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, 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 Type ( splitFunTy_maybe, isUnLiftedType )
61 import Unique ( Unique, buildIdKey, augmentIdKey, hasKey )
62 import Maybes ( maybeToBool )
64 import List ( maximumBy )
65 import Util ( isIn, lengthExceeds )
68 #if __GLASGOW_HASKELL__ >= 404
69 import GlaExts ( fromInt )
74 %************************************************************************
76 \subsection{Making unfoldings}
78 %************************************************************************
81 mkTopUnfolding expr = mkUnfolding True {- Top level -} expr
83 mkUnfolding top_lvl expr
84 = CoreUnfolding (occurAnalyseGlobalExpr expr)
89 (calcUnfoldingGuidance opt_UF_CreationThreshold expr)
90 -- Sometimes during simplification, there's a large let-bound thing
91 -- which has been substituted, and so is now dead; so 'expr' contains
92 -- two copies of the thing while the occurrence-analysed expression doesn't
93 -- Nevertheless, we don't occ-analyse before computing the size because the
94 -- size computation bales out after a while, whereas occurrence analysis does not.
96 -- This can occasionally mean that the guidance is very pessimistic;
97 -- it gets fixed up next round
99 mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
100 = CompulsoryUnfolding (occurAnalyseGlobalExpr expr)
104 %************************************************************************
106 \subsection{The UnfoldingGuidance type}
108 %************************************************************************
111 instance Outputable UnfoldingGuidance where
112 ppr UnfoldNever = ptext SLIT("NEVER")
113 ppr (UnfoldIfGoodArgs v cs size discount)
114 = hsep [ ptext SLIT("IF_ARGS"), int v,
115 brackets (hsep (map int cs)),
122 calcUnfoldingGuidance
123 :: Int -- bomb out if size gets bigger than this
124 -> CoreExpr -- expression to look at
126 calcUnfoldingGuidance bOMB_OUT_SIZE expr
127 = case collect_val_bndrs expr of { (inline, val_binders, body) ->
129 n_val_binders = length val_binders
131 max_inline_size = n_val_binders+2
132 -- The idea is that if there is an INLINE pragma (inline is True)
133 -- and there's a big body, we give a size of n_val_binders+2. This
134 -- This is just enough to fail the no-size-increase test in callSiteInline,
135 -- so that INLINE things don't get inlined into entirely boring contexts,
139 case (sizeExpr bOMB_OUT_SIZE val_binders body) of
142 | not inline -> UnfoldNever
143 -- A big function with an INLINE pragma must
144 -- have an UnfoldIfGoodArgs guidance
145 | inline -> UnfoldIfGoodArgs n_val_binders
146 (map (const 0) val_binders)
149 SizeIs size cased_args scrut_discount
152 (map discount_for val_binders)
158 final_size | inline = boxed_size `min` max_inline_size
159 | otherwise = boxed_size
161 -- Sometimes an INLINE thing is smaller than n_val_binders+2.
162 -- A particular case in point is a constructor, which has size 1.
163 -- We want to inline this regardless, hence the `min`
165 discount_for b = foldlBag (\acc (b',n) -> if b==b' then acc+n else acc)
169 collect_val_bndrs e = go False [] e
170 -- We need to be a bit careful about how we collect the
171 -- value binders. In ptic, if we see
172 -- __inline_me (\x y -> e)
173 -- We want to say "2 value binders". Why? So that
174 -- we take account of information given for the arguments
176 go inline rev_vbs (Note InlineMe e) = go True rev_vbs e
177 go inline rev_vbs (Lam b e) | isId b = go inline (b:rev_vbs) e
178 | otherwise = go inline rev_vbs e
179 go inline rev_vbs e = (inline, reverse rev_vbs, e)
183 sizeExpr :: Int -- Bomb out if it gets bigger than this
184 -> [Id] -- Arguments; we're interested in which of these
189 sizeExpr (I# bOMB_OUT_SIZE) top_args expr
192 size_up (Type t) = sizeZero -- Types cost nothing
193 size_up (Var v) = sizeOne
195 size_up (Note _ body) = size_up body -- Notes cost nothing
197 size_up (App fun (Type t)) = size_up fun
198 size_up (App fun arg) = size_up_app fun [arg]
200 size_up (Lit lit) = sizeOne
202 size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
203 | otherwise = size_up e
205 size_up (Let (NonRec binder rhs) body)
206 = nukeScrutDiscount (size_up rhs) `addSize`
207 size_up body `addSizeN`
208 (if isUnLiftedType (idType binder) then 0 else 1)
209 -- For the allocation
210 -- If the binder has an unlifted type there is no allocation
212 size_up (Let (Rec pairs) body)
213 = nukeScrutDiscount rhs_size `addSize`
214 size_up body `addSizeN`
215 length pairs -- For the allocation
217 rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
219 -- We want to make wrapper-style evaluation look cheap, so that
220 -- when we inline a wrapper it doesn't make call site (much) bigger
221 -- Otherwise we get nasty phase ordering stuff:
224 -- If we inline g's wrapper, f looks big, and doesn't get inlined
225 -- into h; if we inline f first, while it looks small, then g's
226 -- wrapper will get inlined later anyway. To avoid this nasty
227 -- ordering difference, we make (case a of (x,y) -> ...) look free.
