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_FunAppDiscount,
38 opt_UF_CheapOp, opt_UF_DearOp,
39 opt_UnfoldCasms, opt_PprStyle_Debug,
40 DynFlags, dopt_D_dump_inlinings
43 import PprCore ( pprCoreExpr )
44 import OccurAnal ( occurAnalyseGlobalExpr )
45 import CoreUtils ( exprIsValue, exprIsCheap, exprIsBottom, exprIsTrivial )
46 import Id ( Id, idType, idFlavour, isId, idWorkerInfo,
47 idSpecialisation, idInlinePragma, idUnfolding,
51 import Literal ( isLitLitLit, litIsDupable )
52 import PrimOp ( PrimOp(..), primOpIsDupable, primOpOutOfLine, ccallIsCasm )
53 import IdInfo ( ArityInfo(..), InlinePragInfo(..), OccInfo(..), IdFlavour(..), CprInfo(..),
54 insideLam, workerExists, isNeverInlinePrag
56 import Type ( splitFunTy_maybe, isUnLiftedType )
57 import PrelNames ( hasKey, buildIdKey, augmentIdKey )
61 #if __GLASGOW_HASKELL__ >= 404
62 import GlaExts ( fromInt )
67 %************************************************************************
69 \subsection{Making unfoldings}
71 %************************************************************************
74 mkTopUnfolding expr = mkUnfolding True {- Top level -} expr
76 mkUnfolding top_lvl expr
77 = CoreUnfolding (occurAnalyseGlobalExpr expr)
83 -- OK to inline inside a lambda
85 (calcUnfoldingGuidance opt_UF_CreationThreshold expr)
86 -- Sometimes during simplification, there's a large let-bound thing
87 -- which has been substituted, and so is now dead; so 'expr' contains
88 -- two copies of the thing while the occurrence-analysed expression doesn't
89 -- Nevertheless, we don't occ-analyse before computing the size because the
90 -- size computation bales out after a while, whereas occurrence analysis does not.
92 -- This can occasionally mean that the guidance is very pessimistic;
93 -- it gets fixed up next round
95 mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
96 = CompulsoryUnfolding (occurAnalyseGlobalExpr expr)
100 %************************************************************************
102 \subsection{The UnfoldingGuidance type}
104 %************************************************************************
107 instance Outputable UnfoldingGuidance where
108 ppr UnfoldNever = ptext SLIT("NEVER")
109 ppr (UnfoldIfGoodArgs v cs size discount)
110 = hsep [ ptext SLIT("IF_ARGS"), int v,
111 brackets (hsep (map int cs)),
118 calcUnfoldingGuidance
119 :: Int -- bomb out if size gets bigger than this
120 -> CoreExpr -- expression to look at
122 calcUnfoldingGuidance bOMB_OUT_SIZE expr
123 = case collect_val_bndrs expr of { (inline, val_binders, body) ->
125 n_val_binders = length val_binders
127 max_inline_size = n_val_binders+2
128 -- The idea is that if there is an INLINE pragma (inline is True)
129 -- and there's a big body, we give a size of n_val_binders+2. This
130 -- This is just enough to fail the no-size-increase test in callSiteInline,
131 -- so that INLINE things don't get inlined into entirely boring contexts,
135 case (sizeExpr bOMB_OUT_SIZE val_binders body) of
138 | not inline -> UnfoldNever
139 -- A big function with an INLINE pragma must
140 -- have an UnfoldIfGoodArgs guidance
141 | inline -> UnfoldIfGoodArgs n_val_binders
142 (map (const 0) val_binders)
145 SizeIs size cased_args scrut_discount
148 (map discount_for val_binders)
154 final_size | inline = boxed_size `min` max_inline_size
155 | otherwise = boxed_size
157 -- Sometimes an INLINE thing is smaller than n_val_binders+2.
158 -- A particular case in point is a constructor, which has size 1.
