Segfault in complex array comprehensions

[ahonkela]

The below code snippet segfaults Julia 0.4.x, both on Mac (0.4.2 from Homebrew) and on Linux (self-built release-0.4 branch):

N = 51

function f(repeats)
    return zeros(repeats, N)
end

results = [((x, y), f(100)) for x in 1:2, y in 1:2]

Backtrace and full version info are at:
https://gist.github.com/ahonkela/516a7e6b530795c4ade1

Antti

[pfitzseb]

Interestingly enough, declaring N as const N = 51 works fine.

[mauro3]

Crashes my 0.4.3 as well (although with different backtrace). Works fine on 0.3.12

[yuyichao]

Somehow the type of results is not a concrete type ....

(gdb) p jl_(jl_typeof(args[0]))
Array{Tuple{Tuple{Int64, Int64}, Array{#T<:Any, N<:Any}}, 2}

Probably related to #7258

[yuyichao]

This case seems to be fixed now due to improvement in type inference.

[tkelman]

add the test to close then?

[ahonkela]

Just to verify: works for me too now on current master.

I did not check current 0.4 series. If not, it would be great if the fix can be backported there too.

[yuyichao]

The issue is hidden due to unrelated changes (either type inference or definition of zeros, neither is likely backportable). The real bug is still not fixed (i.e. #16420).

[vtjnash]

rendered irrelevant by jb/comprehensions

[tkelman]

still may as well add that test?