[llvm-branch-commits] [clang] [llvm] [GVN] Simple GVN-based hoisting of scalars (PR #210337)

[Antonio Frighetto via llvm-branch-commits]

https://github.com/antoniofrighetto commented:

> RFC/discussion: https://lists.llvm.org/pipermail/llvm-dev/2021-September/152665.html

It seems like the motivating example for this stemmed from an unvectorized loop in SPEC. Does LLVM still fail at vectorizing that code? Would this still be beneficial? (Wondering if the SPEC numbers still hold?)

> All the GVNs are confusing to me. How is this different from GVNHoist?

I share somewhat the same concern. I don't think placing this in GVN is necessarily unprincipled (PRE performs inherently some hoisting as well), but I wonder at this point if we should take the opportunity to acknowledge the current status of GVNHoist (off by default, unmaintained) and discontinue it? One may reasonably find it confusing to discover that both GVN and GVNHoist would perform VN-based hoisting (unless the intent is for them to coexist).

https://github.com/llvm/llvm-project/pull/210337