https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70416
--- Comment #15 from Oleg Endo <olegendo at gcc dot gnu.org> ---
(In reply to Oleg Endo from comment #14)
>
> Maybe at that stage in the reload code it will end up using the last *addsi3
> pattern and not try to look for a new pattern in the .md when it wants to
> change it. In other words, maybe it'll help if the *addsi3 patterns are
> merged into a single pattern somehow. I'll give it a try...
The following seems to fix the problem:
Index: gcc/config/sh/sh.md
===================================================================
--- gcc/config/sh/sh.md (revision 234258)
+++ gcc/config/sh/sh.md (working copy)
@@ -2248,7 +2248,7 @@
(plus:SI (match_operand:SI 1 "arith_reg_operand" "r")
(match_operand:SI 2 "arith_or_int_operand" "rn")))]
"TARGET_SH1 && !sh_lra_p ()
- && reload_completed
+ && (reload_completed || reload_in_progress)
&& !reg_overlap_mentioned_p (operands[0], operands[1])"
"#"
"&& 1"
