gcc/libcilkrts/include/cilk/reducer.h
2013-10-29 11:37:47 -07:00

1901 lines
73 KiB
C++
Raw Blame History

This file contains invisible Unicode characters

This file contains invisible Unicode characters that are indistinguishable to humans but may be processed differently by a computer. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

/* reducer.h -*- C++ -*-
*
* @copyright
* Copyright (C) 2009-2013, Intel Corporation
* All rights reserved.
*
* @copyright
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
*
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in
* the documentation and/or other materials provided with the
* distribution.
* * Neither the name of Intel Corporation nor the names of its
* contributors may be used to endorse or promote products derived
* from this software without specific prior written permission.
*
* @copyright
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
* "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
* LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
* A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
* HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
* BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
* OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
* AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY
* WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
* POSSIBILITY OF SUCH DAMAGE.
*/
/** @file reducer.h
*
* @brief Defines foundation classes for creating Cilk reducers.
*
* @ingroup Reducers
*
* @see @ref pagereducers
*
* @defgroup Reducers Reducers
*/
#ifndef REDUCER_H_INCLUDED
#define REDUCER_H_INCLUDED
#include "cilk/hyperobject_base.h"
#include "cilk/metaprogramming.h"
#ifdef __cplusplus
//===================== C++ interfaces ===================================
#include <new>
namespace cilk {
/** Base class for defining monoids.
*
* The monoid_base class template is useful for creating classes that model
* the monoid concept. It provides the core type and memory management
* functionality. A subclass of monoid_base need only declare and implement
* the `identity` and `reduce` functions.
*
* The monoid_base class also manages the integration between the monoid, the
* reducer class that is based on it, and an optional view class which wraps
* value objects and restricts access to their operations.
*
* @tparam Value The value type for the monoid.
* @tparam View An optional view class that serves as a proxy for the value
* type.
*
* @see monoid_with_view
*/
template <typename Value, typename View = Value>
class monoid_base
{
protected:
/** Class for provisionally constructed objects.
*
* The monoid_base::construct() functions manually construct both a monoid
* and a view. If one of these is constructed successfully, and the
* construction of the other (or some other initialization) fails, then
* the first one must be destroyed to avoid a memory leak. Because the
* construction is explicit, the destruction must be explicit, too.
*
* A provisional_guard object wraps a pointer to a newly constructed
* object. A call to its confirm() function confirms that the object is
* really going to be used. If the guard is destroyed without being
* confirmed, then the pointed-to object is destroyed (but not
* deallocated).
*
* Expected usage:
*
* provisional_guard<T1> x1_provisional( new (x1) T1() );
* … more initialization …
* x1_provisional.confirm();
*
* or
*
* provisional_guard<T1> x1_provisional( new (x1) T1() );
* x1_provisional.confirm_if( new (x2) T2() );
*
* If an exception is thrown in the “more initialization” code in the
* first example, or in the `T2` constructor in the second example, then
* `x1_provisional` will not be confirmed, so when its destructor is
* called during exception unwinding, the `T1` object that was constructed
* in `x1` will be destroyed.
*
* @see provisional()
*
* @tparam Type The type of the provisionally constructed object.
*/
template <typename Type>
class provisional_guard {
Type* m_ptr;
public:
/** Constructor. Creates a guard for a provisionally constructed object.
*
* @param ptr A pointer to the provisionally constructed object.
*/
provisional_guard(Type* ptr) : m_ptr(ptr) {}
/** Destructor. Destroy the object pointed to by the contained pointer
* if it has not been confirmed.
*/
~provisional_guard() { if (m_ptr) m_ptr->~Type(); }
/** Confirm the provisional construction. Do *not* delete the contained
* pointer when the guard is destroyed.
*/
void confirm() { m_ptr = 0; }
/** Confirm provisional construction if argument is non-null. Note that
* if an exception is thrown during evaluation of the argument
* expression, then this function will not be called, and the
* provisional object will not be confirmed. This allows the usage:
*
* x1_provisional.confirm_if( new (x2) T2() );
*
* @param cond An arbitrary pointer. The provisional object will be
* confirmed if @a cond is not null.
*
* @returns The value of the @a cond argument.
*/
template <typename Cond>
Cond* confirm_if(Cond* cond) { if (cond) m_ptr = 0; return cond; }
};
/** Create a provisional_guard object. This function allows simpler code
* when the only use of a provisional_guard is in a
* provisional_guard::confirm_if() call immediately following its
* creation. Instead of
*
* provisional_guard<T>guard( new (ptr_to_T) T() );
* guard.confirm_if( new (ptr_to_U) U() );
*
* you can just write
*
* provisional( new (ptr_to_T) T() ).confirm_if( new (ptr_to_U) U() );
*
* @tparam Type The type of the provisionally constructed object.
*
* @param ptr A pointer to a provisionally constructed object.
*
* @returns A @ref provisional_guard object that guards the
* provisionally constructed object pointed to by @a ptr.
*/
template <typename Type>
static provisional_guard<Type> provisional(Type* ptr)
{ return provisional_guard<Type>(ptr); }
public:
/** Value type of the monoid.
*/
typedef Value value_type;
/** View type of the monoid. Defaults to be the same as the value type.
* @see monoid_with_view
*/
typedef View view_type;
enum {
/** Should reducers created with this monoid be aligned?
*
* @details
* “Aligned” means that the view is allocated at a cache-line aligned
* offset in the reducer, and the reducer must be cache-line aligned.
* “Unaligned” means that the reducer as a whole is just naturally
* aligned, but it contains a large enough block of uninitialized
* storage for a cache-line aligned view to be allocated in it at
* reducer construction time.
*
* Since the standard heap allocator (new reducer) does not allocate
* cache-line aligned storage, only unaligned reducers can be safely
* allocated on the heap.
*
* Default is false (unaligned) unless overridden in a subclass.
*
* @since 1.02
* (In Cilk library versions 1.0 and 1.01, the default was true.
* In Cilk library versions prior to 1.0, reducers were always aligned,
* and this data member did not exist.)
*/
align_reducer = false
};
/** Destroy a view. Destroys (without deallocating) the @a View object
* pointed to by @a p.
*
* @param p The address of the @a View object to be destroyed.
*/
void destroy(view_type* p) const { p->~view_type(); }
/** Allocate raw memory. Allocate @a s bytes of memory with no
* initialization.
*
* @param s The number of bytes of memory to allocate.
* @return An untyped pointer to the allocated memory.
*/
void* allocate(size_t s) const { return operator new(s); }
/** Deallocate raw memory. Deallocates the memory pointed to by @a p
* without doing any destruction.
*
* @param p Pointer to the memory to be deallocated.
*
* @pre @a p points to a block of memory that was allocated by a
* call to allocate().
*/
void deallocate(void* p) const { operator delete(p); }
/** Create the identity value. Constructs (without allocating) a @a View
* object representing the default value of the @a Value type.
*
* @param p A pointer to a block of raw memory large enough to hold a
* @a View object.
*
* @post The memory pointed to by @a p contains a @a View object that
* represents the default value of the @a View type.
*
* @deprecated This function constructs the @a View object with its default
* constructor, which will often, but not always, yield the
* appropriate identity value. Monoid classes should declare
* their identity function explicitly, rather than relying on
* this default definition.
*/
void identity(View* p) const { new ((void*) p) View(); }
/** @name Construct the monoid and the view with arbitrary arguments.
*
* A @ref reducer object contains monoid and view data members, which are
* declared as raw storage (byte arrays), so that they are not implicitly
* constructed when the reducer is constructed. Instead, a reducer
* constructor calls one of the monoid classs static construct()
* functions with the addresses of the monoid and the view, and the
* construct() function uses placement `new` to construct them.
