This is an archive of the discontinued LLVM Phabricator instance.

Differential D135331

[AMDGPU] Add test coverage to ensure first regallocfast only allocates SGPR
ClosedPublic

Authored by jrbyrnes on Oct 5 2022, 4:17 PM.

Details

Reviewers
arsenm
rampitec
Commits
rG8e5b96cce91e: [AMDGPU] Add test coverage to ensure first regallocfast only allocates SGPR
Summary

Register allocation is split into two passes, and the expected behavior is that the first pass only should only work on virtual SGPRs. Whereas the second pass works on virtual VGPRs. This adds a test case which breaks if the first pass allocates VGPRs.

Diff Detail

Event Timeline

jrbyrnes created this revision.Oct 5 2022, 4:17 PM
Herald added a project: Restricted Project. · View Herald TranscriptOct 5 2022, 4:17 PM
Herald added subscribers: kosarev, kerbowa, t-tye and 7 others. · View Herald Transcript
jrbyrnes requested review of this revision.Oct 5 2022, 4:17 PM
Herald added a project: Restricted Project. · View Herald TranscriptOct 5 2022, 4:17 PM
Herald added subscribers: llvm-commits, wdng. · View Herald Transcript
arsenm added inline comments.Oct 5 2022, 4:19 PM
llvm/test/CodeGen/AMDGPU/fastregalloc-sgpr-only.mir
2

Should specify run with regallocfast,0

Harbormaster completed remote builds in B190632: Diff 465591.Oct 5 2022, 5:07 PM
jrbyrnes updated this revision to Diff 465781.Oct 6 2022, 10:20 AM
jrbyrnes marked an inline comment as done.

Specify regallocfast,0. Add sgpr allocation test

Harbormaster completed remote builds in B190759: Diff 465781.Oct 6 2022, 11:22 AM
arsenm accepted this revision.Oct 6 2022, 11:43 AM
This revision is now accepted and ready to land.Oct 6 2022, 11:43 AM
Closed by commit rG8e5b96cce91e: [AMDGPU] Add test coverage to ensure first regallocfast only allocates SGPR (authored by jrbyrnes). · Explain WhyOct 6 2022, 2:32 PM
This revision was automatically updated to reflect the committed changes.
jrbyrnes added a commit: rG8e5b96cce91e: [AMDGPU] Add test coverage to ensure first regallocfast only allocates SGPR.

Revision Contents

PathSize
llvm/
test/
CodeGen/
AMDGPU/
17 lines

Diff 465591

llvm/test/CodeGen/AMDGPU/fastregalloc-sgpr-only.mir

  • This file was added.
# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
# RUN: llc -verify-machineinstrs -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -O0 -start-before=regallocfast -stop-after=regallocfast -o - %s | FileCheck -check-prefix=GCN %s
arsenmUnsubmitted
Done
ReplyInline Actions

Should specify run with regallocfast,0

arsenm: Should specify run with regallocfast,0
---
name: copy_not_allocated
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1, $vgpr4, $vgpr5
; GCN-LABEL: name: copy_not_allocated
; GCN: liveins: $vgpr0, $vgpr1, $vgpr4, $vgpr5
; GCN-NEXT: {{ $}}
; GCN-NEXT: [[COPY:%[0-9]+]]:vreg_64 = COPY killed $vgpr0_vgpr1
; GCN-NEXT: FLAT_STORE_DWORDX2 killed $vgpr4_vgpr5, [[COPY]], 0, 0, implicit $exec, implicit $flat_scr
%14:vreg_64 = COPY $vgpr0_vgpr1
FLAT_STORE_DWORDX2 $vgpr4_vgpr5, %14:vreg_64, 0, 0, implicit $exec, implicit $flat_scr
...