xref: /freebsd/contrib/llvm-project/compiler-rt/lib/tsan/rtl/tsan_interface_atomic.cpp (revision 7fdf597e96a02165cfe22ff357b857d5fa15ed8a)
1 //===-- tsan_interface_atomic.cpp -----------------------------------------===//
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 // This file is a part of ThreadSanitizer (TSan), a race detector.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 // ThreadSanitizer atomic operations are based on C++11/C1x standards.
14 // For background see C++11 standard.  A slightly older, publicly
15 // available draft of the standard (not entirely up-to-date, but close enough
16 // for casual browsing) is available here:
17 // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf
18 // The following page contains more background information:
19 // http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/
20 
21 #include "sanitizer_common/sanitizer_placement_new.h"
22 #include "sanitizer_common/sanitizer_stacktrace.h"
23 #include "sanitizer_common/sanitizer_mutex.h"
24 #include "tsan_flags.h"
25 #include "tsan_interface.h"
26 #include "tsan_rtl.h"
27 
28 using namespace __tsan;
29 
30 #if !SANITIZER_GO && __TSAN_HAS_INT128
31 // Protects emulation of 128-bit atomic operations.
32 static StaticSpinMutex mutex128;
33 #endif
34 
35 #if SANITIZER_DEBUG
36 static bool IsLoadOrder(morder mo) {
37   return mo == mo_relaxed || mo == mo_consume
38       || mo == mo_acquire || mo == mo_seq_cst;
39 }
40 
41 static bool IsStoreOrder(morder mo) {
42   return mo == mo_relaxed || mo == mo_release || mo == mo_seq_cst;
43 }
44 #endif
45 
46 static bool IsReleaseOrder(morder mo) {
47   return mo == mo_release || mo == mo_acq_rel || mo == mo_seq_cst;
48 }
49 
50 static bool IsAcquireOrder(morder mo) {
51   return mo == mo_consume || mo == mo_acquire
52       || mo == mo_acq_rel || mo == mo_seq_cst;
53 }
54 
55 static bool IsAcqRelOrder(morder mo) {
56   return mo == mo_acq_rel || mo == mo_seq_cst;
57 }
58 
59 template<typename T> T func_xchg(volatile T *v, T op) {
60   T res = __sync_lock_test_and_set(v, op);
61   // __sync_lock_test_and_set does not contain full barrier.
62   __sync_synchronize();
63   return res;
64 }
65 
66 template<typename T> T func_add(volatile T *v, T op) {
67   return __sync_fetch_and_add(v, op);
68 }
69 
70 template<typename T> T func_sub(volatile T *v, T op) {
71   return __sync_fetch_and_sub(v, op);
72 }
73 
74 template<typename T> T func_and(volatile T *v, T op) {
75   return __sync_fetch_and_and(v, op);
76 }
77 
78 template<typename T> T func_or(volatile T *v, T op) {
79   return __sync_fetch_and_or(v, op);
80 }
81 
82 template<typename T> T func_xor(volatile T *v, T op) {
83   return __sync_fetch_and_xor(v, op);
84 }
85 
86 template<typename T> T func_nand(volatile T *v, T op) {
87   // clang does not support __sync_fetch_and_nand.
88   T cmp = *v;
89   for (;;) {
90     T newv = ~(cmp & op);
91     T cur = __sync_val_compare_and_swap(v, cmp, newv);
92     if (cmp == cur)
93       return cmp;
94     cmp = cur;
95   }
96 }
97 
98 template<typename T> T func_cas(volatile T *v, T cmp, T xch) {
99   return __sync_val_compare_and_swap(v, cmp, xch);
100 }
101 
102 // clang does not support 128-bit atomic ops.
103 // Atomic ops are executed under tsan internal mutex,
104 // here we assume that the atomic variables are not accessed
105 // from non-instrumented code.
106 #if !defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_16) && !SANITIZER_GO \
107     && __TSAN_HAS_INT128
108 a128 func_xchg(volatile a128 *v, a128 op) {
109   SpinMutexLock lock(&mutex128);
110   a128 cmp = *v;
111   *v = op;
112   return cmp;
113 }
114 
115 a128 func_add(volatile a128 *v, a128 op) {
116   SpinMutexLock lock(&mutex128);
117   a128 cmp = *v;
118   *v = cmp + op;
119   return cmp;
120 }
121 
122 a128 func_sub(volatile a128 *v, a128 op) {
123   SpinMutexLock lock(&mutex128);
124   a128 cmp = *v;
125   *v = cmp - op;
126   return cmp;
127 }
128 
129 a128 func_and(volatile a128 *v, a128 op) {
130   SpinMutexLock lock(&mutex128);
131   a128 cmp = *v;
132   *v = cmp & op;
133   return cmp;
134 }
135 
136 a128 func_or(volatile a128 *v, a128 op) {
137   SpinMutexLock lock(&mutex128);
138   a128 cmp = *v;
139   *v = cmp | op;
140   return cmp;
141 }
142 
143 a128 func_xor(volatile a128 *v, a128 op) {
144   SpinMutexLock lock(&mutex128);
145   a128 cmp = *v;
146   *v = cmp ^ op;
147   return cmp;
148 }
149 
150 a128 func_nand(volatile a128 *v, a128 op) {
151   SpinMutexLock lock(&mutex128);
152   a128 cmp = *v;
153   *v = ~(cmp & op);
154   return cmp;
155 }
156 
157 a128 func_cas(volatile a128 *v, a128 cmp, a128 xch) {
158   SpinMutexLock lock(&mutex128);
159   a128 cur = *v;
160   if (cur == cmp)
161     *v = xch;
162   return cur;
163 }
164 #endif
165 
166 template <typename T>
167 static int AccessSize() {
168   if (sizeof(T) <= 1)
169     return 1;
170   else if (sizeof(T) <= 2)
171     return 2;
172   else if (sizeof(T) <= 4)
173     return 4;
174   else
175     return 8;
176   // For 16-byte atomics we also use 8-byte memory access,
177   // this leads to false negatives only in very obscure cases.
