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 mkOtherCon, otherCons,
21 unfoldingTemplate, maybeUnfoldingTemplate,
22 isEvaldUnfolding, isValueUnfolding, isCheapUnfolding, isCompulsoryUnfolding,
23 hasUnfolding, hasSomeUnfolding,
25 couldBeSmallEnoughToInline,
29 callSiteInline, blackListed
32 #include "HsVersions.h"
34 import CmdLineOpts ( opt_UF_CreationThreshold,
36 opt_UF_ScrutConDiscount,
37 opt_UF_FunAppDiscount,
38 opt_UF_PrimArgDiscount,
40 opt_UF_CheapOp, opt_UF_DearOp,
41 opt_UnfoldCasms, opt_PprStyle_Debug,
45 import PprCore ( pprCoreExpr )
46 import OccurAnal ( occurAnalyseGlobalExpr )
48 import CoreUtils ( exprIsValue, exprIsCheap, exprIsBottom, exprIsTrivial )
49 import Id ( Id, idType, idFlavour, idUnique, isId, idWorkerInfo,
50 idSpecialisation, idInlinePragma, idUnfolding,
54 import Name ( isLocallyDefined )
55 import Literal ( isLitLitLit )
56 import PrimOp ( PrimOp(..), primOpIsDupable, primOpOutOfLine, ccallIsCasm )
57 import IdInfo ( ArityInfo(..), InlinePragInfo(..), OccInfo(..), IdFlavour(..), CprInfo(..),
58 insideLam, workerExists, isNeverInlinePrag
60 import TyCon ( tyConFamilySize )
61 import Type ( splitFunTy_maybe, isUnLiftedType )
62 import Unique ( Unique, buildIdKey, augmentIdKey )
63 import Maybes ( maybeToBool )
65 import List ( maximumBy )
66 import Util ( isIn, lengthExceeds )
69 #if __GLASGOW_HASKELL__ >= 404
70 import GlaExts ( fromInt )
75 %************************************************************************
77 \subsection{Making unfoldings}
79 %************************************************************************
82 mkTopUnfolding expr = mkUnfolding True {- Top level -} expr
84 mkUnfolding top_lvl expr
85 = CoreUnfolding (occurAnalyseGlobalExpr expr)
90 (calcUnfoldingGuidance opt_UF_CreationThreshold expr)
91 -- Sometimes during simplification, there's a large let-bound thing
92 -- which has been substituted, and so is now dead; so 'expr' contains
93 -- two copies of the thing while the occurrence-analysed expression doesn't
94 -- Nevertheless, we don't occ-analyse before computing the size because the
95 -- size computation bales out after a while, whereas occurrence analysis does not.
97 -- This can occasionally mean that the guidance is very pessimistic;
98 -- it gets fixed up next round
100 mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
101 = CompulsoryUnfolding (occurAnalyseGlobalExpr expr)
105 %************************************************************************
107 \subsection{The UnfoldingGuidance type}
109 %************************************************************************
112 instance Outputable UnfoldingGuidance where
113 ppr UnfoldNever = ptext SLIT("NEVER")
114 ppr (UnfoldIfGoodArgs v cs size discount)
115 = hsep [ ptext SLIT("IF_ARGS"), int v,
116 brackets (hsep (map int cs)),
123 calcUnfoldingGuidance
124 :: Int -- bomb out if size gets bigger than this
125 -> CoreExpr -- expression to look at
127 calcUnfoldingGuidance bOMB_OUT_SIZE expr
128 = case collect_val_bndrs expr of { (inline, val_binders, body) ->
130 n_val_binders = length val_binders
132 max_inline_size = n_val_binders+2
133 -- The idea is that if there is an INLINE pragma (inline is True)
134 -- and there's a big body, we give a size of n_val_binders+2. This
135 -- This is just enough to fail the no-size-increase test in callSiteInline,
136 -- so that INLINE things don't get inlined into entirely boring contexts,
140 case (sizeExpr bOMB_OUT_SIZE val_binders body) of
143 | not inline -> UnfoldNever
144 -- A big function with an INLINE pragma must
145 -- have an UnfoldIfGoodArgs guidance
146 | inline -> UnfoldIfGoodArgs n_val_binders
147 (map (const 0) val_binders)
150 SizeIs size cased_args scrut_discount
153 (map discount_for val_binders)
159 final_size | inline = boxed_size `min` max_inline_size
160 | otherwise = boxed_size
162 -- Sometimes an INLINE thing is smaller than n_val_binders+2.
