2 % (c) The AQUA Project, Glasgow University, 1994-1998
4 \section[CoreUnfold]{Core-syntax unfoldings}
6 Unfoldings (which can travel across module boundaries) are in Core
7 syntax (namely @CoreExpr@s).
9 The type @Unfolding@ sits ``above'' simply-Core-expressions
10 unfoldings, capturing ``higher-level'' things we know about a binding,
11 usually things that the simplifier found out (e.g., ``it's a
12 literal''). In the corner of a @CoreUnfolding@ unfolding, you will
13 find, unsurprisingly, a Core expression.
17 Unfolding, UnfoldingGuidance, -- Abstract types
19 noUnfolding, mkTopUnfolding, mkUnfolding, mkCompulsoryUnfolding, seqUnfolding,
20 evaldUnfolding, mkOtherCon, otherCons,
21 unfoldingTemplate, maybeUnfoldingTemplate,
22 isEvaldUnfolding, isValueUnfolding, isCheapUnfolding, isCompulsoryUnfolding,
23 hasUnfolding, hasSomeUnfolding, neverUnfold,
25 couldBeSmallEnoughToInline,
26 certainlyWillInline, smallEnoughToInline,
31 #include "HsVersions.h"
33 import StaticFlags ( opt_UF_CreationThreshold, opt_UF_UseThreshold,
34 opt_UF_FunAppDiscount, opt_UF_KeenessFactor,
37 import DynFlags ( DynFlags, DynFlag(..), dopt )
39 import PprCore () -- Instances
40 import OccurAnal ( occurAnalyseExpr )
41 import CoreUtils ( exprIsHNF, exprIsCheap, exprIsTrivial )
42 import Id ( Id, idType, isId,
43 idUnfolding, globalIdDetails
45 import DataCon ( isUnboxedTupleCon )
46 import Literal ( litSize )
47 import PrimOp ( primOpIsDupable, primOpOutOfLine )
48 import IdInfo ( OccInfo(..), GlobalIdDetails(..) )
49 import Type ( isUnLiftedType )
50 import PrelNames ( hasKey, buildIdKey, augmentIdKey )
55 #if __GLASGOW_HASKELL__ >= 404
56 import GLAEXTS ( Int# )
61 %************************************************************************
63 \subsection{Making unfoldings}
65 %************************************************************************
68 mkTopUnfolding expr = mkUnfolding True {- Top level -} expr
70 mkUnfolding top_lvl expr
71 = CoreUnfolding (occurAnalyseExpr expr)
78 -- OK to inline inside a lambda
80 (calcUnfoldingGuidance opt_UF_CreationThreshold expr)
81 -- Sometimes during simplification, there's a large let-bound thing
82 -- which has been substituted, and so is now dead; so 'expr' contains
83 -- two copies of the thing while the occurrence-analysed expression doesn't
84 -- Nevertheless, we don't occ-analyse before computing the size because the
85 -- size computation bales out after a while, whereas occurrence analysis does not.
87 -- This can occasionally mean that the guidance is very pessimistic;
88 -- it gets fixed up next round
90 instance Outputable Unfolding where
91 ppr NoUnfolding = ptext SLIT("No unfolding")
92 ppr (OtherCon cs) = ptext SLIT("OtherCon") <+> ppr cs
93 ppr (CompulsoryUnfolding e) = ptext SLIT("Compulsory") <+> ppr e
94 ppr (CoreUnfolding e top hnf cheap g)
95 = ptext SLIT("Unf") <+> sep [ppr top <+> ppr hnf <+> ppr cheap <+> ppr g,
98 mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
99 = CompulsoryUnfolding (occurAnalyseExpr expr)
103 %************************************************************************
105 \subsection{The UnfoldingGuidance type}
107 %************************************************************************
110 instance Outputable UnfoldingGuidance where
111 ppr UnfoldNever = ptext SLIT("NEVER")
112 ppr (UnfoldIfGoodArgs v cs size discount)
113 = hsep [ ptext SLIT("IF_ARGS"), int v,
114 brackets (hsep (map int cs)),
121 calcUnfoldingGuidance
122 :: Int -- bomb out if size gets bigger than this
123 -> CoreExpr -- expression to look at
125 calcUnfoldingGuidance bOMB_OUT_SIZE expr
126 = case collect_val_bndrs expr of { (inline, val_binders, body) ->
128 n_val_binders = length val_binders
130 max_inline_size = n_val_binders+2
131 -- The idea is that if there is an INLINE pragma (inline is True)
132 -- and there's a big body, we give a size of n_val_binders+2. This
133 -- This is just enough to fail the no-size-increase test in callSiteInline,
134 -- so that INLINE things don't get inlined into entirely boring contexts,
138 case (sizeExpr (iUnbox bOMB_OUT_SIZE) val_binders body) of
141 | not inline -> UnfoldNever
142 -- A big function with an INLINE pragma must
143 -- have an UnfoldIfGoodArgs guidance
144 | otherwise -> UnfoldIfGoodArgs n_val_binders
145 (map (const 0) val_binders)
148 SizeIs size cased_args scrut_discount
151 (map discount_for val_binders)
153 (iBox scrut_discount)
155 boxed_size = iBox size
157 final_size | inline = boxed_size `min` max_inline_size
158 | otherwise = boxed_size
160 -- Sometimes an INLINE thing is smaller than n_val_binders+2.
