<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
          "http://www.w3.org/TR/html4/strict.dtd">
<!-- Material used from: HTML 4.01 specs: http://www.w3.org/TR/html401/ -->
<html>
<head>
  <META http-equiv="Content-Type" content="text/html; charset=ISO-8859-1">
  <title>&lt;atomic&gt; design</title>
  <link type="text/css" rel="stylesheet" href="menu.css">
  <link type="text/css" rel="stylesheet" href="content.css">
</head>

<body>
<div id="menu">
  <div>
    <a href="https://llvm.org/">LLVM Home</a>
  </div>

  <div class="submenu">
    <label>libc++ Info</label>
    <a href="/index.html">About</a>
  </div>

  <div class="submenu">
    <label>Quick Links</label>
    <a href="https://lists.llvm.org/mailman/listinfo/cfe-dev">cfe-dev</a>
    <a href="https://lists.llvm.org/mailman/listinfo/cfe-commits">cfe-commits</a>
    <a href="https://bugs.llvm.org/">Bug Reports</a>
    <a href="https://github.com/llvm/llvm-project/tree/master/libcxx/">Browse Sources</a>
  </div>
</div>

<div id="content">
  <!--*********************************************************************-->
  <h1>&lt;atomic&gt; design</h1>
  <!--*********************************************************************-->

<p>
The <tt>&lt;atomic&gt;</tt> header is one of the most closely coupled headers to
the compiler.  Ideally when you invoke any function from
<tt>&lt;atomic&gt;</tt>, it should result in highly optimized assembly being
inserted directly into your application ...  assembly that is not otherwise
representable by higher level C or C++ expressions.  The design of the libc++
<tt>&lt;atomic&gt;</tt> header started with this goal in mind.  A secondary, but
still very important goal is that the compiler should have to do minimal work to
facilitate the implementation of <tt>&lt;atomic&gt;</tt>.  Without this second
goal, then practically speaking, the libc++ <tt>&lt;atomic&gt;</tt> header would
be doomed to be a barely supported, second class citizen on almost every
platform.
</p>

<p>Goals:</p>

<blockquote><ul>
<li>Optimal code generation for atomic operations</li>
<li>Minimal effort for the compiler to achieve goal 1 on any given platform</li>
<li>Conformance to the C++0X draft standard</li>
</ul></blockquote>

<p>
The purpose of this document is to inform compiler writers what they need to do
to enable a high performance libc++ <tt>&lt;atomic&gt;</tt> with minimal effort.
</p>

<h2>The minimal work that must be done for a conforming <tt>&lt;atomic&gt;</tt></h2>

<p>
The only "atomic" operations that must actually be lock free in
<tt>&lt;atomic&gt;</tt> are represented by the following compiler intrinsics:
</p>

<blockquote><pre>
__atomic_flag__
__atomic_exchange_seq_cst(__atomic_flag__ volatile* obj, __atomic_flag__ desr)
{
    unique_lock&lt;mutex&gt; _(some_mutex);
    __atomic_flag__ result = *obj;
    *obj = desr;
    return result;
}

void
__atomic_store_seq_cst(__atomic_flag__ volatile* obj, __atomic_flag__ desr)
{
    unique_lock&lt;mutex&gt; _(some_mutex);
    *obj = desr;
}
</pre></blockquote>

<p>
Where:
</p>

<blockquote><ul>
<li>
If <tt>__has_feature(__atomic_flag)</tt> evaluates to 1 in the preprocessor then
the compiler must define <tt>__atomic_flag__</tt> (e.g. as a typedef to
<tt>int</tt>).
</li>
<li>
If <tt>__has_feature(__atomic_flag)</tt> evaluates to 0 in the preprocessor then
the library defines <tt>__atomic_flag__</tt> as a typedef to <tt>bool</tt>.
</li>
<li>
<p>
To communicate that the above intrinsics are available, the compiler must
arrange for <tt>__has_feature</tt> to return 1 when fed the intrinsic name
appended with an '_' and the mangled type name of <tt>__atomic_flag__</tt>.
</p>
<p>
For example if <tt>__atomic_flag__</tt> is <tt>unsigned int</tt>:
</p>
<blockquote><pre>
__has_feature(__atomic_flag) == 1
__has_feature(__atomic_exchange_seq_cst_j) == 1
__has_feature(__atomic_store_seq_cst_j) == 1

