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, -- types
19 noUnfolding, mkTopUnfolding, mkUnfolding, mkCompulsoryUnfolding, seqUnfolding,
20 mkOtherCon, otherCons,
21 unfoldingTemplate, maybeUnfoldingTemplate,
22 isEvaldUnfolding, isCheapUnfolding,
23 hasUnfolding, hasSomeUnfolding,
25 couldBeSmallEnoughToInline,
26 certainlySmallEnoughToInline,
29 calcUnfoldingGuidance,
31 callSiteInline, blackListed
34 #include "HsVersions.h"
36 import CmdLineOpts ( opt_UF_CreationThreshold,
38 opt_UF_ScrutConDiscount,
39 opt_UF_FunAppDiscount,
40 opt_UF_PrimArgDiscount,
42 opt_UF_CheapOp, opt_UF_DearOp, opt_UF_NoRepLit,
43 opt_UnfoldCasms, opt_PprStyle_Debug,
47 import PprCore ( pprCoreExpr )
48 import OccurAnal ( occurAnalyseGlobalExpr )
50 import CoreUtils ( coreExprType, exprIsTrivial, exprIsValue, exprIsCheap )
51 import Id ( Id, idType, idUnique, isId, getIdWorkerInfo,
52 getIdSpecialisation, getInlinePragma, getIdUnfolding
55 import Name ( isLocallyDefined )
56 import Const ( Con(..), isLitLitLit, isWHNFCon )
57 import PrimOp ( PrimOp(..), primOpIsDupable )
58 import IdInfo ( ArityInfo(..), InlinePragInfo(..), OccInfo(..), insideLam, workerExists )
59 import TyCon ( tyConFamilySize )
60 import Type ( splitAlgTyConApp_maybe, splitFunTy_maybe, isUnLiftedType )
61 import Const ( isNoRepLit )
62 import Unique ( Unique, buildIdKey, augmentIdKey )
63 import Maybes ( maybeToBool )
65 import Util ( isIn, lengthExceeds )
68 #if __GLASGOW_HASKELL__ >= 404
69 import GlaExts ( fromInt )
73 %************************************************************************
75 \subsection{@Unfolding@ and @UnfoldingGuidance@ types}
77 %************************************************************************
83 | OtherCon [Con] -- It ain't one of these
84 -- (OtherCon xs) also indicates that something has been evaluated
85 -- and hence there's no point in re-evaluating it.
86 -- OtherCon [] is used even for non-data-type values
87 -- to indicated evaluated-ness. Notably:
88 -- data C = C !(Int -> Int)
89 -- case x of { C f -> ... }
90 -- Here, f gets an OtherCon [] unfolding.
92 | CompulsoryUnfolding CoreExpr -- There is no "original" definition,
93 -- so you'd better unfold.
