My personal experience is very contradictionary to that statement. Handoptimizing a slow PEG parser, but still using PEG gave me 10% performance increase. Handwriting it on low level libraries, but in the same high level language, gave me a 50,000% performance increase on the average file it had to parse.

Task at hand wasn't a programming language though, but more like YAML with only one level of nesting and a lot of quoted strings (which were the bottleneck for PEG in my case).