gamma_fmpq test fails on 32-bits with arb 2.21.0

[SnarkBoojum]

It has been reported here that there's a problem on 32-bits architectures ; indeed the build logs show quite a few problems.

[fredrik-johansson]

I didn't catch this because the 32-bit tests on appveyor all passed.

Is there some way to get access to a failing machine?

[fredrik-johansson]

I can force a harmless test failure out of t-gamma_fmpq by setting ARB_TEST_MULTIPLIER=100, but this does not seem to be the problem because the build logs above show an abort() without a test failure message.

[fredrik-johansson]

I have fixed the harmless bug; can you check if this makes tests pass on some of the machines?

[tornaria]

$ make BUILD_DIRS=arb TESTS=build/arb/test/t-gamma_fmpq check
[...]
gamma_fmpq....FAIL: poor accuracy

prec = 49
q = -77/19

r = -0.7344533477145 +/- 5.8402e-14

make: *** [Makefile:166: check] Error 134

EDIT: with 450435a84a40acc1871cd84f1caf2ea47c11a18c it passes.

[SnarkBoojum]

It's possible to get access to Debian's machines for such work (see here), but it's only temporary and takes some time to get.

I'm compiling 450435a on an armhf box (abel.debian.org), so I should hopefully confirm it helps.

@tornaria on which type of box did you give it a try?

[tornaria]

@tornaria on which type of box did you give it a try?

Void linux, i686-glibc. I also tested x86_64-glibc and x86_64-musl (see https://github.com/void-linux/void-packages/pull/33170).

[SnarkBoojum]

Everything passed on the armhf box!

[fredrik-johansson]

Great. I will wait another day or two just in case any other bug reports come in and then I'll issue 2.21.1.

[SnarkBoojum]

Can you issue 2.21.1 now or are there problems?

[fredrik-johansson]

Done. Apologies for the delay!