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, litIsDupable )
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) | litIsDupable lit = sizeOne
196 | otherwise = sizeN opt_UF_DearOp -- For lack of anything better
198 size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
199 | otherwise = size_up e
201 size_up (Let (NonRec binder rhs) body)
202 = nukeScrutDiscount (size_up rhs) `addSize`
203 size_up body `addSizeN`
204 (if isUnLiftedType (idType binder) then 0 else 1)
205 -- For the allocation
206 -- If the binder has an unlifted type there is no allocation
208 size_up (Let (Rec pairs) body)
209 = nukeScrutDiscount rhs_size `addSize`
210 size_up body `addSizeN`
211 length pairs -- For the allocation
213 rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
215 size_up (Case (Var v) _ alts)
216 | v `elem` top_args -- We are scrutinising an argument variable
218 {- I'm nuking this special case; BUT see the comment with case alternatives.
220 (a) It's too eager. We don't want to inline a wrapper into a
221 context with no benefit.
222 E.g. \ x. f (x+x) o point in inlining (+) here!
224 (b) It's ineffective. Once g's wrapper is inlined, its case-expressions
225 aren't scrutinising arguments any more
229 [alt] -> size_up_alt alt `addSize` SizeIs 0# (unitBag (v, 1)) 0#
230 -- We want to make wrapper-style evaluation look cheap, so that
231 -- when we inline a wrapper it doesn't make call site (much) bigger
232 -- Otherwise we get nasty phase ordering stuff:
235 -- If we inline g's wrapper, f looks big, and doesn't get inlined
236 -- into h; if we inline f first, while it looks small, then g's
237 -- wrapper will get inlined later anyway. To avoid this nasty
238 -- ordering difference, we make (case a of (x,y) -> ...),
239 -- *where a is one of the arguments* look free.
243 alts_size (foldr addSize sizeOne alt_sizes) -- The 1 is for the scrutinee
244 (foldr1 maxSize alt_sizes)
246 -- Good to inline if an arg is scrutinised, because
247 -- that may eliminate allocation in the caller
248 -- And it eliminates the case itself
251 alt_sizes = map size_up_alt alts
253 -- alts_size tries to compute a good discount for
254 -- the case when we are scrutinising an argument variable
255 alts_size (SizeIs tot tot_disc tot_scrut) -- Size of all alternatives
256 (SizeIs max max_disc max_scrut) -- Size of biggest alternative
257 = SizeIs tot (unitBag (v, I# (1# +# tot -# max)) `unionBags` max_disc) max_scrut
258 -- If the variable is known, we produce a discount that
259 -- will take us back to 'max', the size of rh largest alternative
260 -- The 1+ is a little discount for reduced allocation in the caller
261 alts_size tot_size _ = tot_size
264 size_up (Case e _ alts) = nukeScrutDiscount (size_up e) `addSize`
265 foldr (addSize . size_up_alt) sizeZero alts
266 -- We don't charge for the case itself
267 -- It's a strict thing, and the price of the call
268 -- is paid by scrut. Also consider
269 -- case f x of DEFAULT -> e
270 -- This is just ';'! Don't charge for it.
273 size_up_app (App fun arg) args
274 | isTypeArg arg = size_up_app fun args
275 | otherwise = size_up_app fun (arg:args)
276 size_up_app fun args = foldr (addSize . nukeScrutDiscount . size_up)
277 (size_up_fun fun args)
280 -- A function application with at least one value argument
281 -- so if the function is an argument give it an arg-discount
283 -- Also behave specially if the function is a build
285 -- Also if the function is a constant Id (constr or primop)
286 -- compute discounts specially
287 size_up_fun (Var fun) args
288 | fun `hasKey` buildIdKey = buildSize
289 | fun `hasKey` augmentIdKey = augmentSize
291 = case idFlavour fun of
292 DataConId dc -> conSizeN (valArgCount args)
294 PrimOpId op -> primOpSize op (valArgCount args)
295 -- foldr addSize (primOpSize op) (map arg_discount args)
296 -- At one time I tried giving an arg-discount if a primop
297 -- is applied to one of the function's arguments, but it's
298 -- not good. At the moment, any unlifted-type arg gets a
299 -- 'True' for 'yes I'm evald', so we collect the discount even
300 -- if we know nothing about it. And just having it in a primop
301 -- doesn't help at all if we don't know something more.