228 size_up (Case (Var v) _ [alt])
230 = size_up_alt alt `addSize` SizeIs 0# (unitBag (v, 1)) 0#
231 -- Good to inline if an arg is scrutinised, because
232 -- that may eliminate allocation in the caller
233 -- And it eliminates the case itself
237 -- Scrutinising one of the argument variables,
238 -- with more than one alternative
239 size_up (Case (Var v) _ alts)
241 = alts_size (foldr addSize sizeOne alt_sizes) -- The 1 is for the scrutinee
242 (foldr1 maxSize alt_sizes)
244 v_in_args = v `elem` top_args
245 alt_sizes = map size_up_alt alts
247 alts_size (SizeIs tot tot_disc tot_scrut) -- Size of all alternatives
248 (SizeIs max max_disc max_scrut) -- Size of biggest alternative
249 = SizeIs tot (unitBag (v, I# (1# +# tot -# max)) `unionBags` max_disc) max_scrut
250 -- If the variable is known, we produce a discount that
251 -- will take us back to 'max', the size of rh largest alternative
252 -- The 1+ is a little discount for reduced allocation in the caller
254 alts_size tot_size _ = tot_size
257 size_up (Case e _ alts) = nukeScrutDiscount (size_up e) `addSize`
258 foldr (addSize . size_up_alt) sizeZero alts
259 -- We don't charge for the case itself
260 -- It's a strict thing, and the price of the call
261 -- is paid by scrut. Also consider
262 -- case f x of DEFAULT -> e
263 -- This is just ';'! Don't charge for it.
266 size_up_app (App fun arg) args
267 | isTypeArg arg = size_up_app fun args
268 | otherwise = size_up_app fun (arg:args)
269 size_up_app fun args = foldr (addSize . nukeScrutDiscount . size_up)
270 (size_up_fun fun args)
273 -- A function application with at least one value argument
274 -- so if the function is an argument give it an arg-discount
276 -- Also behave specially if the function is a build
278 -- Also if the function is a constant Id (constr or primop)
279 -- compute discounts specially
280 size_up_fun (Var fun) args
281 | fun `hasKey` buildIdKey = buildSize
282 | fun `hasKey` augmentIdKey = augmentSize
284 = case idFlavour fun of
285 DataConId dc -> conSizeN (valArgCount args)
287 PrimOpId op -> primOpSize op (valArgCount args)
288 -- foldr addSize (primOpSize op) (map arg_discount args)
289 -- At one time I tried giving an arg-discount if a primop
290 -- is applied to one of the function's arguments, but it's
291 -- not good. At the moment, any unlifted-type arg gets a
292 -- 'True' for 'yes I'm evald', so we collect the discount even
293 -- if we know nothing about it. And just having it in a primop
294 -- doesn't help at all if we don't know something more.
296 other -> fun_discount fun `addSizeN`
297 (1 + length (filter (not . exprIsTrivial) args))
298 -- The 1+ is for the function itself
299 -- Add 1 for each non-trivial arg;
300 -- the allocation cost, as in let(rec)
301 -- Slight hack here: for constructors the args are almost always
302 -- trivial; and for primops they are almost always prim typed
303 -- We should really only count for non-prim-typed args in the
304 -- general case, but that seems too much like hard work
306 size_up_fun other args = size_up other
309 size_up_alt (con, bndrs, rhs) = size_up rhs
310 -- Don't charge for args, so that wrappers look cheap
313 -- We want to record if we're case'ing, or applying, an argument
314 fun_discount v | v `elem` top_args = SizeIs 0# (unitBag (v, opt_UF_FunAppDiscount)) 0#
315 fun_discount other = sizeZero
318 -- These addSize things have to be here because
319 -- I don't want to give them bOMB_OUT_SIZE as an argument
321 addSizeN TooBig _ = TooBig
322 addSizeN (SizeIs n xs d) (I# m)
323 | n_tot ># bOMB_OUT_SIZE = TooBig
324 | otherwise = SizeIs n_tot xs d
328 addSize TooBig _ = TooBig
329 addSize _ TooBig = TooBig
330 addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
331 | n_tot ># bOMB_OUT_SIZE = TooBig
332 | otherwise = SizeIs n_tot xys d_tot
336 xys = xs `unionBags` ys
339 Code for manipulating sizes
343 data ExprSize = TooBig
344 | SizeIs Int# -- Size found
345 (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
346 Int# -- Size to subtract if result is scrutinised
347 -- by a case expression
349 isTooBig TooBig = True
352 maxSize TooBig _ = TooBig
353 maxSize _ TooBig = TooBig
354 maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
357 sizeZero = SizeIs 0# emptyBag 0#
358 sizeOne = SizeIs 1# emptyBag 0#
359 sizeTwo = SizeIs 2# emptyBag 0#
360 sizeN (I# n) = SizeIs n emptyBag 0#
361 conSizeN (I# n) = SizeIs 1# emptyBag (n +# 1#)
362 -- Treat constructors as size 1; we are keen to expose them
363 -- (and we charge separately for their args). We can't treat
364 -- them as size zero, else we find that (I# x) has size 1,
365 -- which is the same as a lone variable; and hence 'v' will
366 -- always be replaced by (I# x), where v is bound to I# x.
