Contributing ============ Getting Started --------------- Please refer to the `LLVM Getting Started Guide `_ for general information on how to get started on the LLVM project. A detailed explanation on how to build and test LLDB can be found in the `build instructions `_ and `test instructions `_ respectively. Contributing to LLDB -------------------- Please refer to the `LLVM Developer Policy `_ for information about authoring and uploading a patch. LLDB differs from the LLVM Developer Policy in the following respects. For anything not explicitly listed here, assume that LLDB follows the LLVM policy. Coding Style ++++++++++++ LLDB's code style differs from `LLVM's coding style `_ in a few ways. The 2 main ones are: * `Variable and function naming `_: * Variables are ``snake_case``. * Functions and methods are ``UpperCamelCase``. * Static, global and member variables have ``s_``, ``g_`` and ``m_`` prefixes respectively. * `Use of asserts `_: See the :ref:`section below`. For any other contradictions, consider the `golden rule `_ before choosing to update the style of existing code. All new code in LLDB should be formatted with clang-format. Existing code may not conform and should be updated prior to being modified. Bulk reformatting is discouraged. Test Infrastructure +++++++++++++++++++ Like LLVM it is important to submit tests with your patches, but note that a subset of LLDB tests (the API tests) use a different system. Refer to the `test documentation `_ for more details and the `lldb/test `_ folder for examples. LLDB plugins and their dependencies ----------------------------------- LLDB has a concept of *plugins*, which are used to provide abstraction boundaries over functionality that is specific to a certain architecture, operating system, programming language, etc. A plugin implements an abstract base class (rarely, a set of related base classes), which is a part of LLDB core. This setup allows the LLDB core to remain generic while making it possible to support for new architectures, languages, and so on. For this to work, all code needs to obey certain rules. The principal rule is that LLDB core (defined as: everything under lldb/source *minus* lldb/source/Plugins) must not depend on any specific plugin. The only way it can interact with them is through the abstract interface. Explicit dependencies such as casting the base class to the plugin type are not permitted and neither are more subtle dependencies like checking the name plugin or or other situations where some code in LLDB core is tightly coupled to the implementation details of a specific plugin. The rule for interaction between different plugins is more nuanced. We recognize that some cross-plugin dependencies are unavoidable or even desirable. For example, a plugin may want to extend a plugin of the same kind to add/override/refine some functionality (e.g., Android is a "kind of" Linux, but it handles some things differently). Alternatively, a plugin of one kind may want to build on the functionality offered by a specific plugin of another kind (ELFCore Process plugin uses ELF ObjectFile plugin to create a process out of an ELF core file). In cases such as these, direct dependencies are acceptable. However, to keep the dependency graph manageable, we still have some rules to govern these relationships: * All dependencies between plugins of the same kind must flow in the same direction (if plugin `A1` depends on plugin `B1`, then `B2` must not depend on `A2`) * Dependency graph of plugin kinds must not contain loops (dependencies like `A1->B1`, `B2->C2` and `C3->A3` are forbidden because they induce a cycle in the plugin kind graph even though the plugins themselves are acyclical) The first of these rules is checked via CMake scripts (using the `LLDB_ACCEPTABLE_PLUGIN_DEPENDENCIES` property). Dependencies in this category are expected and permitted (subject to other constraints such as that dependency making sense for the particular pair of plugins). Unfortunately, due to historic reasons, not all plugin dependencies follow this rule, which is why we have another category called `LLDB_TOLERATED_PLUGIN_DEPENDENCIES`. New dependencies are forbidden (even though they are accepted by CMake) and existing ones should be removed wherever possible. .. _Error handling: Error handling and use of assertions in LLDB -------------------------------------------- Contrary to Clang, which is typically a short-lived process, LLDB debuggers stay up and running for a long time, often serving multiple debug sessions initiated by an IDE. For this reason LLDB code needs to be extra thoughtful about how to handle errors. Below are a couple rules of thumb: * Invalid input. To deal with invalid input, such as malformed DWARF, missing object files, or otherwise inconsistent debug info, error handling types such as `llvm::Expected `_ or ``std::optional`` should be used. Functions that may fail should return their result using these wrapper types instead of using a bool to indicate success. Returning a default value when an error occurred is also discouraged. * Assertions. Assertions (from ``assert.h``) should be used liberally to assert internal consistency. Assertions shall **never** be used to detect invalid user input, such as malformed DWARF. An assertion should be placed to assert invariants that the developer is convinced will always hold, regardless what an end-user does with LLDB. Because assertions are not present in release builds, the checks in an assertion may be more expensive than otherwise permissible. In combination with the LLDB test suite, assertions are what allows us to refactor and evolve the LLDB code base. * Logging. LLDB provides a very rich logging API. When recoverable errors cannot reasonably be surfaced to the end user, the error may be written to a topical log channel. * Soft assertions. LLDB provides ``lldbassert()`` as a soft alternative to cover the middle ground of situations that indicate a recoverable bug in LLDB. When asserts are enabled ``lldbassert()`` behaves like ``assert()``. When asserts are disabled, it will print a warning and encourage the user to file a bug report, similar to LLVM's crash handler, and then return execution. Use these sparingly and only if error handling is not otherwise feasible. .. note:: New code should not be using ``lldbassert()`` and existing uses should be replaced by other means of error handling. * Fatal errors. Aborting LLDB's process using ``llvm::report_fatal_error()`` or ``abort()`` should be avoided at all costs. It's acceptable to use ``llvm_unreachable()`` for actually unreachable code such as the default in an otherwise exhaustive switch statement. Overall, please keep in mind that the debugger is often used as a last resort, and a crash in the debugger is rarely appreciated by the end-user.