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, inlineBoringOk,
30 certainlyWillInline, smallEnoughToInline,
32 callSiteInline, CallCtxt(..),
38 #include "HsVersions.h"
43 import PprCore () -- Instances
44 import TcType ( tcSplitDFunTy )
45 import OccurAnal ( occurAnalyseExpr )
46 import CoreSubst hiding( substTy )
47 import CoreFVs ( exprFreeVars )
48 import CoreArity ( manifestArity, exprBotStrictness_maybe )
56 import BasicTypes ( Arity )
60 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 -> [DFunArg CoreExpr] -> Unfolding
95 mkDFunUnfolding dfun_ty ops
96 = DFunUnfolding dfun_nargs data_con ops
98 (tvs, n_theta, cls, _) = tcSplitDFunTy dfun_ty
99 dfun_nargs = length tvs + n_theta
100 data_con = classDataCon cls
102 mkWwInlineRule :: Id -> CoreExpr -> Arity -> Unfolding
103 mkWwInlineRule id expr arity
104 = mkCoreUnfolding (InlineWrapper id) True
105 (simpleOptExpr expr) arity
106 (UnfWhen unSaturatedOk boringCxtNotOk)
108 mkCompulsoryUnfolding :: CoreExpr -> Unfolding
109 mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
110 = mkCoreUnfolding InlineCompulsory True
111 (simpleOptExpr expr) 0 -- Arity of unfolding doesn't matter
112 (UnfWhen unSaturatedOk boringCxtOk)
114 mkInlineUnfolding :: Maybe Arity -> CoreExpr -> Unfolding
115 mkInlineUnfolding mb_arity expr
116 = mkCoreUnfolding InlineStable
117 True -- Note [Top-level flag on inline rules]
119 (UnfWhen unsat_ok boring_ok)
121 expr' = simpleOptExpr expr
122 (unsat_ok, arity) = case mb_arity of
123 Nothing -> (unSaturatedOk, manifestArity expr')
124 Just ar -> (needSaturated, ar)
126 boring_ok = inlineBoringOk expr'
128 mkInlinableUnfolding :: CoreExpr -> Unfolding
129 mkInlinableUnfolding expr
130 = mkUnfolding InlineStable True is_bot expr'
132 expr' = simpleOptExpr expr
133 is_bot = isJust (exprBotStrictness_maybe expr')
139 mkCoreUnfolding :: UnfoldingSource -> Bool -> CoreExpr
140 -> Arity -> UnfoldingGuidance -> Unfolding
141 -- Occurrence-analyses the expression before capturing it
142 mkCoreUnfolding src top_lvl expr arity guidance
143 = CoreUnfolding { uf_tmpl = occurAnalyseExpr expr,
147 uf_is_value = exprIsHNF expr,
148 uf_is_conlike = exprIsConLike expr,
149 uf_is_cheap = exprIsCheap expr,
150 uf_expandable = exprIsExpandable expr,
151 uf_guidance = guidance }
153 mkUnfolding :: UnfoldingSource -> Bool -> Bool -> CoreExpr -> Unfolding
154 -- Calculates unfolding guidance
155 -- Occurrence-analyses the expression before capturing it
156 mkUnfolding src top_lvl is_bottoming expr
157 | top_lvl && is_bottoming
158 , not (exprIsTrivial expr)
159 = NoUnfolding -- See Note [Do not inline top-level bottoming functions]
161 = CoreUnfolding { uf_tmpl = occurAnalyseExpr expr,
165 uf_is_value = exprIsHNF expr,
166 uf_is_conlike = exprIsConLike expr,
167 uf_expandable = exprIsExpandable expr,
168 uf_is_cheap = is_cheap,
169 uf_guidance = guidance }
171 is_cheap = exprIsCheap expr
172 (arity, guidance) = calcUnfoldingGuidance is_cheap
173 opt_UF_CreationThreshold expr
174 -- Sometimes during simplification, there's a large let-bound thing
175 -- which has been substituted, and so is now dead; so 'expr' contains
176 -- two copies of the thing while the occurrence-analysed expression doesn't
177 -- Nevertheless, we *don't* occ-analyse before computing the size because the
178 -- size computation bales out after a while, whereas occurrence analysis does not.
180 -- This can occasionally mean that the guidance is very pessimistic;
181 -- it gets fixed up next round. And it should be rare, because large
182 -- let-bound things that are dead are usually caught by preInlineUnconditionally
185 %************************************************************************
187 \subsection{The UnfoldingGuidance type}
189 %************************************************************************
192 inlineBoringOk :: CoreExpr -> Bool
193 -- See Note [INLINE for small functions]
194 -- True => the result of inlining the expression is
195 -- no bigger than the expression itself
196 -- eg (\x y -> f y x)
197 -- This is a quick and dirty version. It doesn't attempt
198 -- to deal with (\x y z -> x (y z))
199 -- The really important one is (x `cast` c)
203 go :: Int -> CoreExpr -> Bool
204 go credit (Lam x e) | isId x = go (credit+1) e
205 | otherwise = go credit e
206 go credit (App f (Type {})) = go credit f
207 go credit (App f a) | credit > 0
208 , exprIsTrivial a = go (credit-1) f
209 go credit (Note _ e) = go credit e
210 go credit (Cast e _) = go credit e
211 go _ (Var {}) = boringCxtOk
212 go _ _ = boringCxtNotOk
214 calcUnfoldingGuidance
215 :: Bool -- True <=> the rhs is cheap, or we want to treat it
216 -- as cheap (INLINE things)
217 -> Int -- Bomb out if size gets bigger than this
218 -> CoreExpr -- Expression to look at
219 -> (Arity, UnfoldingGuidance)
220 calcUnfoldingGuidance expr_is_cheap bOMB_OUT_SIZE expr
221 = case collectBinders expr of { (bndrs, body) ->
223 val_bndrs = filter isId bndrs
224 n_val_bndrs = length val_bndrs
227 = case (sizeExpr (iUnbox bOMB_OUT_SIZE) val_bndrs body) of
229 SizeIs size cased_bndrs scrut_discount
230 | uncondInline n_val_bndrs (iBox size)
232 -> UnfWhen unSaturatedOk boringCxtOk -- Note [INLINE for small functions]
234 -> UnfIfGoodArgs { ug_args = map (discount cased_bndrs) val_bndrs
235 , ug_size = iBox size
236 , ug_res = iBox scrut_discount }
239 = foldlBag (\acc (b',n) -> if bndr==b' then acc+n else acc)
242 (n_val_bndrs, guidance) }
245 Note [Computing the size of an expression]
246 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
247 The basic idea of sizeExpr is obvious enough: count nodes. But getting the
248 heuristics right has taken a long time. Here's the basic strategy:
250 * Variables, literals: 0
251 (Exception for string literals, see litSize.)
