[SYCL] Allow SYCL_EXTERNAL for class member functions (#1295)

This patch allows to apply sycl_device attribute and therefore SYCL_EXTERNAL macro to class member functions.

The same logic was applied to `[[intel::device_indireclty_callable]]`.

Moved device code outlining to CodeGen because the old scheme in
combination with SYCL_EXTERNAL implementation for methods triggered
assertions

Signed-off-by: Mariya Podchishchaeva <[email protected]>
Co-authored-by: Alexey Sachkov <[email protected]>

Author: Fznamznon

clang/include/clang/Basic/AttrDocs.td

@@ -2117,8 +2117,8 @@ def SYCLDeviceIndirectlyCallableDocs : Documentation {
 This attribute can only be applied to functions and indicates that the
 function must be treated as a device function and must be emitted even if it has
 no direct uses from other SYCL device functions. However, it cannot be applied
-to functions marked as 'static', functions declared within an anonymous
-namespace or class member functions.
+to functions marked as 'static' and functions declared within an anonymous
+namespace.
 
 It also means that function should be available externally and
 cannot be optimized out due to reachability analysis or by any other

clang/include/clang/Basic/DiagnosticSemaKinds.td

@@ -10694,8 +10694,7 @@ def warn_boolean_attribute_argument_is_not_valid: Warning<
    InGroup<AdjustedAttributes>;
 def err_sycl_attibute_cannot_be_applied_here
     : Error<"%0 attribute cannot be applied to a "
-            "%select{static function or function in an anonymous namespace"
-            "|class member function}1">;
+            "static function or function in an anonymous namespace">;
 def warn_sycl_attibute_function_raw_ptr
     : Warning<"SYCL 1.2.1 specification does not allow %0 attribute applied "
               "to a function with a raw pointer "

clang/lib/AST/ASTContext.cpp

@@ -10156,6 +10156,10 @@ bool ASTContext::DeclMustBeEmitted(const Decl *D) {
   if (D->hasAttr<AliasAttr>() || D->hasAttr<UsedAttr>())
     return true;
 
+  if (LangOpts.SYCLIsDevice && !D->hasAttr<OpenCLKernelAttr>() &&
+      !D->hasAttr<SYCLDeviceAttr>())
+    return false;
+
   if (const auto *FD = dyn_cast<FunctionDecl>(D)) {
     // Forward declarations aren't required.
     if (!FD->doesThisDeclarationHaveABody())
@@ -10178,6 +10182,16 @@ bool ASTContext::DeclMustBeEmitted(const Decl *D) {
       }
     }
 
+    // Methods explcitly marked with 'sycl_device' attribute (via SYCL_EXTERNAL)
+    // or `indirectly_callable' attribute must be emitted regardless of number
+    // of actual uses
+    if (LangOpts.SYCLIsDevice && isa<CXXMethodDecl>(D)) {
+      if (auto *A = D->getAttr<SYCLDeviceIndirectlyCallableAttr>())
+        return !A->isImplicit();
+      if (auto *A = D->getAttr<SYCLDeviceAttr>())
+        return !A->isImplicit();
+    }
+
     GVALinkage Linkage = GetGVALinkageForFunction(FD);
 
     // static, static inline, always_inline, and extern inline functions can

clang/lib/CodeGen/CodeGenModule.cpp

@@ -2516,11 +2516,6 @@ void CodeGenModule::EmitGlobal(GlobalDecl GD) {
   if (Global->hasAttr<IFuncAttr>())
     return emitIFuncDefinition(GD);
 
-  if (LangOpts.SYCLIsDevice) {
-    if (!Global->hasAttr<SYCLDeviceAttr>())
-      return;
-  }
-
   // If this is a cpu_dispatch multiversion function, emit the resolver.
   if (Global->hasAttr<CPUDispatchAttr>())
     return emitCPUDispatchDefinition(GD);
@@ -2565,6 +2560,11 @@ void CodeGenModule::EmitGlobal(GlobalDecl GD) {
     }
   }
 
