Fix breakage on 0.5-dev.

[yuyichao]

I'm not sure how much of these can go to Compat.jl so I just implement everything here....

c.c. @timholy

Ref JuliaLang/julia#13803

[timholy]

I'm not sure those are material for Compat anyway since they were unexported functions. But if others feel differently, I'm fine with wherever they end up.

[yuyichao]

Oh SegFault, great .....

Also note that if we backport JuliaLang/julia#13803, the version check in this PR will need to be tweaked....

[yuyichao]

Hmm. I couldn't reproduce the segfault but I'm getting some very strange errors in reducedim.jl test.

Depending on what I do before, I get either

ERROR: LoadError: UndefVarError: N not defined
 in _mapreducedim! at ./no file:4294967295

or

ERROR: LoadError: MethodError: `_nextract` has no method matching _nextract(::Type{Any}, ::Symbol, ::Expr)
Closest candidates are:
  _nextract(::Int64, ::Symbol, ::Expr)
  _nextract(::Int64, ::Symbol, ::Symbol)
 in func_for_method_checked at ./reflection.jl:248

when I try to call @code_warntype Base._mapreducedim!(+, Base.AndFun(), Bool[], BitArray(10))

This only happens if the module is precompiled but not if I add --compilecache=no or define the corresponding method at runtime again.

Can anyone reproduce this?

@vtjnash The method that is causing this issue directly is here. Is there anything in it that looks suspicious for you in terms of precompilation compatibility? Also, is there a way I can check the .ji file and see if there's anything strange in there?

[yuyichao]

The first error UndefVarError(:N) is what I get when running the test locally without --compilecache=no.

[yuyichao]

Force pushed to trigger a travis rerun.

[yuyichao]

.... this time the error is the same one I see locally .....

https://travis-ci.org/JuliaStats/DataArrays.jl/jobs/88360029

[yuyichao]

The precompilation error is caused by JuliaLang/julia#13677 ....

[tkelman]

can you use an isdefined check here rather than a version comparison?

[yuyichao]

Good point, will do when I'm back.

[yuyichao]

Updated the check to use isdefined so that it should work on 0.4 if we end up backporting JuliaLang/julia#13803 .

The nightly CI will fail until commit JuliaLang/julia@5498570 is included in the nightly binary (hopefully tomorrow...)

[tkelman]

should be up already: http://buildbot.e.ip.saba.us:8010/builders/package_tarball64/builds/144, hit restart if the versioninfo() looks too old

[yuyichao]

You are right, the travis run fetched exactly the commit that has the fix. This should be ready to go then.

[tkelman]

@simonster et al, okay to merge this?

[tkelman]

Note that this needs to be tagged before JuliaLang/julia#13803 can be considered for backporting. Not urgent as it's almost certainly going to be for 0.4.2, not 0.4.1.