Draft on 32-bit CI

[albinahlback]

No description provided.

[fredrik-johansson]

Hope this works!

I think it would be prudent to enable asserts for the 32-bit tests.

[albinahlback]

Hope this works!

I think it would be prudent to enable asserts for the 32-bit tests.

Sounds like a good idea!

[albinahlback]

Doesn't seem to pass the tests...

[edgarcosta]

@tornaria do tests pass on your 32bit machine?

[albinahlback]

Looks like fmpz_set_str is the problem.

[albinahlback]

Could it be that DIGITS_PER_LIMB should be 10 for 32-bit fmpz_set_str, @fredrik-johansson?

[tornaria]

@tornaria do tests pass on your 32bit machine?

Everything passes here:

https://github.com/void-linux/void-packages/actions/runs/6700948977/job/18207675113?pr=46832

I'm using released 3.0 + #1545 + #1546.

[tornaria]

Also, btw, the job I posted builds and checks eclib, singular, and sagemath with the new flintlib.

[fredrik-johansson]

Could it be that DIGITS_PER_LIMB should be 10 for 32-bit fmpz_set_str, @fredrik-johansson?

It should be 9, but

    tmp = TMP_ALLOC(sizeof(mp_limb_t) * (slen / DIGITS_PER_LIMB) + 2);

looks like it should be

    tmp = TMP_ALLOC(sizeof(mp_limb_t) * (slen / DIGITS_PER_LIMB + 2));

[tornaria]

I'm now testing 4b27c91 on i686 (glibc). However, if this is an issue with allocation maybe it's triggered in musl libc and not in glibc.

[fredrik-johansson]

It's possible that the bug only manifests itself with --enable-assert which does change the behavior of TMP_ALLOC.

[tornaria]

Could also be a bug in gcc since you are getting 13.2 and I'm running 12.2. Is it feasible to test with an earlier version of alpine?

[tornaria]

Build and check of 4b27c91 finished without any failure.

[fredrik-johansson]

Trying this on a 32-bit debian in VirtualBox. The fmpz_is_prime test passes, but valgrind does show a bunch of errors. They go away when fixing the TMP_ALLOC line mentioned above.

[tornaria]

It's possible that the bug only manifests itself with --enable-assert which does change the behavior of TMP_ALLOC.

I redid the build + check adding --enable-assert to configure, and the test works fine.

[tornaria]

Trying this on a 32-bit debian in VirtualBox. The fmpz_is_prime test passes, but valgrind does show a bunch of errors. They go away when fixing the TMP_ALLOC line mentioned above.

IME, musl allocator is less forgiving to allocation mistakes, so it's good that you test on alpine.

I do test musl, but only x86_64 since void doesn't do i686-musl binaries (I'd have to rebuild the whole system).

While running the whole testsuite on --enable-assert I got:

ecm....t-ecm: src/ulong_extras/mulmod_preinv.c:22: n_mulmod_preinv: Assertion `a < n' failed.

[fredrik-johansson]

The allocation bug is fixed in 6c92f36.

[albinahlback]

I will increase the test multiplier once it passes all the tests so that it matches in time with Ubuntu GCC AVX2 x10, so that we may detect more bugs.

[albinahlback]

Test fails for fmpz_factor_ecm

[fredrik-johansson]

The ECM bug may be fixed in 016de13. Tests now pass for me on 32-bit.

Note to self: the ECM code needs a thorough review.

[fredrik-johansson]

Nice!

[tornaria]

FWIW, I built and checked 8528d83 on i686 with --enable-assert. Everything passes now.