The “evilness” is just that sometimes if you inline aggressively in a microbenchmark things get faster but in real programs things get slower.

As you say: “chances are they are a lot better at it than you”. Infrequently they are not.