4249. The past end issue for lazy_split_view

Section: 25.7.16.3 [range.lazy.split.outer] Status: New Submitter: Hewill Kang Opened: 2025-04-26 Last modified: 2025-04-27

Priority: Not Prioritized

View other active issues in [range.lazy.split.outer].

View all other issues in [range.lazy.split.outer].

View all issues with New status.

Discussion:

Consider (demo):

#include <print>
#include <ranges>
#include <sstream>

int main() {
  std::istringstream is{"1 0 2 0 3"};
  auto r = std::views::istream<int>(is)
         | std::views::lazy_split(0)
         | std::views::stride(2);
  std::println("{}", r); // should print [[1], [3]]
}

The above leads to SIGSEGV in libstdc++, the reason is that we are iterating over the nested range as:

for (auto&& inner : r) {
  for (auto&& elem : inner) {
    // […]
  }
}

which is disassembled as:

auto outer_it = r.begin();
std::default_sentinel_t out_end = r.end();
for(; outer_it != out_end; ++outer_it) {
  auto&& inner_r = *outer_it;
  auto inner_it = inner_r.begin();
  std::default_sentinel_t inner_end = inner_r.end();
  for(; inner_it != inner_end; ++inner_it) {
    auto&& elem = *inner_it;
    // […]
  }
}

Since inner_it and output_it actually update the same iterator, when we back to the outer loop, lazy_split_view::outer-iterator is now equal to default_sentinel, which makes output_it reach the end, so ++outer_it will increment the iterator past end, triggering the assertion.

Note that this also happens in MSVC-STL when _ITERATOR_DEBUG_LEVEL is turned on.

It seems that extra flags are needed to fix this issue because output_it should not be considered to reach the end when we back to the outer loop.

Proposed resolution:

This wording is relative to N5008.

1. Modify 25.7.16.3 [range.lazy.split.outer] as indicated:

   namespace std::ranges {
     template<input_range V, forward_range Pattern>
       requires view<V> && view<Pattern> &&
                indirectly_comparable<iterator_t<V>, iterator_t<Pattern>, ranges::equal_to> &&
                (forward_range<V> || tiny-range<Pattern>)
     template<bool Const>
     struct lazy_split_view<V, Pattern>::outer-iterator {
     private:
       using Parent = maybe-const<Const, lazy_split_view>;     // exposition only
       using Base = maybe-const<Const, V>;                     // exposition only
       Parent* parent_ = nullptr;                              // exposition only
   
       iterator_t<Base> current_ = iterator_t<Base>();         // exposition only, present only
                                                               // if V models forward_range
   
       bool trailing_empty_ = false;                           // exposition only
       bool has_next_ = false;                                 // exposition only, present only
                                                               // if forward_range<V> is false
     public:
       […]
     };
   }
   

   […]

   constexpr explicit outer-iterator(Parent& parent)
     requires (!forward_range<Base>);
   

   -2- Effects: Initializes parent_ with addressof(parent) and has_next_ with current != ranges::end(parent_->base_).

   […]

   constexpr outer-iterator& operator++();
   

   -6- Effects: Equivalent to:

   const auto end = ranges::end(parent_->base_);
   if (current == end) {
     trailing_empty_ = false;
     if constexpr (!forward_range<V>)
       has_next_ = false;
     return *this;
   }
   const auto [pbegin, pend] = subrange{parent_->pattern_};
   if (pbegin == pend) ++current;
   else if constexpr (tiny-range<Pattern>) {
     current = ranges::find(std::move(current), end, *pbegin);
     if (current != end) {
       ++current;
       if (current == end)
         trailing_empty_ = true;
     }
   }
   else {
     do {
       auto [b, p] = ranges::mismatch(current, end, pbegin, pend);
       if (p == pend) {
         current = b;
         if (current == end)
           trailing_empty_ = true;
         break;            // The pattern matched; skip it
       }
     } while (++current != end);
   }
   if constexpr (!forward_range<V>)
     if (current == end)
       has_next_ = false;
   return *this;
   

   […]

   friend constexpr bool operator==(const outer-iterator& x, default_sentinel_t);
   

   -8- Effects: Equivalent to:

   if constexpr (!forward_range<V>)
     return !x.has_next_ && !x.trailing_empty_;
   else
     return x.current == ranges::end(x.parent_->base_) && !x.trailing_empty_;