Summary:
Pull Request resolved: https://github.com/facebook/react-native/pull/41956
By default, generated Cxx sources for components all end up in same directory. However the include declarations in them look like this:
```
#include <react/renderer/components/${libraryName}/ShadowNodes.h>
```
And not like this:
```
#include "ShadowNodes.h"
```
This works fine with Buck because it supports header prefixes.
To get this working with CocoaPods we define additional `HEADER_SEARCH_PATHS` for our `React-Codegen` pod.
This approach will not work if we want to generate code at the library level and check in the artifacts. That's because we don't have control over the Podspec there, and can't inject those additional `HEADER_SEARCH_PATHS`.
This diff adds the `headerPrefix` argument to the codegen entry point. It is `react/renderer/components/${libraryName}` by default, but can become empty if we want to generate code at the library level, and don't want to deal with this nested header structure.
*Note:* `RNCodegen` runs all the generators [in a loop](https://github.com/facebook/react-native/blob/main/packages/react-native-codegen/src/generators/RNCodegen.js#L263-L275), assuming that the all have same function signature So I had to add the `headerPrefix` argument to all the generators, even to the ones that don't really need it.
Changelog: [General][Added] - Introduce "headerPrefix" codegen option.
Reviewed By: zeyap
Differential Revision: D51811596
fbshipit-source-id: c5c3e1e571c7c4ea2f5354eb9a7b0df6b917fc0c
Summary:
Pull Request resolved: https://github.com/facebook/react-native/pull/41771
Changelog: [Internal]
Since we are already enforcing C++20 (and 17), we can set the namespace declaration to the C++17 style
Reviewed By: NickGerleman
Differential Revision: D51789991
fbshipit-source-id: 165d7d4e652d60ab200e2355e084010a02f470a4
Summary:
Pull Request resolved: https://github.com/facebook/react-native/pull/39136
## Changelog:
[Internal] -
`getConstants()` method for TM was enforced to only use object literals for the return type specs.
This limits flexibility, in particular those data structures can't be consequently exported and picked up by codegen (not even mentioning the potential need for copypasting those obejct literals around).
This relaxes this restriction.
Note that I've been digging into the development history in order to find out whether there was any particular historical reason for such a limitation, but couldn't find any, so I assume it was rather incidental.
Reviewed By: christophpurrer
Differential Revision: D48620652
fbshipit-source-id: 92d6ba531fc99fb9b25b4957ae123e7832f44ee4
Summary:
Handling of `EnumDeclaration` was introduced in D38967241 (https://github.com/facebook/react-native/commit/745f3ee8c571560406629bc7af3cf4914ef1b211) so it is no longer a type expected to fail generators.
Changelog: [Internal] remove 'EnumDeclaration' as a type expected to throw error in module generators
Reviewed By: cipolleschi
Differential Revision: D42917947
fbshipit-source-id: 16fcb915ccd42c613ca4d30b815d6365681f5fa1
Summary:
Pull Request resolved: https://github.com/facebook/react-native/pull/36042
Pull Request resolved: https://github.com/facebook/react-native/pull/36028
Use aliasMap both in parsers and generators instead of using aliasMap in half of the places and aliases in the other.
This would also allow us to introduce "enumMap" more easily in the next commit.
