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 )
47 import CoreUtils ( exprIsValue, exprIsCheap, exprIsBottom, exprIsTrivial )
48 import Id ( Id, idType, idFlavour, isId, idWorkerInfo,
49 idSpecialisation, idInlinePragma, idUnfolding,
53 import Literal ( isLitLitLit )
54 import PrimOp ( PrimOp(..), primOpIsDupable, primOpOutOfLine, ccallIsCasm )
55 import IdInfo ( ArityInfo(..), InlinePragInfo(..), OccInfo(..), IdFlavour(..), CprInfo(..),
56 insideLam, workerExists, isNeverInlinePrag
58 import Type ( splitFunTy_maybe, isUnLiftedType )
59 import Unique ( Unique, buildIdKey, augmentIdKey, hasKey )
63 #if __GLASGOW_HASKELL__ >= 404
64 import GlaExts ( fromInt )
69 %************************************************************************
71 \subsection{Making unfoldings}
73 %************************************************************************
76 mkTopUnfolding expr = mkUnfolding True {- Top level -} expr
78 mkUnfolding top_lvl expr
79 = CoreUnfolding (occurAnalyseGlobalExpr expr)
84 (calcUnfoldingGuidance opt_UF_CreationThreshold expr)
85 -- Sometimes during simplification, there's a large let-bound thing
86 -- which has been substituted, and so is now dead; so 'expr' contains
87 -- two copies of the thing while the occurrence-analysed expression doesn't
88 -- Nevertheless, we don't occ-analyse before computing the size because the
89 -- size computation bales out after a while, whereas occurrence analysis does not.
91 -- This can occasionally mean that the guidance is very pessimistic;
92 -- it gets fixed up next round
94 mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
95 = CompulsoryUnfolding (occurAnalyseGlobalExpr expr)
99 %************************************************************************
101 \subsection{The UnfoldingGuidance type}
103 %************************************************************************
106 instance Outputable UnfoldingGuidance where
107 ppr UnfoldNever = ptext SLIT("NEVER")
108 ppr (UnfoldIfGoodArgs v cs size discount)
109 = hsep [ ptext SLIT("IF_ARGS"), int v,
110 brackets (hsep (map int cs)),
117 calcUnfoldingGuidance
118 :: Int -- bomb out if size gets bigger than this
119 -> CoreExpr -- expression to look at
121 calcUnfoldingGuidance bOMB_OUT_SIZE expr
122 = case collect_val_bndrs expr of { (inline, val_binders, body) ->
124 n_val_binders = length val_binders
126 max_inline_size = n_val_binders+2
127 -- The idea is that if there is an INLINE pragma (inline is True)
128 -- and there's a big body, we give a size of n_val_binders+2. This
129 -- This is just enough to fail the no-size-increase test in callSiteInline,
130 -- so that INLINE things don't get inlined into entirely boring contexts,
134 case (sizeExpr bOMB_OUT_SIZE val_binders body) of
137 | not inline -> UnfoldNever
138 -- A big function with an INLINE pragma must
139 -- have an UnfoldIfGoodArgs guidance
140 | inline -> UnfoldIfGoodArgs n_val_binders
141 (map (const 0) val_binders)
144 SizeIs size cased_args scrut_discount
147 (map discount_for val_binders)
153 final_size | inline = boxed_size `min` max_inline_size
154 | otherwise = boxed_size
156 -- Sometimes an INLINE thing is smaller than n_val_binders+2.
157 -- A particular case in point is a constructor, which has size 1.
