Did any benchmarks change?

There are no benchmarks for this yet. However, I will modify the one I created last time so we get some numbers on that, too as soon as I have a free moment.

Might be nice to see a small change in benchmarks. But if it doesn't change much at all, I don't mind. It's just better to use the correct logic either way, IMO.
That said, labs is incorrect on windows since long is 32bits, may need to actually use if/else (I guess with a macro helper). Not sure we already have a macro, it doesn't really look like it.

Figured out some old fashioned ABS macro will do the trick, can't be more portable 😅

Regarding benchmarks: the matmul should take longer than the copying, but I assume one could create some particularly weird slice to make things measureable. What do you think?

I'll just put this in. I think it is unclear that this helps as discussed. But if it doesn't help, chances are we should just never bother doing a transposed copy (because the matmul is much slower anyway).
But, I think it's weir dto use the right transposed copy only for positive strides :).