[Bug target/108832] [13 Regression] ICE in replace_rtx, at rtlanal.cc:3358

["ubizjak at gmail dot com" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108832

--- Comment #4 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Jakub Jelinek from comment #1)

> and so ICEs if we see the same REGNO as from in a different mode.
> I think we actually don't need most of what replace_rtx is doing, we don't
> need to simplify anything etc. because we are just changing one register to
> another and can do it in place.
> So, I think we need a different function for what the backend wants.
> It can avoid all the simplify stuff because replace_rtx was destructive, so
> could be implemented say using FOR_EACH_SUBRTX_PTR.  When seeing *loc ==
> from, it obviously
> should set *loc = to, if it sees REG_P (*loc) && REGNO (*loc) == REGNO
> (from), then
> if the mode is the same, it can also just *loc = to, but if it is a
> different mode,
> I'd say for narrower mode it should *loc = gen_rtx_REG (GET_MODE (*loc),
> REGNO (to));
> and for wider mode (especially if say a multi-register reg) punt.
> Not sure if such a case can occur though, but the punting would be hard if
> we have made some changes already...

There are no multi-registers in flags-setting integer instructions, we only
have instructions with implicit ZERO_EXTEND from SI to DImode in case of x86_64
target.

So, a FOR_EACH_RTX loop that blindly changes REGNOs of the RTX should do the
trick. Perhaps do it on a copied RTX, to avoid nasty surprises.