[PPL-devel] Regression test failures on OpenBSD

[Roberto Bagnara]

On 07/16/12 17:48, Abramo Bagnara wrote:
> Il 14/07/2012 20:40, Edd Barrett ha scritto:
>> On Sat, Jul 14, 2012 at 11:25:53AM +0200, Abramo Bagnara wrote:
>>> Il 04/07/2012 12:27, Edd Barrett ha scritto:
>>>> On Tue, Jul 03, 2012 at 03:23:33PM +0200, Roberto Bagnara wrote:
>>>>> thanks to you for the report.  Could you please send to
>>>>> ppl-devel at cs.unipr.it the config.log file generated at
>>>>> configure time, along with the output of `make check' in
>>>>> `/usr/ports/pobj/ppl-1.0/ppl-1.0/tests/Concrete_Expression'
>>>>> once you have added `PPL_NOISY_TESTS=yes' to the environment?
>>>>
>>>> See attached the gzipped config.log and regression test failures in
>>>> verbose mode.
>>>
>>> We have changed the test to produce more output to help to diagnose what
>>> you're observing.
>>>
>>> Can you retry the above with current git head?
>>
>> I applied the patch you just comitted and re-ran the tests with noisy output.
>
> I've just tried with a VirtualBox image where I've installed a pristine
> OpenBSD 5.1 amd64, gmp-5.0.5 and I'm unable to reproduce what you're
> observing: both polyhedron1 and polyhedron2 succeeds.
>
> Might you have something installed that cause the miscompilation or a
> bogus library?

It is indeed a bogus version of libstdc++ to cause the problem.
Here is what happens (experiments repeated with identical results
with OpenBSD 5.1 both on x86_64 and sparc64):

-bash-4.2$ cat bug.cc
#include <cmath>
#include <cstdlib>
#include <cstdio>

int main(int argc, char **argv) {
   long double x = strtold(argv[1], 0);
   long double y = std::floor(x);
   printf("%.1000Lg\n%.1000Lg\n", x, y);
}
-bash-4.2$ g++ bug.cc
-bash-4.2$ ./a.out  13311002825915415087
13311002825915415087
13311002825915414528
-bash-4.2$

As you can see, std::floor(x) returns a wrong result.
An equivalent C program does the right thing:

-bash-4.2$ cat bug.c
#include <math.h>
#include <stdlib.h>
#include <stdio.h>

int main(int argc, char **argv) {
   long double x = strtold(argv[1], 0);
   long double y = floorl(x);
   printf("%.1000Lg\n%.1000Lg\n", x, y);
}
-bash-4.2$ gcc bug.c -lm
-bash-4.2$ ./a.out  13311002825915415087
13311002825915415087
13311002825915415087
-bash-4.2$

But not if it is linked with -lstdc++ instead of -lm:

-bash-4.2$ gcc bug.c -lstdc++
-bash-4.2$ ./a.out  13311002825915415087
13311002825915415087
13311002825915414528
-bash-4.2$

The reason why libstdc++ on OpenBSD exports a
bogus version of floorl() probably lies with the
following piece of code:

#ifndef HAVE_FLOORL
long double
floorl(long double x)
{
   return floor((double) x);
}
#endif

We have no idea why on OpenBSD libstdc++ is compiled
with HAVE_FLOORL not defined.
Kind regards,

     Roberto

-- 
Roberto Bagnara

BUGSENG srl
http://bugseng.com
mailto:roberto.bagnara at bugseng.com