diff options
| author | pvanhout <pierre.vanhoutryve@amd.com> | 2023-07-13 13:49:43 +0200 |
|---|---|---|
| committer | pvanhout <pierre.vanhoutryve@amd.com> | 2023-07-14 09:02:51 +0200 |
| commit | e5296c52e51bf214da32734d2344c9380c58a347 (patch) | |
| tree | bb26bd2cd1a5c4a8a291655c7d497ff6ca5bbbd6 /llvm/lib/Transforms/Utils/ModuleUtils.cpp | |
| parent | 3277aaae6feca4960b6546d20c17bcd796934a52 (diff) | |
| download | llvm-e5296c52e51bf214da32734d2344c9380c58a347.zip llvm-e5296c52e51bf214da32734d2344c9380c58a347.tar.gz llvm-e5296c52e51bf214da32734d2344c9380c58a347.tar.bz2 | |
[AMDGPU] Relax restrictions on unbreakable PHI users in BreakLargePHis
The previous heuristic rejected a PHI if one of its user was an unbreakable PHI, no matter what the other users were.
This worked well in most cases, but there's one case in rocRAND where
it doesn't work. In that case, a PHI node has 2 PHI users where one is
breakable but not the other. When that PHI node isn't broken performance falls by 35%.
Relaxing the restriction to "require that half of the PHI node users are breakable" fixes the issue, and seems like a sensible change.
Solves SWDEV-409648, SWDEV-398393
Reviewed By: #amdgpu, arsenm
Differential Revision: https://reviews.llvm.org/D155184
Diffstat (limited to 'llvm/lib/Transforms/Utils/ModuleUtils.cpp')
0 files changed, 0 insertions, 0 deletions