report_fatal_error no longer generates preprocessed source + reproducer.sh

[arichardson]

Bugzilla Link	44705
Resolution	FIXED
Resolved on	Feb 12, 2020 07:23
Version	trunk
OS	All
Blocks	#43900
CC	@efriedma-quic,@zmodem,@nico,@zygoloid,@rnk

Extended Description

After the integrated-cc1 report_fatal_error (e.g. machine verifier problems) no longer generated a preprocessed source + reproducer.sh.

In our out-of-tree backend (CHERI) we often run into backend errors and I noticed that we no longer get crash reproducers after I merged to latest upstream HEAD.

I can obviously re-run the command with -fno-integrated-cc1 but it would be nice if -fintegrated-cc1 would also generated the crash reproducers.

[arichardson]

assigned to @nico

[llvmbot]

Thanks for reporting this, I can take a look. This seems to be caused by llvm::report_fatal_error() returning exit(1) in ErrorHandling.cpp.

[rnk]

There is an open patch to make it call abort():
https://reviews.llvm.org/D67847

I think that would solve the problem.

I've forgotten if or why it was blocked, I thought everyone was in favor.

[llvmbot]

I think it would still need something along the lines of RaiseException() in place of exit/abort when integrated-cc1 is on, so that control flow goes back to the CrashRecoveryContext. And it would be even nicer if we had a function sys::ExitProcess() instead of directly calling exit/abort, and handle all the gory stuff in there :)
Quick prototype:

[[noreturn]] void sys::Process::ExitProcess(int retCode) {
if (CrashRecoveryContext::GetCurrent())
#if LLVM_ON_WIN32
RaiseException(0xE0000000 | code, 0, 0, NULL);
#else
CrashRecoveryContext::HandleCrash();
#endif
exit(code);
}

[llvmbot]

Sorry the for the delay. Please see https://reviews.llvm.org/D73742

[llvmbot]

Fixed by https://reviews.llvm.org/rGfaace365088a2a3a4cb1050a9facfc34a7a56577

[slacka]

Reopening for 10.0 backport.

faace36

[zmodem]

Reopening for 10.0 backport.

https://github.com/llvm/llvm-project/commit/
faace36

Pushed to 10.x as fd04cb4