[WaveTransform] Fix FileCheck patterns for more tests#2691
Conversation
|
PSDB Build Link: http://mlse-bdc-20dd129:8065/#/builders/10/builds/387 |
|
There seems to be overlap in tests covered in PR #2625 |
cdevadas
force-pushed
the
public/amd/dev/cdevadas/wave-transform/fix-more-control-flow-tests
branch
from
bcf113d to
d582358
Compare
Removed the duplicates in the new commit. |
|
PSDB Build Link: http://mlse-bdc-20dd129:8065/#/builders/10/builds/389 |
cdevadas
force-pushed
the
public/amd/dev/cdevadas/wave-transform/fix-more-control-flow-tests
branch
from
d582358 to
26d0c43
Compare
|
PSDB Build Link: http://mlse-bdc-20dd129:8065/#/builders/10/builds/427 |
cdevadas
force-pushed
the
public/amd/dev/cdevadas/wave-transform/fix-more-control-flow-tests
branch
from
26d0c43 to
0fb0ebd
Compare
|
PSDB Build Link: http://mlse-bdc-20dd129:8065/#/builders/10/builds/445 |
| # CHECK: v_mov_b64_e32 v[0:1], 0 | ||
| # CHECK: v_mov_b32_e32 v6, v1 | ||
| # CHECK: v_mov_b32_e32 v4, 0 | ||
| # CHECK: v_mov_b32_e32 v5, 0 | ||
| # CHECK: v_mov_b32_e32 v6, 0 | ||
| # CHECK: v_mov_b32_e32 v5, v0 |
There was a problem hiding this comment.
Extra v_mov operation generated since VGPR allocation is reordered and runs under VGPR partitioning, so %4 lands in an aligned pair (v[0:1]) instead of being assigned the misaligned v[5:6] slot directly — forcing v_mov_b64 + two copies instead of an in-place 2×v_mov_b32 expansion.
There was a problem hiding this comment.
Ya, such discrepancies would occur now. They are acceptable due to the pipeline changes. We should later analyze them if something can be done to avoid such scenarios. But not now.
cdevadas
force-pushed
the
public/amd/dev/cdevadas/wave-transform/fix-more-control-flow-tests
branch
from
0fb0ebd to
6494265
Compare
|
PSDB Build Link: http://mlse-bdc-20dd129:8065/#/builders/10/builds/487 |
Fixed the check patterns for some more control flow tests when enabled the wave transform flow. Tests fixed in this commit: - mdt-preserving-crash.ll - memcpy-crash-issue63986.ll - memmove-var-size.ll - memory-legalizer-single-wave-workgroup-memops.ll - memset-param-combinations.ll - memset-pattern.ll - mfma-loop.ll - mfma-no-register-aliasing.ll - misaligned-vgpr-regsequence.mir - mmra.ll Disabled late wave transform: - lower-control-flow-live-variables-update.xfail.mir. Disabled wave transform (-amdgpu-late-wave-transform=0) since test expects legacy SI_IF/SI_END_CF pseudos which are incompatible with wave transform.
cdevadas
force-pushed
the
public/amd/dev/cdevadas/wave-transform/fix-more-control-flow-tests
branch
from
6494265 to
ca59f2e
Compare
|
PSDB Build Link: http://mlse-bdc-20dd129:8065/#/builders/10/builds/578 |
3 participants
Fixed the check patterns for some more control flow tests
when enabled the wave transform flow.
Tests fixed in this commit:
Disabled late wave transform:
Disabled wave transform (-amdgpu-late-wave-transform=0)
since test expects legacy SI_IF/SI_END_CF pseudos which
are incompatible with wave transform.