159 -- We want to inline this regardless, hence the `min`
161 discount_for b = foldlBag (\acc (b',n) -> if b==b' then acc+n else acc)
165 collect_val_bndrs e = go False [] e
166 -- We need to be a bit careful about how we collect the
167 -- value binders. In ptic, if we see
168 -- __inline_me (\x y -> e)
169 -- We want to say "2 value binders". Why? So that
170 -- we take account of information given for the arguments
172 go inline rev_vbs (Note InlineMe e) = go True rev_vbs e
173 go inline rev_vbs (Lam b e) | isId b = go inline (b:rev_vbs) e
174 | otherwise = go inline rev_vbs e
175 go inline rev_vbs e = (inline, reverse rev_vbs, e)
179 sizeExpr :: Int -- Bomb out if it gets bigger than this
180 -> [Id] -- Arguments; we're interested in which of these
185 sizeExpr (I# bOMB_OUT_SIZE) top_args expr
188 size_up (Type t) = sizeZero -- Types cost nothing
189 size_up (Var v) = sizeOne
191 size_up (Note _ body) = size_up body -- Notes cost nothing
193 size_up (App fun (Type t)) = size_up fun
194 size_up (App fun arg) = size_up_app fun [arg]
196 size_up (Lit lit) | litIsDupable lit = sizeOne
197 | otherwise = sizeN opt_UF_DearOp -- For lack of anything better
199 size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
200 | otherwise = size_up e
202 size_up (Let (NonRec binder rhs) body)
203 = nukeScrutDiscount (size_up rhs) `addSize`
204 size_up body `addSizeN`
205 (if isUnLiftedType (idType binder) then 0 else 1)
206 -- For the allocation
207 -- If the binder has an unlifted type there is no allocation
209 size_up (Let (Rec pairs) body)
210 = nukeScrutDiscount rhs_size `addSize`
211 size_up body `addSizeN`
212 length pairs -- For the allocation
214 rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
216 size_up (Case (Var v) _ alts)
217 | v `elem` top_args -- We are scrutinising an argument variable
219 {- I'm nuking this special case; BUT see the comment with case alternatives.
221 (a) It's too eager. We don't want to inline a wrapper into a
222 context with no benefit.
223 E.g. \ x. f (x+x) o point in inlining (+) here!
225 (b) It's ineffective. Once g's wrapper is inlined, its case-expressions
226 aren't scrutinising arguments any more
230 [alt] -> size_up_alt alt `addSize` SizeIs 0# (unitBag (v, 1)) 0#
231 -- We want to make wrapper-style evaluation look cheap, so that
232 -- when we inline a wrapper it doesn't make call site (much) bigger
233 -- Otherwise we get nasty phase ordering stuff:
236 -- If we inline g's wrapper, f looks big, and doesn't get inlined
237 -- into h; if we inline f first, while it looks small, then g's
238 -- wrapper will get inlined later anyway. To avoid this nasty
239 -- ordering difference, we make (case a of (x,y) -> ...),
240 -- *where a is one of the arguments* look free.
244 alts_size (foldr addSize sizeOne alt_sizes) -- The 1 is for the scrutinee
245 (foldr1 maxSize alt_sizes)
247 -- Good to inline if an arg is scrutinised, because
248 -- that may eliminate allocation in the caller
249 -- And it eliminates the case itself
252 alt_sizes = map size_up_alt alts
254 -- alts_size tries to compute a good discount for
255 -- the case when we are scrutinising an argument variable
256 alts_size (SizeIs tot tot_disc tot_scrut) -- Size of all alternatives
257 (SizeIs max max_disc max_scrut) -- Size of biggest alternative
258 = SizeIs tot (unitBag (v, I# (1# +# tot -# max)) `unionBags` max_disc) max_scrut
259 -- If the variable is known, we produce a discount that
260 -- will take us back to 'max', the size of rh largest alternative
261 -- The 1+ is a little discount for reduced allocation in the caller
262 alts_size tot_size _ = tot_size
265 size_up (Case e _ alts) = nukeScrutDiscount (size_up e) `addSize`
266 foldr (addSize . size_up_alt) sizeZero alts
267 -- We don't charge for the case itself
268 -- It's a strict thing, and the price of the call
269 -- is paid by scrut. Also consider
270 -- case f x of DEFAULT -> e
271 -- This is just ';'! Don't charge for it.
