https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70261
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Target Milestone|--- |6.0
Summary|r234265 causes fails on |[6 Regression] r234265
|rs6000 |causes fails on rs6000
--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
This is combiner calling replace_rtx with x = (zero_extend:DI (reg:SI 3 3 [ p1
])) from = (reg:DI 3 3) to = (sign_extend:DI (clobber:SI (const_int 0 [0])))
(it will need better analysis, but perhaps it was just some optimization that
wouldn't break stuff it not everything is replaced).
So maybe we can just limit this to reload_completed (assert the mode is the
same), otherwise simply punt on replacing it at least for GCC 6.x, like we used
to do?
--- gcc/rtlanal.c 2016-03-16 18:51:43.735688164 +0100
+++ gcc/rtlanal.c 2016-03-17 00:24:28.949901218 +0100
@@ -2966,8 +2966,10 @@ replace_rtx (rtx x, rtx from, rtx to)
if (GET_CODE (from) == REG
&& REGNO (x) == REGNO (from))
{
- gcc_assert (GET_MODE (x) == GET_MODE (from));
- return to;
+ if (GET_MODE (x) == GET_MODE (from))
+ return to;
+ else
+ gcc_assert (!reload_completed);
}
}
else if (GET_CODE (x) == SUBREG)