369 | not (primOpIsDupable op) = sizeN opt_UF_DearOp
370 | not (primOpOutOfLine op) = sizeZero -- These are good to inline
371 | otherwise = sizeOne
373 buildSize = SizeIs (-2#) emptyBag 4#
374 -- We really want to inline applications of build
375 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
376 -- Indeed, we should add a result_discount becuause build is
377 -- very like a constructor. We don't bother to check that the
378 -- build is saturated (it usually is). The "-2" discounts for the \c n,
379 -- The "4" is rather arbitrary.
381 augmentSize = SizeIs (-2#) emptyBag 4#
382 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
383 -- e plus ys. The -2 accounts for the \cn
385 nukeScrutDiscount (SizeIs n vs d) = SizeIs n vs 0#
386 nukeScrutDiscount TooBig = TooBig
388 -- When we return a lambda, give a discount if it's used (applied)
389 lamScrutDiscount (SizeIs n vs d) = case opt_UF_FunAppDiscount of { I# d -> SizeIs n vs d }
390 lamScrutDiscount TooBig = TooBig
394 %************************************************************************
396 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
398 %************************************************************************
400 We have very limited information about an unfolding expression: (1)~so
401 many type arguments and so many value arguments expected---for our
402 purposes here, we assume we've got those. (2)~A ``size'' or ``cost,''
403 a single integer. (3)~An ``argument info'' vector. For this, what we
404 have at the moment is a Boolean per argument position that says, ``I
405 will look with great favour on an explicit constructor in this
406 position.'' (4)~The ``discount'' to subtract if the expression
407 is being scrutinised.
409 Assuming we have enough type- and value arguments (if not, we give up
410 immediately), then we see if the ``discounted size'' is below some
411 (semi-arbitrary) threshold. It works like this: for every argument
412 position where we're looking for a constructor AND WE HAVE ONE in our
413 hands, we get a (again, semi-arbitrary) discount [proportion to the
414 number of constructors in the type being scrutinized].
416 If we're in the context of a scrutinee ( \tr{(case <expr > of A .. -> ...;.. )})
417 and the expression in question will evaluate to a constructor, we use
418 the computed discount size *for the result only* rather than
419 computing the argument discounts. Since we know the result of
420 the expression is going to be taken apart, discounting its size
421 is more accurate (see @sizeExpr@ above for how this discount size
424 We use this one to avoid exporting inlinings that we ``couldn't possibly
425 use'' on the other side. Can be overridden w/ flaggery.
426 Just the same as smallEnoughToInline, except that it has no actual arguments.
429 couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
430 couldBeSmallEnoughToInline threshold rhs = case calcUnfoldingGuidance threshold rhs of
434 certainlyWillInline :: Id -> Bool
435 -- Sees if the Id is pretty certain to inline
436 certainlyWillInline v
437 = case idUnfolding v of
439 CoreUnfolding _ _ _ is_value _ g@(UnfoldIfGoodArgs n_vals _ size _)
441 && size - (n_vals +1) <= opt_UF_UseThreshold
446 @okToUnfoldInHifile@ is used when emitting unfolding info into an interface
447 file to determine whether an unfolding candidate really should be unfolded.
448 The predicate is needed to prevent @_casm_@s (+ lit-lits) from being emitted
449 into interface files.
451 The reason for inlining expressions containing _casm_s into interface files
452 is that these fragments of C are likely to mention functions/#defines that
453 will be out-of-scope when inlined into another module. This is not an
454 unfixable problem for the user (just need to -#include the approp. header
455 file), but turning it off seems to the simplest thing to do.
458 okToUnfoldInHiFile :: CoreExpr -> Bool
459 okToUnfoldInHiFile e = opt_UnfoldCasms || go e
461 -- Race over an expression looking for CCalls..
462 go (Var v) = case isPrimOpId_maybe v of
463 Just op -> okToUnfoldPrimOp op
465 go (Lit lit) = not (isLitLitLit lit)
466 go (App fun arg) = go fun && go arg
467 go (Lam _ body) = go body
468 go (Let binds body) = and (map go (body :rhssOfBind binds))
469 go (Case scrut bndr alts) = and (map go (scrut:rhssOfAlts alts)) &&
470 not (any isLitLitLit [ lit | (LitAlt lit, _, _) <- 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.