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, -- types
19 noUnfolding, mkTopUnfolding, mkUnfolding, mkCompulsoryUnfolding, seqUnfolding,
20 mkOtherCon, otherCons,
21 unfoldingTemplate, maybeUnfoldingTemplate,
22 isEvaldUnfolding, isCheapUnfolding, isCompulsoryUnfolding,
23 hasUnfolding, hasSomeUnfolding,
25 couldBeSmallEnoughToInline,
26 certainlySmallEnoughToInline,
29 calcUnfoldingGuidance,
31 callSiteInline, blackListed
34 #include "HsVersions.h"
36 import CmdLineOpts ( opt_UF_CreationThreshold,
38 opt_UF_ScrutConDiscount,
39 opt_UF_FunAppDiscount,
40 opt_UF_PrimArgDiscount,
42 opt_UF_CheapOp, opt_UF_DearOp, opt_UF_NoRepLit,
43 opt_UnfoldCasms, opt_PprStyle_Debug,
47 import PprCore ( pprCoreExpr )
48 import OccurAnal ( occurAnalyseGlobalExpr )
50 import CoreUtils ( coreExprType, exprIsTrivial, exprIsValue, exprIsCheap )
51 import Id ( Id, idType, idUnique, isId, getIdWorkerInfo,
52 getIdSpecialisation, getInlinePragma, getIdUnfolding,
56 import Name ( isLocallyDefined )
57 import Const ( Con(..), isLitLitLit, isWHNFCon )
58 import PrimOp ( PrimOp(..), primOpIsDupable )
59 import IdInfo ( ArityInfo(..), InlinePragInfo(..), OccInfo(..), insideLam, workerExists )
60 import TyCon ( tyConFamilySize )
61 import Type ( splitAlgTyConApp_maybe, splitFunTy_maybe, isUnLiftedType )
62 import Const ( isNoRepLit )
63 import Unique ( Unique, buildIdKey, augmentIdKey )
64 import Maybes ( maybeToBool )
66 import Util ( isIn, lengthExceeds )
69 #if __GLASGOW_HASKELL__ >= 404
70 import GlaExts ( fromInt )
74 %************************************************************************
76 \subsection{@Unfolding@ and @UnfoldingGuidance@ types}
78 %************************************************************************
84 | OtherCon [Con] -- It ain't one of these
85 -- (OtherCon xs) also indicates that something has been evaluated
86 -- and hence there's no point in re-evaluating it.
87 -- OtherCon [] is used even for non-data-type values
88 -- to indicated evaluated-ness. Notably:
89 -- data C = C !(Int -> Int)
90 -- case x of { C f -> ... }
91 -- Here, f gets an OtherCon [] unfolding.
93 | CompulsoryUnfolding CoreExpr -- There is no "original" definition,
94 -- so you'd better unfold.