typedef unsigned int __atomic_flag__;

unsigned int __atomic_exchange_seq_cst(unsigned int volatile*, unsigned int)
{
   // ...
}

void __atomic_store_seq_cst(unsigned int volatile*, unsigned int)
{
   // ...
}
</pre></blockquote>
</li>
</ul></blockquote>

<p>
That's it!  Compiler writers do the above and you've got a fully conforming
(though sub-par performance) <tt>&lt;atomic&gt;</tt> header!
</p>

<h2>Recommended work for a higher performance <tt>&lt;atomic&gt;</tt></h2>

<p>
It would be good if the above intrinsics worked with all integral types plus
<tt>void*</tt>.  Because this may not be possible to do in a lock-free manner for
all integral types on all platforms, a compiler must communicate each type that
an intrinsic works with.  For example if <tt>__atomic_exchange_seq_cst</tt> works
for all types except for <tt>long long</tt> and <tt>unsigned long long</tt>
then:
</p>

<blockquote><pre>
__has_feature(__atomic_exchange_seq_cst_b) == 1  // bool
__has_feature(__atomic_exchange_seq_cst_c) == 1  // char
__has_feature(__atomic_exchange_seq_cst_a) == 1  // signed char
__has_feature(__atomic_exchange_seq_cst_h) == 1  // unsigned char
__has_feature(__atomic_exchange_seq_cst_Ds) == 1 // char16_t
__has_feature(__atomic_exchange_seq_cst_Di) == 1 // char32_t
__has_feature(__atomic_exchange_seq_cst_w) == 1  // wchar_t
__has_feature(__atomic_exchange_seq_cst_s) == 1  // short
__has_feature(__atomic_exchange_seq_cst_t) == 1  // unsigned short
__has_feature(__atomic_exchange_seq_cst_i) == 1  // int
__has_feature(__atomic_exchange_seq_cst_j) == 1  // unsigned int
__has_feature(__atomic_exchange_seq_cst_l) == 1  // long
__has_feature(__atomic_exchange_seq_cst_m) == 1  // unsigned long
__has_feature(__atomic_exchange_seq_cst_Pv) == 1 // void*
</pre></blockquote>

<p>
Note that only the <tt>__has_feature</tt> flag is decorated with the argument
type.  The name of the compiler intrinsic is not decorated, but instead works
like a C++ overloaded function.
</p>

<p>
Additionally there are other intrinsics besides
<tt>__atomic_exchange_seq_cst</tt> and <tt>__atomic_store_seq_cst</tt>.  They
are optional.  But if the compiler can generate faster code than provided by the
library, then clients will benefit from the compiler writer's expertise and
knowledge of the targeted platform.
</p>

<p>
Below is the complete list of <i>sequentially consistent</i> intrinsics, and
their library implementations.  Template syntax is used to indicate the desired
overloading for integral and void* types.  The template does not represent a
requirement that the intrinsic operate on <em>any</em> type!
</p>

<blockquote><pre>
T is one of:  bool, char, signed char, unsigned char, short, unsigned short,
              int, unsigned int, long, unsigned long,
              long long, unsigned long long, char16_t, char32_t, wchar_t, void*

template &lt;class T&gt;
T
__atomic_load_seq_cst(T const volatile* obj)
{
    unique_lock&lt;mutex&gt; _(some_mutex);
    return *obj;
}

template &lt;class T&gt;
void
__atomic_store_seq_cst(T volatile* obj, T desr)
{
    unique_lock&lt;mutex&gt; _(some_mutex);
    *obj = desr;
}

template &lt;class T&gt;
T
__atomic_exchange_seq_cst(T volatile* obj, T desr)
{
    unique_lock&lt;mutex&gt; _(some_mutex);
    T r = *obj;
    *obj = desr;
    return r;
}