274 size_up_app (App fun arg) args
275 | isTypeArg arg = size_up_app fun args
276 | otherwise = size_up_app fun (arg:args)
277 size_up_app fun args = foldr (addSize . nukeScrutDiscount . size_up)
278 (size_up_fun fun args)
281 -- A function application with at least one value argument
282 -- so if the function is an argument give it an arg-discount
284 -- Also behave specially if the function is a build
286 -- Also if the function is a constant Id (constr or primop)
287 -- compute discounts specially
288 size_up_fun (Var fun) args
289 | fun `hasKey` buildIdKey = buildSize
290 | fun `hasKey` augmentIdKey = augmentSize
292 = case idFlavour fun of
293 DataConId dc -> conSizeN (valArgCount args)
295 PrimOpId op -> primOpSize op (valArgCount args)
296 -- foldr addSize (primOpSize op) (map arg_discount args)
297 -- At one time I tried giving an arg-discount if a primop
298 -- is applied to one of the function's arguments, but it's
299 -- not good. At the moment, any unlifted-type arg gets a
300 -- 'True' for 'yes I'm evald', so we collect the discount even
301 -- if we know nothing about it. And just having it in a primop
302 -- doesn't help at all if we don't know something more.
304 other -> fun_discount fun `addSizeN`
305 (1 + length (filter (not . exprIsTrivial) args))
306 -- The 1+ is for the function itself
307 -- Add 1 for each non-trivial arg;
308 -- the allocation cost, as in let(rec)
309 -- Slight hack here: for constructors the args are almost always
310 -- trivial; and for primops they are almost always prim typed
311 -- We should really only count for non-prim-typed args in the
312 -- general case, but that seems too much like hard work
314 size_up_fun other args = size_up other
317 size_up_alt (con, bndrs, rhs) = size_up rhs
318 -- Don't charge for args, so that wrappers look cheap
319 -- (See comments about wrappers with Case)
322 -- We want to record if we're case'ing, or applying, an argument
323 fun_discount v | v `elem` top_args = SizeIs 0# (unitBag (v, opt_UF_FunAppDiscount)) 0#
324 fun_discount other = sizeZero
327 -- These addSize things have to be here because
328 -- I don't want to give them bOMB_OUT_SIZE as an argument
330 addSizeN TooBig _ = TooBig
331 addSizeN (SizeIs n xs d) (I# m)
332 | n_tot ># bOMB_OUT_SIZE = TooBig
333 | otherwise = SizeIs n_tot xs d
337 addSize TooBig _ = TooBig
338 addSize _ TooBig = TooBig
339 addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
340 | n_tot ># bOMB_OUT_SIZE = TooBig
341 | otherwise = SizeIs n_tot xys d_tot
345 xys = xs `unionBags` ys
348 Code for manipulating sizes
352 data ExprSize = TooBig
353 | SizeIs Int# -- Size found
354 (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
355 Int# -- Size to subtract if result is scrutinised
356 -- by a case expression
358 isTooBig TooBig = True
361 maxSize TooBig _ = TooBig
362 maxSize _ TooBig = TooBig
363 maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
366 sizeZero = SizeIs 0# emptyBag 0#
367 sizeOne = SizeIs 1# emptyBag 0#
368 sizeTwo = SizeIs 2# emptyBag 0#
369 sizeN (I# n) = SizeIs n emptyBag 0#
370 conSizeN (I# n) = SizeIs 1# emptyBag (n +# 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 (I# x) has size 1,
374 -- which is the same as a lone variable; and hence 'v' will
375 -- always be replaced by (I# x), where v is bound to I# x.
378 | not (primOpIsDupable op) = sizeN opt_UF_DearOp
379 | not (primOpOutOfLine op) = sizeZero -- These are good to inline
380 | otherwise = sizeOne
382 buildSize = SizeIs (-2#) emptyBag 4#
383 -- We really want to inline applications of build
384 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
385 -- Indeed, we should add a result_discount becuause build is
386 -- very like a constructor. We don't bother to check that the
387 -- build is saturated (it usually is). The "-2" discounts for the \c n,
388 -- The "4" is rather arbitrary.
