head	1.1;
branch	1.1.1;
access;
symbols
	netbsd-11-0-RC5:1.1.1.3
	netbsd-11-0-RC4:1.1.1.3
	netbsd-11-0-RC3:1.1.1.3
	netbsd-11-0-RC2:1.1.1.3
	netbsd-11-0-RC1:1.1.1.3
	gcc-14-3-0:1.1.1.4
	perseant-exfatfs-base-20250801:1.1.1.3
	netbsd-11:1.1.1.3.0.4
	netbsd-11-base:1.1.1.3
	gcc-12-5-0:1.1.1.3
	netbsd-10-1-RELEASE:1.1.1.2
	perseant-exfatfs-base-20240630:1.1.1.3
	gcc-12-4-0:1.1.1.3
	perseant-exfatfs:1.1.1.3.0.2
	perseant-exfatfs-base:1.1.1.3
	netbsd-10-0-RELEASE:1.1.1.2
	netbsd-10-0-RC6:1.1.1.2
	netbsd-10-0-RC5:1.1.1.2
	netbsd-10-0-RC4:1.1.1.2
	netbsd-10-0-RC3:1.1.1.2
	netbsd-10-0-RC2:1.1.1.2
	netbsd-10-0-RC1:1.1.1.2
	gcc-12-3-0:1.1.1.3
	gcc-10-5-0:1.1.1.2
	netbsd-10:1.1.1.2.0.6
	netbsd-10-base:1.1.1.2
	gcc-10-4-0:1.1.1.2
	cjep_sun2x-base1:1.1.1.2
	cjep_sun2x:1.1.1.2.0.4
	cjep_sun2x-base:1.1.1.2
	cjep_staticlib_x-base1:1.1.1.2
	cjep_staticlib_x:1.1.1.2.0.2
	cjep_staticlib_x-base:1.1.1.2
	gcc-10-3-0:1.1.1.2
	gcc-9-3-0:1.1.1.1
	FSF:1.1.1;
locks; strict;
comment	@ * @;


1.1
date	2020.09.05.07.52.10;	author mrg;	state Exp;
branches
	1.1.1.1;
next	;
commitid	ZRYA7IOuwfMjAPmC;

1.1.1.1
date	2020.09.05.07.52.10;	author mrg;	state Exp;
branches;
next	1.1.1.2;
commitid	ZRYA7IOuwfMjAPmC;

1.1.1.2
date	2021.04.10.22.10.04;	author mrg;	state Exp;
branches;
next	1.1.1.3;
commitid	eC4g0MRpqTvEkNOC;

1.1.1.3
date	2023.07.30.05.21.20;	author mrg;	state Exp;
branches;
next	1.1.1.4;
commitid	tk6nV4mbc9nVEMyE;

1.1.1.4
date	2025.09.13.23.45.48;	author mrg;	state Exp;
branches;
next	;
commitid	KwhwN4krNWa6XBaG;


desc
@@


1.1
log
@Initial revision
@
text
@// -*- C++ -*-
//===-- unseq_backend_simd.h ----------------------------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//

#ifndef __PSTL_unseq_backend_simd_H
#define __PSTL_unseq_backend_simd_H

#include <type_traits>

#include "utils.h"

// This header defines the minimum set of vector routines required
// to support parallel STL.
namespace __pstl
{
namespace __unseq_backend
{

// Expect vector width up to 64 (or 512 bit)
const std::size_t __lane_size = 64;

template <class _Iterator, class _DifferenceType, class _Function>
_Iterator
__simd_walk_1(_Iterator __first, _DifferenceType __n, _Function __f) noexcept
{
    __PSTL_PRAGMA_SIMD
    for (_DifferenceType __i = 0; __i < __n; ++__i)
        __f(__first[__i]);

    return __first + __n;
}

template <class _Iterator1, class _DifferenceType, class _Iterator2, class _Function>
_Iterator2
__simd_walk_2(_Iterator1 __first1, _DifferenceType __n, _Iterator2 __first2, _Function __f) noexcept
{
    __PSTL_PRAGMA_SIMD
    for (_DifferenceType __i = 0; __i < __n; ++__i)
        __f(__first1[__i], __first2[__i]);
    return __first2 + __n;
}

template <class _Iterator1, class _DifferenceType, class _Iterator2, class _Iterator3, class _Function>
_Iterator3
__simd_walk_3(_Iterator1 __first1, _DifferenceType __n, _Iterator2 __first2, _Iterator3 __first3,
              _Function __f) noexcept
{
    __PSTL_PRAGMA_SIMD
    for (_DifferenceType __i = 0; __i < __n; ++__i)
        __f(__first1[__i], __first2[__i], __first3[__i]);
    return __first3 + __n;
}

// TODO: check whether __simd_first() can be used here
template <class _Index, class _DifferenceType, class _Pred>
bool
__simd_or(_Index __first, _DifferenceType __n, _Pred __pred) noexcept
{
#if __PSTL_EARLYEXIT_PRESENT
    _DifferenceType __i;
    __PSTL_PRAGMA_VECTOR_UNALIGNED
    __PSTL_PRAGMA_SIMD_EARLYEXIT
    for (__i = 0; __i < __n; ++__i)
        if (__pred(__first[__i]))
            break;
    return __i < __n;
#else
    _DifferenceType __block_size = 4 < __n ? 4 : __n;
    const _Index __last = __first + __n;
    while (__last != __first)
    {
        int32_t __flag = 1;
        __PSTL_PRAGMA_SIMD_REDUCTION(& : __flag)
        for (_DifferenceType __i = 0; __i < __block_size; ++__i)
            if (__pred(*(__first + __i)))
                __flag = 0;
        if (!__flag)
            return true;