95 | CoreUnfolding -- An unfolding with redundant cached information
96 CoreExpr -- Template; binder-info is correct
97 Bool -- This is a top-level binding
98 Bool -- exprIsCheap template (cached); it won't duplicate (much) work
99 -- if you inline this in more than one place
100 Bool -- exprIsValue template (cached); it is ok to discard a `seq` on
102 UnfoldingGuidance -- Tells about the *size* of the template.
104 seqUnfolding :: Unfolding -> ()
105 seqUnfolding (CoreUnfolding e top b1 b2 g)
106 = seqExpr e `seq` top `seq` b1 `seq` b2 `seq` seqGuidance g
107 seqUnfolding other = ()
111 noUnfolding = NoUnfolding
112 mkOtherCon = OtherCon
114 mkTopUnfolding expr = mkUnfolding True expr
116 mkUnfolding top_lvl expr
117 = CoreUnfolding (occurAnalyseGlobalExpr expr)
121 (calcUnfoldingGuidance opt_UF_CreationThreshold expr)
123 mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
124 = CompulsoryUnfolding (occurAnalyseGlobalExpr expr)
126 unfoldingTemplate :: Unfolding -> CoreExpr
127 unfoldingTemplate (CoreUnfolding expr _ _ _ _) = expr
128 unfoldingTemplate (CompulsoryUnfolding expr) = expr
129 unfoldingTemplate other = panic "getUnfoldingTemplate"
131 maybeUnfoldingTemplate :: Unfolding -> Maybe CoreExpr
132 maybeUnfoldingTemplate (CoreUnfolding expr _ _ _ _) = Just expr
133 maybeUnfoldingTemplate (CompulsoryUnfolding expr) = Just expr
134 maybeUnfoldingTemplate other = Nothing
136 otherCons (OtherCon cons) = cons
139 isEvaldUnfolding :: Unfolding -> Bool
140 isEvaldUnfolding (OtherCon _) = True
141 isEvaldUnfolding (CoreUnfolding _ _ _ is_evald _) = is_evald
142 isEvaldUnfolding other = False
144 isCheapUnfolding :: Unfolding -> Bool
145 isCheapUnfolding (CoreUnfolding _ _ is_cheap _ _) = is_cheap
146 isCheapUnfolding other = False
148 hasUnfolding :: Unfolding -> Bool
149 hasUnfolding (CoreUnfolding _ _ _ _ _) = True
150 hasUnfolding (CompulsoryUnfolding _) = True
151 hasUnfolding other = False
153 hasSomeUnfolding :: Unfolding -> Bool
154 hasSomeUnfolding NoUnfolding = False
155 hasSomeUnfolding other = True
157 data UnfoldingGuidance
159 | UnfoldIfGoodArgs Int -- and "n" value args
161 [Int] -- Discount if the argument is evaluated.
162 -- (i.e., a simplification will definitely
163 -- be possible). One elt of the list per *value* arg.
165 Int -- The "size" of the unfolding; to be elaborated
168 Int -- Scrutinee discount: the discount to substract if the thing is in
169 -- a context (case (thing args) of ...),
170 -- (where there are the right number of arguments.)
172 seqGuidance (UnfoldIfGoodArgs n ns a b) = n `seq` sum ns `seq` a `seq` b `seq` ()
173 seqGuidance other = ()
177 instance Outputable UnfoldingGuidance where
178 ppr UnfoldNever = ptext SLIT("NEVER")
179 ppr (UnfoldIfGoodArgs v cs size discount)
180 = hsep [ ptext SLIT("IF_ARGS"), int v,
181 brackets (hsep (map int cs)),
187 %************************************************************************
189 \subsection[calcUnfoldingGuidance]{Calculate ``unfolding guidance'' for an expression}
191 %************************************************************************
194 calcUnfoldingGuidance
195 :: Int -- bomb out if size gets bigger than this
196 -> CoreExpr -- expression to look at
198 calcUnfoldingGuidance bOMB_OUT_SIZE expr
199 = case collect_val_bndrs expr of { (inline, val_binders, body) ->
201 n_val_binders = length val_binders
203 case (sizeExpr bOMB_OUT_SIZE val_binders body) of
206 | not inline -> UnfoldNever
207 -- A big function with an INLINE pragma must
208 -- have an UnfoldIfGoodArgs guidance
209 | inline -> UnfoldIfGoodArgs n_val_binders
210 (map (const 0) val_binders)
211 (n_val_binders + 2) 0
212 -- See comments with final_size below
214 SizeIs size cased_args scrut_discount
217 (map discount_for val_binders)
223 final_size | inline = 0 -- Trying very agresssive inlining of INLINE things.
224 -- Reason: we don't want to call the un-inlined version,
225 -- because its body is awful
226 -- boxed_size `min` (n_val_binders + 2) -- Trying "+2" again...
227 | otherwise = boxed_size
228 -- The idea is that if there is an INLINE pragma (inline is True)
229 -- and there's a big body, we give a size of n_val_binders+1. This
230 -- This is enough to pass the no-size-increase test in callSiteInline,
232 -- I tried n_val_binders+2, to just defeat the test, on the grounds that
233 -- we don't want to inline an INLINE thing into a totally boring context,
234 -- but I found that some wrappers (notably one for a join point) weren't
235 -- getting inlined, and that was terrible. In that particular case, the
236 -- call site applied the wrapper to realWorld#, so if we made that an
237 -- "interesting" value the inlining would have happened... but it was
238 -- simpler to inline wrappers a little more eagerly instead.
