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, mkImplicitUnfolding, mkUnfolding,
22 mkCompulsoryUnfolding, seqUnfolding,
23 evaldUnfolding, mkOtherCon, otherCons,
24 unfoldingTemplate, maybeUnfoldingTemplate,
25 isEvaldUnfolding, isValueUnfolding, isCheapUnfolding, isCompulsoryUnfolding,
26 hasUnfolding, hasSomeUnfolding, neverUnfold,
28 couldBeSmallEnoughToInline,
29 certainlyWillInline, smallEnoughToInline,
31 callSiteInline, CallCtxt(..)
38 import PprCore () -- Instances
40 import CoreSubst ( Subst, emptySubst, substTy, extendIdSubst, extendTvSubst
41 , lookupIdSubst, substBndr, substBndrs, substRecBndrs )
48 import Type hiding( substTy, extendTvSubst )
58 %************************************************************************
60 \subsection{Making unfoldings}
62 %************************************************************************
65 mkTopUnfolding :: CoreExpr -> Unfolding
66 mkTopUnfolding expr = mkUnfolding True {- Top level -} expr
68 mkImplicitUnfolding :: CoreExpr -> Unfolding
69 -- For implicit Ids, do a tiny bit of optimising first
70 mkImplicitUnfolding expr
71 = CoreUnfolding (simpleOptExpr emptySubst expr)
75 (calcUnfoldingGuidance opt_UF_CreationThreshold expr)
77 mkUnfolding :: Bool -> CoreExpr -> Unfolding
78 mkUnfolding top_lvl expr
79 = CoreUnfolding (occurAnalyseExpr expr)
86 -- OK to inline inside a lambda
88 (calcUnfoldingGuidance opt_UF_CreationThreshold expr)
89 -- Sometimes during simplification, there's a large let-bound thing
90 -- which has been substituted, and so is now dead; so 'expr' contains
91 -- two copies of the thing while the occurrence-analysed expression doesn't
92 -- Nevertheless, we don't occ-analyse before computing the size because the
93 -- size computation bales out after a while, whereas occurrence analysis does not.
95 -- This can occasionally mean that the guidance is very pessimistic;
96 -- it gets fixed up next round
98 instance Outputable Unfolding where
99 ppr NoUnfolding = ptext (sLit "No unfolding")
100 ppr (OtherCon cs) = ptext (sLit "OtherCon") <+> ppr cs
101 ppr (CompulsoryUnfolding e) = ptext (sLit "Compulsory") <+> ppr e
102 ppr (CoreUnfolding e top hnf cheap g)
103 = ptext (sLit "Unf") <+> sep [ppr top <+> ppr hnf <+> ppr cheap <+> ppr g,
106 mkCompulsoryUnfolding :: CoreExpr -> Unfolding
107 mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
108 = CompulsoryUnfolding (occurAnalyseExpr expr)
112 %************************************************************************
114 \subsection{The UnfoldingGuidance type}
116 %************************************************************************
119 instance Outputable UnfoldingGuidance where
120 ppr UnfoldNever = ptext (sLit "NEVER")
121 ppr (UnfoldIfGoodArgs v cs size discount)
122 = hsep [ ptext (sLit "IF_ARGS"), int v,
123 brackets (hsep (map int cs)),
130 calcUnfoldingGuidance
131 :: Int -- bomb out if size gets bigger than this
132 -> CoreExpr -- expression to look at
134 calcUnfoldingGuidance bOMB_OUT_SIZE expr
135 = case collect_val_bndrs expr of { (inline, val_binders, body) ->
137 n_val_binders = length val_binders
139 max_inline_size = n_val_binders+2
140 -- The idea is that if there is an INLINE pragma (inline is True)
141 -- and there's a big body, we give a size of n_val_binders+2. This
142 -- This is just enough to fail the no-size-increase test in callSiteInline,
143 -- so that INLINE things don't get inlined into entirely boring contexts,
147 case (sizeExpr (iUnbox bOMB_OUT_SIZE) val_binders body) of
150 | not inline -> UnfoldNever
151 -- A big function with an INLINE pragma must
152 -- have an UnfoldIfGoodArgs guidance
153 | otherwise -> UnfoldIfGoodArgs n_val_binders
154 (map (const 0) val_binders)
157 SizeIs size cased_args scrut_discount
160 (map discount_for val_binders)
162 (iBox scrut_discount)
164 boxed_size = iBox size
166 final_size | inline = boxed_size `min` max_inline_size
167 | otherwise = boxed_size
169 -- Sometimes an INLINE thing is smaller than n_val_binders+2.
