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 ( pprCoreExpr )
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 mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
91 = CompulsoryUnfolding (occurAnalyseExpr expr)
95 %************************************************************************
97 \subsection{The UnfoldingGuidance type}
99 %************************************************************************
102 instance Outputable UnfoldingGuidance where
103 ppr UnfoldNever = ptext SLIT("NEVER")
104 ppr (UnfoldIfGoodArgs v cs size discount)
105 = hsep [ ptext SLIT("IF_ARGS"), int v,
106 brackets (hsep (map int cs)),
113 calcUnfoldingGuidance
114 :: Int -- bomb out if size gets bigger than this
115 -> CoreExpr -- expression to look at
117 calcUnfoldingGuidance bOMB_OUT_SIZE expr
118 = case collect_val_bndrs expr of { (inline, val_binders, body) ->
120 n_val_binders = length val_binders
122 max_inline_size = n_val_binders+2
123 -- The idea is that if there is an INLINE pragma (inline is True)
124 -- and there's a big body, we give a size of n_val_binders+2. This
125 -- This is just enough to fail the no-size-increase test in callSiteInline,
126 -- so that INLINE things don't get inlined into entirely boring contexts,
130 case (sizeExpr (iUnbox bOMB_OUT_SIZE) val_binders body) of
133 | not inline -> UnfoldNever
134 -- A big function with an INLINE pragma must
135 -- have an UnfoldIfGoodArgs guidance
136 | otherwise -> UnfoldIfGoodArgs n_val_binders
137 (map (const 0) val_binders)
140 SizeIs size cased_args scrut_discount
143 (map discount_for val_binders)
145 (iBox scrut_discount)
147 boxed_size = iBox size
149 final_size | inline = boxed_size `min` max_inline_size
150 | otherwise = boxed_size
152 -- Sometimes an INLINE thing is smaller than n_val_binders+2.
153 -- A particular case in point is a constructor, which has size 1.
154 -- We want to inline this regardless, hence the `min`
156 discount_for b = foldlBag (\acc (b',n) -> if b==b' then acc+n else acc)
160 collect_val_bndrs e = go False [] e
161 -- We need to be a bit careful about how we collect the
162 -- value binders. In ptic, if we see
163 -- __inline_me (\x y -> e)
164 -- We want to say "2 value binders". Why? So that
165 -- we take account of information given for the arguments
167 go inline rev_vbs (Note InlineMe e) = go True rev_vbs e
168 go inline rev_vbs (Lam b e) | isId b = go inline (b:rev_vbs) e
169 | otherwise = go inline rev_vbs e
170 go inline rev_vbs e = (inline, reverse rev_vbs, e)
174 sizeExpr :: Int# -- Bomb out if it gets bigger than this
175 -> [Id] -- Arguments; we're interested in which of these
180 sizeExpr bOMB_OUT_SIZE top_args expr
183 size_up (Type t) = sizeZero -- Types cost nothing
184 size_up (Var v) = sizeOne
186 size_up (Note InlineMe body) = sizeOne -- Inline notes make it look very small
187 -- This can be important. If you have an instance decl like this:
188 -- instance Foo a => Foo [a] where
189 -- {-# INLINE op1, op2 #-}
192 -- then we'll get a dfun which is a pair of two INLINE lambdas
194 size_up (Note _ body) = size_up body -- Other notes cost nothing
196 size_up (App fun (Type t)) = size_up fun
197 size_up (App fun arg) = size_up_app fun [arg]
199 size_up (Lit lit) = sizeN (litSize lit)
201 size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
202 | otherwise = size_up e
204 size_up (Let (NonRec binder rhs) body)
205 = nukeScrutDiscount (size_up rhs) `addSize`
206 size_up body `addSizeN`
207 (if isUnLiftedType (idType binder) then 0 else 1)
208 -- For the allocation
209 -- If the binder has an unlifted type there is no allocation
211 size_up (Let (Rec pairs) body)
212 = nukeScrutDiscount rhs_size `addSize`
213 size_up body `addSizeN`
214 length pairs -- For the allocation
216 rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
218 size_up (Case (Var v) _ _ alts)
219 | v `elem` top_args -- We are scrutinising an argument variable
221 {- I'm nuking this special case; BUT see the comment with case alternatives.