240 -- Sometimes, though, an INLINE thing is smaller than n_val_binders+2.
241 -- A particular case in point is a constructor, which has size 1.
242 -- We want to inline this regardless, hence the `min`
246 | is_fun_ty = num_cases * opt_UF_FunAppDiscount
247 | is_data_ty = num_cases * opt_UF_ScrutConDiscount
248 | otherwise = num_cases * opt_UF_PrimArgDiscount
250 num_cases = foldlBag (\n b' -> if b==b' then n+1 else n) 0 cased_args
251 -- Count occurrences of b in cased_args
253 is_fun_ty = maybeToBool (splitFunTy_maybe arg_ty)
254 (is_data_ty, tycon) = case (splitAlgTyConApp_maybe (idType b)) of
255 Nothing -> (False, panic "discount")
256 Just (tc,_,_) -> (True, tc)
260 collect_val_bndrs e = go False [] e
261 -- We need to be a bit careful about how we collect the
262 -- value binders. In ptic, if we see
263 -- __inline_me (\x y -> e)
264 -- We want to say "2 value binders". Why? So that
265 -- we take account of information given for the arguments
267 go inline rev_vbs (Note InlineMe e) = go True rev_vbs e
268 go inline rev_vbs (Lam b e) | isId b = go inline (b:rev_vbs) e
269 | otherwise = go inline rev_vbs e
270 go inline rev_vbs e = (inline, reverse rev_vbs, e)
274 sizeExpr :: Int -- Bomb out if it gets bigger than this
275 -> [Id] -- Arguments; we're interested in which of these
280 sizeExpr (I# bOMB_OUT_SIZE) args expr
283 size_up (Type t) = sizeZero -- Types cost nothing
284 size_up (Var v) = sizeOne
286 size_up (Note _ body) = size_up body -- Notes cost nothing
288 size_up (App fun (Type t)) = size_up fun
289 size_up (App fun arg) = size_up_app fun [arg]
291 size_up (Con con args) = foldr (addSize . size_up)
292 (size_up_con con args)
295 size_up (Lam b e) | isId b = size_up e `addSizeN` 1
296 | otherwise = size_up e
298 size_up (Let (NonRec binder rhs) body)
299 = nukeScrutDiscount (size_up rhs) `addSize`
300 size_up body `addSizeN`
301 (if isUnLiftedType (idType binder) then 0 else 1)
302 -- For the allocation
303 -- If the binder has an unlifted type there is no allocation
305 size_up (Let (Rec pairs) body)
306 = nukeScrutDiscount rhs_size `addSize`
307 size_up body `addSizeN`
308 length pairs -- For the allocation
310 rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
312 size_up (Case scrut _ alts)
313 = nukeScrutDiscount (size_up scrut) `addSize`
314 arg_discount scrut `addSize`
315 foldr (addSize . size_up_alt) sizeZero alts
316 `addSizeN` 1 -- charge one for the case itself.
