Opened 4 years ago

Last modified 4 years ago

#10319 new task

Eta expand PAPs

Reported by: nomeata Owned by:
Priority: normal Milestone:
Component: Compiler Version: 7.10.1
Keywords: Cc:
Operating System: Unknown/Multiple Architecture: Unknown/Multiple
Type of failure: Runtime performance bug Test Case:
Blocked By: Blocking:
Related Tickets: Differential Rev(s):
Wiki Page:


SPJ writes in ticket:10260#comment:11:

But see this note in SimplUtils:

Note [Do not eta-expand PAPs]
We used to have old_arity = manifestArity rhs, which meant that we
would eta-expand even PAPs.  But this gives no particular advantage,
and can lead to a massive blow-up in code size, exhibited by Trac #9020.
Suppose we have a PAP
    foo :: IO ()
    foo = returnIO ()
Then we can eta-expand do
    foo = (\eta. (returnIO () |> sym g) eta) |> g
    g :: IO () ~ State# RealWorld -> (# State# RealWorld, () #)

But there is really no point in doing this, and it generates masses of
coercions and whatnot that eventually disappear again. For T9020, GHC
allocated 6.6G beore, and 0.8G afterwards; and residency dropped from
1.8G to 45M.

But note that this won't eta-expand, say
  f = \g -> map g
Does it matter not eta-expanding such functions?  I'm not sure.  Perhaps
strictness analysis will have less to bite on?

The worse code for the un-eta-expanded code for last is perhaps an example of the last paragraph!

And yet the comment is reasonably convincing. I'm not sure what to do here.

One thing that might be worth trying is to eta-expand as far as poss without bumping into a coercion. That would expand last but not foo. Want to try that? We may fix last but there are sure to be other cases, and the more robust the optimisations the better.

Change History (2)

comment:1 Changed 4 years ago by nomeata

I’ll defer this until the idea in ticket:9020#comment:4 has been evaluated, as it would make this task moot.

comment:2 Changed 4 years ago by thomie

Type of failure: None/UnknownRuntime performance bug
Note: See TracTickets for help on using tickets.