blob: c529395358a636fa44fed94539018a2055a8729d [file] [log] [blame]
<!DOCTYPE html>
<html lang="en">
<meta charset="utf-8">
<title>Google C++ Style Guide</title>
<link rel="stylesheet" href="include/styleguide.css">
<script src="include/styleguide.js"></script>
<link rel="shortcut icon" href="">
<body onload="initStyleGuide();">
<div id="content">
<h1>Google C++ Style Guide</h1>
<div class="horizontal_toc" id="tocDiv"></div>
<h2 id="Background" class="ignoreLink">Background</h2>
<p>C++ is one of the main development languages used by
many of Google's open-source projects. As every C++
programmer knows, the language has many powerful features, but
this power brings with it complexity, which in turn can make
code more bug-prone and harder to read and maintain.</p>
<p>The goal of this guide is to manage this complexity by
describing in detail the dos and don'ts of writing C++ code
. These rules exist to
keep the code base manageable while still allowing
coders to use C++ language features productively.</p>
<p><em>Style</em>, also known as readability, is what we call
the conventions that govern our C++ code. The term Style is a
bit of a misnomer, since these conventions cover far more than
just source file formatting.</p>
Most open-source projects developed by
Google conform to the requirements in this guide.
<p>Note that this guide is not a C++ tutorial: we assume that
the reader is familiar with the language. </p>
<h3 id="Goals">Goals of the Style Guide</h3>
<p>Why do we have this document?</p>
<p>There are a few core goals that we believe this guide should
serve. These are the fundamental <b>why</b>s that
underlie all of the individual rules. By bringing these ideas to
the fore, we hope to ground discussions and make it clearer to our
broader community why the rules are in place and why particular
decisions have been made. If you understand what goals each rule is
serving, it should be clearer to everyone when a rule may be waived
(some can be), and what sort of argument or alternative would be
necessary to change a rule in the guide.</p>
<p>The goals of the style guide as we currently see them are as follows:</p>
<dt>Style rules should pull their weight</dt>
<dd>The benefit of a style rule
must be large enough to justify asking all of our engineers to
remember it. The benefit is measured relative to the codebase we would
get without the rule, so a rule against a very harmful practice may
still have a small benefit if people are unlikely to do it
anyway. This principle mostly explains the rules we don’t have, rather
than the rules we do: for example, <code>goto</code> contravenes many
of the following principles, but is already vanishingly rare, so the Style
Guide doesn’t discuss it.</dd>
<dt>Optimize for the reader, not the writer</dt>
<dd>Our codebase (and most individual components submitted to it) is
expected to continue for quite some time. As a result, more time will
be spent reading most of our code than writing it. We explicitly
choose to optimize for the experience of our average software engineer
reading, maintaining, and debugging code in our codebase rather than
ease when writing said code. "Leave a trace for the reader" is a
particularly common sub-point of this principle: When something
surprising or unusual is happening in a snippet of code (for example,
transfer of pointer ownership), leaving textual hints for the reader
at the point of use is valuable (<code>std::unique_ptr</code>
demonstrates the ownership transfer unambiguously at the call
site). </dd>
<dt>Be consistent with existing code</dt>
<dd>Using one style consistently through our codebase lets us focus on
other (more important) issues. Consistency also allows for automation:
tools that format your code or adjust your <code>#include</code>s only
work properly when your code is consistent with the expectations of
the tooling. In many cases, rules that are attributed to "Be
Consistent" boil down to "Just pick one and stop worrying about it";
the potential value of allowing flexibility on these points is
outweighed by the cost of having people argue over them. However,
there are limits to consistency; it is a good tie breaker when there
is no clear technical argument, nor a long-term direction. It applies
more heavily locally (per file, or for a tightly-related set of
interfaces). Consistency should not generally be used as a
justification to do things in an old style without considering the
benefits of the new style, or the tendency of the codebase to converge
on newer styles over time.</dd>
<dt>Be consistent with the broader C++ community when appropriate</dt>
<dd>Consistency with the way other organizations use C++ has value for
the same reasons as consistency within our code base. If a feature in
the C++ standard solves a problem, or if some idiom is widely known
and accepted, that's an argument for using it. However, sometimes
standard features and idioms are flawed, or were just designed without
our codebase's needs in mind. In those cases (as described below) it's
appropriate to constrain or ban standard features. In some cases we
prefer a homegrown or third-party library over a library defined in
the C++ Standard, either out of perceived superiority or insufficient
value to transition the codebase to the standard interface.</dd>
<dt>Avoid surprising or dangerous constructs</dt>
<dd>C++ has features that are more surprising or dangerous than one
might think at a glance. Some style guide restrictions are in place to
prevent falling into these pitfalls. There is a high bar for style
guide waivers on such restrictions, because waiving such rules often
directly risks compromising program correctness.
<dt>Avoid constructs that our average C++ programmer would find tricky
or hard to maintain</dt>
<dd>C++ has features that may not be generally appropriate because of
the complexity they introduce to the code. In widely used
code, it may be more acceptable to use
trickier language constructs, because any benefits of more complex
implementation are multiplied widely by usage, and the cost in understanding
the complexity does not need to be paid again when working with new
portions of the codebase. When in doubt, waivers to rules of this type
can be sought by asking
your project leads. This is specifically
important for our codebase because code ownership and team membership
changes over time: even if everyone that works with some piece of code
currently understands it, such understanding is not guaranteed to hold a
few years from now.</dd>
<dt>Be mindful of our scale</dt>
<dd>With a codebase of 100+ million lines and thousands of engineers,
some mistakes and simplifications for one engineer can become costly
for many. For instance it's particularly important to
avoid polluting the global namespace: name collisions across a
codebase of hundreds of millions of lines are difficult to work with
and hard to avoid if everyone puts things into the global
<dt>Concede to optimization when necessary</dt>
<dd>Performance optimizations can sometimes be necessary and
appropriate, even when they conflict with the other principles of this
<p>The intent of this document is to provide maximal guidance with
reasonable restriction. As always, common sense and good taste should
prevail. By this we specifically refer to the established conventions
of the entire Google C++ community, not just your personal preferences
or those of your team. Be skeptical about and reluctant to use
clever or unusual constructs: the absence of a prohibition is not the
same as a license to proceed. Use your judgment, and if you are
unsure, please don't hesitate to ask your project leads to get additional
<h2 id="C++_Version">C++ Version</h2>
<p>Currently, code should target C++17, i.e., should not use C++2x
features. The C++ version targeted by this guide will advance
(aggressively) over time.</p>
<p>Do not use
<a href="#Nonstandard_Extensions">non-standard extensions</a>.</p>
<div>Consider portability to other environments
before using features from C++14 and C++17 in your project.
<h2 id="Header_Files">Header Files</h2>
<p>In general, every <code>.cc</code> file should have an
associated <code>.h</code> file. There are some common
exceptions, such as unit tests and small <code>.cc</code> files containing
just a <code>main()</code> function.</p>
<p>Correct use of header files can make a huge difference to
the readability, size and performance of your code.</p>
<p>The following rules will guide you through the various
pitfalls of using header files.</p>
<a id="The_-inl.h_Files"></a>
<h3 id="Self_contained_Headers">Self-contained Headers</h3>
<p>Header files should be self-contained (compile on their own) and
end in <code>.h</code>. Non-header files that are meant for inclusion
should end in <code>.inc</code> and be used sparingly.</p>
<p>All header files should be self-contained. Users and refactoring
tools should not have to adhere to special conditions to include the
header. Specifically, a header should
have <a href="#The__define_Guard">header guards</a> and include all
other headers it needs.</p>
<p>Prefer placing the definitions for template and inline functions in
the same file as their declarations. The definitions of these
constructs must be included into every <code>.cc</code> file that uses
them, or the program may fail to link in some build configurations. If
declarations and definitions are in different files, including the
former should transitively include the latter. Do not move these
definitions to separately included header files (<code>-inl.h</code>);
this practice was common in the past, but is no longer allowed.</p>
<p>As an exception, a template that is explicitly instantiated for
all relevant sets of template arguments, or that is a private
implementation detail of a class, is allowed to be defined in the one
and only <code>.cc</code> file that instantiates the template.</p>
<p>There are rare cases where a file designed to be included is not
self-contained. These are typically intended to be included at unusual
locations, such as the middle of another file. They might not
use <a href="#The__define_Guard">header guards</a>, and might not include
their prerequisites. Name such files with the <code>.inc</code>
extension. Use sparingly, and prefer self-contained headers when
<h3 id="The__define_Guard">The #define Guard</h3>
<p>All header files should have <code>#define</code> guards to
prevent multiple inclusion. The format of the symbol name
should be
<p>To guarantee uniqueness, they should
be based on the full path in a project's source tree. For
example, the file <code>foo/src/bar/baz.h</code> in
project <code>foo</code> should have the following
<pre>#ifndef FOO_BAR_BAZ_H_
#define FOO_BAR_BAZ_H_
#endif // FOO_BAR_BAZ_H_
<h3 id="Include_What_You_Use">Include What You Use</h3>
<p>If a source or header file refers to a symbol defined elsewhere,
the file should directly include a header file which properly intends
to provide a declaration or definition of that symbol. It should not
include header files for any other reason.
<p>Do not rely on transitive inclusions. This allows people to remove
no-longer-needed <code>#include</code> statements from their headers without
breaking clients. This also applies to related headers
- <code></code> should include <code>bar.h</code> if it uses a
symbol from it even if <code>foo.h</code>
includes <code>bar.h</code>.</p>
<h3 id="Forward_Declarations">Forward Declarations</h3>
<p>Avoid using forward declarations where possible.
Instead, <a href="#Include_What_You_Use">include the headers you need</a>.
<p class="definition"></p>
<p>A "forward declaration" is a declaration of an entity
without an associated definition.</p>
<pre>// In a C++ source file:
class B;
void FuncInB();
extern int variable_in_b;
<p class="pros"></p>
<li>Forward declarations can save compile time, as
<code>#include</code>s force the compiler to open
more files and process more input.</li>
<li>Forward declarations can save on unnecessary
recompilation. <code>#include</code>s can force
your code to be recompiled more often, due to unrelated
changes in the header.</li>
<p class="cons"></p>
<li>Forward declarations can hide a dependency, allowing
user code to skip necessary recompilation when headers
<li>A forward declaration as opposed to an #include statement
makes it difficult for automatic tooling to discover the module
defining the symbol.</li>
<li>A forward declaration may be broken by subsequent
changes to the library. Forward declarations of functions
and templates can prevent the header owners from making
otherwise-compatible changes to their APIs, such as
widening a parameter type, adding a template parameter
with a default value, or migrating to a new namespace.</li>
<li>Forward declaring symbols from namespace
<code>std::</code> yields undefined behavior.</li>
<li>It can be difficult to determine whether a forward
declaration or a full <code>#include</code> is needed.
Replacing an <code>#include</code> with a forward
declaration can silently change the meaning of
<pre>// b.h:
struct B {};
struct D : B {};
#include "b.h"
void f(B*);
void f(void*);
void test(D* x) { f(x); } // calls f(B*)
If the <code>#include</code> was replaced with forward
decls for <code>B</code> and <code>D</code>,
<code>test()</code> would call <code>f(void*)</code>.
<li>Forward declaring multiple symbols from a header
can be more verbose than simply
<code>#include</code>ing the header.</li>
<li>Structuring code to enable forward declarations
(e.g., using pointer members instead of object members)
can make the code slower and more complex.</li>
<p class="decision"></p>
<p>Try to avoid forward declarations of entities
defined in another project.</p>
<h3 id="Inline_Functions">Inline Functions</h3>
<p>Define functions inline only when they are small, say, 10
lines or fewer.</p>
<p class="definition"></p>
<p>You can declare functions in a way that allows the compiler to expand
them inline rather than calling them through the usual
function call mechanism.</p>
<p class="pros"></p>
<p>Inlining a function can generate more efficient object
code, as long as the inlined function is small. Feel free
to inline accessors and mutators, and other short,
performance-critical functions.</p>
<p class="cons"></p>
<p>Overuse of inlining can actually make programs slower.
Depending on a function's size, inlining it can cause the
code size to increase or decrease. Inlining a very small
accessor function will usually decrease code size while
inlining a very large function can dramatically increase
code size. On modern processors smaller code usually runs
faster due to better use of the instruction cache.</p>
<p class="decision"></p>
<p>A decent rule of thumb is to not inline a function if
it is more than 10 lines long. Beware of destructors,
which are often longer than they appear because of
implicit member- and base-destructor calls!</p>
<p>Another useful rule of thumb: it's typically not cost
effective to inline functions with loops or switch
statements (unless, in the common case, the loop or
switch statement is never executed).</p>
<p>It is important to know that functions are not always
inlined even if they are declared as such; for example,
virtual and recursive functions are not normally inlined.
