);
306 unsigned char __lwpins64(unsigned __int64
, unsigned int, unsigned int);
307 void __lwpval64(unsigned __int64
, unsigned int, unsigned int);
308 unsigned __int64
__lzcnt64(unsigned __int64
);
310 void __movsq(unsigned long long *, unsigned long long const *, size_t);
311 __int64
__mulh(__int64
, __int64
);
313 unsigned __int64
__popcnt64(unsigned __int64
);
315 unsigned char __readgsbyte(unsigned long);
317 unsigned long __readgsdword(unsigned long);
319 unsigned __int64
__readgsqword(unsigned long);
320 unsigned short __readgsword(unsigned long);
321 unsigned __int64
__shiftleft128(unsigned __int64 _LowPart
,
322 unsigned __int64 _HighPart
,
323 unsigned char _Shift
);
324 unsigned __int64
__shiftright128(unsigned __int64 _LowPart
,
325 unsigned __int64 _HighPart
,
326 unsigned char _Shift
);
328 void __stosq(unsigned __int64
*, unsigned __int64
, size_t);
329 unsigned char __vmx_on(unsigned __int64
*);
330 unsigned char __vmx_vmclear(unsigned __int64
*);
331 unsigned char __vmx_vmlaunch(void);
332 unsigned char __vmx_vmptrld(unsigned __int64
*);
333 unsigned char __vmx_vmread(size_t, size_t *);
334 unsigned char __vmx_vmresume(void);
335 unsigned char __vmx_vmwrite(size_t, size_t);
336 void __writegsbyte(unsigned long, unsigned char);
337 void __writegsdword(unsigned long, unsigned long);
338 void __writegsqword(unsigned long, unsigned __int64
);
339 void __writegsword(unsigned long, unsigned short);
341 unsigned char _BitScanForward64(unsigned long *_Index
, unsigned __int64 _Mask
);
343 unsigned char _BitScanReverse64(unsigned long *_Index
, unsigned __int64 _Mask
);
345 unsigned char _bittest64(__int64
const *, __int64
);
347 unsigned char _bittestandcomplement64(__int64
*, __int64
);
349 unsigned char _bittestandreset64(__int64
*, __int64
);
351 unsigned char _bittestandset64(__int64
*, __int64
);
352 unsigned __int64 __cdecl
_byteswap_uint64(unsigned __int64
);
353 long _InterlockedAnd_np(long volatile *_Value
, long _Mask
);
354 short _InterlockedAnd16_np(short volatile *_Value
, short _Mask
);
355 __int64
_InterlockedAnd64_np(__int64
volatile *_Value
, __int64 _Mask
);
356 char _InterlockedAnd8_np(char volatile *_Value
, char _Mask
);
357 unsigned char _interlockedbittestandreset64(__int64
volatile *, __int64
);
359 unsigned char _interlockedbittestandset64(__int64
volatile *, __int64
);
360 long _InterlockedCompareExchange_np(long volatile *_Destination
, long _Exchange
,
362 unsigned char _InterlockedCompareExchange128(__int64
volatile *_Destination
,
363 __int64 _ExchangeHigh
,
364 __int64 _ExchangeLow
,
365 __int64
*_CompareandResult
);
366 unsigned char _InterlockedCompareExchange128_np(__int64
volatile *_Destination
,
367 __int64 _ExchangeHigh
,
368 __int64 _ExchangeLow
,
369 __int64
*_ComparandResult
);
370 short _InterlockedCompareExchange16_np(short volatile *_Destination
,
371 short _Exchange
, short _Comparand
);
372 __int64
_InterlockedCompareExchange64_HLEAcquire(__int64
volatile *, __int64
,
374 __int64
_InterlockedCompareExchange64_HLERelease(__int64
volatile *, __int64
,
376 __int64
_InterlockedCompareExchange64_np(__int64
volatile *_Destination
,
377 __int64 _Exchange
, __int64 _Comparand
);
378 void *_InterlockedCompareExchangePointer(void *volatile *_Destination
,
379 void *_Exchange
, void *_Comparand
);
