freebsd-nq/docs/AttributeReference.rst

..
  -------------------------------------------------------------------
  NOTE: This file is automatically generated by running clang-tblgen
  -gen-attr-docs. Do not edit this file by hand!!
  -------------------------------------------------------------------

===================
Attributes in Clang
===================
.. contents::
   :local:

Introduction
============

This page lists the attributes currently supported by Clang.

Function Attributes
===================


interrupt
---------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

Clang supports the GNU style ``__attribute__((interrupt("TYPE")))`` attribute on
ARM targets. This attribute may be attached to a function definition and
instructs the backend to generate appropriate function entry/exit code so that
it can be used directly as an interrupt service routine.

The parameter passed to the interrupt attribute is optional, but if
provided it must be a string literal with one of the following values: "IRQ",
"FIQ", "SWI", "ABORT", "UNDEF".

The semantics are as follows:

- If the function is AAPCS, Clang instructs the backend to realign the stack to
  8 bytes on entry. This is a general requirement of the AAPCS at public
  interfaces, but may not hold when an exception is taken. Doing this allows
  other AAPCS functions to be called.
- If the CPU is M-class this is all that needs to be done since the architecture
  itself is designed in such a way that functions obeying the normal AAPCS ABI
  constraints are valid exception handlers.
- If the CPU is not M-class, the prologue and epilogue are modified to save all
  non-banked registers that are used, so that upon return the user-mode state
  will not be corrupted. Note that to avoid unnecessary overhead, only
  general-purpose (integer) registers are saved in this way. If VFP operations
  are needed, that state must be saved manually.

  Specifically, interrupt kinds other than "FIQ" will save all core registers
  except "lr" and "sp". "FIQ" interrupts will save r0-r7.
- If the CPU is not M-class, the return instruction is changed to one of the
  canonical sequences permitted by the architecture for exception return. Where
  possible the function itself will make the necessary "lr" adjustments so that
  the "preferred return address" is selected.

  Unfortunately the compiler is unable to make this guarantee for an "UNDEF"
  handler, where the offset from "lr" to the preferred return address depends on
  the execution state of the code which generated the exception. In this case
  a sequence equivalent to "movs pc, lr" will be used.


abi_tag (gnu::abi_tag)
----------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

The ``abi_tag`` attribute can be applied to a function, variable, class or
inline namespace declaration to modify the mangled name of the entity. It gives
the ability to distinguish between different versions of the same entity but
with different ABI versions supported. For example, a newer version of a class
could have a different set of data members and thus have a different size. Using
the ``abi_tag`` attribute, it is possible to have different mangled names for
a global variable of the class type. Therefor, the old code could keep using
the old manged name and the new code will use the new mangled name with tags.


acquire_capability (acquire_shared_capability, clang::acquire_capability, clang::acquire_shared_capability)
-----------------------------------------------------------------------------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

Marks a function as acquiring a capability.


alloc_size (gnu::alloc_size)
----------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

The ``alloc_size`` attribute can be placed on functions that return pointers in
order to hint to the compiler how many bytes of memory will be available at the
returned poiner. ``alloc_size`` takes one or two arguments.

- ``alloc_size(N)`` implies that argument number N equals the number of
  available bytes at the returned pointer.
- ``alloc_size(N, M)`` implies that the product of argument number N and
  argument number M equals the number of available bytes at the returned
  pointer.

Argument numbers are 1-based.

An example of how to use ``alloc_size``

.. code-block:: c

  void *my_malloc(int a) __attribute__((alloc_size(1)));
  void *my_calloc(int a, int b) __attribute__((alloc_size(1, 2)));

  int main() {
    void *const p = my_malloc(100);
    assert(__builtin_object_size(p, 0) == 100);
    void *const a = my_calloc(20, 5);
    assert(__builtin_object_size(a, 0) == 100);
  }

.. Note:: This attribute works differently in clang than it does in GCC.
  Specifically, clang will only trace ``const`` pointers (as above); we give up
  on pointers that are not marked as ``const``. In the vast majority of cases,
  this is unimportant, because LLVM has support for the ``alloc_size``
  attribute. However, this may cause mildly unintuitive behavior when used with
  other attributes, such as ``enable_if``.


interrupt
---------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

Clang supports the GNU style ``__attribute__((interrupt))`` attribute on
x86/x86-64 targets.The compiler generates function entry and exit sequences
suitable for use in an interrupt handler when this attribute is present.
The 'IRET' instruction, instead of the 'RET' instruction, is used to return
from interrupt or exception handlers.  All registers, except for the EFLAGS
register which is restored by the 'IRET' instruction, are preserved by the
compiler.

Any interruptible-without-stack-switch code must be compiled with
-mno-red-zone since interrupt handlers can and will, because of the
hardware design, touch the red zone.

1. interrupt handler must be declared with a mandatory pointer argument:

  .. code-block:: c

    struct interrupt_frame
    {
      uword_t ip;
      uword_t cs;
      uword_t flags;
      uword_t sp;
      uword_t ss;
    };

    __attribute__ ((interrupt))
    void f (struct interrupt_frame *frame) {
      ...
    }

2. exception handler:

  The exception handler is very similar to the interrupt handler with
  a different mandatory function signature:

  .. code-block:: c

    __attribute__ ((interrupt))
    void f (struct interrupt_frame *frame, uword_t error_code) {
      ...
    }

  and compiler pops 'ERROR_CODE' off stack before the 'IRET' instruction.

  The exception handler should only be used for exceptions which push an
  error code and all other exceptions must use the interrupt handler.
  The system will crash if the wrong handler is used.


assert_capability (assert_shared_capability, clang::assert_capability, clang::assert_shared_capability)
-------------------------------------------------------------------------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

Marks a function that dynamically tests whether a capability is held, and halts
the program if it is not held.


assume_aligned (gnu::assume_aligned)
------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

Use ``__attribute__((assume_aligned(<alignment>[,<offset>]))`` on a function
declaration to specify that the return value of the function (which must be a
pointer type) has the specified offset, in bytes, from an address with the
specified alignment. The offset is taken to be zero if omitted.

.. code-block:: c++

  // The returned pointer value has 32-byte alignment.
  void *a() __attribute__((assume_aligned (32)));

  // The returned pointer value is 4 bytes greater than an address having
  // 32-byte alignment.
  void *b() __attribute__((assume_aligned (32, 4)));

Note that this attribute provides information to the compiler regarding a
condition that the code already ensures is true. It does not cause the compiler
to enforce the provided alignment assumption.


availability
------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

The ``availability`` attribute can be placed on declarations to describe the
lifecycle of that declaration relative to operating system versions.  Consider
the function declaration for a hypothetical function ``f``:

.. code-block:: c++

  void f(void) __attribute__((availability(macos,introduced=10.4,deprecated=10.6,obsoleted=10.7)));

The availability attribute states that ``f`` was introduced in Mac OS X 10.4,
deprecated in Mac OS X 10.6, and obsoleted in Mac OS X 10.7.  This information
is used by Clang to determine when it is safe to use ``f``: for example, if
Clang is instructed to compile code for Mac OS X 10.5, a call to ``f()``
succeeds.  If Clang is instructed to compile code for Mac OS X 10.6, the call
succeeds but Clang emits a warning specifying that the function is deprecated.
Finally, if Clang is instructed to compile code for Mac OS X 10.7, the call
fails because ``f()`` is no longer available.

The availability attribute is a comma-separated list starting with the
platform name and then including clauses specifying important milestones in the
declaration's lifetime (in any order) along with additional information.  Those
clauses can be:

introduced=\ *version*
  The first version in which this declaration was introduced.

deprecated=\ *version*
  The first version in which this declaration was deprecated, meaning that
  users should migrate away from this API.

obsoleted=\ *version*
  The first version in which this declaration was obsoleted, meaning that it
  was removed completely and can no longer be used.

unavailable
  This declaration is never available on this platform.

message=\ *string-literal*
  Additional message text that Clang will provide when emitting a warning or
  error about use of a deprecated or obsoleted declaration.  Useful to direct
  users to replacement APIs.

replacement=\ *string-literal*
  Additional message text that Clang will use to provide Fix-It when emitting
  a warning about use of a deprecated declaration. The Fix-It will replace
  the deprecated declaration with the new declaration specified.

Multiple availability attributes can be placed on a declaration, which may
correspond to different platforms.  Only the availability attribute with the
platform corresponding to the target platform will be used; any others will be
ignored.  If no availability attribute specifies availability for the current
target platform, the availability attributes are ignored.  Supported platforms
are:

``ios``
  Apple's iOS operating system.  The minimum deployment target is specified by
  the ``-mios-version-min=*version*`` or ``-miphoneos-version-min=*version*``
  command-line arguments.

``macos``
  Apple's Mac OS X operating system.  The minimum deployment target is
  specified by the ``-mmacosx-version-min=*version*`` command-line argument.
  ``macosx`` is supported for backward-compatibility reasons, but it is
  deprecated.

``tvos``
  Apple's tvOS operating system.  The minimum deployment target is specified by
  the ``-mtvos-version-min=*version*`` command-line argument.

``watchos``
  Apple's watchOS operating system.  The minimum deployment target is specified by
  the ``-mwatchos-version-min=*version*`` command-line argument.

A declaration can typically be used even when deploying back to a platform
version prior to when the declaration was introduced.  When this happens, the
declaration is `weakly linked
<https://developer.apple.com/library/mac/#documentation/MacOSX/Conceptual/BPFrameworks/Concepts/WeakLinking.html>`_,
as if the ``weak_import`` attribute were added to the declaration.  A
weakly-linked declaration may or may not be present a run-time, and a program
can determine whether the declaration is present by checking whether the
address of that declaration is non-NULL.

The flag ``strict`` disallows using API when deploying back to a
platform version prior to when the declaration was introduced.  An
attempt to use such API before its introduction causes a hard error.
Weakly-linking is almost always a better API choice, since it allows
users to query availability at runtime.

If there are multiple declarations of the same entity, the availability
attributes must either match on a per-platform basis or later
declarations must not have availability attributes for that
platform. For example:

.. code-block:: c

  void g(void) __attribute__((availability(macos,introduced=10.4)));
  void g(void) __attribute__((availability(macos,introduced=10.4))); // okay, matches
  void g(void) __attribute__((availability(ios,introduced=4.0))); // okay, adds a new platform
  void g(void); // okay, inherits both macos and ios availability from above.
  void g(void) __attribute__((availability(macos,introduced=10.5))); // error: mismatch

When one method overrides another, the overriding method can be more widely available than the overridden method, e.g.,:

.. code-block:: objc

  @interface A
  - (id)method __attribute__((availability(macos,introduced=10.4)));
  - (id)method2 __attribute__((availability(macos,introduced=10.4)));
  @end

  @interface B : A
  - (id)method __attribute__((availability(macos,introduced=10.3))); // okay: method moved into base class later
  - (id)method __attribute__((availability(macos,introduced=10.5))); // error: this method was available via the base class in 10.4
  @end


_Noreturn
---------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","","X", ""

A function declared as ``_Noreturn`` shall not return to its caller. The
compiler will generate a diagnostic for a function declared as ``_Noreturn``
that appears to be capable of returning to its caller.


noreturn
--------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","X","","", ""

A function declared as ``[[noreturn]]`` shall not return to its caller. The
compiler will generate a diagnostic for a function declared as ``[[noreturn]]``
that appears to be capable of returning to its caller.


carries_dependency
------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

The ``carries_dependency`` attribute specifies dependency propagation into and
out of functions.

When specified on a function or Objective-C method, the ``carries_dependency``
attribute means that the return value carries a dependency out of the function,
so that the implementation need not constrain ordering upon return from that
function. Implementations of the function and its caller may choose to preserve
dependencies instead of emitting memory ordering instructions such as fences.

