Opened 2 years ago

Last modified 2 years ago

#14275 new bug

Large Haskell value unexpectedly gets an unfolding

Reported by: bgamari Owned by:
Priority: low Milestone:
Component: Compiler Version: 8.2.1
Keywords: Inlining Cc:
Operating System: Unknown/Multiple Architecture: Unknown/Multiple
Type of failure: None/Unknown Test Case:
Blocked By: Blocking:
Related Tickets: Differential Rev(s):
Wiki Page:

Description

While working on #14272 I was surprised to find that modifying PrelRules resulted in a great deal of recompilation. Afterall, it has only three exports:

  • primOpRules: A function headed by a rather large case analysis totaling a few hundred lines of code
  • builtinRules: A large [CoreRule] (consisting of a literal list with six entries concatenated with the much-larger builtInIntegerRules)
  • caseRules: A function of moderate size

Intuitively, none of these things seemed particularly beneficial to inline. This is why I was slightly surprised to find that builtinRules (and all of its floated entries) had an unfolding,

  builtinRules :: [CoreRule]
  {- Strictness: m2,
     Unfolding: (: @ CoreRule builtinRules255 builtinRules1) -}

Of course, determining whether unfoldings are helpful is in general quite difficult. However, I can't help but wonder whether our heuristic isn't quite right. Afterall, the Haskell for builtinRules is quite large and consequently most users would be surprised to see GHC try to inline it. The only reason it looks so small is that GHC broke up the structure via float-out.

I don't have any concrete ideas for addressing this at the moment but felt like I should write down the concern so it isn't lost.

Change History (4)

comment:1 Changed 2 years ago by mpickering

Keywords: Inlining added

comment:2 Changed 2 years ago by Ben Gamari <ben@…>

In d11611f/ghc:

Add NOINLINE pragma to builtinRules

As mentioned in #14275, GHC will otherwise decide to produce unfoldings
for this rather large binding, making recompilation more expensive than
necessary. Since inlining is almost certainly not fruitful mark it as
NOINLINE.

[skip ci]

Test Plan: Validate

Reviewers: austin

Subscribers: rwbarton, thomie

GHC Trac Issues: #14275

Differential Revision: https://phabricator.haskell.org/D4023

comment:3 Changed 2 years ago by simonpj

Fine. It's exposed because we might see case builtinRules of ... somewhere. But we never do.. we just fold or map over it.

comment:4 Changed 2 years ago by bgamari

Fine. It's exposed because we might see case builtinRules of ... somewhere. But we never do.. we just fold or map over it.

Of course, I just wonder how often this sort of unfolding will really be useful in practice.

Note: See TracTickets for help on using tickets.