Convert iOS native module event arguments to JS without UTF-8 by hryhoriiK97 · Pull Request #59449 · react/react-native

hryhoriiK97 · 2026-10-24T12:20:22Z

Summary:

Native modules send events to JS with RCTEventEmitter sendEventWithName:body:, which ends in RCTInstance -callFunctionOnJSModule:method:args:. Core modules use it too: for example, RCTNetworking sends the whole text of an XHR response in didReceiveNetworkData, and RCTWebSocketModule sends each message.

Current behavior. The arguments go through folly::dynamic. On the sending thread, convertIdToFollyDynamic encodes every string as UTF-8. On the JS thread, jsi::valueFromDynamic decodes it again into a JS string. Text that isn't ASCII is the expensive case: a 1 MB response with accented text costs about 3.9 ms on the JS thread and 3.2 ms on the sending thread.

With this change.

On the sending thread, RCTInstance reads the arguments into a small tree with the structure that convertIdToFollyDynamic creates. Strings keep their own ASCII or UTF-16 characters instead of being encoded as UTF-8. When CoreFoundation exposes the characters, the tree points at them in an immutable copy of the string; otherwise it copies them.
On the JS thread, the JS values are created from the tree, without Objective-C calls, the same way valueFromDynamic does.
ReactInstance gets a callFunctionOnModule overload that takes an ArgFactory, which creates the arguments on the JS thread. It has the same shape as TurboModule::ArgFactory, which emitDeviceEvent uses. Both overloads share the module lookup, so Android and other callers of the folly::dynamic overload run the same code as before.

JS receives the same values as before:

Numbers use the same rules as convertIdToFollyDynamic.
Number keys become strings, and other non-string keys are skipped.
Values of unsupported types become null.
Strings with a lone surrogate, which UTF-8 can't encode, keep the UTF-8 path, so they still arrive as "".

The order of properties in event objects now follows the NSDictionary. Before, it followed the folly::dynamic hash map, so it was arbitrary in both cases. On JSC, a string with both a NUL and non-ASCII characters is no longer cut off at the NUL, as JSStringCreateWithUTF8CString did. Hermes output is the same.

Changelog:

[IOS] [CHANGED] - Events sent by native modules are converted to JS values without a UTF-8 round trip

Test Plan:

clang-format --dry-run --Werror on the changed files passes.
RNTester built with xcodebuild in Release and run on a physical iPhone 15 Pro (iOS 26.6.1).
Added ReactInstanceTest.testCallFunctionOnModule_withArgFactory and testCallFunctionOnModule_withEmptyArgFactory for the new overload. I couldn't run the C++ tests locally. I updated the C++ API snapshots by hand, because I couldn't run Doxygen here.
Same output.
- A test payload was sent through the old and the new conversion in the same app run, and JS compared the results deeply. They were identical. The payload had:
  - empty, NUL, BOM, emoji, accented, CJK and lone-surrogate strings, and a mutable string changed after sending;
  - BOOL, char, unsigned char, short, int, negative, unsigned long long, 2^60, float, double, NaN and NSDecimalNumber numbers;
  - NSNull, NSDate and NSURL values;
  - nested arrays and dictionaries;
  - number, bool and NSNull keys.
- Every XHR response text in the measurement below matched the original string.
Measurement. RNTester Playground, XHR GET of data: URLs with ASCII or accented text: 1000 requests at 1 KB, 30 at 100 KB and 10 at 1 MB. Also 1000 small events sent from a background queue: a keyboard-like payload, and one with short keys and values created at run time.
- Timed on the JS thread: creating the JS arguments, valueFromDynamic before and the tree conversion after.
- Timed on the sending thread: convertIdToFollyDynamic before and building the tree after.
- The two arms are the same binary with the conversion switched by a launch argument, in the order before, after, after, before, 4 launches per arm. The presses were driven by Argent on the physical iPhone.
- iPhone 15 Pro, Release, medians per event. The ranges of the two arms are separate in every row.