380 void *_InterlockedCompareExchangePointer_np(void *volatile *_Destination
,
381 void *_Exchange
, void *_Comparand
);
383 __int64
_InterlockedDecrement64(__int64
volatile *_Addend
);
385 __int64
_InterlockedExchange64(__int64
volatile *_Target
, __int64 _Value
);
387 __int64
_InterlockedExchangeAdd64(__int64
volatile *_Addend
, __int64 _Value
);
388 void *_InterlockedExchangePointer(void *volatile *_Target
, void *_Value
);
390 __int64
_InterlockedIncrement64(__int64
volatile *_Addend
);
391 long _InterlockedOr_np(long volatile *_Value
, long _Mask
);
392 short _InterlockedOr16_np(short volatile *_Value
, short _Mask
);
394 __int64
_InterlockedOr64(__int64
volatile *_Value
, __int64 _Mask
);
395 __int64
_InterlockedOr64_np(__int64
volatile *_Value
, __int64 _Mask
);
396 char _InterlockedOr8_np(char volatile *_Value
, char _Mask
);
397 long _InterlockedXor_np(long volatile *_Value
, long _Mask
);
398 short _InterlockedXor16_np(short volatile *_Value
, short _Mask
);
400 __int64
_InterlockedXor64(__int64
volatile *_Value
, __int64 _Mask
);
401 __int64
_InterlockedXor64_np(__int64
volatile *_Value
, __int64 _Mask
);
402 char _InterlockedXor8_np(char volatile *_Value
, char _Mask
);
404 __int64
_mul128(__int64 _Multiplier
, __int64 _Multiplicand
,
405 __int64
*_HighProduct
);
406 unsigned __int64
_rorx_u64(unsigned __int64
, const unsigned int);
407 __int64
_sarx_i64(__int64
, unsigned int);
409 int __cdecl
_setjmpex(jmp_buf);
411 unsigned __int64
_shlx_u64(unsigned __int64
, unsigned int);
412 unsigned __int64
_shrx_u64(unsigned __int64
, unsigned int);
414 * Multiply two 64-bit integers and obtain a 64-bit result.
415 * The low-half is returned directly and the high half is in an out parameter.
417 static __inline__
unsigned __int64 __DEFAULT_FN_ATTRS
418 _umul128(unsigned __int64 _Multiplier
, unsigned __int64 _Multiplicand
,
419 unsigned __int64
*_HighProduct
) {
420 unsigned __int128 _FullProduct
=
421 (unsigned __int128
)_Multiplier
* (unsigned __int128
)_Multiplicand
;
422 *_HighProduct
= _FullProduct
>> 64;
425 static __inline__
unsigned __int64 __DEFAULT_FN_ATTRS
426 __umulh(unsigned __int64 _Multiplier
, unsigned __int64 _Multiplicand
) {
427 unsigned __int128 _FullProduct
=
428 (unsigned __int128
)_Multiplier
* (unsigned __int128
)_Multiplicand
;
429 return _FullProduct
>> 64;
432 #endif /* __x86_64__ */
434 /*----------------------------------------------------------------------------*\
436 \*----------------------------------------------------------------------------*/
437 static __inline__ __int64 __DEFAULT_FN_ATTRS
438 __emul(int __in1
, int __in2
) {
439 return (__int64
)__in1
* (__int64
)__in2
;
441 static __inline__
unsigned __int64 __DEFAULT_FN_ATTRS
442 __emulu(unsigned int __in1
, unsigned int __in2
) {
443 return (unsigned __int64
)__in1
* (unsigned __int64
)__in2
;
445 /*----------------------------------------------------------------------------*\
447 \*----------------------------------------------------------------------------*/
448 static __inline__
unsigned char __DEFAULT_FN_ATTRS
449 _rotl8(unsigned char _Value
, unsigned char _Shift
) {
451 return _Shift
? (_Value
<< _Shift
) | (_Value
>> (8 - _Shift
)) : _Value
;
453 static __inline__
unsigned char __DEFAULT_FN_ATTRS
454 _rotr8(unsigned char _Value
, unsigned char _Shift