Changelog: [Internal] rename module exports from "aliases" to "aliasMap"
Reviewed By: christophpurrer, cipolleschi
Differential Revision: D42888752
fbshipit-source-id: cf1929fcebde994d07e5c6bda5ab71106d417b21
Summary:
This PR is a task from https://github.com/facebook/react-native/issues/34872
_Extract the modules/utils.js from the flow and typescript folders in a shared parsers-commons.js file. Then, have the two parsers use the same wrapModuleSchema function for modules._
(`wrapModuleSchema` is a copy-paste mistake, in this case it is `wrapNullable` and `unwrapNullable`)
## Changelog
<!-- Help reviewers and the release process by writing your own changelog entry. For an example, see:
https://reactnative.dev/contributing/changelogs-in-pull-requests
-->
[Internal] [Changed] - Extracts Codegen's modules/utils.js files from the flow and typescript folders in parsers-commons
Pull Request resolved: https://github.com/facebook/react-native/pull/34898
Test Plan:
I ran `yarn jest react-native-codegen`:
<img width="775" alt="Capture d’écran 2022-10-07 à 21 29 48" src="https://user-images.githubusercontent.com/17070498/194639515-a446c2cf-daf3-43a1-9833-cd546ca5865e.png">
Reviewed By: cipolleschi
Differential Revision: D40193740
Pulled By: cipolleschi
fbshipit-source-id: 02cbacc215fe5dd9bdd0839d8796587ab2821906
Summary:
There are cases where we want to pass an enum type into a TurboModule method which is handy to restrict certain arguments to a restricted set of values, e.g. restricting quality to ```enum Quality { SD, HD, }```
Approach:
- We are not generating an ```enum``` type in C++ but rather just cast type safe to the corresponding member type.
- We don't support mixed enum types at this time, e.g. ```export enum StringOption { One = 'one', Two = 2, Three = 'three', };``` will not work. See: https://www.typescriptlang.org/docs/handbook/enums.html#heterogeneous-enums
- We only support untyped (default to String), String, and Number enum properties
This is for C++ only, Java and ObjC are not supported atm.
Changelog:
[General][Added] - react-native-code-gen Add Enum Type support for C++ TurboModules
Reviewed By: RSNara
Differential Revision: D38880963
fbshipit-source-id: f2399b29948306bc555429b6f96c43ea4c39c46e
Summary:
There are cases where we want to pass a union type into a TurboModule method which is handy to restrict certain arguments to a restricted set of values, e.g. restricting quality to ```'low'```, ```high``` or resolution to ```720```, ```1080```.
- We are not generating an ```union``` type in C++ but rather just cast type safe to the corresponding member type.
- We don't support mixed primitive union types at this time, e.g. ```export type ChooseSomething = 'One' | 'Two' | 3 | false;``` does not work.
- We can support mixed object union types such as ... ```export type ChooseObject = {} | {low: string};``` - which need special logic in the C++ TM to correctly parse the resulting jsi::Object
This is for C++ only, Java and ObjC are not supported atm.
Changelog:
[General][Added] - react-native-code-gen Add Union Type support for C++ TurboModules
Reviewed By: javache
Differential Revision: D38919688
fbshipit-source-id: 0fd37545b32b4f2059a8babda62dab4a85de37a9
Summary:
Sometime over the past few months (and with changes such as migrating to the `hermes-eslint` parser), a bunch of lint warnings crept into the codebase.
This does a pass to clean them all up, ignore generated files, and refactor some code to be... better.
There should be no observable behavior changes as a result of this.
Changelog:
[Internal]
Reviewed By: NickGerleman
Differential Revision: D38646643
fbshipit-source-id: a7b55d1e4cd5700340cc5c21f928baf3ea1d5a58
Summary:
There are cases where we want to pass arbitrary types to a TurboModule, which may then handle the values appropriately, but we haven't supported this use case. Since C++ TurboModules can accept any `jsi::Value` (unlike Java/ObjC) and we have real-world need for this (otherwise we must require JSON serialization), this now allows `mixed` (`unknown` in TypeScript) for C++-only TurboModules.
Changelog:
[General][Added] C++ TurboModule methods can now use mixed types
Reviewed By: RSNara
Differential Revision: D36611299
fbshipit-source-id: bbf29dfcc6aed67e213bb3eab06537c18c7db1fe
Summary:
Pull Request resolved: https://github.com/facebook/react-native/pull/32840Closes#31516
I've cherry-picked the original PR that had merge conficts + updated all
the headers as the one for the TurboModule generators were not handled.
Original Commit Message from acoates
The codegen generates a Facebook copyright notice at the top of the generated files.
While this might make sense on the core files, this codegen will be run on external components too.
The notice also refers to a LICENSE file in the root of this project, which might not be there if this is run on another project.
I did a quick look at some of the codegen that we ship within windows dev tools, and it looks like we normally just have comments
saying the file was codegen'd and so the file shouldn't be manually edited.
Open to suggestions on what the comment header should say.