158 -- We want to inline this regardless, hence the `min`
160 discount_for b = foldlBag (\acc (b',n) -> if b==b' then acc+n else acc)
164 collect_val_bndrs e = go False [] e
165 -- We need to be a bit careful about how we collect the
166 -- value binders. In ptic, if we see
167 -- __inline_me (\x y -> e)
168 -- We want to say "2 value binders". Why? So that
169 -- we take account of information given for the arguments
171 go inline rev_vbs (Note InlineMe e) = go True rev_vbs e
172 go inline rev_vbs (Lam b e) | isId b = go inline (b:rev_vbs) e
173 | otherwise = go inline rev_vbs e
174 go inline rev_vbs e = (inline, reverse rev_vbs, e)
178 sizeExpr :: Int -- Bomb out if it gets bigger than this
179 -> [Id] -- Arguments; we're interested in which of these
184 sizeExpr (I# bOMB_OUT_SIZE) top_args expr
187 size_up (Type t) = sizeZero -- Types cost nothing
188 size_up (Var v) = sizeOne
190 size_up (Note _ body) = size_up body -- Notes cost nothing
192 size_up (App fun (Type t)) = size_up fun
193 size_up (App fun arg) = size_up_app fun [arg]
195 size_up (Lit lit) = sizeOne
197 size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
198 | otherwise = size_up e
200 size_up (Let (NonRec binder rhs) body)
201 = nukeScrutDiscount (size_up rhs) `addSize`
202 size_up body `addSizeN`
203 (if isUnLiftedType (idType binder) then 0 else 1)
204 -- For the allocation
205 -- If the binder has an unlifted type there is no allocation
207 size_up (Let (Rec pairs) body)
208 = nukeScrutDiscount rhs_size `addSize`
209 size_up body `addSizeN`
210 length pairs -- For the allocation
212 rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
214 -- We want to make wrapper-style evaluation look cheap, so that
215 -- when we inline a wrapper it doesn't make call site (much) bigger
216 -- Otherwise we get nasty phase ordering stuff:
219 -- If we inline g's wrapper, f looks big, and doesn't get inlined
220 -- into h; if we inline f first, while it looks small, then g's
221 -- wrapper will get inlined later anyway. To avoid this nasty
222 -- ordering difference, we make (case a of (x,y) -> ...) look free.
223 size_up (Case (Var v) _ [alt])
225 = size_up_alt alt `addSize` SizeIs 0# (unitBag (v, 1)) 0#
226 -- Good to inline if an arg is scrutinised, because
227 -- that may eliminate allocation in the caller
228 -- And it eliminates the case itself
232 -- Scrutinising one of the argument variables,
233 -- with more than one alternative
234 size_up (Case (Var v) _ alts)
236 = alts_size (foldr addSize sizeOne alt_sizes) -- The 1 is for the scrutinee
237 (foldr1 maxSize alt_sizes)
239 v_in_args = v `elem` top_args
240 alt_sizes = map size_up_alt alts
242 alts_size (SizeIs tot tot_disc tot_scrut) -- Size of all alternatives
243 (SizeIs max max_disc max_scrut) -- Size of biggest alternative
244 = SizeIs tot (unitBag (v, I# (1# +# tot -# max)) `unionBags` max_disc) max_scrut
245 -- If the variable is known, we produce a discount that
246 -- will take us back to 'max', the size of rh largest alternative
247 -- The 1+ is a little discount for reduced allocation in the caller
249 alts_size tot_size _ = tot_size
252 size_up (Case e _ alts) = nukeScrutDiscount (size_up e) `addSize`
253 foldr (addSize . size_up_alt) sizeZero alts
254 -- We don't charge for the case itself
255 -- It's a strict thing, and the price of the call
256 -- is paid by scrut. Also consider
257 -- case f x of DEFAULT -> e
258 -- This is just ';'! Don't charge for it.
261 size_up_app (App fun arg) args
262 | isTypeArg arg = size_up_app fun args
263 | otherwise = size_up_app fun (arg:args)
264 size_up_app fun args = foldr (addSize . nukeScrutDiscount . size_up)
265 (size_up_fun fun args)
268 -- A function application with at least one value argument
269 -- so if the function is an argument give it an arg-discount
271 -- Also behave specially if the function is a build
273 -- Also if the function is a constant Id (constr or primop)
274 -- compute discounts specially
275 size_up_fun (Var fun) args
276 | fun `hasKey` buildIdKey = buildSize
277 | fun `hasKey` augmentIdKey = augmentSize
279 = case idFlavour fun of
280 DataConId dc -> conSizeN (valArgCount args)
282 PrimOpId op -> primOpSize op (valArgCount args)
283 -- foldr addSize (primOpSize op) (map arg_discount args)
284 -- At one time I tried giving an arg-discount if a primop
285 -- is applied to one of the function's arguments, but it's
286 -- not good. At the moment, any unlifted-type arg gets a
287 -- 'True' for 'yes I'm evald', so we collect the discount even
288 -- if we know nothing about it. And just having it in a primop
289 -- doesn't help at all if we don't know something more.
