|
| 1 | +.. _CodingGuidelines: |
| 2 | + |
| 3 | +======================== |
| 4 | +libc++ Coding Guidelines |
| 5 | +======================== |
| 6 | + |
| 7 | +.. contents:: |
| 8 | + :local: |
| 9 | + |
| 10 | +Use ``__ugly_names`` for implementation details |
| 11 | +=============================================== |
| 12 | + |
| 13 | +Libc++ uses ``__ugly_names`` or ``_UglyNames`` for implementation details. These names are reserved for implementations, |
| 14 | +so users may not use them in their own applications. When using a name like ``T``, a user may have defined a macro that |
| 15 | +changes the meaning of ``T``. By using ``__ugly_names`` we avoid that problem. |
| 16 | + |
| 17 | +This is partially enforced by the clang-tidy check ``readability-identifier-naming`` and |
| 18 | +``libcxx/test/libcxx/system_reserved_names.gen.py``. |
| 19 | + |
| 20 | +Don't use argument-dependent lookup unless required by the standard |
| 21 | +=================================================================== |
| 22 | + |
| 23 | +Unqualified function calls are susceptible to |
| 24 | +`argument-dependent lookup (ADL) <https://en.cppreference.com/w/cpp/language/adl>`_. This means calling |
| 25 | +``move(UserType)`` might not call ``std::move``. Therefore, function calls must use qualified names to avoid ADL. Some |
| 26 | +functions in the standard library `require ADL usage <http://eel.is/c++draft/contents#3>`_. Names of classes, variables, |
| 27 | +concepts, and type aliases are not subject to ADL. They don't need to be qualified. |
| 28 | + |
| 29 | +Function overloading also applies to operators. Using ``&user_object`` may call a user-defined ``operator&``. Use |
| 30 | +``std::addressof`` instead. Similarly, to avoid invoking a user-defined ``operator,``, make sure to cast the result to |
| 31 | +``void`` when using the ``,`` or avoid it in the first place. For example: |
| 32 | + |
| 33 | +.. code-block:: cpp |
| 34 | +
|
| 35 | + for (; __first1 != __last1; ++__first1, (void)++__first2) { |
| 36 | + ... |
| 37 | + } |
| 38 | +
|
| 39 | +This is mostly enforced by the clang-tidy checks ``libcpp-robust-against-adl`` and ``libcpp-qualify-declval``. |
| 40 | + |
| 41 | +Avoid including public headers |
| 42 | +============================== |
| 43 | + |
| 44 | +libc++ uses implementation-detail headers for most code. These are in a directory that starts with two underscores |
| 45 | +(e.g. ``<__type_traits/decay.h>``). These detail headers are significantly smaller than their public counterparts. |
| 46 | +This reduces the amount of code that is included in a single public header, which reduces compile times. |
| 47 | + |
| 48 | +Add ``_LIBCPP_HIDE_FROM_ABI`` unless you know better |
| 49 | +==================================================== |
| 50 | + |
| 51 | +``_LIBCPP_HIDE_FROM_ABI`` should be on every function in the library unless there is a reason not to do so. The main |
| 52 | +reason not to add ``_LIBCPP_HIDE_FROM_ABI`` is if a function is exported from the libc++ built library. In that case the |
| 53 | +function should be marked with ``_LIBCPP_EXPORTED_FROM_ABI``. Virtual functions should be marked with |
| 54 | +``_LIBCPP_HIDE_FROM_ABI_VIRTUAL`` instead. |
| 55 | + |
| 56 | +This is mostly enforced by the clang-tidy checks ``libcpp-hide-from-abi`` and ``libcpp-avoid-abi-tag-on-virtual``. |
| 57 | + |
| 58 | +Define configuration macros to 0 or 1 |
| 59 | +===================================== |
| 60 | + |
| 61 | +Macros should usually be defined in all configurations, instead of defining them when they're enabled and leaving them |
| 62 | +undefined otherwise. For example, use |
| 63 | + |
| 64 | +.. code-block:: cpp |
| 65 | +
|
| 66 | + #if SOMETHING |
| 67 | + # define _LIBCPP_SOMETHING_ENABLED 1 |
| 68 | + #else |
| 69 | + # define _LIBCPP_SOMETHING_ENABLED 0 |
| 70 | + #endif |
| 71 | +
|
| 72 | +and then check for ``#if _LIBCPP_SOMETHING_ENABLED`` instead of |
| 73 | + |
| 74 | +.. code-block:: cpp |
| 75 | +
|
| 76 | + #if SOMETHING |
| 77 | + # define _LIBCPP_SOMETHING_ENABLED |
| 78 | + #endif |
| 79 | +
|
| 80 | +and then checking for ``#ifdef _LIBCPP_SOMETHING_ENABLED``. |
| 81 | + |
| 82 | +This makes it significantly easier to catch missing includes, since Clang and GCC will warn when using and undefined |
| 83 | +marco inside an ``#if`` statement when using ``-Wundef``. Some macros in libc++ don't use this style yet, so this only |
| 84 | +applies when introducing a new macro. |
| 85 | + |
| 86 | +This is partially enforced by the clang-tidy check ``libcpp-internal-ftms``. |
| 87 | + |
| 88 | +Use ``_LIBCPP_STD_VER`` |
| 89 | +======================= |
| 90 | + |
| 91 | +libc++ defines the macro ``_LIBCPP_STD_VER`` for the different libc++ dialects. This should be used instead of |
| 92 | +``__cplusplus``. |
| 93 | + |
| 94 | +This is mostly enforced by the clang-tidy check ``libcpp-cpp-version-check``. |
| 95 | + |
