arsenm wrote: > My glossary might be lacking some definitions, but what I really meant by > lowering is
The IR is not a direct vehicle for language semantics. This case should explicitly not be representable in the target IR, and the frontend needs to implement the calling convention in terms of valid IR for the target. > What I'm trying to understand if it's really should be done in clang for > OpenCL for every target. This is implementing a stupid OpenCL feature nobody asked for, so clang should deal with it. Every other language can happily ignore this. AMDGPU, NVPTX, and SPIRV all have the same behavior where kernels / entry points are a separate calling convention that cannot be called from device code. This call should be unrepresentable in the IR, and forcing it to be such just makes the backends worse. We cannot move the calling convention lowering forward without removing this artificial usage. https://github.com/llvm/llvm-project/pull/115821 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
