Opened 4 years ago

Closed 3 years ago

#11272 closed bug (fixed)

Overloaded state-monadic function is not specialised

Reported by: NickSmallbone Owned by:
Priority: normal Milestone: 8.2.1
Component: Compiler Version: 7.10.3
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): Phab:D3561
Wiki Page:


I have a simple typeclass-polymorphic function which fails to specialise. Here is module A which defines the function overloaded:

module A where

import Control.Monad.Trans.State
import Control.Monad

overloaded :: Ord a => a -> a -> State () ()
overloaded x y = do
  () <- get
  when (x <= y) (overloaded y x)

In module B I use overloaded on Ints:

module B where

import A
import Control.Monad.Trans.State

specialised :: Int -> Int -> ()
specialised x y = execState (A.overloaded x y) ()

Unfortunately the generated code is not specialised but passes an Ord dictionary around. It doesn't make any difference if I mark overloaded as INLINEABLE or not.

In the core file, overloaded has been worker-wrapper transformed but the worker is marked INLINEABLE[0] - so I'm not sure why it's not being specialised. Curiously, if I make overloaded be a normal function instead of one in the state monad, or if I replace () <- get with simply get, specialisation goes through fine.

Attachments (4)

A.hs (181 bytes) - added by NickSmallbone 4 years ago.
B.hs (141 bytes) - added by NickSmallbone 4 years ago.
A.hcr (5.4 KB) - added by NickSmallbone 4 years ago.
B.hcr (956 bytes) - added by NickSmallbone 4 years ago.

Download all attachments as: .zip

Change History (11)

Changed 4 years ago by NickSmallbone

Attachment: A.hs added

Changed 4 years ago by NickSmallbone

Attachment: B.hs added

Changed 4 years ago by NickSmallbone

Attachment: A.hcr added

Changed 4 years ago by NickSmallbone

Attachment: B.hcr added

comment:1 Changed 4 years ago by simonpj

That is terrible. I understand what is happening.

  • Since overloaded is recursive you must have an INLINEABLE pragma to have a chance of specialising it in an importing module
  • Sadly, when compiling A.hs with {-# INLINEABLE overloaded #-} I see a mutually recursive group of two functions:
    Rec {
    -- RHS size: {terms: 15, types: 14, coercions: 4}
    T11272a.$woverloaded [InlPrag=INLINABLE[0], Occ=LoopBreaker]
      :: forall a_arY. Ord a_arY => a_arY -> a_arY -> () -> (# (), () #)
     Str=DmdType <S(LLLC(C(S))LLLL),U(A,A,A,C(C1(U)),A,A,A,A)><L,U><L,U><S,1*H>,
     Unf=Unf{Src=InlineStable, TopLvl=True, Value=True, ConLike=True,
             WorkFree=True, Expandable=True, Guidance=IF_ARGS [30 0 0 20] 120 30
             Tmpl= \ (@ a8_Xst)
                     (w_X1Me :: Ord a8_Xst)
                     (w1_X1Mg :: a8_Xst)
                     (w2_X1Mi :: a8_Xst)
                     (w3_s1LL [Occ=Once!] :: ()) ->
                     case w3_s1LL of _ [Occ=Dead] { () ->
                     case <= @ a8_Xst w_X1Me w1_X1Mg w2_X1Mi of _ [Occ=Dead] {
                       False -> (# (), () #);
                       True -> T11272a.$woverloaded @ a8_Xst w_X1Me w2_X1Mi w1_X1Mg ()
    T11272a.$woverloaded =
      \ (@ a8_Xst)
        (w_X1Me :: Ord a8_Xst)
        (w1_X1Mg :: a8_Xst)
        (w2_X1Mi :: a8_Xst)
           :: ()
           Unf=OtherCon []) ->
        case (a_r1OI @ a8_Xst w_X1Me w1_X1Mg w2_X1Mi w3_s1LL)
             `cast` (NTCo:Identity[0] <((), ())>_R
                     :: Identity ((), ()) ~R# ((), ()))
        of _ [Occ=Dead] { (ww1_s1LX, ww2_s1LY) ->
        (# ww1_s1LX, ww2_s1LY #)
    -- RHS size: {terms: 26, types: 17, coercions: 10}
      :: forall a_a1zi.
         Ord a_a1zi =>
         a_a1zi -> a_a1zi -> () -> Identity ((), ())
     Str=DmdType <S(LLLC(C(S))LLLL),U(A,A,A,C(C1(U)),A,A,A,A)><L,U><L,U><S,1*H>m]
    a_r1OI =
      \ (@ a8_a1zi)
        (w_s1LP :: Ord a8_a1zi)
        (w1_s1LQ :: a8_a1zi)
        (w2_s1LR :: a8_a1zi)
        (w3_s1LS :: ()) ->
        case w3_s1LS of _ [Occ=Dead] { () ->
        case <= @ a8_a1zi w_s1LP w1_s1LQ w2_s1LR of _ [Occ=Dead] {
          False ->
            `cast` (Sym (NTCo:Identity[0] <((), ())>_R)
                    :: ((), ()) ~R# Identity ((), ()));
          True ->
            case T11272a.$woverloaded @ a8_a1zi w_s1LP w2_s1LR w1_s1LQ ()
            of _ [Occ=Dead] { (# ww1_s1M1, ww2_s1M2 #) ->
            (ww1_s1M1, ww2_s1M2)
            `cast` (Sym (NTCo:Identity[0] <((), ())>_R)
                    :: ((), ()) ~R# Identity ((), ()))
    end Rec }
    That's bad; we'd prefer a tight self-recursive function.
  • Worse, it doesn't specialise, because the cast makes it look as if it doesn't have top-level lambdas.