253 * Function applications (f e1 .. en): 1 + #value args
255 * Constructor applications: 1, regardless of #args
257 * Let(rec): 1 + size of components
272 Notice that 'x' counts 0, while (f x) counts 2. That's deliberate: there's
273 a function call to account for. Notice also that constructor applications
274 are very cheap, because exposing them to a caller is so valuable.
277 Note [Do not inline top-level bottoming functions]
278 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
279 The FloatOut pass has gone to some trouble to float out calls to 'error'
280 and similar friends. See Note [Bottoming floats] in SetLevels.
281 Do not re-inline them! But we *do* still inline if they are very small
282 (the uncondInline stuff).
285 Note [INLINE for small functions]
286 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
287 Consider {-# INLINE f #-}
290 Then f's RHS is no larger than its LHS, so we should inline it into
291 even the most boring context. In general, f the function is
292 sufficiently small that its body is as small as the call itself, the
293 inline unconditionally, regardless of how boring the context is.
297 * We inline *unconditionally* if inlined thing is smaller (using sizeExpr)
298 than the thing it's replacing. Notice that
299 (f x) --> (g 3) -- YES, unconditionally
300 (f x) --> x : [] -- YES, *even though* there are two
301 -- arguments to the cons
305 It's very important not to unconditionally replace a variable by
308 * We do this even if the thing isn't saturated, else we end up with the
312 doesn't inline. Even in a boring context, inlining without being
313 saturated will give a lambda instead of a PAP, and will be more
314 efficient at runtime.
316 * However, when the function's arity > 0, we do insist that it
317 has at least one value argument at the call site. Otherwise we find this:
320 If we inline f here we get
321 d = /\b. MkD (\x:b. x)
322 and then prepareRhs floats out the argument, abstracting the type
323 variables, so we end up with the original again!
327 uncondInline :: Arity -> Int -> Bool
328 -- Inline unconditionally if there no size increase
329 -- Size of call is arity (+1 for the function)
330 -- See Note [INLINE for small functions]
331 uncondInline arity size
332 | arity == 0 = size == 0
333 | otherwise = size <= arity + 1
338 sizeExpr :: FastInt -- Bomb out if it gets bigger than this
339 -> [Id] -- Arguments; we're interested in which of these
344 -- Note [Computing the size of an expression]
346 sizeExpr bOMB_OUT_SIZE top_args expr
349 size_up (Cast e _) = size_up e
350 size_up (Note _ e) = size_up e
351 size_up (Type _) = sizeZero -- Types cost nothing
352 size_up (Coercion _) = sizeZero
353 size_up (Lit lit) = sizeN (litSize lit)
354 size_up (Var f) = size_up_call f [] -- Make sure we get constructor
355 -- discounts even on nullary constructors
357 size_up (App fun (Type _)) = size_up fun
358 size_up (App fun (Coercion _)) = size_up fun
359 size_up (App fun arg) = size_up arg `addSizeNSD`
360 size_up_app fun [arg]
362 size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
363 | otherwise = size_up e
365 size_up (Let (NonRec binder rhs) body)
366 = size_up rhs `addSizeNSD`
367 size_up body `addSizeN`
368 (if isUnLiftedType (idType binder) then 0 else 1)
369 -- For the allocation
370 -- If the binder has an unlifted type there is no allocation
372 size_up (Let (Rec pairs) body)
373 = foldr (addSizeNSD . size_up . snd)
374 (size_up body `addSizeN` length pairs) -- (length pairs) for the allocation
377 size_up (Case (Var v) _ _ alts)
378 | v `elem` top_args -- We are scrutinising an argument variable
379 = alts_size (foldr1 addAltSize alt_sizes)
380 (foldr1 maxSize alt_sizes)
381 -- Good to inline if an arg is scrutinised, because
382 -- that may eliminate allocation in the caller
383 -- And it eliminates the case itself
385 alt_sizes = map size_up_alt alts
387 -- alts_size tries to compute a good discount for
388 -- the case when we are scrutinising an argument variable
389 alts_size (SizeIs tot tot_disc tot_scrut) -- Size of all alternatives
390 (SizeIs max _ _) -- Size of biggest alternative
391 = SizeIs tot (unitBag (v, iBox (_ILIT(2) +# tot -# max)) `unionBags` tot_disc) tot_scrut
392 -- If the variable is known, we produce a discount that
393 -- will take us back to 'max', the size of the largest alternative
394 -- The 1+ is a little discount for reduced allocation in the caller
396 -- Notice though, that we return tot_disc, the total discount from
397 -- all branches. I think that's right.
399 alts_size tot_size _ = tot_size
401 size_up (Case e _ _ alts) = size_up e `addSizeNSD`
402 foldr (addAltSize . size_up_alt) sizeZero alts
403 -- We don't charge for the case itself
404 -- It's a strict thing, and the price of the call
405 -- is paid by scrut. Also consider
406 -- case f x of DEFAULT -> e
407 -- This is just ';'! Don't charge for it.
409 -- Moreover, we charge one per alternative.
