Should be addressed after merging #2960, not before.

We're currently inconsistent about whether parallel algorithms use _REQUIRE_PARALLEL_ITERATOR (for reading) and _REQUIRE_CPP17_MUTABLE_ITERATOR (for writing).

The following overloads (found by searching for all occurrences of _ExPo&&) are missing enforcement. Some are just calling the classic serial algorithms, others are doing real parallel work:

• <algorithm>:
  • swap_ranges(_ExPo&&, _FwdIt1 _First1, _FwdIt1 _Last1, _FwdIt2 _Dest)
  • generate(_ExPo&&, _FwdIt _First, _FwdIt _Last, _Fn _Func)
  • unique(_ExPo&&, _FwdIt _First, _FwdIt _Last, _Pr _Pred)
  • unique(_ExPo&&, _FwdIt _First, _FwdIt _Last)
  • shift_left(_ExPo&&, _FwdIt _First, _FwdIt _Last, _Iter_diff_t<_FwdIt> _Pos_to_shift)
  • shift_right(_ExPo&&, _FwdIt _First, _FwdIt _Last, _Iter_diff_t<_FwdIt> _Pos_to_shift)
  • stable_partition(_ExPo&&, _BidIt _First, _BidIt _Last, _Pr _Pred)
  • inplace_merge(_ExPo&&, _BidIt _First, _BidIt _Mid, _BidIt _Last, _Pr _Pred)
  • inplace_merge(_ExPo&&, _BidIt _First, _BidIt _Mid, _BidIt _Last)
  • partial_sort(_ExPo&&, _RanIt _First, _RanIt _Mid, _RanIt _Last, _Pr _Pred)
  • partial_sort(_ExPo&&, _RanIt _First, _RanIt _Mid, _RanIt _Last)
  • nth_element(_ExPo&&, _RanIt _First, _RanIt _Nth, _RanIt _Last, _Pr _Pred)
  • nth_element(_ExPo&&, _RanIt _First, _RanIt _Nth, _RanIt _Last)
  • max_element(_ExPo&&, _FwdIt _First, _FwdIt _Last, _Pr _Pred)
  • max_element(_ExPo&&, _FwdIt _First, _FwdIt _Last)
  • min_element(_ExPo&&, _FwdIt _First, _FwdIt _Last, _Pr _Pred)
  • min_element(_ExPo&&, _FwdIt _First, _FwdIt _Last)
  • minmax_element(_ExPo&&, _FwdIt _First, _FwdIt _Last, _Pr _Pred)
  • minmax_element(_ExPo&&, _FwdIt _First, _FwdIt _Last)
• <execution>:
  • adjacent_find(_ExPo&&, _FwdIt _First, _FwdIt _Last, _Pr _Pred)
  • search_n(_ExPo&&, const _FwdIt _First, _FwdIt _Last, const _Diff _Count_raw, const _Ty& _Val, _Pr _Pred)
  • remove_if(_ExPo&&, _FwdIt _First, const _FwdIt _Last, _Pr _Pred)
  • sort(_ExPo&&, const _RanIt _First, const _RanIt _Last, _Pr _Pred)
  • stable_sort(_ExPo&&, const _BidIt _First, const _BidIt _Last, _Pr _Pred)
  • partition(_ExPo&&, _FwdIt _First, const _FwdIt _Last, _Pr _Pred)
• <xutility>:
  • reverse(_ExPo&&, _BidIt _First, _BidIt _Last)
  • rotate(_ExPo&&, _FwdIt _First, _FwdIt _Mid, _FwdIt _Last)

Additionally (although less concerning), there are some _ExPo&& overloads that check the requirements and then call another _ExPo&& overload. This is inconsistently done (more often, it seems that we directly call another _ExPo&& overload and expect it to perform checking).

I think that a reasonable policy would be: all _ExPo&& overloads (of publicly usable functions) should immediately perform checking, whether they're calling classic serial or doing real work, unless they are immediately calling another _ExPo&& overload (typically to provide defaulted comparators etc).