NaN leak with selectdim view

[michaelsachs]

I came across an example where NaNs "leak" into the result of a nanmean operation. This occurs when I average on a view created using selectdim.

# create a matrix and set one element to NaN
a = rand(10,5)
a[3,1] = NaN

When I create a view using the appropriate macro everything looks fine:

av = @view a[:,1:3]
avMean = nanmean(av,dims=2)
avMean[3,1] 
0.8755526000081276

However, when I create the same view using selectdim the NaN ends up in the final vector:

av = selectdim(a,2,[true,true,true,false,false])
avMean = nanmean(av,dims=2)
avMean[3,1]
NaN

I am using Julia 1.9.0 and NaNStatistics v0.6.28.

[brenhinkeller]

Thanks for the report! I think the immediate solution will be to just be stricter about what non-Array array types we use @turbo on

[brenhinkeller]

Should be fixed in v0.6.29 and onwards