yxsamliu wrote: > > Would it be better to pass the full `--save-temps=cwd/--save-temps=obj` > > option through to the linker wrapper instead of collapsing it to a boolean, > > and then preserve the old (CWD) behavior for `--save-temps=cwd` while using > > the new (output-dir) behavior for `--save-temps=obj`? That would keep the > > linker wrapper consistent with how clang handles both modes. As a bonus, > > lld's intermediate files would follow automatically since they derive their > > location from the .img output path. > > I'm leaning towards using `lld`'s semantics just because these are both > functionally the linker. This is a bit of a strange case because normal > `-save-temps` doesn't imply LLD to emit temp files, but we do that with > offloading just because it's the expected behavior.
Makes sense. Aligning with lld's semantics seems to be a consistent choice for the linker wrapper. In practice, saving to output directory is usually better than saving to CWD. We may want to document -save-temps behavior for device intermediate files in clang HIP support documentation but that could be done in a separate PR. https://github.com/llvm/llvm-project/pull/198679 _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
