--
rguenth at gcc dot gnu dot org changed:
What|Removed |Added
Summary|glibc regex testsuite |[4.5 Regression] glibc regex
|failures
--- Comment #1 from rguenth at gcc dot gnu dot org 2010-03-17 15:19 ---
Waiting for testcase.
--
rguenth at gcc dot gnu dot org changed:
What|Removed |Added
S
--- Comment #2 from rguenth at gcc dot gnu dot org 2010-03-23 13:39 ---
David?
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43385
--- Comment #3 from davem at gcc dot gnu dot org 2010-03-23 20:40 ---
Sorry, it's taking a long time to sort out the test case. The
GLIBC regex code is huge and non-trivial.
However I did track down that it might be dependent upon optimization
options, for example I can reproduce with
--- Comment #4 from jakub at gcc dot gnu dot org 2010-03-23 22:41 ---
Wouldn't this be the same issue as
http://gcc.gnu.org/ml/gcc-patches/2010-03/msg01086.html
?
On the trunk it wasn't just one commit, but several:
156888, 156889, 156893, 156903, 156913
If yes, would be interesting to k
--- Comment #5 from davem at davemloft dot net 2010-03-24 03:01 ---
Yes, the failures mentioned in:
http://gcc.gnu.org/ml/gcc-patches/2010-03/msg01086.html
are the same exact ones I am seeing.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43385
--- Comment #6 from davem at davemloft dot net 2010-03-24 06:00 ---
In the regex cases, glibc/posix/regexec.c:merge_state_with_log() is what
gets miscompiled.
I will attach match_good.s and match_bad.s
match_good.s is a working compile of this function, using gcc-4.3.2
match_bad.s is
--- Comment #7 from davem at davemloft dot net 2010-03-24 06:01 ---
Created an attachment (id=20177)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=20177&action=view)
correctly compiled function
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43385
--- Comment #8 from davem at davemloft dot net 2010-03-24 06:02 ---
Created an attachment (id=20178)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=20178&action=view)
miscompiled function
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43385
--- Comment #9 from jakub at gcc dot gnu dot org 2010-03-24 06:38 ---
Could you please also attach regexec.i (and what gcc options were used to
compile it), so that I can inspect it with a cross-compiler? Thanks.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43385
--- Comment #10 from davem at davemloft dot net 2010-03-24 06:43 ---
Created an attachment (id=20179)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=20179&action=view)
source that emitted miscompiled function
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43385
--- Comment #11 from jakub at gcc dot gnu dot org 2010-03-24 07:05 ---
Created an attachment (id=20180)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=20180&action=view)
hack
Minimal patch that should allow playing with reversion of both parts of PR42233
just by changing that 1 to
--- Comment #12 from jakub at gcc dot gnu dot org 2010-03-24 07:08 ---
FYI, the only change I'm seeing is:
- subcc %g0, %g1, %g0
- addx%g0, 0, %g1
with 1 vs. 0 in gimple_boolify and no effect with the other changes. That
said, I'm probably using different cc1 options t
--- Comment #13 from davem at davemloft dot net 2010-03-24 16:55 ---
I'm not passing anything special to the build, just stock "-O2" with
a 32-bit compiler.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43385
--- Comment #14 from jakub at gcc dot gnu dot org 2010-03-24 17:02 ---
Can you please try the #c11 patch with 0 && instead of 1 && in gimple_boolify
and see whether you see any difference beyond those two insns (and whether it
works)?
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=4
--- Comment #15 from jakub at gcc dot gnu dot org 2010-03-24 17:04 ---
Perhaps also with -fverbose-asm see what options were actually used during the
compilation?
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43385
--- Comment #16 from davem at davemloft dot net 2010-03-24 17:08 ---
I'm trying to distill a test case currently and also something broke bootstrap
on sparc in the past day or two (I think it's the IRA change) which I want
to track down first.
