Multi-threading doesn't work on master

[giordano]

I reported on discourse a problem, but didn't get any feedback, so I don't know if anyone else can reproduce it.

Currently, multi-threading doesn't seem to work on master branch (at least on my system). Using the example proposed in the manual:

julia> versioninfo()
Julia Version 0.6.0-dev.1928
Commit cbc6670 (2017-01-05 11:46 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: Intel(R) Core(TM) i7-4700MQ CPU @ 2.40GHz
  WORD_SIZE: 64
  BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Haswell)
  LAPACK: libopenblas64_
  LIBM: libopenlibm
  LLVM: libLLVM-3.9.1 (ORCJIT, haswell)

julia> Threads.nthreads()
4

julia> a = zeros(10);

julia> Threads.@threads for i = 1:10
           a[i] = Threads.threadid()
       end

julia> a
10-element Array{Float64,1}:
 1.0
 1.0
 1.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0

As show in the discourse message, the same example works on my system with Julia 0.5.

I could not find any reference to this problem in the bug tracker, I assume it's a new issue.

[tkelman]

Would you (or anyone else who can reproduce) be able to bisect?

[giordano]

I'd like to do that, but compiling Julia from scratch takes about half an hour on my box with make -j4, it may take a while (probably some days) to find the offending revision. If someone wants to undertake this task, please go ahead.

@tkelman Out of curiosity: can you reproduce it, is it just me?

[KristofferC]

I can reproduce. I'll try bisect.

[kpamnany]

From debug prints, I find that the threads are running, getting forked, and attempting to apply the generated function containing the loop body. However, every thread except the first is getting an exception on calling jl_apply() in ti_run_fun(). Not sure why. @yuyichao, any ideas?

[vtjnash]

They're all getting no-method errors from trying to run in world 0

[tkelman]

Oh, then it's #17057 / #265. How did the threads test not catch this?

[kpamnany]

$ ./julia test/threads.jl
ERROR: LoadError: BoundsError: attempt to access 10000-element Array{Bool,1} at index [0]
Stacktrace:
 [1] test_threaded_loop_and_atomic_add() at /scratch/kpamnany/julia.intel/test/threads.jl:23
 [2] include_from_node1(::String) at ./loading.jl:532
 [3] include(::String) at ./sysimg.jl:14
 [4] process_options(::Base.JLOptions) at ./client.jl:298
 [5] _start() at ./client.jl:364
while loading /scratch/kpamnany/julia.intel/test/threads.jl, in expression starting on line 34

It should have? Is this test run with JULIA_NUM_THREADS set to >1?

[tkelman]

Ah right, we still have JULIA_NUM_THREADS defaulting to 1. We have a threading buildbot but nobody was paying attention (apologies to @KristofferC that I forgot about it): https://build.julialang.org/builders/nightly_threading-x64/builds/579

[KristofferC]

I gave up on bisecting after LLVM wanted to rebuild for the fourth time ;)