Changelog:
[General] [Changed] - Do not include Facebook license on users codegen'd code
Reviewed By: ShikaSD
Differential Revision: D33455176
fbshipit-source-id: b247e72efb242e79d99b388c80e4126633e5234d
Summary:
This diff runs the codemod to add type annotations to function parameters in preparation for Flow's local type inference (LTI) project. I ran the codemod over xplat/js and reverted any files that had flow errors in them. See the list of commands run to see the regeneration of various files.
Changelog:
[Internal][Changed] - Added type annotations
Reviewed By: yungsters
Differential Revision: D32075270
fbshipit-source-id: 6a9cd85aab120b4d9e690bac142a415525dbf298
Summary:
This pre-suppresses the 154 error diff ahead of its release, since it is large.
Changelog: [Internal]
Reviewed By: samwgoldman
Differential Revision: D29065246
fbshipit-source-id: f418041305a46df410dcbe3d9a4db81a61ac7014
Summary:
Pull Request resolved: https://github.com/facebook/react-native/pull/31543
Changelog:
[iOS][Added] - Description
When compiling iOS apps with flag `-Wnullability-completeness` (like Lightspeed app and soon Instagram), Objective-C headers are required to either have full *explicit* nullability annotations on all members of its public API, or none at all; partially annotated headers will fail to build that module.
RN native modules are currently generated with *partial* annotations. This works today because most apps are not compiled with `-Wnullability-completeness` turned on. But when we flip the switch for Instagram, the app doesn't build due to importing these RN partially annotated modules.
JavsScript Flow types are implied nonnull, and the current RN codegen translates Flow's [maybe/optional](https://flow.org/en/docs/types/maybe/) type to Obj-C `_Nullable` annotation, and everything else without an explicit Obj-C annotation. However this creates a mismatch with the Obj-C type system, where the implied default is *unannotated*, which is handled differently from nonnull when built with the nullability compiler flags.
There is a simple Obj-C macro that automatically adds *explicit nonnull* annotations to all members in a header: `NS_ASSUME_NONNULL_BEGIN` / `NS_ASSUME_NONNULL_END`. If we add this to *all* RN-generated headers, however, we run into issues:
1) We may erroneously assume any previously-unannotated header was meant to be nonnull and cause future bugs
2) Another compiler flag (`-Wnullable-to-nonnull-conversion`) statically analyzes Obj-C implementation code to prevent us from ever passing null to one of these headers. Much existing Obj-C code will break here, and it's ambiguous if these are true or false positives because of the first point.
Instead, in this diff we add a new BUCK flag `ios_assume_nonnull` to let module authors opt into automatic nonnull for unannotated members so that Obj-C headers are generated correctly in alignment with Flow's type system. We can migrate all libraries individually as needed and eventually make this the RN native codegen default.
Reviewed By: RSNara
Differential Revision: D28396446
fbshipit-source-id: ad3a3a97ab19183df4ef504b1c3140596c8f69ca
Summary:
## What This Does
- **Phase 1:** Given a component, convert its props (and all its ObjectTypeAnnotations) into this Pojo data structure under some namespace. This is a recursive operation.
- **Phase 2:** Loop over each Pojo data structure, and serialize it to Java class.
So...
- Each Component has its own namespace (i.e: its hasteModuleName) for Java Pojo objects.
- Each Component generates 1 Pojo object, for its props.
- Each Component generates 1 Pojo object for every ObjectTypeAnnotation in its props.
## Decisions
By design, [JNI can read/write to private properties on Java objects](https://stackoverflow.com/a/12208643). So, each Pojo, which represents an ObjectTypeAnnotation, contains only a private member variable for each of its properties, and a getter to retrieve the data from each of its private members.
## Todos
- Improved type-safety:
- ReservedTypeAnnotation (e.g: PointPrimitive). These currently map to ReadableMap.
- String enums, and Int enums don't actually generate Java enums.
- Verify if there are any compilation issues by wiring this up to the Codegen buck infra.
