[HLSL] Vector Usual Arithmetic Conversions (llvm#110195)

HLSL has a different set of usual arithmetic conversions for vector
types to resolve a common type for binary operator expressions.

This PR implements the current spec proposal from:
microsoft/hlsl-specs#311

There is one case that may need additional handling for implicitly
truncating vector<T,1> to T early to allow other transformations.

Fixes llvm#106253

Re-lands llvm#108659

clang/include/clang/Basic/DiagnosticSemaKinds.td

@@ -12395,6 +12395,9 @@ def err_hlsl_operator_unsupported : Error<
 
 def err_hlsl_param_qualifier_mismatch :
   Error<"conflicting parameter qualifier %0 on parameter %1">;
+def err_hlsl_vector_compound_assignment_truncation : Error<
+  "left hand operand of type %0 to compound assignment cannot be truncated "
+  "when used with right hand operand of type %1">;
 
 def warn_hlsl_impcast_vector_truncation : Warning<
   "implicit conversion truncates vector: %0 to %1">, InGroup<Conversion>;

clang/include/clang/Driver/Options.td

@@ -2979,7 +2979,7 @@ def flax_vector_conversions_EQ : Joined<["-"], "flax-vector-conversions=">, Grou
                     "LangOptions::LaxVectorConversionKind::Integer",
                     "LangOptions::LaxVectorConversionKind::All"]>,
   MarshallingInfoEnum<LangOpts<"LaxVectorConversions">,
-                      open_cl.KeyPath #
+                      !strconcat("(", open_cl.KeyPath, " || ", hlsl.KeyPath, ")") #
                           " ? LangOptions::LaxVectorConversionKind::None" #
                           " : LangOptions::LaxVectorConversionKind::All">;
 def flax_vector_conversions : Flag<["-"], "flax-vector-conversions">, Group<f_Group>,

clang/include/clang/Sema/Sema.h

@@ -7423,7 +7423,8 @@ class Sema final : public SemaBase {
                                               SourceLocation Loc,
                                               BinaryOperatorKind Opc);
   QualType CheckVectorLogicalOperands(ExprResult &LHS, ExprResult &RHS,
-                                      SourceLocation Loc);
+                                      SourceLocation Loc,
+                                      BinaryOperatorKind Opc);
 
   /// Context in which we're performing a usual arithmetic conversion.
   enum ArithConvKind {

clang/include/clang/Sema/SemaHLSL.h

@@ -63,6 +63,11 @@ class SemaHLSL : public SemaBase {
       std::initializer_list<llvm::Triple::EnvironmentType> AllowedStages);
   void DiagnoseAvailabilityViolations(TranslationUnitDecl *TU);
 
+  QualType handleVectorBinOpConversion(ExprResult &LHS, ExprResult &RHS,
+                                       QualType LHSType, QualType RHSType,
+                                       bool IsCompAssign);
+  void emitLogicalOperatorFixIt(Expr *LHS, Expr *RHS, BinaryOperatorKind Opc);
+
   void handleNumThreadsAttr(Decl *D, const ParsedAttr &AL);
   void handleWaveSizeAttr(Decl *D, const ParsedAttr &AL);
   void handleSV_DispatchThreadIDAttr(Decl *D, const ParsedAttr &AL);

clang/lib/Sema/SemaExpr.cpp

@@ -10133,6 +10133,10 @@ QualType Sema::CheckVectorOperands(ExprResult &LHS, ExprResult &RHS,
   const VectorType *RHSVecType = RHSType->getAs<VectorType>();
   assert(LHSVecType || RHSVecType);
 
+  if (getLangOpts().HLSL)
+    return HLSL().handleVectorBinOpConversion(LHS, RHS, LHSType, RHSType,
+                                              IsCompAssign);
+
   // AltiVec-style "vector bool op vector bool" combinations are allowed
   // for some operators but not others.
   if (!AllowBothBool && LHSVecType &&
@@ -12863,7 +12867,8 @@ static void diagnoseXorMisusedAsPow(Sema &S, const ExprResult &XorLHS,
 }
 
