Re: MacOs X Lion compiler and PARI

[Bill Allombert <Bill.Allombert@math.u-bordeaux1.fr>]

On Mon, Aug 15, 2011 at 12:59:14AM +0200, Bill Allombert wrote:
> On Sat, Aug 06, 2011 at 09:15:01AM -0700, William Stein wrote:
> > On Fri, Aug 5, 2011 at 5:30 AM, Bill Allombert
> > <Bill.Allombert@math.u-bordeaux1.fr> wrote:
> > > Hello PARI developers,
> > >
> > > I received reports that MacOs X Lion default compiler is unable to compile PARI
> > > correctly. Precisely this version:
> > >
> > > GNU compiler version 4.2.1 (Based on Apple Inc. build 5658) (LLVM build 2335.15.00)
> > >
> > > So far I got two kind of issues:
> > > With --with-gmp, PARI build, but the resulting code is broken (see bug #1222)
> > > With --without-gmp, PARI fails to build (see
> > > <http://pari.math.u-bordeaux.fr/buildlogs/dac98dec-af1e-4e54-97ba-f001b69bf7a9>)
> > >
> > > Concerning, the first issue, I do not have access to such a system and in any
> > > case I cannot fix Apple compiler, so I would appreciate if one of you Mac OS X
> > > users would report the problem to Apple.
> 
> > To encourage porting efforts to Lion, I setup a computer with the
> > latest OS and XCode 4.1 on it.
> 
> Thanks a lot, this has allowed to confirm that the problem 
> is a compiler issue that cause the file 'src/kernel/none/gcdll.c'
> (included in mp.c) to be miscompiled with -O1 and up.
> The symptom is that
> bezout(7,2) return [0, 1, 2] instead of [1, -3, 1].

So I finally get access to a another working OS X system (snow leopard) and I
have got to the root of the problem (I hope).

"Apple clang version 1.7 (tags/Apple/clang-77) (based on LLVM 2.9svn)"
miscompiles the call to divll in src/kernel/none/divll.c.

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).

However, both variants are correct, so this is a compiler bug.

Cheers,
Bill.