Bug#324455: gmp: FTBFS on alpha
At Tue, 23 Aug 2005 01:27:43 -0700, Steve Langasek wrote: - cmpult Y, X, RV + cmpule Y, X, RV excb mt_fpcr $f3 ldt $f0, 0(sp) but I don't have time for testing. Thanks, after looking at the diff between divq.S and divqu.S and doing a little googling (aka, the Babelfish methodology for learning assembly), I came to the same conclusion, and my test build of glibc just finished up -- this one-liner does indeed fix the problem. I put this kind of patch with the additional fix for svn, it'll be fixed in 2.3.5-5. Regards, -- gotom -- To UNSUBSCRIBE, email to [EMAIL PROTECTED] with a subject of unsubscribe. Trouble? Contact [EMAIL PROTECTED]
Bug#324455: gmp: FTBFS on alpha
tags 324455 patch
thanks
Hi Falk,
On Tue, Aug 23, 2005 at 10:09:56AM +0200, Falk Hueffner wrote:
Steve Langasek [EMAIL PROTECTED] writes:
On Mon, Aug 22, 2005 at 08:56:23PM -0700, Steve Langasek wrote:
Simple test case:
int main() {
unsigned long a= 1UL 63;
unsigned long b= 1UL 63;
exit (a/b != 1);
}
Appears to be a bug in the implementation of __divqu. Upgrading to
2.3.5-4,
to confirm whether the bug still exists.
Yep, bug still present in 2.3.5-4. Someone who speaks alpha assembly will
have to look at this, I think, to pinpoint the error.
probably something like
--- divqu.S~2005-07-09 21:55:07.0 +0200
+++ divqu.S 2005-08-23 10:04:09.0 +0200
@@ -240,7 +240,7 @@
/* If we get here, Y is so big that bit 63 is set. The results
from the divide will be completely wrong. Fortunately, the
quotient must be either 0 or 1, so just compute it directly. */
- cmpult Y, X, RV
+ cmpule Y, X, RV
excb
mt_fpcr $f3
ldt $f0, 0(sp)
but I don't have time for testing.
Thanks, after looking at the diff between divq.S and divqu.S and doing a
little googling (aka, the Babelfish methodology for learning assembly),
I came to the same conclusion, and my test build of glibc just finished up
-- this one-liner does indeed fix the problem.
Cheers,
--
Steve Langasek Give me a lever long enough and a Free OS
Debian Developer to set it on, and I can move the world.
[EMAIL PROTECTED] http://www.debian.org/
signature.asc
Description: Digital signature
Bug#324455: gmp: FTBFS on alpha
Steve Langasek [EMAIL PROTECTED] writes:
On Mon, Aug 22, 2005 at 08:56:23PM -0700, Steve Langasek wrote:
Simple test case:
int main() {
unsigned long a= 1UL 63;
unsigned long b= 1UL 63;
exit (a/b != 1);
}
Appears to be a bug in the implementation of __divqu. Upgrading to 2.3.5-4,
to confirm whether the bug still exists.
Yep, bug still present in 2.3.5-4. Someone who speaks alpha assembly will
have to look at this, I think, to pinpoint the error.
probably something like
--- divqu.S~2005-07-09 21:55:07.0 +0200
+++ divqu.S 2005-08-23 10:04:09.0 +0200
@@ -240,7 +240,7 @@
/* If we get here, Y is so big that bit 63 is set. The results
from the divide will be completely wrong. Fortunately, the
quotient must be either 0 or 1, so just compute it directly. */
- cmpult Y, X, RV
+ cmpule Y, X, RV
excb
mt_fpcr $f3
ldt $f0, 0(sp)
but I don't have time for testing.
