Make print_matrix O(1) again

[iamed2]

Thanks @sbromberger for reporting this on Slack

Previously, printing an array A at the REPL would allocate O(sum(size(A))) memory. This change was made to support OffsetArrays but OffsetArrays appears to work with this new version too.

[iamed2]

Key benchmark:

display(rand(1800000000))

[iamed2]

I believe this commit introduced the performance regression: 0d688dd

[StefanKarpinski]

So I guess the extra credit here would be either a test that takes a short time when this is O(1) but effectively hangs when this is O(sum(size(A)). Or a benchmark in BaseBenchmarks.

[iamed2]

I think the benchmark is the best choice but I've never done that before, so I'll have to read up on BaseBenchmarks.jl. I managed to find a call that calls print_matrix but sends everything to DevNull so I just need to add it.

[iamed2]

Well, I ran into a regression which I think is because of #22932

The vector and column matrix cases are faster and use less memory as expected. The square matrix case regresses. I think this is because in the matrix case, print_matrix calls (n - length(Ralign)) .+ colsA. My hunch is that previously Julia was smart about reusing the allocated array each time, as it was essentially just a copy of a Vector{Int} plus some Int.

Now that colsA is some linear indices, it makes sense to just shift the indices. Unfortunately + has been deprecated and .+ allocates an array. broadcast(+, (n - length(Ralign)), colsA) would have the intended behaviour but would be pretty ugly. Which do you think I should do?

[iamed2]

Now that regression is gone. Just noise I guess? I feel like that shouldn't be an issue with BenchmarkTools but oh well.

[sbromberger]

Any possibility of a backport?