390 augmentSize = SizeIs (-2#) emptyBag 4#
391 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
392 -- e plus ys. The -2 accounts for the \cn
394 nukeScrutDiscount (SizeIs n vs d) = SizeIs n vs 0#
395 nukeScrutDiscount TooBig = TooBig
397 -- When we return a lambda, give a discount if it's used (applied)
398 lamScrutDiscount (SizeIs n vs d) = case opt_UF_FunAppDiscount of { I# d -> SizeIs n vs d }
399 lamScrutDiscount TooBig = TooBig
403 %************************************************************************
405 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
407 %************************************************************************
409 We have very limited information about an unfolding expression: (1)~so
410 many type arguments and so many value arguments expected---for our
411 purposes here, we assume we've got those. (2)~A ``size'' or ``cost,''
412 a single integer. (3)~An ``argument info'' vector. For this, what we
413 have at the moment is a Boolean per argument position that says, ``I
414 will look with great favour on an explicit constructor in this
415 position.'' (4)~The ``discount'' to subtract if the expression
416 is being scrutinised.
418 Assuming we have enough type- and value arguments (if not, we give up
419 immediately), then we see if the ``discounted size'' is below some
420 (semi-arbitrary) threshold. It works like this: for every argument
421 position where we're looking for a constructor AND WE HAVE ONE in our
422 hands, we get a (again, semi-arbitrary) discount [proportion to the
423 number of constructors in the type being scrutinized].
425 If we're in the context of a scrutinee ( \tr{(case <expr > of A .. -> ...;.. )})
426 and the expression in question will evaluate to a constructor, we use
427 the computed discount size *for the result only* rather than
428 computing the argument discounts. Since we know the result of
429 the expression is going to be taken apart, discounting its size
430 is more accurate (see @sizeExpr@ above for how this discount size
433 We use this one to avoid exporting inlinings that we ``couldn't possibly
434 use'' on the other side. Can be overridden w/ flaggery.
435 Just the same as smallEnoughToInline, except that it has no actual arguments.
438 couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
439 couldBeSmallEnoughToInline threshold rhs = case calcUnfoldingGuidance threshold rhs of
443 certainlyWillInline :: Id -> Bool
444 -- Sees if the Id is pretty certain to inline
445 certainlyWillInline v
446 = case idUnfolding v of
448 CoreUnfolding _ _ is_value _ g@(UnfoldIfGoodArgs n_vals _ size _)
450 && size - (n_vals +1) <= opt_UF_UseThreshold
455 @okToUnfoldInHifile@ is used when emitting unfolding info into an interface
456 file to determine whether an unfolding candidate really should be unfolded.
457 The predicate is needed to prevent @_casm_@s (+ lit-lits) from being emitted
458 into interface files.
460 The reason for inlining expressions containing _casm_s into interface files
461 is that these fragments of C are likely to mention functions/#defines that
462 will be out-of-scope when inlined into another module. This is not an
463 unfixable problem for the user (just need to -#include the approp. header
464 file), but turning it off seems to the simplest thing to do.
467 okToUnfoldInHiFile :: CoreExpr -> Bool
468 okToUnfoldInHiFile e = opt_UnfoldCasms || go e
470 -- Race over an expression looking for CCalls..
471 go (Var v) = case isPrimOpId_maybe v of
472 Just op -> okToUnfoldPrimOp op
474 go (Lit lit) = not (isLitLitLit lit)
475 go (App fun arg) = go fun && go arg
476 go (Lam _ body) = go body
477 go (Let binds body) = and (map go (body :rhssOfBind binds))
478 go (Case scrut bndr alts) = and (map go (scrut:rhssOfAlts alts)) &&
479 not (any isLitLitLit [ lit | (LitAlt lit, _, _) <- alts ])
480 go (Note _ body) = go body
483 -- ok to unfold a PrimOp as long as it's not a _casm_
484 okToUnfoldPrimOp (CCallOp ccall) = not (ccallIsCasm ccall)
485 okToUnfoldPrimOp _ = True
489 %************************************************************************
491 \subsection{callSiteInline}
493 %************************************************************************
495 This is the key function. It decides whether to inline a variable at a call site
497 callSiteInline is used at call sites, so it is a bit more generous.
