A few days ago I started working on a compiler refactor, which is in this branch . Right now the compiler uses String all over the place: a method, a call, an argument, etc., their names are strings. Variables tracked by the compiler also use strings. They are put in a Hash for lookups. This lookup …

Compiled the new compiler and I’m getting the following error when running the spec ➜ crystal-1 git:(compiler/ident) time bin/crystal build -s spec/std/array_spec.cr Using compiled compiler at .build/crystal Parse: 00:00:00.003804375 ( 3.55MB) Semantic (top level): …

Oops, thanks! I forgot to push the latest changes. Could you try now?

Just ran the spec successfully and you’re right your compiler is just a bit slower, 1.264 vs 1.209 ran on an M1 master bin/crystal build -s spec/std/array_spec.cr 1.48s user 0.25s system 142% cpu 1.209 total compiler/ident bin/crystal build -s spec/std/array_spec.cr 1.55s user 0.26s system 143…

Going from a String to a Crystal::Ident still involves hashing inside StringPool, doesn’t it?

Correct! But the way things are, the Idents are created early by the parser. There’s a bit more Ident creation during literal expansion and normalization. And there are a few more entries during semantic. But the idea is that once the pool of strings is created, putting these Idents in a Hash, look…

Put another way: every lookup on String involves a Hash. But with a pool it involves a hash once, but never again.

Yeah, but even in the worst case that every string occurs only once, I think there really shouldn’t be too much of an overhead. So something’s probably going wrong here. We could do some performance measurements to see where the time is spent (especially in comparison to before/after the patch). B…

Throwing a profiler at it I get [Screenshot from 2022-07-20 22-16-55] This is to be compared to the same when running unodified master: [Screenshot from 2022-07-20 22-17-13] There is a couple of things to note there (holy GC batman!), but the thing that really stands out to me is that in the rem…

[image] yxhuvud: Other than that I’m skeptical this approach with a hash based string pool will gain big improvements as it will still perform the hashes on every lookup Right! But the hash of Ident is just the hash of an integer. In fact I also tried making the hash method of Ident just be t…

On the topic of memory reduction, did you see Andrew Kelley - Practical DOD on Vimeo (which was posted in #beginners in discord some week ago or so)? It may give some inspiration for that kind of work and goes through a bunch of techniques for doing that. First half is general talk about how allocat…

Ongoing experiment: ident pool

Community

bcardiff July 21, 2022, 6:19pm 14

Linux only

Topic		Replies	Views
Does omitting types in method parameters affect compiler speed? Help & Support	6	667	October 12, 2019
Digging into struct initialization performance Website	4	466	September 19, 2024
StringPool: make the hash key part of the public API Crystal Contrib	8	286	April 24, 2024
Build with --release performance is slow than the 2017 crystal version? Help & Support	16	827	June 13, 2022
Very slow build speeds for hello world Help & Support	48	1927	December 25, 2025

Ongoing experiment: ident pool

Related topics