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
56 %************************************************************************
58 \subsection{Making unfoldings}
60 %************************************************************************
63 mkTopUnfolding :: CoreExpr -> Unfolding
64 mkTopUnfolding expr = mkUnfolding True {- Top level -} expr
66 mkUnfolding :: Bool -> CoreExpr -> Unfolding
67 mkUnfolding top_lvl expr
68 = CoreUnfolding (occurAnalyseExpr expr)
75 -- OK to inline inside a lambda
77 (calcUnfoldingGuidance opt_UF_CreationThreshold expr)
78 -- Sometimes during simplification, there's a large let-bound thing
79 -- which has been substituted, and so is now dead; so 'expr' contains
80 -- two copies of the thing while the occurrence-analysed expression doesn't
81 -- Nevertheless, we don't occ-analyse before computing the size because the
82 -- size computation bales out after a while, whereas occurrence analysis does not.
84 -- This can occasionally mean that the guidance is very pessimistic;
85 -- it gets fixed up next round
87 instance Outputable Unfolding where
88 ppr NoUnfolding = ptext SLIT("No unfolding")
89 ppr (OtherCon cs) = ptext SLIT("OtherCon") <+> ppr cs
90 ppr (CompulsoryUnfolding e) = ptext SLIT("Compulsory") <+> ppr e
91 ppr (CoreUnfolding e top hnf cheap g)
92 = ptext SLIT("Unf") <+> sep [ppr top <+> ppr hnf <+> ppr cheap <+> ppr g,
95 mkCompulsoryUnfolding :: CoreExpr -> Unfolding
96 mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
97 = CompulsoryUnfolding (occurAnalyseExpr expr)
101 %************************************************************************
103 \subsection{The UnfoldingGuidance type}
105 %************************************************************************
108 instance Outputable UnfoldingGuidance where
109 ppr UnfoldNever = ptext SLIT("NEVER")
110 ppr (UnfoldIfGoodArgs v cs size discount)
111 = hsep [ ptext SLIT("IF_ARGS"), int v,
112 brackets (hsep (map int cs)),
119 calcUnfoldingGuidance
120 :: Int -- bomb out if size gets bigger than this
121 -> CoreExpr -- expression to look at
123 calcUnfoldingGuidance bOMB_OUT_SIZE expr
124 = case collect_val_bndrs expr of { (inline, val_binders, body) ->
126 n_val_binders = length val_binders
128 max_inline_size = n_val_binders+2
129 -- The idea is that if there is an INLINE pragma (inline is True)
130 -- and there's a big body, we give a size of n_val_binders+2. This
131 -- This is just enough to fail the no-size-increase test in callSiteInline,
132 -- so that INLINE things don't get inlined into entirely boring contexts,
136 case (sizeExpr (iUnbox bOMB_OUT_SIZE) val_binders body) of
139 | not inline -> UnfoldNever
140 -- A big function with an INLINE pragma must
141 -- have an UnfoldIfGoodArgs guidance
142 | otherwise -> UnfoldIfGoodArgs n_val_binders
143 (map (const 0) val_binders)
146 SizeIs size cased_args scrut_discount
149 (map discount_for val_binders)
151 (iBox scrut_discount)
153 boxed_size = iBox size
155 final_size | inline = boxed_size `min` max_inline_size
156 | otherwise = boxed_size
158 -- Sometimes an INLINE thing is smaller than n_val_binders+2.
159 -- A particular case in point is a constructor, which has size 1.