96 | CoreUnfolding -- An unfolding with redundant cached information
97 CoreExpr -- Template; binder-info is correct
98 Bool -- This is a top-level binding
99 Bool -- exprIsCheap template (cached); it won't duplicate (much) work
100 -- if you inline this in more than one place
101 Bool -- exprIsValue template (cached); it is ok to discard a `seq` on
103 UnfoldingGuidance -- Tells about the *size* of the template.
105 seqUnfolding :: Unfolding -> ()
106 seqUnfolding (CoreUnfolding e top b1 b2 g)
107 = seqExpr e `seq` top `seq` b1 `seq` b2 `seq` seqGuidance g
108 seqUnfolding other = ()
112 noUnfolding = NoUnfolding
113 mkOtherCon = OtherCon
115 mkTopUnfolding expr = mkUnfolding True expr
117 mkUnfolding top_lvl expr
118 = CoreUnfolding (occurAnalyseGlobalExpr expr)
122 (calcUnfoldingGuidance opt_UF_CreationThreshold expr)
124 mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
125 = CompulsoryUnfolding (occurAnalyseGlobalExpr expr)
127 unfoldingTemplate :: Unfolding -> CoreExpr
128 unfoldingTemplate (CoreUnfolding expr _ _ _ _) = expr
129 unfoldingTemplate (CompulsoryUnfolding expr) = expr
130 unfoldingTemplate other = panic "getUnfoldingTemplate"
132 maybeUnfoldingTemplate :: Unfolding -> Maybe CoreExpr
133 maybeUnfoldingTemplate (CoreUnfolding expr _ _ _ _) = Just expr
134 maybeUnfoldingTemplate (CompulsoryUnfolding expr) = Just expr
135 maybeUnfoldingTemplate other = Nothing
137 otherCons (OtherCon cons) = cons
140 isEvaldUnfolding :: Unfolding -> Bool
141 isEvaldUnfolding (OtherCon _) = True
142 isEvaldUnfolding (CoreUnfolding _ _ _ is_evald _) = is_evald
143 isEvaldUnfolding other = False
145 isCheapUnfolding :: Unfolding -> Bool
146 isCheapUnfolding (CoreUnfolding _ _ is_cheap _ _) = is_cheap
147 isCheapUnfolding other = False
149 isCompulsoryUnfolding :: Unfolding -> Bool
150 isCompulsoryUnfolding (CompulsoryUnfolding _) = True
151 isCompulsoryUnfolding other = False
153 hasUnfolding :: Unfolding -> Bool
154 hasUnfolding (CoreUnfolding _ _ _ _ _) = True
155 hasUnfolding (CompulsoryUnfolding _) = True
156 hasUnfolding other = False
158 hasSomeUnfolding :: Unfolding -> Bool
159 hasSomeUnfolding NoUnfolding = False
160 hasSomeUnfolding other = True
162 data UnfoldingGuidance
164 | UnfoldIfGoodArgs Int -- and "n" value args
166 [Int] -- Discount if the argument is evaluated.
167 -- (i.e., a simplification will definitely
168 -- be possible). One elt of the list per *value* arg.
170 Int -- The "size" of the unfolding; to be elaborated
173 Int -- Scrutinee discount: the discount to substract if the thing is in
174 -- a context (case (thing args) of ...),
175 -- (where there are the right number of arguments.)
177 seqGuidance (UnfoldIfGoodArgs n ns a b) = n `seq` sum ns `seq` a `seq` b `seq` ()
178 seqGuidance other = ()
182 instance Outputable UnfoldingGuidance where
183 ppr UnfoldNever = ptext SLIT("NEVER")
184 ppr (UnfoldIfGoodArgs v cs size discount)
185 = hsep [ ptext SLIT("IF_ARGS"), int v,
186 brackets (hsep (map int cs)),
192 %************************************************************************
194 \subsection[calcUnfoldingGuidance]{Calculate ``unfolding guidance'' for an expression}
196 %************************************************************************
199 calcUnfoldingGuidance
200 :: Int -- bomb out if size gets bigger than this
201 -> CoreExpr -- expression to look at
203 calcUnfoldingGuidance bOMB_OUT_SIZE expr
204 = case collect_val_bndrs expr of { (inline, val_binders, body) ->
206 n_val_binders = length val_binders
208 case (sizeExpr bOMB_OUT_SIZE val_binders body) of
211 | not inline -> UnfoldNever
212 -- A big function with an INLINE pragma must
213 -- have an UnfoldIfGoodArgs guidance
214 | inline -> UnfoldIfGoodArgs n_val_binders
215 (map (const 0) val_binders)
216 (n_val_binders + 2) 0
217 -- See comments with final_size below
219 SizeIs size cased_args scrut_discount
222 (map discount_for val_binders)
228 final_size | inline = 0 -- Trying very agresssive inlining of INLINE things.
229 -- Reason: we don't want to call the un-inlined version,
230 -- because its body is awful
231 -- boxed_size `min` (n_val_binders + 2) -- Trying "+2" again...