163 -- A particular case in point is a constructor, which has size 1.
164 -- We want to inline this regardless, hence the `min`
166 discount_for b = foldlBag (\acc (b',n) -> if b==b' then acc+n else acc)
170 collect_val_bndrs e = go False [] e
171 -- We need to be a bit careful about how we collect the
172 -- value binders. In ptic, if we see
173 -- __inline_me (\x y -> e)
174 -- We want to say "2 value binders". Why? So that
175 -- we take account of information given for the arguments
177 go inline rev_vbs (Note InlineMe e) = go True rev_vbs e
178 go inline rev_vbs (Lam b e) | isId b = go inline (b:rev_vbs) e
179 | otherwise = go inline rev_vbs e
180 go inline rev_vbs e = (inline, reverse rev_vbs, e)
184 sizeExpr :: Int -- Bomb out if it gets bigger than this
185 -> [Id] -- Arguments; we're interested in which of these
190 sizeExpr (I# bOMB_OUT_SIZE) top_args expr
193 size_up (Type t) = sizeZero -- Types cost nothing
194 size_up (Var v) = sizeOne
196 size_up (Note _ body) = size_up body -- Notes cost nothing
198 size_up (App fun (Type t)) = size_up fun
199 size_up (App fun arg) = size_up_app fun [arg]
201 size_up (Lit lit) = sizeOne
203 size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
204 | otherwise = size_up e
206 size_up (Let (NonRec binder rhs) body)
207 = nukeScrutDiscount (size_up rhs) `addSize`
208 size_up body `addSizeN`
209 (if isUnLiftedType (idType binder) then 0 else 1)
210 -- For the allocation
211 -- If the binder has an unlifted type there is no allocation
213 size_up (Let (Rec pairs) body)
214 = nukeScrutDiscount rhs_size `addSize`
215 size_up body `addSizeN`
216 length pairs -- For the allocation
218 rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
220 -- We want to make wrapper-style evaluation look cheap, so that
221 -- when we inline a wrapper it doesn't make call site (much) bigger
222 -- Otherwise we get nasty phase ordering stuff:
225 -- If we inline g's wrapper, f looks big, and doesn't get inlined
226 -- into h; if we inline f first, while it looks small, then g's
227 -- wrapper will get inlined later anyway. To avoid this nasty
228 -- ordering difference, we make (case a of (x,y) -> ...) look free.
229 size_up (Case (Var v) _ [alt])
231 = size_up_alt alt `addSize` SizeIs 0# (unitBag (v, 1)) 0#
232 -- Good to inline if an arg is scrutinised, because
233 -- that may eliminate allocation in the caller
234 -- And it eliminates the case itself
238 -- Scrutinising one of the argument variables,
239 -- with more than one alternative
240 size_up (Case (Var v) _ alts)
242 = alts_size (foldr addSize sizeOne alt_sizes) -- The 1 is for the scrutinee
243 (foldr1 maxSize alt_sizes)
245 v_in_args = v `elem` top_args
246 alt_sizes = map size_up_alt alts
248 alts_size (SizeIs tot tot_disc tot_scrut) -- Size of all alternatives
249 (SizeIs max max_disc max_scrut) -- Size of biggest alternative
250 = SizeIs tot (unitBag (v, I# (1# +# tot -# max)) `unionBags` max_disc) max_scrut
251 -- If the variable is known, we produce a discount that
252 -- will take us back to 'max', the size of rh largest alternative
253 -- The 1+ is a little discount for reduced allocation in the caller
255 alts_size tot_size _ = tot_size
258 size_up (Case e _ alts) = nukeScrutDiscount (size_up e) `addSize`
259 foldr (addSize . size_up_alt) sizeZero alts
260 -- We don't charge for the case itself
261 -- It's a strict thing, and the price of the call
262 -- is paid by scrut. Also consider
263 -- case f x of DEFAULT -> e
264 -- This is just ';'! Don't charge for it.