Event	JS thread before	after	change	Sending thread before	after	change
`didReceiveNetworkData`, accented, 1 KB	4.82 µs	0.36 µs	−92%	4.11 µs	0.38 µs	−91%
`didReceiveNetworkData`, accented, 100 KB	382 µs	8.6 µs	−98%	311 µs	3.1 µs	−99%
`didReceiveNetworkData`, accented, 1 MB	3.87 ms	0.09 ms	−98%	3.18 ms	0.03 ms	−99%
`didReceiveNetworkData`, ASCII, 1 KB	0.65 µs	0.34 µs	−47%	0.97 µs	0.38 µs	−61%
`didReceiveNetworkData`, ASCII, 100 KB	12.1 µs	7.8 µs	−36%	5.8 µs	0.35 µs	−94%
`didReceiveNetworkData`, ASCII, 1 MB	97.6 µs	71.8 µs	−26%	49.3 µs	0.33 µs	−99%
`didReceiveNetworkResponse` (headers)	1.08 µs	0.74 µs	−31%	2.45 µs	0.78 µs	−68%
keyboard-like event	1.79 µs	1.17 µs	−35%	5.83 µs	2.28 µs	−61%
short strings created at run time	1.04 µs	0.61 µs	−41%	2.58 µs	0.79 µs	−70%

🤖 Generated with Claude Code

hryhoriiK97 · 2026-10-24T12:23:53Z

Hi @javache, this one is similar to #59325, but for events. On iOS, event arguments don't go through folly::dynamic anymore, so strings aren't converted to UTF-8 and back. I used the same string order you suggested in #59325. Android isn't changed.

If it's easier to review, I can move CallArgument to a separate file. And #59450 is a smaller change to the TurboModule converter, it doesn't depend on this one.

github-actions · 2026-10-24T19:18:53Z

Visual regression

1 visual difference(s) at 22dbf66562a2. Visual differences need maintainer approval. Review the report, then apply visual-change-approved.

ios `button/default-styling/default`: Changed, 38223 px (1.533%)

Expected and Captured

Expected	Captured

hryhoriiK97 · 2026-10-25T21:48:12Z

Hi @javache, the iOS visual diff isn't from this change: a system overlay covers the top of the crop, and the content matches the baseline. On an iPhone 15 Pro, before and after are pixel-identical. Could you apply visual-change-approved and take a look when you have time?

javache · 2026-10-26T11:18:18Z

+    if ([value isKindOfClass:[NSString class]]) {
+      kind_ = Kind::String;
+      setString(value);
+    } else if ([value isKindOfClass:[NSNumber class]]) {
+      setNumber(value);
+    } else if ([value isKindOfClass:[NSArray class]]) {
+      kind_ = Kind::Array;
+      children_.reserve([value count]);
+      for (id element in value) {
+        children_.emplace_back(element);
+      }
+    } else if ([value isKindOfClass:[NSDictionary class]]) {
+      kind_ = Kind::Object;
+      children_.reserve([value count] * 2);
+      for (id key in value) {
+        if ([key isKindOfClass:[NSString class]]) {
+          children_.emplace_back(key);
+        } else if (NSString *numberKey = numberKeyString(key)) {
+          children_.emplace_back(numberKey);
+        } else {
+          continue;
+        }
+        children_.emplace_back(value[key]);
+      }
+    }


This is very similar to convertObjCObjectToJSIValue in RCTTurboModule.

Can we extract it a single shared helper that we use across both paths.

Thank you for the comment! Done 👍

javache · 2026-10-26T11:18:42Z

+        } else {
+          continue;
+        }
+        children_.emplace_back(value[key]);


Why do this conversion up-front (requiring additional allocation) instead of lazily in the toJSValue path?

@javache good point, it's lazy now: the args are converted on the JS thread in the arg factory, so the sending thread doesn't build anything anymore - d05ed19

javache · 2026-10-26T13:04:52Z

+      [moduleName UTF8String] ?: "",
+      [method UTF8String] ?: "",


Early return if either moduleName or method is the empty string.

