[llvm] [AMDGPU][True16][CodeGen] Do not fold hi16 SGPR copy (PR #220858)

[via llvm-commits]

llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-amdgpu

Author: Arseniy Obolenskiy (aobolensk)

<details>
<summary>Changes</summary>

Dropping the subreg on a hi16 copy would read the low half instead

---
Full diff: https://github.com/llvm/llvm-project/pull/220858.diff


2 Files Affected:

- (modified) llvm/lib/Target/AMDGPU/SIFoldOperands.cpp (+2-1) 
- (modified) llvm/test/CodeGen/AMDGPU/true16-fold.mir (+18) 


``````````diff
diff --git a/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp b/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
index d55067c963a89..8a1f748a93f17 100644
--- a/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
+++ b/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
@@ -2215,7 +2215,8 @@ bool SIFoldOperandsImpl::tryFoldFoldableCopy(
       OpToFold.getSubReg()) {
     if (DstRC == &AMDGPU::SReg_32RegClass &&
         DstRC == MRI->getRegClass(OpToFold.getReg())) {
-      assert(OpToFold.getSubReg() == AMDGPU::lo16);
+      if (OpToFold.getSubReg() != AMDGPU::lo16)
+        return false;
       OpToFold.setSubReg(0);
     }
   }
diff --git a/llvm/test/CodeGen/AMDGPU/true16-fold.mir b/llvm/test/CodeGen/AMDGPU/true16-fold.mir
index 29be6c55e6f30..5dfabb938f1f9 100644
--- a/llvm/test/CodeGen/AMDGPU/true16-fold.mir
+++ b/llvm/test/CodeGen/AMDGPU/true16-fold.mir
@@ -265,3 +265,21 @@ body:             |
     $vgpr0 = COPY %5
     SI_RETURN implicit $vgpr0
 ...
+
+---
+name:            sgpr_hi16_copy_not_folded
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    ; CHECK-LABEL: name: sgpr_hi16_copy_not_folded
+    ; CHECK: [[DEF:%[0-9]+]]:sreg_32 = IMPLICIT_DEF
+    ; CHECK-NEXT: [[DEF1:%[0-9]+]]:sreg_32 = IMPLICIT_DEF
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY [[DEF]].hi16
+    ; CHECK-NEXT: [[S_PACK_LL_B32_B16_:%[0-9]+]]:sreg_32 = S_PACK_LL_B32_B16 [[DEF1]], killed [[COPY]], implicit-def dead $scc
+    ; CHECK-NEXT: S_ENDPGM 0, implicit [[S_PACK_LL_B32_B16_]]
+    %0:sreg_32 = IMPLICIT_DEF
+    %1:sreg_32 = IMPLICIT_DEF
+    %2:sreg_32 = COPY %0.hi16
+    %3:sreg_32 = S_PACK_LL_B32_B16 %1, killed %2, implicit-def dead $scc
+    S_ENDPGM 0, implicit %3
+...

``````````

</details>


https://github.com/llvm/llvm-project/pull/220858