178 }
179 
180 #if !SANITIZER_GO
181 static atomic_uint8_t *to_atomic(const volatile a8 *a) {
182   return reinterpret_cast<atomic_uint8_t *>(const_cast<a8 *>(a));
183 }
184 
185 static atomic_uint16_t *to_atomic(const volatile a16 *a) {
186   return reinterpret_cast<atomic_uint16_t *>(const_cast<a16 *>(a));
187 }
188 #endif
189 
190 static atomic_uint32_t *to_atomic(const volatile a32 *a) {
191   return reinterpret_cast<atomic_uint32_t *>(const_cast<a32 *>(a));
192 }
193 
194 static atomic_uint64_t *to_atomic(const volatile a64 *a) {
195   return reinterpret_cast<atomic_uint64_t *>(const_cast<a64 *>(a));
196 }
197 
198 static memory_order to_mo(morder mo) {
199   switch (mo) {
200   case mo_relaxed: return memory_order_relaxed;
201   case mo_consume: return memory_order_consume;
202   case mo_acquire: return memory_order_acquire;
203   case mo_release: return memory_order_release;
204   case mo_acq_rel: return memory_order_acq_rel;
205   case mo_seq_cst: return memory_order_seq_cst;
206   }
207   DCHECK(0);
208   return memory_order_seq_cst;
209 }
210 
211 template<typename T>
212 static T NoTsanAtomicLoad(const volatile T *a, morder mo) {
213   return atomic_load(to_atomic(a), to_mo(mo));
214 }
215 
216 #if __TSAN_HAS_INT128 && !SANITIZER_GO
217 static a128 NoTsanAtomicLoad(const volatile a128 *a, morder mo) {
218   SpinMutexLock lock(&mutex128);
219   return *a;
220 }
221 #endif
222 
223 template <typename T>
224 static T AtomicLoad(ThreadState *thr, uptr pc, const volatile T *a, morder mo) {
225   DCHECK(IsLoadOrder(mo));
226   // This fast-path is critical for performance.
227   // Assume the access is atomic.
228   if (!IsAcquireOrder(mo)) {
229     MemoryAccess(thr, pc, (uptr)a, AccessSize<T>(),
230                  kAccessRead | kAccessAtomic);
231     return NoTsanAtomicLoad(a, mo);
232   }
233   // Don't create sync object if it does not exist yet. For example, an atomic
234   // pointer is initialized to nullptr and then periodically acquire-loaded.
235   T v = NoTsanAtomicLoad(a, mo);
236   SyncVar *s = ctx->metamap.GetSyncIfExists((uptr)a);
237   if (s) {
238     SlotLocker locker(thr);
239     ReadLock lock(&s->mtx);
240     thr->clock.Acquire(s->clock);
241     // Re-read under sync mutex because we need a consistent snapshot
242     // of the value and the clock we acquire.
243     v = NoTsanAtomicLoad(a, mo);
244   }
245   MemoryAccess(thr, pc, (uptr)a, AccessSize<T>(), kAccessRead | kAccessAtomic);
246   return v;
247 }
248 
249 template<typename T>
250 static void NoTsanAtomicStore(volatile T *a, T v, morder mo) {
251   atomic_store(to_atomic(a), v, to_mo(mo));
252 }
253 
254 #if __TSAN_HAS_INT128 && !SANITIZER_GO
255 static void NoTsanAtomicStore(volatile a128 *a, a128 v, morder mo) {
256   SpinMutexLock lock(&mutex128);
257   *a = v;
258 }
259 #endif
260 
261 template <typename T>
262 static void AtomicStore(ThreadState *thr, uptr pc, volatile T *a, T v,
263                         morder mo) {
264   DCHECK(IsStoreOrder(mo));
265   MemoryAccess(thr, pc, (uptr)a, AccessSize<T>(), kAccessWrite | kAccessAtomic);
266   // This fast-path is critical for performance.
267   // Assume the access is atomic.
268   // Strictly saying even relaxed store cuts off release sequence,
269   // so must reset the clock.