- To actually use the Pojos, we'll need C++/Jni codegen to transform C++ props into these Pojos. Building this out will give a more accurate assessment of the app-size cost of Pojos. However, we can do that if we deem that the app-size increase from adding just the Pojo classes is negligible enough to **not** rule out this entire approach.
- ~~Add some annotations to prevent these Pojo classes from being stripped at compile-time.~~ D26041103
Changelog: [Internal]
Reviewed By: mdvacca
Differential Revision: D26038189
fbshipit-source-id: c137c4ca6b043ee76adb354105aff6e0f270df86
Summary:
## Problem
Suppose we have this NativeModule spec, that uses Object literals that contain a property called "id":
```
export interface Spec extends TurboModule {
// Exported methods.
+getConstants: () => {|
id: string,
|};
+getObject: (arg: {id: Object}) => Object;
}
```
For both object literals, we'll generate C++ structs, backed by NSDictionaries. The method in each struct will be named "id_" to avoid a name clash with ObjC's `id` identifier. However, calling that id_ method should still access the "id" property in the corresponding NSDictionary.
## Expected Output
```
inline id<NSObject> JS::NativeSampleTurboModule::SpecGetObjectArg::id_() const
{
id const p = _v[@"id"];
return p;
}
inline JS::NativeSampleTurboModule::Constants::Builder::Builder(const Input i) : _factory(^{
NSMutableDictionary *d = [NSMutableDictionary new];
auto id_ = i.id_.get();
d[@"id"] = id_;
return d;
}) {}
```
## Actual Output
```
inline id<NSObject> JS::NativeSampleTurboModule::SpecGetObjectArg::id_() const
{
id const p = _v[@"id_"]; // <-- HERE!
return p;
}
inline JS::NativeSampleTurboModule::Constants::Builder::Builder(const Input i) : _factory(^{
NSMutableDictionary *d = [NSMutableDictionary new];
auto id_ = i.id_.get();
d[@"id_"] = id_; // <-- HERE!
return d;
}) {}
```
NOTE: This code was generated by running `jf get --version 119805822 && buck build //xplat/js:FBReactNativeSpec_Sample-flow-types-ios --show-output`
This diff fixes this mistake.
Changelog: [Internal]
Reviewed By: fkgozali
Differential Revision: D25907493
fbshipit-source-id: cb37cbf49db4f871b3f4046f7397a7b1b7df0357
Summary:
Changelog: [internal]
Codegen was generating code with return value number instead of boolean.
Reviewed By: RSNara
Differential Revision: D25863062
fbshipit-source-id: 780f88dd2d83e303b03d1ed9cc837ac6733f1702
Summary:
## Changes
{| ... |} -> { ... }
**Motivation:** In Flow, object literals are exact by default. So, there's no need for the pipes. Also: Now, the syntax for object literals is consistent across react-native-codegen.
Changelog: [Internal]
Reviewed By: hramos
Differential Revision: D24774771
fbshipit-source-id: 24ceb6f5876122aa8ad9e08c7e903215864ad6f5
Summary:
Everywhere else in the CodegenSchema, type annotation partials are suffixed with "Shape". In the NativeModule schema, we were using the suffix "Schema". In this diff, we standardize on the "Shape" suffix.
Changelog: [Internal]
Reviewed By: yungsters
Differential Revision: D24719395
fbshipit-source-id: 307935f5fe0681c31cd52e9cf4ae579f61c1ae68
Summary:
CodegenSchema exports `NativeModuleMethodParamSchema` and `NativeModuleObjectTypeAnnotationPropertySchema`, which are partials of NativeModule type annotations. This creates unnecessary coupling between the type annotations of CodegenSchema and the files that depend on it.
**Actual Problem:** Suppose that we want to rename one of these partials. Then, all imports in all files would have to be updated, even when the actual shape of the composed type annotation wasn't changed.
This diff removes these partials, which reduces the surface area of the exports of CodegenSchema.js
Changelog: [Internal]
Reviewed By: yungsters
Differential Revision: D24719396
fbshipit-source-id: c822aaa252f156c524f4ef4917ebb61b1a39ff9e
Summary:
Reserved type annotations can appear in three different contexts: commands, props, and NativeModules. For now, commands and NativeModules share the same reserved type annotations. In the future, we may want to merge these reserved type annotations with the props reserved type annotations.