Note, this attribute does not change the meaning of the program, but may result
in generation of more efficient code.


convergent (clang::convergent)
------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

The ``convergent`` attribute can be placed on a function declaration. It is
translated into the LLVM ``convergent`` attribute, which indicates that the call
instructions of a function with this attribute cannot be made control-dependent
on any additional values.

In languages designed for SPMD/SIMT programming model, e.g. OpenCL or CUDA,
the call instructions of a function with this attribute must be executed by
all work items or threads in a work group or sub group.

This attribute is different from ``noduplicate`` because it allows duplicating
function calls if it can be proved that the duplicated function calls are
not made control-dependent on any additional values, e.g., unrolling a loop
executed by all work items.

Sample usage:
.. code-block:: c

  void convfunc(void) __attribute__((convergent));
  // Setting it as a C++11 attribute is also valid in a C++ program.
  // void convfunc(void) [[clang::convergent]];


deprecated (gnu::deprecated)
----------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","X","", ""

The ``deprecated`` attribute can be applied to a function, a variable, or a
type. This is useful when identifying functions, variables, or types that are
expected to be removed in a future version of a program.

Consider the function declaration for a hypothetical function ``f``:

.. code-block:: c++

  void f(void) __attribute__((deprecated("message", "replacement")));

When spelled as `__attribute__((deprecated))`, the deprecated attribute can have
two optional string arguments. The first one is the message to display when
emitting the warning; the second one enables the compiler to provide a Fix-It
to replace the deprecated name with a new name. Otherwise, when spelled as
`[[gnu::deprecated]] or [[deprecated]]`, the attribute can have one optional
string argument which is the message to display when emitting the warning.


diagnose_if
-----------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

The ``diagnose_if`` attribute can be placed on function declarations to emit
warnings or errors at compile-time if calls to the attributed function meet
certain user-defined criteria. For example:

.. code-block:: c
  void abs(int a)
    __attribute__((diagnose_if(a >= 0, "Redundant abs call", "warning")));
  void must_abs(int a)
    __attribute__((diagnose_if(a >= 0, "Redundant abs call", "error")));

  int val = abs(1); // warning: Redundant abs call
  int val2 = must_abs(1); // error: Redundant abs call
  int val3 = abs(val);
  int val4 = must_abs(val); // Because run-time checks are not emitted for
                            // diagnose_if attributes, this executes without
                            // issue.


``diagnose_if`` is closely related to ``enable_if``, with a few key differences:

* Overload resolution is not aware of ``diagnose_if`` attributes: they're
  considered only after we select the best candidate from a given candidate set.
* Function declarations that differ only in their ``diagnose_if`` attributes are
  considered to be redeclarations of the same function (not overloads).
* If the condition provided to ``diagnose_if`` cannot be evaluated, no
  diagnostic will be emitted.

Otherwise, ``diagnose_if`` is essentially the logical negation of ``enable_if``.

As a result of bullet number two, ``diagnose_if`` attributes will stack on the
same function. For example:

.. code-block:: c

  int foo() __attribute__((diagnose_if(1, "diag1", "warning")));
  int foo() __attribute__((diagnose_if(1, "diag2", "warning")));

  int bar = foo(); // warning: diag1
                   // warning: diag2
  int (*fooptr)(void) = foo; // warning: diag1
                             // warning: diag2

  constexpr int supportsAPILevel(int N) { return N < 5; }
  int baz(int a)
    __attribute__((diagnose_if(!supportsAPILevel(10),
                               "Upgrade to API level 10 to use baz", "error")));
  int baz(int a)
    __attribute__((diagnose_if(!a, "0 is not recommended.", "warning")));

  int (*bazptr)(int) = baz; // error: Upgrade to API level 10 to use baz
  int v = baz(0); // error: Upgrade to API level 10 to use baz

Query for this feature with ``__has_attribute(diagnose_if)``.


disable_tail_calls (clang::disable_tail_calls)
----------------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

The ``disable_tail_calls`` attribute instructs the backend to not perform tail call optimization inside the marked function.

For example:

  .. code-block:: c

    int callee(int);

    int foo(int a) __attribute__((disable_tail_calls)) {
      return callee(a); // This call is not tail-call optimized.
    }

Marking virtual functions as ``disable_tail_calls`` is legal.

  .. code-block:: c++

    int callee(int);

    class Base {
    public:
      [[clang::disable_tail_calls]] virtual int foo1() {
        return callee(); // This call is not tail-call optimized.
      }
    };

    class Derived1 : public Base {
    public:
      int foo1() override {
        return callee(); // This call is tail-call optimized.
      }
    };


enable_if
---------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

.. Note:: Some features of this attribute are experimental. The meaning of
  multiple enable_if attributes on a single declaration is subject to change in
  a future version of clang. Also, the ABI is not standardized and the name
  mangling may change in future versions. To avoid that, use asm labels.

The ``enable_if`` attribute can be placed on function declarations to control
which overload is selected based on the values of the function's arguments.
When combined with the ``overloadable`` attribute, this feature is also
available in C.

.. code-block:: c++

  int isdigit(int c);
  int isdigit(int c) __attribute__((enable_if(c <= -1 || c > 255, "chosen when 'c' is out of range"))) __attribute__((unavailable("'c' must have the value of an unsigned char or EOF")));
  
  void foo(char c) {
    isdigit(c);
    isdigit(10);
    isdigit(-10);  // results in a compile-time error.
  }

The enable_if attribute takes two arguments, the first is an expression written
in terms of the function parameters, the second is a string explaining why this
overload candidate could not be selected to be displayed in diagnostics. The
expression is part of the function signature for the purposes of determining
whether it is a redeclaration (following the rules used when determining
whether a C++ template specialization is ODR-equivalent), but is not part of
the type.

The enable_if expression is evaluated as if it were the body of a
bool-returning constexpr function declared with the arguments of the function
it is being applied to, then called with the parameters at the call site. If the
result is false or could not be determined through constant expression
evaluation, then this overload will not be chosen and the provided string may
be used in a diagnostic if the compile fails as a result.

Because the enable_if expression is an unevaluated context, there are no global
state changes, nor the ability to pass information from the enable_if
expression to the function body. For example, suppose we want calls to
strnlen(strbuf, maxlen) to resolve to strnlen_chk(strbuf, maxlen, size of
strbuf) only if the size of strbuf can be determined:

.. code-block:: c++

  __attribute__((always_inline))
  static inline size_t strnlen(const char *s, size_t maxlen)
    __attribute__((overloadable))
    __attribute__((enable_if(__builtin_object_size(s, 0) != -1))),
                             "chosen when the buffer size is known but 'maxlen' is not")))
  {
    return strnlen_chk(s, maxlen, __builtin_object_size(s, 0));
  }

Multiple enable_if attributes may be applied to a single declaration. In this
case, the enable_if expressions are evaluated from left to right in the
following manner. First, the candidates whose enable_if expressions evaluate to
false or cannot be evaluated are discarded. If the remaining candidates do not
share ODR-equivalent enable_if expressions, the overload resolution is
ambiguous. Otherwise, enable_if overload resolution continues with the next
enable_if attribute on the candidates that have not been discarded and have
remaining enable_if attributes. In this way, we pick the most specific
overload out of a number of viable overloads using enable_if.

.. code-block:: c++

  void f() __attribute__((enable_if(true, "")));  // #1
  void f() __attribute__((enable_if(true, ""))) __attribute__((enable_if(true, "")));  // #2
  
  void g(int i, int j) __attribute__((enable_if(i, "")));  // #1
  void g(int i, int j) __attribute__((enable_if(j, ""))) __attribute__((enable_if(true)));  // #2

In this example, a call to f() is always resolved to #2, as the first enable_if
expression is ODR-equivalent for both declarations, but #1 does not have another
enable_if expression to continue evaluating, so the next round of evaluation has
only a single candidate. In a call to g(1, 1), the call is ambiguous even though
#2 has more enable_if attributes, because the first enable_if expressions are
not ODR-equivalent.

Query for this feature with ``__has_attribute(enable_if)``.

Note that functions with one or more ``enable_if`` attributes may not have
their address taken, unless all of the conditions specified by said
``enable_if`` are constants that evaluate to ``true``. For example:

.. code-block:: c

  const int TrueConstant = 1;
  const int FalseConstant = 0;
  int f(int a) __attribute__((enable_if(a > 0, "")));
  int g(int a) __attribute__((enable_if(a == 0 || a != 0, "")));
  int h(int a) __attribute__((enable_if(1, "")));
  int i(int a) __attribute__((enable_if(TrueConstant, "")));
  int j(int a) __attribute__((enable_if(FalseConstant, "")));

  void fn() {
    int (*ptr)(int);
    ptr = &f; // error: 'a > 0' is not always true
    ptr = &g; // error: 'a == 0 || a != 0' is not a truthy constant
    ptr = &h; // OK: 1 is a truthy constant
    ptr = &i; // OK: 'TrueConstant' is a truthy constant
    ptr = &j; // error: 'FalseConstant' is a constant, but not truthy
  }

Because ``enable_if`` evaluation happens during overload resolution,
``enable_if`` may give unintuitive results when used with templates, depending
on when overloads are resolved. In the example below, clang will emit a
diagnostic about no viable overloads for ``foo`` in ``bar``, but not in ``baz``:

.. code-block:: c++

  double foo(int i) __attribute__((enable_if(i > 0, "")));
  void *foo(int i) __attribute__((enable_if(i <= 0, "")));
  template <int I>
  auto bar() { return foo(I); }

  template <typename T>
  auto baz() { return foo(T::number); }

  struct WithNumber { constexpr static int number = 1; };
  void callThem() {
    bar<sizeof(WithNumber)>();
    baz<WithNumber>();
  }

This is because, in ``bar``, ``foo`` is resolved prior to template
instantiation, so the value for ``I`` isn't known (thus, both ``enable_if``
conditions for ``foo`` fail). However, in ``baz``, ``foo`` is resolved during
template instantiation, so the value for ``T::number`` is known.


flatten (gnu::flatten)
----------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

The ``flatten`` attribute causes calls within the attributed function to
be inlined unless it is impossible to do so, for example if the body of the
callee is unavailable or if the callee has the ``noinline`` attribute.


format (gnu::format)
--------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

Clang supports the ``format`` attribute, which indicates that the function
accepts a ``printf`` or ``scanf``-like format string and corresponding
arguments or a ``va_list`` that contains these arguments.

Please see `GCC documentation about format attribute
<http://gcc.gnu.org/onlinedocs/gcc/Function-Attributes.html>`_ to find details
about attribute syntax.

Clang implements two kinds of checks with this attribute.

#. Clang checks that the function with the ``format`` attribute is called with
   a format string that uses format specifiers that are allowed, and that
   arguments match the format string.  This is the ``-Wformat`` warning, it is
   on by default.

#. Clang checks that the format string argument is a literal string.  This is
   the ``-Wformat-nonliteral`` warning, it is off by default.

   Clang implements this mostly the same way as GCC, but there is a difference
   for functions that accept a ``va_list`` argument (for example, ``vprintf``).
   GCC does not emit ``-Wformat-nonliteral`` warning for calls to such
   functions.  Clang does not warn if the format string comes from a function
   parameter, where the function is annotated with a compatible attribute,
   otherwise it warns.  For example:

   .. code-block:: c

     __attribute__((__format__ (__scanf__, 1, 3)))
     void foo(const char* s, char *buf, ...) {
       va_list ap;
       va_start(ap, buf);

       vprintf(s, ap); // warning: format string is not a string literal
     }

   In this case we warn because ``s`` contains a format string for a
   ``scanf``-like function, but it is passed to a ``printf``-like function.

