[SYCL] Add Clang support for FPGA loop fusion function attributes

clang/include/clang/Basic/Attr.td

@@ -1313,6 +1313,25 @@ def SYCLIntelNoGlobalWorkOffset : InheritableAttr {
   let PragmaAttributeSupport = 0;
 }
 
+def SYCLIntelLoopFuse : InheritableAttr {
+  let Spellings = [CXX11<"intel", "loop_fuse">,
+                   CXX11<"intel", "loop_fuse_independent">];
+  let Args = [ExprArgument<"Value", /*optional=*/ 1>];
+  let LangOpts = [SYCLIsDevice, SYCLIsHost];
+  let Subjects = SubjectList<[Function], ErrorDiag>;
+  let Accessors = [Accessor<"isIndependent",
+                  [CXX11<"intel", "loop_fuse_independent">]>];
+  let Documentation = [SYCLIntelLoopFuseDocs];
+  let AdditionalMembers = [{
+    static unsigned getMinValue() {
+      return 0;
+    }
+    static unsigned getMaxValue() {
+      return 1024*1024;
+    }
+  }];
+}
+
 def C11NoReturn : InheritableAttr {
   let Spellings = [Keyword<"_Noreturn">];
   let Subjects = SubjectList<[Function], ErrorDiag>;

clang/include/clang/Basic/AttrDocs.td

@@ -2608,6 +2608,34 @@ loop should not be fused with any adjacent loop.
   }];
 }
 
+def SYCLIntelLoopFuseDocs : Documentation {
+  let Category = DocCatFunction;
+  let Heading = "loop_fuse, loop_fuse_independent";
+  let Content = [{
+``[[intel::loop_fuse(N)]]`` and ``[[intel::loop_fuse_independent(N)]]`` attributes apply
+to a function/lambda function. It is a strong request, to the extent possible, to fuse
+the loops within the function, that are contained in at most N-1 other loops within the
+function. If the optional parameter N is omitted, it is a strong request, to the extent
+possible, to fuse loops within the function that are not contained in any other loop
+within the function. ``[[intel::loop_fuse_independent(N)]]`` also guarantees that fusion
+safety analysis can ignore negative-distance dependences between these loops.
+
+.. code-block:: c++
+
+  [[intel::loop_fuse(N)]]
+  int foo() {}
+
+  [[intel::loop_fuse_independent(N)]]
+  int foo() {}
+
+
+``[[intel::loop_fuse(N)]]`` and ``[[intel::loop_fuse_independent(N)]]`` takes one optional
+parameter that is a constant unsigned integer expression. The parameter N may be a template
+parameter.
+
+  }];
+}
+
 def SYCLDeviceIndirectlyCallableDocs : Documentation {
   let Category = DocCatFunction;
   let Heading = "intel::device_indirectly_callable";

clang/include/clang/Basic/AttributeCommonInfo.h

@@ -166,7 +166,8 @@ class AttributeCommonInfo {
         ParsedAttr == AT_SYCLIntelMaxWorkGroupSize ||
         ParsedAttr == AT_SYCLIntelMaxGlobalWorkDim ||
         ParsedAttr == AT_SYCLIntelNoGlobalWorkOffset ||
-        ParsedAttr == AT_SYCLIntelUseStallEnableClusters)
+        ParsedAttr == AT_SYCLIntelUseStallEnableClusters ||
+        ParsedAttr == AT_SYCLIntelLoopFuse)
       return true;
 
     return false;

clang/include/clang/Sema/Sema.h

@@ -3363,6 +3363,8 @@ class Sema final {
   WebAssemblyImportModuleAttr *mergeImportModuleAttr(
       Decl *D, const WebAssemblyImportModuleAttr &AL);
 