+  if (LangOpts.SYCLIsDevice && MustBeEmitted(Global)) {
+    addDeferredDeclToEmit(GD);
+    return;
+  }
+
   // Ignore declarations, they will be emitted on their first use.
   if (const auto *FD = dyn_cast<FunctionDecl>(Global)) {
     // Forward declarations are emitted lazily on first use.

clang/lib/Sema/SemaDeclAttr.cpp

@@ -4428,13 +4428,7 @@ static void handleOptimizeNoneAttr(Sema &S, Decl *D, const ParsedAttr &AL) {
 static void handleSYCLDeviceAttr(Sema &S, Decl *D, const ParsedAttr &AL) {
   auto *FD = cast<FunctionDecl>(D);
   if (!FD->isExternallyVisible()) {
-    S.Diag(AL.getLoc(), diag::err_sycl_attibute_cannot_be_applied_here)
-        << AL << 0 /* static function or anonymous namespace */;
-    return;
-  }
-  if (isa<CXXMethodDecl>(FD)) {
-    S.Diag(AL.getLoc(), diag::err_sycl_attibute_cannot_be_applied_here)
-        << AL << 1 /* class member function */;
+    S.Diag(AL.getLoc(), diag::err_sycl_attibute_cannot_be_applied_here) << AL;
     return;
   }
   if (FD->getReturnType()->isPointerType()) {
@@ -4447,25 +4441,17 @@ static void handleSYCLDeviceAttr(Sema &S, Decl *D, const ParsedAttr &AL) {
           << AL << 1 /* function with a raw pointer parameter type */;
     }
 
-  S.addSyclDeviceDecl(D);
   handleSimpleAttribute<SYCLDeviceAttr>(S, D, AL);
 }
 
 static void handleSYCLDeviceIndirectlyCallableAttr(Sema &S, Decl *D,
                                                    const ParsedAttr &AL) {
   auto *FD = cast<FunctionDecl>(D);
   if (!FD->isExternallyVisible()) {
-    S.Diag(AL.getLoc(), diag::err_sycl_attibute_cannot_be_applied_here)
-        << AL << 0 /* static function or anonymous namespace */;
-    return;
-  }
-  if (isa<CXXMethodDecl>(FD)) {
-    S.Diag(AL.getLoc(), diag::err_sycl_attibute_cannot_be_applied_here)
-        << AL << 1 /* class member function */;
+    S.Diag(AL.getLoc(), diag::err_sycl_attibute_cannot_be_applied_here) << AL;
     return;
   }
 
-  S.addSyclDeviceDecl(D);
   D->addAttr(SYCLDeviceAttr::CreateImplicit(S.Context));
   handleSimpleAttribute<SYCLDeviceIndirectlyCallableAttr>(S, D, AL);
 }

clang/lib/Sema/SemaSYCL.cpp

@@ -228,12 +228,6 @@ class MarkDeviceFunction : public RecursiveASTVisitor<MarkDeviceFunction> {
 
       CheckSYCLType(Callee->getReturnType(), Callee->getSourceRange());
 
-      if (FunctionDecl *Def = Callee->getDefinition()) {
-        if (!Def->hasAttr<SYCLDeviceAttr>()) {
-          Def->addAttr(SYCLDeviceAttr::CreateImplicit(SemaRef.Context));
-          SemaRef.addSyclDeviceDecl(Def);
-        }
-      }
       if (auto const *FD = dyn_cast<FunctionDecl>(Callee)) {
         // FIXME: We need check all target specified attributes for error if
         // that function with attribute can not be called from sycl kernel.  The
@@ -265,23 +259,6 @@ class MarkDeviceFunction : public RecursiveASTVisitor<MarkDeviceFunction> {
   bool VisitCXXConstructExpr(CXXConstructExpr *E) {
     for (const auto &Arg : E->arguments())
       CheckSYCLType(Arg->getType(), Arg->getSourceRange());
-
-    CXXConstructorDecl *Ctor = E->getConstructor();
-
-    if (FunctionDecl *Def = Ctor->getDefinition()) {
-      Def->addAttr(SYCLDeviceAttr::CreateImplicit(SemaRef.Context));
-      SemaRef.addSyclDeviceDecl(Def);
-    }
-
-    const auto *ConstructedType = Ctor->getParent();
-    if (ConstructedType->hasUserDeclaredDestructor()) {
-      CXXDestructorDecl *Dtor = ConstructedType->getDestructor();
-
-      if (FunctionDecl *Def = Dtor->getDefinition()) {
-        Def->addAttr(SYCLDeviceAttr::CreateImplicit(SemaRef.Context));
-        SemaRef.addSyclDeviceDecl(Def);
-      }
-    }
     return true;
   }
 