291 other -> fun_discount fun `addSizeN`
292 (1 + length (filter (not . exprIsTrivial) args))
293 -- The 1+ is for the function itself
294 -- Add 1 for each non-trivial arg;
295 -- the allocation cost, as in let(rec)
296 -- Slight hack here: for constructors the args are almost always
297 -- trivial; and for primops they are almost always prim typed
298 -- We should really only count for non-prim-typed args in the
299 -- general case, but that seems too much like hard work
301 size_up_fun other args = size_up other
304 size_up_alt (con, bndrs, rhs) = size_up rhs
305 -- Don't charge for args, so that wrappers look cheap
308 -- We want to record if we're case'ing, or applying, an argument
309 fun_discount v | v `elem` top_args = SizeIs 0# (unitBag (v, opt_UF_FunAppDiscount)) 0#
310 fun_discount other = sizeZero
313 -- These addSize things have to be here because
314 -- I don't want to give them bOMB_OUT_SIZE as an argument
316 addSizeN TooBig _ = TooBig
317 addSizeN (SizeIs n xs d) (I# m)
318 | n_tot ># bOMB_OUT_SIZE = TooBig
319 | otherwise = SizeIs n_tot xs d
323 addSize TooBig _ = TooBig
324 addSize _ TooBig = TooBig
325 addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
326 | n_tot ># bOMB_OUT_SIZE = TooBig
327 | otherwise = SizeIs n_tot xys d_tot
331 xys = xs `unionBags` ys
334 Code for manipulating sizes
338 data ExprSize = TooBig
339 | SizeIs Int# -- Size found
340 (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
341 Int# -- Size to subtract if result is scrutinised
342 -- by a case expression
344 isTooBig TooBig = True
347 maxSize TooBig _ = TooBig
348 maxSize _ TooBig = TooBig
349 maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
352 sizeZero = SizeIs 0# emptyBag 0#
353 sizeOne = SizeIs 1# emptyBag 0#
354 sizeTwo = SizeIs 2# emptyBag 0#
355 sizeN (I# n) = SizeIs n emptyBag 0#
356 conSizeN (I# n) = SizeIs 1# emptyBag (n +# 1#)
357 -- Treat constructors as size 1; we are keen to expose them
358 -- (and we charge separately for their args). We can't treat
359 -- them as size zero, else we find that (I# x) has size 1,
360 -- which is the same as a lone variable; and hence 'v' will
361 -- always be replaced by (I# x), where v is bound to I# x.
364 | not (primOpIsDupable op) = sizeN opt_UF_DearOp
365 | not (primOpOutOfLine op) = sizeZero -- These are good to inline
366 | otherwise = sizeOne
368 buildSize = SizeIs (-2#) emptyBag 4#
369 -- We really want to inline applications of build
370 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
371 -- Indeed, we should add a result_discount becuause build is
372 -- very like a constructor. We don't bother to check that the
373 -- build is saturated (it usually is). The "-2" discounts for the \c n,
374 -- The "4" is rather arbitrary.
376 augmentSize = SizeIs (-2#) emptyBag 4#
377 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
378 -- e plus ys. The -2 accounts for the \cn
380 nukeScrutDiscount (SizeIs n vs d) = SizeIs n vs 0#
381 nukeScrutDiscount TooBig = TooBig
383 -- When we return a lambda, give a discount if it's used (applied)
384 lamScrutDiscount (SizeIs n vs d) = case opt_UF_FunAppDiscount of { I# d -> SizeIs n vs d }
385 lamScrutDiscount TooBig = TooBig
389 %************************************************************************
391 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
393 %************************************************************************
395 We have very limited information about an unfolding expression: (1)~so
396 many type arguments and so many value arguments expected---for our
397 purposes here, we assume we've got those. (2)~A ``size'' or ``cost,''
398 a single integer. (3)~An ``argument info'' vector. For this, what we
399 have at the moment is a Boolean per argument position that says, ``I
400 will look with great favour on an explicit constructor in this
401 position.'' (4)~The ``discount'' to subtract if the expression
402 is being scrutinised.
