[llvm] [AMDGPU] Sink a single-use fmul into the block of its fadd/fsub user (PR #215810)

Dmitry Sidorov via llvm-commits llvm-commits at lists.llvm.org
Wed Aug 19 11:49:36 PDT 2026


MrSidims wrote:

Friendly ping. Please help finding reviewers if you believe, that you are not the correct one.

This patch is first in line of PRs in AMDGPU TTI to help slp vectorizer not to vectorize code which is more profitable to leave scalar and make sure, that fmul and fadd are fused.

For the contexts, there are several benchmarks along HeCBench and llama.cpp bench that perform better with slp-vectorizer not enabled (will provide a table once machine with perf data is back online) and I would like to address this perf difference fixing few things in the vectorizer as well as TTI.

https://github.com/llvm/llvm-project/pull/215810


More information about the llvm-commits mailing list