[SYCL] Add support for __registered_kernels__ (#16485)

This change adds support for a new attribute
``__sycl_detail::__registered_kernels__``, which appears at translation
unit scope.  The parameter for this attribute is a list of pairs like:

```
[[__sycl_detail__::__registered_kernels__(
  {"foo", foo},
  {"(void(*)(int, int*))iota", (void(*)(int, int*))iota},
  {"kernel<float>", kernel<float>}
)]];
```

The first element in each pair is a string, and the second element is a
constant expressiton for a pointer to a SYCL free function kernel.

The change creates the kernel's wrapper function and generates
module-level metadata of the form:

```
!sycl_registered_kernels = !{!0, !1}
!0 = !{!"foo", !"mangled-name-of-wrapper-for-foo"}
!1 = !{!"kernel<float>", !"mangled-name-of-wrapper-for-kernel<float>"}
```

where the first element in the pair of strings, is the first element
of the pair in ``__registered_kernels__`` and the second element is the
mangled named of the wrapper corresponding to the free function.

clang/include/clang/Basic/Attr.td

@@ -2147,6 +2147,22 @@ def SYCLAddIRAnnotationsMember : InheritableAttr {
   let Documentation = [SYCLAddIRAnnotationsMemberDocs];
 }
 
+def SYCLRegisteredKernels : InheritableAttr {
+  let Spellings = [CXX11<"__sycl_detail__", "__registered_kernels__">];
+  let Args = [VariadicExprArgument<"Args">];
+  let LangOpts = [SYCLIsDevice, SilentlyIgnoreSYCLIsHost];
+  let Subjects = SubjectList<[Empty], ErrorDiag, "Translation Unit Scope">;
+  let AdditionalMembers = SYCLAddIRAttrCommonMembers.MemberCode;
+  let Documentation = [SYCLAddIRAnnotationsMemberDocs];
+}
+
+def SYCLRegisteredKernelName : InheritableAttr {
+  let Spellings = [];
+  let Subjects = SubjectList<[Function]>;
+  let Args = [StringArgument<"RegName">];
+  let Documentation = [InternalOnly];
+}
+
 def C11NoReturn : InheritableAttr {
   let Spellings = [CustomKeyword<"_Noreturn">];
   let Subjects = SubjectList<[Function], ErrorDiag>;

clang/include/clang/Basic/DiagnosticSemaKinds.td

@@ -12546,6 +12546,20 @@ def err_sycl_special_type_num_init_method : Error<
 def warn_launch_bounds_is_cuda_specific : Warning<
   "%0 attribute ignored, only applicable when targeting Nvidia devices">,
   InGroup<IgnoredAttributes>;
+def err_registered_kernels_num_of_args : Error<
+  "'__registered_kernels__' attribute must have at least one argument">;
+def err_registered_kernels_init_list : Error<
+  "argument to the '__registered_kernels__' attribute must be an "
+  "initializer list expression">;
+def err_registered_kernels_init_list_pair_values : Error<
+  "each initializer list argument to the '__registered_kernels__' attribute "
+  "must contain a pair of values">;
+def err_registered_kernels_resolve_function : Error<
+  "unable to resolve free function kernel '%0'">;
+def err_registered_kernels_name_already_registered : Error<
+  "free function kernel has already been registered with '%0'; cannot register with '%1'">;
+def err_not_sycl_free_function : Error<
+  "attempting to register a function that is not a SYCL free function as '%0'">;
 
 def warn_cuda_maxclusterrank_sm_90 : Warning<
   "'maxclusterrank' requires sm_90 or higher, CUDA arch provided: %0, ignoring "

clang/include/clang/Sema/SemaSYCL.h

@@ -252,8 +252,9 @@ class SemaSYCL : public SemaBase {
   // We need to store the list of the sycl_kernel functions and their associated
   // generated OpenCL Kernels so we can go back and re-name these after the
   // fact.
-  llvm::SmallVector<std::pair<const FunctionDecl *, FunctionDecl *>>
-      SyclKernelsToOpenCLKernels;
+  using KernelFDPairs =
+      llvm::SmallVector<std::pair<const FunctionDecl *, FunctionDecl *>>;
+  KernelFDPairs SyclKernelsToOpenCLKernels;
 
