2 ms·
The essay is missing a "usually", but it's true that monomorphisation is a gain in the vast majority of situations because of the data locality and optimisation
by masklinn 5y ago
The essay is missing a "usually", but it's true that monomorphisation is a gain in the vast majority of situations because of the data locality and optimisation opportunities offered by all the calls being static. Though obviously that assumes a pretty heavy optimisation pipeline (so languages like C++ or Rust benefit a lot more than a language with a lighter AOT optimisation pipeline like Java).
Much as with JITs (though probably with higher thresholds), issues occur for megamorphic callsites (when a generic function has a ton of instances), but that should be possible to dump for visibility, and there are common and pretty easy solutions for at least some cases e.g. trampolining through a small generic function (which will almost certainly be inlined) to one that's already monomorphic is pretty common when the generic bits are mostly a few conversions at the head of the function (this sort of trampolining is common in Rust, where "conversion" generics are often used for convenience purposes so e.g. a function will take an `T: AsRef<str>` so the caller doesn't have to extract an `&str` themselves).