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, mkUnfolding, 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,
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(..) )
59 import TyCon ( tyConFamilySize )
60 import Type ( splitAlgTyConApp_maybe, splitFunTy_maybe, isUnLiftedType )
61 import Const ( isNoRepLit )
62 import Unique ( Unique, buildIdKey, augmentIdKey, runSTRepIdKey )
63 import Maybes ( maybeToBool )
65 import Util ( isIn, lengthExceeds )
69 %************************************************************************
71 \subsection{@Unfolding@ and @UnfoldingGuidance@ types}
73 %************************************************************************
79 | OtherCon [Con] -- It ain't one of these
80 -- (OtherCon xs) also indicates that something has been evaluated
81 -- and hence there's no point in re-evaluating it.
82 -- OtherCon [] is used even for non-data-type values
83 -- to indicated evaluated-ness. Notably:
84 -- data C = C !(Int -> Int)
85 -- case x of { C f -> ... }
86 -- Here, f gets an OtherCon [] unfolding.
88 | CoreUnfolding -- An unfolding with redundant cached information
89 CoreExpr -- Template; binder-info is correct
90 Bool -- exprIsCheap template (cached); it won't duplicate (much) work
91 -- if you inline this in more than one place
92 Bool -- exprIsValue template (cached); it is ok to discard a `seq` on
94 UnfoldingGuidance -- Tells about the *size* of the template.
96 seqUnfolding :: Unfolding -> ()
97 seqUnfolding (CoreUnfolding e b1 b2 g)
98 = seqExpr e `seq` b1 `seq` b2 `seq` seqGuidance g
99 seqUnfolding other = ()
103 noUnfolding = NoUnfolding
104 mkOtherCon = OtherCon
107 = CoreUnfolding (occurAnalyseGlobalExpr expr)
110 (calcUnfoldingGuidance opt_UF_CreationThreshold expr)
112 unfoldingTemplate :: Unfolding -> CoreExpr
113 unfoldingTemplate (CoreUnfolding expr _ _ _) = expr
114 unfoldingTemplate other = panic "getUnfoldingTemplate"
116 maybeUnfoldingTemplate :: Unfolding -> Maybe CoreExpr
117 maybeUnfoldingTemplate (CoreUnfolding expr _ _ _) = Just expr
118 maybeUnfoldingTemplate other = Nothing
120 otherCons (OtherCon cons) = cons
123 isEvaldUnfolding :: Unfolding -> Bool
124 isEvaldUnfolding (OtherCon _) = True
125 isEvaldUnfolding (CoreUnfolding _ _ is_evald _) = is_evald
126 isEvaldUnfolding other = False
128 isCheapUnfolding :: Unfolding -> Bool
129 isCheapUnfolding (CoreUnfolding _ is_cheap _ _) = is_cheap
130 isCheapUnfolding other = False
132 hasUnfolding :: Unfolding -> Bool
133 hasUnfolding (CoreUnfolding _ _ _ _) = True
134 hasUnfolding other = False
136 hasSomeUnfolding :: Unfolding -> Bool
137 hasSomeUnfolding NoUnfolding = False
138 hasSomeUnfolding other = True
140 data UnfoldingGuidance
142 | UnfoldAlways -- There is no "original" definition,
143 -- so you'd better unfold. Or: something
144 -- so cheap to unfold (e.g., 1#) that
145 -- you should do it absolutely always.
147 | UnfoldIfGoodArgs Int -- and "n" value args
149 [Int] -- Discount if the argument is evaluated.
150 -- (i.e., a simplification will definitely
151 -- be possible). One elt of the list per *value* arg.
153 Int -- The "size" of the unfolding; to be elaborated
156 Int -- Scrutinee discount: the discount to substract if the thing is in
157 -- a context (case (thing args) of ...),
158 -- (where there are the right number of arguments.)
160 seqGuidance (UnfoldIfGoodArgs n ns a b) = n `seq` sum ns `seq` a `seq` b `seq` ()
161 seqGuidance other = ()
165 instance Outputable UnfoldingGuidance where
166 ppr UnfoldAlways = ptext SLIT("ALWAYS")
167 ppr UnfoldNever = ptext SLIT("NEVER")
168 ppr (UnfoldIfGoodArgs v cs size discount)
169 = hsep [ptext SLIT("IF_ARGS"), int v,
170 if null cs -- always print *something*
172 else hcat (map (text . show) cs),
178 %************************************************************************
180 \subsection[calcUnfoldingGuidance]{Calculate ``unfolding guidance'' for an expression}
182 %************************************************************************
185 calcUnfoldingGuidance
186 :: Int -- bomb out if size gets bigger than this
187 -> CoreExpr -- expression to look at
189 calcUnfoldingGuidance bOMB_OUT_SIZE expr
190 | exprIsTrivial expr -- Often trivial expressions are never bound
191 -- to an expression, but it can happen. For
192 -- example, the Id for a nullary constructor has
193 -- a trivial expression as its unfolding, and
194 -- we want to make sure that we always unfold it.
