Theory and practice can diverge and it's easy to over-conclude based on either with such complex systems. For example, I have seen gcc PGO make the very same t

Theory and practice can diverge and it's easy to over-conclude based on either with such complex systems. For example, I have seen gcc PGO make the very same training case used to measure the profile run more slowly. One might think that impossible naively, but maybe it sounds more plausible if I put it differently - "steering the many code generation heuristics with the profile failed in practice in that case". As with almost everything in computer systems, "it all depends...."