https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125986
--- Comment #8 from Aldy Hernandez <aldyh at gcc dot gnu.org> --- (In reply to Andrew Macleod from comment #5) > Created attachment 64868 [details] > possible patch > > Way cleaner version.. if the path query fails, it tries a root query with > just the ssa-names, not the equivalences. > > In theory, I think that is OK because it should only be relations which > dominate the entry block, so any killing defs within the block shouldn't > matter, This is how things look with the instrumentation patch applied while compiling the .ii files from a stage1 bootstrap: 1. Baseline: 152 failures. 2. Patch1 (Proof of concept tickling incoming relations to a path): 135 failures. 3. Your cleaned up patch: 92 failures. This is definitely a big improvement. Thanks. I'll continue looking at the other discrepancies.
