2 % (c) The AQUA Project, Glasgow University, 1994-1998
4 \section[CoreUnfold]{Core-syntax unfoldings}
6 Unfoldings (which can travel across module boundaries) are in Core
7 syntax (namely @CoreExpr@s).
9 The type @Unfolding@ sits ``above'' simply-Core-expressions
10 unfoldings, capturing ``higher-level'' things we know about a binding,
11 usually things that the simplifier found out (e.g., ``it's a
12 literal''). In the corner of a @CoreUnfolding@ unfolding, you will
13 find, unsurprisingly, a Core expression.
17 Unfolding, UnfoldingGuidance, -- Abstract types
19 noUnfolding, mkTopUnfolding, mkUnfolding, mkCompulsoryUnfolding, seqUnfolding,
20 mkOtherCon, otherCons,
21 unfoldingTemplate, maybeUnfoldingTemplate,
22 isEvaldUnfolding, isValueUnfolding, isCheapUnfolding, isCompulsoryUnfolding,
23 hasUnfolding, hasSomeUnfolding,
25 couldBeSmallEnoughToInline,
29 callSiteInline, blackListed
32 #include "HsVersions.h"
34 import CmdLineOpts ( opt_UF_CreationThreshold,
36 opt_UF_ScrutConDiscount,
37 opt_UF_FunAppDiscount,
38 opt_UF_PrimArgDiscount,
40 opt_UF_CheapOp, opt_UF_DearOp,
41 opt_UnfoldCasms, opt_PprStyle_Debug,
45 import PprCore ( pprCoreExpr )
46 import OccurAnal ( occurAnalyseGlobalExpr )
48 import CoreUtils ( exprIsValue, exprIsCheap, exprIsBottom, exprIsTrivial )
49 import Id ( Id, idType, idFlavour, idUnique, isId, idWorkerInfo,
50 idSpecialisation, idInlinePragma, idUnfolding,
54 import Name ( isLocallyDefined )
55 import Literal ( isLitLitLit )
56 import PrimOp ( PrimOp(..), primOpIsDupable, primOpOutOfLine, ccallIsCasm )
57 import IdInfo ( ArityInfo(..), InlinePragInfo(..), OccInfo(..), IdFlavour(..), CprInfo(..), insideLam, workerExists )
58 import TyCon ( tyConFamilySize )
59 import Type ( splitAlgTyConApp_maybe, splitFunTy_maybe, isUnLiftedType )
60 import Unique ( Unique, buildIdKey, augmentIdKey )
61 import Maybes ( maybeToBool )
63 import List ( maximumBy )
64 import Util ( isIn, lengthExceeds )
67 #if __GLASGOW_HASKELL__ >= 404
68 import GlaExts ( fromInt )
73 %************************************************************************
75 \subsection{Making unfoldings}
77 %************************************************************************
80 mkTopUnfolding cpr_info expr = mkUnfolding True {- Top level -} cpr_info expr
82 mkUnfolding top_lvl cpr_info expr
83 = CoreUnfolding (occurAnalyseGlobalExpr expr)
88 (calcUnfoldingGuidance opt_UF_CreationThreshold cpr_info expr)
89 -- Sometimes during simplification, there's a large let-bound thing
90 -- which has been substituted, and so is now dead; so 'expr' contains
91 -- two copies of the thing while the occurrence-analysed expression doesn't
92 -- Nevertheless, we don't occ-analyse before computing the size because the
93 -- size computation bales out after a while, whereas occurrence analysis does not.
95 -- This can occasionally mean that the guidance is very pessimistic;
96 -- it gets fixed up next round
98 mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
99 = CompulsoryUnfolding (occurAnalyseGlobalExpr expr)
103 %************************************************************************
105 \subsection{The UnfoldingGuidance type}
107 %************************************************************************
110 instance Outputable UnfoldingGuidance where
111 ppr UnfoldNever = ptext SLIT("NEVER")
112 ppr (UnfoldIfGoodArgs v cs size discount)
113 = hsep [ ptext SLIT("IF_ARGS"), int v,
114 brackets (hsep (map int cs)),
121 calcUnfoldingGuidance
122 :: Int -- bomb out if size gets bigger than this
123 -> CprInfo -- CPR info for this RHS
124 -> CoreExpr -- expression to look at
126 calcUnfoldingGuidance bOMB_OUT_SIZE cpr_info 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,
138 -- Experimental thing commented in for now
139 -- max_inline_size = case cpr_info of
140 -- NoCPRInfo -> n_val_binders + 2
141 -- ReturnsCPR -> n_val_binders + 1
143 -- However, the wrapper for a CPR'd function is particularly good to inline,
144 -- even in a boring context, because we may get to do update in place:
145 -- let x = case y of { I# y# -> I# (y# +# 1#) }
146 -- Hence the case on cpr_info
149 case (sizeExpr bOMB_OUT_SIZE val_binders body) of
152 | not inline -> UnfoldNever
153 -- A big function with an INLINE pragma must
154 -- have an UnfoldIfGoodArgs guidance
155 | inline -> UnfoldIfGoodArgs n_val_binders
156 (map (const 0) val_binders)
159 SizeIs size cased_args scrut_discount
162 (map discount_for val_binders)
168 final_size | inline = boxed_size `min` max_inline_size
169 | otherwise = boxed_size
171 -- Sometimes an INLINE thing is smaller than n_val_binders+2.
