* [Bug other/24757] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
@ 2005-11-09 17:00 ` pcarlini at suse dot de
2005-11-09 17:35 ` pinskia at gcc dot gnu dot org
` (33 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pcarlini at suse dot de @ 2005-11-09 17:00 UTC (permalink / raw)
To: gcc-bugs
------- Comment #1 from pcarlini at suse dot de 2005-11-09 17:00 -------
(In reply to comment #0)
> Those tests *never* fail in 4_0-branch, which doesn't use the builtins, and
> never did in mainline before the below of mine (and a simultaneous one to
> the compiler, which emptied ia64intrin.h)
Whereas it *is* true that those tests never fail in 4_0-branch, the second part
is not strictly true: sometimes, in mainline the tests failed also before that
date. I would say, much less consistently. This is all I know, for now.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug other/24757] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
2005-11-09 17:00 ` [Bug other/24757] " pcarlini at suse dot de
@ 2005-11-09 17:35 ` pinskia at gcc dot gnu dot org
2005-11-09 17:45 ` pcarlini at suse dot de
` (32 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pinskia at gcc dot gnu dot org @ 2005-11-09 17:35 UTC (permalink / raw)
To: gcc-bugs
------- Comment #2 from pinskia at gcc dot gnu dot org 2005-11-09 17:35 -------
Hmm you said in:
http://gcc.gnu.org/ml/libstdc++/2005-11/msg00149.html
That was really a glibc bug.
And actually 4.0 and before uses the builtins for ia64, this is where the
builtins came from in the first place.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug other/24757] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
2005-11-09 17:00 ` [Bug other/24757] " pcarlini at suse dot de
2005-11-09 17:35 ` pinskia at gcc dot gnu dot org
@ 2005-11-09 17:45 ` pcarlini at suse dot de
2005-11-09 17:51 ` pcarlini at suse dot de
` (31 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pcarlini at suse dot de @ 2005-11-09 17:45 UTC (permalink / raw)
To: gcc-bugs
------- Comment #3 from pcarlini at suse dot de 2005-11-09 17:45 -------
(In reply to comment #2)
> Hmm you said in:
> http://gcc.gnu.org/ml/libstdc++/2005-11/msg00149.html
>
> That was really a glibc bug.
Exactly *was*. Ehi, do you think I'm stupid? Of course in the meanwhile
I have checked that the problem is present in machines equipped with
glibc2.3.5 and is not, on the same machines, for 4_0-branch.
> And actually 4.0 and before uses the builtins for ia64, this is where the
> builtins came from in the first place.
Exactly. But, given the glibc bug above, masking the gcc behavior well into
2005 for some of our machines, at least, isn't possible that when we switched
to builtins everywhere something got broken? Alternately, the ia64 builtins
themselves can be defective, but that seems much less likely to me, because
we are talking about a very consistent behavior for relatively simple usages
of one single mutex and one single atomic counter.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug other/24757] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (2 preceding siblings ...)
2005-11-09 17:45 ` pcarlini at suse dot de
@ 2005-11-09 17:51 ` pcarlini at suse dot de
2005-11-09 19:30 ` pcarlini at suse dot de
` (30 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pcarlini at suse dot de @ 2005-11-09 17:51 UTC (permalink / raw)
To: gcc-bugs
------- Comment #4 from pcarlini at suse dot de 2005-11-09 17:51 -------
(In reply to comment #3)
> .............................................Alternately, the ia64 builtins
> themselves can be defective, but that seems much less likely to me, because
> we are talking about a very consistent behavior for relatively simple usages
> of one single mutex and one single atomic counter.
And this is also inconsistent with 4_0-branch, where we have no problems, never
during the last 6 months, at least.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug other/24757] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (3 preceding siblings ...)