404 Assuming we have enough type- and value arguments (if not, we give up
405 immediately), then we see if the ``discounted size'' is below some
406 (semi-arbitrary) threshold. It works like this: for every argument
407 position where we're looking for a constructor AND WE HAVE ONE in our
408 hands, we get a (again, semi-arbitrary) discount [proportion to the
409 number of constructors in the type being scrutinized].
411 If we're in the context of a scrutinee ( \tr{(case <expr > of A .. -> ...;.. )})
412 and the expression in question will evaluate to a constructor, we use
413 the computed discount size *for the result only* rather than
414 computing the argument discounts. Since we know the result of
415 the expression is going to be taken apart, discounting its size
416 is more accurate (see @sizeExpr@ above for how this discount size
419 We use this one to avoid exporting inlinings that we ``couldn't possibly
420 use'' on the other side. Can be overridden w/ flaggery.
421 Just the same as smallEnoughToInline, except that it has no actual arguments.
424 couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
425 couldBeSmallEnoughToInline threshold rhs = case calcUnfoldingGuidance threshold rhs of
429 certainlyWillInline :: Id -> Bool
430 -- Sees if the Id is pretty certain to inline
431 certainlyWillInline v
432 = case idUnfolding v of
434 CoreUnfolding _ _ _ is_value _ g@(UnfoldIfGoodArgs n_vals _ size _)
436 && size - (n_vals +1) <= opt_UF_UseThreshold
441 @okToUnfoldInHifile@ is used when emitting unfolding info into an interface
442 file to determine whether an unfolding candidate really should be unfolded.
443 The predicate is needed to prevent @_casm_@s (+ lit-lits) from being emitted
444 into interface files.
446 The reason for inlining expressions containing _casm_s into interface files
447 is that these fragments of C are likely to mention functions/#defines that
448 will be out-of-scope when inlined into another module. This is not an
449 unfixable problem for the user (just need to -#include the approp. header
450 file), but turning it off seems to the simplest thing to do.
453 okToUnfoldInHiFile :: CoreExpr -> Bool
454 okToUnfoldInHiFile e = opt_UnfoldCasms || go e
456 -- Race over an expression looking for CCalls..
457 go (Var v) = case isPrimOpId_maybe v of
458 Just op -> okToUnfoldPrimOp op
460 go (Lit lit) = not (isLitLitLit lit)
461 go (App fun arg) = go fun && go arg
462 go (Lam _ body) = go body
463 go (Let binds body) = and (map go (body :rhssOfBind binds))
464 go (Case scrut bndr alts) = and (map go (scrut:rhssOfAlts alts)) &&
465 not (any isLitLitLit [ lit | (LitAlt lit, _, _) <- alts ])
466 go (Note _ body) = go body
469 -- ok to unfold a PrimOp as long as it's not a _casm_
470 okToUnfoldPrimOp (CCallOp ccall) = not (ccallIsCasm ccall)
471 okToUnfoldPrimOp _ = True
475 %************************************************************************
477 \subsection{callSiteInline}
479 %************************************************************************
481 This is the key function. It decides whether to inline a variable at a call site
483 callSiteInline is used at call sites, so it is a bit more generous.
