When we save/load state in the parser, we preserve the lexer state by
simply making a copy of it. This was made extremely heavy by the lexer
keeping a cache of all parsed identifiers.
It keeps the cache to ensure that StringViews into parsed Unicode escape
sequences don't become dangling views when the Token goes out of scope.
This patch solves the problem by replacing the Vector<FlyString> which
was used to cache the identifiers with a ref-counted
HashTable<FlyString> instead.
Since the purpose of the cache is just to keep FlyStrings alive, it's
fine for all Lexer instances to share the cache. And as a bonus, using a
HashTable instead of a Vector replaces the O(n) accesses with O(1) ones.
This makes a 1.9 MiB JavaScript file parse in 0.6s instead of 24s. :^)
d7578ddebb