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,
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 ( pprCoreExpr )
40 import OccurAnal ( occurAnalyseExpr )
41 import CoreUtils ( exprIsValue, 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 )
56 #if __GLASGOW_HASKELL__ >= 404
57 import GLAEXTS ( Int# )
62 %************************************************************************
64 \subsection{Making unfoldings}
66 %************************************************************************
69 mkTopUnfolding expr = mkUnfolding True {- Top level -} expr
71 mkUnfolding top_lvl expr
72 = CoreUnfolding (occurAnalyseExpr expr)
79 -- OK to inline inside a lambda
81 (calcUnfoldingGuidance opt_UF_CreationThreshold expr)
82 -- Sometimes during simplification, there's a large let-bound thing
83 -- which has been substituted, and so is now dead; so 'expr' contains
84 -- two copies of the thing while the occurrence-analysed expression doesn't
85 -- Nevertheless, we don't occ-analyse before computing the size because the
86 -- size computation bales out after a while, whereas occurrence analysis does not.
88 -- This can occasionally mean that the guidance is very pessimistic;
89 -- it gets fixed up next round
91 mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
92 = CompulsoryUnfolding (occurAnalyseExpr expr)
96 %************************************************************************
98 \subsection{The UnfoldingGuidance type}
100 %************************************************************************
103 instance Outputable UnfoldingGuidance where
104 ppr UnfoldNever = ptext SLIT("NEVER")
105 ppr (UnfoldIfGoodArgs v cs size discount)
106 = hsep [ ptext SLIT("IF_ARGS"), int v,
107 brackets (hsep (map int cs)),
114 calcUnfoldingGuidance
115 :: Int -- bomb out if size gets bigger than this
116 -> CoreExpr -- expression to look at
118 calcUnfoldingGuidance bOMB_OUT_SIZE expr
119 = case collect_val_bndrs expr of { (inline, val_binders, body) ->
121 n_val_binders = length val_binders
123 max_inline_size = n_val_binders+2
124 -- The idea is that if there is an INLINE pragma (inline is True)
125 -- and there's a big body, we give a size of n_val_binders+2. This
126 -- This is just enough to fail the no-size-increase test in callSiteInline,
127 -- so that INLINE things don't get inlined into entirely boring contexts,
131 case (sizeExpr (iUnbox bOMB_OUT_SIZE) val_binders body) of
134 | not inline -> UnfoldNever
135 -- A big function with an INLINE pragma must
136 -- have an UnfoldIfGoodArgs guidance
137 | otherwise -> UnfoldIfGoodArgs n_val_binders
138 (map (const 0) val_binders)
141 SizeIs size cased_args scrut_discount
144 (map discount_for val_binders)
146 (iBox scrut_discount)
148 boxed_size = iBox size
150 final_size | inline = boxed_size `min` max_inline_size
151 | otherwise = boxed_size
153 -- Sometimes an INLINE thing is smaller than n_val_binders+2.
154 -- A particular case in point is a constructor, which has size 1.
155 -- We want to inline this regardless, hence the `min`
157 discount_for b = foldlBag (\acc (b',n) -> if b==b' then acc+n else acc)
161 collect_val_bndrs e = go False [] e
162 -- We need to be a bit careful about how we collect the
163 -- value binders. In ptic, if we see
164 -- __inline_me (\x y -> e)
165 -- We want to say "2 value binders". Why? So that
166 -- we take account of information given for the arguments
168 go inline rev_vbs (Note InlineMe e) = go True rev_vbs e
169 go inline rev_vbs (Lam b e) | isId b = go inline (b:rev_vbs) e
170 | otherwise = go inline rev_vbs e
171 go inline rev_vbs e = (inline, reverse rev_vbs, e)
175 sizeExpr :: Int# -- Bomb out if it gets bigger than this
176 -> [Id] -- Arguments; we're interested in which of these
181 sizeExpr bOMB_OUT_SIZE top_args expr
184 size_up (Type t) = sizeZero -- Types cost nothing
185 size_up (Var v) = sizeOne
187 size_up (Note InlineMe body) = sizeOne -- Inline notes make it look very small
188 -- This can be important. If you have an instance decl like this:
189 -- instance Foo a => Foo [a] where
190 -- {-# INLINE op1, op2 #-}
193 -- then we'll get a dfun which is a pair of two INLINE lambdas
195 size_up (Note _ body) = size_up body -- Other notes cost nothing
197 size_up (App fun (Type t)) = size_up fun
198 size_up (App fun arg) = size_up_app fun [arg]
200 size_up (Lit lit) = sizeN (litSize lit)
202 size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
203 | otherwise = size_up e
205 size_up (Let (NonRec binder rhs) body)
206 = nukeScrutDiscount (size_up rhs) `addSize`
207 size_up body `addSizeN`
208 (if isUnLiftedType (idType binder) then 0 else 1)
209 -- For the allocation
210 -- If the binder has an unlifted type there is no allocation
212 size_up (Let (Rec pairs) body)
213 = nukeScrutDiscount rhs_size `addSize`
214 size_up body `addSizeN`
215 length pairs -- For the allocation
217 rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
219 size_up (Case (Var v) _ _ alts)
220 | v `elem` top_args -- We are scrutinising an argument variable
222 {- I'm nuking this special case; BUT see the comment with case alternatives.