270   if (!IsReleaseOrder(mo)) {
271     NoTsanAtomicStore(a, v, mo);
272     return;
273   }
274   SlotLocker locker(thr);
275   {
276     auto s = ctx->metamap.GetSyncOrCreate(thr, pc, (uptr)a, false);
277     Lock lock(&s->mtx);
278     thr->clock.ReleaseStore(&s->clock);
279     NoTsanAtomicStore(a, v, mo);
280   }
281   IncrementEpoch(thr);
282 }
283 
284 template <typename T, T (*F)(volatile T *v, T op)>
285 static T AtomicRMW(ThreadState *thr, uptr pc, volatile T *a, T v, morder mo) {
286   MemoryAccess(thr, pc, (uptr)a, AccessSize<T>(), kAccessWrite | kAccessAtomic);
287   if (LIKELY(mo == mo_relaxed))
288     return F(a, v);
289   SlotLocker locker(thr);
290   {
291     auto s = ctx->metamap.GetSyncOrCreate(thr, pc, (uptr)a, false);
292     RWLock lock(&s->mtx, IsReleaseOrder(mo));
293     if (IsAcqRelOrder(mo))
294       thr->clock.ReleaseAcquire(&s->clock);
295     else if (IsReleaseOrder(mo))
296       thr->clock.Release(&s->clock);
297     else if (IsAcquireOrder(mo))
298       thr->clock.Acquire(s->clock);
299     v = F(a, v);
300   }
301   if (IsReleaseOrder(mo))
302     IncrementEpoch(thr);
303   return v;
304 }
305 
306 template<typename T>
307 static T NoTsanAtomicExchange(volatile T *a, T v, morder mo) {
308   return func_xchg(a, v);
309 }
310 
311 template<typename T>
312 static T NoTsanAtomicFetchAdd(volatile T *a, T v, morder mo) {
313   return func_add(a, v);
314 }
315 
316 template<typename T>
317 static T NoTsanAtomicFetchSub(volatile T *a, T v, morder mo) {
318   return func_sub(a, v);
319 }
320 
321 template<typename T>
322 static T NoTsanAtomicFetchAnd(volatile T *a, T v, morder mo) {
323   return func_and(a, v);
324 }
325 
326 template<typename T>
327 static T NoTsanAtomicFetchOr(volatile T *a, T v, morder mo) {
328   return func_or(a, v);
329 }
330 
331 template<typename T>
332 static T NoTsanAtomicFetchXor(volatile T *a, T v, morder mo) {
333   return func_xor(a, v);
334 }
335 
336 template<typename T>
337 static T NoTsanAtomicFetchNand(volatile T *a, T v, morder mo) {
338   return func_nand(a, v);
339 }
340 
341 template<typename T>
342 static T AtomicExchange(ThreadState *thr, uptr pc, volatile T *a, T v,
343     morder mo) {
344   return AtomicRMW<T, func_xchg>(thr, pc, a, v, mo);
345 }
346 
347 template<typename T>
348 static T AtomicFetchAdd(ThreadState *thr, uptr pc, volatile T *a, T v,
349     morder mo) {
350   return AtomicRMW<T, func_add>(thr, pc, a, v, mo);
351 }
352 
353 template<typename T>
354 static T AtomicFetchSub(ThreadState *thr, uptr pc, volatile T *a, T v,
355     morder mo) {
356   return AtomicRMW<T, func_sub>(thr, pc, a, v, mo);
357 }
358 
359 template<typename T>
360 static T AtomicFetchAnd(ThreadState *thr, uptr pc, volatile T *a, T v,
361     morder mo) {
362   return AtomicRMW<T, func_and>(thr, pc, a, v, mo);
363 }
364 
365 template<typename T>
366 static T AtomicFetchOr(ThreadState *thr, uptr pc, volatile T *a, T v,
367     morder mo) {
368   return AtomicRMW<T, func_or>(thr, pc, a, v, mo);
369 }
370 
371 template<typename T>
372 static T AtomicFetchXor(ThreadState *thr, uptr pc, volatile T *a, T v,
373     morder mo) {
374   return AtomicRMW<T, func_xor>(thr, pc, a, v, mo);
375 }
376 
377 template<typename T>
378 static T AtomicFetchNand(ThreadState *thr, uptr pc, volatile T *a, T v,
379     morder mo) {
380   return AtomicRMW<T, func_nand>(thr, pc, a, v, mo);
381 }
382 
383 template<typename T>
384 static bool NoTsanAtomicCAS(volatile T *a, T *c, T v, morder mo, morder fmo) {
385   return atomic_compare_exchange_strong(to_atomic(a), c, v, to_mo(mo));
386 }
387 
388 #if __TSAN_HAS_INT128
389 static bool NoTsanAtomicCAS(volatile a128 *a, a128 *c, a128 v,
390     morder mo, morder fmo) {
391   a128 old = *c;
392   a128 cur = func_cas(a, old, v);
393   if (cur == old)
394     return true;
395   *c = cur;
396   return false;
397 }
398 #endif
399 
400 template<typename T>
401 static T NoTsanAtomicCAS(volatile T *a, T c, T v, morder mo, morder fmo) {
402   NoTsanAtomicCAS(a, &c, v, mo, fmo);
403   return c;
404 }
405 
406 template <typename T>
407 static bool AtomicCAS(ThreadState *thr, uptr pc, volatile T *a, T *c, T v,
408                       morder mo, morder fmo) {
409   // 31.7.2.18: "The failure argument shall not be memory_order_release
410   // nor memory_order_acq_rel". LLVM (2021-05) fallbacks to Monotonic
411   // (mo_relaxed) when those are used.