Usually recursive functions should not be inline. The
main reason for making a virtual function inline is to
place its definition in the class, either for convenience
or to document its behavior, e.g., for accessors and
<h3 id="Names_and_Order_of_Includes">Names and Order of Includes</h3>
<p>Include headers in the following order: Related header, C system headers,
C++ standard library headers,
other libraries' headers, your project's
All of a project's header files should be
listed as descendants of the project's source
directory without use of UNIX directory aliases
<code>.</code> (the current directory) or <code>..</code>
(the parent directory). For example,
should be included as:</p>
<pre>#include "base/logging.h"
<p>In <code><var>dir/foo</var>.cc</code> or
<code><var>dir/foo_test</var>.cc</code>, whose main
purpose is to implement or test the stuff in
<code><var>dir2/foo2</var>.h</code>, order your includes
as follows:</p>
<li>A blank line</li>
<li>C system headers (more precisely: headers in angle brackets with the
<code>.h</code> extension), e.g., <code>&lt;unistd.h&gt;</code>,
<li>A blank line</li>
<li>C++ standard library headers (without file extension), e.g.,
<code>&lt;algorithm&gt;</code>, <code>&lt;cstddef&gt;</code>.</li>
<li>A blank line</li>
<li>Other libraries' <code>.h</code> files.</li>
Your project's <code>.h</code>
<p>Separate each non-empty group with one blank line.</p>
<p>With the preferred ordering, if the related header
<code><var>dir2/foo2</var>.h</code> omits any necessary
includes, the build of <code><var>dir/foo</var>.cc</code>
or <code><var>dir/foo</var></code> will break.
Thus, this rule ensures that build breaks show up first
for the people working on these files, not for innocent
people in other packages.</p>
<p><code><var>dir/foo</var>.cc</code> and
<code><var>dir2/foo2</var>.h</code> are usually in the same
directory (e.g., <code>base/</code> and
<code>base/basictypes.h</code>), but may sometimes be in different
directories too.</p>
<p>Note that the C headers such as <code>stddef.h</code>
are essentially interchangeable with their C++ counterparts
Either style is acceptable, but prefer consistency with existing code.</p>
<p>Within each section the includes should be ordered
alphabetically. Note that older code might not conform to
this rule and should be fixed when convenient.</p>
<p>For example, the includes in
might look like this:</p>
<pre>#include "foo/server/fooserver.h"
#include &lt;sys/types.h&gt;
#include &lt;unistd.h&gt;
#include &lt;string&gt;
#include &lt;vector&gt;
#include "base/basictypes.h"
#include "base/commandlineflags.h"
#include "foo/server/bar.h"
<p>Sometimes, system-specific code needs
conditional includes. Such code can put conditional
includes after other includes. Of course, keep your
system-specific code small and localized. Example:</p>
<pre>#include "foo/public/fooserver.h"
#include "base/port.h" // For LANG_CXX11.
#ifdef LANG_CXX11
#include &lt;initializer_list&gt;
#endif // LANG_CXX11
<h2 id="Scoping">Scoping</h2>
<h3 id="Namespaces">Namespaces</h3>
<p>With few exceptions, place code in a namespace. Namespaces
should have unique names based on the project name, and possibly
its path. Do not use <i>using-directives</i> (e.g.,
<code>using namespace foo</code>). Do not use
inline namespaces. For unnamed namespaces, see
<a href="#Internal_Linkage">Internal Linkage</a>.
</p><p class="definition"></p>
<p>Namespaces subdivide the global scope
into distinct, named scopes, and so are useful for preventing
name collisions in the global scope.</p>
<p class="pros"></p>
<p>Namespaces provide a method for preventing name conflicts
in large programs while allowing most code to use reasonably
short names.</p>
<p>For example, if two different projects have a class
<code>Foo</code> in the global scope, these symbols may
collide at compile time or at runtime. If each project
places their code in a namespace, <code>project1::Foo</code>
and <code>project2::Foo</code> are now distinct symbols that
do not collide, and code within each project's namespace
can continue to refer to <code>Foo</code> without the prefix.</p>
<p>Inline namespaces automatically place their names in
the enclosing scope. Consider the following snippet, for
<pre class="neutralcode">namespace outer {
inline namespace inner {
void foo();
} // namespace inner
} // namespace outer
<p>The expressions <code>outer::inner::foo()</code> and
<code>outer::foo()</code> are interchangeable. Inline
namespaces are primarily intended for ABI compatibility
across versions.</p>
<p class="cons"></p>
<p>Namespaces can be confusing, because they complicate
the mechanics of figuring out what definition a name refers
<p>Inline namespaces, in particular, can be confusing
because names aren't actually restricted to the namespace
where they are declared. They are only useful as part of
some larger versioning policy.</p>
<p>In some contexts, it's necessary to repeatedly refer to
symbols by their fully-qualified names. For deeply-nested
namespaces, this can add a lot of clutter.</p>
<p class="decision"></p>
<p>Namespaces should be used as follows:</p>
<li>Follow the rules on <a href="#Namespace_Names">Namespace Names</a>.
</li><li>Terminate multi-line namespaces with comments as shown in the given examples.
<p>Namespaces wrap the entire source file after
<a href="">
gflags</a> definitions/declarations
and forward declarations of classes from other namespaces.</p>
<pre>// In the .h file
namespace mynamespace {
// All declarations are within the namespace scope.
// Notice the lack of indentation.
class MyClass {
void Foo();
} // namespace mynamespace
<pre>// In the .cc file
namespace mynamespace {
// Definition of functions is within scope of the namespace.
void MyClass::Foo() {
} // namespace mynamespace
<p>More complex <code>.cc</code> files might have additional details,
like flags or using-declarations.</p>
<pre>#include "a.h"
ABSL_FLAG(bool, someflag, false, "dummy flag");
namespace mynamespace {
using ::foo::Bar;
...code for mynamespace... // Code goes against the left margin.
} // namespace mynamespace
<li>To place generated protocol
message code in a namespace, use the
<code>package</code> specifier in the
<code>.proto</code> file. See
<a href="">
Protocol Buffer Packages</a>
for details.</li>
<li>Do not declare anything in namespace
<code>std</code>, including forward declarations of
standard library classes. Declaring entities in
namespace <code>std</code> is undefined behavior, i.e.,
not portable. To declare entities from the standard
library, include the appropriate header file.</li>
<li><p>You may not use a <i>using-directive</i>
to make all names from a namespace available.</p>
<pre class="badcode">// Forbidden -- This pollutes the namespace.
using namespace foo;
<li><p>Do not use <i>Namespace aliases</i> at namespace scope
in header files except in explicitly marked
internal-only namespaces, because anything imported into a namespace
in a header file becomes part of the public
API exported by that file.</p>
<pre>// Shorten access to some commonly used names in .cc files.
namespace baz = ::foo::bar::baz;
<pre>// Shorten access to some commonly used names (in a .h file).
namespace librarian {
namespace impl { // Internal, not part of the API.
namespace sidetable = ::pipeline_diagnostics::sidetable;
} // namespace impl
inline void my_inline_function() {
// namespace alias local to a function (or method).
namespace baz = ::foo::bar::baz;
} // namespace librarian
</li><li>Do not use inline namespaces.</li>
<a id="Unnamed_Namespaces_and_Static_Variables"></a>
<h3 id="Internal_Linkage">Internal Linkage</h3>
<p>When definitions in a <code>.cc</code> file do not need to be
referenced outside that file, give them internal linkage by placing
them in an unnamed namespace or declaring them <code>static</code>.
Do not use either of these constructs in <code>.h</code> files.
</p><p class="definition"></p>
<p>All declarations can be given internal linkage by placing them in unnamed
namespaces. Functions and variables can also be given internal linkage by
declaring them <code>static</code>. This means that anything you're declaring
can't be accessed from another file. If a different file declares something with
the same name, then the two entities are completely independent.</p>
<p class="decision"></p>
<p>Use of internal linkage in <code>.cc</code> files is encouraged
for all code that does not need to be referenced elsewhere.
Do not use internal linkage in <code>.h</code> files.</p>
<p>Format unnamed namespaces like named namespaces. In the
terminating comment, leave the namespace name empty:</p>
<pre>namespace {
} // namespace
<h3 id="Nonmember,_Static_Member,_and_Global_Functions">Nonmember, Static Member, and Global Functions</h3>
<p>Prefer placing nonmember functions in a namespace; use completely global
functions rarely. Do not use a class simply to group static members. Static
methods of a class should generally be closely related to instances of the
class or the class's static data.</p>
<p class="pros"></p>
<p>Nonmember and static member functions can be useful in
some situations. Putting nonmember functions in a
namespace avoids polluting the global namespace.</p>
<p class="cons"></p>
<p>Nonmember and static member functions may make more sense
as members of a new class, especially if they access
external resources or have significant dependencies.</p>
<p class="decision"></p>
<p>Sometimes it is useful to define a
function not bound to a class instance. Such a function
can be either a static member or a nonmember function.
Nonmember functions should not depend on external
variables, and should nearly always exist in a namespace.
Do not create classes only to group static members;
this is no different than just giving the names a
common prefix, and such grouping is usually unnecessary anyway.</p>
<p>If you define a nonmember function and it is only
needed in its <code>.cc</code> file, use
<a href="#Internal_Linkage">internal linkage</a> to limit
its scope.</p>
<h3 id="Local_Variables">Local Variables</h3>
<p>Place a function's variables in the narrowest scope
possible, and initialize variables in the declaration.</p>
<p>C++ allows you to declare variables anywhere in a
function. We encourage you to declare them in as local a
scope as possible, and as close to the first use as
possible. This makes it easier for the reader to find the
declaration and see what type the variable is and what it
was initialized to. In particular, initialization should
be used instead of declaration and assignment, e.g.,:</p>
<pre class="badcode">int i;
i = f(); // Bad -- initialization separate from declaration.
<pre>int j = g(); // Good -- declaration has initialization.
<pre class="badcode">std::vector&lt;int&gt; v;
v.push_back(1); // Prefer initializing using brace initialization.
<pre>std::vector&lt;int&gt; v = {1, 2}; // Good -- v starts initialized.
<p>Variables needed for <code>if</code>, <code>while</code>
and <code>for</code> statements should normally be declared
within those statements, so that such variables are confined
to those scopes. E.g.:</p>
<pre>while (const char* p = strchr(str, '/')) str = p + 1;
<p>There is one caveat: if the variable is an object, its
constructor is invoked every time it enters scope and is
created, and its destructor is invoked every time it goes
out of scope.</p>
<pre class="badcode">// Inefficient implementation:
for (int i = 0; i &lt; 1000000; ++i) {
Foo f; // My ctor and dtor get called 1000000 times each.
<p>It may be more efficient to declare such a variable
used in a loop outside that loop:</p>
<pre>Foo f; // My ctor and dtor get called once each.
for (int i = 0; i &lt; 1000000; ++i) {
<h3 id="Static_and_Global_Variables">Static and Global Variables</h3>
<p>Objects with
<a href="">
static storage duration</a> are forbidden unless they are
<a href="">trivially
destructible</a>. Informally this means that the destructor does not do
anything, even taking member and base destructors into account. More formally it
means that the type has no user-defined or virtual destructor and that all bases
and non-static members are trivially destructible.
Static function-local variables may use dynamic initialization.
Use of dynamic initialization for static class member variables or variables at
namespace scope is discouraged, but allowed in limited circumstances; see below
for details.</p>
<p>As a rule of thumb: a global variable satisfies these requirements if its
declaration, considered in isolation, could be <code>constexpr</code>.</p>
<p class="definition"></p>
<p>Every object has a <dfn>storage duration</dfn>, which correlates with its
lifetime. Objects with static storage duration live from the point of their
initialization until the end of the program. Such objects appear as variables at
namespace scope ("global variables"), as static data members of classes, or as
function-local variables that are declared with the <code>static</code>
specifier. Function-local static variables are initialized when control first
passes through their declaration; all other objects with static storage duration
are initialized as part of program start-up. All objects with static storage
duration are destroyed at program exit (which happens before unjoined threads
are terminated).</p>
<p>Initialization may be <dfn>dynamic</dfn>, which means that something
non-trivial happens during initialization. (For example, consider a constructor
that allocates memory, or a variable that is initialized with the current
process ID.) The other kind of initialization is <dfn>static</dfn>
initialization. The two aren't quite opposites, though: static
initialization <em>always</em> happens to objects with static storage duration
(initializing the object either to a given constant or to a representation
consisting of all bytes set to zero), whereas dynamic initialization happens
after that, if required.</p>
<p class="pros"></p>
<p>Global and static variables are very useful for a large number of
applications: named constants, auxiliary data structures internal to some
translation unit, command-line flags, logging, registration mechanisms,
background infrastructure, etc.</p>
<p class="cons"></p>
<p>Global and static variables that use dynamic initialization or have
non-trivial destructors create complexity that can easily lead to hard-to-find
bugs. Dynamic initialization is not ordered across translation units, and
neither is destruction (except that destruction
happens in reverse order of initialization). When one initialization refers to
another variable with static storage duration, it is possible that this causes
an object to be accessed before its lifetime has begun (or after its lifetime
has ended). Moreover, when a program starts threads that are not joined at exit,
those threads may attempt to access objects after their lifetime has ended if
their destructor has already run.</p>
<p class="decision"></p>
<h4>Decision on destruction</h4>
<p>When destructors are trivial, their execution is not subject to ordering at
all (they are effectively not "run"); otherwise we are exposed to the risk of
accessing objects after the end of their lifetime. Therefore, we only allow
objects with static storage duration if they are trivially destructible.
