[PATCH] D21367: AMDGPU: Set amdgpu_kernel calling convention for OpenCL kernels.

[Nikolay Haustov via cfe-commits]

nhaustov added a comment.

In http://reviews.llvm.org/D21367#467894, @yaxunl wrote:

> I am wondering what's the difference between this calling convention and spir_kernel.


spir_kernel has only effect and was created only for SPIR target. In theory it could perhaps be reused for AMDGPU target, however it seems not very clear to me.

In http://reviews.llvm.org/D21367#468057, @yaxunl wrote:

> In http://reviews.llvm.org/D21367#467982, @nhaustov wrote:
>
> > In http://reviews.llvm.org/D21367#467894, @yaxunl wrote:
> >
> > > I am wondering what's the difference between this calling convention and spir_kernel.
> >
> >
> > spir_kernel has only effect and was created only for SPIR target. In theory it could perhaps be reused for AMDGPU target, however it seems not very clear to me.
>
>
> If the new calling convention is to indicate a function is an OpenCL kernel, then it has exactly the same meaning as spir_kernel. Can we just use it in AMDGPU target?


This looks like a hack to me. Calling conventions in LLVM IR are really target specific. spir_kernel really doesn't make sense for anything other than SPIR.

Maybe we can use the fact that calling conventions are separate in AST and in LLVM IR. So in AST maybe we can have opencl_kernel and in LLVM it can map to spir_kernel for SPIR and amdgpu_kernel for AMDGPU.


http://reviews.llvm.org/D21367