267 size_up_app (App fun arg) args
268 | isTypeArg arg = size_up_app fun args
269 | otherwise = size_up_app fun (arg:args)
270 size_up_app fun args = foldr (addSize . nukeScrutDiscount . size_up)
271 (size_up_fun fun args)
274 -- A function application with at least one value argument
275 -- so if the function is an argument give it an arg-discount
277 -- Also behave specially if the function is a build
279 -- Also if the function is a constant Id (constr or primop)
280 -- compute discounts specially
281 size_up_fun (Var fun) args
282 | idUnique fun == buildIdKey = buildSize
283 | idUnique fun == augmentIdKey = augmentSize
285 = case idFlavour fun of
286 DataConId dc -> conSizeN (valArgCount args)
288 PrimOpId op -> primOpSize op (valArgCount args)
289 -- foldr addSize (primOpSize op) (map arg_discount args)
290 -- At one time I tried giving an arg-discount if a primop
291 -- is applied to one of the function's arguments, but it's
292 -- not good. At the moment, any unlifted-type arg gets a
293 -- 'True' for 'yes I'm evald', so we collect the discount even
294 -- if we know nothing about it. And just having it in a primop
295 -- doesn't help at all if we don't know something more.
297 other -> fun_discount fun `addSizeN`
298 (1 + length (filter (not . exprIsTrivial) args))
299 -- The 1+ is for the function itself
300 -- Add 1 for each non-trivial arg;
301 -- the allocation cost, as in let(rec)
302 -- Slight hack here: for constructors the args are almost always
303 -- trivial; and for primops they are almost always prim typed
304 -- We should really only count for non-prim-typed args in the
305 -- general case, but that seems too much like hard work
307 size_up_fun other args = size_up other
310 size_up_alt (con, bndrs, rhs) = size_up rhs
311 -- Don't charge for args, so that wrappers look cheap
314 -- We want to record if we're case'ing, or applying, an argument
315 fun_discount v | v `elem` top_args = SizeIs 0# (unitBag (v, opt_UF_FunAppDiscount)) 0#
316 fun_discount other = sizeZero
319 -- These addSize things have to be here because
320 -- I don't want to give them bOMB_OUT_SIZE as an argument
322 addSizeN TooBig _ = TooBig
323 addSizeN (SizeIs n xs d) (I# m)
324 | n_tot ># bOMB_OUT_SIZE = TooBig
325 | otherwise = SizeIs n_tot xs d
329 addSize TooBig _ = TooBig
330 addSize _ TooBig = TooBig
331 addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
332 | n_tot ># bOMB_OUT_SIZE = TooBig
333 | otherwise = SizeIs n_tot xys d_tot
337 xys = xs `unionBags` ys
340 Code for manipulating sizes
344 data ExprSize = TooBig
345 | SizeIs Int# -- Size found
346 (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
347 Int# -- Size to subtract if result is scrutinised
348 -- by a case expression
350 isTooBig TooBig = True
353 maxSize TooBig _ = TooBig
354 maxSize _ TooBig = TooBig
355 maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
358 sizeZero = SizeIs 0# emptyBag 0#
359 sizeOne = SizeIs 1# emptyBag 0#
360 sizeTwo = SizeIs 2# emptyBag 0#
361 sizeN (I# n) = SizeIs n emptyBag 0#
362 conSizeN (I# n) = SizeIs 1# emptyBag (n +# 1#)
363 -- Treat constructors as size 1; we are keen to expose them
364 -- (and we charge separately for their args). We can't treat
365 -- them as size zero, else we find that (I# x) has size 1,
366 -- which is the same as a lone variable; and hence 'v' will
367 -- always be replaced by (I# x), where v is bound to I# x.
370 | not (primOpIsDupable op) = sizeN opt_UF_DearOp
371 | not (primOpOutOfLine op) = sizeZero -- These are good to inline
372 | otherwise = sizeOne
374 buildSize = SizeIs (-2#) emptyBag 4#
375 -- We really want to inline applications of build
376 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
377 -- Indeed, we should add a result_discount becuause build is
378 -- very like a constructor. We don't bother to check that the
379 -- build is saturated (it usually is). The "-2" discounts for the \c n,
380 -- The "4" is rather arbitrary.
