We like to keep the source consistent and readable. Herein are some guidelines that should help with that.
libgit2
runs on many different platforms with many different compilers.
It is written in ANSI C (a.k.a. C89)
with some specific standards for function and type naming, code formatting,
and testing.
We try to avoid more recent extensions to maximize portability. We also, to
the greatest extent possible, try to avoid lots of #ifdef
s inside the core
code base. This is somewhat unavoidable, but since it can really hamper
maintainability, we keep it to a minimum.
If there is one rule to take away from this document, it is new code should match the surrounding code in a way that makes it impossible to distinguish the new from the old. Consistency is more important to us than anyone's personal opinion about where braces should be placed or spaces vs. tabs.
If a section of code is being completely rewritten, it is okay to bring it in line with the standards that are laid out here, but we will not accept submissions that contain a large number of changes that are merely reformatting.
All external types and functions start with git_
and all #define
macros
start with GIT_
. The libgit2
API is mostly broken into related
functional modules each with a corresponding header. All functions in a
module should be named like git_modulename_functioname()
(e.g. git_repository_open()
).
Functions with a single output parameter should name that parameter out
.
Multiple outputs should be named foo_out
, bar_out
, etc.
Parameters of type git_oid
should be named id
, or foo_id
. Calls that
return an OID should be named git_foo_id
.
Where a callback function is used, the function should also include a
user-supplied extra input that is a void *
named "payload" that will be
passed through to the callback at each invocation.
Wherever possible, use typedef
. In some cases, if a structure is just a
collection of function pointers, the pointer types don't need to be
separately typedef'd, but loose function pointer types should be.
All exported functions must be declared as:
GIT_EXTERN(result_type) git_modulename_functionname(arg_list);
Functions whose modulename is followed by two underscores,
for example git_odb__read_packed
, are semi-private functions.
They are primarily intended for use within the library itself,
and may disappear or change their signature in a future release.
Out parameters come first.
Whenever possible, pass argument pointers as const
. Some structures (such
as git_repository
and git_index
) have mutable internal structure that
prevents this.
Callbacks should always take a void *
payload as their last parameter.
Callback pointers are grouped with their payloads, and typically come last
when passed as arguments:
int git_foo(git_repository *repo, git_foo_cb callback, void *payload);
Some APIs allocate memory which the caller is responsible for freeing; others return a pointer into a buffer that's owned by some other object. Make this explicit in the documentation.
Most public APIs should return an int
error code. As is typical with most
C library functions, a zero value indicates success and a negative value
indicates failure.
Some bindings will transform these returned error codes into exception
types, so returning a semantically appropriate error code is important.
Check
include/git2/errors.h
for the return codes already defined.
In your implementation, use giterr_set()
to provide extended error
information to callers.
If a libgit2
function internally invokes another function that reports an
error, but the error is not propagated up, use giterr_clear()
to prevent
callers from getting the wrong error message later on.
Most public types should be opaque, e.g.:
typedef struct git_odb git_odb;
...with allocation functions returning an "instance" created within the library, and not within the application. This allows the type to grow (or shrink) in size without rebuilding client code.
To preserve ABI compatibility, include an int version
field in all opaque
structures, and initialize to the latest version in the construction call.
Increment the "latest" version whenever the structure changes, and try to only
append to the end of the structure.
If a function's parameter count is too high, it may be desirable to package up the options in a structure. Make them transparent, include a version field, and provide an initializer constant or constructor. Using these structures should be this easy:
git_foo_options opts = GIT_FOO_OPTIONS_INIT;
opts.baz = BAZ_OPTION_ONE;
git_foo(&opts);
Typedef all enumerated types. If each option stands alone, use the enum
type for passing them as parameters; if they are flags to be OR'ed together,
pass them as unsigned int
or uint32_t
or some appropriate type.
Try to keep lines less than 80 characters long. This is a loose requirement, but going significantly over 80 columns is not nice.
Use common sense to wrap most code lines; public function declarations can use a couple of different styles:
/** All on one line is okay if it fits */
GIT_EXTERN(int) git_foo_simple(git_oid *id);
/** Otherwise one argument per line is a good next step */
GIT_EXTERN(int) git_foo_id(
git_oid **out,
int a,
int b);
Indent with tabs; set your editor's tab width to 4 for best effect.
Avoid trailing whitespace and only commit Unix-style newlines (i.e. no CRLF
in the repository - just set core.autocrlf
to true if you are writing code
on a Windows machine).
All comments should conform to Doxygen "javadoc" style conventions for formatting the public API documentation. Try to document every parameter, and keep the comments up to date if you change the parameter list.
Use this template when creating a new public header.
#ifndef INCLUDE_git_${filename}_h__
#define INCLUDE_git_${filename}_h__
#include "git/common.h"
/**
* @file git/${filename}.h
* @brief Git some description
* @defgroup git_${filename} some description routines
* @ingroup Git
* @{
*/
GIT_BEGIN_DECL
/* ... definitions ... */
/** @} */
GIT_END_DECL
#endif
All inlined functions must be declared as:
GIT_INLINE(result_type) git_modulename_functionname(arg_list);
libgit2
uses the clar testing framework.
All PRs should have corresponding tests.
- If the PR fixes an existing issue, the test should fail prior to applying the PR and succeed after applying it.
- If the PR is for new functionality, then the tests should exercise that new functionality to a certain extent. We don't require 100% coverage right now (although we are getting stricter over time).
When adding new tests, we prefer if you attempt to reuse existing test data
(in tests-clar/resources/
) if possible. If you are going to add new test
repositories, please try to strip them of unnecessary files (e.g. sample
hooks, etc).