faster indexing with UInt for Arrays

[KristofferC]

Potential idea for fixing #17103 (at least for UInt).

Not sure if it is worth it though.

@mbauman, thoughts?

[mbauman]

So in the error case this would trade an InexactError with the value provided for a bounds error with a strange negative number. Have you measured how big of a performance improvement we get in exchange?

[KristofferC]

Yes, that is the trade-off.

Regarding performance

function mysum_uint(data)
    sum = zero(eltype(data))
    for i in UInt(1):UInt(lastindex(data))
        @inbounds sum += data[i]
    end
    return sum
end

using BenchmarkTools

siz = 128 * 2^20;
data = zeros(UInt8, siz);
@btime mysum_uint(data);

gives on master

julia> @btime mysum_uint(data);
  59.974 ms (0 allocations: 0 bytes)

but with the PR

julia> @btime mysum_uint(data)
 7.860 ms (0 allocations: 0 bytes)

it is, of course, the extreme case.

[mbauman]

Seems worthwhile.