Using a perfect hash table is definitely cool, but it's hard to imagine a query with so many keywords that this would have a measurable impact on total query execution time.
Due to the desire to have keywords that aren't fully reserved (we have keyword classes that are fully reserved, can appear as type names, can appear as column names, or can appear nearly everywhere) we/PG need to look up whether a specific string is a keyword or an identifier in a number of places in the scanner/parser.
Addendum:
And that indeed means that column references, unless quoted with "", used to go through the binary search and now go through the perfect hash.
Addendum: And that indeed means that column references, unless quoted with "", used to go through the binary search and now go through the perfect hash.
Edit: Typo