   If the attribute is removed, clang still warns, because the format string is
   not a string literal.

   Another example:

   .. code-block:: c

     __attribute__((__format__ (__printf__, 1, 3)))
     void foo(const char* s, char *buf, ...) {
       va_list ap;
       va_start(ap, buf);

       vprintf(s, ap); // warning
     }

   In this case Clang does not warn because the format string ``s`` and
   the corresponding arguments are annotated.  If the arguments are
   incorrect, the caller of ``foo`` will receive a warning.


ifunc (gnu::ifunc)
------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

``__attribute__((ifunc("resolver")))`` is used to mark that the address of a declaration should be resolved at runtime by calling a resolver function.

The symbol name of the resolver function is given in quotes.  A function with this name (after mangling) must be defined in the current translation unit; it may be ``static``.  The resolver function should take no arguments and return a pointer.

The ``ifunc`` attribute may only be used on a function declaration.  A function declaration with an ``ifunc`` attribute is considered to be a definition of the declared entity.  The entity must not have weak linkage; for example, in C++, it cannot be applied to a declaration if a definition at that location would be considered inline.

Not all targets support this attribute.  ELF targets support this attribute when using binutils v2.20.1 or higher and glibc v2.11.1 or higher.  Non-ELF targets currently do not support this attribute.


internal_linkage (clang::internal_linkage)
------------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

The ``internal_linkage`` attribute changes the linkage type of the declaration to internal.
This is similar to C-style ``static``, but can be used on classes and class methods. When applied to a class definition,
this attribute affects all methods and static data members of that class.
This can be used to contain the ABI of a C++ library by excluding unwanted class methods from the export tables.


interrupt
---------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

Clang supports the GNU style ``__attribute__((interrupt("ARGUMENT")))`` attribute on
MIPS targets. This attribute may be attached to a function definition and instructs
the backend to generate appropriate function entry/exit code so that it can be used
directly as an interrupt service routine.

By default, the compiler will produce a function prologue and epilogue suitable for
an interrupt service routine that handles an External Interrupt Controller (eic)
generated interrupt. This behaviour can be explicitly requested with the "eic"
argument.

Otherwise, for use with vectored interrupt mode, the argument passed should be
of the form "vector=LEVEL" where LEVEL is one of the following values:
"sw0", "sw1", "hw0", "hw1", "hw2", "hw3", "hw4", "hw5". The compiler will
then set the interrupt mask to the corresponding level which will mask all
interrupts up to and including the argument.

The semantics are as follows:

- The prologue is modified so that the Exception Program Counter (EPC) and
  Status coprocessor registers are saved to the stack. The interrupt mask is
  set so that the function can only be interrupted by a higher priority
  interrupt. The epilogue will restore the previous values of EPC and Status.

- The prologue and epilogue are modified to save and restore all non-kernel
  registers as necessary.

- The FPU is disabled in the prologue, as the floating pointer registers are not
  spilled to the stack.

- The function return sequence is changed to use an exception return instruction.

- The parameter sets the interrupt mask for the function corresponding to the
  interrupt level specified. If no mask is specified the interrupt mask
  defaults to "eic".


noalias
-------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","X","", ""

The ``noalias`` attribute indicates that the only memory accesses inside
function are loads and stores from objects pointed to by its pointer-typed
arguments, with arbitrary offsets.


noduplicate (clang::noduplicate)
--------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

The ``noduplicate`` attribute can be placed on function declarations to control
whether function calls to this function can be duplicated or not as a result of
optimizations. This is required for the implementation of functions with
certain special requirements, like the OpenCL "barrier" function, that might
need to be run concurrently by all the threads that are executing in lockstep
on the hardware. For example this attribute applied on the function
"nodupfunc" in the code below avoids that:

.. code-block:: c

  void nodupfunc() __attribute__((noduplicate));
  // Setting it as a C++11 attribute is also valid
  // void nodupfunc() [[clang::noduplicate]];
  void foo();
  void bar();

  nodupfunc();
  if (a > n) {
    foo();
  } else {
    bar();
  }

gets possibly modified by some optimizations into code similar to this:

.. code-block:: c

  if (a > n) {
    nodupfunc();
    foo();
  } else {
    nodupfunc();
    bar();
  }

where the call to "nodupfunc" is duplicated and sunk into the two branches
of the condition.


no_sanitize (clang::no_sanitize)
--------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

Use the ``no_sanitize`` attribute on a function declaration to specify
that a particular instrumentation or set of instrumentations should not be
applied to that function. The attribute takes a list of string literals,
which have the same meaning as values accepted by the ``-fno-sanitize=``
flag. For example, ``__attribute__((no_sanitize("address", "thread")))``
specifies that AddressSanitizer and ThreadSanitizer should not be applied
to the function.

See :ref:`Controlling Code Generation <controlling-code-generation>` for a
full list of supported sanitizer flags.


no_sanitize_address (no_address_safety_analysis, gnu::no_address_safety_analysis, gnu::no_sanitize_address)
-----------------------------------------------------------------------------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

.. _langext-address_sanitizer:

Use ``__attribute__((no_sanitize_address))`` on a function declaration to
specify that address safety instrumentation (e.g. AddressSanitizer) should
not be applied to that function.


no_sanitize_thread
------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

.. _langext-thread_sanitizer:

Use ``__attribute__((no_sanitize_thread))`` on a function declaration to
specify that checks for data races on plain (non-atomic) memory accesses should
not be inserted by ThreadSanitizer. The function is still instrumented by the
tool to avoid false positives and provide meaningful stack traces.


no_sanitize_memory
------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

.. _langext-memory_sanitizer:

Use ``__attribute__((no_sanitize_memory))`` on a function declaration to
specify that checks for uninitialized memory should not be inserted
(e.g. by MemorySanitizer). The function may still be instrumented by the tool
to avoid false positives in other places.


no_split_stack (gnu::no_split_stack)
------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

The ``no_split_stack`` attribute disables the emission of the split stack
preamble for a particular function. It has no effect if ``-fsplit-stack``
is not specified.


not_tail_called (clang::not_tail_called)
----------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

The ``not_tail_called`` attribute prevents tail-call optimization on statically bound calls. It has no effect on indirect calls. Virtual functions, objective-c methods, and functions marked as ``always_inline`` cannot be marked as ``not_tail_called``.

For example, it prevents tail-call optimization in the following case:

  .. code-block:: c

    int __attribute__((not_tail_called)) foo1(int);

    int foo2(int a) {
      return foo1(a); // No tail-call optimization on direct calls.
    }

However, it doesn't prevent tail-call optimization in this case:

  .. code-block:: c

    int __attribute__((not_tail_called)) foo1(int);

    int foo2(int a) {
      int (*fn)(int) = &foo1;

      // not_tail_called has no effect on an indirect call even if the call can be
      // resolved at compile time.
      return (*fn)(a);
    }

Marking virtual functions as ``not_tail_called`` is an error:

  .. code-block:: c++

    class Base {
    public:
      // not_tail_called on a virtual function is an error.
      [[clang::not_tail_called]] virtual int foo1();

      virtual int foo2();

      // Non-virtual functions can be marked ``not_tail_called``.
      [[clang::not_tail_called]] int foo3();
    };

    class Derived1 : public Base {
    public:
      int foo1() override;

      // not_tail_called on a virtual function is an error.
      [[clang::not_tail_called]] int foo2() override;
    };


#pragma omp declare simd
------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","","", "X"

The `declare simd` construct can be applied to a function to enable the creation
of one or more versions that can process multiple arguments using SIMD
instructions from a single invocation in a SIMD loop. The `declare simd`
directive is a declarative directive. There may be multiple `declare simd`
directives for a function. The use of a `declare simd` construct on a function
enables the creation of SIMD versions of the associated function that can be
used to process multiple arguments from a single invocation from a SIMD loop
concurrently.
The syntax of the `declare simd` construct is as follows:

  .. code-block:: c

  #pragma omp declare simd [clause[[,] clause] ...] new-line
  [#pragma omp declare simd [clause[[,] clause] ...] new-line]
  [...]
  function definition or declaration

where clause is one of the following:

  .. code-block:: c

  simdlen(length)
  linear(argument-list[:constant-linear-step])
  aligned(argument-list[:alignment])
  uniform(argument-list)
  inbranch
  notinbranch


#pragma omp declare target
--------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","","", "X"

The `declare target` directive specifies that variables and functions are mapped
to a device for OpenMP offload mechanism.

The syntax of the declare target directive is as follows:

  .. code-block:: c

  #pragma omp declare target new-line
  declarations-definition-seq
  #pragma omp end declare target new-line


objc_boxable
------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

Structs and unions marked with the ``objc_boxable`` attribute can be used
with the Objective-C boxed expression syntax, ``@(...)``.

**Usage**: ``__attribute__((objc_boxable))``. This attribute
can only be placed on a declaration of a trivially-copyable struct or union:

.. code-block:: objc

  struct __attribute__((objc_boxable)) some_struct {
    int i;
  };
  union __attribute__((objc_boxable)) some_union {
    int i;
    float f;
  };
  typedef struct __attribute__((objc_boxable)) _some_struct some_struct;

  // ...

  some_struct ss;
  NSValue *boxed = @(ss);


objc_method_family
------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

Many methods in Objective-C have conventional meanings determined by their
selectors. It is sometimes useful to be able to mark a method as having a
particular conventional meaning despite not having the right selector, or as
not having the conventional meaning that its selector would suggest. For these
use cases, we provide an attribute to specifically describe the "method family"
that a method belongs to.

**Usage**: ``__attribute__((objc_method_family(X)))``, where ``X`` is one of
``none``, ``alloc``, ``copy``, ``init``, ``mutableCopy``, or ``new``.  This
attribute can only be placed at the end of a method declaration:

.. code-block:: objc

  - (NSString *)initMyStringValue __attribute__((objc_method_family(none)));

Users who do not wish to change the conventional meaning of a method, and who
merely want to document its non-standard retain and release semantics, should
use the retaining behavior attributes (``ns_returns_retained``,
``ns_returns_not_retained``, etc).

Query for this feature with ``__has_attribute(objc_method_family)``.


objc_requires_super
-------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

Some Objective-C classes allow a subclass to override a particular method in a
parent class but expect that the overriding method also calls the overridden
method in the parent class. For these cases, we provide an attribute to
designate that a method requires a "call to ``super``" in the overriding
method in the subclass.

**Usage**: ``__attribute__((objc_requires_super))``.  This attribute can only
be placed at the end of a method declaration:

.. code-block:: objc

  - (void)foo __attribute__((objc_requires_super));

This attribute can only be applied the method declarations within a class, and
not a protocol.  Currently this attribute does not enforce any placement of
where the call occurs in the overriding method (such as in the case of
``-dealloc`` where the call must appear at the end).  It checks only that it
exists.

Note that on both OS X and iOS that the Foundation framework provides a
convenience macro ``NS_REQUIRES_SUPER`` that provides syntactic sugar for this
attribute:

.. code-block:: objc

  - (void)foo NS_REQUIRES_SUPER;

This macro is conditionally defined depending on the compiler's support for
this attribute.  If the compiler does not support the attribute the macro
expands to nothing.

Operationally, when a method has this annotation the compiler will warn if the
implementation of an override in a subclass does not call super.  For example:

.. code-block:: objc

   warning: method possibly missing a [super AnnotMeth] call
   - (void) AnnotMeth{};
                      ^


objc_runtime_name
-----------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

By default, the Objective-C interface or protocol identifier is used
in the metadata name for that object. The `objc_runtime_name`
attribute allows annotated interfaces or protocols to use the
specified string argument in the object's metadata name instead of the
default name.
        