498 It's a very important function that embodies lots of heuristics.
499 A non-WHNF can be inlined if it doesn't occur inside a lambda,
500 and occurs exactly once or
501 occurs once in each branch of a case and is small
503 If the thing is in WHNF, there's no danger of duplicating work,
504 so we can inline if it occurs once, or is small
506 NOTE: we don't want to inline top-level functions that always diverge.
507 It just makes the code bigger. Tt turns out that the convenient way to prevent
508 them inlining is to give them a NOINLINE pragma, which we do in
509 StrictAnal.addStrictnessInfoToTopId
512 callSiteInline :: DynFlags
513 -> Bool -- True <=> the Id is black listed
514 -> Bool -- 'inline' note at call site
517 -> [Bool] -- One for each value arg; True if it is interesting
518 -> Bool -- True <=> continuation is interesting
519 -> Maybe CoreExpr -- Unfolding, if any
522 callSiteInline dflags black_listed inline_call occ id arg_infos interesting_cont
523 = case idUnfolding id of {
524 NoUnfolding -> Nothing ;
525 OtherCon cs -> Nothing ;
526 CompulsoryUnfolding unf_template | black_listed -> Nothing
527 | otherwise -> Just unf_template ;
528 -- Constructors have compulsory unfoldings, but
529 -- may have rules, in which case they are
530 -- black listed till later
531 CoreUnfolding unf_template is_top is_value is_cheap guidance ->
534 result | yes_or_no = Just unf_template
535 | otherwise = Nothing
537 n_val_args = length arg_infos
540 | black_listed = False
541 | otherwise = case occ of
542 IAmDead -> pprTrace "callSiteInline: dead" (ppr id) False
543 IAmALoopBreaker -> False
544 OneOcc in_lam one_br -> (not in_lam || is_cheap) && consider_safe in_lam True one_br
545 NoOccInfo -> is_cheap && consider_safe True False False
547 consider_safe in_lam once once_in_one_branch
548 -- consider_safe decides whether it's a good idea to inline something,
549 -- given that there's no work-duplication issue (the caller checks that).
550 -- once_in_one_branch = True means there's a unique textual occurrence
554 -- Be very keen to inline something if this is its unique occurrence:
556 -- a) Inlining gives a good chance of eliminating the original
557 -- binding (and hence the allocation) for the thing.
558 -- (Provided it's not a top level binding, in which case the
559 -- allocation costs nothing.)
561 -- b) Inlining a function that is called only once exposes the
562 -- body function to the call site.
564 -- The only time we hold back is when substituting inside a lambda;
565 -- then if the context is totally uninteresting (not applied, not scrutinised)
566 -- there is no point in substituting because it might just increase allocation,
567 -- by allocating the function itself many times
569 -- Note: there used to be a '&& not top_level' in the guard above,
570 -- but that stopped us inlining top-level functions used only once,
572 = not in_lam || not (null arg_infos) || interesting_cont
576 UnfoldNever -> False ;
577 UnfoldIfGoodArgs n_vals_wanted arg_discounts size res_discount
579 | enough_args && size <= (n_vals_wanted + 1)
581 -- Size of call is n_vals_wanted (+1 for the function)
585 -> some_benefit && small_enough
588 some_benefit = or arg_infos || really_interesting_cont ||
589 (not is_top && (once || (n_vals_wanted > 0 && enough_args)))
590 -- If it occurs more than once, there must be something interesting
591 -- about some argument, or the result context, to make it worth inlining
593 -- If a function has a nested defn we also record some-benefit,
594 -- on the grounds that we are often able to eliminate the binding,
595 -- and hence the allocation, for the function altogether; this is good
596 -- for join points. But this only makes sense for *functions*;
597 -- inlining a constructor doesn't help allocation unless the result is
598 -- scrutinised. UNLESS the constructor occurs just once, albeit possibly
599 -- in multiple case branches. Then inlining it doesn't increase allocation,
600 -- but it does increase the chance that the constructor won't be allocated at all
601 -- in the branches that don't use it.
603 enough_args = n_val_args >= n_vals_wanted
604 really_interesting_cont | n_val_args < n_vals_wanted = False -- Too few args
605 | n_val_args == n_vals_wanted = interesting_cont
606 | otherwise = True -- Extra args
607 -- really_interesting_cont tells if the result of the
608 -- call is in an interesting context.
