mysterious assertion failure when reinterpreting subarrays

[staticfloat]

In my naive attempts to chase the performance rabbit, I came across this interesting case:

function breakme!(data)
    data = reinterpret(Uint32, data)
    a = data[1]
end

a = ones(Uint8, 10)
sa = sub(a,4:6)
breakme!(sa)

This causes the following assertion failure:

$ julia breakme.jl 
Assertion failed: (jb->instance != NULL), function static_void_instance, file ./cgutils.cpp, line 1436.

signal (6): Abort trap: 6
__pthread_kill at /usr/lib/system/libsystem_kernel.dylib (unknown line)
Abort trap: 6

Surprisingly, if I comment out the a = data[1] line, I get something much more sane:

$ julia breakme.jl 
ERROR: auto_unbox: unable to determine argument type
 in breakme! at /Users/sabae/breakme.jl:2
 in include at /usr/local/Cellar/julia/HEAD/lib/julia/sys.dylib
 in include_from_node1 at loading.jl:128
 in process_options at /usr/local/Cellar/julia/HEAD/lib/julia/sys.dylib
 in _start at /usr/local/Cellar/julia/HEAD/lib/julia/sys.dylib (repeats 2 times)
while loading /Users/sabae/breakme.jl, in expression starting on line 8

And it's not the fact that it's trying to read in from data that's confusing it; it's the fact that there is an array read at all; it could be a completely different array, and Julia still throws a hissy fit:

function breakme!(data, arr)
    data = reinterpret(Uint32, data)
    a = arr[1]
end

a = ones(Uint8, 10)
sa = sub(a,4:6)
breakme!(sa, zeros(10))

[Keno]

I fixed the assertion failure (not pushed yet though), but I'm wondering what reinterpret on a subarray should do. Opinions @JeffBezanson @StefanKarpinski?

[staticfloat]

I think it's reasonable to just plain disallow that for now (it would be really really nice if we could reinterpret subarrays or array views, or whatever you want to call that though). I opened this issue because of the strange failure though.

[Keno]

Alright, I'll fix the assertion failure and we can discuss subarray reinterpret later.

[sebastiang]

What is the status of SubArray reinterpretation? If regular array views (x[a:b]) start being treated this way, it will be a regular occurrence.

[timholy]

It seems fairly likely that this will be doable, but I haven't really thought about it seriously yet.