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 )
73 %************************************************************************
75 \subsection{Making unfoldings}
77 %************************************************************************
80 mkTopUnfolding :: Bool -> CoreExpr -> Unfolding
81 mkTopUnfolding = mkUnfolding InlineRhs True {- Top level -}
83 mkImplicitUnfolding :: CoreExpr -> Unfolding
84 -- For implicit Ids, do a tiny bit of optimising first
85 mkImplicitUnfolding expr = mkTopUnfolding False (simpleOptExpr expr)
87 -- Note [Top-level flag on inline rules]
88 -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
89 -- Slight hack: note that mk_inline_rules conservatively sets the
90 -- top-level flag to True. It gets set more accurately by the simplifier
91 -- Simplify.simplUnfolding.
93 mkSimpleUnfolding :: CoreExpr -> Unfolding
94 mkSimpleUnfolding = mkUnfolding InlineRhs False False
96 mkDFunUnfolding :: Type -> [DFunArg CoreExpr] -> Unfolding
97 mkDFunUnfolding dfun_ty ops
98 = DFunUnfolding dfun_nargs data_con ops
100 (tvs, n_theta, cls, _) = tcSplitDFunTy dfun_ty
101 dfun_nargs = length tvs + n_theta
102 data_con = classDataCon cls
104 mkWwInlineRule :: Id -> CoreExpr -> Arity -> Unfolding
105 mkWwInlineRule id expr arity
106 = mkCoreUnfolding (InlineWrapper id) True
107 (simpleOptExpr expr) arity
108 (UnfWhen unSaturatedOk boringCxtNotOk)
110 mkCompulsoryUnfolding :: CoreExpr -> Unfolding
111 mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
112 = mkCoreUnfolding InlineCompulsory True
113 (simpleOptExpr expr) 0 -- Arity of unfolding doesn't matter
114 (UnfWhen unSaturatedOk boringCxtOk)
116 mkInlineUnfolding :: Maybe Arity -> CoreExpr -> Unfolding
117 mkInlineUnfolding mb_arity expr
118 = mkCoreUnfolding InlineStable
119 True -- Note [Top-level flag on inline rules]
121 (UnfWhen unsat_ok boring_ok)
123 expr' = simpleOptExpr expr
124 (unsat_ok, arity) = case mb_arity of
125 Nothing -> (unSaturatedOk, manifestArity expr')
126 Just ar -> (needSaturated, ar)
128 boring_ok = inlineBoringOk expr'
130 mkInlinableUnfolding :: CoreExpr -> Unfolding
131 mkInlinableUnfolding expr
132 = mkUnfolding InlineStable True is_bot expr'
134 expr' = simpleOptExpr expr
135 is_bot = isJust (exprBotStrictness_maybe expr')
141 mkCoreUnfolding :: UnfoldingSource -> Bool -> CoreExpr
142 -> Arity -> UnfoldingGuidance -> Unfolding
143 -- Occurrence-analyses the expression before capturing it
144 mkCoreUnfolding src top_lvl expr arity guidance
145 = CoreUnfolding { uf_tmpl = occurAnalyseExpr expr,
149 uf_is_value = exprIsHNF expr,
150 uf_is_conlike = exprIsConLike expr,
151 uf_is_cheap = exprIsCheap expr,
152 uf_expandable = exprIsExpandable expr,
153 uf_guidance = guidance }
155 mkUnfolding :: UnfoldingSource -> Bool -> Bool -> CoreExpr -> Unfolding
156 -- Calculates unfolding guidance
157 -- Occurrence-analyses the expression before capturing it
158 mkUnfolding src top_lvl is_bottoming expr
159 | top_lvl && is_bottoming
160 , not (exprIsTrivial expr)
161 = NoUnfolding -- See Note [Do not inline top-level bottoming functions]
163 = CoreUnfolding { uf_tmpl = occurAnalyseExpr expr,
167 uf_is_value = exprIsHNF expr,
168 uf_is_conlike = exprIsConLike expr,
169 uf_expandable = exprIsExpandable expr,
170 uf_is_cheap = is_cheap,
171 uf_guidance = guidance }
173 is_cheap = exprIsCheap expr
174 (arity, guidance) = calcUnfoldingGuidance is_cheap
175 opt_UF_CreationThreshold expr
176 -- Sometimes during simplification, there's a large let-bound thing
177 -- which has been substituted, and so is now dead; so 'expr' contains
178 -- two copies of the thing while the occurrence-analysed expression doesn't
179 -- Nevertheless, we *don't* occ-analyse before computing the size because the
180 -- size computation bales out after a while, whereas occurrence analysis does not.
182 -- This can occasionally mean that the guidance is very pessimistic;
183 -- it gets fixed up next round. And it should be rare, because large
184 -- let-bound things that are dead are usually caught by preInlineUnconditionally
187 %************************************************************************
189 \subsection{The UnfoldingGuidance type}
191 %************************************************************************
194 inlineBoringOk :: CoreExpr -> Bool
195 -- See Note [INLINE for small functions]
196 -- True => the result of inlining the expression is
197 -- no bigger than the expression itself
198 -- eg (\x y -> f y x)
199 -- This is a quick and dirty version. It doesn't attempt
200 -- to deal with (\x y z -> x (y z))
201 -- The really important one is (x `cast` c)
205 go :: Int -> CoreExpr -> Bool
206 go credit (Lam x e) | isId x = go (credit+1) e
207 | otherwise = go credit e
208 go credit (App f (Type {})) = go credit f
209 go credit (App f a) | credit > 0
210 , exprIsTrivial a = go (credit-1) f
211 go credit (Note _ e) = go credit e
212 go credit (Cast e _) = go credit e
213 go _ (Var {}) = boringCxtOk
214 go _ _ = boringCxtNotOk
216 calcUnfoldingGuidance
217 :: Bool -- True <=> the rhs is cheap, or we want to treat it
218 -- as cheap (INLINE things)
219 -> Int -- Bomb out if size gets bigger than this
220 -> CoreExpr -- Expression to look at
221 -> (Arity, UnfoldingGuidance)
222 calcUnfoldingGuidance expr_is_cheap bOMB_OUT_SIZE expr
223 = case collectBinders expr of { (bndrs, body) ->
225 val_bndrs = filter isId bndrs
226 n_val_bndrs = length val_bndrs
229 = case (sizeExpr (iUnbox bOMB_OUT_SIZE) val_bndrs body) of
231 SizeIs size cased_bndrs scrut_discount
232 | uncondInline n_val_bndrs (iBox size)
234 -> UnfWhen unSaturatedOk boringCxtOk -- Note [INLINE for small functions]
236 -> UnfIfGoodArgs { ug_args = map (discount cased_bndrs) val_bndrs
237 , ug_size = iBox size
238 , ug_res = iBox scrut_discount }
241 = foldlBag (\acc (b',n) -> if bndr==b' then acc+n else acc)
244 (n_val_bndrs, guidance) }
247 Note [Computing the size of an expression]
248 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
249 The basic idea of sizeExpr is obvious enough: count nodes. But getting the
250 heuristics right has taken a long time. Here's the basic strategy:
252 * Variables, literals: 0
253 (Exception for string literals, see litSize.)