Fundamental types (like pointers and <code>int</code>) are trivially
destructible, as are arrays of trivially destructible types. Note that
variables marked with <code>constexpr</code> are trivially destructible.</p>
<pre>const int kNum = 10; // allowed
struct X { int n; };
const X kX[] = {{1}, {2}, {3}}; // allowed
void foo() {
static const char* const kMessages[] = {"hello", "world"}; // allowed
// allowed: constexpr guarantees trivial destructor
constexpr std::array&lt;int, 3&gt; kArray = {{1, 2, 3}};</pre>
<pre class="badcode">// bad: non-trivial destructor
const std::string kFoo = "foo";
// bad for the same reason, even though kBar is a reference (the
// rule also applies to lifetime-extended temporary objects)
const std::string&amp; kBar = StrCat("a", "b", "c");
void bar() {
// bad: non-trivial destructor
static std::map&lt;int, int&gt; kData = {{1, 0}, {2, 0}, {3, 0}};
<p>Note that references are not objects, and thus they are not subject to the
constraints on destructibility. The constraint on dynamic initialization still
applies, though. In particular, a function-local static reference of the form
<code>static T&amp; t = *new T;</code> is allowed.</p>
<h4>Decision on initialization</h4>
<p>Initialization is a more complex topic. This is because we must not only
consider whether class constructors execute, but we must also consider the
evaluation of the initializer:</p>
<pre class="neutralcode">int n = 5; // fine
int m = f(); // ? (depends on f)
Foo x; // ? (depends on Foo::Foo)
Bar y = g(); // ? (depends on g and on Bar::Bar)
<p>All but the first statement expose us to indeterminate initialization
<p>The concept we are looking for is called <em>constant initialization</em> in
the formal language of the C++ standard. It means that the initializing
expression is a constant expression, and if the object is initialized by a
constructor call, then the constructor must be specified as
<code>constexpr</code>, too:</p>
<pre>struct Foo { constexpr Foo(int) {} };
int n = 5; // fine, 5 is a constant expression
Foo x(2); // fine, 2 is a constant expression and the chosen constructor is constexpr
Foo a[] = { Foo(1), Foo(2), Foo(3) }; // fine</pre>
<p>Constant initialization is always allowed. Constant initialization of
static storage duration variables should be marked with <code>constexpr</code>
or where possible the
<a href="">
attribute. Any non-local static storage
duration variable that is not so marked should be presumed to have
dynamic initialization, and reviewed very carefully.</p>
<p>By contrast, the following initializations are problematic:</p>
<pre class="badcode">// Some declarations used below.
time_t time(time_t*); // not constexpr!
int f(); // not constexpr!
struct Bar { Bar() {} };
// Problematic initializations.
time_t m = time(nullptr); // initializing expression not a constant expression
Foo y(f()); // ditto
Bar b; // chosen constructor Bar::Bar() not constexpr</pre>
<p>Dynamic initialization of nonlocal variables is discouraged, and in general
it is forbidden. However, we do permit it if no aspect of the program depends
on the sequencing of this initialization with respect to all other
initializations. Under those restrictions, the ordering of the initialization
does not make an observable difference. For example:</p>
<pre>int p = getpid(); // allowed, as long as no other static variable
// uses p in its own initialization</pre>
<p>Dynamic initialization of static local variables is allowed (and common).</p>
<h4>Common patterns</h4>
<li>Global strings: if you require a global or static string constant,
consider using a simple character array, or a char pointer to the first
element of a string literal. String literals have static storage duration
already and are usually sufficient.</li>
<li>Maps, sets, and other dynamic containers: if you require a static, fixed
collection, such as a set to search against or a lookup table, you cannot
use the dynamic containers from the standard library as a static variable,
since they have non-trivial destructors. Instead, consider a simple array of
trivial types, e.g., an array of arrays of ints (for a "map from int to
int"), or an array of pairs (e.g., pairs of <code>int</code> and <code>const
char*</code>). For small collections, linear search is entirely sufficient
(and efficient, due to memory locality); consider using the facilities from
<a href="">absl/algorithm/container.h</a>
for the standard operations. If necessary, keep the collection in sorted
order and use a binary search algorithm. If you do really prefer a dynamic
container from the standard library, consider using a function-local static
pointer, as described below.</li>
<li>Smart pointers (<code>unique_ptr</code>, <code>shared_ptr</code>): smart
pointers execute cleanup during destruction and are therefore forbidden.
Consider whether your use case fits into one of the other patterns described
in this section. One simple solution is to use a plain pointer to a
dynamically allocated object and never delete it (see last item).</li>
<li>Static variables of custom types: if you require static, constant data of
a type that you need to define yourself, give the type a trivial destructor
and a <code>constexpr</code> constructor.</li>
<li>If all else fails, you can create an object dynamically and never delete
it by using a function-local static pointer or reference (e.g., <code>static
const auto&amp; impl = *new T(args...);</code>).</li>
<h3 id="thread_local">thread_local Variables</h3>
<p><code>thread_local</code> variables that aren't declared inside a function
must be initialized with a true compile-time constant,
and this must be enforced by using the
<a href="">
attribute. Prefer
<code>thread_local</code> over other ways of defining thread-local data.</p>
<p class="definition"></p>
<p>Starting with C++11, variables can be declared with the
<code>thread_local</code> specifier:</p>
<pre>thread_local Foo foo = ...;
<p>Such a variable is actually a collection of objects, so that when different
threads access it, they are actually accessing different objects.
<code>thread_local</code> variables are much like
<a href="#Static_and_Global_Variables">static storage duration variables</a>
in many respects. For instance, they can be declared at namespace scope,
inside functions, or as static class members, but not as ordinary class
<p><code>thread_local</code> variable instances are initialized much like
static variables, except that they must be initialized separately for each
thread, rather than once at program startup. This means that
<code>thread_local</code> variables declared within a function are safe, but
other <code>thread_local</code> variables are subject to the same
initialization-order issues as static variables (and more besides).</p>
<p><code>thread_local</code> variable instances are destroyed when their thread
terminates, so they do not have the destruction-order issues of static
<p class="pros"></p>
<li>Thread-local data is inherently safe from races (because only one thread
can ordinarily access it), which makes <code>thread_local</code> useful for
concurrent programming.</li>
<li><code>thread_local</code> is the only standard-supported way of creating
thread-local data.</li>
<p class="cons"></p>
<li>Accessing a <code>thread_local</code> variable may trigger execution of
an unpredictable and uncontrollable amount of other code.</li>
<li><code>thread_local</code> variables are effectively global variables,
and have all the drawbacks of global variables other than lack of
<li>The memory consumed by a <code>thread_local</code> variable scales with
the number of running threads (in the worst case), which can be quite large
in a program.</li>
<li>An ordinary class member cannot be <code>thread_local</code>.</li>
<li><code>thread_local</code> may not be as efficient as certain compiler
<p class="decision"></p>
<p><code>thread_local</code> variables inside a function have no safety
concerns, so they can be used without restriction. Note that you can use
a function-scope <code>thread_local</code> to simulate a class- or
namespace-scope <code>thread_local</code> by defining a function or
static method that exposes it:</p>
<pre>Foo&amp; MyThreadLocalFoo() {
thread_local Foo result = ComplicatedInitialization();
return result;
<p><code>thread_local</code> variables at class or namespace scope must be
initialized with a true compile-time constant (i.e., they must have no
dynamic initialization). To enforce this, <code>thread_local</code> variables
at class or namespace scope must be annotated with
<a href="">
(or <code>constexpr</code>, but that should be rare):</p>
<pre>ABSL_CONST_INIT thread_local Foo foo = ...;
<p><code>thread_local</code> should be preferred over other mechanisms for
defining thread-local data.</p>
<h2 id="Classes">Classes</h2>
<p>Classes are the fundamental unit of code in C++. Naturally,
we use them extensively. This section lists the main dos and
don'ts you should follow when writing a class.</p>
<h3 id="Doing_Work_in_Constructors">Doing Work in Constructors</h3>
<p>Avoid virtual method calls in constructors, and avoid
initialization that can fail if you can't signal an error.</p>
<p class="definition"></p>
<p>It is possible to perform arbitrary initialization in the body
of the constructor.</p>
<p class="pros"></p>
<li>No need to worry about whether the class has been initialized or
<li>Objects that are fully initialized by constructor call can
be <code>const</code> and may also be easier to use with standard containers
or algorithms.</li>
<p class="cons"></p>
<li>If the work calls virtual functions, these calls
will not get dispatched to the subclass
implementations. Future modification to your class can
quietly introduce this problem even if your class is
not currently subclassed, causing much confusion.</li>
<li>There is no easy way for constructors to signal errors, short of
crashing the program (not always appropriate) or using exceptions
(which are <a href="#Exceptions">forbidden</a>).</li>
<li>If the work fails, we now have an object whose initialization
code failed, so it may be an unusual state requiring a <code>bool
IsValid()</code> state checking mechanism (or similar) which is easy
to forget to call.</li>
<li>You cannot take the address of a constructor, so whatever work
is done in the constructor cannot easily be handed off to, for
example, another thread.</li>
<p class="decision"></p>
<p>Constructors should never call virtual functions. If appropriate
for your code ,
terminating the program may be an appropriate error handling
response. Otherwise, consider a factory function
or <code>Init()</code> method as described in
<a href="">TotW #42</a>
Avoid <code>Init()</code> methods on objects with
no other states that affect which public methods may be called
(semi-constructed objects of this form are particularly hard to work
with correctly).</p>
<a id="Explicit_Constructors"></a>
<h3 id="Implicit_Conversions">Implicit Conversions</h3>
<p>Do not define implicit conversions. Use the <code>explicit</code>
keyword for conversion operators and single-argument
<p class="definition"></p>
<p>Implicit conversions allow an
object of one type (called the <dfn>source type</dfn>) to
be used where a different type (called the <dfn>destination
type</dfn>) is expected, such as when passing an
<code>int</code> argument to a function that takes a
<code>double</code> parameter.</p>
<p>In addition to the implicit conversions defined by the language,
users can define their own, by adding appropriate members to the
class definition of the source or destination type. An implicit
conversion in the source type is defined by a type conversion operator
named after the destination type (e.g., <code>operator
bool()</code>). An implicit conversion in the destination
type is defined by a constructor that can take the source type as
its only argument (or only argument with no default value).</p>
<p>The <code>explicit</code> keyword can be applied to a constructor
or (since C++11) a conversion operator, to ensure that it can only be
used when the destination type is explicit at the point of use,
e.g., with a cast. This applies not only to implicit conversions, but to
C++11's list initialization syntax:</p>
<pre>class Foo {
explicit Foo(int x, double y);
void Func(Foo f);
<pre class="badcode">Func({42, 3.14}); // Error
This kind of code isn't technically an implicit conversion, but the
language treats it as one as far as <code>explicit</code> is concerned.
<p class="pros"></p>
<li>Implicit conversions can make a type more usable and
expressive by eliminating the need to explicitly name a type
when it's obvious.</li>
<li>Implicit conversions can be a simpler alternative to
overloading, such as when a single
function with a <code>string_view</code> parameter takes the
place of separate overloads for <code>std::string</code> and
<code>const char*</code>.</li>
<li>List initialization syntax is a concise and expressive
way of initializing objects.</li>
<p class="cons"></p>
<li>Implicit conversions can hide type-mismatch bugs, where the
destination type does not match the user's expectation, or
the user is unaware that any conversion will take place.</li>
<li>Implicit conversions can make code harder to read, particularly
in the presence of overloading, by making it less obvious what
code is actually getting called.</li>
<li>Constructors that take a single argument may accidentally
be usable as implicit type conversions, even if they are not
intended to do so.</li>
<li>When a single-argument constructor is not marked
<code>explicit</code>, there's no reliable way to tell whether
it's intended to define an implicit conversion, or the author
simply forgot to mark it.</li>
<li>Implicit conversions can lead to call-site ambiguities, especially
when there are bidirectional implicit conversions. This can be caused
either by having two types that both provide an implicit conversion,
or by a single type that has both an implicit constructor and an
implicit type conversion operator.</li>
<li>List initialization can suffer from the same problems if
the destination type is implicit, particularly if the
list has only a single element.</li>
<p class="decision"></p>
<p>Type conversion operators, and constructors that are
callable with a single argument, must be marked
<code>explicit</code> in the class definition. As an
exception, copy and move constructors should not be
<code>explicit</code>, since they do not perform type
<p>Implicit conversions can sometimes be necessary and appropriate for
types that are designed to be interchangeable, for example when objects
of two types are just different representations of the same underlying
value. In that case, contact
your project leads to request a waiver
of this rule.
<p>Constructors that cannot be called with a single argument
may omit <code>explicit</code>. Constructors that
take a single <code>std::initializer_list</code> parameter should
also omit <code>explicit</code>, in order to support copy-initialization
(e.g., <code>MyType m = {1, 2};</code>).</p>
<h3 id="Copyable_Movable_Types">Copyable and Movable Types</h3>
<a id="Copy_Constructors"></a>
<p>A class's public API must make clear whether the class is copyable,
move-only, or neither copyable nor movable. Support copying and/or
moving if these operations are clear and meaningful for your type.</p>
<p class="definition"></p>
<p>A movable type is one that can be initialized and assigned
from temporaries.</p>
<p>A copyable type is one that can be initialized or assigned from
any other object of the same type (so is also movable by definition), with the
stipulation that the value of the source does not change.
