2 % (c) The University of Glasgow 2006
3 % (c) The AQUA Project, Glasgow University, 1994-1998
8 Unfoldings (which can travel across module boundaries) are in Core
9 syntax (namely @CoreExpr@s).
11 The type @Unfolding@ sits ``above'' simply-Core-expressions
12 unfoldings, capturing ``higher-level'' things we know about a binding,
13 usually things that the simplifier found out (e.g., ``it's a
14 literal''). In the corner of a @CoreUnfolding@ unfolding, you will
15 find, unsurprisingly, a Core expression.
19 Unfolding, UnfoldingGuidance, -- Abstract types
21 noUnfolding, mkImplicitUnfolding,
22 mkUnfolding, mkCoreUnfolding,
23 mkTopUnfolding, mkSimpleUnfolding,
24 mkInlineUnfolding, mkInlinableUnfolding, mkWwInlineRule,
25 mkCompulsoryUnfolding, mkDFunUnfolding,
27 interestingArg, ArgSummary(..),
29 couldBeSmallEnoughToInline,
30 certainlyWillInline, smallEnoughToInline,
32 callSiteInline, CallCtxt(..),
38 #include "HsVersions.h"
43 import PprCore () -- Instances
44 import TcType ( tcSplitSigmaTy, tcSplitDFunHead )
46 import CoreSubst hiding( substTy )
47 import CoreFVs ( exprFreeVars )
48 import CoreArity ( manifestArity, exprBotStrictness_maybe )
56 import BasicTypes ( Arity )
57 import TcType ( tcSplitDFunTy )
61 import VarEnv ( mkInScopeSet )
71 %************************************************************************
73 \subsection{Making unfoldings}
75 %************************************************************************
78 mkTopUnfolding :: Bool -> CoreExpr -> Unfolding
79 mkTopUnfolding = mkUnfolding InlineRhs True {- Top level -}
81 mkImplicitUnfolding :: CoreExpr -> Unfolding
82 -- For implicit Ids, do a tiny bit of optimising first
83 mkImplicitUnfolding expr = mkTopUnfolding False (simpleOptExpr expr)
85 -- Note [Top-level flag on inline rules]
86 -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
87 -- Slight hack: note that mk_inline_rules conservatively sets the
88 -- top-level flag to True. It gets set more accurately by the simplifier
89 -- Simplify.simplUnfolding.
91 mkSimpleUnfolding :: CoreExpr -> Unfolding
92 mkSimpleUnfolding = mkUnfolding InlineRhs False False
94 mkDFunUnfolding :: Type -> [CoreExpr] -> Unfolding
95 mkDFunUnfolding dfun_ty ops
96 = DFunUnfolding dfun_nargs data_con ops
98 (tvs, theta, head_ty) = tcSplitSigmaTy dfun_ty
99 -- NB: tcSplitSigmaTy: do not look through a newtype
100 -- when the dictionary type is a newtype
101 (cls, _) = tcSplitDFunHead head_ty
102 dfun_nargs = length tvs + length theta
103 data_con = classDataCon cls
105 mkWwInlineRule :: Id -> CoreExpr -> Arity -> Unfolding
106 mkWwInlineRule id expr arity
107 = mkCoreUnfolding (InlineWrapper id) True
108 (simpleOptExpr expr) arity
109 (UnfWhen unSaturatedOk boringCxtNotOk)
111 mkCompulsoryUnfolding :: CoreExpr -> Unfolding
112 mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
113 = mkCoreUnfolding InlineCompulsory True
114 expr 0 -- Arity of unfolding doesn't matter
115 (UnfWhen unSaturatedOk boringCxtOk)
117 mkInlineUnfolding :: Maybe Arity -> CoreExpr -> Unfolding
118 mkInlineUnfolding mb_arity expr
119 = mkCoreUnfolding InlineStable
120 True -- Note [Top-level flag on inline rules]
122 (UnfWhen unsat_ok boring_ok)
124 expr' = simpleOptExpr expr
125 (unsat_ok, arity) = case mb_arity of
126 Nothing -> (unSaturatedOk, manifestArity expr')
127 Just ar -> (needSaturated, ar)
129 boring_ok = case calcUnfoldingGuidance True -- Treat as cheap
130 False -- But not bottoming
132 (_, UnfWhen _ boring_ok) -> boring_ok
133 _other -> boringCxtNotOk
134 -- See Note [INLINE for small functions]
136 mkInlinableUnfolding :: CoreExpr -> Unfolding
137 mkInlinableUnfolding expr
138 = mkUnfolding InlineStable True is_bot expr'
140 expr' = simpleOptExpr expr
141 is_bot = isJust (exprBotStrictness_maybe expr')
147 mkCoreUnfolding :: UnfoldingSource -> Bool -> CoreExpr
148 -> Arity -> UnfoldingGuidance -> Unfolding
149 -- Occurrence-analyses the expression before capturing it
150 mkCoreUnfolding src top_lvl expr arity guidance
151 = CoreUnfolding { uf_tmpl = occurAnalyseExpr expr,
155 uf_is_value = exprIsHNF expr,
156 uf_is_conlike = exprIsConLike expr,
157 uf_is_cheap = exprIsCheap expr,
158 uf_expandable = exprIsExpandable expr,
159 uf_guidance = guidance }
161 mkUnfolding :: UnfoldingSource -> Bool -> Bool -> CoreExpr -> Unfolding
162 -- Calculates unfolding guidance
163 -- Occurrence-analyses the expression before capturing it
164 mkUnfolding src top_lvl is_bottoming expr
165 = CoreUnfolding { uf_tmpl = occurAnalyseExpr expr,
169 uf_is_value = exprIsHNF expr,
170 uf_is_conlike = exprIsConLike expr,
171 uf_expandable = exprIsExpandable expr,
172 uf_is_cheap = is_cheap,
173 uf_guidance = guidance }
175 is_cheap = exprIsCheap expr
176 (arity, guidance) = calcUnfoldingGuidance is_cheap (top_lvl && is_bottoming)
177 opt_UF_CreationThreshold expr
178 -- Sometimes during simplification, there's a large let-bound thing
179 -- which has been substituted, and so is now dead; so 'expr' contains
180 -- two copies of the thing while the occurrence-analysed expression doesn't
181 -- Nevertheless, we *don't* occ-analyse before computing the size because the
182 -- size computation bales out after a while, whereas occurrence analysis does not.