) {
456 return _Shift
? (_Value
>> _Shift
) | (_Value
<< (8 - _Shift
)) : _Value
;
458 static __inline__
unsigned short __DEFAULT_FN_ATTRS
459 _rotl16(unsigned short _Value
, unsigned char _Shift
) {
461 return _Shift
? (_Value
<< _Shift
) | (_Value
>> (16 - _Shift
)) : _Value
;
463 static __inline__
unsigned short __DEFAULT_FN_ATTRS
464 _rotr16(unsigned short _Value
, unsigned char _Shift
) {
466 return _Shift
? (_Value
>> _Shift
) | (_Value
<< (16 - _Shift
)) : _Value
;
468 static __inline__
unsigned int __DEFAULT_FN_ATTRS
469 _rotl(unsigned int _Value
, int _Shift
) {
471 return _Shift
? (_Value
<< _Shift
) | (_Value
>> (32 - _Shift
)) : _Value
;
473 static __inline__
unsigned int __DEFAULT_FN_ATTRS
474 _rotr(unsigned int _Value
, int _Shift
) {
476 return _Shift
? (_Value
>> _Shift
) | (_Value
<< (32 - _Shift
)) : _Value
;
478 static __inline__
unsigned long __DEFAULT_FN_ATTRS
479 _lrotl(unsigned long _Value
, int _Shift
) {
481 return _Shift
? (_Value
<< _Shift
) | (_Value
>> (32 - _Shift
)) : _Value
;
483 static __inline__
unsigned long __DEFAULT_FN_ATTRS
484 _lrotr(unsigned long _Value
, int _Shift
) {
486 return _Shift
? (_Value
>> _Shift
) | (_Value
<< (32 - _Shift
)) : _Value
;
489 __inline__
unsigned __int64 __DEFAULT_FN_ATTRS
490 _rotl64(unsigned __int64 _Value
, int _Shift
) {
492 return _Shift
? (_Value
<< _Shift
) | (_Value
>> (64 - _Shift
)) : _Value
;
495 __inline__
unsigned __int64 __DEFAULT_FN_ATTRS
496 _rotr64(unsigned __int64 _Value
, int _Shift
) {
498 return _Shift
? (_Value
>> _Shift
) | (_Value
<< (64 - _Shift
)) : _Value
;
500 /*----------------------------------------------------------------------------*\
501 |* Bit Counting and Testing
502 \*----------------------------------------------------------------------------*/
503 static __inline__
unsigned char __DEFAULT_FN_ATTRS
504 _BitScanForward(unsigned long *_Index
, unsigned long _Mask
) {
507 *_Index
= __builtin_ctzl(_Mask
);
510 static __inline__
unsigned char __DEFAULT_FN_ATTRS
511 _BitScanReverse(unsigned long *_Index
, unsigned long _Mask
) {
514 *_Index
= 31 - __builtin_clzl(_Mask
);
517 static __inline__
unsigned short __DEFAULT_FN_ATTRS
518 __popcnt16(unsigned short _Value
) {
519 return __builtin_popcount((int)_Value
);
521 static __inline__
unsigned int __DEFAULT_FN_ATTRS
522 __popcnt(unsigned int _Value
) {
523 return __builtin_popcount(_Value
);
525 static __inline__
unsigned char __DEFAULT_FN_ATTRS
526 _bittest(long const *_BitBase
, long _BitPos
) {
527 return (*_BitBase
>> _BitPos
) & 1;
529 static __inline__
unsigned char __DEFAULT_FN_ATTRS
530 _bittestandcomplement(long *_BitBase
, long _BitPos
) {
531 unsigned char _Res
= (*_BitBase
>> _BitPos
) & 1;
532 *_BitBase
= *_BitBase
^ (1 << _BitPos
);
535 static __inline__
unsigned char __DEFAULT_FN_ATTRS
536 _bittestandreset(long *_BitBase
, long _BitPos
) {
537 unsigned char _Res
= (*_BitBase
>> _BitPos
) & 1;
538 *_BitBase
= *_BitBase
& ~(1 << _BitPos
);
541 static __inline__
unsigned char __DEFAULT_FN_ATTRS
542 _bittestandset(long *_BitBase
, long _BitPos
) {
543 unsigned char _Res
= (*_BitBase
>> _BitPos
) & 1;
544 *_BitBase
= *_BitBase
| (1 << _BitPos
);
547 static __inline__
unsigned char __DEFAULT_FN_ATTRS
548 _interlockedbittestandset(long volatile *_BitBase
, long _BitPos
) {
549 long _PrevVal
