[Bug rtl-optimization/70261] [6 Regression] r234265 causes fails on rs6000

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70261

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rguenth at gcc dot gnu.org

--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
But all the replace_rtx behavior seen here is IMHO bogus and might cause latent
issues.

An option besides asserting is to fail replacement (as we did before - the
assert was to catch other cases where not replacing might just hide a bug).

Of course as far as I understand the PPC regressions show cases where
replacing more (the REGNO == REGNO with equal mode case) actually _breaks_
things?

Rather than adding a new function I'd add a param to replace_rtx, like
'bool all_regs' guarding the new REG case.