julia: 0.3.9 -> 0.3.10

[vbgl]

Tests on i686 are also disabled since some of them fail and these failures do not reveal packaging bugs (rather bugs in the test-suite or in julia itself). AFAICT julia is properly built and behaves well on NixOS-i686.

[vcunat]

/cc maintainers: @7c6f434c, @ttuegel.

[ttuegel]

I think a fix for the failing test has already been committed upstream. I'm trying to backport it now...

[vbgl]

I tried that too. Patching numbers.jl is easy. Then I hit next failing test.

IIRC, it’s ishermitian(A'A) in arrayops.jl. I’m not sure it has been patched upstream (it seems unclear what this function is supposed to do wrt numerical inaccuracy).

If this test is dropped, then the next failing one is in reduce.jl, also patched upstream.

After this one is fixed, next one is in random.jl.

There I realize all of this was pointless.

[ttuegel]

I see. Mentioning that there were several test failures all related to numerical accuracy would have saved some time. (As a general rule, there is no point in writing tests that depend on numerical accuracy, as they will always fail on some OS, on some platform, somewhere, unless the tolerance is pointlessly large. I had hoped the Julia folks would be acquainted with this fact.)

[vbgl]

Ok. Sorry for wasting your time.

[ttuegel]

@vbgl Oh, it's no trouble, don't worry about it. I'm mostly irritated that upstream would do something like test numerical equality because (as someone who does numerical computing for a living) I know how useless it is! :)

[vcunat]

Hmm, I thought IEEE-754 defines the exact results of operations, but that's rather off-topic...

[ttuegel]

@vcunat Yes, but bit-perfect results still depend on knowing details about the hardware implementation and choice of C library. In theory, it should work predictably, in practice it's a crapshoot.