484 It's a very important function that embodies lots of heuristics.
485 A non-WHNF can be inlined if it doesn't occur inside a lambda,
486 and occurs exactly once or
487 occurs once in each branch of a case and is small
489 If the thing is in WHNF, there's no danger of duplicating work,
490 so we can inline if it occurs once, or is small
492 NOTE: we don't want to inline top-level functions that always diverge.
493 It just makes the code bigger. Tt turns out that the convenient way to prevent
494 them inlining is to give them a NOINLINE pragma, which we do in
495 StrictAnal.addStrictnessInfoToTopId
498 callSiteInline :: Bool -- True <=> the Id is black listed
499 -> Bool -- 'inline' note at call site
502 -> [Bool] -- One for each value arg; True if it is interesting
503 -> Bool -- True <=> continuation is interesting
504 -> Maybe CoreExpr -- Unfolding, if any
507 callSiteInline black_listed inline_call occ id arg_infos interesting_cont
508 = case idUnfolding id of {
509 NoUnfolding -> Nothing ;
510 OtherCon cs -> Nothing ;
511 CompulsoryUnfolding unf_template | black_listed -> Nothing
512 | otherwise -> Just unf_template ;
513 -- Constructors have compulsory unfoldings, but
514 -- may have rules, in which case they are
515 -- black listed till later
516 CoreUnfolding unf_template is_top is_cheap is_value is_bot guidance ->
519 result | yes_or_no = Just unf_template
520 | otherwise = Nothing
522 n_val_args = length arg_infos
524 ok_inside_lam = is_value || is_bot || (is_cheap && not is_top)
525 -- I'm experimenting with is_cheap && not is_top
528 | black_listed = False
529 | otherwise = case occ of
530 IAmDead -> pprTrace "callSiteInline: dead" (ppr id) False
531 IAmALoopBreaker -> False
532 OneOcc in_lam one_br -> (not in_lam || ok_inside_lam) && consider_safe in_lam True one_br
533 NoOccInfo -> ok_inside_lam && consider_safe True False False
535 consider_safe in_lam once once_in_one_branch
536 -- consider_safe decides whether it's a good idea to inline something,
537 -- given that there's no work-duplication issue (the caller checks that).
538 -- once_in_one_branch = True means there's a unique textual occurrence
542 -- Be very keen to inline something if this is its unique occurrence:
544 -- a) Inlining gives a good chance of eliminating the original
545 -- binding (and hence the allocation) for the thing.
546 -- (Provided it's not a top level binding, in which case the
547 -- allocation costs nothing.)
549 -- b) Inlining a function that is called only once exposes the
550 -- body function to the call site.
552 -- The only time we hold back is when substituting inside a lambda;
553 -- then if the context is totally uninteresting (not applied, not scrutinised)
554 -- there is no point in substituting because it might just increase allocation,
555 -- by allocating the function itself many times
557 -- Note: there used to be a '&& not top_level' in the guard above,
558 -- but that stopped us inlining top-level functions used only once,
560 = not in_lam || not (null arg_infos) || interesting_cont
564 UnfoldNever -> False ;
565 UnfoldIfGoodArgs n_vals_wanted arg_discounts size res_discount
567 | enough_args && size <= (n_vals_wanted + 1)
569 -- Size of call is n_vals_wanted (+1 for the function)
573 -> some_benefit && small_enough
576 some_benefit = or arg_infos || really_interesting_cont ||
577 (not is_top && (once || (n_vals_wanted > 0 && enough_args)))
578 -- If it occurs more than once, there must be something interesting
579 -- about some argument, or the result context, to make it worth inlining
581 -- If a function has a nested defn we also record some-benefit,
582 -- on the grounds that we are often able to eliminate the binding,
583 -- and hence the allocation, for the function altogether; this is good
584 -- for join points. But this only makes sense for *functions*;
585 -- inlining a constructor doesn't help allocation unless the result is
586 -- scrutinised. UNLESS the constructor occurs just once, albeit possibly
587 -- in multiple case branches. Then inlining it doesn't increase allocation,
588 -- but it does increase the chance that the constructor won't be allocated at all
589 -- in the branches that don't use it.
591 enough_args = n_val_args >= n_vals_wanted
592 really_interesting_cont | n_val_args < n_vals_wanted = False -- Too few args
593 | n_val_args == n_vals_wanted = interesting_cont
594 | otherwise = True -- Extra args
595 -- really_interesting_cont tells if the result of the
596 -- call is in an interesting context.