412 -- size_up_app is used when there's ONE OR MORE value args
413 size_up_app (App fun arg) args
414 | isTyCoArg arg = size_up_app fun args
415 | otherwise = size_up arg `addSizeNSD`
416 size_up_app fun (arg:args)
417 size_up_app (Var fun) args = size_up_call fun args
418 size_up_app other args = size_up other `addSizeN` length args
421 size_up_call :: Id -> [CoreExpr] -> ExprSize
422 size_up_call fun val_args
423 = case idDetails fun of
424 FCallId _ -> sizeN opt_UF_DearOp
425 DataConWorkId dc -> conSize dc (length val_args)
426 PrimOpId op -> primOpSize op (length val_args)
427 ClassOpId _ -> classOpSize top_args val_args
428 _ -> funSize top_args fun (length val_args)
431 size_up_alt (_con, _bndrs, rhs) = size_up rhs `addSizeN` 1
432 -- Don't charge for args, so that wrappers look cheap
433 -- (See comments about wrappers with Case)
435 -- IMPORATANT: *do* charge 1 for the alternative, else we
436 -- find that giant case nests are treated as practically free
437 -- A good example is Foreign.C.Error.errrnoToIOError
440 -- These addSize things have to be here because
441 -- I don't want to give them bOMB_OUT_SIZE as an argument
442 addSizeN TooBig _ = TooBig
443 addSizeN (SizeIs n xs d) m = mkSizeIs bOMB_OUT_SIZE (n +# iUnbox m) xs d
445 -- addAltSize is used to add the sizes of case alternatives
446 addAltSize TooBig _ = TooBig
447 addAltSize _ TooBig = TooBig
448 addAltSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
449 = mkSizeIs bOMB_OUT_SIZE (n1 +# n2)
451 (d1 +# d2) -- Note [addAltSize result discounts]
453 -- This variant ignores the result discount from its LEFT argument
454 -- It's used when the second argument isn't part of the result
455 addSizeNSD TooBig _ = TooBig
456 addSizeNSD _ TooBig = TooBig
457 addSizeNSD (SizeIs n1 xs _) (SizeIs n2 ys d2)
458 = mkSizeIs bOMB_OUT_SIZE (n1 +# n2)
464 -- | Finds a nominal size of a string literal.
465 litSize :: Literal -> Int
466 -- Used by CoreUnfold.sizeExpr
467 litSize (MachStr str) = 1 + ((lengthFS str + 3) `div` 4)
468 -- If size could be 0 then @f "x"@ might be too small
469 -- [Sept03: make literal strings a bit bigger to avoid fruitless
470 -- duplication of little strings]
471 litSize _other = 0 -- Must match size of nullary constructors
472 -- Key point: if x |-> 4, then x must inline unconditionally
473 -- (eg via case binding)
475 classOpSize :: [Id] -> [CoreExpr] -> ExprSize
476 -- See Note [Conlike is interesting]
479 classOpSize top_args (arg1 : other_args)
480 = SizeIs (iUnbox size) arg_discount (_ILIT(0))
482 size = 2 + length other_args
483 -- If the class op is scrutinising a lambda bound dictionary then
484 -- give it a discount, to encourage the inlining of this function
485 -- The actual discount is rather arbitrarily chosen
486 arg_discount = case arg1 of
487 Var dict | dict `elem` top_args
488 -> unitBag (dict, opt_UF_DictDiscount)
491 funSize :: [Id] -> Id -> Int -> ExprSize
492 -- Size for functions that are not constructors or primops
493 -- Note [Function applications]
494 funSize top_args fun n_val_args
495 | fun `hasKey` buildIdKey = buildSize
496 | fun `hasKey` augmentIdKey = augmentSize
497 | otherwise = SizeIs (iUnbox size) arg_discount (iUnbox res_discount)
499 some_val_args = n_val_args > 0
501 arg_discount | some_val_args && fun `elem` top_args
502 = unitBag (fun, opt_UF_FunAppDiscount)
503 | otherwise = emptyBag
504 -- If the function is an argument and is applied
505 -- to some values, give it an arg-discount
507 res_discount | idArity fun > n_val_args = opt_UF_FunAppDiscount
509 -- If the function is partially applied, show a result discount
511 size | some_val_args = 1 + n_val_args
513 -- The 1+ is for the function itself
514 -- Add 1 for each non-trivial arg;
515 -- the allocation cost, as in let(rec)
518 conSize :: DataCon -> Int -> ExprSize
519 conSize dc n_val_args
520 | n_val_args == 0 = SizeIs (_ILIT(0)) emptyBag (_ILIT(1)) -- Like variables
522 -- See Note [Constructor size]
523 | isUnboxedTupleCon dc = SizeIs (_ILIT(0)) emptyBag (iUnbox n_val_args +# _ILIT(1))
525 -- See Note [Unboxed tuple result discount]
526 -- | isUnboxedTupleCon dc = SizeIs (_ILIT(0)) emptyBag (_ILIT(0))
528 -- See Note [Constructor size]
529 | otherwise = SizeIs (_ILIT(1)) emptyBag (iUnbox n_val_args +# _ILIT(1))
532 Note [Constructor size]
533 ~~~~~~~~~~~~~~~~~~~~~~~
534 Treat a constructors application as size 1, regardless of how many
535 arguments it has; we are keen to expose them (and we charge separately
536 for their args). We can't treat them as size zero, else we find that
537 (Just x) has size 0, which is the same as a lone variable; and hence
538 'v' will always be replaced by (Just x), where v is bound to Just x.
540 However, unboxed tuples count as size zero. I found occasions where we had
541 f x y z = case op# x y z of { s -> (# s, () #) }
542 and f wasn't getting inlined.
544 Note [Unboxed tuple result discount]
545 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
546 I tried giving unboxed tuples a *result discount* of zero (see the
547 commented-out line). Why? When returned as a result they do not
548 allocate, so maybe we don't want to charge so much for them If you
549 have a non-zero discount here, we find that workers often get inlined
550 back into wrappers, because it look like
551 f x = case $wf x of (# a,b #) -> (a,b)
552 and we are keener because of the case. However while this change
553 shrank binary sizes by 0.5% it also made spectral/boyer allocate 5%
554 more. All other changes were very small. So it's not a big deal but I
555 didn't adopt the idea.
558 primOpSize :: PrimOp -> Int -> ExprSize
559 primOpSize op n_val_args
560 | not (primOpIsDupable op) = sizeN opt_UF_DearOp
561 | not (primOpOutOfLine op) = sizeN 1
562 -- Be very keen to inline simple primops.
