[CodeGenPrepare] Replace undef use with poison

[pedroclobo]

When generating a constant vector, if UseSplat is false, the indices different from the index of the extract can be filled with poison instead of undef.

[llvmbot]

@llvm/pr-subscribers-backend-arm

Author: Pedro Lobo (pedroclobo)

Changes

When generating a constant vector, if UseSplat is false, the indices different from the index of the extract can be filled with poison instead of undef.

---

Full diff: https://github.com/llvm/llvm-project/pull/123111.diff

2 Files Affected:

• (modified) llvm/lib/CodeGen/CodeGenPrepare.cpp (+4-4)
• (modified) llvm/test/CodeGen/ARM/vector-promotion.ll (+13-13)

diff --git a/llvm/lib/CodeGen/CodeGenPrepare.cpp b/llvm/lib/CodeGen/CodeGenPrepare.cpp
index ba1b10ec8b9b1b..9abcbc0393e484 100644
--- a/llvm/lib/CodeGen/CodeGenPrepare.cpp
+++ b/llvm/lib/CodeGen/CodeGenPrepare.cpp
@@ -7983,8 +7983,8 @@ class VectorPromoteHelper {
   /// \p UseSplat defines whether or not \p Val should be replicated
   /// across the whole vector.
   /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
-  /// otherwise we generate a vector with as many undef as possible:
-  /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
+  /// otherwise we generate a vector with as many poison as possible:
+  /// <poison, ..., poison, Val, poison, ..., poison> where \p Val is only
   /// used at the index of the extract.
   Value *getConstantVector(Constant *Val, bool UseSplat) const {
     unsigned ExtractIdx = std::numeric_limits<unsigned>::max();
@@ -8004,12 +8004,12 @@ class VectorPromoteHelper {
 
     if (!EC.isScalable()) {
       SmallVector<Constant *, 4> ConstVec;
-      UndefValue *UndefVal = UndefValue::get(Val->getType());
+      PoisonValue *PoisonVal = PoisonValue::get(Val->getType());
       for (unsigned Idx = 0; Idx != EC.getKnownMinValue(); ++Idx) {
         if (Idx == ExtractIdx)
           ConstVec.push_back(Val);
         else
-          ConstVec.push_back(UndefVal);
+          ConstVec.push_back(PoisonVal);
       }
       return ConstantVector::get(ConstVec);
     } else
diff --git a/llvm/test/CodeGen/ARM/vector-promotion.ll b/llvm/test/CodeGen/ARM/vector-promotion.ll
index a9a8f58963a374..344014ad804495 100644
--- a/llvm/test/CodeGen/ARM/vector-promotion.ll
+++ b/llvm/test/CodeGen/ARM/vector-promotion.ll
@@ -4,7 +4,7 @@
 
 ; IR-BOTH-LABEL: @simpleOneInstructionPromotion
 ; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <2 x i32>, ptr %addr1
-; IR-BOTH-NEXT: [[VECTOR_OR:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[LOAD]], <i32 undef, i32 1>
+; IR-BOTH-NEXT: [[VECTOR_OR:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[LOAD]], <i32 poison, i32 1>
 ; IR-BOTH-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x i32> [[VECTOR_OR]], i32 1
 ; IR-BOTH-NEXT: store i32 [[EXTRACT]], ptr %dest
 ; IR-BOTH-NEXT: ret
@@ -71,13 +71,13 @@ end:
 