<code>std::unique_ptr&lt;int&gt;</code> is an example of a movable but not
copyable type (since the value of the source
<code>std::unique_ptr&lt;int&gt;</code> must be modified during assignment to
the destination). <code>int</code> and <code>std::string</code> are examples of
movable types that are also copyable. (For <code>int</code>, the move and copy
operations are the same; for <code>std::string</code>, there exists a move operation
that is less expensive than a copy.)</p>
<p>For user-defined types, the copy behavior is defined by the copy
constructor and the copy-assignment operator. Move behavior is defined by the
move constructor and the move-assignment operator, if they exist, or by the
copy constructor and the copy-assignment operator otherwise.</p>
<p>The copy/move constructors can be implicitly invoked by the compiler
in some situations, e.g., when passing objects by value.</p>
<p class="pros"></p>
<p>Objects of copyable and movable types can be passed and returned by value,
which makes APIs simpler, safer, and more general. Unlike when passing objects
by pointer or reference, there's no risk of confusion over ownership,
lifetime, mutability, and similar issues, and no need to specify them in the
contract. It also prevents non-local interactions between the client and the
implementation, which makes them easier to understand, maintain, and optimize by
the compiler. Further, such objects can be used with generic APIs that
require pass-by-value, such as most containers, and they allow for additional
flexibility in e.g., type composition.</p>
<p>Copy/move constructors and assignment operators are usually
easier to define correctly than alternatives
like <code>Clone()</code>, <code>CopyFrom()</code> or <code>Swap()</code>,
because they can be generated by the compiler, either implicitly or
with <code>= default</code>. They are concise, and ensure
that all data members are copied. Copy and move
constructors are also generally more efficient, because they don't
require heap allocation or separate initialization and assignment
steps, and they're eligible for optimizations such as
<a href="">
copy elision</a>.</p>
<p>Move operations allow the implicit and efficient transfer of
resources out of rvalue objects. This allows a plainer coding style
in some cases.</p>
<p class="cons"></p>
<p>Some types do not need to be copyable, and providing copy
operations for such types can be confusing, nonsensical, or outright
incorrect. Types representing singleton objects (<code>Registerer</code>),
objects tied to a specific scope (<code>Cleanup</code>), or closely coupled to
object identity (<code>Mutex</code>) cannot be copied meaningfully.
Copy operations for base class types that are to be used
polymorphically are hazardous, because use of them can lead to
<a href="">object slicing</a>.
Defaulted or carelessly-implemented copy operations can be incorrect, and the
resulting bugs can be confusing and difficult to diagnose.</p>
<p>Copy constructors are invoked implicitly, which makes the
invocation easy to miss. This may cause confusion for programmers used to
languages where pass-by-reference is conventional or mandatory. It may also
encourage excessive copying, which can cause performance problems.</p>
<p class="decision"></p>
<p>Every class's public interface must make clear which copy and move
operations the class supports. This should usually take the form of explicitly
declaring and/or deleting the appropriate operations in the <code>public</code>
section of the declaration.</p>
<p>Specifically, a copyable class should explicitly declare the copy
operations, a move-only class should explicitly declare the move operations, and
a non-copyable/movable class should explicitly delete the copy operations. A
copyable class may also declare move operations in order to support efficient
moves. Explicitly declaring or deleting all four copy/move operations is
permitted, but not required. If you provide a copy or move assignment operator,
you must also provide the corresponding constructor.</p>
<pre>class Copyable {
Copyable(const Copyable&amp; other) = default;
Copyable&amp; operator=(const Copyable&amp; other) = default;
// The implicit move operations are suppressed by the declarations above.
// You may explicitly declare move operations to support efficient moves.
class MoveOnly {
MoveOnly(MoveOnly&amp;&amp; other) = default;
MoveOnly&amp; operator=(MoveOnly&amp;&amp; other) = default;
// The copy operations are implicitly deleted, but you can
// spell that out explicitly if you want:
MoveOnly(const MoveOnly&amp;) = delete;
MoveOnly&amp; operator=(const MoveOnly&amp;) = delete;
class NotCopyableOrMovable {
// Not copyable or movable
NotCopyableOrMovable(const NotCopyableOrMovable&amp;) = delete;
NotCopyableOrMovable&amp; operator=(const NotCopyableOrMovable&amp;)
= delete;
// The move operations are implicitly disabled, but you can
// spell that out explicitly if you want:
NotCopyableOrMovable(NotCopyableOrMovable&amp;&amp;) = delete;
NotCopyableOrMovable&amp; operator=(NotCopyableOrMovable&amp;&amp;)
= delete;
<p>These declarations/deletions can be omitted only if they are obvious:</p>
<li>If the class has no <code>private</code> section, like a
<a href="#Structs_vs._Classes">struct</a> or an interface-only base class,
then the copyability/movability can be determined by the
copyability/movability of any public data members.
</li><li>If a base class clearly isn't copyable or movable, derived classes
naturally won't be either. An interface-only base class that leaves these
operations implicit is not sufficient to make concrete subclasses clear.
</li><li>Note that if you explicitly declare or delete either the constructor or
assignment operation for copy, the other copy operation is not obvious and
must be declared or deleted. Likewise for move operations.
<p>A type should not be copyable/movable if the meaning of
copying/moving is unclear to a casual user, or if it incurs unexpected
costs. Move operations for copyable types are strictly a performance
optimization and are a potential source of bugs and complexity, so
avoid defining them unless they are significantly more efficient than
the corresponding copy operations. If your type provides copy operations, it is
recommended that you design your class so that the default implementation of
those operations is correct. Remember to review the correctness of any
defaulted operations as you would any other code.</p>
<p>Due to the risk of slicing, prefer to avoid providing a public assignment
operator or copy/move constructor for a class that's
intended to be derived from (and prefer to avoid deriving from a class
with such members). If your base class needs to be
copyable, provide a public virtual <code>Clone()</code>
method, and a protected copy constructor that derived classes
can use to implement it.</p>
<h3 id="Structs_vs._Classes">Structs vs. Classes</h3>
<p>Use a <code>struct</code> only for passive objects that
carry data; everything else is a <code>class</code>.</p>
<p>The <code>struct</code> and <code>class</code>
keywords behave almost identically in C++. We add our own
semantic meanings to each keyword, so you should use the
appropriate keyword for the data-type you're
<p><code>structs</code> should be used for passive objects that carry
data, and may have associated constants. All fields must be public. The
struct must not have invariants that imply relationships between
different fields, since direct user access to those fields may
break those invariants. Constructors, destructors, and helper methods may
be present; however, these methods must not require or enforce any
<p>If more functionality or invariants are required, a
<code>class</code> is more appropriate. If in doubt, make
it a <code>class</code>.</p>
<p>For consistency with STL, you can use
<code>struct</code> instead of <code>class</code> for
stateless types, such as traits,
<a href="#Template_metaprogramming">template metafunctions</a>,
and some functors.</p>
<p>Note that member variables in structs and classes have
<a href="#Variable_Names">different naming rules</a>.</p>
<h3 id="Structs_vs._Tuples">Structs vs. Pairs and Tuples</h3>
<p>Prefer to use a <code>struct</code> instead of a pair or a
tuple whenever the elements can have meaningful names.</p>
While using pairs and tuples can avoid the need to define a custom type,
potentially saving work when <em>writing</em> code, a meaningful field
name will almost always be much clearer when <em>reading</em> code than
<code>.first</code>, <code>.second</code>, or <code>std::get&lt;X&gt;</code>.
While C++14's introduction of <code>std::get&lt;Type&gt;</code> to access a
tuple element by type rather than index (when the type is unique) can
sometimes partially mitigate this, a field name is usually substantially
clearer and more informative than a type.
Pairs and tuples may be appropriate in generic code where there are not
specific meanings for the elements of the pair or tuple. Their use may
also be required in order to interoperate with existing code or APIs.
<a id="Multiple_Inheritance"></a>
<h3 id="Inheritance">Inheritance</h3>
<p>Composition is often more appropriate than inheritance.
When using inheritance, make it <code>public</code>.</p>
<p class="definition"></p>
<p> When a sub-class
inherits from a base class, it includes the definitions
of all the data and operations that the base class
defines. "Interface inheritance" is inheritance from a
pure abstract base class (one with no state or defined
methods); all other inheritance is "implementation
<p class="pros"></p>
<p>Implementation inheritance reduces code size by re-using
the base class code as it specializes an existing type.
Because inheritance is a compile-time declaration, you
and the compiler can understand the operation and detect
errors. Interface inheritance can be used to
programmatically enforce that a class expose a particular
API. Again, the compiler can detect errors, in this case,
when a class does not define a necessary method of the
<p class="cons"></p>
<p>For implementation inheritance, because the code
implementing a sub-class is spread between the base and
the sub-class, it can be more difficult to understand an
implementation. The sub-class cannot override functions
that are not virtual, so the sub-class cannot change
<p>Multiple inheritance is especially problematic, because
it often imposes a higher performance overhead (in fact,
the performance drop from single inheritance to multiple
inheritance can often be greater than the performance
drop from ordinary to virtual dispatch), and because
it risks leading to "diamond" inheritance patterns,
which are prone to ambiguity, confusion, and outright bugs.</p>
<p class="decision"></p>
<p>All inheritance should be <code>public</code>. If you
want to do private inheritance, you should be including
an instance of the base class as a member instead.</p>
<p>Do not overuse implementation inheritance. Composition
is often more appropriate. Try to restrict use of
inheritance to the "is-a" case: <code>Bar</code>
subclasses <code>Foo</code> if it can reasonably be said
that <code>Bar</code> "is a kind of"
<p>Limit the use of <code>protected</code> to those
member functions that might need to be accessed from
subclasses. Note that <a href="#Access_Control">data
members should be private</a>.</p>
<p>Explicitly annotate overrides of virtual functions or virtual
destructors with exactly one of an <code>override</code> or (less
frequently) <code>final</code> specifier. Do not
use <code>virtual</code> when declaring an override.
Rationale: A function or destructor marked
<code>override</code> or <code>final</code> that is
not an override of a base class virtual function will
not compile, and this helps catch common errors. The
specifiers serve as documentation; if no specifier is
present, the reader has to check all ancestors of the
class in question to determine if the function or
destructor is virtual or not.</p>
<p>Multiple inheritance is permitted, but multiple <em>implementation</em>
inheritance is strongly discouraged.</p>
<h3 id="Operator_Overloading">Operator Overloading</h3>
<p>Overload operators judiciously. Do not use user-defined literals.</p>
<p class="definition"></p>
<p>C++ permits user code to
<a href="">declare
overloaded versions of the built-in operators</a> using the
<code>operator</code> keyword, so long as one of the parameters
is a user-defined type. The <code>operator</code> keyword also
permits user code to define new kinds of literals using
<code>operator""</code>, and to define type-conversion functions
such as <code>operator bool()</code>.</p>
<p class="pros"></p>
<p>Operator overloading can make code more concise and
intuitive by enabling user-defined types to behave the same
as built-in types. Overloaded operators are the idiomatic names
for certain operations (e.g., <code>==</code>, <code>&lt;</code>,
<code>=</code>, and <code>&lt;&lt;</code>), and adhering to
those conventions can make user-defined types more readable
and enable them to interoperate with libraries that expect
those names.</p>
<p>User-defined literals are a very concise notation for
creating objects of user-defined types.</p>
<p class="cons"></p>
<li>Providing a correct, consistent, and unsurprising
set of operator overloads requires some care, and failure
to do so can lead to confusion and bugs.</li>
<li>Overuse of operators can lead to obfuscated code,
particularly if the overloaded operator's semantics
don't follow convention.</li>
<li>The hazards of function overloading apply just as
much to operator overloading, if not more so.</li>
<li>Operator overloads can fool our intuition into
thinking that expensive operations are cheap, built-in
<li>Finding the call sites for overloaded operators may
require a search tool that's aware of C++ syntax, rather
than e.g., grep.</li>
<li>If you get the argument type of an overloaded operator
wrong, you may get a different overload rather than a
compiler error. For example, <code>foo &lt; bar</code>
may do one thing, while <code>&amp;foo &lt; &amp;bar</code>
does something totally different.</li>
<li>Certain operator overloads are inherently hazardous.
Overloading unary <code>&amp;</code> can cause the same
code to have different meanings depending on whether
the overload declaration is visible. Overloads of
<code>&amp;&amp;</code>, <code>||</code>, and <code>,</code>
(comma) cannot match the evaluation-order semantics of the
built-in operators.</li>
<li>Operators are often defined outside the class,
so there's a risk of different files introducing
different definitions of the same operator. If both
definitions are linked into the same binary, this results
in undefined behavior, which can manifest as subtle
run-time bugs.</li>
<li>User-defined literals (UDLs) allow the creation of new
syntactic forms that are unfamiliar even to experienced C++
programmers, such as <code>"Hello World"sv</code> as a
shorthand for <code>std::string_view("Hello World")</code>.