224 (a) It's too eager. We don't want to inline a wrapper into a
225 context with no benefit.
226 E.g. \ x. f (x+x) no point in inlining (+) here!
228 (b) It's ineffective. Once g's wrapper is inlined, its case-expressions
229 aren't scrutinising arguments any more
233 [alt] -> size_up_alt alt `addSize` SizeIs 0# (unitBag (v, 1)) 0#
234 -- We want to make wrapper-style evaluation look cheap, so that
235 -- when we inline a wrapper it doesn't make call site (much) bigger
236 -- Otherwise we get nasty phase ordering stuff:
239 -- If we inline g's wrapper, f looks big, and doesn't get inlined
240 -- into h; if we inline f first, while it looks small, then g's
241 -- wrapper will get inlined later anyway. To avoid this nasty
242 -- ordering difference, we make (case a of (x,y) -> ...),
243 -- *where a is one of the arguments* look free.
247 alts_size (foldr addSize sizeOne alt_sizes) -- The 1 is for the scrutinee
248 (foldr1 maxSize alt_sizes)
250 -- Good to inline if an arg is scrutinised, because
251 -- that may eliminate allocation in the caller
252 -- And it eliminates the case itself
255 alt_sizes = map size_up_alt alts
257 -- alts_size tries to compute a good discount for
258 -- the case when we are scrutinising an argument variable
259 alts_size (SizeIs tot tot_disc tot_scrut) -- Size of all alternatives
260 (SizeIs max max_disc max_scrut) -- Size of biggest alternative
261 = SizeIs tot (unitBag (v, iBox (_ILIT 1 +# tot -# max)) `unionBags` max_disc) max_scrut
262 -- If the variable is known, we produce a discount that
263 -- will take us back to 'max', the size of rh largest alternative
264 -- The 1+ is a little discount for reduced allocation in the caller
265 alts_size tot_size _ = tot_size
268 size_up (Case e _ _ alts) = nukeScrutDiscount (size_up e) `addSize`
269 foldr (addSize . size_up_alt) sizeZero alts
270 -- We don't charge for the case itself
271 -- It's a strict thing, and the price of the call
272 -- is paid by scrut. Also consider
273 -- case f x of DEFAULT -> e
274 -- This is just ';'! Don't charge for it.
277 size_up_app (App fun arg) args
278 | isTypeArg arg = size_up_app fun args
279 | otherwise = size_up_app fun (arg:args)
280 size_up_app fun args = foldr (addSize . nukeScrutDiscount . size_up)
281 (size_up_fun fun args)
284 -- A function application with at least one value argument
285 -- so if the function is an argument give it an arg-discount
287 -- Also behave specially if the function is a build
289 -- Also if the function is a constant Id (constr or primop)
290 -- compute discounts specially
291 size_up_fun (Var fun) args
292 | fun `hasKey` buildIdKey = buildSize
293 | fun `hasKey` augmentIdKey = augmentSize
295 = case globalIdDetails fun of
296 DataConWorkId dc -> conSizeN dc (valArgCount args)
298 FCallId fc -> sizeN opt_UF_DearOp
299 PrimOpId op -> primOpSize op (valArgCount args)
300 -- foldr addSize (primOpSize op) (map arg_discount args)
301 -- At one time I tried giving an arg-discount if a primop
302 -- is applied to one of the function's arguments, but it's
303 -- not good. At the moment, any unlifted-type arg gets a
304 -- 'True' for 'yes I'm evald', so we collect the discount even
305 -- if we know nothing about it. And just having it in a primop
306 -- doesn't help at all if we don't know something more.