160 -- We want to inline this regardless, hence the `min`
162 discount_for b = foldlBag (\acc (b',n) -> if b==b' then acc+n else acc)
166 collect_val_bndrs e = go False [] e
167 -- We need to be a bit careful about how we collect the
168 -- value binders. In ptic, if we see
169 -- __inline_me (\x y -> e)
170 -- We want to say "2 value binders". Why? So that
171 -- we take account of information given for the arguments
173 go _ rev_vbs (Note InlineMe e) = go True rev_vbs e
174 go inline rev_vbs (Lam b e) | isId b = go inline (b:rev_vbs) e
175 | otherwise = go inline rev_vbs e
176 go inline rev_vbs e = (inline, reverse rev_vbs, e)
180 sizeExpr :: FastInt -- Bomb out if it gets bigger than this
181 -> [Id] -- Arguments; we're interested in which of these
186 sizeExpr bOMB_OUT_SIZE top_args expr
189 size_up (Type _) = sizeZero -- Types cost nothing
190 size_up (Var _) = sizeOne
192 size_up (Note InlineMe _) = sizeOne -- Inline notes make it look very small
193 -- This can be important. If you have an instance decl like this:
194 -- instance Foo a => Foo [a] where
195 -- {-# INLINE op1, op2 #-}
198 -- then we'll get a dfun which is a pair of two INLINE lambdas
200 size_up (Note _ body) = size_up body -- Other notes cost nothing
202 size_up (Cast e _) = size_up e
204 size_up (App fun (Type _)) = size_up fun
205 size_up (App fun arg) = size_up_app fun [arg]
207 size_up (Lit lit) = sizeN (litSize lit)
209 size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
210 | otherwise = size_up e
212 size_up (Let (NonRec binder rhs) body)
213 = nukeScrutDiscount (size_up rhs) `addSize`
214 size_up body `addSizeN`
215 (if isUnLiftedType (idType binder) then 0 else 1)
216 -- For the allocation
217 -- If the binder has an unlifted type there is no allocation
219 size_up (Let (Rec pairs) body)
220 = nukeScrutDiscount rhs_size `addSize`
221 size_up body `addSizeN`
222 length pairs -- For the allocation
224 rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
226 size_up (Case (Var v) _ _ alts)
227 | v `elem` top_args -- We are scrutinising an argument variable
229 {- I'm nuking this special case; BUT see the comment with case alternatives.
231 (a) It's too eager. We don't want to inline a wrapper into a
232 context with no benefit.
233 E.g. \ x. f (x+x) no point in inlining (+) here!
235 (b) It's ineffective. Once g's wrapper is inlined, its case-expressions
236 aren't scrutinising arguments any more
240 [alt] -> size_up_alt alt `addSize` SizeIs (_ILIT(0)) (unitBag (v, 1)) (_ILIT(0))
241 -- We want to make wrapper-style evaluation look cheap, so that
242 -- when we inline a wrapper it doesn't make call site (much) bigger
243 -- Otherwise we get nasty phase ordering stuff:
246 -- If we inline g's wrapper, f looks big, and doesn't get inlined
247 -- into h; if we inline f first, while it looks small, then g's
248 -- wrapper will get inlined later anyway. To avoid this nasty
249 -- ordering difference, we make (case a of (x,y) -> ...),
250 -- *where a is one of the arguments* look free.
254 alts_size (foldr addSize sizeOne alt_sizes) -- The 1 is for the scrutinee
255 (foldr1 maxSize alt_sizes)
257 -- Good to inline if an arg is scrutinised, because
258 -- that may eliminate allocation in the caller
259 -- And it eliminates the case itself
262 alt_sizes = map size_up_alt alts
264 -- alts_size tries to compute a good discount for
265 -- the case when we are scrutinising an argument variable
266 alts_size (SizeIs tot _tot_disc _tot_scrut) -- Size of all alternatives
267 (SizeIs max max_disc max_scrut) -- Size of biggest alternative
268 = SizeIs tot (unitBag (v, iBox (_ILIT(1) +# tot -# max)) `unionBags` max_disc) max_scrut
269 -- If the variable is known, we produce a discount that
270 -- will take us back to 'max', the size of rh largest alternative
271 -- The 1+ is a little discount for reduced allocation in the caller
272 alts_size tot_size _ = tot_size
274 size_up (Case e _ _ alts) = nukeScrutDiscount (size_up e) `addSize`
275 foldr (addSize . size_up_alt) sizeZero alts
276 -- We don't charge for the case itself
277 -- It's a strict thing, and the price of the call
278 -- is paid by scrut. Also consider
279 -- case f x of DEFAULT -> e
280 -- This is just ';'! Don't charge for it.