Done - 843adb2

Sorry, I wanted to avoid the ?: "" fallback here entirely, that should never happen.

That can be an early return inside callFunctionOnModule?

@javache sure! done - f05f675

javache · 2026-10-26T13:06:39Z

+            moduleName,
+            methodName,
+            [&](std::vector<jsi::Value>& jsArgs) {
+              for (auto& arg : args) {


jsArgs.reserve?

@javache done, thanks! 843adb2

meta-codesync · 2026-10-26T14:47:27Z

@javache has imported this pull request. If you are a Meta employee, you can view this in D127583929.

…omodule-faster-object-conversion Keeps react#59449's key handling in convertNSDictionaryToJSIObject: number keys become strings, and other keys that aren't strings are skipped. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Events sent with RCTEventEmitter went through folly::dynamic: every string was encoded as UTF-8 on the sending thread and decoded again on the JS thread. RCTInstance now reads the arguments on the sending thread into a small tree with the same structure, where strings keep their ASCII or UTF-16 characters, and the JS thread creates the JS values from it. ReactInstance gets a callFunctionOnModule overload that takes an ArgFactory, which creates the arguments on the JS thread. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Drop CallArgument and convert the arguments on the JS thread with the TurboModule helper, like codegen's AsyncEventEmitter does. The helper now turns number keys into strings and skips other keys that aren't strings, as convertIdToFollyDynamic does, instead of treating every key as an NSString. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Move the empty-name check from RCTInstance into both ReactInstance::callFunctionOnModule overloads, so it returns before anything is scheduled on the JS thread for every caller. RCTInstance's parameters are nonnull, so it passes UTF8String directly without a fallback. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

hryhoriiK97 · 2026-10-29T11:58:39Z

@javache I rebased on main and resolved the conflict, ready for another look 🙂

christophpurrer

Review automatically exported from Phabricator review in Meta.

javache · 2026-10-30T19:55:49Z

  }
  _reactInstance->callFunctionOnModule(
-      [moduleName UTF8String] ?: "", [method UTF8String] ?: "", convertIdToFollyDynamic(args ? args : @[]));
+      [moduleName UTF8String], [method UTF8String], [callArgs](jsi::Runtime &runtime, std::vector<jsi::Value> &jsArgs) {


Let's extract these locals and do the bailout here - otherwise it will crash turning a nullptr into an std::string

const char *moduleNameStr = [moduleName UTF8String];
const char *methodNameStr = [method UTF8String];
If (moduleNameStr == nullptr || methodNameStr == nullptr) {
return;
}

@javache thank you for the comment! Done - bad4fb2

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

hryhoriiK97 · 2026-11-05T12:41:24Z

@javache @christophpurrer the nullptr bailout is in bad4fb2. Could you take another look when you have a moment?

javache · 2026-11-13T10:38:42Z

+    __unsafe_unretained id key = keysAndObjects[i];
+    NSString *numberKey = nil;
+    if (![key isKindOfClass:[NSString class]]) {
+      if (![key isKindOfClass:[NSNumber class]] || [key isKindOfClass:[@YES class]]) {
+        continue;
+      }
+      numberKey = [key stringValue];
+      key = numberKey;
+    }


Simplify?

Suggested change

__unsafe_unretained id key = keysAndObjects[i];

NSString *numberKey = nil;

if (![key isKindOfClass:[NSString class]]) {

if (![key isKindOfClass:[NSNumber class]] || [key isKindOfClass:[@YES class]]) {

continue;

}

numberKey = [key stringValue];

key = numberKey;

}

id key = keysAndObjects[i];

if (![key isKindOfClass:[NSString class]]) {

if (![key isKindOfClass:[NSNumber class]] || [key isKindOfClass:[@YES class]]) {

continue;

}

key = [key stringValue];

}

@javache thanks, done in 22dbf66. It adds one retain/release per key, which costs about 2–5% on small events on an iPhone, but it's much simpler

