Re: [PATCH] libstdc++: optimize std::for_each for segmented iterators

On Mon, 3 Aug 2026, Yuao Ma wrote:

> On Mon, Aug 3, 2026 at 9:22 PM Patrick Palka <[email protected]> wrote:
> >
> > On Sun, 2 Aug 2026, Yuao Ma wrote:
> >
> > > Hi!
> > >
> > > Similar to ranges::distance, this patch optimizes std::for_each for
> > > segmented iterators.
> >
> > Thanks for working on this.
> >
> > >
> > > Fully tested on x86_64-linux with no regressions.
> > >
> > > Using the newly added benchmark, it shows a 3x improvement when using
> > > std::for_each with std::deque.
> > >
> > > === Sun Aug  2 09:30:52 AM UTC 2026 ===
> > > for_each.cc              std::for_each vector<int>    2r    1u    0s
> > >       0mem    0pf
> > > for_each.cc              std::for_each deque<int>    6r    6u    0s
> > >      0mem    0pf
> > > for_each.cc              std::for_each list<int>    13r   13u    0s
> > >      0mem    0pf
> > > === Sun Aug  2 09:31:08 AM UTC 2026 ===
> > > for_each.cc              std::for_each vector<int>    2r    1u    0s
> > >       0mem    0pf
> > > for_each.cc              std::for_each deque<int>    2r    2u    0s
> > >      0mem    0pf
> > > for_each.cc              std::for_each list<int>    13r   13u    0s
> > >      0mem    0pf
> >
> > Interesting that even for_each benefits from this!  So this assumes
> > it's significantly cheaper to iterate over the inner/local iterators
> > than over the segmented iterators in general, makes sense.
> >
> > >
> > > Please take a look when you are available, thanks!
> > >
> > > Yuao
> > >
> >
> > > Subject: [PATCH] libstdc++: optimize std::for_each for segmented iterators
> > >
> > > Similar to r17-2859-g15505d012872dd, we can optimize std::for_each for
> > > segemented iterators.
> > >
> > > libstdc++-v3/ChangeLog:
> > >
> > >       * include/bits/stl_algo.h (__for_each): Add segemented
> > >       iterators logic. Split out naive for-loop from ...
> > >       (for_each): ... here.
> > >       * testsuite/performance/25_algorithms/for_each.cc: New test.
> > > ---
> > >  libstdc++-v3/include/bits/stl_algo.h          | 29 ++++++++++++++--
> > >  .../performance/25_algorithms/for_each.cc     | 34 +++++++++++++++++++
> > >  2 files changed, 61 insertions(+), 2 deletions(-)
> > >  create mode 100644 libstdc++-v3/testsuite/performance/25_algorithms/for_each.cc
> > >
> > > diff --git a/libstdc++-v3/include/bits/stl_algo.h b/libstdc++-v3/include/bits/stl_algo.h
> > > index 800c176cd5b..54d73ae8c62 100644
> > > --- a/libstdc++-v3/include/bits/stl_algo.h
> > > +++ b/libstdc++-v3/include/bits/stl_algo.h
> > > @@ -132,6 +132,32 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
> > >        return __first;
> > >      }
> > >
> > > +  /// Apply __f to each element in [__first, __last)
> > > +  /// Dispatches to __for_each_segment for segmented iterators
> > > +  /// (e.g. deque::iterator).
> > > +  /// Returns an iterator equal to __last.
> > > +  template<typename _InputIterator, typename _Function>
> > > +    _GLIBCXX20_CONSTEXPR
> > > +    _InputIterator
> > > +    __for_each(_InputIterator __first, _InputIterator __last, _Function& __f)
> > > +    {
> > > +#if __cplusplus >= 201703L
> > > +      if constexpr (__enable_for_each_segment<_InputIterator>)
> > > +     {
> > > +       std::__for_each_segment(__first, __last,
> > > +         [&](auto __lfirst, auto __llast)
> > > +         { return std::__for_each(__lfirst, __llast, __f); });
> > > +       return __last;
> > > +     }
> > > +      else
> > > +#endif // C++17
> >
> > I think we could enable this optimization as far back as C++11 if we
> > use a lambda template here (and #pragma GCC diagnostic ignored etc):
> >
> >   [&]<template _It>(_It __lfirst, _It __llast)
> >
> 
> Thanks for the suggestions. I use this approach in the new patch.
> 
> > Besides that, looks good.
> >
> > (GCC and Clang accept lambda templates even in C++98 mode so in theory
> > we could enable the optimization unconditionally if we resort to using
> > C++98 compatible SFINAE, but I don't think it's worth it personally.)
> >
> > > +     {
> > > +       for (; __first != __last; ++__first)
> > > +         __f(*__first);
> > > +       return __first;
> > > +     }
> > > +    }
> > > +
> > >    // set_difference
> > >    // set_intersection
> > >    // set_symmetric_difference
> > > @@ -3813,8 +3839,7 @@ _GLIBCXX_BEGIN_NAMESPACE_ALGO
> > >        // concept requirements
> > >        __glibcxx_function_requires(_InputIteratorConcept<_InputIterator>)
> > >        __glibcxx_requires_valid_range(__first, __last);
> > > -      for (; __first != __last; ++__first)
> > > -     __f(*__first);
> > > +      std::__for_each(__first, __last, __f);
> >
> > I guess you're factoring this out to use elsewhere, e.g. ranges::for_each?
> >
> 
> Hmm, it is mostly for support the stateful functor passed in the
> std::for_each, in which case we really need the parameter to be a
> reference type.