255 * Function applications (f e1 .. en): 1 + #value args
257 * Constructor applications: 1, regardless of #args
259 * Let(rec): 1 + size of components
274 Notice that 'x' counts 0, while (f x) counts 2. That's deliberate: there's
275 a function call to account for. Notice also that constructor applications
276 are very cheap, because exposing them to a caller is so valuable.
279 Note [Do not inline top-level bottoming functions]
280 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
281 The FloatOut pass has gone to some trouble to float out calls to 'error'
282 and similar friends. See Note [Bottoming floats] in SetLevels.
283 Do not re-inline them! But we *do* still inline if they are very small
284 (the uncondInline stuff).
287 Note [INLINE for small functions]
288 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
289 Consider {-# INLINE f #-}
292 Then f's RHS is no larger than its LHS, so we should inline it into
293 even the most boring context. In general, f the function is
294 sufficiently small that its body is as small as the call itself, the
295 inline unconditionally, regardless of how boring the context is.
299 * We inline *unconditionally* if inlined thing is smaller (using sizeExpr)
300 than the thing it's replacing. Notice that
301 (f x) --> (g 3) -- YES, unconditionally
302 (f x) --> x : [] -- YES, *even though* there are two
303 -- arguments to the cons
307 It's very important not to unconditionally replace a variable by
310 * We do this even if the thing isn't saturated, else we end up with the
314 doesn't inline. Even in a boring context, inlining without being
315 saturated will give a lambda instead of a PAP, and will be more
316 efficient at runtime.
318 * However, when the function's arity > 0, we do insist that it
319 has at least one value argument at the call site. Otherwise we find this:
322 If we inline f here we get
323 d = /\b. MkD (\x:b. x)
324 and then prepareRhs floats out the argument, abstracting the type
325 variables, so we end up with the original again!
329 uncondInline :: Arity -> Int -> Bool
330 -- Inline unconditionally if there no size increase
331 -- Size of call is arity (+1 for the function)
332 -- See Note [INLINE for small functions]
333 uncondInline arity size
334 | arity == 0 = size == 0
335 | otherwise = size <= arity + 1
340 sizeExpr :: FastInt -- Bomb out if it gets bigger than this
341 -> [Id] -- Arguments; we're interested in which of these
346 -- Note [Computing the size of an expression]
348 sizeExpr bOMB_OUT_SIZE top_args expr
351 size_up (Cast e _) = size_up e
352 size_up (Note _ e) = size_up e
353 size_up (Type _) = sizeZero -- Types cost nothing
354 size_up (Coercion _) = sizeZero
355 size_up (Lit lit) = sizeN (litSize lit)
356 size_up (Var f) = size_up_call f [] -- Make sure we get constructor
357 -- discounts even on nullary constructors
359 size_up (App fun (Type _)) = size_up fun
360 size_up (App fun (Coercion _)) = size_up fun
361 size_up (App fun arg) = size_up arg `addSizeNSD`
362 size_up_app fun [arg]
364 size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
365 | otherwise = size_up e
367 size_up (Let (NonRec binder rhs) body)
368 = size_up rhs `addSizeNSD`
369 size_up body `addSizeN`
370 (if isUnLiftedType (idType binder) then 0 else 1)
371 -- For the allocation
372 -- If the binder has an unlifted type there is no allocation
374 size_up (Let (Rec pairs) body)
375 = foldr (addSizeNSD . size_up . snd)
376 (size_up body `addSizeN` length pairs) -- (length pairs) for the allocation
379 size_up (Case (Var v) _ _ alts)
380 | v `elem` top_args -- We are scrutinising an argument variable
381 = alts_size (foldr1 addAltSize alt_sizes)
382 (foldr1 maxSize alt_sizes)
383 -- Good to inline if an arg is scrutinised, because
384 -- that may eliminate allocation in the caller
385 -- And it eliminates the case itself
387 alt_sizes = map size_up_alt alts
389 -- alts_size tries to compute a good discount for
390 -- the case when we are scrutinising an argument variable
391 alts_size (SizeIs tot tot_disc tot_scrut) -- Size of all alternatives
392 (SizeIs max _ _) -- Size of biggest alternative
393 = SizeIs tot (unitBag (v, iBox (_ILIT(2) +# tot -# max)) `unionBags` tot_disc) tot_scrut
394 -- If the variable is known, we produce a discount that
395 -- will take us back to 'max', the size of the largest alternative
396 -- The 1+ is a little discount for reduced allocation in the caller
398 -- Notice though, that we return tot_disc, the total discount from
399 -- all branches. I think that's right.
401 alts_size tot_size _ = tot_size
403 size_up (Case e b _ alts) = size_up e `addSizeNSD`
404 foldr (addAltSize . size_up_alt) case_size alts
407 | is_inline_scrut e, not (lengthExceeds alts 1) = sizeN (-1)
408 | otherwise = sizeZero
409 -- Normally we don't charge for the case itself, but
410 -- we charge one per alternative (see size_up_alt,
411 -- below) to account for the cost of the info table
414 -- However, in certain cases (see is_inline_scrut
415 -- below), no code is generated for the case unless
416 -- there are multiple alts. In these cases we
417 -- subtract one, making the first alt free.
