[PATCH] D45878: [DEBUG INFO] Fixing cases where debug info (-g) causes changes in the program.

[Vedant Kumar via Phabricator via llvm-commits]

vsk added a subscriber: gramanas.
vsk added a comment.

In https://reviews.llvm.org/D45878#1083950, @probinson wrote:

> In https://reviews.llvm.org/D45878#1083482, @davide wrote:
>
> > FWIW, I hacked up a fuzzer which found already around 20 (possibly with duplicate) changes in generated code due to `-g` being passed around. (cc: @zhendongsu)
>
>
> There are lots of minor cases.  The .cfi_* directives being a scheduling barrier creates a lot of noise in such testing (PR37240).
>
> > It was prodded by the bug Geoff reported on the SLPVectorizer which has been fixed on Friday.
> >  I can try to reduce the issues reported and also try to fix them.
>
> @davide can you coordinate with @vsk and the GSOC intern on these?


+ @gramanas. Anastasis is currently working on getting up to speed with LLVM IR and debug info handling in the SROA pass. We can certainly explore tackling some of these issues as a part of his project. I've asked @jonpa to re-send his notes to llvm-dev so we can continue the discussion there.


https://reviews.llvm.org/D45878