303 other -> fun_discount fun `addSizeN`
304 (1 + length (filter (not . exprIsTrivial) args))
305 -- The 1+ is for the function itself
306 -- Add 1 for each non-trivial arg;
307 -- the allocation cost, as in let(rec)
308 -- Slight hack here: for constructors the args are almost always
309 -- trivial; and for primops they are almost always prim typed
310 -- We should really only count for non-prim-typed args in the
311 -- general case, but that seems too much like hard work
313 size_up_fun other args = size_up other
316 size_up_alt (con, bndrs, rhs) = size_up rhs
317 -- Don't charge for args, so that wrappers look cheap
318 -- (See comments about wrappers with Case)
321 -- We want to record if we're case'ing, or applying, an argument
322 fun_discount v | v `elem` top_args = SizeIs 0# (unitBag (v, opt_UF_FunAppDiscount)) 0#
323 fun_discount other = sizeZero
326 -- These addSize things have to be here because
327 -- I don't want to give them bOMB_OUT_SIZE as an argument
329 addSizeN TooBig _ = TooBig
330 addSizeN (SizeIs n xs d) (I# m)
331 | n_tot ># bOMB_OUT_SIZE = TooBig
332 | otherwise = SizeIs n_tot xs d
336 addSize TooBig _ = TooBig
337 addSize _ TooBig = TooBig
338 addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
339 | n_tot ># bOMB_OUT_SIZE = TooBig
340 | otherwise = SizeIs n_tot xys d_tot
344 xys = xs `unionBags` ys
347 Code for manipulating sizes
351 data ExprSize = TooBig
352 | SizeIs Int# -- Size found
353 (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
354 Int# -- Size to subtract if result is scrutinised
355 -- by a case expression
357 isTooBig TooBig = True
360 maxSize TooBig _ = TooBig
361 maxSize _ TooBig = TooBig
362 maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
365 sizeZero = SizeIs 0# emptyBag 0#
366 sizeOne = SizeIs 1# emptyBag 0#
367 sizeTwo = SizeIs 2# emptyBag 0#
368 sizeN (I# n) = SizeIs n emptyBag 0#
369 conSizeN (I# n) = SizeIs 1# emptyBag (n +# 1#)
370 -- Treat constructors as size 1; we are keen to expose them
371 -- (and we charge separately for their args). We can't treat
372 -- them as size zero, else we find that (I# x) has size 1,
373 -- which is the same as a lone variable; and hence 'v' will
374 -- always be replaced by (I# x), where v is bound to I# x.
377 | not (primOpIsDupable op) = sizeN opt_UF_DearOp
378 | not (primOpOutOfLine op) = sizeZero -- These are good to inline
379 | otherwise = sizeOne
381 buildSize = SizeIs (-2#) emptyBag 4#
382 -- We really want to inline applications of build
383 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
384 -- Indeed, we should add a result_discount becuause build is
385 -- very like a constructor. We don't bother to check that the
386 -- build is saturated (it usually is). The "-2" discounts for the \c n,
387 -- The "4" is rather arbitrary.
389 augmentSize = SizeIs (-2#) emptyBag 4#
390 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
391 -- e plus ys. The -2 accounts for the \cn
393 nukeScrutDiscount (SizeIs n vs d) = SizeIs n vs 0#
394 nukeScrutDiscount TooBig = TooBig
396 -- When we return a lambda, give a discount if it's used (applied)
397 lamScrutDiscount (SizeIs n vs d) = case opt_UF_FunAppDiscount of { I# d -> SizeIs n vs d }
398 lamScrutDiscount TooBig = TooBig
402 %************************************************************************
404 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
406 %************************************************************************
408 We have very limited information about an unfolding expression: (1)~so
409 many type arguments and so many value arguments expected---for our
410 purposes here, we assume we've got those. (2)~A ``size'' or ``cost,''
411 a single integer. (3)~An ``argument info'' vector. For this, what we
412 have at the moment is a Boolean per argument position that says, ``I
413 will look with great favour on an explicit constructor in this
414 position.'' (4)~The ``discount'' to subtract if the expression
415 is being scrutinised.