232 | otherwise = boxed_size
233 -- The idea is that if there is an INLINE pragma (inline is True)
234 -- and there's a big body, we give a size of n_val_binders+1. This
235 -- This is enough to pass the no-size-increase test in callSiteInline,
237 -- I tried n_val_binders+2, to just defeat the test, on the grounds that
238 -- we don't want to inline an INLINE thing into a totally boring context,
239 -- but I found that some wrappers (notably one for a join point) weren't
240 -- getting inlined, and that was terrible. In that particular case, the
241 -- call site applied the wrapper to realWorld#, so if we made that an
242 -- "interesting" value the inlining would have happened... but it was
243 -- simpler to inline wrappers a little more eagerly instead.
245 -- Sometimes, though, an INLINE thing is smaller than n_val_binders+2.
246 -- A particular case in point is a constructor, which has size 1.
247 -- We want to inline this regardless, hence the `min`
251 | is_fun_ty = num_cases * opt_UF_FunAppDiscount
252 | is_data_ty = num_cases * opt_UF_ScrutConDiscount
253 | otherwise = num_cases * opt_UF_PrimArgDiscount
255 num_cases = foldlBag (\n b' -> if b==b' then n+1 else n) 0 cased_args
256 -- Count occurrences of b in cased_args
258 is_fun_ty = maybeToBool (splitFunTy_maybe arg_ty)
259 (is_data_ty, tycon) = case (splitAlgTyConApp_maybe (idType b)) of
260 Nothing -> (False, panic "discount")
261 Just (tc,_,_) -> (True, tc)
265 collect_val_bndrs e = go False [] e
266 -- We need to be a bit careful about how we collect the
267 -- value binders. In ptic, if we see
268 -- __inline_me (\x y -> e)
269 -- We want to say "2 value binders". Why? So that
270 -- we take account of information given for the arguments
272 go inline rev_vbs (Note InlineMe e) = go True rev_vbs e
273 go inline rev_vbs (Lam b e) | isId b = go inline (b:rev_vbs) e
274 | otherwise = go inline rev_vbs e
275 go inline rev_vbs e = (inline, reverse rev_vbs, e)
279 sizeExpr :: Int -- Bomb out if it gets bigger than this
280 -> [Id] -- Arguments; we're interested in which of these
285 sizeExpr (I# bOMB_OUT_SIZE) top_args expr
288 size_up (Type t) = sizeZero -- Types cost nothing
289 size_up (Var v) = sizeOne
291 size_up (Note _ body) = size_up body -- Notes cost nothing
293 size_up (App fun (Type t)) = size_up fun
294 size_up (App fun arg) = size_up_app fun [arg]
296 size_up (Con con args) = foldr (addSize . nukeScrutDiscount . size_up)
297 (size_up_con con args)
300 size_up (Lam b e) | isId b = size_up e `addSizeN` 1
301 | otherwise = size_up e
303 size_up (Let (NonRec binder rhs) body)
304 = nukeScrutDiscount (size_up rhs) `addSize`
305 size_up body `addSizeN`
306 (if isUnLiftedType (idType binder) then 0 else 1)
307 -- For the allocation
308 -- If the binder has an unlifted type there is no allocation
310 size_up (Let (Rec pairs) body)
311 = nukeScrutDiscount rhs_size `addSize`
312 size_up body `addSizeN`
313 length pairs -- For the allocation
315 rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
317 size_up (Case scrut _ alts)
318 = nukeScrutDiscount (size_up scrut) `addSize`
319 arg_discount scrut `addSize`
320 foldr (addSize . size_up_alt) sizeZero alts
321 `addSizeN` 1 -- charge one for the case itself.