*
* This allows the monoid to determine the order in which the monoid and
* view are constructed, and to make one of them dependent on the other.
*
* Any arguments to the reducer constructor are just passed on as
* additional arguments to the construct() function (after the monoid
* and view addresses).
*
* Any monoid whose needs are satisfied by the suite of construct()
* functions below, such as @ref monoid_with_view, can just inherit them
* from monoid_base. Other monoids will need to provide their own versions
* to override the monoid_base functions.
*/
//@{
/** Default-construct the monoid, and pass zero to five const reference
* arguments to the view constructor.
*/
//@{
template <typename Monoid>
static void construct(Monoid* monoid, View* view)
{ provisional( new ((void*)monoid) Monoid() ).confirm_if(
(monoid->identity(view), view) ); }
template <typename Monoid, typename T1>
static void construct(Monoid* monoid, View* view, const T1& x1)
{ provisional( new ((void*)monoid) Monoid() ).confirm_if(
new ((void*)view) View(x1) ); }
template <typename Monoid, typename T1, typename T2>
static void construct(Monoid* monoid, View* view,
const T1& x1, const T2& x2)
{ provisional( new ((void*)monoid) Monoid() ).confirm_if(
new ((void*)view) View(x1, x2) ); }
template <typename Monoid, typename T1, typename T2, typename T3>
static void construct(Monoid* monoid, View* view,
const T1& x1, const T2& x2, const T3& x3)
{ provisional( new ((void*)monoid) Monoid() ).confirm_if(
new ((void*)view) View(x1, x2, x3) ); }
template <typename Monoid, typename T1, typename T2, typename T3,
typename T4>
static void construct(Monoid* monoid, View* view,
const T1& x1, const T2& x2, const T3& x3,
const T4& x4)
{ provisional( new ((void*)monoid) Monoid() ).confirm_if(
new ((void*)view) View(x1, x2, x3, x4) ); }
template <typename Monoid, typename T1, typename T2, typename T3,
typename T4, typename T5>
static void construct(Monoid* monoid, View* view,
const T1& x1, const T2& x2, const T3& x3,
const T4& x4, const T5& x5)
{ provisional( new ((void*)monoid) Monoid() ).confirm_if(
new ((void*)view) View(x1, x2, x3, x4, x5) ); }
//@}
/** Default-construct the monoid, and pass one non-const reference argument
* to the view constructor.
*/
//@{
template <typename Monoid, typename T1>
static void construct(Monoid* monoid, View* view, T1& x1)
{ provisional( new ((void*)monoid) Monoid() ).confirm_if(
new ((void*)view) View(x1) ); }
//@}
/** Copy-construct the monoid, and pass zero to four const reference
* arguments to the view constructor.
*/
//@{
template <typename Monoid>
static void construct(Monoid* monoid, View* view, const Monoid& m)
{ provisional( new ((void*)monoid) Monoid(m) ).confirm_if(
new ((void*)view) View() ); }
template <typename Monoid, typename T1>
static void construct(Monoid* monoid, View* view, const Monoid& m,
const T1& x1)
{ provisional( new ((void*)monoid) Monoid(m) ).confirm_if(
new ((void*)view) View(x1) ); }
template <typename Monoid, typename T1, typename T2>
static void construct(Monoid* monoid, View* view, const Monoid& m,
const T1& x1, const T2& x2)
{ provisional( new ((void*)monoid) Monoid(m) ).confirm_if(
new ((void*)view) View(x1, x2) ); }
template <typename Monoid, typename T1, typename T2, typename T3>
static void construct(Monoid* monoid, View* view, const Monoid& m,
const T1& x1, const T2& x2, const T3& x3)
{
provisional( new ((void*)monoid) Monoid(m) ).confirm_if(
new ((void*)view) View(x1, x2, x3) );
}
template <typename Monoid, typename T1, typename T2, typename T3,
typename T4>
static void construct(Monoid* monoid, View* view, const Monoid& m,
const T1& x1, const T2& x2, const T3& x3,
const T4& x4)
{
provisional( new ((void*)monoid) Monoid(m) ).confirm_if(
new ((void*)view) View(x1, x2, x3, x4) );
}
//@}
//@}
};
/** Monoid class that gets its value type and identity and reduce operations
* from its view.
*
* A simple implementation of the monoid-view-reducer architecture would
* distribute knowledge about the type and operations for the reduction
* between the monoid and the view — the identity and reduction operations are
* specified in the monoid, the reduction operations are implemented in the
* view, and the value type is specified in both the monoid and the view.
* This is inelegant.
*
* monoid_with_view is a subclass of @ref monoid_base that gets its value type
* and its identity and reduction operations from its view class. No
* customization of the monoid_with_view class itself is needed beyond
* instantiating it with an appropriate view class. (Customized subclasses of
* monoid_with_view may be needed for other reasons, such as to keep some
* state for the reducer.) All of the Cilk predefined reducers use
* monoid_with_view or one of its subclasses.
*
* The view class `View` of a monoid_with_view must provide the following public definitions:
*
* Definition | Meaning
* ---------------------------------|--------
* `value_type` | a typedef of the value type for the reduction
* `View()` | a default constructor which constructs the identity value for the reduction
* `void reduce(const View* other)` | a member function which applies the reduction operation to the values of `this` view and the `other` view, leaving the result as the value of `this` view, and leaving the value of the `other` view undefined (but valid)
*
* @tparam View The view class for the monoid.
* @tparam Align If true, reducers instantiated on this monoid will be
* cache-aligned. By default, library reducers (unlike legacy
* library reducer _wrappers_) are aligned only as required by
* contents.
*/
template <class View, bool Align = false>
class monoid_with_view : public monoid_base<typename View::value_type, View>
{
public:
/** Should reducers created with this monoid be aligned?
*/
enum { align_reducer = Align };
/** Create the identity value.
*
* Implements the monoid `identity` operation by using the @a View classs
* default constructor.
*
* @param p A pointer to a block of raw memory large enough to hold a
* @p View object.
*/
void identity(View* p) const { new ((void*)p) View(); }
/** Reduce the values of two views.
*
* Implements the monoid `reduce` operation by calling the left views
* `%reduce()` function with the right view as an operand.
*
* @param left The left operand of the reduce operation.
* @param right The right operand of the reduce operation.
* @post The left view contains the result of the reduce
* operation, and the right view is undefined.
*/
void reduce(View* left, View* right) const { left->reduce(right); }
};
/** Base class for simple views with (usually) scalar values.
*
* The scalar_view class is intended as a base class which provides about half
* of the required definitions for simple views. It defines the `value_type`
* required by a @ref monoid_with_view (but not the identity constructor and
* reduce operation, which are inherently specific to a particular kind of
* reduction). It also defines the value access functions which will be called
* by the corresponding @ref reducer functions. (It uses copy semantics for
* the view_move_in() and view_move_out() functions, which is appropriate
* for simple scalar types, but not necessarily for more complex types like
* STL containers.
*
* @tparam Type The type of value wrapped by the view.
*/
template <typename Type>
class scalar_view
{
protected:
Type m_value; ///< The wrapped accumulator variable.
public:
/** Value type definition required by @ref monoid_with_view.
*/
typedef Type value_type;
/** Default constructor.
*/
scalar_view() : m_value() {}
/** Value constructor.
*/
scalar_view(const Type& v) : m_value(v) {}
/** @name Value functions required by the reducer class.
*
* Note that the move in/out functions use simple assignment semantics.
*/
//@{
/** Set the value of the view.
*/
void view_move_in(Type& v) { m_value = v; }
/** Get the value of the view.
*/
void view_move_out(Type& v) { v = m_value; }
/** Set the value of the view.