172 -- A particular case in point is a constructor, which has size 1.
173 -- We want to inline this regardless, hence the `min`
175 discount_for b = foldlBag (\acc (b',n) -> if b==b' then acc+n else acc)
179 collect_val_bndrs e = go False [] e
180 -- We need to be a bit careful about how we collect the
181 -- value binders. In ptic, if we see
182 -- __inline_me (\x y -> e)
183 -- We want to say "2 value binders". Why? So that
184 -- we take account of information given for the arguments
186 go inline rev_vbs (Note InlineMe e) = go True rev_vbs e
187 go inline rev_vbs (Lam b e) | isId b = go inline (b:rev_vbs) e
188 | otherwise = go inline rev_vbs e
189 go inline rev_vbs e = (inline, reverse rev_vbs, e)
193 sizeExpr :: Int -- Bomb out if it gets bigger than this
194 -> [Id] -- Arguments; we're interested in which of these
199 sizeExpr (I# bOMB_OUT_SIZE) top_args expr
202 size_up (Type t) = sizeZero -- Types cost nothing
203 size_up (Var v) = sizeOne
205 size_up (Note _ body) = size_up body -- Notes cost nothing
207 size_up (App fun (Type t)) = size_up fun
208 size_up (App fun arg) = size_up_app fun [arg]
210 size_up (Lit lit) = sizeOne
212 size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
213 | otherwise = size_up e
215 size_up (Let (NonRec binder rhs) body)
216 = nukeScrutDiscount (size_up rhs) `addSize`
217 size_up body `addSizeN`
218 (if isUnLiftedType (idType binder) then 0 else 1)
219 -- For the allocation
220 -- If the binder has an unlifted type there is no allocation
222 size_up (Let (Rec pairs) body)
223 = nukeScrutDiscount rhs_size `addSize`
224 size_up body `addSizeN`
225 length pairs -- For the allocation
227 rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
229 -- We want to make wrapper-style evaluation look cheap, so that
230 -- when we inline a wrapper it doesn't make call site (much) bigger
231 -- Otherwise we get nasty phase ordering stuff:
234 -- If we inline g's wrapper, f looks big, and doesn't get inlined
235 -- into h; if we inline f first, while it looks small, then g's
236 -- wrapper will get inlined later anyway. To avoid this nasty
237 -- ordering difference, we make (case a of (x,y) -> ...) look free.
238 size_up (Case (Var v) _ [alt])
240 = size_up_alt alt `addSize` SizeIs 0# (unitBag (v, 1)) 0#
241 -- Good to inline if an arg is scrutinised, because
242 -- that may eliminate allocation in the caller
243 -- And it eliminates the case itself
247 -- Scrutinising one of the argument variables,
248 -- with more than one alternative
249 size_up (Case (Var v) _ alts)
251 = alts_size (foldr addSize sizeOne alt_sizes) -- The 1 is for the scrutinee
252 (foldr1 maxSize alt_sizes)
254 v_in_args = v `elem` top_args
255 alt_sizes = map size_up_alt alts
257 alts_size (SizeIs tot tot_disc tot_scrut) -- Size of all alternatives
258 (SizeIs max max_disc max_scrut) -- Size of biggest alternative
259 = SizeIs tot (unitBag (v, I# (1# +# tot -# max)) `unionBags` max_disc) max_scrut
260 -- If the variable is known, we produce a discount that
261 -- will take us back to 'max', the size of rh largest alternative
262 -- The 1+ is a little discount for reduced allocation in the caller
264 alts_size tot_size _ = tot_size
267 size_up (Case e _ alts) = nukeScrutDiscount (size_up e) `addSize`
268 foldr (addSize . size_up_alt) sizeZero alts
269 -- We don't charge for the case itself
270 -- It's a strict thing, and the price of the call
271 -- is paid by scrut. Also consider
272 -- case f x of DEFAULT -> e
273 -- This is just ';'! Don't charge for it.