598 small_enough = (size - discount) <= opt_UF_UseThreshold
599 discount = computeDiscount n_vals_wanted arg_discounts res_discount
600 arg_infos really_interesting_cont
604 if opt_D_dump_inlinings then
605 pprTrace "Considering inlining"
606 (ppr id <+> vcat [text "black listed" <+> ppr black_listed,
607 text "occ info:" <+> ppr occ,
608 text "arg infos" <+> ppr arg_infos,
609 text "interesting continuation" <+> ppr interesting_cont,
610 text "is value:" <+> ppr is_value,
611 text "is cheap:" <+> ppr is_cheap,
612 text "is bottom:" <+> ppr is_bot,
613 text "is top-level:" <+> ppr is_top,
614 text "guidance" <+> ppr guidance,
615 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO",
617 text "Unfolding =" <+> pprCoreExpr unf_template
625 computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Bool -> Int
626 computeDiscount n_vals_wanted arg_discounts res_discount arg_infos result_used
627 -- We multiple the raw discounts (args_discount and result_discount)
628 -- ty opt_UnfoldingKeenessFactor because the former have to do with
629 -- *size* whereas the discounts imply that there's some extra
630 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
633 -- we also discount 1 for each argument passed, because these will
634 -- reduce with the lambdas in the function (we count 1 for a lambda
636 = 1 + -- Discount of 1 because the result replaces the call
637 -- so we count 1 for the function itself
638 length (take n_vals_wanted arg_infos) +
639 -- Discount of 1 for each arg supplied, because the
640 -- result replaces the call
641 round (opt_UF_KeenessFactor *
642 fromInt (arg_discount + result_discount))
644 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
646 mk_arg_discount discount is_evald | is_evald = discount
649 -- Don't give a result discount unless there are enough args
650 result_discount | result_used = res_discount -- Over-applied, or case scrut
655 %************************************************************************
657 \subsection{Black-listing}
659 %************************************************************************
661 Inlining is controlled by the "Inline phase" number, which is set
662 by the per-simplification-pass '-finline-phase' flag.
664 For optimisation we use phase 1,2 and nothing (i.e. no -finline-phase flag)
665 in that order. The meanings of these are determined by the @blackListed@ function
668 The final simplification doesn't have a phase number.
674 (least black listing, most inlining)
675 INLINE n foo phase is Just p *and* p<n *and* foo appears on LHS of rule
676 INLINE foo phase is Just p *and* foo appears on LHS of rule
677 NOINLINE n foo phase is Just p *and* (p<n *or* foo appears on LHS of rule)
679 (most black listing, least inlining)
682 blackListed :: IdSet -- Used in transformation rules
683 -> Maybe Int -- Inline phase
684 -> Id -> Bool -- True <=> blacklisted
686 -- The blackListed function sees whether a variable should *not* be
687 -- inlined because of the inline phase we are in. This is the sole
688 -- place that the inline phase number is looked at.
690 blackListed rule_vars Nothing -- Last phase
691 = \v -> isNeverInlinePrag (idInlinePragma v)
693 blackListed rule_vars (Just phase)
694 = \v -> normal_case rule_vars phase v
696 normal_case rule_vars phase v
697 = case idInlinePragma v of
698 NoInlinePragInfo -> has_rules
700 IMustNotBeINLINEd from_INLINE Nothing
701 | from_INLINE -> has_rules -- Black list until final phase
702 | otherwise -> True -- Always blacklisted
704 IMustNotBeINLINEd from_inline (Just threshold)
705 | from_inline -> (phase < threshold && has_rules)
706 | otherwise -> (phase < threshold || has_rules)
708 has_rules = v `elemVarSet` rule_vars
709 || not (isEmptyCoreRules (idSpecialisation v))
713 SLPJ 95/04: Why @runST@ must be inlined very late:
717 (a, s') = newArray# 100 [] s
718 (_, s'') = fill_in_array_or_something a x s'
722 If we inline @runST@, we'll get:
725 (a, s') = newArray# 100 [] realWorld#{-NB-}
726 (_, s'') = fill_in_array_or_something a x s'
730 And now the @newArray#@ binding can be floated to become a CAF, which
731 is totally and utterly wrong:
734 (a, s') = newArray# 100 [] realWorld#{-NB-} -- YIKES!!!
737 let (_, s'') = fill_in_array_or_something a x s' in
740 All calls to @f@ will share a {\em single} array!
742 Yet we do want to inline runST sometime, so we can avoid
743 needless code. Solution: black list it until the last moment.