In vector case do more heavy lifting in sema so codegen can reuse VectorSplat codegen

Author: spall

clang/include/clang/Sema/SemaHLSL.h

@@ -144,7 +144,7 @@ class SemaHLSL : public SemaBase {
   bool CanPerformScalarCast(QualType SrcTy, QualType DestTy);
   bool ContainsBitField(QualType BaseTy);
   bool CanPerformElementwiseCast(Expr *Src, QualType DestType);
-  bool CanPerformSplatCast(Expr *Src, QualType DestType);
+  bool CanPerformAggregateSplatCast(Expr *Src, QualType DestType);
   ExprResult ActOnOutParamExpr(ParmVarDecl *Param, Expr *Arg);
 
   QualType getInoutParameterType(QualType Ty);

clang/lib/CodeGen/CGExprScalar.cpp

@@ -2643,6 +2643,11 @@ Value *ScalarExprEmitter::VisitCastExpr(CastExpr *CE) {
     return EmitScalarConversion(Visit(E), E->getType(), DestTy,
                                 CE->getExprLoc());
   }
+    // CK_HLSLAggregateSplatCast only handles splatting to vectors from a vec1
+    // Casts were inserted in Sema to Cast the Src Expr to a Scalar and
+    // To perform any necessary Scalar Cast, so this Cast can be handled
+    // by the regular Vector Splat cast code.
+  case CK_HLSLAggregateSplatCast:
   case CK_VectorSplat: {
     llvm::Type *DstTy = ConvertType(DestTy);
     Value *Elt = Visit(const_cast<Expr *>(E));
@@ -2795,22 +2800,6 @@ Value *ScalarExprEmitter::VisitCastExpr(CastExpr *CE) {
     llvm::Value *Zero = llvm::Constant::getNullValue(CGF.SizeTy);
     return Builder.CreateExtractElement(Vec, Zero, "cast.vtrunc");
   }
-  case CK_HLSLAggregateSplatCast: {
-    // This cast should only handle splatting from vectors of length 1.
-    // But in Sema a cast should have been inserted to convert the vec1 to a
-    // scalar
-    assert(DestTy->isVectorType() && "Destination type must be a vector.");
-    auto *DestVecTy = DestTy->getAs<VectorType>();
-    QualType SrcTy = E->getType();
-    SourceLocation Loc = CE->getExprLoc();
-    Value *V = Visit(const_cast<Expr *>(E));
-    assert(SrcTy->isBuiltinType() && "Invalid HLSL Aggregate splat cast.");
-
-    Value *Cast =
-        EmitScalarConversion(V, SrcTy, DestVecTy->getElementType(), Loc);
-    return Builder.CreateVectorSplat(DestVecTy->getNumElements(), Cast,
-                                     "splat");
-  }
   case CK_HLSLElementwiseCast: {
     RValue RV = CGF.EmitAnyExpr(E);
     SourceLocation Loc = CE->getExprLoc();

clang/lib/Sema/SemaCast.cpp

@@ -2793,13 +2793,20 @@ void CastOperation::CheckCXXCStyleCast(bool FunctionalStyle,
   // If the relative order of this and the HLSLElementWise cast checks
   // are changed, it might change which cast handles what in a few cases
   if (Self.getLangOpts().HLSL &&
-      Self.HLSL().CanPerformSplatCast(SrcExpr.get(), DestType)) {
+      Self.HLSL().CanPerformAggregateSplatCast(SrcExpr.get(), DestType)) {
     const VectorType *VT = SrcTy->getAs<VectorType>();
     // change splat from vec1 case to splat from scalar
     if (VT && VT->getNumElements() == 1)
       SrcExpr = Self.ImpCastExprToType(
           SrcExpr.get(), VT->getElementType(), CK_HLSLVectorTruncation,
           SrcExpr.get()->getValueKind(), nullptr, CCK);
+    // Inserting a scalar cast here allows for a simplified codegen in
+    // the case the destTy is a vector
+    if (const VectorType *DVT = DestType->getAs<VectorType>())
+      SrcExpr = Self.ImpCastExprToType(
+          SrcExpr.get(), DVT->getElementType(),
+          Self.PrepareScalarCast(SrcExpr, DVT->getElementType()),
+          SrcExpr.get()->getValueKind(), nullptr, CCK);
     Kind = CK_HLSLAggregateSplatCast;
     return;
   }

clang/test/SemaHLSL/Language/AggregateSplatCasts.hlsl

@@ -3,6 +3,7 @@
 // splat from vec1 to vec
 // CHECK-LABEL: call1
 // CHECK: CStyleCastExpr {{.*}} 'int3':'vector<int, 3>' <HLSLAggregateSplatCast>
+// CHECK-NEXT: ImplicitCastExpr {{.*}} 'int' lvalue <FloatingToIntegral> part_of_explicit_cast
 // CHECK-NEXT: ImplicitCastExpr {{.*}} 'float' lvalue <HLSLVectorTruncation> part_of_explicit_cast
 // CHECK-NEXT: DeclRefExpr {{.*}} 'float1':'vector<float, 1>' lvalue Var {{.*}} 'A' 'float1':'vector<float, 1>'
 export void call1() {