        __first += __block_size;
        if (__last - __first >= __block_size << 1)
        {
            // Double the block _Size.  Any unnecessary iterations can be amortized against work done so far.
            __block_size <<= 1;
        }
        else
        {
            __block_size = __last - __first;
        }
    }
    return false;
#endif
}

template <class _Index, class _DifferenceType, class _Compare>
_Index
__simd_first(_Index __first, _DifferenceType __begin, _DifferenceType __end, _Compare __comp) noexcept
{
#if __PSTL_EARLYEXIT_PRESENT
    _DifferenceType __i = __begin;
    __PSTL_PRAGMA_VECTOR_UNALIGNED // Do not generate peel loop part
        __PSTL_PRAGMA_SIMD_EARLYEXIT for (; __i < __end; ++__i)
    {
        if (__comp(__first, __i))
        {
            break;
        }
    }
    return __first + __i;
#else
    // Experiments show good block sizes like this
    const _DifferenceType __block_size = 8;
    alignas(__lane_size) _DifferenceType __lane[__block_size] = {0};
    while (__end - __begin >= __block_size)
    {
        _DifferenceType __found = 0;
        __PSTL_PRAGMA_VECTOR_UNALIGNED // Do not generate peel loop part
            __PSTL_PRAGMA_SIMD_REDUCTION(|
                                         : __found) for (_DifferenceType __i = __begin; __i < __begin + __block_size;
                                                         ++__i)
        {
            const _DifferenceType __t = __comp(__first, __i);
            __lane[__i - __begin] = __t;
            __found |= __t;
        }
        if (__found)
        {
            _DifferenceType __i;
            // This will vectorize
            for (__i = 0; __i < __block_size; ++__i)
            {
                if (__lane[__i])
                {
                    break;
                }
            }
            return __first + __begin + __i;
        }
        __begin += __block_size;
    }

    //Keep remainder scalar
    while (__begin != __end)
    {
        if (__comp(__first, __begin))
        {
            return __first + __begin;
        }
        ++__begin;
    }
    return __first + __end;
#endif //__PSTL_EARLYEXIT_PRESENT
}

template <class _Index1, class _DifferenceType, class _Index2, class _Pred>
std::pair<_Index1, _Index2>
__simd_first(_Index1 __first1, _DifferenceType __n, _Index2 __first2, _Pred __pred) noexcept
{
#if __PSTL_EARLYEXIT_PRESENT
    _DifferenceType __i = 0;
    __PSTL_PRAGMA_VECTOR_UNALIGNED
    __PSTL_PRAGMA_SIMD_EARLYEXIT
    for (; __i < __n; ++__i)
        if (__pred(__first1[__i], __first2[__i]))
            break;
    return std::make_pair(__first1 + __i, __first2 + __i);
#else
    const _Index1 __last1 = __first1 + __n;
    const _Index2 __last2 = __first2 + __n;
    // Experiments show good block sizes like this
    const _DifferenceType __block_size = 8;
    alignas(__lane_size) _DifferenceType __lane[__block_size] = {0};
    while (__last1 - __first1 >= __block_size)
    {
        _DifferenceType __found = 0;
        _DifferenceType __i;
        __PSTL_PRAGMA_VECTOR_UNALIGNED // Do not generate peel loop part
            __PSTL_PRAGMA_SIMD_REDUCTION(|
                                         : __found) for (__i = 0; __i < __block_size; ++__i)
        {
            const _DifferenceType __t = __pred(__first1[__i], __first2[__i]);
            __lane[__i] = __t;
            __found |= __t;
        }
        if (__found)
        {
            _DifferenceType __i;
            // This will vectorize
            for (__i = 0; __i < __block_size; ++__i)
            {
                if (__lane[__i])
                    break;
            }
            return std::make_pair(__first1 + __i, __first2 + __i);
        }
        __first1 += __block_size;
        __first2 += __block_size;
    }

    //Keep remainder scalar
    for (; __last1 != __first1; ++__first1, ++__first2)
        if (__pred(*(__first1), *(__first2)))
            return std::make_pair(__first1, __first2);

    return std::make_pair(__last1, __last2);
#endif //__PSTL_EARLYEXIT_PRESENT
}

template <class _Index, class _DifferenceType, class _Pred>
_DifferenceType
__simd_count(_Index __index, _DifferenceType __n, _Pred __pred) noexcept
{
    _DifferenceType __count = 0;
    __PSTL_PRAGMA_SIMD_REDUCTION(+ : __count)
    for (_DifferenceType __i = 0; __i < __n; ++__i)
        if (__pred(*(__index + __i)))
            ++__count;

    return __count;
}

template <class _InputIterator, class _DifferenceType, class _OutputIterator, class _BinaryPredicate>
_OutputIterator
__simd_unique_copy(_InputIterator __first, _DifferenceType __n, _OutputIterator __result,
                   _BinaryPredicate __pred) noexcept
{
    if (__n == 0)
        return __result;

    _DifferenceType __cnt = 1;
    __result[0] = __first[0];

    __PSTL_PRAGMA_SIMD
    for (_DifferenceType __i = 1; __i < __n; ++__i)
    {
        __PSTL_PRAGMA_SIMD_ORDERED_MONOTONIC(__cnt : 1)
        if (!__pred(__first[__i], __first[__i - 1]))
        {
            __result[__cnt] = __first[__i];
            ++__cnt;
        }
    }
    return __result + __cnt;
}

template <class _InputIterator, class _DifferenceType, class _OutputIterator, class _Assigner>
_OutputIterator
__simd_assign(_InputIterator __first, _DifferenceType __n, _OutputIterator __result, _Assigner __assigner) noexcept
{
    __PSTL_USE_NONTEMPORAL_STORES_IF_ALLOWED
    __PSTL_PRAGMA_SIMD
    for (_DifferenceType __i = 0; __i < __n; ++__i)
        __assigner(__first + __i, __result + __i);
    return __result + __n;
}

template <class _InputIterator, class _DifferenceType, class _OutputIterator, class _UnaryPredicate>
_OutputIterator
__simd_copy_if(_InputIterator __first, _DifferenceType __n, _OutputIterator __result, _UnaryPredicate __pred) noexcept
{
    _DifferenceType __cnt = 0;

    __PSTL_PRAGMA_SIMD
    for (_DifferenceType __i = 0; __i < __n; ++__i)
    {
        __PSTL_PRAGMA_SIMD_ORDERED_MONOTONIC(__cnt : 1)
        if (__pred(__first[__i]))
        {
            __result[__cnt] = __first[__i];
            ++__cnt;
        }
    }
    return __result + __cnt;
}

template <class _InputIterator, class _DifferenceType, class _BinaryPredicate>
_DifferenceType
__simd_calc_mask_2(_InputIterator __first, _DifferenceType __n, bool* __mask, _BinaryPredicate __pred) noexcept
{
    _DifferenceType __count = 0;