382 augmentSize = SizeIs (-2#) emptyBag 4#
383 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
384 -- e plus ys. The -2 accounts for the \cn
386 nukeScrutDiscount (SizeIs n vs d) = SizeIs n vs 0#
387 nukeScrutDiscount TooBig = TooBig
389 -- When we return a lambda, give a discount if it's used (applied)
390 lamScrutDiscount (SizeIs n vs d) = case opt_UF_FunAppDiscount of { I# d -> SizeIs n vs d }
391 lamScrutDiscount TooBig = TooBig
395 %************************************************************************
397 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
399 %************************************************************************
401 We have very limited information about an unfolding expression: (1)~so
402 many type arguments and so many value arguments expected---for our
403 purposes here, we assume we've got those. (2)~A ``size'' or ``cost,''
404 a single integer. (3)~An ``argument info'' vector. For this, what we
405 have at the moment is a Boolean per argument position that says, ``I
406 will look with great favour on an explicit constructor in this
407 position.'' (4)~The ``discount'' to subtract if the expression
408 is being scrutinised.
410 Assuming we have enough type- and value arguments (if not, we give up
411 immediately), then we see if the ``discounted size'' is below some
412 (semi-arbitrary) threshold. It works like this: for every argument
413 position where we're looking for a constructor AND WE HAVE ONE in our
414 hands, we get a (again, semi-arbitrary) discount [proportion to the
415 number of constructors in the type being scrutinized].
417 If we're in the context of a scrutinee ( \tr{(case <expr > of A .. -> ...;.. )})
418 and the expression in question will evaluate to a constructor, we use
419 the computed discount size *for the result only* rather than
420 computing the argument discounts. Since we know the result of
421 the expression is going to be taken apart, discounting its size
422 is more accurate (see @sizeExpr@ above for how this discount size
425 We use this one to avoid exporting inlinings that we ``couldn't possibly
426 use'' on the other side. Can be overridden w/ flaggery.
427 Just the same as smallEnoughToInline, except that it has no actual arguments.
430 couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
431 couldBeSmallEnoughToInline threshold rhs = case calcUnfoldingGuidance threshold rhs of
435 certainlyWillInline :: Id -> Bool
436 -- Sees if the Id is pretty certain to inline
437 certainlyWillInline v
438 = case idUnfolding v of
440 CoreUnfolding _ _ _ is_value _ g@(UnfoldIfGoodArgs n_vals _ size _)
442 && size - (n_vals +1) <= opt_UF_UseThreshold
447 @okToUnfoldInHifile@ is used when emitting unfolding info into an interface
448 file to determine whether an unfolding candidate really should be unfolded.
449 The predicate is needed to prevent @_casm_@s (+ lit-lits) from being emitted
450 into interface files.
452 The reason for inlining expressions containing _casm_s into interface files
453 is that these fragments of C are likely to mention functions/#defines that
454 will be out-of-scope when inlined into another module. This is not an
455 unfixable problem for the user (just need to -#include the approp. header
456 file), but turning it off seems to the simplest thing to do.
459 okToUnfoldInHiFile :: CoreExpr -> Bool
460 okToUnfoldInHiFile e = opt_UnfoldCasms || go e
462 -- Race over an expression looking for CCalls..
463 go (Var v) = case isPrimOpId_maybe v of
464 Just op -> okToUnfoldPrimOp op
466 go (Lit lit) = not (isLitLitLit lit)
467 go (App fun arg) = go fun && go arg
468 go (Lam _ body) = go body
469 go (Let binds body) = and (map go (body :rhssOfBind binds))
470 go (Case scrut bndr alts) = and (map go (scrut:rhssOfAlts alts))
471 go (Note _ body) = go body
474 -- ok to unfold a PrimOp as long as it's not a _casm_
475 okToUnfoldPrimOp (CCallOp ccall) = not (ccallIsCasm ccall)
476 okToUnfoldPrimOp _ = True
480 %************************************************************************
482 \subsection{callSiteInline}
484 %************************************************************************
486 This is the key function. It decides whether to inline a variable at a call site
488 callSiteInline is used at call sites, so it is a bit more generous.