161 -- A particular case in point is a constructor, which has size 1.
162 -- We want to inline this regardless, hence the `min`
164 discount_for b = foldlBag (\acc (b',n) -> if b==b' then acc+n else acc)
168 collect_val_bndrs e = go False [] e
169 -- We need to be a bit careful about how we collect the
170 -- value binders. In ptic, if we see
171 -- __inline_me (\x y -> e)
172 -- We want to say "2 value binders". Why? So that
173 -- we take account of information given for the arguments
175 go inline rev_vbs (Note InlineMe e) = go True rev_vbs e
176 go inline rev_vbs (Lam b e) | isId b = go inline (b:rev_vbs) e
177 | otherwise = go inline rev_vbs e
178 go inline rev_vbs e = (inline, reverse rev_vbs, e)
182 sizeExpr :: Int# -- Bomb out if it gets bigger than this
183 -> [Id] -- Arguments; we're interested in which of these
188 sizeExpr bOMB_OUT_SIZE top_args expr
191 size_up (Type t) = sizeZero -- Types cost nothing
192 size_up (Var v) = sizeOne
194 size_up (Note InlineMe body) = sizeOne -- Inline notes make it look very small
195 -- This can be important. If you have an instance decl like this:
196 -- instance Foo a => Foo [a] where
197 -- {-# INLINE op1, op2 #-}
200 -- then we'll get a dfun which is a pair of two INLINE lambdas
202 size_up (Note _ body) = size_up body -- Other notes cost nothing
204 size_up (App fun (Type t)) = 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 0# (unitBag (v, 1)) 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
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 fc -> 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 other -> 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 args = 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 0# (unitBag (v, opt_UF_FunAppDiscount)) 0#
335 fun_discount other = 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 max n xs d | (n -# d) ># max = TooBig
365 | otherwise = SizeIs n xs d
367 maxSize TooBig _ = TooBig
368 maxSize _ TooBig = TooBig
369 maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
372 sizeZero = SizeIs (_ILIT 0) emptyBag (_ILIT 0)
373 sizeOne = SizeIs (_ILIT 1) emptyBag (_ILIT 0)
374 sizeN n = SizeIs (iUnbox n) emptyBag (_ILIT 0)
376 | isUnboxedTupleCon dc = SizeIs (_ILIT 0) emptyBag (iUnbox n +# _ILIT 1)
377 | otherwise = SizeIs (_ILIT 1) emptyBag (iUnbox n +# _ILIT 1)
378 -- Treat constructors as size 1; we are keen to expose them
379 -- (and we charge separately for their args). We can't treat
380 -- them as size zero, else we find that (iBox x) has size 1,
381 -- which is the same as a lone variable; and hence 'v' will
382 -- always be replaced by (iBox x), where v is bound to iBox x.
384 -- However, unboxed tuples count as size zero
385 -- I found occasions where we had
386 -- f x y z = case op# x y z of { s -> (# s, () #) }
387 -- and f wasn't getting inlined
390 | not (primOpIsDupable op) = sizeN opt_UF_DearOp
391 | not (primOpOutOfLine op) = sizeN (2 - n_args)
392 -- Be very keen to inline simple primops.
393 -- We give a discount of 1 for each arg so that (op# x y z) costs 2.
394 -- We can't make it cost 1, else we'll inline let v = (op# x y z)
395 -- at every use of v, which is excessive.
397 -- A good example is:
398 -- let x = +# p q in C {x}
399 -- Even though x get's an occurrence of 'many', its RHS looks cheap,
400 -- and there's a good chance it'll get inlined back into C's RHS. Urgh!
