yxsamliu wrote: the major blocker of using the non-lto pipeline is that clang driver does not support parallel compilation for mulit-GPU through -offload-jobs. we used to have an out-of-tree parallel jobs implementation but we have removed that after we switch to new offload driver and rely on the parallel jobs support in the linker wrapper for parallel compilation for multi-GPU.
If we switch to non-lto pipeline, we need to have a way to support parallel compilation for multi-GPU for the non-lto pipeline, otherwise we are facing serious compilation time regressions. one way to support that might be to implement the classic non-lto pipeline in the linker wrapper like the one clang use by default. maybe linker wrapper already have that and we only need to wire it properly. https://github.com/llvm/llvm-project/pull/200847 _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