323 -- Just charge for the alts that exist, not the ones that might exist
325 -- case (splitAlgTyConApp_maybe (coreExprType scrut)) of
327 -- Just (tc,_,_) -> tyConFamilySize tc
330 size_up_app (App fun arg) args = size_up_app fun (arg:args)
331 size_up_app fun args = foldr (addSize . nukeScrutDiscount . size_up)
332 (size_up_fun fun args)
335 -- A function application with at least one value argument
336 -- so if the function is an argument give it an arg-discount
337 -- Also behave specially if the function is a build
338 -- Also if the function is a constant Id (constr or primop)
339 -- compute discounts as if it were actually a Con; in the early
340 -- stages these constructors and primops may not yet be inlined
341 size_up_fun (Var fun) args | idUnique fun == buildIdKey = buildSize
342 | idUnique fun == augmentIdKey = augmentSize
343 | fun `is_elem` top_args = scrutArg fun `addSize` fun_size
344 | otherwise = fun_size
346 fun_size = case isConstantId_maybe fun of
347 Just con -> size_up_con con args
350 size_up_fun other args = size_up other
353 size_up_alt (con, bndrs, rhs) = size_up rhs
354 -- Don't charge for args, so that wrappers look cheap
357 size_up_con (Literal lit) args | isNoRepLit lit = sizeN opt_UF_NoRepLit
358 | otherwise = sizeOne
360 size_up_con (DataCon dc) args = conSizeN (valArgCount args)
362 size_up_con (PrimOp op) args = foldr addSize (sizeN op_cost) (map arg_discount args)
363 -- Give an arg-discount if a primop is applies to
364 -- one of the function's arguments
366 op_cost | primOpIsDupable op = opt_UF_CheapOp
367 | otherwise = opt_UF_DearOp
369 -- We want to record if we're case'ing, or applying, an argument
370 arg_discount (Var v) | v `is_elem` top_args = scrutArg v
371 arg_discount other = sizeZero
374 is_elem :: Id -> [Id] -> Bool
375 is_elem = isIn "size_up_scrut"
378 -- These addSize things have to be here because
379 -- I don't want to give them bOMB_OUT_SIZE as an argument
381 addSizeN TooBig _ = TooBig
382 addSizeN (SizeIs n xs d) (I# m)
383 | n_tot -# d <# bOMB_OUT_SIZE = SizeIs n_tot xs d
388 addSize TooBig _ = TooBig
389 addSize _ TooBig = TooBig
390 addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
391 | (n_tot -# d_tot) <# bOMB_OUT_SIZE = SizeIs n_tot xys d_tot
396 xys = xs `unionBags` ys
399 Code for manipulating sizes
403 data ExprSize = TooBig
404 | SizeIs Int# -- Size found
405 (Bag Id) -- Arguments cased herein
406 Int# -- Size to subtract if result is scrutinised
407 -- by a case expression
409 sizeZero = SizeIs 0# emptyBag 0#
410 sizeOne = SizeIs 1# emptyBag 0#
411 sizeTwo = SizeIs 2# emptyBag 0#
412 sizeN (I# n) = SizeIs n emptyBag 0#
413 conSizeN (I# n) = SizeIs 1# emptyBag (n +# 1#)
414 -- Treat constructors as size 1, that unfoldAlways responsds 'False'
415 -- when asked about 'x' when x is bound to (C 3#).
416 -- This avoids gratuitous 'ticks' when x itself appears as an
417 -- atomic constructor argument.
419 buildSize = SizeIs (-2#) emptyBag 4#
420 -- We really want to inline applications of build
421 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
422 -- Indeed, we should add a result_discount becuause build is
423 -- very like a constructor. We don't bother to check that the
424 -- build is saturated (it usually is). The "-2" discounts for the \c n,
425 -- The "4" is rather arbitrary.
427 augmentSize = SizeIs (-2#) emptyBag 4#
428 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
429 -- e plus ys. The -2 accounts for the \cn
431 scrutArg v = SizeIs 0# (unitBag v) 0#
433 nukeScrutDiscount (SizeIs n vs d) = SizeIs n vs 0#
434 nukeScrutDiscount TooBig = TooBig
438 %************************************************************************
440 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
442 %************************************************************************
444 We have very limited information about an unfolding expression: (1)~so
445 many type arguments and so many value arguments expected---for our
446 purposes here, we assume we've got those. (2)~A ``size'' or ``cost,''
447 a single integer. (3)~An ``argument info'' vector. For this, what we
448 have at the moment is a Boolean per argument position that says, ``I
449 will look with great favour on an explicit constructor in this
450 position.'' (4)~The ``discount'' to subtract if the expression
451 is being scrutinised.