412   DCHECK(IsLoadOrder(fmo));
413 
414   MemoryAccess(thr, pc, (uptr)a, AccessSize<T>(), kAccessWrite | kAccessAtomic);
415   if (LIKELY(mo == mo_relaxed && fmo == mo_relaxed)) {
416     T cc = *c;
417     T pr = func_cas(a, cc, v);
418     if (pr == cc)
419       return true;
420     *c = pr;
421     return false;
422   }
423   SlotLocker locker(thr);
424   bool release = IsReleaseOrder(mo);
425   bool success;
426   {
427     auto s = ctx->metamap.GetSyncOrCreate(thr, pc, (uptr)a, false);
428     RWLock lock(&s->mtx, release);
429     T cc = *c;
430     T pr = func_cas(a, cc, v);
431     success = pr == cc;
432     if (!success) {
433       *c = pr;
434       mo = fmo;
435     }
436     if (success && IsAcqRelOrder(mo))
437       thr->clock.ReleaseAcquire(&s->clock);
438     else if (success && IsReleaseOrder(mo))
439       thr->clock.Release(&s->clock);
440     else if (IsAcquireOrder(mo))
441       thr->clock.Acquire(s->clock);
442   }
443   if (success && release)
444     IncrementEpoch(thr);
445   return success;
446 }
447 
448 template<typename T>
449 static T AtomicCAS(ThreadState *thr, uptr pc,
450     volatile T *a, T c, T v, morder mo, morder fmo) {
451   AtomicCAS(thr, pc, a, &c, v, mo, fmo);
452   return c;
453 }
454 
455 #if !SANITIZER_GO
456 static void NoTsanAtomicFence(morder mo) {
457   __sync_synchronize();
458 }
459 
460 static void AtomicFence(ThreadState *thr, uptr pc, morder mo) {
461   // FIXME(dvyukov): not implemented.
462   __sync_synchronize();
463 }
464 #endif
465 
466 // Interface functions follow.
467 #if !SANITIZER_GO
468 
469 // C/C++
470 
471 static morder convert_morder(morder mo) {
472   if (flags()->force_seq_cst_atomics)
473     return (morder)mo_seq_cst;
474 
475   // Filter out additional memory order flags:
476   // MEMMODEL_SYNC        = 1 << 15
477   // __ATOMIC_HLE_ACQUIRE = 1 << 16
478   // __ATOMIC_HLE_RELEASE = 1 << 17
479   //
480   // HLE is an optimization, and we pretend that elision always fails.
481   // MEMMODEL_SYNC is used when lowering __sync_ atomics,
482   // since we use __sync_ atomics for actual atomic operations,
483   // we can safely ignore it as well. It also subtly affects semantics,
484   // but we don't model the difference.
485   return (morder)(mo & 0x7fff);
486 }
487 
488 #  define ATOMIC_IMPL(func, ...)                                \
489     ThreadState *const thr = cur_thread();                      \
490     ProcessPendingSignals(thr);                                 \
491     if (UNLIKELY(thr->ignore_sync || thr->ignore_interceptors)) \
492       return NoTsanAtomic##func(__VA_ARGS__);                   \
493     mo = convert_morder(mo);                                    \
494     return Atomic##func(thr, GET_CALLER_PC(), __VA_ARGS__);
495 
496 extern "C" {
497 SANITIZER_INTERFACE_ATTRIBUTE
498 a8 __tsan_atomic8_load(const volatile a8 *a, morder mo) {
499   ATOMIC_IMPL(Load, a, mo);
500 }
501 
502 SANITIZER_INTERFACE_ATTRIBUTE
503 a16 __tsan_atomic16_load(const volatile a16 *a, morder mo) {
504   ATOMIC_IMPL(Load, a, mo);
505 }
506 
507 SANITIZER_INTERFACE_ATTRIBUTE
508 a32 __tsan_atomic32_load(const volatile a32 *a, morder mo) {
509   ATOMIC_IMPL(Load, a, mo);
510 }
511 
512 SANITIZER_INTERFACE_ATTRIBUTE
513 a64 __tsan_atomic64_load(const volatile a64 *a, morder mo) {
514   ATOMIC_IMPL(Load, a, mo);
515 }
516 
517 #if __TSAN_HAS_INT128
518 SANITIZER_INTERFACE_ATTRIBUTE
519 a128 __tsan_atomic128_load(const volatile a128 *a, morder mo) {