 QualType Sema::CheckVectorLogicalOperands(ExprResult &LHS, ExprResult &RHS,
-                                          SourceLocation Loc) {
+                                          SourceLocation Loc,
+                                          BinaryOperatorKind Opc) {
   // Ensure that either both operands are of the same vector type, or
   // one operand is of a vector type and the other is of its element type.
   QualType vType = CheckVectorOperands(LHS, RHS, Loc, false,
@@ -12883,6 +12888,15 @@ QualType Sema::CheckVectorLogicalOperands(ExprResult &LHS, ExprResult &RHS,
   if (!getLangOpts().CPlusPlus &&
       !(isa<ExtVectorType>(vType->getAs<VectorType>())))
     return InvalidLogicalVectorOperands(Loc, LHS, RHS);
+  // Beginning with HLSL 2021, HLSL disallows logical operators on vector
+  // operands and instead requires the use of the `and`, `or`, `any`, `all`, and
+  // `select` functions.
+  if (getLangOpts().HLSL &&
+      getLangOpts().getHLSLVersion() >= LangOptionsBase::HLSL_2021) {
+    (void)InvalidOperands(Loc, LHS, RHS);
+    HLSL().emitLogicalOperatorFixIt(LHS.get(), RHS.get(), Opc);
+    return QualType();
+  }
 
   return GetSignedVectorType(LHS.get()->getType());
 }
@@ -13054,7 +13068,7 @@ inline QualType Sema::CheckLogicalOperands(ExprResult &LHS, ExprResult &RHS,
   // Check vector operands differently.
   if (LHS.get()->getType()->isVectorType() ||
       RHS.get()->getType()->isVectorType())
-    return CheckVectorLogicalOperands(LHS, RHS, Loc);
+    return CheckVectorLogicalOperands(LHS, RHS, Loc, Opc);
 
   bool EnumConstantInBoolContext = false;
   for (const ExprResult &HS : {LHS, RHS}) {

clang/lib/Sema/SemaHLSL.cpp

@@ -401,6 +401,194 @@ void SemaHLSL::DiagnoseAttrStageMismatch(
       << (AllowedStages.size() != 1) << join(StageStrings, ", ");
 }
 