308 other -> fun_discount fun `addSizeN`
309 (1 + length (filter (not . exprIsTrivial) args))
310 -- The 1+ is for the function itself
311 -- Add 1 for each non-trivial arg;
312 -- the allocation cost, as in let(rec)
313 -- Slight hack here: for constructors the args are almost always
314 -- trivial; and for primops they are almost always prim typed
315 -- We should really only count for non-prim-typed args in the
316 -- general case, but that seems too much like hard work
318 size_up_fun other args = size_up other
321 size_up_alt (con, bndrs, rhs) = size_up rhs
322 -- Don't charge for args, so that wrappers look cheap
323 -- (See comments about wrappers with Case)
326 -- We want to record if we're case'ing, or applying, an argument
327 fun_discount v | v `elem` top_args = SizeIs 0# (unitBag (v, opt_UF_FunAppDiscount)) 0#
328 fun_discount other = sizeZero
331 -- These addSize things have to be here because
332 -- I don't want to give them bOMB_OUT_SIZE as an argument
334 addSizeN TooBig _ = TooBig
335 addSizeN (SizeIs n xs d) m = mkSizeIs bOMB_OUT_SIZE (n +# iUnbox m) xs d
337 addSize TooBig _ = TooBig
338 addSize _ TooBig = TooBig
339 addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
340 = mkSizeIs bOMB_OUT_SIZE (n1 +# n2) (xs `unionBags` ys) (d1 +# d2)
343 Code for manipulating sizes
346 data ExprSize = TooBig
347 | SizeIs FastInt -- Size found
348 (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
349 FastInt -- Size to subtract if result is scrutinised
350 -- by a case expression
352 -- subtract the discount before deciding whether to bale out. eg. we
353 -- want to inline a large constructor application into a selector:
354 -- tup = (a_1, ..., a_99)
355 -- x = case tup of ...
357 mkSizeIs max n xs d | (n -# d) ># max = TooBig
358 | otherwise = SizeIs n xs d
360 maxSize TooBig _ = TooBig
361 maxSize _ TooBig = TooBig
362 maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
365 sizeZero = SizeIs (_ILIT 0) emptyBag (_ILIT 0)
366 sizeOne = SizeIs (_ILIT 1) emptyBag (_ILIT 0)
367 sizeN n = SizeIs (iUnbox n) emptyBag (_ILIT 0)
369 | isUnboxedTupleCon dc = SizeIs (_ILIT 0) emptyBag (iUnbox n +# _ILIT 1)
370 | otherwise = SizeIs (_ILIT 1) emptyBag (iUnbox n +# _ILIT 1)
371 -- Treat constructors as size 1; we are keen to expose them
372 -- (and we charge separately for their args). We can't treat
373 -- them as size zero, else we find that (iBox x) has size 1,
374 -- which is the same as a lone variable; and hence 'v' will
375 -- always be replaced by (iBox x), where v is bound to iBox x.
377 -- However, unboxed tuples count as size zero
378 -- I found occasions where we had
379 -- f x y z = case op# x y z of { s -> (# s, () #) }
380 -- and f wasn't getting inlined
383 | not (primOpIsDupable op) = sizeN opt_UF_DearOp
384 | not (primOpOutOfLine op) = sizeN (2 - n_args)
385 -- Be very keen to inline simple primops.
386 -- We give a discount of 1 for each arg so that (op# x y z) costs 2.
387 -- We can't make it cost 1, else we'll inline let v = (op# x y z)
388 -- at every use of v, which is excessive.
390 -- A good example is:
391 -- let x = +# p q in C {x}
392 -- Even though x get's an occurrence of 'many', its RHS looks cheap,
393 -- and there's a good chance it'll get inlined back into C's RHS. Urgh!
394 | otherwise = sizeOne
396 buildSize = SizeIs (-2#) emptyBag 4#
397 -- We really want to inline applications of build
398 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
399 -- Indeed, we should add a result_discount becuause build is
400 -- very like a constructor. We don't bother to check that the
401 -- build is saturated (it usually is). The "-2" discounts for the \c n,
402 -- The "4" is rather arbitrary.