520   ATOMIC_IMPL(Load, a, mo);
521 }
522 #endif
523 
524 SANITIZER_INTERFACE_ATTRIBUTE
525 void __tsan_atomic8_store(volatile a8 *a, a8 v, morder mo) {
526   ATOMIC_IMPL(Store, a, v, mo);
527 }
528 
529 SANITIZER_INTERFACE_ATTRIBUTE
530 void __tsan_atomic16_store(volatile a16 *a, a16 v, morder mo) {
531   ATOMIC_IMPL(Store, a, v, mo);
532 }
533 
534 SANITIZER_INTERFACE_ATTRIBUTE
535 void __tsan_atomic32_store(volatile a32 *a, a32 v, morder mo) {
536   ATOMIC_IMPL(Store, a, v, mo);
537 }
538 
539 SANITIZER_INTERFACE_ATTRIBUTE
540 void __tsan_atomic64_store(volatile a64 *a, a64 v, morder mo) {
541   ATOMIC_IMPL(Store, a, v, mo);
542 }
543 
544 #if __TSAN_HAS_INT128
545 SANITIZER_INTERFACE_ATTRIBUTE
546 void __tsan_atomic128_store(volatile a128 *a, a128 v, morder mo) {
547   ATOMIC_IMPL(Store, a, v, mo);
548 }
549 #endif
550 
551 SANITIZER_INTERFACE_ATTRIBUTE
552 a8 __tsan_atomic8_exchange(volatile a8 *a, a8 v, morder mo) {
553   ATOMIC_IMPL(Exchange, a, v, mo);
554 }
555 
556 SANITIZER_INTERFACE_ATTRIBUTE
557 a16 __tsan_atomic16_exchange(volatile a16 *a, a16 v, morder mo) {
558   ATOMIC_IMPL(Exchange, a, v, mo);
559 }
560 
561 SANITIZER_INTERFACE_ATTRIBUTE
562 a32 __tsan_atomic32_exchange(volatile a32 *a, a32 v, morder mo) {
563   ATOMIC_IMPL(Exchange, a, v, mo);
564 }
565 
566 SANITIZER_INTERFACE_ATTRIBUTE
567 a64 __tsan_atomic64_exchange(volatile a64 *a, a64 v, morder mo) {
568   ATOMIC_IMPL(Exchange, a, v, mo);
569 }
570 
571 #if __TSAN_HAS_INT128
572 SANITIZER_INTERFACE_ATTRIBUTE
573 a128 __tsan_atomic128_exchange(volatile a128 *a, a128 v, morder mo) {
574   ATOMIC_IMPL(Exchange, a, v, mo);
575 }
576 #endif
577 
578 SANITIZER_INTERFACE_ATTRIBUTE
579 a8 __tsan_atomic8_fetch_add(volatile a8 *a, a8 v, morder mo) {
580   ATOMIC_IMPL(FetchAdd, a, v, mo);
581 }
582 
583 SANITIZER_INTERFACE_ATTRIBUTE
584 a16 __tsan_atomic16_fetch_add(volatile a16 *a, a16 v, morder mo) {
585   ATOMIC_IMPL(FetchAdd, a, v, mo);
586 }
587 
588 SANITIZER_INTERFACE_ATTRIBUTE
589 a32 __tsan_atomic32_fetch_add(volatile a32 *a, a32 v, morder mo) {
590   ATOMIC_IMPL(FetchAdd, a, v, mo);
591 }
592 
593 SANITIZER_INTERFACE_ATTRIBUTE
594 a64 __tsan_atomic64_fetch_add(volatile a64 *a, a64 v, morder mo) {
595   ATOMIC_IMPL(FetchAdd, a, v, mo);
596 }
597 
598 #if __TSAN_HAS_INT128
599 SANITIZER_INTERFACE_ATTRIBUTE
600 a128 __tsan_atomic128_fetch_add(volatile a128 *a, a128 v, morder mo) {
601   ATOMIC_IMPL(FetchAdd, a, v, mo);
602 }
603 #endif
604 
605 SANITIZER_INTERFACE_ATTRIBUTE
606 a8 __tsan_atomic8_fetch_sub(volatile a8 *a, a8 v, morder mo) {
607   ATOMIC_IMPL(FetchSub, a, v, mo);
608 }
609 
610 SANITIZER_INTERFACE_ATTRIBUTE
611 a16 __tsan_atomic16_fetch_sub(volatile a16 *a, a16 v, morder mo) {
612   ATOMIC_IMPL(FetchSub, a, v, mo);
613 }
614 
615 SANITIZER_INTERFACE_ATTRIBUTE
616 a32 __tsan_atomic32_fetch_sub(volatile a32 *a, a32 v, morder mo) {
617   ATOMIC_IMPL(FetchSub, a, v, mo);
618 }
619 
620 SANITIZER_INTERFACE_ATTRIBUTE
621 a64 __tsan_atomic64_fetch_sub(volatile a64 *a, a64 v, morder mo) {
622   ATOMIC_IMPL(FetchSub, a, v, mo);
623 }
624 
625 #if __TSAN_HAS_INT128
626 SANITIZER_INTERFACE_ATTRIBUTE
627 a128 __tsan_atomic128_fetch_sub(volatile a128 *a, a128 v, morder mo) {
628   ATOMIC_IMPL(FetchSub, a, v, mo);
629 }
630 #endif
631 
632 SANITIZER_INTERFACE_ATTRIBUTE
633 a8 __tsan_atomic8_fetch_and(volatile a8 *a, a8 v, morder mo) {
634   ATOMIC_IMPL(FetchAnd, a, v, mo);
635 }
636 
637 SANITIZER_INTERFACE_ATTRIBUTE