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

react-native-bot · 2026-11-13T14:48:30Z

This pull request was successfully merged by @hryhoriiK97 in 9b0200e

^{When will my fix make it into a release? | How to file a pick request?}

## Summary: Native modules send events to JS with `RCTEventEmitter sendEventWithName:body:`, which ends in `RCTInstance -callFunctionOnJSModule:method:args:`. Core modules use it too: for example, `RCTNetworking` sends the whole text of an XHR response in `didReceiveNetworkData`, and `RCTWebSocketModule` sends each message. **Current behavior.** The arguments go through `folly::dynamic`. On the sending thread, `convertIdToFollyDynamic` encodes every string as UTF-8. On the JS thread, `jsi::valueFromDynamic` decodes it again into a JS string. Text that isn't ASCII is the expensive case: a 1 MB response with accented text costs about 3.9 ms on the JS thread and 3.2 ms on the sending thread. **With this change.** - On the sending thread, `RCTInstance` reads the arguments into a small tree with the structure that `convertIdToFollyDynamic` creates. Strings keep their own ASCII or UTF-16 characters instead of being encoded as UTF-8. When CoreFoundation exposes the characters, the tree points at them in an immutable copy of the string; otherwise it copies them. - On the JS thread, the JS values are created from the tree, without Objective-C calls, the same way `valueFromDynamic` does. - `ReactInstance` gets a `callFunctionOnModule` overload that takes an `ArgFactory`, which creates the arguments on the JS thread. It has the same shape as `TurboModule::ArgFactory`, which `emitDeviceEvent` uses. Both overloads share the module lookup, so Android and other callers of the `folly::dynamic` overload run the same code as before. JS receives the same values as before: - Numbers use the same rules as `convertIdToFollyDynamic`. - Number keys become strings, and other non-string keys are skipped. - Values of unsupported types become `null`. - Strings with a lone surrogate, which UTF-8 can't encode, keep the UTF-8 path, so they still arrive as `""`. The order of properties in event objects now follows the `NSDictionary`. Before, it followed the `folly::dynamic` hash map, so it was arbitrary in both cases. On JSC, a string with both a NUL and non-ASCII characters is no longer cut off at the NUL, as `JSStringCreateWithUTF8CString` did. Hermes output is the same. ## Changelog: [IOS] [CHANGED] - Events sent by native modules are converted to JS values without a UTF-8 round trip ## Test Plan: - `clang-format --dry-run --Werror` on the changed files passes. - RNTester built with `xcodebuild` in Release and run on a physical iPhone 15 Pro (iOS 26.6.1). - Added `ReactInstanceTest.testCallFunctionOnModule_withArgFactory` and `testCallFunctionOnModule_withEmptyArgFactory` for the new overload. I couldn't run the C++ tests locally. I updated the C++ API snapshots by hand, because I couldn't run Doxygen here. - **Same output.** - A test payload was sent through the old and the new conversion in the same app run, and JS compared the results deeply. They were identical. The payload had: - empty, NUL, BOM, emoji, accented, CJK and lone-surrogate strings, and a mutable string changed after sending; - `BOOL`, `char`, `unsigned char`, `short`, `int`, negative, `unsigned long long`, 2^60, `float`, `double`, NaN and `NSDecimalNumber` numbers; - `NSNull`, `NSDate` and `NSURL` values; - nested arrays and dictionaries; - number, bool and `NSNull` keys. - Every XHR response text in the measurement below matched the original string. - **Measurement.** RNTester Playground, XHR `GET` of `data:` URLs with ASCII or accented text: 1000 requests at 1 KB, 30 at 100 KB and 10 at 1 MB. Also 1000 small events sent from a background queue: a keyboard-like payload, and one with short keys and values created at run time. - Timed on the JS thread: creating the JS arguments, `valueFromDynamic` before and the tree conversion after. - Timed on the sending thread: `convertIdToFollyDynamic` before and building the tree after. - The two arms are the same binary with the conversion switched by a launch argument, in the order before, after, after, before, 4 launches per arm. The presses were driven by [Argent](https://github.com/software-mansion/argent) on the physical iPhone. - iPhone 15 Pro, Release, medians per event. The ranges of the two arms are separate in every row. | Event | JS thread before | after | change | Sending thread before | after | change | |---|--:|--:|--:|--:|--:|--:| | `didReceiveNetworkData`, accented, 1 KB | 4.82 µs | 0.36 µs | −92% | 4.11 µs | 0.38 µs | −91% | | `didReceiveNetworkData`, accented, 100 KB | 382 µs | 8.6 µs | −98% | 311 µs | 3.1 µs | −99% | | `didReceiveNetworkData`, accented, 1 MB | 3.87 ms | 0.09 ms | −98% | 3.18 ms | 0.03 ms | −99% | | `didReceiveNetworkData`, ASCII, 1 KB | 0.65 µs | 0.34 µs | −47% | 0.97 µs | 0.38 µs | −61% | | `didReceiveNetworkData`, ASCII, 100 KB | 12.1 µs | 7.8 µs | −36% | 5.8 µs | 0.35 µs | −94% | | `didReceiveNetworkData`, ASCII, 1 MB | 97.6 µs | 71.8 µs | −26% | 49.3 µs | 0.33 µs | −99% | | `didReceiveNetworkResponse` (headers) | 1.08 µs | 0.74 µs | −31% | 2.45 µs | 0.78 µs | −68% | | keyboard-like event | 1.79 µs | 1.17 µs | −35% | 5.83 µs | 2.28 µs | −61% | | short strings created at run time | 1.04 µs | 0.61 µs | −41% | 2.58 µs | 0.79 µs | −70% | 🤖 Generated with [Claude Code](https://claude.com/claude-code) Differential Revision: D127583929 Pull Request resolved: react#59449

