https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81763
Alexander Monakov <amonakov at gcc dot gnu.org> changed: What |Removed |Added ---------------------------------------------------------------------------- CC| |amonakov at gcc dot gnu.org --- Comment #8 from Alexander Monakov <amonakov at gcc dot gnu.org> --- Mike, you can start by preparing two build trees, one with faulty mesa compiled with -march=native, the other with working-fine mesa compiled with -march=native -mno-bmi. You should be able to collect: - .o files from each tree, and - link commands from build logs and be able to re-link mesa libraries by hand and verify they still exhibit the same behavior (one fails, the other doesn't). >From there you can proceed by building "hybrid" libraries by taking a set of good .o files and a complementary set of bad .o files. This will allow you to find a single .o file that makes the library behave wrongly. More explanation and a helper script for bisecting is available at https://gcc.gnu.org/wiki/Analysing_Large_Testcases At that point please share your status (once you're down to one file there's no generic recipe, you'd have to get creative).