638 a16 __tsan_atomic16_fetch_and(volatile a16 *a, a16 v, morder mo) {
639   ATOMIC_IMPL(FetchAnd, a, v, mo);
640 }
641 
642 SANITIZER_INTERFACE_ATTRIBUTE
643 a32 __tsan_atomic32_fetch_and(volatile a32 *a, a32 v, morder mo) {
644   ATOMIC_IMPL(FetchAnd, a, v, mo);
645 }
646 
647 SANITIZER_INTERFACE_ATTRIBUTE
648 a64 __tsan_atomic64_fetch_and(volatile a64 *a, a64 v, morder mo) {
649   ATOMIC_IMPL(FetchAnd, a, v, mo);
650 }
651 
652 #if __TSAN_HAS_INT128
653 SANITIZER_INTERFACE_ATTRIBUTE
654 a128 __tsan_atomic128_fetch_and(volatile a128 *a, a128 v, morder mo) {
655   ATOMIC_IMPL(FetchAnd, a, v, mo);
656 }
657 #endif
658 
659 SANITIZER_INTERFACE_ATTRIBUTE
660 a8 __tsan_atomic8_fetch_or(volatile a8 *a, a8 v, morder mo) {
661   ATOMIC_IMPL(FetchOr, a, v, mo);
662 }
663 
664 SANITIZER_INTERFACE_ATTRIBUTE
665 a16 __tsan_atomic16_fetch_or(volatile a16 *a, a16 v, morder mo) {
666   ATOMIC_IMPL(FetchOr, a, v, mo);
667 }
668 
669 SANITIZER_INTERFACE_ATTRIBUTE
670 a32 __tsan_atomic32_fetch_or(volatile a32 *a, a32 v, morder mo) {
671   ATOMIC_IMPL(FetchOr, a, v, mo);
672 }
673 
674 SANITIZER_INTERFACE_ATTRIBUTE
675 a64 __tsan_atomic64_fetch_or(volatile a64 *a, a64 v, morder mo) {
676   ATOMIC_IMPL(FetchOr, a, v, mo);
677 }
678 
679 #if __TSAN_HAS_INT128
680 SANITIZER_INTERFACE_ATTRIBUTE
681 a128 __tsan_atomic128_fetch_or(volatile a128 *a, a128 v, morder mo) {
682   ATOMIC_IMPL(FetchOr, a, v, mo);
683 }
684 #endif
685 
686 SANITIZER_INTERFACE_ATTRIBUTE
687 a8 __tsan_atomic8_fetch_xor(volatile a8 *a, a8 v, morder mo) {
688   ATOMIC_IMPL(FetchXor, a, v, mo);
689 }
690 
691 SANITIZER_INTERFACE_ATTRIBUTE
692 a16 __tsan_atomic16_fetch_xor(volatile a16 *a, a16 v, morder mo) {
693   ATOMIC_IMPL(FetchXor, a, v, mo);
694 }
695 
696 SANITIZER_INTERFACE_ATTRIBUTE
697 a32 __tsan_atomic32_fetch_xor(volatile a32 *a, a32 v, morder mo) {
698   ATOMIC_IMPL(FetchXor, a, v, mo);
699 }
700 
701 SANITIZER_INTERFACE_ATTRIBUTE
702 a64 __tsan_atomic64_fetch_xor(volatile a64 *a, a64 v, morder mo) {
703   ATOMIC_IMPL(FetchXor, a, v, mo);
704 }
705 
706 #if __TSAN_HAS_INT128
707 SANITIZER_INTERFACE_ATTRIBUTE
708 a128 __tsan_atomic128_fetch_xor(volatile a128 *a, a128 v, morder mo) {
709   ATOMIC_IMPL(FetchXor, a, v, mo);
710 }
711 #endif
712 
713 SANITIZER_INTERFACE_ATTRIBUTE
714 a8 __tsan_atomic8_fetch_nand(volatile a8 *a, a8 v, morder mo) {
715   ATOMIC_IMPL(FetchNand, a, v, mo);
716 }
717 
718 SANITIZER_INTERFACE_ATTRIBUTE
719 a16 __tsan_atomic16_fetch_nand(volatile a16 *a, a16 v, morder mo) {
720   ATOMIC_IMPL(FetchNand, a, v, mo);
721 }
722 
723 SANITIZER_INTERFACE_ATTRIBUTE
724 a32 __tsan_atomic32_fetch_nand(volatile a32 *a, a32 v, morder mo) {
725   ATOMIC_IMPL(FetchNand, a, v, mo);
726 }
727 
728 SANITIZER_INTERFACE_ATTRIBUTE
729 a64 __tsan_atomic64_fetch_nand(volatile a64 *a, a64 v, morder mo) {
730   ATOMIC_IMPL(FetchNand, a, v, mo);
731 }
732 
733 #if __TSAN_HAS_INT128
734 SANITIZER_INTERFACE_ATTRIBUTE
735 a128 __tsan_atomic128_fetch_nand(volatile a128 *a, a128 v, morder mo) {
736   ATOMIC_IMPL(FetchNand, a, v, mo);
737 }
738 #endif
739 
740 SANITIZER_INTERFACE_ATTRIBUTE
741 int __tsan_atomic8_compare_exchange_strong(volatile a8 *a, a8 *c, a8 v,
742     morder mo, morder fmo) {
743   ATOMIC_IMPL(CAS, a, c, v, mo, fmo);
744 }
745 
746 SANITIZER_INTERFACE_ATTRIBUTE
747 int __tsan_atomic16_compare_exchange_strong(volatile a16 *a, a16 *c, a16 v,
748     morder mo, morder fmo) {
749   ATOMIC_IMPL(CAS, a, c, v, mo, fmo);
750 }