283 size_up_app (App fun arg) args
284 | isTypeArg arg = size_up_app fun args
285 | otherwise = size_up_app fun (arg:args)
286 size_up_app fun args = foldr (addSize . nukeScrutDiscount . size_up)
287 (size_up_fun fun args)
290 -- A function application with at least one value argument
291 -- so if the function is an argument give it an arg-discount
293 -- Also behave specially if the function is a build
295 -- Also if the function is a constant Id (constr or primop)
296 -- compute discounts specially
297 size_up_fun (Var fun) args
298 | fun `hasKey` buildIdKey = buildSize
299 | fun `hasKey` augmentIdKey = augmentSize
301 = case globalIdDetails fun of
302 DataConWorkId dc -> conSizeN dc (valArgCount args)
304 FCallId _ -> sizeN opt_UF_DearOp
305 PrimOpId op -> primOpSize op (valArgCount args)
306 -- foldr addSize (primOpSize op) (map arg_discount args)
307 -- At one time I tried giving an arg-discount if a primop
308 -- is applied to one of the function's arguments, but it's
309 -- not good. At the moment, any unlifted-type arg gets a
310 -- 'True' for 'yes I'm evald', so we collect the discount even
311 -- if we know nothing about it. And just having it in a primop
312 -- doesn't help at all if we don't know something more.
314 _ -> fun_discount fun `addSizeN`
315 (1 + length (filter (not . exprIsTrivial) args))
316 -- The 1+ is for the function itself
317 -- Add 1 for each non-trivial arg;
318 -- the allocation cost, as in let(rec)
319 -- Slight hack here: for constructors the args are almost always
320 -- trivial; and for primops they are almost always prim typed
321 -- We should really only count for non-prim-typed args in the
322 -- general case, but that seems too much like hard work
324 size_up_fun other _ = size_up other
327 size_up_alt (_con, _bndrs, rhs) = size_up rhs
328 -- Don't charge for args, so that wrappers look cheap
329 -- (See comments about wrappers with Case)
332 -- We want to record if we're case'ing, or applying, an argument
333 fun_discount v | v `elem` top_args = SizeIs (_ILIT(0)) (unitBag (v, opt_UF_FunAppDiscount)) (_ILIT(0))
334 fun_discount _ = sizeZero
337 -- These addSize things have to be here because
338 -- I don't want to give them bOMB_OUT_SIZE as an argument
340 addSizeN TooBig _ = TooBig
341 addSizeN (SizeIs n xs d) m = mkSizeIs bOMB_OUT_SIZE (n +# iUnbox m) xs d
343 addSize TooBig _ = TooBig
344 addSize _ TooBig = TooBig
345 addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
346 = mkSizeIs bOMB_OUT_SIZE (n1 +# n2) (xs `unionBags` ys) (d1 +# d2)
349 Code for manipulating sizes
352 data ExprSize = TooBig
353 | SizeIs FastInt -- Size found
354 (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
355 FastInt -- Size to subtract if result is scrutinised
356 -- by a case expression
358 -- subtract the discount before deciding whether to bale out. eg. we
359 -- want to inline a large constructor application into a selector:
360 -- tup = (a_1, ..., a_99)
361 -- x = case tup of ...
363 mkSizeIs :: FastInt -> FastInt -> Bag (Id, Int) -> FastInt -> ExprSize
364 mkSizeIs max n xs d | (n -# d) ># max = TooBig
365 | otherwise = SizeIs n xs d
367 maxSize :: ExprSize -> ExprSize -> ExprSize
368 maxSize TooBig _ = TooBig
369 maxSize _ TooBig = TooBig
370 maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
373 sizeZero, sizeOne :: ExprSize
374 sizeN :: Int -> ExprSize
375 conSizeN :: DataCon ->Int -> ExprSize
377 sizeZero = SizeIs (_ILIT(0)) emptyBag (_ILIT(0))
378 sizeOne = SizeIs (_ILIT(1)) emptyBag (_ILIT(0))
379 sizeN n = SizeIs (iUnbox n) emptyBag (_ILIT(0))
381 | isUnboxedTupleCon dc = SizeIs (_ILIT(0)) emptyBag (iUnbox n +# _ILIT(1))
382 | otherwise = SizeIs (_ILIT(1)) emptyBag (iUnbox n +# _ILIT(1))
383 -- Treat constructors as size 1; we are keen to expose them
384 -- (and we charge separately for their args). We can't treat
385 -- them as size zero, else we find that (iBox x) has size 1,
386 -- which is the same as a lone variable; and hence 'v' will
387 -- always be replaced by (iBox x), where v is bound to iBox x.