Existing notations are clearer, though less terse.</li>
<li>Because they can't be namespace-qualified, uses of UDLs also require
use of either using-directives (which <a href="#Namespaces">we ban</a>) or
using-declarations (which <a href="#Aliases">we ban in header files</a> except
when the imported names are part of the interface exposed by the header
file in question). Given that header files would have to avoid UDL
suffixes, we prefer to avoid having conventions for literals differ
between header files and source files.
<p class="decision"></p>
<p>Define overloaded operators only if their meaning is
obvious, unsurprising, and consistent with the corresponding
built-in operators. For example, use <code>|</code> as a
bitwise- or logical-or, not as a shell-style pipe.</p>
<p>Define operators only on your own types. More precisely,
define them in the same headers, .cc files, and namespaces
as the types they operate on. That way, the operators are available
wherever the type is, minimizing the risk of multiple
definitions. If possible, avoid defining operators as templates,
because they must satisfy this rule for any possible template
arguments. If you define an operator, also define
any related operators that make sense, and make sure they
are defined consistently. For example, if you overload
<code>&lt;</code>, overload all the comparison operators,
and make sure <code>&lt;</code> and <code>&gt;</code> never
return true for the same arguments.</p>
<p>Prefer to define non-modifying binary operators as
non-member functions. If a binary operator is defined as a
class member, implicit conversions will apply to the
right-hand argument, but not the left-hand one. It will
confuse your users if <code>a &lt; b</code> compiles but
<code>b &lt; a</code> doesn't.</p>
<p>Don't go out of your way to avoid defining operator
overloads. For example, prefer to define <code>==</code>,
<code>=</code>, and <code>&lt;&lt;</code>, rather than
<code>Equals()</code>, <code>CopyFrom()</code>, and
<code>PrintTo()</code>. Conversely, don't define
operator overloads just because other libraries expect
them. For example, if your type doesn't have a natural
ordering, but you want to store it in a <code>std::set</code>,
use a custom comparator rather than overloading
<p>Do not overload <code>&amp;&amp;</code>, <code>||</code>,
<code>,</code> (comma), or unary <code>&amp;</code>. Do not overload
<code>operator""</code>, i.e., do not introduce user-defined
literals. Do not use any such literals provided by others
(including the standard library).</p>
<p>Type conversion operators are covered in the section on
<a href="#Implicit_Conversions">implicit conversions</a>.
The <code>=</code> operator is covered in the section on
<a href="#Copy_Constructors">copy constructors</a>. Overloading
<code>&lt;&lt;</code> for use with streams is covered in the
section on <a href="#Streams">streams</a>. See also the rules on
<a href="#Function_Overloading">function overloading</a>, which
apply to operator overloading as well.</p>
<h3 id="Access_Control">Access Control</h3>
<p>Make classes' data members <code>private</code>, unless they are
<a href="#Constant_Names">constants</a>. This simplifies reasoning about invariants, at the cost
of some easy boilerplate in the form of accessors (usually <code>const</code>) if necessary.</p>
<p>For technical
reasons, we allow data members of a test fixture class defined in a .cc file to
be <code>protected</code> when using
<a href="">Google
If a test fixture class is defined outside of the .cc file it is used in, for example in a .h file,
make data members <code>private</code>.</p>
<h3 id="Declaration_Order">Declaration Order</h3>
<p>Group similar declarations together, placing public parts
<p>A class definition should usually start with a
<code>public:</code> section, followed by
<code>protected:</code>, then <code>private:</code>. Omit
sections that would be empty.</p>
<p>Within each section, prefer grouping similar
kinds of declarations together, and prefer the
following order: types (including <code>typedef</code>,
<code>using</code>, <code>enum</code>, and nested structs and classes),
constants, factory functions, constructors and assignment
operators, destructor, all other methods, data members.</p>
<p>Do not put large method definitions inline in the
class definition. Usually, only trivial or
performance-critical, and very short, methods may be
defined inline. See <a href="#Inline_Functions">Inline
Functions</a> for more details.</p>
<h2 id="Functions">Functions</h2>
<a id="Function_Parameter_Ordering"></a>
<a id="Output_Parameters"></a>
<h3 id="Inputs_and_Outputs">Inputs and Outputs</h3>
<p>The output of a C++ function is naturally provided via
a return value and sometimes via output parameters (or in/out parameters).</p>
<p>Prefer using return values over output parameters: they
improve readability, and often provide the same or better
<p>Prefer to return by value or, failing that, return by reference.
Avoid returning a pointer unless it can be null.</p>
<p>Parameters are either inputs to the function, outputs from the
function, or both. Non-optional input parameters should usually be values
or <code>const</code> references, while non-optional output and
input/output parameters should usually be references (which cannot be null).
Generally, use <code>absl::optional</code> to represent optional by-value
inputs, and use a <code>const</code> pointer when the non-optional form would
have used a reference. Use non-<code>const</code> pointers to represent
optional outputs and optional input/output parameters.</p>
Avoid defining functions that require a <code>const</code> reference parameter
to outlive the call, because <code>const</code> reference parameters bind
to temporaries. Instead, find a way to eliminate the lifetime requirement
(for example, by copying the parameter), or pass it by <code>const</code>
pointer and document the lifetime and non-null requirements.
<p>When ordering function parameters, put all input-only
parameters before any output parameters. In particular,
do not add new parameters to the end of the function just
because they are new; place new input-only parameters before
the output parameters. This is not a hard-and-fast rule. Parameters that
are both input and output muddy the waters, and, as always,
consistency with related functions may require you to bend the rule.
Variadic functions may also require unusual parameter ordering.</p>
<h3 id="Write_Short_Functions">Write Short Functions</h3>
<p>Prefer small and focused functions.</p>
<p>We recognize that long functions are sometimes
appropriate, so no hard limit is placed on functions
length. If a function exceeds about 40 lines, think about
whether it can be broken up without harming the structure
of the program.</p>
<p>Even if your long function works perfectly now,
someone modifying it in a few months may add new
behavior. This could result in bugs that are hard to
find. Keeping your functions short and simple makes it
easier for other people to read and modify your code.
Small functions are also easier to test.</p>
<p>You could find long and complicated functions when
working with
some code. Do not be
intimidated by modifying existing code: if working with
such a function proves to be difficult, you find that
errors are hard to debug, or you want to use a piece of
it in several different contexts, consider breaking up
the function into smaller and more manageable pieces.</p>
<h3 id="Function_Overloading">Function Overloading</h3>
<p>Use overloaded functions (including constructors) only if a
reader looking at a call site can get a good idea of what
is happening without having to first figure out exactly
which overload is being called.</p>
<p class="definition"></p>
<p>You may write a function that takes a <code>const
std::string&amp;</code> and overload it with another that
takes <code>const char*</code>. However, in this case consider
<pre>class MyClass {
void Analyze(const std::string &amp;text);
void Analyze(const char *text, size_t textlen);
<p class="pros"></p>
<p>Overloading can make code more intuitive by allowing an
identically-named function to take different arguments.
It may be necessary for templatized code, and it can be
convenient for Visitors.</p>
<p>Overloading based on const or ref qualification may make utility
code more usable, more efficient, or both.
(See <a href="">TotW 148</a> for more.)
<p class="cons"></p>
<p>If a function is overloaded by the argument types alone,
a reader may have to understand C++'s complex matching
rules in order to tell what's going on. Also many people
are confused by the semantics of inheritance if a derived
class overrides only some of the variants of a
<p class="decision"></p>
<p>You may overload a function when there are no semantic differences
between variants. These overloads may vary in types, qualifiers, or
argument count. However, a reader of such a call must not need to know
which member of the overload set is chosen, only that <b>something</b>
from the set is being called. If you can document all entries in the
overload set with a single comment in the header, that is a good sign
that it is a well-designed overload set.</p>
<h3 id="Default_Arguments">Default Arguments</h3>
<p>Default arguments are allowed on non-virtual functions
when the default is guaranteed to always have the same
value. Follow the same restrictions as for <a href="#Function_Overloading">function overloading</a>, and
prefer overloaded functions if the readability gained with
default arguments doesn't outweigh the downsides below.</p>
<p class="pros"></p>
<p>Often you have a function that uses default values, but
occasionally you want to override the defaults. Default
parameters allow an easy way to do this without having to
define many functions for the rare exceptions. Compared
to overloading the function, default arguments have a
cleaner syntax, with less boilerplate and a clearer
distinction between 'required' and 'optional'
<p class="cons"></p>
<p>Defaulted arguments are another way to achieve the
semantics of overloaded functions, so all the <a href="#Function_Overloading">reasons not to overload
functions</a> apply.</p>
<p>The defaults for arguments in a virtual function call are
determined by the static type of the target object, and
there's no guarantee that all overrides of a given function
declare the same defaults.</p>
<p>Default parameters are re-evaluated at each call site,
which can bloat the generated code. Readers may also expect
the default's value to be fixed at the declaration instead
of varying at each call.</p>
<p>Function pointers are confusing in the presence of
default arguments, since the function signature often
doesn't match the call signature. Adding
function overloads avoids these problems.</p>
<p class="decision"></p>
<p>Default arguments are banned on virtual functions, where
they don't work properly, and in cases where the specified
default might not evaluate to the same value depending on
when it was evaluated. (For example, don't write <code>void
f(int n = counter++);</code>.)</p>
<p>In some other cases, default arguments can improve the
readability of their function declarations enough to
overcome the downsides above, so they are allowed. When in
doubt, use overloads.</p>
<h3 id="trailing_return">Trailing Return Type Syntax</h3>
<p>Use trailing return types only where using the ordinary syntax (leading
return types) is impractical or much less readable.</p>
<p class="definition"></p>
<p>C++ allows two different forms of function declarations. In the older
form, the return type appears before the function name. For example:</p>
<pre>int foo(int x);
<p>The newer form, introduced in C++11, uses the <code>auto</code>
keyword before the function name and a trailing return type after
the argument list. For example, the declaration above could
equivalently be written:</p>
<pre>auto foo(int x) -&gt; int;
<p>The trailing return type is in the function's scope. This doesn't
make a difference for a simple case like <code>int</code> but it matters
for more complicated cases, like types declared in class scope or
types written in terms of the function parameters.</p>
<p class="pros"></p>
<p>Trailing return types are the only way to explicitly specify the
return type of a <a href="#Lambda_expressions">lambda expression</a>.
In some cases the compiler is able to deduce a lambda's return type,
but not in all cases. Even when the compiler can deduce it automatically,
sometimes specifying it explicitly would be clearer for readers.
<p>Sometimes it's easier and more readable to specify a return type
after the function's parameter list has already appeared. This is
particularly true when the return type depends on template parameters.
For example:</p>
<pre> template &lt;typename T, typename U&gt;
auto add(T t, U u) -&gt; decltype(t + u);
<pre> template &lt;typename T, typename U&gt;
decltype(declval&lt;T&amp;&gt;() + declval&lt;U&amp;&gt;()) add(T t, U u);
<p class="cons"></p>
<p>Trailing return type syntax is relatively new and it has no
analogue in C++-like languages such as C and Java, so some readers may
find it unfamiliar.</p>
<p>Existing code bases have an enormous number of function
declarations that aren't going to get changed to use the new syntax,
so the realistic choices are using the old syntax only or using a mixture
of the two. Using a single version is better for uniformity of style.</p>
<p class="decision"></p>
<p>In most cases, continue to use the older style of function
declaration where the return type goes before the function name.
Use the new trailing-return-type form only in cases where it's
required (such as lambdas) or where, by putting the type after the
function's parameter list, it allows you to write the type in a much
more readable way. The latter case should be rare; it's mostly an
issue in fairly complicated template code, which is
<a href="#Template_metaprogramming">discouraged in most cases</a>.</p>
<h2 id="Google-Specific_Magic">Google-Specific Magic</h2>
<p>There are various tricks and utilities that
we use to make C++ code more robust, and various ways we use
C++ that may differ from what you see elsewhere.</p>
<h3 id="Ownership_and_Smart_Pointers">Ownership and Smart Pointers</h3>
<p>Prefer to have single, fixed owners for dynamically
allocated objects. Prefer to transfer ownership with smart
<p class="definition"></p>
<p>"Ownership" is a bookkeeping technique for managing
dynamically allocated memory (and other resources). The
owner of a dynamically allocated object is an object or
function that is responsible for ensuring that it is
deleted when no longer needed. Ownership can sometimes be
shared, in which case the last owner is typically
responsible for deleting it. Even when ownership is not
shared, it can be transferred from one piece of code to
<p>"Smart" pointers are classes that act like pointers,
e.g., by overloading the <code>*</code> and
<code>-&gt;</code> operators. Some smart pointer types
can be used to automate ownership bookkeeping, to ensure
these responsibilities are met.
<a href="">
<code>std::unique_ptr</code></a> is a smart pointer type
introduced in C++11, which expresses exclusive ownership
of a dynamically allocated object; the object is deleted
when the <code>std::unique_ptr</code> goes out of scope.
It cannot be copied, but can be <em>moved</em> to
represent ownership transfer.