184 -- This can occasionally mean that the guidance is very pessimistic;
185 -- it gets fixed up next round. And it should be rare, because large
186 -- let-bound things that are dead are usually caught by preInlineUnconditionally
189 %************************************************************************
191 \subsection{The UnfoldingGuidance type}
193 %************************************************************************
196 calcUnfoldingGuidance
197 :: Bool -- True <=> the rhs is cheap, or we want to treat it
198 -- as cheap (INLINE things)
199 -> Bool -- True <=> this is a top-level unfolding for a
200 -- diverging function; don't inline this
201 -> Int -- Bomb out if size gets bigger than this
202 -> CoreExpr -- Expression to look at
203 -> (Arity, UnfoldingGuidance)
204 calcUnfoldingGuidance expr_is_cheap top_bot bOMB_OUT_SIZE expr
205 = case collectBinders expr of { (bndrs, body) ->
207 val_bndrs = filter isId bndrs
208 n_val_bndrs = length val_bndrs
211 = case (sizeExpr (iUnbox bOMB_OUT_SIZE) val_bndrs body) of
213 SizeIs size cased_bndrs scrut_discount
214 | uncondInline n_val_bndrs (iBox size)
216 -> UnfWhen unSaturatedOk boringCxtOk -- Note [INLINE for small functions]
217 | top_bot -- See Note [Do not inline top-level bottoming functions]
221 -> UnfIfGoodArgs { ug_args = map (discount cased_bndrs) val_bndrs
222 , ug_size = iBox size
223 , ug_res = iBox scrut_discount }
226 = foldlBag (\acc (b',n) -> if bndr==b' then acc+n else acc)
229 (n_val_bndrs, guidance) }
232 Note [Computing the size of an expression]
233 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
234 The basic idea of sizeExpr is obvious enough: count nodes. But getting the
235 heuristics right has taken a long time. Here's the basic strategy:
237 * Variables, literals: 0
238 (Exception for string literals, see litSize.)
240 * Function applications (f e1 .. en): 1 + #value args
242 * Constructor applications: 1, regardless of #args
244 * Let(rec): 1 + size of components
259 Notice that 'x' counts 0, while (f x) counts 2. That's deliberate: there's
260 a function call to account for. Notice also that constructor applications
261 are very cheap, because exposing them to a caller is so valuable.
264 Note [Do not inline top-level bottoming functions]
265 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
266 The FloatOut pass has gone to some trouble to float out calls to 'error'
267 and similar friends. See Note [Bottoming floats] in SetLevels.
268 Do not re-inline them! But we *do* still inline if they are very small
269 (the uncondInline stuff).
272 Note [INLINE for small functions]
273 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
274 Consider {-# INLINE f #-}
277 Then f's RHS is no larger than its LHS, so we should inline it into
278 even the most boring context. In general, f the function is
279 sufficiently small that its body is as small as the call itself, the
280 inline unconditionally, regardless of how boring the context is.
284 * We inline *unconditionally* if inlined thing is smaller (using sizeExpr)
285 than the thing it's replacing. Notice that
286 (f x) --> (g 3) -- YES, unconditionally
287 (f x) --> x : [] -- YES, *even though* there are two
288 -- arguments to the cons
292 It's very important not to unconditionally replace a variable by
295 * We do this even if the thing isn't saturated, else we end up with the
299 doesn't inline. Even in a boring context, inlining without being
300 saturated will give a lambda instead of a PAP, and will be more
301 efficient at runtime.
303 * However, when the function's arity > 0, we do insist that it
304 has at least one value argument at the call site. Otherwise we find this:
307 If we inline f here we get
308 d = /\b. MkD (\x:b. x)
309 and then prepareRhs floats out the argument, abstracting the type
310 variables, so we end up with the original again!
314 uncondInline :: Arity -> Int -> Bool
315 -- Inline unconditionally if there no size increase
316 -- Size of call is arity (+1 for the function)
317 -- See Note [INLINE for small functions]
318 uncondInline arity size
319 | arity == 0 = size == 0
320 | otherwise = size <= arity + 1
325 sizeExpr :: FastInt -- Bomb out if it gets bigger than this
326 -> [Id] -- Arguments; we're interested in which of these
331 -- Note [Computing the size of an expression]
333 sizeExpr bOMB_OUT_SIZE top_args expr
336 size_up (Cast e _) = size_up e
337 size_up (Note _ e) = size_up e
338 size_up (Type _) = sizeZero -- Types cost nothing
339 size_up (Lit lit) = sizeN (litSize lit)
340 size_up (Var f) = size_up_call f [] -- Make sure we get constructor
341 -- discounts even on nullary constructors
343 size_up (App fun (Type _)) = size_up fun
344 size_up (App fun arg) = size_up arg `addSizeNSD`
345 size_up_app fun [arg]
347 size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
348 | otherwise = size_up e
350 size_up (Let (NonRec binder rhs) body)
351 = size_up rhs `addSizeNSD`
352 size_up body `addSizeN`
353 (if isUnLiftedType (idType binder) then 0 else 1)
354 -- For the allocation
355 -- If the binder has an unlifted type there is no allocation
357 size_up (Let (Rec pairs) body)
358 = foldr (addSizeNSD . size_up . snd)
359 (size_up body `addSizeN` length pairs) -- (length pairs) for the allocation
362 size_up (Case (Var v) _ _ alts)
363 | v `elem` top_args -- We are scrutinising an argument variable
364 = alts_size (foldr1 addAltSize alt_sizes)
365 (foldr1 maxSize alt_sizes)
366 -- Good to inline if an arg is scrutinised, because
367 -- that may eliminate allocation in the caller
368 -- And it eliminates the case itself
370 alt_sizes = map size_up_alt alts
372 -- alts_size tries to compute a good discount for
373 -- the case when we are scrutinising an argument variable
374 alts_size (SizeIs tot tot_disc tot_scrut) -- Size of all alternatives
375 (SizeIs max _ _) -- Size of biggest alternative
376 = SizeIs tot (unitBag (v, iBox (_ILIT(2) +# tot -# max)) `unionBags` tot_disc) tot_scrut
377 -- If the variable is known, we produce a discount that
378 -- will take us back to 'max', the size of the largest alternative
379 -- The 1+ is a little discount for reduced allocation in the caller
381 -- Notice though, that we return tot_disc, the total discount from
382 -- all branches. I think that's right.