template &lt;class T&gt;
bool
__atomic_compare_exchange_strong_seq_cst_seq_cst(T volatile* obj, T* exp, T desr)
{
    unique_lock&lt;mutex&gt; _(some_mutex);
    if (std::memcmp(const_cast&lt;T*&gt;(obj), exp, sizeof(T)) == 0)
    {
        std::memcpy(const_cast&lt;T*&gt;(obj), &amp;desr, sizeof(T));
        return true;
    }
    std::memcpy(exp, const_cast&lt;T*&gt;(obj), sizeof(T));
    return false;
}

template &lt;class T&gt;
bool
__atomic_compare_exchange_weak_seq_cst_seq_cst(T volatile* obj, T* exp, T desr)
{
    unique_lock&lt;mutex&gt; _(some_mutex);
    if (std::memcmp(const_cast&lt;T*&gt;(obj), exp, sizeof(T)) == 0)
    {
        std::memcpy(const_cast&lt;T*&gt;(obj), &amp;desr, sizeof(T));
        return true;
    }
    std::memcpy(exp, const_cast&lt;T*&gt;(obj), sizeof(T));
    return false;
}

T is one of:  char, signed char, unsigned char, short, unsigned short,
              int, unsigned int, long, unsigned long,
              long long, unsigned long long, char16_t, char32_t, wchar_t

template &lt;class T&gt;
T
__atomic_fetch_add_seq_cst(T volatile* obj, T operand)
{
    unique_lock&lt;mutex&gt; _(some_mutex);
    T r = *obj;
    *obj += operand;
    return r;
}

template &lt;class T&gt;
T
__atomic_fetch_sub_seq_cst(T volatile* obj, T operand)
{
    unique_lock&lt;mutex&gt; _(some_mutex);
    T r = *obj;
    *obj -= operand;
    return r;
}

template &lt;class T&gt;
T
__atomic_fetch_and_seq_cst(T volatile* obj, T operand)
{
    unique_lock&lt;mutex&gt; _(some_mutex);
    T r = *obj;
    *obj &amp;= operand;
    return r;
}

template &lt;class T&gt;
T
__atomic_fetch_or_seq_cst(T volatile* obj, T operand)
{
    unique_lock&lt;mutex&gt; _(some_mutex);
    T r = *obj;
    *obj |= operand;
    return r;
}

template &lt;class T&gt;
T
__atomic_fetch_xor_seq_cst(T volatile* obj, T operand)
{
    unique_lock&lt;mutex&gt; _(some_mutex);
    T r = *obj;
    *obj ^= operand;
    return r;
}

void*
__atomic_fetch_add_seq_cst(void* volatile* obj, ptrdiff_t operand)
{
    unique_lock&lt;mutex&gt; _(some_mutex);
    void* r = *obj;
    (char*&amp;)(*obj) += operand;
    return r;
}

void*
__atomic_fetch_sub_seq_cst(void* volatile* obj, ptrdiff_t operand)
{
    unique_lock&lt;mutex&gt; _(some_mutex);
    void* r = *obj;
    (char*&amp;)(*obj) -= operand;
    return r;
}

void __atomic_thread_fence_seq_cst()
{
    unique_lock&lt;mutex&gt; _(some_mutex);
}

void __atomic_signal_fence_seq_cst()
{
    unique_lock&lt;mutex&gt; _(some_mutex);
}
</pre></blockquote>

<p>
One should consult the (currently draft)
<a href="https://wg21.link/n3126">C++ standard</a>
for the details of the definitions for these operations.  For example
<tt>__atomic_compare_exchange_weak_seq_cst_seq_cst</tt> is allowed to fail
spuriously while <tt>__atomic_compare_exchange_strong_seq_cst_seq_cst</tt> is
not.
</p>