+  SYCLIntelLoopFuseAttr *
+  mergeSYCLIntelLoopFuseAttr(Decl *D, const AttributeCommonInfo &CI, Expr *E);
   void mergeDeclAttributes(NamedDecl *New, Decl *Old,
                            AvailabilityMergeKind AMK = AMK_Redeclaration);
   void MergeTypedefNameDecl(Scope *S, TypedefNameDecl *New,
@@ -10204,6 +10206,8 @@ class Sema final {
   /// addSYCLIntelPipeIOAttr - Adds a pipe I/O attribute to a particular
   /// declaration.
   void addSYCLIntelPipeIOAttr(Decl *D, const AttributeCommonInfo &CI, Expr *ID);
+  void addSYCLIntelLoopFuseAttr(Decl *D, const AttributeCommonInfo &CI,
+                                Expr *E);
 
   bool checkNSReturnsRetainedReturnType(SourceLocation loc, QualType type);
   bool checkAllowedSYCLInitializer(VarDecl *VD,

clang/lib/CodeGen/CodeGenFunction.cpp

@@ -955,6 +955,19 @@ void CodeGenFunction::StartFunction(GlobalDecl GD, QualType RetTy,
   if (getLangOpts().SYCLIsHost && D && D->hasAttr<SYCLKernelAttr>())
     Fn->addFnAttr("sycl_kernel");
 
+  if (getLangOpts().SYCLIsDevice && D) {
+    if (const auto *A = D->getAttr<SYCLIntelLoopFuseAttr>()) {
+      Expr *E = A->getValue();
+      llvm::Metadata *AttrMDArgs[] = {
+          llvm::ConstantAsMetadata::get(Builder.getInt32(
+              E->getIntegerConstantExpr(D->getASTContext())->getZExtValue())),
+          llvm::ConstantAsMetadata::get(
+              A->isIndependent() ? Builder.getInt32(1) : Builder.getInt32(0))};
+      Fn->setMetadata("loop_fuse",
+                      llvm::MDNode::get(getLLVMContext(), AttrMDArgs));
+    }
+  }
+
   if (getLangOpts().OpenCL || getLangOpts().SYCLIsDevice) {
     // Add metadata for a kernel function.
     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {

clang/lib/Sema/SemaDecl.cpp

@@ -2614,6 +2614,8 @@ static bool mergeDeclAttribute(Sema &S, NamedDecl *D,
     NewAttr = S.mergeImportModuleAttr(D, *IMA);
   else if (const auto *INA = dyn_cast<WebAssemblyImportNameAttr>(Attr))
     NewAttr = S.mergeImportNameAttr(D, *INA);
+  else if (const auto *LFA = dyn_cast<SYCLIntelLoopFuseAttr>(Attr))
+    NewAttr = S.mergeSYCLIntelLoopFuseAttr(D, *LFA, LFA->getValue());
   else if (Attr->shouldInheritEvenIfAlreadyPresent() || !DeclHasAttr(D, Attr))
     NewAttr = cast<InheritableAttr>(Attr->clone(S.Context));
 

clang/lib/Sema/SemaDeclAttr.cpp

@@ -3091,6 +3091,90 @@ static void handleMaxGlobalWorkDimAttr(Sema &S, Decl *D,
                                                                      E);
 }
 