2005-11-09 17:51 ` pcarlini at suse dot de
@ 2005-11-09 19:30 ` pcarlini at suse dot de
2005-11-10 1:41 ` [Bug target/24757] " pcarlini at suse dot de
` (29 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pcarlini at suse dot de @ 2005-11-09 19:30 UTC (permalink / raw)
To: gcc-bugs
------- Comment #5 from pcarlini at suse dot de 2005-11-09 19:30 -------
Some additional details from testresults: the bulk of the builtin atomics
changes
went in around mid of April, the ia64 specific bits, on April, 14. All the
results that Andreas sent at the beginning of the month (for instance april, 7,
8, 9) are fine. Then, at the end of the month (for instance, april, 22, 23, 24)
the results are not clean anymore.
Lately, never less than 2 failures, often 4, in threading tests.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (4 preceding siblings ...)
2005-11-09 19:30 ` pcarlini at suse dot de
@ 2005-11-10 1:41 ` pcarlini at suse dot de
2005-11-10 9:59 ` [Bug target/24757] [4.1 Regression] " pcarlini at suse dot de
` (28 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pcarlini at suse dot de @ 2005-11-10 1:41 UTC (permalink / raw)
To: gcc-bugs
------- Comment #6 from pcarlini at suse dot de 2005-11-10 01:41 -------
Actually, I'm pretty much worried by this issue: whatever is at fault, as a
matter of fact, for this target the library testsuite shows a severe
misbehavior
in a MT environment. Raising the priority/severity seems an appropriate action.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (5 preceding siblings ...)
2005-11-10 1:41 ` [Bug target/24757] " pcarlini at suse dot de
@ 2005-11-10 9:59 ` pcarlini at suse dot de
2005-11-17 13:50 ` pcarlini at suse dot de
` (27 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pcarlini at suse dot de @ 2005-11-10 9:59 UTC (permalink / raw)
To: gcc-bugs
------- Comment #7 from pcarlini at suse dot de 2005-11-10 09:59 -------
I'm marking this as a 4.1 Regression: certainly it is. If people can figure a
way to workaround it at the library level, I'm ok with that.
--
pcarlini at suse dot de changed:
What |Removed |Added
----------------------------------------------------------------------------
Severity|normal |major
Status|UNCONFIRMED |NEW
Ever Confirmed|0 |1
Last reconfirmed|0000-00-00 00:00:00 |2005-11-10 09:59:06
date| |
Summary|__sync_fetch_and_add on ia64|[4.1 Regression]
| |__sync_fetch_and_add on ia64
Target Milestone|--- |4.1.0
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (6 preceding siblings ...)
2005-11-10 9:59 ` [Bug target/24757] [4.1 Regression] " pcarlini at suse dot de
@ 2005-11-17 13:50 ` pcarlini at suse dot de
2005-11-18 15:20 ` pinskia at gcc dot gnu dot org
` (26 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pcarlini at suse dot de @ 2005-11-17 13:50 UTC (permalink / raw)
To: gcc-bugs
------- Comment #8 from pcarlini at suse dot de 2005-11-17 13:50 -------
I have got additional evidence that __sync_fetch_and_add does not work
correctly. If, for example, in this code, stressed in the testcase
12658_thread-1.cc, and using atomic operations:
const locale&
locale::operator=(const locale& __other) throw()
{
__other._M_impl->_M_add_reference();
_M_impl->_M_remove_reference();
_M_impl = __other._M_impl;
return *this;
}
I wrap everything in the locale_cons mutex then the testcase doesn't fail
anymore.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (7 preceding siblings ...)
2005-11-17 13:50 ` pcarlini at suse dot de
@ 2005-11-18 15:20 ` pinskia at gcc dot gnu dot org
2005-11-18 15:23 ` pcarlini at suse dot de
` (25 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pinskia at gcc dot gnu dot org @ 2005-11-18 15:20 UTC (permalink / raw)
To: gcc-bugs
------- Comment #9 from pinskia at gcc dot gnu dot org 2005-11-18 15:20 -------
I should note that
ld.acq+cmpxchg.rel is a full barrier.
as noted in ia64.c so a mf is not required.
See also http://www.ussg.iu.edu/hypermail/linux/kernel/0205.1/0226.html
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (8 preceding siblings ...)