*/
void view_set_value(const Type& v) { m_value = v; }
/** Get the value of the view.
*/
Type const& view_get_value() const { return m_value; }
/** Get a reference to the value contained in the view. For legacy
* reducer support only.
*/
Type & view_get_reference() { return m_value; }
/** Get a reference to the value contained in the view. For legacy
* reducer support only.
*/
Type const& view_get_reference() const { return m_value; }
//@}
};
/** Wrapper class for move-in construction.
*
* Some types allow their values to be _moved_ as an alternative to copying.
* Moving a value may be much faster than copying it, but may leave the value
* of the moves source undefined. Consider the `swap` operation provided by
* many STL container classes:
*
* list<T> x, y;
* x = y; // Copy
* x.swap(y); // Move
*
* The assignment _copies_ the value of `y` into `x` in time linear in the
* size of `y`, leaving `y` unchanged. The `swap` _moves_ the value of `y`
* into `x` in constant time, but it also moves the value of `x` into `y`,
* potentially leaving `y` undefined.
*
* A move_in_wrapper simply wraps a pointer to an object. It is created by a
* call to cilk::move_in(). Passing a move_in_wrapper to a view constructor
* (actually, passing it to a reducer constructor, which passes it to the
* monoid `construct()` function, which passes it to the view constructor)
* allows, but does not require, the value pointed to by the wrapper to be
* moved into the view instead of copied.
*
* A view class exercises this option by defining a _move-in constructor_,
* i.e., a constructor with a move_in_wrapper parameter. The constructor calls
* the wrappers `value()` function to get a reference to its pointed-to
* value, and can then use that reference in a move operation.
*
* A move_in_wrapper also has an implicit conversion to its pointed-to value,
* so if a view class does not define a move-in constructor, its ordinary
* value constructor will be called with the wrapped value. For example, an
* @ref ReducersAdd "op_add" view does not have a move-in constructor, so
*
* int x;
* reducer< op_add<int> > xr(move_in(x));
*
* will simply call the `op_add_view(const int &)` constructor. But an
* @ref ReducersList "op_list_append" view does have a move-in constructor,
* so
*
* list<int> x;
* reducer< op_list_append<int> > xr(move_in(x));
*
* will call the `op_list_append_view(move_in_wrapper<int>)` constructor,
* which can `swap` the value of `x` into the view.
*
* @note Remember that passing the value of a variable to a reducer
* constructor using a move_in_wrapper leaves the variable undefined.
* You cannot assume that the constructor either will or will not copy
* or move the value.
*
* @tparam Type The type of the wrapped value.
*
* @see cilk::move_in()
*/
template <typename Type>
class move_in_wrapper
{
Type *m_pointer;
public:
/** Constructor that captures the address of its argument. This is almost
* always called from the @ref move_in function.
*/
explicit move_in_wrapper(Type& ref) : m_pointer(&ref) { }
/** Implicit conversion to the wrapped value. This allows a move_in_wrapper
* to be used where a value of the wrapped type is expected, in which case
* the wrapper is completely transparent.
*/
operator Type&() const { return *m_pointer; }
/** Get a reference to the pointed-to value. This has the same effect as
* the implicit conversion, but makes the intent clearer in a move-in
* constructor.
*/
Type& value() const { return *m_pointer; }
};
/** Function to create a move_in_wrapper for a value.
*
* @tparam Type The type of the argument, which will be the `type` of the
* created wrapper.
*
* @see move_in_wrapper
*/
template <typename Type>
inline
move_in_wrapper<Type> move_in(Type& ref)
{ return move_in_wrapper<Type>(ref); }
/** @copydoc move_in(Type&)
*
* @note Applying a function that is explicitly specified as modifying its
* argument to a const argument is obviously an irrational thing to
* do. This move_in() variant is just provided to allow calling a
* move-in constructor with a function return value, which the
* language treats as a const. Using it for any other purpose will
* probably end in tears.
*/
template <typename Type>
inline
move_in_wrapper<Type> move_in(const Type& ref)
{ return move_in_wrapper<Type>(ref); }
/** Wrapper class to allow implicit downcasts to reducer subclasses.
*
* The Cilk library contains a collection of reducer wrapper classes which
* were created before the `cilk::reducer<Monoid>` style was developed. For
* example, `cilk::reducer_opadd<Type>` provided essentially the same
* functionality that is now provided by
* `cilk::reducer< cilk::op_add<Type> >`. These legacy reducer classes are
* deprecated, but still supported, and they have been reimplemented as
* subclasses of the corresponding `cilk::reducer` classes. For example:
*
* template <class T>
* reducer_opadd<T> : public reducer< op_add<T> > { ... };
*
* This reimplementation allows transparent conversion between legacy and
* new reducers. That is, a `reducer<op_add>*` or `reducer<op_add>&` can be
* used anywhere that a `reducer_opadd*` or `reducer_opadd&` is expected,
* and vice versa.
*
* The conversion from the legacy reducer to the new reducer is just an
* up-cast, which is provided for free by C++. The conversion from the new
* reducer to the legacy reducer is a down-cast, though, which requires an
* explicit conversion member function in the `reducer` class. The challenge
* is to define a function in the reducer template class which will convert
* each cilk::reducer specialization to the corresponding legacy reducer,
* if there is one.
*
* The trick is in the legacy_reducer_downcast template class, which provides
* a mapping from `cilk::reducer` specializations to legacy reducer classes.
* `reducer<Monoid>` has a conversion function to convert itself to
* `legacy_reducer_downcast< reducer<Monoid> >::%type`. By default,
* `legacy_reducer_downcast<Reducer>::%type` is just a trivial subclass of
* `Reducer`, which is uninteresting, but a reducer with a legacy counterpart
* will have a specialization of `legacy_reducer_downcast` whose `type` is
* the corresponding legacy reducer. For example:
*
* template <typename Type>
* struct legacy_reducer_downcast< reducer< op_add<Type> > >
* {
* typedef reducer_opadd<Type> type;
* };
*
*
* @tparam Reducer The new-style reducer class whose corresponding legacy reducer class
* is `type`, if there is such a legacy reducer class.
*/
template <typename Reducer>
struct legacy_reducer_downcast
{
/** The related legacy reducer class.
*
* By default, this is just a trivial subclass of Reducer, but it can be
* overridden in the specialization of legacy_reducer_downcast for
* a reducer that has a corresponding legacy reducers.
*/
struct type : Reducer { };
};
namespace internal {
/// @cond internal
template <typename Value, typename View>
struct reducer_set_get
{
static View theView; // Declared but not defined
// sizeof(notchar) is guaranteed larger than 1
struct notchar { char x[2]; };
// check_for_ref returns char if 'get_value' returns by value and notchar
// if 'get_value' returns by reference.
static char check_for_ref(Value, ...);
static notchar check_for_ref(Value&, int);
enum { GET_VALUE_BY_VALUE =
(1 == sizeof(check_for_ref(theView.view_get_value(), 0))) } ;
typedef typename condition<GET_VALUE_BY_VALUE,
Value, const Value&>::type get_value_type;
static void move_in(View& view, Value& v) { view.view_move_in(v); }
static void move_out(View& view, Value& v) { view.view_move_out(v); }
static void set_value(View& view, const Value& v)
{ view.view_set_value(v); }
static get_value_type get_value(const View& view)
{ return view.view_get_value(); }
};
template <typename Value>
struct reducer_set_get<Value, Value>
{
typedef const Value& get_value_type;
static void move_in(Value& view, Value& v) { view = v; }
static void move_out(Value& view, Value& v) { v = view; }
static void set_value(Value& view, const Value& v) { view = v; }
static get_value_type get_value(const Value& view) { return view; }
};
/// @endcond
/** Base class defining the data layout that is common to all reducers.
