Inconsistent errors with SparseArray construction (new in 1.2)

[sbromberger]

In 1.1:

julia> a = spzeros(Float64, UInt8, 255, 255)
255×255 SparseMatrixCSC{Float64,UInt8} with 0 stored entries

julia> a = spzeros(Float64, UInt8, 0xff, 0xff)
255×255 SparseMatrixCSC{Float64,UInt8} with 0 stored entries

But in 1.2, there apparently was a change to ensure that the number of nonzero entries could fit:

julia> a = spzeros(Float64, UInt8, 255, 255)
ERROR: ArgumentError: maximal nnz+1 (m*n+1 = 65026) does not fit in Ti = UInt8

However, the check is not applied consistently:

julia> a = spzeros(Float64, UInt8, 0xff, 0xff)
255×255 SparseMatrixCSC{Float64,UInt8} with 0 stored entries

[StefanKarpinski]

Note that this might be a 1.2 regression so we should take a look at it on triage.

[sbromberger]

(Just a guess that the error checks by multiplying the dimensions together to get the maximum possible number of nnzs, and 0xff * 0xff overflows to 0x01 which, when increased by 1, is still less than typemax(UInt8).)

[KristofferC]

This was reverted for 1.2 because it was breaking (#31667). The PR introducing this was then re-merged 2 weeks ago #31724 and has not been backported so I am pretty sure this only exists on master.

[StefanKarpinski]

This was reverted for 1.2

That's a relief. Sorry for the false alarm.

[JeffBezanson]

I believe this and #32539 were caused by the same set of changes.

[sbromberger]

Thank you very much for the quick analysis.

[JeffBezanson]

Fixed by #31724.

[StefanKarpinski]

Seems to be fixed.