[libcxx-commits] [libcxx] [libc++] Speed up set_intersection() by fast-forwarding over ranges of non-matching elements with one-sided binary search. (PR #75230)

[Louis Dionne via libcxx-commits]

================
@@ -46,6 +48,48 @@ __lower_bound(_Iter __first, _Sent __last, const _Type& __value, _Comp& __comp,
   return __first;
 }
 
+// One-sided binary search, aka meta binary search, has been in the public domain for decades, and has the general
+// advantage of being \Omega(1) rather than the classic algorithm's \Omega(log(n)), with the downside of executing at
+// most 2*log(n) comparisons vs the classic algorithm's exact log(n). There are two scenarios in which it really shines:
+// the first one is when operating over non-random iterators, because the classic algorithm requires knowing the
+// container's size upfront, which adds \Omega(n) iterator increments to the complexity. The second one is when you're
+// traversing the container in order, trying to fast-forward to the next value: in that case, the classic algorithm
+// would yield \Omega(n*log(n)) comparisons and, for non-random iterators, \Omega(n^2) iterator increments, whereas the
+// one-sided version will yield O(n) operations on both counts, with a \Omega(log(n)) bound on the number of
+// comparisons.
+template <class _AlgPolicy, class _ForwardIterator, class _Sent, class _Type, class _Proj, class _Comp>
+_LIBCPP_NODISCARD _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 _ForwardIterator
+__lower_bound_onesided(_ForwardIterator __first, _Sent __last, const _Type& __value, _Comp& __comp, _Proj& __proj) {
+  // step = 0, ensuring we can always short-circuit when distance is 1 later on
+  if (__first == __last || !std::__invoke(__comp, std::__invoke(__proj, *__first), __value))
+    return __first;
+
+  using _Distance = typename iterator_traits<_ForwardIterator>::difference_type;
+  for (_Distance __step = 1; __first != __last; __step <<= 1) {
+    auto __it   = __first;
+    auto __dist = __step - _IterOps<_AlgPolicy>::__advance_to(__it, __step, __last);
+    // once we reach the last range where needle can be we must start
+    // looking inwards, bisecting that range
+    if (__it == __last || !std::__invoke(__comp, std::__invoke(__proj, *__it), __value)) {
+      // we've already checked the previous value and it was less, we can save
+      // one comparison by skipping bisection
+      if (__dist == 1)
+        return __it;
+      return std::__lower_bound_bisecting<_AlgPolicy>(__first, __value, __dist, __comp, __proj);
+    }
+    // range not found, move forward!
+    __first = __it;
+  }
+  return __first;
+}
+
+template <class _AlgPolicy, class _ForwardIterator, class _Sent, class _Type, class _Proj, class _Comp>
+_LIBCPP_NODISCARD inline _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 _ForwardIterator
+__lower_bound(_ForwardIterator __first, _Sent __last, const _Type& __value, _Comp& __comp, _Proj& __proj) {
----------------
ldionne wrote:

That's also my thinking.

@ichaer After this patch has landed, you could consider a patch like the above if you want -- I think we'd take it.

https://github.com/llvm/llvm-project/pull/75230