+SYCLIntelLoopFuseAttr *
+Sema::mergeSYCLIntelLoopFuseAttr(Decl *D, const AttributeCommonInfo &CI,
+                                 Expr *E) {
+
+  if (const auto ExistingAttr = D->getAttr<SYCLIntelLoopFuseAttr>()) {
+    // [[intel::loop_fuse]] and [[intel::loop_fuse_independent]] are
+    // incompatible.
+    // FIXME: If additional spellings are provided for this attribute,
+    // this code will do the wrong thing.
+    if (ExistingAttr->getAttributeSpellingListIndex() !=
+        CI.getAttributeSpellingListIndex()) {
+      Diag(CI.getLoc(), diag::err_attributes_are_not_compatible)
+          << CI << ExistingAttr;
+      Diag(ExistingAttr->getLocation(), diag::note_conflicting_attribute);
+      return nullptr;
+    }
+
+    if (!E->isValueDependent()) {
+      Optional<llvm::APSInt> ArgVal = E->getIntegerConstantExpr(Context);
+      Optional<llvm::APSInt> ExistingArgVal =
+          ExistingAttr->getValue()->getIntegerConstantExpr(Context);
+
+      assert(ArgVal && ExistingArgVal &&
+             "Argument should be an integer constant expression");
+      // Compare attribute argument value and warn if there is a mismatch.
+      if (ArgVal->getExtValue() != ExistingArgVal->getExtValue())
+        Diag(ExistingAttr->getLoc(), diag::warn_duplicate_attribute)
+            << ExistingAttr;
+    }
+
+    // If there is no mismatch, silently ignore duplicate attribute.
+    return nullptr;
+  }
+  return ::new (Context) SYCLIntelLoopFuseAttr(Context, CI, E);
+}
+
+static bool checkSYCLIntelLoopFuseArgument(Sema &S,
+                                           const AttributeCommonInfo &CI,
+                                           Expr *E) {
+  // Dependent expressions are checked when instantiated.
+  if (E->isValueDependent())
+    return false;
+
+  Optional<llvm::APSInt> ArgVal = E->getIntegerConstantExpr(S.Context);
+  if (!ArgVal) {
+    S.Diag(E->getExprLoc(), diag::err_attribute_argument_type)
+        << CI << AANT_ArgumentIntegerConstant << E->getSourceRange();
+    return true;
+  }
+
+  SYCLIntelLoopFuseAttr TmpAttr(S.Context, CI, E);
+  ExprResult ICE;
+
+  return S.checkRangedIntegralArgument<SYCLIntelLoopFuseAttr>(E, &TmpAttr, ICE);
+}
+
+void Sema::addSYCLIntelLoopFuseAttr(Decl *D, const AttributeCommonInfo &CI,
+                                    Expr *E) {
+  assert(E && "argument has unexpected null value");
+
+  if (checkSYCLIntelLoopFuseArgument(*this, CI, E))
+    return;
+
+  // Attribute should not be added during host compilation.
+  if (getLangOpts().SYCLIsHost)
+    return;
+
+  SYCLIntelLoopFuseAttr *NewAttr = mergeSYCLIntelLoopFuseAttr(D, CI, E);
+
+  if (NewAttr)
+    D->addAttr(NewAttr);
+}
+
+// Handles [[intel::loop_fuse]] and [[intel::loop_fuse_independent]].
+static void handleLoopFuseAttr(Sema &S, Decl *D, const ParsedAttr &Attr) {
+  // Default argument value is set to 1.
+  Expr *E = Attr.isArgExpr(0)
+                ? Attr.getArgAsExpr(0)
+                : IntegerLiteral::Create(S.Context, llvm::APInt(32, 1),
+                                         S.Context.IntTy, Attr.getLoc());
+
+  S.addSYCLIntelLoopFuseAttr(D, Attr, E);
+}
+
 static void handleVecTypeHint(Sema &S, Decl *D, const ParsedAttr &AL) {
   if (!AL.hasParsedType()) {
     S.Diag(AL.getLoc(), diag::err_attribute_wrong_number_arguments) << AL << 1;
@@ -8379,6 +8463,9 @@ static void ProcessDeclAttribute(Sema &S, Scope *scope, Decl *D,
   case ParsedAttr::AT_SYCLIntelUseStallEnableClusters:
     handleUseStallEnableClustersAttr(S, D, AL);
     break;
+  case ParsedAttr::AT_SYCLIntelLoopFuse:
+    handleLoopFuseAttr(S, D, AL);
+    break;
   case ParsedAttr::AT_VecTypeHint:
     handleVecTypeHint(S, D, AL);
     break;

clang/lib/Sema/SemaSYCL.cpp

@@ -570,6 +570,14 @@ class MarkDeviceFunction : public RecursiveASTVisitor<MarkDeviceFunction> {
           (KernelBody != FD) && !FD->hasAttr<SYCLSimdAttr>())
         FD->addAttr(SYCLSimdAttr::CreateImplicit(SemaRef.getASTContext()));
 
+      // Attribute "loop_fuse" can be applied explicitly on kernel function.
+      // Attribute should not be propagated from device functions to kernel.
+      if (auto *A = FD->getAttr<SYCLIntelLoopFuseAttr>()) {
+        if (ParentFD == SYCLKernel) {
+          Attrs.insert(A);
+        }
+      }
+
       // TODO: vec_len_hint should be handled here
 
       CallGraphNode *N = SYCLCG.getNode(FD);
@@ -3335,6 +3343,7 @@ void Sema::MarkDevice(void) {
         case attr::Kind::SYCLIntelMaxGlobalWorkDim:
         case attr::Kind::SYCLIntelNoGlobalWorkOffset:
         case attr::Kind::SYCLIntelUseStallEnableClusters:
+        case attr::Kind::SYCLIntelLoopFuse:
         case attr::Kind::SYCLSimd: {
           if ((A->getKind() == attr::Kind::SYCLSimd) && KernelBody &&
               !KernelBody->getAttr<SYCLSimdAttr>()) {

clang/lib/Sema/SemaTemplateInstantiateDecl.cpp

@@ -592,6 +592,16 @@ static void instantiateSYCLIntelPipeIOAttr(
     S.addSYCLIntelPipeIOAttr(New, *Attr, Result.getAs<Expr>());
 }
 