= __atomic_fetch_or(_BitBase
, 1l << _BitPos
, __ATOMIC_SEQ_CST
);
550 return (_PrevVal
>> _BitPos
) & 1;
553 static __inline__
unsigned char __DEFAULT_FN_ATTRS
554 _BitScanForward64(unsigned long *_Index
, unsigned __int64 _Mask
) {
557 *_Index
= __builtin_ctzll(_Mask
);
560 static __inline__
unsigned char __DEFAULT_FN_ATTRS
561 _BitScanReverse64(unsigned long *_Index
, unsigned __int64 _Mask
) {
564 *_Index
= 63 - __builtin_clzll(_Mask
);
568 unsigned __int64 __DEFAULT_FN_ATTRS
569 __popcnt64(unsigned __int64 _Value
) {
570 return __builtin_popcountll(_Value
);
572 static __inline__
unsigned char __DEFAULT_FN_ATTRS
573 _bittest64(__int64
const *_BitBase
, __int64 _BitPos
) {
574 return (*_BitBase
>> _BitPos
) & 1;
576 static __inline__
unsigned char __DEFAULT_FN_ATTRS
577 _bittestandcomplement64(__int64
*_BitBase
, __int64 _BitPos
) {
578 unsigned char _Res
= (*_BitBase
>> _BitPos
) & 1;
579 *_BitBase
= *_BitBase
^ (1ll << _BitPos
);
582 static __inline__
unsigned char __DEFAULT_FN_ATTRS
583 _bittestandreset64(__int64
*_BitBase
, __int64 _BitPos
) {
584 unsigned char _Res
= (*_BitBase
>> _BitPos
) & 1;
585 *_BitBase
= *_BitBase
& ~(1ll << _BitPos
);
588 static __inline__
unsigned char __DEFAULT_FN_ATTRS
589 _bittestandset64(__int64
*_BitBase
, __int64 _BitPos
) {
590 unsigned char _Res
= (*_BitBase
>> _BitPos
) & 1;
591 *_BitBase
= *_BitBase
| (1ll << _BitPos
);
594 static __inline__
unsigned char __DEFAULT_FN_ATTRS
595 _interlockedbittestandset64(__int64
volatile *_BitBase
, __int64 _BitPos
) {
597 __atomic_fetch_or(_BitBase
, 1ll << _BitPos
, __ATOMIC_SEQ_CST
);
598 return (_PrevVal
>> _BitPos
) & 1;
601 /*----------------------------------------------------------------------------*\
602 |* Interlocked Exchange Add
603 \*----------------------------------------------------------------------------*/
604 static __inline__
char __DEFAULT_FN_ATTRS
605 _InterlockedExchangeAdd8(char volatile *_Addend
, char _Value
) {
606 return __atomic_fetch_add(_Addend
, _Value
, __ATOMIC_SEQ_CST
);
608 static __inline__
short __DEFAULT_FN_ATTRS
609 _InterlockedExchangeAdd16(short volatile *_Addend
, short _Value
) {
610 return __atomic_fetch_add(_Addend
, _Value
, __ATOMIC_SEQ_CST
);
613 static __inline__ __int64 __DEFAULT_FN_ATTRS
614 _InterlockedExchangeAdd64(__int64
volatile *_Addend
, __int64 _Value
) {
615 return __atomic_fetch_add(_Addend
, _Value
, __ATOMIC_SEQ_CST
);
618 /*----------------------------------------------------------------------------*\
619 |* Interlocked Exchange Sub
620 \*----------------------------------------------------------------------------*/
621 static __inline__
char __DEFAULT_FN_ATTRS
622 _InterlockedExchangeSub8(char volatile *_Subend
, char _Value
) {
623 return __atomic_fetch_sub(_Subend
, _Value
, __ATOMIC_SEQ_CST
);
625 static __inline__
short __DEFAULT_FN_ATTRS
626 _InterlockedExchangeSub16(short volatile *_Subend
, short _Value
) {
627 return __atomic_fetch_sub(_Subend
, _Value
, __ATOMIC_SEQ_CST
);
629 static __inline__
long __DEFAULT_FN_ATTRS
630 _InterlockedExchangeSub(long volatile *_Subend
, long _Value
) {
631 return __atomic_fetch_sub(_Subend
, _Value
, __ATOMIC_SEQ_CST
);
634 static __inline__ __int64 __DEFAULT_FN_ATTRS
635 _InterlockedExchangeSub64(__int64
volatile *_Subend
, __int64 _Value
) {