223 (a) It's too eager. We don't want to inline a wrapper into a
224 context with no benefit.
225 E.g. \ x. f (x+x) no point in inlining (+) here!
227 (b) It's ineffective. Once g's wrapper is inlined, its case-expressions
228 aren't scrutinising arguments any more
232 [alt] -> size_up_alt alt `addSize` SizeIs 0# (unitBag (v, 1)) 0#
233 -- We want to make wrapper-style evaluation look cheap, so that
234 -- when we inline a wrapper it doesn't make call site (much) bigger
235 -- Otherwise we get nasty phase ordering stuff:
238 -- If we inline g's wrapper, f looks big, and doesn't get inlined
239 -- into h; if we inline f first, while it looks small, then g's
240 -- wrapper will get inlined later anyway. To avoid this nasty
241 -- ordering difference, we make (case a of (x,y) -> ...),
242 -- *where a is one of the arguments* look free.
246 alts_size (foldr addSize sizeOne alt_sizes) -- The 1 is for the scrutinee
247 (foldr1 maxSize alt_sizes)
249 -- Good to inline if an arg is scrutinised, because
250 -- that may eliminate allocation in the caller
251 -- And it eliminates the case itself
254 alt_sizes = map size_up_alt alts
256 -- alts_size tries to compute a good discount for
257 -- the case when we are scrutinising an argument variable
258 alts_size (SizeIs tot tot_disc tot_scrut) -- Size of all alternatives
259 (SizeIs max max_disc max_scrut) -- Size of biggest alternative
260 = SizeIs tot (unitBag (v, iBox (_ILIT 1 +# tot -# max)) `unionBags` max_disc) max_scrut
261 -- If the variable is known, we produce a discount that
262 -- will take us back to 'max', the size of rh largest alternative
263 -- The 1+ is a little discount for reduced allocation in the caller
264 alts_size tot_size _ = tot_size
267 size_up (Case e _ _ alts) = nukeScrutDiscount (size_up e) `addSize`
268 foldr (addSize . size_up_alt) sizeZero alts
269 -- We don't charge for the case itself
270 -- It's a strict thing, and the price of the call
271 -- is paid by scrut. Also consider
272 -- case f x of DEFAULT -> e
273 -- This is just ';'! Don't charge for it.
276 size_up_app (App fun arg) args
277 | isTypeArg arg = size_up_app fun args
278 | otherwise = size_up_app fun (arg:args)
279 size_up_app fun args = foldr (addSize . nukeScrutDiscount . size_up)
280 (size_up_fun fun args)
283 -- A function application with at least one value argument
284 -- so if the function is an argument give it an arg-discount
286 -- Also behave specially if the function is a build
288 -- Also if the function is a constant Id (constr or primop)
289 -- compute discounts specially
290 size_up_fun (Var fun) args
291 | fun `hasKey` buildIdKey = buildSize
292 | fun `hasKey` augmentIdKey = augmentSize
294 = case globalIdDetails fun of
295 DataConWorkId dc -> conSizeN dc (valArgCount args)
297 FCallId fc -> sizeN opt_UF_DearOp
298 PrimOpId op -> primOpSize op (valArgCount args)
299 -- foldr addSize (primOpSize op) (map arg_discount args)
300 -- At one time I tried giving an arg-discount if a primop
301 -- is applied to one of the function's arguments, but it's
302 -- not good. At the moment, any unlifted-type arg gets a
303 -- 'True' for 'yes I'm evald', so we collect the discount even
304 -- if we know nothing about it. And just having it in a primop
305 -- doesn't help at all if we don't know something more.
