Skip to content

Latest commit

 

History

History
368 lines (266 loc) · 12.9 KB

README.snapchat.md

File metadata and controls

368 lines (266 loc) · 12.9 KB

Transit / Djinni

Djinni

Djinni is a project originally created by Dropbox that generates bridging code between C++ and other programming languages. While Dropbox has stopped maintaining the open source Djinni repository on Github, we at Snapchat have developed our internal fork since then. After contacting Dropbox, we are now officially taking over the maintenance of Djinni from Dropbox and will keep developing it here.

This file only covers the parts that have been changed. Please see the original dropbox readme for the full Djinni documentation.

Building

Bazel Build

Both the Djinni code generator and test suite are built with Bazel. You can use either plain Bazel or Bazelisk.

Building and running the test suite

./ci/generate.sh generates the examples sources.

Use bazel test //test-suite:djinni-objc-tests //test-suite:djinni-java-tests to build and run Objective-C and Java tests.

Building and running the mobile example apps

The Android example app can be build with bazel: bazel build //examples:android-app, and then install to a device with adb install bazel-bin/examples/android-app.apk

The iOS example app are built with Xcode. Simply open the project in Xcode and it should work.

Working on the Djinni code generator

You can load the project via Bazel

  • Intellij with the Bazel and Scala plugin is required
  • Configure the bazel binary. If you use Bazelisk, set it as the binary in the IDEA bazel settings.
  • In Intellij, import a new Bazel project.
    • Workspace directory: /Users/$HOME/path-to-djinni-directory
    • Import project view file: WORKSPACE/bzl/ide/djinni.bazelproject
  • Similarly you can also use CLion if you wish to edit the C++ code
    • You can set up any of the cc_* targets after importing the workspace.

You can also use bazel build //src:djinni and bazel run //src:djinni to verify the build and binary from the command line.

Modifications

  • Replaced sbt and gyp with Bazel
  • Added move assigment operator to GlobalRef in all djinni_support.hpp files
  • Made JniClassInitializer constructor public to allow arbitrary additional initialization in JNI_OnLoad()
  • Speed up string passing between Java and C++ (about 5-10x faster depending on string size)
  • Eliminate CppProxy finalizers (better stability)
  • Injecting code with DJINNI_FUNCTION_PROLOGUE
  • Option to generate ObjC protocols
  • Option to generate function prologue
  • Option to omit objc helper methods
  • Option to disable exception translation in ObjC
  • array<> type support
  • outcome<> type support
  • Protobuf type support
  • Local flags with @flag directive
  • DataView for copy free data passing
  • DateRef for copy free data passing with ownership
  • Generating string names for C++ enums
  • Bug fixes

Using new features

Injecting code to Djinni calls

Two new switches are introduced:

--jni-function-prologue-file "path/to/header_file.h"
--objcpp-function-prologue-file "path/to/header_file.h"

If they are specified at the command line, then all generated Djinni calls will contain a macro at the beginning of the bridge function:

DJINNI_FUNCTION_PROLOGUE("DjinniInterfaceName.methodName");

You should then #define the DJINNI_FUNCTION_PROLOGUE macro in the specified header file to inject whatever code you need there.

For example you could make it log the method name. Or you could instantiate a scoped object to trace the duration of the call.

Generate all interfaces as ObjC protocols

By default, Djinni only generates interfaces as ObjC @protocol when these interfaces are tagged with +o. Otherwise they become @interface in ObjC.

A new switch is introduced to override the default:

--objc-gen-protocol true

Enabling this will make all Djinni interfaces @protocol in ObjC. This means you can build your code without linking with the native code implementations, for example, when you want to mock the interfaces in pure ObjC for testing.

Prohibit some ObjC helper methods to minimize binary size

You can use --objc-disable-class-ctor true to remove the static constructors from Djinni generated classes.

You can also remove the description method in Djinni generated classes by defining the DJINNI_DISABLE_DESCRIPTION_METHODS preprocessor symbol.

These help reducing the binary size of the app slightly.

Disable C++ exception translation in ObjC

You can use --objcpp-disable-exception-translation true to disable the translation of C++ exceptions to ObjC. This means ObjC code will see C++ exceptions directly.

This is sometimes preferrable compared to the default behaviour which converts C++ exceptions to ObjC NSException, because the original exception stack is preserved.

outcome<> builtin type

The outcome<> type provides a standard way to explicitly define error conditions in Djinni IDL. We believe this is a superior approach than throwing exceptions across language boundaries.

An outcome is defined by combining a "result" type and an "error" type:

outcome<RESULT, ERROR>

Where RESULT and ERROR can be any Djinni types (primitives or records).

In C++, the outcome<> type maps to the template class djinni::expected<>. In Java, it maps to the generic class com.snapchat.djinni.Outcome<>. In ObjC, it maps to the generic class DJOutcome<>.

Example:

Use @extern directive to include support of outcome type to your idl file.

@extern "../djinni/support-lib/outcome.yaml"

my_interface = interface +c {
    query(): outcome<string, i32>;
}

Use Protobuf types in Djinni

We support using Protobuf types directly in Djinni IDL. In order to use Protobuf types, you first need to declare them in a yaml manifest file, in which we specify

  • How to find the header file and namespace/class/prefix
  • List the Protobuf messages you want to use in Djinni
cpp:
    header: '"proto/cpp/test.pb.h"'
    namespace: 'djinni::test'
java:
    class: 'djinni.test.Test'
objc:
    header: '"proto/objc/test.pbobjc.h"'
    prefix: 'DJTest'
 
messages:
    - Person
    - AddressBook

After that you can reference it in Djinni IDL and use the types as if they are regular Djinni types:

@protobuf "my_proto.yaml"
...
MyRecordWithProto = record {
    // Yay! proto!
    person: Person;
}

MyInterface = interface +c {
   // Yay! proto!
    static foo(x: AddressBook);
}

Optimize primitive array with array<>

The new array<> type is similar to list<>, but optimized for Java primitive arrays. It is mapped to regular Java arrays (T[]) instead of ArrayList. Since it doesn't have to box elements as objects, it's order of magnitude faster than list<> when the element type is primitive.

array<> is identical to list<> in Objective-C because there is no managed primitive array in Objective-C.