198 = case collectBinders expr of { (binders, body) ->
200 val_binders = filter isId binders
202 case (sizeExpr bOMB_OUT_SIZE val_binders body) of
204 TooBig -> UnfoldNever
206 SizeIs size cased_args scrut_discount
209 (map discount_for val_binders)
215 | is_fun_ty = num_cases * opt_UF_FunAppDiscount
216 | is_data_ty = num_cases * opt_UF_ScrutConDiscount
217 | otherwise = num_cases * opt_UF_PrimArgDiscount
219 num_cases = foldlBag (\n b' -> if b==b' then n+1 else n) 0 cased_args
220 -- Count occurrences of b in cased_args
222 is_fun_ty = maybeToBool (splitFunTy_maybe arg_ty)
223 (is_data_ty, tycon) = case (splitAlgTyConApp_maybe (idType b)) of
224 Nothing -> (False, panic "discount")
225 Just (tc,_,_) -> (True, tc)
230 sizeExpr :: Int -- Bomb out if it gets bigger than this
231 -> [Id] -- Arguments; we're interested in which of these
236 sizeExpr (I# bOMB_OUT_SIZE) args expr
239 size_up (Type t) = sizeZero -- Types cost nothing
240 size_up (Var v) = sizeOne
242 size_up (Note InlineMe _) = sizeTwo -- The idea is that this is one more
243 -- than the size of the "call" (i.e. 1)
244 -- We want to reply "no" to noSizeIncrease
245 -- for a bare reference (i.e. applied to no args)
246 -- to an INLINE thing
248 size_up (Note _ body) = size_up body -- Notes cost nothing
250 size_up (App fun (Type t)) = size_up fun
251 size_up (App fun arg) = size_up_app fun [arg]
253 size_up (Con con args) = foldr (addSize . size_up)
254 (size_up_con con args)
257 size_up (Lam b e) | isId b = size_up e `addSizeN` 1
258 | otherwise = size_up e
260 size_up (Let (NonRec binder rhs) body)
261 = nukeScrutDiscount (size_up rhs) `addSize`
262 size_up body `addSizeN`
263 (if isUnLiftedType (idType binder) then 0 else 1)
264 -- For the allocation
265 -- If the binder has an unlifted type there is no allocation
267 size_up (Let (Rec pairs) body)
268 = nukeScrutDiscount rhs_size `addSize`
269 size_up body `addSizeN`
270 length pairs -- For the allocation
272 rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
274 size_up (Case scrut _ alts)
275 = nukeScrutDiscount (size_up scrut) `addSize`
276 arg_discount scrut `addSize`
277 foldr (addSize . size_up_alt) sizeZero alts
278 `addSizeN` 1 -- charge one for the case itself.
280 -- Just charge for the alts that exist, not the ones that might exist
282 -- case (splitAlgTyConApp_maybe (coreExprType scrut)) of
284 -- Just (tc,_,_) -> tyConFamilySize tc
287 size_up_app (App fun arg) args = size_up_app fun (arg:args)
288 size_up_app fun args = foldr (addSize . size_up) (fun_discount fun) args
290 -- A function application with at least one value argument
291 -- so if the function is an argument give it an arg-discount
292 -- Also behave specially if the function is a build
293 fun_discount (Var fun) | idUnique fun == buildIdKey = buildSize
294 | idUnique fun == augmentIdKey = augmentSize
295 | fun `is_elem` args = scrutArg fun
296 fun_discount other = sizeZero
299 size_up_alt (con, bndrs, rhs) = size_up rhs
300 -- Don't charge for args, so that wrappers look cheap
303 size_up_con (Literal lit) args | isNoRepLit lit = sizeN opt_UF_NoRepLit
304 | otherwise = sizeOne
306 size_up_con (DataCon dc) args = conSizeN (valArgCount args)
308 size_up_con (PrimOp op) args = foldr addSize (sizeN op_cost) (map arg_discount args)
309 -- Give an arg-discount if a primop is applies to
310 -- one of the function's arguments
312 op_cost | primOpIsDupable op = opt_UF_CheapOp
313 | otherwise = opt_UF_DearOp
315 -- We want to record if we're case'ing, or applying, an argument
316 arg_discount (Var v) | v `is_elem` args = scrutArg v
317 arg_discount other = sizeZero
320 is_elem :: Id -> [Id] -> Bool
321 is_elem = isIn "size_up_scrut"
324 -- These addSize things have to be here because
325 -- I don't want to give them bOMB_OUT_SIZE as an argument
327 addSizeN TooBig _ = TooBig
328 addSizeN (SizeIs n xs d) (I# m)
329 | n_tot -# d <# bOMB_OUT_SIZE = SizeIs n_tot xs d
