[llvm] [NVPTX] Reland `mad.wide` combine under (default off) CLI option (PR #160214)
Justin Fargnoli via llvm-commits
llvm-commits at lists.llvm.org
Wed Sep 24 16:52:44 PDT 2025
justinfargnoli wrote:
> Do we have a good idea what exactly triggered the performance regression?
No, I haven't had the bandwidth to diagnose the issue properly.
---
> It appears to be "ptxas can't optimize loads that use mad.wide result as the source address".
Apologies if I'm being too pedantic, but I'd *guess* that `ptxas` *can* do this, but for whatever reason, it doesn't do it right now.
---
> we should make it a prominent FIXME/TODO next to the knob controlling it now.
Added a FIXME with 204513fb583b4778081e768bbaf0ea6872303821.
https://github.com/llvm/llvm-project/pull/160214
More information about the llvm-commits
mailing list