**Usage**: ``__attribute__((objc_runtime_name("MyLocalName")))``.  This attribute
can only be placed before an @protocol or @interface declaration:
        
.. code-block:: objc
        
  __attribute__((objc_runtime_name("MyLocalName")))
  @interface Message
  @end


objc_runtime_visible
--------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

This attribute specifies that the Objective-C class to which it applies is visible to the Objective-C runtime but not to the linker. Classes annotated with this attribute cannot be subclassed and cannot have categories defined for them.


optnone (clang::optnone)
------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

The ``optnone`` attribute suppresses essentially all optimizations
on a function or method, regardless of the optimization level applied to
the compilation unit as a whole.  This is particularly useful when you
need to debug a particular function, but it is infeasible to build the
entire application without optimization.  Avoiding optimization on the
specified function can improve the quality of the debugging information
for that function.

This attribute is incompatible with the ``always_inline`` and ``minsize``
attributes.


overloadable
------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

Clang provides support for C++ function overloading in C.  Function overloading
in C is introduced using the ``overloadable`` attribute.  For example, one
might provide several overloaded versions of a ``tgsin`` function that invokes
the appropriate standard function computing the sine of a value with ``float``,
``double``, or ``long double`` precision:

.. code-block:: c

  #include <math.h>
  float __attribute__((overloadable)) tgsin(float x) { return sinf(x); }
  double __attribute__((overloadable)) tgsin(double x) { return sin(x); }
  long double __attribute__((overloadable)) tgsin(long double x) { return sinl(x); }

Given these declarations, one can call ``tgsin`` with a ``float`` value to
receive a ``float`` result, with a ``double`` to receive a ``double`` result,
etc.  Function overloading in C follows the rules of C++ function overloading
to pick the best overload given the call arguments, with a few C-specific
semantics:

* Conversion from ``float`` or ``double`` to ``long double`` is ranked as a
  floating-point promotion (per C99) rather than as a floating-point conversion
  (as in C++).

* A conversion from a pointer of type ``T*`` to a pointer of type ``U*`` is
  considered a pointer conversion (with conversion rank) if ``T`` and ``U`` are
  compatible types.

* A conversion from type ``T`` to a value of type ``U`` is permitted if ``T``
  and ``U`` are compatible types.  This conversion is given "conversion" rank.

* If no viable candidates are otherwise available, we allow a conversion from a
  pointer of type ``T*`` to a pointer of type ``U*``, where ``T`` and ``U`` are
  incompatible. This conversion is ranked below all other types of conversions.
  Please note: ``U`` lacking qualifiers that are present on ``T`` is sufficient
  for ``T`` and ``U`` to be incompatible.

The declaration of ``overloadable`` functions is restricted to function
declarations and definitions.  Most importantly, if any function with a given
name is given the ``overloadable`` attribute, then all function declarations
and definitions with that name (and in that scope) must have the
``overloadable`` attribute.  This rule even applies to redeclarations of
functions whose original declaration had the ``overloadable`` attribute, e.g.,

.. code-block:: c

  int f(int) __attribute__((overloadable));
  float f(float); // error: declaration of "f" must have the "overloadable" attribute

  int g(int) __attribute__((overloadable));
  int g(int) { } // error: redeclaration of "g" must also have the "overloadable" attribute

Functions marked ``overloadable`` must have prototypes.  Therefore, the
following code is ill-formed:

.. code-block:: c

  int h() __attribute__((overloadable)); // error: h does not have a prototype

However, ``overloadable`` functions are allowed to use a ellipsis even if there
are no named parameters (as is permitted in C++).  This feature is particularly
useful when combined with the ``unavailable`` attribute:

.. code-block:: c++

  void honeypot(...) __attribute__((overloadable, unavailable)); // calling me is an error

Functions declared with the ``overloadable`` attribute have their names mangled
according to the same rules as C++ function names.  For example, the three
``tgsin`` functions in our motivating example get the mangled names
``_Z5tgsinf``, ``_Z5tgsind``, and ``_Z5tgsine``, respectively.  There are two
caveats to this use of name mangling:

* Future versions of Clang may change the name mangling of functions overloaded
  in C, so you should not depend on an specific mangling.  To be completely
  safe, we strongly urge the use of ``static inline`` with ``overloadable``
  functions.

* The ``overloadable`` attribute has almost no meaning when used in C++,
  because names will already be mangled and functions are already overloadable.
  However, when an ``overloadable`` function occurs within an ``extern "C"``
  linkage specification, it's name *will* be mangled in the same way as it
  would in C.

Query for this feature with ``__has_extension(attribute_overloadable)``.


release_capability (release_shared_capability, clang::release_capability, clang::release_shared_capability)
-----------------------------------------------------------------------------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

Marks a function as releasing a capability.


kernel
------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

``__attribute__((kernel))`` is used to mark a ``kernel`` function in
RenderScript.

In RenderScript, ``kernel`` functions are used to express data-parallel
computations.  The RenderScript runtime efficiently parallelizes ``kernel``
functions to run on computational resources such as multi-core CPUs and GPUs.
See the RenderScript_ documentation for more information.

.. _RenderScript: https://developer.android.com/guide/topics/renderscript/compute.html


target (gnu::target)
--------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

Clang supports the GNU style ``__attribute__((target("OPTIONS")))`` attribute.
This attribute may be attached to a function definition and instructs
the backend to use different code generation options than were passed on the
command line.

The current set of options correspond to the existing "subtarget features" for
the target with or without a "-mno-" in front corresponding to the absence
of the feature, as well as ``arch="CPU"`` which will change the default "CPU"
for the function.

Example "subtarget features" from the x86 backend include: "mmx", "sse", "sse4.2",
"avx", "xop" and largely correspond to the machine specific options handled by
the front end.


try_acquire_capability (try_acquire_shared_capability, clang::try_acquire_capability, clang::try_acquire_shared_capability)
---------------------------------------------------------------------------------------------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

Marks a function that attempts to acquire a capability. This function may fail to
actually acquire the capability; they accept a Boolean value determining
whether acquiring the capability means success (true), or failing to acquire
the capability means success (false).


nodiscard, warn_unused_result, clang::warn_unused_result, gnu::warn_unused_result
---------------------------------------------------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

Clang supports the ability to diagnose when the results of a function call
expression are discarded under suspicious circumstances. A diagnostic is
generated when a function or its return type is marked with ``[[nodiscard]]``
(or ``__attribute__((warn_unused_result))``) and the function call appears as a
potentially-evaluated discarded-value expression that is not explicitly cast to
`void`.

.. code-block: c++
  struct [[nodiscard]] error_info { /*...*/ };
  error_info enable_missile_safety_mode();
  
  void launch_missiles();
  void test_missiles() {
    enable_missile_safety_mode(); // diagnoses
    launch_missiles();
  }
  error_info &foo();
  void f() { foo(); } // Does not diagnose, error_info is a reference.


xray_always_instrument (clang::xray_always_instrument), xray_never_instrument (clang::xray_never_instrument)
------------------------------------------------------------------------------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

``__attribute__((xray_always_instrument))`` or ``[[clang::xray_always_instrument]]`` is used to mark member functions (in C++), methods (in Objective C), and free functions (in C, C++, and Objective C) to be instrumented with XRay. This will cause the function to always have space at the beginning and exit points to allow for runtime patching.

Conversely, ``__attribute__((xray_never_instrument))`` or ``[[clang::xray_never_instrument]]`` will inhibit the insertion of these instrumentation points.

If a function has neither of these attributes, they become subject to the XRay heuristics used to determine whether a function should be instrumented or otherwise.


Variable Attributes
===================


dllexport (gnu::dllexport)
--------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","X","", ""

The ``__declspec(dllexport)`` attribute declares a variable, function, or
Objective-C interface to be exported from the module.  It is available under the
``-fdeclspec`` flag for compatibility with various compilers.  The primary use
is for COFF object files which explicitly specify what interfaces are available
for external use.  See the dllexport_ documentation on MSDN for more
information.

.. _dllexport: https://msdn.microsoft.com/en-us/library/3y1sfaz2.aspx


dllimport (gnu::dllimport)
--------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","X","", ""

The ``__declspec(dllimport)`` attribute declares a variable, function, or
Objective-C interface to be imported from an external module.  It is available
under the ``-fdeclspec`` flag for compatibility with various compilers.  The
primary use is for COFF object files which explicitly specify what interfaces
are imported from external modules.  See the dllimport_ documentation on MSDN
for more information.

.. _dllimport: https://msdn.microsoft.com/en-us/library/3y1sfaz2.aspx


init_seg
--------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","","", "X"

The attribute applied by ``pragma init_seg()`` controls the section into
which global initialization function pointers are emitted.  It is only
available with ``-fms-extensions``.  Typically, this function pointer is
emitted into ``.CRT$XCU`` on Windows.  The user can change the order of
initialization by using a different section name with the same
``.CRT$XC`` prefix and a suffix that sorts lexicographically before or
after the standard ``.CRT$XCU`` sections.  See the init_seg_
documentation on MSDN for more information.

.. _init_seg: http://msdn.microsoft.com/en-us/library/7977wcck(v=vs.110).aspx


nodebug (gnu::nodebug)
----------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

The ``nodebug`` attribute allows you to suppress debugging information for a
function or method, or for a variable that is not a parameter or a non-static
data member.


nosvm
-----
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

OpenCL 2.0 supports the optional ``__attribute__((nosvm))`` qualifier for
pointer variable. It informs the compiler that the pointer does not refer
to a shared virtual memory region. See OpenCL v2.0 s6.7.2 for details.

Since it is not widely used and has been removed from OpenCL 2.1, it is ignored
by Clang.


pass_object_size
----------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

.. Note:: The mangling of functions with parameters that are annotated with
  ``pass_object_size`` is subject to change. You can get around this by
  using ``__asm__("foo")`` to explicitly name your functions, thus preserving
  your ABI; also, non-overloadable C functions with ``pass_object_size`` are
  not mangled.

The ``pass_object_size(Type)`` attribute can be placed on function parameters to
instruct clang to call ``__builtin_object_size(param, Type)`` at each callsite
of said function, and implicitly pass the result of this call in as an invisible
argument of type ``size_t`` directly after the parameter annotated with
``pass_object_size``. Clang will also replace any calls to
``__builtin_object_size(param, Type)`` in the function by said implicit
parameter.

Example usage:

.. code-block:: c

  int bzero1(char *const p __attribute__((pass_object_size(0))))
      __attribute__((noinline)) {
    int i = 0;
    for (/**/; i < (int)__builtin_object_size(p, 0); ++i) {
      p[i] = 0;
    }
    return i;
  }

  int main() {
    char chars[100];
    int n = bzero1(&chars[0]);
    assert(n == sizeof(chars));
    return 0;
  }

If successfully evaluating ``__builtin_object_size(param, Type)`` at the
callsite is not possible, then the "failed" value is passed in. So, using the
definition of ``bzero1`` from above, the following code would exit cleanly:

.. code-block:: c

  int main2(int argc, char *argv[]) {
    int n = bzero1(argv);
    assert(n == -1);
    return 0;
  }

``pass_object_size`` plays a part in overload resolution. If two overload
candidates are otherwise equally good, then the overload with one or more
parameters with ``pass_object_size`` is preferred. This implies that the choice
between two identical overloads both with ``pass_object_size`` on one or more
parameters will always be ambiguous; for this reason, having two such overloads
is illegal. For example:

.. code-block:: c++

  #define PS(N) __attribute__((pass_object_size(N)))
  // OK
  void Foo(char *a, char *b); // Overload A
  // OK -- overload A has no parameters with pass_object_size.
  void Foo(char *a PS(0), char *b PS(0)); // Overload B
  // Error -- Same signature (sans pass_object_size) as overload B, and both
  // overloads have one or more parameters with the pass_object_size attribute.
  void Foo(void *a PS(0), void *b);

  // OK
  void Bar(void *a PS(0)); // Overload C
  // OK
  void Bar(char *c PS(1)); // Overload D

  void main() {
    char known[10], *unknown;
    Foo(unknown, unknown); // Calls overload B
    Foo(known, unknown); // Calls overload B
    Foo(unknown, known); // Calls overload B
    Foo(known, known); // Calls overload B

    Bar(known); // Calls overload D
    Bar(unknown); // Calls overload D
  }

Currently, ``pass_object_size`` is a bit restricted in terms of its usage:

* Only one use of ``pass_object_size`` is allowed per parameter.

