Failure on 32-bit Fedora/RHEL nightlies

[nalimilan]

The Fedora/RHEL nightlies fail on 32-bit since 2014-11-21 (last good build happened on 2014-11-19 at 18:48 UTC):

/builddir/build/BUILD/julia/build/usr/bin/julia --check-bounds=yes -f ./runtests.jl arrayops
     �[1m*�[0m �[31marrayops�[0m
exception on 1: ERROR: InexactError()
while loading arrayops.jl, in expression starting on line 298
ERROR: InexactError()
while loading arrayops.jl, in expression starting on line 298
while loading /builddir/build/BUILD/julia/test/runtests.jl, in expression starting on line 42
make: *** [arrayops] Error 1

http://copr-be.cloud.fedoraproject.org/results/nalimilan/julia-nightlies/fedora-20-i386/julia-0.4.0-0.20141121.el7.centos/build.log

EDIT: the last successful build was not one, but two days before

[tkelman]

I can't reproduce this. It might be specific to the build flags you're using on Fedora, or maybe the system libraries you're using (LLVM maybe?).

[nalimilan]

This is with system LLVM 3.4 and 3.5. All libraries are the system's, gcc is 4.8. You can see the flags in the above link, but nothing fancy.

Ah, with LLVM 3.5 the stracktrace is much better:

/builddir/build/BUILD/julia/build/usr/bin/julia --check-bounds=yes -f ./runtests.jl arrayops
     �[1m*�[0m �[31marrayops�[0m
exception on 1: ERROR: InexactError()
 in rand! at ./random.jl:252
 in anonymous at ./no file:299
 in runtests at /builddir/build/BUILD/julia/test/testdefs.jl:6
 in anonymous at ./multi.jl:632
 in run_work_thunk at ./multi.jl:593
 in remotecall_fetch at ./multi.jl:666
 in remotecall_fetch at ./multi.jl:681
 in anonymous at ./task.jl:1450
while loading arrayops.jl, in expression starting on line 298
ERROR: InexactError()
 in rand! at ./random.jl:252
 in anonymous at ./no file:299
 in runtests at /builddir/build/BUILD/julia/test/testdefs.jl:6
 in anonymous at ./multi.jl:632
 in run_work_thunk at ./multi.jl:593
 in remotecall_fetch at ./multi.jl:666
 in remotecall_fetch at ./multi.jl:681
 in anonymous at ./task.jl:1450
while loading arrayops.jl, in expression starting on line 298
while loading /builddir/build/BUILD/julia/test/runtests.jl, in expression starting on line 42
Makefile:9: recipe for target 'arrayops' failed
make: *** [arrayops] Error 1

So looks like it's related to random numbers. @rfourquet?

[tkelman]

Ah yes, there have been a bunch of changes there recently. See the recently added RNG label https://github.com/JuliaLang/julia/issues?q=label%3ARNG+. I wonder whether any of those changes involve the way we build dsfmt? You package that library separately right?

[nalimilan]

You package that library separately right?

Right.

[nalimilan]

random.jl:252 is this:

 rand!(mt, A) # must not segfault even if Int(pointer(A)) % 16 != 0

julia/test/random.jl

Line 252 in 4045c52

rand!(mt, A) # must not segfault even if Int(pointer(A)) % 16 != 0

So this may well be related to #9037

[ViralBShah]

We only patch dsfmt to change a few static methods to extern so that they can be called.

Cc @rfourquet

[nalimilan]

We only patch dsfmt to change a few static methods to extern so that they can be called.

Ah, so this means I need to export them from Fedora's dSFMT too. Will have a look.

[rfourquet]

@nalimilan Could you tell for which value(s) of A (loop variable) the test fails? Also, could it be that the random tests failed before 2014-11-19, but are noticed only now because of 4125f4c ?

[nalimilan]

@rfourquet Actually my reading of the backtrace was wrong. ./random.jl points to base/random.jl, not tests/random.jl. The random test was already run and passes, both before and after commit 4125f4c, as I needed to use a hardcoded list of tests. The only failure is in the arrayops test.

So the real failing call is unsafe_copy!(pA, pA2, n2):

julia/base/random.jl

Line 252 in 4045c52

unsafe_copy!(pA, pA2, n2)

[rfourquet]

I also suspected unsafe_copy! for its line number, but I think it gets called only via line 252 of test/random.jl anyway... So I have no idea why InexactError() is triggered in this call. Could you run this minimal test:

 c8 = zeros(UInt8, 8001)
f1 = Ptr{Float64}(pointer(c8)) + 1
f8 = Ptr{Float64}(pointer(c8)) + 8
f16 = Ptr{Float64}(pointer(c8)) + 16
unsafe_copy!(f1, f16, 998)
unsafe_copy!(f8, f16, 998)

It reproduces the conditions of the random tests WRT unsafe_copy!.

[nalimilan]

@rfourquet The test in your comment passes. But I don't think the problem is related to tests/random.jl:252, since that test passes too.

[nalimilan]

Fixed by 8b63c51 (cf. #9040)

[rfourquet]

@nalimilan so this can be closed?

[nalimilan]

Right.