On Sun, 2016-09-25 at 16:06 +0900, Oleg Endo wrote:
>
> This fixes a fallout that actually goes back to 5.0 but went
> unnoticed. The costs for movt and movrt type of insns were not
> correctly reported and ifcvt thus made some bad choices for SH. A
> new cset_zero pattern variant is also required to fix the matching
> for some recent changes in the middle end.
>
> Tested on sh-elf with
> make -k check RUNTESTFLAGS="--target_board=sh-sim\{-m2/-ml,-m2/-mb,-
> m2a/-mb,-m4/-ml,-m4/-mb,-m4a/-ml,-m4a/-mb}"
>
> Committed as r240471.
> Backports to GCC 6 and GCC 5 will follow.
>
While working on the backport, I noticed that I've screwed up the costs
return value. Funny that it had no effect in this case.
Maybe rtx_costs functions should be made to return something like
std::optional<int> to avoid these kind of mistakes. It would also cut
the code by half in most cases.
Anyway, fixed as attached, re-tested as above and committed as r240533.
Cheers,
Oleg
gcc/ChangeLog
PR target/51244
* config/sh/sh.c (sh_rtx_costs): Fix return value of SET of movt and
movrt patterns. Match them before anything else in the SET case.
Index: gcc/config/sh/sh.c
===================================================================
--- gcc/config/sh/sh.c (revision 240471)
+++ gcc/config/sh/sh.c (working copy)
@@ -3199,6 +3199,12 @@
vector-move, so we have to provide the correct cost in the number
of move insns to load/store the reg of the mode in question. */
case SET:
+ if (sh_movt_set_dest (x) != NULL || sh_movrt_set_dest (x) != NULL)
+ {
+ *total = COSTS_N_INSNS (1);
+ return true;
+ }
+
if (register_operand (SET_DEST (x), VOIDmode)
&& (register_operand (SET_SRC (x), VOIDmode)
|| satisfies_constraint_Z (SET_SRC (x))))
@@ -3208,10 +3214,6 @@
/ mov_insn_size (mode, TARGET_SH2A));
return true;
}
-
- if (sh_movt_set_dest (x) != NULL || sh_movrt_set_dest (x) != NULL)
- return COSTS_N_INSNS (1);
-
return false;
/* The cost of a mem access is mainly the cost of the address mode. */