276 size_up_app (App fun arg) args
277 | isTypeArg arg = size_up_app fun args
278 | otherwise = size_up_app fun (arg:args)
279 size_up_app fun args = foldr (addSize . nukeScrutDiscount . size_up)
280 (size_up_fun fun args)
283 -- A function application with at least one value argument
284 -- so if the function is an argument give it an arg-discount
286 -- Also behave specially if the function is a build
288 -- Also if the function is a constant Id (constr or primop)
289 -- compute discounts specially
290 size_up_fun (Var fun) args
291 | idUnique fun == buildIdKey = buildSize
292 | idUnique fun == augmentIdKey = augmentSize
294 = case idFlavour fun of
295 DataConId dc -> conSizeN (valArgCount args)
297 PrimOpId op -> primOpSize op (valArgCount args)
298 -- foldr addSize (primOpSize op) (map arg_discount args)
299 -- At one time I tried giving an arg-discount if a primop
300 -- is applied to one of the function's arguments, but it's
301 -- not good. At the moment, any unlifted-type arg gets a
302 -- 'True' for 'yes I'm evald', so we collect the discount even
303 -- if we know nothing about it. And just having it in a primop
304 -- doesn't help at all if we don't know something more.
306 other -> fun_discount fun `addSizeN`
307 (1 + length (filter (not . exprIsTrivial) args))
308 -- The 1+ is for the function itself
309 -- Add 1 for each non-trivial arg;
310 -- the allocation cost, as in let(rec)
311 -- Slight hack here: for constructors the args are almost always
312 -- trivial; and for primops they are almost always prim typed
313 -- We should really only count for non-prim-typed args in the
314 -- general case, but that seems too much like hard work
316 size_up_fun other args = size_up other
319 size_up_alt (con, bndrs, rhs) = size_up rhs
320 -- Don't charge for args, so that wrappers look cheap
323 -- We want to record if we're case'ing, or applying, an argument
324 fun_discount v | v `elem` top_args = SizeIs 0# (unitBag (v, opt_UF_FunAppDiscount)) 0#
325 fun_discount other = sizeZero
328 -- These addSize things have to be here because
329 -- I don't want to give them bOMB_OUT_SIZE as an argument
331 addSizeN TooBig _ = TooBig
332 addSizeN (SizeIs n xs d) (I# m)
333 | n_tot ># bOMB_OUT_SIZE = TooBig
334 | otherwise = SizeIs n_tot xs d
338 addSize TooBig _ = TooBig
339 addSize _ TooBig = TooBig
340 addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
341 | n_tot ># bOMB_OUT_SIZE = TooBig
342 | otherwise = SizeIs n_tot xys d_tot
346 xys = xs `unionBags` ys
349 Code for manipulating sizes
353 data ExprSize = TooBig
354 | SizeIs Int# -- Size found
355 (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
356 Int# -- Size to subtract if result is scrutinised
357 -- by a case expression
359 isTooBig TooBig = True
362 maxSize TooBig _ = TooBig
363 maxSize _ TooBig = TooBig
364 maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
367 sizeZero = SizeIs 0# emptyBag 0#
368 sizeOne = SizeIs 1# emptyBag 0#
369 sizeTwo = SizeIs 2# emptyBag 0#
370 sizeN (I# n) = SizeIs n emptyBag 0#
371 conSizeN (I# n) = SizeIs 1# emptyBag (n +# 1#)
372 -- Treat constructors as size 1; we are keen to expose them
373 -- (and we charge separately for their args). We can't treat
374 -- them as size zero, else we find that (I# x) has size 1,
375 -- which is the same as a lone variable; and hence 'v' will
376 -- always be replaced by (I# x), where v is bound to I# x.
379 | not (primOpIsDupable op) = sizeN opt_UF_DearOp
380 | not (primOpOutOfLine op) = sizeZero -- These are good to inline
381 | otherwise = sizeOne
383 buildSize = SizeIs (-2#) emptyBag 4#
384 -- We really want to inline applications of build
385 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
386 -- Indeed, we should add a result_discount becuause build is
387 -- very like a constructor. We don't bother to check that the
388 -- build is saturated (it usually is). The "-2" discounts for the \c n,
389 -- The "4" is rather arbitrary.
