================
@@ -229,6 +229,12 @@ class CodeGenTypes {
const CGFunctionInfo &arrangeBuiltinFunctionCall(QualType resultType,
const CallArgList &args);
+ /// A SYCL device kernel function is a free standing function with
+ /// spir_kernel calling convention
----------------
bader wrote:
```suggestion
/// target specific calling convention (e.g. spir_kernel for the SPIR target).
```
Technically we use OpenCL language calling convention, but I'm not sure if we
should specify that in the comments. Typically, each offload target has its own
calling convention. For example, OpenCL compiler emits `amdgpu_kernel` calling
convention for AMD GPU targets and `ptx_kernel` for NVIDIA GPU targets.
NOTE: Setting calling convention might be not enough for decorating kernel
functions properly. E.g. NVPTX target uses named metadata to identify kernel
functions.
https://docs.nvidia.com/cuda/nvvm-ir-spec/index.html#representation-of-properties
I suppose this means that SYCL compiler for NVPTX target must emit named
metadata referring to the kernel in addition to using `ptx_kernel` calling
convention.
Some SPIR tools consuming LLVM IR doesn't require additional metadata, so
setting calling convention is enough. I don't know how NVPTX backend will
handle LLVM w/o the metadata.
@npmiller, FYI.
https://github.com/llvm/llvm-project/pull/133030
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
