Segfault when adding new zeros method

[martinholters]

Discovered this in #16740.

The segfault happens on Travis' 64bit linux, no other CI, not my own 64bit linux, but I could nail it down using a branch in my repo, and it is actually sufficient to add a definition for zeros{T<:AbstractArray}(::Type{T}, dims...) (which is not even called).

The test triggering the segfault is arrayops.jl:177, and it occurs in sig_match_simple at typemap.c:127, see the travis output.

[martinholters]

Setting up a (virtual) machine with Ubuntu 12.04 and following the build steps from the travis configuration, I can now reproduce the problem. That allowed me to narrow it down to the following code resulting in a segfault:

import Base.zeros
zeros{T<:AbstractArray}(::Type{T}, dims...) = error("What, me? I'm not even being called!")

rng = (2,2:3,2:2:5)
tmp = zeros(Int,2,3,5)
tmp[rng...] = 1 # needs the splatting
B = cat(3,zeros(Int,2,3),zeros(Int,2,3),zeros(Int,2,3)) # needs three matrices cat'ed

I will try to bisect (assuming there is a version where this does not segfault) to figure out what might be going on...

[martinholters]

Ah, good news: Segfault is gone in current master, specifically since e03f01c. However, considering how much the appearance of this bug depended on the build environment, I wonder whether the underlying cause has indeed been fixed or it is just better hidden now. CC @vtjnash
Should we close this, or do some more investigation?

[vtjnash]

That commit seems a likely candidate for the fix. Since that commit was on master before this issue was found on travis, that may also suggest why it was hard to reproduce elsewhere?

[martinholters]

Since that commit was on master before this issue was found on travis, that may also suggest why it was hard to reproduce elsewhere?

No, the problem first occurred in #16740 before that commit (including the hard reproducibility) and I started my investigations from there. Anyway, I'm glad it's fixed! (And having it nailed down to that specific commit gives the confidence that it is fixed indeed.)