* It is an error to take the address of a function with ``pass_object_size`` on
  any of its parameters. If you wish to do this, you can create an overload
  without ``pass_object_size`` on any parameters.

* It is an error to apply the ``pass_object_size`` attribute to parameters that
  are not pointers. Additionally, any parameter that ``pass_object_size`` is
  applied to must be marked ``const`` at its function's definition.


require_constant_initialization (clang::require_constant_initialization)
------------------------------------------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

This attribute specifies that the variable to which it is attached is intended
to have a `constant initializer <http://en.cppreference.com/w/cpp/language/constant_initialization>`_
according to the rules of [basic.start.static]. The variable is required to
have static or thread storage duration. If the initialization of the variable
is not a constant initializer an error will be produced. This attribute may
only be used in C++.

Note that in C++03 strict constant expression checking is not done. Instead
the attribute reports if Clang can emit the variable as a constant, even if it's
not technically a 'constant initializer'. This behavior is non-portable.

Static storage duration variables with constant initializers avoid hard-to-find
bugs caused by the indeterminate order of dynamic initialization. They can also
be safely used during dynamic initialization across translation units.

This attribute acts as a compile time assertion that the requirements
for constant initialization have been met. Since these requirements change
between dialects and have subtle pitfalls it's important to fail fast instead
of silently falling back on dynamic initialization.

.. code-block:: c++

  // -std=c++14
  #define SAFE_STATIC [[clang::require_constant_initialization]]
  struct T {
    constexpr T(int) {}
    ~T(); // non-trivial
  };
  SAFE_STATIC T x = {42}; // Initialization OK. Doesn't check destructor.
  SAFE_STATIC T y = 42; // error: variable does not have a constant initializer
  // copy initialization is not a constant expression on a non-literal type.


section (gnu::section, __declspec(allocate))
--------------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","X","", ""

The ``section`` attribute allows you to specify a specific section a
global variable or function should be in after translation.


swiftcall (gnu::swiftcall)
--------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

The ``swiftcall`` attribute indicates that a function should be called
using the Swift calling convention for a function or function pointer.

The lowering for the Swift calling convention, as described by the Swift
ABI documentation, occurs in multiple phases.  The first, "high-level"
phase breaks down the formal parameters and results into innately direct
and indirect components, adds implicit paraameters for the generic
signature, and assigns the context and error ABI treatments to parameters
where applicable.  The second phase breaks down the direct parameters
and results from the first phase and assigns them to registers or the
stack.  The ``swiftcall`` convention only handles this second phase of
lowering; the C function type must accurately reflect the results
of the first phase, as follows:

- Results classified as indirect by high-level lowering should be
  represented as parameters with the ``swift_indirect_result`` attribute.

- Results classified as direct by high-level lowering should be represented
  as follows:

  - First, remove any empty direct results.

  - If there are no direct results, the C result type should be ``void``.

  - If there is one direct result, the C result type should be a type with
    the exact layout of that result type.

  - If there are a multiple direct results, the C result type should be
    a struct type with the exact layout of a tuple of those results.

- Parameters classified as indirect by high-level lowering should be
  represented as parameters of pointer type.

- Parameters classified as direct by high-level lowering should be
  omitted if they are empty types; otherwise, they should be represented
  as a parameter type with a layout exactly matching the layout of the
  Swift parameter type.

- The context parameter, if present, should be represented as a trailing
  parameter with the ``swift_context`` attribute.

- The error result parameter, if present, should be represented as a
  trailing parameter (always following a context parameter) with the
  ``swift_error_result`` attribute.

``swiftcall`` does not support variadic arguments or unprototyped functions.

The parameter ABI treatment attributes are aspects of the function type.
A function type which which applies an ABI treatment attribute to a
parameter is a different type from an otherwise-identical function type
that does not.  A single parameter may not have multiple ABI treatment
attributes.

Support for this feature is target-dependent, although it should be
supported on every target that Swift supports.  Query for this support
with ``__has_attribute(swiftcall)``.  This implies support for the
``swift_context``, ``swift_error_result``, and ``swift_indirect_result``
attributes.


swift_context (gnu::swift_context)
----------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

The ``swift_context`` attribute marks a parameter of a ``swiftcall``
function as having the special context-parameter ABI treatment.

This treatment generally passes the context value in a special register
which is normally callee-preserved.

A ``swift_context`` parameter must either be the last parameter or must be
followed by a ``swift_error_result`` parameter (which itself must always be
the last parameter).

A context parameter must have pointer or reference type.


swift_error_result (gnu::swift_error_result)
--------------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

The ``swift_error_result`` attribute marks a parameter of a ``swiftcall``
function as having the special error-result ABI treatment.

This treatment generally passes the underlying error value in and out of
the function through a special register which is normally callee-preserved.
This is modeled in C by pretending that the register is addressable memory:

- The caller appears to pass the address of a variable of pointer type.
  The current value of this variable is copied into the register before
  the call; if the call returns normally, the value is copied back into the
  variable.

- The callee appears to receive the address of a variable.  This address
  is actually a hidden location in its own stack, initialized with the
  value of the register upon entry.  When the function returns normally,
  the value in that hidden location is written back to the register.

A ``swift_error_result`` parameter must be the last parameter, and it must be
preceded by a ``swift_context`` parameter.

A ``swift_error_result`` parameter must have type ``T**`` or ``T*&`` for some
type T.  Note that no qualifiers are permitted on the intermediate level.

It is undefined behavior if the caller does not pass a pointer or
reference to a valid object.

The standard convention is that the error value itself (that is, the
value stored in the apparent argument) will be null upon function entry,
but this is not enforced by the ABI.


swift_indirect_result (gnu::swift_indirect_result)
--------------------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

The ``swift_indirect_result`` attribute marks a parameter of a ``swiftcall``
function as having the special indirect-result ABI treatment.

This treatment gives the parameter the target's normal indirect-result
ABI treatment, which may involve passing it differently from an ordinary
parameter.  However, only the first indirect result will receive this
treatment.  Furthermore, low-level lowering may decide that a direct result
must be returned indirectly; if so, this will take priority over the
``swift_indirect_result`` parameters.

A ``swift_indirect_result`` parameter must either be the first parameter or
follow another ``swift_indirect_result`` parameter.

A ``swift_indirect_result`` parameter must have type ``T*`` or ``T&`` for
some object type ``T``.  If ``T`` is a complete type at the point of
definition of a function, it is undefined behavior if the argument
value does not point to storage of adequate size and alignment for a
value of type ``T``.

Making indirect results explicit in the signature allows C functions to
directly construct objects into them without relying on language
optimizations like C++'s named return value optimization (NRVO).


tls_model (gnu::tls_model)
--------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

The ``tls_model`` attribute allows you to specify which thread-local storage
model to use. It accepts the following strings:

* global-dynamic
* local-dynamic
* initial-exec
* local-exec

TLS models are mutually exclusive.


thread
------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","X","", ""

The ``__declspec(thread)`` attribute declares a variable with thread local
storage.  It is available under the ``-fms-extensions`` flag for MSVC
compatibility.  See the documentation for `__declspec(thread)`_ on MSDN.

.. _`__declspec(thread)`: http://msdn.microsoft.com/en-us/library/9w1sdazb.aspx

In Clang, ``__declspec(thread)`` is generally equivalent in functionality to the
GNU ``__thread`` keyword.  The variable must not have a destructor and must have
a constant initializer, if any.  The attribute only applies to variables
declared with static storage duration, such as globals, class static data
members, and static locals.


maybe_unused, unused, gnu::unused
---------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

When passing the ``-Wunused`` flag to Clang, entities that are unused by the
program may be diagnosed. The ``[[maybe_unused]]`` (or
``__attribute__((unused))``) attribute can be used to silence such diagnostics
when the entity cannot be removed. For instance, a local variable may exist
solely for use in an ``assert()`` statement, which makes the local variable
unused when ``NDEBUG`` is defined.

The attribute may be applied to the declaration of a class, a typedef, a
variable, a function or method, a function parameter, an enumeration, an
enumerator, a non-static data member, or a label.

.. code-block: c++
  #include <cassert>

  [[maybe_unused]] void f([[maybe_unused]] bool thing1,
                          [[maybe_unused]] bool thing2) {
    [[maybe_unused]] bool b = thing1 && thing2;
    assert(b);
  }


Type Attributes
===============


align_value
-----------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

The align_value attribute can be added to the typedef of a pointer type or the
declaration of a variable of pointer or reference type. It specifies that the
pointer will point to, or the reference will bind to, only objects with at
least the provided alignment. This alignment value must be some positive power
of 2.

   .. code-block:: c

     typedef double * aligned_double_ptr __attribute__((align_value(64)));
     void foo(double & x  __attribute__((align_value(128)),
              aligned_double_ptr y) { ... }

If the pointer value does not have the specified alignment at runtime, the
behavior of the program is undefined.


empty_bases
-----------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","X","", ""

The empty_bases attribute permits the compiler to utilize the
empty-base-optimization more frequently.
This attribute only applies to struct, class, and union types.
It is only supported when using the Microsoft C++ ABI.


flag_enum
---------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

This attribute can be added to an enumerator to signal to the compiler that it
is intended to be used as a flag type. This will cause the compiler to assume
that the range of the type includes all of the values that you can get by
manipulating bits of the enumerator when issuing warnings.


lto_visibility_public (clang::lto_visibility_public)
----------------------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","X","","", ""

See :doc:`LTOVisibility`.


layout_version
--------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","X","", ""

The layout_version attribute requests that the compiler utilize the class
layout rules of a particular compiler version.
This attribute only applies to struct, class, and union types.
It is only supported when using the Microsoft C++ ABI.


__single_inhertiance, __multiple_inheritance, __virtual_inheritance
-------------------------------------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","","X", ""

This collection of keywords is enabled under ``-fms-extensions`` and controls
the pointer-to-member representation used on ``*-*-win32`` targets.

The ``*-*-win32`` targets utilize a pointer-to-member representation which
varies in size and alignment depending on the definition of the underlying
class.

However, this is problematic when a forward declaration is only available and
no definition has been made yet.  In such cases, Clang is forced to utilize the
most general representation that is available to it.

These keywords make it possible to use a pointer-to-member representation other
than the most general one regardless of whether or not the definition will ever
be present in the current translation unit.