318 -- Just charge for the alts that exist, not the ones that might exist
320 -- case (splitAlgTyConApp_maybe (coreExprType scrut)) of
322 -- Just (tc,_,_) -> tyConFamilySize tc
325 size_up_app (App fun arg) args = size_up_app fun (arg:args)
326 size_up_app fun args = foldr (addSize . nukeScrutDiscount . size_up)
330 -- A function application with at least one value argument
331 -- so if the function is an argument give it an arg-discount
332 -- Also behave specially if the function is a build
333 size_up_fun (Var fun) | idUnique fun == buildIdKey = buildSize
334 | idUnique fun == augmentIdKey = augmentSize
335 | fun `is_elem` args = scrutArg fun `addSize` sizeOne
336 size_up_fun other = size_up other
339 size_up_alt (con, bndrs, rhs) = size_up rhs
340 -- Don't charge for args, so that wrappers look cheap
343 size_up_con (Literal lit) args | isNoRepLit lit = sizeN opt_UF_NoRepLit
344 | otherwise = sizeOne
346 size_up_con (DataCon dc) args = conSizeN (valArgCount args)
348 size_up_con (PrimOp op) args = foldr addSize (sizeN op_cost) (map arg_discount args)
349 -- Give an arg-discount if a primop is applies to
350 -- one of the function's arguments
352 op_cost | primOpIsDupable op = opt_UF_CheapOp
353 | otherwise = opt_UF_DearOp
355 -- We want to record if we're case'ing, or applying, an argument
356 arg_discount (Var v) | v `is_elem` args = scrutArg v
357 arg_discount other = sizeZero
360 is_elem :: Id -> [Id] -> Bool
361 is_elem = isIn "size_up_scrut"
364 -- These addSize things have to be here because
365 -- I don't want to give them bOMB_OUT_SIZE as an argument
367 addSizeN TooBig _ = TooBig
368 addSizeN (SizeIs n xs d) (I# m)
369 | n_tot -# d <# bOMB_OUT_SIZE = SizeIs n_tot xs d
374 addSize TooBig _ = TooBig
375 addSize _ TooBig = TooBig
376 addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
377 | (n_tot -# d_tot) <# bOMB_OUT_SIZE = SizeIs n_tot xys d_tot
382 xys = xs `unionBags` ys
385 Code for manipulating sizes
389 data ExprSize = TooBig
390 | SizeIs Int# -- Size found
391 (Bag Id) -- Arguments cased herein
392 Int# -- Size to subtract if result is scrutinised
393 -- by a case expression
395 sizeZero = SizeIs 0# emptyBag 0#
396 sizeOne = SizeIs 1# emptyBag 0#
397 sizeTwo = SizeIs 2# emptyBag 0#
398 sizeN (I# n) = SizeIs n emptyBag 0#
399 conSizeN (I# n) = SizeIs 1# emptyBag (n +# 1#)
400 -- Treat constructors as size 1, that unfoldAlways responsds 'False'
401 -- when asked about 'x' when x is bound to (C 3#).
402 -- This avoids gratuitous 'ticks' when x itself appears as an
403 -- atomic constructor argument.
405 buildSize = SizeIs (-2#) emptyBag 4#
406 -- We really want to inline applications of build
407 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
408 -- Indeed, we should add a result_discount becuause build is
409 -- very like a constructor. We don't bother to check that the
410 -- build is saturated (it usually is). The "-2" discounts for the \c n,
411 -- The "4" is rather arbitrary.
413 augmentSize = SizeIs (-2#) emptyBag 4#
414 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
415 -- e plus ys. The -2 accounts for the \cn
417 scrutArg v = SizeIs 0# (unitBag v) 0#
419 nukeScrutDiscount (SizeIs n vs d) = SizeIs n vs 0#
420 nukeScrutDiscount TooBig = TooBig
424 %************************************************************************
426 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
428 %************************************************************************
430 We have very limited information about an unfolding expression: (1)~so
431 many type arguments and so many value arguments expected---for our
432 purposes here, we assume we've got those. (2)~A ``size'' or ``cost,''
433 a single integer. (3)~An ``argument info'' vector. For this, what we
434 have at the moment is a Boolean per argument position that says, ``I
435 will look with great favour on an explicit constructor in this
436 position.'' (4)~The ``discount'' to subtract if the expression
437 is being scrutinised.
439 Assuming we have enough type- and value arguments (if not, we give up
440 immediately), then we see if the ``discounted size'' is below some
441 (semi-arbitrary) threshold. It works like this: for every argument
442 position where we're looking for a constructor AND WE HAVE ONE in our
443 hands, we get a (again, semi-arbitrary) discount [proportion to the
444 number of constructors in the type being scrutinized].