489 It's a very important function that embodies lots of heuristics.
490 A non-WHNF can be inlined if it doesn't occur inside a lambda,
491 and occurs exactly once or
492 occurs once in each branch of a case and is small
494 If the thing is in WHNF, there's no danger of duplicating work,
495 so we can inline if it occurs once, or is small
497 NOTE: we don't want to inline top-level functions that always diverge.
498 It just makes the code bigger. Tt turns out that the convenient way to prevent
499 them inlining is to give them a NOINLINE pragma, which we do in
500 StrictAnal.addStrictnessInfoToTopId
503 callSiteInline :: Bool -- True <=> the Id is black listed
504 -> Bool -- 'inline' note at call site
507 -> [Bool] -- One for each value arg; True if it is interesting
508 -> Bool -- True <=> continuation is interesting
509 -> Maybe CoreExpr -- Unfolding, if any
512 callSiteInline black_listed inline_call occ id arg_infos interesting_cont
513 = case idUnfolding id of {
514 NoUnfolding -> Nothing ;
515 OtherCon cs -> Nothing ;
516 CompulsoryUnfolding unf_template | black_listed -> Nothing
517 | otherwise -> Just unf_template ;
518 -- Constructors have compulsory unfoldings, but
519 -- may have rules, in which case they are
520 -- black listed till later
521 CoreUnfolding unf_template is_top is_cheap is_value is_bot guidance ->
524 result | yes_or_no = Just unf_template
525 | otherwise = Nothing
527 n_val_args = length arg_infos
529 ok_inside_lam = is_value || is_bot || (is_cheap && not is_top)
530 -- I'm experimenting with is_cheap && not is_top
533 | black_listed = False
534 | otherwise = case occ of
535 IAmDead -> pprTrace "callSiteInline: dead" (ppr id) False
536 IAmALoopBreaker -> False
537 OneOcc in_lam one_br -> (not in_lam || ok_inside_lam) && consider_safe in_lam True one_br
538 NoOccInfo -> ok_inside_lam && consider_safe True False False
540 consider_safe in_lam once once_in_one_branch
541 -- consider_safe decides whether it's a good idea to inline something,
542 -- given that there's no work-duplication issue (the caller checks that).
543 -- once_in_one_branch = True means there's a unique textual occurrence
547 -- Be very keen to inline something if this is its unique occurrence:
549 -- a) Inlining gives a good chance of eliminating the original
550 -- binding (and hence the allocation) for the thing.
551 -- (Provided it's not a top level binding, in which case the
552 -- allocation costs nothing.)
554 -- b) Inlining a function that is called only once exposes the
555 -- body function to the call site.
557 -- The only time we hold back is when substituting inside a lambda;
558 -- then if the context is totally uninteresting (not applied, not scrutinised)
559 -- there is no point in substituting because it might just increase allocation,
560 -- by allocating the function itself many times
562 -- Note: there used to be a '&& not top_level' in the guard above,
563 -- but that stopped us inlining top-level functions used only once,
565 = not in_lam || not (null arg_infos) || interesting_cont
569 UnfoldNever -> False ;
570 UnfoldIfGoodArgs n_vals_wanted arg_discounts size res_discount
572 | enough_args && size <= (n_vals_wanted + 1)
574 -- Size of call is n_vals_wanted (+1 for the function)
578 -> some_benefit && small_enough
581 some_benefit = or arg_infos || really_interesting_cont ||
582 (not is_top && (once || (n_vals_wanted > 0 && enough_args)))
583 -- If it occurs more than once, there must be something interesting
584 -- about some argument, or the result context, to make it worth inlining
586 -- If a function has a nested defn we also record some-benefit,
587 -- on the grounds that we are often able to eliminate the binding,
588 -- and hence the allocation, for the function altogether; this is good
589 -- for join points. But this only makes sense for *functions*;
590 -- inlining a constructor doesn't help allocation unless the result is
591 -- scrutinised. UNLESS the constructor occurs just once, albeit possibly
592 -- in multiple case branches. Then inlining it doesn't increase allocation,
593 -- but it does increase the chance that the constructor won't be allocated at all
594 -- in the branches that don't use it.
596 enough_args = n_val_args >= n_vals_wanted
597 really_interesting_cont | n_val_args < n_vals_wanted = False -- Too few args
598 | n_val_args == n_vals_wanted = interesting_cont
599 | otherwise = True -- Extra args
600 -- really_interesting_cont tells if the result of the
601 -- call is in an interesting context.