563 -- We give a discount of 1 for each arg so that (op# x y z) costs 2.
564 -- We can't make it cost 1, else we'll inline let v = (op# x y z)
565 -- at every use of v, which is excessive.
567 -- A good example is:
568 -- let x = +# p q in C {x}
569 -- Even though x get's an occurrence of 'many', its RHS looks cheap,
570 -- and there's a good chance it'll get inlined back into C's RHS. Urgh!
572 | otherwise = sizeN n_val_args
575 buildSize :: ExprSize
576 buildSize = SizeIs (_ILIT(0)) emptyBag (_ILIT(4))
577 -- We really want to inline applications of build
578 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
579 -- Indeed, we should add a result_discount becuause build is
580 -- very like a constructor. We don't bother to check that the
581 -- build is saturated (it usually is). The "-2" discounts for the \c n,
582 -- The "4" is rather arbitrary.
584 augmentSize :: ExprSize
585 augmentSize = SizeIs (_ILIT(0)) emptyBag (_ILIT(4))
586 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
587 -- e plus ys. The -2 accounts for the \cn
589 -- When we return a lambda, give a discount if it's used (applied)
590 lamScrutDiscount :: ExprSize -> ExprSize
591 lamScrutDiscount (SizeIs n vs _) = SizeIs n vs (iUnbox opt_UF_FunAppDiscount)
592 lamScrutDiscount TooBig = TooBig
595 Note [addAltSize result discounts]
596 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
597 When adding the size of alternatives, we *add* the result discounts
598 too, rather than take the *maximum*. For a multi-branch case, this
599 gives a discount for each branch that returns a constructor, making us
600 keener to inline. I did try using 'max' instead, but it makes nofib
601 'rewrite' and 'puzzle' allocate significantly more, and didn't make
602 binary sizes shrink significantly either.
604 Note [Discounts and thresholds]
605 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
606 Constants for discounts and thesholds are defined in main/StaticFlags,
607 all of form opt_UF_xxxx. They are:
609 opt_UF_CreationThreshold (45)
610 At a definition site, if the unfolding is bigger than this, we
611 may discard it altogether
613 opt_UF_UseThreshold (6)
614 At a call site, if the unfolding, less discounts, is smaller than
615 this, then it's small enough inline
617 opt_UF_KeennessFactor (1.5)
618 Factor by which the discounts are multiplied before
619 subtracting from size
621 opt_UF_DictDiscount (1)
622 The discount for each occurrence of a dictionary argument
623 as an argument of a class method. Should be pretty small
624 else big functions may get inlined
626 opt_UF_FunAppDiscount (6)
627 Discount for a function argument that is applied. Quite
628 large, because if we inline we avoid the higher-order call.
631 The size of a foreign call or not-dupable PrimOp
634 Note [Function applications]
635 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
636 In a function application (f a b)
638 - If 'f' is an argument to the function being analysed,
639 and there's at least one value arg, record a FunAppDiscount for f
641 - If the application if a PAP (arity > 2 in this example)
642 record a *result* discount (because inlining
643 with "extra" args in the call may mean that we now
644 get a saturated application)
646 Code for manipulating sizes
649 data ExprSize = TooBig
650 | SizeIs FastInt -- Size found
651 (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
652 FastInt -- Size to subtract if result is scrutinised
653 -- by a case expression
655 instance Outputable ExprSize where
656 ppr TooBig = ptext (sLit "TooBig")
657 ppr (SizeIs a _ c) = brackets (int (iBox a) <+> int (iBox c))
659 -- subtract the discount before deciding whether to bale out. eg. we
660 -- want to inline a large constructor application into a selector:
661 -- tup = (a_1, ..., a_99)
662 -- x = case tup of ...
664 mkSizeIs :: FastInt -> FastInt -> Bag (Id, Int) -> FastInt -> ExprSize
665 mkSizeIs max n xs d | (n -# d) ># max = TooBig
666 | otherwise = SizeIs n xs d
668 maxSize :: ExprSize -> ExprSize -> ExprSize
669 maxSize TooBig _ = TooBig
670 maxSize _ TooBig = TooBig
671 maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
675 sizeN :: Int -> ExprSize
677 sizeZero = SizeIs (_ILIT(0)) emptyBag (_ILIT(0))
678 sizeN n = SizeIs (iUnbox n) emptyBag (_ILIT(0))
682 %************************************************************************
684 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
686 %************************************************************************
688 We use 'couldBeSmallEnoughToInline' to avoid exporting inlinings that
689 we ``couldn't possibly use'' on the other side. Can be overridden w/
690 flaggery. Just the same as smallEnoughToInline, except that it has no
694 couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
695 couldBeSmallEnoughToInline threshold rhs
696 = case sizeExpr (iUnbox threshold) [] body of
700 (_, body) = collectBinders rhs
703 smallEnoughToInline :: Unfolding -> Bool
704 smallEnoughToInline (CoreUnfolding {uf_guidance = UnfIfGoodArgs {ug_size = size}})
705 = size <= opt_UF_UseThreshold
706 smallEnoughToInline _
710 certainlyWillInline :: Unfolding -> Bool
711 -- Sees if the unfolding is pretty certain to inline
712 certainlyWillInline (CoreUnfolding { uf_is_cheap = is_cheap, uf_arity = n_vals, uf_guidance = guidance })
716 UnfIfGoodArgs { ug_size = size}
717 -> is_cheap && size - (n_vals +1) <= opt_UF_UseThreshold
719 certainlyWillInline _
723 %************************************************************************
725 \subsection{callSiteInline}
727 %************************************************************************
729 This is the key function. It decides whether to inline a variable at a call site
731 callSiteInline is used at call sites, so it is a bit more generous.
