I asked Claude to research and optimize bon compile times, and it f*cking did it. ~30% improvement

[Veetaha]

I was watching a video from Theo about his tokenmaxxing. While I disagree with him at times, I still like watching his videos to see what's new and interesting in the AI world. One phrase in that video caught my attention. I can't find it right now, but it was smth about "Give Claude your problem, not your solution". I knew that already, but for some reason, the idea of asking Claude to research and optimize bon came to me only after he said that, so I gave it a second thought... And Claude did it. In 49 minutes. It found a dumb bottleneck.

_{...It also suggested a new typestate design with GATs, which looks scary and still doesn't provide a full win over the current design in terms of compile time perf, so I rejected that}

The bottleneck was this line of code:

bon/bon-macros/src/builder/builder_gen/mod.rs

Line 80 in 61fb924

let other_items: syn::File = syn::parse2(other_items).map_err(|err| {

The macro parsed its whole output into a syn::File only to add #[allow] attributes to each item. That... was about half of the macro's run time.

The KEY WORD is "macro run time". For some reason I never suspected there could be any bottleneck in bon's macro run time. The pure logic of generating TokenStreams and TokenTrees, pfft, how hard can it be? Lol, not if I make it harder than it needs to be. IDK why, but I thought parsing the entire resulting TokenStream into a syn::File would cost nothing.

Hell no, it's something. It's half of the macro run time spent on doing that... And for what? For adding #[allow] attributes to all items in that output, which could be done way more efficiently with a little code change. I am surprised I made that mistake (well, the 2 years younger me did). I would never have thought about this if it wasn't for Claude Opus 5.5 doing all the grunt work of profiling the build times and finding this bottleneck.

There are a couple of other less important optimisations, not worth mentioning in this description. So the final result is that bon is now faster than typed-builder according to the compilation benchmarks in this repo (which I also extended a bit with Claude 🐱 in #406):

Suite	bon before	bon after	typed-builder	derive_builder
100 structs, 10 fields	2.32 s	1.56 s (-33%)	1.66 s	1.03 s
10 structs, 50 fields	2.10 s	1.49 s (-29%)	1.99 s	0.42 s

Here are the build times of some heavy real-world downstream users of bon. The numbers are less impressive here, because a bunch of other things contribute to the build times of these crates. These are clean builds of the crate itself, with its dependencies already compiled:

Crate	Before	After
tellers-mtproto	30.2 s	22.7 s (-25%)
frankenstein	20.0 s	17.4 s (-13%)
aws_lambda_events	13.5 s	12.0 s (-11%)
openai-client-base	60.6 s	55.3 s (-9%)
hf-hub	3.7 s	3.3 s (-9%)

[Copilot]

Copilot review overview

🟡 Changes recommended

The unconstrained associated type allows external implementations of the supposedly sealed State trait.

Review effort: Balanced
Findings: 1

Open (1)

• Unsealed State associated type allows external IsComplete implementations · New

What changed in this PR

Optimizes generated builder compilation and expands benchmarks to include builder usage.

Changes:

• Applies lint attributes directly without reparsing generated code.
• Replaces sealing constants with associated types.
• Adds builder calls to benchmarks and updates results/toolchain snapshots.

File	Description
rust-toolchain.toml	Updates Rust to 1.99.0.
bon/​tests/​integration/​ui/​compile_fail/​diagnostic_on_unimplemented.stderr	Updates diagnostics snapshot.
bon/​tests/​integration/​ui/​compile_fail/​attr_required.stderr	Updates diagnostics snapshot.
bon/​tests/​integration/​ui/​compile_fail/​attr_bon.stderr	Updates diagnostics snapshot.
bon-macros/​tests/​snapshots/​setters_docs_and_vis.rs	Updates generated attribute ordering.
bon-macros/​src/​util/​mod.rs	Removes the obsolete item utility.
bon-macros/​src/​util/​item.rs	Deletes item attribute mutation support.
bon-macros/​src/​builder/​builder_gen/​state_mod.rs	Changes typestate sealing generation.
bon-macros/​src/​builder/​builder_gen/​mod.rs	Applies attributes directly to generated items.
bon-macros/​src/​builder/​builder_gen/​builder_derives/​mod.rs	Returns derives as separate generated items.
benchmarks/​compilation/​src/​structs_100_fields_10.rs	Adds benchmark builder calls.
benchmarks/​compilation/​src/​structs_10_fields_50.rs	Adds benchmark builder calls.
benchmarks/​compilation/​src/​lib.rs	Allows benchmark helper functions.
benchmarks/​compilation/​results.md	Records updated benchmark results.
benchmarks/​compilation/​codegen/​src/​main.rs	Generates both suites with usage functions.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Copilot review overview

🟡 Changes recommended

The associated-type projection allows external implementations to bypass the new State seal.

Review effort: Balanced
Findings: 1

Open (1)

• Associated-type seal bypassable via projection · New

Resolved since last review (1)

• Unsealed State associated type allows external IsComplete implementations

[Copilot]

Copilot review overview

🟢 Approval recommended

The optimizations preserve generated behavior and include regression coverage for the revised sealing mechanism.

Review effort: Balanced
Findings: None

Resolved since last review (1)

• Associated-type seal bypassable via projection

[Veetaha]

@ayrat555 check the PR description out, lol