404 augmentSize = SizeIs (-2#) emptyBag 4#
405 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
406 -- e plus ys. The -2 accounts for the \cn
408 nukeScrutDiscount (SizeIs n vs d) = SizeIs n vs 0#
409 nukeScrutDiscount TooBig = TooBig
411 -- When we return a lambda, give a discount if it's used (applied)
412 lamScrutDiscount (SizeIs n vs d) = case opt_UF_FunAppDiscount of { d -> SizeIs n vs (iUnbox d) }
413 lamScrutDiscount TooBig = TooBig
417 %************************************************************************
419 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
421 %************************************************************************
423 We have very limited information about an unfolding expression: (1)~so
424 many type arguments and so many value arguments expected---for our
425 purposes here, we assume we've got those. (2)~A ``size'' or ``cost,''
426 a single integer. (3)~An ``argument info'' vector. For this, what we
427 have at the moment is a Boolean per argument position that says, ``I
428 will look with great favour on an explicit constructor in this
429 position.'' (4)~The ``discount'' to subtract if the expression
430 is being scrutinised.
432 Assuming we have enough type- and value arguments (if not, we give up
433 immediately), then we see if the ``discounted size'' is below some
434 (semi-arbitrary) threshold. It works like this: for every argument
435 position where we're looking for a constructor AND WE HAVE ONE in our
436 hands, we get a (again, semi-arbitrary) discount [proportion to the
437 number of constructors in the type being scrutinized].
439 If we're in the context of a scrutinee ( \tr{(case <expr > of A .. -> ...;.. )})
440 and the expression in question will evaluate to a constructor, we use
441 the computed discount size *for the result only* rather than
442 computing the argument discounts. Since we know the result of
443 the expression is going to be taken apart, discounting its size
444 is more accurate (see @sizeExpr@ above for how this discount size
447 We use this one to avoid exporting inlinings that we ``couldn't possibly
448 use'' on the other side. Can be overridden w/ flaggery.
449 Just the same as smallEnoughToInline, except that it has no actual arguments.
452 couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
453 couldBeSmallEnoughToInline threshold rhs = case calcUnfoldingGuidance threshold rhs of
457 certainlyWillInline :: Unfolding -> Bool
458 -- Sees if the unfolding is pretty certain to inline
459 certainlyWillInline (CoreUnfolding _ _ _ is_cheap (UnfoldIfGoodArgs n_vals _ size _))
460 = is_cheap && size - (n_vals +1) <= opt_UF_UseThreshold
461 certainlyWillInline other
465 %************************************************************************
467 \subsection{callSiteInline}
469 %************************************************************************
471 This is the key function. It decides whether to inline a variable at a call site
473 callSiteInline is used at call sites, so it is a bit more generous.
474 It's a very important function that embodies lots of heuristics.
475 A non-WHNF can be inlined if it doesn't occur inside a lambda,
476 and occurs exactly once or
477 occurs once in each branch of a case and is small
479 If the thing is in WHNF, there's no danger of duplicating work,
480 so we can inline if it occurs once, or is small
482 NOTE: we don't want to inline top-level functions that always diverge.
483 It just makes the code bigger. Tt turns out that the convenient way to prevent
484 them inlining is to give them a NOINLINE pragma, which we do in
485 StrictAnal.addStrictnessInfoToTopId
488 callSiteInline :: DynFlags
489 -> Bool -- True <=> the Id can be inlined
490 -> Bool -- 'inline' note at call site
493 -> [Bool] -- One for each value arg; True if it is interesting
494 -> Bool -- True <=> continuation is interesting
495 -> Maybe CoreExpr -- Unfolding, if any
498 callSiteInline dflags active_inline inline_call occ id arg_infos interesting_cont
499 = case idUnfolding id of {
500 NoUnfolding -> Nothing ;
501 OtherCon cs -> Nothing ;
503 CompulsoryUnfolding unf_template -> Just unf_template ;
504 -- CompulsoryUnfolding => there is no top-level binding
505 -- for these things, so we must inline it.
506 -- Only a couple of primop-like things have
507 -- compulsory unfoldings (see MkId.lhs).