170 -- A particular case in point is a constructor, which has size 1.
171 -- We want to inline this regardless, hence the `min`
173 discount_for b = foldlBag (\acc (b',n) -> if b==b' then acc+n else acc)
177 collect_val_bndrs e = go False [] e
178 -- We need to be a bit careful about how we collect the
179 -- value binders. In ptic, if we see
180 -- __inline_me (\x y -> e)
181 -- We want to say "2 value binders". Why? So that
182 -- we take account of information given for the arguments
184 go _ rev_vbs (Note InlineMe e) = go True rev_vbs e
185 go inline rev_vbs (Lam b e) | isId b = go inline (b:rev_vbs) e
186 | otherwise = go inline rev_vbs e
187 go inline rev_vbs e = (inline, reverse rev_vbs, e)
191 sizeExpr :: FastInt -- Bomb out if it gets bigger than this
192 -> [Id] -- Arguments; we're interested in which of these
197 sizeExpr bOMB_OUT_SIZE top_args expr
200 size_up (Type _) = sizeZero -- Types cost nothing
201 size_up (Var _) = sizeOne
203 size_up (Note InlineMe _) = sizeOne -- Inline notes make it look very small
204 -- This can be important. If you have an instance decl like this:
205 -- instance Foo a => Foo [a] where
206 -- {-# INLINE op1, op2 #-}
209 -- then we'll get a dfun which is a pair of two INLINE lambdas
211 size_up (Note _ body) = size_up body -- Other notes cost nothing
213 size_up (Cast e _) = size_up e
215 size_up (App fun (Type _)) = size_up fun
216 size_up (App fun arg) = size_up_app fun [arg]
218 size_up (Lit lit) = sizeN (litSize lit)
220 size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
221 | otherwise = size_up e
223 size_up (Let (NonRec binder rhs) body)
224 = nukeScrutDiscount (size_up rhs) `addSize`
225 size_up body `addSizeN`
226 (if isUnLiftedType (idType binder) then 0 else 1)
227 -- For the allocation
228 -- If the binder has an unlifted type there is no allocation
230 size_up (Let (Rec pairs) body)
231 = nukeScrutDiscount rhs_size `addSize`
232 size_up body `addSizeN`
233 length pairs -- For the allocation
235 rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
237 size_up (Case (Var v) _ _ alts)
238 | v `elem` top_args -- We are scrutinising an argument variable
240 {- I'm nuking this special case; BUT see the comment with case alternatives.
242 (a) It's too eager. We don't want to inline a wrapper into a
243 context with no benefit.
244 E.g. \ x. f (x+x) no point in inlining (+) here!
246 (b) It's ineffective. Once g's wrapper is inlined, its case-expressions
247 aren't scrutinising arguments any more
251 [alt] -> size_up_alt alt `addSize` SizeIs (_ILIT(0)) (unitBag (v, 1)) (_ILIT(0))
252 -- We want to make wrapper-style evaluation look cheap, so that
253 -- when we inline a wrapper it doesn't make call site (much) bigger
254 -- Otherwise we get nasty phase ordering stuff:
257 -- If we inline g's wrapper, f looks big, and doesn't get inlined
258 -- into h; if we inline f first, while it looks small, then g's
259 -- wrapper will get inlined later anyway. To avoid this nasty
260 -- ordering difference, we make (case a of (x,y) -> ...),
261 -- *where a is one of the arguments* look free.
