Opened 5 years ago

Last modified 3 years ago

#9418 new bug

Warnings about "INLINE binder is (non-rule) loop breaker"

Reported by: simonpj Owned by:
Priority: normal Milestone:
Component: Compiler Version: 7.8.2
Keywords: Inlining Cc:
Operating System: Unknown/Multiple Architecture: Unknown/Multiple
Type of failure: Incorrect warning at compile-time Test Case:
Blocked By: Blocking:
Related Tickets: #12137, #3073 Differential Rev(s):
Wiki Page:


Gabor rightly complains: I see literally thousands of these warnings (in yesterday's and) today's bootstraps:

  HC [stage 1] libraries/base/dist-install/build/GHC/Base.o
*** Core Lint warnings : in result of Desugar (after optimization) *** {-# LINE 261 "libraries/base/GHC/Base.lhs #-}: Warning:
    [RHS of $c>>_arr :: forall r_agf a_adQ b_adR.
                        (r_agf -> a_adQ) -> (r_agf -> b_adR) -> r_agf -> b_adR]
    INLINE binder is (non-rule) loop breaker: $c>>_arr {-# LINE 632 "libraries/base/GHC/Base.lhs #-}: Warning:
    [RHS of $c>>_apH :: forall a_adQ b_adR.
                        GHC.Types.IO a_adQ -> GHC.Types.IO b_adR -> GHC.Types.IO b_adR]
    INLINE binder is (non-rule) loop breaker: $c>>_apH }}}

This is clearly unsatisfactory, because it is unsettling, and perhaps conceals more important warnings.

Change History (6)

comment:1 Changed 5 years ago by simonpj

What is happening is this. Consider

class C a where
  op1 :: a -> a -> (a,a)
  op2 :: a -> a -> (a,a)
  {-# INLINE op2 #-}
  op2 x y = op1 y x

instance C Int where
  op1 x y = (x,y)

The intent of the INLINE pragma is obviously to ensure that in every instance declaration we get an INLINEd definition for op2.

These declarations desugar to something like this:

op1 :: forall a. C a => a -> a -> (a,a,)
op1 c = case c of MkC op1 op2 -> op1

Rec { $dCInt :: C Int 
         = MkC $cop1 $cop2

      $cop1 :: Int->Int->(Int,Int) 
         = \xy. (x,y)

      $cop2 :: Int->Int->(Int,Int) {-# INLINE #-} 
         = \xy -> op1 $dCInt y x


  • op1 is the method selector, which picks the op1 field out of a C dictionary.
  • $dCInt is the dictionary for C Int.
  • $cop1 and $cop2 are the implementations of op1 and op2 at type Int.

Notice that $dCInt and $cop2 are apparently mutually recursive; each uses the other. We break the mutual recursion by inlining the op1 selector and $dCInt (in the rhs of $cop2) so that now we can do the record selection. But do to this, we must inline $dCInt, so it can't be a loop breaker.

So $cop2 must be picked as the loop breaker, even though it has an INLINE pragma. That's fine, and it's what GHC does. But Lint is just warning that, as a result, the INLINE pragma isn't going to do anything. This might be useful to the programmer. If you wrote

f x = ....(f y)....
{-# INLINE f #-}

then f won't be inlnined (despite the pragma) because it's recursive. So you might want to re-think your definitions.

The Lint warning is just that: a warning. The unravelling of the mutual recursion happens early, otherwise the warning would be repeated after in every subsequent run of the simplifier. So one possible fix would be to have a flag for Core Lint to control whether the warning was enabled, and only switch it on later in the pipeline. But that doesn't seem very satisfactory.

Well then. That's why it is the way it is. I'd like it to be better!


comment:2 Changed 4 years ago by thomie

Resolution: fixed
Status: newclosed

I don't see those warnings in the build logs, and also the example from comment:1 compiles without warnings. I guess this is fixed.

comment:3 Changed 4 years ago by thomie

Resolution: fixed
Status: closednew
Type of failure: None/UnknownIncorrect warning at compile-time

I closed this prematurely. The example from comment:1 requires -dcore-lint.

Also reported as #12137.

comment:4 Changed 3 years ago by mpickering

comment:5 Changed 3 years ago by mpickering

Keywords: Inlining added

comment:6 Changed 3 years ago by simonpj

See this thread, where Mikolaj wants the inline-loop-breaker warning to happen when not in a debug build.

Note: See TracTickets for help on using tickets.