[clang] [llvm] [InstCombine] Relax singleUse for GEP(p,add(x,constant)) to 2 GEPs (PR #209551)

[Florian Hahn via llvm-commits]

https://github.com/fhahn commented:

> > As for the change itself, I expect that this is going to be a very high impact change, so this should get some performance evaluation in advance.
> 
> On internal tests, we see this does enable more loop vectorization opportunities, we also noticed SLP vectorizer choosing to not widen some references (due to its cost model) when this fold triggers on a vectorized loop. But it is improvement in general.

Could you also take a look at https://github.com/dtcxzyw/llvm-opt-benchmark-nightly/pull/918 and check for regressions? There are a lot of changes where only the GEP count increases which is somewhat expected, but maybe we could be more targeted?

https://github.com/llvm/llvm-project/pull/209551