265 alts_size (foldr addSize sizeOne alt_sizes) -- The 1 is for the scrutinee
266 (foldr1 maxSize alt_sizes)
268 -- Good to inline if an arg is scrutinised, because
269 -- that may eliminate allocation in the caller
270 -- And it eliminates the case itself
273 alt_sizes = map size_up_alt alts
275 -- alts_size tries to compute a good discount for
276 -- the case when we are scrutinising an argument variable
277 alts_size (SizeIs tot _tot_disc _tot_scrut) -- Size of all alternatives
278 (SizeIs max max_disc max_scrut) -- Size of biggest alternative
279 = SizeIs tot (unitBag (v, iBox (_ILIT(1) +# tot -# max)) `unionBags` max_disc) max_scrut
280 -- If the variable is known, we produce a discount that
281 -- will take us back to 'max', the size of rh largest alternative
282 -- The 1+ is a little discount for reduced allocation in the caller
283 alts_size tot_size _ = tot_size
285 size_up (Case e _ _ alts) = nukeScrutDiscount (size_up e) `addSize`
286 foldr (addSize . size_up_alt) sizeZero alts
287 -- We don't charge for the case itself
288 -- It's a strict thing, and the price of the call
289 -- is paid by scrut. Also consider
290 -- case f x of DEFAULT -> e
291 -- This is just ';'! Don't charge for it.
294 size_up_app (App fun arg) args
295 | isTypeArg arg = size_up_app fun args
296 | otherwise = size_up_app fun (arg:args)
297 size_up_app fun args = foldr (addSize . nukeScrutDiscount . size_up)
298 (size_up_fun fun args)
301 -- A function application with at least one value argument
302 -- so if the function is an argument give it an arg-discount
304 -- Also behave specially if the function is a build
306 -- Also if the function is a constant Id (constr or primop)
307 -- compute discounts specially
308 size_up_fun (Var fun) args
309 | fun `hasKey` buildIdKey = buildSize
310 | fun `hasKey` augmentIdKey = augmentSize
312 = case globalIdDetails fun of
313 DataConWorkId dc -> conSizeN dc (valArgCount args)
315 FCallId _ -> sizeN opt_UF_DearOp
316 PrimOpId op -> primOpSize op (valArgCount args)
317 -- foldr addSize (primOpSize op) (map arg_discount args)
318 -- At one time I tried giving an arg-discount if a primop
319 -- is applied to one of the function's arguments, but it's
320 -- not good. At the moment, any unlifted-type arg gets a
321 -- 'True' for 'yes I'm evald', so we collect the discount even
322 -- if we know nothing about it. And just having it in a primop
323 -- doesn't help at all if we don't know something more.
325 _ -> fun_discount fun `addSizeN`
326 (1 + length (filter (not . exprIsTrivial) args))
327 -- The 1+ is for the function itself
328 -- Add 1 for each non-trivial arg;
329 -- the allocation cost, as in let(rec)
330 -- Slight hack here: for constructors the args are almost always
331 -- trivial; and for primops they are almost always prim typed
332 -- We should really only count for non-prim-typed args in the
333 -- general case, but that seems too much like hard work
335 size_up_fun other _ = size_up other
338 size_up_alt (_con, _bndrs, rhs) = size_up rhs
339 -- Don't charge for args, so that wrappers look cheap
340 -- (See comments about wrappers with Case)
343 -- We want to record if we're case'ing, or applying, an argument
344 fun_discount v | v `elem` top_args = SizeIs (_ILIT(0)) (unitBag (v, opt_UF_FunAppDiscount)) (_ILIT(0))
345 fun_discount _ = sizeZero
348 -- These addSize things have to be here because
349 -- I don't want to give them bOMB_OUT_SIZE as an argument
351 addSizeN TooBig _ = TooBig
352 addSizeN (SizeIs n xs d) m = mkSizeIs bOMB_OUT_SIZE (n +# iUnbox m) xs d
354 addSize TooBig _ = TooBig
355 addSize _ TooBig = TooBig
356 addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
357 = mkSizeIs bOMB_OUT_SIZE (n1 +# n2) (xs `unionBags` ys) (d1 +# d2)
360 Code for manipulating sizes
363 data ExprSize = TooBig
364 | SizeIs FastInt -- Size found
365 (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
366 FastInt -- Size to subtract if result is scrutinised
367 -- by a case expression
369 -- subtract the discount before deciding whether to bale out. eg. we
370 -- want to inline a large constructor application into a selector:
371 -- tup = (a_1, ..., a_99)
372 -- x = case tup of ...
