------- Comment #3 from kkojima at gcc dot gnu dot org 2008-02-18 04:36 -------
It seems that sh.c:find_barrier doesn't handle a corner case
which occurs with the testcase. Although the given testcases
don't fail on 4.3, it'd be a hidden issue even on trunk. Now
I'm testing the patch below.
I'll apply it on trunk first and then 4.2/4.3 if it survives
with usual tests after we are out of a current critical section
for branching.
diff -uprN ORIG/gcc-4_2-branch/gcc/config/sh/sh.c
LOCAL/gcc-4_2-branch/gcc/config/sh/sh.c
--- ORIG/gcc-4_2-branch/gcc/config/sh/sh.c 2008-01-18 22:11:17.000000000
+0900
+++ LOCAL/gcc-4_2-branch/gcc/config/sh/sh.c 2008-02-18 11:25:56.000000000
+0900
@@ -3530,6 +3530,7 @@ find_barrier (int num_mova, rtx mova, rt
rtx barrier_before_mova = 0, found_barrier = 0, good_barrier = 0;
int si_limit;
int hi_limit;
+ rtx orig = from;
/* For HImode: range is 510, add 4 because pc counts from address of
second instruction after this one, subtract 2 for the jump instruction
@@ -3753,7 +3754,8 @@ find_barrier (int num_mova, rtx mova, rt
/* If we exceeded the range, then we must back up over the last
instruction we looked at. Otherwise, we just need to undo the
NEXT_INSN at the end of the loop. */
- if (count_hi > hi_limit || count_si > si_limit)
+ if (PREV_INSN (from) != orig
+ && (count_hi > hi_limit || count_si > si_limit))
from = PREV_INSN (PREV_INSN (from));
else
from = PREV_INSN (from);
--
kkojima at gcc dot gnu dot org changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |kkojima at gcc dot gnu dot
| |org
Status|UNCONFIRMED |NEW
Ever Confirmed|0 |1
Last reconfirmed|0000-00-00 00:00:00 |2008-02-18 04:36:05
date| |
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35225