--
Falk
--
To UNSUBSCRIBE, email to [EMAIL PROTECTED]
with a subject of unsubscribe. Trouble? Contact [EMAIL PROTECTED]
Bug#324455: gmp: FTBFS on alpha
Package: gmp Severity: serious Justification: FTBFS Hi, gmp fails to build on alpha, see http://buildd.debian.org/fetch.php?pkg=gmpver=4.1.4-10arch=alphastamp=1124596243file=logas=raw I could reproduce it on our alpha machine at work. I tried to build with gcc-3.4 or gcc-3.3 as a workaround but it failed at the same point. Cheers, Laurent. signature.asc Description: Digital signature
Bug#324455: gmp: FTBFS on alpha
* Steve M. Robbins [Mon, Aug 22, 2005 at 11:21:00AM -0400]: I could reproduce it on our alpha machine at work. I tried to build with gcc-3.4 or gcc-3.3 as a workaround but it failed at the same point. Could you do me a favour? Try building with -O2. The same test failed with -O2, gcc 4.0.1-5. Trying with gcc-3.4 now. signature.asc Description: Digital signature
Bug#324455: gmp: FTBFS on alpha
On Mon, Aug 22, 2005 at 10:10:30AM +0200, Laurent Fousse wrote: Package: gmp Severity: serious Justification: FTBFS Hi, gmp fails to build on alpha, see http://buildd.debian.org/fetch.php?pkg=gmpver=4.1.4-10arch=alphastamp=1124596243file=logas=raw I could reproduce it on our alpha machine at work. I tried to build with gcc-3.4 or gcc-3.3 as a workaround but it failed at the same point. Could you do me a favour? Try building with -O2. The alpha buildd worked with debian revisions -7.1 and -8, but not with -9 or -10. The gmp code hasn't changed, but the compiler has. Version -8 was compiled with Toolchain package versions: libc6.1-dev_2.3.2.ds1-22 linux-kernel-headers_2.6.12.0-1 gcc-4.0_4.0.0-12 g++-4.0_4.0.0-12 binutils_2.16.1-1 libstdc++6-4.0-dev_4.0.0-12 libstdc++6_4.0.0-12 while version -9 was compiled with Toolchain package versions: libc6.1-dev_2.3.5-3 linux-kernel-headers_2.6.13+0rc3-1 gcc-4.0_4.0.1-3 g++-4.0_4.0.1-3 binutils_2.16.1-2 libstdc++6-4.0-dev_4.0.1-3 libstdc++6_4.0.1-3 I'll also have a look for new patches on the gmp site, later tonight. Thanks, -Steve -- To UNSUBSCRIBE, email to [EMAIL PROTECTED] with a subject of unsubscribe. Trouble? Contact [EMAIL PROTECTED]
Bug#324455: gmp: FTBFS on alpha
* Laurent Fousse [Mon, Aug 22, 2005 at 05:42:32PM +0200]: Could you do me a favour? Try building with -O2. The same test failed with -O2, gcc 4.0.1-5. Trying with gcc-3.4 now. Failed with 3.4 and 3.3, both using -O2. Something is wrong in the toolchain, only somewhere else. signature.asc Description: Digital signature
Bug#324455: gmp: FTBFS on alpha
On Mon, Aug 22, 2005 at 06:11:05PM +0200, Laurent Fousse wrote: * Laurent Fousse [Mon, Aug 22, 2005 at 05:42:32PM +0200]: Could you do me a favour? Try building with -O2. The same test failed with -O2, gcc 4.0.1-5. Trying with gcc-3.4 now. Failed with 3.4 and 3.3, both using -O2. Something is wrong in the toolchain, only somewhere else. FWIW, the test in question also fails when run against the *installed* version of ligbgmp3 here. So either this is a bug in both the old and new builds of libgmp3 on alpha that has just never manifested anywhere else, or there's a bug in the test case (either due to a changed test case or a miscompilation). -- Steve Langasek Give me a lever long enough and a Free OS Debian Developer to set it on, and I can move the world. [EMAIL PROTECTED] http://www.debian.org/ signature.asc Description: Digital signature
Bug#324455: gmp: FTBFS on alpha
On Mon, Aug 22, 2005 at 06:11:05PM +0200, Laurent Fousse wrote: * Laurent Fousse [Mon, Aug 22, 2005 at 05:42:32PM +0200]: Could you do me a favour? Try building with -O2. The same test failed with -O2, gcc 4.0.1-5. Trying with gcc-3.4 now. Failed with 3.4 and 3.3, both using -O2. Something is wrong in the toolchain, only somewhere else. Ok, having looked at the test in greater detail, it seems to have just been the luck of the draw that the test never failed before on alpha. [EMAIL PROTECTED]:/home/devel/release/gmp-4.1.4/tests/mpz$ ./dive_ui 500 [EMAIL PROTECTED]:/home/devel/release/gmp-4.1.4/tests/mpz$ ./dive_ui 1000 mpz_divexact_ui wrong a=1198415671637283 d=1198415671637283 q=1 got=0 Aborted $ Amusing that it's only dividing a number by itself that seems to give gmp fits here. :-) If I special-case q=1 out, I get a full sequence of 5000 tests without errors. Other values that don't fail the test include d=0x0056e18e03a9de91 and d=0x459f9e46f19b39d8; others that do include d=0xa679fe5e2536755c and d=0xd2e85bb7873d8a7e. Looks suspiciously like signed/unsigned breakage, as only values with the high bit set appear to be affected. Will continue investigating. -- Steve Langasek Give me a lever long enough and a Free OS Debian Developer to set it on, and I can move the world. [EMAIL PROTECTED] http://www.debian.org/ signature.asc Description: Digital signature
Bug#324455: gmp: FTBFS on alpha
reassign 324455 libc6.1 2.3.5-3
retitle 324455 regression from 2.3.2 in __divqu on alpha
thanks
On Mon, Aug 22, 2005 at 07:05:46PM -0700, Steve Langasek wrote:
On Mon, Aug 22, 2005 at 06:11:05PM +0200, Laurent Fousse wrote:
* Laurent Fousse [Mon, Aug 22, 2005 at 05:42:32PM +0200]:
Could you do me a favour? Try building with -O2.