   // Used to suppress diagnostics during kernel construction, since these were
   // already emitted earlier. Diagnosing during Kernel emissions also skips the
@@ -296,11 +297,15 @@ class SemaSYCL : public SemaBase {
                               llvm::DenseSet<QualType> Visited,
                               ValueDecl *DeclToCheck);
 
+  const KernelFDPairs &getKernelFDPairs() { return SyclKernelsToOpenCLKernels; }
+
   void addSyclOpenCLKernel(const FunctionDecl *SyclKernel,
                            FunctionDecl *OpenCLKernel) {
     SyclKernelsToOpenCLKernels.emplace_back(SyclKernel, OpenCLKernel);
   }
 
+  void constructFreeFunctionKernel(FunctionDecl *FD, StringRef NameStr = "");
+
   void addSyclDeviceDecl(Decl *d) { SyclDeviceDecls.insert(d); }
   llvm::SetVector<Decl *> &syclDeviceDecls() { return SyclDeviceDecls; }
 
@@ -480,6 +485,7 @@ class SemaSYCL : public SemaBase {
   void handleSYCLIntelMaxWorkGroupsPerMultiprocessor(Decl *D,
                                                      const ParsedAttr &AL);
   void handleSYCLScopeAttr(Decl *D, const ParsedAttr &AL);
+  void handleSYCLRegisteredKernels(Decl *D, const ParsedAttr &AL);
 
   void checkSYCLAddIRAttributesFunctionAttrConflicts(Decl *D);
 
@@ -655,6 +661,10 @@ class SemaSYCL : public SemaBase {
   void addIntelReqdSubGroupSizeAttr(Decl *D, const AttributeCommonInfo &CI,
                                     Expr *E);
   void handleKernelEntryPointAttr(Decl *D, const ParsedAttr &AL);
+
+  // Used to check whether the function represented by FD is a SYCL
+  // free function kernel or not.
+  bool isFreeFunction(const FunctionDecl *FD);
 };
 
 } // namespace clang

clang/lib/CodeGen/CodeGenFunction.cpp

@@ -641,6 +641,12 @@ void CodeGenFunction::EmitKernelMetadata(const FunctionDecl *FD,
 
   llvm::LLVMContext &Context = getLLVMContext();
 
+  if (getLangOpts().SYCLIsDevice)
+    if (FD->hasAttr<SYCLRegisteredKernelNameAttr>())
+      CGM.SYCLAddRegKernelNamePairs(
+          FD->getAttr<SYCLRegisteredKernelNameAttr>()->getRegName(),
+          FD->getNameAsString());
+
   if (FD->hasAttr<OpenCLKernelAttr>() || FD->hasAttr<CUDAGlobalAttr>())
     CGM.GenKernelArgMetadata(Fn, FD, this);
 

clang/lib/CodeGen/CodeGenModule.cpp

@@ -1431,6 +1431,19 @@ void CodeGenModule::Release() {
         AspectEnumValsMD->addOperand(
             getAspectEnumValueMD(Context, TheModule.getContext(), ECD));
     }
+
+    if (!SYCLRegKernelNames.empty()) {
+      std::vector<llvm::Metadata *> Nodes;
+      llvm::LLVMContext &Ctx = TheModule.getContext();
+      for (auto MDKernelNames : SYCLRegKernelNames) {
+        llvm::Metadata *Vals[] = {MDKernelNames.first, MDKernelNames.second};
+        Nodes.push_back(llvm::MDTuple::get(Ctx, Vals));
+      }
+
+      llvm::NamedMDNode *SYCLRegKernelsMD =
+          TheModule.getOrInsertNamedMetadata("sycl_registered_kernels");
+      SYCLRegKernelsMD->addOperand(llvm::MDNode::get(Ctx, Nodes));
+    }
   }
 
   // HLSL related end of code gen work items.

clang/lib/CodeGen/CodeGenModule.h

@@ -456,6 +456,9 @@ class CodeGenModule : public CodeGenTypeCache {
   /// handled differently than regular annotations so they cannot share map.
   llvm::DenseMap<unsigned, llvm::Constant *> SYCLAnnotationArgs;
 
