Unify the "regular" and "bounded" implementations for copy_if

[akukanov]

On top of #2499, this patch unifies the copy_if implementation for parallel and vector policies. combining much of the code previously separate for the "bounded" ranges::copy_if and the "regular" C++17 copy_if.

• Support bounded output in __simd_copy_by_mask
• Simplify __brick_bounded_copy_by_mask and make it stop according to the ranges::copy_if semantics
• Subsume __brick_copy_by_mask into __brick_bounded_copy_by_mask
• Support bounded output in __parallel_selective_copy
• Make __pattern_bounded_copy_if "redirecting" to __parallel_selective_copy

[SergeyKopienko]

I think the declaration of

template <bool, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Bound, class _Assigner>
_Bound
__brick_copy_by_mask(_RandomAccessIterator1, _Bound, _RandomAccessIterator2, _Bound, bool*, _Assigner,
                     /*vector=*/std::false_type) noexcept;

template <bool, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Bound, class _Assigner>
_Bound
__brick_copy_by_mask(_RandomAccessIterator1, _Bound, _RandomAccessIterator2, _Bound, bool*, _Assigner,
                     /*vector=*/std::true_type) noexcept;

not aligned with the implementations:

template <bool __Bounded, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Bound, class _Assigner>
_Bound
__brick_copy_by_mask(_RandomAccessIterator1 __first, _Bound __in_len, _RandomAccessIterator2 __result, _Bound __out_len,
                     bool* __mask, _Assigner __assigner, /*vector=*/std::false_type) noexcept
{
   // ,,,
}

template <bool __Bounded, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Bound, class _Assigner>
_Bound
__brick_copy_by_mask(_RandomAccessIterator1 __first, _Bound __in_len, _RandomAccessIterator2 __result, _Bound __out_len,
                     bool* __mask, _Assigner __assigner, /*vector=*/std::true_type) noexcept
{
    return __unseq_backend::__simd_copy_by_mask<__Bounded>(__first, __in_len, __result, __out_len, __mask, __assigner);
}

[SergeyKopienko]

One more question.
For what we have iterator and size in

template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _DifferenceType,
          class _RandomAccessIterator2, class _UnaryPredicate>
std::pair<_RandomAccessIterator1, _RandomAccessIterator2>
__pattern_bounded_copy_if(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _DifferenceType,
                          _RandomAccessIterator2, _DifferenceType, _UnaryPredicate);

Why we can't use two iterators (begin and end) ?
As I tried to find this is the first example of this approach on our pattern's level.

[akukanov]

I think the declaration of ... __brick_copy_by_mask ... not aligned with the implementations

I have checked and have not found any misalignment, aside from omitted names in the forward declarations.

For what we have iterator and size in ... __pattern_bounded_copy_if Why we can't use two iterators (begin and end) ?

The only caller of __pattern_bounded_copy_if has already computed the sizes, which the callee needs. Why would we re-compute them again?

As I tried to find this is the first example of this approach on our pattern's level.

Most (if not all) other __pattern_ functions have two overloads - for serial and parallel implementations. The serial variant does not require random access iterators and therefore takes iterator pairs; the parallel variant has the same API and computes sizes internally. However __pattern_bounded_copy_if does not need a serial overload, and so can have a different API.

[SergeyKopienko]

LGTM

[MikeDvorskiy]

LGTM