374 mkSizeIs :: FastInt -> FastInt -> Bag (Id, Int) -> FastInt -> ExprSize
375 mkSizeIs max n xs d | (n -# d) ># max = TooBig
376 | otherwise = SizeIs n xs d
378 maxSize :: ExprSize -> ExprSize -> ExprSize
379 maxSize TooBig _ = TooBig
380 maxSize _ TooBig = TooBig
381 maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
384 sizeZero, sizeOne :: ExprSize
385 sizeN :: Int -> ExprSize
386 conSizeN :: DataCon ->Int -> ExprSize
388 sizeZero = SizeIs (_ILIT(0)) emptyBag (_ILIT(0))
389 sizeOne = SizeIs (_ILIT(1)) emptyBag (_ILIT(0))
390 sizeN n = SizeIs (iUnbox n) emptyBag (_ILIT(0))
392 | isUnboxedTupleCon dc = SizeIs (_ILIT(0)) emptyBag (iUnbox n +# _ILIT(1))
393 | otherwise = SizeIs (_ILIT(1)) emptyBag (iUnbox n +# _ILIT(1))
394 -- Treat constructors as size 1; we are keen to expose them
395 -- (and we charge separately for their args). We can't treat
396 -- them as size zero, else we find that (iBox x) has size 1,
397 -- which is the same as a lone variable; and hence 'v' will
398 -- always be replaced by (iBox x), where v is bound to iBox x.
400 -- However, unboxed tuples count as size zero
401 -- I found occasions where we had
402 -- f x y z = case op# x y z of { s -> (# s, () #) }
403 -- and f wasn't getting inlined
405 primOpSize :: PrimOp -> Int -> ExprSize
407 | not (primOpIsDupable op) = sizeN opt_UF_DearOp
408 | not (primOpOutOfLine op) = sizeN (2 - n_args)
409 -- Be very keen to inline simple primops.
410 -- We give a discount of 1 for each arg so that (op# x y z) costs 2.
411 -- We can't make it cost 1, else we'll inline let v = (op# x y z)
412 -- at every use of v, which is excessive.
414 -- A good example is:
415 -- let x = +# p q in C {x}
416 -- Even though x get's an occurrence of 'many', its RHS looks cheap,
417 -- and there's a good chance it'll get inlined back into C's RHS. Urgh!
418 | otherwise = sizeOne
420 buildSize :: ExprSize
421 buildSize = SizeIs (_ILIT(-2)) emptyBag (_ILIT(4))
422 -- We really want to inline applications of build
423 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
424 -- Indeed, we should add a result_discount becuause build is
425 -- very like a constructor. We don't bother to check that the
426 -- build is saturated (it usually is). The "-2" discounts for the \c n,
427 -- The "4" is rather arbitrary.
429 augmentSize :: ExprSize
430 augmentSize = SizeIs (_ILIT(-2)) emptyBag (_ILIT(4))
431 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
432 -- e plus ys. The -2 accounts for the \cn
434 nukeScrutDiscount :: ExprSize -> ExprSize
435 nukeScrutDiscount (SizeIs n vs _) = SizeIs n vs (_ILIT(0))
436 nukeScrutDiscount TooBig = TooBig
438 -- When we return a lambda, give a discount if it's used (applied)
439 lamScrutDiscount :: ExprSize -> ExprSize
440 lamScrutDiscount (SizeIs n vs _) = case opt_UF_FunAppDiscount of { d -> SizeIs n vs (iUnbox d) }
441 lamScrutDiscount TooBig = TooBig
445 %************************************************************************
447 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
449 %************************************************************************
451 We have very limited information about an unfolding expression: (1)~so
452 many type arguments and so many value arguments expected---for our
453 purposes here, we assume we've got those. (2)~A ``size'' or ``cost,''
454 a single integer. (3)~An ``argument info'' vector. For this, what we
455 have at the moment is a Boolean per argument position that says, ``I
456 will look with great favour on an explicit constructor in this
457 position.'' (4)~The ``discount'' to subtract if the expression
458 is being scrutinised.
460 Assuming we have enough type- and value arguments (if not, we give up
461 immediately), then we see if the ``discounted size'' is below some
462 (semi-arbitrary) threshold. It works like this: for every argument
463 position where we're looking for a constructor AND WE HAVE ONE in our
464 hands, we get a (again, semi-arbitrary) discount [proportion to the
465 number of constructors in the type being scrutinized].
467 If we're in the context of a scrutinee ( \tr{(case <expr > of A .. -> ...;.. )})
468 and the expression in question will evaluate to a constructor, we use
469 the computed discount size *for the result only* rather than
470 computing the argument discounts. Since we know the result of
471 the expression is going to be taken apart, discounting its size
472 is more accurate (see @sizeExpr@ above for how this discount size
475 We use this one to avoid exporting inlinings that we ``couldn't possibly
476 use'' on the other side. Can be overridden w/ flaggery.
