arsenm added a comment. In D139627#3981475 <https://reviews.llvm.org/D139627#3981475>, @pengfei wrote:
> The use of `min-legal-vector-width` doesn't look great to me either. I'm more > than glad if we can remove it totally without any user perceivable affects. > I cannot agree with this change because it neither eliminates the clutter > (but makes it even worse [1]) nor is NFC to end user. > I think we used `UINT32_MAX` just to be compatible with BCs that generated > before introduing the attribute. This change definitely breaks the > compatibility. What I'm getting is this is only a performance hint, and definitively doesn't matter for ABI purposes. Bitcode backwards performance compatibility is not important. That would also be recovered by having a proper attribute propagation done as an optimization. I think all of clang's handling of this should be purged, except for the part where it's passing through the user attribute. > Placing a `"min-legal-vector-width" = "512"` doesn't make any sense either. > For one thing, we cannot place the attribute in pre-built BC files, for > another `512` is the current max vector suppoted on X86, we cannot guarantee > no `1024`, `2048` etc. in future and we cannot change it too once compiled > into BC files. It's a test for specific behavior, with a specific configuration that exists today. It doesn't matter what this would be in the future for larger testcases > [1] `"min-legal-vector-width" = "0"` was clear to indicate there are only > scalar operations. It's not remotely clear what this means CHANGES SINCE LAST ACTION https://reviews.llvm.org/D139627/new/ https://reviews.llvm.org/D139627 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