2005-11-18 15:20 ` pinskia at gcc dot gnu dot org
@ 2005-11-18 15:23 ` pcarlini at suse dot de
2005-11-18 15:24 ` pinskia at gcc dot gnu dot org
` (24 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pcarlini at suse dot de @ 2005-11-18 15:23 UTC (permalink / raw)
To: gcc-bugs
------- Comment #10 from pcarlini at suse dot de 2005-11-18 15:23 -------
(In reply to comment #9)
> I should note that
> ld.acq+cmpxchg.rel is a full barrier.
> as noted in ia64.c so a mf is not required.
Did you read the last message from Alexander on the libstdc++ mailing list, in
particular the last paragraph?
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (9 preceding siblings ...)
2005-11-18 15:23 ` pcarlini at suse dot de
@ 2005-11-18 15:24 ` pinskia at gcc dot gnu dot org
2005-11-18 15:27 ` pinskia at gcc dot gnu dot org
` (23 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pinskia at gcc dot gnu dot org @ 2005-11-18 15:24 UTC (permalink / raw)
To: gcc-bugs
------- Comment #11 from pinskia at gcc dot gnu dot org 2005-11-18 15:24 -------
(In reply to comment #9)
> I should note that
> ld.acq+cmpxchg.rel is a full barrier.
> as noted in ia64.c so a mf is not required.
If that is really the issue, then
http://gcc.gnu.org/ml/gcc-patches/2005-05/msg00872.html casued it but I really
doubt it from reading that linux kernel message.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (10 preceding siblings ...)
2005-11-18 15:24 ` pinskia at gcc dot gnu dot org
@ 2005-11-18 15:27 ` pinskia at gcc dot gnu dot org
2005-11-18 15:29 ` pinskia at gcc dot gnu dot org
` (22 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pinskia at gcc dot gnu dot org @ 2005-11-18 15:27 UTC (permalink / raw)
To: gcc-bugs
------- Comment #12 from pinskia at gcc dot gnu dot org 2005-11-18 15:27 -------
(In reply to comment #10)
> Did you read the last message from Alexander on the libstdc++ mailing list, in
> particular the last paragraph?
Yes but I trust RTH better than Alexander on this issue.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (11 preceding siblings ...)
2005-11-18 15:27 ` pinskia at gcc dot gnu dot org
@ 2005-11-18 15:29 ` pinskia at gcc dot gnu dot org
2005-11-18 15:32 ` pcarlini at suse dot de
` (21 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pinskia at gcc dot gnu dot org @ 2005-11-18 15:29 UTC (permalink / raw)
To: gcc-bugs
------- Comment #13 from pinskia at gcc dot gnu dot org 2005-11-18 15:29 -------
(In reply to comment #12)
> Yes but I trust RTH better than Alexander on this issue.
And if you read the message which I pointed to:
The semaphore instructions also have implicit ordering.
So if I am reading this wrong, then there is something else wrong.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (12 preceding siblings ...)
2005-11-18 15:29 ` pinskia at gcc dot gnu dot org
@ 2005-11-18 15:32 ` pcarlini at suse dot de
2005-11-18 15:35 ` pinskia at gcc dot gnu dot org
` (20 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pcarlini at suse dot de @ 2005-11-18 15:32 UTC (permalink / raw)
To: gcc-bugs
------- Comment #14 from pcarlini at suse dot de 2005-11-18 15:32 -------
I'm not sure that specific patch is the culprit, because (see Comment #5) the
threaded testcases started failing before that date. I would like to see the
assembly before and after that date. In any case, it would be nice if Richard
Henderson could analyze a bit this PR. Also relevant:
http://gcc.gnu.org/ml/libstdc++/2005-11/msg00206.html
http://gcc.gnu.org/ml/libstdc++/2005-11/msg00207.html
--
pcarlini at suse dot de changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |rth at gcc dot gnu dot org
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (13 preceding siblings ...)
