[AMDGPU] Use poison instead of undef for non-demanded elements

Author: nikic

llvm/lib/Target/AMDGPU/AMDGPUInstCombineIntrinsic.cpp

@@ -1261,7 +1261,7 @@ static Value *simplifyAMDGCNMemoryIntrinsicDemanded(InstCombiner &IC,
 
   unsigned NewNumElts = DemandedElts.popcount();
   if (!NewNumElts)
-    return UndefValue::get(IIVTy);
+    return PoisonValue::get(IIVTy);
 
   if (NewNumElts >= VWidth && DemandedElts.isMask()) {
     if (DMaskIdx >= 0)
@@ -1299,7 +1299,7 @@ static Value *simplifyAMDGCNMemoryIntrinsicDemanded(InstCombiner &IC,
 
   if (IsLoad) {
     if (NewNumElts == 1) {
-      return IC.Builder.CreateInsertElement(UndefValue::get(IIVTy), NewCall,
+      return IC.Builder.CreateInsertElement(PoisonValue::get(IIVTy), NewCall,
                                             DemandedElts.countr_zero());
     }
 

llvm/test/Transforms/InstCombine/AMDGPU/amdgcn-demanded-vector-elts-inseltpoison.ll

@@ -4792,7 +4792,7 @@ define amdgpu_ps float @extract_elt0_image_sample_2d_v4f32_f32(float %s, float %
 
 define amdgpu_ps float @extract_elt0_dmask_0000_image_sample_3d_v4f32_f32(float %s, float %t, float %r, <8 x i32> inreg %sampler, <4 x i32> inreg %rsrc) #0 {
 ; CHECK-LABEL: @extract_elt0_dmask_0000_image_sample_3d_v4f32_f32(
-; CHECK-NEXT:    ret float undef
+; CHECK-NEXT:    ret float poison
 ;
   %data = call <4 x float> @llvm.amdgcn.image.sample.3d.v4f32.f32(i32 0, float %s, float %t, float %r, <8 x i32> %sampler, <4 x i32> %rsrc, i1 false, i32 0, i32 0)
   %elt0 = extractelement <4 x float> %data, i32 0

llvm/test/Transforms/InstCombine/AMDGPU/amdgcn-demanded-vector-elts.ll

@@ -4791,7 +4791,7 @@ define amdgpu_ps float @extract_elt0_image_sample_2d_v4f32_f32(float %s, float %
 
 define amdgpu_ps float @extract_elt0_dmask_0000_image_sample_3d_v4f32_f32(float %s, float %t, float %r, <8 x i32> inreg %sampler, <4 x i32> inreg %rsrc) #0 {
 ; CHECK-LABEL: @extract_elt0_dmask_0000_image_sample_3d_v4f32_f32(
-; CHECK-NEXT:    ret float undef
+; CHECK-NEXT:    ret float poison
 ;
   %data = call <4 x float> @llvm.amdgcn.image.sample.3d.v4f32.f32(i32 0, float %s, float %t, float %r, <8 x i32> %sampler, <4 x i32> %rsrc, i1 false, i32 0, i32 0)
   %elt0 = extractelement <4 x float> %data, i32 0