384 alts_size tot_size _ = tot_size
386 size_up (Case e _ _ alts) = size_up e `addSizeNSD`
387 foldr (addAltSize . size_up_alt) sizeZero alts
388 -- We don't charge for the case itself
389 -- It's a strict thing, and the price of the call
390 -- is paid by scrut. Also consider
391 -- case f x of DEFAULT -> e
392 -- This is just ';'! Don't charge for it.
394 -- Moreover, we charge one per alternative.
397 -- size_up_app is used when there's ONE OR MORE value args
398 size_up_app (App fun arg) args
399 | isTypeArg arg = size_up_app fun args
400 | otherwise = size_up arg `addSizeNSD`
401 size_up_app fun (arg:args)
402 size_up_app (Var fun) args = size_up_call fun args
403 size_up_app other args = size_up other `addSizeN` length args
406 size_up_call :: Id -> [CoreExpr] -> ExprSize
407 size_up_call fun val_args
408 = case idDetails fun of
409 FCallId _ -> sizeN opt_UF_DearOp
410 DataConWorkId dc -> conSize dc (length val_args)
411 PrimOpId op -> primOpSize op (length val_args)
412 ClassOpId _ -> classOpSize top_args val_args
413 _ -> funSize top_args fun (length val_args)
416 size_up_alt (_con, _bndrs, rhs) = size_up rhs `addSizeN` 1
417 -- Don't charge for args, so that wrappers look cheap
418 -- (See comments about wrappers with Case)
420 -- IMPORATANT: *do* charge 1 for the alternative, else we
421 -- find that giant case nests are treated as practically free
422 -- A good example is Foreign.C.Error.errrnoToIOError
425 -- These addSize things have to be here because
426 -- I don't want to give them bOMB_OUT_SIZE as an argument
427 addSizeN TooBig _ = TooBig
428 addSizeN (SizeIs n xs d) m = mkSizeIs bOMB_OUT_SIZE (n +# iUnbox m) xs d
430 -- addAltSize is used to add the sizes of case alternatives
431 addAltSize TooBig _ = TooBig
432 addAltSize _ TooBig = TooBig
433 addAltSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
434 = mkSizeIs bOMB_OUT_SIZE (n1 +# n2)
436 (d1 +# d2) -- Note [addAltSize result discounts]
438 -- This variant ignores the result discount from its LEFT argument
439 -- It's used when the second argument isn't part of the result
440 addSizeNSD TooBig _ = TooBig
441 addSizeNSD _ TooBig = TooBig
442 addSizeNSD (SizeIs n1 xs _) (SizeIs n2 ys d2)
443 = mkSizeIs bOMB_OUT_SIZE (n1 +# n2)
449 -- | Finds a nominal size of a string literal.
450 litSize :: Literal -> Int
451 -- Used by CoreUnfold.sizeExpr
452 litSize (MachStr str) = 1 + ((lengthFS str + 3) `div` 4)
453 -- If size could be 0 then @f "x"@ might be too small
454 -- [Sept03: make literal strings a bit bigger to avoid fruitless
455 -- duplication of little strings]
456 litSize _other = 0 -- Must match size of nullary constructors
457 -- Key point: if x |-> 4, then x must inline unconditionally
458 -- (eg via case binding)
460 classOpSize :: [Id] -> [CoreExpr] -> ExprSize
461 -- See Note [Conlike is interesting]
464 classOpSize top_args (arg1 : other_args)
465 = SizeIs (iUnbox size) arg_discount (_ILIT(0))
467 size = 2 + length other_args
468 -- If the class op is scrutinising a lambda bound dictionary then
469 -- give it a discount, to encourage the inlining of this function
470 -- The actual discount is rather arbitrarily chosen
471 arg_discount = case arg1 of
472 Var dict | dict `elem` top_args
473 -> unitBag (dict, opt_UF_DictDiscount)
476 funSize :: [Id] -> Id -> Int -> ExprSize
477 -- Size for functions that are not constructors or primops
478 -- Note [Function applications]
479 funSize top_args fun n_val_args
480 | fun `hasKey` buildIdKey = buildSize
481 | fun `hasKey` augmentIdKey = augmentSize
482 | otherwise = SizeIs (iUnbox size) arg_discount (iUnbox res_discount)
484 some_val_args = n_val_args > 0
486 arg_discount | some_val_args && fun `elem` top_args
487 = unitBag (fun, opt_UF_FunAppDiscount)
488 | otherwise = emptyBag
489 -- If the function is an argument and is applied
490 -- to some values, give it an arg-discount
492 res_discount | idArity fun > n_val_args = opt_UF_FunAppDiscount
494 -- If the function is partially applied, show a result discount
496 size | some_val_args = 1 + n_val_args
498 -- The 1+ is for the function itself
499 -- Add 1 for each non-trivial arg;
500 -- the allocation cost, as in let(rec)
503 conSize :: DataCon -> Int -> ExprSize
504 conSize dc n_val_args
505 | n_val_args == 0 = SizeIs (_ILIT(0)) emptyBag (_ILIT(1)) -- Like variables
507 -- See Note [Constructor size]
508 | isUnboxedTupleCon dc = SizeIs (_ILIT(0)) emptyBag (iUnbox n_val_args +# _ILIT(1))
510 -- See Note [Unboxed tuple result discount]
511 -- | isUnboxedTupleCon dc = SizeIs (_ILIT(0)) emptyBag (_ILIT(0))
513 -- See Note [Constructor size]
514 | otherwise = SizeIs (_ILIT(1)) emptyBag (iUnbox n_val_args +# _ILIT(1))
517 Note [Constructor size]
518 ~~~~~~~~~~~~~~~~~~~~~~~
519 Treat a constructors application as size 1, regardless of how many
520 arguments it has; we are keen to expose them (and we charge separately
521 for their args). We can't treat them as size zero, else we find that
522 (Just x) has size 0, which is the same as a lone variable; and hence
523 'v' will always be replaced by (Just x), where v is bound to Just x.