334 addSize TooBig _ = TooBig
335 addSize _ TooBig = TooBig
336 addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
337 | (n_tot -# d_tot) <# bOMB_OUT_SIZE = SizeIs n_tot xys d_tot
342 xys = xs `unionBags` ys
345 Code for manipulating sizes
349 data ExprSize = TooBig
350 | SizeIs Int# -- Size found
351 (Bag Id) -- Arguments cased herein
352 Int# -- Size to subtract if result is scrutinised
353 -- by a case expression
355 sizeZero = SizeIs 0# emptyBag 0#
356 sizeOne = SizeIs 1# emptyBag 0#
357 sizeTwo = SizeIs 2# emptyBag 0#
358 sizeN (I# n) = SizeIs n emptyBag 0#
359 conSizeN (I# n) = SizeIs 1# emptyBag (n +# 1#)
360 -- Treat constructors as size 1, that unfoldAlways responsds 'False'
361 -- when asked about 'x' when x is bound to (C 3#).
362 -- This avoids gratuitous 'ticks' when x itself appears as an
363 -- atomic constructor argument.
365 buildSize = SizeIs (-2#) emptyBag 4#
366 -- We really want to inline applications of build
367 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
368 -- Indeed, we should add a result_discount becuause build is
369 -- very like a constructor. We don't bother to check that the
370 -- build is saturated (it usually is). The "-2" discounts for the \c n,
371 -- The "4" is rather arbitrary.
373 augmentSize = SizeIs (-2#) emptyBag 4#
374 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
375 -- e plus ys. The -2 accounts for the \cn
377 scrutArg v = SizeIs 0# (unitBag v) 0#
379 nukeScrutDiscount (SizeIs n vs d) = SizeIs n vs 0#
380 nukeScrutDiscount TooBig = TooBig
384 %************************************************************************
386 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
388 %************************************************************************
390 We have very limited information about an unfolding expression: (1)~so
391 many type arguments and so many value arguments expected---for our
392 purposes here, we assume we've got those. (2)~A ``size'' or ``cost,''
393 a single integer. (3)~An ``argument info'' vector. For this, what we
394 have at the moment is a Boolean per argument position that says, ``I
395 will look with great favour on an explicit constructor in this
396 position.'' (4)~The ``discount'' to subtract if the expression
397 is being scrutinised.
399 Assuming we have enough type- and value arguments (if not, we give up
400 immediately), then we see if the ``discounted size'' is below some
401 (semi-arbitrary) threshold. It works like this: for every argument
402 position where we're looking for a constructor AND WE HAVE ONE in our
403 hands, we get a (again, semi-arbitrary) discount [proportion to the
404 number of constructors in the type being scrutinized].
406 If we're in the context of a scrutinee ( \tr{(case <expr > of A .. -> ...;.. )})
407 and the expression in question will evaluate to a constructor, we use
408 the computed discount size *for the result only* rather than
409 computing the argument discounts. Since we know the result of
410 the expression is going to be taken apart, discounting its size
411 is more accurate (see @sizeExpr@ above for how this discount size
414 We use this one to avoid exporting inlinings that we ``couldn't possibly
415 use'' on the other side. Can be overridden w/ flaggery.
416 Just the same as smallEnoughToInline, except that it has no actual arguments.
419 couldBeSmallEnoughToInline :: UnfoldingGuidance -> Bool
420 couldBeSmallEnoughToInline UnfoldNever = False
421 couldBeSmallEnoughToInline other = True
423 certainlySmallEnoughToInline :: UnfoldingGuidance -> Bool
424 certainlySmallEnoughToInline UnfoldNever = False
425 certainlySmallEnoughToInline UnfoldAlways = True
426 certainlySmallEnoughToInline (UnfoldIfGoodArgs _ _ size _) = size <= opt_UF_UseThreshold
429 @okToUnfoldInHifile@ is used when emitting unfolding info into an interface
430 file to determine whether an unfolding candidate really should be unfolded.
