[PATCH] D155868: [AMDGPU] Add patterns for v_dot*_IU for GFX11

Jeffrey Byrnes via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Thu Jul 20 11:13:08 PDT 2023

jrbyrnes added a comment.

In D155868#4519802 <https://reviews.llvm.org/D155868#4519802>, @arsenm wrote:

> In D155868#4519799 <https://reviews.llvm.org/D155868#4519799>, @arsenm wrote:
>> We used to pattern match all the dot operations, but stopped because of a ridiculous blow up in compile time. Have you tried measuring that?
> Also look at the generated selection tables. This shouldn't be one of the first patterns tried

I'll take a look.

I have not directly measured compile time of this patch, but I will. I was thinking of deleting these patterns as they are supported by DAGCombining -- which shouldn't be too expensive with early exits on missed V_MUL_*24 operands. Worth mentioning is that there is also a huge compile time cost for not selecting into these instructions when we should -- we are seeing kernels getting stuck in RA for hours due to code bloat.

  rG LLVM Github Monorepo



More information about the llvm-commits mailing list