Specialize Interleave[Shortest]::fold

[Philippe-Cholet]

Related to #755

cargo bench --bench specializations "interleave(_shortest)?/fold"

interleave/fold            [2.1749 µs 2.1782 µs 2.1817 µs]
                           [754.64 ns 757.16 ns 760.51 ns]
                           [-66.092% -65.545% -65.107%]

interleave_shortest/fold   [1.6414 µs 1.6460 µs 1.6511 µs]
                           [412.05 ns 413.60 ns 415.41 ns]
                           [-74.904% -74.580% -74.173%]

[phimuemue]

I see we must check if ‘b.next‘ returns something. But if it returns None, couldn’t we exploit this by using try_fold and early-out, leaving the rest to a simple a.fold?

I could imagine that this might be better when b has very few elements compared to a, and I would hope that the compiler is able to generate the early-out without additional checks (it needs to check for None, anyway).

[Philippe-Cholet]

This makes me think of one the first specializations done in this large streak: https://github.com/rust-itertools/itertools/pull/774/files#diff-5423381b620ccb5a8ab55224d730c8729d676199d27701ac61043cd6773e4617R63-R66
There are probably some others that could possibly benefit from this idea (anytime there are two iterators I guess).

I'm not sure benchmarks would show improvements but it's definitely worth considering.

[Philippe-Cholet]

Benchmark is now -65% instead of -75% initially but I pushed it as it seems to be a better decision for the general case.

[phimuemue]

As always, LGTM.

Note: We should clean up these adaptors.

flag/phase should probably have the same name (maybe next_is_from_it1 or similar), and we may want to fuse the iterators also for the shortest-variant, or am I missing something?

[Philippe-Cholet]

They could surely be a bit more similar (in another PR) at least the names. I'm pretty sure that fuse the iterators in the shortest-variant would be a breaking change for unfused iterators though.