[llvm] [RISCV][VLOPT] Allow users that are passthrus if tail elements aren't demanded (PR #124066)
Michael Maitland via llvm-commits
llvm-commits at lists.llvm.org
Thu Jan 23 11:58:37 PST 2025
michaelmaitland wrote:
> > Do we get any impact if we remove `MI.getNumDefs() != 1` in is candidate? Not sure if that would be related here.
>
> Off the top of my head, vsadd etc. have extra implicit defs, but I'm not sure if it's safe to change their VL. Fault-first loads have an extra explicit def, but I don't think they're safe to reduce either
> Fault-first loads have an extra explicit def, but I don't think they're safe to reduce either
I had played around with a patch in the past that changed this line to `MI.getNumExplicitDefs() != 1` to avoid the fault-first loads. `vsadd` and others who mess with vxsat are not part of `isSupportedInstr`. I think we could probably remove this line. It was placed here to guard against passthru. I think this would give you more optimization.
https://github.com/llvm/llvm-project/pull/124066
More information about the llvm-commits
mailing list