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 24, 2026

facebook-github-tools Bot added p: Software Mansion Partner: Software Mansion Partner labels Oct 24, 2026

facebook-github-tools Bot added the Shared with Meta Applied via automation to indicate that an Issue or Pull Request has been shared with the team. label Oct 24, 2026

javache requested changes Oct 26, 2026

View reviewed changes

javache mentioned this pull request Oct 26, 2026

Convert iOS native module objects to JS with fewer lookups and retains #59450

Merged

hryhoriiK97 requested a review from javache October 26, 2026 12:47

javache reviewed Oct 26, 2026

View reviewed changes

javache approved these changes Oct 26, 2026

View reviewed changes

hryhoriiK97 and others added 4 commits October 29, 2026 12:55

Skip calls without a module or method name, and reserve the arguments

5106a80

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

hryhoriiK97 force-pushed the ios-event-arguments-without-dynamic branch from f05f675 to 506ede7 Compare October 29, 2026 11:57

christophpurrer requested changes Oct 30, 2026

View reviewed changes

javache reviewed Oct 30, 2026

View reviewed changes

Bail out before converting nil module or method names

bad4fb2

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

javache reviewed Nov 13, 2026

View reviewed changes

Hold number keys in a strong key variable

22dbf66

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

javache approved these changes Nov 13, 2026

View reviewed changes

meta-codesync Bot merged commit 9b0200e into react:main Nov 13, 2026
93 of 95 checks passed

react-native-bot added the Merged This PR has been merged. label Nov 13, 2026

Conversation

hryhoriiK97 commented Oct 24, 2026

Summary:

Changelog:

Test Plan:

Uh oh!

hryhoriiK97 commented Oct 24, 2026

Uh oh!

github-actions Bot commented Oct 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Visual regression

ios button/default-styling/default: Changed, 38223 px (1.533%)

Uh oh!

hryhoriiK97 commented Oct 25, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

meta-codesync Bot commented Oct 26, 2026

Uh oh!

hryhoriiK97 commented Oct 29, 2026

Uh oh!

christophpurrer left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

hryhoriiK97 commented Nov 5, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

hryhoriiK97 Nov 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

react-native-bot commented Nov 13, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

github-actions Bot commented Oct 24, 2026 •

edited

Loading

ios `button/default-styling/default`: Changed, 38223 px (1.533%)

hryhoriiK97 Nov 13, 2026 •

edited

Loading