Ah good point. Alternatively I think we could pass std::ref(__f)
in the recursive for_each calls.  I'm not sure which approach is
preferable.

> To use it for ranges::for_each at least it needs two
> refactor: for function parameter it needs _Sentinel and _Proj, for the
> function body it needs std::invoke. Not sure how to correctly includes
> these given the current header structure, I think I will have a
> standalone optimization for ranges::for_each, and let other algorithms
> that could benefit from segmented iterators depend on either of them.

Makes sense.

> 
> > >        return __f; // N.B. [alg.foreach] says std::move(f) but it's redundant.
> > >      }
> > >
> > > diff --git a/libstdc++-v3/testsuite/performance/25_algorithms/for_each.cc b/libstdc++-v3/testsuite/performance/25_algorithms/for_each.cc
> > > new file mode 100644
> > > index 00000000000..22981eb9b64
> > > --- /dev/null
> > > +++ b/libstdc++-v3/testsuite/performance/25_algorithms/for_each.cc
> > > @@ -0,0 +1,34 @@
> > > +// STD=gnu++17
> > > +
> > > +#include <testsuite_performance.h>
> > > +
> > > +#include <algorithm>
> > > +#include <deque>
> > > +#include <list>
> > > +#include <vector>
> > > +
> > > +const std::size_t size = 8192;
> > > +
> > > +template <typename Container>
> > > +void bench_seq(const char* label, __gnu_test::time_counter& time,
> > > +               __gnu_test::resource_counter& resource) {
> > > +  using T = typename Container::value_type;
> > > +  Container c(size, 1);
> > > +  start_counters(time, resource);
> > > +  for (int i = 0; i < 20000; ++i)
> > > +    std::for_each(c.begin(), c.end(),
> > > +                  [](T& x) { x = std::clamp<T>(x, 10, 100); });
> > > +  stop_counters(time, resource);
> > > +  report_performance(__FILE__, label, time, resource);
> > > +  clear_counters(time, resource);
> > > +}
> > > +
> > > +int main() {
> > > +  using namespace __gnu_test;
> > > +  time_counter time;
> > > +  resource_counter resource;
> > > +
> > > +  bench_seq<std::vector<int>>("std::for_each vector<int>", time, resource);
> > > +  bench_seq<std::deque<int>>("std::for_each deque<int>", time, resource);
> > > +  bench_seq<std::list<int>>("std::for_each list<int>", time, resource);
> > > +}
> > > --
> > > 2.54.0
> > >
> >
>