https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116949
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> --- Yeah, but note the vectorizer internally (up to recent Tamars patch?) made use of walk_stmt specifically for pattern stmts so it was(?) still needed. There's also GENERIC comparison handling in genmatch because of phiopt (and ifcombine?) building COND_EXPRs with tcc_comparison ops ...; I think the genmatch support for (match ... of PHIs now also does that. So double-check tree-vect-patterns.cc no longer builds such COND_EXPRs (the failure mode would be bogus hybrid SLP and costing, so not necessarily ICEs). Maybe add gcc_assert()s to gimple_build_assign ...
