[CrashRecovery] Don't treat OutputDebugString as a crash

[zturner]

Summary

After spending 2 days tracking down a bizarre silent program termination when using the SetupApi code path for locating the MSVC toolchain, we finally traced it down to this. Internal to the SetupApi com library, someone is calling OutputDebugString, which raises an exception so that the debugger can handle.

Honestly, we probably should not be using AddVectoredExceptionHandler in the first place. There are too many cases where RaiseException is used on purpose, and a library intends to catch it itself. By using AddVectoredExceptionHandler, we break all of these applications.

Regardless, this makes things better right now, and perhaps in the future we can look into using SetUnhandledExceptionFilter.

[zturner]

Also add the exception codes for printing a wide debugger string and setting a thread name.

[rnk]

I went ahead and did the "right" thing of using SEH instead: https://reviews.llvm.org/D33261

Unfortunately, mingw doesn't have SEH, so maybe we need this patch too. :(

I guess let's land it and I'll un-disable the unittest I wrote for this?

[rnk]

lgtm

llvm/trunk/lib/Support/CrashRecoveryContext.cpp

 // it will have to be added.
 //
 // If we want to support as far back as Win2k, we could use the
 // SetUnhandledExceptionFilter API, but there's a risk of that
 // being entirely overwritten (it's not a chain).
 
 static LONG CALLBACK ExceptionHandler(PEXCEPTION_POINTERS ExceptionInfo)
 {
+  switch (ExceptionInfo->ExceptionRecord->ExceptionCode)
+  {
+  case DBG_PRINTEXCEPTION_C:
+  case DBG_PRINTEXCEPTION_WIDE_C:
+  case 0x406D1388:  // set debugger thread name
+    return EXCEPTION_CONTINUE_EXECUTION;
+  }
+
   // Lookup the current thread local recovery object.
   const CrashRecoveryContextImpl *CRCI = CurrentContext->get();
 
   if (!CRCI) {
     // Something has gone horribly wrong, so let's just tell everyone
     // to keep searching
     CrashRecoveryContext::Disable();
     return EXCEPTION_CONTINUE_SEARCH;