431 The predicate is needed to prevent @_casm_@s (+ lit-lits) from being emitted
432 into interface files.
434 The reason for inlining expressions containing _casm_s into interface files
435 is that these fragments of C are likely to mention functions/#defines that
436 will be out-of-scope when inlined into another module. This is not an
437 unfixable problem for the user (just need to -#include the approp. header
438 file), but turning it off seems to the simplest thing to do.
441 okToUnfoldInHiFile :: CoreExpr -> Bool
442 okToUnfoldInHiFile e = opt_UnfoldCasms || go e
444 -- Race over an expression looking for CCalls..
446 go (Con (Literal lit) _) = not (isLitLitLit lit)
447 go (Con (PrimOp op) args) = okToUnfoldPrimOp op && all go args
448 go (Con con args) = True -- con args are always atomic
449 go (App fun arg) = go fun && go arg
450 go (Lam _ body) = go body
451 go (Let binds body) = and (map go (body :rhssOfBind binds))
452 go (Case scrut bndr alts) = and (map go (scrut:rhssOfAlts alts))
453 go (Note _ body) = go body
456 -- ok to unfold a PrimOp as long as it's not a _casm_
457 okToUnfoldPrimOp (CCallOp _ is_casm _ _) = not is_casm
458 okToUnfoldPrimOp _ = True
462 %************************************************************************
464 \subsection{callSiteInline}
466 %************************************************************************
468 This is the key function. It decides whether to inline a variable at a call site
470 callSiteInline is used at call sites, so it is a bit more generous.
471 It's a very important function that embodies lots of heuristics.
472 A non-WHNF can be inlined if it doesn't occur inside a lambda,
473 and occurs exactly once or
474 occurs once in each branch of a case and is small
476 If the thing is in WHNF, there's no danger of duplicating work,
477 so we can inline if it occurs once, or is small
480 callSiteInline :: Bool -- True <=> the Id is black listed
481 -> Bool -- 'inline' note at call site
483 -> [Bool] -- One for each value arg; True if it is interesting
484 -> Bool -- True <=> continuation is interesting
485 -> Maybe CoreExpr -- Unfolding, if any
488 callSiteInline black_listed inline_call id arg_infos interesting_cont
489 = case getIdUnfolding id of {
490 NoUnfolding -> Nothing ;
491 OtherCon _ -> Nothing ;
492 CoreUnfolding unf_template is_cheap _ guidance ->
495 result | yes_or_no = Just unf_template
496 | otherwise = Nothing
498 inline_prag = getInlinePragma id
499 n_val_args = length arg_infos
503 IAmDead -> pprTrace "callSiteInline: dead" (ppr id) False
504 IMustNotBeINLINEd -> False
505 IAmALoopBreaker -> False
506 IMustBeINLINEd -> True -- Overrides absolutely everything, including the black list
507 ICanSafelyBeINLINEd in_lam one_br -> consider in_lam True one_br
508 NoInlinePragInfo -> consider InsideLam False False
510 consider in_lam once once_in_one_branch
511 | black_listed = False
513 | once_in_one_branch -- Be very keen to inline something if this is its unique occurrence; that
514 -- gives a good chance of eliminating the original binding for the thing.
515 -- The only time we hold back is when substituting inside a lambda;
516 -- then if the context is totally uninteresting (not applied, not scrutinised)
517 -- there is no point in substituting because it might just increase allocation.
518 = WARN( case in_lam of { NotInsideLam -> True; other -> False },
519 text "callSiteInline:oneOcc" <+> ppr id )
520 -- If it has one occurrence, not inside a lambda, PreInlineUnconditionally
521 -- should have zapped it already
522 is_cheap && (not (null arg_infos) || interesting_cont)
524 | otherwise -- Occurs (textually) more than once, so look at its size
528 UnfoldIfGoodArgs n_vals_wanted arg_discounts size res_discount
529 | enough_args && size <= (n_vals_wanted + 1)
531 -- Size of call is n_vals_wanted (+1 for the function)
534 InsideLam -> is_cheap
536 | not (or arg_infos || really_interesting_cont || once)
537 -- If it occurs more than once, there must be something interesting
538 -- about some argument, or the result, to make it worth inlining
539 -- We also drop this case if the thing occurs once, although perhaps in
540 -- several branches. In this case we are keener about inlining in the hope
541 -- that we'll be able to drop the allocation for the function altogether.