389 -- However, unboxed tuples count as size zero
390 -- I found occasions where we had
391 -- f x y z = case op# x y z of { s -> (# s, () #) }
392 -- and f wasn't getting inlined
394 primOpSize :: PrimOp -> Int -> ExprSize
396 | not (primOpIsDupable op) = sizeN opt_UF_DearOp
397 | not (primOpOutOfLine op) = sizeN (2 - n_args)
398 -- Be very keen to inline simple primops.
399 -- We give a discount of 1 for each arg so that (op# x y z) costs 2.
400 -- We can't make it cost 1, else we'll inline let v = (op# x y z)
401 -- at every use of v, which is excessive.
403 -- A good example is:
404 -- let x = +# p q in C {x}
405 -- Even though x get's an occurrence of 'many', its RHS looks cheap,
406 -- and there's a good chance it'll get inlined back into C's RHS. Urgh!
407 | otherwise = sizeOne
409 buildSize :: ExprSize
410 buildSize = SizeIs (_ILIT(-2)) emptyBag (_ILIT(4))
411 -- We really want to inline applications of build
412 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
413 -- Indeed, we should add a result_discount becuause build is
414 -- very like a constructor. We don't bother to check that the
415 -- build is saturated (it usually is). The "-2" discounts for the \c n,
416 -- The "4" is rather arbitrary.
418 augmentSize :: ExprSize
419 augmentSize = SizeIs (_ILIT(-2)) emptyBag (_ILIT(4))
420 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
421 -- e plus ys. The -2 accounts for the \cn
423 nukeScrutDiscount :: ExprSize -> ExprSize
424 nukeScrutDiscount (SizeIs n vs _) = SizeIs n vs (_ILIT(0))
425 nukeScrutDiscount TooBig = TooBig
427 -- When we return a lambda, give a discount if it's used (applied)
428 lamScrutDiscount :: ExprSize -> ExprSize
429 lamScrutDiscount (SizeIs n vs _) = case opt_UF_FunAppDiscount of { d -> SizeIs n vs (iUnbox d) }
430 lamScrutDiscount TooBig = TooBig
434 %************************************************************************
436 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
438 %************************************************************************
440 We have very limited information about an unfolding expression: (1)~so
441 many type arguments and so many value arguments expected---for our
442 purposes here, we assume we've got those. (2)~A ``size'' or ``cost,''
443 a single integer. (3)~An ``argument info'' vector. For this, what we
444 have at the moment is a Boolean per argument position that says, ``I
445 will look with great favour on an explicit constructor in this
446 position.'' (4)~The ``discount'' to subtract if the expression
447 is being scrutinised.
449 Assuming we have enough type- and value arguments (if not, we give up
450 immediately), then we see if the ``discounted size'' is below some
451 (semi-arbitrary) threshold. It works like this: for every argument
452 position where we're looking for a constructor AND WE HAVE ONE in our
453 hands, we get a (again, semi-arbitrary) discount [proportion to the
454 number of constructors in the type being scrutinized].
456 If we're in the context of a scrutinee ( \tr{(case <expr > of A .. -> ...;.. )})
457 and the expression in question will evaluate to a constructor, we use
458 the computed discount size *for the result only* rather than
459 computing the argument discounts. Since we know the result of
460 the expression is going to be taken apart, discounting its size
461 is more accurate (see @sizeExpr@ above for how this discount size
464 We use this one to avoid exporting inlinings that we ``couldn't possibly
465 use'' on the other side. Can be overridden w/ flaggery.
466 Just the same as smallEnoughToInline, except that it has no actual arguments.
469 couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
470 couldBeSmallEnoughToInline threshold rhs = case calcUnfoldingGuidance threshold rhs of
474 certainlyWillInline :: Unfolding -> Bool
475 -- Sees if the unfolding is pretty certain to inline
476 certainlyWillInline (CoreUnfolding _ _ _ is_cheap (UnfoldIfGoodArgs n_vals _ size _))
477 = is_cheap && size - (n_vals +1) <= opt_UF_UseThreshold
478 certainlyWillInline _
481 smallEnoughToInline :: Unfolding -> Bool
482 smallEnoughToInline (CoreUnfolding _ _ _ _ (UnfoldIfGoodArgs _ _ size _))
483 = size <= opt_UF_UseThreshold
484 smallEnoughToInline _
488 %************************************************************************
490 \subsection{callSiteInline}
492 %************************************************************************
494 This is the key function. It decides whether to inline a variable at a call site
496 callSiteInline is used at call sites, so it is a bit more generous.
