[HLSL] Implement HLSL Elementwise casting (excluding splat cases); Re-land #118842 (#126258)

Implement HLSLElementwiseCast excluding support for splat cases
Do not support casting types that contain bitfields.
Partly closes #100609 and
partly closes #100619
Re-land #118842 after fixing warning as an error, found by a buildbot.

Author: spall

clang/include/clang/AST/OperationKinds.def

@@ -367,6 +367,9 @@ CAST_OPERATION(HLSLVectorTruncation)
 // Non-decaying array RValue cast (HLSL only).
 CAST_OPERATION(HLSLArrayRValue)
 
+// Aggregate by Value cast (HLSL only).
+CAST_OPERATION(HLSLElementwiseCast)
+
 //===- Binary Operations  -------------------------------------------------===//
 // Operators listed in order of precedence.
 // Note that additions to this should also update the StmtVisitor class,

clang/include/clang/Sema/SemaHLSL.h

@@ -141,6 +141,9 @@ class SemaHLSL : public SemaBase {
   // Diagnose whether the input ID is uint/unit2/uint3 type.
   bool diagnoseInputIDType(QualType T, const ParsedAttr &AL);
 
+  bool CanPerformScalarCast(QualType SrcTy, QualType DestTy);
+  bool ContainsBitField(QualType BaseTy);
+  bool CanPerformElementwiseCast(Expr *Src, QualType DestType);
   ExprResult ActOnOutParamExpr(ParmVarDecl *Param, Expr *Arg);
 
   QualType getInoutParameterType(QualType Ty);

clang/lib/AST/Expr.cpp

@@ -1956,6 +1956,7 @@ bool CastExpr::CastConsistency() const {
   case CK_FixedPointToBoolean:
   case CK_HLSLArrayRValue:
   case CK_HLSLVectorTruncation:
+  case CK_HLSLElementwiseCast:
   CheckNoBasePath:
     assert(path_empty() && "Cast kind should not have a base path!");
     break;

clang/lib/AST/ExprConstant.cpp

@@ -15047,6 +15047,7 @@ bool IntExprEvaluator::VisitCastExpr(const CastExpr *E) {
   case CK_NoOp:
   case CK_LValueToRValueBitCast:
   case CK_HLSLArrayRValue:
+  case CK_HLSLElementwiseCast:
     return ExprEvaluatorBaseTy::VisitCastExpr(E);
 
   case CK_MemberPointerToBoolean:
@@ -15905,6 +15906,7 @@ bool ComplexExprEvaluator::VisitCastExpr(const CastExpr *E) {
   case CK_IntegralToFixedPoint:
   case CK_MatrixCast:
   case CK_HLSLVectorTruncation:
+  case CK_HLSLElementwiseCast:
     llvm_unreachable("invalid cast kind for complex value");
 
   case CK_LValueToRValue:

clang/lib/CodeGen/CGExpr.cpp

@@ -5338,6 +5338,7 @@ LValue CodeGenFunction::EmitCastLValue(const CastExpr *E) {
   case CK_MatrixCast:
   case CK_HLSLVectorTruncation:
   case CK_HLSLArrayRValue:
+  case CK_HLSLElementwiseCast:
     return EmitUnsupportedLValue(E, "unexpected cast lvalue");
 
