jhuber6 wrote: I applied this locally and ran it against the [LLVM libc GPU test suite](https://libc.llvm.org/gpu/testing.html). This is probably the test suite that best exercises real object linking. All tests pass except one, which is admittedly broken in theory by mixing wave32 and wave64 compiled modules, so I just removed that one. This yields a nice increase in compile times without the need for monolithic LTO for everything.
| Phase | LTO (status quo) | Object linking | Speedup | |---|---|---|---| | Compile + link | ~145 s | ~44.5 s | **3.3× faster** (−101 s) | | Test execution (lit) | 63.6 s | 64.5 s | ~1.01x | | **Total wall (clean + tests done)** | **209 s** | **109 s** | **1.9× faster (−48%)** | | Tests | 714 / 714 | 713 / 713 | — | https://github.com/llvm/llvm-project/pull/206787 _______________________________________________ llvm-branch-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
