It's not just a title, it's what the article is about, making llvm fast again is the whole point of author's effort. He presents reason for regression in this area throughout the time (regressions on compile time are not captured vs ie. regressions on produced binary performance are). This is a problem ie. julia is heavily dependant on llvm compile times and julia authors mention that every time they're bumping llvm version they're scared about how much performance will degrade.