    __PSTL_PRAGMA_SIMD_REDUCTION(+ : __count)
    for (_DifferenceType __i = 0; __i < __n; ++__i)
    {
        __mask[__i] = !__pred(__first[__i], __first[__i - 1]);
        __count += __mask[__i];
    }
    return __count;
}

template <class _InputIterator, class _DifferenceType, class _UnaryPredicate>
_DifferenceType
__simd_calc_mask_1(_InputIterator __first, _DifferenceType __n, bool* __mask, _UnaryPredicate __pred) noexcept
{
    _DifferenceType __count = 0;

    __PSTL_PRAGMA_SIMD_REDUCTION(+ : __count)
    for (_DifferenceType __i = 0; __i < __n; ++__i)
    {
        __mask[__i] = __pred(__first[__i]);
        __count += __mask[__i];
    }
    return __count;
}

template <class _InputIterator, class _DifferenceType, class _OutputIterator, class _Assigner>
void
__simd_copy_by_mask(_InputIterator __first, _DifferenceType __n, _OutputIterator __result, bool* __mask,
                    _Assigner __assigner) noexcept
{
    _DifferenceType __cnt = 0;
    __PSTL_PRAGMA_SIMD
    for (_DifferenceType __i = 0; __i < __n; ++__i)
    {
        if (__mask[__i])
        {
            __PSTL_PRAGMA_SIMD_ORDERED_MONOTONIC(__cnt : 1)
            {
                __assigner(__first + __i, __result + __cnt);
                ++__cnt;
            }
        }
    }
}

template <class _InputIterator, class _DifferenceType, class _OutputIterator1, class _OutputIterator2>
void
__simd_partition_by_mask(_InputIterator __first, _DifferenceType __n, _OutputIterator1 __out_true,
                         _OutputIterator2 __out_false, bool* __mask) noexcept
{
    _DifferenceType __cnt_true = 0, __cnt_false = 0;
    __PSTL_PRAGMA_SIMD
    for (_DifferenceType __i = 0; __i < __n; ++__i)
    {
        __PSTL_PRAGMA_SIMD_ORDERED_MONOTONIC_2ARGS(__cnt_true : 1, __cnt_false : 1)
        if (__mask[__i])
        {
            __out_true[__cnt_true] = __first[__i];
            ++__cnt_true;
        }
        else
        {
            __out_false[__cnt_false] = __first[__i];
            ++__cnt_false;
        }
    }
}

template <class _Index, class _DifferenceType, class _Tp>
_Index
__simd_fill_n(_Index __first, _DifferenceType __n, const _Tp& __value) noexcept
{
    __PSTL_USE_NONTEMPORAL_STORES_IF_ALLOWED
    __PSTL_PRAGMA_SIMD
    for (_DifferenceType __i = 0; __i < __n; ++__i)
        __first[__i] = __value;
    return __first + __n;
}

template <class _Index, class _DifferenceType, class _Generator>
_Index
__simd_generate_n(_Index __first, _DifferenceType __size, _Generator __g) noexcept
{
    __PSTL_USE_NONTEMPORAL_STORES_IF_ALLOWED
    __PSTL_PRAGMA_SIMD
    for (_DifferenceType __i = 0; __i < __size; ++__i)
        __first[__i] = __g();
    return __first + __size;
}

template <class _Index, class _BinaryPredicate>
_Index
__simd_adjacent_find(_Index __first, _Index __last, _BinaryPredicate __pred, bool __or_semantic) noexcept
{
    if (__last - __first < 2)
        return __last;

    typedef typename std::iterator_traits<_Index>::difference_type _DifferenceType;
    _DifferenceType __i = 0;

#if __PSTL_EARLYEXIT_PRESENT
    //Some compiler versions fail to compile the following loop when iterators are used. Indices are used instead
    const _DifferenceType __n = __last - __first - 1;
    __PSTL_PRAGMA_VECTOR_UNALIGNED
    __PSTL_PRAGMA_SIMD_EARLYEXIT
    for (; __i < __n; ++__i)
        if (__pred(__first[__i], __first[__i + 1]))
            break;

    return __i < __n ? __first + __i : __last;
#else
    // Experiments show good block sizes like this
    //TODO: to consider tuning block_size for various data types
    const _DifferenceType __block_size = 8;
    alignas(__lane_size) _DifferenceType __lane[__block_size] = {0};
    while (__last - __first >= __block_size)
    {
        _DifferenceType __found = 0;
        __PSTL_PRAGMA_VECTOR_UNALIGNED // Do not generate peel loop part
            __PSTL_PRAGMA_SIMD_REDUCTION(|
                                         : __found) for (__i = 0; __i < __block_size - 1; ++__i)
        {
            //TODO: to improve SIMD vectorization
            const _DifferenceType __t = __pred(*(__first + __i), *(__first + __i + 1));
            __lane[__i] = __t;
            __found |= __t;
        }

        //Process a pair of elements on a boundary of a data block
        if (__first + __block_size < __last && __pred(*(__first + __i), *(__first + __i + 1)))
            __lane[__i] = __found = 1;

        if (__found)
        {
            if (__or_semantic)
                return __first;

            // This will vectorize
            for (__i = 0; __i < __block_size; ++__i)
                if (__lane[__i])
                    break;
            return __first + __i; //As far as found is true a __result (__lane[__i] is true) is guaranteed
        }
        __first += __block_size;
    }
    //Process the rest elements
    for (; __last - __first > 1; ++__first)
        if (__pred(*__first, *(__first + 1)))
            return __first;

    return __last;
#endif
}

// It was created to reduce the code inside std::enable_if
template <typename _Tp, typename _BinaryOperation>
using is_arithmetic_plus = std::integral_constant<bool, std::is_arithmetic<_Tp>::value &&
                                                            std::is_same<_BinaryOperation, std::plus<_Tp>>::value>;

template <typename _DifferenceType, typename _Tp, typename _BinaryOperation, typename _UnaryOperation>
typename std::enable_if<is_arithmetic_plus<_Tp, _BinaryOperation>::value, _Tp>::type
__simd_transform_reduce(_DifferenceType __n, _Tp __init, _BinaryOperation, _UnaryOperation __f) noexcept
{
    __PSTL_PRAGMA_SIMD_REDUCTION(+ : __init)
    for (_DifferenceType __i = 0; __i < __n; ++__i)
        __init += __f(__i);
    return __init;
}

template <typename _Size, typename _Tp, typename _BinaryOperation, typename _UnaryOperation>
typename std::enable_if<!is_arithmetic_plus<_Tp, _BinaryOperation>::value, _Tp>::type
__simd_transform_reduce(_Size __n, _Tp __init, _BinaryOperation __binary_op, _UnaryOperation __f) noexcept
{
    const _Size __block_size = __lane_size / sizeof(_Tp);
    if (__n > 2 * __block_size && __block_size > 1)
    {
        alignas(__lane_size) char __lane_[__lane_size];
        _Tp* __lane = reinterpret_cast<_Tp*>(__lane_);