391 augmentSize = SizeIs (-2#) emptyBag 4#
392 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
393 -- e plus ys. The -2 accounts for the \cn
395 nukeScrutDiscount (SizeIs n vs d) = SizeIs n vs 0#
396 nukeScrutDiscount TooBig = TooBig
398 -- When we return a lambda, give a discount if it's used (applied)
399 lamScrutDiscount (SizeIs n vs d) = case opt_UF_FunAppDiscount of { I# d -> SizeIs n vs d }
400 lamScrutDiscount TooBig = TooBig
404 %************************************************************************
406 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
408 %************************************************************************
410 We have very limited information about an unfolding expression: (1)~so
411 many type arguments and so many value arguments expected---for our
412 purposes here, we assume we've got those. (2)~A ``size'' or ``cost,''
413 a single integer. (3)~An ``argument info'' vector. For this, what we
414 have at the moment is a Boolean per argument position that says, ``I
415 will look with great favour on an explicit constructor in this
416 position.'' (4)~The ``discount'' to subtract if the expression
417 is being scrutinised.
419 Assuming we have enough type- and value arguments (if not, we give up
420 immediately), then we see if the ``discounted size'' is below some
421 (semi-arbitrary) threshold. It works like this: for every argument
422 position where we're looking for a constructor AND WE HAVE ONE in our
423 hands, we get a (again, semi-arbitrary) discount [proportion to the
424 number of constructors in the type being scrutinized].
426 If we're in the context of a scrutinee ( \tr{(case <expr > of A .. -> ...;.. )})
427 and the expression in question will evaluate to a constructor, we use
428 the computed discount size *for the result only* rather than
429 computing the argument discounts. Since we know the result of
430 the expression is going to be taken apart, discounting its size
431 is more accurate (see @sizeExpr@ above for how this discount size
434 We use this one to avoid exporting inlinings that we ``couldn't possibly
435 use'' on the other side. Can be overridden w/ flaggery.
436 Just the same as smallEnoughToInline, except that it has no actual arguments.
439 couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
440 couldBeSmallEnoughToInline threshold rhs = case calcUnfoldingGuidance threshold NoCPRInfo rhs of
444 certainlyWillInline :: Id -> Bool
445 -- Sees if the Id is pretty certain to inline
446 certainlyWillInline v
447 = case idUnfolding v of
449 CoreUnfolding _ _ _ is_value _ (UnfoldIfGoodArgs n_vals _ size _)
451 && size - (n_vals +1) <= opt_UF_UseThreshold
456 never_inline = case idInlinePragma v of
457 IMustNotBeINLINEd False Nothing -> True
461 @okToUnfoldInHifile@ is used when emitting unfolding info into an interface
462 file to determine whether an unfolding candidate really should be unfolded.
463 The predicate is needed to prevent @_casm_@s (+ lit-lits) from being emitted
464 into interface files.
466 The reason for inlining expressions containing _casm_s into interface files
467 is that these fragments of C are likely to mention functions/#defines that
468 will be out-of-scope when inlined into another module. This is not an
469 unfixable problem for the user (just need to -#include the approp. header
470 file), but turning it off seems to the simplest thing to do.
473 okToUnfoldInHiFile :: CoreExpr -> Bool
474 okToUnfoldInHiFile e = opt_UnfoldCasms || go e
476 -- Race over an expression looking for CCalls..
477 go (Var v) = case isPrimOpId_maybe v of
478 Just op -> okToUnfoldPrimOp op
480 go (Lit lit) = not (isLitLitLit lit)
481 go (App fun arg) = go fun && go arg
482 go (Lam _ body) = go body
483 go (Let binds body) = and (map go (body :rhssOfBind binds))
484 go (Case scrut bndr alts) = and (map go (scrut:rhssOfAlts alts))
485 go (Note _ body) = go body
488 -- ok to unfold a PrimOp as long as it's not a _casm_
489 okToUnfoldPrimOp (CCallOp ccall) = not (ccallIsCasm ccall)
490 okToUnfoldPrimOp _ = True
494 %************************************************************************
496 \subsection{callSiteInline}
498 %************************************************************************
500 This is the key function. It decides whether to inline a variable at a call site
502 callSiteInline is used at call sites, so it is a bit more generous.