307 other -> fun_discount fun `addSizeN`
308 (1 + length (filter (not . exprIsTrivial) args))
309 -- The 1+ is for the function itself
310 -- Add 1 for each non-trivial arg;
311 -- the allocation cost, as in let(rec)
312 -- Slight hack here: for constructors the args are almost always
313 -- trivial; and for primops they are almost always prim typed
314 -- We should really only count for non-prim-typed args in the
315 -- general case, but that seems too much like hard work
317 size_up_fun other args = size_up other
320 size_up_alt (con, bndrs, rhs) = size_up rhs
321 -- Don't charge for args, so that wrappers look cheap
322 -- (See comments about wrappers with Case)
325 -- We want to record if we're case'ing, or applying, an argument
326 fun_discount v | v `elem` top_args = SizeIs 0# (unitBag (v, opt_UF_FunAppDiscount)) 0#
327 fun_discount other = sizeZero
330 -- These addSize things have to be here because
331 -- I don't want to give them bOMB_OUT_SIZE as an argument
333 addSizeN TooBig _ = TooBig
334 addSizeN (SizeIs n xs d) m = mkSizeIs bOMB_OUT_SIZE (n +# iUnbox m) xs d
336 addSize TooBig _ = TooBig
337 addSize _ TooBig = TooBig
338 addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
339 = mkSizeIs bOMB_OUT_SIZE (n1 +# n2) (xs `unionBags` ys) (d1 +# d2)
342 Code for manipulating sizes
345 data ExprSize = TooBig
346 | SizeIs FastInt -- Size found
347 (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
348 FastInt -- Size to subtract if result is scrutinised
349 -- by a case expression
351 -- subtract the discount before deciding whether to bale out. eg. we
352 -- want to inline a large constructor application into a selector:
353 -- tup = (a_1, ..., a_99)
354 -- x = case tup of ...
356 mkSizeIs max n xs d | (n -# d) ># max = TooBig
357 | otherwise = SizeIs n xs d
359 maxSize TooBig _ = TooBig
360 maxSize _ TooBig = TooBig
361 maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
364 sizeZero = SizeIs (_ILIT 0) emptyBag (_ILIT 0)
365 sizeOne = SizeIs (_ILIT 1) emptyBag (_ILIT 0)
366 sizeN n = SizeIs (iUnbox n) emptyBag (_ILIT 0)
368 | isUnboxedTupleCon dc = SizeIs (_ILIT 0) emptyBag (iUnbox n +# _ILIT 1)
369 | otherwise = SizeIs (_ILIT 1) emptyBag (iUnbox n +# _ILIT 1)
370 -- Treat constructors as size 1; we are keen to expose them
371 -- (and we charge separately for their args). We can't treat
372 -- them as size zero, else we find that (iBox x) has size 1,
373 -- which is the same as a lone variable; and hence 'v' will
374 -- always be replaced by (iBox x), where v is bound to iBox x.
376 -- However, unboxed tuples count as size zero
377 -- I found occasions where we had
378 -- f x y z = case op# x y z of { s -> (# s, () #) }
379 -- and f wasn't getting inlined
382 | not (primOpIsDupable op) = sizeN opt_UF_DearOp
383 | not (primOpOutOfLine op) = sizeN (2 - n_args)
384 -- Be very keen to inline simple primops.
385 -- We give a discount of 1 for each arg so that (op# x y z) costs 2.
386 -- We can't make it cost 1, else we'll inline let v = (op# x y z)
387 -- at every use of v, which is excessive.
389 -- A good example is:
390 -- let x = +# p q in C {x}
391 -- Even though x get's an occurrence of 'many', its RHS looks cheap,
392 -- and there's a good chance it'll get inlined back into C's RHS. Urgh!
393 | otherwise = sizeOne
395 buildSize = SizeIs (-2#) emptyBag 4#
396 -- We really want to inline applications of build
397 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
398 -- Indeed, we should add a result_discount becuause build is
399 -- very like a constructor. We don't bother to check that the
400 -- build is saturated (it usually is). The "-2" discounts for the \c n,
401 -- The "4" is rather arbitrary.
