In Enumerable#to_a, use size to set array capa when possible
Cross-post from https://github.com/ruby/ruby/pull/444.
Enumerable#to_a works by creating an empty array with small capacity, then populating it and expanding the capacity as it goes. For large enumerables, this causes several resizes, which can hurt performance. When an enumerable exposes a size method, we can guess that the resulting array's size will usually be equal to the enumerable's size. If we're right, we only have to set capacity once, and if we're wrong, we don't lose anything.
The attached file (or linked PR) adjusts enum.c's to_a method to take advantage of the size method when it's there. In my tests this makes Range#to_a about 10% faster, and doesn't have any significant effect on a vanilla enum with no size method. I couldn't find any existing benchmark that this consistently made better or worse.
If you like this idea, this could also be done in other classes with custom to_a, like Hash.
#2 [ruby-core:58381] Updated by HonoreDB (Aaron Weiner) about 4 years ago
Ah, right! This seems like an opportunity to improve on existing behavior: right now that just silently hangs forever. Do you think we should warn, then hang, or just raise? I'd lean towards the warn because it's possible size is returning the wrong thing.
#3 [ruby-core:58382] Updated by mame (Yusuke Endoh) about 4 years ago
I think the proposal will break the compatibility of the following code:
C.new.size #=> expected: 0, with the proposal: stack level too deep
Examples in the wild:
In addition, #each and #size does not necessarily have a common semantics.
In fact, IO#each yields strings in lines, but IO#size returns a count in bytes.
Yusuke Endoh email@example.com
#4 [ruby-core:58383] Updated by HonoreDB (Aaron Weiner) about 4 years ago
It definitely breaks that usage, but that's bad usage--we're supposed to use Enumerable#count for that, not size.
In cases where size doesn't correctly predict the array, this doesn't really break anything, it just switches out one bad guess at capa for another.