Skip to content

[0.88] Compile React Native C/C++ at -O2 on iOS to match Android - #59478

Open
chrfalch wants to merge 1 commit into
0.88-stablefrom
chrfalch/0.88-port-59226
Open

chrfalch wants to merge 1 commit into
0.88-stablefrom
chrfalch/0.88-port-59226

Conversation

@chrfalch

@chrfalch chrfalch commented Oct 26, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

Port of 60c817e (#59226, "Compile React Native C/C++ at -O2 on iOS to match Android") to 0.88-stable. That commit already landed in main. A plain cherry-pick conflicts, so this PR resolves the conflicts.

iOS Release builds compile React Native's C, C++ and Objective-C code with -Os. Android compiles the same code with -O2. This change sets -O2 for React Native's own code in Release on iOS, and for Yoga and glog in all builds, as on Android. Hermes and Swift code do not change. See #59226 for the benchmark: a Yoga layout takes 21–54% less time in Release, and the stripped React binary grows by 6.2%.

The conflicts come from two later PRs that are not on 0.88-stable:

  • Package.swift: the Yoga target on main also has excludedPaths: ["test"]. This port adds only the -O2 setting.
  • scripts/codegen/templates/Package.swift.spm-template: main also defines RN_ALLOW_FRAMEWORKS next to the new flag. This port adds only the Release -O2 flag.

Everything else is the same as the commit on main.

Pick request: reactwg/react-native-releases#1463

Changelog:

[IOS] [CHANGED] - Compile React Native's C/C++/Objective-C code with -O2 instead of -Os in Release, and Yoga and glog with -O2 in all builds, to match Android

Test Plan:

  • git range-diff against 60c817e: the only differences are the two conflicts described above.
  • yarn jest --no-cache -i packages/react-native/scripts/spm packages/react-native/scripts/ios-prebuild/__tests__/build-test.js scripts/releases/ios-prebuild/__tests__/build-test.js on this branch: 1018 of 1018 tests pass.
  • swift package dump-package --package-path packages/react-native: exits 0.
  • ruby -c on utils.rb and react_native_pods.rb: Syntax OK.
  • Not run: an iOS build with Xcode or CocoaPods. CI on this PR covers it.

Summary:
iOS Release builds compile React Native's C, C++ and Objective-C code with `-Os` (optimize for size). React Native never sets `GCC_OPTIMIZATION_LEVEL`, so Xcode's default applies. Android compiles the same code with `-O2` (optimize for speed) in Release. Yoga even sets `-O2` explicitly ("prioritize speed", `ReactCommon/yoga/cmake/project-defaults.cmake`).

This PR sets `GCC_OPTIMIZATION_LEVEL = 2` for Release builds of React Native's own code on iOS, to match Android. Android also builds Yoga and glog at `-O2` in Debug builds, so this PR does the same on iOS. All other code keeps `-O0` in Debug. Hermes keeps its size-optimized `MinSizeRel` build, as on Android. Swift code is not affected.

In a layout benchmark on Apple Silicon, a Yoga layout takes 21–54% less time at `-O2` than at `-Os` in Release. In Debug it takes about 94% less time at `-O2` than at `-O0` (16–19 times faster). The stripped React binary grows by 6.2%.

After this PR, every library that both platforms compile uses the same optimization level: React Native core, Yoga, folly, glog, fmt, boost and double-conversion. Two differences stay out of scope: Android also uses link-time optimization, and Android turns off Yoga's assertions in Debug builds.

I found no commit that chose `-Os` for iOS on purpose. The earlier optimization decisions were all for Android: `c04ab921924` (2016, "Use -O3 to compile CSSLayout in open source … This is how we compile internally") and #37371 (2023, Yoga's reference CMake build at `-O2`).

The question came up in the community. A user profiling Bluesky found Yoga about 20% faster at `-O2` and layout 2–5% faster in the full app: bluesky-social/social-app#12061. Their app-level workaround is expo/expo#51770. An app-level setting cannot reach the prebuilt React Native frameworks, so the fix belongs here.