418 -- e.g. case x# +# y# of _ -> ... should cost 1
419 -- case touch# x# of _ -> ... should cost 0
422 -- I would like to not have the "not (lengthExceeds alts 1)"
423 -- condition above, but without that some programs got worse
424 -- (spectral/hartel/event and spectral/para). I don't fully
425 -- understand why. (SDM 24/5/11)
427 -- unboxed variables, inline primops and unsafe foreign calls
428 -- are all "inline" things:
429 is_inline_scrut (Var v) = isUnLiftedType (idType v)
430 is_inline_scrut scrut
431 | (Var f, _) <- collectArgs scrut
432 = case idDetails f of
433 FCallId fc -> not (isSafeForeignCall fc)
434 PrimOpId op -> not (primOpOutOfLine op)
440 -- size_up_app is used when there's ONE OR MORE value args
441 size_up_app (App fun arg) args
442 | isTyCoArg arg = size_up_app fun args
443 | otherwise = size_up arg `addSizeNSD`
444 size_up_app fun (arg:args)
445 size_up_app (Var fun) args = size_up_call fun args
446 size_up_app other args = size_up other `addSizeN` length args
449 size_up_call :: Id -> [CoreExpr] -> ExprSize
450 size_up_call fun val_args
451 = case idDetails fun of
452 FCallId _ -> sizeN opt_UF_DearOp
453 DataConWorkId dc -> conSize dc (length val_args)
454 PrimOpId op -> primOpSize op (length val_args)
455 ClassOpId _ -> classOpSize top_args val_args
456 _ -> funSize top_args fun (length val_args)
459 size_up_alt (_con, _bndrs, rhs) = size_up rhs `addSizeN` 1
460 -- Don't charge for args, so that wrappers look cheap
461 -- (See comments about wrappers with Case)
463 -- IMPORATANT: *do* charge 1 for the alternative, else we
464 -- find that giant case nests are treated as practically free
465 -- A good example is Foreign.C.Error.errrnoToIOError
468 -- These addSize things have to be here because
469 -- I don't want to give them bOMB_OUT_SIZE as an argument
470 addSizeN TooBig _ = TooBig
471 addSizeN (SizeIs n xs d) m = mkSizeIs bOMB_OUT_SIZE (n +# iUnbox m) xs d
473 -- addAltSize is used to add the sizes of case alternatives
474 addAltSize TooBig _ = TooBig
475 addAltSize _ TooBig = TooBig
476 addAltSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
477 = mkSizeIs bOMB_OUT_SIZE (n1 +# n2)
479 (d1 +# d2) -- Note [addAltSize result discounts]
481 -- This variant ignores the result discount from its LEFT argument
482 -- It's used when the second argument isn't part of the result
483 addSizeNSD TooBig _ = TooBig
484 addSizeNSD _ TooBig = TooBig
485 addSizeNSD (SizeIs n1 xs _) (SizeIs n2 ys d2)
486 = mkSizeIs bOMB_OUT_SIZE (n1 +# n2)
492 -- | Finds a nominal size of a string literal.
493 litSize :: Literal -> Int
494 -- Used by CoreUnfold.sizeExpr
495 litSize (MachStr str) = 1 + ((lengthFS str + 3) `div` 4)
496 -- If size could be 0 then @f "x"@ might be too small
497 -- [Sept03: make literal strings a bit bigger to avoid fruitless
498 -- duplication of little strings]
499 litSize _other = 0 -- Must match size of nullary constructors
500 -- Key point: if x |-> 4, then x must inline unconditionally
501 -- (eg via case binding)
503 classOpSize :: [Id] -> [CoreExpr] -> ExprSize
504 -- See Note [Conlike is interesting]
507 classOpSize top_args (arg1 : other_args)
508 = SizeIs (iUnbox size) arg_discount (_ILIT(0))
510 size = 2 + length other_args
511 -- If the class op is scrutinising a lambda bound dictionary then
512 -- give it a discount, to encourage the inlining of this function
513 -- The actual discount is rather arbitrarily chosen
514 arg_discount = case arg1 of
515 Var dict | dict `elem` top_args
516 -> unitBag (dict, opt_UF_DictDiscount)
519 funSize :: [Id] -> Id -> Int -> ExprSize
520 -- Size for functions that are not constructors or primops
521 -- Note [Function applications]
522 funSize top_args fun n_val_args
523 | fun `hasKey` buildIdKey = buildSize
524 | fun `hasKey` augmentIdKey = augmentSize
525 | otherwise = SizeIs (iUnbox size) arg_discount (iUnbox res_discount)
527 some_val_args = n_val_args > 0
529 arg_discount | some_val_args && fun `elem` top_args
530 = unitBag (fun, opt_UF_FunAppDiscount)
531 | otherwise = emptyBag
532 -- If the function is an argument and is applied
533 -- to some values, give it an arg-discount
535 res_discount | idArity fun > n_val_args = opt_UF_FunAppDiscount
537 -- If the function is partially applied, show a result discount
539 size | some_val_args = 1 + n_val_args
541 -- The 1+ is for the function itself
542 -- Add 1 for each non-trivial arg;
543 -- the allocation cost, as in let(rec)
546 conSize :: DataCon -> Int -> ExprSize
547 conSize dc n_val_args
548 | n_val_args == 0 = SizeIs (_ILIT(0)) emptyBag (_ILIT(1)) -- Like variables
550 -- See Note [Constructor size]
551 | isUnboxedTupleCon dc = SizeIs (_ILIT(0)) emptyBag (iUnbox n_val_args +# _ILIT(1))
553 -- See Note [Unboxed tuple result discount]
554 -- | isUnboxedTupleCon dc = SizeIs (_ILIT(0)) emptyBag (_ILIT(0))
556 -- See Note [Constructor size]
557 | otherwise = SizeIs (_ILIT(1)) emptyBag (iUnbox n_val_args +# _ILIT(1))
560 Note [Constructor size]
561 ~~~~~~~~~~~~~~~~~~~~~~~
562 Treat a constructors application as size 1, regardless of how many
563 arguments it has; we are keen to expose them (and we charge separately
564 for their args). We can't treat them as size zero, else we find that
565 (Just x) has size 0, which is the same as a lone variable; and hence
566 'v' will always be replaced by (Just x), where v is bound to Just x.
