tsan: inline ProcessPendingSignals check

[dvyukov]

Summary

ProcessPendingSignals is called in all interceptors
and user atomic operations. Make the fast-path check
(no pending signals) inlinable.

compiler-rt/lib/tsan/rtl/tsan_interceptors_posix.cpp

struct SignalDesc {
   bool armed;
   __sanitizer_siginfo siginfo;
   ucontext_t ctx;
 };
 
 struct ThreadSignalContext {
   int int_signal_send;
   atomic_uintptr_t in_blocking_func;
-  atomic_uintptr_t have_pending_signals;
   SignalDesc pending_signals[kSigCount];
   // emptyset and oldset are too big for stack.
   __sanitizer_sigset_t emptyset;
   __sanitizer_sigset_t oldset;
 };
 
 // The sole reason tsan wraps atexit callbacks is to establish synchronization
 // between callback setup and callback execution.

 #define BLOCK_REAL(name) (BlockingCall(thr), REAL(name))
 
 struct BlockingCall {
   explicit BlockingCall(ThreadState *thr)
       : thr(thr)
       , ctx(SigCtx(thr)) {
     for (;;) {
       atomic_store(&ctx->in_blocking_func, 1, memory_order_relaxed);
-      if (atomic_load(&ctx->have_pending_signals, memory_order_relaxed) == 0)
+      if (atomic_load(&thr->pending_signals, memory_order_relaxed) == 0)
         break;
       atomic_store(&ctx->in_blocking_func, 0, memory_order_relaxed);
       ProcessPendingSignals(thr);
     }
     // When we are in a "blocking call", we process signals asynchronously
     // (right when they arrive). In this context we do not expect to be
     // executing any user/runtime code. The known interceptor sequence when
     // this is not true is: pthread_join -> munmap(stack). It's fine
if (ShouldReport(thr, ReportTypeErrnoInSignal) && !sync && sig != SIGTERM &&
     if (!IsFiredSuppression(ctx, ReportTypeErrnoInSignal, stack)) {
       rep.AddStack(stack, true);
       OutputReport(thr, rep);
     }
   }
   errno = saved_errno;
 }
 
-void ProcessPendingSignals(ThreadState *thr) {
+void ProcessPendingSignalsImpl(ThreadState *thr) {
+  atomic_store(&thr->pending_signals, 0, memory_order_relaxed);
   ThreadSignalContext *sctx = SigCtx(thr);
-  if (sctx == 0 ||
-      atomic_load(&sctx->have_pending_signals, memory_order_relaxed) == 0)
+  if (sctx == 0)
     return;
-  atomic_store(&sctx->have_pending_signals, 0, memory_order_relaxed);
   atomic_fetch_add(&thr->in_signal_handler, 1, memory_order_relaxed);
   internal_sigfillset(&sctx->emptyset);
   int res = REAL(pthread_sigmask)(SIG_SETMASK, &sctx->emptyset, &sctx->oldset);
   CHECK_EQ(res, 0);
   for (int sig = 0; sig < kSigCount; sig++) {
     SignalDesc *signal = &sctx->pending_signals[sig];
     if (signal->armed) {
       signal->armed = false;
void sighandler(int sig, __sanitizer_siginfo *info, void *ctx) {
 
   if (sctx == 0)
     return;
   SignalDesc *signal = &sctx->pending_signals[sig];
   if (signal->armed == false) {
     signal->armed = true;
     internal_memcpy(&signal->siginfo, info, sizeof(*info));
     internal_memcpy(&signal->ctx, ctx, sizeof(signal->ctx));
-    atomic_store(&sctx->have_pending_signals, 1, memory_order_relaxed);
+    atomic_store(&thr->pending_signals, 1, memory_order_relaxed);
   }
 }
 
 TSAN_INTERCEPTOR(int, raise, int sig) {
   SCOPED_TSAN_INTERCEPTOR(raise, sig);
   ThreadSignalContext *sctx = SigCtx(thr);
   CHECK_NE(sctx, 0);
   int prev = sctx->int_signal_send;

compiler-rt/lib/tsan/rtl/tsan_rtl.h

struct ThreadState {
   u64 fast_synch_epoch;
   // Technically `current` should be a separate THREADLOCAL variable;
   // but it is placed here in order to share cache line with previous fields.
   ThreadState* current;
   // This is a slow path flag. On fast path, fast_state.GetIgnoreBit() is read.
   // We do not distinguish beteween ignoring reads and writes
   // for better performance.
   int ignore_reads_and_writes;
+  atomic_sint32_t pending_signals;
   int ignore_sync;
   int suppress_reports;
   // Go does not support ignores.
 #if !SANITIZER_GO
   IgnoreSet mop_ignore_set;
   IgnoreSet sync_ignore_set;
 #endif
   // C/C++ uses fixed size shadow stack embed into Trace.
void ThreadStart(ThreadState *thr, Tid tid, tid_t os_id,
                  ThreadType thread_type);
 void ThreadFinish(ThreadState *thr);
 Tid ThreadConsumeTid(ThreadState *thr, uptr pc, uptr uid);
 void ThreadJoin(ThreadState *thr, uptr pc, Tid tid);
 void ThreadDetach(ThreadState *thr, uptr pc, Tid tid);
 void ThreadFinalize(ThreadState *thr);
 void ThreadSetName(ThreadState *thr, const char *name);
 int ThreadCount(ThreadState *thr);
-void ProcessPendingSignals(ThreadState *thr);
+void ProcessPendingSignalsImpl(ThreadState *thr);
 void ThreadNotJoined(ThreadState *thr, uptr pc, Tid tid, uptr uid);
 
 Processor *ProcCreate();
 void ProcDestroy(Processor *proc);
 void ProcWire(Processor *proc, ThreadState *thr);
 void ProcUnwire(Processor *proc, ThreadState *thr);
 
 // Note: the parameter is called flagz, because flags is already taken

 void FiberSwitch(ThreadState *thr, uptr pc, ThreadState *fiber, unsigned flags);
 
 // These need to match __tsan_switch_to_fiber_* flags defined in
 // tsan_interface.h. See documentation there as well.
 enum FiberSwitchFlags {
   FiberSwitchFlagNoSync = 1 << 0, // __tsan_switch_to_fiber_no_sync
 };
 
+ALWAYS_INLINE void ProcessPendingSignals(ThreadState *thr) {
+  if (UNLIKELY(atomic_load_relaxed(&thr->pending_signals)))
+    ProcessPendingSignalsImpl(thr);
+}
+
 extern bool is_initialized;
 
 ALWAYS_INLINE
 void LazyInitialize(ThreadState *thr) {
   // If we can use .preinit_array, assume that __tsan_init
   // called from .preinit_array initializes runtime before
   // any instrumented code.
 #if !SANITIZER_CAN_USE_PREINIT_ARRAY
   if (UNLIKELY(!is_initialized))
     Initialize(thr);
 #endif
 }
 
 }  // namespace __tsan
 
 #endif  // TSAN_RTL_H