<a href="">
<code>std::shared_ptr</code></a> is a smart pointer type
that expresses shared ownership of
a dynamically allocated object. <code>std::shared_ptr</code>s
can be copied; ownership of the object is shared among
all copies, and the object is deleted when the last
<code>std::shared_ptr</code> is destroyed. </p>
<p class="pros"></p>
<li>It's virtually impossible to manage dynamically
allocated memory without some sort of ownership
<li>Transferring ownership of an object can be cheaper
than copying it (if copying it is even possible).</li>
<li>Transferring ownership can be simpler than
'borrowing' a pointer or reference, because it reduces
the need to coordinate the lifetime of the object
between the two users.</li>
<li>Smart pointers can improve readability by making
ownership logic explicit, self-documenting, and
<li>Smart pointers can eliminate manual ownership
bookkeeping, simplifying the code and ruling out large
classes of errors.</li>
<li>For const objects, shared ownership can be a simple
and efficient alternative to deep copying.</li>
<p class="cons"></p>
<li>Ownership must be represented and transferred via
pointers (whether smart or plain). Pointer semantics
are more complicated than value semantics, especially
in APIs: you have to worry not just about ownership,
but also aliasing, lifetime, and mutability, among
other issues.</li>
<li>The performance costs of value semantics are often
overestimated, so the performance benefits of ownership
transfer might not justify the readability and
complexity costs.</li>
<li>APIs that transfer ownership force their clients
into a single memory management model.</li>
<li>Code using smart pointers is less explicit about
where the resource releases take place.</li>
<li><code>std::unique_ptr</code> expresses ownership
transfer using C++11's move semantics, which are
relatively new and may confuse some programmers.</li>
<li>Shared ownership can be a tempting alternative to
careful ownership design, obfuscating the design of a
<li>Shared ownership requires explicit bookkeeping at
run-time, which can be costly.</li>
<li>In some cases (e.g., cyclic references), objects
with shared ownership may never be deleted.</li>
<li>Smart pointers are not perfect substitutes for
plain pointers.</li>
<p class="decision"></p>
<p>If dynamic allocation is necessary, prefer to keep
ownership with the code that allocated it. If other code
needs access to the object, consider passing it a copy,
or passing a pointer or reference without transferring
ownership. Prefer to use <code>std::unique_ptr</code> to
make ownership transfer explicit. For example:</p>
<pre>std::unique_ptr&lt;Foo&gt; FooFactory();
void FooConsumer(std::unique_ptr&lt;Foo&gt; ptr);
<p>Do not design your code to use shared ownership
without a very good reason. One such reason is to avoid
expensive copy operations, but you should only do this if
the performance benefits are significant, and the
underlying object is immutable (i.e.,
<code>std::shared_ptr&lt;const Foo&gt;</code>). If you
do use shared ownership, prefer to use
<p>Never use <code>std::auto_ptr</code>. Instead, use
<h3 id="cpplint">cpplint</h3>
<p>Use <code></code> to detect style errors.</p>
is a tool that reads a source file and identifies many
style errors. It is not perfect, and has both false
positives and false negatives, but it is still a valuable
tool. </p>
<p>Some projects have instructions on
how to run <code></code> from their project
tools. If the project you are contributing to does not,
you can download
<a href="">
<code></code></a> separately.</p>
<h2 id="Other_C++_Features">Other C++ Features</h2>
<h3 id="Rvalue_references">Rvalue References</h3>
<p>Use rvalue references only in certain special cases listed below.</p>
<p class="definition"></p>
<p> Rvalue references
are a type of reference that can only bind to temporary
objects. The syntax is similar to traditional reference
syntax. For example, <code>void f(std::string&amp;&amp;
s);</code> declares a function whose argument is an
rvalue reference to a std::string.</p>
<p id="Forwarding_references"> When the token '&amp;&amp;' is applied to
an unqualified template argument in a function
parameter, special template argument deduction
rules apply. Such a reference is called forwarding reference.</p>
<p class="pros"></p>
<li>Defining a move constructor (a constructor taking
an rvalue reference to the class type) makes it
possible to move a value instead of copying it. If
<code>v1</code> is a <code>std::vector&lt;std::string&gt;</code>,
for example, then <code>auto v2(std::move(v1))</code>
will probably just result in some simple pointer
manipulation instead of copying a large amount of data.
In many cases this can result in a major performance
<li>Rvalue references make it possible to implement
types that are movable but not copyable, which can be
useful for types that have no sensible definition of
copying but where you might still want to pass them as
function arguments, put them in containers, etc.</li>
<li><code>std::move</code> is necessary to make
effective use of some standard-library types, such as
<li><a href="#Forwarding_references">Forwarding references</a> which
use the rvalue reference token, make it possible to write a
generic function wrapper that forwards its arguments to
another function, and works whether or not its
arguments are temporary objects and/or const.
This is called 'perfect forwarding'.</li>
<p class="cons"></p>
<li>Rvalue references are not yet widely understood. Rules like reference
collapsing and the special deduction rule for forwarding references
are somewhat obscure.</li>
<li>Rvalue references are often misused. Using rvalue
references is counter-intuitive in signatures where the argument is expected
to have a valid specified state after the function call, or where no move
operation is performed.</li>
<p class="decision"></p>
<p>Do not use rvalue references (or apply the <code>&amp;&amp;</code>
qualifier to methods), except as follows:</p>
<li>You may use them to define move constructors and move assignment
operators (as described in
<a href="#Copyable_Movable_Types">Copyable and Movable Types</a>).
<li>You may use them to define <code>&amp;&amp;</code>-qualified methods that
logically "consume" <code>*this</code>, leaving it in an unusable
or empty state. Note that this applies only to method qualifiers (which come
after the closing parenthesis of the function signature); if you want to
"consume" an ordinary function parameter, prefer to pass it by value.</li>
<li>You may use forwarding references in conjunction with <code>
<a href="">std::forward</a></code>,
to support perfect forwarding.</li>
<li>You may use them to define pairs of overloads, such as one taking
<code>Foo&amp;&amp;</code> and the other taking <code>const Foo&amp;</code>.
Usually the preferred solution is just to pass by value, but an overloaded
pair of functions sometimes yields better performance and is sometimes
necessary in generic code that needs to support a wide variety of types.
As always: if you're writing more complicated code for the sake of
performance, make sure you have evidence that it actually helps.</li>
<h3 id="Friends">Friends</h3>
<p>We allow use of <code>friend</code> classes and functions,
within reason.</p>
<p>Friends should usually be defined in the same file so
that the reader does not have to look in another file to
find uses of the private members of a class. A common use
of <code>friend</code> is to have a
<code>FooBuilder</code> class be a friend of
<code>Foo</code> so that it can construct the inner state
of <code>Foo</code> correctly, without exposing this
state to the world. In some cases it may be useful to
make a unittest class a friend of the class it tests.</p>
<p>Friends extend, but do not break, the encapsulation
boundary of a class. In some cases this is better than
making a member public when you want to give only one
other class access to it. However, most classes should
interact with other classes solely through their public
<h3 id="Exceptions">Exceptions</h3>
<p>We do not use C++ exceptions.</p>
<p class="pros"></p>
<li>Exceptions allow higher levels of an application to
decide how to handle "can't happen" failures in deeply
nested functions, without the obscuring and error-prone
bookkeeping of error codes.</li>
<li>Exceptions are used by most other
modern languages. Using them in C++ would make it more
consistent with Python, Java, and the C++ that others
are familiar with.</li>
<li>Some third-party C++ libraries use exceptions, and
turning them off internally makes it harder to
integrate with those libraries.</li>
<li>Exceptions are the only way for a constructor to
fail. We can simulate this with a factory function or
an <code>Init()</code> method, but these require heap
allocation or a new "invalid" state, respectively.</li>
<li>Exceptions are really handy in testing
<p class="cons"></p>
<li>When you add a <code>throw</code> statement to an
existing function, you must examine all of its
transitive callers. Either they must make at least the
basic exception safety guarantee, or they must never
catch the exception and be happy with the program
terminating as a result. For instance, if
<code>f()</code> calls <code>g()</code> calls
<code>h()</code>, and <code>h</code> throws an
exception that <code>f</code> catches, <code>g</code>
has to be careful or it may not clean up properly.</li>
<li>More generally, exceptions make the control flow of
programs difficult to evaluate by looking at code:
functions may return in places you don't expect. This
causes maintainability and debugging difficulties. You
can minimize this cost via some rules on how and where
exceptions can be used, but at the cost of more that a
developer needs to know and understand.</li>
<li>Exception safety requires both RAII and different
coding practices. Lots of supporting machinery is
needed to make writing correct exception-safe code
easy. Further, to avoid requiring readers to understand
the entire call graph, exception-safe code must isolate
logic that writes to persistent state into a "commit"
phase. This will have both benefits and costs (perhaps
where you're forced to obfuscate code to isolate the
commit). Allowing exceptions would force us to always
pay those costs even when they're not worth it.</li>
<li>Turning on exceptions adds data to each binary
produced, increasing compile time (probably slightly)
and possibly increasing address space pressure.
<li>The availability of exceptions may encourage
developers to throw them when they are not appropriate
or recover from them when it's not safe to do so. For
example, invalid user input should not cause exceptions
to be thrown. We would need to make the style guide
even longer to document these restrictions!</li>
<p class="decision"></p>
<p>On their face, the benefits of using exceptions
outweigh the costs, especially in new projects. However,
for existing code, the introduction of exceptions has
implications on all dependent code. If exceptions can be
propagated beyond a new project, it also becomes
problematic to integrate the new project into existing
exception-free code. Because most existing C++ code at
Google is not prepared to deal with exceptions, it is
comparatively difficult to adopt new code that generates
<p>Given that Google's existing code is not
exception-tolerant, the costs of using exceptions are
somewhat greater than the costs in a new project. The
conversion process would be slow and error-prone. We
don't believe that the available alternatives to
exceptions, such as error codes and assertions, introduce
a significant burden. </p>
<p>Our advice against using exceptions is not predicated
on philosophical or moral grounds, but practical ones.
Because we'd like to use our open-source
projects at Google and it's difficult to do so if those
projects use exceptions, we need to advise against
exceptions in Google open-source projects as well.
Things would probably be different if we had to do it all
over again from scratch.</p>
<p>This prohibition also applies to the exception handling related
features added in C++11, such as
<code>std::exception_ptr</code> and
<p>There is an <a href="#Windows_Code">exception</a> to
this rule (no pun intended) for Windows code.</p>
<h3 id="noexcept"><code>noexcept</code></h3>
<p>Specify <code>noexcept</code> when it is useful and correct.</p>
<p class="definition"></p>
<p>The <code>noexcept</code> specifier is used to specify whether
a function will throw exceptions or not. If an exception
escapes from a function marked <code>noexcept</code>, the program
crashes via <code>std::terminate</code>.</p>
<p>The <code>noexcept</code> operator performs a compile-time
check that returns true if an expression is declared to not
throw any exceptions.</p>
<p class="pros"></p>
<li>Specifying move constructors as <code>noexcept</code>
improves performance in some cases, e.g.,
<code>std::vector&lt;T&gt;::resize()</code> moves rather than
copies the objects if T's move constructor is
<li>Specifying <code>noexcept</code> on a function can
trigger compiler optimizations in environments where
exceptions are enabled, e.g., compiler does not have to
generate extra code for stack-unwinding, if it knows
that no exceptions can be thrown due to a
<code>noexcept</code> specifier.</li>
<p class="cons"></p>
In projects following this guide
that have exceptions disabled it is hard
to ensure that <code>noexcept</code>
specifiers are correct, and hard to define what
correctness even means.</li>
<li>It's hard, if not impossible, to undo <code>noexcept</code>
because it eliminates a guarantee that callers may be relying
on, in ways that are hard to detect.</li>
<p class="decision"></p>
<p>You may use <code>noexcept</code> when it is useful for
performance if it accurately reflects the intended semantics
of your function, i.e., that if an exception is somehow thrown
from within the function body then it represents a fatal error.
You can assume that <code>noexcept</code> on move constructors
has a meaningful performance benefit. If you think
there is significant performance benefit from specifying
<code>noexcept</code> on some other function, please discuss it
your project leads.</p>
<p>Prefer unconditional <code>noexcept</code> if exceptions are
completely disabled (i.e., most Google C++ environments).
Otherwise, use conditional <code>noexcept</code> specifiers
with simple conditions, in ways that evaluate false only in
the few cases where the function could potentially throw.