### What changes

| Build path | Change |
|---|---|
| Prebuilt `React.xcframework` (`packages/react-native/scripts/ios-prebuild/build.js`) | Adds `GCC_OPTIMIZATION_LEVEL=2` to the `xcodebuild` command for the Release flavor only. |
| Prebuilt `ReactNativeDependencies.xcframework` (`scripts/releases/ios-prebuild/build.js`) | Same, for folly, glog, fmt, boost, double-conversion, fast_float and SocketRocket. |
| CI cache (`.github/workflows/prebuild-ios-dependencies.yml`) | The slice and xcframework cache keys now include `scripts/releases/ios-prebuild/build.js`. Without this, CI reuses the cached `-Os` build. |
| Yoga in `React.xcframework` (`packages/react-native/Package.swift`) | The `yoga` target gets `-O2` in every configuration, as on Android (`ReactCommon/yoga/CMakeLists.txt`). |
| glog in `ReactNativeDependencies.xcframework` (`scripts/releases/ios-prebuild/configuration.js`) | glog gets `-O2` in every configuration, as on Android (`ReactAndroid/src/main/jni/third-party/glog/CMakeLists.txt`). |
| CocoaPods builds from source (`react_native_post_install`) | New step `set_optimization_level_for_react_native_pods`. It sets level 2 on Release-type configurations of React Native pods, and on every configuration of `Yoga` and `glog`. |
| SwiftPM codegen package (`packages/react-native/scripts/codegen/templates/Package.swift.spm-template`) | `ReactCodegen` and `ReactAppDependencyProvider` get `-O2` in Release through `unsafeFlags`, as the CocoaPods generated pods do. SwiftPM allows `unsafeFlags` here because the app always uses this package from a local path. |

The command-line setting applies to every configuration, so the prebuild scripts add it only when they build Release. The per-target `-O2` for Yoga and glog comes after the configuration's `-O0` on the compiler command line, so it wins in Debug.

In CocoaPods, a React Native pod is one of these:
- a local pod inside the `react-native` package
- one of React Native's third-party dependency pods (`RCT-Folly`, `glog`, `boost`, `DoubleConversion`, `fmt`, `fast_float`, `SocketRocket`)
- one of the generated pods `ReactCodegen` and `ReactAppDependencyProvider`

Community library pods, the app's own pods and `hermes-engine` do not change. Release-type configurations are found by `config.type == :release`, as `add_ndebug_flag_to_pods_in_release` already does, so custom configurations such as "Staging" are included.

An app that wants the old behaviour can set `GCC_OPTIMIZATION_LEVEL` back to `s` in its own `post_install` after `react_native_post_install`.

### Benchmark: layout speed

Yoga built twice from this branch with Apple clang 21.0.0 (`-std=c++20 -DNDEBUG -arch arm64`); only the optimization level differs. The benchmark harness is the same object file in both builds. Every iteration forces a full relayout (root width alternates 390/391). There were 10 runs of each build, alternating. Both builds produce identical layout results.

Machine: Apple M4 Max, macOS 27.0.1. This is a Mac, not an iPhone, and the trees are synthetic.

| Scenario | `-Os` (µs per layout, median) | `-O2` (µs per layout, median) | Time reduction |
|---|---|---|---|
| Feed: 500 rows, each with an avatar, 3 measured text nodes and 4 buttons | 2,879 | 2,261 | 21% |
| Deep: 15 levels, mixed row/column, padding, margin, percent widths | 1,098 | 805 | 27% |
| Grid: 1,000 wrapped items with `aspectRatio` | 1,018 | 472 | 54% |

"Time reduction" = 1 − median(`-O2`) / median(`-Os`). The spread between the fastest and slowest of the 10 runs was 3–11%.

Debug builds, same harness, built without `NDEBUG` so Yoga's assertions stay on, as in an iOS Debug build. Median of 3 alternating runs:

| Scenario | `-O0` (µs per layout) | `-O2` (µs per layout) | Time reduction |
|---|---|---|---|
| Feed | 36,350 | 2,310 | 94% (15.7× faster) |
| Deep | 15,184 | 816 | 95% (18.6× faster) |
| Grid | 8,190 | 486 | 94% (16.9× faster) |

### Benchmark: binary size

Release, `ios-arm64` device slice. Sizes are uncompressed; App Store download size will differ.

| Binary | `-Os` (bytes) | `-O2` (bytes) | Difference |
|---|---|---|---|
| `React.framework/React`, stripped (`strip -x`) | 6,154,912 | 6,537,576 | +382,664 (+6.2%) |
| `React.framework/React`, `__TEXT` segment | 4,931,584 | 5,324,800 | +393,216 (+8.0%) |
| `ReactNativeDependencies.framework`, stripped | 778,240 | 873,296 | +95,056 (+12.2%) |
| Yoga only, `__text` section | 77,140 | 112,188 | +35,048 (+45%) |

The unstripped React binary gets *smaller* (10,785,672 → 10,335,408 bytes). More inlining at `-O2` leaves about 4,950 fewer symbols, so the link-edit data shrinks more than the code grows. The stripped size is the relevant number.