*/
template <typename Monoid>
class reducer_base {
typedef typename Monoid::view_type view_type;
// This makes the reducer a hyper-object. (Partially initialized in
// the derived reducer_content class.)
//
__cilkrts_hyperobject_base m_base;
// The monoid is allocated here as raw bytes, and is constructed explicitly
// by a call to the monoid_type::construct() function in the constructor of
// the `reducer` subclass.
//
storage_for_object<Monoid> m_monoid;
// Used for sanity checking at destruction.
//
void* m_initialThis;
// The leftmost view comes next. It is defined in the derived
// reducer_content class.
/** @name C-callable wrappers for the C++-coded monoid dispatch functions.
*/
//@{
static void reduce_wrapper(void* r, void* lhs, void* rhs);
static void identity_wrapper(void* r, void* view);
static void destroy_wrapper(void* r, void* view);
static void* allocate_wrapper(void* r, __STDNS size_t bytes);
static void deallocate_wrapper(void* r, void* view);
//@}
protected:
/** Constructor.
*
* @param leftmost The address of the leftmost view in the reducer.
*/
reducer_base(char* leftmost)
{
static const cilk_c_monoid c_monoid_initializer = {
(cilk_c_reducer_reduce_fn_t) &reduce_wrapper,
(cilk_c_reducer_identity_fn_t) &identity_wrapper,
(cilk_c_reducer_destroy_fn_t) &destroy_wrapper,
(cilk_c_reducer_allocate_fn_t) &allocate_wrapper,
(cilk_c_reducer_deallocate_fn_t) &deallocate_wrapper
};
m_base.__c_monoid = c_monoid_initializer;
m_base.__flags = 0;
m_base.__view_offset = (char*)leftmost - (char*)this;
m_base.__view_size = sizeof(view_type);
m_initialThis = this;
__cilkrts_hyper_create(&m_base);
}
/** Destructor.
*/
__CILKRTS_STRAND_STALE(~reducer_base())
{
// Make sure we haven't been memcopy'd or corrupted
__CILKRTS_ASSERT(
this == m_initialThis ||
// Allow for a layout bug that may put the initialThis field one
// word later in 1.0 reducers than in 0.9 and 1.1 reducers.
this == *(&m_initialThis + 1)
);
__cilkrts_hyper_destroy(&m_base);
}
/** Monoid data member.
*
* @return A pointer to the reducers monoid data member.
*/
Monoid* monoid_ptr() { return &m_monoid.object(); }
/** Leftmost view data member.
*
* @return A pointer to the reducers leftmost view data member.
*
* @note This function returns the address of the *leftmost* view,
* which is unique for the lifetime of the reducer. It is
* intended to be used in constructors and destructors.
* Use the reducer::view() function to access the per-strand
* view instance.
*/
view_type* leftmost_ptr()
{
char* view_addr = (char*)this + m_base.__view_offset;
return reinterpret_cast<view_type*>(view_addr);
}
public:
/** @name Access the current view.
*
* These functions return a reference to the instance of the reducers
* view that was created for the current strand of a parallel computation
* (and create it if it doesnt already exist). Note the difference from
* the (private) leftmost_ptr() function, which returns a pointer to the
* _leftmost_ view, which is the same in all strands.
*/
//@{
/** Per-strand view instance.
*
* @return A reference to the per-strand view instance.
*/
view_type& view()
{
return *static_cast<view_type *>(__cilkrts_hyper_lookup(&m_base));
}
/** @copydoc view()
*/
const view_type& view() const
{
return const_cast<reducer_base*>(this)->view();
}
//@}
/** Initial view pointer field.
*
* @internal
*
* @return a reference to the m_initialThis field.
*
* @note This function is provided for “white-box” testing of the
* reducer layout code. There is never any reason for user code
* to call it.
*/
const void* const & initial_this() const { return m_initialThis; }
};
template <typename Monoid>
void reducer_base<Monoid>::reduce_wrapper(void* r, void* lhs, void* rhs)
{
Monoid* monoid = static_cast<reducer_base*>(r)->monoid_ptr();
monoid->reduce(static_cast<view_type*>(lhs),
static_cast<view_type*>(rhs));
}
template <typename Monoid>
void reducer_base<Monoid>::identity_wrapper(void* r, void* view)
{
Monoid* monoid = static_cast<reducer_base*>(r)->monoid_ptr();
monoid->identity(static_cast<view_type*>(view));
}
template <typename Monoid>
void reducer_base<Monoid>::destroy_wrapper(void* r, void* view)
{
Monoid* monoid = static_cast<reducer_base*>(r)->monoid_ptr();
monoid->destroy(static_cast<view_type*>(view));
}
template <typename Monoid>
void* reducer_base<Monoid>::allocate_wrapper(void* r, __STDNS size_t bytes)
{
Monoid* monoid = static_cast<reducer_base*>(r)->monoid_ptr();
return monoid->allocate(bytes);
}
template <typename Monoid>
void reducer_base<Monoid>::deallocate_wrapper(void* r, void* view)
{
Monoid* monoid = static_cast<reducer_base*>(r)->monoid_ptr();
monoid->deallocate(static_cast<view_type*>(view));
}
/** Base class defining the data members of a reducer.
*
* @tparam Aligned The `m_view` data member, and therefore the entire
* structure, are cache-line aligned if this parameter
* is `true'.
*/
template <typename Monoid, bool Aligned = Monoid::align_reducer>
class reducer_content;
/** Base class defining the data members of an aligned reducer.
*/
template <typename Monoid>
class reducer_content<Monoid, true> : public reducer_base<Monoid>
{
typedef typename Monoid::view_type view_type;
// The leftmost view is defined as raw bytes. It will be constructed
// by the monoid `construct` function. It is cache-aligned, which
// will push it into a new cache line. Furthermore, its alignment causes
// the reducer as a whole to be cache-aligned, which makes the reducer
// size a multiple of a cache line. Since there is nothing in the reducer
// after the view, all this means that the leftmost view gets one or more
// cache lines all to itself, which prevents false sharing.
//
__CILKRTS_CACHE_ALIGN
char m_leftmost[sizeof(view_type)];
/** Test if the reducer is cache-line-aligned.
*
* Used in assertions.
*/
bool reducer_is_cache_aligned() const
{ return 0 == ((std::size_t) this & (__CILKRTS_CACHE_LINE__ - 1)); }
protected:
/** Constructor.
*/
reducer_content() : reducer_base<Monoid>((char*)&m_leftmost)
{
#ifndef CILK_IGNORE_REDUCER_ALIGNMENT
assert(reducer_is_cache_aligned() &&
"Reducer should be cache aligned. Please see comments following this assertion for explanation and fixes.");
#endif
/* "REDUCER SHOULD BE CACHE ALIGNED" ASSERTION.
*
* This Reducer class instantiation specifies cache-line alignment of the
* leftmost view field (and, implicitly, of the reducer itself). You got
* this assertion because a reducer with this class was allocated at a
* non-cache-aligned address, probably because it was allocated on the
* heap with `new`. This can be a problem for two reasons:
*
* 1. If the leftmost view is not on a cache line by itself, there might
* be a slowdown resulting from accesses to the same cache line from
* different threads.
*
* 2. The compiler thinks that reducer is cache-line aligned, but it
* really isn't. If the reducer is contained in a structure, then the
* compiler will believe that the containing structure, and other
* fields contained in it, are also more aligned than they really
* are. In particular, if the structure contains a numeric array that
* is used in a vectorizable loop, then the compiler might generate
* invalid vector instructions, resulting in a runtime error.
