Prevent Array(foo, n) from returning an array of the wrong size

[danluu]

julia> Array(Int, -14195763867060736356299439484696781493)
4611694853239872843-element Array{Int64,1}:
...

julia> x = int128(2)^65+10
36893488147419103242

julia> Array(Int, x)
10-element Array{Int64,1}:
...

The first statement only fails on my mac (10.9.2), but the second returns a mis-sized array on both linux and mac.

This could also be fixed by passing an Int128 to jl_alloc_array* and changing _new_array_. But, currently, both of my machines use uint64_t for the internal calculations because of this typedef:

#if defined(_P64) && defined(UINT128MAX)
typedef __uint128_t wideint_t;
#else
typedef uint64_t wideint_t;
#endif

Sorry for the all of the weird corner case issue spam; I spent a few minutes writing a really simplistic fuzzer. It's found a couple of bugs I could imagine actually running into, and a bunch of obscure issues like this one. IME, it's more effective to fix these really unlikely bugs than it is to try to target "good" bugs, and I've heard that mozilla had the same experience with jsfunfuzz.

[pao]

I spent a few minutes a really simplistic fuzzer

Awesome. I know I identified a few bugs early on by "fuzzing" the REPL via fatfingering inputs.

[sjkelly]

For people like myself that are working to use Julia in industry, issues (with fixes!) like this are really great to see. Thanks @danluu !

[rfourquet]

I think you want || instead of && in the checks?

[ivarne]

All errors worth making, is also worth fixing.

I could not help but noticing that your code is broken for abs(typemin(Int128)) > typemax(Int) 😄

The issue here is that ccall implicitly calls convert(T, x) for parameters where the type is specified as T, and it was forgotten that convert throws InexactError for BigInt, but not for Int128 (and Int64 on 32 bit platforms).

[danluu]

Argh! typemin(Int128) == -typemin(Int128) strikes again. Added a fix for that covers that case. Thanks @ivarne and @rfourquet .

[rfourquet]

I would personally prefer not add more repetive code to Julia and have a function like check_array_dims(d::Integer...).
Also, does someone know why the exact same versions of Array constructors for 1,2 or 3 dimensions are written for both Int and Integer? efficiency?

[JeffBezanson]

I think #5413 is a better way to fix this. That will handle all such cases.