[CUDA] Do not merge CUDA target attributes.

[tra]

Summary

CUDA target attributes are used for function overloading and must not be merged.

This fixes a bug where attributes were inherited during function template
specialization in CUDA and made it impossible for specialized function
to provide its own target attributes.

[jlebar]

Yay, this is great.

test/SemaCUDA/target_attr_inheritance.cu

4 ↗

(On Diff #71238)

Other reviewers have pointed out to me that we don't usually (ever?) need this. I think these have to do with llvm's ability to generate code for our targets, but it's not relevant to clang here.

[tra]

Removed REQUIRED lines.

cfe/trunk/lib/Sema/SemaDecl.cpp

static bool mergeDeclAttribute(Sema &S, NamedDecl *D,
   else if (const auto *IA = dyn_cast<MSInheritanceAttr>(Attr))
     NewAttr = S.mergeMSInheritanceAttr(D, IA->getRange(), IA->getBestCase(),
                                        AttrSpellingListIndex,
                                        IA->getSemanticSpelling());
   else if (const auto *AA = dyn_cast<AlwaysInlineAttr>(Attr))
     NewAttr = S.mergeAlwaysInlineAttr(D, AA->getRange(),
                                       &S.Context.Idents.get(AA->getSpelling()),
                                       AttrSpellingListIndex);
-  else if (const auto *MA = dyn_cast<MinSizeAttr>(Attr))
+  else if (S.getLangOpts().CUDA && isa<FunctionDecl>(D) &&
+           (isa<CUDAHostAttr>(Attr) || isa<CUDADeviceAttr>(Attr) ||
+            isa<CUDAGlobalAttr>(Attr))) {
+    // CUDA target attributes are part of function signature for
+    // overloading purposes and must not be merged.
+    return false;
+  } else if (const auto *MA = dyn_cast<MinSizeAttr>(Attr))
     NewAttr = S.mergeMinSizeAttr(D, MA->getRange(), AttrSpellingListIndex);
   else if (const auto *OA = dyn_cast<OptimizeNoneAttr>(Attr))
     NewAttr = S.mergeOptimizeNoneAttr(D, OA->getRange(), AttrSpellingListIndex);
   else if (const auto *InternalLinkageA = dyn_cast<InternalLinkageAttr>(Attr))
     NewAttr = S.mergeInternalLinkageAttr(
         D, InternalLinkageA->getRange(),
         &S.Context.Idents.get(InternalLinkageA->getSpelling()),
         AttrSpellingListIndex);

cfe/trunk/test/SemaCUDA/function-overload.cu

 __host__ HostReturnTy2 host_only_function(float arg) { return HostReturnTy2(); }
 
 __host__ __device__ void test_host_device_single_side_overloading() {
   DeviceReturnTy ret1 = device_only_function(1);
   DeviceReturnTy2 ret2 = device_only_function(1.0f);
   HostReturnTy ret3 = host_only_function(1);
   HostReturnTy2 ret4 = host_only_function(1.0f);
 }
+
+// Verify that we allow overloading function templates.
+template <typename T> __host__ T template_overload(const T &a) { return a; };
+template <typename T> __device__ T template_overload(const T &a) { return a; };
+
+__host__ void test_host_template_overload() {
+  template_overload(1); // OK. Attribute-based overloading picks __host__ variant.
+}
+__device__ void test_device_template_overload() {
+  template_overload(1); // OK. Attribute-based overloading picks __device__ variant.
+}

cfe/trunk/test/SemaCUDA/target_attr_inheritance.cu

+// Verifies correct inheritance of target attributes during template
+// instantiation and specialization.
+
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -fsyntax-only -verify %s
+// RUN: %clang_cc1 -triple nvptx64-nvidia-cuda -fsyntax-only -fcuda-is-device -verify %s
+
+#include "Inputs/cuda.h"
+
+// Function must inherit target attributes during instantiation, but not during
+// specialization.
+template <typename T> __host__ __device__ T function_template(const T &a);
+
+// Specialized functions have their own attributes.
+// expected-note@+1 {{candidate function not viable: call to __host__ function from __device__ function}}
+template <> __host__ float function_template<float>(const float &from);
+
+// expected-note@+1 {{candidate function not viable: call to __device__ function from __host__ function}}
+template <> __device__ double function_template<double>(const double &from);
+
+__host__ void hf() {
+  function_template<float>(1.0f); // OK. Specialization is __host__.
+  function_template<double>(2.0); // expected-error {{no matching function for call to 'function_template'}}
+  function_template(1);           // OK. Instantiated function template is HD.
+}
+__device__ void df() {
+  function_template<float>(3.0f); // expected-error {{no matching function for call to 'function_template'}}
+  function_template<double>(4.0); // OK. Specialization is __device__.
+  function_template(1);           // OK. Instantiated function template is HD.
+}