[turbopack] Reduce native React Compiler work - #96820
Merged
Merged
Conversation
Co-Authored-By: Marcos Hernanz <96699542+marcoshernanz@users.noreply.github.com>
Contributor
Tests PassedCommit: 01a3e98 |
Contributor
Stats from current PR✅ No significant changes detected📊 All Metrics📖 Metrics GlossaryDev Server Metrics:
Build Metrics:
Change Thresholds:
⚡ Dev Server
📦 Dev Server (Webpack) (Legacy)📦 Dev Server (Webpack)
⚡ Production Builds
📦 Production Builds (Webpack) (Legacy)📦 Production Builds (Webpack)
📦 Bundle SizesBundle Sizes⚡ TurbopackClient Main Bundles
Server Middleware
Build DetailsBuild Manifests
Build Cache
📦 WebpackClient Main Bundles
Polyfills
Pages
Server Edge SSR
Middleware
Build DetailsBuild Manifests
Build Cache
🔄 Shared (bundler-independent)Runtimes
📎 Tarball URLCommit: 01a3e98 |
![vercel[bot]](https://avatars.githubusercontent.com/in/8329?s=60&v=4){kind=small}
Contributor
Author
|
Upstream follow-up is now open as swc-project/swc#12105. I compared SWC\047s existing \140is_required\140 against this PR\047s conservative predicate on the full v0 corpus. The existing SWC heuristic is not safe as a skip gate: it missed 10 of 257 actually transformed modules (nine \140forwardRef\140 modules plus one member-hook module). The new upstream \140may_require\140 had zero false negatives, selected the same 315 files as this PR, and was 1.52% faster than the local visitor across five detector-only runs. It also moves the Node binding to the conservative contract and treats parse failures conservatively. This Next.js draft should keep the local implementation until the SWC crate release containing that API is available, then switch both Turbopack and next-napi to the upstream predicate and delete the duplicate. |
lukesandberg
reviewed
Aug 6, 2026
lukesandberg
reviewed
Aug 6, 2026
kdy1
pushed a commit
to swc-project/swc
that referenced
this pull request
Aug 8, 2026
## Description
This fixes false negatives in the existing React Compiler
`fast_check::is_required` predicate and documents its conservative
skip-gate contract.
Before this PR, `is_required` was a low-false-positive heuristic, but it
was not safe for callers that skip React Compiler when it returns
`false`. On the real v0 corpus it missed anonymous `forwardRef`
callbacks and a module using member hook calls. This PR improves the
existing method rather than adding a second public predicate, following
maintainer feedback.
The new implementation allows false positives but not false negatives:
`false` means compilation cannot change the program. It scans every
function context for JSX, actual hook calls matching React Compiler
HIR's own hook-name rule, member hook calls, and opt-in/dynamic-gating
directives in real directive prologues.
The Node `isReactCompilerRequired{,Sync}` APIs continue to call
`is_required`. Fatal and recovered parse errors return `true`, because
an unsuccessful fixed-syntax parse cannot prove that compilation is
unnecessary.
The package test command enumerates both test files explicitly so it
remains compatible with the existing Windows + Node 18 CI job.
## v0 correctness corpus
I benchmarked a fixed v0 checkout
(`1ab042a47c053bf5ddad5ffb9c6af153ce8e5b56`) on a 16-vCPU / 32-GB Vercel
DevBox:
- 1,816 tracked JS/TS source files
- 10,464,109 source bytes
- zero parser failures
- 257 modules actually transformed by the Rust React Compiler
- 67 compiler diagnostics
The harness parses the corpus once, runs detectors repeatedly, and uses
the actual compiler result as ground truth. It asserts that gated and
ungated transformed paths are identical.
| Detector | Selected files | False negatives |
| --- | ---: | ---: |
| `is_required` before this PR | 279 | 10 / 257 (3.89%) |
| Current Next.js conservative visitor | 315 | 0 / 257 |
| This PR: improved `is_required` | **302** | **0 / 257** |
The old implementation missed nine real `React.forwardRef` UI modules
and one hook module using `React.useState` / `React.useEffect`. The
improved implementation retains all 257 transforms while rejecting 13
more false positives than the current Next.js visitor.
## End-to-end compiler effect
A fresh optimized release run after applying all review feedback used
five interleaved gated/ungated samples on the exact corpus:
| Pipeline | Median | Transformed modules |
| --- | ---: | ---: |
| No gate | 2,371.956 ms | 257 |
| Improved `is_required` gate | **1,906.180 ms** | 257 |
That removes **465.776 ms / 19.64%** from the React Compiler pipeline on
the real v0 corpus without changing compiler output. The detector itself
took 20.624 ms for the full 10.46-MB corpus (mean of 100 passes).