636 return __atomic_fetch_sub(_Subend
, _Value
, __ATOMIC_SEQ_CST
);
639 /*----------------------------------------------------------------------------*\
640 |* Interlocked Increment
641 \*----------------------------------------------------------------------------*/
642 static __inline__
short __DEFAULT_FN_ATTRS
643 _InterlockedIncrement16(short volatile *_Value
) {
644 return __atomic_add_fetch(_Value
, 1, __ATOMIC_SEQ_CST
);
647 static __inline__ __int64 __DEFAULT_FN_ATTRS
648 _InterlockedIncrement64(__int64
volatile *_Value
) {
649 return __atomic_add_fetch(_Value
, 1, __ATOMIC_SEQ_CST
);
652 /*----------------------------------------------------------------------------*\
653 |* Interlocked Decrement
654 \*----------------------------------------------------------------------------*/
655 static __inline__
short __DEFAULT_FN_ATTRS
656 _InterlockedDecrement16(short volatile *_Value
) {
657 return __atomic_sub_fetch(_Value
, 1, __ATOMIC_SEQ_CST
);
660 static __inline__ __int64 __DEFAULT_FN_ATTRS
661 _InterlockedDecrement64(__int64
volatile *_Value
) {
662 return __atomic_sub_fetch(_Value
, 1, __ATOMIC_SEQ_CST
);
665 /*----------------------------------------------------------------------------*\
667 \*----------------------------------------------------------------------------*/
668 static __inline__
char __DEFAULT_FN_ATTRS
669 _InterlockedAnd8(char volatile *_Value
, char _Mask
) {
670 return __atomic_and_fetch(_Value
, _Mask
, __ATOMIC_SEQ_CST
);
672 static __inline__
short __DEFAULT_FN_ATTRS
673 _InterlockedAnd16(short volatile *_Value
, short _Mask
) {
674 return __atomic_and_fetch(_Value
, _Mask
, __ATOMIC_SEQ_CST
);
676 static __inline__
long __DEFAULT_FN_ATTRS
677 _InterlockedAnd(long volatile *_Value
, long _Mask
) {
678 return __atomic_and_fetch(_Value
, _Mask
, __ATOMIC_SEQ_CST
);
681 static __inline__ __int64 __DEFAULT_FN_ATTRS
682 _InterlockedAnd64(__int64
volatile *_Value
, __int64 _Mask
) {
683 return __atomic_and_fetch(_Value
, _Mask
, __ATOMIC_SEQ_CST
);
686 /*----------------------------------------------------------------------------*\
688 \*----------------------------------------------------------------------------*/
689 static __inline__
char __DEFAULT_FN_ATTRS
690 _InterlockedOr8(char volatile *_Value
, char _Mask
) {
691 return __atomic_or_fetch(_Value
, _Mask
, __ATOMIC_SEQ_CST
);
693 static __inline__
short __DEFAULT_FN_ATTRS
694 _InterlockedOr16(short volatile *_Value
, short _Mask
) {
695 return __atomic_or_fetch(_Value
, _Mask
, __ATOMIC_SEQ_CST
);
697 static __inline__
long __DEFAULT_FN_ATTRS
698 _InterlockedOr(long volatile *_Value
, long _Mask
) {
699 return __atomic_or_fetch(_Value
, _Mask
, __ATOMIC_SEQ_CST
);
702 static __inline__ __int64 __DEFAULT_FN_ATTRS
703 _InterlockedOr64(__int64
volatile *_Value
, __int64 _Mask
) {
704 return __atomic_or_fetch(_Value
, _Mask
, __ATOMIC_SEQ_CST
);
707 /*----------------------------------------------------------------------------*\
709 \*----------------------------------------------------------------------------*/
710 static __inline__
char __DEFAULT_FN_ATTRS
711 _InterlockedXor8(char volatile *_Value
, char _Mask
) {
712 return __atomic_xor_fetch(_Value
, _Mask
, __ATOMIC_SEQ_CST
);
714 static __inline__
short __DEFAULT_FN_ATTRS
715 _InterlockedXor16(short volatile *_Value