508 -- We don't allow them to be inactive
510 CoreUnfolding unf_template is_top is_value is_cheap guidance ->
513 result | yes_or_no = Just unf_template
514 | otherwise = Nothing
516 n_val_args = length arg_infos
519 | not active_inline = False
520 | otherwise = case occ of
521 IAmDead -> pprTrace "callSiteInline: dead" (ppr id) False
522 IAmALoopBreaker -> False
523 OneOcc in_lam _ _ -> (not in_lam || is_cheap) && consider_safe True
524 other -> is_cheap && consider_safe False
525 -- we consider even the once-in-one-branch
526 -- occurrences, because they won't all have been
527 -- caught by preInlineUnconditionally. In particular,
528 -- if the occurrence is once inside a lambda, and the
529 -- rhs is cheap but not a manifest lambda, then
530 -- pre-inline will not have inlined it for fear of
531 -- invalidating the occurrence info in the rhs.
534 -- consider_safe decides whether it's a good idea to
535 -- inline something, given that there's no
536 -- work-duplication issue (the caller checks that).
542 UnfoldIfGoodArgs n_vals_wanted arg_discounts size res_discount
544 | enough_args && size <= (n_vals_wanted + 1)
545 -- Inline unconditionally if there no size increase
546 -- Size of call is n_vals_wanted (+1 for the function)
550 -> some_benefit && small_enough
553 some_benefit = or arg_infos || really_interesting_cont ||
554 (not is_top && (once || (n_vals_wanted > 0 && enough_args)))
555 -- If it occurs more than once, there must be
556 -- something interesting about some argument, or the
557 -- result context, to make it worth inlining
559 -- If a function has a nested defn we also record
560 -- some-benefit, on the grounds that we are often able
561 -- to eliminate the binding, and hence the allocation,
562 -- for the function altogether; this is good for join
563 -- points. But this only makes sense for *functions*;
564 -- inlining a constructor doesn't help allocation
565 -- unless the result is scrutinised. UNLESS the
566 -- constructor occurs just once, albeit possibly in
567 -- multiple case branches. Then inlining it doesn't
568 -- increase allocation, but it does increase the
569 -- chance that the constructor won't be allocated at
570 -- all in the branches that don't use it.
572 enough_args = n_val_args >= n_vals_wanted
573 really_interesting_cont | n_val_args < n_vals_wanted = False -- Too few args
574 | n_val_args == n_vals_wanted = interesting_cont
575 | otherwise = True -- Extra args
576 -- really_interesting_cont tells if the result of the
577 -- call is in an interesting context.
579 small_enough = (size - discount) <= opt_UF_UseThreshold
580 discount = computeDiscount n_vals_wanted arg_discounts res_discount
581 arg_infos really_interesting_cont
584 if dopt Opt_D_dump_inlinings dflags then
585 pprTrace "Considering inlining"
586 (ppr id <+> vcat [text "active:" <+> ppr active_inline,
587 text "occ info:" <+> ppr occ,
588 text "arg infos" <+> ppr arg_infos,
589 text "interesting continuation" <+> ppr interesting_cont,
590 text "is value:" <+> ppr is_value,
591 text "is cheap:" <+> ppr is_cheap,
592 text "guidance" <+> ppr guidance,
593 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO"])
599 computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Bool -> Int
600 computeDiscount n_vals_wanted arg_discounts res_discount arg_infos result_used
601 -- We multiple the raw discounts (args_discount and result_discount)
602 -- ty opt_UnfoldingKeenessFactor because the former have to do with
603 -- *size* whereas the discounts imply that there's some extra
604 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
607 -- we also discount 1 for each argument passed, because these will
608 -- reduce with the lambdas in the function (we count 1 for a lambda
610 = 1 + -- Discount of 1 because the result replaces the call
611 -- so we count 1 for the function itself
612 length (take n_vals_wanted arg_infos) +
613 -- Discount of 1 for each arg supplied, because the
614 -- result replaces the call
615 round (opt_UF_KeenessFactor *
616 fromIntegral (arg_discount + result_discount))
618 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
620 mk_arg_discount discount is_evald | is_evald = discount
623 -- Don't give a result discount unless there are enough args
624 result_discount | result_used = res_discount -- Over-applied, or case scrut