401 | otherwise = sizeOne
403 buildSize = SizeIs (-2#) emptyBag 4#
404 -- We really want to inline applications of build
405 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
406 -- Indeed, we should add a result_discount becuause build is
407 -- very like a constructor. We don't bother to check that the
408 -- build is saturated (it usually is). The "-2" discounts for the \c n,
409 -- The "4" is rather arbitrary.
411 augmentSize = SizeIs (-2#) emptyBag 4#
412 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
413 -- e plus ys. The -2 accounts for the \cn
415 nukeScrutDiscount (SizeIs n vs d) = SizeIs n vs 0#
416 nukeScrutDiscount TooBig = TooBig
418 -- When we return a lambda, give a discount if it's used (applied)
419 lamScrutDiscount (SizeIs n vs d) = case opt_UF_FunAppDiscount of { d -> SizeIs n vs (iUnbox d) }
420 lamScrutDiscount TooBig = TooBig
424 %************************************************************************
426 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
428 %************************************************************************
430 We have very limited information about an unfolding expression: (1)~so
431 many type arguments and so many value arguments expected---for our
432 purposes here, we assume we've got those. (2)~A ``size'' or ``cost,''
433 a single integer. (3)~An ``argument info'' vector. For this, what we
434 have at the moment is a Boolean per argument position that says, ``I
435 will look with great favour on an explicit constructor in this
436 position.'' (4)~The ``discount'' to subtract if the expression
437 is being scrutinised.
439 Assuming we have enough type- and value arguments (if not, we give up
440 immediately), then we see if the ``discounted size'' is below some
441 (semi-arbitrary) threshold. It works like this: for every argument
442 position where we're looking for a constructor AND WE HAVE ONE in our
443 hands, we get a (again, semi-arbitrary) discount [proportion to the
444 number of constructors in the type being scrutinized].
446 If we're in the context of a scrutinee ( \tr{(case <expr > of A .. -> ...;.. )})
447 and the expression in question will evaluate to a constructor, we use
448 the computed discount size *for the result only* rather than
449 computing the argument discounts. Since we know the result of
450 the expression is going to be taken apart, discounting its size
451 is more accurate (see @sizeExpr@ above for how this discount size
454 We use this one to avoid exporting inlinings that we ``couldn't possibly
455 use'' on the other side. Can be overridden w/ flaggery.
456 Just the same as smallEnoughToInline, except that it has no actual arguments.
459 couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
460 couldBeSmallEnoughToInline threshold rhs = case calcUnfoldingGuidance threshold rhs of
464 certainlyWillInline :: Unfolding -> Bool
465 -- Sees if the unfolding is pretty certain to inline
466 certainlyWillInline (CoreUnfolding _ _ _ is_cheap (UnfoldIfGoodArgs n_vals _ size _))
467 = is_cheap && size - (n_vals +1) <= opt_UF_UseThreshold
468 certainlyWillInline other
471 smallEnoughToInline :: Unfolding -> Bool
472 smallEnoughToInline (CoreUnfolding _ _ _ _ (UnfoldIfGoodArgs _ _ size _))
473 = size <= opt_UF_UseThreshold
474 smallEnoughToInline other
478 %************************************************************************
480 \subsection{callSiteInline}
482 %************************************************************************
484 This is the key function. It decides whether to inline a variable at a call site
486 callSiteInline is used at call sites, so it is a bit more generous.
487 It's a very important function that embodies lots of heuristics.
488 A non-WHNF can be inlined if it doesn't occur inside a lambda,
489 and occurs exactly once or
490 occurs once in each branch of a case and is small
492 If the thing is in WHNF, there's no danger of duplicating work,
493 so we can inline if it occurs once, or is small
495 NOTE: we don't want to inline top-level functions that always diverge.
496 It just makes the code bigger. Tt turns out that the convenient way to prevent
497 them inlining is to give them a NOINLINE pragma, which we do in
498 StrictAnal.addStrictnessInfoToTopId
501 callSiteInline :: DynFlags
502 -> Bool -- True <=> the Id can be inlined
505 -> [Bool] -- One for each value arg; True if it is interesting
506 -> Bool -- True <=> continuation is interesting
507 -> Maybe CoreExpr -- Unfolding, if any
510 callSiteInline dflags active_inline occ id arg_infos interesting_cont
511 = case idUnfolding id of {
512 NoUnfolding -> Nothing ;
513 OtherCon cs -> Nothing ;
515 CompulsoryUnfolding unf_template -> Just unf_template ;
516 -- CompulsoryUnfolding => there is no top-level binding
517 -- for these things, so we must inline it.
