recint seems to abuse 32 bits asm instructions

[ClementPernet]

The problem arose here:
https://trac.sagemath.org/ticket/26932#comment:56

I could reproduce by building sage on this branch on the i386 machine on ci.inria.
From what I understand, recint hard codes limb to be int64_t. However in recint/rudiv.h line 222, it calls umul_ppmm with 4 arguments of limb type. And accordind to reclonglong.h line 866, the implementation of umul_ppmm on intel 32 bits (i386) is an asm call to mull which expects 32 bits ints.
I don't know why this bug only shows up in the sage build and not in givaro's test-suite on this machine.

[ClementPernet]

Note that the compilation failure is reproducible by typing make in /builds/sage on the 32 bit ubuntu ci.inria slave.

[Breush]

@ClementPernet Interesting, I'm going to check that.

[Breush]

@ClementPernet We used #define NO_ASM and #define W_TYPE_SIZE 64 in recdefine.h.
So we probably shouldn't go to the umul_ppmm line 866 in any case.

Looking for it, I think that, for me, the umul_ppmm is the plain C one defined line 1971.

On the 32 bit machine, however, I don't really know.

[ClementPernet]

For the record, I have identified the bug: this is because in sage, givaro's reclonglong.h yields for flint's longlong.h which does not include the NO_ASM section.
See: https://trac.sagemath.org/ticket/26932#comment:64
Workaround in progress...

[ClementPernet]

The workaround used in https://trac.sagemath.org/ticket/26932 consists in avoiding name conflicts between givaro's reclonglong.h and flint's longlong.h. I ended up adding a recint_ prefix to all macros defined in reclonglong.h. For the moment, this is just a patch that sage applies to givaro. If nobody objects, I'll apply this patch upstream.

[tobihan]

We also hit this issue in Debian now. From our side it would be welcome if you apply the patch.