**Motivation:** The meaning of FunctionValue in FunctionValueTypeAnnotation isn't clear - in fact, it's downright confusing. Therefore, this diff renames this Flow type to ReservedTypeAnnotation, which I believe sufficiently captures the intent of the type annotation.
Changelog: [Internal]
Reviewed By: yungsters
Differential Revision: D24701322
fbshipit-source-id: bde0273b4a89c9e7175c60ed3468ed870b320044
Summary:
For now, separate the definition of `modules` generator per platform to avoid file output collision. Additionally:
* For Android, produce files under java/ (plus nested subdirs based on packageName) and jni/ (for C++ files) - JavaPoet version already does it
* Allow configuring packageName for Android - JavaPoet version has this
* Avoid tmp directory dance in the CLI script, given the proper modules separation
Changelog: [Internal]
Reviewed By: hramos
Differential Revision: D24410864
fbshipit-source-id: 9bd6bc1d65bec037bfca32ec478f3af50d72e927
Summary:
In our Codegen generators, we were using `codegenModuleName` to refer to the name of the spec file. Calling this `hasteModuleName` makes it more clear what this name refers to (i.e: the name of the spec file). This diff performs that rename.
Changelog: [Internal]
Reviewed By: fkgozali
Differential Revision: D24386282
fbshipit-source-id: fe2beda9a0abf63a5cf88fa0664f83416c9f1aa2
Summary:
If a native module schema has `excludedPlatforms` defined, honor it and skip the module that doesn't belong to the platform.
E.g. NativeImagePickerIOS shouldn't generate anything for Android codegen output.
Similarly, IntentAndroid shouldn't generate anything for iOS codegen output.
Changelog: [Internal]
Reviewed By: RSNara
Differential Revision: D24373092
fbshipit-source-id: cfeb455a18c92f60191d988af2e9ce7ea5021304
Summary:
NOTE: Flow and Jest won't pass on this diff. Sandcastle, should, however, be green on D24236405 (i.e: the tip of this stack).
## Changes
1. NativeModule generators now use the new RN Codegen NativeModule schema.
2. Tangential: We're no longer removing the `Native` prefix from the NativeModule filename, assuming that that's the module name (problem: wrong), and prefixing again with Native (problem: redundant), when we're generating code. Instead, like the internal codegen, we simply pass the filename to the Codegen output. Our linters enforce that all NativeModule specs are contained with files that start off with `Native`.
3. `GenerateModuleCpp` was fixed to use the actual module name as opposed to the spec name. I added a comment inline.
Changelog: [Internal]
(Note: this ignores all push blocking failures!)
Reviewed By: PeteTheHeat
Differential Revision: D24236405
fbshipit-source-id: ccd6b5674d252c350be0ec8a86e7ca5f2f614778
Summary:
Remove extraneous newlines before and after structs, before copyright headers, and other locations.
Changelog: [Internal]
Reviewed By: fkgozali
Differential Revision: D24117137
fbshipit-source-id: 194996019b4cadef9239a78334f31c0bc89e3901
Summary:
Fixes an issue where, if a spec uses an array of elements where the element is a type alias, the generated code would use the wrong variable name.
An example of such a spec can be found in `NativeExceptionsManager.js`:
```
+reportSoftException: (
message: string,
stack: Array<StackFrame>,
exceptionId: number,
) => void;
```
The fix ensures the local variable name is passed through, ensuring that either p or itemValue_N is used when appropriate.
Changelog: [Internal]
Reviewed By: RSNara
Differential Revision: D24116429
fbshipit-source-id: f39b75adb604c751d70a284a11a7fa6649b1344d
Summary:
Adjust generated ObjC++ code to resolve a few build time and run time errors:
* Suppress CONSTANTS struct implementations
* Use type alias name as struct name when serializing arguments that involve a type alias
* Use actual number of arguments for a method when generating method map.
With these changes in place, RNTester can be built and run using the code that is generated by the new codegen.
