(Deprecated, will be removed from polysemy at some point)
Performance of libraries like polysemy depends on code being aggresively inlined. Problem is that GHC is not very keen on inlining self-recursive definitions. Luckily, there's a way we can trick compiler to do so: by introducing intermediate loopbreaker that simply calls our original function:
fact :: Int -> Int
fact 0 = 1
fact n = n * fact' (n - 1)
{-# INLINE fact #-}
fact' :: Int -> Int
fact' = fact
{-# NOINLINE fact' #-}
But because this is ugly boilerplate nobody wants to write, we created GHC plugin that searches for such recursive definitions and automatically inserts loopbreakers during compilation.
To quote isovector:
As described in Writing Custom Optimization Passes, The
polysemy-plugin
has had support for creating explicit loopbreakers for self-recursive functions. The result is pretty dramatic code improvements in a lot of cases when-O2
is turned on.
Rather embarrassingly, after publishing that post, it turned out that my implementation didn't in fact improve optimizations. Sure, it spit out the correct code, but it was being done too late, and some special analysis passes had already run. The result: we'd generate loopbreakers, but they wouldn't be used.
TheMatten took it upon himself to fix this. There's no trick --- just do the same transformations after renaming, rather than after typechecking. We realized this plugin is useful outside of polysemy, so it's been released as a standalone package.
Add -fplugin=Loopbreaker
into your package.yaml
or specific module. This
will cause the plugin to generate explicit loopbreakers for any self-recursive
functions marked as INLINE
, which can dramatically improve their
performance. See documentation for more info.