525 However, unboxed tuples count as size zero. I found occasions where we had
526 f x y z = case op# x y z of { s -> (# s, () #) }
527 and f wasn't getting inlined.
529 Note [Unboxed tuple result discount]
530 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
531 I tried giving unboxed tuples a *result discount* of zero (see the
532 commented-out line). Why? When returned as a result they do not
533 allocate, so maybe we don't want to charge so much for them If you
534 have a non-zero discount here, we find that workers often get inlined
535 back into wrappers, because it look like
536 f x = case $wf x of (# a,b #) -> (a,b)
537 and we are keener because of the case. However while this change
538 shrank binary sizes by 0.5% it also made spectral/boyer allocate 5%
539 more. All other changes were very small. So it's not a big deal but I
540 didn't adopt the idea.
543 primOpSize :: PrimOp -> Int -> ExprSize
544 primOpSize op n_val_args
545 | not (primOpIsDupable op) = sizeN opt_UF_DearOp
546 | not (primOpOutOfLine op) = sizeN 1
547 -- Be very keen to inline simple primops.
548 -- We give a discount of 1 for each arg so that (op# x y z) costs 2.
549 -- We can't make it cost 1, else we'll inline let v = (op# x y z)
550 -- at every use of v, which is excessive.
552 -- A good example is:
553 -- let x = +# p q in C {x}
554 -- Even though x get's an occurrence of 'many', its RHS looks cheap,
555 -- and there's a good chance it'll get inlined back into C's RHS. Urgh!
557 | otherwise = sizeN n_val_args
560 buildSize :: ExprSize
561 buildSize = SizeIs (_ILIT(0)) emptyBag (_ILIT(4))
562 -- We really want to inline applications of build
563 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
564 -- Indeed, we should add a result_discount becuause build is
565 -- very like a constructor. We don't bother to check that the
566 -- build is saturated (it usually is). The "-2" discounts for the \c n,
567 -- The "4" is rather arbitrary.
569 augmentSize :: ExprSize
570 augmentSize = SizeIs (_ILIT(0)) emptyBag (_ILIT(4))
571 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
572 -- e plus ys. The -2 accounts for the \cn
574 -- When we return a lambda, give a discount if it's used (applied)
575 lamScrutDiscount :: ExprSize -> ExprSize
576 lamScrutDiscount (SizeIs n vs _) = SizeIs n vs (iUnbox opt_UF_FunAppDiscount)
577 lamScrutDiscount TooBig = TooBig
580 Note [addAltSize result discounts]
581 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
582 When adding the size of alternatives, we *add* the result discounts
583 too, rather than take the *maximum*. For a multi-branch case, this
584 gives a discount for each branch that returns a constructor, making us
585 keener to inline. I did try using 'max' instead, but it makes nofib
586 'rewrite' and 'puzzle' allocate significantly more, and didn't make
587 binary sizes shrink significantly either.
589 Note [Discounts and thresholds]
590 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
591 Constants for discounts and thesholds are defined in main/StaticFlags,
592 all of form opt_UF_xxxx. They are:
594 opt_UF_CreationThreshold (45)
595 At a definition site, if the unfolding is bigger than this, we
596 may discard it altogether
598 opt_UF_UseThreshold (6)
599 At a call site, if the unfolding, less discounts, is smaller than
600 this, then it's small enough inline
602 opt_UF_KeennessFactor (1.5)
603 Factor by which the discounts are multiplied before
604 subtracting from size
606 opt_UF_DictDiscount (1)
607 The discount for each occurrence of a dictionary argument
608 as an argument of a class method. Should be pretty small
609 else big functions may get inlined
611 opt_UF_FunAppDiscount (6)
612 Discount for a function argument that is applied. Quite
613 large, because if we inline we avoid the higher-order call.
616 The size of a foreign call or not-dupable PrimOp
619 Note [Function applications]
620 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
621 In a function application (f a b)
623 - If 'f' is an argument to the function being analysed,
624 and there's at least one value arg, record a FunAppDiscount for f
626 - If the application if a PAP (arity > 2 in this example)
627 record a *result* discount (because inlining
628 with "extra" args in the call may mean that we now
629 get a saturated application)
631 Code for manipulating sizes
634 data ExprSize = TooBig
635 | SizeIs FastInt -- Size found
636 (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
637 FastInt -- Size to subtract if result is scrutinised
638 -- by a case expression
640 instance Outputable ExprSize where
641 ppr TooBig = ptext (sLit "TooBig")
642 ppr (SizeIs a _ c) = brackets (int (iBox a) <+> int (iBox c))
644 -- subtract the discount before deciding whether to bale out. eg. we
645 -- want to inline a large constructor application into a selector:
646 -- tup = (a_1, ..., a_99)
647 -- x = case tup of ...
649 mkSizeIs :: FastInt -> FastInt -> Bag (Id, Int) -> FastInt -> ExprSize
650 mkSizeIs max n xs d | (n -# d) ># max = TooBig
651 | otherwise = SizeIs n xs d
653 maxSize :: ExprSize -> ExprSize -> ExprSize
654 maxSize TooBig _ = TooBig
655 maxSize _ TooBig = TooBig
656 maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
660 sizeN :: Int -> ExprSize
662 sizeZero = SizeIs (_ILIT(0)) emptyBag (_ILIT(0))
663 sizeN n = SizeIs (iUnbox n) emptyBag (_ILIT(0))
667 %************************************************************************
669 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
671 %************************************************************************
673 We use 'couldBeSmallEnoughToInline' to avoid exporting inlinings that
674 we ``couldn't possibly use'' on the other side. Can be overridden w/
675 flaggery. Just the same as smallEnoughToInline, except that it has no
679 couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
680 couldBeSmallEnoughToInline threshold rhs
681 = case sizeExpr (iUnbox threshold) [] body of
685 (_, body) = collectBinders rhs
688 smallEnoughToInline :: Unfolding -> Bool
689 smallEnoughToInline (CoreUnfolding {uf_guidance = UnfIfGoodArgs {ug_size = size}})
690 = size <= opt_UF_UseThreshold
691 smallEnoughToInline _
695 certainlyWillInline :: Unfolding -> Bool
696 -- Sees if the unfolding is pretty certain to inline
697 certainlyWillInline (CoreUnfolding { uf_is_cheap = is_cheap, uf_arity = n_vals, uf_guidance = guidance })
701 UnfIfGoodArgs { ug_size = size}
702 -> is_cheap && size - (n_vals +1) <= opt_UF_UseThreshold
704 certainlyWillInline _
708 %************************************************************************
710 \subsection{callSiteInline}
712 %************************************************************************
714 This is the key function. It decides whether to inline a variable at a call site
716 callSiteInline is used at call sites, so it is a bit more generous.