403 augmentSize = SizeIs (-2#) emptyBag 4#
404 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
405 -- e plus ys. The -2 accounts for the \cn
407 nukeScrutDiscount (SizeIs n vs d) = SizeIs n vs 0#
408 nukeScrutDiscount TooBig = TooBig
410 -- When we return a lambda, give a discount if it's used (applied)
411 lamScrutDiscount (SizeIs n vs d) = case opt_UF_FunAppDiscount of { d -> SizeIs n vs (iUnbox d) }
412 lamScrutDiscount TooBig = TooBig
416 %************************************************************************
418 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
420 %************************************************************************
422 We have very limited information about an unfolding expression: (1)~so
423 many type arguments and so many value arguments expected---for our
424 purposes here, we assume we've got those. (2)~A ``size'' or ``cost,''
425 a single integer. (3)~An ``argument info'' vector. For this, what we
426 have at the moment is a Boolean per argument position that says, ``I
427 will look with great favour on an explicit constructor in this
428 position.'' (4)~The ``discount'' to subtract if the expression
429 is being scrutinised.
431 Assuming we have enough type- and value arguments (if not, we give up
432 immediately), then we see if the ``discounted size'' is below some
433 (semi-arbitrary) threshold. It works like this: for every argument
434 position where we're looking for a constructor AND WE HAVE ONE in our
435 hands, we get a (again, semi-arbitrary) discount [proportion to the
436 number of constructors in the type being scrutinized].
438 If we're in the context of a scrutinee ( \tr{(case <expr > of A .. -> ...;.. )})
439 and the expression in question will evaluate to a constructor, we use
440 the computed discount size *for the result only* rather than
441 computing the argument discounts. Since we know the result of
442 the expression is going to be taken apart, discounting its size
443 is more accurate (see @sizeExpr@ above for how this discount size
446 We use this one to avoid exporting inlinings that we ``couldn't possibly
447 use'' on the other side. Can be overridden w/ flaggery.
448 Just the same as smallEnoughToInline, except that it has no actual arguments.
451 couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
452 couldBeSmallEnoughToInline threshold rhs = case calcUnfoldingGuidance threshold rhs of
456 certainlyWillInline :: Unfolding -> Bool
457 -- Sees if the unfolding is pretty certain to inline
458 certainlyWillInline (CoreUnfolding _ _ _ is_cheap (UnfoldIfGoodArgs n_vals _ size _))
459 = is_cheap && size - (n_vals +1) <= opt_UF_UseThreshold
460 certainlyWillInline other
463 smallEnoughToInline :: Unfolding -> Bool
464 smallEnoughToInline (CoreUnfolding _ _ _ _ (UnfoldIfGoodArgs _ _ size _))
465 = size <= opt_UF_UseThreshold
466 smallEnoughToInline other
470 %************************************************************************
472 \subsection{callSiteInline}
474 %************************************************************************
476 This is the key function. It decides whether to inline a variable at a call site
478 callSiteInline is used at call sites, so it is a bit more generous.
479 It's a very important function that embodies lots of heuristics.
480 A non-WHNF can be inlined if it doesn't occur inside a lambda,
481 and occurs exactly once or
482 occurs once in each branch of a case and is small
484 If the thing is in WHNF, there's no danger of duplicating work,
485 so we can inline if it occurs once, or is small
487 NOTE: we don't want to inline top-level functions that always diverge.
488 It just makes the code bigger. Tt turns out that the convenient way to prevent
489 them inlining is to give them a NOINLINE pragma, which we do in
490 StrictAnal.addStrictnessInfoToTopId
493 callSiteInline :: DynFlags
494 -> Bool -- True <=> the Id can be inlined
495 -> Bool -- 'inline' note at call site
498 -> [Bool] -- One for each value arg; True if it is interesting
499 -> Bool -- True <=> continuation is interesting
500 -> Maybe CoreExpr -- Unfolding, if any
503 callSiteInline dflags active_inline inline_call occ id arg_infos interesting_cont
504 = case idUnfolding id of {
505 NoUnfolding -> Nothing ;
506 OtherCon cs -> Nothing ;
508 CompulsoryUnfolding unf_template -> Just unf_template ;
509 -- CompulsoryUnfolding => there is no top-level binding
510 -- for these things, so we must inline it.
511 -- Only a couple of primop-like things have
512 -- compulsory unfoldings (see MkId.lhs).