+  typedef std::pair<llvm::Metadata *, llvm::Metadata *> MetadataPair;
+  SmallVector<MetadataPair, 4> SYCLRegKernelNames;
+
   llvm::StringMap<llvm::GlobalVariable *> CFConstantStringMap;
 
   llvm::DenseMap<llvm::Constant *, llvm::GlobalVariable *> ConstantStringMap;
@@ -1483,6 +1486,12 @@ class CodeGenModule : public CodeGenTypeCache {
   llvm::Constant *EmitSYCLAnnotationArgs(
       SmallVectorImpl<std::pair<std::string, std::string>> &Pairs);
 
+  void SYCLAddRegKernelNamePairs(StringRef First, StringRef Second) {
+    SYCLRegKernelNames.push_back(
+        std::make_pair(llvm::MDString::get(getLLVMContext(), First),
+                       llvm::MDString::get(getLLVMContext(), Second)));
+  }
+
   /// Add attributes from add_ir_attributes_global_variable on TND to GV.
   void AddGlobalSYCLIRAttributes(llvm::GlobalVariable *GV,
                                  const RecordDecl *RD);

clang/lib/Sema/SemaDeclAttr.cpp

@@ -7479,6 +7479,9 @@ ProcessDeclAttribute(Sema &S, Scope *scope, Decl *D, const ParsedAttr &AL,
   case ParsedAttr::AT_SYCLAddIRAnnotationsMember:
     S.SYCL().handleSYCLAddIRAnnotationsMemberAttr(D, AL);
     break;
+  case ParsedAttr::AT_SYCLRegisteredKernels:
+    S.SYCL().handleSYCLRegisteredKernels(D, AL);
+    break;
 
   // Swift attributes.
   case ParsedAttr::AT_SwiftAsyncName:

clang/lib/Sema/SemaSYCL.cpp

@@ -1148,10 +1148,10 @@ static target getAccessTarget(QualType FieldTy,
 
 // FIXME: Free functions must have void return type and be declared at file
 // scope, outside any namespaces.
-static bool isFreeFunction(SemaSYCL &SemaSYCLRef, const FunctionDecl *FD) {
+bool SemaSYCL::isFreeFunction(const FunctionDecl *FD) {
   for (auto *IRAttr : FD->specific_attrs<SYCLAddIRAttributesFunctionAttr>()) {
     SmallVector<std::pair<std::string, std::string>, 4> NameValuePairs =
-        IRAttr->getAttributeNameValuePairs(SemaSYCLRef.getASTContext());
+        IRAttr->getAttributeNameValuePairs(getASTContext());
     for (const auto &NameValuePair : NameValuePairs) {
       if (NameValuePair.first == "sycl-nd-range-kernel" ||
           NameValuePair.first == "sycl-single-task-kernel") {
@@ -5291,7 +5291,7 @@ void SemaSYCL::SetSYCLKernelNames() {
        SyclKernelsToOpenCLKernels) {
     std::string CalculatedName, StableName;
     StringRef KernelName;
-    if (isFreeFunction(*this, Pair.first)) {
+    if (isFreeFunction(Pair.first)) {
       std::tie(CalculatedName, StableName) =
           constructFreeFunctionKernelName(*this, Pair.first, *MangleCtx);
       KernelName = CalculatedName;
@@ -5414,24 +5414,66 @@ void SemaSYCL::ConstructOpenCLKernel(FunctionDecl *KernelCallerFunc,
   }
 }
 