        // initializer
        __PSTL_PRAGMA_SIMD
        for (_Size __i = 0; __i < __block_size; ++__i)
        {
            ::new (__lane + __i) _Tp(__binary_op(__f(__i), __f(__block_size + __i)));
        }
        // main loop
        _Size __i = 2 * __block_size;
        const _Size last_iteration = __block_size * (__n / __block_size);
        for (; __i < last_iteration; __i += __block_size)
        {
            __PSTL_PRAGMA_SIMD
            for (_Size __j = 0; __j < __block_size; ++__j)
            {
                __lane[__j] = __binary_op(__lane[__j], __f(__i + __j));
            }
        }
        // remainder
        __PSTL_PRAGMA_SIMD
        for (_Size __j = 0; __j < __n - last_iteration; ++__j)
        {
            __lane[__j] = __binary_op(__lane[__j], __f(last_iteration + __j));
        }
        // combiner
        for (_Size __i = 0; __i < __block_size; ++__i)
        {
            __init = __binary_op(__init, __lane[__i]);
        }
        // destroyer
        __PSTL_PRAGMA_SIMD
        for (_Size __i = 0; __i < __block_size; ++__i)
        {
            __lane[__i].~_Tp();
        }
    }
    else
    {
        for (_Size __i = 0; __i < __n; ++__i)
        {
            __init = __binary_op(__init, __f(__i));
        }
    }
    return __init;
}

// Exclusive scan for "+" and arithmetic types
template <class _InputIterator, class _Size, class _OutputIterator, class _UnaryOperation, class _Tp,
          class _BinaryOperation>
typename std::enable_if<is_arithmetic_plus<_Tp, _BinaryOperation>::value, std::pair<_OutputIterator, _Tp>>::type
__simd_scan(_InputIterator __first, _Size __n, _OutputIterator __result, _UnaryOperation __unary_op, _Tp __init,
            _BinaryOperation, /*Inclusive*/ std::false_type)
{
    __PSTL_PRAGMA_SIMD_SCAN(+ : __init)
    for (_Size __i = 0; __i < __n; ++__i)
    {
        __result[__i] = __init;
        __PSTL_PRAGMA_SIMD_EXCLUSIVE_SCAN(__init)
        __init += __unary_op(__first[__i]);
    }
    return std::make_pair(__result + __n, __init);
}

// As soon as we cannot call __binary_op in "combiner" we create a wrapper over _Tp to encapsulate __binary_op
template <typename _Tp, typename _BinaryOp>
struct _Combiner
{
    _Tp __value;
    _BinaryOp* __bin_op; // Here is a pointer to function because of default ctor

    _Combiner() : __value{}, __bin_op(nullptr) {}
    _Combiner(const _Tp& value, const _BinaryOp* bin_op) : __value(value), __bin_op(const_cast<_BinaryOp*>(bin_op)) {}
    _Combiner(const _Combiner& __obj) : __value{}, __bin_op(__obj.__bin_op) {}

    void
    operator()(const _Combiner& __obj)
    {
        __value = (*__bin_op)(__value, __obj.__value);
    }
};

// Exclusive scan for other binary operations and types
template <class _InputIterator, class _Size, class _OutputIterator, class _UnaryOperation, class _Tp,
          class _BinaryOperation>
typename std::enable_if<!is_arithmetic_plus<_Tp, _BinaryOperation>::value, std::pair<_OutputIterator, _Tp>>::type
__simd_scan(_InputIterator __first, _Size __n, _OutputIterator __result, _UnaryOperation __unary_op, _Tp __init,
            _BinaryOperation __binary_op, /*Inclusive*/ std::false_type)
{
    typedef _Combiner<_Tp, _BinaryOperation> _CombinerType;
    _CombinerType __init_{__init, &__binary_op};

    __PSTL_PRAGMA_DECLARE_REDUCTION(__bin_op, _CombinerType)

    __PSTL_PRAGMA_SIMD_SCAN(__bin_op : __init_)
    for (_Size __i = 0; __i < __n; ++__i)
    {
        __result[__i] = __init_.__value;
        __PSTL_PRAGMA_SIMD_EXCLUSIVE_SCAN(__init_)
        __PSTL_PRAGMA_FORCEINLINE
        __init_.__value = __binary_op(__init_.__value, __unary_op(__first[__i]));
    }
    return std::make_pair(__result + __n, __init_.__value);
}

// Inclusive scan for "+" and arithmetic types
template <class _InputIterator, class _Size, class _OutputIterator, class _UnaryOperation, class _Tp,
          class _BinaryOperation>
typename std::enable_if<is_arithmetic_plus<_Tp, _BinaryOperation>::value, std::pair<_OutputIterator, _Tp>>::type
__simd_scan(_InputIterator __first, _Size __n, _OutputIterator __result, _UnaryOperation __unary_op, _Tp __init,
            _BinaryOperation, /*Inclusive*/ std::true_type)
{
    __PSTL_PRAGMA_SIMD_SCAN(+ : __init)
    for (_Size __i = 0; __i < __n; ++__i)
    {
        __init += __unary_op(__first[__i]);
        __PSTL_PRAGMA_SIMD_INCLUSIVE_SCAN(__init)
        __result[__i] = __init;
    }
    return std::make_pair(__result + __n, __init);
}

// Inclusive scan for other binary operations and types
template <class _InputIterator, class _Size, class _OutputIterator, class _UnaryOperation, class _Tp,
          class _BinaryOperation>
typename std::enable_if<!is_arithmetic_plus<_Tp, _BinaryOperation>::value, std::pair<_OutputIterator, _Tp>>::type
__simd_scan(_InputIterator __first, _Size __n, _OutputIterator __result, _UnaryOperation __unary_op, _Tp __init,
            _BinaryOperation __binary_op, std::true_type)
{
    typedef _Combiner<_Tp, _BinaryOperation> _CombinerType;
    _CombinerType __init_{__init, &__binary_op};