751 
752 SANITIZER_INTERFACE_ATTRIBUTE
753 int __tsan_atomic32_compare_exchange_strong(volatile a32 *a, a32 *c, a32 v,
754     morder mo, morder fmo) {
755   ATOMIC_IMPL(CAS, a, c, v, mo, fmo);
756 }
757 
758 SANITIZER_INTERFACE_ATTRIBUTE
759 int __tsan_atomic64_compare_exchange_strong(volatile a64 *a, a64 *c, a64 v,
760     morder mo, morder fmo) {
761   ATOMIC_IMPL(CAS, a, c, v, mo, fmo);
762 }
763 
764 #if __TSAN_HAS_INT128
765 SANITIZER_INTERFACE_ATTRIBUTE
766 int __tsan_atomic128_compare_exchange_strong(volatile a128 *a, a128 *c, a128 v,
767     morder mo, morder fmo) {
768   ATOMIC_IMPL(CAS, a, c, v, mo, fmo);
769 }
770 #endif
771 
772 SANITIZER_INTERFACE_ATTRIBUTE
773 int __tsan_atomic8_compare_exchange_weak(volatile a8 *a, a8 *c, a8 v,
774     morder mo, morder fmo) {
775   ATOMIC_IMPL(CAS, a, c, v, mo, fmo);
776 }
777 
778 SANITIZER_INTERFACE_ATTRIBUTE
779 int __tsan_atomic16_compare_exchange_weak(volatile a16 *a, a16 *c, a16 v,
780     morder mo, morder fmo) {
781   ATOMIC_IMPL(CAS, a, c, v, mo, fmo);
782 }
783 
784 SANITIZER_INTERFACE_ATTRIBUTE
785 int __tsan_atomic32_compare_exchange_weak(volatile a32 *a, a32 *c, a32 v,
786     morder mo, morder fmo) {
787   ATOMIC_IMPL(CAS, a, c, v, mo, fmo);
788 }
789 
790 SANITIZER_INTERFACE_ATTRIBUTE
791 int __tsan_atomic64_compare_exchange_weak(volatile a64 *a, a64 *c, a64 v,
792     morder mo, morder fmo) {
793   ATOMIC_IMPL(CAS, a, c, v, mo, fmo);
794 }
795 
796 #if __TSAN_HAS_INT128
797 SANITIZER_INTERFACE_ATTRIBUTE
798 int __tsan_atomic128_compare_exchange_weak(volatile a128 *a, a128 *c, a128 v,
799     morder mo, morder fmo) {
800   ATOMIC_IMPL(CAS, a, c, v, mo, fmo);
801 }
802 #endif
803 
804 SANITIZER_INTERFACE_ATTRIBUTE
805 a8 __tsan_atomic8_compare_exchange_val(volatile a8 *a, a8 c, a8 v,
806     morder mo, morder fmo) {
807   ATOMIC_IMPL(CAS, a, c, v, mo, fmo);
808 }
809 
810 SANITIZER_INTERFACE_ATTRIBUTE
811 a16 __tsan_atomic16_compare_exchange_val(volatile a16 *a, a16 c, a16 v,
812     morder mo, morder fmo) {
813   ATOMIC_IMPL(CAS, a, c, v, mo, fmo);
814 }
815 
816 SANITIZER_INTERFACE_ATTRIBUTE
817 a32 __tsan_atomic32_compare_exchange_val(volatile a32 *a, a32 c, a32 v,
818     morder mo, morder fmo) {
819   ATOMIC_IMPL(CAS, a, c, v, mo, fmo);
820 }
821 
822 SANITIZER_INTERFACE_ATTRIBUTE
823 a64 __tsan_atomic64_compare_exchange_val(volatile a64 *a, a64 c, a64 v,
824     morder mo, morder fmo) {
825   ATOMIC_IMPL(CAS, a, c, v, mo, fmo);
826 }
827 
828 #if __TSAN_HAS_INT128
829 SANITIZER_INTERFACE_ATTRIBUTE
830 a128 __tsan_atomic128_compare_exchange_val(volatile a128 *a, a128 c, a128 v,
831     morder mo, morder fmo) {
832   ATOMIC_IMPL(CAS, a, c, v, mo, fmo);
833 }
834 #endif
835 
836 SANITIZER_INTERFACE_ATTRIBUTE
837 void __tsan_atomic_thread_fence(morder mo) { ATOMIC_IMPL(Fence, mo); }
838 
839 SANITIZER_INTERFACE_ATTRIBUTE
840 void __tsan_atomic_signal_fence(morder mo) {
841 }
842 }  // extern "C"
843 
844 #else  // #if !SANITIZER_GO
845 
846 // Go
847 
848 #  define ATOMIC(func, ...)               \
849     if (thr->ignore_sync) {               \
850       NoTsanAtomic##func(__VA_ARGS__);    \
851     } else {                              \
852       FuncEntry(thr, cpc);                \
853       Atomic##func(thr, pc, __VA_ARGS__); \
854       FuncExit(thr);                      \
855     }
856 
857 #  define ATOMIC_RET(func, ret, ...)              \
858     if (thr->ignore_sync) {                       \