-void ConstructFreeFunctionKernel(SemaSYCL &SemaSYCLRef, FunctionDecl *FD) {
-  SyclKernelArgsSizeChecker argsSizeChecker(SemaSYCLRef, FD->getLocation(),
+static void addRegisteredKernelName(SemaSYCL &S, StringRef Str,
+                                    FunctionDecl *FD, SourceLocation Loc) {
+  if (!Str.empty())
+    FD->addAttr(SYCLRegisteredKernelNameAttr::CreateImplicit(S.getASTContext(),
+                                                             Str, Loc));
+}
+
+static bool checkAndAddRegisteredKernelName(SemaSYCL &S, FunctionDecl *FD,
+                                            StringRef Str) {
+  using KernelPair = std::pair<const FunctionDecl *, FunctionDecl *>;
+  for (const KernelPair &Pair : S.getKernelFDPairs()) {
+    if (Pair.first == FD) {
+      // If the current list of free function entries already contains this
+      // free function, apply the name Str as an attribute.  But if it already
+      // has an attribute name, issue a diagnostic instead.
+      if (!Str.empty()) {
+        if (!Pair.second->hasAttr<SYCLRegisteredKernelNameAttr>())
+          addRegisteredKernelName(S, Str, Pair.second, FD->getLocation());
+        else
+          S.Diag(FD->getLocation(),
+                 diag::err_registered_kernels_name_already_registered)
+              << Pair.second->getAttr<SYCLRegisteredKernelNameAttr>()
+                     ->getRegName()
+              << Str;
+      }
+      // An empty name string implies a regular free kernel construction
+      // call, so simply return.
+      return false;
+    }
+  }
+  return true;
+}
+
+void SemaSYCL::constructFreeFunctionKernel(FunctionDecl *FD,
+                                           StringRef NameStr) {
+  if (!checkAndAddRegisteredKernelName(*this, FD, NameStr))
+    return;
+
+  SyclKernelArgsSizeChecker argsSizeChecker(*this, FD->getLocation(),
                                             false /*IsSIMDKernel*/);
-  SyclKernelDeclCreator kernel_decl(SemaSYCLRef, FD->getLocation(),
-                                    FD->isInlined(), false /*IsSIMDKernel */,
-                                    FD);
+  SyclKernelDeclCreator kernel_decl(*this, FD->getLocation(), FD->isInlined(),
+                                    false /*IsSIMDKernel */, FD);
 
-  FreeFunctionKernelBodyCreator kernel_body(SemaSYCLRef, kernel_decl, FD);
+  FreeFunctionKernelBodyCreator kernel_body(*this, kernel_decl, FD);
 
-  SyclKernelIntHeaderCreator int_header(
-      SemaSYCLRef, SemaSYCLRef.getSyclIntegrationHeader(), FD->getType(), FD);
+  SyclKernelIntHeaderCreator int_header(*this, getSyclIntegrationHeader(),
+                                        FD->getType(), FD);
 
-  SyclKernelIntFooterCreator int_footer(SemaSYCLRef,
-                                        SemaSYCLRef.getSyclIntegrationFooter());
-  KernelObjVisitor Visitor{SemaSYCLRef};
+  SyclKernelIntFooterCreator int_footer(*this, getSyclIntegrationFooter());
+  KernelObjVisitor Visitor{*this};
 
   Visitor.VisitFunctionParameters(FD, argsSizeChecker, kernel_decl, kernel_body,
                                   int_header, int_footer);
+
+  assert(getKernelFDPairs().back().first == FD &&
+         "OpenCL Kernel not found for free function entry");
+  // Register the kernel name with the OpenCL kernel generated for the
+  // free function.
+  addRegisteredKernelName(*this, NameStr, getKernelFDPairs().back().second,
+                          FD->getLocation());
 }
 
 // Figure out the sub-group for the this function.  First we check the
@@ -5717,7 +5759,7 @@ void SemaSYCL::MarkDevices() {
 }
 
 void SemaSYCL::ProcessFreeFunction(FunctionDecl *FD) {
-  if (isFreeFunction(*this, FD)) {
+  if (isFreeFunction(FD)) {
     SyclKernelDecompMarker DecompMarker(*this);
     SyclKernelFieldChecker FieldChecker(*this);
     SyclKernelUnionChecker UnionChecker(*this);
@@ -5736,7 +5778,7 @@ void SemaSYCL::ProcessFreeFunction(FunctionDecl *FD) {
     if (!FieldChecker.isValid() || !UnionChecker.isValid())
       return;
 
-    ConstructFreeFunctionKernel(*this, FD);
+    constructFreeFunctionKernel(FD);
   }
 }
 