732 It's a very important function that embodies lots of heuristics.
733 A non-WHNF can be inlined if it doesn't occur inside a lambda,
734 and occurs exactly once or
735 occurs once in each branch of a case and is small
737 If the thing is in WHNF, there's no danger of duplicating work,
738 so we can inline if it occurs once, or is small
740 NOTE: we don't want to inline top-level functions that always diverge.
741 It just makes the code bigger. Tt turns out that the convenient way to prevent
742 them inlining is to give them a NOINLINE pragma, which we do in
743 StrictAnal.addStrictnessInfoToTopId
746 callSiteInline :: DynFlags
748 -> Bool -- True <=> unfolding is active
749 -> Bool -- True if there are are no arguments at all (incl type args)
750 -> [ArgSummary] -- One for each value arg; True if it is interesting
751 -> CallCtxt -- True <=> continuation is interesting
752 -> Maybe CoreExpr -- Unfolding, if any
754 instance Outputable ArgSummary where
755 ppr TrivArg = ptext (sLit "TrivArg")
756 ppr NonTrivArg = ptext (sLit "NonTrivArg")
757 ppr ValueArg = ptext (sLit "ValueArg")
759 data CallCtxt = BoringCtxt
761 | ArgCtxt -- We are somewhere in the argument of a function
762 Bool -- True <=> we're somewhere in the RHS of function with rules
763 -- False <=> we *are* the argument of a function with non-zero
766 -- we *are* the RHS of a let Note [RHS of lets]
767 -- In both cases, be a little keener to inline
769 | ValAppCtxt -- We're applied to at least one value arg
770 -- This arises when we have ((f x |> co) y)
771 -- Then the (f x) has argument 'x' but in a ValAppCtxt
773 | CaseCtxt -- We're the scrutinee of a case
774 -- that decomposes its scrutinee
776 instance Outputable CallCtxt where
777 ppr BoringCtxt = ptext (sLit "BoringCtxt")
778 ppr (ArgCtxt rules) = ptext (sLit "ArgCtxt") <+> ppr rules
779 ppr CaseCtxt = ptext (sLit "CaseCtxt")
780 ppr ValAppCtxt = ptext (sLit "ValAppCtxt")
782 callSiteInline dflags id active_unfolding lone_variable arg_infos cont_info
783 = case idUnfolding id of
784 -- idUnfolding checks for loop-breakers, returning NoUnfolding
785 -- Things with an INLINE pragma may have an unfolding *and*
786 -- be a loop breaker (maybe the knot is not yet untied)
787 CoreUnfolding { uf_tmpl = unf_template, uf_is_top = is_top
788 , uf_is_cheap = is_cheap, uf_arity = uf_arity
789 , uf_guidance = guidance, uf_expandable = is_exp }
790 | active_unfolding -> tryUnfolding dflags id lone_variable
791 arg_infos cont_info unf_template is_top
792 is_cheap is_exp uf_arity guidance
793 | otherwise -> Nothing
794 NoUnfolding -> Nothing
795 OtherCon {} -> Nothing
796 DFunUnfolding {} -> Nothing -- Never unfold a DFun
798 tryUnfolding :: DynFlags -> Id -> Bool -> [ArgSummary] -> CallCtxt
799 -> CoreExpr -> Bool -> Bool -> Bool -> Arity -> UnfoldingGuidance
801 tryUnfolding dflags id lone_variable
802 arg_infos cont_info unf_template is_top
803 is_cheap is_exp uf_arity guidance
804 -- uf_arity will typically be equal to (idArity id),
805 -- but may be less for InlineRules
806 | dopt Opt_D_dump_inlinings dflags && dopt Opt_D_verbose_core2core dflags
807 = pprTrace ("Considering inlining: " ++ showSDoc (ppr id))
808 (vcat [text "arg infos" <+> ppr arg_infos,
809 text "uf arity" <+> ppr uf_arity,
810 text "interesting continuation" <+> ppr cont_info,
811 text "some_benefit" <+> ppr some_benefit,
812 text "is exp:" <+> ppr is_exp,
813 text "is cheap:" <+> ppr is_cheap,
814 text "guidance" <+> ppr guidance,
816 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO"])
821 n_val_args = length arg_infos
822 saturated = n_val_args >= uf_arity
824 result | yes_or_no = Just unf_template
825 | otherwise = Nothing
827 interesting_args = any nonTriv arg_infos
828 -- NB: (any nonTriv arg_infos) looks at the
829 -- over-saturated args too which is "wrong";
830 -- but if over-saturated we inline anyway.
832 -- some_benefit is used when the RHS is small enough
833 -- and the call has enough (or too many) value
834 -- arguments (ie n_val_args >= arity). But there must
835 -- be *something* interesting about some argument, or the
836 -- result context, to make it worth inlining
838 | not saturated = interesting_args -- Under-saturated
839 -- Note [Unsaturated applications]
840 | n_val_args > uf_arity = True -- Over-saturated
841 | otherwise = interesting_args -- Saturated
842 || interesting_saturated_call
844 interesting_saturated_call
846 BoringCtxt -> not is_top && uf_arity > 0 -- Note [Nested functions]
847 CaseCtxt -> not (lone_variable && is_cheap) -- Note [Lone variables]
848 ArgCtxt {} -> uf_arity > 0 -- Note [Inlining in ArgCtxt]
849 ValAppCtxt -> True -- Note [Cast then apply]
851 (yes_or_no, extra_doc)
853 UnfNever -> (False, empty)
855 UnfWhen unsat_ok boring_ok
856 -> (enough_args && (boring_ok || some_benefit), empty )
857 where -- See Note [INLINE for small functions]
858 enough_args = saturated || (unsat_ok && n_val_args > 0)
860 UnfIfGoodArgs { ug_args = arg_discounts, ug_res = res_discount, ug_size = size }
861 -> ( is_cheap && some_benefit && small_enough
862 , (text "discounted size =" <+> int discounted_size) )
864 discounted_size = size - discount
865 small_enough = discounted_size <= opt_UF_UseThreshold
866 discount = computeDiscount uf_arity arg_discounts
867 res_discount arg_infos cont_info
872 Be a tiny bit keener to inline in the RHS of a let, because that might
873 lead to good thing later
875 g y = let x = f y in ...(case x of (a,b,c) -> ...) ...
876 We'd inline 'f' if the call was in a case context, and it kind-of-is,
877 only we can't see it. So we treat the RHS of a let as not-totally-boring.
