[all-commits] [llvm/llvm-project] d3a8ae: [AMDGPU][NFC] Harden scheduler tests balanced on p...
Dhruva Chakrabarti via All-commits
all-commits at lists.llvm.org
Thu Oct 1 18:00:20 PDT 2026
Branch: refs/heads/users/dhruvachak/harden_sched_tests
Home: https://github.com/llvm/llvm-project
Commit: d3a8ae8b43d5768f4c3f4b3f9b11e951826d2542
https://github.com/llvm/llvm-project/commit/d3a8ae8b43d5768f4c3f4b3f9b11e951826d2542
Author: Dhruva Chakrabarti <Dhruva.Chakrabarti at amd.com>
Date: 2026-09-30 (Wed, 30 Sep 2026)
Changed paths:
M llvm/test/CodeGen/AMDGPU/coexec-rewrite-mfma.ll
M llvm/test/CodeGen/AMDGPU/rewrite-vgpr-mfma-to-agpr-spill-multi-store-mir.mir
M llvm/test/CodeGen/AMDGPU/sched-dynamic-vgpr-blocks.mir
M llvm/test/CodeGen/AMDGPU/sema-v-unsched-bundle.ll
Log Message:
-----------
[AMDGPU][NFC] Harden scheduler tests balanced on pressure boundaries
Several AMDGPU scheduler tests could stop exercising the intended code
path for small changes in certain factors, e.g. register pressure, a
VGPR block, etc. Move the affected tests away from those boundaries.
- sema-v-unsched-bundle.ll kept just enough values live across the GWS
bundle to land on an appropriate number of VGPRs for the schedule to
be reverted. Another live vector ensures that unscheduling of bundled
instructions is covered with room to spare.
- sched-dynamic-vgpr-blocks.mir needed the region to grow from one
16-VGPR block to two. One more load/add pair ensures the growth.
- coexec-rewrite-mfma.ll needed ArchVGPR pressure above the addressable
limit for RewriteMFMAFormStage to run at all. Two loop-carried phis and
one more MFMA pair keep it there.
- rewrite-vgpr-mfma-to-agpr-spill-multi-store-mir relied on an
accumulator set that only spilled at one particular pressure. Re-reduced
the test to make it less dependent on the register pressure.
No functional change; all four tests pass before and after.
Assisted-by: Cursor/Claude Opus
To unsubscribe from these emails, change your notification settings at https://github.com/llvm/llvm-project/settings/notifications
More information about the All-commits
mailing list