@@ -321,31 +298,6 @@ class MarkDeviceFunction : public RecursiveASTVisitor<MarkDeviceFunction> {
       return true;
 
     CheckSYCLType(E->getType(), E->getSourceRange());
-    if (VarDecl *VD = dyn_cast<VarDecl>(D)) {
-      if (!VD->isLocalVarDeclOrParm() && VD->hasGlobalStorage()) {
-        VD->addAttr(SYCLDeviceAttr::CreateImplicit(SemaRef.Context));
-        SemaRef.addSyclDeviceDecl(VD);
-      }
-    }
-    return true;
-  }
-
-  bool VisitCXXNewExpr(CXXNewExpr *E) {
-    // Memory storage allocation is not allowed in kernels.
-    // All memory allocation for the device is done on
-    // the host using accessor classes. Consequently, the default
-    // allocation operator new overloads that allocate
-    // storage are disallowed in a SYCL kernel. The placement
-    // new operator and any user-defined overloads that
-    // do not allocate storage are permitted.
-    if (FunctionDecl *FD = E->getOperatorNew()) {
-      if (FunctionDecl *Def = FD->getDefinition()) {
-        if (!Def->hasAttr<SYCLDeviceAttr>()) {
-          Def->addAttr(SYCLDeviceAttr::CreateImplicit(SemaRef.Context));
-          SemaRef.addSyclDeviceDecl(Def);
-        }
-      }
-    }
     return true;
   }
 
@@ -1389,13 +1341,8 @@ void Sema::MarkDevice(void) {
     }
   }
   for (const auto &elt : Marker.KernelSet) {
-    if (FunctionDecl *Def = elt->getDefinition()) {
-      if (!Def->hasAttr<SYCLDeviceAttr>()) {
-        Def->addAttr(SYCLDeviceAttr::CreateImplicit(Context));
-        addSyclDeviceDecl(Def);
-      }
+    if (FunctionDecl *Def = elt->getDefinition())
       Marker.TraverseStmt(Def->getBody());
-    }
   }
 }
 
@@ -1467,10 +1414,6 @@ static void emitCallToUndefinedFnDiag(Sema &SemaRef, const FunctionDecl *Callee,
   if (Callee->getTemplatedKind() == FunctionDecl::TK_FunctionTemplate)
     return;
 
-  // Don't emit diagnostic for functions not called from device code
-  if (!Caller->hasAttr<SYCLDeviceAttr>() && !Caller->hasAttr<SYCLKernelAttr>())
-    return;
-
   bool RedeclHasAttr = false;
 