568 However, unboxed tuples count as size zero. I found occasions where we had
569 f x y z = case op# x y z of { s -> (# s, () #) }
570 and f wasn't getting inlined.
572 Note [Unboxed tuple result discount]
573 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
574 I tried giving unboxed tuples a *result discount* of zero (see the
575 commented-out line). Why? When returned as a result they do not
576 allocate, so maybe we don't want to charge so much for them If you
577 have a non-zero discount here, we find that workers often get inlined
578 back into wrappers, because it look like
579 f x = case $wf x of (# a,b #) -> (a,b)
580 and we are keener because of the case. However while this change
581 shrank binary sizes by 0.5% it also made spectral/boyer allocate 5%
582 more. All other changes were very small. So it's not a big deal but I
583 didn't adopt the idea.
586 primOpSize :: PrimOp -> Int -> ExprSize
587 primOpSize op n_val_args
588 = if primOpOutOfLine op
589 then sizeN (op_size + n_val_args)
592 op_size = primOpCodeSize op
595 buildSize :: ExprSize
596 buildSize = SizeIs (_ILIT(0)) emptyBag (_ILIT(4))
597 -- We really want to inline applications of build
598 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
599 -- Indeed, we should add a result_discount becuause build is
600 -- very like a constructor. We don't bother to check that the
601 -- build is saturated (it usually is). The "-2" discounts for the \c n,
602 -- The "4" is rather arbitrary.
604 augmentSize :: ExprSize
605 augmentSize = SizeIs (_ILIT(0)) emptyBag (_ILIT(4))
606 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
607 -- e plus ys. The -2 accounts for the \cn
609 -- When we return a lambda, give a discount if it's used (applied)
610 lamScrutDiscount :: ExprSize -> ExprSize
611 lamScrutDiscount (SizeIs n vs _) = SizeIs n vs (iUnbox opt_UF_FunAppDiscount)
612 lamScrutDiscount TooBig = TooBig
615 Note [addAltSize result discounts]
616 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
617 When adding the size of alternatives, we *add* the result discounts
618 too, rather than take the *maximum*. For a multi-branch case, this
619 gives a discount for each branch that returns a constructor, making us
620 keener to inline. I did try using 'max' instead, but it makes nofib
621 'rewrite' and 'puzzle' allocate significantly more, and didn't make
622 binary sizes shrink significantly either.
624 Note [Discounts and thresholds]
625 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
626 Constants for discounts and thesholds are defined in main/StaticFlags,
627 all of form opt_UF_xxxx. They are:
629 opt_UF_CreationThreshold (45)
630 At a definition site, if the unfolding is bigger than this, we
631 may discard it altogether
633 opt_UF_UseThreshold (6)
634 At a call site, if the unfolding, less discounts, is smaller than
635 this, then it's small enough inline
637 opt_UF_KeennessFactor (1.5)
638 Factor by which the discounts are multiplied before
639 subtracting from size
641 opt_UF_DictDiscount (1)
642 The discount for each occurrence of a dictionary argument
643 as an argument of a class method. Should be pretty small
644 else big functions may get inlined
646 opt_UF_FunAppDiscount (6)
647 Discount for a function argument that is applied. Quite
648 large, because if we inline we avoid the higher-order call.
651 The size of a foreign call or not-dupable PrimOp
654 Note [Function applications]
655 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
656 In a function application (f a b)
658 - If 'f' is an argument to the function being analysed,
659 and there's at least one value arg, record a FunAppDiscount for f
661 - If the application if a PAP (arity > 2 in this example)
662 record a *result* discount (because inlining
663 with "extra" args in the call may mean that we now
664 get a saturated application)
666 Code for manipulating sizes
669 data ExprSize = TooBig
670 | SizeIs FastInt -- Size found
671 (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
672 FastInt -- Size to subtract if result is scrutinised
673 -- by a case expression
675 instance Outputable ExprSize where
676 ppr TooBig = ptext (sLit "TooBig")
677 ppr (SizeIs a _ c) = brackets (int (iBox a) <+> int (iBox c))
679 -- subtract the discount before deciding whether to bale out. eg. we
680 -- want to inline a large constructor application into a selector:
681 -- tup = (a_1, ..., a_99)
682 -- x = case tup of ...
684 mkSizeIs :: FastInt -> FastInt -> Bag (Id, Int) -> FastInt -> ExprSize
685 mkSizeIs max n xs d | (n -# d) ># max = TooBig
686 | otherwise = SizeIs n xs d
688 maxSize :: ExprSize -> ExprSize -> ExprSize
689 maxSize TooBig _ = TooBig
690 maxSize _ TooBig = TooBig
691 maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
695 sizeN :: Int -> ExprSize
697 sizeZero = SizeIs (_ILIT(0)) emptyBag (_ILIT(0))
698 sizeN n = SizeIs (iUnbox n) emptyBag (_ILIT(0))
702 %************************************************************************
704 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
706 %************************************************************************
708 We use 'couldBeSmallEnoughToInline' to avoid exporting inlinings that
709 we ``couldn't possibly use'' on the other side. Can be overridden w/
710 flaggery. Just the same as smallEnoughToInline, except that it has no
714 couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
715 couldBeSmallEnoughToInline threshold rhs
716 = case sizeExpr (iUnbox threshold) [] body of
720 (_, body) = collectBinders rhs
723 smallEnoughToInline :: Unfolding -> Bool
724 smallEnoughToInline (CoreUnfolding {uf_guidance = UnfIfGoodArgs {ug_size = size}})
725 = size <= opt_UF_UseThreshold
726 smallEnoughToInline _
730 certainlyWillInline :: Unfolding -> Bool
731 -- Sees if the unfolding is pretty certain to inline
732 certainlyWillInline (CoreUnfolding { uf_is_cheap = is_cheap, uf_arity = n_vals, uf_guidance = guidance })
736 UnfIfGoodArgs { ug_size = size}
737 -> is_cheap && size - (n_vals +1) <= opt_UF_UseThreshold
739 certainlyWillInline _
743 %************************************************************************
745 \subsection{callSiteInline}
747 %************************************************************************
749 This is the key function. It decides whether to inline a variable at a call site
751 callSiteInline is used at call sites, so it is a bit more generous.
