Fix performance issue with subset iterators

Previously we were creating a new iterator each time we returned a
value, now we store the previous index in the iterator and update the
full iterator with the difference when needed.

The performance issue was occurring due to transient object creation of
the temporary iterators and moving them to the right location.

Fixes #378

NEWS.md

@@ -5,6 +5,7 @@
 * `vroom(n_max=)` now correctly handles cases when reading from a connection and the file does _not_ end with a newline (https://github.com/tidyverse/readr/issues/1321)
 
 * `vroom()` no longer issues a spurious warning when the parsing needs to be restarted due to the presence of embedded newlines (https://github.com/tidyverse/readr/issues/1313)
+* Fix performance issue when materializing subsetted vectors (#378)
 
 * `vroom_format()` now uses the same internal multi-threaded code as `vroom_write()`, improving its performance in most cases (#377)
 

src/index.h

@@ -12,13 +12,14 @@ class index {
 public:
   class subset_iterator : public base_iterator {
     size_t i_;
-    iterator it_;
+    mutable size_t prev_;
+    mutable iterator it_;
     std::shared_ptr<std::vector<size_t>> indexes_;
 
   public:
     subset_iterator(
         const iterator& it, const std::shared_ptr<std::vector<size_t>>& indexes)
-        : i_(0), it_(it), indexes_(indexes) {}
+        : i_(0), prev_(0), it_(it), indexes_(indexes) {}
     void next() override { ++i_; }
     void prev() override { --i_; }
     void advance(ptrdiff_t n) override { i_ += n; }
@@ -30,7 +31,13 @@ class index {
       auto that_ = static_cast<const subset_iterator*>(&that);
       return that_->i_ - i_;
     };
-    string value() const override { return *(it_ + (*indexes_)[i_]); };
+    string value() const override {
+      size_t cur = (*indexes_)[i_];
+      ptrdiff_t diff = cur - prev_;
+      it_ += diff;
+      prev_ = cur;
+      return *it_;
+    };
     subset_iterator* clone() const override {
       auto copy = new subset_iterator(*this);
       return copy;