jhuber6 wrote: This is not the correct approach. Like I said, we should only track dependencies from the host, as the device phases are an implementation detail and can contain dozens of candidate architectures. There are many places in practice that set `--offload-arch=` to contain several. The device-side compilation has no relevance to the build system.
Right now, HIP deliberately suppresses dependency file generation on the device compilations. This was added in https://github.com/llvm/llvm-project/pull/125646. This should likely be applied to CUDA as well. The `-fdepfile-entry` option does not read from a dependency file, it just appends a literal filename to the output `.d` file. In the CUDA toolchain you should just be able to emit `-fdepfile-entry` on the necessary files if you want to track changes to the `.bc`. https://github.com/llvm/llvm-project/pull/176072 _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
