darkbuck wrote: One motivation for this patch is to be able to mark individual inline asms with 'convergent.' So far, CUDA/HIP assumes all calls and inline-asms are convergent and marks all of them with 'convergent.' This guarantees correctness but loses possible optimizations for inline-asms with ALU instructions only. After allowing 'convergent' to be a statement attribute, we could use 'convergent' mark only relevant inline-asms as convergent.
https://github.com/llvm/llvm-project/pull/100637 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
