https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64835
--- Comment #21 from Richard Biener <rguenth at gcc dot gnu.org> --- Ah, so the actual difference is that we inline hiphip (hooray) during early inlining on x86_64 (and i?86), but not on 32bit sparc. Disabling early inlining shows that IPA-CP plus IPA inlining works on x86_64, the difference being that x86 shows only modification of test() and IPA inline only considering hooray but not hiphip.constprop/5 for some reason. That reason would need to be investigated (likewise the reason for the failed early inlining, which is likely the same - optimize attribute related).