752 It's a very important function that embodies lots of heuristics.
753 A non-WHNF can be inlined if it doesn't occur inside a lambda,
754 and occurs exactly once or
755 occurs once in each branch of a case and is small
757 If the thing is in WHNF, there's no danger of duplicating work,
758 so we can inline if it occurs once, or is small
760 NOTE: we don't want to inline top-level functions that always diverge.
761 It just makes the code bigger. Tt turns out that the convenient way to prevent
762 them inlining is to give them a NOINLINE pragma, which we do in
763 StrictAnal.addStrictnessInfoToTopId
766 callSiteInline :: DynFlags
768 -> Bool -- True <=> unfolding is active
769 -> Bool -- True if there are are no arguments at all (incl type args)
770 -> [ArgSummary] -- One for each value arg; True if it is interesting
771 -> CallCtxt -- True <=> continuation is interesting
772 -> Maybe CoreExpr -- Unfolding, if any
774 instance Outputable ArgSummary where
775 ppr TrivArg = ptext (sLit "TrivArg")
776 ppr NonTrivArg = ptext (sLit "NonTrivArg")
777 ppr ValueArg = ptext (sLit "ValueArg")
779 data CallCtxt = BoringCtxt
781 | ArgCtxt -- We are somewhere in the argument of a function
782 Bool -- True <=> we're somewhere in the RHS of function with rules
783 -- False <=> we *are* the argument of a function with non-zero
786 -- we *are* the RHS of a let Note [RHS of lets]
787 -- In both cases, be a little keener to inline
789 | ValAppCtxt -- We're applied to at least one value arg
790 -- This arises when we have ((f x |> co) y)
791 -- Then the (f x) has argument 'x' but in a ValAppCtxt
793 | CaseCtxt -- We're the scrutinee of a case
794 -- that decomposes its scrutinee
796 instance Outputable CallCtxt where
797 ppr BoringCtxt = ptext (sLit "BoringCtxt")
798 ppr (ArgCtxt rules) = ptext (sLit "ArgCtxt") <+> ppr rules
799 ppr CaseCtxt = ptext (sLit "CaseCtxt")
800 ppr ValAppCtxt = ptext (sLit "ValAppCtxt")
802 callSiteInline dflags id active_unfolding lone_variable arg_infos cont_info
803 = case idUnfolding id of
804 -- idUnfolding checks for loop-breakers, returning NoUnfolding
805 -- Things with an INLINE pragma may have an unfolding *and*
806 -- be a loop breaker (maybe the knot is not yet untied)
807 CoreUnfolding { uf_tmpl = unf_template, uf_is_top = is_top
808 , uf_is_cheap = is_cheap, uf_arity = uf_arity
809 , uf_guidance = guidance, uf_expandable = is_exp }
810 | active_unfolding -> tryUnfolding dflags id lone_variable
811 arg_infos cont_info unf_template is_top
812 is_cheap is_exp uf_arity guidance
813 | otherwise -> Nothing
814 NoUnfolding -> Nothing
815 OtherCon {} -> Nothing
816 DFunUnfolding {} -> Nothing -- Never unfold a DFun
818 tryUnfolding :: DynFlags -> Id -> Bool -> [ArgSummary] -> CallCtxt
819 -> CoreExpr -> Bool -> Bool -> Bool -> Arity -> UnfoldingGuidance
821 tryUnfolding dflags id lone_variable
822 arg_infos cont_info unf_template is_top
823 is_cheap is_exp uf_arity guidance
824 -- uf_arity will typically be equal to (idArity id),
825 -- but may be less for InlineRules
826 | dopt Opt_D_dump_inlinings dflags && dopt Opt_D_verbose_core2core dflags
827 = pprTrace ("Considering inlining: " ++ showSDoc (ppr id))
828 (vcat [text "arg infos" <+> ppr arg_infos,
829 text "uf arity" <+> ppr uf_arity,
830 text "interesting continuation" <+> ppr cont_info,
831 text "some_benefit" <+> ppr some_benefit,
832 text "is exp:" <+> ppr is_exp,
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"])
841 n_val_args = length arg_infos
842 saturated = n_val_args >= uf_arity
844 result | yes_or_no = Just unf_template
845 | otherwise = Nothing
847 interesting_args = any nonTriv arg_infos
848 -- NB: (any nonTriv arg_infos) looks at the
849 -- over-saturated args too which is "wrong";
850 -- but if over-saturated we inline anyway.
852 -- some_benefit is used when the RHS is small enough
853 -- and the call has enough (or too many) value
854 -- arguments (ie n_val_args >= arity). But there must
855 -- be *something* interesting about some argument, or the
856 -- result context, to make it worth inlining
858 | not saturated = interesting_args -- Under-saturated
859 -- Note [Unsaturated applications]
860 | n_val_args > uf_arity = True -- Over-saturated
861 | otherwise = interesting_args -- Saturated
862 || interesting_saturated_call
864 interesting_saturated_call
866 BoringCtxt -> not is_top && uf_arity > 0 -- Note [Nested functions]
867 CaseCtxt -> not (lone_variable && is_cheap) -- Note [Lone variables]
868 ArgCtxt {} -> uf_arity > 0 -- Note [Inlining in ArgCtxt]
869 ValAppCtxt -> True -- Note [Cast then apply]
871 (yes_or_no, extra_doc)
873 UnfNever -> (False, empty)
875 UnfWhen unsat_ok boring_ok
876 -> (enough_args && (boring_ok || some_benefit), empty )
877 where -- See Note [INLINE for small functions]
878 enough_args = saturated || (unsat_ok && n_val_args > 0)
880 UnfIfGoodArgs { ug_args = arg_discounts, ug_res = res_discount, ug_size = size }
881 -> ( is_cheap && some_benefit && small_enough
882 , (text "discounted size =" <+> int discounted_size) )
884 discounted_size = size - discount
885 small_enough = discounted_size <= opt_UF_UseThreshold
886 discount = computeDiscount uf_arity arg_discounts
887 res_discount arg_infos cont_info
892 Be a tiny bit keener to inline in the RHS of a let, because that might
893 lead to good thing later
895 g y = let x = f y in ...(case x of (a,b,c) -> ...) ...
