[compiler-rt] [TSan] Fix deadlocks during TSan error reporting on Apple platforms (PR #151495)

[Thurston Dang via llvm-commits]

================
@@ -2141,13 +2142,23 @@ static void ReportErrnoSpoiling(ThreadState *thr, uptr pc, int sig) {
   // StackTrace::GetNestInstructionPc(pc) is used because return address is
   // expected, OutputReport() will undo this.
   ObtainCurrentStack(thr, StackTrace::GetNextInstructionPc(pc), &stack);
-  ThreadRegistryLock l(&ctx->thread_registry);
-  ScopedReport rep(ReportTypeErrnoInSignal);
-  rep.SetSigNum(sig);
-  if (!IsFiredSuppression(ctx, ReportTypeErrnoInSignal, stack)) {
-    rep.AddStack(stack, true);
-    OutputReport(thr, rep);
-  }
+  ScopedReport *_rep = (ScopedReport *)__builtin_alloca(sizeof(ScopedReport));
+  // Take a new scope as Apple platforms require the below locks released
+  // before symbolizing in order to avoid a deadlock
+  {
+    ThreadRegistryLock l(&ctx->thread_registry);
+    new (_rep) ScopedReport(ReportTypeErrnoInSignal);
+    ScopedReport &rep = *_rep;
+    rep.SetSigNum(sig);
+    if (!IsFiredSuppression(ctx, ReportTypeErrnoInSignal, stack)) {
+      rep.AddStack(stack, true);
+      OutputReport(thr, rep);
+    }
+  }  // Close this scope to release the locks
+
+  OutputReport(thr, *_rep);
+  // Need to manually destroy this because we used placement new to allocate
----------------
thurstond wrote:

I don't know of any reason not to use unique_ptr. If the tests have it, that implies the sanitizers assume a libc++ with unique_ptr support.

Any thoughts/objections from other reviewers?

https://github.com/llvm/llvm-project/pull/151495