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 )
70 %************************************************************************
72 \subsection{Making unfoldings}
74 %************************************************************************
77 mkTopUnfolding :: Bool -> CoreExpr -> Unfolding
78 mkTopUnfolding = mkUnfolding InlineRhs True {- Top level -}
80 mkImplicitUnfolding :: CoreExpr -> Unfolding
81 -- For implicit Ids, do a tiny bit of optimising first
82 mkImplicitUnfolding expr = mkTopUnfolding False (simpleOptExpr expr)
84 -- Note [Top-level flag on inline rules]
85 -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
86 -- Slight hack: note that mk_inline_rules conservatively sets the
87 -- top-level flag to True. It gets set more accurately by the simplifier
88 -- Simplify.simplUnfolding.
90 mkSimpleUnfolding :: CoreExpr -> Unfolding
91 mkSimpleUnfolding = mkUnfolding InlineRhs False False
93 mkDFunUnfolding :: Type -> [DFunArg CoreExpr] -> Unfolding
94 mkDFunUnfolding dfun_ty ops
95 = DFunUnfolding dfun_nargs data_con ops
97 (tvs, n_theta, cls, _) = tcSplitDFunTy dfun_ty
98 dfun_nargs = length tvs + n_theta
99 data_con = classDataCon cls
101 mkWwInlineRule :: Id -> CoreExpr -> Arity -> Unfolding
102 mkWwInlineRule id expr arity
103 = mkCoreUnfolding (InlineWrapper id) True
104 (simpleOptExpr expr) arity
105 (UnfWhen unSaturatedOk boringCxtNotOk)
107 mkCompulsoryUnfolding :: CoreExpr -> Unfolding
108 mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
109 = mkCoreUnfolding InlineCompulsory True
110 expr 0 -- Arity of unfolding doesn't matter
111 (UnfWhen unSaturatedOk boringCxtOk)
113 mkInlineUnfolding :: Maybe Arity -> CoreExpr -> Unfolding
114 mkInlineUnfolding mb_arity expr
115 = mkCoreUnfolding InlineStable
116 True -- Note [Top-level flag on inline rules]
118 (UnfWhen unsat_ok boring_ok)
120 expr' = simpleOptExpr expr
121 (unsat_ok, arity) = case mb_arity of
122 Nothing -> (unSaturatedOk, manifestArity expr')
123 Just ar -> (needSaturated, ar)
125 boring_ok = inlineBoringOk expr'
127 mkInlinableUnfolding :: CoreExpr -> Unfolding
128 mkInlinableUnfolding expr
129 = mkUnfolding InlineStable True is_bot expr'
131 expr' = simpleOptExpr expr
132 is_bot = isJust (exprBotStrictness_maybe expr')
138 mkCoreUnfolding :: UnfoldingSource -> Bool -> CoreExpr
139 -> Arity -> UnfoldingGuidance -> Unfolding
140 -- Occurrence-analyses the expression before capturing it
141 mkCoreUnfolding src top_lvl expr arity guidance
142 = CoreUnfolding { uf_tmpl = occurAnalyseExpr expr,
146 uf_is_value = exprIsHNF expr,
147 uf_is_conlike = exprIsConLike expr,
148 uf_is_cheap = exprIsCheap expr,
149 uf_expandable = exprIsExpandable expr,
150 uf_guidance = guidance }
152 mkUnfolding :: UnfoldingSource -> Bool -> Bool -> CoreExpr -> Unfolding
153 -- Calculates unfolding guidance
154 -- Occurrence-analyses the expression before capturing it
155 mkUnfolding src top_lvl is_bottoming expr
156 | top_lvl && is_bottoming
157 , not (exprIsTrivial expr)
158 = NoUnfolding -- See Note [Do not inline top-level bottoming functions]
160 = CoreUnfolding { uf_tmpl = occurAnalyseExpr expr,
164 uf_is_value = exprIsHNF expr,
165 uf_is_conlike = exprIsConLike expr,
166 uf_expandable = exprIsExpandable expr,
167 uf_is_cheap = is_cheap,
168 uf_guidance = guidance }
170 is_cheap = exprIsCheap expr
171 (arity, guidance) = calcUnfoldingGuidance is_cheap
172 opt_UF_CreationThreshold expr
173 -- Sometimes during simplification, there's a large let-bound thing
174 -- which has been substituted, and so is now dead; so 'expr' contains
175 -- two copies of the thing while the occurrence-analysed expression doesn't
176 -- Nevertheless, we *don't* occ-analyse before computing the size because the
177 -- size computation bales out after a while, whereas occurrence analysis does not.
179 -- This can occasionally mean that the guidance is very pessimistic;
180 -- it gets fixed up next round. And it should be rare, because large
181 -- let-bound things that are dead are usually caught by preInlineUnconditionally
184 %************************************************************************
186 \subsection{The UnfoldingGuidance type}
188 %************************************************************************
191 inlineBoringOk :: CoreExpr -> Bool
192 -- See Note [INLINE for small functions]
193 -- True => the result of inlining the expression is
194 -- no bigger than the expression itself
195 -- eg (\x y -> f y x)
196 -- This is a quick and dirty version. It doesn't attempt
197 -- to deal with (\x y z -> x (y z))
198 -- The really important one is (x `cast` c)
202 go :: Int -> CoreExpr -> Bool
203 go credit (Lam x e) | isId x = go (credit+1) e
204 | otherwise = go credit e
205 go credit (App f (Type {})) = go credit f
206 go credit (App f a) | credit > 0
207 , exprIsTrivial a = go (credit-1) f
208 go credit (Note _ e) = go credit e
209 go credit (Cast e _) = go credit e
210 go _ (Var {}) = boringCxtOk
211 go _ _ = boringCxtNotOk
213 calcUnfoldingGuidance
214 :: Bool -- True <=> the rhs is cheap, or we want to treat it
215 -- as cheap (INLINE things)
216 -> Int -- Bomb out if size gets bigger than this
217 -> CoreExpr -- Expression to look at
218 -> (Arity, UnfoldingGuidance)
219 calcUnfoldingGuidance expr_is_cheap bOMB_OUT_SIZE expr
220 = case collectBinders expr of { (bndrs, body) ->
222 val_bndrs = filter isId bndrs
223 n_val_bndrs = length val_bndrs
226 = case (sizeExpr (iUnbox bOMB_OUT_SIZE) val_bndrs body) of
228 SizeIs size cased_bndrs scrut_discount
229 | uncondInline n_val_bndrs (iBox size)
231 -> UnfWhen unSaturatedOk boringCxtOk -- Note [INLINE for small functions]
233 -> UnfIfGoodArgs { ug_args = map (discount cased_bndrs) val_bndrs
234 , ug_size = iBox size
235 , ug_res = iBox scrut_discount }
238 = foldlBag (\acc (b',n) -> if bndr==b' then acc+n else acc)
241 (n_val_bndrs, guidance) }
244 Note [Computing the size of an expression]
245 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
246 The basic idea of sizeExpr is obvious enough: count nodes. But getting the
247 heuristics right has taken a long time. Here's the basic strategy:
249 * Variables, literals: 0
250 (Exception for string literals, see litSize.)
252 * Function applications (f e1 .. en): 1 + #value args
254 * Constructor applications: 1, regardless of #args
256 * Let(rec): 1 + size of components
271 Notice that 'x' counts 0, while (f x) counts 2. That's deliberate: there's
272 a function call to account for. Notice also that constructor applications
273 are very cheap, because exposing them to a caller is so valuable.
276 Note [Do not inline top-level bottoming functions]
277 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
278 The FloatOut pass has gone to some trouble to float out calls to 'error'
279 and similar friends. See Note [Bottoming floats] in SetLevels.
280 Do not re-inline them! But we *do* still inline if they are very small
281 (the uncondInline stuff).
284 Note [INLINE for small functions]
285 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
286 Consider {-# INLINE f #-}
289 Then f's RHS is no larger than its LHS, so we should inline it into
290 even the most boring context. In general, f the function is
291 sufficiently small that its body is as small as the call itself, the
292 inline unconditionally, regardless of how boring the context is.
296 * We inline *unconditionally* if inlined thing is smaller (using sizeExpr)
297 than the thing it's replacing. Notice that
298 (f x) --> (g 3) -- YES, unconditionally
299 (f x) --> x : [] -- YES, *even though* there are two
300 -- arguments to the cons
304 It's very important not to unconditionally replace a variable by
307 * We do this even if the thing isn't saturated, else we end up with the
311 doesn't inline. Even in a boring context, inlining without being
312 saturated will give a lambda instead of a PAP, and will be more
313 efficient at runtime.
315 * However, when the function's arity > 0, we do insist that it
316 has at least one value argument at the call site. Otherwise we find this:
319 If we inline f here we get
320 d = /\b. MkD (\x:b. x)
321 and then prepareRhs floats out the argument, abstracting the type
322 variables, so we end up with the original again!
326 uncondInline :: Arity -> Int -> Bool
327 -- Inline unconditionally if there no size increase
328 -- Size of call is arity (+1 for the function)
329 -- See Note [INLINE for small functions]
330 uncondInline arity size
331 | arity == 0 = size == 0
332 | otherwise = size <= arity + 1
337 sizeExpr :: FastInt -- Bomb out if it gets bigger than this
338 -> [Id] -- Arguments; we're interested in which of these
343 -- Note [Computing the size of an expression]
345 sizeExpr bOMB_OUT_SIZE top_args expr
348 size_up (Cast e _) = size_up e
349 size_up (Note _ e) = size_up e
350 size_up (Type _) = sizeZero -- Types cost nothing
351 size_up (Lit lit) = sizeN (litSize lit)
352 size_up (Var f) = size_up_call f [] -- Make sure we get constructor
353 -- discounts even on nullary constructors
355 size_up (App fun (Type _)) = size_up fun
356 size_up (App fun arg) = size_up arg `addSizeNSD`
357 size_up_app fun [arg]
359 size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
360 | otherwise = size_up e
362 size_up (Let (NonRec binder rhs) body)
363 = size_up rhs `addSizeNSD`
364 size_up body `addSizeN`
365 (if isUnLiftedType (idType binder) then 0 else 1)
366 -- For the allocation
367 -- If the binder has an unlifted type there is no allocation
369 size_up (Let (Rec pairs) body)
370 = foldr (addSizeNSD . size_up . snd)
371 (size_up body `addSizeN` length pairs) -- (length pairs) for the allocation
374 size_up (Case (Var v) _ _ alts)
375 | v `elem` top_args -- We are scrutinising an argument variable
376 = alts_size (foldr1 addAltSize alt_sizes)
377 (foldr1 maxSize alt_sizes)
378 -- Good to inline if an arg is scrutinised, because
379 -- that may eliminate allocation in the caller
380 -- And it eliminates the case itself
382 alt_sizes = map size_up_alt alts
384 -- alts_size tries to compute a good discount for
385 -- the case when we are scrutinising an argument variable
386 alts_size (SizeIs tot tot_disc tot_scrut) -- Size of all alternatives
387 (SizeIs max _ _) -- Size of biggest alternative
388 = SizeIs tot (unitBag (v, iBox (_ILIT(2) +# tot -# max)) `unionBags` tot_disc) tot_scrut
389 -- If the variable is known, we produce a discount that
390 -- will take us back to 'max', the size of the largest alternative
391 -- The 1+ is a little discount for reduced allocation in the caller
393 -- Notice though, that we return tot_disc, the total discount from
394 -- all branches. I think that's right.
396 alts_size tot_size _ = tot_size
398 size_up (Case e _ _ alts) = size_up e `addSizeNSD`
399 foldr (addAltSize . size_up_alt) sizeZero alts
400 -- We don't charge for the case itself
401 -- It's a strict thing, and the price of the call
402 -- is paid by scrut. Also consider
403 -- case f x of DEFAULT -> e
404 -- This is just ';'! Don't charge for it.
406 -- Moreover, we charge one per alternative.
409 -- size_up_app is used when there's ONE OR MORE value args
410 size_up_app (App fun arg) args
411 | isTypeArg arg = size_up_app fun args
412 | otherwise = size_up arg `addSizeNSD`
413 size_up_app fun (arg:args)
414 size_up_app (Var fun) args = size_up_call fun args
415 size_up_app other args = size_up other `addSizeN` length args
418 size_up_call :: Id -> [CoreExpr] -> ExprSize
419 size_up_call fun val_args
420 = case idDetails fun of
421 FCallId _ -> sizeN opt_UF_DearOp
422 DataConWorkId dc -> conSize dc (length val_args)
423 PrimOpId op -> primOpSize op (length val_args)
424 ClassOpId _ -> classOpSize top_args val_args
425 _ -> funSize top_args fun (length val_args)
428 size_up_alt (_con, _bndrs, rhs) = size_up rhs `addSizeN` 1
429 -- Don't charge for args, so that wrappers look cheap
430 -- (See comments about wrappers with Case)
432 -- IMPORATANT: *do* charge 1 for the alternative, else we
433 -- find that giant case nests are treated as practically free
434 -- A good example is Foreign.C.Error.errrnoToIOError
437 -- These addSize things have to be here because
438 -- I don't want to give them bOMB_OUT_SIZE as an argument
439 addSizeN TooBig _ = TooBig
440 addSizeN (SizeIs n xs d) m = mkSizeIs bOMB_OUT_SIZE (n +# iUnbox m) xs d
442 -- addAltSize is used to add the sizes of case alternatives
443 addAltSize TooBig _ = TooBig
444 addAltSize _ TooBig = TooBig
445 addAltSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
446 = mkSizeIs bOMB_OUT_SIZE (n1 +# n2)
448 (d1 +# d2) -- Note [addAltSize result discounts]
450 -- This variant ignores the result discount from its LEFT argument
451 -- It's used when the second argument isn't part of the result
452 addSizeNSD TooBig _ = TooBig
453 addSizeNSD _ TooBig = TooBig
454 addSizeNSD (SizeIs n1 xs _) (SizeIs n2 ys d2)
455 = mkSizeIs bOMB_OUT_SIZE (n1 +# n2)
461 -- | Finds a nominal size of a string literal.
462 litSize :: Literal -> Int
463 -- Used by CoreUnfold.sizeExpr
464 litSize (MachStr str) = 1 + ((lengthFS str + 3) `div` 4)
465 -- If size could be 0 then @f "x"@ might be too small
466 -- [Sept03: make literal strings a bit bigger to avoid fruitless
467 -- duplication of little strings]
468 litSize _other = 0 -- Must match size of nullary constructors
469 -- Key point: if x |-> 4, then x must inline unconditionally
470 -- (eg via case binding)
472 classOpSize :: [Id] -> [CoreExpr] -> ExprSize
473 -- See Note [Conlike is interesting]
476 classOpSize top_args (arg1 : other_args)
477 = SizeIs (iUnbox size) arg_discount (_ILIT(0))
479 size = 2 + length other_args
480 -- If the class op is scrutinising a lambda bound dictionary then
481 -- give it a discount, to encourage the inlining of this function
482 -- The actual discount is rather arbitrarily chosen
483 arg_discount = case arg1 of
484 Var dict | dict `elem` top_args
485 -> unitBag (dict, opt_UF_DictDiscount)
488 funSize :: [Id] -> Id -> Int -> ExprSize
489 -- Size for functions that are not constructors or primops
490 -- Note [Function applications]
491 funSize top_args fun n_val_args
492 | fun `hasKey` buildIdKey = buildSize
493 | fun `hasKey` augmentIdKey = augmentSize
494 | otherwise = SizeIs (iUnbox size) arg_discount (iUnbox res_discount)
496 some_val_args = n_val_args > 0
498 arg_discount | some_val_args && fun `elem` top_args
499 = unitBag (fun, opt_UF_FunAppDiscount)
500 | otherwise = emptyBag
501 -- If the function is an argument and is applied
502 -- to some values, give it an arg-discount
504 res_discount | idArity fun > n_val_args = opt_UF_FunAppDiscount
506 -- If the function is partially applied, show a result discount
508 size | some_val_args = 1 + n_val_args
510 -- The 1+ is for the function itself
511 -- Add 1 for each non-trivial arg;
512 -- the allocation cost, as in let(rec)
515 conSize :: DataCon -> Int -> ExprSize
516 conSize dc n_val_args
517 | n_val_args == 0 = SizeIs (_ILIT(0)) emptyBag (_ILIT(1)) -- Like variables
519 -- See Note [Constructor size]
520 | isUnboxedTupleCon dc = SizeIs (_ILIT(0)) emptyBag (iUnbox n_val_args +# _ILIT(1))
522 -- See Note [Unboxed tuple result discount]
523 -- | isUnboxedTupleCon dc = SizeIs (_ILIT(0)) emptyBag (_ILIT(0))
525 -- See Note [Constructor size]
526 | otherwise = SizeIs (_ILIT(1)) emptyBag (iUnbox n_val_args +# _ILIT(1))
529 Note [Constructor size]
530 ~~~~~~~~~~~~~~~~~~~~~~~
531 Treat a constructors application as size 1, regardless of how many
532 arguments it has; we are keen to expose them (and we charge separately
533 for their args). We can't treat them as size zero, else we find that
534 (Just x) has size 0, which is the same as a lone variable; and hence
535 'v' will always be replaced by (Just x), where v is bound to Just x.
537 However, unboxed tuples count as size zero. I found occasions where we had
538 f x y z = case op# x y z of { s -> (# s, () #) }
539 and f wasn't getting inlined.
541 Note [Unboxed tuple result discount]
542 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
543 I tried giving unboxed tuples a *result discount* of zero (see the
544 commented-out line). Why? When returned as a result they do not
545 allocate, so maybe we don't want to charge so much for them If you
546 have a non-zero discount here, we find that workers often get inlined
547 back into wrappers, because it look like
548 f x = case $wf x of (# a,b #) -> (a,b)
549 and we are keener because of the case. However while this change
550 shrank binary sizes by 0.5% it also made spectral/boyer allocate 5%
551 more. All other changes were very small. So it's not a big deal but I
552 didn't adopt the idea.
555 primOpSize :: PrimOp -> Int -> ExprSize
556 primOpSize op n_val_args
557 | not (primOpIsDupable op) = sizeN opt_UF_DearOp
558 | not (primOpOutOfLine op) = sizeN 1
559 -- Be very keen to inline simple primops.
560 -- We give a discount of 1 for each arg so that (op# x y z) costs 2.
561 -- We can't make it cost 1, else we'll inline let v = (op# x y z)
562 -- at every use of v, which is excessive.
564 -- A good example is:
565 -- let x = +# p q in C {x}
566 -- Even though x get's an occurrence of 'many', its RHS looks cheap,
567 -- and there's a good chance it'll get inlined back into C's RHS. Urgh!
569 | otherwise = sizeN n_val_args
572 buildSize :: ExprSize
573 buildSize = SizeIs (_ILIT(0)) emptyBag (_ILIT(4))
574 -- We really want to inline applications of build
575 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
576 -- Indeed, we should add a result_discount becuause build is
577 -- very like a constructor. We don't bother to check that the
578 -- build is saturated (it usually is). The "-2" discounts for the \c n,
579 -- The "4" is rather arbitrary.
581 augmentSize :: ExprSize
582 augmentSize = SizeIs (_ILIT(0)) emptyBag (_ILIT(4))
583 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
584 -- e plus ys. The -2 accounts for the \cn
586 -- When we return a lambda, give a discount if it's used (applied)
587 lamScrutDiscount :: ExprSize -> ExprSize
588 lamScrutDiscount (SizeIs n vs _) = SizeIs n vs (iUnbox opt_UF_FunAppDiscount)
589 lamScrutDiscount TooBig = TooBig
592 Note [addAltSize result discounts]
593 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
594 When adding the size of alternatives, we *add* the result discounts
595 too, rather than take the *maximum*. For a multi-branch case, this
596 gives a discount for each branch that returns a constructor, making us
597 keener to inline. I did try using 'max' instead, but it makes nofib
598 'rewrite' and 'puzzle' allocate significantly more, and didn't make
599 binary sizes shrink significantly either.
601 Note [Discounts and thresholds]
602 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
603 Constants for discounts and thesholds are defined in main/StaticFlags,
604 all of form opt_UF_xxxx. They are:
606 opt_UF_CreationThreshold (45)
607 At a definition site, if the unfolding is bigger than this, we
608 may discard it altogether
610 opt_UF_UseThreshold (6)
611 At a call site, if the unfolding, less discounts, is smaller than
612 this, then it's small enough inline
614 opt_UF_KeennessFactor (1.5)
615 Factor by which the discounts are multiplied before
616 subtracting from size
618 opt_UF_DictDiscount (1)
619 The discount for each occurrence of a dictionary argument
620 as an argument of a class method. Should be pretty small
621 else big functions may get inlined
623 opt_UF_FunAppDiscount (6)
624 Discount for a function argument that is applied. Quite
625 large, because if we inline we avoid the higher-order call.
628 The size of a foreign call or not-dupable PrimOp
631 Note [Function applications]
632 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
633 In a function application (f a b)
635 - If 'f' is an argument to the function being analysed,
636 and there's at least one value arg, record a FunAppDiscount for f
638 - If the application if a PAP (arity > 2 in this example)
639 record a *result* discount (because inlining
640 with "extra" args in the call may mean that we now
641 get a saturated application)
643 Code for manipulating sizes
646 data ExprSize = TooBig
647 | SizeIs FastInt -- Size found
648 (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
649 FastInt -- Size to subtract if result is scrutinised
650 -- by a case expression
652 instance Outputable ExprSize where
653 ppr TooBig = ptext (sLit "TooBig")
654 ppr (SizeIs a _ c) = brackets (int (iBox a) <+> int (iBox c))
656 -- subtract the discount before deciding whether to bale out. eg. we
657 -- want to inline a large constructor application into a selector:
658 -- tup = (a_1, ..., a_99)
659 -- x = case tup of ...
661 mkSizeIs :: FastInt -> FastInt -> Bag (Id, Int) -> FastInt -> ExprSize
662 mkSizeIs max n xs d | (n -# d) ># max = TooBig
663 | otherwise = SizeIs n xs d
665 maxSize :: ExprSize -> ExprSize -> ExprSize
666 maxSize TooBig _ = TooBig
667 maxSize _ TooBig = TooBig
668 maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
672 sizeN :: Int -> ExprSize
674 sizeZero = SizeIs (_ILIT(0)) emptyBag (_ILIT(0))
675 sizeN n = SizeIs (iUnbox n) emptyBag (_ILIT(0))
679 %************************************************************************
681 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
683 %************************************************************************
685 We use 'couldBeSmallEnoughToInline' to avoid exporting inlinings that
686 we ``couldn't possibly use'' on the other side. Can be overridden w/
687 flaggery. Just the same as smallEnoughToInline, except that it has no
691 couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
692 couldBeSmallEnoughToInline threshold rhs
693 = case sizeExpr (iUnbox threshold) [] body of
697 (_, body) = collectBinders rhs
700 smallEnoughToInline :: Unfolding -> Bool
701 smallEnoughToInline (CoreUnfolding {uf_guidance = UnfIfGoodArgs {ug_size = size}})
702 = size <= opt_UF_UseThreshold
703 smallEnoughToInline _
707 certainlyWillInline :: Unfolding -> Bool
708 -- Sees if the unfolding is pretty certain to inline
709 certainlyWillInline (CoreUnfolding { uf_is_cheap = is_cheap, uf_arity = n_vals, uf_guidance = guidance })
713 UnfIfGoodArgs { ug_size = size}
714 -> is_cheap && size - (n_vals +1) <= opt_UF_UseThreshold
716 certainlyWillInline _
720 %************************************************************************
722 \subsection{callSiteInline}
724 %************************************************************************
726 This is the key function. It decides whether to inline a variable at a call site
728 callSiteInline is used at call sites, so it is a bit more generous.
729 It's a very important function that embodies lots of heuristics.
730 A non-WHNF can be inlined if it doesn't occur inside a lambda,
731 and occurs exactly once or
732 occurs once in each branch of a case and is small
734 If the thing is in WHNF, there's no danger of duplicating work,
735 so we can inline if it occurs once, or is small
737 NOTE: we don't want to inline top-level functions that always diverge.
738 It just makes the code bigger. Tt turns out that the convenient way to prevent
739 them inlining is to give them a NOINLINE pragma, which we do in
740 StrictAnal.addStrictnessInfoToTopId
743 callSiteInline :: DynFlags
745 -> Bool -- True <=> unfolding is active
746 -> Bool -- True if there are are no arguments at all (incl type args)
747 -> [ArgSummary] -- One for each value arg; True if it is interesting
748 -> CallCtxt -- True <=> continuation is interesting
749 -> Maybe CoreExpr -- Unfolding, if any
751 instance Outputable ArgSummary where
752 ppr TrivArg = ptext (sLit "TrivArg")
753 ppr NonTrivArg = ptext (sLit "NonTrivArg")
754 ppr ValueArg = ptext (sLit "ValueArg")
756 data CallCtxt = BoringCtxt
758 | ArgCtxt -- We are somewhere in the argument of a function
759 Bool -- True <=> we're somewhere in the RHS of function with rules
760 -- False <=> we *are* the argument of a function with non-zero
763 -- we *are* the RHS of a let Note [RHS of lets]
764 -- In both cases, be a little keener to inline
766 | ValAppCtxt -- We're applied to at least one value arg
767 -- This arises when we have ((f x |> co) y)
768 -- Then the (f x) has argument 'x' but in a ValAppCtxt
770 | CaseCtxt -- We're the scrutinee of a case
771 -- that decomposes its scrutinee
773 instance Outputable CallCtxt where
774 ppr BoringCtxt = ptext (sLit "BoringCtxt")
775 ppr (ArgCtxt rules) = ptext (sLit "ArgCtxt") <+> ppr rules
776 ppr CaseCtxt = ptext (sLit "CaseCtxt")
777 ppr ValAppCtxt = ptext (sLit "ValAppCtxt")
779 callSiteInline dflags id active_unfolding lone_variable arg_infos cont_info
780 = case idUnfolding id of
781 -- idUnfolding checks for loop-breakers, returning NoUnfolding
782 -- Things with an INLINE pragma may have an unfolding *and*
783 -- be a loop breaker (maybe the knot is not yet untied)
784 CoreUnfolding { uf_tmpl = unf_template, uf_is_top = is_top
785 , uf_is_cheap = is_cheap, uf_arity = uf_arity
786 , uf_guidance = guidance, uf_expandable = is_exp }
787 | active_unfolding -> tryUnfolding dflags id lone_variable
788 arg_infos cont_info unf_template is_top
789 is_cheap is_exp uf_arity guidance
790 | otherwise -> Nothing
791 NoUnfolding -> Nothing
792 OtherCon {} -> Nothing
793 DFunUnfolding {} -> Nothing -- Never unfold a DFun
795 tryUnfolding :: DynFlags -> Id -> Bool -> [ArgSummary] -> CallCtxt
796 -> CoreExpr -> Bool -> Bool -> Bool -> Arity -> UnfoldingGuidance
798 tryUnfolding dflags id lone_variable
799 arg_infos cont_info unf_template is_top
800 is_cheap is_exp uf_arity guidance
801 -- uf_arity will typically be equal to (idArity id),
802 -- but may be less for InlineRules
803 | dopt Opt_D_dump_inlinings dflags && dopt Opt_D_verbose_core2core dflags
804 = pprTrace ("Considering inlining: " ++ showSDoc (ppr id))
805 (vcat [text "arg infos" <+> ppr arg_infos,
806 text "uf arity" <+> ppr uf_arity,
807 text "interesting continuation" <+> ppr cont_info,
808 text "some_benefit" <+> ppr some_benefit,
809 text "is exp:" <+> ppr is_exp,
810 text "is cheap:" <+> ppr is_cheap,
811 text "guidance" <+> ppr guidance,
813 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO"])
818 n_val_args = length arg_infos
819 saturated = n_val_args >= uf_arity
821 result | yes_or_no = Just unf_template
822 | otherwise = Nothing
824 interesting_args = any nonTriv arg_infos
825 -- NB: (any nonTriv arg_infos) looks at the
826 -- over-saturated args too which is "wrong";
827 -- but if over-saturated we inline anyway.
829 -- some_benefit is used when the RHS is small enough
830 -- and the call has enough (or too many) value
831 -- arguments (ie n_val_args >= arity). But there must
832 -- be *something* interesting about some argument, or the
833 -- result context, to make it worth inlining
835 | not saturated = interesting_args -- Under-saturated
836 -- Note [Unsaturated applications]
837 | n_val_args > uf_arity = True -- Over-saturated
838 | otherwise = interesting_args -- Saturated
839 || interesting_saturated_call
841 interesting_saturated_call
843 BoringCtxt -> not is_top && uf_arity > 0 -- Note [Nested functions]
844 CaseCtxt -> not (lone_variable && is_cheap) -- Note [Lone variables]
845 ArgCtxt {} -> uf_arity > 0 -- Note [Inlining in ArgCtxt]
846 ValAppCtxt -> True -- Note [Cast then apply]
848 (yes_or_no, extra_doc)
850 UnfNever -> (False, empty)
852 UnfWhen unsat_ok boring_ok
853 -> (enough_args && (boring_ok || some_benefit), empty )
854 where -- See Note [INLINE for small functions]
855 enough_args = saturated || (unsat_ok && n_val_args > 0)
857 UnfIfGoodArgs { ug_args = arg_discounts, ug_res = res_discount, ug_size = size }
858 -> ( is_cheap && some_benefit && small_enough
859 , (text "discounted size =" <+> int discounted_size) )
861 discounted_size = size - discount
862 small_enough = discounted_size <= opt_UF_UseThreshold
863 discount = computeDiscount uf_arity arg_discounts
864 res_discount arg_infos cont_info
869 Be a tiny bit keener to inline in the RHS of a let, because that might
870 lead to good thing later
872 g y = let x = f y in ...(case x of (a,b,c) -> ...) ...
873 We'd inline 'f' if the call was in a case context, and it kind-of-is,
874 only we can't see it. So we treat the RHS of a let as not-totally-boring.
876 Note [Unsaturated applications]
877 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
878 When a call is not saturated, we *still* inline if one of the
879 arguments has interesting structure. That's sometimes very important.
880 A good example is the Ord instance for Bool in Base:
883 $fOrdBool =GHC.Classes.D:Ord
888 $cmin_ajX [Occ=LoopBreaker] :: Bool -> Bool -> Bool
889 $cmin_ajX = GHC.Classes.$dmmin @ Bool $fOrdBool
892 But the defn of GHC.Classes.$dmmin is:
894 $dmmin :: forall a. GHC.Classes.Ord a => a -> a -> a
895 {- Arity: 3, HasNoCafRefs, Strictness: SLL,
896 Unfolding: (\ @ a $dOrd :: GHC.Classes.Ord a x :: a y :: a ->
897 case @ a GHC.Classes.<= @ a $dOrd x y of wild {
898 GHC.Types.False -> y GHC.Types.True -> x }) -}
900 We *really* want to inline $dmmin, even though it has arity 3, in
901 order to unravel the recursion.
904 Note [Things to watch]
905 ~~~~~~~~~~~~~~~~~~~~~~
906 * { y = I# 3; x = y `cast` co; ...case (x `cast` co) of ... }
907 Assume x is exported, so not inlined unconditionally.
908 Then we want x to inline unconditionally; no reason for it
909 not to, and doing so avoids an indirection.
911 * { x = I# 3; ....f x.... }
912 Make sure that x does not inline unconditionally!
913 Lest we get extra allocation.
915 Note [Inlining an InlineRule]
916 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
917 An InlineRules is used for
918 (a) programmer INLINE pragmas
919 (b) inlinings from worker/wrapper
921 For (a) the RHS may be large, and our contract is that we *only* inline
922 when the function is applied to all the arguments on the LHS of the
923 source-code defn. (The uf_arity in the rule.)
925 However for worker/wrapper it may be worth inlining even if the
926 arity is not satisfied (as we do in the CoreUnfolding case) so we don't
930 Note [Nested functions]
931 ~~~~~~~~~~~~~~~~~~~~~~~
932 If a function has a nested defn we also record some-benefit, on the
933 grounds that we are often able to eliminate the binding, and hence the
934 allocation, for the function altogether; this is good for join points.
935 But this only makes sense for *functions*; inlining a constructor
936 doesn't help allocation unless the result is scrutinised. UNLESS the
937 constructor occurs just once, albeit possibly in multiple case
938 branches. Then inlining it doesn't increase allocation, but it does
939 increase the chance that the constructor won't be allocated at all in
940 the branches that don't use it.
942 Note [Cast then apply]
943 ~~~~~~~~~~~~~~~~~~~~~~
945 myIndex = __inline_me ( (/\a. <blah>) |> co )
946 co :: (forall a. a -> a) ~ (forall a. T a)
947 ... /\a.\x. case ((myIndex a) |> sym co) x of { ... } ...
949 We need to inline myIndex to unravel this; but the actual call (myIndex a) has
950 no value arguments. The ValAppCtxt gives it enough incentive to inline.
952 Note [Inlining in ArgCtxt]
953 ~~~~~~~~~~~~~~~~~~~~~~~~~~
954 The condition (arity > 0) here is very important, because otherwise
955 we end up inlining top-level stuff into useless places; eg
958 This can make a very big difference: it adds 16% to nofib 'integer' allocs,
961 At one stage I replaced this condition by 'True' (leading to the above
962 slow-down). The motivation was test eyeball/inline1.hs; but that seems
965 NOTE: arguably, we should inline in ArgCtxt only if the result of the
966 call is at least CONLIKE. At least for the cases where we use ArgCtxt
967 for the RHS of a 'let', we only profit from the inlining if we get a
968 CONLIKE thing (modulo lets).
970 Note [Lone variables] See also Note [Interaction of exprIsCheap and lone variables]
971 ~~~~~~~~~~~~~~~~~~~~~ which appears below
972 The "lone-variable" case is important. I spent ages messing about
973 with unsatisfactory varaints, but this is nice. The idea is that if a
974 variable appears all alone
976 as an arg of lazy fn, or rhs BoringCtxt
977 as scrutinee of a case CaseCtxt
978 as arg of a fn ArgCtxt
980 it is bound to a cheap expression
982 then we should not inline it (unless there is some other reason,
983 e.g. is is the sole occurrence). That is what is happening at
984 the use of 'lone_variable' in 'interesting_saturated_call'.
986 Why? At least in the case-scrutinee situation, turning
987 let x = (a,b) in case x of y -> ...
989 let x = (a,b) in case (a,b) of y -> ...
991 let x = (a,b) in let y = (a,b) in ...
992 is bad if the binding for x will remain.
994 Another example: I discovered that strings
995 were getting inlined straight back into applications of 'error'
996 because the latter is strict.
998 f = \x -> ...(error s)...
1000 Fundamentally such contexts should not encourage inlining because the
1001 context can ``see'' the unfolding of the variable (e.g. case or a
1002 RULE) so there's no gain. If the thing is bound to a value.
1007 foo = _inline_ (\n. [n])
1008 bar = _inline_ (foo 20)
1009 baz = \n. case bar of { (m:_) -> m + n }
1010 Here we really want to inline 'bar' so that we can inline 'foo'
1011 and the whole thing unravels as it should obviously do. This is
1012 important: in the NDP project, 'bar' generates a closure data
1013 structure rather than a list.
1015 So the non-inlining of lone_variables should only apply if the
1016 unfolding is regarded as cheap; because that is when exprIsConApp_maybe
1017 looks through the unfolding. Hence the "&& is_cheap" in the
1020 * Even a type application or coercion isn't a lone variable.
1022 case $fMonadST @ RealWorld of { :DMonad a b c -> c }
1023 We had better inline that sucker! The case won't see through it.
1025 For now, I'm treating treating a variable applied to types
1026 in a *lazy* context "lone". The motivating example was
1028 g = /\a. \y. h (f a)
1029 There's no advantage in inlining f here, and perhaps
1030 a significant disadvantage. Hence some_val_args in the Stop case
1032 Note [Interaction of exprIsCheap and lone variables]
1033 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1034 The lone-variable test says "don't inline if a case expression
1035 scrutines a lone variable whose unfolding is cheap". It's very
1036 important that, under these circumstances, exprIsConApp_maybe
1037 can spot a constructor application. So, for example, we don't
1040 to be cheap, and that's good because exprIsConApp_maybe doesn't
1041 think that expression is a constructor application.
1043 I used to test is_value rather than is_cheap, which was utterly
1044 wrong, because the above expression responds True to exprIsHNF.
1046 This kind of thing can occur if you have
1049 foo = let x = e in (x,x)
1054 computeDiscount :: Int -> [Int] -> Int -> [ArgSummary] -> CallCtxt -> Int
1055 computeDiscount n_vals_wanted arg_discounts res_discount arg_infos cont_info
1056 -- We multiple the raw discounts (args_discount and result_discount)
1057 -- ty opt_UnfoldingKeenessFactor because the former have to do with
1058 -- *size* whereas the discounts imply that there's some extra
1059 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
1062 = 1 -- Discount of 1 because the result replaces the call
1063 -- so we count 1 for the function itself
1065 + length (take n_vals_wanted arg_infos)
1066 -- Discount of (un-scaled) 1 for each arg supplied,
1067 -- because the result replaces the call
1069 + round (opt_UF_KeenessFactor *
1070 fromIntegral (arg_discount + res_discount'))
1072 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
1074 mk_arg_discount _ TrivArg = 0
1075 mk_arg_discount _ NonTrivArg = 1
1076 mk_arg_discount discount ValueArg = discount
1078 res_discount' = case cont_info of
1080 CaseCtxt -> res_discount
1081 _other -> 4 `min` res_discount
1082 -- res_discount can be very large when a function returns
1083 -- constructors; but we only want to invoke that large discount
1084 -- when there's a case continuation.
1085 -- Otherwise we, rather arbitrarily, threshold it. Yuk.
1086 -- But we want to aovid inlining large functions that return
1087 -- constructors into contexts that are simply "interesting"
1090 %************************************************************************
1092 Interesting arguments
1094 %************************************************************************
1096 Note [Interesting arguments]
1097 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1098 An argument is interesting if it deserves a discount for unfoldings
1099 with a discount in that argument position. The idea is to avoid
1100 unfolding a function that is applied only to variables that have no
1101 unfolding (i.e. they are probably lambda bound): f x y z There is
1102 little point in inlining f here.
1104 Generally, *values* (like (C a b) and (\x.e)) deserve discounts. But
1105 we must look through lets, eg (let x = e in C a b), because the let will
1106 float, exposing the value, if we inline. That makes it different to
1109 Before 2009 we said it was interesting if the argument had *any* structure
1110 at all; i.e. (hasSomeUnfolding v). But does too much inlining; see Trac #3016.
1112 But we don't regard (f x y) as interesting, unless f is unsaturated.
1113 If it's saturated and f hasn't inlined, then it's probably not going
1116 Note [Conlike is interesting]
1117 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1119 f d = ...((*) d x y)...
1121 where df is con-like. Then we'd really like to inline 'f' so that the
1122 rule for (*) (df d) can fire. To do this
1123 a) we give a discount for being an argument of a class-op (eg (*) d)
1124 b) we say that a con-like argument (eg (df d)) is interesting
1127 data ArgSummary = TrivArg -- Nothing interesting
1128 | NonTrivArg -- Arg has structure
1129 | ValueArg -- Arg is a con-app or PAP
1130 -- ..or con-like. Note [Conlike is interesting]
1132 interestingArg :: CoreExpr -> ArgSummary
1133 -- See Note [Interesting arguments]
1134 interestingArg e = go e 0
1136 -- n is # value args to which the expression is applied
1137 go (Lit {}) _ = ValueArg
1139 | isConLikeId v = ValueArg -- Experimenting with 'conlike' rather that
1140 -- data constructors here
1141 | idArity v > n = ValueArg -- Catches (eg) primops with arity but no unfolding
1142 | n > 0 = NonTrivArg -- Saturated or unknown call
1143 | conlike_unfolding = ValueArg -- n==0; look for an interesting unfolding
1144 -- See Note [Conlike is interesting]
1145 | otherwise = TrivArg -- n==0, no useful unfolding
1147 conlike_unfolding = isConLikeUnfolding (idUnfolding v)
1149 go (Type _) _ = TrivArg
1150 go (App fn (Type _)) n = go fn n
1151 go (App fn _) n = go fn (n+1)
1152 go (Note _ a) n = go a n
1153 go (Cast e _) n = go e n
1155 | isTyCoVar v = go e n
1157 | otherwise = ValueArg
1158 go (Let _ e) n = case go e n of { ValueArg -> ValueArg; _ -> NonTrivArg }
1159 go (Case {}) _ = NonTrivArg
1161 nonTriv :: ArgSummary -> Bool
1162 nonTriv TrivArg = False
1166 %************************************************************************
1170 %************************************************************************
1172 Note [exprIsConApp_maybe]
1173 ~~~~~~~~~~~~~~~~~~~~~~~~~
1174 exprIsConApp_maybe is a very important function. There are two principal
1176 * case e of { .... }
1177 * cls_op e, where cls_op is a class operation
1179 In both cases you want to know if e is of form (C e1..en) where C is
1182 However e might not *look* as if
1185 -- | Returns @Just (dc, [t1..tk], [x1..xn])@ if the argument expression is
1186 -- a *saturated* constructor application of the form @dc t1..tk x1 .. xn@,
1187 -- where t1..tk are the *universally-qantified* type args of 'dc'
1188 exprIsConApp_maybe :: IdUnfoldingFun -> CoreExpr -> Maybe (DataCon, [Type], [CoreExpr])
1190 exprIsConApp_maybe id_unf (Note note expr)
1192 = exprIsConApp_maybe id_unf expr
1193 -- We ignore all notes except SCCs. For example,
1194 -- case _scc_ "foo" (C a b) of
1196 -- should not be optimised away, because we'll lose the
1197 -- entry count on 'foo'; see Trac #4414
1199 exprIsConApp_maybe id_unf (Cast expr co)
1200 = -- Here we do the KPush reduction rule as described in the FC paper
1201 -- The transformation applies iff we have
1202 -- (C e1 ... en) `cast` co
1203 -- where co :: (T t1 .. tn) ~ to_ty
1204 -- The left-hand one must be a T, because exprIsConApp returned True
1205 -- but the right-hand one might not be. (Though it usually will.)
1207 case exprIsConApp_maybe id_unf expr of {
1208 Nothing -> Nothing ;
1209 Just (dc, _dc_univ_args, dc_args) ->
1211 let (_from_ty, to_ty) = coercionKind co
1212 dc_tc = dataConTyCon dc
1214 case splitTyConApp_maybe to_ty of {
1215 Nothing -> Nothing ;
1216 Just (to_tc, to_tc_arg_tys)
1217 | dc_tc /= to_tc -> Nothing
1218 -- These two Nothing cases are possible; we might see
1219 -- (C x y) `cast` (g :: T a ~ S [a]),
1220 -- where S is a type function. In fact, exprIsConApp
1221 -- will probably not be called in such circumstances,
1222 -- but there't nothing wrong with it
1226 tc_arity = tyConArity dc_tc
1227 dc_univ_tyvars = dataConUnivTyVars dc
1228 dc_ex_tyvars = dataConExTyVars dc
1229 arg_tys = dataConRepArgTys dc
1231 dc_eqs :: [(Type,Type)] -- All equalities from the DataCon
1232 dc_eqs = [(mkTyVarTy tv, ty) | (tv,ty) <- dataConEqSpec dc] ++
1233 [getEqPredTys eq_pred | eq_pred <- dataConEqTheta dc]
1235 (ex_args, rest1) = splitAtList dc_ex_tyvars dc_args
1236 (co_args, val_args) = splitAtList dc_eqs rest1
1238 -- Make the "theta" from Fig 3 of the paper
1239 gammas = decomposeCo tc_arity co
1240 theta = zipOpenTvSubst (dc_univ_tyvars ++ dc_ex_tyvars)
1241 (gammas ++ stripTypeArgs ex_args)
1243 -- Cast the existential coercion arguments
1244 cast_co (ty1, ty2) (Type co)
1245 = Type $ mkSymCoercion (substTy theta ty1)
1246 `mkTransCoercion` co
1247 `mkTransCoercion` (substTy theta ty2)
1248 cast_co _ other_arg = pprPanic "cast_co" (ppr other_arg)
1249 new_co_args = zipWith cast_co dc_eqs co_args
1251 -- Cast the value arguments (which include dictionaries)
1252 new_val_args = zipWith cast_arg arg_tys val_args
1253 cast_arg arg_ty arg = mkCoerce (substTy theta arg_ty) arg
1256 let dump_doc = vcat [ppr dc, ppr dc_univ_tyvars, ppr dc_ex_tyvars,
1257 ppr arg_tys, ppr dc_args, ppr _dc_univ_args,
1258 ppr ex_args, ppr val_args]
1260 ASSERT2( coreEqType _from_ty (mkTyConApp dc_tc _dc_univ_args), dump_doc )
1261 ASSERT2( all isTypeArg (ex_args ++ co_args), dump_doc )
1262 ASSERT2( equalLength val_args arg_tys, dump_doc )
1265 Just (dc, to_tc_arg_tys, ex_args ++ new_co_args ++ new_val_args)
1268 exprIsConApp_maybe id_unf expr
1271 analyse (App fun arg) args = analyse fun (arg:args)
1272 analyse fun@(Lam {}) args = beta fun [] args
1274 analyse (Var fun) args
1275 | Just con <- isDataConWorkId_maybe fun
1276 , count isValArg args == idArity fun
1277 , let (univ_ty_args, rest_args) = splitAtList (dataConUnivTyVars con) args
1278 = Just (con, stripTypeArgs univ_ty_args, rest_args)
1280 -- Look through dictionary functions; see Note [Unfolding DFuns]
1281 | DFunUnfolding dfun_nargs con ops <- unfolding
1282 , let sat = length args == dfun_nargs -- See Note [DFun arity check]
1283 in if sat then True else
1284 pprTrace "Unsaturated dfun" (ppr fun <+> int dfun_nargs $$ ppr args) False
1285 , let (dfun_tvs, _n_theta, _cls, dfun_res_tys) = tcSplitDFunTy (idType fun)
1286 subst = zipOpenTvSubst dfun_tvs (stripTypeArgs (takeList dfun_tvs args))
1287 mk_arg (DFunConstArg e) = e
1288 mk_arg (DFunLamArg i) = args !! i
1289 mk_arg (DFunPolyArg e) = mkApps e args
1290 = Just (con, substTys subst dfun_res_tys, map mk_arg ops)
1292 -- Look through unfoldings, but only cheap ones, because
1293 -- we are effectively duplicating the unfolding
1294 | Just rhs <- expandUnfolding_maybe unfolding
1295 = -- pprTrace "expanding" (ppr fun $$ ppr rhs) $
1298 unfolding = id_unf fun
1300 analyse _ _ = Nothing
1303 beta (Lam v body) pairs (arg : args)
1305 = beta body ((v,arg):pairs) args
1307 beta (Lam {}) _ _ -- Un-saturated, or not a type lambda
1311 = analyse (substExpr (text "subst-expr-is-con-app") subst fun) args
1313 subst = mkOpenSubst (mkInScopeSet (exprFreeVars fun)) pairs
1314 -- doc = vcat [ppr fun, ppr expr, ppr pairs, ppr args]
1317 stripTypeArgs :: [CoreExpr] -> [Type]
1318 stripTypeArgs args = ASSERT2( all isTypeArg args, ppr args )
1319 [ty | Type ty <- args]
1322 Note [Unfolding DFuns]
1323 ~~~~~~~~~~~~~~~~~~~~~~
1326 df :: forall a b. (Eq a, Eq b) -> Eq (a,b)
1327 df a b d_a d_b = MkEqD (a,b) ($c1 a b d_a d_b)
1330 So to split it up we just need to apply the ops $c1, $c2 etc
1331 to the very same args as the dfun. It takes a little more work
1332 to compute the type arguments to the dictionary constructor.
1334 Note [DFun arity check]
1335 ~~~~~~~~~~~~~~~~~~~~~~~
1336 Here we check that the total number of supplied arguments (inclding
1337 type args) matches what the dfun is expecting. This may be *less*
1338 than the ordinary arity of the dfun: see Note [DFun unfoldings] in CoreSyn