    __PSTL_PRAGMA_DECLARE_REDUCTION(__bin_op, _CombinerType)

    __PSTL_PRAGMA_SIMD_SCAN(__bin_op : __init_)
    for (_Size __i = 0; __i < __n; ++__i)
    {
        __PSTL_PRAGMA_FORCEINLINE
        __init_.__value = __binary_op(__init_.__value, __unary_op(__first[__i]));
        __PSTL_PRAGMA_SIMD_INCLUSIVE_SCAN(__init_)
        __result[__i] = __init_.__value;
    }
    return std::make_pair(__result + __n, __init_.__value);
}

// [restriction] - std::iterator_traits<_ForwardIterator>::value_type should be DefaultConstructible.
// complexity [violation] - We will have at most (__n-1 + number_of_lanes) comparisons instead of at most __n-1.
template <typename _ForwardIterator, typename _Size, typename _Compare>
_ForwardIterator
__simd_min_element(_ForwardIterator __first, _Size __n, _Compare __comp) noexcept
{
    if (__n == 0)
    {
        return __first;
    }

    typedef typename std::iterator_traits<_ForwardIterator>::value_type _ValueType;
    struct _ComplexType
    {
        _ValueType __min_val;
        _Size __min_ind;
        _Compare* __min_comp;

        _ComplexType() : __min_val{}, __min_ind{}, __min_comp(nullptr) {}
        _ComplexType(const _ValueType& val, const _Compare* comp)
            : __min_val(val), __min_ind(0), __min_comp(const_cast<_Compare*>(comp))
        {
        }
        _ComplexType(const _ComplexType& __obj)
            : __min_val(__obj.__min_val), __min_ind(__obj.__min_ind), __min_comp(__obj.__min_comp)
        {
        }

        __PSTL_PRAGMA_DECLARE_SIMD
        void
        operator()(const _ComplexType& __obj)
        {
            if (!(*__min_comp)(__min_val, __obj.__min_val) &&
                ((*__min_comp)(__obj.__min_val, __min_val) || __obj.__min_ind - __min_ind < 0))
            {
                __min_val = __obj.__min_val;
                __min_ind = __obj.__min_ind;
            }
        }
    };

    _ComplexType __init{*__first, &__comp};

    __PSTL_PRAGMA_DECLARE_REDUCTION(__min_func, _ComplexType)

    __PSTL_PRAGMA_SIMD_REDUCTION(__min_func : __init)
    for (_Size __i = 1; __i < __n; ++__i)
    {
        const _ValueType __min_val = __init.__min_val;
        const _ValueType __current = __first[__i];
        if (__comp(__current, __min_val))
        {
            __init.__min_val = __current;
            __init.__min_ind = __i;
        }
    }
    return __first + __init.__min_ind;
}

// [restriction] - std::iterator_traits<_ForwardIterator>::value_type should be DefaultConstructible.
// complexity [violation] - We will have at most (2*(__n-1) + 4*number_of_lanes) comparisons instead of at most [1.5*(__n-1)].
template <typename _ForwardIterator, typename _Size, typename _Compare>
std::pair<_ForwardIterator, _ForwardIterator>
__simd_minmax_element(_ForwardIterator __first, _Size __n, _Compare __comp) noexcept
{
    if (__n == 0)
    {
        return std::make_pair(__first, __first);
    }
    typedef typename std::iterator_traits<_ForwardIterator>::value_type _ValueType;

    struct _ComplexType
    {
        _ValueType __min_val;
        _ValueType __max_val;
        _Size __min_ind;
        _Size __max_ind;
        _Compare* __minmax_comp;

        _ComplexType() : __min_val{}, __max_val{}, __min_ind{}, __max_ind{}, __minmax_comp(nullptr) {}
        _ComplexType(const _ValueType& min_val, const _ValueType& max_val, const _Compare* comp)
            : __min_val(min_val), __max_val(max_val), __min_ind(0), __max_ind(0),
              __minmax_comp(const_cast<_Compare*>(comp))
        {
        }
        _ComplexType(const _ComplexType& __obj)
            : __min_val(__obj.__min_val), __max_val(__obj.__max_val), __min_ind(__obj.__min_ind),
              __max_ind(__obj.__max_ind), __minmax_comp(__obj.__minmax_comp)
        {
        }

        void
        operator()(const _ComplexType& __obj)
        {
            // min
            if ((*__minmax_comp)(__obj.__min_val, __min_val))
            {
                __min_val = __obj.__min_val;
                __min_ind = __obj.__min_ind;
            }
            else if (!(*__minmax_comp)(__min_val, __obj.__min_val))
            {
                __min_val = __obj.__min_val;
                __min_ind = (__min_ind - __obj.__min_ind < 0) ? __min_ind : __obj.__min_ind;
            }

            // max
            if ((*__minmax_comp)(__max_val, __obj.__max_val))
            {
                __max_val = __obj.__max_val;
                __max_ind = __obj.__max_ind;
            }
            else if (!(*__minmax_comp)(__obj.__max_val, __max_val))
            {
                __max_val = __obj.__max_val;
                __max_ind = (__max_ind - __obj.__max_ind < 0) ? __obj.__max_ind : __max_ind;
            }
        }
    };

    _ComplexType __init{*__first, *__first, &__comp};

    __PSTL_PRAGMA_DECLARE_REDUCTION(__min_func, _ComplexType);

    __PSTL_PRAGMA_SIMD_REDUCTION(__min_func : __init)
    for (_Size __i = 1; __i < __n; ++__i)
    {
        auto __min_val = __init.__min_val;
        auto __max_val = __init.__max_val;
        auto __current = __first + __i;
        if (__comp(*__current, __min_val))
        {
            __init.__min_val = *__current;
            __init.__min_ind = __i;
        }
        else if (!__comp(*__current, __max_val))
        {
            __init.__max_val = *__current;
            __init.__max_ind = __i;
        }
    }
    return std::make_pair(__first + __init.__min_ind, __first + __init.__max_ind);
}

template <class _InputIterator, class _DifferenceType, class _OutputIterator1, class _OutputIterator2,
          class _UnaryPredicate>
std::pair<_OutputIterator1, _OutputIterator2>
__simd_partition_copy(_InputIterator __first, _DifferenceType __n, _OutputIterator1 __out_true,
                      _OutputIterator2 __out_false, _UnaryPredicate __pred) noexcept
{
    _DifferenceType __cnt_true = 0, __cnt_false = 0;

