[robhogan]

Motivation

• Dev/release consistency in correctness and performance shape.
• ~10% faster dev cold builds
• More responsive dev apps (~2x faster promise handling)
• Improved debugging support

Summary

Since #52651, Hermes profiles have lowered generators (with enableBabelRuntime: false), for...of, ?., ?? and optional catch binding in dev builds only, because Hermes couldn't show local variables when paused on a breakpoint in a native generator.

That turned out to be a Hermes bug: a breakpoint on that a yield line resolved to scope-less entry code before the body had even run. It was fixed in facebook/hermes@d1238c8 (hermes-v260318099.0.2), which main has used since #58397.

This removes enableRegenerator, so Hermes-profile dev output matches release. It also fixes a dev-only divergence - loose for-of never calls iterator.return(), so breaking out of a for...of over a generator skipped its finally in dev but not release.

Note that with the default enableBabelRuntime, generators were already native in dev, so in practice the generator change mostly affects enableBabelRuntime: false setups, including RN's own Jest. (There's no good reason generator was ever conditional on enableBabelRuntime === false - it seems to date from #44668 and probably a misunderstanding of what the regenerator argument to @babel/plugin-transform-runtime did - it doesn't make that plugin transform generators, it just rewrites the helper import.)

Native generator debugging still isn't perfect on .0.2 - stepping through try/catch can stop on a catch line that doesn't run, and breakpoints on yield* lines or on a throw after a resume don't fire. Those already affect every default config though, so there's no regression in most cases. Regenerator has its own problems, eg breakpoints often stop a line early because statements merge into its case blocks. Overall, native is now a better experience.

Detailed comparison in https://robhogan.github.io/reports/react-native-regenerator-removal/

Changelog

[GENERAL] [FIXED] - Babel preset: Hermes profiles no longer lower generators, for...of, optional chaining, nullish coalescing or optional catch binding in dev, matching release

Test Plan

Reproduced the original bug through Hermes's CDPAgent test harness with Debugger.setBreakpointByUrl. On hermes-v250829098.0.16 and v260318099.0.1, a breakpoint on a generator's last yield line pauses at the generator call with only the global scope, and every local evaluates to undefined. On v260318099.0.2 and .0.4 it pauses at the real code with all locals.

AI-driven breakpoint sweep: this preset with dev: true, enableBabelRuntime: false, before and after this change, over 248 source lines of generator, async function, async generator, exception and loop fixtures. A breakpoint was set on every line, each pause was mapped back through the source map, and every local was evaluated. Same results with eager and lazy compilation.

Hermes	Identical	Different, equivalent	Native better	Native worse	Mixed (loop headers)	Total
250829098.0.16	164	10	49	22	3	248
260318099.0.2	183	11	47	3	3	247
260318099.0.4	184	11	47	3	3	248

[robhogan]

Jest failures are legitimate due to changing some inlining behaviour - looking into that now. I might need to stack a fix under this to do it cleanly. Converting to draft for the moment.

[robhogan]

Prerequisite: #59307

[robhogan]

@javache (from #59307)

Will #59303 cause changes to our internal inline-requires transform?

Only in dev, and only to align with prod.

The difference is destructuring of CJS require:

const {foo, bar} = require('./stuff')

Is not inlined by inlineRequires, which runs after the RN preset. So with Babel lowered destructuring this was previously inlined and now isn't in release, since #58965.

It's still lowered and therefore inlined in dev && !enableBabelRuntime && enableRegenerator because lowered destructuring is a prerequisite of regenerator, dropped by this PR.

I didn't spot this in #58965. It's likely to have minimal impact at Meta because there's no change to ESM syntax, but probably a handful of 3P CJS libraries will have lost some inlining.

[robhogan]

I missed some snapshot failures for set-version-test from #59307. Added in a new commit here and also separately in #59352 if we want to land that first so this is clean.