+template <CastKind Kind>
+static void castVector(Sema &S, ExprResult &E, QualType &Ty, unsigned Sz) {
+  if (const auto *VTy = Ty->getAs<VectorType>())
+    Ty = VTy->getElementType();
+  Ty = S.getASTContext().getExtVectorType(Ty, Sz);
+  E = S.ImpCastExprToType(E.get(), Ty, Kind);
+}
+
+template <CastKind Kind>
+static QualType castElement(Sema &S, ExprResult &E, QualType Ty) {
+  E = S.ImpCastExprToType(E.get(), Ty, Kind);
+  return Ty;
+}
+
+static QualType handleFloatVectorBinOpConversion(
+    Sema &SemaRef, ExprResult &LHS, ExprResult &RHS, QualType LHSType,
+    QualType RHSType, QualType LElTy, QualType RElTy, bool IsCompAssign) {
+  bool LHSFloat = LElTy->isRealFloatingType();
+  bool RHSFloat = RElTy->isRealFloatingType();
+
+  if (LHSFloat && RHSFloat) {
+    if (IsCompAssign ||
+        SemaRef.getASTContext().getFloatingTypeOrder(LElTy, RElTy) > 0)
+      return castElement<CK_FloatingCast>(SemaRef, RHS, LHSType);
+
+    return castElement<CK_FloatingCast>(SemaRef, LHS, RHSType);
+  }
+
+  if (LHSFloat)
+    return castElement<CK_IntegralToFloating>(SemaRef, RHS, LHSType);
+
+  assert(RHSFloat);
+  if (IsCompAssign)
+    return castElement<clang::CK_FloatingToIntegral>(SemaRef, RHS, LHSType);
+
+  return castElement<CK_IntegralToFloating>(SemaRef, LHS, RHSType);
+}
+
+static QualType handleIntegerVectorBinOpConversion(
+    Sema &SemaRef, ExprResult &LHS, ExprResult &RHS, QualType LHSType,
+    QualType RHSType, QualType LElTy, QualType RElTy, bool IsCompAssign) {
+
+  int IntOrder = SemaRef.Context.getIntegerTypeOrder(LElTy, RElTy);
+  bool LHSSigned = LElTy->hasSignedIntegerRepresentation();
+  bool RHSSigned = RElTy->hasSignedIntegerRepresentation();
+  auto &Ctx = SemaRef.getASTContext();
+
+  // If both types have the same signedness, use the higher ranked type.
+  if (LHSSigned == RHSSigned) {
+    if (IsCompAssign || IntOrder >= 0)
+      return castElement<CK_IntegralCast>(SemaRef, RHS, LHSType);
+
+    return castElement<CK_IntegralCast>(SemaRef, LHS, RHSType);
+  }
+
+  // If the unsigned type has greater than or equal rank of the signed type, use
+  // the unsigned type.
+  if (IntOrder != (LHSSigned ? 1 : -1)) {
+    if (IsCompAssign || RHSSigned)
+      return castElement<CK_IntegralCast>(SemaRef, RHS, LHSType);
+    return castElement<CK_IntegralCast>(SemaRef, LHS, RHSType);
+  }
+
+  // At this point the signed type has higher rank than the unsigned type, which
+  // means it will be the same size or bigger. If the signed type is bigger, it
+  // can represent all the values of the unsigned type, so select it.
+  if (Ctx.getIntWidth(LElTy) != Ctx.getIntWidth(RElTy)) {
+    if (IsCompAssign || LHSSigned)
+      return castElement<CK_IntegralCast>(SemaRef, RHS, LHSType);
+    return castElement<CK_IntegralCast>(SemaRef, LHS, RHSType);
+  }
+
+  // This is a bit of an odd duck case in HLSL. It shouldn't happen, but can due
+  // to C/C++ leaking through. The place this happens today is long vs long
+  // long. When arguments are vector<unsigned long, N> and vector<long long, N>,
+  // the long long has higher rank than long even though they are the same size.
+
+  // If this is a compound assignment cast the right hand side to the left hand
+  // side's type.
+  if (IsCompAssign)
+    return castElement<CK_IntegralCast>(SemaRef, RHS, LHSType);
+
+  // If this isn't a compound assignment we convert to unsigned long long.
+  QualType ElTy = Ctx.getCorrespondingUnsignedType(LHSSigned ? LElTy : RElTy);
+  QualType NewTy = Ctx.getExtVectorType(
+      ElTy, RHSType->castAs<VectorType>()->getNumElements());
+  (void)castElement<CK_IntegralCast>(SemaRef, RHS, NewTy);
+
+  return castElement<CK_IntegralCast>(SemaRef, LHS, NewTy);
+}
+
+static CastKind getScalarCastKind(ASTContext &Ctx, QualType DestTy,
+                                  QualType SrcTy) {
+  if (DestTy->isRealFloatingType() && SrcTy->isRealFloatingType())
+    return CK_FloatingCast;
+  if (DestTy->isIntegralType(Ctx) && SrcTy->isIntegralType(Ctx))
+    return CK_IntegralCast;
+  if (DestTy->isRealFloatingType())
+    return CK_IntegralToFloating;