879 Note [Unsaturated applications]
880 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
881 When a call is not saturated, we *still* inline if one of the
882 arguments has interesting structure. That's sometimes very important.
883 A good example is the Ord instance for Bool in Base:
886 $fOrdBool =GHC.Classes.D:Ord
891 $cmin_ajX [Occ=LoopBreaker] :: Bool -> Bool -> Bool
892 $cmin_ajX = GHC.Classes.$dmmin @ Bool $fOrdBool
895 But the defn of GHC.Classes.$dmmin is:
897 $dmmin :: forall a. GHC.Classes.Ord a => a -> a -> a
898 {- Arity: 3, HasNoCafRefs, Strictness: SLL,
899 Unfolding: (\ @ a $dOrd :: GHC.Classes.Ord a x :: a y :: a ->
900 case @ a GHC.Classes.<= @ a $dOrd x y of wild {
901 GHC.Types.False -> y GHC.Types.True -> x }) -}
903 We *really* want to inline $dmmin, even though it has arity 3, in
904 order to unravel the recursion.
907 Note [Things to watch]
908 ~~~~~~~~~~~~~~~~~~~~~~
909 * { y = I# 3; x = y `cast` co; ...case (x `cast` co) of ... }
910 Assume x is exported, so not inlined unconditionally.
911 Then we want x to inline unconditionally; no reason for it
912 not to, and doing so avoids an indirection.
914 * { x = I# 3; ....f x.... }
915 Make sure that x does not inline unconditionally!
916 Lest we get extra allocation.
918 Note [Inlining an InlineRule]
919 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
920 An InlineRules is used for
921 (a) programmer INLINE pragmas
922 (b) inlinings from worker/wrapper
924 For (a) the RHS may be large, and our contract is that we *only* inline
925 when the function is applied to all the arguments on the LHS of the
926 source-code defn. (The uf_arity in the rule.)
928 However for worker/wrapper it may be worth inlining even if the
929 arity is not satisfied (as we do in the CoreUnfolding case) so we don't
933 Note [Nested functions]
934 ~~~~~~~~~~~~~~~~~~~~~~~
935 If a function has a nested defn we also record some-benefit, on the
936 grounds that we are often able to eliminate the binding, and hence the
937 allocation, for the function altogether; this is good for join points.
938 But this only makes sense for *functions*; inlining a constructor
939 doesn't help allocation unless the result is scrutinised. UNLESS the
940 constructor occurs just once, albeit possibly in multiple case
941 branches. Then inlining it doesn't increase allocation, but it does
942 increase the chance that the constructor won't be allocated at all in
943 the branches that don't use it.
945 Note [Cast then apply]
946 ~~~~~~~~~~~~~~~~~~~~~~
948 myIndex = __inline_me ( (/\a. <blah>) |> co )
949 co :: (forall a. a -> a) ~ (forall a. T a)
950 ... /\a.\x. case ((myIndex a) |> sym co) x of { ... } ...
952 We need to inline myIndex to unravel this; but the actual call (myIndex a) has
953 no value arguments. The ValAppCtxt gives it enough incentive to inline.
955 Note [Inlining in ArgCtxt]
956 ~~~~~~~~~~~~~~~~~~~~~~~~~~
957 The condition (arity > 0) here is very important, because otherwise
958 we end up inlining top-level stuff into useless places; eg
961 This can make a very big difference: it adds 16% to nofib 'integer' allocs,
964 At one stage I replaced this condition by 'True' (leading to the above
965 slow-down). The motivation was test eyeball/inline1.hs; but that seems
968 NOTE: arguably, we should inline in ArgCtxt only if the result of the
969 call is at least CONLIKE. At least for the cases where we use ArgCtxt
970 for the RHS of a 'let', we only profit from the inlining if we get a
971 CONLIKE thing (modulo lets).
973 Note [Lone variables] See also Note [Interaction of exprIsCheap and lone variables]
974 ~~~~~~~~~~~~~~~~~~~~~ which appears below
975 The "lone-variable" case is important. I spent ages messing about
976 with unsatisfactory varaints, but this is nice. The idea is that if a
977 variable appears all alone
979 as an arg of lazy fn, or rhs BoringCtxt
980 as scrutinee of a case CaseCtxt
981 as arg of a fn ArgCtxt
983 it is bound to a cheap expression
985 then we should not inline it (unless there is some other reason,
986 e.g. is is the sole occurrence). That is what is happening at
987 the use of 'lone_variable' in 'interesting_saturated_call'.
989 Why? At least in the case-scrutinee situation, turning
990 let x = (a,b) in case x of y -> ...
992 let x = (a,b) in case (a,b) of y -> ...
994 let x = (a,b) in let y = (a,b) in ...
995 is bad if the binding for x will remain.
997 Another example: I discovered that strings
998 were getting inlined straight back into applications of 'error'
999 because the latter is strict.
1001 f = \x -> ...(error s)...
1003 Fundamentally such contexts should not encourage inlining because the
1004 context can ``see'' the unfolding of the variable (e.g. case or a
1005 RULE) so there's no gain. If the thing is bound to a value.
1010 foo = _inline_ (\n. [n])
1011 bar = _inline_ (foo 20)
1012 baz = \n. case bar of { (m:_) -> m + n }
1013 Here we really want to inline 'bar' so that we can inline 'foo'
1014 and the whole thing unravels as it should obviously do. This is
1015 important: in the NDP project, 'bar' generates a closure data
1016 structure rather than a list.
1018 So the non-inlining of lone_variables should only apply if the
1019 unfolding is regarded as cheap; because that is when exprIsConApp_maybe
1020 looks through the unfolding. Hence the "&& is_cheap" in the
1023 * Even a type application or coercion isn't a lone variable.