518 -- Only a couple of primop-like things have
519 -- compulsory unfoldings (see MkId.lhs).
520 -- We don't allow them to be inactive
522 CoreUnfolding unf_template is_top is_value is_cheap guidance ->
525 result | yes_or_no = Just unf_template
526 | otherwise = Nothing
528 n_val_args = length arg_infos
531 | not active_inline = False
532 | otherwise = case occ of
533 IAmDead -> pprTrace "callSiteInline: dead" (ppr id) False
534 IAmALoopBreaker -> False
535 --OneOcc in_lam _ _ -> (not in_lam || is_cheap) && consider_safe True
536 other -> is_cheap && consider_safe False
537 -- we consider even the once-in-one-branch
538 -- occurrences, because they won't all have been
539 -- caught by preInlineUnconditionally. In particular,
540 -- if the occurrence is once inside a lambda, and the
541 -- rhs is cheap but not a manifest lambda, then
542 -- pre-inline will not have inlined it for fear of
543 -- invalidating the occurrence info in the rhs.
546 -- consider_safe decides whether it's a good idea to
547 -- inline something, given that there's no
548 -- work-duplication issue (the caller checks that).
551 UnfoldIfGoodArgs n_vals_wanted arg_discounts size res_discount
553 | enough_args && size <= (n_vals_wanted + 1)
554 -- Inline unconditionally if there no size increase
555 -- Size of call is n_vals_wanted (+1 for the function)
559 -> some_benefit && small_enough
562 some_benefit = or arg_infos || really_interesting_cont ||
563 (not is_top && ({- once || -} (n_vals_wanted > 0 && enough_args)))
564 -- [was (once && not in_lam)]
565 -- If it occurs more than once, there must be
566 -- something interesting about some argument, or the
567 -- result context, to make it worth inlining
569 -- If a function has a nested defn we also record
570 -- some-benefit, on the grounds that we are often able
571 -- to eliminate the binding, and hence the allocation,
572 -- for the function altogether; this is good for join
573 -- points. But this only makes sense for *functions*;
574 -- inlining a constructor doesn't help allocation
575 -- unless the result is scrutinised. UNLESS the
576 -- constructor occurs just once, albeit possibly in
577 -- multiple case branches. Then inlining it doesn't
578 -- increase allocation, but it does increase the
579 -- chance that the constructor won't be allocated at
580 -- all in the branches that don't use it.
582 enough_args = n_val_args >= n_vals_wanted
583 really_interesting_cont | n_val_args < n_vals_wanted = False -- Too few args
584 | n_val_args == n_vals_wanted = interesting_cont
585 | otherwise = True -- Extra args
586 -- really_interesting_cont tells if the result of the
587 -- call is in an interesting context.
589 small_enough = (size - discount) <= opt_UF_UseThreshold
590 discount = computeDiscount n_vals_wanted arg_discounts res_discount
591 arg_infos really_interesting_cont
594 if dopt Opt_D_dump_inlinings dflags then
595 pprTrace "Considering inlining"
596 (ppr id <+> vcat [text "active:" <+> ppr active_inline,
597 text "occ info:" <+> ppr occ,
598 text "arg infos" <+> ppr arg_infos,
599 text "interesting continuation" <+> ppr interesting_cont,
600 text "is value:" <+> ppr is_value,
601 text "is cheap:" <+> ppr is_cheap,
602 text "guidance" <+> ppr guidance,
603 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO"])
609 computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Bool -> Int
610 computeDiscount n_vals_wanted arg_discounts res_discount arg_infos result_used
611 -- We multiple the raw discounts (args_discount and result_discount)
612 -- ty opt_UnfoldingKeenessFactor because the former have to do with
613 -- *size* whereas the discounts imply that there's some extra
614 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
617 -- we also discount 1 for each argument passed, because these will
618 -- reduce with the lambdas in the function (we count 1 for a lambda
620 = 1 + -- Discount of 1 because the result replaces the call
621 -- so we count 1 for the function itself
622 length (take n_vals_wanted arg_infos) +
623 -- Discount of 1 for each arg supplied, because the
624 -- result replaces the call
625 round (opt_UF_KeenessFactor *
626 fromIntegral (arg_discount + result_discount))
628 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
630 mk_arg_discount discount is_evald | is_evald = discount
633 -- Don't give a result discount unless there are enough args
634 result_discount | result_used = res_discount -- Over-applied, or case scrut