446 If we're in the context of a scrutinee ( \tr{(case <expr > of A .. -> ...;.. )})
447 and the expression in question will evaluate to a constructor, we use
448 the computed discount size *for the result only* rather than
449 computing the argument discounts. Since we know the result of
450 the expression is going to be taken apart, discounting its size
451 is more accurate (see @sizeExpr@ above for how this discount size
454 We use this one to avoid exporting inlinings that we ``couldn't possibly
455 use'' on the other side. Can be overridden w/ flaggery.
456 Just the same as smallEnoughToInline, except that it has no actual arguments.
459 couldBeSmallEnoughToInline :: UnfoldingGuidance -> Bool
460 couldBeSmallEnoughToInline UnfoldNever = False
461 couldBeSmallEnoughToInline other = True
463 certainlySmallEnoughToInline :: UnfoldingGuidance -> Bool
464 certainlySmallEnoughToInline UnfoldNever = False
465 certainlySmallEnoughToInline (UnfoldIfGoodArgs _ _ size _) = size <= opt_UF_UseThreshold
468 @okToUnfoldInHifile@ is used when emitting unfolding info into an interface
469 file to determine whether an unfolding candidate really should be unfolded.
470 The predicate is needed to prevent @_casm_@s (+ lit-lits) from being emitted
471 into interface files.
473 The reason for inlining expressions containing _casm_s into interface files
474 is that these fragments of C are likely to mention functions/#defines that
475 will be out-of-scope when inlined into another module. This is not an
476 unfixable problem for the user (just need to -#include the approp. header
477 file), but turning it off seems to the simplest thing to do.
480 okToUnfoldInHiFile :: CoreExpr -> Bool
481 okToUnfoldInHiFile e = opt_UnfoldCasms || go e
483 -- Race over an expression looking for CCalls..
485 go (Con (Literal lit) _) = not (isLitLitLit lit)
486 go (Con (PrimOp op) args) = okToUnfoldPrimOp op && all go args
487 go (Con con args) = True -- con args are always atomic
488 go (App fun arg) = go fun && go arg
489 go (Lam _ body) = go body
490 go (Let binds body) = and (map go (body :rhssOfBind binds))
491 go (Case scrut bndr alts) = and (map go (scrut:rhssOfAlts alts))
492 go (Note _ body) = go body
495 -- ok to unfold a PrimOp as long as it's not a _casm_
496 okToUnfoldPrimOp (CCallOp _ is_casm _ _) = not is_casm
497 okToUnfoldPrimOp _ = True
501 %************************************************************************
503 \subsection{callSiteInline}
505 %************************************************************************
507 This is the key function. It decides whether to inline a variable at a call site
509 callSiteInline is used at call sites, so it is a bit more generous.
510 It's a very important function that embodies lots of heuristics.
511 A non-WHNF can be inlined if it doesn't occur inside a lambda,
512 and occurs exactly once or
513 occurs once in each branch of a case and is small
515 If the thing is in WHNF, there's no danger of duplicating work,
516 so we can inline if it occurs once, or is small
519 callSiteInline :: Bool -- True <=> the Id is black listed
520 -> Bool -- 'inline' note at call site
523 -> [Bool] -- One for each value arg; True if it is interesting
524 -> Bool -- True <=> continuation is interesting
525 -> Maybe CoreExpr -- Unfolding, if any
528 callSiteInline black_listed inline_call occ id arg_infos interesting_cont
529 = case getIdUnfolding id of {
530 NoUnfolding -> Nothing ;
531 OtherCon _ -> Nothing ;
532 CompulsoryUnfolding unf_template -> Just unf_template ;
533 CoreUnfolding unf_template is_top is_cheap _ guidance ->
536 result | yes_or_no = Just unf_template
537 | otherwise = Nothing
539 n_val_args = length arg_infos
542 | black_listed = False
543 | otherwise = case occ of
544 IAmDead -> pprTrace "callSiteInline: dead" (ppr id) False
545 IAmALoopBreaker -> False
546 OneOcc in_lam one_br -> (not in_lam || is_cheap) && consider_safe in_lam True one_br
547 NoOccInfo -> is_cheap && consider_safe True False False
549 consider_safe in_lam once once_in_one_branch
550 -- consider_safe decides whether it's a good idea to inline something,
551 -- given that there's no work-duplication issue (the caller checks that).