717 It's a very important function that embodies lots of heuristics.
718 A non-WHNF can be inlined if it doesn't occur inside a lambda,
719 and occurs exactly once or
720 occurs once in each branch of a case and is small
722 If the thing is in WHNF, there's no danger of duplicating work,
723 so we can inline if it occurs once, or is small
725 NOTE: we don't want to inline top-level functions that always diverge.
726 It just makes the code bigger. Tt turns out that the convenient way to prevent
727 them inlining is to give them a NOINLINE pragma, which we do in
728 StrictAnal.addStrictnessInfoToTopId
731 callSiteInline :: DynFlags
733 -> Unfolding -- Its unfolding (if active)
734 -> Bool -- True if there are are no arguments at all (incl type args)
735 -> [ArgSummary] -- One for each value arg; True if it is interesting
736 -> CallCtxt -- True <=> continuation is interesting
737 -> Maybe CoreExpr -- Unfolding, if any
740 instance Outputable ArgSummary where
741 ppr TrivArg = ptext (sLit "TrivArg")
742 ppr NonTrivArg = ptext (sLit "NonTrivArg")
743 ppr ValueArg = ptext (sLit "ValueArg")
745 data CallCtxt = BoringCtxt
747 | ArgCtxt -- We are somewhere in the argument of a function
748 Bool -- True <=> we're somewhere in the RHS of function with rules
749 -- False <=> we *are* the argument of a function with non-zero
752 -- we *are* the RHS of a let Note [RHS of lets]
753 -- In both cases, be a little keener to inline
755 | ValAppCtxt -- We're applied to at least one value arg
756 -- This arises when we have ((f x |> co) y)
757 -- Then the (f x) has argument 'x' but in a ValAppCtxt
759 | CaseCtxt -- We're the scrutinee of a case
760 -- that decomposes its scrutinee
762 instance Outputable CallCtxt where
763 ppr BoringCtxt = ptext (sLit "BoringCtxt")
764 ppr (ArgCtxt rules) = ptext (sLit "ArgCtxt") <+> ppr rules
765 ppr CaseCtxt = ptext (sLit "CaseCtxt")
766 ppr ValAppCtxt = ptext (sLit "ValAppCtxt")
768 callSiteInline dflags id unfolding lone_variable arg_infos cont_info
769 = case unfolding of {
770 NoUnfolding -> Nothing ;
771 OtherCon _ -> Nothing ;
772 DFunUnfolding {} -> Nothing ; -- Never unfold a DFun
773 CoreUnfolding { uf_tmpl = unf_template, uf_is_top = is_top,
774 uf_is_cheap = is_cheap, uf_arity = uf_arity, uf_guidance = guidance } ->
775 -- uf_arity will typically be equal to (idArity id),
776 -- but may be less for InlineRules
778 n_val_args = length arg_infos
779 saturated = n_val_args >= uf_arity
781 result | yes_or_no = Just unf_template
782 | otherwise = Nothing
784 interesting_args = any nonTriv arg_infos
785 -- NB: (any nonTriv arg_infos) looks at the
786 -- over-saturated args too which is "wrong";
787 -- but if over-saturated we inline anyway.
789 -- some_benefit is used when the RHS is small enough
790 -- and the call has enough (or too many) value
791 -- arguments (ie n_val_args >= arity). But there must
792 -- be *something* interesting about some argument, or the
793 -- result context, to make it worth inlining
795 | not saturated = interesting_args -- Under-saturated
796 -- Note [Unsaturated applications]
797 | n_val_args > uf_arity = True -- Over-saturated
798 | otherwise = interesting_args -- Saturated
799 || interesting_saturated_call
801 interesting_saturated_call
803 BoringCtxt -> not is_top && uf_arity > 0 -- Note [Nested functions]
804 CaseCtxt -> not (lone_variable && is_cheap) -- Note [Lone variables]
805 ArgCtxt {} -> uf_arity > 0 -- Note [Inlining in ArgCtxt]
806 ValAppCtxt -> True -- Note [Cast then apply]
808 (yes_or_no, extra_doc)
810 UnfNever -> (False, empty)
812 UnfWhen unsat_ok boring_ok
813 -> (enough_args && (boring_ok || some_benefit), empty )
814 where -- See Note [INLINE for small functions]
815 enough_args = saturated || (unsat_ok && n_val_args > 0)
817 UnfIfGoodArgs { ug_args = arg_discounts, ug_res = res_discount, ug_size = size }
818 -> ( is_cheap && some_benefit && small_enough
819 , (text "discounted size =" <+> int discounted_size) )
821 discounted_size = size - discount
822 small_enough = discounted_size <= opt_UF_UseThreshold
823 discount = computeDiscount uf_arity arg_discounts
824 res_discount arg_infos cont_info
827 if (dopt Opt_D_dump_inlinings dflags && dopt Opt_D_verbose_core2core dflags) then
828 pprTrace ("Considering inlining: " ++ showSDoc (ppr id))
829 (vcat [text "arg infos" <+> ppr arg_infos,
830 text "uf arity" <+> ppr uf_arity,
831 text "interesting continuation" <+> ppr cont_info,
832 text "some_benefit" <+> ppr some_benefit,
833 text "is cheap:" <+> ppr is_cheap,
834 text "guidance" <+> ppr guidance,
836 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO"])
845 Be a tiny bit keener to inline in the RHS of a let, because that might
846 lead to good thing later
848 g y = let x = f y in ...(case x of (a,b,c) -> ...) ...
849 We'd inline 'f' if the call was in a case context, and it kind-of-is,
850 only we can't see it. So we treat the RHS of a let as not-totally-boring.