417 Assuming we have enough type- and value arguments (if not, we give up
418 immediately), then we see if the ``discounted size'' is below some
419 (semi-arbitrary) threshold. It works like this: for every argument
420 position where we're looking for a constructor AND WE HAVE ONE in our
421 hands, we get a (again, semi-arbitrary) discount [proportion to the
422 number of constructors in the type being scrutinized].
424 If we're in the context of a scrutinee ( \tr{(case <expr > of A .. -> ...;.. )})
425 and the expression in question will evaluate to a constructor, we use
426 the computed discount size *for the result only* rather than
427 computing the argument discounts. Since we know the result of
428 the expression is going to be taken apart, discounting its size
429 is more accurate (see @sizeExpr@ above for how this discount size
432 We use this one to avoid exporting inlinings that we ``couldn't possibly
433 use'' on the other side. Can be overridden w/ flaggery.
434 Just the same as smallEnoughToInline, except that it has no actual arguments.
437 couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
438 couldBeSmallEnoughToInline threshold rhs = case calcUnfoldingGuidance threshold rhs of
442 certainlyWillInline :: Id -> Bool
443 -- Sees if the Id is pretty certain to inline
444 certainlyWillInline v
445 = case idUnfolding v of
447 CoreUnfolding _ _ _ is_value _ g@(UnfoldIfGoodArgs n_vals _ size _)
449 && size - (n_vals +1) <= opt_UF_UseThreshold
454 @okToUnfoldInHifile@ is used when emitting unfolding info into an interface
455 file to determine whether an unfolding candidate really should be unfolded.
456 The predicate is needed to prevent @_casm_@s (+ lit-lits) from being emitted
457 into interface files.
459 The reason for inlining expressions containing _casm_s into interface files
460 is that these fragments of C are likely to mention functions/#defines that
461 will be out-of-scope when inlined into another module. This is not an
462 unfixable problem for the user (just need to -#include the approp. header
463 file), but turning it off seems to the simplest thing to do.
466 okToUnfoldInHiFile :: CoreExpr -> Bool
467 okToUnfoldInHiFile e = opt_UnfoldCasms || go e
469 -- Race over an expression looking for CCalls..
470 go (Var v) = case isPrimOpId_maybe v of
471 Just op -> okToUnfoldPrimOp op
473 go (Lit lit) = not (isLitLitLit lit)
474 go (App fun arg) = go fun && go arg
475 go (Lam _ body) = go body
476 go (Let binds body) = and (map go (body :rhssOfBind binds))
477 go (Case scrut bndr alts) = and (map go (scrut:rhssOfAlts alts)) &&
478 not (any isLitLitLit [ lit | (LitAlt lit, _, _) <- alts ])
479 go (Note _ body) = go body
482 -- ok to unfold a PrimOp as long as it's not a _casm_
483 okToUnfoldPrimOp (CCallOp ccall) = not (ccallIsCasm ccall)
484 okToUnfoldPrimOp _ = True
488 %************************************************************************
490 \subsection{callSiteInline}
492 %************************************************************************
494 This is the key function. It decides whether to inline a variable at a call site
496 callSiteInline is used at call sites, so it is a bit more generous.