, short _Mask
) {
716 return __atomic_xor_fetch(_Value
, _Mask
, __ATOMIC_SEQ_CST
);
718 static __inline__
long __DEFAULT_FN_ATTRS
719 _InterlockedXor(long volatile *_Value
, long _Mask
) {
720 return __atomic_xor_fetch(_Value
, _Mask
, __ATOMIC_SEQ_CST
);
723 static __inline__ __int64 __DEFAULT_FN_ATTRS
724 _InterlockedXor64(__int64
volatile *_Value
, __int64 _Mask
) {
725 return __atomic_xor_fetch(_Value
, _Mask
, __ATOMIC_SEQ_CST
);
728 /*----------------------------------------------------------------------------*\
729 |* Interlocked Exchange
730 \*----------------------------------------------------------------------------*/
731 static __inline__
char __DEFAULT_FN_ATTRS
732 _InterlockedExchange8(char volatile *_Target
, char _Value
) {
733 __atomic_exchange(_Target
, &_Value
, &_Value
, __ATOMIC_SEQ_CST
);
736 static __inline__
short __DEFAULT_FN_ATTRS
737 _InterlockedExchange16(short volatile *_Target
, short _Value
) {
738 __atomic_exchange(_Target
, &_Value
, &_Value
, __ATOMIC_SEQ_CST
);
742 static __inline__ __int64 __DEFAULT_FN_ATTRS
743 _InterlockedExchange64(__int64
volatile *_Target
, __int64 _Value
) {
744 __atomic_exchange(_Target
, &_Value
, &_Value
, __ATOMIC_SEQ_CST
);
748 /*----------------------------------------------------------------------------*\
749 |* Interlocked Compare Exchange
750 \*----------------------------------------------------------------------------*/
751 static __inline__
char __DEFAULT_FN_ATTRS
752 _InterlockedCompareExchange8(char volatile *_Destination
,
753 char _Exchange
, char _Comparand
) {
754 __atomic_compare_exchange(_Destination
, &_Comparand
, &_Exchange
, 0,
755 __ATOMIC_SEQ_CST
, __ATOMIC_SEQ_CST
);
758 static __inline__
short __DEFAULT_FN_ATTRS
759 _InterlockedCompareExchange16(short volatile *_Destination
,
760 short _Exchange
, short _Comparand
) {
761 __atomic_compare_exchange(_Destination
, &_Comparand
, &_Exchange
, 0,
762 __ATOMIC_SEQ_CST
, __ATOMIC_SEQ_CST
);
765 static __inline__ __int64 __DEFAULT_FN_ATTRS
766 _InterlockedCompareExchange64(__int64
volatile *_Destination
,
767 __int64 _Exchange
, __int64 _Comparand
) {
768 __atomic_compare_exchange(_Destination
, &_Comparand
, &_Exchange
, 0,
769 __ATOMIC_SEQ_CST
, __ATOMIC_SEQ_CST
);
772 /*----------------------------------------------------------------------------*\
774 \*----------------------------------------------------------------------------*/
775 static __inline__
void __DEFAULT_FN_ATTRS
776 __attribute__((__deprecated__("use other intrinsics or C++11 atomics instead")))
777 _ReadWriteBarrier(void) {
778 __atomic_signal_fence(__ATOMIC_SEQ_CST
);
780 static __inline__
void __DEFAULT_FN_ATTRS
781 __attribute__((__deprecated__("use other intrinsics or C++11 atomics instead")))
783 __atomic_signal_fence(__ATOMIC_SEQ_CST
);
785 static __inline__
void __DEFAULT_FN_ATTRS
786 __attribute__((__deprecated__("use other intrinsics or C++11 atomics instead")))
787 _WriteBarrier(void) {
788 __atomic_signal_fence(__ATOMIC_SEQ_CST
);
791 static __inline__
void __DEFAULT_FN_ATTRS
792 __faststorefence(void) {
793 __atomic_thread_fence(__ATOMIC_SEQ_CST
);
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 movsw" : : "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 stosw" : : "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 */