PyCall master trips LLVM assertion on precompile

[staticfloat]

Using the Homebrew build of Julia, which has LLVM assertions enabled by default, we get:

julia> using PyCall
INFO: Precompiling module PyCall...
Assertion failed: (ctx->gensym_assigned.at(idx)), function emit_expr, file codegen.cpp, line 3155.

signal (6): Abort trap: 6
__pthread_kill at /usr/lib/system/libsystem_kernel.dylib (unknown line)
ERROR: Failed to precompile PyCall to /Users/sabae/.julia/lib/v0.4/PyCall.ji
 in error at /usr/local/Cellar/julia/0.4.5/lib/julia/sys.dylib
 in compilecache at loading.jl:400
 in require at /usr/local/Cellar/julia/0.4.5/lib/julia/sys.dylib

[stevengj]

Related to #254?

Maybe this is related to my use of global const foo = ... in the init, which I should really change to Refs or similar?

[zaburo-ch]

I'm finding this problem too.
I used homebrew build julia 0.4.5 on Mac OSX 10.11.2.

[zaburo-ch]

something wrong in the homebrew build of julia.
I tried to build myself, and successfully build PyCall.

git clone git://github.com/JuliaLang/julia.git
cd julia
git checkout release-0.4
make -j (num threads)

[stevengj]

@zaburo-ch, good to know; I wonder what is wrong with Homebrew?

[zaburo-ch]

@stevengj sorry... I couldn't find what is wrong.
julia 0.4.6-pre was installed by that way. So there could possibly be problem in 0.4.5

[staticfloat]

It's not something wrong with the Homebrew build, it's that Homebrew LLVM
has assertions enabled, whereas julia-built LLVM does not.

On Wed, May 18, 2016, 08:03 Kaizaburo Chubachi [email protected]
wrote:

@stevengj https://github.com/stevengj sorry... I couldn't find what is
wrong.
julia 0.4.6-pre was installed by that way. So there could possibly be
problem in 0.4.5

—
You are receiving this because you authored the thread.
Reply to this email directly or view it on GitHub
#269 (comment)

-E

[stevengj]

@staticfloat, thanks for the clarification. So there is a Julia problem that is being hidden by the lack of assertions? Or is there an incorrect assertion somewhere?

[staticfloat]

I think there are many julia problems hidden by the lack of assertions. Pkg.test("Nettle") also fails when assertions are enabled; although both Nettle and PyCall have global const values, so perhaps these problems are all symptoms of that root issue.

[oschulz]

I see this too, with a self-compiled Julia-0.4.5 on Ubuntu-14.04 (built with override LLVM_VER=3.7.1 for Cxx.jl support).

[gasagna]

Is there any temporary workaround for this issue?

[oschulz]

My current workaround is

Pkg.pin("PyCall", v"1.4.0")

;-)

[gasagna]

Thanks! This works for me. PyPlot was unusable after the update.

[oschulz]

Thanks! This works for me. PyPlot was unusable after the update.

Pinning is quickly getting out of hand, though - now I have to pin SymPy too, because the latest version needs a current PyCall, and the last version that works with PyCall 1.4.0 has a problem with Plots.

Any new on this issue, @stevengj ?

[stevengj]

I'm unable to reproduce the problem. I just tried the Homebrew Julia 0.4.5 (on OSX 10.11.4) with the latest PyCall master and the Conda Python (ENV["PYTHON"]=""; Pkg.build("PyCall")) and it works fine for me.

[stevengj]

I also tried it with /usr/bin/python and a global Anaconda Python install. In all cases, Pkg.test("PyCall") passed.

This is with brew install Caskroom/cask/julia ... is there another julia build I should be trying?

[staticfloat]

Ah, I had forgotten about the cask-based Julia build. That's not the homebrew build we're talking about, it doesn't have LLVM assertions enabled. We're talking about this one,

[stevengj]

Okay, I can reproduce the problem.

[stevengj]

Nothing to do with global const assignments in __init__ and precompilation, sincd #279 does not eliminate the problem.

[stevengj]

Seems to be a Julia bug. Here is a minimal module that exhibits the error:

__precompile__()

module PyCall

type PyObject
    o::Ptr{Void} # the actual PyObject*
end

PyNULL() = PyObject(C_NULL)

function jl_Function_call(self_::Ptr{Void}, args_::Ptr{Void}, kw_::Ptr{Void})
    args = PyNULL()
    try
      return C_NULL
    catch e
    finally
        args.o = C_NULL # don't decref
    end
    return C_NULL
end

precompile(jl_Function_call, (Ptr{Void},Ptr{Void},Ptr{Void}))

end # module PyCall

[oschulz]

Thanks for the quick analysis, @stevengj!

[oschulz]

As a workaround (until JuliaLang/julia/#16546 is fixed), it's sufficient to disable line 3155 in julia/src/codegen.cpp:

        // assert(ctx->gensym_assigned.at(idx));

Of course it just hides the underlying issue, but PyCall (and PyPlot, SymPy, etc. on top of it) work fine again so far.

[stevengj]

The easiest workaround seems like it would be to download the official Julia binaries for MacOS, rather than using the Homebrew build. If you want a command-line Julia in your path, you can just do

ln -s /Applications/Julia-0.4.5.app/Contents/Resources/julia/bin/julia /usr/local/bin/julia-0.4

[oschulz]

The easiest workaround seems like it would be to download the official Julia binaries for MacOS

@stevengj true, however, I have the problem on Linux, too. I need to enable assertions because Cxx.jl doesn't work if they're disabled.