[CUDA][HIP] Set kernel calling convention before arrange function

Currently clang set kernel calling convention for CUDA/HIP after
arranging function, which causes incorrect kernel function type since
it depends on calling convention.

This patch moves setting kernel convention before arranging
function.

Differential Revision: https://reviews.llvm.org/D47733

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@334457 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/CGCall.cpp b/lib/CodeGen/CGCall.cpp
index 2d8339ef697010440824ab4f3608b9e89ecf9d5f..064b5fcb102ff4438703bf2e9d61ac99ace4cfd1 100644 (file)
--- a/lib/CodeGen/CGCall.cpp
+++ b/lib/CodeGen/CGCall.cpp
@@ -255,6 +255,16 @@ CodeGenTypes::arrangeCXXMethodType(const CXXRecordDecl *RD,
       FTP->getCanonicalTypeUnqualified().getAs<FunctionProtoType>(), MD);
 }
 
+/// Set calling convention for CUDA/HIP kernel.
+static void setCUDAKernelCallingConvention(CanQualType &FTy, CodeGenModule &CGM,
+                                           const FunctionDecl *FD) {
+  if (FD->hasAttr<CUDAGlobalAttr>()) {
+    const FunctionType *FT = FTy->getAs<FunctionType>();
+    CGM.getTargetCodeGenInfo().setCUDAKernelCallingConvention(FT);
+    FTy = FT->getCanonicalTypeUnqualified();
+  }
+}
+
 /// Arrange the argument and result information for a declaration or
 /// definition of the given C++ non-static member function.  The
 /// member function must be an ordinary function, i.e. not a
@@ -264,7 +274,9 @@ CodeGenTypes::arrangeCXXMethodDeclaration(const CXXMethodDecl *MD) {
   assert(!isa<CXXConstructorDecl>(MD) && "wrong method for constructors!");
   assert(!isa<CXXDestructorDecl>(MD) && "wrong method for destructors!");
 
-  CanQual<FunctionProtoType> prototype = GetFormalType(MD);
+  CanQualType FT = GetFormalType(MD).getAs<Type>();
+  setCUDAKernelCallingConvention(FT, CGM, MD);
+  auto prototype = FT.getAs<FunctionProtoType>();
 
   if (MD->isInstance()) {
     // The abstract case is perfectly fine.
@@ -424,6 +436,7 @@ CodeGenTypes::arrangeFunctionDeclaration(const FunctionDecl *FD) {
   CanQualType FTy = FD->getType()->getCanonicalTypeUnqualified();
 
   assert(isa<FunctionType>(FTy));
+  setCUDAKernelCallingConvention(FTy, CGM, FD);
 
   // When declaring a function without a prototype, always use a
   // non-variadic type.

diff --git a/lib/CodeGen/CodeGenModule.cpp b/lib/CodeGen/CodeGenModule.cpp
index 814eda4381b546ca179bf9f5b404fa915e08af5b..6595694ebd533e6a8c3ab92ffa033037d21ac3b2 100644 (file)
--- a/lib/CodeGen/CodeGenModule.cpp
+++ b/lib/CodeGen/CodeGenModule.cpp
@@ -3671,8 +3671,6 @@ void CodeGenModule::EmitGlobalFunctionDefinition(GlobalDecl GD,
 
   MaybeHandleStaticInExternC(D, Fn);
 
-  if (D->hasAttr<CUDAGlobalAttr>())
-    getTargetCodeGenInfo().setCUDAKernelCallingConvention(Fn);
 
   maybeSetTrivialComdat(*D, *Fn);
 

diff --git a/lib/CodeGen/TargetInfo.cpp b/lib/CodeGen/TargetInfo.cpp
index 3ec00553607cd116aaf0f0234230d9c2ec5ee213..b29bcce237cffd232f75ba300aa900ae398939eb 100644 (file)
--- a/lib/CodeGen/TargetInfo.cpp
+++ b/lib/CodeGen/TargetInfo.cpp
@@ -7646,7 +7646,7 @@ public:
                             llvm::Function *BlockInvokeFunc,
                             llvm::Value *BlockLiteral) const override;
   bool shouldEmitStaticExternCAliases() const override;
-  void setCUDAKernelCallingConvention(llvm::Function *F) const override;
+  void setCUDAKernelCallingConvention(const FunctionType *&FT) const override;
 };
 }
 