497 It's a very important function that embodies lots of heuristics.
498 A non-WHNF can be inlined if it doesn't occur inside a lambda,
499 and occurs exactly once or
500 occurs once in each branch of a case and is small
502 If the thing is in WHNF, there's no danger of duplicating work,
503 so we can inline if it occurs once, or is small
505 NOTE: we don't want to inline top-level functions that always diverge.
506 It just makes the code bigger. Tt turns out that the convenient way to prevent
507 them inlining is to give them a NOINLINE pragma, which we do in
508 StrictAnal.addStrictnessInfoToTopId
511 callSiteInline :: DynFlags
512 -> Bool -- True <=> the Id can be inlined
514 -> Bool -- True if there are are no arguments at all (incl type args)
515 -> [Bool] -- One for each value arg; True if it is interesting
516 -> CallCtxt -- True <=> continuation is interesting
517 -> Maybe CoreExpr -- Unfolding, if any
520 data CallCtxt = BoringCtxt
522 | ArgCtxt Bool -- We're somewhere in the RHS of function with rules
523 -- => be keener to inline
524 Int -- We *are* the argument of a function with this arg discount
525 -- => be keener to inline
526 -- INVARIANT: ArgCtxt False 0 ==> BoringCtxt
528 | CaseCtxt -- We're the scrutinee of a case
529 -- that decomposes its scrutinee
531 instance Outputable CallCtxt where
532 ppr BoringCtxt = ptext SLIT("BoringCtxt")
533 ppr (ArgCtxt _ _) = ptext SLIT("ArgCtxt")
534 ppr CaseCtxt = ptext SLIT("CaseCtxt")
536 callSiteInline dflags active_inline id lone_variable arg_infos cont_info
537 = case idUnfolding id of {
538 NoUnfolding -> Nothing ;
539 OtherCon _ -> Nothing ;
541 CompulsoryUnfolding unf_template -> Just unf_template ;
542 -- CompulsoryUnfolding => there is no top-level binding
543 -- for these things, so we must inline it.
544 -- Only a couple of primop-like things have
545 -- compulsory unfoldings (see MkId.lhs).
546 -- We don't allow them to be inactive
548 CoreUnfolding unf_template is_top is_value is_cheap guidance ->
551 result | yes_or_no = Just unf_template
552 | otherwise = Nothing
554 n_val_args = length arg_infos
556 yes_or_no = active_inline && is_cheap && consider_safe
557 -- We consider even the once-in-one-branch
558 -- occurrences, because they won't all have been
559 -- caught by preInlineUnconditionally. In particular,
560 -- if the occurrence is once inside a lambda, and the
561 -- rhs is cheap but not a manifest lambda, then
562 -- pre-inline will not have inlined it for fear of
563 -- invalidating the occurrence info in the rhs.
566 -- consider_safe decides whether it's a good idea to
567 -- inline something, given that there's no
568 -- work-duplication issue (the caller checks that).
571 UnfoldIfGoodArgs n_vals_wanted arg_discounts size res_discount
572 | enough_args && size <= (n_vals_wanted + 1)
573 -- Inline unconditionally if there no size increase
574 -- Size of call is n_vals_wanted (+1 for the function)
578 -> some_benefit && small_enough
581 enough_args = n_val_args >= n_vals_wanted
583 some_benefit = or arg_infos || really_interesting_cont
584 -- There must be something interesting
585 -- about some argument, or the result
586 -- context, to make it worth inlining
588 really_interesting_cont
589 | n_val_args < n_vals_wanted = False -- Too few args
590 | n_val_args == n_vals_wanted = interesting_saturated_call
591 | otherwise = True -- Extra args
592 -- really_interesting_cont tells if the result of the
593 -- call is in an interesting context.
595 interesting_saturated_call
597 BoringCtxt -> not is_top && n_vals_wanted > 0 -- Note [Nested functions]
598 CaseCtxt -> not lone_variable || not is_value -- Note [Lone variables]
600 -- Was: n_vals_wanted > 0; but see test eyeball/inline1.hs
602 small_enough = (size - discount) <= opt_UF_UseThreshold
603 discount = computeDiscount n_vals_wanted arg_discounts
604 res_discount' arg_infos
605 res_discount' = case cont_info of
607 CaseCtxt -> res_discount
608 ArgCtxt _ _ -> 4 `min` res_discount
609 -- res_discount can be very large when a function returns
610 -- construtors; but we only want to invoke that large discount
611 -- when there's a case continuation.