503 It's a very important function that embodies lots of heuristics.
504 A non-WHNF can be inlined if it doesn't occur inside a lambda,
505 and occurs exactly once or
506 occurs once in each branch of a case and is small
508 If the thing is in WHNF, there's no danger of duplicating work,
509 so we can inline if it occurs once, or is small
511 NOTE: we don't want to inline top-level functions that always diverge.
512 It just makes the code bigger. Tt turns out that the convenient way to prevent
513 them inlining is to give them a NOINLINE pragma, which we do in
514 StrictAnal.addStrictnessInfoToTopId
517 callSiteInline :: Bool -- True <=> the Id is black listed
518 -> Bool -- 'inline' note at call site
521 -> [Bool] -- One for each value arg; True if it is interesting
522 -> Bool -- True <=> continuation is interesting
523 -> Maybe CoreExpr -- Unfolding, if any
526 callSiteInline black_listed inline_call occ id arg_infos interesting_cont
527 = case idUnfolding id of {
528 NoUnfolding -> Nothing ;
529 OtherCon _ -> Nothing ;
530 CompulsoryUnfolding unf_template | black_listed -> Nothing
531 | otherwise -> Just unf_template ;
532 -- Constructors have compulsory unfoldings, but
533 -- may have rules, in which case they are
534 -- black listed till later
535 CoreUnfolding unf_template is_top is_cheap is_value is_bot guidance ->
538 result | yes_or_no = Just unf_template
539 | otherwise = Nothing
541 n_val_args = length arg_infos
543 ok_inside_lam = is_value || is_bot || (is_cheap && not is_top)
544 -- I'm experimenting with is_cheap && not is_top
547 | black_listed = False
548 | otherwise = case occ of
549 IAmDead -> pprTrace "callSiteInline: dead" (ppr id) False
550 IAmALoopBreaker -> False
551 OneOcc in_lam one_br -> (not in_lam || ok_inside_lam) && consider_safe in_lam True one_br
552 NoOccInfo -> ok_inside_lam && consider_safe True False False
554 consider_safe in_lam once once_in_one_branch
555 -- consider_safe decides whether it's a good idea to inline something,
556 -- given that there's no work-duplication issue (the caller checks that).
557 -- once_in_one_branch = True means there's a unique textual occurrence
561 -- Be very keen to inline something if this is its unique occurrence:
563 -- a) Inlining gives a good chance of eliminating the original
564 -- binding (and hence the allocation) for the thing.
565 -- (Provided it's not a top level binding, in which case the
566 -- allocation costs nothing.)
568 -- b) Inlining a function that is called only once exposes the
569 -- body function to the call site.
571 -- The only time we hold back is when substituting inside a lambda;
572 -- then if the context is totally uninteresting (not applied, not scrutinised)
573 -- there is no point in substituting because it might just increase allocation,
574 -- by allocating the function itself many times
576 -- Note: there used to be a '&& not top_level' in the guard above,
577 -- but that stopped us inlining top-level functions used only once,
579 = not in_lam || not (null arg_infos) || interesting_cont
583 UnfoldNever -> False ;
584 UnfoldIfGoodArgs n_vals_wanted arg_discounts size res_discount
586 | enough_args && size <= (n_vals_wanted + 1)
588 -- Size of call is n_vals_wanted (+1 for the function)
592 -> some_benefit && small_enough
595 some_benefit = or arg_infos || really_interesting_cont ||
596 (not is_top && (once || (n_vals_wanted > 0 && enough_args)))
597 -- If it occurs more than once, there must be something interesting
598 -- about some argument, or the result context, to make it worth inlining
600 -- If a function has a nested defn we also record some-benefit,
601 -- on the grounds that we are often able to eliminate the binding,
602 -- and hence the allocation, for the function altogether; this is good
603 -- for join points. But this only makes sense for *functions*;
604 -- inlining a constructor doesn't help allocation unless the result is
605 -- scrutinised. UNLESS the constructor occurs just once, albeit possibly
606 -- in multiple case branches. Then inlining it doesn't increase allocation,
607 -- but it does increase the chance that the constructor won't be allocated at all
608 -- in the branches that don't use it.
610 enough_args = n_val_args >= n_vals_wanted
611 really_interesting_cont | n_val_args < n_vals_wanted = False -- Too few args
612 | n_val_args == n_vals_wanted = interesting_cont
613 | otherwise = True -- Extra args
614 -- really_interesting_cont tells if the result of the
615 -- call is in an interesting context.
