[PATCH] D107859: [AMDGPU] MachineLICM cannot hoist VALU

Stanislav Mekhanoshin via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Fri Aug 27 10:42:55 PDT 2021


rampitec added a comment.

In D107859#2969153 <https://reviews.llvm.org/D107859#2969153>, @foad wrote:

> Are there any tests that show a benefit from this, not just churn?

Most of the changed tests show progressions. For sure most of the loop invariants are hoisted yet in the IR LICM pass, so MachineLICM only picks what was created during or after lowering and that is mostly service or index registers with constants.
For example in the dag-divergence-atomic.ll that is `v_mov_b32_e32 v0, 0` hoisted out of the loop at `BB5_1`, and similar changes in the rest of the tests.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D107859/new/

https://reviews.llvm.org/D107859



More information about the llvm-commits mailing list