1025 case $fMonadST @ RealWorld of { :DMonad a b c -> c }
1026 We had better inline that sucker! The case won't see through it.
1028 For now, I'm treating treating a variable applied to types
1029 in a *lazy* context "lone". The motivating example was
1031 g = /\a. \y. h (f a)
1032 There's no advantage in inlining f here, and perhaps
1033 a significant disadvantage. Hence some_val_args in the Stop case
1035 Note [Interaction of exprIsCheap and lone variables]
1036 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1037 The lone-variable test says "don't inline if a case expression
1038 scrutines a lone variable whose unfolding is cheap". It's very
1039 important that, under these circumstances, exprIsConApp_maybe
1040 can spot a constructor application. So, for example, we don't
1043 to be cheap, and that's good because exprIsConApp_maybe doesn't
1044 think that expression is a constructor application.
1046 I used to test is_value rather than is_cheap, which was utterly
1047 wrong, because the above expression responds True to exprIsHNF.
1049 This kind of thing can occur if you have
1052 foo = let x = e in (x,x)
1057 computeDiscount :: Int -> [Int] -> Int -> [ArgSummary] -> CallCtxt -> Int
1058 computeDiscount n_vals_wanted arg_discounts res_discount arg_infos cont_info
1059 -- We multiple the raw discounts (args_discount and result_discount)
1060 -- ty opt_UnfoldingKeenessFactor because the former have to do with
1061 -- *size* whereas the discounts imply that there's some extra
1062 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
1065 = 1 -- Discount of 1 because the result replaces the call
1066 -- so we count 1 for the function itself
1068 + length (take n_vals_wanted arg_infos)
1069 -- Discount of (un-scaled) 1 for each arg supplied,
1070 -- because the result replaces the call
1072 + round (opt_UF_KeenessFactor *
1073 fromIntegral (arg_discount + res_discount'))
1075 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
1077 mk_arg_discount _ TrivArg = 0
1078 mk_arg_discount _ NonTrivArg = 1
1079 mk_arg_discount discount ValueArg = discount
1081 res_discount' = case cont_info of
1083 CaseCtxt -> res_discount
1084 _other -> 4 `min` res_discount
1085 -- res_discount can be very large when a function returns
1086 -- constructors; but we only want to invoke that large discount
1087 -- when there's a case continuation.
1088 -- Otherwise we, rather arbitrarily, threshold it. Yuk.
1089 -- But we want to aovid inlining large functions that return
1090 -- constructors into contexts that are simply "interesting"
1093 %************************************************************************
1095 Interesting arguments
1097 %************************************************************************
1099 Note [Interesting arguments]
1100 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1101 An argument is interesting if it deserves a discount for unfoldings
1102 with a discount in that argument position. The idea is to avoid
1103 unfolding a function that is applied only to variables that have no
1104 unfolding (i.e. they are probably lambda bound): f x y z There is
1105 little point in inlining f here.
1107 Generally, *values* (like (C a b) and (\x.e)) deserve discounts. But
1108 we must look through lets, eg (let x = e in C a b), because the let will
1109 float, exposing the value, if we inline. That makes it different to
1112 Before 2009 we said it was interesting if the argument had *any* structure
1113 at all; i.e. (hasSomeUnfolding v). But does too much inlining; see Trac #3016.
1115 But we don't regard (f x y) as interesting, unless f is unsaturated.
1116 If it's saturated and f hasn't inlined, then it's probably not going
1119 Note [Conlike is interesting]
1120 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1122 f d = ...((*) d x y)...
1124 where df is con-like. Then we'd really like to inline 'f' so that the
1125 rule for (*) (df d) can fire. To do this
1126 a) we give a discount for being an argument of a class-op (eg (*) d)
1127 b) we say that a con-like argument (eg (df d)) is interesting
1130 data ArgSummary = TrivArg -- Nothing interesting
1131 | NonTrivArg -- Arg has structure
1132 | ValueArg -- Arg is a con-app or PAP
1133 -- ..or con-like. Note [Conlike is interesting]
1135 interestingArg :: CoreExpr -> ArgSummary
1136 -- See Note [Interesting arguments]
1137 interestingArg e = go e 0
1139 -- n is # value args to which the expression is applied
1140 go (Lit {}) _ = ValueArg
1142 | isConLikeId v = ValueArg -- Experimenting with 'conlike' rather that
1143 -- data constructors here
1144 | idArity v > n = ValueArg -- Catches (eg) primops with arity but no unfolding
1145 | n > 0 = NonTrivArg -- Saturated or unknown call
1146 | conlike_unfolding = ValueArg -- n==0; look for an interesting unfolding
1147 -- See Note [Conlike is interesting]
1148 | otherwise = TrivArg -- n==0, no useful unfolding
1150 conlike_unfolding = isConLikeUnfolding (idUnfolding v)
1152 go (Type _) _ = TrivArg
1153 go (Coercion _) _ = TrivArg
1154 go (App fn (Type _)) n = go fn n
1155 go (App fn (Coercion _)) n = go fn n
1156 go (App fn _) n = go fn (n+1)
1157 go (Note _ a) n = go a n
1158 go (Cast e _) n = go e n
1160 | isTyVar v = go e n
1162 | otherwise = ValueArg
1163 go (Let _ e) n = case go e n of { ValueArg -> ValueArg; _ -> NonTrivArg }
1164 go (Case {}) _ = NonTrivArg
1166 nonTriv :: ArgSummary -> Bool
1167 nonTriv TrivArg = False
1171 %************************************************************************
1175 %************************************************************************
1177 Note [exprIsConApp_maybe]
1178 ~~~~~~~~~~~~~~~~~~~~~~~~~
1179 exprIsConApp_maybe is a very important function. There are two principal
1181 * case e of { .... }
1182 * cls_op e, where cls_op is a class operation
1184 In both cases you want to know if e is of form (C e1..en) where C is
1187 However e might not *look* as if
1190 -- | Returns @Just (dc, [t1..tk], [x1..xn])@ if the argument expression is
1191 -- a *saturated* constructor application of the form @dc t1..tk x1 .. xn@,
1192 -- where t1..tk are the *universally-qantified* type args of 'dc'
1193 exprIsConApp_maybe :: IdUnfoldingFun -> CoreExpr -> Maybe (DataCon, [Type], [CoreExpr])
1195 exprIsConApp_maybe id_unf (Note note expr)
1197 = exprIsConApp_maybe id_unf expr
1198 -- We ignore all notes except SCCs. For example,
1199 -- case _scc_ "foo" (C a b) of
1201 -- should not be optimised away, because we'll lose the
1202 -- entry count on 'foo'; see Trac #4414
1204 exprIsConApp_maybe id_unf (Cast expr co)
1205 = -- Here we do the KPush reduction rule as described in the FC paper
1206 -- The transformation applies iff we have
1207 -- (C e1 ... en) `cast` co
1208 -- where co :: (T t1 .. tn) ~ to_ty
1209 -- The left-hand one must be a T, because exprIsConApp returned True
1210 -- but the right-hand one might not be. (Though it usually will.)