@@ -6621,7 +6663,7 @@ void SYCLIntegrationHeader::emit(raw_ostream &O) {
   unsigned ShimCounter = 1;
   int FreeFunctionCount = 0;
   for (const KernelDesc &K : KernelDescs) {
-    if (!isFreeFunction(S, K.SyclKernel))
+    if (!S.isFreeFunction(K.SyclKernel))
       continue;
     ++FreeFunctionCount;
     // Generate forward declaration for free function.
@@ -6739,7 +6781,7 @@ void SYCLIntegrationHeader::emit(raw_ostream &O) {
   }
   ShimCounter = 1;
   for (const KernelDesc &K : KernelDescs) {
-    if (!isFreeFunction(S, K.SyclKernel))
+    if (!S.isFreeFunction(K.SyclKernel))
       continue;
 
     O << "\n// Definition of kernel_id of " << K.Name << "\n";

clang/lib/Sema/SemaSYCLDeclAttr.cpp

@@ -3162,3 +3162,68 @@ void SemaSYCL::checkSYCLAddIRAttributesFunctionAttrConflicts(Decl *D) {
       Diag(Attr->getLoc(), diag::warn_sycl_old_and_new_kernel_attributes)
           << Attr;
 }
+
+void SemaSYCL::handleSYCLRegisteredKernels(Decl *D, const ParsedAttr &A) {
+  // Check for SYCL device compilation context.
+  if (!getLangOpts().SYCLIsDevice)
+    return;
+
+  unsigned NumArgs = A.getNumArgs();
+  // When declared, we expect at least one item in the list.
+  if (NumArgs == 0) {
+    Diag(A.getLoc(), diag::err_registered_kernels_num_of_args);
+    return;
+  }
+
+  // Traverse through the items in the list.
+  for (unsigned I = 0; I < NumArgs; I++) {
+    assert(A.isArgExpr(I) && "Expected expression argument");
+    // Each item in the list must be an initializer list expression.
+    Expr *ArgExpr = A.getArgAsExpr(I);
+    if (!isa<InitListExpr>(ArgExpr)) {
+      Diag(ArgExpr->getExprLoc(), diag::err_registered_kernels_init_list);
+      return;
+    }
+
+    auto *ArgListE = cast<InitListExpr>(ArgExpr);
+    unsigned NumInits = ArgListE->getNumInits();
+    // Each init-list expression must have a pair of values.
+    if (NumInits != 2) {
+      Diag(ArgExpr->getExprLoc(),
+           diag::err_registered_kernels_init_list_pair_values);
+      return;
+    }
+
+    // The first value of the pair must be a string.
+    Expr *FirstExpr = ArgListE->getInit(0);
+    StringRef CurStr;
+    SourceLocation Loc = FirstExpr->getExprLoc();
+    if (!SemaRef.checkStringLiteralArgumentAttr(A, FirstExpr, CurStr, &Loc))
+      return;
+
+    // Resolve the FunctionDecl from the second value of the pair.
+    Expr *SecondE = ArgListE->getInit(1);
+    FunctionDecl *FD = nullptr;
+    if (auto *ULE = dyn_cast<UnresolvedLookupExpr>(SecondE)) {
+      FD = SemaRef.ResolveSingleFunctionTemplateSpecialization(ULE, true);
+      Loc = ULE->getExprLoc();
+    } else {
+      SecondE = SecondE->IgnoreParenCasts();
+      if (auto *DRE = dyn_cast<DeclRefExpr>(SecondE))
+        FD = dyn_cast<FunctionDecl>(DRE->getDecl());
+      Loc = SecondE->getExprLoc();
+    }
+    // Issue a diagnostic if we are unable to resolve the FunctionDecl.
+    if (!FD) {
+      Diag(Loc, diag::err_registered_kernels_resolve_function) << CurStr;
+      return;
+    }
+    // Issue a diagnostic is the FunctionDecl is not a SYCL free function.
+    if (!isFreeFunction(FD)) {
+      Diag(FD->getLocation(), diag::err_not_sycl_free_function) << CurStr;
+      return;
+    }
+    // Construct a free function kernel.
+    constructFreeFunctionKernel(FD, CurStr);
+  }
+}