*
* The compiler will always allocate reducer variables, and structure
* variables containing reducers, with their required alignment.
* Reducers, and structures containing a reducer, which are allocated
* on the heap with `new` will _not_ be properly aligned.
*
* There are three ways that you can fix this assertion failure.
*
* A. Rewrite your code to use the new-style `reducer< op_XXX<Type> >`
* instead of the legacy `reducer_XXX<type>`. The new-style reducers
* are not declared to be cache-aligned, and will work properly if
* they are not cache-aligned.
*
* B. If you must allocate an old-style reducer or a structure containing
* a reducer on the heap, figure out how to align it correctly. The
* suggested fix is to use `cilk::aligned_new()` and
* `cilk::aligned_delete()` instead of `new` and `delete`, as follows:
*
* Type* ptr = cilk::aligned_new<Type>(constructor-arguments);
* cilk::aligned_delete(ptr);
*
* C. Define the macro CILK_IGNORE_REDUCER_ALIGNMENT, which will suppress
* the assertion check. Do this only if you are comfortable that
* problem (2) above will not occur.
*/
}
};
/** Base class defining the data members of an unaligned reducer.
*/
template <typename Monoid>
class reducer_content<Monoid, false> : public reducer_base<Monoid>
{
typedef typename Monoid::view_type view_type; ///< The view type.
// Reserve space for the leftmost view. The view will be allocated at an
// aligned offset in this space at runtime, to guarantee that the view
// will get one or more cache lines all to itself, to prevent false
// sharing.
//
// The number of bytes to reserve is determined as follows:
// * Start with the view size.
// * Round up to a multiple of the cache line size, to get the total size
// of the cache lines that will be dedicated to the view.
// * Add (cache line size - 1) filler bytes to guarantee that the reserved
// area will contain a cache-aligned block of the required cache lines,
// no matter where the reserved area starts.
//
char m_leftmost[
// View size rounded up to multiple cache lines
( (sizeof(view_type) + __CILKRTS_CACHE_LINE__ - 1)
& ~ (__CILKRTS_CACHE_LINE__ - 1)
)
// plus filler to allow alignment.
+ __CILKRTS_CACHE_LINE__ - 1
];
protected:
/** Constructor. Find the first cache-aligned position in the reserved
* area, and pass it to the base constructor as the leftmost view
* address.
*/
reducer_content() :
reducer_base<Monoid>(
(char*)( ((std::size_t)&m_leftmost + __CILKRTS_CACHE_LINE__ - 1)
& ~ (__CILKRTS_CACHE_LINE__ - 1) ) )
{}
};
} // namespace internal
// The __cilkrts_hyperobject_ functions are defined differently depending on
// whether a file is compiled with or without the CILK_STUB option. Therefore,
// reducers compiled in the two modes should be link-time incompatible, so that
// object files compiled with stubbed reducers won't be linked into an
// unstubbed program, or vice versa. We achieve this by putting the reducer
// class definition into the cilk::stub namespace in a stubbed compilation.
#ifdef CILK_STUB
namespace stub {
#endif
/** Reducer class.
*
* A reducer is instantiated on a Monoid. The Monoid provides the value
* type, associative reduce function, and identity for the reducer.
*
* @tparam Monoid The monoid class that the reducer is instantiated on. It must model
* the @ref reducers_monoid_concept "monoid concept".
*
* @see @ref pagereducers
*/
template <class Monoid>
class reducer : public internal::reducer_content<Monoid>
{
typedef internal::reducer_content<Monoid> base;
using base::monoid_ptr;
using base::leftmost_ptr;
public:
typedef Monoid monoid_type; ///< The monoid type.
typedef typename Monoid::value_type value_type; ///< The value type.
typedef typename Monoid::view_type view_type; ///< The view type.
private:
typedef internal::reducer_set_get<value_type, view_type> set_get;
reducer(const reducer&); ///< Disallow copying.
reducer& operator=(const reducer&); ///< Disallow assignment.
public:
/** @name Constructors
*
* All reducer constructors call the static `construct()` function of the monoid class to
* construct the reducer's monoid and leftmost view.
*
* The reducer constructor arguments are simply passed through to the construct() function.
* Thus, the constructor parameters accepted by a particular reducer class are determined
* by its monoid class.
*/
//@{
/** 0 6 const reference parameters.
*/
//@{
reducer()
{
monoid_type::construct(monoid_ptr(), leftmost_ptr());
}
template <typename T1>
reducer(const T1& x1)
{
monoid_type::construct(monoid_ptr(), leftmost_ptr(), x1);
}
template <typename T1, typename T2>
reducer(const T1& x1, const T2& x2)
{
monoid_type::construct(monoid_ptr(), leftmost_ptr(), x1, x2);
}
template <typename T1, typename T2, typename T3>
reducer(const T1& x1, const T2& x2, const T3& x3)
{
monoid_type::construct(monoid_ptr(), leftmost_ptr(), x1, x2, x3);
}
template <typename T1, typename T2, typename T3, typename T4>
reducer(const T1& x1, const T2& x2, const T3& x3, const T4& x4)
{
monoid_type::construct(monoid_ptr(), leftmost_ptr(), x1, x2, x3, x4);
}
template <typename T1, typename T2, typename T3, typename T4, typename T5>
reducer(const T1& x1, const T2& x2, const T3& x3, const T4& x4, const T5& x5)
{
monoid_type::construct(monoid_ptr(), leftmost_ptr(), x1, x2, x3, x4, x5);
}
template <typename T1, typename T2, typename T3, typename T4, typename T5, typename T6>
reducer(const T1& x1, const T2& x2, const T3& x3, const T4& x4, const T5& x5, const T6& x6)
{
monoid_type::construct(monoid_ptr(), leftmost_ptr(), x1, x2, x3, x4, x5, x6);
}
//@}
/** 1 non-const reference parameter.
*/
//@{
template <typename T1>
reducer(T1& x1)
{
monoid_type::construct(monoid_ptr(), leftmost_ptr(), x1);
}
//@}
/** Destructor.
*/
__CILKRTS_STRAND_STALE(~reducer())
{
leftmost_ptr()->~view_type();
monoid_ptr()->~monoid_type();
}
//@{
/** Get the monoid.
*
* @return A reference to the monoid object belonging to this reducer.
*/
Monoid& monoid() { return *monoid_ptr(); }
const Monoid& monoid() const
{ return const_cast<reducer*>(this)->monoid(); }
//@}
//@{
/** Access the current view.
*
* Return a reference to the instance of the reducers view that was
* created for the current strand of a parallel computation (and create
* it if it doesnt already exist).
*/
view_type& view() { return base::view(); }
const view_type& view() const { return base::view(); }
//@}
/** @name Dereference the reducer to get the view.
*
* “Dereferencing” a reducer yields the view for the current strand. The
* view, in turn, acts as a proxy for its contained value, exposing only
* those operations which are consistent with the reducers monoid. Thus,
* all modifications of the reducers accumulator variable are written as
*
* *reducer OP ...
*
* or
*
* reducer->func(...)
*
* (The permitted operations on a reducers accumulator are listed in the
* documentation for that particular kind of reducer.)
*
* @note `*r` is a synonym for `r.view()`. Recommended style is to use
* `*r` (or `r->`) in the common case where code is simply
* updating the accumulator variable wrapped in the view, and to
* use `r.view()` in the unusual case where it is desirable to
* call attention to the view itself.
*/
//@{
//@{
/** Dereference operator.
*
* @return A reference to the per-strand view instance.
*/
view_type& operator*() { return view(); }
view_type const& operator*() const { return view(); }
//@}
//@{
/** Pointer operator.
*
* @return A pointer to the per-strand view instance.