453 Assuming we have enough type- and value arguments (if not, we give up
454 immediately), then we see if the ``discounted size'' is below some
455 (semi-arbitrary) threshold. It works like this: for every argument
456 position where we're looking for a constructor AND WE HAVE ONE in our
457 hands, we get a (again, semi-arbitrary) discount [proportion to the
458 number of constructors in the type being scrutinized].
460 If we're in the context of a scrutinee ( \tr{(case <expr > of A .. -> ...;.. )})
461 and the expression in question will evaluate to a constructor, we use
462 the computed discount size *for the result only* rather than
463 computing the argument discounts. Since we know the result of
464 the expression is going to be taken apart, discounting its size
465 is more accurate (see @sizeExpr@ above for how this discount size
468 We use this one to avoid exporting inlinings that we ``couldn't possibly
469 use'' on the other side. Can be overridden w/ flaggery.
470 Just the same as smallEnoughToInline, except that it has no actual arguments.
473 couldBeSmallEnoughToInline :: UnfoldingGuidance -> Bool
474 couldBeSmallEnoughToInline UnfoldNever = False
475 couldBeSmallEnoughToInline other = True
477 certainlySmallEnoughToInline :: UnfoldingGuidance -> Bool
478 certainlySmallEnoughToInline UnfoldNever = False
479 certainlySmallEnoughToInline (UnfoldIfGoodArgs _ _ size _) = size <= opt_UF_UseThreshold
482 @okToUnfoldInHifile@ is used when emitting unfolding info into an interface
483 file to determine whether an unfolding candidate really should be unfolded.
484 The predicate is needed to prevent @_casm_@s (+ lit-lits) from being emitted
485 into interface files.
487 The reason for inlining expressions containing _casm_s into interface files
488 is that these fragments of C are likely to mention functions/#defines that
489 will be out-of-scope when inlined into another module. This is not an
490 unfixable problem for the user (just need to -#include the approp. header
491 file), but turning it off seems to the simplest thing to do.
494 okToUnfoldInHiFile :: CoreExpr -> Bool
495 okToUnfoldInHiFile e = opt_UnfoldCasms || go e
497 -- Race over an expression looking for CCalls..
499 go (Con (Literal lit) _) = not (isLitLitLit lit)
500 go (Con (PrimOp op) args) = okToUnfoldPrimOp op && all go args
501 go (Con con args) = all go args -- might be litlits in here
502 go (App fun arg) = go fun && go arg
503 go (Lam _ body) = go body
504 go (Let binds body) = and (map go (body :rhssOfBind binds))
505 go (Case scrut bndr alts) = and (map go (scrut:rhssOfAlts alts))
506 go (Note _ body) = go body
509 -- ok to unfold a PrimOp as long as it's not a _casm_
510 okToUnfoldPrimOp (CCallOp _ is_casm _ _) = not is_casm
511 okToUnfoldPrimOp _ = True
515 %************************************************************************
517 \subsection{callSiteInline}
519 %************************************************************************
521 This is the key function. It decides whether to inline a variable at a call site
523 callSiteInline is used at call sites, so it is a bit more generous.