   for (const Decl *Redecl : Callee->redecls()) {

clang/test/CodeGenSYCL/device-indirectly-callable-attr.cpp

@@ -7,10 +7,63 @@ void foo() {
   helper();
 }
 
-// CHECK: define spir_func void @{{.*foo.*}}() #[[ATTRS_FOO:[0-9]+]]
+// CHECK: define spir_func void @{{.*foo.*}}() #[[ATTRS_INDIR_CALL:[0-9]+]]
 // CHECK: call spir_func void @{{.*helper.*}}()
 //
-// CHECK: define spir_func void @{{.*helper.*}}() #[[ATTRS_HELPER:[0-9]+]]
+// CHECK: define spir_func void @{{.*helper.*}}() #[[ATTRS_NOT_INDIR_CALL:[0-9]+]]
 //
-// CHECK: attributes #[[ATTRS_FOO]] = { {{.*}} "referenced-indirectly"
-// CHECK-NOT: attributes #[[ATTRS_HELPER]] = { {{.*}} "referenced-indirectly"
+
+int bar20(int a) { return a + 20; }
+
+class A {
+public:
+  // CHECK-DAG: define linkonce_odr spir_func void @_ZN1A3fooEv{{.*}}#[[ATTRS_INDIR_CALL]]
+  // CHECK-DAG: define spir_func i32 @_Z5bar20{{.*}}#[[ATTRS_NOT_INDIR_CALL]]
+  [[intel::device_indirectly_callable]] void foo() { bar20(10); }
+
+  // CHECK-DAG: define linkonce_odr spir_func void @_ZN1AC1Ev{{.*}}#[[ATTRS_INDIR_CALL]]
+  [[intel::device_indirectly_callable]] A() {}
+  // CHECK-DAG: define linkonce_odr spir_func void @_ZN1AD1Ev{{.*}}#[[ATTRS_INDIR_CALL]]
+  [[intel::device_indirectly_callable]] ~A() {}
+
+  template <typename T>
+  [[intel::device_indirectly_callable]] void AFoo(T t) {}
+
+  // Templates are emitted when they are instantiated
+  // CHECK-DAG: define linkonce_odr spir_func void @_ZN1A4AFooIiEEvT_{{.*}}#[[ATTRS_INDIR_CALL]]
+  template <>
+  [[intel::device_indirectly_callable]] void AFoo<int>(int t) {}
+};
+
+struct Base {
+  // CHECK-DAG: define linkonce_odr spir_func void @_ZN4Base12BaseWithAttrEv{{.*}}#[[ATTRS_INDIR_CALL]]
+  [[intel::device_indirectly_callable]] virtual void BaseWithAttr() { int a = 10; }
+  virtual void BaseWithoutAttr() { int b = 20; }
+};
+
+struct Overrider : Base {
+  // CHECK-DAG: define linkonce_odr spir_func void @_ZN9Overrider12BaseWithAttrEv{{.*}}#[[ATTRS_INDIR_CALL]]
+  [[intel::device_indirectly_callable]] void BaseWithAttr() override { int a = 20; }
+  // CHECK-DAG: define linkonce_odr spir_func void @_ZN9Overrider15BaseWithoutAttrEv{{.*}}#[[ATTRS_INDIR_CALL]]
+  [[intel::device_indirectly_callable]] void BaseWithoutAttr() override { int b = 30; }
+};
+
+struct Overrider1 : Base {
+  // CHECK-NOT: define linkonce_odr spir_func void @_ZN10Overrider112BaseWithAttrEv
+  void BaseWithAttr() override { int a = 20; }
+};
+
+struct Finalizer : Base {
+  // CHECK-DAG: define linkonce_odr spir_func void @_ZN9Finalizer12BaseWithAttrEv{{.*}}#[[ATTRS_INDIR_CALL]]
+  [[intel::device_indirectly_callable]] void BaseWithAttr() final { int a = 20; }
+  // CHECK-DAG: define linkonce_odr spir_func void @_ZN9Finalizer15BaseWithoutAttrEv{{.*}}#[[ATTRS_INDIR_CALL]]
+  [[intel::device_indirectly_callable]] void BaseWithoutAttr() final { int b = 30; }
+};
+
+struct Finalizer1 : Base {
+  // CHECK-NOT: define linkonce_odr spir_func void @_ZN10Finalizer112BaseWithAttrEv
+  void BaseWithAttr() final { int a = 20; }
+};
+
+// CHECK: attributes #[[ATTRS_INDIR_CALL]] = { {{.*}} "referenced-indirectly"
+// CHECK-NOT: attributes #[[ATTRS_NOT_INDIR_CALL]] = { {{.*}} "referenced-indirectly"

clang/test/CodeGenSYCL/sycl-device.cpp

@@ -3,6 +3,84 @@
 
 int bar(int b);
 