   case CK_Dependent:
@@ -6376,3 +6377,75 @@ RValue CodeGenFunction::EmitPseudoObjectRValue(const PseudoObjectExpr *E,
 LValue CodeGenFunction::EmitPseudoObjectLValue(const PseudoObjectExpr *E) {
   return emitPseudoObjectExpr(*this, E, true, AggValueSlot::ignored()).LV;
 }
+
+void CodeGenFunction::FlattenAccessAndType(
+    Address Addr, QualType AddrType,
+    SmallVectorImpl<std::pair<Address, llvm::Value *>> &AccessList,
+    SmallVectorImpl<QualType> &FlatTypes) {
+  // WorkList is list of type we are processing + the Index List to access
+  // the field of that type in Addr for use in a GEP
+  llvm::SmallVector<std::pair<QualType, llvm::SmallVector<llvm::Value *, 4>>,
+                    16>
+      WorkList;
+  llvm::IntegerType *IdxTy = llvm::IntegerType::get(getLLVMContext(), 32);
+  // Addr should be a pointer so we need to 'dereference' it
+  WorkList.push_back({AddrType, {llvm::ConstantInt::get(IdxTy, 0)}});
+
+  while (!WorkList.empty()) {
+    auto [T, IdxList] = WorkList.pop_back_val();
+    T = T.getCanonicalType().getUnqualifiedType();
+    assert(!isa<MatrixType>(T) && "Matrix types not yet supported in HLSL");
+    if (const auto *CAT = dyn_cast<ConstantArrayType>(T)) {
+      uint64_t Size = CAT->getZExtSize();
+      for (int64_t I = Size - 1; I > -1; I--) {
+        llvm::SmallVector<llvm::Value *, 4> IdxListCopy = IdxList;
+        IdxListCopy.push_back(llvm::ConstantInt::get(IdxTy, I));
+        WorkList.emplace_back(CAT->getElementType(), IdxListCopy);
+      }
+    } else if (const auto *RT = dyn_cast<RecordType>(T)) {
+      const RecordDecl *Record = RT->getDecl();
+      assert(!Record->isUnion() && "Union types not supported in flat cast.");
+
+      const CXXRecordDecl *CXXD = dyn_cast<CXXRecordDecl>(Record);
+
+      llvm::SmallVector<QualType, 16> FieldTypes;
+      if (CXXD && CXXD->isStandardLayout())
+        Record = CXXD->getStandardLayoutBaseWithFields();
+
+      // deal with potential base classes
+      if (CXXD && !CXXD->isStandardLayout()) {
+        for (auto &Base : CXXD->bases())
+          FieldTypes.push_back(Base.getType());
+      }
+
+      for (auto *FD : Record->fields())
+        FieldTypes.push_back(FD->getType());
+
+      for (int64_t I = FieldTypes.size() - 1; I > -1; I--) {
+        llvm::SmallVector<llvm::Value *, 4> IdxListCopy = IdxList;
+        IdxListCopy.push_back(llvm::ConstantInt::get(IdxTy, I));
+        WorkList.insert(WorkList.end(), {FieldTypes[I], IdxListCopy});
+      }
+    } else if (const auto *VT = dyn_cast<VectorType>(T)) {
+      llvm::Type *LLVMT = ConvertTypeForMem(T);
+      CharUnits Align = getContext().getTypeAlignInChars(T);
+      Address GEP =
+          Builder.CreateInBoundsGEP(Addr, IdxList, LLVMT, Align, "vector.gep");
+      for (unsigned I = 0, E = VT->getNumElements(); I < E; I++) {
+        llvm::Value *Idx = llvm::ConstantInt::get(IdxTy, I);
+        // gep on vector fields is not recommended so combine gep with
+        // extract/insert
+        AccessList.emplace_back(GEP, Idx);
+        FlatTypes.push_back(VT->getElementType());
+      }
+    } else {
+      // a scalar/builtin type
+      llvm::Type *LLVMT = ConvertTypeForMem(T);
+      CharUnits Align = getContext().getTypeAlignInChars(T);
+      Address GEP =
+          Builder.CreateInBoundsGEP(Addr, IdxList, LLVMT, Align, "gep");
+      AccessList.emplace_back(GEP, nullptr);
+      FlatTypes.push_back(T);
+    }
+  }
+}

