fix up in prep for PR.

Author: farzonl

llvm/lib/Transforms/Scalar/Scalarizer.cpp

@@ -200,11 +200,17 @@ struct VectorLayout {
 static bool isStructAllVectors(Type *Ty) {
   if (!isa<StructType>(Ty))
     return false;
-
-  for(unsigned I = 0; I < Ty->getNumContainedTypes(); I++)
-    if (!isa<FixedVectorType>(Ty->getContainedType(I)))
+  if (Ty->getNumContainedTypes() < 1)
+    return false;
+  FixedVectorType *VecTy = dyn_cast<FixedVectorType>(Ty->getContainedType(0));
+  if (!VecTy)
+    return false;
+  unsigned VecSize = VecTy->getNumElements();
+  for (unsigned I = 1; I < Ty->getNumContainedTypes(); I++) {
+    VecTy = dyn_cast<FixedVectorType>(Ty->getContainedType(I));
+    if (!VecTy || VecSize != VecTy->getNumElements())
       return false;
-
+  }
   return true;
 }
 
@@ -679,8 +685,9 @@ bool ScalarizerVisitor::splitBinary(Instruction &I, const Splitter &Split) {
 bool ScalarizerVisitor::isTriviallyScalarizable(Intrinsic::ID ID) {
   if (isTriviallyVectorizable(ID))
     return true;
+  // TODO: investigate vectorizable frexp
   switch (ID) {
-    case Intrinsic::frexp:
+  case Intrinsic::frexp:
     return true;
   }
   return Intrinsic::isTargetIntrinsic(ID) &&
@@ -690,10 +697,10 @@ bool ScalarizerVisitor::isTriviallyScalarizable(Intrinsic::ID ID) {
 /// If a call to a vector typed intrinsic function, split into a scalar call per
 /// element if possible for the intrinsic.
 bool ScalarizerVisitor::splitCall(CallInst &CI) {
-  Type* CallType = CI.getType();
-  bool areAllVectors = isStructAllVectors(CallType);
-   std::optional<VectorSplit> VS;
-  if (areAllVectors)
+  Type *CallType = CI.getType();
+  bool AreAllVectors = isStructAllVectors(CallType);
+  std::optional<VectorSplit> VS;
+  if (AreAllVectors)
     VS = getVectorSplit(CallType->getContainedType(0));
   else
     VS = getVectorSplit(CallType);
@@ -721,12 +728,12 @@ bool ScalarizerVisitor::splitCall(CallInst &CI) {
   if (isVectorIntrinsicWithOverloadTypeAtArg(ID, -1))
     Tys.push_back(VS->SplitTy);
 
-  if(areAllVectors) {
-    Type* PrevType = CallType->getContainedType(0);
-    Type* CallType = CI.getType();
-    for(unsigned I = 1; I < CallType->getNumContainedTypes(); I++) {
-      Type* CurrType = cast<FixedVectorType>(CallType->getContainedType(I));
-      if(PrevType != CurrType) {
+  if (AreAllVectors) {
+    Type *PrevType = CallType->getContainedType(0);
+    Type *CallType = CI.getType();
+    for (unsigned I = 1; I < CallType->getNumContainedTypes(); I++) {
+      Type *CurrType = cast<FixedVectorType>(CallType->getContainedType(I));
+      if (PrevType != CurrType) {
         std::optional<VectorSplit> CurrVS = getVectorSplit(CurrType);
         Tys.push_back(CurrVS->SplitTy);
         PrevType = CurrType;
@@ -1070,7 +1077,7 @@ bool ScalarizerVisitor::visitExtractValueInst(ExtractValueInst &EVI) {
   ValueVector Res;
   if (!isStructAllVectors(OpTy))
     return false;
-  Type* VecType = cast<FixedVectorType>(OpTy->getContainedType(0));
+  Type *VecType = cast<FixedVectorType>(OpTy->getContainedType(0));
   std::optional<VectorSplit> VS = getVectorSplit(VecType);
   if (!VS)
     return false;
@@ -1084,7 +1091,7 @@ bool ScalarizerVisitor::visitExtractValueInst(ExtractValueInst &EVI) {
         Op0[OpIdx], Index, EVI.getName() + ".elem" + std::to_string(Index));
     Res.push_back(ResElem);
   }
-  // replaceUses(&EVI, Res);
+
   gather(&EVI, Res, *VS);
   return true;
 }

llvm/test/CodeGen/DirectX/split-double.ll

@@ -1,10 +1,18 @@
-; RUN: opt -S -scalarizer  -mtriple=dxil-pc-shadermodel6.3-library %s | FileCheck %s
+; RUN: opt -passes='function(scalarizer<load-store>)' -S -mtriple=dxil-pc-shadermodel6.3-library %s | FileCheck %s
 
-define void @test_vector_double_split_void(<3 x double> noundef %d) {
-  %hlsl.asuint = call { <3 x i32>, <3 x i32> }  @llvm.dx.splitdouble.v3i32(<3 x double> %d)
+; CHECK-LABEL: @test_vector_double_split_void
+define void @test_vector_double_split_void(<2 x double> noundef %d) {
+  ; CHECK: [[ee0:%.*]] = extractelement <2 x double> %d, i64 0
+  ; CHECK: [[ie0:%.*]] = call { i32, i32 } @llvm.dx.splitdouble.i32(double [[ee0]])
+  ; CHECK: [[ee1:%.*]] = extractelement <2 x double> %d, i64 1
+  ; CHECK: [[ie1:%.*]] = call { i32, i32 } @llvm.dx.splitdouble.i32(double [[ee1]])
+  ; CHECK-NOT: extractvalue { i32, i32 } {{.*}}, 0
+  ; CHECK-NOT: insertelement <2 x i32> {{.*}}, i32 {{.*}}, i64 0
+  %hlsl.asuint = call { <2 x i32>, <2 x i32> }  @llvm.dx.splitdouble.v2i32(<2 x double> %d)
   ret void
 }
 
+; CHECK-LABEL: @test_vector_double_split
 define noundef <3 x i32> @test_vector_double_split(<3 x double> noundef %d) {
   ; CHECK: [[ee0:%.*]] = extractelement <3 x double> %d, i64 0
   ; CHECK: [[ie0:%.*]] = call { i32, i32 } @llvm.dx.splitdouble.i32(double [[ee0]])

llvm/test/Transforms/Scalarizer/frexp.ll

@@ -30,7 +30,7 @@ define noundef <2 x i32> @test_vector_half_frexp_int(<2 x half> noundef %h) {
   ret <2 x i32> %e1
 }
 
-
+; CHECK-LABEL: @test_vector_float_frexp_int
 define noundef <2 x float> @test_vector_float_frexp_int(<2 x float> noundef %f) {
   ; CHECK: [[ee0:%.*]] = extractelement <2 x float> %f, i64 0
   ; CHECK-NEXT: [[ie0:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[ee0]])
@@ -48,6 +48,7 @@ define noundef <2 x float> @test_vector_float_frexp_int(<2 x float> noundef %f)
   ret <2 x float> %2
 }
 
+; CHECK-LABEL: @test_vector_double_frexp_int
 define noundef <2 x double> @test_vector_double_frexp_int(<2 x double> noundef %d) {
   ; CHECK: [[ee0:%.*]] = extractelement <2 x double> %d, i64 0
   ; CHECK-NEXT: [[ie0:%.*]] = call { double, i32 } @llvm.frexp.f64.i32(double [[ee0]])