[SYCL][FPGA] Refactor [[intel::max_work_group_size()]] attribute implementation

clang/include/clang/Basic/Attr.td

@@ -1406,17 +1406,20 @@ def SYCLIntelMaxWorkGroupSize : InheritableAttr {
   let LangOpts = [SYCLIsDevice, SYCLIsHost];
   let Subjects = SubjectList<[Function], ErrorDiag>;
   let AdditionalMembers = [{
-    ArrayRef<const Expr *> dimensions() const {
-      return {getXDim(), getYDim(), getZDim()};
-    }
-    Optional<llvm::APSInt> getXDimVal(ASTContext &Ctx) const {
-      return getXDim()->getIntegerConstantExpr(Ctx);
+    Optional<llvm::APSInt> getXDimVal() const {
+      if (const auto *CE = dyn_cast<ConstantExpr>(getXDim()))
+        return CE->getResultAsAPSInt();
+      return None;
     }
-    Optional<llvm::APSInt> getYDimVal(ASTContext &Ctx) const {
-      return getYDim()->getIntegerConstantExpr(Ctx);
+    Optional<llvm::APSInt> getYDimVal() const {
+      if (const auto *CE = dyn_cast<ConstantExpr>(getYDim()))
+        return CE->getResultAsAPSInt();
+      return None;
     }
-    Optional<llvm::APSInt> getZDimVal(ASTContext &Ctx) const {
-      return getZDim()->getIntegerConstantExpr(Ctx);
+    Optional<llvm::APSInt> getZDimVal() const {
+      if (const auto *CE = dyn_cast<ConstantExpr>(getZDim()))
+        return CE->getResultAsAPSInt();
+      return None;
     }
   }];
   let Documentation = [SYCLIntelMaxWorkGroupSizeAttrDocs];

clang/include/clang/Sema/Sema.h

@@ -10559,6 +10559,11 @@ class Sema final {
                                                 const SYCLUsesAspectsAttr &A);
   void AddSYCLUsesAspectsAttr(Decl *D, const AttributeCommonInfo &CI,
                               Expr **Exprs, unsigned Size);
+  void AddSYCLIntelMaxWorkGroupSizeAttr(Decl *D, const AttributeCommonInfo &CI,
+                                        Expr *XDim, Expr *YDim, Expr *ZDim);
+  SYCLIntelMaxWorkGroupSizeAttr *
+  MergeSYCLIntelMaxWorkGroupSizeAttr(Decl *D,
+                                     const SYCLIntelMaxWorkGroupSizeAttr &A);
   /// AddAlignedAttr - Adds an aligned attribute to a particular declaration.
   void AddAlignedAttr(Decl *D, const AttributeCommonInfo &CI, Expr *E,
                       bool IsPackExpansion);

clang/lib/CodeGen/CodeGenFunction.cpp

@@ -743,24 +743,16 @@ void CodeGenFunction::EmitOpenCLKernelMetadata(const FunctionDecl *FD,
 
   if (const SYCLIntelMaxWorkGroupSizeAttr *A =
           FD->getAttr<SYCLIntelMaxWorkGroupSizeAttr>()) {
-    ASTContext &ClangCtx = FD->getASTContext();
-    Optional<llvm::APSInt> XDimVal = A->getXDimVal(ClangCtx);
-    Optional<llvm::APSInt> YDimVal = A->getYDimVal(ClangCtx);
-    Optional<llvm::APSInt> ZDimVal = A->getZDimVal(ClangCtx);
 
-    // For a SYCLDevice SYCLIntelMaxWorkGroupSizeAttr arguments are reversed.
-    if (getLangOpts().SYCLIsDevice)
-      std::swap(XDimVal, ZDimVal);
-
-    llvm::Metadata *AttrMDArgs[] = {
-        llvm::ConstantAsMetadata::get(
-            Builder.getInt32(XDimVal->getZExtValue())),
-        llvm::ConstantAsMetadata::get(
-            Builder.getInt32(YDimVal->getZExtValue())),
-        llvm::ConstantAsMetadata::get(
-            Builder.getInt32(ZDimVal->getZExtValue()))};
-    Fn->setMetadata("max_work_group_size",
-                    llvm::MDNode::get(Context, AttrMDArgs));
+    // Attributes arguments (first and third) are reversed on SYCLDevice.
+    if (getLangOpts().SYCLIsDevice) {
+      llvm::Metadata *AttrMDArgs[] = {
+          llvm::ConstantAsMetadata::get(Builder.getInt(*A->getZDimVal())),
+          llvm::ConstantAsMetadata::get(Builder.getInt(*A->getYDimVal())),
+          llvm::ConstantAsMetadata::get(Builder.getInt(*A->getXDimVal()))};
+      Fn->setMetadata("max_work_group_size",
+                      llvm::MDNode::get(Context, AttrMDArgs));
+    }
   }
 