clang/lib/CodeGen/CGExprAgg.cpp

@@ -491,6 +491,79 @@ static bool isTrivialFiller(Expr *E) {
   return false;
 }
 
+// emit a flat cast where the RHS is a scalar, including vector
+static void EmitHLSLScalarFlatCast(CodeGenFunction &CGF, Address DestVal,
+                                   QualType DestTy, llvm::Value *SrcVal,
+                                   QualType SrcTy, SourceLocation Loc) {
+  // Flatten our destination
+  SmallVector<QualType, 16> DestTypes; // Flattened type
+  SmallVector<std::pair<Address, llvm::Value *>, 16> StoreGEPList;
+  // ^^ Flattened accesses to DestVal we want to store into
+  CGF.FlattenAccessAndType(DestVal, DestTy, StoreGEPList, DestTypes);
+
+  assert(SrcTy->isVectorType() && "HLSL Flat cast doesn't handle splatting.");
+  const VectorType *VT = SrcTy->getAs<VectorType>();
+  SrcTy = VT->getElementType();
+  assert(StoreGEPList.size() <= VT->getNumElements() &&
+         "Cannot perform HLSL flat cast when vector source \
+         object has less elements than flattened destination \
+         object.");
+  for (unsigned I = 0, Size = StoreGEPList.size(); I < Size; I++) {
+    llvm::Value *Load = CGF.Builder.CreateExtractElement(SrcVal, I, "vec.load");
+    llvm::Value *Cast =
+        CGF.EmitScalarConversion(Load, SrcTy, DestTypes[I], Loc);
+
+    // store back
+    llvm::Value *Idx = StoreGEPList[I].second;
+    if (Idx) {
+      llvm::Value *V =
+          CGF.Builder.CreateLoad(StoreGEPList[I].first, "load.for.insert");
+      Cast = CGF.Builder.CreateInsertElement(V, Cast, Idx);
+    }
+    CGF.Builder.CreateStore(Cast, StoreGEPList[I].first);
+  }
+  return;
+}
+
+// emit a flat cast where the RHS is an aggregate
+static void EmitHLSLElementwiseCast(CodeGenFunction &CGF, Address DestVal,
+                                    QualType DestTy, Address SrcVal,
+                                    QualType SrcTy, SourceLocation Loc) {
+  // Flatten our destination
+  SmallVector<QualType, 16> DestTypes; // Flattened type
+  SmallVector<std::pair<Address, llvm::Value *>, 16> StoreGEPList;
+  // ^^ Flattened accesses to DestVal we want to store into
+  CGF.FlattenAccessAndType(DestVal, DestTy, StoreGEPList, DestTypes);
+  // Flatten our src
+  SmallVector<QualType, 16> SrcTypes; // Flattened type
+  SmallVector<std::pair<Address, llvm::Value *>, 16> LoadGEPList;
+  // ^^ Flattened accesses to SrcVal we want to load from
+  CGF.FlattenAccessAndType(SrcVal, SrcTy, LoadGEPList, SrcTypes);
+
+  assert(StoreGEPList.size() <= LoadGEPList.size() &&
+         "Cannot perform HLSL flat cast when flattened source object \
+          has less elements than flattened destination object.");
+  // apply casts to what we load from LoadGEPList
+  // and store result in Dest
+  for (unsigned I = 0, E = StoreGEPList.size(); I < E; I++) {
+    llvm::Value *Idx = LoadGEPList[I].second;
+    llvm::Value *Load = CGF.Builder.CreateLoad(LoadGEPList[I].first, "load");
+    Load =
+        Idx ? CGF.Builder.CreateExtractElement(Load, Idx, "vec.extract") : Load;
+    llvm::Value *Cast =
+        CGF.EmitScalarConversion(Load, SrcTypes[I], DestTypes[I], Loc);
+
+    // store back
+    Idx = StoreGEPList[I].second;
+    if (Idx) {
+      llvm::Value *V =
+          CGF.Builder.CreateLoad(StoreGEPList[I].first, "load.for.insert");
+      Cast = CGF.Builder.CreateInsertElement(V, Cast, Idx);
+    }
+    CGF.Builder.CreateStore(Cast, StoreGEPList[I].first);
+  }
+}
+
 /// Emit initialization of an array from an initializer list. ExprToVisit must
 /// be either an InitListEpxr a CXXParenInitListExpr.
 void AggExprEmitter::EmitArrayInit(Address DestPtr, llvm::ArrayType *AType,
@@ -890,7 +963,25 @@ void AggExprEmitter::VisitCastExpr(CastExpr *E) {
   case CK_HLSLArrayRValue:
     Visit(E->getSubExpr());
     break;
-
+  case CK_HLSLElementwiseCast: {
+    Expr *Src = E->getSubExpr();
+    QualType SrcTy = Src->getType();
+    RValue RV = CGF.EmitAnyExpr(Src);
+    QualType DestTy = E->getType();
+    Address DestVal = Dest.getAddress();
+    SourceLocation Loc = E->getExprLoc();
+
+    if (RV.isScalar()) {
+      llvm::Value *SrcVal = RV.getScalarVal();
+      EmitHLSLScalarFlatCast(CGF, DestVal, DestTy, SrcVal, SrcTy, Loc);
+    } else {
+      assert(RV.isAggregate() &&
+             "Can't perform HLSL Aggregate cast on a complex type.");
+      Address SrcVal = RV.getAggregateAddress();
+      EmitHLSLElementwiseCast(CGF, DestVal, DestTy, SrcVal, SrcTy, Loc);
+    }
+    break;
+  }
   case CK_NoOp:
   case CK_UserDefinedConversion:
   case CK_ConstructorConversion:
@@ -1461,6 +1552,7 @@ static bool castPreservesZero(const CastExpr *CE) {
   case CK_NonAtomicToAtomic:
   case CK_AtomicToNonAtomic:
   case CK_HLSLVectorTruncation:
+  case CK_HLSLElementwiseCast:
     return true;
 
   case CK_BaseToDerivedMemberPointer:

clang/lib/CodeGen/CGExprComplex.cpp

@@ -610,6 +610,7 @@ ComplexPairTy ComplexExprEmitter::EmitCast(CastKind CK, Expr *Op,
   case CK_MatrixCast:
   case CK_HLSLVectorTruncation:
   case CK_HLSLArrayRValue:
+  case CK_HLSLElementwiseCast:
     llvm_unreachable("invalid cast kind for complex value");
 
   case CK_FloatingRealToComplex:

clang/lib/CodeGen/CGExprConstant.cpp

@@ -1335,6 +1335,7 @@ class ConstExprEmitter
     case CK_MatrixCast:
     case CK_HLSLVectorTruncation:
     case CK_HLSLArrayRValue:
+    case CK_HLSLElementwiseCast:
       return nullptr;
     }
     llvm_unreachable("Invalid CastKind");

clang/lib/CodeGen/CGExprScalar.cpp

@@ -2269,6 +2269,42 @@ bool CodeGenFunction::ShouldNullCheckClassCastValue(const CastExpr *CE) {
   return true;
 }
 