896 We'd inline 'f' if the call was in a case context, and it kind-of-is,
897 only we can't see it. So we treat the RHS of a let as not-totally-boring.
899 Note [Unsaturated applications]
900 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
901 When a call is not saturated, we *still* inline if one of the
902 arguments has interesting structure. That's sometimes very important.
903 A good example is the Ord instance for Bool in Base:
906 $fOrdBool =GHC.Classes.D:Ord
911 $cmin_ajX [Occ=LoopBreaker] :: Bool -> Bool -> Bool
912 $cmin_ajX = GHC.Classes.$dmmin @ Bool $fOrdBool
915 But the defn of GHC.Classes.$dmmin is:
917 $dmmin :: forall a. GHC.Classes.Ord a => a -> a -> a
918 {- Arity: 3, HasNoCafRefs, Strictness: SLL,
919 Unfolding: (\ @ a $dOrd :: GHC.Classes.Ord a x :: a y :: a ->
920 case @ a GHC.Classes.<= @ a $dOrd x y of wild {
921 GHC.Types.False -> y GHC.Types.True -> x }) -}
923 We *really* want to inline $dmmin, even though it has arity 3, in
924 order to unravel the recursion.
927 Note [Things to watch]
928 ~~~~~~~~~~~~~~~~~~~~~~
929 * { y = I# 3; x = y `cast` co; ...case (x `cast` co) of ... }
930 Assume x is exported, so not inlined unconditionally.
931 Then we want x to inline unconditionally; no reason for it
932 not to, and doing so avoids an indirection.
934 * { x = I# 3; ....f x.... }
935 Make sure that x does not inline unconditionally!
936 Lest we get extra allocation.
938 Note [Inlining an InlineRule]
939 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
940 An InlineRules is used for
941 (a) programmer INLINE pragmas
942 (b) inlinings from worker/wrapper
944 For (a) the RHS may be large, and our contract is that we *only* inline
945 when the function is applied to all the arguments on the LHS of the
946 source-code defn. (The uf_arity in the rule.)
948 However for worker/wrapper it may be worth inlining even if the
949 arity is not satisfied (as we do in the CoreUnfolding case) so we don't
953 Note [Nested functions]
954 ~~~~~~~~~~~~~~~~~~~~~~~
955 If a function has a nested defn we also record some-benefit, on the
956 grounds that we are often able to eliminate the binding, and hence the
957 allocation, for the function altogether; this is good for join points.
958 But this only makes sense for *functions*; inlining a constructor
959 doesn't help allocation unless the result is scrutinised. UNLESS the
960 constructor occurs just once, albeit possibly in multiple case
961 branches. Then inlining it doesn't increase allocation, but it does
962 increase the chance that the constructor won't be allocated at all in
963 the branches that don't use it.
965 Note [Cast then apply]
966 ~~~~~~~~~~~~~~~~~~~~~~
968 myIndex = __inline_me ( (/\a. <blah>) |> co )
969 co :: (forall a. a -> a) ~ (forall a. T a)
970 ... /\a.\x. case ((myIndex a) |> sym co) x of { ... } ...
972 We need to inline myIndex to unravel this; but the actual call (myIndex a) has
973 no value arguments. The ValAppCtxt gives it enough incentive to inline.
975 Note [Inlining in ArgCtxt]
976 ~~~~~~~~~~~~~~~~~~~~~~~~~~
977 The condition (arity > 0) here is very important, because otherwise
978 we end up inlining top-level stuff into useless places; eg
981 This can make a very big difference: it adds 16% to nofib 'integer' allocs,
984 At one stage I replaced this condition by 'True' (leading to the above
985 slow-down). The motivation was test eyeball/inline1.hs; but that seems
988 NOTE: arguably, we should inline in ArgCtxt only if the result of the
989 call is at least CONLIKE. At least for the cases where we use ArgCtxt
990 for the RHS of a 'let', we only profit from the inlining if we get a
991 CONLIKE thing (modulo lets).
993 Note [Lone variables] See also Note [Interaction of exprIsCheap and lone variables]
994 ~~~~~~~~~~~~~~~~~~~~~ which appears below
995 The "lone-variable" case is important. I spent ages messing about
996 with unsatisfactory varaints, but this is nice. The idea is that if a
997 variable appears all alone
999 as an arg of lazy fn, or rhs BoringCtxt
1000 as scrutinee of a case CaseCtxt
1001 as arg of a fn ArgCtxt
1003 it is bound to a cheap expression
1005 then we should not inline it (unless there is some other reason,
1006 e.g. is is the sole occurrence). That is what is happening at
1007 the use of 'lone_variable' in 'interesting_saturated_call'.
1009 Why? At least in the case-scrutinee situation, turning
1010 let x = (a,b) in case x of y -> ...
1012 let x = (a,b) in case (a,b) of y -> ...
1014 let x = (a,b) in let y = (a,b) in ...
1015 is bad if the binding for x will remain.
1017 Another example: I discovered that strings
1018 were getting inlined straight back into applications of 'error'
1019 because the latter is strict.
1021 f = \x -> ...(error s)...
1023 Fundamentally such contexts should not encourage inlining because the
1024 context can ``see'' the unfolding of the variable (e.g. case or a
1025 RULE) so there's no gain. If the thing is bound to a value.
1030 foo = _inline_ (\n. [n])
1031 bar = _inline_ (foo 20)
1032 baz = \n. case bar of { (m:_) -> m + n }
1033 Here we really want to inline 'bar' so that we can inline 'foo'
1034 and the whole thing unravels as it should obviously do. This is
1035 important: in the NDP project, 'bar' generates a closure data
1036 structure rather than a list.
1038 So the non-inlining of lone_variables should only apply if the
1039 unfolding is regarded as cheap; because that is when exprIsConApp_maybe
1040 looks through the unfolding. Hence the "&& is_cheap" in the
1043 * Even a type application or coercion isn't a lone variable.
1045 case $fMonadST @ RealWorld of { :DMonad a b c -> c }
1046 We had better inline that sucker! The case won't see through it.
1048 For now, I'm treating treating a variable applied to types
1049 in a *lazy* context "lone". The motivating example was
1051 g = /\a. \y. h (f a)
1052 There's no advantage in inlining f here, and perhaps
1053 a significant disadvantage. Hence some_val_args in the Stop case
1055 Note [Interaction of exprIsCheap and lone variables]
1056 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1057 The lone-variable test says "don't inline if a case expression
1058 scrutines a lone variable whose unfolding is cheap". It's very
1059 important that, under these circumstances, exprIsConApp_maybe
1060 can spot a constructor application. So, for example, we don't
1063 to be cheap, and that's good because exprIsConApp_maybe doesn't
1064 think that expression is a constructor application.
1066 I used to test is_value rather than is_cheap, which was utterly
1067 wrong, because the above expression responds True to exprIsHNF.
1069 This kind of thing can occur if you have
1072 foo = let x = e in (x,x)
1077 computeDiscount :: Int -> [Int] -> Int -> [ArgSummary] -> CallCtxt -> Int
1078 computeDiscount n_vals_wanted arg_discounts res_discount arg_infos cont_info
1079 -- We multiple the raw discounts (args_discount and result_discount)
1080 -- ty opt_UnfoldingKeenessFactor because the former have to do with
1081 -- *size* whereas the discounts imply that there's some extra
1082 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
1085 = 1 -- Discount of 1 because the result replaces the call
1086 -- so we count 1 for the function itself
1088 + length (take n_vals_wanted arg_infos)
1089 -- Discount of (un-scaled) 1 for each arg supplied,
1090 -- because the result replaces the call
1092 + round (opt_UF_KeenessFactor *
1093 fromIntegral (arg_discount + res_discount'))
1095 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
1097 mk_arg_discount _ TrivArg = 0
1098 mk_arg_discount _ NonTrivArg = 1
1099 mk_arg_discount discount ValueArg = discount
1101 res_discount' = case cont_info of
1103 CaseCtxt -> res_discount
1104 _other -> 4 `min` res_discount
1105 -- res_discount can be very large when a function returns
1106 -- constructors; but we only want to invoke that large discount
1107 -- when there's a case continuation.
1108 -- Otherwise we, rather arbitrarily, threshold it. Yuk.
1109 -- But we want to aovid inlining large functions that return
1110 -- constructors into contexts that are simply "interesting"
1113 %************************************************************************
1115 Interesting arguments
1117 %************************************************************************
1119 Note [Interesting arguments]
1120 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1121 An argument is interesting if it deserves a discount for unfoldings
1122 with a discount in that argument position. The idea is to avoid
1123 unfolding a function that is applied only to variables that have no
1124 unfolding (i.e. they are probably lambda bound): f x y z There is
1125 little point in inlining f here.
1127 Generally, *values* (like (C a b) and (\x.e)) deserve discounts. But
1128 we must look through lets, eg (let x = e in C a b), because the let will
1129 float, exposing the value, if we inline. That makes it different to
1132 Before 2009 we said it was interesting if the argument had *any* structure
1133 at all; i.e. (hasSomeUnfolding v). But does too much inlining; see Trac #3016.
1135 But we don't regard (f x y) as interesting, unless f is unsaturated.
1136 If it's saturated and f hasn't inlined, then it's probably not going
1139 Note [Conlike is interesting]
1140 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1142 f d = ...((*) d x y)...
1144 where df is con-like. Then we'd really like to inline 'f' so that the
1145 rule for (*) (df d) can fire. To do this
1146 a) we give a discount for being an argument of a class-op (eg (*) d)
1147 b) we say that a con-like argument (eg (df d)) is interesting
1150 data ArgSummary = TrivArg -- Nothing interesting
1151 | NonTrivArg -- Arg has structure
1152 | ValueArg -- Arg is a con-app or PAP
1153 -- ..or con-like. Note [Conlike is interesting]
1155 interestingArg :: CoreExpr -> ArgSummary
1156 -- See Note [Interesting arguments]
1157 interestingArg e = go e 0
1159 -- n is # value args to which the expression is applied
1160 go (Lit {}) _ = ValueArg
1162 | isConLikeId v = ValueArg -- Experimenting with 'conlike' rather that
1163 -- data constructors here
1164 | idArity v > n = ValueArg -- Catches (eg) primops with arity but no unfolding
1165 | n > 0 = NonTrivArg -- Saturated or unknown call
1166 | conlike_unfolding = ValueArg -- n==0; look for an interesting unfolding
1167 -- See Note [Conlike is interesting]
1168 | otherwise = TrivArg -- n==0, no useful unfolding
1170 conlike_unfolding = isConLikeUnfolding (idUnfolding v)
1172 go (Type _) _ = TrivArg
1173 go (Coercion _) _ = TrivArg
1174 go (App fn (Type _)) n = go fn n
1175 go (App fn (Coercion _)) n = go fn n
1176 go (App fn _) n = go fn (n+1)
1177 go (Note _ a) n = go a n
1178 go (Cast e _) n = go e n
1180 | isTyVar v = go e n
1182 | otherwise = ValueArg
1183 go (Let _ e) n = case go e n of { ValueArg -> ValueArg; _ -> NonTrivArg }
1184 go (Case {}) _ = NonTrivArg
1186 nonTriv :: ArgSummary -> Bool
1187 nonTriv TrivArg = False
1191 %************************************************************************
1195 %************************************************************************
1197 Note [exprIsConApp_maybe]
1198 ~~~~~~~~~~~~~~~~~~~~~~~~~
1199 exprIsConApp_maybe is a very important function. There are two principal
1201 * case e of { .... }
1202 * cls_op e, where cls_op is a class operation
1204 In both cases you want to know if e is of form (C e1..en) where C is
1207 However e might not *look* as if
1210 -- | Returns @Just (dc, [t1..tk], [x1..xn])@ if the argument expression is
1211 -- a *saturated* constructor application of the form @dc t1..tk x1 .. xn@,
1212 -- where t1..tk are the *universally-qantified* type args of 'dc'
1213 exprIsConApp_maybe :: IdUnfoldingFun -> CoreExpr -> Maybe (DataCon, [Type], [CoreExpr])
1215 exprIsConApp_maybe id_unf (Note note expr)
1217 = exprIsConApp_maybe id_unf expr
1218 -- We ignore all notes except SCCs. For example,
1219 -- case _scc_ "foo" (C a b) of
1221 -- should not be optimised away, because we'll lose the
1222 -- entry count on 'foo'; see Trac #4414
1224 exprIsConApp_maybe id_unf (Cast expr co)
1225 = -- Here we do the KPush reduction rule as described in the FC paper
1226 -- The transformation applies iff we have
1227 -- (C e1 ... en) `cast` co
1228 -- where co :: (T t1 .. tn) ~ to_ty
1229 -- The left-hand one must be a T, because exprIsConApp returned True
1230 -- but the right-hand one might not be. (Though it usually will.)
1232 case exprIsConApp_maybe id_unf expr of {
1233 Nothing -> Nothing ;
1234 Just (dc, _dc_univ_args, dc_args) ->
1236 let Pair _from_ty to_ty = coercionKind co
1237 dc_tc = dataConTyCon dc
1239 case splitTyConApp_maybe to_ty of {
1240 Nothing -> Nothing ;
1241 Just (to_tc, to_tc_arg_tys)
1242 | dc_tc /= to_tc -> Nothing
1243 -- These two Nothing cases are possible; we might see
1244 -- (C x y) `cast` (g :: T a ~ S [a]),
1245 -- where S is a type function. In fact, exprIsConApp
1246 -- will probably not be called in such circumstances,
1247 -- but there't nothing wrong with it
1251 tc_arity = tyConArity dc_tc
1252 dc_univ_tyvars = dataConUnivTyVars dc
1253 dc_ex_tyvars = dataConExTyVars dc
1254 arg_tys = dataConRepArgTys dc
1256 (ex_args, val_args) = splitAtList dc_ex_tyvars dc_args
1258 -- Make the "theta" from Fig 3 of the paper
1259 gammas = decomposeCo tc_arity co
1260 theta = zipOpenCvSubst (dc_univ_tyvars ++ dc_ex_tyvars)
1261 (gammas ++ map mkReflCo (stripTypeArgs ex_args))
1263 -- Cast the value arguments (which include dictionaries)
1264 new_val_args = zipWith cast_arg arg_tys val_args
1265 cast_arg arg_ty arg = mkCoerce (liftCoSubst theta arg_ty) arg
1268 let dump_doc = vcat [ppr dc, ppr dc_univ_tyvars, ppr dc_ex_tyvars,
1269 ppr arg_tys, ppr dc_args, ppr _dc_univ_args,
1270 ppr ex_args, ppr val_args]
1272 ASSERT2( eqType _from_ty (mkTyConApp dc_tc _dc_univ_args), dump_doc )
1273 ASSERT2( all isTypeArg ex_args, dump_doc )
1274 ASSERT2( equalLength val_args arg_tys, dump_doc )
1277 Just (dc, to_tc_arg_tys, ex_args ++ new_val_args)
1280 exprIsConApp_maybe id_unf expr
1283 analyse (App fun arg) args = analyse fun (arg:args)
1284 analyse fun@(Lam {}) args = beta fun [] args
1286 analyse (Var fun) args
1287 | Just con <- isDataConWorkId_maybe fun
1288 , count isValArg args == idArity fun
1289 , let (univ_ty_args, rest_args) = splitAtList (dataConUnivTyVars con) args
1290 = Just (con, stripTypeArgs univ_ty_args, rest_args)
1292 -- Look through dictionary functions; see Note [Unfolding DFuns]
1293 | DFunUnfolding dfun_nargs con ops <- unfolding
1294 , let sat = length args == dfun_nargs -- See Note [DFun arity check]
1295 in if sat then True else
1296 pprTrace "Unsaturated dfun" (ppr fun <+> int dfun_nargs $$ ppr args) False
1297 , let (dfun_tvs, _n_theta, _cls, dfun_res_tys) = tcSplitDFunTy (idType fun)
1298 subst = zipOpenTvSubst dfun_tvs (stripTypeArgs (takeList dfun_tvs args))
1299 mk_arg (DFunConstArg e) = e
1300 mk_arg (DFunLamArg i) = args !! i
1301 mk_arg (DFunPolyArg e) = mkApps e args
1302 = Just (con, substTys subst dfun_res_tys, map mk_arg ops)
1304 -- Look through unfoldings, but only cheap ones, because
1305 -- we are effectively duplicating the unfolding
1306 | Just rhs <- expandUnfolding_maybe unfolding
1307 = -- pprTrace "expanding" (ppr fun $$ ppr rhs) $
1310 unfolding = id_unf fun
1312 analyse _ _ = Nothing
1315 beta (Lam v body) pairs (arg : args)
1317 = beta body ((v,arg):pairs) args
1319 beta (Lam {}) _ _ -- Un-saturated, or not a type lambda
1323 = analyse (substExpr (text "subst-expr-is-con-app") subst fun) args
1325 subst = mkOpenSubst (mkInScopeSet (exprFreeVars fun)) pairs
1326 -- doc = vcat [ppr fun, ppr expr, ppr pairs, ppr args]
1328 stripTypeArgs :: [CoreExpr] -> [Type]
1329 stripTypeArgs args = ASSERT2( all isTypeArg args, ppr args )
1330 [ty | Type ty <- args]
1331 -- We really do want isTypeArg here, not isTyCoArg!
1334 Note [Unfolding DFuns]
1335 ~~~~~~~~~~~~~~~~~~~~~~
1338 df :: forall a b. (Eq a, Eq b) -> Eq (a,b)
1339 df a b d_a d_b = MkEqD (a,b) ($c1 a b d_a d_b)
1342 So to split it up we just need to apply the ops $c1, $c2 etc
1343 to the very same args as the dfun. It takes a little more work
1344 to compute the type arguments to the dictionary constructor.
1346 Note [DFun arity check]
1347 ~~~~~~~~~~~~~~~~~~~~~~~
1348 Here we check that the total number of supplied arguments (inclding
1349 type args) matches what the dfun is expecting. This may be *less*
1350 than the ordinary arity of the dfun: see Note [DFun unfoldings] in CoreSyn