Pierre-vh wrote: @arsenm Given that we can't assume all accesses to addrspace(5) are thread-local, what is a good way to proceed here ? I was thinking of doing something a bit more generic. e.g. add a separate pass to analyze all load/stores in addrspace(5), and if they're thread-local, add some MD to them (e.g. `!amdgpu.thread_private`)
I'd like to at least mark some scratch accesses. I don't think we need to mark all of them to get good performance, but the trivial cases (like basic allocas, or byvals) would be nice. https://github.com/llvm/llvm-project/pull/179415 _______________________________________________ llvm-branch-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
