jhuber6 wrote: > > Sorry for the delayed response - I have been on a long vacation. > > I’m not familiar with the new offload driver internals to give pointers on > > what needs to be changed and where for retaining the custom plugin and > > linker behavior needed by the > > [chipStar](https://github.com/CHIP-SPV/chipStar) nor I can devote time to > > look at it currently. > > Perhaps, the HIPSPV toolchain could default to the old offload driver in > > the meantime until I or someone else takes a look at it. > > cc @pjaaskel and @pvelesko for awareness. > > The steps are roughly outlined in > https://clang.llvm.org/docs/OffloadingDesign.html, though it hasn't been > updated for the other targets. Generally it means you'll need to do some more > plumbing. The problem with HIPSPV here is that it still uses the `OFK_HIP` > type, which means it's not easy to determine whether or not to use the new > driver from the compiler stages, and even so I'd _really_ prefer to be able > to delete all the old handling soon. > > More or less, the main change is that all offloading targets go through the > same function. For non-RDC compilation it's not too different, but for RDC > compilation everything gets put into a fat binary using a custom binary > format inside LLVM. This then gets passed to the linker wrapper, whose job is > to extract those objects and generate a linker job. So, what you need is for > `clang --target=<whatever your triple> a b c` to work.
I'd suggest you just try doing `--offload-new-driver` with `-###` and `-v` and see what doesn't work. https://github.com/llvm/llvm-project/pull/123359 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