 ; IR-LABEL: @chainOfInstructionsToPromote
 ; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <2 x i32>, ptr %addr1
-; IR-BOTH-NEXT: [[VECTOR_OR1:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[LOAD]], <i32 1, i32 undef>
-; IR-BOTH-NEXT: [[VECTOR_OR2:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[VECTOR_OR1]], <i32 1, i32 undef>
-; IR-BOTH-NEXT: [[VECTOR_OR3:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[VECTOR_OR2]], <i32 1, i32 undef>
-; IR-BOTH-NEXT: [[VECTOR_OR4:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[VECTOR_OR3]], <i32 1, i32 undef>
-; IR-BOTH-NEXT: [[VECTOR_OR5:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[VECTOR_OR4]], <i32 1, i32 undef>
-; IR-BOTH-NEXT: [[VECTOR_OR6:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[VECTOR_OR5]], <i32 1, i32 undef>
-; IR-BOTH-NEXT: [[VECTOR_OR7:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[VECTOR_OR6]], <i32 1, i32 undef>
+; IR-BOTH-NEXT: [[VECTOR_OR1:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[LOAD]], <i32 1, i32 poison>
+; IR-BOTH-NEXT: [[VECTOR_OR2:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[VECTOR_OR1]], <i32 1, i32 poison>
+; IR-BOTH-NEXT: [[VECTOR_OR3:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[VECTOR_OR2]], <i32 1, i32 poison>
+; IR-BOTH-NEXT: [[VECTOR_OR4:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[VECTOR_OR3]], <i32 1, i32 poison>
+; IR-BOTH-NEXT: [[VECTOR_OR5:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[VECTOR_OR4]], <i32 1, i32 poison>
+; IR-BOTH-NEXT: [[VECTOR_OR6:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[VECTOR_OR5]], <i32 1, i32 poison>
+; IR-BOTH-NEXT: [[VECTOR_OR7:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[VECTOR_OR6]], <i32 1, i32 poison>
 ; IR-BOTH-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x i32> [[VECTOR_OR7]], i32 0
 ; IR-BOTH-NEXT: store i32 [[EXTRACT]], ptr %dest
 ; IR-BOTH-NEXT: ret
@@ -276,7 +276,7 @@ define void @undefRemCase(ptr %addr1, ptr %dest) {
 ; IR-NORMAL-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x float> [[LOAD]], i32 1
 ; IR-NORMAL-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = frem nnan float [[EXTRACT]], 7.0
 ; Vector version:
-; IR-STRESS-NEXT: [[DIV:%[a-zA-Z_0-9-]+]] = frem nnan <2 x float> [[LOAD]], <float undef, float 7.000000e+00>
+; IR-STRESS-NEXT: [[DIV:%[a-zA-Z_0-9-]+]] = frem nnan <2 x float> [[LOAD]], <float poison, float 7.000000e+00>
 ; IR-STRESS-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = extractelement <2 x float> [[DIV]], i32 1
 ;
 ; IR-BOTH-NEXT: store float [[RES]], ptr %dest
@@ -297,7 +297,7 @@ define void @undefConstantFRemCaseWithFastMath(ptr %addr1, ptr %dest) {
 ; IR-NORMAL-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x float> [[LOAD]], i32 1
 ; IR-NORMAL-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = frem nnan float 7.000000e+00, [[EXTRACT]]
 ; Vector version:
-; IR-STRESS-NEXT: [[DIV:%[a-zA-Z_0-9-]+]] = frem nnan <2 x float> <float undef, float 7.000000e+00>, [[LOAD]]
+; IR-STRESS-NEXT: [[DIV:%[a-zA-Z_0-9-]+]] = frem nnan <2 x float> <float poison, float 7.000000e+00>, [[LOAD]]
 ; IR-STRESS-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = extractelement <2 x float> [[DIV]], i32 1
 ;
 ; IR-BOTH-NEXT: store float [[RES]], ptr %dest
@@ -319,7 +319,7 @@ define void @undefVectorFRemCaseWithFastMath(ptr %addr1, ptr %dest) {
 ; IR-NORMAL-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x float> [[LOAD]], i32 1
 ; IR-NORMAL-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = fadd float [[EXTRACT]], 1.0
 ; Vector version:
-; IR-STRESS-NEXT: [[DIV:%[a-zA-Z_0-9-]+]] = fadd <2 x float> [[LOAD]], <float undef, float 1.000000e+00>
+; IR-STRESS-NEXT: [[DIV:%[a-zA-Z_0-9-]+]] = fadd <2 x float> [[LOAD]], <float poison, float 1.000000e+00>
 ; IR-STRESS-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = extractelement <2 x float> [[DIV]], i32 1
 ;
 ; IR-BOTH-NEXT: store float [[RES]], ptr %dest
@@ -358,7 +358,7 @@ define void @simpleOneInstructionPromotionVariableIdx(ptr %addr1, ptr %dest, i32
 ; Check a vector with more than 2 elements.
 ; IR-BOTH-LABEL: @simpleOneInstructionPromotion8x8
 ; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <8 x i8>, ptr %addr1
-; IR-BOTH-NEXT: [[OR:%[a-zA-Z_0-9-]+]] = or <8 x i8> [[LOAD]], <i8 undef, i8 1, i8 undef, i8 undef, i8 undef, i8 undef, i8 undef, i8 undef>
+; IR-BOTH-NEXT: [[OR:%[a-zA-Z_0-9-]+]] = or <8 x i8> [[LOAD]], <i8 poison, i8 1, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison>
 ; IR-BOTH-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = extractelement <8 x i8> [[OR]], i32 1
 ; IR-BOTH-NEXT: store i8 [[RES]], ptr %dest
 ; IR-BOTH-NEXT: ret
@@ -374,7 +374,7 @@ define void @simpleOneInstructionPromotion8x8(ptr %addr1, ptr %dest) {
 ; lowered on a Q register.
 ; IR-BOTH-LABEL: @simpleOneInstructionPromotion
 ; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <4 x i32>, ptr %addr1
-; IR-BOTH-NEXT: [[VECTOR_OR:%[a-zA-Z_0-9-]+]] = or <4 x i32> [[LOAD]], <i32 undef, i32 1, i32 undef, i32 undef>
+; IR-BOTH-NEXT: [[VECTOR_OR:%[a-zA-Z_0-9-]+]] = or <4 x i32> [[LOAD]], <i32 poison, i32 1, i32 poison, i32 poison>
 ; IR-BOTH-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <4 x i32> [[VECTOR_OR]], i32 1
 ; IR-BOTH-NEXT: store i32 [[EXTRACT]], ptr %dest
 ; IR-BOTH-NEXT: ret

[pedroclobo]

cc @nunoplopes