603 small_enough = (size - discount) <= opt_UF_UseThreshold
604 discount = computeDiscount n_vals_wanted arg_discounts res_discount
605 arg_infos really_interesting_cont
609 if opt_D_dump_inlinings then
610 pprTrace "Considering inlining"
611 (ppr id <+> vcat [text "black listed" <+> ppr black_listed,
612 text "occ info:" <+> ppr occ,
613 text "arg infos" <+> ppr arg_infos,
614 text "interesting continuation" <+> ppr interesting_cont,
615 text "is value:" <+> ppr is_value,
616 text "is cheap:" <+> ppr is_cheap,
617 text "is bottom:" <+> ppr is_bot,
618 text "is top-level:" <+> ppr is_top,
619 text "guidance" <+> ppr guidance,
620 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO",
622 text "Unfolding =" <+> pprCoreExpr unf_template
630 computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Bool -> Int
631 computeDiscount n_vals_wanted arg_discounts res_discount arg_infos result_used
632 -- We multiple the raw discounts (args_discount and result_discount)
633 -- ty opt_UnfoldingKeenessFactor because the former have to do with
634 -- *size* whereas the discounts imply that there's some extra
635 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
638 -- we also discount 1 for each argument passed, because these will
639 -- reduce with the lambdas in the function (we count 1 for a lambda
641 = 1 + -- Discount of 1 because the result replaces the call
642 -- so we count 1 for the function itself
643 length (take n_vals_wanted arg_infos) +
644 -- Discount of 1 for each arg supplied, because the
645 -- result replaces the call
646 round (opt_UF_KeenessFactor *
647 fromInt (arg_discount + result_discount))
649 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
651 mk_arg_discount discount is_evald | is_evald = discount
654 -- Don't give a result discount unless there are enough args
655 result_discount | result_used = res_discount -- Over-applied, or case scrut
660 %************************************************************************
662 \subsection{Black-listing}
664 %************************************************************************
666 Inlining is controlled by the "Inline phase" number, which is set
667 by the per-simplification-pass '-finline-phase' flag.
669 For optimisation we use phase 1,2 and nothing (i.e. no -finline-phase flag)
670 in that order. The meanings of these are determined by the @blackListed@ function
673 The final simplification doesn't have a phase number.
679 (least black listing, most inlining)
680 INLINE n foo phase is Just p *and* p<n *and* foo appears on LHS of rule
681 INLINE foo phase is Just p *and* foo appears on LHS of rule
682 NOINLINE n foo phase is Just p *and* (p<n *or* foo appears on LHS of rule)
684 (most black listing, least inlining)
687 blackListed :: IdSet -- Used in transformation rules
688 -> Maybe Int -- Inline phase
689 -> Id -> Bool -- True <=> blacklisted
691 -- The blackListed function sees whether a variable should *not* be
692 -- inlined because of the inline phase we are in. This is the sole
693 -- place that the inline phase number is looked at.
695 blackListed rule_vars Nothing -- Last phase
696 = \v -> isNeverInlinePrag (idInlinePragma v)
698 blackListed rule_vars (Just phase)
699 = \v -> normal_case rule_vars phase v
701 normal_case rule_vars phase v
702 = case idInlinePragma v of
703 NoInlinePragInfo -> has_rules
705 IMustNotBeINLINEd from_INLINE Nothing
706 | from_INLINE -> has_rules -- Black list until final phase
707 | otherwise -> True -- Always blacklisted
709 IMustNotBeINLINEd from_inline (Just threshold)
710 | from_inline -> (phase < threshold && has_rules)
711 | otherwise -> (phase < threshold || has_rules)
713 has_rules = v `elemVarSet` rule_vars
714 || not (isEmptyCoreRules (idSpecialisation v))
718 SLPJ 95/04: Why @runST@ must be inlined very late:
722 (a, s') = newArray# 100 [] s
723 (_, s'') = fill_in_array_or_something a x s'
727 If we inline @runST@, we'll get:
730 (a, s') = newArray# 100 [] realWorld#{-NB-}
731 (_, s'') = fill_in_array_or_something a x s'
735 And now the @newArray#@ binding can be floated to become a CAF, which
736 is totally and utterly wrong:
739 (a, s') = newArray# 100 [] realWorld#{-NB-} -- YIKES!!!
742 let (_, s'') = fill_in_array_or_something a x s' in
745 All calls to @f@ will share a {\em single} array!
747 Yet we do want to inline runST sometime, so we can avoid
748 needless code. Solution: black list it until the last moment.