   if (const auto *A = FD->getAttr<SYCLIntelNoGlobalWorkOffsetAttr>()) {

clang/lib/Sema/SemaDecl.cpp

@@ -2732,6 +2732,8 @@ static bool mergeDeclAttribute(Sema &S, NamedDecl *D,
     NewAttr = S.MergeSYCLUsesAspectsAttr(D, *A);
   else if (const auto *A = dyn_cast<SYCLIntelPipeIOAttr>(Attr))
     NewAttr = S.MergeSYCLIntelPipeIOAttr(D, *A);
+  else if (const auto *A = dyn_cast<SYCLIntelMaxWorkGroupSizeAttr>(Attr))
+    NewAttr = S.MergeSYCLIntelMaxWorkGroupSizeAttr(D, *A);
   else if (Attr->shouldInheritEvenIfAlreadyPresent() || !DeclHasAttr(D, Attr))
     NewAttr = cast<InheritableAttr>(Attr->clone(S.Context));
 
@@ -3433,8 +3435,6 @@ bool Sema::MergeFunctionDecl(FunctionDecl *New, NamedDecl *&OldD,
 
     checkDimensionsAndSetDiagnostics<ReqdWorkGroupSizeAttr>(*this, New, Old);
 
-    checkDimensionsAndSetDiagnostics<SYCLIntelMaxWorkGroupSizeAttr>(*this, New,
-                                                                    Old);
   if (const auto *ILA = New->getAttr<InternalLinkageAttr>())
     if (!Old->hasAttr<InternalLinkageAttr>()) {
       Diag(New->getLocation(), diag::err_attribute_missing_on_first_decl)

clang/lib/Sema/SemaDeclAttr.cpp

@@ -3205,12 +3205,9 @@ static bool checkWorkGroupSizeValues(Sema &S, Decl *D, const ParsedAttr &AL) {
   ASTContext &Ctx = S.getASTContext();
 
   if (const auto *A = D->getAttr<SYCLIntelMaxWorkGroupSizeAttr>()) {
-    if (!((getExprValue(AL.getArgAsExpr(0), Ctx) <=
-           getExprValue(A->getXDim(), Ctx)) &&
-          (getExprValue(AL.getArgAsExpr(1), Ctx) <=
-           getExprValue(A->getYDim(), Ctx)) &&
-          (getExprValue(AL.getArgAsExpr(2), Ctx) <=
-           getExprValue(A->getZDim(), Ctx)))) {
+    if (!((getExprValue(AL.getArgAsExpr(0), Ctx) <= *A->getXDimVal()) &&
+          (getExprValue(AL.getArgAsExpr(1), Ctx) <= *A->getYDimVal()) &&
+          (getExprValue(AL.getArgAsExpr(2), Ctx) <= *A->getZDimVal()))) {
       S.Diag(AL.getLoc(), diag::err_conflicting_sycl_function_attributes)
           << AL << A->getSpelling();
       Result &= false;
@@ -3232,19 +3229,18 @@ static bool checkWorkGroupSizeValues(Sema &S, Decl *D, const ParsedAttr &AL) {
   return Result;
 }
 
-// Handles reqd_work_group_size and max_work_group_size.
+// Handles reqd_work_group_size.
 template <typename WorkGroupAttr>
 static void handleWorkGroupSize(Sema &S, Decl *D, const ParsedAttr &AL) {
   if (D->isInvalidDecl())
     return;
 
   S.CheckDeprecatedSYCLAttributeSpelling(AL);
-  // __attribute__((reqd_work_group_size)), [[cl::reqd_work_group_size]], and
-  // [[intel::max_work_group_size]] all require exactly three arguments.
+  // __attribute__((reqd_work_group_size)) and [[cl::reqd_work_group_size]]
+  // all require exactly three arguments.
   if ((AL.getKind() == ParsedAttr::AT_ReqdWorkGroupSize &&
        AL.getAttributeSpellingListIndex() ==
            ReqdWorkGroupSizeAttr::CXX11_cl_reqd_work_group_size) ||
-      AL.getKind() == ParsedAttr::AT_SYCLIntelMaxWorkGroupSize ||
       AL.getSyntax() == ParsedAttr::AS_GNU) {
     if (!AL.checkExactlyNumArgs(S, 3))
       return;
@@ -3308,8 +3304,8 @@ static void handleWorkGroupSize(Sema &S, Decl *D, const ParsedAttr &AL) {
       }
     }
 
