[libcxx-commits] [libcxx] [libc++] Optimize lexicographical_compare (PR #65279)

[Louis Dionne via libcxx-commits]

================
@@ -11,39 +11,96 @@
 
 #include <__algorithm/comp.h>
 #include <__algorithm/comp_ref_type.h>
+#include <__algorithm/max.h>
+#include <__algorithm/mismatch.h>
+#include <__algorithm/simd_utils.h>
 #include <__config>
+#include <__functional/identity.h>
 #include <__iterator/iterator_traits.h>
+#include <__string/constexpr_c_functions.h>
+#include <__type_traits/desugars_to.h>
+#include <__type_traits/invoke.h>
+#include <__type_traits/is_equality_comparable.h>
+#include <__type_traits/is_integral.h>
+#include <__type_traits/is_volatile.h>
+#include <cstring>
+#include <cwchar>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header
 #endif
 
 _LIBCPP_BEGIN_NAMESPACE_STD
 
-template <class _Compare, class _InputIterator1, class _InputIterator2>
+template <class _Iter1, class _Sent1, class _Iter2, class _Sent2, class _Proj1, class _Proj2, class _Comp>
 _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 bool __lexicographical_compare(
-    _InputIterator1 __first1,
-    _InputIterator1 __last1,
-    _InputIterator2 __first2,
-    _InputIterator2 __last2,
-    _Compare __comp) {
-  for (; __first2 != __last2; ++__first1, (void)++__first2) {
-    if (__first1 == __last1 || __comp(*__first1, *__first2))
+    _Iter1 __first1, _Sent1 __last1, _Iter2 __first2, _Sent2 __last2, _Comp& __comp, _Proj1& __proj1, _Proj2& __proj2) {
+  while (__first2 != __last2) {
+    if (__first1 == __last1 ||
+        std::__invoke(__comp, std::__invoke(__proj1, *__first1), std::__invoke(__proj2, *__first2)))
       return true;
-    if (__comp(*__first2, *__first1))
+    if (std::__invoke(__comp, std::__invoke(__proj2, *__first2), std::__invoke(__proj1, *__first1)))
       return false;
+    ++__first1;
+    ++__first2;
   }
   return false;
 }
 
+#if _LIBCPP_VECTORIZE_ALGORITHMS
----------------
ldionne wrote:

It would make sense to only guard the place where you call `mismatch` with this. Along with a comment saying e.g. "the naive implementation is faster than calling mismatch when we don't vectorize mismatch".

https://github.com/llvm/llvm-project/pull/65279