    __PSTL_PRAGMA_SIMD
    for (_DifferenceType __i = 0; __i < __n; ++__i)
    {
        __PSTL_PRAGMA_SIMD_ORDERED_MONOTONIC_2ARGS(__cnt_true : 1, __cnt_false : 1)
        if (__pred(__first[__i]))
        {
            __out_true[__cnt_true] = __first[__i];
            ++__cnt_true;
        }
        else
        {
            __out_false[__cnt_false] = __first[__i];
            ++__cnt_false;
        }
    }
    return std::make_pair(__out_true + __cnt_true, __out_false + __cnt_false);
}

template <class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
_ForwardIterator1
__simd_find_first_of(_ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __s_first,
                     _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
{
    typedef typename std::iterator_traits<_ForwardIterator1>::difference_type _DifferencType;

    const _DifferencType __n1 = __last - __first;
    const _DifferencType __n2 = __s_last - __s_first;
    if (__n1 == 0 || __n2 == 0)
    {
        return __last; // according to the standard
    }

    // Common case
    // If first sequence larger than second then we'll run simd_first with parameters of first sequence.
    // Otherwise, vice versa.
    if (__n1 < __n2)
    {
        for (; __first != __last; ++__first)
        {
  	    if (__unseq_backend::__simd_or(__s_first, __n2,
                          __internal::__equal_value_by_pred<decltype(*__first), _BinaryPredicate>(*__first, __pred)))
            {
                return __first;
            }
        }
    }
    else
    {
        for (; __s_first != __s_last; ++__s_first)
        {
  	    const auto __result = __unseq_backend::__simd_first(__first, _DifferencType(0), __n1,
                                               [__s_first, &__pred](_ForwardIterator1 __it, _DifferencType __i) {
                                                   return __pred(__it[__i], *__s_first);
                                               });
            if (__result != __last)
            {
                return __result;
            }
        }
    }
    return __last;
}

template <class _RandomAccessIterator, class _DifferenceType, class _UnaryPredicate>
_RandomAccessIterator
__simd_remove_if(_RandomAccessIterator __first, _DifferenceType __n, _UnaryPredicate __pred) noexcept
{
    // find first element we need to remove
    auto __current =
        __unseq_backend::__simd_first(__first, _DifferenceType(0), __n,
                     [&__pred](_RandomAccessIterator __it, _DifferenceType __i) { return __pred(__it[__i]); });
    __n -= __current - __first;

    // if we have in sequence only one element that pred(__current[1]) != false we can exit the function
    if (__n < 2)
    {
        return __current;
    }

    _DifferenceType __cnt = 0;
    __PSTL_PRAGMA_SIMD
    for (_DifferenceType __i = 1; __i < __n; ++__i)
    {
        __PSTL_PRAGMA_SIMD_ORDERED_MONOTONIC(__cnt : 1)
        if (!__pred(__current[__i]))
        {
            __current[__cnt] = std::move(__current[__i]);
            ++__cnt;
        }
    }
    return __current + __cnt;
}
} // namespace __unseq_backend
} // namespace __pstl

#endif /* __PSTL_unseq_backend_simd_H */
@


1.1.1.1
log
@initial import of GCC 9.3.0.  changes include:

- live patching support
- shell completion help
- generally better diagnostic output (less verbose/more useful)
- diagnostics and optimisation choices can be emitted in json
- asan memory usage reduction
- many general, and specific to switch, inter-procedure,
  profile and link-time optimisations.  from the release notes:
  "Overall compile time of Firefox 66 and LibreOffice 6.2.3 on
  an 8-core machine was reduced by about 5% compared to GCC 8.3"
- OpenMP 5.0 support
- better spell-guesser
- partial experimental support for c2x and c++2a
- c++17 is no longer experimental
- arm AAPCS GCC 6-8 structure passing bug fixed, may cause
  incompatibility (restored compat with GCC 5 and earlier.)
- openrisc support
@
text
@@


1.1.1.2
log
@initial import of GCC 10.3.0.  main changes include:

caveats:
- ABI issue between c++14 and c++17 fixed
- profile mode is removed from libstdc++
- -fno-common is now the default

new features:
- new flags -fallocation-dce, -fprofile-partial-training,
  -fprofile-reproducible, -fprofile-prefix-path, and -fanalyzer
- many new compile and link time optimisations
- enhanced drive optimisations
- openacc 2.6 support
- openmp 5.0 features
- new warnings: -Wstring-compare and -Wzero-length-bounds
- extended warnings: -Warray-bounds, -Wformat-overflow,
  -Wrestrict, -Wreturn-local-addr, -Wstringop-overflow,
  -Warith-conversion, -Wmismatched-tags, and -Wredundant-tags
- some likely C2X features implemented
- more C++20 implemented
- many new arm & intel CPUs known

hundreds of reported bugs are fixed.  full list of changes
can be found at:

   https://gcc.gnu.org/gcc-10/changes.html
@
text
@d10 2
a11 2
#ifndef _PSTL_UNSEQ_BACKEND_SIMD_H
#define _PSTL_UNSEQ_BACKEND_SIMD_H
d31 1
a31 1
    _PSTL_PRAGMA_SIMD
d42 1
a42 1
    _PSTL_PRAGMA_SIMD
d53 1
a53 1
    _PSTL_PRAGMA_SIMD
d64 1
a64 1
#if _PSTL_EARLYEXIT_PRESENT
d66 2
a67 2
    _PSTL_PRAGMA_VECTOR_UNALIGNED
    _PSTL_PRAGMA_SIMD_EARLYEXIT
d78 1
a78 1
        _PSTL_PRAGMA_SIMD_REDUCTION(& : __flag)
d104 1
a104 1
#if _PSTL_EARLYEXIT_PRESENT
d106 2
a107 2
    _PSTL_PRAGMA_VECTOR_UNALIGNED // Do not generate peel loop part
        _PSTL_PRAGMA_SIMD_EARLYEXIT for (; __i < __end; ++__i)
d122 4
a125 4
        _PSTL_PRAGMA_VECTOR_UNALIGNED // Do not generate peel loop part
            _PSTL_PRAGMA_SIMD_REDUCTION(|
                                        : __found) for (_DifferenceType __i = __begin; __i < __begin + __block_size;
                                                        ++__i)
d157 1
a157 1
#endif //_PSTL_EARLYEXIT_PRESENT
d164 1
a164 1
#if _PSTL_EARLYEXIT_PRESENT
d166 2
a167 2
    _PSTL_PRAGMA_VECTOR_UNALIGNED
    _PSTL_PRAGMA_SIMD_EARLYEXIT
d182 2
a183 2
        _PSTL_PRAGMA_VECTOR_UNALIGNED // Do not generate peel loop part
            _PSTL_PRAGMA_SIMD_REDUCTION(|
d211 1
a211 1
#endif //_PSTL_EARLYEXIT_PRESENT
d219 1
a219 1
    _PSTL_PRAGMA_SIMD_REDUCTION(+ : __count)
d238 1
a238 1
    _PSTL_PRAGMA_SIMD
d241 1
a241 1
        _PSTL_PRAGMA_SIMD_ORDERED_MONOTONIC(__cnt : 1)
d255 2
a256 2
    _PSTL_USE_NONTEMPORAL_STORES_IF_ALLOWED
    _PSTL_PRAGMA_SIMD
d268 1
a268 1
    _PSTL_PRAGMA_SIMD
d271 1
a271 1
        _PSTL_PRAGMA_SIMD_ORDERED_MONOTONIC(__cnt : 1)
d287 1
a287 1
    _PSTL_PRAGMA_SIMD_REDUCTION(+ : __count)
d302 1
a302 1
    _PSTL_PRAGMA_SIMD_REDUCTION(+ : __count)
d317 1
a317 1
    _PSTL_PRAGMA_SIMD
d322 1
a322 1
            _PSTL_PRAGMA_SIMD_ORDERED_MONOTONIC(__cnt : 1)
d337 1
a337 1
    _PSTL_PRAGMA_SIMD
d340 1
a340 1
        _PSTL_PRAGMA_SIMD_ORDERED_MONOTONIC_2ARGS(__cnt_true : 1, __cnt_false : 1)
d358 2
a359 2
    _PSTL_USE_NONTEMPORAL_STORES_IF_ALLOWED
    _PSTL_PRAGMA_SIMD
d369 2
a370 2
    _PSTL_USE_NONTEMPORAL_STORES_IF_ALLOWED
    _PSTL_PRAGMA_SIMD
d386 1
a386 1
#if _PSTL_EARLYEXIT_PRESENT
d389 2
a390 2
    _PSTL_PRAGMA_VECTOR_UNALIGNED
    _PSTL_PRAGMA_SIMD_EARLYEXIT
d404 2
a405 2
        _PSTL_PRAGMA_VECTOR_UNALIGNED // Do not generate peel loop part
            _PSTL_PRAGMA_SIMD_REDUCTION(|
d449 1
a449 1
    _PSTL_PRAGMA_SIMD_REDUCTION(+ : __init)
d466 1
a466 1
        _PSTL_PRAGMA_SIMD
d476 1
a476 1
            _PSTL_PRAGMA_SIMD
d483 1
a483 1
        _PSTL_PRAGMA_SIMD
d494 1
a494 1
        _PSTL_PRAGMA_SIMD
d517 1
a517 1
    _PSTL_PRAGMA_SIMD_SCAN(+ : __init)
d521 1
a521 1
        _PSTL_PRAGMA_SIMD_EXCLUSIVE_SCAN(__init)
d555 1
a555 1
    _PSTL_PRAGMA_DECLARE_REDUCTION(__bin_op, _CombinerType)
d557 1
a557 1
    _PSTL_PRAGMA_SIMD_SCAN(__bin_op : __init_)
d561 2
a562 2
        _PSTL_PRAGMA_SIMD_EXCLUSIVE_SCAN(__init_)
        _PSTL_PRAGMA_FORCEINLINE
d575 1
a575 1
    _PSTL_PRAGMA_SIMD_SCAN(+ : __init)
d579 1
a579 1
        _PSTL_PRAGMA_SIMD_INCLUSIVE_SCAN(__init)
d595 1
a595 1
    _PSTL_PRAGMA_DECLARE_REDUCTION(__bin_op, _CombinerType)
d597 1
a597 1
    _PSTL_PRAGMA_SIMD_SCAN(__bin_op : __init_)
d600 1
a600 1
        _PSTL_PRAGMA_FORCEINLINE
d602 1
a602 1
        _PSTL_PRAGMA_SIMD_INCLUSIVE_SCAN(__init_)
d636 1
a636 1
        _PSTL_PRAGMA_DECLARE_SIMD
d651 1
a651 1
    _PSTL_PRAGMA_DECLARE_REDUCTION(__min_func, _ComplexType)
d653 1
a653 1
    _PSTL_PRAGMA_SIMD_REDUCTION(__min_func : __init)
d730 1
a730 1
    _PSTL_PRAGMA_DECLARE_REDUCTION(__min_func, _ComplexType);
d732 1
a732 1
    _PSTL_PRAGMA_SIMD_REDUCTION(__min_func : __init)
d760 1
a760 1
    _PSTL_PRAGMA_SIMD
d763 1
a763 1
        _PSTL_PRAGMA_SIMD_ORDERED_MONOTONIC_2ARGS(__cnt_true : 1, __cnt_false : 1)
d840 1
a840 1
    _PSTL_PRAGMA_SIMD
d843 1
a843 1
        _PSTL_PRAGMA_SIMD_ORDERED_MONOTONIC(__cnt : 1)
d855 1
a855 1
#endif /* _PSTL_UNSEQ_BACKEND_SIMD_H */
@


1.1.1.3
log
@initial import of GCC 12.3.0.

major changes in GCC 11 included:

- The default mode for C++ is now -std=gnu++17 instead of -std=gnu++14.
- When building GCC itself, the host compiler must now support C++11,
  rather than C++98.
- Some short options of the gcov tool have been renamed: -i to -j and
  -j to -H.
- ThreadSanitizer improvements.
- Introduce Hardware-assisted AddressSanitizer support.
- For targets that produce DWARF debugging information GCC now defaults
  to DWARF version 5. This can produce up to 25% more compact debug
  information compared to earlier versions.
- Many optimisations.
- The existing malloc attribute has been extended so that it can be
  used to identify allocator/deallocator API pairs. A pair of new
  -Wmismatched-dealloc and -Wmismatched-new-delete warnings are added.
- Other new warnings:
  -Wsizeof-array-div, enabled by -Wall, warns about divisions of two
    sizeof operators when the first one is applied to an array and the
    divisor does not equal the size of the array element.
  -Wstringop-overread, enabled by default, warns about calls to string
    functions reading past the end of the arrays passed to them as
    arguments.
  -Wtsan, enabled by default, warns about unsupported features in
    ThreadSanitizer (currently std::atomic_thread_fence).
- Enchanced warnings:
  -Wfree-nonheap-object detects many more instances of calls to
    deallocation functions with pointers not returned from a dynamic
    memory allocation function.
  -Wmaybe-uninitialized diagnoses passing pointers or references to
    uninitialized memory to functions taking const-qualified arguments.
  -Wuninitialized detects reads from uninitialized dynamically
    allocated memory.
  -Warray-parameter warns about functions with inconsistent array forms.
  -Wvla-parameter warns about functions with inconsistent VLA forms.
- Several new features from the upcoming C2X revision of the ISO C
  standard are supported with -std=c2x and -std=gnu2x.
- Several C++20 features have been implemented.
- The C++ front end has experimental support for some of the upcoming
  C++23 draft.
- Several new C++ warnings.
- Enhanced Arm, AArch64, x86, and RISC-V CPU support.
- The implementation of how program state is tracked within
  -fanalyzer has been completely rewritten with many enhancements.

