Permalink
488 lines (360 sloc) 17.6 KB

Contributing to bindgen

Hi! We'd love to have your contributions! If you want help or mentorship, reach out to us in a GitHub issue, or stop by #servo on irc.mozilla.org and introduce yourself.

Code of Conduct

We abide by the Rust Code of Conduct and ask that you do as well.

Filing an Issue

Think you've found a bug? File an issue! To help us understand and reproduce the issue, provide us with:

  • A (preferably reduced) C/C++ header file that reproduces the issue
  • The bindgen flags used to reproduce the issue with the header file
  • The expected bindgen output
  • The actual bindgen output
  • The debugging logs generated when running bindgen on this testcase

Looking to Start Contributing to bindgen?

Building

To build the bindgen library and the bindgen executable:

$ cargo build

If you installed multiple versions of llvm, it may not be able to locate the latest version of libclang. In that case, you may want to either uninstall other versions of llvm, or specify the path of the desired libclang explicitly:

$ export LIBCLANG_PATH=path/to/clang-3.9/lib

Additionally, you may want to build and test with the testing_only_docs feature to ensure that you aren't forgetting to document types and functions. CI will catch it if you forget, but the turn around will be a lot slower ;)

$ cargo build --features testing_only_docs

Testing

Overview

Input C/C++ test headers reside in the tests/headers directory. Expected output Rust bindings live in tests/expectations/tests. For example, tests/headers/my_header.h's expected generated Rust bindings would be tests/expectations/tests/my_header.rs.

There are also some integration tests in the ./bindgen-integration crate, which uses bindgen to generate bindings to some C++ code, and then uses the bindings, asserting that values are what we expect them to be, both on the Rust and C++ side.

The generated and expected bindings are run through rustfmt before they are compared. Make sure you have rustfmt up to date:

$ rustup update nightly
$ rustup run nightly cargo install -f rustfmt-nightly

Note: running cargo test from the root directory of bindgen's repository does not automatically test the generated bindings or run the integration tests. These steps must be performed manually when needed.

Testing Bindings Generation

To regenerate bindings from the corpus of test headers in tests/headers and compare them against the expected bindings in tests/expectations/tests, run:

$ cargo test

As long as you aren't making any changes to bindgen's output, running this should be sufficient to test your local modifications.

Testing Generated Bindings

If your local changes are introducing expected modifications in the tests/expectations/tests/* bindings files, then you should test that the generated bindings files still compile, and that their struct layout tests still pass. Also, run the integration tests (see below).

You can do this with these commands:

$ cd tests/expectations
$ cargo test

Testing a Single Header's Bindings Generation and Compiling its Bindings

Sometimes its useful to work with one test header from start (generating bindings for it) to finish (compiling the bindings and running their layout tests). This can be done with the tests/test-one.sh script. It supports fuzzy searching for test headers. For example, to test tests/headers/what_is_going_on.hpp, execute this command:

$ ./tests/test-one.sh going

Note that test-one.sh does not recompile bindgen, so if you change the code, you'll need to rebuild it before running the script again.

Authoring New Tests

To add a new test header to the suite, simply put it in the tests/headers directory. Next, run bindgen to generate the initial expected output Rust bindings. Put those in tests/expectations/tests.

If your new test requires certain flags to be passed to bindgen, you can specify them at the top of the test header, with a comment like this:

new_test_header.hpp:

// bindgen-flags: --enable-cxx-namespaces -- -std=c++14

Then verify the new Rust bindings compile and pass their layout tests:

$ cd tests/expectations
$ cargo test new_test_header

Test Expectations and libclang Versions

If a test generates different bindings across different libclang versions (for example, because we take advantage of better/newer APIs when possible), then you can add multiple test expectations, one for each supported libclang version. Instead of having a single tests/expectations/tests/my_test.rs file, add each of:

  • tests/expectations/tests/libclang-4/my_test.rs
  • tests/expectations/tests/libclang-3.9/my_test.rs
  • tests/expectations/tests/libclang-3.8/my_test.rs

If you need to update the test expectations for a test file that generates different bindings for different libclang versions, you don't need to have many version of libclang installed locally. Just make a work-in-progress pull request, and then when Travis CI fails, it will log a diff of the expectations. Use the diff to patch the appropriate expectation file locally and then update your pull request.