2005-11-18 15:32 ` pcarlini at suse dot de
@ 2005-11-18 15:35 ` pinskia at gcc dot gnu dot org
2005-11-18 23:32 ` rth at gcc dot gnu dot org
` (19 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pinskia at gcc dot gnu dot org @ 2005-11-18 15:35 UTC (permalink / raw)
To: gcc-bugs
------- Comment #15 from pinskia at gcc dot gnu dot org 2005-11-18 15:35 -------
See also:
http://www.linuxbase.org/spec/refspecs/IA64-softdevman-vol1.pdf
Page73.
Which is where the comment in comment #13. So If I read table 4-20 currently
then the memory access will always be ordered in that loop.
And if it started before that date, then there is something wrong as that is
when the major difference between 4.0 and 4.1 started.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (14 preceding siblings ...)
2005-11-18 15:35 ` pinskia at gcc dot gnu dot org
@ 2005-11-18 23:32 ` rth at gcc dot gnu dot org
2005-11-18 23:41 ` pcarlini at suse dot de
` (18 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: rth at gcc dot gnu dot org @ 2005-11-18 23:32 UTC (permalink / raw)
To: gcc-bugs
------- Comment #16 from rth at gcc dot gnu dot org 2005-11-18 23:32 -------
Alex's explanation can't be all of it. I hacked the compiler to emit two
"mf" instructions before and after the sequence here, and it made no
difference to the two tests failing. I really have no idea what's going on.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (15 preceding siblings ...)
2005-11-18 23:32 ` rth at gcc dot gnu dot org
@ 2005-11-18 23:41 ` pcarlini at suse dot de
2005-11-19 0:56 ` rth at gcc dot gnu dot org
` (17 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pcarlini at suse dot de @ 2005-11-18 23:41 UTC (permalink / raw)
To: gcc-bugs
------- Comment #17 from pcarlini at suse dot de 2005-11-18 23:41 -------
(In reply to comment #16)
> Alex's explanation can't be all of it. I hacked the compiler to emit two
> "mf" instructions before and after the sequence here, and it made no
> difference to the two tests failing. I really have no idea what's going on.
Thanks for your experiments. I'll try to help further, come up with a reduced
piece of code actually failing, investigate whether something else important
went in mid of April, maybe. I don't have much more, right now :(
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (16 preceding siblings ...)
2005-11-18 23:41 ` pcarlini at suse dot de
@ 2005-11-19 0:56 ` rth at gcc dot gnu dot org
2005-11-19 1:05 ` pcarlini at suse dot de
` (16 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: rth at gcc dot gnu dot org @ 2005-11-19 0:56 UTC (permalink / raw)
To: gcc-bugs
------- Comment #18 from rth at gcc dot gnu dot org 2005-11-19 00:56 -------
Dammit, lemme try again; I patched a different tree than I tested.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (17 preceding siblings ...)
2005-11-19 0:56 ` rth at gcc dot gnu dot org
@ 2005-11-19 1:05 ` pcarlini at suse dot de
2005-11-19 1:06 ` pcarlini at suse dot de
` (15 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pcarlini at suse dot de @ 2005-11-19 1:05 UTC (permalink / raw)
To: gcc-bugs
------- Comment #19 from pcarlini at suse dot de 2005-11-19 01:05 -------
Created an attachment (id=10287)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=10287&action=view)
Slightly hacked, self contained test (for use outside the testsuite)
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (18 preceding siblings ...)
2005-11-19 1:05 ` pcarlini at suse dot de
@ 2005-11-19 1:06 ` pcarlini at suse dot de
2005-11-19 1:08 ` pcarlini at suse dot de
` (14 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pcarlini at suse dot de @ 2005-11-19 1:06 UTC (permalink / raw)
To: gcc-bugs
------- Comment #20 from pcarlini at suse dot de 2005-11-19 01:06 -------
Created an attachment (id=10288)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=10288&action=view)
Slightly hacked, self contained test (for use outside the testsuite)
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (19 preceding siblings ...)