see https://gcc.gnu.org/gcc-11/changes.html for a full list.

major changes in GCC 12 include:

- An ABI incompatibility between C and C++ when passing or returning
  by value certain aggregates containing zero width bit-fields has
  been discovered on various targets. x86-64, ARM and AArch64
  will always ignore them (so there is a C ABI incompatibility
  between GCC 11 and earlier with GCC 12 or later), PowerPC64 ELFv2
  always take them into account (so there is a C++ ABI
  incompatibility, GCC 4.4 and earlier compatible with GCC 12 or
  later, incompatible with GCC 4.5 through GCC 11). RISC-V has
  changed the handling of these already starting with GCC 10. As
  the ABI requires, MIPS takes them into account handling function
  return values so there is a C++ ABI incompatibility with GCC 4.5
  through 11.
- STABS: Support for emitting the STABS debugging format is
  deprecated and will be removed in the next release. All ports now
  default to emit DWARF (version 2 or later) debugging info or are
  obsoleted.
- Vectorization is enabled at -O2 which is now equivalent to the
  original -O2 -ftree-vectorize -fvect-cost-model=very-cheap.
- GCC now supports the ShadowCallStack sanitizer.
- Support for __builtin_shufflevector compatible with the clang
  language extension was added.
- Support for attribute unavailable was added.
- Support for __builtin_dynamic_object_size compatible with the
  clang language extension was added.
- New warnings:
  -Wbidi-chars warns about potentially misleading UTF-8
    bidirectional control characters.
  -Warray-compare warns about comparisons between two operands of
    array type.
- Some new features from the upcoming C2X revision of the ISO C
  standard are supported with -std=c2x and -std=gnu2x.
- Several C++23 features have been implemented.
- Many C++ enhancements across warnings and -f options.

see https://gcc.gnu.org/gcc-12/changes.html for a full list.
@
text
@d184 1
a184 1
                                        : __found) for (__i = 0; __i < __block_size; ++__i)
d192 1
a192 1
            _DifferenceType __i2;
d194 1
a194 1
            for (__i2 = 0; __i2 < __block_size; ++__i2)
d196 1
a196 1
                if (__lane[__i2])
d199 1
a199 1
            return std::make_pair(__first1 + __i2, __first2 + __i2);
d406 1
a406 1
                                        : __found) for (__i = 0; __i < __block_size - 1; ++__i)
d489 1
a489 1
        for (_Size __j = 0; __j < __block_size; ++__j)
d491 1
a491 1
            __init = __binary_op(__init, __lane[__j]);
d495 1
a495 1
        for (_Size __j = 0; __j < __block_size; ++__j)
d497 1
a497 1
            __lane[__j].~_Tp();
d799 2
a800 3
            if (__unseq_backend::__simd_or(
                    __s_first, __n2,
                    __internal::__equal_value_by_pred<decltype(*__first), _BinaryPredicate>(*__first, __pred)))
d810 4
a813 4
            const auto __result = __unseq_backend::__simd_first(
                __first, _DifferencType(0), __n1, [__s_first, &__pred](_ForwardIterator1 __it, _DifferencType __i) {
                    return __pred(__it[__i], *__s_first);
                });
d828 3
a830 3
    auto __current = __unseq_backend::__simd_first(
        __first, _DifferenceType(0), __n,
        [&__pred](_RandomAccessIterator __it, _DifferenceType __i) { return __pred(__it[__i]); });
@


1.1.1.4
log
@initial import of GCC 14.3.0.

major changes in GCC 13:
- improved sanitizer
- zstd debug info compression
- LTO improvements
- SARIF based diagnostic support
- new warnings: -Wxor-used-as-pow, -Wenum-int-mismatch, -Wself-move,
  -Wdangling-reference
- many new -Wanalyzer* specific warnings
- enhanced warnings: -Wpessimizing-move, -Wredundant-move
- new attributes to mark file descriptors, c++23 "assume"
- several C23 features added
- several C++23 features added
- many new features for Arm, x86, RISC-V

major changes in GCC 14:
- more strict C99 or newer support
- ia64* marked deprecated (but seemingly still in GCC 15.)
- several new hardening features
- support for "hardbool", which can have user supplied values of true/false
- explicit support for stack scrubbing upon function exit
- better auto-vectorisation support
- added clang-compatible __has_feature and __has_extension
- more C23, including -std=c23
- several C++26 features added
- better diagnostics in C++ templates
- new warnings: -Wnrvo, Welaborated-enum-base
- many new features for Arm, x86, RISC-V
- possible ABI breaking change for SPARC64 and small structures with arrays
  of floats.
@
text
@d64 1
a64 1
#if defined(_PSTL_EARLYEXIT_PRESENT)
d77 1
a77 1
        __INT32_TYPE__ __flag = 1;
d104 1
a104 1
#if defined(_PSTL_EARLYEXIT_PRESENT)
d164 1
a164 1
#if defined(_PSTL_EARLYEXIT_PRESENT)
d386 1
a386 1
#if defined(_PSTL_EARLYEXIT_PRESENT)
d535 1
a535 1
    _Combiner(const _Tp& __v, const _BinaryOp* __b) : __value(__v), __bin_op(const_cast<_BinaryOp*>(__b)) {}
d688 3
a690 3
        _ComplexType(const _ValueType& __min, const _ValueType& __max, const _Compare* __comp)
            : __min_val(__min), __max_val(__max), __min_ind(0), __max_ind(0),
              __minmax_comp(const_cast<_Compare*>(__comp))
@