This family of keywords belong between the ``class-key`` and ``class-name``:

.. code-block:: c++

  struct __single_inheritance S;
  int S::*i;
  struct S {};

This keyword can be applied to class templates but only has an effect when used
on full specializations:

.. code-block:: c++

  template <typename T, typename U> struct __single_inheritance A; // warning: inheritance model ignored on primary template
  template <typename T> struct __multiple_inheritance A<T, T>; // warning: inheritance model ignored on partial specialization
  template <> struct __single_inheritance A<int, float>;

Note that choosing an inheritance model less general than strictly necessary is
an error:

.. code-block:: c++

  struct __multiple_inheritance S; // error: inheritance model does not match definition
  int S::*i;
  struct S {};


novtable
--------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","X","", ""

This attribute can be added to a class declaration or definition to signal to
the compiler that constructors and destructors will not reference the virtual
function table. It is only supported when using the Microsoft C++ ABI.


objc_subclassing_restricted
---------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

This attribute can be added to an Objective-C ``@interface`` declaration to
ensure that this class cannot be subclassed.


transparent_union (gnu::transparent_union)
------------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

This attribute can be applied to a union to change the behaviour of calls to
functions that have an argument with a transparent union type. The compiler
behaviour is changed in the following manner:

- A value whose type is any member of the transparent union can be passed as an
  argument without the need to cast that value.

- The argument is passed to the function using the calling convention of the
  first member of the transparent union. Consequently, all the members of the
  transparent union should have the same calling convention as its first member.

Transparent unions are not supported in C++.


Statement Attributes
====================


fallthrough, clang::fallthrough
-------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","X","","", ""

The ``fallthrough`` (or ``clang::fallthrough``) attribute is used
to annotate intentional fall-through
between switch labels.  It can only be applied to a null statement placed at a
point of execution between any statement and the next switch label.  It is
common to mark these places with a specific comment, but this attribute is
meant to replace comments with a more strict annotation, which can be checked
by the compiler.  This attribute doesn't change semantics of the code and can
be used wherever an intended fall-through occurs.  It is designed to mimic
control-flow statements like ``break;``, so it can be placed in most places
where ``break;`` can, but only if there are no statements on the execution path
between it and the next switch label.

By default, Clang does not warn on unannotated fallthrough from one ``switch``
case to another. Diagnostics on fallthrough without a corresponding annotation
can be enabled with the ``-Wimplicit-fallthrough`` argument.

Here is an example:

.. code-block:: c++

  // compile with -Wimplicit-fallthrough
  switch (n) {
  case 22:
  case 33:  // no warning: no statements between case labels
    f();
  case 44:  // warning: unannotated fall-through
    g();
    [[clang::fallthrough]];
  case 55:  // no warning
    if (x) {
      h();
      break;
    }
    else {
      i();
      [[clang::fallthrough]];
    }
  case 66:  // no warning
    p();
    [[clang::fallthrough]]; // warning: fallthrough annotation does not
                            //          directly precede case label
    q();
  case 77:  // warning: unannotated fall-through
    r();
  }


#pragma clang loop
------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","","", "X"

The ``#pragma clang loop`` directive allows loop optimization hints to be
specified for the subsequent loop. The directive allows vectorization,
interleaving, and unrolling to be enabled or disabled. Vector width as well
as interleave and unrolling count can be manually specified. See
`language extensions
<http://clang.llvm.org/docs/LanguageExtensions.html#extensions-for-loop-hint-optimizations>`_
for details.


#pragma unroll, #pragma nounroll
--------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","","", "X"

Loop unrolling optimization hints can be specified with ``#pragma unroll`` and
``#pragma nounroll``. The pragma is placed immediately before a for, while,
do-while, or c++11 range-based for loop.

Specifying ``#pragma unroll`` without a parameter directs the loop unroller to
attempt to fully unroll the loop if the trip count is known at compile time and
attempt to partially unroll the loop if the trip count is not known at compile
time:

.. code-block:: c++

  #pragma unroll
  for (...) {
    ...
  }

Specifying the optional parameter, ``#pragma unroll _value_``, directs the
unroller to unroll the loop ``_value_`` times.  The parameter may optionally be
enclosed in parentheses:

.. code-block:: c++

  #pragma unroll 16
  for (...) {
    ...
  }

  #pragma unroll(16)
  for (...) {
    ...
  }

Specifying ``#pragma nounroll`` indicates that the loop should not be unrolled:

.. code-block:: c++

  #pragma nounroll
  for (...) {
    ...
  }

``#pragma unroll`` and ``#pragma unroll _value_`` have identical semantics to
``#pragma clang loop unroll(full)`` and
``#pragma clang loop unroll_count(_value_)`` respectively. ``#pragma nounroll``
is equivalent to ``#pragma clang loop unroll(disable)``.  See
`language extensions
<http://clang.llvm.org/docs/LanguageExtensions.html#extensions-for-loop-hint-optimizations>`_
for further details including limitations of the unroll hints.


__read_only, __write_only, __read_write (read_only, write_only, read_write)
---------------------------------------------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","","X", ""

The access qualifiers must be used with image object arguments or pipe arguments
to declare if they are being read or written by a kernel or function.

The read_only/__read_only, write_only/__write_only and read_write/__read_write
names are reserved for use as access qualifiers and shall not be used otherwise.

.. code-block:: c

  kernel void
  foo (read_only image2d_t imageA,
       write_only image2d_t imageB) {
    ...
  }

In the above example imageA is a read-only 2D image object, and imageB is a
write-only 2D image object.

The read_write (or __read_write) qualifier can not be used with pipe.

More details can be found in the OpenCL C language Spec v2.0, Section 6.6.


__attribute__((opencl_unroll_hint))
-----------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

The opencl_unroll_hint attribute qualifier can be used to specify that a loop
(for, while and do loops) can be unrolled. This attribute qualifier can be
used to specify full unrolling or partial unrolling by a specified amount.
This is a compiler hint and the compiler may ignore this directive. See
`OpenCL v2.0 <https://www.khronos.org/registry/cl/specs/opencl-2.0.pdf>`_
s6.11.5 for details.


Type Safety Checking
====================
Clang supports additional attributes to enable checking type safety properties
that can't be enforced by the C type system. To see warnings produced by these
checks, ensure that -Wtype-safety is enabled. Use cases include:

* MPI library implementations, where these attributes enable checking that
  the buffer type matches the passed ``MPI_Datatype``;
* for HDF5 library there is a similar use case to MPI;
* checking types of variadic functions' arguments for functions like
  ``fcntl()`` and ``ioctl()``.

You can detect support for these attributes with ``__has_attribute()``.  For
example:

.. code-block:: c++

  #if defined(__has_attribute)
  #  if __has_attribute(argument_with_type_tag) && \
        __has_attribute(pointer_with_type_tag) && \
        __has_attribute(type_tag_for_datatype)
  #    define ATTR_MPI_PWT(buffer_idx, type_idx) __attribute__((pointer_with_type_tag(mpi,buffer_idx,type_idx)))
  /* ... other macros ...  */
  #  endif
  #endif

  #if !defined(ATTR_MPI_PWT)
  # define ATTR_MPI_PWT(buffer_idx, type_idx)
  #endif

  int MPI_Send(void *buf, int count, MPI_Datatype datatype /*, other args omitted */)
      ATTR_MPI_PWT(1,3);

argument_with_type_tag
----------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

Use ``__attribute__((argument_with_type_tag(arg_kind, arg_idx,
type_tag_idx)))`` on a function declaration to specify that the function
accepts a type tag that determines the type of some other argument.

This attribute is primarily useful for checking arguments of variadic functions
(``pointer_with_type_tag`` can be used in most non-variadic cases).

In the attribute prototype above:
  * ``arg_kind`` is an identifier that should be used when annotating all
    applicable type tags.
  * ``arg_idx`` provides the position of a function argument. The expected type of
    this function argument will be determined by the function argument specified
    by ``type_tag_idx``. In the code example below, "3" means that the type of the
    function's third argument will be determined by ``type_tag_idx``.
  * ``type_tag_idx`` provides the position of a function argument. This function
    argument will be a type tag. The type tag will determine the expected type of
    the argument specified by ``arg_idx``. In the code example below, "2" means
    that the type tag associated with the function's second argument should agree
    with the type of the argument specified by ``arg_idx``.

For example:

.. code-block:: c++

  int fcntl(int fd, int cmd, ...)
      __attribute__(( argument_with_type_tag(fcntl,3,2) ));
  // The function's second argument will be a type tag; this type tag will
  // determine the expected type of the function's third argument.


pointer_with_type_tag
---------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

Use ``__attribute__((pointer_with_type_tag(ptr_kind, ptr_idx, type_tag_idx)))``
on a function declaration to specify that the function accepts a type tag that
determines the pointee type of some other pointer argument.

In the attribute prototype above:
  * ``ptr_kind`` is an identifier that should be used when annotating all
    applicable type tags.
  * ``ptr_idx`` provides the position of a function argument; this function
    argument will have a pointer type. The expected pointee type of this pointer
    type will be determined by the function argument specified by
    ``type_tag_idx``. In the code example below, "1" means that the pointee type
    of the function's first argument will be determined by ``type_tag_idx``.
  * ``type_tag_idx`` provides the position of a function argument; this function
    argument will be a type tag. The type tag will determine the expected pointee
    type of the pointer argument specified by ``ptr_idx``. In the code example
    below, "3" means that the type tag associated with the function's third
    argument should agree with the pointee type of the pointer argument specified
    by ``ptr_idx``.

For example:

.. code-block:: c++

  typedef int MPI_Datatype;
  int MPI_Send(void *buf, int count, MPI_Datatype datatype /*, other args omitted */)
      __attribute__(( pointer_with_type_tag(mpi,1,3) ));
  // The function's 3rd argument will be a type tag; this type tag will
  // determine the expected pointee type of the function's 1st argument.


type_tag_for_datatype
---------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

When declaring a variable, use
``__attribute__((type_tag_for_datatype(kind, type)))`` to create a type tag that
is tied to the ``type`` argument given to the attribute.

In the attribute prototype above:
  * ``kind`` is an identifier that should be used when annotating all applicable
    type tags.
  * ``type`` indicates the name of the type.

Clang supports annotating type tags of two forms.

  * **Type tag that is a reference to a declared identifier.**
    Use ``__attribute__((type_tag_for_datatype(kind, type)))`` when declaring that
    identifier:

    .. code-block:: c++

      typedef int MPI_Datatype;
      extern struct mpi_datatype mpi_datatype_int
          __attribute__(( type_tag_for_datatype(mpi,int) ));
      #define MPI_INT ((MPI_Datatype) &mpi_datatype_int)
      // &mpi_datatype_int is a type tag. It is tied to type "int".

  * **Type tag that is an integral literal.**
    Declare a ``static const`` variable with an initializer value and attach
    ``__attribute__((type_tag_for_datatype(kind, type)))`` on that declaration:

    .. code-block:: c++

      typedef int MPI_Datatype;
      static const MPI_Datatype mpi_datatype_int
          __attribute__(( type_tag_for_datatype(mpi,int) )) = 42;
      #define MPI_INT ((MPI_Datatype) 42)
      // The number 42 is a type tag. It is tied to type "int".


