Richard Feldman (Nov 03 2026 at 18:10): Richard Feldman (Nov 03 2026 at 18:10):I haven't really looked into string interpolation perf - do you know offhand what's inefficient about it at the moment?
Richard Feldman (Nov 03 2026 at 18:11):I'm doing perf stuff now but plan to stop that in favor of bugfixes only pretty soon, so now's the time for fixes like that :smile:
Zak Kohler (Nov 03 2026 at 20:35):Richard Feldman said:
I haven't really looked into string interpolation perf - do you know offhand what's inefficient about it at the moment?
Its all mentioned here https://github.com/roc-lang/roc/issues/12824:
2 have PRs (#12750 + #12823) that are changes to built-ins only.
Those two alone close the gap:
┌─────────────────────────────────┬────────────────────────┬──────────────────┐
│ competitor │ instructions a request │ cycles a request │
├─────────────────────────────────┼────────────────────────┼──────────────────┤
│ fourneau-zig │ 344k │ 216k │
├─────────────────────────────────┼────────────────────────┼──────────────────┤
│ roux, nightly + #12750 + #12823 │ 502k │ 383k │
├─────────────────────────────────┼────────────────────────┼──────────────────┤
│ axum │ 510k │ 410k │
├─────────────────────────────────┼────────────────────────┼──────────────────┤
│ roux, nightly-2026-10-24 │ 687k │ 511k │
├─────────────────────────────────┼────────────────────────┼──────────────────┤
│ go (templ) │ 2,231k │ 1,889k │
├─────────────────────────────────┼────────────────────────┼──────────────────┤
│ basic-webserver │ 3,640k │ 3,924k │
└─────────────────────────────────┴────────────────────────┴──────────────────┘
The last is doing something more complicated for string interpolation.
Zak Kohler (Nov 03 2026 at 20:44):Richard Feldman said:
I haven't really looked into string interpolation perf - do you know offhand what's inefficient about it at the moment?
Sorry I didn't remember that i already cross posted that summary....
specifically the string interpolation compilers to a List(Str) doing allocation and holding a closure. against a handwritten chain of concats difference was 37% in cpu cycle, but i did not disambiguate how much of that comes just from the optimizing away zero length strings in the interpolation string.
Zak Kohler (Nov 03 2026 at 20:48):turns out, it was 100% the empty string optimization.
Zak Kohler (Nov 03 2026 at 21:01):For the other 2 PRs, I'm currently doing some perf regression testing on how it affects largers strings.
Zak Kohler (Nov 04 2026 at 00:16):https://github.com/roc-lang/roc/pull/12830
There's a proof of concept for a slight improvement. It implicitly eliminates empty strings and that's the majority of improvement.
I think the other two have more bang for the buck
Richard Feldman (Nov 04 2026 at 02:05):hrm, this investigation makes me realize that from_interpolation should take an Iter, not a List, now that Iter exists
Richard Feldman (Nov 04 2026 at 02:05):because it doesn't require a heap allocation the way List does
Richard Feldman (Nov 04 2026 at 02:10):@Zak Kohler thanks for opening all this! I'm gonna merge your other two PRs and then open a replacement PR for https://github.com/roc-lang/roc/pull/12823 that makes the API change, along with the change you have in there, and some other related changes
Zak Kohler (Nov 04 2026 at 04:03):Updated the small str concat: https://github.com/roc-lang/roc/pull/12823 not sure which version you saw, but early experiments has some small regression in long string perf. This final version seems best of both worlds.
I had the agents summarize all the experi toments for the small-str concat optimization and update the PR message body.
Be sure check out the comparison it is quite interesting. I definitely am request some keen review on this as well since I'm steering the LLMs beyond what I fully grok here.
Finally, the small string in register perf is sensitive to inlining. Do you have any thoughts on perf regression testing/codegen asserts?
Richard Feldman (Nov 04 2026 at 04:07):yeah https://github.com/roc-lang/roc/pull/12823#pullrequestreview-5763068602 is the one I had already approved. Re-approved!
Richard Feldman (Nov 04 2026 at 04:39):Zak Kohler said:
Finally, the small string in register perf is sensitive to inlining. Do you have any thoughts on perf regression testing/codegen asserts?
can you say more about this? I need more specifics to understand the question! :smile:
Zak Kohler (Nov 04 2026 at 05:03):Basically we are right at threshold of llvm's inlining discretion. Had to fiddle/measure to get out just right.
LLM summary:
Sure! The fast path in #12823 only pays off when LLVM inlines roc_builtins_str_concat at the call site. That keeps the small strings' words in registers across a chain of concats (about 15 cycles a concat, against about 90 on main). Inlining is LLVM's cost-model call, and the final wrapper is close to the line: at 74 IR instructions it's inlined at 44 of 44 sites on my test page, while an earlier 88-instruction version was inlined at only 16 of 43. So an innocent edit to the wrapper, an LLVM upgrade, or a change in how codegen calls it could quietly drop it out of line. Strings would still be correct, and it would still beat main (variants inlined nowhere measured −13% on the page against −30%), but most of the win would be gone and nothing would fail.
So the question is whether you'd want either:
a test that builds a small app at --opt=speed and asserts the optimized IR has no call @roc_builtins_str_concat in the hot function (the IR tests in MonoLlvmCodeGen.zig check pre-optimization IR, so this would need a post-opt hook), or
codegen marking that builtin alwaysinline, so it holds by construction, with the trade-off you already cap in #11451 if the wrapper ever grows.
Zak Kohler (Nov 04 2026 at 05:14):https://github.com/roc-lang/roc/pull/12823#issuecomment-6622877160
Well your iter interpolation made the small-str concat superflous for at least my bench. I'm going to see if I could get the same small-str trick on top of the new interpolation mechanism.
Zak Kohler (Nov 04 2026 at 11:15):Zak Kohler said:
. I'm going to see if I could get the same small-str trick on top of the new interpolation mechanism
A: nope
Zak Kohler (Nov 04 2026 at 14:32):Its not a bad nope either, your implementation already lets it all happen naturally.
Zak Kohler (Nov 04 2026 at 17:47):
I expect the three string PRs to put us back above Rust/Axum based on cycle counts. I'm excited for the next Nightly! I'll call it now 12.4k req/s
A tangent:
The new harder template bench makes Go-lang look even weaker.
I decided to make it a more fair and switch from the std-lib reflection based templates to the fairly popular third party compiled templates (.templ)...
...just benchmarked it...it only boosts Go from 600->2500 req/s
Zak Kohler (Nov 04 2026 at 17:56):Could someone move this to performance instead of show and tell?
Anton (Nov 04 2026 at 17:57):Should I move all messages there or starting from a specific message?
Zak Kohler (Nov 04 2026 at 17:58):I already split the tangent and renamed it, so this whole thread can move. i just couldn't move it to differnent area
Notification Bot (Nov 04 2026 at 18:00):This topic was moved here from #show and tell > Rocstache: small string concat perf by Anton.
Last updated: Nov 04 2026 at 18:54 UTC