Changelog: [Internal]
Reviewed By: hramos
Differential Revision: D23926500
fbshipit-source-id: 88fcbb795fd71dc8155eb26348db943975e13e84
Summary:
* Removed extraneous closing brace.
* Fixed static method signature, replacing double colon with an underscore (`static facebook::jsi::Value __hostFunction_Native${moduleName}SpecJSI::${methodName}()` -> `static facebook::jsi::Value __hostFunction_Native${moduleName}SpecJSI_${methodName}()`).
* Wrap `getConstants` selector name with `selector()`.
* Pass through `getConstants` and `constantsToExport` to allow de-duping of `getConstants` method in generator output.
Note that the FBReactNativeSpec that is output by the generator still has some issues that need to be addressed before it can be used.
Changelog: [Internal]
Reviewed By: fkgozali
Differential Revision: D23910505
fbshipit-source-id: 37d884885b8878f38d40637377c2a74a728c3a13
Summary:
## Misc. Improvements
* We now have 95%+ flow coverage in all generator files. Henceforth, we can make changes to these files with more confidence, and trust flow to catch more errors. This should also improve the DevX of working on these files.
* Better templates: Instead of doing string replace with RegExps, we instead use functions and leverage JS template literals to generate our code. A few benefits: (1) data dependencies of templates are clearly visible, and statically checked by flow, (2) the templates are more readable in VSCode.
* Merged the GenerateModuleHObjCpp.js and GenerateModuleMm.js generators. They can share a lot of logic, so it's not a good idea to keep them separate.
* The ObjC++ module generator no longer generates “dead” structs (i.e structs that aren’t used by type-safety infra). In fact, it explicitly only supports the types in our Wiki. (I know this wasn’t the case with the legacy codegen, because we were generating native code for enums in the legacy codegen). This is a mixed bag. The test to verify correctness will be more difficult to write. However, keeping structs in the codegen needlessly complicates the parsers + generators, and creates technical debt for us to clean up later.
## Abstractions
- **StructCollector:** As we serialize NativeModule methods, when we detect an ObjectTypeAnnotation in the return type of `getConstants()` or inside a method param, we must create a Struct JS object for it. When we detect a type-alias (also in the same locations), we must look up that type-alias and create a Struct from its RHS. A Struct is basically an ObjectTypeAnnotation with a context (i.e: used in getConstants() vs as a method param), that cannot contain other ObjectTypeAnnotations.
- **serializeMethod.js** Given a NativeModule method type annotation, output the protocol method, JS return type, selector, a record of which params were structs, and which structs. Basically, this is all the information necessary to generate the declaration and implementation codegen for a partiular NativeModule method.
- **serializeStruct/*.js**: After creating all these Structs, we need to loop over all of them, and tranform them into ObjC++ code.
- **serializeStruct.js**: Depending on the struct context, calls either `serializeRegularStruct.js` or `serializeConstantsStruct.js`. Both of these files have the same layout/abstractions. They look very similar.
- **serializeModule.js:** Outputs RCTCxxConvert categories for transforming `NSDictionary *` into C++ structs. Outputs ObjCTurboModule subclass.
## Algorithm
```
for spec in NativeModuleSpecs
structCollector = new StructCollector
resolveAlias = (aliasName) => nullthrows(spec.aliases[aliasName])
methodDatas = []
for method in methods(spec)
methodData.push(serializeMethod(method, structCollector, resolveAlias))
end
structs = structCollector.getStructs()
output generateImplCodegen(methodDatas, structs)
output generateHeaderCodegen(methodDatas, structs)
end
```
Changelog: [Internal]
Reviewed By: hramos
Differential Revision: D23633940
fbshipit-source-id: 7c29f458b65434f4865ef1993061b0f0dc7d04ce
Dmitry Rykun
Christoph Purrer
Moti Zilberman
Ruslan Shestopalyuk
Vitali Zaidman
Mike Vitousek
Antoine Doubovetzky
Christoph Purrer
Tim Yung
Scott Kyle
Erich Graham
Nicola Corti
Andres Suarez
Evan Yeung
Daniel Sainati
Ramanpreet Nara
Samuel Susla
Kevin Gozali
Héctor Ramos
