https://gcc.gnu.org/bugzilla/show_bug.cgi?id=55212
--- Comment #414 from Oleg Endo <olegendo at gcc dot gnu.org> --- (In reply to Kazumoto Kojima from comment #410) > Created attachment 59432 [details] > a trial patch for c#404 > > It's difficult to see what is going on, because the test case is too huge. > Looking at the postmortem debug, it seems that reload failed with f - r reg > move generated by LRA, perhaps in starvation of registers including fpu regs > and fpul. That case might require reloading with memories which isn't > handled by the current movsf logic for LRA. > > The patch falls back to movsf_ie for f - r moves when lra_in_progress is > true. > Tested with c/c++ testsuite on devel/sh-lra. I've merged this patch into commit "SH: Try to workaround fp-reg related move insns pt.2" and added the reduced test case from comment #413. It looks like the two commits "SH: Try to workaround fp-reg related ..." should be squashed. Will do that in the next iteration. Updated branch: https://github.com/olegendo/gcc/tree/devel/sh-lra
