[CUDA/SPIR-V] Force passing aggregate type byval

[shangwuyao]

Summary

This patch forces copying aggregate type in kernel arguments by value when
compiling CUDA targeting SPIR-V. The original behavior is not passing by value
when there is any of destructor, copy constructor and move constructor defined
by user. This patch makes the behavior of SPIR-V generated from CUDA follow
the CUDA spec
(https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#global-function-argument-processing),
and matches the NVPTX
implementation (
https://github.com/llvm/llvm-project/blob/41958f76d8a2c47484fa176cba1de565cfe84de7/clang/lib/CodeGen/TargetInfo.cpp#L7241).

[tra]

LGTM with a couple of nits.

clang/lib/CodeGen/TargetInfo.cpp

10450

I would add a comment that this is a CUDA-specific behavior with the pointer to the relevant CUDA doc explaining what's going on.

10450

Style nit: LLVM does not use braces for single-statement body.
https://llvm.org/docs/CodingStandards.html#don-t-use-braces-on-simple-single-statement-bodies-of-if-else-loop-statements

[shangwuyao]

Accidentally submitted early...

[tra]

In D130387#3672961, @shangwuyao wrote:

Accidentally submitted early...

The landed revision seems to have my comments addressed. Was there something missing?

[shangwuyao]

In D130387#3672969, @tra wrote:

In D130387#3672961, @shangwuyao wrote:

Accidentally submitted early...

The landed revision seems to have my comments addressed. Was there something missing?

No, just want to wait a bit longer to see if someone else has comments.

[tra]

It's OK. If something comes up you can address it in a follow-up patch.

clang/lib/CodeGen/TargetInfo.cpp

if (getContext().getLangOpts().CUDAIsDevice) {
     llvm::Type *LTy = CGT.ConvertType(Ty);
     auto DefaultAS = getContext().getTargetAddressSpace(LangAS::Default);
     auto GlobalAS = getContext().getTargetAddressSpace(LangAS::cuda_device);
     auto *PtrTy = llvm::dyn_cast<llvm::PointerType>(LTy);
     if (PtrTy && PtrTy->getAddressSpace() == DefaultAS) {
       LTy = llvm::PointerType::getWithSamePointeeType(PtrTy, GlobalAS);
       return ABIArgInfo::getDirect(LTy, 0, nullptr, false);
     }
+
+    if (isAggregateTypeForABI(Ty)) {
+      return getNaturalAlignIndirect(Ty, /* byval */ true);
+    }
   }
   return classifyArgumentType(Ty);
 }
 
 void SPIRVABIInfo::computeInfo(CGFunctionInfo &FI) const {
   // The logic is same as in DefaultABIInfo with an exception on the kernel
   // arguments handling.
   llvm::CallingConv::ID CC = FI.getCallingConvention();

clang/test/CodeGenCUDASPIRV/copy-aggregate-byval.cu

+// Tests CUDA kernel arguments get copied by value when targeting SPIR-V, even with
+// destructor, copy constructor or move constructor defined by user.
+
+// RUN: %clang -Xclang -no-opaque-pointers -emit-llvm --cuda-device-only --offload=spirv32 \
+// RUN:   -nocudalib -nocudainc %s -o %t.bc -c 2>&1
+// RUN: llvm-dis %t.bc -o %t.ll
+// RUN: FileCheck %s --input-file=%t.ll
+
+// RUN: %clang -Xclang -no-opaque-pointers -emit-llvm --cuda-device-only --offload=spirv64 \
+// RUN:   -nocudalib -nocudainc %s -o %t.bc -c 2>&1
+// RUN: llvm-dis %t.bc -o %t.ll
+// RUN: FileCheck %s --input-file=%t.ll
+
+class GpuData {
+ public:
+  __attribute__((host)) __attribute__((device)) GpuData(int* src) {}
+  __attribute__((host)) __attribute__((device)) ~GpuData() {}
+  __attribute__((host)) __attribute__((device)) GpuData(const GpuData& other) {}
+  __attribute__((host)) __attribute__((device)) GpuData(GpuData&& other) {}
+};
+
+// CHECK: define
+// CHECK-SAME: spir_kernel void @_Z6kernel7GpuData(%class.GpuData* noundef byval(%class.GpuData) align
+
+__attribute__((global)) void kernel(GpuData output) {}