1212 case exprIsConApp_maybe id_unf expr of {
1213 Nothing -> Nothing ;
1214 Just (dc, _dc_univ_args, dc_args) ->
1216 let Pair _from_ty to_ty = coercionKind co
1217 dc_tc = dataConTyCon dc
1219 case splitTyConApp_maybe to_ty of {
1220 Nothing -> Nothing ;
1221 Just (to_tc, to_tc_arg_tys)
1222 | dc_tc /= to_tc -> Nothing
1223 -- These two Nothing cases are possible; we might see
1224 -- (C x y) `cast` (g :: T a ~ S [a]),
1225 -- where S is a type function. In fact, exprIsConApp
1226 -- will probably not be called in such circumstances,
1227 -- but there't nothing wrong with it
1231 tc_arity = tyConArity dc_tc
1232 dc_univ_tyvars = dataConUnivTyVars dc
1233 dc_ex_tyvars = dataConExTyVars dc
1234 arg_tys = dataConRepArgTys dc
1236 (ex_args, val_args) = splitAtList dc_ex_tyvars dc_args
1238 -- Make the "theta" from Fig 3 of the paper
1239 gammas = decomposeCo tc_arity co
1240 theta = zipOpenCvSubst (dc_univ_tyvars ++ dc_ex_tyvars)
1241 (gammas ++ map mkReflCo (stripTypeArgs ex_args))
1243 -- Cast the value arguments (which include dictionaries)
1244 new_val_args = zipWith cast_arg arg_tys val_args
1245 cast_arg arg_ty arg = mkCoerce (liftCoSubst theta arg_ty) arg
1248 let dump_doc = vcat [ppr dc, ppr dc_univ_tyvars, ppr dc_ex_tyvars,
1249 ppr arg_tys, ppr dc_args, ppr _dc_univ_args,
1250 ppr ex_args, ppr val_args]
1252 ASSERT2( eqType _from_ty (mkTyConApp dc_tc _dc_univ_args), dump_doc )
1253 ASSERT2( all isTypeArg ex_args, dump_doc )
1254 ASSERT2( equalLength val_args arg_tys, dump_doc )
1257 Just (dc, to_tc_arg_tys, ex_args ++ new_val_args)
1260 exprIsConApp_maybe id_unf expr
1263 analyse (App fun arg) args = analyse fun (arg:args)
1264 analyse fun@(Lam {}) args = beta fun [] args
1266 analyse (Var fun) args
1267 | Just con <- isDataConWorkId_maybe fun
1268 , count isValArg args == idArity fun
1269 , let (univ_ty_args, rest_args) = splitAtList (dataConUnivTyVars con) args
1270 = Just (con, stripTypeArgs univ_ty_args, rest_args)
1272 -- Look through dictionary functions; see Note [Unfolding DFuns]
1273 | DFunUnfolding dfun_nargs con ops <- unfolding
1274 , let sat = length args == dfun_nargs -- See Note [DFun arity check]
1275 in if sat then True else
1276 pprTrace "Unsaturated dfun" (ppr fun <+> int dfun_nargs $$ ppr args) False
1277 , let (dfun_tvs, _n_theta, _cls, dfun_res_tys) = tcSplitDFunTy (idType fun)
1278 subst = zipOpenTvSubst dfun_tvs (stripTypeArgs (takeList dfun_tvs args))
1279 mk_arg (DFunConstArg e) = e
1280 mk_arg (DFunLamArg i) = args !! i
1281 mk_arg (DFunPolyArg e) = mkApps e args
1282 = Just (con, substTys subst dfun_res_tys, map mk_arg ops)
1284 -- Look through unfoldings, but only cheap ones, because
1285 -- we are effectively duplicating the unfolding
1286 | Just rhs <- expandUnfolding_maybe unfolding
1287 = -- pprTrace "expanding" (ppr fun $$ ppr rhs) $
1290 unfolding = id_unf fun
1292 analyse _ _ = Nothing
1295 beta (Lam v body) pairs (arg : args)
1297 = beta body ((v,arg):pairs) args
1299 beta (Lam {}) _ _ -- Un-saturated, or not a type lambda
1303 = analyse (substExpr (text "subst-expr-is-con-app") subst fun) args
1305 subst = mkOpenSubst (mkInScopeSet (exprFreeVars fun)) pairs
1306 -- doc = vcat [ppr fun, ppr expr, ppr pairs, ppr args]
1308 stripTypeArgs :: [CoreExpr] -> [Type]
1309 stripTypeArgs args = ASSERT2( all isTypeArg args, ppr args )
1310 [ty | Type ty <- args]
1311 -- We really do want isTypeArg here, not isTyCoArg!
1314 Note [Unfolding DFuns]
1315 ~~~~~~~~~~~~~~~~~~~~~~
1318 df :: forall a b. (Eq a, Eq b) -> Eq (a,b)
1319 df a b d_a d_b = MkEqD (a,b) ($c1 a b d_a d_b)
1322 So to split it up we just need to apply the ops $c1, $c2 etc
1323 to the very same args as the dfun. It takes a little more work
1324 to compute the type arguments to the dictionary constructor.
1326 Note [DFun arity check]
1327 ~~~~~~~~~~~~~~~~~~~~~~~
1328 Here we check that the total number of supplied arguments (inclding
1329 type args) matches what the dfun is expecting. This may be *less*
1330 than the ordinary arity of the dfun: see Note [DFun unfoldings] in CoreSyn