852 Note [Unsaturated applications]
853 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
854 When a call is not saturated, we *still* inline if one of the
855 arguments has interesting structure. That's sometimes very important.
856 A good example is the Ord instance for Bool in Base:
859 $fOrdBool =GHC.Classes.D:Ord
864 $cmin_ajX [Occ=LoopBreaker] :: Bool -> Bool -> Bool
865 $cmin_ajX = GHC.Classes.$dmmin @ Bool $fOrdBool
868 But the defn of GHC.Classes.$dmmin is:
870 $dmmin :: forall a. GHC.Classes.Ord a => a -> a -> a
871 {- Arity: 3, HasNoCafRefs, Strictness: SLL,
872 Unfolding: (\ @ a $dOrd :: GHC.Classes.Ord a x :: a y :: a ->
873 case @ a GHC.Classes.<= @ a $dOrd x y of wild {
874 GHC.Bool.False -> y GHC.Bool.True -> x }) -}
876 We *really* want to inline $dmmin, even though it has arity 3, in
877 order to unravel the recursion.
880 Note [Things to watch]
881 ~~~~~~~~~~~~~~~~~~~~~~
882 * { y = I# 3; x = y `cast` co; ...case (x `cast` co) of ... }
883 Assume x is exported, so not inlined unconditionally.
884 Then we want x to inline unconditionally; no reason for it
885 not to, and doing so avoids an indirection.
887 * { x = I# 3; ....f x.... }
888 Make sure that x does not inline unconditionally!
889 Lest we get extra allocation.
891 Note [Inlining an InlineRule]
892 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
893 An InlineRules is used for
894 (a) programmer INLINE pragmas
895 (b) inlinings from worker/wrapper
897 For (a) the RHS may be large, and our contract is that we *only* inline
898 when the function is applied to all the arguments on the LHS of the
899 source-code defn. (The uf_arity in the rule.)
901 However for worker/wrapper it may be worth inlining even if the
902 arity is not satisfied (as we do in the CoreUnfolding case) so we don't
906 Note [Nested functions]
907 ~~~~~~~~~~~~~~~~~~~~~~~
908 If a function has a nested defn we also record some-benefit, on the
909 grounds that we are often able to eliminate the binding, and hence the
910 allocation, for the function altogether; this is good for join points.
911 But this only makes sense for *functions*; inlining a constructor
912 doesn't help allocation unless the result is scrutinised. UNLESS the
913 constructor occurs just once, albeit possibly in multiple case
914 branches. Then inlining it doesn't increase allocation, but it does
915 increase the chance that the constructor won't be allocated at all in
916 the branches that don't use it.
918 Note [Cast then apply]
919 ~~~~~~~~~~~~~~~~~~~~~~
921 myIndex = __inline_me ( (/\a. <blah>) |> co )
922 co :: (forall a. a -> a) ~ (forall a. T a)
923 ... /\a.\x. case ((myIndex a) |> sym co) x of { ... } ...
925 We need to inline myIndex to unravel this; but the actual call (myIndex a) has
926 no value arguments. The ValAppCtxt gives it enough incentive to inline.
928 Note [Inlining in ArgCtxt]
929 ~~~~~~~~~~~~~~~~~~~~~~~~~~
930 The condition (arity > 0) here is very important, because otherwise
931 we end up inlining top-level stuff into useless places; eg
934 This can make a very big difference: it adds 16% to nofib 'integer' allocs,
937 At one stage I replaced this condition by 'True' (leading to the above
938 slow-down). The motivation was test eyeball/inline1.hs; but that seems
941 NOTE: arguably, we should inline in ArgCtxt only if the result of the
942 call is at least CONLIKE. At least for the cases where we use ArgCtxt
943 for the RHS of a 'let', we only profit from the inlining if we get a
944 CONLIKE thing (modulo lets).
946 Note [Lone variables] See also Note [Interaction of exprIsCheap and lone variables]
947 ~~~~~~~~~~~~~~~~~~~~~ which appears below
948 The "lone-variable" case is important. I spent ages messing about
949 with unsatisfactory varaints, but this is nice. The idea is that if a
950 variable appears all alone
952 as an arg of lazy fn, or rhs BoringCtxt
953 as scrutinee of a case CaseCtxt
954 as arg of a fn ArgCtxt
956 it is bound to a cheap expression
958 then we should not inline it (unless there is some other reason,
959 e.g. is is the sole occurrence). That is what is happening at
960 the use of 'lone_variable' in 'interesting_saturated_call'.
962 Why? At least in the case-scrutinee situation, turning
963 let x = (a,b) in case x of y -> ...
965 let x = (a,b) in case (a,b) of y -> ...
967 let x = (a,b) in let y = (a,b) in ...
968 is bad if the binding for x will remain.
970 Another example: I discovered that strings
971 were getting inlined straight back into applications of 'error'
972 because the latter is strict.
974 f = \x -> ...(error s)...
976 Fundamentally such contexts should not encourage inlining because the
977 context can ``see'' the unfolding of the variable (e.g. case or a
978 RULE) so there's no gain. If the thing is bound to a value.
983 foo = _inline_ (\n. [n])
984 bar = _inline_ (foo 20)
985 baz = \n. case bar of { (m:_) -> m + n }
986 Here we really want to inline 'bar' so that we can inline 'foo'
987 and the whole thing unravels as it should obviously do. This is
988 important: in the NDP project, 'bar' generates a closure data
989 structure rather than a list.
991 So the non-inlining of lone_variables should only apply if the
992 unfolding is regarded as cheap; because that is when exprIsConApp_maybe
993 looks through the unfolding. Hence the "&& is_cheap" in the
996 * Even a type application or coercion isn't a lone variable.
998 case $fMonadST @ RealWorld of { :DMonad a b c -> c }
999 We had better inline that sucker! The case won't see through it.
1001 For now, I'm treating treating a variable applied to types
1002 in a *lazy* context "lone". The motivating example was
1004 g = /\a. \y. h (f a)
1005 There's no advantage in inlining f here, and perhaps
1006 a significant disadvantage. Hence some_val_args in the Stop case
1008 Note [Interaction of exprIsCheap and lone variables]
1009 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1010 The lone-variable test says "don't inline if a case expression
1011 scrutines a lone variable whose unfolding is cheap". It's very
1012 important that, under these circumstances, exprIsConApp_maybe
1013 can spot a constructor application. So, for example, we don't
1016 to be cheap, and that's good because exprIsConApp_maybe doesn't
1017 think that expression is a constructor application.