+  assert(SrcTy->isRealFloatingType() && DestTy->isIntegralType(Ctx));
+  return CK_FloatingToIntegral;
+}
+
+QualType SemaHLSL::handleVectorBinOpConversion(ExprResult &LHS, ExprResult &RHS,
+                                               QualType LHSType,
+                                               QualType RHSType,
+                                               bool IsCompAssign) {
+  const auto *LVecTy = LHSType->getAs<VectorType>();
+  const auto *RVecTy = RHSType->getAs<VectorType>();
+  auto &Ctx = getASTContext();
+
+  // If the LHS is not a vector and this is a compound assignment, we truncate
+  // the argument to a scalar then convert it to the LHS's type.
+  if (!LVecTy && IsCompAssign) {
+    QualType RElTy = RHSType->castAs<VectorType>()->getElementType();
+    RHS = SemaRef.ImpCastExprToType(RHS.get(), RElTy, CK_HLSLVectorTruncation);
+    RHSType = RHS.get()->getType();
+    if (Ctx.hasSameUnqualifiedType(LHSType, RHSType))
+      return LHSType;
+    RHS = SemaRef.ImpCastExprToType(RHS.get(), LHSType,
+                                    getScalarCastKind(Ctx, LHSType, RHSType));
+    return LHSType;
+  }
+
+  unsigned EndSz = std::numeric_limits<unsigned>::max();
+  unsigned LSz = 0;
+  if (LVecTy)
+    LSz = EndSz = LVecTy->getNumElements();
+  if (RVecTy)
+    EndSz = std::min(RVecTy->getNumElements(), EndSz);
+  assert(EndSz != std::numeric_limits<unsigned>::max() &&
+         "one of the above should have had a value");
+
+  // In a compound assignment, the left operand does not change type, the right
+  // operand is converted to the type of the left operand.
+  if (IsCompAssign && LSz != EndSz) {
+    Diag(LHS.get()->getBeginLoc(),
+         diag::err_hlsl_vector_compound_assignment_truncation)
+        << LHSType << RHSType;
+    return QualType();
+  }
+
+  if (RVecTy && RVecTy->getNumElements() > EndSz)
+    castVector<CK_HLSLVectorTruncation>(SemaRef, RHS, RHSType, EndSz);
+  if (!IsCompAssign && LVecTy && LVecTy->getNumElements() > EndSz)
+    castVector<CK_HLSLVectorTruncation>(SemaRef, LHS, LHSType, EndSz);
+
+  if (!RVecTy)
+    castVector<CK_VectorSplat>(SemaRef, RHS, RHSType, EndSz);
+  if (!IsCompAssign && !LVecTy)
+    castVector<CK_VectorSplat>(SemaRef, LHS, LHSType, EndSz);
+
+  // If we're at the same type after resizing we can stop here.
+  if (Ctx.hasSameUnqualifiedType(LHSType, RHSType))
+    return Ctx.getCommonSugaredType(LHSType, RHSType);
+
+  QualType LElTy = LHSType->castAs<VectorType>()->getElementType();
+  QualType RElTy = RHSType->castAs<VectorType>()->getElementType();
+
+  // Handle conversion for floating point vectors.
+  if (LElTy->isRealFloatingType() || RElTy->isRealFloatingType())
+    return handleFloatVectorBinOpConversion(SemaRef, LHS, RHS, LHSType, RHSType,
+                                            LElTy, RElTy, IsCompAssign);
+
+  assert(LElTy->isIntegralType(Ctx) && RElTy->isIntegralType(Ctx) &&
+         "HLSL Vectors can only contain integer or floating point types");
+  return handleIntegerVectorBinOpConversion(SemaRef, LHS, RHS, LHSType, RHSType,
+                                            LElTy, RElTy, IsCompAssign);
+}
+
+void SemaHLSL::emitLogicalOperatorFixIt(Expr *LHS, Expr *RHS,
+                                        BinaryOperatorKind Opc) {
+  assert((Opc == BO_LOr || Opc == BO_LAnd) &&
+         "Called with non-logical operator");
+  llvm::SmallVector<char, 256> Buff;
+  llvm::raw_svector_ostream OS(Buff);
+  PrintingPolicy PP(SemaRef.getLangOpts());
+  StringRef NewFnName = Opc == BO_LOr ? "or" : "and";
+  OS << NewFnName << "(";
+  LHS->printPretty(OS, nullptr, PP);
+  OS << ", ";
+  RHS->printPretty(OS, nullptr, PP);
+  OS << ")";
+  SourceRange FullRange = SourceRange(LHS->getBeginLoc(), RHS->getEndLoc());
+  SemaRef.Diag(LHS->getBeginLoc(), diag::note_function_suggestion)
+      << NewFnName << FixItHint::CreateReplacement(FullRange, OS.str());
+}
+
 void SemaHLSL::handleNumThreadsAttr(Decl *D, const ParsedAttr &AL) {
   llvm::VersionTuple SMVersion =
       getASTContext().getTargetInfo().getTriple().getOSVersion();