617 small_enough = (size - discount) <= opt_UF_UseThreshold
618 discount = computeDiscount n_vals_wanted arg_discounts res_discount
619 arg_infos really_interesting_cont
623 if opt_D_dump_inlinings then
624 pprTrace "Considering inlining"
625 (ppr id <+> vcat [text "black listed" <+> ppr black_listed,
626 text "occ info:" <+> ppr occ,
627 text "arg infos" <+> ppr arg_infos,
628 text "interesting continuation" <+> ppr interesting_cont,
629 text "is value:" <+> ppr is_value,
630 text "is cheap:" <+> ppr is_cheap,
631 text "is bottom:" <+> ppr is_bot,
632 text "is top-level:" <+> ppr is_top,
633 text "guidance" <+> ppr guidance,
634 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO",
636 text "Unfolding =" <+> pprCoreExpr unf_template
644 computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Bool -> Int
645 computeDiscount n_vals_wanted arg_discounts res_discount arg_infos result_used
646 -- We multiple the raw discounts (args_discount and result_discount)
647 -- ty opt_UnfoldingKeenessFactor because the former have to do with
648 -- *size* whereas the discounts imply that there's some extra
649 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
652 -- we also discount 1 for each argument passed, because these will
653 -- reduce with the lambdas in the function (we count 1 for a lambda
655 = 1 + -- Discount of 1 because the result replaces the call
656 -- so we count 1 for the function itself
657 length (take n_vals_wanted arg_infos) +
658 -- Discount of 1 for each arg supplied, because the
659 -- result replaces the call
660 round (opt_UF_KeenessFactor *
661 fromInt (arg_discount + result_discount))
663 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
665 mk_arg_discount discount is_evald | is_evald = discount
668 -- Don't give a result discount unless there are enough args
669 result_discount | result_used = res_discount -- Over-applied, or case scrut
674 %************************************************************************
676 \subsection{Black-listing}
678 %************************************************************************
680 Inlining is controlled by the "Inline phase" number, which is set
681 by the per-simplification-pass '-finline-phase' flag.
683 For optimisation we use phase 1,2 and nothing (i.e. no -finline-phase flag)
684 in that order. The meanings of these are determined by the @blackListed@ function
687 The final simplification doesn't have a phase number
693 (least black listing, most inlining)
694 INLINE n foo phase is Just p *and* p<n *and* foo appears on LHS of rule
695 INLINE foo phase is Just p *and* foo appears on LHS of rule
696 NOINLINE n foo phase is Just p *and* (p<n *or* foo appears on LHS of rule)
698 (most black listing, least inlining)
701 blackListed :: IdSet -- Used in transformation rules
702 -> Maybe Int -- Inline phase
703 -> Id -> Bool -- True <=> blacklisted
705 -- The blackListed function sees whether a variable should *not* be
706 -- inlined because of the inline phase we are in. This is the sole
707 -- place that the inline phase number is looked at.
709 blackListed rule_vars Nothing -- Last phase
710 = \v -> case idInlinePragma v of
711 IMustNotBeINLINEd False Nothing -> True -- An unconditional NOINLINE pragma
714 blackListed rule_vars (Just phase)
715 = \v -> normal_case rule_vars phase v
717 normal_case rule_vars phase v
718 = case idInlinePragma v of
719 NoInlinePragInfo -> has_rules
721 IMustNotBeINLINEd from_INLINE Nothing
722 | from_INLINE -> has_rules -- Black list until final phase
723 | otherwise -> True -- Always blacklisted
725 IMustNotBeINLINEd from_inline (Just threshold)
726 | from_inline -> phase < threshold && has_rules
727 | otherwise -> phase < threshold || has_rules
729 has_rules = v `elemVarSet` rule_vars
730 || not (isEmptyCoreRules (idSpecialisation v))
734 SLPJ 95/04: Why @runST@ must be inlined very late:
738 (a, s') = newArray# 100 [] s
739 (_, s'') = fill_in_array_or_something a x s'
743 If we inline @runST@, we'll get:
746 (a, s') = newArray# 100 [] realWorld#{-NB-}
747 (_, s'') = fill_in_array_or_something a x s'
751 And now the @newArray#@ binding can be floated to become a CAF, which
752 is totally and utterly wrong:
755 (a, s') = newArray# 100 [] realWorld#{-NB-} -- YIKES!!!
758 let (_, s'') = fill_in_array_or_something a x s' in
761 All calls to @f@ will share a {\em single} array!
763 Yet we do want to inline runST sometime, so we can avoid
764 needless code. Solution: black list it until the last moment.