+static void instantiateSYCLIntelLoopFuseAttr(
+    Sema &S, const MultiLevelTemplateArgumentList &TemplateArgs,
+    const SYCLIntelLoopFuseAttr *Attr, Decl *New) {
+  EnterExpressionEvaluationContext Unevaluated(
+      S, Sema::ExpressionEvaluationContext::ConstantEvaluated);
+  ExprResult Result = S.SubstExpr(Attr->getValue(), TemplateArgs);
+  if (!Result.isInvalid())
+    S.addSYCLIntelLoopFuseAttr(New, *Attr, Result.getAs<Expr>());
+}
+
 template <typename AttrName>
 static void instantiateIntelSYCLFunctionAttr(
     Sema &S, const MultiLevelTemplateArgumentList &TemplateArgs,
@@ -775,6 +785,12 @@ void Sema::InstantiateAttrs(const MultiLevelTemplateArgumentList &TemplateArgs,
           *this, TemplateArgs, SYCLIntelMaxGlobalWorkDim, New);
       continue;
     }
+    if (const auto *SYCLIntelLoopFuse =
+            dyn_cast<SYCLIntelLoopFuseAttr>(TmplAttr)) {
+      instantiateSYCLIntelLoopFuseAttr(*this, TemplateArgs, SYCLIntelLoopFuse,
+                                       New);
+      continue;
+    }
     if (const auto *SYCLIntelNoGlobalWorkOffset =
             dyn_cast<SYCLIntelNoGlobalWorkOffsetAttr>(TmplAttr)) {
       instantiateIntelSYCLFunctionAttr<SYCLIntelNoGlobalWorkOffsetAttr>(
@@ -6204,7 +6220,10 @@ static void processSYCLKernel(Sema &S, FunctionDecl *FD, MangleContext &MC) {
   if (S.LangOpts.SYCLIsDevice) {
     S.ConstructOpenCLKernel(FD, MC);
   } else if (S.LangOpts.SYCLIsHost) {
-    CXXRecordDecl *CRD = (*FD->param_begin())->getType()->getAsCXXRecordDecl();
+    QualType KernelParamTy = (*FD->param_begin())->getType();
+    const CXXRecordDecl *CRD = (KernelParamTy->isReferenceType()
+                                    ? KernelParamTy->getPointeeCXXRecordDecl()
+                                    : KernelParamTy->getAsCXXRecordDecl());
     for (auto *Method : CRD->methods())
       if (Method->getOverloadedOperator() == OO_Call &&
           !Method->hasAttr<AlwaysInlineAttr>())