-    // If the declaration has a SYCLIntelMaxWorkGroupSizeAttr or
-    // ReqdWorkGroupSizeAttr, check to see if they hold equal values
+    // If the declaration has a ReqdWorkGroupSizeAttr,
+    // check to see if they hold equal values
     // (1, 1, 1) in case the value of SYCLIntelMaxGlobalWorkDimAttr
     // equals to 0.
     if (const auto *DeclAttr = D->getAttr<SYCLIntelMaxGlobalWorkDimAttr>()) {
@@ -3454,6 +3450,146 @@ static void handleWorkGroupSizeHint(Sema &S, Decl *D, const ParsedAttr &AL) {
                              AL.getArgAsExpr(2));
 }
 
+// Handles max_work_group_size attribute.
+// If the [[intel::max_work_group_size(X, Y, Z)]] attribute is specified on a
+// declaration along with [[intel::max_global_work_dim()]] attribute,
+// check to see if all arguments of [[intel::max_work_group_size(X, Y, Z)]]
+// attribute hold value 1 in case the argument of
+// [[intel::max_global_work_dim()]] attribute equals to 0.
+static bool InvalidWorkGroupSizeAttrs(const Expr *MGValue, const Expr *XDim,
+                                      const Expr *YDim, const Expr *ZDim) {
+  // If any of the operand is still value dependent, we can't test anything.
+  const auto *MGValueExpr = dyn_cast<ConstantExpr>(MGValue);
+  const auto *XDimExpr = dyn_cast<ConstantExpr>(XDim);
+  const auto *YDimExpr = dyn_cast<ConstantExpr>(YDim);
+  const auto *ZDimExpr = dyn_cast<ConstantExpr>(ZDim);
+
+  if (!MGValueExpr || !XDimExpr || !YDimExpr || !ZDimExpr)
+    return false;
+
+  // Otherwise, check if the attribute values are equal to one.
+  return (MGValueExpr->getResultAsAPSInt() == 0 &&
+          (XDimExpr->getResultAsAPSInt() != 1 ||
+           YDimExpr->getResultAsAPSInt() != 1 ||
+           ZDimExpr->getResultAsAPSInt() != 1));
+}
+
+void Sema::AddSYCLIntelMaxWorkGroupSizeAttr(Decl *D,
+                                            const AttributeCommonInfo &CI,
+                                            Expr *XDim, Expr *YDim,
+                                            Expr *ZDim) {
+  // Returns nullptr if diagnosing, otherwise returns the original expression
+  // or the original expression converted to a constant expression.
+  auto CheckAndConvertArg = [&](Expr *E) -> Expr * {
+    // Check if the expression is not value dependent.
+    if (!E->isValueDependent()) {
+      llvm::APSInt ArgVal;
+      ExprResult Res = VerifyIntegerConstantExpression(E, &ArgVal);
+      if (Res.isInvalid())
+        return nullptr;
+      E = Res.get();
+
+      // This attribute requires a strictly positive value.
+      if (ArgVal <= 0) {
+        Diag(E->getExprLoc(), diag::err_attribute_requires_positive_integer)
+            << CI << /*positive*/ 0;
+        return nullptr;
+      }
+    }
+    return E;
+  };
+
+  // Check all three argument values, and if any are bad, bail out. This will
+  // convert the given expressions into constant expressions when possible.
+  XDim = CheckAndConvertArg(XDim);
+  YDim = CheckAndConvertArg(YDim);
+  ZDim = CheckAndConvertArg(ZDim);
+  if (!XDim || !YDim || !ZDim)
+    return;
+
+  // If the declaration has a SYCLIntelMaxWorkGroupSizeAttr, check to see if
+  // the attribute holds equal values to (1, 1, 1) in case the value of
+  // SYCLIntelMaxGlobalWorkDimAttr equals to 0.
+  if (const auto *DeclAttr = D->getAttr<SYCLIntelMaxGlobalWorkDimAttr>()) {
+    if (InvalidWorkGroupSizeAttrs(DeclAttr->getValue(), XDim, YDim, ZDim)) {
+      Diag(CI.getLoc(), diag::err_sycl_x_y_z_arguments_must_be_one)
+          << CI << DeclAttr;
+      return;
+    }
+  }
+
+  // If the attribute was already applied with different arguments, then
+  // diagnose the second attribute as a duplicate and don't add it.
+  if (const auto *Existing = D->getAttr<SYCLIntelMaxWorkGroupSizeAttr>()) {