497 It's a very important function that embodies lots of heuristics.
498 A non-WHNF can be inlined if it doesn't occur inside a lambda,
499 and occurs exactly once or
500 occurs once in each branch of a case and is small
502 If the thing is in WHNF, there's no danger of duplicating work,
503 so we can inline if it occurs once, or is small
505 NOTE: we don't want to inline top-level functions that always diverge.
506 It just makes the code bigger. Tt turns out that the convenient way to prevent
507 them inlining is to give them a NOINLINE pragma, which we do in
508 StrictAnal.addStrictnessInfoToTopId
511 callSiteInline :: Bool -- True <=> the Id is black listed
512 -> Bool -- 'inline' note at call site
515 -> [Bool] -- One for each value arg; True if it is interesting
516 -> Bool -- True <=> continuation is interesting
517 -> Maybe CoreExpr -- Unfolding, if any
520 callSiteInline black_listed inline_call occ id arg_infos interesting_cont
521 = case idUnfolding id of {
522 NoUnfolding -> Nothing ;
523 OtherCon cs -> Nothing ;
524 CompulsoryUnfolding unf_template | black_listed -> Nothing
525 | otherwise -> Just unf_template ;
526 -- Constructors have compulsory unfoldings, but
527 -- may have rules, in which case they are
528 -- black listed till later
529 CoreUnfolding unf_template is_top is_cheap is_value is_bot guidance ->
532 result | yes_or_no = Just unf_template
533 | otherwise = Nothing
535 n_val_args = length arg_infos
537 ok_inside_lam = is_value || is_bot || (is_cheap && not is_top)
538 -- I'm experimenting with is_cheap && not is_top
541 | black_listed = False
542 | otherwise = case occ of
543 IAmDead -> pprTrace "callSiteInline: dead" (ppr id) False
544 IAmALoopBreaker -> False
545 OneOcc in_lam one_br -> (not in_lam || ok_inside_lam) && consider_safe in_lam True one_br
546 NoOccInfo -> ok_inside_lam && consider_safe True False False
548 consider_safe in_lam once once_in_one_branch
549 -- consider_safe decides whether it's a good idea to inline something,
550 -- given that there's no work-duplication issue (the caller checks that).
551 -- once_in_one_branch = True means there's a unique textual occurrence
555 -- Be very keen to inline something if this is its unique occurrence:
557 -- a) Inlining gives a good chance of eliminating the original
558 -- binding (and hence the allocation) for the thing.
559 -- (Provided it's not a top level binding, in which case the
560 -- allocation costs nothing.)
562 -- b) Inlining a function that is called only once exposes the
563 -- body function to the call site.
565 -- The only time we hold back is when substituting inside a lambda;
566 -- then if the context is totally uninteresting (not applied, not scrutinised)
567 -- there is no point in substituting because it might just increase allocation,
568 -- by allocating the function itself many times
570 -- Note: there used to be a '&& not top_level' in the guard above,
571 -- but that stopped us inlining top-level functions used only once,
573 = not in_lam || not (null arg_infos) || interesting_cont
577 UnfoldNever -> False ;
578 UnfoldIfGoodArgs n_vals_wanted arg_discounts size res_discount
580 | enough_args && size <= (n_vals_wanted + 1)
582 -- Size of call is n_vals_wanted (+1 for the function)
586 -> some_benefit && small_enough
589 some_benefit = or arg_infos || really_interesting_cont ||
590 (not is_top && (once || (n_vals_wanted > 0 && enough_args)))
591 -- If it occurs more than once, there must be something interesting
592 -- about some argument, or the result context, to make it worth inlining
594 -- If a function has a nested defn we also record some-benefit,
595 -- on the grounds that we are often able to eliminate the binding,
596 -- and hence the allocation, for the function altogether; this is good
597 -- for join points. But this only makes sense for *functions*;
598 -- inlining a constructor doesn't help allocation unless the result is
599 -- scrutinised. UNLESS the constructor occurs just once, albeit possibly
600 -- in multiple case branches. Then inlining it doesn't increase allocation,
601 -- but it does increase the chance that the constructor won't be allocated at all
602 -- in the branches that don't use it.
604 enough_args = n_val_args >= n_vals_wanted
605 really_interesting_cont | n_val_args < n_vals_wanted = False -- Too few args
606 | n_val_args == n_vals_wanted = interesting_cont
607 | otherwise = True -- Extra args
608 -- really_interesting_cont tells if the result of the
609 -- call is in an interesting context.