Local flags with @flag directive

In addition to supplying switches on the Djinni command line, it's also possible to specify them in the idl file with the @flag directive.

For example, adding the line

@flag "--objc-gen-protocol true"

to the djinni idl file will enable generation of objc protocols for interfaces. This is equivalent to adding the same switches to the djinni command line.

  • Multiple @flag lines can appear in the same file.
  • @flag lines must appear at the top of the djinni idl file before anything else.
  • @imported files can include @flag lines too, and they will apply to the parent file. (this means you can put common flags in a file and @import it)

Exposing data across language boundary with DataView

The builtin binary type always copies data across language boundary. It is simple to use but can be expensive when the size of the data is large or when the same data is passed back and force many times.

The DataView type is a viewport into a buffer owned by the other language. It maps to the custom DataView class in C++. In Java it is mapped to java.nio.ByteBuffer (direct buffer). And in Objective-C, it is mapped to NSData.

It is important to keep in mind that DataView does not carry ownership. If the underlying data is destroyed by the owner, then it will point to an invalid memory location. It is your responsibility to make sure the buffer remains valid while you use it.

Use DataRef to pass data across language boundary with ownership

DataRef is a buffer in Java or Objective-C heap but accessible from C++. It is also mapped to java.nio.ByteBuffer and NSData. But unlike DataView, it owns the buffer so destroying the object on one side of the language boundary does not destroy the underlying buffer as long as the other side still holds it.

Since the underlying data object is in Java or Object-C, it is more expensive to create than DataView. But the ownership allows you to hold on to the object without worrying about memory safety.

DataRef has a special optimization to take over the data from std::vector<uint8_t> and std::string. If you construct a DataRef with an R-value reference of these types, then DataRef can steal the buffer from them without copying the bytes.

String names for C++ enums

Djinni now generates a to_string() function that you can use to convert C++ enums to their string names. This makes printing enums in debugging traces a lot more convenient. This function is constexpr so if you call it with an enum value known to the compiler at compile time, then there is no runtime overhead. You may also call it with a dynamic value, in that case it's a fast array indexing operation.

Experimantal WASM support

Djinni can generate code that bridges C++ (that compiles to Web Assembly) and Javascript/TypeScript in web browsers. This feature is currently experimental.

For WASM, Djinni generates:

  • C++ code, which should be compiled into the WASM bindary
  • TypeScript code, provides optional TypeScript interface definitions

The generated code can be used with both plain Javascript and TypeScript.

New command line switches:

  • --wasm-out: wasm bridge code output folder
  • --wasm-include-prefix: path prefix to be added to include lines for WASM bridge class header files
  • --wasm-include-cpp-prefix: path prefix to be added to include lines for main C++ class header files
  • --wasm-base-lib-include-prefix: path prefix to be added to djinni support library inlcude lines in generated files
  • --ts-out: typescript output folder
  • --ts-module: typescript module(file) name

Almost all Djinni features are supported, including importing external types via yaml and zero-copy buffer passing.

Notable differences when comparing to the Java/ObjC support:

  • deriving(ord, eq) is not applicable to Javascript because JS doesn't support overloading standard comparison methods.
  • Extended records generate the same code as regular records. Because JS can easily add extension methods (by add functions to prototype) without having to derive from a base class.

The command to run Wasm/TypeScript unit tests is bazel run //test-suite:server-ts. You will need the tsc compiler and the browserify tool to run these tests.

Experimental async interface support

With the new yaml type future<> we can now write Djinni interfaces that return results asynchronously more easily.

Instead of writing this:

FooCb = interface {
  complete(res: i32): void;
}
Foo = interface {
  bar(cb: FooCb): void;
}

We can now write:

(use @extern directive to include support of future type to your idl file)

@extern "../djinni/support-lib/future.yaml"

Foo = interface {
  bar(): future<i32>;
}

The future<> djinni type is mapped to the Future types defined in the C++, Java and ObjC djinni support library. In Javascript, future<> is mapped to the builtin Promise type (and therefore supports the await syntax).

The C++ Future type has optional support for coroutines. If coroutines are availble (eg. compiling with C++20 or C++17 with -fcoroutines-ts), then you can use co_await on future objects.

FAQ

Q. Do I need to use Bazel to build my project?

A. No. You may use whatever build system or IDE you like. All you need for your project is including the generated files in it. We use Bazel to build the code generator and unit tests, but it's not needed for building user projects. You still need to have Bazel installed if you want to run the code generator though, because the run_djinni.sh script indirectly uses it to ensure the code generator is built and up to date.

Q. Can we include arbitrary bytes in the string type?

A. No! iOS and Java strings are unicode, so arbitrary bytes won't necessarily be able to get translated into such strings.

Q. How many Djinni objects can I have?

A. There is no limit in iOS. But on Android, JNI has a limit on the number of global references you can create (usually ~64k). Each interface type Djinni object (either native object in java or java object in native) takes up one global reference. record type objects do not take up global references.