https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116189
Andrew Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords|needs-bisection |
Summary|[14/15 Regression] compare |[14/15 Regression] compare
|debug failure at -O2 on sh |debug failure at -O2 on sh
| |since
| |r14-8319-g86de9b66480b71
--- Comment #11 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
So rtx_costs started to be called from fwprop by r14-8319-g86de9b66480b71 . So
that exposed the latent bug in the sh backend.
As for r15-2069-g9f9faebb8ebfc0 exposing the issue is most likely a similar
reason, you accidently get different regnums from GCSE but before the patch it
would just happen the RA would get the same result but now you don't.
Anyways the bug is definitely in the backend here.
Let me try to figure out the best way to test this patch.