546 NotInsideLam -> small_enough
547 InsideLam -> is_cheap && small_enough
550 enough_args = n_val_args >= n_vals_wanted
551 really_interesting_cont | n_val_args < n_vals_wanted = False -- Too few args
552 | n_val_args == n_vals_wanted = interesting_cont
553 | otherwise = True -- Extra args
554 -- This rather elaborate defn for really_interesting_cont is important
555 -- Consider an I# = INLINE (\x -> I# {x})
556 -- The unfolding guidance deems it to have size 2, and no arguments.
557 -- So in an application (I# y) we must take the extra arg 'y' as
558 -- evidence of an interesting context!
560 small_enough = (size - discount) <= opt_UF_UseThreshold
561 discount = computeDiscount n_vals_wanted arg_discounts res_discount
562 arg_infos really_interesting_cont
567 if opt_D_dump_inlinings then
568 pprTrace "Considering inlining"
569 (ppr id <+> vcat [text "black listed" <+> ppr black_listed,
570 text "inline prag:" <+> ppr inline_prag,
571 text "arg infos" <+> ppr arg_infos,
572 text "interesting continuation" <+> ppr interesting_cont,
573 text "is cheap" <+> ppr is_cheap,
574 text "guidance" <+> ppr guidance,
575 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO",
577 text "Unfolding =" <+> pprCoreExpr unf_template
585 computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Bool -> Int
586 computeDiscount n_vals_wanted arg_discounts res_discount arg_infos result_used
587 -- We multiple the raw discounts (args_discount and result_discount)
588 -- ty opt_UnfoldingKeenessFactor because the former have to do with
589 -- *size* whereas the discounts imply that there's some extra
590 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
593 -- we also discount 1 for each argument passed, because these will
594 -- reduce with the lambdas in the function (we count 1 for a lambda
596 = length (take n_vals_wanted arg_infos) +
597 -- Discount of 1 for each arg supplied, because the
598 -- result replaces the call
599 round (opt_UF_KeenessFactor *
600 fromInt (arg_discount + result_discount))
602 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
604 mk_arg_discount discount is_evald | is_evald = discount
607 -- Don't give a result discount unless there are enough args
608 result_discount | result_used = res_discount -- Over-applied, or case scrut
613 %************************************************************************
615 \subsection{Black-listing}
617 %************************************************************************
619 Inlining is controlled by the "Inline phase" number, which is set
620 by the per-simplification-pass '-finline-phase' flag.
622 For optimisation we use phase 1,2 and nothing (i.e. no -finline-phase flag)
623 in that order. The meanings of these are determined by the @blackListed@ function
627 blackListed :: IdSet -- Used in transformation rules
628 -> Maybe Int -- Inline phase
629 -> Id -> Bool -- True <=> blacklisted
631 -- The blackListed function sees whether a variable should *not* be
632 -- inlined because of the inline phase we are in. This is the sole
633 -- place that the inline phase number is looked at.
635 -- Phase 0: used for 'no imported inlinings please'
636 -- This prevents wrappers getting inlined which in turn is bad for full laziness
637 blackListed rule_vars (Just 0)
638 = \v -> not (isLocallyDefined v)
640 -- Phase 1: don't inline any rule-y things or things with specialisations
641 blackListed rule_vars (Just 1)
642 = \v -> let v_uniq = idUnique v
643 in v `elemVarSet` rule_vars
644 || not (isEmptyCoreRules (getIdSpecialisation v))
645 || v_uniq == runSTRepIdKey
647 -- Phase 2: allow build/augment to inline, and specialisations
648 blackListed rule_vars (Just 2)
649 = \v -> let v_uniq = idUnique v
650 in (v `elemVarSet` rule_vars && not (v_uniq == buildIdKey ||
651 v_uniq == augmentIdKey))
652 || v_uniq == runSTRepIdKey
654 -- Otherwise just go for it
655 blackListed rule_vars phase
660 SLPJ 95/04: Why @runST@ must be inlined very late:
664 (a, s') = newArray# 100 [] s
665 (_, s'') = fill_in_array_or_something a x s'
669 If we inline @runST@, we'll get:
672 (a, s') = newArray# 100 [] realWorld#{-NB-}
673 (_, s'') = fill_in_array_or_something a x s'
677 And now the @newArray#@ binding can be floated to become a CAF, which
678 is totally and utterly wrong:
681 (a, s') = newArray# 100 [] realWorld#{-NB-} -- YIKES!!!
684 let (_, s'') = fill_in_array_or_something a x s' in
687 All calls to @f@ will share a {\em single} array!
689 Yet we do want to inline runST sometime, so we can avoid
690 needless code. Solution: black list it until the last moment.