+int bar10(int a) { return a + 10; }
+int bar20(int a) { return a + 20; }
+
+class A {
+public:
+  // CHECK-DAG: define linkonce_odr spir_func void @_ZN1A3fooEv
+  // CHECK-DAG: define spir_func i32 @_Z5bar20i
+  __attribute__((sycl_device)) void foo() { bar20(10); }
+
+  // CHECK-DAG: define linkonce_odr spir_func void @_ZN1AC1Ev
+  // CHECK-DAG: define spir_func i32 @_Z5bar10i
+  __attribute__((sycl_device))
+  A() { bar10(10); }
+  // CHECK-DAG: define linkonce_odr spir_func void @_ZN1AD1Ev
+  __attribute__((sycl_device)) ~A() {}
+
+  template <typename T>
+  __attribute__((sycl_device)) void AFoo(T t) {}
+
+  // Templates are emitted when they are instantiated
+  // CHECK-DAG: define linkonce_odr spir_func void @_ZN1A4AFooIiEEvT_
+  template <>
+  __attribute__((sycl_device)) void AFoo<int>(int t) {}
+
+  // CHECK-DAG: define linkonce_odr spir_func i32 @_ZN1A13non_annotatedEv
+  int non_annotated() { return 1; }
+
+  // CHECK-DAG: define linkonce_odr spir_func i32 @_ZN1A9annotatedEv
+  __attribute__((sycl_device)) int annotated() { return non_annotated() + 1; }
+};
+
+template <typename T>
+struct B {
+  T data;
+  B(T _data) : data(_data) {}
+
+  __attribute__((sycl_device)) void BFoo(T t) {}
+};
+
+template <>
+struct B<int> {
+  int data;
+  B(int _data) : data(_data) {}
+
+  // CHECK-DAG: define linkonce_odr spir_func void @_ZN1BIiE4BFooEi
+  __attribute__((sycl_device)) void BFoo(int t) {}
+};
+
+struct Base {
+  // CHECK-DAG: define linkonce_odr spir_func void @_ZN4Base12BaseWithAttrEv
+  __attribute__((sycl_device)) virtual void BaseWithAttr() { int a = 10; }
+  virtual void BaseWithoutAttr() { int b = 20; }
+};
+
+struct Overrider : Base {
+  // CHECK-DAG: define linkonce_odr spir_func void @_ZN9Overrider12BaseWithAttrEv
+  __attribute__((sycl_device)) void BaseWithAttr() override { int a = 20; }
+  // CHECK-DAG: define linkonce_odr spir_func void @_ZN9Overrider15BaseWithoutAttrEv
+  __attribute__((sycl_device)) void BaseWithoutAttr() override { int b = 30; }
+};
+
+struct Overrider1 : Base {
+  // CHECK-NOT: define linkonce_odr spir_func void @_ZN10Overrider112BaseWithAttrEv
+  void BaseWithAttr() override { int a = 20; }
+};
+
+struct Finalizer : Base {
+  // CHECK-DAG: define linkonce_odr spir_func void @_ZN9Finalizer12BaseWithAttrEv
+  __attribute__((sycl_device)) void BaseWithAttr() final { int a = 20; }
+  // CHECK-DAG: define linkonce_odr spir_func void @_ZN9Finalizer15BaseWithoutAttrEv
+  __attribute__((sycl_device)) void BaseWithoutAttr() final { int b = 30; }
+};
+
+struct Finalizer1 : Base {
+  // CHECK-NOT: define linkonce_odr spir_func void @_ZN10Finalizer112BaseWithAttrEv
+  void BaseWithAttr() final { int a = 20; }
+};
+
 // CHECK-DAG: define spir_func i32 @_Z3fooii
 __attribute__((sycl_device))
 int foo(int a, int b) { return a + bar(b); }

clang/test/SemaSYCL/call-to-undefined-function.cpp

@@ -92,6 +92,9 @@ void forwardDeclFn() {
 }
 
 int main() {
+  // No problems in host code
+  undefined();
+
   kernel_single_task<class CallToUndefinedFnTester>([]() {
     // expected-note@-1 {{called by 'operator()'}}
     // expected-note@-2 {{called by 'operator()'}}