https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88545

Jonathan Wakely <redi at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization

--- Comment #2 from Jonathan Wakely <redi at gcc dot gnu.org> ---
(In reply to Georg Sauthoff from comment #0)
> Created attachment 45259 [details]
> specialize std::find to memchr for character searches in continous memory
> 
> If std::find() is called with continuous random access iterators and a
> trivial char sized value, then calling memchr() is much more efficient than
> calling into the generic __find_if().
> 
> The attached patch implements this optimization.
> 
> That means it specializes a std::find helper on the iterator category and
> the value and calls __builtin_memchr() if possible.

Why specialize on the iterator category, when the __is_simple boolean already
checks if the iterator is a pointer?

The condition of a trivial byte-sized type seem insufficient, because you could
have:

struct B {
  char c;
  bool operator==(const B& b) const { return true; }
};



I would prefer to do simply:

--- a/libstdc++-v3/include/bits/stl_algo.h
+++ b/libstdc++-v3/include/bits/stl_algo.h
@@ -3846,6 +3846,32 @@ _GLIBCXX_BEGIN_NAMESPACE_ALGO
       __glibcxx_function_requires(_EqualOpConcept<
                typename iterator_traits<_InputIterator>::value_type, _Tp>)
       __glibcxx_requires_valid_range(__first, __last);
+
+#if __cpp_if_constexpr
+      using _ValT = typename iterator_traits<_InputIterator>::value_type;
+      if constexpr (is_same_v<_ValT, _Tp>)
+       if constexpr (__is_byte<_ValT>::__value)
+#if __cpp_lib_concepts
+         if constexpr (contiguous_iterator<_InputIterator>)
+           {
+             if (const size_t __n = __last - __first)
+               {
+                 auto __p0 = std::to_address(__first);
+                 if (auto __p1 = __builtin_memchr(__p0, __val, __n))
+                   return __first + (__p1 - __p0);
+               }
+             return __last;
+           }
+#else
+         if constexpr (is_pointer_v<_InputIterator>)
+           {
+             if (const size_t __n = __last - __first)
+               if (auto __p = __builtin_memchr(__first, __val, __n))
+                 return __p;
+             return __last;
+           }
+#endif
+#endif
       return std::__find_if(__first, __last,
                            __gnu_cxx::__ops::__iter_equals_val(__val));
     }


I think we're going to remove the manual loop unrolling in __find_if for GCC
15, which should allow the compiler to optimize it better, potentially
auto-vectorizing. That might make memchr less advantageous, but I think it's
worth doing anyway.

Reply via email to