The tests might include type traits check on whether the
involved operation might throw (e.g.,
<code>std::is_nothrow_move_constructible</code> for
move-constructing objects), or on whether allocation can throw
(e.g., <code>absl::default_allocator_is_nothrow</code> for
standard default allocation). Note in many cases the only
possible cause for an exception is allocation failure (we
believe move constructors should not throw except due to
allocation failure), and there are many applications where it’s
appropriate to treat memory exhaustion as a fatal error rather
than an exceptional condition that your program should attempt
to recover from. Even for other
potential failures you should prioritize interface simplicity
over supporting all possible exception throwing scenarios:
instead of writing a complicated <code>noexcept</code> clause
that depends on whether a hash function can throw, for example,
simply document that your component doesn’t support hash
functions throwing and make it unconditionally
<h3 id="Run-Time_Type_Information__RTTI_">Run-Time Type
Information (RTTI)</h3>
<p>Avoid using run-time type information (RTTI).</p>
<p class="definition"></p>
<p> RTTI allows a
programmer to query the C++ class of an object at
run-time. This is done by use of <code>typeid</code> or
<p class="pros"></p>
<p>The standard alternatives to RTTI (described below)
require modification or redesign of the class hierarchy
in question. Sometimes such modifications are infeasible
or undesirable, particularly in widely-used or mature
<p>RTTI can be useful in some unit tests. For example, it
is useful in tests of factory classes where the test has
to verify that a newly created object has the expected
dynamic type. It is also useful in managing the
relationship between objects and their mocks.</p>
<p>RTTI is useful when considering multiple abstract
objects. Consider</p>
<pre>bool Base::Equal(Base* other) = 0;
bool Derived::Equal(Base* other) {
Derived* that = dynamic_cast&lt;Derived*&gt;(other);
if (that == nullptr)
return false;
<p class="cons"></p>
<p>Querying the type of an object at run-time frequently
means a design problem. Needing to know the type of an
object at runtime is often an indication that the design
of your class hierarchy is flawed.</p>
<p>Undisciplined use of RTTI makes code hard to maintain.
It can lead to type-based decision trees or switch
statements scattered throughout the code, all of which
must be examined when making further changes.</p>
<p class="decision"></p>
<p>RTTI has legitimate uses but is prone to abuse, so you
must be careful when using it. You may use it freely in
unittests, but avoid it when possible in other code. In
particular, think twice before using RTTI in new code. If
you find yourself needing to write code that behaves
differently based on the class of an object, consider one
of the following alternatives to querying the type:</p>
<li>Virtual methods are the preferred way of executing
different code paths depending on a specific subclass
type. This puts the work within the object itself.</li>
<li>If the work belongs outside the object and instead
in some processing code, consider a double-dispatch
solution, such as the Visitor design pattern. This
allows a facility outside the object itself to
determine the type of class using the built-in type
<p>When the logic of a program guarantees that a given
instance of a base class is in fact an instance of a
particular derived class, then a
<code>dynamic_cast</code> may be used freely on the
object. Usually one
can use a <code>static_cast</code> as an alternative in
such situations.</p>
<p>Decision trees based on type are a strong indication
that your code is on the wrong track.</p>
<pre class="badcode">if (typeid(*data) == typeid(D1)) {
} else if (typeid(*data) == typeid(D2)) {
} else if (typeid(*data) == typeid(D3)) {
<p>Code such as this usually breaks when additional
subclasses are added to the class hierarchy. Moreover,
when properties of a subclass change, it is difficult to
find and modify all the affected code segments.</p>
<p>Do not hand-implement an RTTI-like workaround. The
arguments against RTTI apply just as much to workarounds
like class hierarchies with type tags. Moreover,
workarounds disguise your true intent.</p>
<h3 id="Casting">Casting</h3>
<p>Use C++-style casts
like <code>static_cast&lt;float&gt;(double_value)</code>, or brace
initialization for conversion of arithmetic types like
<code>int64_t y = int64_t{1} &lt;&lt; 42</code>. Do not use
cast formats like <code>(int)x</code> unless the cast is to
<code>void</code>. You may use cast formats like `T(x)` only when
`T` is a class type.</p>
<p class="definition"></p>
<p> C++ introduced a
different cast system from C that distinguishes the types
of cast operations.</p>
<p class="pros"></p>
<p>The problem with C casts is the ambiguity of the operation;
sometimes you are doing a <em>conversion</em>
(e.g., <code>(int)3.5</code>) and sometimes you are doing
a <em>cast</em> (e.g., <code>(int)"hello"</code>). Brace
initialization and C++ casts can often help avoid this
ambiguity. Additionally, C++ casts are more visible when searching for
<p class="cons"></p>
<p>The C++-style cast syntax is verbose and cumbersome.</p>
<p class="decision"></p>
<p>In general, do not use C-style casts. Instead, use these C++-style
casts when explicit type conversion is necessary.
<li>Use brace initialization to convert arithmetic types
(e.g., <code>int64_t{x}</code>). This is the safest approach because code
will not compile if conversion can result in information loss. The
syntax is also concise.</li>
<li>Use <code>static_cast</code> as the equivalent of a C-style cast
that does value conversion, when you need to
explicitly up-cast a pointer from a class to its superclass, or when
you need to explicitly cast a pointer from a superclass to a
subclass. In this last case, you must be sure your object is
actually an instance of the subclass.</li>
<li>Use <code>const_cast</code> to remove the
<code>const</code> qualifier (see <a href="#Use_of_const">const</a>).</li>
<li>Use <code>reinterpret_cast</code> to do unsafe conversions of
pointer types to and from integer and other pointer
including <code>void*</code>. Use this
only if you know what you are doing and you understand the aliasing
issues. Also, consider the alternative
<li>Use <code>absl::bit_cast</code> to interpret the raw bits of a
value using a different type of the same size (a type pun), such as
interpreting the bits of a <code>double</code> as
<p>See the <a href="#Run-Time_Type_Information__RTTI_">
RTTI section</a> for guidance on the use of
<h3 id="Streams">Streams</h3>
<p>Use streams where appropriate, and stick to "simple"
usages. Overload <code>&lt;&lt;</code> for streaming only for types
representing values, and write only the user-visible value, not any
implementation details.</p>
<p class="definition"></p>
<p>Streams are the standard I/O abstraction in C++, as
exemplified by the standard header <code>&lt;iostream&gt;</code>.
They are widely used in Google code, mostly for debug logging
and test diagnostics.</p>
<p class="pros"></p>
<p>The <code>&lt;&lt;</code> and <code>&gt;&gt;</code>
stream operators provide an API for formatted I/O that
is easily learned, portable, reusable, and extensible.
<code>printf</code>, by contrast, doesn't even support
<code>std::string</code>, to say nothing of user-defined types,
and is very difficult to use portably.
<code>printf</code> also obliges you to choose among the
numerous slightly different versions of that function,
and navigate the dozens of conversion specifiers.</p>
<p>Streams provide first-class support for console I/O
via <code>std::cin</code>, <code>std::cout</code>,
<code>std::cerr</code>, and <code>std::clog</code>.
The C APIs do as well, but are hampered by the need to
manually buffer the input. </p>
<p class="cons"></p>
<li>Stream formatting can be configured by mutating the
state of the stream. Such mutations are persistent, so
the behavior of your code can be affected by the entire
previous history of the stream, unless you go out of your
way to restore it to a known state every time other code
might have touched it. User code can not only modify the
built-in state, it can add new state variables and behaviors
through a registration system.</li>
<li>It is difficult to precisely control stream output, due
to the above issues, the way code and data are mixed in
streaming code, and the use of operator overloading (which
may select a different overload than you expect).</li>
<li>The practice of building up output through chains
of <code>&lt;&lt;</code> operators interferes with
internationalization, because it bakes word order into the
code, and streams' support for localization is <a href="">
<li>The streams API is subtle and complex, so programmers must
develop experience with it in order to use it effectively.</li>
<li>Resolving the many overloads of <code>&lt;&lt;</code> is
extremely costly for the compiler. When used pervasively in a
large code base, it can consume as much as 20% of the parsing
and semantic analysis time.</li>
<p class="decision"></p>
<p>Use streams only when they are the best tool for the job.
This is typically the case when the I/O is ad-hoc, local,
human-readable, and targeted at other developers rather than
end-users. Be consistent with the code around you, and with the
codebase as a whole; if there's an established tool for
your problem, use that tool instead.
In particular,
logging libraries are usually a better
choice than <code>std::cerr</code> or <code>std::clog</code>
for diagnostic output, and the libraries in
or the equivalent are usually a
better choice than <code>std::stringstream</code>.</p>
<p>Avoid using streams for I/O that faces external users or
handles untrusted data. Instead, find and use the appropriate
templating libraries to handle issues like internationalization,
localization, and security hardening.</p>
<p>If you do use streams, avoid the stateful parts of the
streams API (other than error state), such as <code>imbue()</code>,
<code>xalloc()</code>, and <code>register_callback()</code>.
Use explicit formatting functions (see e.g.,
rather than
stream manipulators or formatting flags to control formatting
details such as number base, precision, or padding.</p>
<p>Overload <code>&lt;&lt;</code> as a streaming operator
for your type only if your type represents a value, and
<code>&lt;&lt;</code> writes out a human-readable string
representation of that value. Avoid exposing implementation
details in the output of <code>&lt;&lt;</code>; if you need to print
object internals for debugging, use named functions instead
(a method named <code>DebugString()</code> is the most common
<h3 id="Preincrement_and_Predecrement">Preincrement and Predecrement</h3>
<p>Use the prefix form (<code>++i</code>) of the increment
and decrement operators unless you need postfix semantics.</p>
<p class="definition"></p>
<p> When a variable
is incremented (<code>++i</code> or <code>i++</code>) or
decremented (<code>--i</code> or <code>i--</code>) and
the value of the expression is not used, one must decide
whether to preincrement (decrement) or postincrement
<p class="pros"></p>
<p>A postfix increment/decrement expression evaluates to the value
<i>as it was before it was modified</i>. This can result in code that is more
compact but harder to read. The prefix form is generally more readable, is
never less efficient, and can be more efficient because it doesn't need to
make a copy of the value as it was before the operation.
<p class="cons"></p>
<p>The tradition developed, in C, of using post-increment, even
when the expression value is not used, especially in
<code>for</code> loops.</p>
<p class="decision"></p>
<p>Use prefix increment/decrement, unless the code explicitly
needs the result of the postfix increment/decrement expression.</p>
<h3 id="Use_of_const">Use of const</h3>
<p>In APIs, use <code>const</code> whenever it makes sense.
<code>constexpr</code> is a better choice for some uses of
<p class="definition"></p>
<p> Declared variables and parameters can be preceded
by the keyword <code>const</code> to indicate the variables
are not changed (e.g., <code>const int foo</code>). Class
functions can have the <code>const</code> qualifier to
indicate the function does not change the state of the
class member variables (e.g., <code>class Foo { int
Bar(char c) const; };</code>).</p>
<p class="pros"></p>
<p>Easier for people to understand how variables are being
used. Allows the compiler to do better type checking,
and, conceivably, generate better code. Helps people
convince themselves of program correctness because they
know the functions they call are limited in how they can
modify your variables. Helps people know what functions
are safe to use without locks in multi-threaded
<p class="cons"></p>
<p><code>const</code> is viral: if you pass a
<code>const</code> variable to a function, that function
must have <code>const</code> in its prototype (or the
variable will need a <code>const_cast</code>). This can
be a particular problem when calling library
<p class="decision"></p>
<p>We strongly recommend using <code>const</code>
in APIs (i.e., on function parameters, methods, and
non-local variables) wherever it is meaningful and accurate. This
provides consistent, mostly compiler-verified documentation
of what objects an operation can mutate. Having
a consistent and reliable way to distinguish reads from writes
is critical to writing thread-safe code, and is useful in
many other contexts as well. In particular:</p>
<li>If a function guarantees that it will not modify an argument
passed by reference or by pointer, the corresponding function parameter
should be a reference-to-const (<code>const T&amp;</code>) or
pointer-to-const (<code>const T*</code>), respectively.</li>
<li>For a function parameter passed by value, <code>const</code> has
no effect on the caller, thus is not recommended in function
declarations. See
<a href="">TotW #109</a>.
</li><li>Declare methods to be <code>const</code> unless they
alter the logical state of the object (or enable the user to modify
that state, e.g., by returning a non-const reference, but that's
rare), or they can't safely be invoked concurrently.</li>
<p>Using <code>const</code> on local variables is neither encouraged
nor discouraged.</p>
<p>All of a class's <code>const</code> operations should be safe
to invoke concurrently with each other. If that's not feasible, the class must
be clearly documented as "thread-unsafe".</p>
<h4>Where to put the const</h4>
<p>Some people favor the form <code>int const *foo</code>
to <code>const int* foo</code>. They argue that this is
more readable because it's more consistent: it keeps the
rule that <code>const</code> always follows the object
it's describing. However, this consistency argument
doesn't apply in codebases with few deeply-nested pointer
expressions since most <code>const</code> expressions
have only one <code>const</code>, and it applies to the
underlying value. In such cases, there's no consistency
to maintain. Putting the <code>const</code> first is
arguably more readable, since it follows English in
putting the "adjective" (<code>const</code>) before the
"noun" (<code>int</code>).</p>
<p>That said, while we encourage putting
<code>const</code> first, we do not require it. But be
consistent with the code around you!</p>
<h3 id="Use_of_constexpr">Use of constexpr</h3>
<p>Use <code>constexpr</code> to define true
constants or to ensure constant initialization.</p>
<p class="definition"></p>
<p> Some variables can be declared <code>constexpr</code>
to indicate the variables are true constants, i.e., fixed at
compilation/link time. Some functions and constructors
can be declared <code>constexpr</code> which enables them
to be used in defining a <code>constexpr</code>
<p class="pros"></p>
<p>Use of <code>constexpr</code> enables definition of
constants with floating-point expressions rather than
just literals; definition of constants of user-defined
types; and definition of constants with function
<p class="cons"></p>
<p>Prematurely marking something as constexpr may cause
migration problems if later on it has to be downgraded.
