Pitch: Efficient `Collection.startIndex` throughout the standard library

[Karl]

@timv

Thanks for raising this concern, this is definitely good to keep in mind when adding new types and deprecating the old ones.

I think generic value parameters have kind of gone a bit under-the-radar next to variadic generics, which are much sexier, but they would give us tremendous capabilities for any kind of wrapper which includes a closure. As things stand, you can get significantly better, more efficient wrappers than any general-purpose type by copy/pasting the implementation and swapping the closure for a function. In a project I'm working on, I did this with LazyFilterCollection and saw enormous performance improvements (I assume because the compiler didn't need to retain a closure, and generic specialisation told it exactly what the filter predicate was, so it could inline it and optimise around it).

Just re-ran my benchmarks

benchmark                                              column      baseline    latest       %
---------------------------------------------------------------------------------------------
Constructor.SpecialNonFile.AverageURLs filtered        std            31.94     22.33   30.09
Constructor.SpecialNonFile.AverageURLs filtered        iterations  20704.00   5217.00   74.80
Constructor.SpecialNonFile.AverageURLs filtered        time        59565.00 250315.00 -320.24
...
Constructor.SpecialNonFile.IPv4 host filtered          std            33.79     40.77  -20.64
Constructor.SpecialNonFile.IPv4 host filtered          iterations  26756.00   8477.00   68.32
Constructor.SpecialNonFile.IPv4 host filtered          time        46838.00 154797.00 -230.49
...
Constructor.SpecialNonFile.IPv6 host filtered          std            29.79     28.72    3.62
Constructor.SpecialNonFile.IPv6 host filtered          iterations  22074.00   7611.00   65.52
Constructor.SpecialNonFile.IPv6 host filtered          time        57169.00 168267.00 -194.33

"baseline" is using my copy/pasted lazy-filter, "latest" is using the standard library's one. It's anywhere from ~200% to 300% slower to use the standard library.

I'm not sure how widely-known this issue is, or how the core team rates its severity. Personally I think it is more pressing than variadic generics, such that it's probably worth holding off on touching the lazy collections until they can be fixed.

As a bonus, generic value parameters would mean we can finally make @lazy a property wrapper, without having to actually store a closure inside of it.