clang/test/CodeGenSYCL/loop_fuse_device.cpp

@@ -0,0 +1,52 @@
+// RUN: %clang_cc1 -fsycl -fsycl-is-device -internal-isystem %S/Inputs -triple spir64-unknown-unknown-sycldevice -emit-llvm -o - %s | FileCheck %s
+
+#include "sycl.hpp"
+
+using namespace cl::sycl;
+queue q;
+
+[[intel::loop_fuse(5)]] void foo() {}
+
+template <int SIZE>
+class KernelFunctor5 {
+public:
+  [[intel::loop_fuse(SIZE)]] void operator()() const {}
+};
+
+void bar() {
+
+  q.submit([&](handler &h) {
+    // Test template argument.
+    KernelFunctor5<5> f5;
+    h.single_task<class kernel_name_1>(f5);
+
+    // Test different argument sizes.
+    // Emit 1 if there is no argument.
+    h.single_task<class kernel_name_2>(
+        []() [[intel::loop_fuse]]{});
+    h.single_task<class kernel_name_3>(
+        []() [[intel::loop_fuse(0)]]{});
+    h.single_task<class kernel_name_4>(
+        []() [[intel::loop_fuse(1)]]{});
+    h.single_task<class kernel_name_5>(
+        []() [[intel::loop_fuse(10)]]{});
+
+    // Test attribute is not propagated.
+    h.single_task<class kernel_name_6>(
+        []() { foo(); });
+  });
+}
+
+// CHECK: define spir_kernel void @"{{.*}}kernel_name_1"() {{.*}} !loop_fuse ![[LF5:[0-9]+]]
+// CHECK: define spir_kernel void @"{{.*}}kernel_name_2"() {{.*}} !loop_fuse ![[LF1:[0-9]+]]
+// CHECK: define spir_kernel void @"{{.*}}kernel_name_3"() {{.*}} !loop_fuse ![[LF0:[0-9]+]]
+// CHECK: define spir_kernel void @"{{.*}}kernel_name_4"() {{.*}} !loop_fuse ![[LF1]]
+// CHECK: define spir_kernel void @"{{.*}}kernel_name_5"() {{.*}} !loop_fuse ![[LF10:[0-9]+]]
+// CHECK: define spir_kernel void @"{{.*}}kernel_name_6"()
+// CHECK-NOT: !loop_fuse
+// CHECK-SAME: {
+// CHECK: define spir_func void @{{.*}}foo{{.*}} !loop_fuse ![[LF5]]
+// CHECK: ![[LF5]] = !{i32 5, i32 0}
+// CHECK: ![[LF1]] = !{i32 1, i32 0}
+// CHECK: ![[LF0]] = !{i32 0, i32 0}
+// CHECK: ![[LF10]] = !{i32 10, i32 0}

clang/test/CodeGenSYCL/loop_fuse_ind_device.cpp

@@ -0,0 +1,52 @@
+// RUN: %clang_cc1 -fsycl -fsycl-is-device -internal-isystem %S/Inputs -triple spir64-unknown-unknown-sycldevice -emit-llvm -o - %s | FileCheck %s
+
+#include "sycl.hpp"
+
+using namespace cl::sycl;
+queue q;
+
+[[intel::loop_fuse_independent(5)]] void foo() {}
+
+template <int SIZE>
+class KernelFunctor5 {
+public:
+  [[intel::loop_fuse_independent(SIZE)]] void operator()() const {}
+};
+
+void bar() {
+
+  q.submit([&](handler &h) {
+    // Test template argument.
+    KernelFunctor5<5> f5;
+    h.single_task<class kernel_name_1>(f5);
+
+    // Test different argument sizes.
+    // Emit 1 if there is no argument.
+    h.single_task<class kernel_name_2>(
+        []() [[intel::loop_fuse_independent]]{});
+    h.single_task<class kernel_name_3>(
+        []() [[intel::loop_fuse_independent(0)]]{});
+    h.single_task<class kernel_name_4>(
+        []() [[intel::loop_fuse_independent(1)]]{});
+    h.single_task<class kernel_name_5>(
+        []() [[intel::loop_fuse_independent(10)]]{});
+
+    // Test attribute is not propagated.
+    h.single_task<class kernel_name_6>(
+        []() { foo(); });
+  });
+}
+
+// CHECK: define spir_kernel void @"{{.*}}kernel_name_1"() {{.*}} !loop_fuse ![[LFI5:[0-9]+]]
+// CHECK: define spir_kernel void @"{{.*}}kernel_name_2"() {{.*}} !loop_fuse ![[LFI1:[0-9]+]]
+// CHECK: define spir_kernel void @"{{.*}}kernel_name_3"() {{.*}} !loop_fuse ![[LFI0:[0-9]+]]
+// CHECK: define spir_kernel void @"{{.*}}kernel_name_4"() {{.*}} !loop_fuse ![[LFI1]]
+// CHECK: define spir_kernel void @"{{.*}}kernel_name_5"() {{.*}} !loop_fuse ![[LFI10:[0-9]+]]
+// CHECK: define spir_kernel void @"{{.*}}kernel_name_6"()
+// CHECK-NOT: !loop_fuse
+// CHECK-SAME: {
+// CHECK: define spir_func void @{{.*}}foo{{.*}} !loop_fuse ![[LFI5]]
+// CHECK: ![[LFI5]] = !{i32 5, i32 1}
+// CHECK: ![[LFI1]] = !{i32 1, i32 1}
+// CHECK: ![[LFI0]] = !{i32 0, i32 1}
+// CHECK: ![[LFI10]] = !{i32 10, i32 1}