Optimize swap_ranges for contigious, replaceable ranges

cor3ntin · Oct 7, 2024 · 1eb88fc · 1eb88fc
1 parent 4c4b7b0
commit 1eb88fc
Showing 1 changed file with 42 additions and 0 deletions.
diff --git a/libcxx/include/__algorithm/swap_ranges.h b/libcxx/include/__algorithm/swap_ranges.h
@@ -40,6 +40,48 @@ __swap_ranges(_ForwardIterator1 __first1, _Sentinel1 __last1, _ForwardIterator2
 template <class _AlgPolicy, class _ForwardIterator1, class _Sentinel1, class _ForwardIterator2>
 _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 pair<_ForwardIterator1, _ForwardIterator2>
 __swap_ranges(_ForwardIterator1 __first1, _Sentinel1 __last1, _ForwardIterator2 __first2) {
+#if _LIBCPP_STD_VER >= 26
+  if ! consteval {
+    using _V1Type = iterator_traits<_ForwardIterator1>::value_type;
+    using _V2Type = iterator_traits<_ForwardIterator2>::value_type;
+    if constexpr(__libcpp_is_contiguous_iterator<_ForwardIterator1>::value &&
+                 __libcpp_is_contiguous_iterator<_ForwardIterator2>::value &&
+                 is_same_v<_V1Type, _V2Type> && is_replaceable_v<_V1Type>) {
+      size_t __distance = distance(__first1, __last1);
+      if (__distance == 1) {
+        using std::swap;
+        swap(*__first1, *__first2);
+        }
+      else if (__distance > 0) {
+        size_t __numBytes = __distance * sizeof(_V1Type);
+        byte *__aptr = reinterpret_cast<byte *> (addressof(*__first1));
+        byte *__bptr = reinterpret_cast<byte *> (addressof(*__first2));
+
+        size_t __size  = __numBytes < 512 ? __numBytes : 512;
+        size_t __chunk = __numBytes / __size;
+        size_t __rem   = __numBytes % __size;
+
+        char __buffer[__size];
+        if (__chunk > 0) {
+          for (std::size_t __n = 0; __n < __chunk; __n++, __aptr += __size, __bptr += __size) {
+            __builtin_memcpy(__buffer, __aptr,   __size);
+            __builtin_memcpy(__aptr,   __bptr,   __size);
+            __builtin_memcpy(__bptr,   __buffer, __size);
+            }
+          }
+
+          if (__rem > 0) {
+            __builtin_memcpy(__buffer, __aptr,   __rem);
+            __builtin_memcpy(__aptr,   __bptr,   __rem);
+            __builtin_memcpy(__bptr,   __buffer, __rem);
+            }
+          }
+          return pair<_ForwardIterator1, _ForwardIterator2>(__first1 + __distance, __first2 + __distance);
+        }
+    }
+#endif
+
+// else not replaceable, not contiguous, or constexpr
   while (__first1 != __last1) {
     _IterOps<_AlgPolicy>::iter_swap(__first1, __first2);
     ++__first1;