https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54488
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> --- I had patches to implement the suggestion, work on outermost loops separately. Of course it won't help once you wrap all of them in a single outer loop (which is why I didn't end up committing this). Re-measuring state of current trunk would be still nice (I think we must have improved over the 1GB).