1019 I used to test is_value rather than is_cheap, which was utterly
1020 wrong, because the above expression responds True to exprIsHNF.
1022 This kind of thing can occur if you have
1025 foo = let x = e in (x,x)
1030 computeDiscount :: Int -> [Int] -> Int -> [ArgSummary] -> CallCtxt -> Int
1031 computeDiscount n_vals_wanted arg_discounts res_discount arg_infos cont_info
1032 -- We multiple the raw discounts (args_discount and result_discount)
1033 -- ty opt_UnfoldingKeenessFactor because the former have to do with
1034 -- *size* whereas the discounts imply that there's some extra
1035 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
1038 = 1 -- Discount of 1 because the result replaces the call
1039 -- so we count 1 for the function itself
1041 + length (take n_vals_wanted arg_infos)
1042 -- Discount of (un-scaled) 1 for each arg supplied,
1043 -- because the result replaces the call
1045 + round (opt_UF_KeenessFactor *
1046 fromIntegral (arg_discount + res_discount'))
1048 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
1050 mk_arg_discount _ TrivArg = 0
1051 mk_arg_discount _ NonTrivArg = 1
1052 mk_arg_discount discount ValueArg = discount
1054 res_discount' = case cont_info of
1056 CaseCtxt -> res_discount
1057 _other -> 4 `min` res_discount
1058 -- res_discount can be very large when a function returns
1059 -- constructors; but we only want to invoke that large discount
1060 -- when there's a case continuation.
1061 -- Otherwise we, rather arbitrarily, threshold it. Yuk.
1062 -- But we want to aovid inlining large functions that return
1063 -- constructors into contexts that are simply "interesting"
1066 %************************************************************************
1068 Interesting arguments
1070 %************************************************************************
1072 Note [Interesting arguments]
1073 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1074 An argument is interesting if it deserves a discount for unfoldings
1075 with a discount in that argument position. The idea is to avoid
1076 unfolding a function that is applied only to variables that have no
1077 unfolding (i.e. they are probably lambda bound): f x y z There is
1078 little point in inlining f here.
1080 Generally, *values* (like (C a b) and (\x.e)) deserve discounts. But
1081 we must look through lets, eg (let x = e in C a b), because the let will
1082 float, exposing the value, if we inline. That makes it different to
1085 Before 2009 we said it was interesting if the argument had *any* structure
1086 at all; i.e. (hasSomeUnfolding v). But does too much inlining; see Trac #3016.
1088 But we don't regard (f x y) as interesting, unless f is unsaturated.
1089 If it's saturated and f hasn't inlined, then it's probably not going
1092 Note [Conlike is interesting]
1093 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1095 f d = ...((*) d x y)...
1097 where df is con-like. Then we'd really like to inline 'f' so that the
1098 rule for (*) (df d) can fire. To do this
1099 a) we give a discount for being an argument of a class-op (eg (*) d)
1100 b) we say that a con-like argument (eg (df d)) is interesting
1103 data ArgSummary = TrivArg -- Nothing interesting
1104 | NonTrivArg -- Arg has structure
1105 | ValueArg -- Arg is a con-app or PAP
1106 -- ..or con-like. Note [Conlike is interesting]
1108 interestingArg :: CoreExpr -> ArgSummary
1109 -- See Note [Interesting arguments]
1110 interestingArg e = go e 0
1112 -- n is # value args to which the expression is applied
1113 go (Lit {}) _ = ValueArg
1115 | isConLikeId v = ValueArg -- Experimenting with 'conlike' rather that
1116 -- data constructors here
1117 | idArity v > n = ValueArg -- Catches (eg) primops with arity but no unfolding
1118 | n > 0 = NonTrivArg -- Saturated or unknown call
1119 | conlike_unfolding = ValueArg -- n==0; look for an interesting unfolding
1120 -- See Note [Conlike is interesting]
1121 | otherwise = TrivArg -- n==0, no useful unfolding
1123 conlike_unfolding = isConLikeUnfolding (idUnfolding v)
1125 go (Type _) _ = TrivArg
1126 go (App fn (Type _)) n = go fn n
1127 go (App fn _) n = go fn (n+1)
1128 go (Note _ a) n = go a n
1129 go (Cast e _) n = go e n
1131 | isTyCoVar v = go e n
1133 | otherwise = ValueArg
1134 go (Let _ e) n = case go e n of { ValueArg -> ValueArg; _ -> NonTrivArg }
1135 go (Case {}) _ = NonTrivArg
1137 nonTriv :: ArgSummary -> Bool
1138 nonTriv TrivArg = False
1142 %************************************************************************
1146 %************************************************************************
1148 Note [exprIsConApp_maybe]
1149 ~~~~~~~~~~~~~~~~~~~~~~~~~
1150 exprIsConApp_maybe is a very important function. There are two principal
1152 * case e of { .... }
1153 * cls_op e, where cls_op is a class operation
1155 In both cases you want to know if e is of form (C e1..en) where C is
1158 However e might not *look* as if
1161 -- | Returns @Just (dc, [t1..tk], [x1..xn])@ if the argument expression is
1162 -- a *saturated* constructor application of the form @dc t1..tk x1 .. xn@,
1163 -- where t1..tk are the *universally-qantified* type args of 'dc'
1164 exprIsConApp_maybe :: IdUnfoldingFun -> CoreExpr -> Maybe (DataCon, [Type], [CoreExpr])
1166 exprIsConApp_maybe id_unf (Note _ expr)
1167 = exprIsConApp_maybe id_unf expr
1168 -- We ignore all notes. For example,
1169 -- case _scc_ "foo" (C a b) of
1171 -- should be optimised away, but it will be only if we look
1172 -- through the SCC note.
