Optimize quadratic-time insertion in QSortFilterProxyModel

Let N = proxy_to_source.size() before the code modified in this commit.
Let M = (N - proxy_start). Let K = source_items.size(). The algorithmic
complexity of the removed loop is O(N+K+K*M), assuming the number of
O(N+K) reallocations is a constant. The complexity of the QList::insert
and std::copy implementation is O(N+K). This is much faster in practice
when K and M are of the same order of magnitude as N.

For example, this quadratic complexity issue results in noticeable
slowdown in the following scenario:
  * a QSortFilterProxyModel is used only for filtering, not sorting;
  * first set a filter that matches a single item in the middle of a
    huge number of items (about one million) - this is reasonably fast
    (takes about a second);
  * then clear the filter (i.e. set an empty filter so that no item is
    filtered out) and watch your application's UI freeze for a minute.

The "Add QSortFilterProxyModel clear-filter benchmark" commit (with
Change-Id I419a5521dd0be7676fbb09b34b4069d4a76423b1) adds a benchmark
that runs much faster with this performance fix.

Pick-to: 6.0 6.1
Change-Id: Ieaec173e6910f5d21eaee49402087f7711abbedf
Reviewed-by: Qt CI Bot <qt_ci_bot@qt-project.org>
Reviewed-by: Giuseppe D'Angelo <giuseppe.dangelo@kdab.com>
This commit is contained in:
Igor Kushnir 2021-03-29 18:58:19 +03:00
parent 885eff0537
commit 7d92ef63d7

View File

@ -834,8 +834,9 @@ void QSortFilterProxyModelPrivate::insert_source_items(
q->beginInsertColumns(proxy_parent, proxy_start, proxy_end);
}
for (int i = 0; i < source_items.size(); ++i)
proxy_to_source.insert(proxy_start + i, source_items.at(i));
// TODO: use the range QList::insert() overload once it is implemented (QTBUG-58633).
proxy_to_source.insert(proxy_start, source_items.size(), 0);
std::copy(source_items.cbegin(), source_items.cend(), proxy_to_source.begin() + proxy_start);
build_source_to_proxy_mapping(proxy_to_source, source_to_proxy, proxy_start);