2005-11-19 1:06 ` pcarlini at suse dot de
@ 2005-11-19 1:08 ` pcarlini at suse dot de
2005-11-19 2:08 ` mmitchel at gcc dot gnu dot org
` (13 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pcarlini at suse dot de @ 2005-11-19 1:08 UTC (permalink / raw)
To: gcc-bugs
------- Comment #21 from pcarlini at suse dot de 2005-11-19 01:08 -------
Just in case, those are ready to use outside the testsuite... For me,
especially
the second one, fails very quickly and consistently on 4-way and 16-way.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (20 preceding siblings ...)
2005-11-19 1:08 ` pcarlini at suse dot de
@ 2005-11-19 2:08 ` mmitchel at gcc dot gnu dot org
2005-11-19 2:09 ` rth at gcc dot gnu dot org
` (12 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: mmitchel at gcc dot gnu dot org @ 2005-11-19 2:08 UTC (permalink / raw)
To: gcc-bugs
------- Comment #22 from mmitchel at gcc dot gnu dot org 2005-11-19 02:08 -------
I think we need to understand this problem before we release. It might be that
we downgrade the priority if it turns out to be a problem outside our control.
--
mmitchel at gcc dot gnu dot org changed:
What |Removed |Added
----------------------------------------------------------------------------
Priority|P3 |P1
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (21 preceding siblings ...)
2005-11-19 2:08 ` mmitchel at gcc dot gnu dot org
@ 2005-11-19 2:09 ` rth at gcc dot gnu dot org
2005-11-19 2:15 ` pcarlini at suse dot de
` (11 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: rth at gcc dot gnu dot org @ 2005-11-19 2:09 UTC (permalink / raw)
To: gcc-bugs
------- Comment #23 from rth at gcc dot gnu dot org 2005-11-19 02:09 -------
Created an attachment (id=10289)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=10289&action=view)
mf hack
For the record, here's the aforementioned hack.
And it does appear to change the behaviour of one of the testsuite tests,
but it doesn't fix both of them. I'll claim, then, that it didn't fix either
of them, but merely changed the timing. Feel free to tell me what's still
not right, but it seems like, with a hammer this big, that it isn't the
synchronization that's the problem.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (22 preceding siblings ...)
2005-11-19 2:09 ` rth at gcc dot gnu dot org
@ 2005-11-19 2:15 ` pcarlini at suse dot de
2005-11-19 2:33 ` pcarlini at suse dot de
` (10 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pcarlini at suse dot de @ 2005-11-19 2:15 UTC (permalink / raw)
To: gcc-bugs
------- Comment #24 from pcarlini at suse dot de 2005-11-19 02:15 -------
(In reply to comment #23)
> Created an attachment (id=10289)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=10289&action=view) [edit]
> mf hack
>
> For the record, here's the aforementioned hack.
Ok, thanks. Over the WE I'll play a bit with it. Among other things, I want
to check its effects on other tests, like the thread/pthread*.cc, and also
for a different version of the library, which uses more refcounting in the
string class and for me exposed many more troubles during the last days.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (23 preceding siblings ...)
2005-11-19 2:15 ` pcarlini at suse dot de
@ 2005-11-19 2:33 ` pcarlini at suse dot de
2005-11-19 22:12 ` pcarlini at suse dot de
` (9 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pcarlini at suse dot de @ 2005-11-19 2:33 UTC (permalink / raw)
To: gcc-bugs
------- Comment #25 from pcarlini at suse dot de 2005-11-19 02:33 -------
(In reply to comment #23)
> And it does appear to change the behaviour of one of the testsuite tests,
> but it doesn't fix both of them. I'll claim, then, that it didn't fix either
> of them, but merely changed the timing.
Confirmed here. Actually, for me, the "improvement" is barely noticeable. It's
something else...
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (24 preceding siblings ...)