859       (ret) = NoTsanAtomic##func(__VA_ARGS__);    \
860     } else {                                      \
861       FuncEntry(thr, cpc);                        \
862       (ret) = Atomic##func(thr, pc, __VA_ARGS__); \
863       FuncExit(thr);                              \
864     }
865 
866 extern "C" {
867 SANITIZER_INTERFACE_ATTRIBUTE
868 void __tsan_go_atomic32_load(ThreadState *thr, uptr cpc, uptr pc, u8 *a) {
869   ATOMIC_RET(Load, *(a32*)(a+8), *(a32**)a, mo_acquire);
870 }
871 
872 SANITIZER_INTERFACE_ATTRIBUTE
873 void __tsan_go_atomic64_load(ThreadState *thr, uptr cpc, uptr pc, u8 *a) {
874   ATOMIC_RET(Load, *(a64*)(a+8), *(a64**)a, mo_acquire);
875 }
876 
877 SANITIZER_INTERFACE_ATTRIBUTE
878 void __tsan_go_atomic32_store(ThreadState *thr, uptr cpc, uptr pc, u8 *a) {
879   ATOMIC(Store, *(a32**)a, *(a32*)(a+8), mo_release);
880 }
881 
882 SANITIZER_INTERFACE_ATTRIBUTE
883 void __tsan_go_atomic64_store(ThreadState *thr, uptr cpc, uptr pc, u8 *a) {
884   ATOMIC(Store, *(a64**)a, *(a64*)(a+8), mo_release);
885 }
886 
887 SANITIZER_INTERFACE_ATTRIBUTE
888 void __tsan_go_atomic32_fetch_add(ThreadState *thr, uptr cpc, uptr pc, u8 *a) {
889   ATOMIC_RET(FetchAdd, *(a32*)(a+16), *(a32**)a, *(a32*)(a+8), mo_acq_rel);
890 }
891 
892 SANITIZER_INTERFACE_ATTRIBUTE
893 void __tsan_go_atomic64_fetch_add(ThreadState *thr, uptr cpc, uptr pc, u8 *a) {
894   ATOMIC_RET(FetchAdd, *(a64*)(a+16), *(a64**)a, *(a64*)(a+8), mo_acq_rel);
895 }
896 
897 SANITIZER_INTERFACE_ATTRIBUTE
898 void __tsan_go_atomic32_fetch_and(ThreadState *thr, uptr cpc, uptr pc, u8 *a) {
899   ATOMIC_RET(FetchAnd, *(a32 *)(a + 16), *(a32 **)a, *(a32 *)(a + 8),
900              mo_acq_rel);
901 }
902 
903 SANITIZER_INTERFACE_ATTRIBUTE
904 void __tsan_go_atomic64_fetch_and(ThreadState *thr, uptr cpc, uptr pc, u8 *a) {
905   ATOMIC_RET(FetchAnd, *(a64 *)(a + 16), *(a64 **)a, *(a64 *)(a + 8),
906              mo_acq_rel);
907 }
908 
909 SANITIZER_INTERFACE_ATTRIBUTE
910 void __tsan_go_atomic32_fetch_or(ThreadState *thr, uptr cpc, uptr pc, u8 *a) {
911   ATOMIC_RET(FetchOr, *(a32 *)(a + 16), *(a32 **)a, *(a32 *)(a + 8),
912              mo_acq_rel);
913 }
914 
915 SANITIZER_INTERFACE_ATTRIBUTE
916 void __tsan_go_atomic64_fetch_or(ThreadState *thr, uptr cpc, uptr pc, u8 *a) {
917   ATOMIC_RET(FetchOr, *(a64 *)(a + 16), *(a64 **)a, *(a64 *)(a + 8),
918              mo_acq_rel);
919 }
920 
921 SANITIZER_INTERFACE_ATTRIBUTE
922 void __tsan_go_atomic32_exchange(ThreadState *thr, uptr cpc, uptr pc, u8 *a) {
923   ATOMIC_RET(Exchange, *(a32*)(a+16), *(a32**)a, *(a32*)(a+8), mo_acq_rel);
924 }
925 
926 SANITIZER_INTERFACE_ATTRIBUTE
927 void __tsan_go_atomic64_exchange(ThreadState *thr, uptr cpc, uptr pc, u8 *a) {
928   ATOMIC_RET(Exchange, *(a64*)(a+16), *(a64**)a, *(a64*)(a+8), mo_acq_rel);
929 }
930 
931 SANITIZER_INTERFACE_ATTRIBUTE
932 void __tsan_go_atomic32_compare_exchange(
933     ThreadState *thr, uptr cpc, uptr pc, u8 *a) {
934   a32 cur = 0;
935   a32 cmp = *(a32*)(a+8);
936   ATOMIC_RET(CAS, cur, *(a32**)a, cmp, *(a32*)(a+12), mo_acq_rel, mo_acquire);
937   *(bool*)(a+16) = (cur == cmp);
938 }
939 
940 SANITIZER_INTERFACE_ATTRIBUTE
941 void __tsan_go_atomic64_compare_exchange(
942     ThreadState *thr, uptr cpc, uptr pc, u8 *a) {
943   a64 cur = 0;
944   a64 cmp = *(a64*)(a+8);
945   ATOMIC_RET(CAS, cur, *(a64**)a, cmp, *(a64*)(a+16), mo_acq_rel, mo_acquire);
946   *(bool*)(a+24) = (cur == cmp);
947 }
948 }  // extern "C"
949 #endif  // #if !SANITIZER_GO
950