477 Just the same as smallEnoughToInline, except that it has no actual arguments.
480 couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
481 couldBeSmallEnoughToInline threshold rhs = case calcUnfoldingGuidance threshold rhs of
485 certainlyWillInline :: Unfolding -> Bool
486 -- Sees if the unfolding is pretty certain to inline
487 certainlyWillInline (CoreUnfolding _ _ _ is_cheap (UnfoldIfGoodArgs n_vals _ size _))
488 = is_cheap && size - (n_vals +1) <= opt_UF_UseThreshold
489 certainlyWillInline _
492 smallEnoughToInline :: Unfolding -> Bool
493 smallEnoughToInline (CoreUnfolding _ _ _ _ (UnfoldIfGoodArgs _ _ size _))
494 = size <= opt_UF_UseThreshold
495 smallEnoughToInline _
499 %************************************************************************
501 \subsection{callSiteInline}
503 %************************************************************************
505 This is the key function. It decides whether to inline a variable at a call site
507 callSiteInline is used at call sites, so it is a bit more generous.
508 It's a very important function that embodies lots of heuristics.
509 A non-WHNF can be inlined if it doesn't occur inside a lambda,
510 and occurs exactly once or
511 occurs once in each branch of a case and is small
513 If the thing is in WHNF, there's no danger of duplicating work,
514 so we can inline if it occurs once, or is small
516 NOTE: we don't want to inline top-level functions that always diverge.
517 It just makes the code bigger. Tt turns out that the convenient way to prevent
518 them inlining is to give them a NOINLINE pragma, which we do in
519 StrictAnal.addStrictnessInfoToTopId
522 callSiteInline :: DynFlags
523 -> Bool -- True <=> the Id can be inlined
525 -> Bool -- True if there are are no arguments at all (incl type args)
526 -> [Bool] -- One for each value arg; True if it is interesting
527 -> CallCtxt -- True <=> continuation is interesting
528 -> Maybe CoreExpr -- Unfolding, if any
531 data CallCtxt = BoringCtxt
533 | ArgCtxt Bool -- We're somewhere in the RHS of function with rules
534 -- => be keener to inline
535 Int -- We *are* the argument of a function with this arg discount
536 -- => be keener to inline
537 -- INVARIANT: ArgCtxt False 0 ==> BoringCtxt
539 | CaseCtxt -- We're the scrutinee of a case
540 -- that decomposes its scrutinee
542 instance Outputable CallCtxt where
543 ppr BoringCtxt = ptext (sLit "BoringCtxt")
544 ppr (ArgCtxt _ _) = ptext (sLit "ArgCtxt")
545 ppr CaseCtxt = ptext (sLit "CaseCtxt")
547 callSiteInline dflags active_inline id lone_variable arg_infos cont_info
548 = case idUnfolding id of {
549 NoUnfolding -> Nothing ;
550 OtherCon _ -> Nothing ;
552 CompulsoryUnfolding unf_template -> Just unf_template ;
553 -- CompulsoryUnfolding => there is no top-level binding
554 -- for these things, so we must inline it.
555 -- Only a couple of primop-like things have
556 -- compulsory unfoldings (see MkId.lhs).
557 -- We don't allow them to be inactive
559 CoreUnfolding unf_template is_top is_value is_cheap guidance ->
562 result | yes_or_no = Just unf_template
563 | otherwise = Nothing
565 n_val_args = length arg_infos
567 yes_or_no = active_inline && is_cheap && consider_safe
568 -- We consider even the once-in-one-branch
569 -- occurrences, because they won't all have been
570 -- caught by preInlineUnconditionally. In particular,
571 -- if the occurrence is once inside a lambda, and the
572 -- rhs is cheap but not a manifest lambda, then
573 -- pre-inline will not have inlined it for fear of
574 -- invalidating the occurrence info in the rhs.
577 -- consider_safe decides whether it's a good idea to
578 -- inline something, given that there's no
579 -- work-duplication issue (the caller checks that).