524 It's a very important function that embodies lots of heuristics.
525 A non-WHNF can be inlined if it doesn't occur inside a lambda,
526 and occurs exactly once or
527 occurs once in each branch of a case and is small
529 If the thing is in WHNF, there's no danger of duplicating work,
530 so we can inline if it occurs once, or is small
533 callSiteInline :: Bool -- True <=> the Id is black listed
534 -> Bool -- 'inline' note at call site
537 -> [Bool] -- One for each value arg; True if it is interesting
538 -> Bool -- True <=> continuation is interesting
539 -> Maybe CoreExpr -- Unfolding, if any
542 callSiteInline black_listed inline_call occ id arg_infos interesting_cont
543 = case getIdUnfolding id of {
544 NoUnfolding -> Nothing ;
545 OtherCon _ -> Nothing ;
546 CompulsoryUnfolding unf_template | black_listed -> Nothing
547 | otherwise -> Just unf_template ;
548 -- Primops have compulsory unfoldings, but
549 -- may have rules, in which case they are
550 -- black listed till later
551 CoreUnfolding unf_template is_top is_cheap _ guidance ->
554 result | yes_or_no = Just unf_template
555 | otherwise = Nothing
557 n_val_args = length arg_infos
560 | black_listed = False
561 | otherwise = case occ of
562 IAmDead -> pprTrace "callSiteInline: dead" (ppr id) False
563 IAmALoopBreaker -> False
564 OneOcc in_lam one_br -> (not in_lam || is_cheap) && consider_safe in_lam True one_br
565 NoOccInfo -> is_cheap && consider_safe True False False
567 consider_safe in_lam once once_in_one_branch
568 -- consider_safe decides whether it's a good idea to inline something,
569 -- given that there's no work-duplication issue (the caller checks that).
570 -- once_in_one_branch = True means there's a unique textual occurrence
573 | once_in_one_branch -- Be very keen to inline something if this is its unique occurrence; that
574 -- gives a good chance of eliminating the original binding for the thing.
575 -- The only time we hold back is when substituting inside a lambda;
576 -- then if the context is totally uninteresting (not applied, not scrutinised)
577 -- there is no point in substituting because it might just increase allocation.
578 = not in_lam || not (null arg_infos) || interesting_cont
582 UnfoldNever -> False ;
583 UnfoldIfGoodArgs n_vals_wanted arg_discounts size res_discount
585 | enough_args && size <= (n_vals_wanted + 1)
587 -- Size of call is n_vals_wanted (+1 for the function)
591 -> some_benefit && small_enough
594 some_benefit = or arg_infos || really_interesting_cont ||
595 (not is_top && (once || (n_vals_wanted > 0 && enough_args)))
596 -- If it occurs more than once, there must be something interesting
597 -- about some argument, or the result context, to make it worth inlining
599 -- If a function has a nested defn we also record some-benefit,
600 -- on the grounds that we are often able to eliminate the binding,
601 -- and hence the allocation, for the function altogether; this is good
602 -- for join points. But this only makes sense for *functions*;
603 -- inlining a constructor doesn't help allocation unless the result is
604 -- scrutinised. UNLESS the constructor occurs just once, albeit possibly
605 -- in multiple case branches. Then inlining it doesn't increase allocation,
606 -- but it does increase the chance that the constructor won't be allocated at all
607 -- in the branches that don't use it.
609 enough_args = n_val_args >= n_vals_wanted
610 really_interesting_cont | n_val_args < n_vals_wanted = False -- Too few args
611 | n_val_args == n_vals_wanted = interesting_cont
612 | otherwise = True -- Extra args
613 -- really_interesting_cont tells if the result of the
614 -- call is in an interesting context.
616 small_enough = (size - discount) <= opt_UF_UseThreshold
617 discount = computeDiscount n_vals_wanted arg_discounts res_discount
618 arg_infos really_interesting_cont
622 if opt_D_dump_inlinings then
623 pprTrace "Considering inlining"
624 (ppr id <+> vcat [text "black listed" <+> ppr black_listed,
625 text "occ info:" <+> ppr occ,
626 text "arg infos" <+> ppr arg_infos,
627 text "interesting continuation" <+> ppr interesting_cont,
628 text "is cheap" <+> ppr is_cheap,
629 text "guidance" <+> ppr guidance,
630 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO",
632 text "Unfolding =" <+> pprCoreExpr unf_template
640 computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Bool -> Int
641 computeDiscount n_vals_wanted arg_discounts res_discount arg_infos result_used
642 -- We multiple the raw discounts (args_discount and result_discount)
643 -- ty opt_UnfoldingKeenessFactor because the former have to do with
644 -- *size* whereas the discounts imply that there's some extra
645 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
648 -- we also discount 1 for each argument passed, because these will
649 -- reduce with the lambdas in the function (we count 1 for a lambda
651 = 1 + -- Discount of 1 because the result replaces the call
652 -- so we count 1 for the function itself
653 length (take n_vals_wanted arg_infos) +
654 -- Discount of 1 for each arg supplied, because the
655 -- result replaces the call
656 round (opt_UF_KeenessFactor *
657 fromInt (arg_discount + result_discount))
659 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
661 mk_arg_discount discount is_evald | is_evald = discount
664 -- Don't give a result discount unless there are enough args
665 result_discount | result_used = res_discount -- Over-applied, or case scrut
670 %************************************************************************
672 \subsection{Black-listing}
674 %************************************************************************
676 Inlining is controlled by the "Inline phase" number, which is set
677 by the per-simplification-pass '-finline-phase' flag.
