[PATCH] D107502: [WebAssembly] Legalize vector types by widening
Thomas Lively via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Tue Aug 17 20:59:43 PDT 2021
tlively added a comment.
In D107502#2927980 <https://reviews.llvm.org/D107502#2927980>, @aheejin wrote:
> Looks very nice! It looks it will have a significant impact on the performance. Are there possibly code patterns that will suffer from this change? If not, I wonder why was the default option in `TargetLowering` set up that way.. Maybe other architectures have more benefits when doing integer promotion? Anyway I feel I don't fully understand things yet so I added some questions.
Thanks! I haven't seen any code patterns that looks worse after this change, but it's possible it exposes new missed opportunities where we don't have patterns in place yet, like with the scalar_to_vector stuff. I'm not sure why the defaults are set up this way, either.
> And why were the two tests deleted?
I deleted the two regression tests because after this change the code paths that contained their relevant bugs are either no longer used at all (e.g. simd-nonconst-sext.ll) or have become well-tested by other tests (e.g. simd-scalar-to-vector.ll). In both cases, the regression tests no longer seemed useful to keep around.
Repository:
rG LLVM Github Monorepo
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D107502/new/
https://reviews.llvm.org/D107502
More information about the llvm-commits
mailing list