I'll play with your patch once I get past t
--- Comment #17 from davem at davemloft dot net 2010-03-24 17:22 ---
I get the same two instruction change you saw with "0 &&" and it
makes the test pass.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43385
--- Comment #18 from doko at ubuntu dot com 2010-03-24 17:31 ---
seen on the 4.4 branch as well on
ia64-linux-gnu, configured --with-system-libunwind
sparc-linux-gnu, building with -m32 -mcpu=ultrasparc
arm-linux-gnueabi, configured with -with-arch=armv7-a --with-float=softfp
--with-fpu
--- Comment #19 from davem at davemloft dot net 2010-03-24 17:59 ---
Created an attachment (id=20186)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=20186&action=view)
Distilled test case.
The expression that causes problems is:
if (__builtin_expect (integer, 0) && ptr != NULL)
--- Comment #20 from jakub at gcc dot gnu dot org 2010-03-24 20:54 ---
Thanks, will debug tomorrow. At least it is in gimple_boolify which is much
shorter than the other changes...
--
jakub at gcc dot gnu dot org changed:
What|Removed |Added
-
--- Comment #21 from jakub at gcc dot gnu dot org 2010-03-24 22:48 ---
I guess we want to call gimple_boolify on the __builtin_expect argument only if
it is (inside of the (long) cast) one of the handled cases
(TRUTH_*_EXPR, comparisons), but not in the case it is neither of these - then
--- Comment #22 from jakub at gcc dot gnu dot org 2010-03-24 23:41 ---
Created an attachment (id=20190)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=20190&action=view)
gcc45-pr43385.patch
So far only very lightly tested fix.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=433
--- Comment #23 from jakub at gcc dot gnu dot org 2010-03-24 23:54 ---
Created an attachment (id=20191)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=20191&action=view)
gcc45-pr43385.patch
Simpler version of the patch.
--
jakub at gcc dot gnu dot org changed:
What
--- Comment #24 from jakub at gcc dot gnu dot org 2010-03-25 00:15 ---
Created an attachment (id=20192)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=20192&action=view)
gcc45-pr43385.patch
And now with a testcase (hopefully it fails on sparc* without the patch).
--
jakub at gc
--- Comment #25 from doko at ubuntu dot com 2010-03-25 00:23 ---
the testcase fails without the patch (on sparc, 4.4 branch)
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43385
--- Comment #26 from davem at davemloft dot net 2010-03-25 03:41 ---
I'll run this patch through my tests, thanks Jakub.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43385
--- Comment #27 from jakub at gcc dot gnu dot org 2010-03-25 06:53 ---
More complete testcase:
/* PR c/43385 */
extern void abort (void);
int e;
__attribute__((noinline, noclone)) void
foo (int x, int y)
{
if (__builtin_expect (x, 0) && y != 0)
e++;
}
__attribute__((noinline, n
--- Comment #28 from jakub at gcc dot gnu dot org 2010-03-25 08:16 ---
Subject: Bug 43385
Author: jakub
Date: Thu Mar 25 08:16:18 2010
New Revision: 157721
URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=157721
Log:
PR c/43385
* gimplify.c (gimple_boolify): Only r
--- Comment #29 from sezeroz at gmail dot com 2010-03-25 08:54 ---
The testcase fails with gcc-4.4 at -O1 and higher, too (x86_64-pc-linux-gnu,
with or without -m32.) It would be nice to have the fix backported to 4.4.
--
sezeroz at gmail dot com changed:
What|Removed
--- Comment #30 from doko at ubuntu dot com 2010-03-25 11:45 ---
checked on ia64 and sparc, that the patch fixes the problem with 4.4.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43385
--- Comment #31 from jakub at gcc dot gnu dot org 2010-03-25 11:55 ---
Subject: Bug 43385
Author: jakub
Date: Thu Mar 25 11:55:08 2010
New Revision: 157722
URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=157722
Log:
PR c/43385
* gimplify.c (gimple_boolify): Only r
32 matches
Mail list logo