If only Yoga were moved to `-O2`, the stripped React binary would grow by 32,776 bytes (+0.53%).

## Changelog:

[IOS] [CHANGED] - Compile React Native's C/C++/Objective-C code with `-O2` instead of `-Os` in Release, and Yoga and glog with `-O2` in all builds, to match Android

Pull Request resolved: #59226

Test Plan:
**Unit tests.** New Jest tests check that the Release command has `GCC_OPTIMIZATION_LEVEL=2`, the Debug command does not, and the rest of each command is unchanged. They failed before the change and pass after it.

```
yarn jest packages/react-native/scripts/ios-prebuild/__tests__/build-test.js scripts/releases/ios-prebuild/__tests__/build-test.js
Tests: 8 passed, 8 total
```

A new Jest test checks that the SwiftPM codegen template gives `ReactCodegen` and `ReactAppDependencyProvider` the Release-only `-O2` flag in both `cSettings` and `cxxSettings`, and that `ReactAppHeaders` does not get it. It failed before the change and passes after it.

```
yarn jest packages/react-native/scripts/spm/__tests__/codegen-template-optimization-test.js
Tests: 5 passed, 5 total
```

All tests in `packages/react-native/scripts/ios-prebuild` and `scripts/releases`: 24 suites, 196 tests passed. All tests in `packages/react-native/scripts/spm`: 23 suites, 1060 tests passed. ESLint, Prettier and Flow report no errors on the changed files.

**Prebuilt frameworks.** I built `React.xcframework` (`node scripts/ios-prebuild --setup --build --compose --platforms ios --flavor <Debug|Release>`) and `ReactNativeDependencies.xcframework` (`node scripts/releases/prepare-ios-prebuilds.js --setup --build --compose --platforms ios --configuration <Debug|Release>`), then read the last `-O` flag in each target's compiler argument file:

| Build | React Native code | Yoga | glog | Other dependencies |
|---|---|---|---|---|
| Release, before this PR | `-Os` | `-Os` | `-Os` | `-Os` |
| Release, with this PR | `-O2` | `-O2` | `-O2` | `-O2` |
| Debug, with this PR | `-O0` | `-O2` | `-O2` | `-O0` |

**CocoaPods.** `RCT_USE_PREBUILT_RNCORE=0 RCT_USE_RN_DEP=0 bundle exec pod install` in `packages/rn-tester`, then read `Pods.xcodeproj`:
- `Yoga` and `glog`: `GCC_OPTIMIZATION_LEVEL = 2` in Debug and Release.
- Other React Native targets (`React-*`, `RCT-Folly`, `ReactCodegen`, `SocketRocket`, …): 2 in Release, unset in Debug.
- RNTester's own pods (`ScreenshotManager`, `MyNativeView`, `NativeCxxModuleExample`, …) and the app targets: unset in both configurations.

**SwiftPM.** In `private/helloworld`: `node ../../packages/react-native/cli.js spm add --deintegrate --version nightly`, then `xcodebuild -configuration Release -sdk iphonesimulator build`. The build succeeded, and Xcode accepted `unsafeFlags` in the generated package. The `-O` flags in each target's compiler argument file:
- `ReactCodegen` and `ReactAppDependencyProvider`, Release: `-Os`, then `-O2`. The last flag wins.
- The app target and `ReactAppHeaders`, Release: `-Os` only.
- `ReactCodegen` and `ReactAppDependencyProvider`, Debug: `-O0` only.

**Not tested:**
- `pod install` in prebuilt mode. No prebuilt frameworks exist for the development version, so it fell back to source. In prebuilt mode the React pods contain only binaries, so the setting has no effect there.
- Layout speed on a physical iPhone, and in RNTester.
- The Ruby CocoaPods unit tests, which are not run in CI.

Reviewed By: andrewdacenko

Differential Revision: D126452701

Pulled By: cipolleschi

fbshipit-source-id: d1b4296200a37300f0d6134cc8f922b4c3754c61

(cherry picked from commit 60c817e)

Conflict resolution for 0.88-stable: added only the -O2 flags, without the main-only Yoga excludedPaths and codegen RN_ALLOW_FRAMEWORKS define.
@meta-cla meta-cla Bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Oct 26, 2026

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. p: Expo Partner: Expo Partner Pick Request

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant