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, mkTopUnfolding, mkUnfolding, mkCompulsoryUnfolding, seqUnfolding,
22 evaldUnfolding, mkOtherCon, otherCons,
23 unfoldingTemplate, maybeUnfoldingTemplate,
24 isEvaldUnfolding, isValueUnfolding, isCheapUnfolding, isCompulsoryUnfolding,
25 hasUnfolding, hasSomeUnfolding, neverUnfold,
27 couldBeSmallEnoughToInline,
28 certainlyWillInline, smallEnoughToInline,
30 callSiteInline, CallCtxt(..)
34 #include "HsVersions.h"
39 import PprCore () -- Instances
57 %************************************************************************
59 \subsection{Making unfoldings}
61 %************************************************************************
64 mkTopUnfolding :: CoreExpr -> Unfolding
65 mkTopUnfolding expr = mkUnfolding True {- Top level -} expr
67 mkUnfolding :: Bool -> CoreExpr -> Unfolding
68 mkUnfolding top_lvl expr
69 = CoreUnfolding (occurAnalyseExpr expr)
76 -- OK to inline inside a lambda
78 (calcUnfoldingGuidance opt_UF_CreationThreshold expr)
79 -- Sometimes during simplification, there's a large let-bound thing
80 -- which has been substituted, and so is now dead; so 'expr' contains
81 -- two copies of the thing while the occurrence-analysed expression doesn't
82 -- Nevertheless, we don't occ-analyse before computing the size because the
83 -- size computation bales out after a while, whereas occurrence analysis does not.
85 -- This can occasionally mean that the guidance is very pessimistic;
86 -- it gets fixed up next round
88 instance Outputable Unfolding where
89 ppr NoUnfolding = ptext SLIT("No unfolding")
90 ppr (OtherCon cs) = ptext SLIT("OtherCon") <+> ppr cs
91 ppr (CompulsoryUnfolding e) = ptext SLIT("Compulsory") <+> ppr e
92 ppr (CoreUnfolding e top hnf cheap g)
93 = ptext SLIT("Unf") <+> sep [ppr top <+> ppr hnf <+> ppr cheap <+> ppr g,
96 mkCompulsoryUnfolding :: CoreExpr -> Unfolding
97 mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
98 = CompulsoryUnfolding (occurAnalyseExpr expr)
102 %************************************************************************
104 \subsection{The UnfoldingGuidance type}
106 %************************************************************************
109 instance Outputable UnfoldingGuidance where
110 ppr UnfoldNever = ptext SLIT("NEVER")
111 ppr (UnfoldIfGoodArgs v cs size discount)
112 = hsep [ ptext SLIT("IF_ARGS"), int v,
113 brackets (hsep (map int cs)),
120 calcUnfoldingGuidance
121 :: Int -- bomb out if size gets bigger than this
122 -> CoreExpr -- expression to look at
124 calcUnfoldingGuidance bOMB_OUT_SIZE expr
125 = case collect_val_bndrs expr of { (inline, val_binders, body) ->
127 n_val_binders = length val_binders
129 max_inline_size = n_val_binders+2
130 -- The idea is that if there is an INLINE pragma (inline is True)
131 -- and there's a big body, we give a size of n_val_binders+2. This
132 -- This is just enough to fail the no-size-increase test in callSiteInline,
133 -- so that INLINE things don't get inlined into entirely boring contexts,
137 case (sizeExpr (iUnbox bOMB_OUT_SIZE) val_binders body) of
140 | not inline -> UnfoldNever
141 -- A big function with an INLINE pragma must
142 -- have an UnfoldIfGoodArgs guidance
143 | otherwise -> UnfoldIfGoodArgs n_val_binders
144 (map (const 0) val_binders)
147 SizeIs size cased_args scrut_discount
150 (map discount_for val_binders)
152 (iBox scrut_discount)
154 boxed_size = iBox size
156 final_size | inline = boxed_size `min` max_inline_size
157 | otherwise = boxed_size
159 -- Sometimes an INLINE thing is smaller than n_val_binders+2.
160 -- A particular case in point is a constructor, which has size 1.
161 -- We want to inline this regardless, hence the `min`
163 discount_for b = foldlBag (\acc (b',n) -> if b==b' then acc+n else acc)
167 collect_val_bndrs e = go False [] e
168 -- We need to be a bit careful about how we collect the
169 -- value binders. In ptic, if we see
170 -- __inline_me (\x y -> e)
171 -- We want to say "2 value binders". Why? So that
172 -- we take account of information given for the arguments
174 go _ rev_vbs (Note InlineMe e) = go True rev_vbs e
175 go inline rev_vbs (Lam b e) | isId b = go inline (b:rev_vbs) e
176 | otherwise = go inline rev_vbs e
177 go inline rev_vbs e = (inline, reverse rev_vbs, e)
181 sizeExpr :: FastInt -- Bomb out if it gets bigger than this
182 -> [Id] -- Arguments; we're interested in which of these
187 sizeExpr bOMB_OUT_SIZE top_args expr
190 size_up (Type _) = sizeZero -- Types cost nothing
191 size_up (Var _) = sizeOne
193 size_up (Note InlineMe _) = sizeOne -- Inline notes make it look very small
194 -- This can be important. If you have an instance decl like this:
195 -- instance Foo a => Foo [a] where
196 -- {-# INLINE op1, op2 #-}
199 -- then we'll get a dfun which is a pair of two INLINE lambdas
201 size_up (Note _ body) = size_up body -- Other notes cost nothing
203 size_up (Cast e _) = size_up e
205 size_up (App fun (Type _)) = size_up fun
206 size_up (App fun arg) = size_up_app fun [arg]
208 size_up (Lit lit) = sizeN (litSize lit)
210 size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
211 | otherwise = size_up e
213 size_up (Let (NonRec binder rhs) body)
214 = nukeScrutDiscount (size_up rhs) `addSize`
215 size_up body `addSizeN`
216 (if isUnLiftedType (idType binder) then 0 else 1)
217 -- For the allocation
218 -- If the binder has an unlifted type there is no allocation
220 size_up (Let (Rec pairs) body)
221 = nukeScrutDiscount rhs_size `addSize`
222 size_up body `addSizeN`
223 length pairs -- For the allocation
225 rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
227 size_up (Case (Var v) _ _ alts)
228 | v `elem` top_args -- We are scrutinising an argument variable
230 {- I'm nuking this special case; BUT see the comment with case alternatives.
232 (a) It's too eager. We don't want to inline a wrapper into a
233 context with no benefit.
234 E.g. \ x. f (x+x) no point in inlining (+) here!
236 (b) It's ineffective. Once g's wrapper is inlined, its case-expressions
237 aren't scrutinising arguments any more
241 [alt] -> size_up_alt alt `addSize` SizeIs (_ILIT(0)) (unitBag (v, 1)) (_ILIT(0))
242 -- We want to make wrapper-style evaluation look cheap, so that
243 -- when we inline a wrapper it doesn't make call site (much) bigger
244 -- Otherwise we get nasty phase ordering stuff:
247 -- If we inline g's wrapper, f looks big, and doesn't get inlined
248 -- into h; if we inline f first, while it looks small, then g's
249 -- wrapper will get inlined later anyway. To avoid this nasty
250 -- ordering difference, we make (case a of (x,y) -> ...),
251 -- *where a is one of the arguments* look free.
255 alts_size (foldr addSize sizeOne alt_sizes) -- The 1 is for the scrutinee
256 (foldr1 maxSize alt_sizes)
258 -- Good to inline if an arg is scrutinised, because
259 -- that may eliminate allocation in the caller
260 -- And it eliminates the case itself
263 alt_sizes = map size_up_alt alts
265 -- alts_size tries to compute a good discount for
266 -- the case when we are scrutinising an argument variable
267 alts_size (SizeIs tot _tot_disc _tot_scrut) -- Size of all alternatives
268 (SizeIs max max_disc max_scrut) -- Size of biggest alternative
269 = SizeIs tot (unitBag (v, iBox (_ILIT(1) +# tot -# max)) `unionBags` max_disc) max_scrut
270 -- If the variable is known, we produce a discount that
271 -- will take us back to 'max', the size of rh largest alternative
272 -- The 1+ is a little discount for reduced allocation in the caller
273 alts_size tot_size _ = tot_size
275 size_up (Case e _ _ alts) = nukeScrutDiscount (size_up e) `addSize`
276 foldr (addSize . size_up_alt) sizeZero alts
277 -- We don't charge for the case itself
278 -- It's a strict thing, and the price of the call
279 -- is paid by scrut. Also consider
280 -- case f x of DEFAULT -> e
281 -- This is just ';'! Don't charge for it.
284 size_up_app (App fun arg) args
285 | isTypeArg arg = size_up_app fun args
286 | otherwise = size_up_app fun (arg:args)
287 size_up_app fun args = foldr (addSize . nukeScrutDiscount . size_up)
288 (size_up_fun fun args)
291 -- A function application with at least one value argument
292 -- so if the function is an argument give it an arg-discount
294 -- Also behave specially if the function is a build
296 -- Also if the function is a constant Id (constr or primop)
297 -- compute discounts specially
298 size_up_fun (Var fun) args
299 | fun `hasKey` buildIdKey = buildSize
300 | fun `hasKey` augmentIdKey = augmentSize
302 = case globalIdDetails fun of
303 DataConWorkId dc -> conSizeN dc (valArgCount args)
305 FCallId _ -> sizeN opt_UF_DearOp
306 PrimOpId op -> primOpSize op (valArgCount args)
307 -- foldr addSize (primOpSize op) (map arg_discount args)
308 -- At one time I tried giving an arg-discount if a primop
309 -- is applied to one of the function's arguments, but it's
310 -- not good. At the moment, any unlifted-type arg gets a
311 -- 'True' for 'yes I'm evald', so we collect the discount even
312 -- if we know nothing about it. And just having it in a primop
313 -- doesn't help at all if we don't know something more.
315 _ -> fun_discount fun `addSizeN`
316 (1 + length (filter (not . exprIsTrivial) args))
317 -- The 1+ is for the function itself
318 -- Add 1 for each non-trivial arg;
319 -- the allocation cost, as in let(rec)
320 -- Slight hack here: for constructors the args are almost always
321 -- trivial; and for primops they are almost always prim typed
322 -- We should really only count for non-prim-typed args in the
323 -- general case, but that seems too much like hard work
325 size_up_fun other _ = size_up other
328 size_up_alt (_con, _bndrs, rhs) = size_up rhs
329 -- Don't charge for args, so that wrappers look cheap
330 -- (See comments about wrappers with Case)
333 -- We want to record if we're case'ing, or applying, an argument
334 fun_discount v | v `elem` top_args = SizeIs (_ILIT(0)) (unitBag (v, opt_UF_FunAppDiscount)) (_ILIT(0))
335 fun_discount _ = sizeZero
338 -- These addSize things have to be here because
339 -- I don't want to give them bOMB_OUT_SIZE as an argument
341 addSizeN TooBig _ = TooBig
342 addSizeN (SizeIs n xs d) m = mkSizeIs bOMB_OUT_SIZE (n +# iUnbox m) xs d
344 addSize TooBig _ = TooBig
345 addSize _ TooBig = TooBig
346 addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
347 = mkSizeIs bOMB_OUT_SIZE (n1 +# n2) (xs `unionBags` ys) (d1 +# d2)
350 Code for manipulating sizes
353 data ExprSize = TooBig
354 | SizeIs FastInt -- Size found
355 (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
356 FastInt -- Size to subtract if result is scrutinised
357 -- by a case expression
359 -- subtract the discount before deciding whether to bale out. eg. we
360 -- want to inline a large constructor application into a selector:
361 -- tup = (a_1, ..., a_99)
362 -- x = case tup of ...
364 mkSizeIs :: FastInt -> FastInt -> Bag (Id, Int) -> FastInt -> ExprSize
365 mkSizeIs max n xs d | (n -# d) ># max = TooBig
366 | otherwise = SizeIs n xs d
368 maxSize :: ExprSize -> ExprSize -> ExprSize
369 maxSize TooBig _ = TooBig
370 maxSize _ TooBig = TooBig
371 maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
374 sizeZero, sizeOne :: ExprSize
375 sizeN :: Int -> ExprSize
376 conSizeN :: DataCon ->Int -> ExprSize
378 sizeZero = SizeIs (_ILIT(0)) emptyBag (_ILIT(0))
379 sizeOne = SizeIs (_ILIT(1)) emptyBag (_ILIT(0))
380 sizeN n = SizeIs (iUnbox n) emptyBag (_ILIT(0))
382 | isUnboxedTupleCon dc = SizeIs (_ILIT(0)) emptyBag (iUnbox n +# _ILIT(1))
383 | otherwise = SizeIs (_ILIT(1)) emptyBag (iUnbox n +# _ILIT(1))
384 -- Treat constructors as size 1; we are keen to expose them
385 -- (and we charge separately for their args). We can't treat
386 -- them as size zero, else we find that (iBox x) has size 1,
387 -- which is the same as a lone variable; and hence 'v' will
388 -- always be replaced by (iBox x), where v is bound to iBox x.
390 -- However, unboxed tuples count as size zero
391 -- I found occasions where we had
392 -- f x y z = case op# x y z of { s -> (# s, () #) }
393 -- and f wasn't getting inlined
395 primOpSize :: PrimOp -> Int -> ExprSize
397 | not (primOpIsDupable op) = sizeN opt_UF_DearOp
398 | not (primOpOutOfLine op) = sizeN (2 - n_args)
399 -- Be very keen to inline simple primops.
400 -- We give a discount of 1 for each arg so that (op# x y z) costs 2.
401 -- We can't make it cost 1, else we'll inline let v = (op# x y z)
402 -- at every use of v, which is excessive.
404 -- A good example is:
405 -- let x = +# p q in C {x}
406 -- Even though x get's an occurrence of 'many', its RHS looks cheap,
407 -- and there's a good chance it'll get inlined back into C's RHS. Urgh!
408 | otherwise = sizeOne
410 buildSize :: ExprSize
411 buildSize = SizeIs (_ILIT(-2)) emptyBag (_ILIT(4))
412 -- We really want to inline applications of build
413 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
414 -- Indeed, we should add a result_discount becuause build is
415 -- very like a constructor. We don't bother to check that the
416 -- build is saturated (it usually is). The "-2" discounts for the \c n,
417 -- The "4" is rather arbitrary.
419 augmentSize :: ExprSize
420 augmentSize = SizeIs (_ILIT(-2)) emptyBag (_ILIT(4))
421 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
422 -- e plus ys. The -2 accounts for the \cn
424 nukeScrutDiscount :: ExprSize -> ExprSize
425 nukeScrutDiscount (SizeIs n vs _) = SizeIs n vs (_ILIT(0))
426 nukeScrutDiscount TooBig = TooBig
428 -- When we return a lambda, give a discount if it's used (applied)
429 lamScrutDiscount :: ExprSize -> ExprSize
430 lamScrutDiscount (SizeIs n vs _) = case opt_UF_FunAppDiscount of { d -> SizeIs n vs (iUnbox d) }
431 lamScrutDiscount TooBig = TooBig
435 %************************************************************************
437 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
439 %************************************************************************
441 We have very limited information about an unfolding expression: (1)~so
442 many type arguments and so many value arguments expected---for our
443 purposes here, we assume we've got those. (2)~A ``size'' or ``cost,''
444 a single integer. (3)~An ``argument info'' vector. For this, what we
445 have at the moment is a Boolean per argument position that says, ``I
446 will look with great favour on an explicit constructor in this
447 position.'' (4)~The ``discount'' to subtract if the expression
448 is being scrutinised.
450 Assuming we have enough type- and value arguments (if not, we give up
451 immediately), then we see if the ``discounted size'' is below some
452 (semi-arbitrary) threshold. It works like this: for every argument
453 position where we're looking for a constructor AND WE HAVE ONE in our
454 hands, we get a (again, semi-arbitrary) discount [proportion to the
455 number of constructors in the type being scrutinized].
457 If we're in the context of a scrutinee ( \tr{(case <expr > of A .. -> ...;.. )})
458 and the expression in question will evaluate to a constructor, we use
459 the computed discount size *for the result only* rather than
460 computing the argument discounts. Since we know the result of
461 the expression is going to be taken apart, discounting its size
462 is more accurate (see @sizeExpr@ above for how this discount size
465 We use this one to avoid exporting inlinings that we ``couldn't possibly
466 use'' on the other side. Can be overridden w/ flaggery.
467 Just the same as smallEnoughToInline, except that it has no actual arguments.
470 couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
471 couldBeSmallEnoughToInline threshold rhs = case calcUnfoldingGuidance threshold rhs of
475 certainlyWillInline :: Unfolding -> Bool
476 -- Sees if the unfolding is pretty certain to inline
477 certainlyWillInline (CoreUnfolding _ _ _ is_cheap (UnfoldIfGoodArgs n_vals _ size _))
478 = is_cheap && size - (n_vals +1) <= opt_UF_UseThreshold
479 certainlyWillInline _
482 smallEnoughToInline :: Unfolding -> Bool
483 smallEnoughToInline (CoreUnfolding _ _ _ _ (UnfoldIfGoodArgs _ _ size _))
484 = size <= opt_UF_UseThreshold
485 smallEnoughToInline _
489 %************************************************************************
491 \subsection{callSiteInline}
493 %************************************************************************
495 This is the key function. It decides whether to inline a variable at a call site
497 callSiteInline is used at call sites, so it is a bit more generous.
498 It's a very important function that embodies lots of heuristics.
499 A non-WHNF can be inlined if it doesn't occur inside a lambda,
500 and occurs exactly once or
501 occurs once in each branch of a case and is small
503 If the thing is in WHNF, there's no danger of duplicating work,
504 so we can inline if it occurs once, or is small
506 NOTE: we don't want to inline top-level functions that always diverge.
507 It just makes the code bigger. Tt turns out that the convenient way to prevent
508 them inlining is to give them a NOINLINE pragma, which we do in
509 StrictAnal.addStrictnessInfoToTopId
512 callSiteInline :: DynFlags
513 -> Bool -- True <=> the Id can be inlined
515 -> Bool -- True if there are are no arguments at all (incl type args)
516 -> [Bool] -- One for each value arg; True if it is interesting
517 -> CallCtxt -- True <=> continuation is interesting
518 -> Maybe CoreExpr -- Unfolding, if any
521 data CallCtxt = BoringCtxt
523 | ArgCtxt Bool -- We're somewhere in the RHS of function with rules
524 -- => be keener to inline
525 Int -- We *are* the argument of a function with this arg discount
526 -- => be keener to inline
527 -- INVARIANT: ArgCtxt False 0 ==> BoringCtxt
529 | CaseCtxt -- We're the scrutinee of a case
530 -- that decomposes its scrutinee
532 instance Outputable CallCtxt where
533 ppr BoringCtxt = ptext SLIT("BoringCtxt")
534 ppr (ArgCtxt _ _) = ptext SLIT("ArgCtxt")
535 ppr CaseCtxt = ptext SLIT("CaseCtxt")
537 callSiteInline dflags active_inline id lone_variable arg_infos cont_info
538 = case idUnfolding id of {
539 NoUnfolding -> Nothing ;
540 OtherCon _ -> Nothing ;
542 CompulsoryUnfolding unf_template -> Just unf_template ;
543 -- CompulsoryUnfolding => there is no top-level binding
544 -- for these things, so we must inline it.
545 -- Only a couple of primop-like things have
546 -- compulsory unfoldings (see MkId.lhs).
547 -- We don't allow them to be inactive
549 CoreUnfolding unf_template is_top is_value is_cheap guidance ->
552 result | yes_or_no = Just unf_template
553 | otherwise = Nothing
555 n_val_args = length arg_infos
557 yes_or_no = active_inline && is_cheap && consider_safe
558 -- We consider even the once-in-one-branch
559 -- occurrences, because they won't all have been
560 -- caught by preInlineUnconditionally. In particular,
561 -- if the occurrence is once inside a lambda, and the
562 -- rhs is cheap but not a manifest lambda, then
563 -- pre-inline will not have inlined it for fear of
564 -- invalidating the occurrence info in the rhs.
567 -- consider_safe decides whether it's a good idea to
568 -- inline something, given that there's no
569 -- work-duplication issue (the caller checks that).
572 UnfoldIfGoodArgs n_vals_wanted arg_discounts size res_discount
573 | enough_args && size <= (n_vals_wanted + 1)
574 -- Inline unconditionally if there no size increase
575 -- Size of call is n_vals_wanted (+1 for the function)
579 -> some_benefit && small_enough
582 enough_args = n_val_args >= n_vals_wanted
584 some_benefit = or arg_infos || really_interesting_cont
585 -- There must be something interesting
586 -- about some argument, or the result
587 -- context, to make it worth inlining
589 really_interesting_cont
590 | n_val_args < n_vals_wanted = False -- Too few args
591 | n_val_args == n_vals_wanted = interesting_saturated_call
592 | otherwise = True -- Extra args
593 -- really_interesting_cont tells if the result of the
594 -- call is in an interesting context.