<p>
If on your platform the lock-free definition of
<tt>__atomic_compare_exchange_weak_seq_cst_seq_cst</tt> would be the same as
<tt>__atomic_compare_exchange_strong_seq_cst_seq_cst</tt>, you may omit the
<tt>__atomic_compare_exchange_weak_seq_cst_seq_cst</tt> intrinsic without a
performance cost.  The library will prefer your implementation of
<tt>__atomic_compare_exchange_strong_seq_cst_seq_cst</tt> over its own
definition for implementing
<tt>__atomic_compare_exchange_weak_seq_cst_seq_cst</tt>.  That is, the library
will arrange for <tt>__atomic_compare_exchange_weak_seq_cst_seq_cst</tt> to call
<tt>__atomic_compare_exchange_strong_seq_cst_seq_cst</tt> if you supply an
intrinsic for the strong version but not the weak.
</p>

<h2>Taking advantage of weaker memory synchronization</h2>

<p>
So far all of the intrinsics presented require a <em>sequentially
consistent</em> memory ordering.  That is, no loads or stores can move across
the operation (just as if the library had locked that internal mutex).  But
<tt>&lt;atomic&gt;</tt> supports weaker memory ordering operations.  In all,
there are six memory orderings (listed here from strongest to weakest):
</p>

<blockquote><pre>
memory_order_seq_cst
memory_order_acq_rel
memory_order_release
memory_order_acquire
memory_order_consume
memory_order_relaxed
</pre></blockquote>

<p>
(See the
<a href="https://wg21.link/n3126">C++ standard</a>
for the detailed definitions of each of these orderings).
</p>

<p>
On some platforms, the compiler vendor can offer some or even all of the above
intrinsics at one or more weaker levels of memory synchronization.  This might
lead for example to not issuing an <tt>mfence</tt> instruction on the x86.
</p>

<p>
If the compiler does not offer any given operation, at any given memory ordering
level, the library will automatically attempt to call the next highest memory
ordering operation.  This continues up to <tt>seq_cst</tt>, and if that doesn't
exist, then the library takes over and does the job with a <tt>mutex</tt>.  This
is a compile-time search &amp; selection operation.  At run time, the
application will only see the few inlined assembly instructions for the selected
intrinsic.
</p>

<p>
Each intrinsic is appended with the 7-letter name of the memory ordering it
addresses.  For example a <tt>load</tt> with <tt>relaxed</tt> ordering is
defined by:
</p>

<blockquote><pre>
T __atomic_load_relaxed(const volatile T* obj);
</pre></blockquote>

<p>
And announced with:
</p>

<blockquote><pre>
__has_feature(__atomic_load_relaxed_b) == 1  // bool
__has_feature(__atomic_load_relaxed_c) == 1  // char
__has_feature(__atomic_load_relaxed_a) == 1  // signed char
...
</pre></blockquote>

<p>
The <tt>__atomic_compare_exchange_strong(weak)</tt> intrinsics are parameterized
on two memory orderings.  The first ordering applies when the operation returns
<tt>true</tt> and the second ordering applies when the operation returns
<tt>false</tt>.
</p>

<p>
Not every memory ordering is appropriate for every operation.  <tt>exchange</tt>
and the <tt>fetch_<i>op</i></tt> operations support all 6.  But <tt>load</tt>
only supports <tt>relaxed</tt>, <tt>consume</tt>, <tt>acquire</tt> and <tt>seq_cst</tt>.
<tt>store</tt>
only supports <tt>relaxed</tt>, <tt>release</tt>, and <tt>seq_cst</tt>.  The
<tt>compare_exchange</tt> operations support the following 16 combinations out
of the possible 36:
</p>

<blockquote><pre>
relaxed_relaxed
consume_relaxed
consume_consume
acquire_relaxed
acquire_consume
acquire_acquire
release_relaxed
release_consume
release_acquire
acq_rel_relaxed
acq_rel_consume
acq_rel_acquire
seq_cst_relaxed
seq_cst_consume
seq_cst_acquire
seq_cst_seq_cst
</pre></blockquote>

<p>
Again, the compiler supplies intrinsics only for the strongest orderings where
it can make a difference.  The library takes care of calling the weakest
supplied intrinsic that is as strong or stronger than the customer asked for.
</p>

</div>
</body>
</html>