582 UnfoldIfGoodArgs n_vals_wanted arg_discounts size res_discount
583 | enough_args && size <= (n_vals_wanted + 1)
584 -- Inline unconditionally if there no size increase
585 -- Size of call is n_vals_wanted (+1 for the function)
589 -> some_benefit && small_enough && inline_enough_args
592 enough_args = n_val_args >= n_vals_wanted
594 not (dopt Opt_InlineIfEnoughArgs dflags) || enough_args
597 some_benefit = or arg_infos || really_interesting_cont
598 -- There must be something interesting
599 -- about some argument, or the result
600 -- context, to make it worth inlining
602 really_interesting_cont
603 | n_val_args < n_vals_wanted = False -- Too few args
604 | n_val_args == n_vals_wanted = interesting_saturated_call
605 | otherwise = True -- Extra args
606 -- really_interesting_cont tells if the result of the
607 -- call is in an interesting context.
609 interesting_saturated_call
611 BoringCtxt -> not is_top && n_vals_wanted > 0 -- Note [Nested functions]
612 CaseCtxt -> not lone_variable || not is_value -- Note [Lone variables]
613 ArgCtxt {} -> n_vals_wanted > 0
614 -- See Note [Inlining in ArgCtxt]
616 small_enough = (size - discount) <= opt_UF_UseThreshold
617 discount = computeDiscount n_vals_wanted arg_discounts
618 res_discount' arg_infos
619 res_discount' = case cont_info of
621 CaseCtxt -> res_discount
622 ArgCtxt _ _ -> 4 `min` res_discount
623 -- res_discount can be very large when a function returns
624 -- construtors; but we only want to invoke that large discount
625 -- when there's a case continuation.
626 -- Otherwise we, rather arbitrarily, threshold it. Yuk.
627 -- But we want to aovid inlining large functions that return
628 -- constructors into contexts that are simply "interesting"
631 if dopt Opt_D_dump_inlinings dflags then
632 pprTrace "Considering inlining"
633 (ppr id <+> vcat [text "active:" <+> ppr active_inline,
634 text "arg infos" <+> ppr arg_infos,
635 text "interesting continuation" <+> ppr cont_info,
636 text "is value:" <+> ppr is_value,
637 text "is cheap:" <+> ppr is_cheap,
638 text "guidance" <+> ppr guidance,
639 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO"])
646 Note [Nested functions]
647 ~~~~~~~~~~~~~~~~~~~~~~~
648 If a function has a nested defn we also record some-benefit, on the
649 grounds that we are often able to eliminate the binding, and hence the
650 allocation, for the function altogether; this is good for join points.
651 But this only makes sense for *functions*; inlining a constructor
652 doesn't help allocation unless the result is scrutinised. UNLESS the
653 constructor occurs just once, albeit possibly in multiple case
654 branches. Then inlining it doesn't increase allocation, but it does
655 increase the chance that the constructor won't be allocated at all in
656 the branches that don't use it.
658 Note [Inlining in ArgCtxt]
659 ~~~~~~~~~~~~~~~~~~~~~~~~~~
660 The condition (n_vals_wanted > 0) here is very important, because otherwise
661 we end up inlining top-level stuff into useless places; eg
664 This can make a very big difference: it adds 16% to nofib 'integer' allocs,
667 At one stage I replaced this condition by 'True' (leading to the above
668 slow-down). The motivation was test eyeball/inline1.hs; but that seems
671 Note [Lone variables]
672 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
673 The "lone-variable" case is important. I spent ages messing about
674 with unsatisfactory varaints, but this is nice. The idea is that if a
675 variable appears all alone
676 as an arg of lazy fn, or rhs Stop
677 as scrutinee of a case Select
678 as arg of a strict fn ArgOf
680 it is bound to a value
681 then we should not inline it (unless there is some other reason,
682 e.g. is is the sole occurrence). That is what is happening at
683 the use of 'lone_variable' in 'interesting_saturated_call'.
685 Why? At least in the case-scrutinee situation, turning
686 let x = (a,b) in case x of y -> ...
688 let x = (a,b) in case (a,b) of y -> ...
690 let x = (a,b) in let y = (a,b) in ...
691 is bad if the binding for x will remain.
693 Another example: I discovered that strings
694 were getting inlined straight back into applications of 'error'
695 because the latter is strict.
697 f = \x -> ...(error s)...
