[llvm] [RISCV] Check only demanded VTYPE fields in needVSETVLIPHI (PR #90168)

Luke Lau via llvm-commits llvm-commits at lists.llvm.org
Tue May 21 23:54:31 PDT 2024


lukel97 wrote:

> This is unsound because the CurInfo state during rewriting will differ from the CurInfo computed during the forward data flow. The first instruction may not care, but later ones in the block (or following blocks!) may depend on the other fields.
> 
> There is a correct variant of this, but you end of with different states flowing through the block during emission, and the need to compensate. Not sure if it's worth the complexity?

Woops, agreeed I don't think its worth the complexity. The original motivation for this was to split out a change from #89089 needed to avoid a regression, but since RISCVInsertVSETVLI has changed plenty in the meantime I'll check if this is still actually needed first.

https://github.com/llvm/llvm-project/pull/90168


More information about the llvm-commits mailing list