The ``type_tag_for_datatype`` attribute also accepts an optional third argument
that determines how the type of the function argument specified by either
``arg_idx`` or ``ptr_idx`` is compared against the type associated with the type
tag. (Recall that for the ``argument_with_type_tag`` attribute, the type of the
function argument specified by ``arg_idx`` is compared against the type
associated with the type tag. Also recall that for the ``pointer_with_type_tag``
attribute, the pointee type of the function argument specified by ``ptr_idx`` is
compared against the type associated with the type tag.) There are two supported
values for this optional third argument:

  * ``layout_compatible`` will cause types to be compared according to
    layout-compatibility rules (In C++11 [class.mem] p 17, 18, see the
    layout-compatibility rules for two standard-layout struct types and for two
    standard-layout union types). This is useful when creating a type tag
    associated with a struct or union type. For example:

    .. code-block:: c++

      /* In mpi.h */
      typedef int MPI_Datatype;
      struct internal_mpi_double_int { double d; int i; };
      extern struct mpi_datatype mpi_datatype_double_int
          __attribute__(( type_tag_for_datatype(mpi,
                          struct internal_mpi_double_int, layout_compatible) ));

      #define MPI_DOUBLE_INT ((MPI_Datatype) &mpi_datatype_double_int)

      int MPI_Send(void *buf, int count, MPI_Datatype datatype, ...)
          __attribute__(( pointer_with_type_tag(mpi,1,3) ));

      /* In user code */
      struct my_pair { double a; int b; };
      struct my_pair *buffer;
      MPI_Send(buffer, 1, MPI_DOUBLE_INT /*, ...  */); // no warning because the
                                                       // layout of my_pair is
                                                       // compatible with that of
                                                       // internal_mpi_double_int

      struct my_int_pair { int a; int b; }
      struct my_int_pair *buffer2;
      MPI_Send(buffer2, 1, MPI_DOUBLE_INT /*, ...  */); // warning because the
                                                        // layout of my_int_pair
                                                        // does not match that of
                                                        // internal_mpi_double_int

  * ``must_be_null`` specifies that the function argument specified by either
    ``arg_idx`` (for the ``argument_with_type_tag`` attribute) or ``ptr_idx`` (for
    the ``pointer_with_type_tag`` attribute) should be a null pointer constant.
    The second argument to the ``type_tag_for_datatype`` attribute is ignored. For
    example:

    .. code-block:: c++

      /* In mpi.h */
      typedef int MPI_Datatype;
      extern struct mpi_datatype mpi_datatype_null
          __attribute__(( type_tag_for_datatype(mpi, void, must_be_null) ));

      #define MPI_DATATYPE_NULL ((MPI_Datatype) &mpi_datatype_null)
      int MPI_Send(void *buf, int count, MPI_Datatype datatype, ...)
          __attribute__(( pointer_with_type_tag(mpi,1,3) ));

      /* In user code */
      struct my_pair { double a; int b; };
      struct my_pair *buffer;
      MPI_Send(buffer, 1, MPI_DATATYPE_NULL /*, ...  */); // warning: MPI_DATATYPE_NULL
                                                          // was specified but buffer
                                                          // is not a null pointer


AMD GPU Attributes
==================


amdgpu_flat_work_group_size
---------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

The flat work-group size is the number of work-items in the work-group size
specified when the kernel is dispatched. It is the product of the sizes of the
x, y, and z dimension of the work-group.

Clang supports the
``__attribute__((amdgpu_flat_work_group_size(<min>, <max>)))`` attribute for the
AMDGPU target. This attribute may be attached to a kernel function definition
and is an optimization hint.

``<min>`` parameter specifies the minimum flat work-group size, and ``<max>``
parameter specifies the maximum flat work-group size (must be greater than
``<min>``) to which all dispatches of the kernel will conform. Passing ``0, 0``
as ``<min>, <max>`` implies the default behavior (``128, 256``).

If specified, the AMDGPU target backend might be able to produce better machine
code for barriers and perform scratch promotion by estimating available group
segment size.

An error will be given if:
  - Specified values violate subtarget specifications;
  - Specified values are not compatible with values provided through other
    attributes.


amdgpu_num_sgpr
---------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

Clang supports the ``__attribute__((amdgpu_num_sgpr(<num_sgpr>)))`` and
``__attribute__((amdgpu_num_vgpr(<num_vgpr>)))`` attributes for the AMDGPU
target. These attributes may be attached to a kernel function definition and are
an optimization hint.

If these attributes are specified, then the AMDGPU target backend will attempt
to limit the number of SGPRs and/or VGPRs used to the specified value(s). The
number of used SGPRs and/or VGPRs may further be rounded up to satisfy the
allocation requirements or constraints of the subtarget. Passing ``0`` as
``num_sgpr`` and/or ``num_vgpr`` implies the default behavior (no limits).

These attributes can be used to test the AMDGPU target backend. It is
recommended that the ``amdgpu_waves_per_eu`` attribute be used to control
resources such as SGPRs and VGPRs since it is aware of the limits for different
subtargets.

An error will be given if:
  - Specified values violate subtarget specifications;
  - Specified values are not compatible with values provided through other
    attributes;
  - The AMDGPU target backend is unable to create machine code that can meet the
    request.


amdgpu_num_vgpr
---------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

Clang supports the ``__attribute__((amdgpu_num_sgpr(<num_sgpr>)))`` and
``__attribute__((amdgpu_num_vgpr(<num_vgpr>)))`` attributes for the AMDGPU
target. These attributes may be attached to a kernel function definition and are
an optimization hint.

If these attributes are specified, then the AMDGPU target backend will attempt
to limit the number of SGPRs and/or VGPRs used to the specified value(s). The
number of used SGPRs and/or VGPRs may further be rounded up to satisfy the
allocation requirements or constraints of the subtarget. Passing ``0`` as
``num_sgpr`` and/or ``num_vgpr`` implies the default behavior (no limits).

These attributes can be used to test the AMDGPU target backend. It is
recommended that the ``amdgpu_waves_per_eu`` attribute be used to control
resources such as SGPRs and VGPRs since it is aware of the limits for different
subtargets.

An error will be given if:
  - Specified values violate subtarget specifications;
  - Specified values are not compatible with values provided through other
    attributes;
  - The AMDGPU target backend is unable to create machine code that can meet the
    request.


amdgpu_waves_per_eu
-------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

A compute unit (CU) is responsible for executing the wavefronts of a work-group.
It is composed of one or more execution units (EU), which are responsible for
executing the wavefronts. An EU can have enough resources to maintain the state
of more than one executing wavefront. This allows an EU to hide latency by
switching between wavefronts in a similar way to symmetric multithreading on a
CPU. In order to allow the state for multiple wavefronts to fit on an EU, the
resources used by a single wavefront have to be limited. For example, the number
of SGPRs and VGPRs. Limiting such resources can allow greater latency hiding,
but can result in having to spill some register state to memory.

Clang supports the ``__attribute__((amdgpu_waves_per_eu(<min>[, <max>])))``
attribute for the AMDGPU target. This attribute may be attached to a kernel
function definition and is an optimization hint.

``<min>`` parameter specifies the requested minimum number of waves per EU, and
*optional* ``<max>`` parameter specifies the requested maximum number of waves
per EU (must be greater than ``<min>`` if specified). If ``<max>`` is omitted,
then there is no restriction on the maximum number of waves per EU other than
the one dictated by the hardware for which the kernel is compiled. Passing
``0, 0`` as ``<min>, <max>`` implies the default behavior (no limits).

If specified, this attribute allows an advanced developer to tune the number of
wavefronts that are capable of fitting within the resources of an EU. The AMDGPU
target backend can use this information to limit resources, such as number of
SGPRs, number of VGPRs, size of available group and private memory segments, in
such a way that guarantees that at least ``<min>`` wavefronts and at most
``<max>`` wavefronts are able to fit within the resources of an EU. Requesting
more wavefronts can hide memory latency but limits available registers which
can result in spilling. Requesting fewer wavefronts can help reduce cache
thrashing, but can reduce memory latency hiding.

This attribute controls the machine code generated by the AMDGPU target backend
to ensure it is capable of meeting the requested values. However, when the
kernel is executed, there may be other reasons that prevent meeting the request,
for example, there may be wavefronts from other kernels executing on the EU.

An error will be given if:
  - Specified values violate subtarget specifications;
  - Specified values are not compatible with values provided through other
    attributes;
  - The AMDGPU target backend is unable to create machine code that can meet the
    request.


Calling Conventions
===================
Clang supports several different calling conventions, depending on the target
platform and architecture. The calling convention used for a function determines
how parameters are passed, how results are returned to the caller, and other
low-level details of calling a function.

fastcall (gnu::fastcall, __fastcall, _fastcall)
-----------------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","X", ""

On 32-bit x86 targets, this attribute changes the calling convention of a
function to use ECX and EDX as register parameters and clear parameters off of
the stack on return. This convention does not support variadic calls or
unprototyped functions in C, and has no effect on x86_64 targets. This calling
convention is supported primarily for compatibility with existing code. Users
seeking register parameters should use the ``regparm`` attribute, which does
not require callee-cleanup.  See the documentation for `__fastcall`_ on MSDN.

.. _`__fastcall`: http://msdn.microsoft.com/en-us/library/6xa169sk.aspx


ms_abi (gnu::ms_abi)
--------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

On non-Windows x86_64 targets, this attribute changes the calling convention of
a function to match the default convention used on Windows x86_64. This
attribute has no effect on Windows targets or non-x86_64 targets.


pcs (gnu::pcs)
--------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

On ARM targets, this attribute can be used to select calling conventions
similar to ``stdcall`` on x86. Valid parameter values are "aapcs" and
"aapcs-vfp".


preserve_all
------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

On X86-64 and AArch64 targets, this attribute changes the calling convention of
a function. The ``preserve_all`` calling convention attempts to make the code
in the caller even less intrusive than the ``preserve_most`` calling convention.
This calling convention also behaves identical to the ``C`` calling convention
on how arguments and return values are passed, but it uses a different set of
caller/callee-saved registers. This removes the burden of saving and
recovering a large register set before and after the call in the caller. If
the arguments are passed in callee-saved registers, then they will be
preserved by the callee across the call. This doesn't apply for values
returned in callee-saved registers.

- On X86-64 the callee preserves all general purpose registers, except for
  R11. R11 can be used as a scratch register. Furthermore it also preserves
  all floating-point registers (XMMs/YMMs).

The idea behind this convention is to support calls to runtime functions
that don't need to call out to any other functions.

This calling convention, like the ``preserve_most`` calling convention, will be
used by a future version of the Objective-C runtime and should be considered
experimental at this time.


preserve_most
-------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

On X86-64 and AArch64 targets, this attribute changes the calling convention of
a function. The ``preserve_most`` calling convention attempts to make the code
in the caller as unintrusive as possible. This convention behaves identically
to the ``C`` calling convention on how arguments and return values are passed,
but it uses a different set of caller/callee-saved registers. This alleviates
the burden of saving and recovering a large register set before and after the
call in the caller. If the arguments are passed in callee-saved registers,
then they will be preserved by the callee across the call. This doesn't
apply for values returned in callee-saved registers.

- On X86-64 the callee preserves all general purpose registers, except for
  R11. R11 can be used as a scratch register. Floating-point registers
  (XMMs/YMMs) are not preserved and need to be saved by the caller.

The idea behind this convention is to support calls to runtime functions
that have a hot path and a cold path. The hot path is usually a small piece
of code that doesn't use many registers. The cold path might need to call out to
another function and therefore only needs to preserve the caller-saved
registers, which haven't already been saved by the caller. The
`preserve_most` calling convention is very similar to the ``cold`` calling
convention in terms of caller/callee-saved registers, but they are used for
different types of function calls. ``coldcc`` is for function calls that are
rarely executed, whereas `preserve_most` function calls are intended to be
on the hot path and definitely executed a lot. Furthermore ``preserve_most``
doesn't prevent the inliner from inlining the function call.

This calling convention will be used by a future version of the Objective-C
runtime and should therefore still be considered experimental at this time.
Although this convention was created to optimize certain runtime calls to
the Objective-C runtime, it is not limited to this runtime and might be used
by other runtimes in the future too. The current implementation only
supports X86-64 and AArch64, but the intention is to support more architectures
in the future.


regcall (gnu::regcall, __regcall)
---------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","X", ""

On x86 targets, this attribute changes the calling convention to
`__regcall`_ convention. This convention aims to pass as many arguments
as possible in registers. It also tries to utilize registers for the
return value whenever it is possible.

