Pierre-vh wrote: > mapping LangAS to LangAS is a fundamentally dodgy operation.
This is not the goal of this patch (and #195613 does not do that at any point either), but the added TargetInfo hook could be abused for that indeed. This patch just changes the flow from conditionally calling TargetInfo to determine the AS of a GV to always calling it. > Make GetGlobalVarAddressSpace return that address space for barrier globals This patch is simply trying to enable that. I did not think it was okay to check an AMDGPU-specific type in the middle of `lib/CodeGen/CodeGenModule.cpp`. Is it the recommended approach? Then this patch can just be closed and I can do that directly in #195613 indeed. > Define a LangAS::hip_barrier language-level address space That's fine if we want to avoid the `getLangASFromTargetAS(llvm::AMDGPUAS::BARRIER)` call later, but it's orthogonal to this patch. It's a concern for #195613 Suggestion: How about I do a very targeted hook instead, something like `getCUDADeviceGlobalVarAddressSpace` that'd always be called? Seems like that'd be less abuse-able. I can do it directly in #195613 as well. https://github.com/llvm/llvm-project/pull/195612 _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