Current restrictions on what is allowed in constexpr
functions and constructors may invite obscure workarounds
in these definitions.</p>
<p class="decision"></p>
<p><code>constexpr</code> definitions enable a more
robust specification of the constant parts of an
interface. Use <code>constexpr</code> to specify true
constants and the functions that support their
definitions. Avoid complexifying function definitions to
enable their use with <code>constexpr</code>. Do not use
<code>constexpr</code> to force inlining.</p>
<h3 id="Integer_Types">Integer Types</h3>
<p>Of the built-in C++ integer types, the only one used
<code>int</code>. If a program needs a variable of a
different size, use a precise-width integer type from
<code>&lt;stdint.h&gt;</code>, such as
<code>int16_t</code>. If your variable represents a
value that could ever be greater than or equal to 2^31
(2GiB), use a 64-bit type such as <code>int64_t</code>.
Keep in mind that even if your value won't ever be too large
for an <code>int</code>, it may be used in intermediate
calculations which may require a larger type. When in doubt,
choose a larger type.</p>
<p class="definition"></p>
<p> C++ does not specify the sizes of integer types
like <code>int</code>. Typically people assume
that <code>short</code> is 16 bits,
<code>int</code> is 32 bits, <code>long</code> is 32 bits
and <code>long long</code> is 64 bits.</p>
<p class="pros"></p>
<p>Uniformity of declaration.</p>
<p class="cons"></p>
<p>The sizes of integral types in C++ can vary based on
compiler and architecture.</p>
<p class="decision"></p>
<p>We use <code>int</code> very often, for integers we
know are not going to be too big, e.g., loop counters.
Use plain old <code>int</code> for such things. You
should assume that an <code>int</code> is
at least 32 bits, but don't
assume that it has more than 32 bits. If you need a 64-bit
integer type, use <code>int64_t</code> or <code>uint64_t</code>.
</p><p>For integers we know can be "big",
<p>You should not use the unsigned integer types such as
<code>uint32_t</code>, unless there is a valid
reason such as representing a bit pattern rather than a
number, or you need defined overflow modulo 2^N. In
particular, do not use unsigned types to say a number
will never be negative. Instead, use
assertions for this.</p>
<p>If your code is a container that returns a size, be
sure to use a type that will accommodate any possible
usage of your container. When in doubt, use a larger type
rather than a smaller type.</p>
<p>Use care when converting integer types. Integer conversions and
promotions can cause undefined behavior, leading to security bugs and
other problems.</p>
<h4>On Unsigned Integers</h4>
<p>Unsigned integers are good for representing bitfields and modular
arithmetic. Because of historical accident, the C++ standard also uses
unsigned integers to represent the size of containers - many members
of the standards body believe this to be a mistake, but it is
effectively impossible to fix at this point. The fact that unsigned
arithmetic doesn't model the behavior of a simple integer, but is
instead defined by the standard to model modular arithmetic (wrapping
around on overflow/underflow), means that a significant class of bugs
cannot be diagnosed by the compiler. In other cases, the defined
behavior impedes optimization.</p>
<p>That said, mixing signedness of integer types is responsible for an
equally large class of problems. The best advice we can provide: try
to use iterators and containers rather than pointers and sizes, try
not to mix signedness, and try to avoid unsigned types (except for
representing bitfields or modular arithmetic). Do not use an unsigned
type merely to assert that a variable is non-negative.</p>
<h3 id="64-bit_Portability">64-bit Portability</h3>
<p>Code should be 64-bit and 32-bit friendly. Bear in mind
problems of printing, comparisons, and structure alignment.</p>
<p>Correct portable <code>printf()</code> conversion specifiers for
some integral typedefs rely on macro expansions that we find unpleasant to
use and impractical to require (the <code>PRI</code> macros from
<code>&lt;cinttypes&gt;</code>). Unless there is no reasonable alternative
for your particular case, try to avoid or even upgrade APIs that rely on the
<code>printf</code> family. Instead use a library supporting typesafe numeric
formatting, such as
<a href=""><code>StrCat</code></a>
<a href=""><code>Substitute</code></a>
for fast simple conversions,
or <a href="#Streams"><code>std::ostream</code></a>.</p>
<p>Unfortunately, the <code>PRI</code> macros are the only portable way to
specify a conversion for the standard bitwidth typedefs (e.g.,
<code>int64_t</code>, <code>uint64_t</code>, <code>int32_t</code>,
<code>uint32_t</code>, etc).
Where possible, avoid passing arguments of types specified by bitwidth
typedefs to <code>printf</code>-based APIs. Note that it is acceptable
to use typedefs for which printf has dedicated length modifiers, such as
<code>size_t</code> (<code>z</code>),
<code>ptrdiff_t</code> (<code>t</code>), and
<code>maxint_t</code> (<code>j</code>).</p>
<li>Remember that <code>sizeof(void *)</code> !=
<code>sizeof(int)</code>. Use <code>intptr_t</code> if
you want a pointer-sized integer.</li>
<li>You may need to be careful with structure
alignments, particularly for structures being stored on
disk. Any class/structure with a <code>int64_t</code>/<code>uint64_t</code>
member will by default end up being 8-byte aligned on a
64-bit system. If you have such structures being shared
on disk between 32-bit and 64-bit code, you will need
to ensure that they are packed the same on both
Most compilers offer a way to
alter structure alignment. For gcc, you can use
<code>__attribute__((packed))</code>. MSVC offers
<code>#pragma pack()</code> and
<p>Use <a href="#Casting">braced-initialization</a> as needed to create
64-bit constants. For example:</p>
<pre>int64_t my_value{0x123456789};
uint64_t my_mask{3ULL &lt;&lt; 48};
<h3 id="Preprocessor_Macros">Preprocessor Macros</h3>
<p>Avoid defining macros, especially in headers; prefer
inline functions, enums, and <code>const</code> variables.
Name macros with a project-specific prefix. Do not use
macros to define pieces of a C++ API.</p>
<p>Macros mean that the code you see is not the same as
the code the compiler sees. This can introduce unexpected
behavior, especially since macros have global scope.</p>
<p>The problems introduced by macros are especially severe
when they are used to define pieces of a C++ API,
and still more so for public APIs. Every error message from
the compiler when developers incorrectly use that interface
now must explain how the macros formed the interface.
Refactoring and analysis tools have a dramatically harder
time updating the interface. As a consequence, we
specifically disallow using macros in this way.
For example, avoid patterns like:</p>
<pre class="badcode">class WOMBAT_TYPE(Foo) {
// ...
<p>Luckily, macros are not nearly as necessary in C++ as
they are in C. Instead of using a macro to inline
performance-critical code, use an inline function.
Instead of using a macro to store a constant, use a
<code>const</code> variable. Instead of using a macro to
"abbreviate" a long variable name, use a reference.
Instead of using a macro to conditionally compile code
... well, don't do that at all (except, of course, for
the <code>#define</code> guards to prevent double
inclusion of header files). It makes testing much more
<p>Macros can do things these other techniques cannot,
and you do see them in the codebase, especially in the
lower-level libraries. And some of their special features
(like stringifying, concatenation, and so forth) are not
available through the language proper. But before using a
macro, consider carefully whether there's a non-macro way
to achieve the same result. If you need to use a macro to
define an interface, contact
your project leads to request
a waiver of this rule.</p>
<p>The following usage pattern will avoid many problems
with macros; if you use macros, follow it whenever
<li>Don't define macros in a <code>.h</code> file.</li>
<li><code>#define</code> macros right before you use
them, and <code>#undef</code> them right after.</li>
<li>Do not just <code>#undef</code> an existing macro
before replacing it with your own; instead, pick a name
that's likely to be unique.</li>
<li>Try not to use macros that expand to unbalanced C++
constructs, or at least document that behavior
<li>Prefer not using <code>##</code> to generate
function/class/variable names.</li>
<p>Exporting macros from headers (i.e., defining them in a header
without <code>#undef</code>ing them before the end of the header)
is extremely strongly discouraged. If you do export a macro from a
header, it must have a globally unique name. To achieve this, it
must be named with a prefix consisting of your project's namespace
name (but upper case). </p>
<h3 id="0_and_nullptr/NULL">0 and nullptr/NULL</h3>
<p>Use <code>nullptr</code> for pointers, and <code>'\0'</code> for chars (and
not the <code>0</code> literal).</p>
<p>For pointers (address values), use <code>nullptr</code>, as this
provides type-safety.</p>
<p>For C++03 projects, prefer <code>NULL</code> to <code>0</code>. While the
values are equivalent, <code>NULL</code> looks more like a pointer to the
reader, and some C++ compilers provide special definitions of <code>NULL</code>
which enable them to give useful warnings. Never use <code>NULL</code> for
numeric (integer or floating-point) values.</p>
<p>Use <code>'\0'</code> for the null character. Using the correct type makes
the code more readable.</p>
<h3 id="sizeof">sizeof</h3>
<p>Prefer <code>sizeof(<var>varname</var>)</code> to
<p>Use <code>sizeof(<var>varname</var>)</code> when you
take the size of a particular variable.
<code>sizeof(<var>varname</var>)</code> will update
appropriately if someone changes the variable type either
now or later. You may use
<code>sizeof(<var>type</var>)</code> for code unrelated
to any particular variable, such as code that manages an
external or internal data format where a variable of an
appropriate C++ type is not convenient.</p>
<pre>MyStruct data;
memset(&amp;data, 0, sizeof(data));
<pre class="badcode">memset(&amp;data, 0, sizeof(MyStruct));
<pre>if (raw_size &lt; sizeof(int)) {
LOG(ERROR) &lt;&lt; "compressed record not big enough for count: " &lt;&lt; raw_size;
return false;
<a id="auto"></a>
<h3 id="Type_deduction">Type Deduction (including auto)</h3>
<p>Use type deduction only if it makes the code clearer to readers who aren't
familiar with the project, or if it makes the code safer. Do not use it
merely to avoid the inconvenience of writing an explicit type.</p>
<p class="definition"></p>
<p>There are several contexts in which C++ allows (or even requires) types to
be deduced by the compiler, rather than spelled out explicitly in the code:</p>
<dt><a href="">Function template argument deduction</a></dt>
<dd>A function template can be invoked without explicit template arguments.
The compiler deduces those arguments from the types of the function
<pre class="neutralcode">template &lt;typename T&gt;
void f(T t);
f(0); // Invokes f&lt;int&gt;(0)</pre>
<dt><a href=""><code>auto</code> variable declarations</a></dt>
<dd>A variable declaration can use the <code>auto</code> keyword in place
of the type. The compiler deduces the type from the variable's
initializer, following the same rules as function template argument
deduction with the same initializer (so long as you don't use curly braces
instead of parentheses).
<pre class="neutralcode">auto a = 42; // a is an int
auto&amp; b = a; // b is an int&amp;
auto c = b; // c is an int
auto d{42}; // d is an int, not a std::initializer_list&lt;int&gt;
<code>auto</code> can be qualified with <code>const</code>, and can be
used as part of a pointer or reference type, but it can't be used as a
template argument. A rare variant of this syntax uses
<code>decltype(auto)</code> instead of <code>auto</code>, in which case
the deduced type is the result of applying
<a href=""><code>decltype</code></a>
to the initializer.
<dt><a href="">Function return type deduction</a></dt>
<dd><code>auto</code> (and <code>decltype(auto)</code>) can also be used in
place of a function return type. The compiler deduces the return type from
the <code>return</code> statements in the function body, following the same
rules as for variable declarations:
<pre class="neutralcode">auto f() { return 0; } // The return type of f is int</pre>
<a href="#Lambda_expressions">Lambda expression</a> return types can be
deduced in the same way, but this is triggered by omitting the return type,
rather than by an explicit <code>auto</code>. Confusingly,
<a href="#trailing_return">trailing return type</a> syntax for functions
also uses <code>auto</code> in the return-type position, but that doesn't
rely on type deduction; it's just an alternate syntax for an explicit
return type.
<dt><a href="">Generic lambdas</a></dt>
<dd>A lambda expression can use the <code>auto</code> keyword in place of
one or more of its parameter types. This causes the lambda's call operator
to be a function template instead of an ordinary function, with a separate
template parameter for each <code>auto</code> function parameter:
<pre class="neutralcode">// Sort `vec` in decreasing order
std::sort(vec.begin(), vec.end(), [](auto lhs, auto rhs) { return lhs &gt; rhs; });</pre>
<dt><a href="">Lambda init captures</a></dt>
<dd>Lambda captures can have explicit initializers, which can be used to
declare wholly new variables rather than only capturing existing ones:
<pre class="neutralcode">[x = 42, y = "foo"] { ... } // x is an int, and y is a const char*</pre>
This syntax doesn't allow the type to be specified; instead, it's deduced
using the rules for <code>auto</code> variables.
<dt><a href="">Class template argument deduction</a></dt>
<dd>See <a href="#CTAD">below</a>.</dd>
<dt><a href