2005-11-19 2:33 ` pcarlini at suse dot de
@ 2005-11-19 22:12 ` pcarlini at suse dot de
2005-11-19 22:38 ` rguenth at gcc dot gnu dot org
` (8 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pcarlini at suse dot de @ 2005-11-19 22:12 UTC (permalink / raw)
To: gcc-bugs
------- Comment #26 from pcarlini at suse dot de 2005-11-19 22:12 -------
Created an attachment (id=10297)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=10297&action=view)
Reduced from thread/pthread3
Richard, I'm attaching a really minimal testcase, which fails very quickly
for me with mainline or 4_1-branch and doesn't with 4_0-branch. It's just
two threads and many constructions and destructions of a trivial class
which increases (descreases, respectively) atomically a static counter.
Of course the abort should never trigger.
Can you debug further with this? Otherwise, just tell me, ok?
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (25 preceding siblings ...)
2005-11-19 22:12 ` pcarlini at suse dot de
@ 2005-11-19 22:38 ` rguenth at gcc dot gnu dot org
2005-11-19 22:45 ` pcarlini at suse dot de
` (7 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: rguenth at gcc dot gnu dot org @ 2005-11-19 22:38 UTC (permalink / raw)
To: gcc-bugs
------- Comment #27 from rguenth at gcc dot gnu dot org 2005-11-19 22:38 -------
The only thing in the pthread3_reduced.cc testcase that could make it going
wrong is wrong alignment of aw. And I remember seeing a lot of unaligned traps
in dmesg during libstdc++ testing on ia64.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (26 preceding siblings ...)
2005-11-19 22:38 ` rguenth at gcc dot gnu dot org
@ 2005-11-19 22:45 ` pcarlini at suse dot de
2005-11-19 23:02 ` pcarlini at suse dot de
` (6 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pcarlini at suse dot de @ 2005-11-19 22:45 UTC (permalink / raw)
To: gcc-bugs
------- Comment #28 from pcarlini at suse dot de 2005-11-19 22:45 -------
Thanks Richard G. In fact, from time to time I saw warnings on the shell,
which really puzzled me, because I was sure that no alignment issues were
present anymore in the higher lever library code proper.
I'll try to add an alignment attribute to the _Atomic_word typedef, as cris
is already doing for instance, and see whether something changes for the
better. Interestingly no such issue existed with 4_0-branch... Any idea why?
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (27 preceding siblings ...)
2005-11-19 22:45 ` pcarlini at suse dot de
@ 2005-11-19 23:02 ` pcarlini at suse dot de
2005-11-19 23:06 ` pcarlini at suse dot de
` (5 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pcarlini at suse dot de @ 2005-11-19 23:02 UTC (permalink / raw)
To: gcc-bugs
------- Comment #29 from pcarlini at suse dot de 2005-11-19 23:02 -------
(In reply to comment #28)
> I'll try to add an alignment attribute to the _Atomic_word typedef, as cris
> is already doing for instance, and see whether something changes for the
> better.
Nope. Maybe the alignment is for some reason broken to the point that not
even an explicit __attribute__ ((aligned (X))) can fix it... I don't know.
Anyway, the problem is very easy to reproduce, now. I'm also attaching a
further reduced, pure C testcase.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (28 preceding siblings ...)
2005-11-19 23:02 ` pcarlini at suse dot de
@ 2005-11-19 23:06 ` pcarlini at suse dot de
2005-11-20 0:22 ` schwab at suse dot de
` (4 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: pcarlini at suse dot de @ 2005-11-19 23:06 UTC (permalink / raw)
To: gcc-bugs
------- Comment #30 from pcarlini at suse dot de 2005-11-19 23:06 -------
Created an attachment (id=10298)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=10298&action=view)
Further reduced, pure C
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (29 preceding siblings ...)