611 small_enough = (size - discount) <= opt_UF_UseThreshold
612 discount = computeDiscount n_vals_wanted arg_discounts res_discount
613 arg_infos really_interesting_cont
617 if opt_D_dump_inlinings then
618 pprTrace "Considering inlining"
619 (ppr id <+> vcat [text "black listed:" <+> ppr black_listed,
620 text "occ info:" <+> ppr occ,
621 text "arg infos" <+> ppr arg_infos,
622 text "interesting continuation" <+> ppr interesting_cont,
623 text "is value:" <+> ppr is_value,
624 text "is cheap:" <+> ppr is_cheap,
625 text "is bottom:" <+> ppr is_bot,
626 text "is top-level:" <+> ppr is_top,
627 text "guidance" <+> ppr guidance,
628 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO",
630 text "Unfolding =" <+> pprCoreExpr unf_template
638 computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Bool -> Int
639 computeDiscount n_vals_wanted arg_discounts res_discount arg_infos result_used
640 -- We multiple the raw discounts (args_discount and result_discount)
641 -- ty opt_UnfoldingKeenessFactor because the former have to do with
642 -- *size* whereas the discounts imply that there's some extra
643 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
646 -- we also discount 1 for each argument passed, because these will
647 -- reduce with the lambdas in the function (we count 1 for a lambda
649 = 1 + -- Discount of 1 because the result replaces the call
650 -- so we count 1 for the function itself
651 length (take n_vals_wanted arg_infos) +
652 -- Discount of 1 for each arg supplied, because the
653 -- result replaces the call
654 round (opt_UF_KeenessFactor *
655 fromInt (arg_discount + result_discount))
657 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
659 mk_arg_discount discount is_evald | is_evald = discount
662 -- Don't give a result discount unless there are enough args
663 result_discount | result_used = res_discount -- Over-applied, or case scrut
668 %************************************************************************
670 \subsection{Black-listing}
672 %************************************************************************
674 Inlining is controlled by the "Inline phase" number, which is set
675 by the per-simplification-pass '-finline-phase' flag.
677 For optimisation we use phase 1,2 and nothing (i.e. no -finline-phase flag)
678 in that order. The meanings of these are determined by the @blackListed@ function
681 The final simplification doesn't have a phase number.
687 (least black listing, most inlining)
688 INLINE n foo phase is Just p *and* p<n *and* foo appears on LHS of rule
689 INLINE foo phase is Just p *and* foo appears on LHS of rule
690 NOINLINE n foo phase is Just p *and* (p<n *or* foo appears on LHS of rule)
692 (most black listing, least inlining)
695 blackListed :: IdSet -- Used in transformation rules
696 -> Maybe Int -- Inline phase
697 -> Id -> Bool -- True <=> blacklisted
699 -- The blackListed function sees whether a variable should *not* be
700 -- inlined because of the inline phase we are in. This is the sole
701 -- place that the inline phase number is looked at.
703 blackListed rule_vars Nothing -- Last phase
704 = \v -> isNeverInlinePrag (idInlinePragma v)
706 blackListed rule_vars (Just phase)
707 = \v -> normal_case rule_vars phase v
709 normal_case rule_vars phase v
710 = case idInlinePragma v of
711 NoInlinePragInfo -> has_rules
713 IMustNotBeINLINEd from_INLINE Nothing
714 | from_INLINE -> has_rules -- Black list until final phase
715 | otherwise -> True -- Always blacklisted
717 IMustNotBeINLINEd from_INLINE (Just threshold)
718 | from_INLINE -> (phase < threshold && has_rules)
719 | otherwise -> (phase < threshold || has_rules)
721 has_rules = v `elemVarSet` rule_vars
722 || not (isEmptyCoreRules (idSpecialisation v))
726 SLPJ 95/04: Why @runST@ must be inlined very late:
730 (a, s') = newArray# 100 [] s
731 (_, s'') = fill_in_array_or_something a x s'
735 If we inline @runST@, we'll get:
738 (a, s') = newArray# 100 [] realWorld#{-NB-}
739 (_, s'') = fill_in_array_or_something a x s'
743 And now the @newArray#@ binding can be floated to become a CAF, which
744 is totally and utterly wrong:
747 (a, s') = newArray# 100 [] realWorld#{-NB-} -- YIKES!!!
750 let (_, s'') = fill_in_array_or_something a x s' in
753 All calls to @f@ will share a {\em single} array!
755 Yet we do want to inline runST sometime, so we can avoid
756 needless code. Solution: black list it until the last moment.