610 small_enough = (size - discount) <= opt_UF_UseThreshold
611 discount = computeDiscount n_vals_wanted arg_discounts res_discount
612 arg_infos really_interesting_cont
616 if dopt_D_dump_inlinings dflags then
617 pprTrace "Considering inlining"
618 (ppr id <+> vcat [text "black listed:" <+> ppr black_listed,
619 text "occ info:" <+> ppr occ,
620 text "arg infos" <+> ppr arg_infos,
621 text "interesting continuation" <+> ppr interesting_cont,
622 text "is value:" <+> ppr is_value,
623 text "is cheap:" <+> ppr is_cheap,
624 text "guidance" <+> ppr guidance,
625 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO",
627 text "Unfolding =" <+> pprCoreExpr unf_template
635 computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Bool -> Int
636 computeDiscount n_vals_wanted arg_discounts res_discount arg_infos result_used
637 -- We multiple the raw discounts (args_discount and result_discount)
638 -- ty opt_UnfoldingKeenessFactor because the former have to do with
639 -- *size* whereas the discounts imply that there's some extra
640 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
643 -- we also discount 1 for each argument passed, because these will
644 -- reduce with the lambdas in the function (we count 1 for a lambda
646 = 1 + -- Discount of 1 because the result replaces the call
647 -- so we count 1 for the function itself
648 length (take n_vals_wanted arg_infos) +
649 -- Discount of 1 for each arg supplied, because the
650 -- result replaces the call
651 round (opt_UF_KeenessFactor *
652 fromInt (arg_discount + result_discount))
654 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
656 mk_arg_discount discount is_evald | is_evald = discount
659 -- Don't give a result discount unless there are enough args
660 result_discount | result_used = res_discount -- Over-applied, or case scrut
665 %************************************************************************
667 \subsection{Black-listing}
669 %************************************************************************
671 Inlining is controlled by the "Inline phase" number, which is set
672 by the per-simplification-pass '-finline-phase' flag.
674 For optimisation we use phase 1,2 and nothing (i.e. no -finline-phase flag)
675 in that order. The meanings of these are determined by the @blackListed@ function
678 The final simplification doesn't have a phase number.
684 (least black listing, most inlining)
685 INLINE n foo phase is Just p *and* p<n *and* foo appears on LHS of rule
686 INLINE foo phase is Just p *and* foo appears on LHS of rule
687 NOINLINE n foo phase is Just p *and* (p<n *or* foo appears on LHS of rule)
689 (most black listing, least inlining)
692 blackListed :: IdSet -- Used in transformation rules
693 -> Maybe Int -- Inline phase
694 -> Id -> Bool -- True <=> blacklisted
696 -- The blackListed function sees whether a variable should *not* be
697 -- inlined because of the inline phase we are in. This is the sole
698 -- place that the inline phase number is looked at.
700 blackListed rule_vars Nothing -- Last phase
701 = \v -> isNeverInlinePrag (idInlinePragma v)
703 blackListed rule_vars (Just phase)
704 = \v -> normal_case rule_vars phase v
706 normal_case rule_vars phase v
707 = case idInlinePragma v of
708 NoInlinePragInfo -> has_rules
710 IMustNotBeINLINEd from_INLINE Nothing
711 | from_INLINE -> has_rules -- Black list until final phase
712 | otherwise -> True -- Always blacklisted
714 IMustNotBeINLINEd from_INLINE (Just threshold)
715 | from_INLINE -> (phase < threshold && has_rules)
716 | otherwise -> (phase < threshold || has_rules)
718 has_rules = v `elemVarSet` rule_vars
719 || not (isEmptyCoreRules (idSpecialisation v))
723 SLPJ 95/04: Why @runST@ must be inlined very late:
727 (a, s') = newArray# 100 [] s
728 (_, s'') = fill_in_array_or_something a x s'
732 If we inline @runST@, we'll get:
735 (a, s') = newArray# 100 [] realWorld#{-NB-}
736 (_, s'') = fill_in_array_or_something a x s'
740 And now the @newArray#@ binding can be floated to become a CAF, which
741 is totally and utterly wrong:
744 (a, s') = newArray# 100 [] realWorld#{-NB-} -- YIKES!!!
747 let (_, s'') = fill_in_array_or_something a x s' in
750 All calls to @f@ will share a {\em single} array!
752 Yet we do want to inline runST sometime, so we can avoid
753 needless code. Solution: black list it until the last moment.