> So, barring the removal of the get_initial_register_offset stuff, the only > simple fix is probably to prevent it from calling into the back-end too > early, for example with the attached fixlet. Tested on x86-64 and > PowerPC/Linux. > > Thoughts? Where do we want to fix this? > > > * rtlanal.c (get_initial_register_offset): Fall back to the raw estimate > as long as the epilogue isn't completed.
I have installed it on mainline only for now. -- Eric Botcazou