Need to think about this, but thought I'd jot down what I know so far.

comment:2 Changed 3 years ago by nfrisby

If I add the INLINABLE overloaded pragma in A and place a SPECIALIZE overloaded :: Int -> Int -> State () () pragma in B, then the desired specialization ultimately happens (with GHC 7.10.2). I mention this as a possible workaround for people stuck on this bug.

Forcing the overloaded wrapper to be specialized creates a "SPEC" rule whose RHS gets gently simplified enough before the Specialise pass such that GHC then notices the specializable call to the $woverloaded worker. Since that worker inherits the original INLINABLE pragma (see #6056), GHC automatically specializes the worker. The specialized worker eventually ends up in the RHS of specialised.

This approach was not obvious to me because I hadn't yet realized that forcibly specializing a w/w wrapper function tends to cause the worker to be automatically specialized. Two shortcomings come immediately to mind.

  1. I get the SPECIALISE pragma on INLINE function probably won't fire: ‘overloaded’ warning, even though the rule does fire.
  2. In my experience, SPECIALIZE pragmas are often rejected with "LHS too complicated to desugar" (e.g. #10555), so you may need to use GHC.Exts.inline instead.
    • Note that inline crucially causes the w/w wrapper to be inlined before the Specialise pass (during Phase = InitialPhase [Gentle]), whereas the w/w wrapper's INLINE[0] activation delays its natural inlining until after the Specialise pass.
    • This is awkward: we only want to inline overloaded if it was replaced by a w/w wrapper! A "safer" but more troublesome alternative is to have inline overloaded somewhere else in the module, but then you must ensure that it's exported from this module, else GHC will cull it and the SPEC rule before the rule can usefully affect specialise. This risks binary bloat, but it also prevents inline from ever accidentally inlining overloaded into the genuine code.

It's remarkable that the w/w transform creates the cast that prevents the wrapper from being automatically specialized (that cast floats to the top of the unfolding, which blocks specialization; see Note [Specialisation shape] in Specialise.hs). The demand analyzer and w/w transform in this example see through the State type synonym, StateT newtype, and Identity newtype all the way to the s -> (# a, s #) worker type -- the Specialise pass does not yet have the complementary X-ray vision to see through the corresponding cast to the desired lambdas in the wrapper's unfolding.

(#9509 is also specialization-blocked-by-a-cast-atop-the-unfolding, but I think the cast in that case comes from an actual bug.)

comment:3 Changed 3 years ago by dfeuer

This appears to be fixed in HEAD (as long as overloaded has an INLINABLE pragma. I'm not sure how test cases for this sort of thing are supposed to be added to the test suite; should I be digging through -ddump-rule-firings or somethinhg?

comment:4 Changed 3 years ago by bgamari

should I be digging through -ddump-rule-firings or somethinhg?

Either that or test -ddump-simpl. Both are sadly rather fragile.

comment:5 Changed 3 years ago by dfeuer

Differential Rev(s): Phab:D3561
Status: newpatch

comment:6 Changed 3 years ago by David Feuer <David.Feuer@…>

In 70191f5/ghc:

Add a test for #11272

Reviewers: austin, bgamari

Reviewed By: bgamari

Subscribers: rwbarton, thomie

GHC Trac Issues: #11272

Differential Revision:

comment:7 Changed 3 years ago by dfeuer

Milestone: 8.2.1
Resolution: fixed
Status: patchclosed
Note: See TracTickets for help on using tickets.