.. _`__regcall`: https://software.intel.com/en-us/node/693069


regparm (gnu::regparm)
----------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

On 32-bit x86 targets, the regparm attribute causes the compiler to pass
the first three integer parameters in EAX, EDX, and ECX instead of on the
stack. This attribute has no effect on variadic functions, and all parameters
are passed via the stack as normal.


stdcall (gnu::stdcall, __stdcall, _stdcall)
-------------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","X", ""

On 32-bit x86 targets, this attribute changes the calling convention of a
function to clear parameters off of the stack on return. This convention does
not support variadic calls or unprototyped functions in C, and has no effect on
x86_64 targets. This calling convention is used widely by the Windows API and
COM applications.  See the documentation for `__stdcall`_ on MSDN.

.. _`__stdcall`: http://msdn.microsoft.com/en-us/library/zxk0tw93.aspx


thiscall (gnu::thiscall, __thiscall, _thiscall)
-----------------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","X", ""

On 32-bit x86 targets, this attribute changes the calling convention of a
function to use ECX for the first parameter (typically the implicit ``this``
parameter of C++ methods) and clear parameters off of the stack on return. This
convention does not support variadic calls or unprototyped functions in C, and
has no effect on x86_64 targets. See the documentation for `__thiscall`_ on
MSDN.

.. _`__thiscall`: http://msdn.microsoft.com/en-us/library/ek8tkfbw.aspx


vectorcall (__vectorcall, _vectorcall)
--------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","X", ""

On 32-bit x86 *and* x86_64 targets, this attribute changes the calling
convention of a function to pass vector parameters in SSE registers.

On 32-bit x86 targets, this calling convention is similar to ``__fastcall``.
The first two integer parameters are passed in ECX and EDX. Subsequent integer
parameters are passed in memory, and callee clears the stack.  On x86_64
targets, the callee does *not* clear the stack, and integer parameters are
passed in RCX, RDX, R8, and R9 as is done for the default Windows x64 calling
convention.

On both 32-bit x86 and x86_64 targets, vector and floating point arguments are
passed in XMM0-XMM5. Homogeneous vector aggregates of up to four elements are
passed in sequential SSE registers if enough are available. If AVX is enabled,
256 bit vectors are passed in YMM0-YMM5. Any vector or aggregate type that
cannot be passed in registers for any reason is passed by reference, which
allows the caller to align the parameter memory.

See the documentation for `__vectorcall`_ on MSDN for more details.

.. _`__vectorcall`: http://msdn.microsoft.com/en-us/library/dn375768.aspx


Consumed Annotation Checking
============================
Clang supports additional attributes for checking basic resource management
properties, specifically for unique objects that have a single owning reference.
The following attributes are currently supported, although **the implementation
for these annotations is currently in development and are subject to change.**

callable_when
-------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

Use ``__attribute__((callable_when(...)))`` to indicate what states a method
may be called in.  Valid states are unconsumed, consumed, or unknown.  Each
argument to this attribute must be a quoted string.  E.g.:

``__attribute__((callable_when("unconsumed", "unknown")))``


consumable
----------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

Each ``class`` that uses any of the typestate annotations must first be marked
using the ``consumable`` attribute.  Failure to do so will result in a warning.

This attribute accepts a single parameter that must be one of the following:
``unknown``, ``consumed``, or ``unconsumed``.


param_typestate
---------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

This attribute specifies expectations about function parameters.  Calls to an
function with annotated parameters will issue a warning if the corresponding
argument isn't in the expected state.  The attribute is also used to set the
initial state of the parameter when analyzing the function's body.


return_typestate
----------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

The ``return_typestate`` attribute can be applied to functions or parameters.
When applied to a function the attribute specifies the state of the returned
value.  The function's body is checked to ensure that it always returns a value
in the specified state.  On the caller side, values returned by the annotated
function are initialized to the given state.

When applied to a function parameter it modifies the state of an argument after
a call to the function returns.  The function's body is checked to ensure that
the parameter is in the expected state before returning.


set_typestate
-------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

Annotate methods that transition an object into a new state with
``__attribute__((set_typestate(new_state)))``.  The new state must be
unconsumed, consumed, or unknown.


test_typestate
--------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","","","", ""

Use ``__attribute__((test_typestate(tested_state)))`` to indicate that a method
returns true if the object is in the specified state..


OpenCL Address Spaces
=====================
The address space qualifier may be used to specify the region of memory that is
used to allocate the object. OpenCL supports the following address spaces:
__generic(generic), __global(global), __local(local), __private(private),
__constant(constant).

  .. code-block:: c

    __constant int c = ...;

    __generic int* foo(global int* g) {
      __local int* l;
      private int p;
      ...
      return l;
    }

More details can be found in the OpenCL C language Spec v2.0, Section 6.5.

constant (__constant)
---------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","","X", ""

The constant address space attribute signals that an object is located in
a constant (non-modifiable) memory region. It is available to all work items.
Any type can be annotated with the constant address space attribute. Objects
with the constant address space qualifier can be declared in any scope and must
have an initializer.


generic (__generic)
-------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","","X", ""

The generic address space attribute is only available with OpenCL v2.0 and later.
It can be used with pointer types. Variables in global and local scope and
function parameters in non-kernel functions can have the generic address space
type attribute. It is intended to be a placeholder for any other address space
except for '__constant' in OpenCL code which can be used with multiple address
spaces.


global (__global)
-----------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","","X", ""

The global address space attribute specifies that an object is allocated in
global memory, which is accessible by all work items. The content stored in this
memory area persists between kernel executions. Pointer types to the global
address space are allowed as function parameters or local variables. Starting
with OpenCL v2.0, the global address space can be used with global (program
scope) variables and static local variable as well.


local (__local)
---------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","","X", ""

The local address space specifies that an object is allocated in the local (work
group) memory area, which is accessible to all work items in the same work
group. The content stored in this memory region is not accessible after
the kernel execution ends. In a kernel function scope, any variable can be in
the local address space. In other scopes, only pointer types to the local address
space are allowed. Local address space variables cannot have an initializer.


private (__private)
-------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","","X", ""

The private address space specifies that an object is allocated in the private
(work item) memory. Other work items cannot access the same memory area and its
content is destroyed after work item execution ends. Local variables can be
declared in the private address space. Function arguments are always in the
private address space. Kernel function arguments of a pointer or an array type
cannot point to the private address space.


Nullability Attributes
======================
Whether a particular pointer may be "null" is an important concern when working with pointers in the C family of languages. The various nullability attributes indicate whether a particular pointer can be null or not, which makes APIs more expressive and can help static analysis tools identify bugs involving null pointers. Clang supports several kinds of nullability attributes: the ``nonnull`` and ``returns_nonnull`` attributes indicate which function or method parameters and result types can never be null, while nullability type qualifiers indicate which pointer types can be null (``_Nullable``) or cannot be null (``_Nonnull``).

The nullability (type) qualifiers express whether a value of a given pointer type can be null (the ``_Nullable`` qualifier), doesn't have a defined meaning for null (the ``_Nonnull`` qualifier), or for which the purpose of null is unclear (the ``_Null_unspecified`` qualifier). Because nullability qualifiers are expressed within the type system, they are more general than the ``nonnull`` and ``returns_nonnull`` attributes, allowing one to express (for example) a nullable pointer to an array of nonnull pointers. Nullability qualifiers are written to the right of the pointer to which they apply. For example:

  .. code-block:: c

    // No meaningful result when 'ptr' is null (here, it happens to be undefined behavior).
    int fetch(int * _Nonnull ptr) { return *ptr; }

    // 'ptr' may be null.
    int fetch_or_zero(int * _Nullable ptr) {
      return ptr ? *ptr : 0;
    }

    // A nullable pointer to non-null pointers to const characters.
    const char *join_strings(const char * _Nonnull * _Nullable strings, unsigned n);

In Objective-C, there is an alternate spelling for the nullability qualifiers that can be used in Objective-C methods and properties using context-sensitive, non-underscored keywords. For example:

  .. code-block:: objective-c

    @interface NSView : NSResponder
      - (nullable NSView *)ancestorSharedWithView:(nonnull NSView *)aView;
      @property (assign, nullable) NSView *superview;
      @property (readonly, nonnull) NSArray *subviews;
    @end

nonnull (gnu::nonnull)
----------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

The ``nonnull`` attribute indicates that some function parameters must not be null, and can be used in several different ways. It's original usage (`from GCC <https://gcc.gnu.org/onlinedocs/gcc/Common-Function-Attributes.html#Common-Function-Attributes>`_) is as a function (or Objective-C method) attribute that specifies which parameters of the function are nonnull in a comma-separated list. For example:

  .. code-block:: c

    extern void * my_memcpy (void *dest, const void *src, size_t len)
                    __attribute__((nonnull (1, 2)));

Here, the ``nonnull`` attribute indicates that parameters 1 and 2
cannot have a null value. Omitting the parenthesized list of parameter indices means that all parameters of pointer type cannot be null:

  .. code-block:: c

    extern void * my_memcpy (void *dest, const void *src, size_t len)
                    __attribute__((nonnull));

Clang also allows the ``nonnull`` attribute to be placed directly on a function (or Objective-C method) parameter, eliminating the need to specify the parameter index ahead of type. For example:

  .. code-block:: c

    extern void * my_memcpy (void *dest __attribute__((nonnull)),
                             const void *src __attribute__((nonnull)), size_t len);

Note that the ``nonnull`` attribute indicates that passing null to a non-null parameter is undefined behavior, which the optimizer may take advantage of to, e.g., remove null checks. The ``_Nonnull`` type qualifier indicates that a pointer cannot be null in a more general manner (because it is part of the type system) and does not imply undefined behavior, making it more widely applicable.


returns_nonnull (gnu::returns_nonnull)
--------------------------------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "X","X","","", ""

The ``returns_nonnull`` attribute indicates that a particular function (or Objective-C method) always returns a non-null pointer. For example, a particular system ``malloc`` might be defined to terminate a process when memory is not available rather than returning a null pointer:

  .. code-block:: c

    extern void * malloc (size_t size) __attribute__((returns_nonnull));

The ``returns_nonnull`` attribute implies that returning a null pointer is undefined behavior, which the optimizer may take advantage of. The ``_Nonnull`` type qualifier indicates that a pointer cannot be null in a more general manner (because it is part of the type system) and does not imply undefined behavior, making it more widely applicable


_Nonnull
--------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","","X", ""

The ``_Nonnull`` nullability qualifier indicates that null is not a meaningful value for a value of the ``_Nonnull`` pointer type. For example, given a declaration such as:

  .. code-block:: c

    int fetch(int * _Nonnull ptr);

a caller of ``fetch`` should not provide a null value, and the compiler will produce a warning if it sees a literal null value passed to ``fetch``. Note that, unlike the declaration attribute ``nonnull``, the presence of ``_Nonnull`` does not imply that passing null is undefined behavior: ``fetch`` is free to consider null undefined behavior or (perhaps for backward-compatibility reasons) defensively handle null.


_Null_unspecified
-----------------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","","X", ""

The ``_Null_unspecified`` nullability qualifier indicates that neither the ``_Nonnull`` nor ``_Nullable`` qualifiers make sense for a particular pointer type. It is used primarily to indicate that the role of null with specific pointers in a nullability-annotated header is unclear, e.g., due to overly-complex implementations or historical factors with a long-lived API.


_Nullable
---------
.. csv-table:: Supported Syntaxes
   :header: "GNU", "C++11", "__declspec", "Keyword", "Pragma"

   "","","","X", ""

The ``_Nullable`` nullability qualifier indicates that a value of the ``_Nullable`` pointer type can be null. For example, given:

  .. code-block:: c

    int fetch_or_zero(int * _Nullable ptr);

a caller of ``fetch_or_zero`` can provide null.