612 -- Otherwise we, rather arbitrarily, threshold it. Yuk.
613 -- But we want to aovid inlining large functions that return
614 -- constructors into contexts that are simply "interesting"
617 if dopt Opt_D_dump_inlinings dflags then
618 pprTrace "Considering inlining"
619 (ppr id <+> vcat [text "active:" <+> ppr active_inline,
620 text "arg infos" <+> ppr arg_infos,
621 text "interesting continuation" <+> ppr cont_info,
622 text "is value:" <+> ppr is_value,
623 text "is cheap:" <+> ppr is_cheap,
624 text "guidance" <+> ppr guidance,
625 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO"])
632 Note [Nested functions]
633 ~~~~~~~~~~~~~~~~~~~~~~~
634 If a function has a nested defn we also record some-benefit, on the
635 grounds that we are often able to eliminate the binding, and hence the
636 allocation, for the function altogether; this is good for join points.
637 But this only makes sense for *functions*; inlining a constructor
638 doesn't help allocation unless the result is scrutinised. UNLESS the
639 constructor occurs just once, albeit possibly in multiple case
640 branches. Then inlining it doesn't increase allocation, but it does
641 increase the chance that the constructor won't be allocated at all in
642 the branches that don't use it.
644 Note [Lone variables]
645 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
646 The "lone-variable" case is important. I spent ages messing about
647 with unsatisfactory varaints, but this is nice. The idea is that if a
648 variable appears all alone
649 as an arg of lazy fn, or rhs Stop
650 as scrutinee of a case Select
651 as arg of a strict fn ArgOf
653 it is bound to a value
654 then we should not inline it (unless there is some other reason,
655 e.g. is is the sole occurrence). That is what is happening at
656 the use of 'lone_variable' in 'interesting_saturated_call'.
658 Why? At least in the case-scrutinee situation, turning
659 let x = (a,b) in case x of y -> ...
661 let x = (a,b) in case (a,b) of y -> ...
663 let x = (a,b) in let y = (a,b) in ...
664 is bad if the binding for x will remain.
666 Another example: I discovered that strings
667 were getting inlined straight back into applications of 'error'
668 because the latter is strict.
670 f = \x -> ...(error s)...
672 Fundamentally such contexts should not encourage inlining because the
673 context can ``see'' the unfolding of the variable (e.g. case or a
674 RULE) so there's no gain. If the thing is bound to a value.
679 foo = _inline_ (\n. [n])
680 bar = _inline_ (foo 20)
681 baz = \n. case bar of { (m:_) -> m + n }
682 Here we really want to inline 'bar' so that we can inline 'foo'
683 and the whole thing unravels as it should obviously do. This is
684 important: in the NDP project, 'bar' generates a closure data
685 structure rather than a list.
687 * Even a type application or coercion isn't a lone variable.
689 case $fMonadST @ RealWorld of { :DMonad a b c -> c }
690 We had better inline that sucker! The case won't see through it.
692 For now, I'm treating treating a variable applied to types
693 in a *lazy* context "lone". The motivating example was
696 There's no advantage in inlining f here, and perhaps
697 a significant disadvantage. Hence some_val_args in the Stop case
700 computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Int
701 computeDiscount n_vals_wanted arg_discounts result_discount arg_infos
702 -- We multiple the raw discounts (args_discount and result_discount)
703 -- ty opt_UnfoldingKeenessFactor because the former have to do with
704 -- *size* whereas the discounts imply that there's some extra
705 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
708 -- we also discount 1 for each argument passed, because these will
709 -- reduce with the lambdas in the function (we count 1 for a lambda
711 = 1 + -- Discount of 1 because the result replaces the call
712 -- so we count 1 for the function itself
713 length (take n_vals_wanted arg_infos) +
714 -- Discount of 1 for each arg supplied, because the
715 -- result replaces the call
716 round (opt_UF_KeenessFactor *
717 fromIntegral (arg_discount + result_discount))
719 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
721 mk_arg_discount discount is_evald | is_evald = discount