@@ -7783,8 +7783,9 @@ bool AMDGPUTargetCodeGenInfo::shouldEmitStaticExternCAliases() const {
 }
 
 void AMDGPUTargetCodeGenInfo::setCUDAKernelCallingConvention(
-    llvm::Function *F) const {
-  F->setCallingConv(llvm::CallingConv::AMDGPU_KERNEL);
+    const FunctionType *&FT) const {
+  FT = getABIInfo().getContext().adjustFunctionType(
+      FT, FT->getExtInfo().withCallingConv(CC_OpenCLKernel));
 }
 
 //===----------------------------------------------------------------------===//

diff --git a/lib/CodeGen/TargetInfo.h b/lib/CodeGen/TargetInfo.h
index 5c19c7141243cccee8ab97cf247be1690b7153a8..b530260ea48f8df258e16fe744ee800579f9b14d 100644 (file)
--- a/lib/CodeGen/TargetInfo.h
+++ b/lib/CodeGen/TargetInfo.h
@@ -302,7 +302,7 @@ public:
   /// as 'used', and having internal linkage.
   virtual bool shouldEmitStaticExternCAliases() const { return true; }
 
-  virtual void setCUDAKernelCallingConvention(llvm::Function *F) const {}
+  virtual void setCUDAKernelCallingConvention(const FunctionType *&FT) const {}
 };
 
 } // namespace CodeGen

diff --git a/test/CodeGenCUDA/kernel-args.cu b/test/CodeGenCUDA/kernel-args.cu
new file mode 100644 (file)
index 0000000..d098662
--- /dev/null
+++ b/test/CodeGenCUDA/kernel-args.cu
@@ -0,0 +1,39 @@
+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa -fcuda-is-device \
+// RUN:     -emit-llvm %s -o - | FileCheck -check-prefix=AMDGCN %s
+// RUN: %clang_cc1 -triple nvptx64-nvidia-cuda- -fcuda-is-device \
+// RUN:     -emit-llvm %s -o - | FileCheck -check-prefix=NVPTX %s
+#include "Inputs/cuda.h"
+
+struct A {
+  int a[32];
+};
+
+// AMDGCN: define amdgpu_kernel void @_Z6kernel1A(%struct.A %x.coerce)
+// NVPTX: define void @_Z6kernel1A(%struct.A* byval align 4 %x)
+__global__ void kernel(A x) {
+}
+
+class Kernel {
+public:
+  // AMDGCN: define amdgpu_kernel void @_ZN6Kernel12memberKernelE1A(%struct.A %x.coerce)
+  // NVPTX: define void @_ZN6Kernel12memberKernelE1A(%struct.A* byval align 4 %x)
+  static __global__ void memberKernel(A x){}
+  template<typename T> static __global__ void templateMemberKernel(T x) {}
+};
+
+
+template <typename T>
+__global__ void templateKernel(T x) {}
+
+void launch(void*);
+
+void test() {
+  Kernel K;
+  // AMDGCN: define amdgpu_kernel void @_Z14templateKernelI1AEvT_(%struct.A %x.coerce)
+  // NVPTX: define void @_Z14templateKernelI1AEvT_(%struct.A* byval align 4 %x)
+  launch((void*)templateKernel<A>);
+
+  // AMDGCN: define amdgpu_kernel void @_ZN6Kernel20templateMemberKernelI1AEEvT_(%struct.A %x.coerce)
+  // NVPTX: define void @_ZN6Kernel20templateMemberKernelI1AEEvT_(%struct.A* byval align 4 %x)
+  launch((void*)Kernel::templateMemberKernel<A>);
+}