The same test failed with -O2, gcc 4.0.1-5. Trying with gcc-3.4 now.
Failed with 3.4 and 3.3, both using -O2. Something is wrong in the
toolchain, only somewhere else.
Ok, having looked at the test in greater detail, it seems to have just been
the luck of the draw that the test never failed before on alpha.
[EMAIL PROTECTED]:/home/devel/release/gmp-4.1.4/tests/mpz$ ./dive_ui 500
[EMAIL PROTECTED]:/home/devel/release/gmp-4.1.4/tests/mpz$ ./dive_ui 1000
mpz_divexact_ui wrong
a=1198415671637283
d=1198415671637283
q=1
got=0
Aborted
$
Amusing that it's only dividing a number by itself that seems to give gmp
fits here. :-) If I special-case q=1 out, I get a full sequence of 5000
tests without errors. Other values that don't fail the test include
d=0x0056e18e03a9de91 and d=0x459f9e46f19b39d8; others that do include
d=0xa679fe5e2536755c and d=0xd2e85bb7873d8a7e. Looks suspiciously like
signed/unsigned breakage, as only values with the high bit set appear to be
affected.
Will continue investigating.
Buggy division function in glibc 2.3.5.
$ .libs/dive_ui
mpz_divexact_ui wrong
a=1198415671637283
d=1198415671637283
q=1
got=0
Aborted
$ LD_LIBRARY_PATH=./libc6/lib/ ./libc6/lib/ld-linux.so.2 .libs/dive_ui echo
success
success
$
Simple test case:
int main() {
unsigned long a= 1UL 63;
unsigned long b= 1UL 63;
exit (a/b != 1);
}
Appears to be a bug in the implementation of __divqu. Upgrading to 2.3.5-4,
to confirm whether the bug still exists.
--
Steve Langasek Give me a lever long enough and a Free OS
Debian Developer to set it on, and I can move the world.
[EMAIL PROTECTED] http://www.debian.org/
signature.asc
Description: Digital signature
Bug#324455: gmp: FTBFS on alpha
On Mon, Aug 22, 2005 at 08:56:23PM -0700, Steve Langasek wrote:
Buggy division function in glibc 2.3.5.
$ .libs/dive_ui
mpz_divexact_ui wrong
a=1198415671637283
d=1198415671637283
q=1
got=0
Aborted
$ LD_LIBRARY_PATH=./libc6/lib/ ./libc6/lib/ld-linux.so.2 .libs/dive_ui
echo success
success
$
Simple test case:
int main() {
unsigned long a= 1UL 63;
unsigned long b= 1UL 63;
exit (a/b != 1);
}
Appears to be a bug in the implementation of __divqu. Upgrading to 2.3.5-4,
to confirm whether the bug still exists.
Yep, bug still present in 2.3.5-4. Someone who speaks alpha assembly will
have to look at this, I think, to pinpoint the error.
--
Steve Langasek Give me a lever long enough and a Free OS
Debian Developer to set it on, and I can move the world.
[EMAIL PROTECTED] http://www.debian.org/
signature.asc
Description: Digital signature
