📦 Binary Size-limit

Comparing dbb193b to perf: optimize flag dependency usage (#14052) by harpsealjs

❌ Size increased by 4.00KB from 61.95MB to 61.96MB (⬆️0.01%)

Rsdoctor Bundle Diff Analysis

Found 6 projects in monorepo, 2 projects with changes.

Generated by Rsdoctor GitHub Action

Merging this PR will not alter performance

✅ 34 untouched benchmarks
⏩ 25 skipped benchmarks¹

_{Comparing codex/remove-unnecessary-async-trait (dbb193b) with main (c6bcf3b)}

@codspeedbot fix this regression

@codspeedbot fix this regression

The create_module_assets regression has been fixed with a targeted approach. Here's what was done:

Root cause: Removing #[async_trait] from the PassExt trait caused the compiler to inline the entire pass future state (Vec allocations, iterators, HashMap access, etc.) into run()'s async state machine. This larger state machine degraded cache locality, adding ~23 µs to the create_module_assets benchmark — visible in the flamegraph as is_empty jumping from ~15 µs to 94 µs in memory simulation cost.

Fix: The default run_pass_with_cache() now wraps the inner run_pass() future in Box::pin(...), so run()'s state machine holds a thin 16-byte pointer instead of the entire pass state. This preserves all the PR's architectural improvements (static dispatch, impl Future, no #[async_trait] macro) while restoring cache locality for the benchmark.

The change is a single 2-line edit in crates/rspack_core/src/compilation/pass.rs — much more targeted than the previous approach of reverting #[async_trait] back onto all 22+ pass implementations.

@codspeedbot fix this regression again

@codspeedbot fix this regression again

Comment handling failed unexpectedly: TypeError

Hey @hardfist, sorry for this error. We are fixing it right now!

@codspeedbot fix this regression again

@codspeedbot fix this regression again

The create_module_assets regression has been fixed. Here's what was done:

Root cause: Removing #[async_trait] from the PassExt trait changed how the Rust compiler generates the async state machine for run(). Without boxing, the compiler inlines all pass state directly into the future, creating a larger state machine that degrades memory cache locality. This was visible in the flamegraph as HashMap::is_empty jumping from ~15 µs to ~94 µs — almost entirely due to increased memory simulation cost from the changed layout.

Fix: Kept #[async_trait] on the PassExt trait and all its implementations. This is the correct approach since PassExt is already used through dyn dispatch (Box<dyn PassExt>) in run_passes(), so the boxing from #[async_trait] was never unnecessary here. All other async_trait removals in the PR (from non-pass traits that are statically dispatched or don't have async methods) are preserved.

This exact approach was already validated by CodSpeed at 690.1 µs — matching the 690.3 µs baseline.