*/
view_type* operator->() { return &view(); }
view_type const* operator->() const { return &view(); }
//@}
//@{
/** Deprecated view access.
*
* `r()` is a synonym for `*r` which was used with early versions of Cilk
* reducers. `*r` is now the preferred usage.
*
* @deprecated Use operator*() instead of operator()().
*
* @return A reference to the per-strand view instance.
*/
view_type& operator()() { return view(); }
view_type const& operator()() const { return view(); }
//@}
//@}
/** @name Set and get the value.
*
* These functions are used to set an initial value for the reducer before
* starting the reduction, or to get the final value after the reduction
* is complete.
*
* @note These functions are completely different from the view
* operations that are made available via operator*() and
* operator->(), which are used to _modify_ the reducers value
* _during_ the reduction.
*
* @warning These functions _can_ be called at any time, and in
* general, they will refer to the value contained in the view
* for the current strand. However, using them other than to
* set the reductions initial value or get its final value
* will almost always result in undefined behavior.
*/
//@{
/** Move a value into the reducer.
*
* This function is used to set the initial value of the reducers
* accumulator variable by either copying or _moving_ the value of @a obj
* into it. Moving a value can often be performed in constant time, even
* for large container objects, but has the side effect of leaving the
* value of @a obj undefined. (See the description of the
* @ref move_in_wrapper class for a discussion of moving values.)
*
* @par Usage
* A move_in() call to initialize a reducer is often paired with a
* move_out() call to get its final value:
*
* reducer<Type> xr;
* xr.move_in(x);
* … do the reduction …
* xr.move_out(x);
*
* @par Assumptions
* - You cannot assume either that this will function will copy its
* value or that it will move it.
* - You must assume that the value of @a obj will be undefined
* after the call to move_in().
* - You can assume that move_in() will be at least as efficient as
* set_value(), and you should therefore prefer move_in() unless
* you need the value of @a obj to be unchanged after the call.
* (But you should usually prefer the move-in constructor over a
* move_in() call — see the note below.)
*
* @note The behavior of a default constructor followed by move-in
* initialization:
*
* reducer<Type> xr;
* xr.move_in(x);
*
* @note is not necessarily the same as a move-in constructor:
*
* reducer<Type> xr(move_in(x));
*
* @note In particular, when @a Type is a container type with a
* non-empty allocator, the move-in constructor will create the
* accumulator variable with the same allocator as the input
* argument @a x, while the default constructor will create the
* accumulator variable with a default allocator. The mismatch of
* allocators in the latter case means that the input argument
* @a x may have to be copied in linear time instead of being
* moved in constant time.
*
* @note Best practice is to prefer the move-in constructor over the
* move-in function unless the move-in function is required for
* some specific reason.
*
* @warning Calling this function other than to set the initial value
* for a reduction will almost always result in undefined
* behavior.
*
* @param obj The object containing the value that will be moved into the
* reducer.
*
* @post The reducer contains the value that was initially in @a obj.
* @post The value of @a obj is undefined.
*
* @see set_value()
*/
void move_in(value_type& obj) { set_get::move_in(view(), obj);}
/** Move the value out of the reducer.
*
* This function is used to retrieve the final value of the reducers
* accumulator variable by either copying or _moving_ the value of @a obj
* into it. Moving a value can often be performed in constant time, even
* for large container objects, but has the side effect of leaving the
* value of the reducers accumulator variable undefined. (See the
* description of the @ref move_in_wrapper class for a discussion of
* moving values.)
*
* @par Usage
* A move_in() call to initialize a reducer is often paired with a
* move_out() call to get its final value:
*
* reducer<Type> xr;
* xr.move_in(x);
* … do the reduction …
* xr.move_out(x);
*
* @par Assumptions
* - You cannot assume either that this will function will copy its
* value or that it will move it.
* - You must assume that the value of the reducers accumulator
* variable will be undefined after the call to move_out().
* - You can assume that move_out() will be at least as efficient as
* get_value(), and you should therefore prefer move_out() unless
* you need the accumulator variable to be preserved after the
* call.
*
* @warning Calling this function other than to retrieve the final
* value of a reduction will almost always result in undefined
* behavior.
*
* @param obj The object that the value of the reducer will be moved into.
*
* @post @a obj contains the value that was initially in the reducer.
* @post The value of the reducer is undefined.
*
* @see get_value()
*/
void move_out(value_type& obj) { set_get::move_out(view(), obj); }
/** Set the value of the reducer.
*
* This function sets the initial value of the reducers accumulator
* variable to the value of @a obj.
*
* @note The behavior of a default constructor followed by
* initialization:
*
* reducer<Type> xr;
* xr.set_value(x);
*
* @note is not necessarily the same as a value constructor:
*
* reducer<Type> xr(x);
*
* @note In particular, when @a Type is a container type with a
* non-empty allocator, the value constructor will create the
* accumulator variable with the same allocator as the input
* argument @a x, while the default constructor will create the
* accumulator variable with a default allocator.
*
* @warning Calling this function other than to set the initial value
* for a reduction will almost always result in undefined
* behavior.
*
* @param obj The object containing the value that will be copied into
* the reducer.
*
* @post The reducer contains a copy of the value in @a obj.
*
* @see move_in()
*/
void set_value(const value_type& obj) { set_get::set_value(view(), obj); }
/** Get the value of the reducer.
*
* This function gets the final value of the reducers accumulator
* variable.
*
* @warning Calling this function other than to retrieve the final
* value of a reduction will almost always result in undefined
* behavior.
*
* @return A reference to the value contained in the reducer.
*
* @see move_out()
*/
typename set_get::get_value_type get_value() const
{ return set_get::get_value(view()); }
//@}
/** Implicit downcast to legacy reducer wrapper, if any.
*
* @see legacy_reducer_downcast
*/
operator typename legacy_reducer_downcast<reducer>::type& ()
{
typedef typename legacy_reducer_downcast<reducer>::type downcast_type;
return *reinterpret_cast<downcast_type*>(this);
}
/** Implicit downcast to legacy reducer wrapper, if any.
*
* @see legacy_reducer_downcast
*/
operator const typename legacy_reducer_downcast<reducer>::type& () const
{
typedef typename legacy_reducer_downcast<reducer>::type downcast_type;
return *reinterpret_cast<const downcast_type*>(this);
}
};
#ifdef CILK_STUB
} // namespace stub
using stub::reducer;
#endif
} // end namespace cilk
#endif /* __cplusplus */
/** @page page_reducers_in_c Creating and Using Reducers in C
*
* @tableofcontents
*
* The Cilk runtime supports reducers written in C as well as in C++. The basic logic is the
* same, but the implementation details are very different. The C++ reducer implementation uses
* templates heavily to create very generic components. The C reducer implementation uses
* macros, which are a much blunter instrument. The most immediate consequence is that the
* monoid/view/reducer architecture is mostly implicit rather than explicit in C reducers.
*
* @section reducers_c_overview Overview of Using Reducers in C
*
* The basic usage pattern for C reducers is:
*
* 1. Create and initialize a reducer object.
* 2. Tell the Cilk runtime about the reducer.
* 3. Update the value contained in the reducer in a parallel computation.
* 4. Tell the Cilk runtime that you are done with the reducer.
* 5. Retrieve the value from the reducer.
*
* @subsection reducers_c_creation Creating and Initializing a C Reducer
*
* The basic pattern for creating and initializing a reducer object in C is
*
* CILK_C_DECLARE_REDUCER(value-type) reducer-name =
* CILK_C_INIT_REDUCER(value-type,
* reduce-function,
* identity-function,
* destroy-function,
* initial-value);
*
* This is simply an initialized definition of a variable named _reducer-name_. The
* @ref CILK_C_DECLARE_REDUCER macro expands to an anonymous `struct` declaration for a reducer
* object containing a view of type _value-type_, and the @ref CILK_C_INIT_REDUCER macro
* expands to a struct initializer.
