[PATCH] D16083: [bugpoint] Teach bugpoint to reduce MDNodes

[Duncan P. N. Exon Smith via llvm-commits]

> On 2016-Jan-11, at 15:06, Keno Fischer <kfischer at college.harvard.edu> wrote:
> 
> Does bugpoint look for any crash or does it look at the backtrace?
> 
> It only looks for a crash of any kind.
>  
>  
> >
> > I think we'd have more luck with something that understood the
> > structure of the graph, at least at a high-level.
> >
> > I've heard Adrian and Pete's ideas on that in the past, maybe they
> > can share here?
> 
> It is true that blindly hacking at MDNodes can leave us with metadata that satisfies the verifier but is corrupt enough to crash the backend. This means that the reduced testcase will not be very useful once the bug has been fixed. That said, I’m leaning towards that this should be an incentive to improve the verifier. We can still add smarter simplifications that run before this, such as replacing type definitions with forward declarations and the like
> 
> It sounds like we're essentially all in agreement then. If we're worried about introducing corruptions, how about we guard this reduction behind a flag for now (disabled by default), such that those who know about the caveats can start playing with it and filing any issues for cases that need to be added to the Verifier. 

A flag is a good idea, but I'd rather have it on-by-default so that we
catch more verifier bugs.