+// RHS is an aggregate type
+static Value *EmitHLSLElementwiseCast(CodeGenFunction &CGF, Address RHSVal,
+                                      QualType RHSTy, QualType LHSTy,
+                                      SourceLocation Loc) {
+  SmallVector<std::pair<Address, llvm::Value *>, 16> LoadGEPList;
+  SmallVector<QualType, 16> SrcTypes; // Flattened type
+  CGF.FlattenAccessAndType(RHSVal, RHSTy, LoadGEPList, SrcTypes);
+  // LHS is either a vector or a builtin?
+  // if its a vector create a temp alloca to store into and return that
+  if (auto *VecTy = LHSTy->getAs<VectorType>()) {
+    assert(SrcTypes.size() >= VecTy->getNumElements() &&
+           "Flattened type on RHS must have more elements than vector on LHS.");
+    llvm::Value *V =
+        CGF.Builder.CreateLoad(CGF.CreateIRTemp(LHSTy, "flatcast.tmp"));
+    // write to V.
+    for (unsigned I = 0, E = VecTy->getNumElements(); I < E; I++) {
+      llvm::Value *Load = CGF.Builder.CreateLoad(LoadGEPList[I].first, "load");
+      llvm::Value *Idx = LoadGEPList[I].second;
+      Load = Idx ? CGF.Builder.CreateExtractElement(Load, Idx, "vec.extract")
+                 : Load;
+      llvm::Value *Cast = CGF.EmitScalarConversion(
+          Load, SrcTypes[I], VecTy->getElementType(), Loc);
+      V = CGF.Builder.CreateInsertElement(V, Cast, I);
+    }
+    return V;
+  }
+  // i its a builtin just do an extract element or load.
+  assert(LHSTy->isBuiltinType() &&
+         "Destination type must be a vector or builtin type.");
+  llvm::Value *Load = CGF.Builder.CreateLoad(LoadGEPList[0].first, "load");
+  llvm::Value *Idx = LoadGEPList[0].second;
+  Load =
+      Idx ? CGF.Builder.CreateExtractElement(Load, Idx, "vec.extract") : Load;
+  return CGF.EmitScalarConversion(Load, LHSTy, SrcTypes[0], Loc);
+}
+
 // VisitCastExpr - Emit code for an explicit or implicit cast.  Implicit casts
 // have to handle a more broad range of conversions than explicit casts, as they
 // handle things like function to ptr-to-function decay etc.
@@ -2759,7 +2795,16 @@ Value *ScalarExprEmitter::VisitCastExpr(CastExpr *CE) {
     llvm::Value *Zero = llvm::Constant::getNullValue(CGF.SizeTy);
     return Builder.CreateExtractElement(Vec, Zero, "cast.vtrunc");
   }
+  case CK_HLSLElementwiseCast: {
+    RValue RV = CGF.EmitAnyExpr(E);
+    SourceLocation Loc = CE->getExprLoc();
+    QualType SrcTy = E->getType();
 
+    assert(RV.isAggregate() && "Not a valid HLSL Flat Cast.");
+    // RHS is an aggregate
+    Address SrcVal = RV.getAggregateAddress();
+    return EmitHLSLElementwiseCast(CGF, SrcVal, SrcTy, DestTy, Loc);
+  }
   } // end of switch
 
   llvm_unreachable("unknown scalar cast");

clang/lib/CodeGen/CodeGenFunction.h

@@ -4439,6 +4439,11 @@ class CodeGenFunction : public CodeGenTypeCache {
                                 AggValueSlot slot = AggValueSlot::ignored());
   LValue EmitPseudoObjectLValue(const PseudoObjectExpr *e);
 
+  void FlattenAccessAndType(
+      Address Addr, QualType AddrTy,
+      SmallVectorImpl<std::pair<Address, llvm::Value *>> &AccessList,
+      SmallVectorImpl<QualType> &FlatTypes);
+
   llvm::Value *EmitIvarOffset(const ObjCInterfaceDecl *Interface,
                               const ObjCIvarDecl *Ivar);
   llvm::Value *EmitIvarOffsetAsPointerDiff(const ObjCInterfaceDecl *Interface,