2005-11-19 23:06 ` pcarlini at suse dot de
@ 2005-11-20 0:22 ` schwab at suse dot de
2005-11-20 5:43 ` rth at gcc dot gnu dot org
` (3 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: schwab at suse dot de @ 2005-11-20 0:22 UTC (permalink / raw)
To: gcc-bugs
------- Comment #31 from schwab at suse dot de 2005-11-20 00:22 -------
Testing this patch:
Index: ia64.c
===================================================================
--- ia64.c (revision 107220)
+++ ia64.c (working copy)
@@ -2113,7 +2113,7 @@
emit_insn (GEN_FCN (icode) (cmp_reg, mem, ar_ccv, new_reg));
- emit_cmp_and_jump_insns (cmp_reg, old_reg, EQ, NULL, DImode, true, label);
+ emit_cmp_and_jump_insns (cmp_reg, old_reg, NE, NULL, DImode, true, label);
}
/* Begin the assembly file. */
--
schwab at suse dot de changed:
What |Removed |Added
----------------------------------------------------------------------------
AssignedTo|unassigned at gcc dot gnu |schwab at suse dot de
|dot org |
Status|NEW |ASSIGNED
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (30 preceding siblings ...)
2005-11-20 0:22 ` schwab at suse dot de
@ 2005-11-20 5:43 ` rth at gcc dot gnu dot org
2005-11-20 10:43 ` schwab at gcc dot gnu dot org
` (2 subsequent siblings)
34 siblings, 0 replies; 36+ messages in thread
From: rth at gcc dot gnu dot org @ 2005-11-20 5:43 UTC (permalink / raw)
To: gcc-bugs
------- Comment #32 from rth at gcc dot gnu dot org 2005-11-20 05:43 -------
Oh good grief. I can't see the forest for the trees. Andreas, I expect
that patch will work. If it tests ok, please commit it.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (31 preceding siblings ...)
2005-11-20 5:43 ` rth at gcc dot gnu dot org
@ 2005-11-20 10:43 ` schwab at gcc dot gnu dot org
2005-11-20 10:44 ` schwab at gcc dot gnu dot org
2005-11-20 10:45 ` schwab at suse dot de
34 siblings, 0 replies; 36+ messages in thread
From: schwab at gcc dot gnu dot org @ 2005-11-20 10:43 UTC (permalink / raw)
To: gcc-bugs
------- Comment #33 from schwab at gcc dot gnu dot org 2005-11-20 10:43 -------
Subject: Bug 24757
Author: schwab
Date: Sun Nov 20 10:43:43 2005
New Revision: 107246
URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=107246
Log:
PR target/24757
* config/ia64/ia64.c (ia64_expand_atomic_op): Fix condition of cmp
insn.
Modified:
trunk/gcc/ChangeLog
trunk/gcc/config/ia64/ia64.c
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (32 preceding siblings ...)
2005-11-20 10:43 ` schwab at gcc dot gnu dot org
@ 2005-11-20 10:44 ` schwab at gcc dot gnu dot org
2005-11-20 10:45 ` schwab at suse dot de
34 siblings, 0 replies; 36+ messages in thread
From: schwab at gcc dot gnu dot org @ 2005-11-20 10:44 UTC (permalink / raw)
To: gcc-bugs
------- Comment #34 from schwab at gcc dot gnu dot org 2005-11-20 10:44 -------
Subject: Bug 24757
Author: schwab
Date: Sun Nov 20 10:44:27 2005
New Revision: 107247
URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=107247
Log:
PR target/24757
* config/ia64/ia64.c (ia64_expand_atomic_op): Fix condition of cmp
insn.
Modified:
branches/gcc-4_1-branch/gcc/ChangeLog
branches/gcc-4_1-branch/gcc/config/ia64/ia64.c
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread* [Bug target/24757] [4.1 Regression] __sync_fetch_and_add on ia64
2005-11-09 16:45 [Bug other/24757] New: __sync_fetch_and_add on ia64 pcarlini at suse dot de
` (33 preceding siblings ...)
2005-11-20 10:44 ` schwab at gcc dot gnu dot org
@ 2005-11-20 10:45 ` schwab at suse dot de
34 siblings, 0 replies; 36+ messages in thread
From: schwab at suse dot de @ 2005-11-20 10:45 UTC (permalink / raw)
To: gcc-bugs
------- Comment #35 from schwab at suse dot de 2005-11-20 10:45 -------
Fixed.
--
schwab at suse dot de changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|ASSIGNED |RESOLVED
Resolution| |FIXED
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24757
^ permalink raw reply [flat|nested] 36+ messages in thread