[AMDGPU][CodeGen] enable D16Writes32BitVgpr for gfx12

[broxigarchen]

The issue in #157795 also reproducible for gfx12.

This should only be used in true16 mode. Although gfx12 not yet has true16 mode enable, there could be downstream branch to test true16 mode manually thus enable this workaround now

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: Brox Chen (broxigarchen)

Changes

The issue in #157795 also producible for gfx12

---

Full diff: https://github.com/llvm/llvm-project/pull/165587.diff

1 Files Affected:

• (modified) llvm/lib/Target/AMDGPU/AMDGPU.td (+2-1)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPU.td b/llvm/lib/Target/AMDGPU/AMDGPU.td
index 54d94b1f8682e..84121138a40df 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPU.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPU.td
@@ -1614,7 +1614,8 @@ def FeatureGFX12 : GCNSubtargetFeatureGeneration<"GFX12",
    FeatureDefaultComponentBroadcast, FeatureMaxHardClauseLength32,
    FeatureAtomicFMinFMaxF32GlobalInsts, FeatureAtomicFMinFMaxF32FlatInsts,
    FeatureIEEEMinimumMaximumInsts, FeatureMinimum3Maximum3F32,
-   FeatureMinimum3Maximum3F16, FeatureAgentScopeFineGrainedRemoteMemoryAtomics
+   FeatureMinimum3Maximum3F16, FeatureAgentScopeFineGrainedRemoteMemoryAtomics,
+   FeatureD16Writes32BitVgpr
   ]
 >;
 

[Sisyph]

There should be a test. For example we should see a new waitcnt on the true16 gfx12 path in some test.

[broxigarchen]

There should be a test. For example we should see a new waitcnt on the true16 gfx12 path in some test.

It seems we don't have any gfx12 runline that generates a 16bit related waitcnt. Update one test runline to generate gfx12 runline with/without the d16writevgpr32 attribute. But still it's not very straightforward to view the diff