+    DupArgResult Results[] = {AreArgValuesIdentical(XDim, Existing->getXDim()),
+                              AreArgValuesIdentical(YDim, Existing->getYDim()),
+                              AreArgValuesIdentical(ZDim, Existing->getZDim())};
+    // If any of the results are known to be different, we can diagnose at this
+    // point and drop the attribute.
+    if (llvm::is_contained(Results, DupArgResult::Different)) {
+      Diag(CI.getLoc(), diag::warn_duplicate_attribute) << CI;
+      Diag(Existing->getLoc(), diag::note_previous_attribute);
+      return;
+    }
+    // If all of the results are known to be the same, we can silently drop the
+    // attribute. Otherwise, we have to add the attribute and resolve its
+    // differences later.
+    if (llvm::all_of(Results,
+                     [](DupArgResult V) { return V == DupArgResult::Same; }))
+      return;
+  }
+
+  D->addAttr(::new (Context)
+                 SYCLIntelMaxWorkGroupSizeAttr(Context, CI, XDim, YDim, ZDim));
+}
+
+SYCLIntelMaxWorkGroupSizeAttr *Sema::MergeSYCLIntelMaxWorkGroupSizeAttr(
+    Decl *D, const SYCLIntelMaxWorkGroupSizeAttr &A) {
+  // Check to see if there's a duplicate attribute already applied.
+  if (const auto *DeclAttr = D->getAttr<SYCLIntelMaxWorkGroupSizeAttr>()) {
+    DupArgResult Results[] = {
+        AreArgValuesIdentical(DeclAttr->getXDim(), A.getXDim()),
+        AreArgValuesIdentical(DeclAttr->getYDim(), A.getYDim()),
+        AreArgValuesIdentical(DeclAttr->getZDim(), A.getZDim())};
+
+    // If any of the results are known to be different, we can diagnose at this
+    // point and drop the attribute.
+    if (llvm::is_contained(Results, DupArgResult::Different)) {
+      Diag(DeclAttr->getLoc(), diag::warn_duplicate_attribute) << &A;
+      Diag(A.getLoc(), diag::note_previous_attribute);
+      return nullptr;
+    }
+    // If all of the results are known to be the same, we can silently drop the
+    // attribute. Otherwise, we have to add the attribute and resolve its
+    // differences later.
+    if (llvm::all_of(Results,
+                     [](DupArgResult V) { return V == DupArgResult::Same; }))
+      return nullptr;
+  }
+
+  // If the declaration has a SYCLIntelMaxWorkGroupSizeAttr,
+  // check to see if the attribute holds equal values to
+  // (1, 1, 1) in case the value of SYCLIntelMaxGlobalWorkDimAttr
+  // equals to 0.
+  if (const auto *DeclAttr = D->getAttr<SYCLIntelMaxGlobalWorkDimAttr>()) {
+    if (InvalidWorkGroupSizeAttrs(DeclAttr->getValue(), A.getXDim(),
+                                  A.getYDim(), A.getZDim())) {
+      Diag(A.getLoc(), diag::err_sycl_x_y_z_arguments_must_be_one)
+          << &A << DeclAttr;
+      return nullptr;
+    }
+  }
+
+  return ::new (Context) SYCLIntelMaxWorkGroupSizeAttr(
+      Context, A, A.getXDim(), A.getYDim(), A.getZDim());
+}
+
+static void handleSYCLIntelMaxWorkGroupSize(Sema &S, Decl *D,
+                                            const ParsedAttr &AL) {
+  S.AddSYCLIntelMaxWorkGroupSizeAttr(D, AL, AL.getArgAsExpr(0),
+                                     AL.getArgAsExpr(1), AL.getArgAsExpr(2));
+}
+
 void Sema::AddIntelReqdSubGroupSize(Decl *D, const AttributeCommonInfo &CI,
                                     Expr *E) {
   if (!E->isValueDependent()) {
@@ -10331,7 +10467,7 @@ static void ProcessDeclAttribute(Sema &S, Scope *scope, Decl *D,
     handleWorkGroupSize<ReqdWorkGroupSizeAttr>(S, D, AL);
     break;
   case ParsedAttr::AT_SYCLIntelMaxWorkGroupSize:
-    handleWorkGroupSize<SYCLIntelMaxWorkGroupSizeAttr>(S, D, AL);
+    handleSYCLIntelMaxWorkGroupSize(S, D, AL);
     break;
   case ParsedAttr::AT_IntelReqdSubGroupSize:
     handleIntelReqdSubGroupSize(S, D, AL);

