Simplify primitives for short cut fusion
|Reported by:||dfeuer||Owned by:||dfeuer|
|Keywords:||fusion||Cc:||hvr, ekmett, core-libraries-committee@…|
|Type of failure:||Other||Test Case:|
|Related Tickets:||Differential Rev(s):|
Currently, there appear to be two production primitives, build and augment (although I may have missed some). There are multiple consumption primitives (at least: foldr, head, and, or, any, and all). The rule sets for some producers seem to forget to handle augment, and the augment/augment rule is omitted as "true, but not, I think, useful".
A number of other functions are instead rewritten into foldr forms, and then written back if they don't fuse.
What to do:
Personally, I'd be very tempted to start by saying build g = augment g  (or possibly even build g = extend  g  if I can ever get that idea to work) and cut the problem in half. The main problem I see with this is if other people are importing GHC.Exts or GHC.Base, writing things with build, and expecting them to fuse. One way to deal with this, perhaps, is to hack a special rule into the RULES compiler to recognize GHC.Base.build on the LHS of RULES and replace it with the appropriate augment form, emitting a warning.
Where to go after that: the question remains whether it's best in general to rewrite a form to foldr to make it fuse, or to fuse directly with augment. The answer presumably depends, at least in part, on whether there are additional foldr-based rules we may want to add that would take advantage of the effort put into the translation back from the foldr form. I would conjecture that most such rules we could want would go beyond what the RULES system actually can do. That said, if we can find a way to use foldr forms without the horrible pain of translating back from them, that would be a very good thing.
Change History (5)
comment:3 in reply to: ↑ 2 Changed 18 months ago by dfeuer
comment:4 Changed 16 months ago by thoughtpolice
- Component changed from libraries/base to Core Libraries