699 Fundamentally such contexts should not encourage inlining because the
700 context can ``see'' the unfolding of the variable (e.g. case or a
701 RULE) so there's no gain. If the thing is bound to a value.
706 foo = _inline_ (\n. [n])
707 bar = _inline_ (foo 20)
708 baz = \n. case bar of { (m:_) -> m + n }
709 Here we really want to inline 'bar' so that we can inline 'foo'
710 and the whole thing unravels as it should obviously do. This is
711 important: in the NDP project, 'bar' generates a closure data
712 structure rather than a list.
714 * Even a type application or coercion isn't a lone variable.
716 case $fMonadST @ RealWorld of { :DMonad a b c -> c }
717 We had better inline that sucker! The case won't see through it.
719 For now, I'm treating treating a variable applied to types
720 in a *lazy* context "lone". The motivating example was
723 There's no advantage in inlining f here, and perhaps
724 a significant disadvantage. Hence some_val_args in the Stop case
727 computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Int
728 computeDiscount n_vals_wanted arg_discounts result_discount arg_infos
729 -- We multiple the raw discounts (args_discount and result_discount)
730 -- ty opt_UnfoldingKeenessFactor because the former have to do with
731 -- *size* whereas the discounts imply that there's some extra
732 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
735 -- we also discount 1 for each argument passed, because these will
736 -- reduce with the lambdas in the function (we count 1 for a lambda
738 = 1 + -- Discount of 1 because the result replaces the call
739 -- so we count 1 for the function itself
740 length (take n_vals_wanted arg_infos) +
741 -- Discount of 1 for each arg supplied, because the
742 -- result replaces the call
743 round (opt_UF_KeenessFactor *
744 fromIntegral (arg_discount + result_discount))
746 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
748 mk_arg_discount discount is_evald | is_evald = discount
752 %************************************************************************
754 The Very Simple Optimiser
756 %************************************************************************
760 simpleOptExpr :: Subst -> CoreExpr -> CoreExpr
761 -- Return an occur-analysed and slightly optimised expression
762 -- The optimisation is very straightforward: just
763 -- inline non-recursive bindings that are used only once,
764 -- or wheere the RHS is trivial
766 simpleOptExpr subst expr
767 = go subst (occurAnalyseExpr expr)
769 go subst (Var v) = lookupIdSubst subst v
770 go subst (App e1 e2) = App (go subst e1) (go subst e2)
771 go subst (Type ty) = Type (substTy subst ty)
772 go _ (Lit lit) = Lit lit
773 go subst (Note note e) = Note note (go subst e)
774 go subst (Cast e co) = Cast (go subst e) (substTy subst co)
775 go subst (Let bind body) = go_bind subst bind body
776 go subst (Lam bndr body) = Lam bndr' (go subst' body)
778 (subst', bndr') = substBndr subst bndr
780 go subst (Case e b ty as) = Case (go subst e) b'
782 (map (go_alt subst') as)
784 (subst', b') = substBndr subst b
787 ----------------------
788 go_alt subst (con, bndrs, rhs) = (con, bndrs', go subst' rhs)
790 (subst', bndrs') = substBndrs subst bndrs
792 ----------------------
793 go_bind subst (Rec prs) body = Let (Rec (bndrs' `zip` rhss'))
796 (bndrs, rhss) = unzip prs
797 (subst', bndrs') = substRecBndrs subst bndrs
798 rhss' = map (go subst') rhss
800 go_bind subst (NonRec b r) body = go_nonrec subst b (go subst r) body
802 ----------------------
803 go_nonrec subst b (Type ty') body
804 | isTyVar b = go (extendTvSubst subst b ty') body
805 -- let a::* = TYPE ty in <body>
806 go_nonrec subst b r' body
807 | isId b -- let x = e in <body>
808 , exprIsTrivial r' || safe_to_inline (idOccInfo b)
809 = go (extendIdSubst subst b r') body
810 go_nonrec subst b r' body
811 = Let (NonRec b' r') (go subst' body)
813 (subst', b') = substBndr subst b
815 ----------------------
816 -- Unconditionally safe to inline
817 safe_to_inline :: OccInfo -> Bool
818 safe_to_inline IAmDead = True
819 safe_to_inline (OneOcc in_lam one_br _) = not in_lam && one_br
820 safe_to_inline (IAmALoopBreaker {}) = False
821 safe_to_inline NoOccInfo = False