*
* @subsection reducers_c_reduce_func Reduce Functions
*
* The reduce function for a reducer is called when a parallel execution strand terminates, to
* combine the values computed by the terminating strand and the strand to its left. It takes
* three arguments:
*
* - `void* reducer` — the address of the reducer.
* - `void* left` — the address of the value for the left strand.
* - `void* right` — the address of the value for the right (terminating) strand.
*
* It must apply the reducers reduction operation to the `left` and `right` values, leaving
* the result in the `left` value. The `right` value is undefined after the reduce function
* call.
*
* @subsection reducers_c_identity_func Identity Functions
*
* The identity function for a reducer is called when a parallel execution strand begins, to
* initialize its value to the reducers identity value. It takes two arguments:
*
* - `void* reducer` — the address of the reducer.
* - `void* v` — the address of a freshly allocated block of memory of size
* `sizeof(value-type)`.
*
* It must initialize the memory pointed to by `v` so that it contains the reducers identity
* value.
*
* @subsection reducers_c_destroy_func Destroy Functions
*
* The destroy function for a reducer is called when a parallel execution strand terminates, to
* do any necessary cleanup before its value is deallocated. It takes two arguments:
*
* - `void* reducer` — the address of the reducer.
* - `void* p` — the address of the value for the terminating strand.
*
* It must release any resources belonging to the value pointed to by `p`, to avoid a resource
* leak when the memory containing the value is deallocated.
*
* The runtime function `__cilkrts_hyperobject_noop_destroy` can be used for the destructor
* function if the reducers values do not need any cleanup.
*
* @subsection reducers_c_register Tell the Cilk Runtime About the Reducer
*
* Call the @ref CILK_C_REGISTER_REDUCER macro to register the reducer with the Cilk runtime:
*
* CILK_C_REGISTER_REDUCER(reducer-name);
*
* The runtime will manage reducer values for all registered reducers when parallel execution
* strands begin and end.
*
* @subsection reducers_c_update Update the Value Contained in the Reducer
*
* The @ref REDUCER_VIEW macro returns a reference to the reducers value for the current
* parallel strand:
*
* REDUCER_VIEW(reducer-name) = REDUCER_VIEW(reducer-name) OP x;
*
* C++ reducer views restrict access to the wrapped value so that it can only be modified in
* ways consistent with the reducers operation. No such protection is provided for C reducers.
* It is
* entirely the responsibility of the user to avoid modifying the value in any
* inappropriate way.
*
* @subsection c_reducers_unregister Tell the Cilk Runtime That You Are Done with the Reducer
*
* When the parallel computation is complete, call the @ref CILK_C_UNREGISTER_REDUCER macro to
* unregister the reducer with the Cilk runtime:
*
* CILK_C_UNREGISTER_REDUCER(reducer-name);
*
* The runtime will stop managing reducer values for the reducer.
*
* @subsection c_reducers_retrieve Retrieve the Value from the Reducer
*
* When the parallel computation is complete, use the @ref REDUCER_VIEW macro to retrieve the
* final value computed by the reducer.
*
* @subsection reducers_c_example_custom Example — Creating and Using a Custom C Reducer
*
* The `IntList` type represents a simple list of integers.
*
* struct _intListNode {
* int value;
* _intListNode* next;
* } IntListNode;
* typedef struct { IntListNode* head; IntListNode* tail; } IntList;
*
* // Initialize a list to be empty
* void IntList_init(IntList* list) { list->head = list->tail = 0; }
*
* // Append an integer to the list
* void IntList_append(IntList* list, int x)
* {
* IntListNode* node = (IntListNode*) malloc(sizeof(IntListNode));
* if (list->tail) list->tail->next = node; else list->head = node;
* list->tail = node;
* }
*
* // Append the right list to the left list, and leave the right list empty
* void IntList_concat(IntList* left, IntList* right)
* {
* if (left->head) {
* left->tail->next = right->head;
* if (right->tail) left->tail = right->tail;
* }
* else {
* *left = *right;
* }
* IntList_init(*right);
* }
*
* This code creates a reducer that supports creating an `IntList` by appending values to it.
*
* void identity_IntList(void* reducer, void* list)
* {
* IntList_init((IntList*)list);
* }
*
* void reduce_IntList(void* reducer, void* left, void* right)
* {
* IntList_concat((IntList*)left, (IntList*)right);
* }
*
* CILK_C_DECLARE_REDUCER(IntList) my_list_int_reducer =
* CILK_C_INIT_REDUCER(IntList,
* reduce_int_list,
* identity_int_list,
* __cilkrts_hyperobject_noop_destroy);
* // Initial value omitted //
* ListInt_init(&REDUCER_VIEW(my_int_list_reducer));
*
* CILK_C_REGISTER_REDUCER(my_int_list_reducer);
* cilk_for (int i = 0; i != n; ++i) {
* IntList_append(&REDUCER_VIEW(my_int_list_reducer), a[i]);
* }
* CILK_C_UNREGISTER_REDUCER(my_int_list_reducer);
*
* IntList result = REDUCER_VIEW(my_int_list_reducer);
*
* @section reducers_c_predefined Predefined C Reducers
*
* Some of the predefined reducer classes in the Cilk library come with a set of predefined
* macros to provide the same capabilities in C. In general, two macros are provided for each
* predefined reducer family:
*
* - `CILK_C_REDUCER_operation(reducer-name, type-name, initial-value)` — Declares a
* reducer object named _reducer-name_ with initial value _initial-value_ to perform
* a reduction using the _operation_ on values of the type specified by _type-name_.
* This is the equivalent of the general code described in @ref reducers_c_creation :
*
* CILK_C_DECLARE_REDUCER(type) reducer-name =
* CILK_C_INIT_REDUCER(type, ..., initial-value);
*
* where _type_ is the C type corresponding to _type_name_. See @ref reducers_c_type_names
* below for the _type-names_ that you can use.
*
* - `CILK_C_REDUCER_operation_TYPE(type-name)` — Expands to the `typedef` name for the type
* of the reducer object declared by
* `CILK_C_REDUCER_operation(reducer-name, type-name, initial-value)`.
*
* See @ref reducers_c_example_predefined.
*
* The predefined C reducers are:
*
* | Operation | Name | Documentation |
* |-------------------|---------------|-------------------------------|
* | addition | `OPADD` | @ref ReducersAdd |
* | bitwise and | `OPAND` | @ref ReducersAnd |
* | bitwise or | `OPOR` | @ref ReducersOr |
* | bitwise xor | `OPXOR` | @ref ReducersXor |
* | multiplication | `OPMUL` | @ref ReducersMul |
* | minimum | `MIN` | @ref ReducersMinMax |
* | minimum & index | `MIN_INDEX` | @ref ReducersMinMax |
* | maximum | `MIN` | @ref ReducersMinMax |
* | maximum & index | `MIN_INDEX` | @ref ReducersMinMax |
*
* @subsection reducers_c_type_names Numeric Type Names
*
* The type and function names created by the C reducer definition macros incorporate both the
* reducer kind (`opadd`, `opxor`, etc.) and the value type of the reducer (`int`, `double`,
* etc.). The value type is represented by a _numeric type name_ string. The types supported
* in C reducers, and their corresponding numeric type names, are given in the following table:
*
* | Type | Numeric Type Name |
* |-----------------------|-------------------------------|
* | `char` | `char` |
* | `unsigned char` | `uchar` |
* | `signed char` | `schar` |
* | `wchar_t` | `wchar_t` |
* | `short` | `short` |
* | `unsigned short` | `ushort` |
* | `int` | `int` |
* | `unsigned int` | `uint` |
* | `unsigned int` | `unsigned` (alternate name) |
* | `long` | `long` |
* | `unsigned long` | `ulong` |
* | `long long` | `longlong` |
* | `unsigned long long` | `ulonglong` |
* | `float` | `float` |
* | `double` | `double` |
* | `long double` | `longdouble` |
*
* @subsection reducers_c_example_predefined Example — Using a Predefined C Reducer
*
* To compute the sum of all the values in an array of `unsigned int`:
*
* CILK_C_REDUCER_OPADD(sum, uint, 0);
* CILK_C_REGISTER_REDUCER(sum);
* cilk_for(int i = 0; i != n; ++i) {
* REDUCER_VIEW(sum) += a[i];
* }
* CILK_C_UNREGISTER_REDUCER(sum);
* printf("The sum is %u\n", REDUCER_VIEW(sum));
*/
/** @name C language reducer macros
*
* These macros are used to declare and work with reducers in C code.