679 For optimisation we use phase 1,2 and nothing (i.e. no -finline-phase flag)
680 in that order. The meanings of these are determined by the @blackListed@ function
683 The final simplification doesn't have a phase number
689 (least black listing, most inlining)
690 INLINE n foo phase is Just p *and* p<n *and* foo appears on LHS of rule
691 INLINE foo phase is Just p *and* foo appears on LHS of rule
692 NOINLINE n foo phase is Just p *and* (p<n *or* foo appears on LHS of rule)
694 (most black listing, least inlining)
697 blackListed :: IdSet -- Used in transformation rules
698 -> Maybe Int -- Inline phase
699 -> Id -> Bool -- True <=> blacklisted
701 -- The blackListed function sees whether a variable should *not* be
702 -- inlined because of the inline phase we are in. This is the sole
703 -- place that the inline phase number is looked at.
705 blackListed rule_vars Nothing -- Last phase
706 = \v -> case getInlinePragma v of
707 IMustNotBeINLINEd False Nothing -> True -- An unconditional NOINLINE pragma
710 blackListed rule_vars (Just 0)
711 -- Phase 0: used for 'no imported inlinings please'
712 -- This prevents wrappers getting inlined which in turn is bad for full laziness
713 -- NEW: try using 'not a wrapper' rather than 'not imported' in this phase.
714 -- This allows a little more inlining, which seems to be important, sometimes.
715 -- For example PrelArr.newIntArr gets better.
716 = \v -> -- workerExists (getIdWorkerInfo v) || normal_case rule_vars 0 v
717 -- True -- Try going back to no inlinings at all
718 -- BUT: I found that there is some advantage in doing
719 -- local inlinings first. For example in fish/Main.hs
720 -- it's advantageous to inline scale_vec2 before inlining
721 -- wrappers from PrelNum that make it look big.
722 not (isLocallyDefined v) || normal_case rule_vars 0 v -- This seems best at the moment
724 blackListed rule_vars (Just phase)
725 = \v -> normal_case rule_vars phase v
727 normal_case rule_vars phase v
728 = case getInlinePragma v of
729 NoInlinePragInfo -> has_rules
731 IMustNotBeINLINEd from_INLINE Nothing
732 | from_INLINE -> has_rules -- Black list until final phase
733 | otherwise -> True -- Always blacklisted
735 IMustNotBeINLINEd from_inline (Just threshold)
736 | from_inline -> phase < threshold && has_rules
737 | otherwise -> phase < threshold || has_rules
739 has_rules = v `elemVarSet` rule_vars
740 || not (isEmptyCoreRules (getIdSpecialisation v))
744 SLPJ 95/04: Why @runST@ must be inlined very late:
748 (a, s') = newArray# 100 [] s
749 (_, s'') = fill_in_array_or_something a x s'
753 If we inline @runST@, we'll get:
756 (a, s') = newArray# 100 [] realWorld#{-NB-}
757 (_, s'') = fill_in_array_or_something a x s'
761 And now the @newArray#@ binding can be floated to become a CAF, which
762 is totally and utterly wrong:
765 (a, s') = newArray# 100 [] realWorld#{-NB-} -- YIKES!!!
768 let (_, s'') = fill_in_array_or_something a x s' in
771 All calls to @f@ will share a {\em single} array!
773 Yet we do want to inline runST sometime, so we can avoid
774 needless code. Solution: black list it until the last moment.