crash in llc with debug info

[efriedma-quic]

Bugzilla Link	2351
Resolution	INVALID
Resolved on	May 23, 2008 19:20
Version	unspecified
OS	Linux
Attachments	Testcase
CC	@asl

Extended Description

Per description, attached testcase crashes running llc over it. (The testcase was generated by clang, and reduced by bugpoint.)

I don't know if the attached bc file is valid; I haven't looked closely, and I know next to nothing about the debugging info format.

[llvmbot]

The debug info is not correct. There are no subprogram or func.start calls, and region.start and region.end should occur in pairs. See here for debug info description.
http://llvm.org/docs/SourceLevelDebugging.html#format_common_intrinsics
I suggest closing this, crashes are not good, but trying to crashproof llc against all possible combinations of bad (binary) IR input doesn't make sense to me.

[efriedma-quic]

I agree that it's not productive to try and crash-proof llc against t incorrect info. That said, would it be possible for the validator to do some level of validation on debug info? Or is debug info difficult to validate? Mysteriously crashing is a bad thing.

Also, how does llvm-gcc deal with jumping out of the middle of a block (i.e. break/continue/return/goto)? Does it emit a bunch of dbg.region.start/end calls before the jump?

[lattner]

That said, would it be possible for the validator to do some level of
validation on debug info?

This is a great idea, I would strongly support it.

Does it emit a bunch of dbg.region.start/end
calls before the jump?

I don't think llvm-gcc emits any region start/end markers.

[efriedma-quic]

I don't think llvm-gcc emits any region start/end markers.

Ah, I see. It does look a bit tricky to implement. Also, gotos could potentially cause an explosion in the number of region start/end markers without a clever implementation. The naive implementation of just emitting all the appropriate markers before the jump potentially requires emitting N*M region markers, where N is the number of gotos and M is the number of scopes. And I have no clue what the clever implementation would look like. It would be a bit easier if there were an intrinsic that could set the current region without having to worry about matching the start/end pairs.

I managed to fix clang to emit valid debug information, so this crash shouldn't show up from clang-generated code anymore.

[lattner]

Nice. Should this be closed then?

[efriedma-quic]

Okay, fine; the only interesting thing here is the enhancement request for a debug info validator.