1174 exprIsConApp_maybe id_unf (Cast expr co)
1175 = -- Here we do the KPush reduction rule as described in the FC paper
1176 -- The transformation applies iff we have
1177 -- (C e1 ... en) `cast` co
1178 -- where co :: (T t1 .. tn) ~ to_ty
1179 -- The left-hand one must be a T, because exprIsConApp returned True
1180 -- but the right-hand one might not be. (Though it usually will.)
1182 case exprIsConApp_maybe id_unf expr of {
1183 Nothing -> Nothing ;
1184 Just (dc, _dc_univ_args, dc_args) ->
1186 let (_from_ty, to_ty) = coercionKind co
1187 dc_tc = dataConTyCon dc
1189 case splitTyConApp_maybe to_ty of {
1190 Nothing -> Nothing ;
1191 Just (to_tc, to_tc_arg_tys)
1192 | dc_tc /= to_tc -> Nothing
1193 -- These two Nothing cases are possible; we might see
1194 -- (C x y) `cast` (g :: T a ~ S [a]),
1195 -- where S is a type function. In fact, exprIsConApp
1196 -- will probably not be called in such circumstances,
1197 -- but there't nothing wrong with it
1201 tc_arity = tyConArity dc_tc
1202 dc_univ_tyvars = dataConUnivTyVars dc
1203 dc_ex_tyvars = dataConExTyVars dc
1204 arg_tys = dataConRepArgTys dc
1206 dc_eqs :: [(Type,Type)] -- All equalities from the DataCon
1207 dc_eqs = [(mkTyVarTy tv, ty) | (tv,ty) <- dataConEqSpec dc] ++
1208 [getEqPredTys eq_pred | eq_pred <- dataConEqTheta dc]
1210 (ex_args, rest1) = splitAtList dc_ex_tyvars dc_args
1211 (co_args, val_args) = splitAtList dc_eqs rest1
1213 -- Make the "theta" from Fig 3 of the paper
1214 gammas = decomposeCo tc_arity co
1215 theta = zipOpenTvSubst (dc_univ_tyvars ++ dc_ex_tyvars)
1216 (gammas ++ stripTypeArgs ex_args)
1218 -- Cast the existential coercion arguments
1219 cast_co (ty1, ty2) (Type co)
1220 = Type $ mkSymCoercion (substTy theta ty1)
1221 `mkTransCoercion` co
1222 `mkTransCoercion` (substTy theta ty2)
1223 cast_co _ other_arg = pprPanic "cast_co" (ppr other_arg)
1224 new_co_args = zipWith cast_co dc_eqs co_args
1226 -- Cast the value arguments (which include dictionaries)
1227 new_val_args = zipWith cast_arg arg_tys val_args
1228 cast_arg arg_ty arg = mkCoerce (substTy theta arg_ty) arg
1231 let dump_doc = vcat [ppr dc, ppr dc_univ_tyvars, ppr dc_ex_tyvars,
1232 ppr arg_tys, ppr dc_args, ppr _dc_univ_args,
1233 ppr ex_args, ppr val_args]
1235 ASSERT2( coreEqType _from_ty (mkTyConApp dc_tc _dc_univ_args), dump_doc )
1236 ASSERT2( all isTypeArg (ex_args ++ co_args), dump_doc )
1237 ASSERT2( equalLength val_args arg_tys, dump_doc )
1240 Just (dc, to_tc_arg_tys, ex_args ++ new_co_args ++ new_val_args)
1243 exprIsConApp_maybe id_unf expr
1246 analyse (App fun arg) args = analyse fun (arg:args)
1247 analyse fun@(Lam {}) args = beta fun [] args
1249 analyse (Var fun) args
1250 | Just con <- isDataConWorkId_maybe fun
1251 , count isValArg args == idArity fun
1252 , let (univ_ty_args, rest_args) = splitAtList (dataConUnivTyVars con) args
1253 = Just (con, stripTypeArgs univ_ty_args, rest_args)
1255 -- Look through dictionary functions; see Note [Unfolding DFuns]
1256 | DFunUnfolding dfun_nargs con ops <- unfolding
1257 , let sat = length args == dfun_nargs -- See Note [DFun arity check]
1258 in if sat then True else
1259 pprTrace "Unsaturated dfun" (ppr fun <+> int dfun_nargs $$ ppr args) False
1260 , let (dfun_tvs, _cls, dfun_res_tys) = tcSplitDFunTy (idType fun)
1261 subst = zipOpenTvSubst dfun_tvs (stripTypeArgs (takeList dfun_tvs args))
1262 = Just (con, substTys subst dfun_res_tys,
1263 [mkApps op args | op <- ops])
1265 -- Look through unfoldings, but only cheap ones, because
1266 -- we are effectively duplicating the unfolding
1267 | Just rhs <- expandUnfolding_maybe unfolding
1268 = -- pprTrace "expanding" (ppr fun $$ ppr rhs) $
1271 unfolding = id_unf fun
1273 analyse _ _ = Nothing
1276 beta (Lam v body) pairs (arg : args)
1278 = beta body ((v,arg):pairs) args
1280 beta (Lam {}) _ _ -- Un-saturated, or not a type lambda
1284 = analyse (substExpr (text "subst-expr-is-con-app") subst fun) args
1286 subst = mkOpenSubst (mkInScopeSet (exprFreeVars fun)) pairs
1287 -- doc = vcat [ppr fun, ppr expr, ppr pairs, ppr args]
1290 stripTypeArgs :: [CoreExpr] -> [Type]
1291 stripTypeArgs args = ASSERT2( all isTypeArg args, ppr args )
1292 [ty | Type ty <- args]
1295 Note [Unfolding DFuns]
1296 ~~~~~~~~~~~~~~~~~~~~~~
1299 df :: forall a b. (Eq a, Eq b) -> Eq (a,b)
1300 df a b d_a d_b = MkEqD (a,b) ($c1 a b d_a d_b)
1303 So to split it up we just need to apply the ops $c1, $c2 etc
1304 to the very same args as the dfun. It takes a little more work
1305 to compute the type arguments to the dictionary constructor.
1307 Note [DFun arity check]
1308 ~~~~~~~~~~~~~~~~~~~~~~~
1309 Here we check that the total number of supplied arguments (inclding
1310 type args) matches what the dfun is expecting. This may be *less*
1311 than the ordinary arity of the dfun: see Note [DFun unfoldings] in CoreSyn