596 interesting_saturated_call
598 BoringCtxt -> not is_top && n_vals_wanted > 0 -- Note [Nested functions]
599 CaseCtxt -> not lone_variable || not is_value -- Note [Lone variables]
601 -- Was: n_vals_wanted > 0; but see test eyeball/inline1.hs
603 small_enough = (size - discount) <= opt_UF_UseThreshold
604 discount = computeDiscount n_vals_wanted arg_discounts
605 res_discount' arg_infos
606 res_discount' = case cont_info of
608 CaseCtxt -> res_discount
609 ArgCtxt _ _ -> 4 `min` res_discount
610 -- res_discount can be very large when a function returns
611 -- construtors; but we only want to invoke that large discount
612 -- when there's a case continuation.
613 -- Otherwise we, rather arbitrarily, threshold it. Yuk.
614 -- But we want to aovid inlining large functions that return
615 -- constructors into contexts that are simply "interesting"
618 if dopt Opt_D_dump_inlinings dflags then
619 pprTrace "Considering inlining"
620 (ppr id <+> vcat [text "active:" <+> ppr active_inline,
621 text "arg infos" <+> ppr arg_infos,
622 text "interesting continuation" <+> ppr cont_info,
623 text "is value:" <+> ppr is_value,
624 text "is cheap:" <+> ppr is_cheap,
625 text "guidance" <+> ppr guidance,
626 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO"])
633 Note [Nested functions]
634 ~~~~~~~~~~~~~~~~~~~~~~~
635 If a function has a nested defn we also record some-benefit, on the
636 grounds that we are often able to eliminate the binding, and hence the
637 allocation, for the function altogether; this is good for join points.
638 But this only makes sense for *functions*; inlining a constructor
639 doesn't help allocation unless the result is scrutinised. UNLESS the
640 constructor occurs just once, albeit possibly in multiple case
641 branches. Then inlining it doesn't increase allocation, but it does
642 increase the chance that the constructor won't be allocated at all in
643 the branches that don't use it.
645 Note [Lone variables]
646 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
647 The "lone-variable" case is important. I spent ages messing about
648 with unsatisfactory varaints, but this is nice. The idea is that if a
649 variable appears all alone
650 as an arg of lazy fn, or rhs Stop
651 as scrutinee of a case Select
652 as arg of a strict fn ArgOf
654 it is bound to a value
655 then we should not inline it (unless there is some other reason,
656 e.g. is is the sole occurrence). That is what is happening at
657 the use of 'lone_variable' in 'interesting_saturated_call'.
659 Why? At least in the case-scrutinee situation, turning
660 let x = (a,b) in case x of y -> ...
662 let x = (a,b) in case (a,b) of y -> ...
664 let x = (a,b) in let y = (a,b) in ...
665 is bad if the binding for x will remain.
667 Another example: I discovered that strings
668 were getting inlined straight back into applications of 'error'
669 because the latter is strict.
671 f = \x -> ...(error s)...
673 Fundamentally such contexts should not encourage inlining because the
674 context can ``see'' the unfolding of the variable (e.g. case or a
675 RULE) so there's no gain. If the thing is bound to a value.
680 foo = _inline_ (\n. [n])
681 bar = _inline_ (foo 20)
682 baz = \n. case bar of { (m:_) -> m + n }
683 Here we really want to inline 'bar' so that we can inline 'foo'
684 and the whole thing unravels as it should obviously do. This is
685 important: in the NDP project, 'bar' generates a closure data
686 structure rather than a list.
688 * Even a type application or coercion isn't a lone variable.
690 case $fMonadST @ RealWorld of { :DMonad a b c -> c }
691 We had better inline that sucker! The case won't see through it.
693 For now, I'm treating treating a variable applied to types
694 in a *lazy* context "lone". The motivating example was
697 There's no advantage in inlining f here, and perhaps
698 a significant disadvantage. Hence some_val_args in the Stop case
701 computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Int
702 computeDiscount n_vals_wanted arg_discounts result_discount arg_infos
703 -- We multiple the raw discounts (args_discount and result_discount)
704 -- ty opt_UnfoldingKeenessFactor because the former have to do with
705 -- *size* whereas the discounts imply that there's some extra
706 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
709 -- we also discount 1 for each argument passed, because these will
710 -- reduce with the lambdas in the function (we count 1 for a lambda
712 = 1 + -- Discount of 1 because the result replaces the call
713 -- so we count 1 for the function itself
714 length (take n_vals_wanted arg_infos) +
715 -- Discount of 1 for each arg supplied, because the
716 -- result replaces the call
717 round (opt_UF_KeenessFactor *
718 fromIntegral (arg_discount + result_discount))
720 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
722 mk_arg_discount discount is_evald | is_evald = discount