PARI/GP Bug report logs - #1230
Grave error in bezout on MacOs X Lion

Display info messages

Message #5 received at submit@pari.math.u-bordeaux.fr (full text, mbox, reply):

Package: pari-stable
Version: 2.5.0
Severity: Grave 

bezout has an incorrect behavior which vanishes when gp is compiled in debug mode.
I isolated one example, but there are many others. The bezout function says that gcd(2,7)=2 !

This of course has consequence on all functions that use modular arithmetic. I first encountered the problem with "chinese" and elliptic curves.


Here is pari's output:

                  GP/PARI CALCULATOR Version 2.5.0 (released)
          i386 running darwin (x86-64/GMP-5.0.2 kernel) 64-bit version
compiled: Aug  4 2011, gcc-4.2.1 (Based on Apple Inc. build 5658) (LLVM build 2335.15.00)
                 (readline v6.2 enabled, extended help enabled)

                     Copyright (C) 2000-2011 The PARI Group

PARI/GP is free software, covered by the GNU General Public License, and comes 
WITHOUT ANY WARRANTY WHATSOEVER.

Type ? for help, \q to quit.
Type ?12 for how to get moral (and possibly technical) support.

parisize = 8000000, primelimit = 500509
? bezout(7,2)
%1 = [0, 1, 2]

INCORRECT
-----------------------------
The same in debug mode:

                  GP/PARI CALCULATOR Version 2.5.0 (released)
   i386 running darwin (x86-64/GMP-5.0.2 kernel) 64-bit version -- debugging
compiled: Aug  4 2011, gcc-4.2.1 (Based on Apple Inc. build 5658) (LLVM build 2335.15.00)
                 (readline v6.2 enabled, extended help enabled)

                     Copyright (C) 2000-2011 The PARI Group

PARI/GP is free software, covered by the GNU General Public License, and comes 
WITHOUT ANY WARRANTY WHATSOEVER.

Type ? for help, \q to quit.
Type ?12 for how to get moral (and possibly technical) support.

parisize = 8000000, primelimit = 500509
? bezout(7,2)
%1 = [1, -3, 1]

CORRECT

--------------------------------

Message #10 received at 1230@pari.math.u-bordeaux.fr (full text, mbox, reply):

On Thu, Aug 04, 2011 at 08:38:38PM +0200, Antoine Joux wrote:
> Package: pari-stable
> Version: 2.5.0
> Severity: Grave 
> 
> bezout has an incorrect behavior which vanishes when gp is compiled in debug mode.
> I isolated one example, but there are many others. The bezout function says that gcd(2,7)=2 !

Hello Antoine,

Does 'make bench' even succeed ?

All the evidence point to a compiler problem. Indeed the compiler version you mention
is not able to compile PARI without GMP support. 

You can try to build with
make clean;
make "CFLAGS=-O0 -Wall -fno-strict-aliasing -fomit-frame-pointer" 'KERNELCFLAGS=' gp

to see if this fix the compiler issue.

Cheers,
Bill.

Message #15 received at 1230@pari.math.u-bordeaux.fr (full text, mbox, reply):

On Thu, Aug 04, 2011 at 11:27:51PM +0200, Antoine Joux wrote:
> Hi Bill,
> 
> No, "make bench" does not succeed !
> 
> 
> Your compile command fixes the problem (and the debug mode also does it). So yes it is probably an optimizer problem.
> What do you do in this case ? Report to the compiler maintainer or just ignore ?

I try to get the attention of the compier maintainers, but it is not always easy.

> By the way, is there a configure option to use a different compiler (already installed in my system) ?

Yes, do
CC=/path/to/compiler ./Configure

Cheers,
Bill.

Message #26 received at 1230@pari.math.u-bordeaux.fr (full text, mbox, reply):

On Thu, Aug 04, 2011 at 11:17:55PM +0200, Bill Allombert wrote:
> On Thu, Aug 04, 2011 at 08:38:38PM +0200, Antoine Joux wrote:
> > Package: pari-stable
> > Version: 2.5.0
> > Severity: Grave 
> > 
> > bezout has an incorrect behavior which vanishes when gp is compiled in debug mode.
> > I isolated one example, but there are many others. The bezout function says that gcd(2,7)=2 !
> 
> Hello Antoine,
> 
> Does 'make bench' even succeed ?
> 
> All the evidence point to a compiler problem. Indeed the compiler version you mention
> is not able to compile PARI without GMP support. 
> 
> You can try to build with
> make clean;
> make "CFLAGS=-O0 -Wall -fno-strict-aliasing -fomit-frame-pointer" 'KERNELCFLAGS=' gp
> 
> to see if this fix the compiler issue.

I finally found a fix (at least for snow leopard), see
<http://pari.math.u-bordeaux.fr/archives/pari-dev-1312/msg00021.html>

If one change the definition of divll in src/kernel/x86_64/asm0.h

#define divll(a,b) \
__extension__ ({ ulong __value, __arg1 = (a), __arg2 = (b); \
   __asm__ ("divq %4" \
        : "=a" /* %eax */ (__value), "=&d" /* %edx */ (hiremainder) \
                                     ^^^^^
        : "0" /* %eax */ (__arg1), "1" /* %edx */ (hiremainder), "mr" (__arg2)); \
   __value; \
})

by replacing "=&d" by "=d", then GP works fine (make test-all succeed).

Cheers,
Bill.

Message #31 received at 1230-done@pari.math.u-bordeaux.fr (full text, mbox, reply):

On Wed, Dec 18, 2013 at 02:13:58PM +0100, Bill Allombert wrote:
> On Thu, Aug 04, 2011 at 11:17:55PM +0200, Bill Allombert wrote:
> > On Thu, Aug 04, 2011 at 08:38:38PM +0200, Antoine Joux wrote:
> > > Package: pari-stable
> > > Version: 2.5.0
> > > Severity: Grave 
> > > 
> > > bezout has an incorrect behavior which vanishes when gp is compiled in debug mode.
> > > I isolated one example, but there are many others. The bezout function says that gcd(2,7)=2 !
> > 
> > Hello Antoine,
> > 
> > Does 'make bench' even succeed ?
> > 
> > All the evidence point to a compiler problem. Indeed the compiler version you mention
> > is not able to compile PARI without GMP support. 
> > 
> > You can try to build with
> > make clean;
> > make "CFLAGS=-O0 -Wall -fno-strict-aliasing -fomit-frame-pointer" 'KERNELCFLAGS=' gp
> > 
> > to see if this fix the compiler issue.
> 
> I finally found a fix (at least for snow leopard), see
> <http://pari.math.u-bordeaux.fr/archives/pari-dev-1312/msg00021.html>
> 
> If one change the definition of divll in src/kernel/x86_64/asm0.h
> 
> #define divll(a,b) \
> __extension__ ({ ulong __value, __arg1 = (a), __arg2 = (b); \
>    __asm__ ("divq %4" \
>         : "=a" /* %eax */ (__value), "=&d" /* %edx */ (hiremainder) \
>                                      ^^^^^
>         : "0" /* %eax */ (__arg1), "1" /* %edx */ (hiremainder), "mr" (__arg2)); \
>    __value; \
> })
> 
> by replacing "=&d" by "=d", then GP works fine (make test-all succeed).

PARI 2.7.0 was released with this fix, so I close this report.

Cheers,
Bill.

Send a report that this bug log contains spam.