552 -- once_in_one_branch = True means there's a unique textual occurrence
555 | once_in_one_branch -- Be very keen to inline something if this is its unique occurrence; that
556 -- gives a good chance of eliminating the original binding for the thing.
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 = 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 cheap" <+> ppr is_cheap,
611 text "guidance" <+> ppr guidance,
612 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO",
614 text "Unfolding =" <+> pprCoreExpr unf_template
622 computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Bool -> Int
623 computeDiscount n_vals_wanted arg_discounts res_discount arg_infos result_used
624 -- We multiple the raw discounts (args_discount and result_discount)
625 -- ty opt_UnfoldingKeenessFactor because the former have to do with
626 -- *size* whereas the discounts imply that there's some extra
627 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
630 -- we also discount 1 for each argument passed, because these will
631 -- reduce with the lambdas in the function (we count 1 for a lambda
633 = 1 + -- Discount of 1 because the result replaces the call
634 -- so we count 1 for the function itself
635 length (take n_vals_wanted arg_infos) +
636 -- Discount of 1 for each arg supplied, because the
637 -- result replaces the call
638 round (opt_UF_KeenessFactor *
639 fromInt (arg_discount + result_discount))
641 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
643 mk_arg_discount discount is_evald | is_evald = discount
646 -- Don't give a result discount unless there are enough args
647 result_discount | result_used = res_discount -- Over-applied, or case scrut
652 %************************************************************************
654 \subsection{Black-listing}
656 %************************************************************************
658 Inlining is controlled by the "Inline phase" number, which is set
659 by the per-simplification-pass '-finline-phase' flag.
661 For optimisation we use phase 1,2 and nothing (i.e. no -finline-phase flag)
662 in that order. The meanings of these are determined by the @blackListed@ function
665 The final simplification doesn't have a phase number
671 (least black listing, most inlining)
672 INLINE n foo phase is Just p *and* p<n *and* foo appears on LHS of rule
673 INLINE foo phase is Just p *and* foo appears on LHS of rule
674 NOINLINE n foo phase is Just p *and* (p<n *or* foo appears on LHS of rule)
676 (most black listing, least inlining)
679 blackListed :: IdSet -- Used in transformation rules
680 -> Maybe Int -- Inline phase
681 -> Id -> Bool -- True <=> blacklisted
683 -- The blackListed function sees whether a variable should *not* be
684 -- inlined because of the inline phase we are in. This is the sole
685 -- place that the inline phase number is looked at.
687 blackListed rule_vars Nothing -- Last phase
688 = \v -> case getInlinePragma v of
689 IMustNotBeINLINEd False Nothing -> True -- An unconditional NOINLINE pragma
692 blackListed rule_vars (Just 0)
693 -- Phase 0: used for 'no imported inlinings please'
694 -- This prevents wrappers getting inlined which in turn is bad for full laziness
695 -- NEW: try using 'not a wrapper' rather than 'not imported' in this phase.
696 -- This allows a little more inlining, which seems to be important, sometimes.
697 -- For example PrelArr.newIntArr gets better.
698 = \v -> -- workerExists (getIdWorkerInfo v) || normal_case rule_vars 0 v
699 -- True -- Try going back to no inlinings at all
700 -- BUT: I found that there is some advantage in doing
701 -- local inlinings first. For example in fish/Main.hs
702 -- it's advantageous to inline scale_vec2 before inlining
703 -- wrappers from PrelNum that make it look big.
704 not (isLocallyDefined v) -- This seems best at the moment
706 blackListed rule_vars (Just phase)
707 = \v -> normal_case rule_vars phase v
709 normal_case rule_vars phase v
710 = case getInlinePragma 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 (getIdSpecialisation 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.