inconsistent errors with negative dimensions in Array constructor

[rfourquet]

This was reported on slack:

julia> Array{Int, 2}(undef, 0, -10)
0×0 Array{Int64,2}

julia> Array{Int, 2}(undef, -10, 0)
ERROR: invalid Array dimensions
Stacktrace:
 [1] Array{Int64,2}(::UndefInitializer, ::Int64, ::Int64) at ./boot.jl:396
 [2] top-level scope at none:0

This comes from the _new_array_ function in "array.c", which computes the product of the dimensions: if 0 comes first, then the product is 0 and doesn't grow out of bounds.

[StefanKarpinski]

Ah, good catch!

[andreasnoack]

There is some int/uint conversion stuff going on here. The array allocators's size argument is size_t, see e.g.

julia/src/array.c

Line 414 in 3fe2d08

JL_DLLEXPORT jl_array_t *jl_alloc_array_1d(jl_value_t *atype, size_t nr)

but we pass the value as an Int, see

julia/base/boot.jl

Lines 394 to 395 in 3fe2d08

	Array{T,1}(::UndefInitializer, m::Int) where {T} =
	ccall(:jl_alloc_array_1d, Array{T,1}, (Any, Int), Array{T,1}, m)

so e.g. -1 becomes a very large number which errors in

julia/src/array.c

Lines 63 to 68 in 3fe2d08

	for(i=0; i < ndims; i++) {
	wideint_t prod = (wideint_t)nel * (wideint_t)dims[i];
	if (prod > (wideint_t) MAXINTVAL)
	jl_error("invalid Array dimensions");
	nel = prod;
	}

unless the previous size was zero, in which case prod becomes zero.