Usually, bindgen's test runner can infer which version of libclang you have. If for some reason it can't, you can force a specific libclang version to check the bindings against with a cargo feature:

$ cargo test --features testing_only_libclang_$VERSION

Where $VERSION is one of:

  • 4
  • 3_9
  • 3_8

depending on which version of libclang you have installed.

Integration Tests

The ./bindgen-integration crate uses bindgen to generate bindings to some C++ code, and then uses the bindings, asserting that values are what we expect them to be, both on the Rust and C++ side.

To run the integration tests, issue the following:

$ cd bindgen-integration
$ cargo test

Fuzzing bindgen with csmith

We <3 finding hidden bugs and the people who help us find them! One way to help uncover hidden bugs is by running csmith to generate random headers to test bindgen against.

See ./csmith-fuzzing/README.md for details.

Property tests for bindgen with quickchecking

The tests/quickchecking crate generates property tests for bindgen. From the crate's directory you can run the tests with cargo run. For details on additional configuration including how to preserve / inspect the generated property tests, see ./tests/quickchecking/README.md.

Code Overview

bindgen takes C and C++ header files as input and generates corresponding Rust #[repr(C)] type definitions and extern foreign function declarations.

First, we use libclang to parse the input headers. See src/clang.rs for our Rust-y wrappers over the raw C libclang API that the clang-sys crate exposes. We walk over libclang's AST and construct our own internal representation (IR). The ir module and submodules (src/ir/*) contain the IR type definitions and libclang AST into IR parsing code.

The umbrella IR type is the Item. It contains various nested enums that let us drill down and get more specific about the kind of construct that we're looking at. Here is a summary of the IR types and their relationships:

  • Item contains:
    • An ItemId to uniquely identify it.
    • An ItemKind, which is one of:
      • A Module, which is originally a C++ namespace and becomes a Rust module. It contains the set of ItemIds of Items that are defined within it.
      • A Type, which contains:
        • A Layout, describing the type's size and alignment.
        • A TypeKind, which is one of:
          • Some integer type.
          • Some float type.
          • A Pointer to another type.
          • A function pointer type, with ItemIds of its parameter types and return type.
          • An Alias to another type (typedef or using X = ...).
          • A fixed size Array of n elements of another type.
          • A Comp compound type, which is either a struct, class, or union. This is potentially a template definition.
          • A TemplateInstantiation referencing some template definition and a set of template argument types.
          • Etc...
      • A Function, which contains:
        • An ABI
        • A mangled name
        • a FunctionKind, which describes whether this function is a plain function, method, static method, constructor, destructor, etc.
        • The ItemId of its function pointer type.
      • A Var representing a static variable or #define constant, which contains:
        • Its type's ItemId
        • Optionally, a mangled name
        • Optionally, a value

The IR forms a graph of interconnected and inter-referencing types and functions. The ir::traversal module provides IR graph traversal infrastructure: edge kind definitions (base member vs field type vs function parameter, etc...), the Trace trait to enumerate an IR thing's outgoing edges, various traversal types.

After constructing the IR, we run a series of analyses on it. These analyses do everything from allocate logical bitfields into physical units, compute for which types we can #[derive(Debug)], to determining which implicit template parameters a given type uses. The analyses are defined in src/ir/analysis/*. They are implemented as fixed-point algorithms, using the ir::analysis::MonotoneFramework trait.

The final phase is generating Rust source text from the analyzed IR, and it is defined in src/codegen/*. We use the quote crate, which provides the quote! { ... } macro for quasi-quoting Rust forms.

Pull Requests and Code Reviews

Ensure that each commit stands alone, and passes tests. This enables better git bisecting when needed. If your commits do not stand on their own, then rebase them on top of the latest master and squash them into a single commit.

All pull requests undergo code review before merging. To request review, comment r? @github_username_of_reviewer. They we will respond with r+ to approve the pull request, or may leave feedback and request changes to the pull request. Any changes should be squashed into the original commit.

Unsure who to ask for review? Ask any of:

  • @emilio
  • @fitzgen
  • @pepyakin

More resources:

Generating Graphviz Dot Files

We can generate Graphviz dot files from our internal representation of a C/C++ input header, and then you can create a PNG or PDF from it with Graphviz's dot program. This is very useful when debugging bindgen!