*
* @see @ref page_reducers_in_c
*/
//@{
/// @cond internal
/** @name Compound identifier macros.
*
* These macros are used to construct an identifier by concatenating two or three identifiers.
*/
//@{
/** Expand to an identifier formed by concatenating two identifiers.
*/
#define __CILKRTS_MKIDENT(a,b) __CILKRTS_MKIDENT_IMP(a,b,)
/** Expand to an identifier formed by concatenating three identifiers.
*/
#define __CILKRTS_MKIDENT3(a,b,c) __CILKRTS_MKIDENT_IMP(a,b,c)
/** Helper macro to do the concatenation.
*/
#define __CILKRTS_MKIDENT_IMP(a,b,c) a ## b ## c
//@}
/** Compiler-specific keyword for the “type of” operator.
*/
#if defined(__GNUC__) && !defined(__INTEL_COMPILER)
# define _Typeof __typeof__
#endif
/** @name Predefined reducer function declaration macros.
*
* These macros are used to create the function headers for the identity, reduction,
* and destructor functions for a builtin reducer family. The macro can be followed by
* a semicolon to create a declaration, or by a brace-enclosed body to create a definition.
*/
//@{
/** Create an identity function header.
*
* @note The name of the functions value pointer parameter will always be `v`.
*
* @param name The reducer family name.
* @param tn The type name.
*/
#define __CILKRTS_DECLARE_REDUCER_IDENTITY(name,tn) CILK_EXPORT \
void __CILKRTS_MKIDENT3(name,_identity_,tn)(void* key, void* v)
/** Create a reduction function header.
*
* @param name The reducer family name.
* @param tn The type name.
* @param l The name to use for the functions left value pointer parameter.
* @param r The name to use for the functions right value pointer parameter.
*/
#define __CILKRTS_DECLARE_REDUCER_REDUCE(name,tn,l,r) CILK_EXPORT \
void __CILKRTS_MKIDENT3(name,_reduce_,tn)(void* key, void* l, void* r)
/** Create a destructor function header.
*
* @param name The reducer family name.
* @param tn The type name.
* @param p The name to use for the functions value pointer parameter.
*/
#define __CILKRTS_DECLARE_REDUCER_DESTROY(name,tn,p) CILK_EXPORT \
void __CILKRTS_MKIDENT3(name,_destroy_,tn)(void* key, void* p)
//@}
/// @endcond
/***************************************************************************
* Real implementation
***************************************************************************/
/** Declaration of a C reducer structure type.
*
* This macro expands into an anonymous structure declaration for a C reducer structure
* which contains a @a Type value. For example:
*
* CILK_C_DECLARE_REDUCER(int) my_add_int_reducer =
* CILK_C_INIT_REDUCER(int, …);
*
* @param Type The type of the value contained in the reducer object.
*
* @see @ref reducers_c_creation
*/
#define CILK_C_DECLARE_REDUCER(Type) struct { \
__cilkrts_hyperobject_base __cilkrts_hyperbase; \
__CILKRTS_CACHE_ALIGN Type value; \
}
/** Initializer for a C reducer structure.
*
* This macro expands into a brace-enclosed structure initializer for a C reducer structure
* that was declared with `CILK_C_DECLARE_REDUCER(Type)`. For example:
*
* CILK_C_DECLARE_REDUCER(int) my_add_int_reducer =
* CILK_C_INIT_REDUCER(int,
* add_int_reduce,
* add_int_identity,
* __cilkrts_hyperobject_noop_destroy,
* 0);
*
* @param Type The type of the value contained in the reducer object. Must be the same as
* the @a Type argument of the CILK_C_DECLARE_REDUCER macro call that created
* the reducer.
* @param Reduce The address of the @ref reducers_c_reduce_func "reduce function" for the
* reducer.
* @param Identity The address of the @ref reducers_c_identity_func "identity function" for
* the reducer.
* @param Destroy The address of the @ref reducers_c_destroy_func "destroy function" for the
* reducer.
* @param ... The initial value for the reducer. (A single expression if @a Type is a
* scalar type; a list of values if @a Type is a struct or array type.)
*
* @see @ref reducers_c_creation
*/
#define CILK_C_INIT_REDUCER(Type, Reduce, Identity, Destroy, ...) \
{ { { Reduce \
, Identity \
, Destroy \
, __cilkrts_hyperobject_alloc \
, __cilkrts_hyperobject_dealloc \
} \
, 0 \
, __CILKRTS_CACHE_LINE__ \
, sizeof(Type) \
} \
, __VA_ARGS__ \
}
/** Register a reducer with the Cilk runtime.
*
* The runtime will manage reducer values for all registered reducers when parallel execution
* strands begin and end. For example:
*
* CILK_C_REGISTER_REDUCER(my_add_int_reducer);
* cilk_for (int i = 0; i != n; ++i) {
* …
* }
*
* @param Expr The reducer to be registered.
*
* @see @ref page_reducers_in_c
*/
#define CILK_C_REGISTER_REDUCER(Expr) \
__cilkrts_hyper_create(&(Expr).__cilkrts_hyperbase)
/** Unregister a reducer with the Cilk runtime.
*
* The runtime will stop managing reducer values for a reducer after it is unregistered. For
* example:
*
* cilk_for (int i = 0; i != n; ++i) {
* …
* }
* CILK_C_UNREGISTER_REDUCER(my_add_int_reducer);
*
* @param Expr The reducer to be unregistered.
*
* @see @ref page_reducers_in_c
*/
#define CILK_C_UNREGISTER_REDUCER(Expr) \
__cilkrts_hyper_destroy(&(Expr).__cilkrts_hyperbase)
/** Get the current view for a reducer.
*
* The `REDUCER_VIEW(reducer-name)` returns a reference to the reducers value for the
* current parallel strand. This can be used to initialize thevalue of the reducer before it
* is used, to modify the value of the reducer on the current parallel strand, or to retrieve
* the final value of the reducer at the end of the parallel computation.
*
* REDUCER_VIEW(my_add_int_reducer) = REDUCER_VIEW(my_add_int_reducer) + x;
*
* @note C++ reducer views restrict access to the wrapped value so that it can only be
* modified in ways consistent with the reducers operation. No such protection is provided
* for C reducers. It is entirely the responsibility of the user to refrain from modifying the
* value in any inappropriate way.
*
* @param Expr The reducer whose value is to be returned.
*
* @see @ref page_reducers_in_c
*/
#define REDUCER_VIEW(Expr) (*(_Typeof((Expr).value)*) \
__cilkrts_hyper_lookup(&(Expr).__cilkrts_hyperbase))
//@} C language reducer macros
#endif // CILK_REDUCER_H_INCLUDED