513 -- We don't allow them to be inactive
515 CoreUnfolding unf_template is_top is_value is_cheap guidance ->
518 result | yes_or_no = Just unf_template
519 | otherwise = Nothing
521 n_val_args = length arg_infos
524 | not active_inline = False
525 | otherwise = case occ of
526 IAmDead -> pprTrace "callSiteInline: dead" (ppr id) False
527 IAmALoopBreaker -> False
528 --OneOcc in_lam _ _ -> (not in_lam || is_cheap) && consider_safe True
529 other -> is_cheap && consider_safe False
530 -- we consider even the once-in-one-branch
531 -- occurrences, because they won't all have been
532 -- caught by preInlineUnconditionally. In particular,
533 -- if the occurrence is once inside a lambda, and the
534 -- rhs is cheap but not a manifest lambda, then
535 -- pre-inline will not have inlined it for fear of
536 -- invalidating the occurrence info in the rhs.
539 -- consider_safe decides whether it's a good idea to
540 -- inline something, given that there's no
541 -- work-duplication issue (the caller checks that).
547 UnfoldIfGoodArgs n_vals_wanted arg_discounts size res_discount
549 | enough_args && size <= (n_vals_wanted + 1)
550 -- Inline unconditionally if there no size increase
551 -- Size of call is n_vals_wanted (+1 for the function)
555 -> some_benefit && small_enough
558 some_benefit = or arg_infos || really_interesting_cont ||
559 (not is_top && ({- once || -} (n_vals_wanted > 0 && enough_args)))
560 -- [was (once && not in_lam)]
561 -- If it occurs more than once, there must be
562 -- something interesting about some argument, or the
563 -- result context, to make it worth inlining
565 -- If a function has a nested defn we also record
566 -- some-benefit, on the grounds that we are often able
567 -- to eliminate the binding, and hence the allocation,
568 -- for the function altogether; this is good for join
569 -- points. But this only makes sense for *functions*;
570 -- inlining a constructor doesn't help allocation
571 -- unless the result is scrutinised. UNLESS the
572 -- constructor occurs just once, albeit possibly in
573 -- multiple case branches. Then inlining it doesn't
574 -- increase allocation, but it does increase the
575 -- chance that the constructor won't be allocated at
576 -- all in the branches that don't use it.
578 enough_args = n_val_args >= n_vals_wanted
579 really_interesting_cont | n_val_args < n_vals_wanted = False -- Too few args
580 | n_val_args == n_vals_wanted = interesting_cont
581 | otherwise = True -- Extra args
582 -- really_interesting_cont tells if the result of the
583 -- call is in an interesting context.
585 small_enough = (size - discount) <= opt_UF_UseThreshold
586 discount = computeDiscount n_vals_wanted arg_discounts res_discount
587 arg_infos really_interesting_cont
590 if dopt Opt_D_dump_inlinings dflags then
591 pprTrace "Considering inlining"
592 (ppr id <+> vcat [text "active:" <+> ppr active_inline,
593 text "occ info:" <+> ppr occ,
594 text "arg infos" <+> ppr arg_infos,
595 text "interesting continuation" <+> ppr interesting_cont,
596 text "is value:" <+> ppr is_value,
597 text "is cheap:" <+> ppr is_cheap,
598 text "guidance" <+> ppr guidance,
599 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO"])
605 computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Bool -> Int
606 computeDiscount n_vals_wanted arg_discounts res_discount arg_infos result_used
607 -- We multiple the raw discounts (args_discount and result_discount)
608 -- ty opt_UnfoldingKeenessFactor because the former have to do with
609 -- *size* whereas the discounts imply that there's some extra
610 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
613 -- we also discount 1 for each argument passed, because these will
614 -- reduce with the lambdas in the function (we count 1 for a lambda
616 = 1 + -- Discount of 1 because the result replaces the call
617 -- so we count 1 for the function itself
618 length (take n_vals_wanted arg_infos) +
619 -- Discount of 1 for each arg supplied, because the
620 -- result replaces the call
621 round (opt_UF_KeenessFactor *
622 fromIntegral (arg_discount + result_discount))
624 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
626 mk_arg_discount discount is_evald | is_evald = discount
629 -- Don't give a result discount unless there are enough args
630 result_discount | result_used = res_discount -- Over-applied, or case scrut