The temporary synthetic CodSpeed benchmark was removed at review
request. The performance evidence remains the real v0 corpus with actual
compiler output as the correctness oracle.
## Review feedback incorporated
- Improved the existing `is_required` API instead of introducing
`may_require`.
- Detect hooks only at call sites and reuse React Compiler's
`is_hook_name` rule (`use` plus uppercase ASCII or a digit).
- Restrict opt-in strings to leading directive prologues using
`Stmt::can_precede_directive`.
- Treat both fatal and recovered parser errors conservatively in the
Node binding.
- Added actual-compiler regressions proving name-only and JSX-free
`React.createElement` wrapper cases do not compile.
- Removed the synthetic benchmark and changed the changeset from a minor
feature to a patch fix.
## Validation
- `cargo fmt --all -- --check`
- `cargo test -p swc_ecma_react_compiler` (107 unit tests + 15 fixtures)
- `cargo test -p binding_react_compiler_node` (14 passed)
- `cargo clippy -p swc_ecma_react_compiler -p
binding_react_compiler_node --all-targets -- -D warnings`
- `pnpm --filter @swc/react-compiler build:dev`
- `pnpm --filter @swc/react-compiler test` (8/8)
- Earlier full-PR validation: `cargo clippy --all --all-targets -- -D
warnings`; `packages/core: pnpm build:dev && pnpm test` (117 passed, 3
skipped)
- Final GPT-5.6 Sol xhigh + Claude Opus 5 xhigh autoreview panel: zero
findings, “patch is correct” (0.99 confidence)
Follow-up to
[vercel/next.js#96820](vercel/next.js#96820).
This comment has been minimized.
This comment has been minimized.
Contributor
Author
|
This is ready for final maintainer review/merge from the implementation and CI side.
The PR is mergeable with current canary. It remains a draft/review-required PR, so maintainer approval and the normal ready-for-review state transition are the only remaining gates. |
marcoshernanz
marked this pull request as ready for review
August 10, 2026 18:50
lukesandberg
approved these changes
Aug 10, 2026
Contributor
Author
|
Luke review follow-up is complete in signed commit 01a3e98. Changes:
Verification:
All review threads are resolved, Luke approval remains valid, and GitHub reports MERGEABLE / CLEAN. Ready to merge. |
lukesandberg
approved these changes
Aug 10, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What?
swc_ecma_react_compiler::fast_check::is_requiredAPI to skip React Compiler work for modules that cannot change in nativeinfermode.annotationandallmodes unconditional.mdxjs-rs-turbopackbranch, with no duplicate SWC 75 stack.This consumes the implementation landed in swc-project/swc#12105, released in
swc_ecma_react_compiler23.0.0. Its contract is conservative: false positives add compiler work, while a negative result guarantees that compilation cannot alter the program.Why?
The native compiler currently runs for every parsed module in infer mode and is configured for both browser and App SSR contexts. This adds avoidable AST conversion/compiler work and compiles hydrated client modules in a server context that the existing Turbopack Babel path excludes.
This matters for large applications such as v0. Their legacy Babel React Compiler path creates a separate machine-sized Node worker pool in addition to PostCSS. Moving to the native path removes that pool; this patch makes the native path cheaper while preserving the established browser-only scope.
Performance
Upstream fast check on the v0 corpus
The final SWC predicate was checked against 1,816 real v0 modules (10.46 MB):
This Next.js patch only
Real v0 homepage, cold
.next, Chromium interaction gate, 16 vCPU / 32 GB, exact Next canary7916855653, exact v0 commit1ab042a47c, three samples per arm. Both arms use the native compiler.All three paired visible, quiet, and CPU deltas favored the patch. The RSS difference is inside noise, so this PR makes no patch-only memory claim.
v0 legacy Babel path to native + this PR
Both arms use
TURBO_ENGINE_EVICT_AFTER_SNAPSHOT=1and remain open for 120 seconds after network quiet.The large memory/process delta comes from switching v0 off the external Babel pool, not from this source diff alone. v0 still needs to enable the native compiler configuration separately.
Every accepted browser run required HTTP 200, a visible nonblank UI with at least ten controls, zero Next overlays/page errors, and a verified prompt interaction. The complete Next process session was sampled every 100 ms.
Validation
cargo fmt --all -- --checkcargo checkswc_ecma_react_compiler; WASI excludes it; noswc_core75 remainsmdxjs-rs-turbopackSWC 76 bump: 188 tests passed