[llvm] [llvm-mca] Abort on parse error without -skip-unsupported-instructions (PR #90474)

[Andrea Di Biagio via llvm-commits]

adibiagio wrote:

> > I understand skipping instructions that lack scheduling information but do we really mean to allow instructions that don’t parse?
> 
> One example use case of this is in the exynos test modified by this PR. That one has instructions in the test case which are not supported (edit: for some variations of -mcpu tested). Previously, MCA would happily continue to report something anyway. Now that I've changed this not to be default, it's necessary to be able to skip them again to make the test work.

The fact that previously MCA still generates a report after a parse error is unfortunate.

Ideally, llvm-mca shouldn't skip instructions that don't parse.
I prefer if this flag is only used for skipping instructions that don't have scheduling information.

> 
> Generally I think this is useful to have. You could find yourself with an assembly you want to analyse which has a rare instruction in it which hasn't yet had LLVM support added, for example if you're using an old version of LLVM at some point in the future and you want to see what MCA would have said if it were not for the (edit: not yet supported at this version) unsupported instructions.

Maybe we should add another flag for this particular use case?



https://github.com/llvm/llvm-project/pull/90474