First, make sure you have Graphviz and dot installed:

$ brew install graphviz         # OS X
$ sudo dnf install graphviz     # Fedora
$ # Etc...

Then, use the --emit-ir-graphviz flag to generate a dot file from our IR:

$ cargo run -- example.hpp --emit-ir-graphviz output.dot

Finally, convert the dot file to an image:

$ dot -Tpng output.dot -o output.png

The final result will look something like this:

An example graphviz rendering of our IR

Debug Logging

To help debug what bindgen is doing, you can define the environment variable RUST_LOG=bindgen to get a bunch of debugging log spew.

$ RUST_LOG=bindgen ./target/debug/bindgen [flags...] ~/path/to/some/header.h

This logging can also be used when debugging failing tests:

$ RUST_LOG=bindgen cargo test

Using creduce to Minimize Test Cases

If you find a test case that triggers an unexpected panic in bindgen, causes bindgen to emit bindings that won't compile, define structs with the wrong size/alignment, or results in any other kind of incorrectness, then using creduce can help reduce the test case to a minimal one that still exhibits that same bad behavior.

Reduced test cases are SUPER helpful when filing bug reports!

Getting creduce

Often, you can install creduce from your OS's package manager:

$ sudo apt install creduce
$ brew install creduce
$ # Etc...

Otherwise, follow these instructions for building and/or installing creduce.

Running creduce requires two things:

  1. Your isolated test case, and

  2. A script to act as a predicate script describing whether the behavior you're trying to isolate occurred.

With those two things in hand, running creduce looks like this:

$ creduce ./predicate.sh ./isolated-test-case.h

Isolating Your Test Case

If you're using bindgen as a command line tool, pass --dump-preprocessed-input flag.

If you're using bindgen as a Rust library, invoke the bindgen::Builder::dump_preprocessed_input method where you call bindgen::Builder::generate.

Afterwards, there should be a __bindgen.i or __bindgen.ii file containing the combined and preprocessed input headers, which is usable as an isolated, standalone test case.

Writing a Predicate Script

Writing a predicate.sh script for a bindgen test case is straightforward. We already have a general purpose predicate script that you can use, you just have to wrap and configure it.

#!/usr/bin/env bash

# Exit the script with a nonzero exit code if:
# * any individual command finishes with a nonzero exit code, or
# * we access any undefined variable.
set -eu

# Invoke the general purpose predicate script that comes in the
# `bindgen` repository.
#
# You'll need to replace `--whatever-flags` with things that are specific to the
# incorrectness you're trying to pin down. See below for details.
path/to/rust-bindgen/csmith-fuzzing/predicate.py \
    --whatever-flags \
    ./isolated-test-case.h

When hunting down a particular panic emanating from inside bindgen, you can invoke predicate.py like this:

path/to/rust-bindgen/csmith-fuzzing/predicate.py \
    --expect-bindgen-fail \
    --bindgen-grep "thread main panicked at '<insert panic message here>'" \
    ./isolated-test-case.h

Alternatively, when hunting down a bad #[derive(Eq)] that is causing rustc to fail to compile bindgen's emitted bindings, you can invoke predicate.py like this:

path/to/rust-bindgen/csmith-fuzzing/predicate.py \
    --bindings-grep NameOfTheStructThatIsErroneouslyDerivingEq \
    --expect-compile-fail \
    --rustc-grep 'error[E0277]: the trait bound `f64: std::cmp::Eq` is not satisfied' \
    ./isolated-test-case.h

Or, when minimizing a failing layout test in the compiled bindings, you can invoke predicate.py like this:

path/to/rust-bindgen/csmith-fuzzing/predicate.py \
    --bindings-grep MyStruct \
    --expect-layout-tests-fail \
    --layout-tests-grep "thread 'bindgen_test_layout_MyStruct' panicked" \
    ./isolated-test-case.h

For details on all the flags that you can pass to predicate.py, run:

$ path/to/rust-bindgen/csmith-fuzzing/predicate.py --help

And you can always write your own, arbitrary predicate script if you prefer. (Although, maybe we should add extra functionality to predicate.py -- file an issue if you think so!)

creduce is really helpful and can cut hundreds of thousands of lines of test case down to 5 lines.

Happy bug hunting and test case reducing!

More information on using creduce.