clang/lib/Sema/SemaSYCL.cpp

@@ -3889,9 +3889,9 @@ static void PropagateAndDiagnoseDeviceAttr(
     } else if (auto *Existing =
                    SYCLKernel->getAttr<SYCLIntelMaxWorkGroupSizeAttr>()) {
       ASTContext &Ctx = S.getASTContext();
-      if (Existing->getXDimVal(Ctx) < RWGSA->getXDimVal(Ctx) ||
-          Existing->getYDimVal(Ctx) < RWGSA->getYDimVal(Ctx) ||
-          Existing->getZDimVal(Ctx) < RWGSA->getZDimVal(Ctx)) {
+      if (*Existing->getXDimVal() < RWGSA->getXDimVal(Ctx) ||
+          *Existing->getYDimVal() < RWGSA->getYDimVal(Ctx) ||
+          *Existing->getZDimVal() < RWGSA->getZDimVal(Ctx)) {
         S.Diag(SYCLKernel->getLocation(),
                diag::err_conflicting_sycl_kernel_attributes);
         S.Diag(Existing->getLocation(), diag::note_conflicting_attribute);
@@ -3909,9 +3909,9 @@ static void PropagateAndDiagnoseDeviceAttr(
     auto *SIMWGSA = cast<SYCLIntelMaxWorkGroupSizeAttr>(A);
     if (auto *Existing = SYCLKernel->getAttr<ReqdWorkGroupSizeAttr>()) {
       ASTContext &Ctx = S.getASTContext();
-      if (Existing->getXDimVal(Ctx) > SIMWGSA->getXDimVal(Ctx) ||
-          Existing->getYDimVal(Ctx) > SIMWGSA->getYDimVal(Ctx) ||
-          Existing->getZDimVal(Ctx) > SIMWGSA->getZDimVal(Ctx)) {
+      if (Existing->getXDimVal(Ctx) > *SIMWGSA->getXDimVal() ||
+          Existing->getYDimVal(Ctx) > *SIMWGSA->getYDimVal() ||
+          Existing->getZDimVal(Ctx) > *SIMWGSA->getZDimVal()) {
         S.Diag(SYCLKernel->getLocation(),
                diag::err_conflicting_sycl_kernel_attributes);
         S.Diag(Existing->getLocation(), diag::note_conflicting_attribute);

clang/lib/Sema/SemaTemplateInstantiateDecl.cpp

@@ -814,6 +814,25 @@ static void instantiateWorkGroupSizeHintAttr(
                              ZResult.get());
 }
 
+static void instantiateSYCLIntelMaxWorkGroupSizeAttr(
+    Sema &S, const MultiLevelTemplateArgumentList &TemplateArgs,
+    const SYCLIntelMaxWorkGroupSizeAttr *A, Decl *New) {
+  EnterExpressionEvaluationContext Unevaluated(
+      S, Sema::ExpressionEvaluationContext::ConstantEvaluated);
+  ExprResult XResult = S.SubstExpr(A->getXDim(), TemplateArgs);
+  if (XResult.isInvalid())
+    return;
+  ExprResult YResult = S.SubstExpr(A->getYDim(), TemplateArgs);
+  if (YResult.isInvalid())
+    return;
+  ExprResult ZResult = S.SubstExpr(A->getZDim(), TemplateArgs);
+  if (ZResult.isInvalid())
+    return;
+
+  S.AddSYCLIntelMaxWorkGroupSizeAttr(New, *A, XResult.get(), YResult.get(),
+                                     ZResult.get());
+}
+
 // This doesn't take any template parameters, but we have a custom action that
 // needs to happen when the kernel itself is instantiated. We need to run the
 // ItaniumMangler to mark the names required to name this kernel.
@@ -1045,8 +1064,8 @@ void Sema::InstantiateAttrs(const MultiLevelTemplateArgumentList &TemplateArgs,
     }
     if (const auto *SYCLIntelMaxWorkGroupSize =
             dyn_cast<SYCLIntelMaxWorkGroupSizeAttr>(TmplAttr)) {
-      instantiateIntelSYCTripleLFunctionAttr<SYCLIntelMaxWorkGroupSizeAttr>(
-          *this, TemplateArgs, SYCLIntelMaxWorkGroupSize, New);
+      instantiateSYCLIntelMaxWorkGroupSizeAttr(*this, TemplateArgs,
+                                               SYCLIntelMaxWorkGroupSize, New);
       continue;
     }
     if (const auto *SYCLIntelMaxConcurrency =