Optimize segmented_copy_if algorithm when the second range is segmented

Author: igaztanaga

include/boost/container/experimental/segmented_copy_if.hpp

@@ -120,6 +120,51 @@ segmented_copy_if_dst_bounded
    return segduo<SrcIter, DstIter>(first, dst_first);
 }
 
+template <class RASrcIter, class RADstIter, class Pred>
+typename iterator_enable_if_tag
+   <RADstIter, std::random_access_iterator_tag, segduo<RASrcIter, RADstIter> >::type
+segmented_copy_if_dst_bounded
+   (RASrcIter first, RASrcIter last, RADstIter dst_first, RADstIter dst_last, Pred pred,
+    const non_segmented_iterator_tag &, const std::random_access_iterator_tag &src_tag)
+{
+   typedef typename iterator_traits<RASrcIter>::difference_type difference_type;
+
+   difference_type src_n = last - first;
+   difference_type dst_n = difference_type(dst_last - dst_first);
+
+   (void)src_tag;
+   RASrcIter cur     = first;
+   RADstIter dst_cur = dst_first;
+   if(src_n <= dst_n) {
+      // Whole source fits the remaining destination capacity: lean pass,
+      // no destination-full check (mirrors the original single-chunk kernel).
+      BOOST_CONTAINER_SEGMENTED_UNROLL(4)
+      while(src_n) {
+         --src_n;
+         if(pred(*cur)) {
+            *dst_cur = *cur;
+            ++dst_cur;
+         }
+         ++cur;
+      }
+   }
+   else {
+      // Source spans more than the destination segment: single bounded pass,
+      // check the destination only on a hit -- avoids repeated chunk set-ups.
+      BOOST_CONTAINER_SEGMENTED_UNROLL(4)
+      while(cur != last) {
+         if(pred(*cur)) {
+            if(dst_cur == dst_last)
+               break;
+            *dst_cur = *cur;
+            ++dst_cur;
+         }
+         ++cur;
+      }
+   }
+   return segduo<RASrcIter, RADstIter>(cur, dst_cur);
+}
+
 template <class SrcIter, class Sent, class SegDstIter, class Pred, class SrcCat>
 segduo<SrcIter, SegDstIter> segmented_copy_if_dst_bounded
    (SrcIter first, Sent last, SegDstIter dst_first, SegDstIter dst_last, Pred pred,