[llvm] [AMDGPU] Fix SIFoldOperands crash on REG_SEQUENCE with physical register (PR #206976)

via llvm-commits llvm-commits at lists.llvm.org
Wed Jul 1 08:11:25 PDT 2026


llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-amdgpu

Author: Frederik Harwath (frederik-h)

<details>
<summary>Changes</summary>

The getRegSeqInit function crashes on REG_SEQUENCE instructions with physical register inputs.

Since both optimizations that use getRegSeqInit do not need to handle such REG_SEQUENCE instructions, this patch changes the function to return nullptr which signals that the optimization should not happen.

---
Full diff: https://github.com/llvm/llvm-project/pull/206976.diff


2 Files Affected:

- (modified) llvm/lib/Target/AMDGPU/SIFoldOperands.cpp (+2) 
- (modified) llvm/test/CodeGen/AMDGPU/si-fold-reg-sequence.mir (+21) 


``````````diff
diff --git a/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp b/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
index 18b04ef58fa49..00cde4c5f7d8e 100644
--- a/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
+++ b/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
@@ -1001,6 +1001,8 @@ const TargetRegisterClass *SIFoldOperandsImpl::getRegSeqInit(
 
   for (unsigned I = 1, E = RegSeq.getNumExplicitOperands(); I != E; I += 2) {
     MachineOperand &SrcOp = RegSeq.getOperand(I);
+    if (SrcOp.isReg() && SrcOp.getReg().isPhysical())
+      return nullptr;
     unsigned SubRegIdx = RegSeq.getOperand(I + 1).getImm();
 
     // Only accept reg_sequence with uniform reg class inputs for simplicity.
diff --git a/llvm/test/CodeGen/AMDGPU/si-fold-reg-sequence.mir b/llvm/test/CodeGen/AMDGPU/si-fold-reg-sequence.mir
index 23b24a22b69bc..2adea7557dcd4 100644
--- a/llvm/test/CodeGen/AMDGPU/si-fold-reg-sequence.mir
+++ b/llvm/test/CodeGen/AMDGPU/si-fold-reg-sequence.mir
@@ -42,3 +42,24 @@ body:             |
     %3:vreg_64_align2 = COPY %1:sreg_64
     %4:vreg_64_align2 = nofpexcept V_PK_ADD_F32 8, %2:sreg_64_xexec, 8, %3:vreg_64_align2, 0, 0, 0, 0, 0, implicit $mode, implicit $exec
 ...
+
+---
+name: regseq_with_phys_source
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $vgpr0, $vgpr2
+
+    ; CHECK-LABEL: name: regseq_with_phys_source
+    ; CHECK: liveins: $vgpr0, $vgpr2
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 1, implicit $exec
+    ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE $vgpr0, %subreg.sub0, [[V_MOV_B32_e32_]], %subreg.sub1
+    ; CHECK-NEXT: [[V_ADD_CO_U32_e32_:%[0-9]+]]:vgpr_32 = V_ADD_CO_U32_e32 [[REG_SEQUENCE]].sub0, $vgpr2, implicit-def $vcc, implicit $exec
+    ; CHECK-NEXT: S_ENDPGM 0, implicit [[V_ADD_CO_U32_e32_]]
+    %0:vgpr_32 = V_MOV_B32_e32 1, implicit $exec
+    %1:vreg_64 = REG_SEQUENCE $vgpr0, %subreg.sub0, %0, %subreg.sub1
+    %2:vgpr_32 = V_ADD_CO_U32_e32 %1.sub0, $vgpr2, implicit-def $vcc, implicit $exec
+    S_ENDPGM 0, implicit %2
+
+...

``````````

</details>


https://github.com/llvm/llvm-project/pull/206976


More information about the llvm-commits mailing list