Optimizing hashtable.grow() for faster hashtables (#459)

* Using a specialized version of Insert to speed up hashmap expansion

* Further improvements

* undo changes and update TODO

Author: SamWheating

starlark/hashtable.go

@@ -155,13 +155,12 @@ func overloaded(elems, buckets int) bool {
 
 func (ht *hashtable) grow() {
 	// Double the number of buckets and rehash.
-	// TODO(adonovan): opt:
-	// - avoid reentrant calls to ht.insert, and specialize it.
-	//   e.g. we know the calls to Equals will return false since
-	//   there are no duplicates among the old keys.
-	// - saving the entire hash in the bucket would avoid the need to
-	//   recompute the hash.
-	// - save the old buckets on a free list.
+	//
+	// Even though this makes reentrant calls to ht.insert,
+	// calls Equals unnecessarily (since there can't be duplicate keys),
+	// and recomputes the hash unnecessarily, the gains from
+	// avoiding these steps were found to be too small to justify
+	// the extra logic: -2% on hashtable benchmark.
 	ht.table = make([]bucket, len(ht.table)<<1)
 	oldhead := ht.head
 	ht.head = nil