| 96 | +Use ``__ugly__`` spellings of vendor attributes |
| 97 | +=============================================== |
| 98 | + |
| 99 | +Vendor attributes should always be ``__uglified__`` to avoid naming clashes with user-defined macros. For gnu-style |
| 100 | +attributes this takes the form ``__attribute__((__foo__))``. C++11-style attributes look like ``[[_Clang::__foo__]]`` or |
| 101 | +``[[__gnu__::__foo__]]`` for Clang or GCC attributes respectively. Clang and GCC also support standard attributes in |
| 102 | +earlier language dialects than they were introduced. These should be spelled as ``[[__foo__]]``. MSVC currently doesn't |
| 103 | +provide alternative spellings for their attributes, so these should be avoided if at all possible. |
| 104 | + |
| 105 | +This is enforced by the clang-tidy check ``libcpp-uglify-attributes``. |
| 106 | + |
| 107 | +Use C++11 extensions in C++03 code if they simplify the code |
| 108 | +============================================================ |
| 109 | + |
| 110 | +libc++ only supports Clang in C++98/03 mode. Clang provides many C++11 features in C++03, making it possible to write a |
| 111 | +lot of code in a simpler way than if we were restricted to C++03 features. Some use of extensions is even mandatory, |
| 112 | +since libc++ supports move semantics in C++03. |
| 113 | + |
| 114 | +Use ``using`` aliases instead of ``typedef`` |
| 115 | +============================================ |
| 116 | + |
| 117 | +``using`` aliases are generally easier to read and support templates. Some code in libc++ uses ``typedef`` for |
| 118 | +historical reasons. |
| 119 | + |
| 120 | +Write SFINAE with ``requires`` clauses in C++20-only code |
| 121 | +========================================================= |
| 122 | + |
| 123 | +``requires`` clauses can be significantly easier to read than ``enable_if`` and friends in some cases, since concepts |
| 124 | +subsume other concepts. This means that overloads based on traits can be written without negating more general cases. |
| 125 | +They also show intent better. |
| 126 | + |
| 127 | +Write ``enable_if`` as ``enable_if_t<conditon, int> = 0`` |
| 128 | +========================================================= |
| 129 | + |
| 130 | +The form ``enable_if_t<condition, int> = 0`` is the only one that works in every language mode and for overload sets |
| 131 | +using the same template arguments otherwise. If the code must work in C++11 or C++03, the libc++-internal alias |
| 132 | +``__enable_if_t`` can be used instead. |
| 133 | + |
| 134 | +Prefer alias templates over class templates |
| 135 | +=========================================== |
| 136 | + |
| 137 | +Alias templates are much more lightweight than class templates, since they don't require new instantiations for |
| 138 | +different types. If the only member of a class is an alias, like in type traits, alias templates should be used if |
| 139 | +possible. They do force more eager evaluation though, which can be a problem in some cases. |
| 140 | + |
| 141 | +Apply ``[[nodiscard]]`` where relevant |
| 142 | +====================================== |
| 143 | + |
| 144 | +Libc++ adds ``[[nodiscard]]`` whenever relevant to catch potential bugs. The standards committee has decided to _not_ |
| 145 | +have a recommended practice where to put them, so libc++ applies it whenever it makes sense to catch potential bugs. |
| 146 | + |
| 147 | +``[[nodiscard]]`` should be applied to functions |
| 148 | + |
| 149 | +- where discarding the return value is most likely a correctness issue. For example a locking constructor in |
| 150 | + ``unique_lock``. |
| 151 | + |
| 152 | +- where discarding the return value likely points to the user wanting to do something different. For example |
| 153 | + ``vector::empty()``, which probably should have been ``vector::clear()``. |
| 154 | + |
| 155 | + This can help spotting bugs easily which otherwise may take a very long time to find. |
| 156 | + |
| 157 | +- which return a constant. For example ``numeric_limits::min()``. |
| 158 | +- which only observe a value. For example ``string::size()``. |
| 159 | + |
| 160 | + Code that discards values from these kinds of functions is dead code. It can either be removed, or the programmer |
| 161 | + meant to do something different. |
| 162 | + |
| 163 | +- where discarding the value is most likely a misuse of the function. For example ``std::find(first, last, val)``. |
| 164 | + |
| 165 | + This protects programmers from assuming too much about how the internals of a function work, making code more robust |
| 166 | + in the presence of future optimizations. |
| 167 | + |
| 168 | +Applications of ``[[nodiscard]]`` are code like any other code, so we aim to test them on public interfaces. This can be |
| 169 | +done with a ``.verify.cpp`` test. Many examples are available. Just look for tests with the suffix |
| 170 | +``.nodiscard.verify.cpp``. |
0 commit comments