AMDGPU: Replace unused permlane inputs with poison instead of undef (#131288)

Author: arsenm

llvm/lib/Target/AMDGPU/AMDGPUInstCombineIntrinsic.cpp

@@ -1115,7 +1115,7 @@ GCNTTIImpl::instCombineIntrinsic(InstCombiner &IC, IntrinsicInst &II) const {
   case Intrinsic::amdgcn_permlanex16_var: {
     // Discard vdst_in if it's not going to be read.
     Value *VDstIn = II.getArgOperand(0);
-    if (isa<UndefValue>(VDstIn))
+    if (isa<PoisonValue>(VDstIn))
       break;
 
     // FetchInvalid operand idx.
@@ -1134,7 +1134,7 @@ GCNTTIImpl::instCombineIntrinsic(InstCombiner &IC, IntrinsicInst &II) const {
     if (!FetchInvalid->getZExtValue() && !BoundCtrl->getZExtValue())
       break;
 
-    return IC.replaceOperand(II, 0, UndefValue::get(VDstIn->getType()));
+    return IC.replaceOperand(II, 0, PoisonValue::get(VDstIn->getType()));
   }
   case Intrinsic::amdgcn_permlane64:
   case Intrinsic::amdgcn_readfirstlane:

llvm/test/Transforms/InstCombine/AMDGPU/amdgcn-intrinsics.ll

@@ -3272,7 +3272,7 @@ define amdgpu_kernel void @permlane16(ptr addrspace(1) %out, i32 %src0, i32 %src
 
 define amdgpu_kernel void @permlane16_bound_ctrl(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {
 ; CHECK-LABEL: @permlane16_bound_ctrl(
-; CHECK-NEXT:    [[RES:%.*]] = call i32 @llvm.amdgcn.permlane16.i32(i32 undef, i32 [[SRC0:%.*]], i32 [[SRC1:%.*]], i32 [[SRC2:%.*]], i1 false, i1 true)
+; CHECK-NEXT:    [[RES:%.*]] = call i32 @llvm.amdgcn.permlane16.i32(i32 poison, i32 [[SRC0:%.*]], i32 [[SRC1:%.*]], i32 [[SRC2:%.*]], i1 false, i1 true)
 ; CHECK-NEXT:    store i32 [[RES]], ptr addrspace(1) [[OUT:%.*]], align 4
 ; CHECK-NEXT:    ret void
 ;
@@ -3283,7 +3283,7 @@ define amdgpu_kernel void @permlane16_bound_ctrl(ptr addrspace(1) %out, i32 %src
 
 define amdgpu_kernel void @permlane16_fetch_invalid_bound_ctrl(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {
 ; CHECK-LABEL: @permlane16_fetch_invalid_bound_ctrl(
-; CHECK-NEXT:    [[RES:%.*]] = call i32 @llvm.amdgcn.permlane16.i32(i32 undef, i32 [[SRC0:%.*]], i32 [[SRC1:%.*]], i32 [[SRC2:%.*]], i1 true, i1 true)
+; CHECK-NEXT:    [[RES:%.*]] = call i32 @llvm.amdgcn.permlane16.i32(i32 poison, i32 [[SRC0:%.*]], i32 [[SRC1:%.*]], i32 [[SRC2:%.*]], i1 true, i1 true)
 ; CHECK-NEXT:    store i32 [[RES]], ptr addrspace(1) [[OUT:%.*]], align 4
 ; CHECK-NEXT:    ret void
 ;
@@ -3311,7 +3311,7 @@ define amdgpu_kernel void @permlanex16(ptr addrspace(1) %out, i32 %src0, i32 %sr
 
 define amdgpu_kernel void @permlanex16_bound_ctrl(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {
 ; CHECK-LABEL: @permlanex16_bound_ctrl(
-; CHECK-NEXT:    [[RES:%.*]] = call i32 @llvm.amdgcn.permlanex16.i32(i32 undef, i32 [[SRC0:%.*]], i32 [[SRC1:%.*]], i32 [[SRC2:%.*]], i1 false, i1 true)
+; CHECK-NEXT:    [[RES:%.*]] = call i32 @llvm.amdgcn.permlanex16.i32(i32 poison, i32 [[SRC0:%.*]], i32 [[SRC1:%.*]], i32 [[SRC2:%.*]], i1 false, i1 true)
 ; CHECK-NEXT:    store i32 [[RES]], ptr addrspace(1) [[OUT:%.*]], align 4
 ; CHECK-NEXT:    ret void
 ;
@@ -3322,7 +3322,7 @@ define amdgpu_kernel void @permlanex16_bound_ctrl(ptr addrspace(1) %out, i32 %sr
 
 define amdgpu_kernel void @permlanex16_fetch_invalid_bound_ctrl(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {
 ; CHECK-LABEL: @permlanex16_fetch_invalid_bound_ctrl(
-; CHECK-NEXT:    [[RES:%.*]] = call i32 @llvm.amdgcn.permlanex16.i32(i32 undef, i32 [[SRC0:%.*]], i32 [[SRC1:%.*]], i32 [[SRC2:%.*]], i1 true, i1 true)
+; CHECK-NEXT:    [[RES:%.*]] = call i32 @llvm.amdgcn.permlanex16.i32(i32 poison, i32 [[SRC0:%.*]], i32 [[SRC1:%.*]], i32 [[SRC2:%.*]], i1 true, i1 true)
 ; CHECK-NEXT:    store i32 [[RES]], ptr addrspace(1) [[OUT:%.*]], align 4
 ; CHECK-NEXT:    ret void
 ;