[PATCH] D148523: [LegalizeTypes][VP] Widen arithmetic ops of fixed length vectors to VP nodes

Luke Lau via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Tue Jun 6 02:23:22 PDT 2023

luke added a comment.

Just a brief update from the last time I visited this patch:
Using VP nodes for unary/binary/ternary ops worked well, but unless we're able to use VP nodes for **all nodes** in a sequence we end up generating extra vsetvli toggles. 
In particular, I haven't been able to come up with a VP equivalent sequence for `build_vector` nodes, and I'm not sure how possible it will be.

After discussing offline with @reames, I think the issue of reducing the VL of arithmetic ops to match the smaller VLs of memory ops from D148713 <https://reviews.llvm.org/D148713> might be better handled in the vsetvli pass.

  rG LLVM Github Monorepo



More information about the llvm-commits mailing list