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,
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.
98 noUnfolding = NoUnfolding
102 = CoreUnfolding (occurAnalyseGlobalExpr expr)
105 (calcUnfoldingGuidance opt_UF_CreationThreshold expr)
107 unfoldingTemplate :: Unfolding -> CoreExpr
108 unfoldingTemplate (CoreUnfolding expr _ _ _) = expr
109 unfoldingTemplate other = panic "getUnfoldingTemplate"
111 maybeUnfoldingTemplate :: Unfolding -> Maybe CoreExpr
112 maybeUnfoldingTemplate (CoreUnfolding expr _ _ _) = Just expr
113 maybeUnfoldingTemplate other = Nothing
115 otherCons (OtherCon cons) = cons
118 isEvaldUnfolding :: Unfolding -> Bool
119 isEvaldUnfolding (OtherCon _) = True
120 isEvaldUnfolding (CoreUnfolding _ _ is_evald _) = is_evald
121 isEvaldUnfolding other = False
123 isCheapUnfolding :: Unfolding -> Bool
124 isCheapUnfolding (CoreUnfolding _ is_cheap _ _) = is_cheap
125 isCheapUnfolding other = False
127 hasUnfolding :: Unfolding -> Bool
128 hasUnfolding (CoreUnfolding _ _ _ _) = True
129 hasUnfolding other = False
131 hasSomeUnfolding :: Unfolding -> Bool
132 hasSomeUnfolding NoUnfolding = False
133 hasSomeUnfolding other = True
135 data UnfoldingGuidance
137 | UnfoldAlways -- There is no "original" definition,
138 -- so you'd better unfold. Or: something
139 -- so cheap to unfold (e.g., 1#) that
140 -- you should do it absolutely always.
142 | UnfoldIfGoodArgs Int -- and "n" value args
144 [Int] -- Discount if the argument is evaluated.
145 -- (i.e., a simplification will definitely
146 -- be possible). One elt of the list per *value* arg.
148 Int -- The "size" of the unfolding; to be elaborated
151 Int -- Scrutinee discount: the discount to substract if the thing is in
152 -- a context (case (thing args) of ...),
153 -- (where there are the right number of arguments.)
157 instance Outputable UnfoldingGuidance where
158 ppr UnfoldAlways = ptext SLIT("ALWAYS")
159 ppr UnfoldNever = ptext SLIT("NEVER")
160 ppr (UnfoldIfGoodArgs v cs size discount)
161 = hsep [ptext SLIT("IF_ARGS"), int v,
162 if null cs -- always print *something*
164 else hcat (map (text . show) cs),
170 %************************************************************************
172 \subsection[calcUnfoldingGuidance]{Calculate ``unfolding guidance'' for an expression}
174 %************************************************************************
177 calcUnfoldingGuidance
178 :: Int -- bomb out if size gets bigger than this
179 -> CoreExpr -- expression to look at
181 calcUnfoldingGuidance bOMB_OUT_SIZE expr
182 | exprIsTrivial expr -- Often trivial expressions are never bound
183 -- to an expression, but it can happen. For
184 -- example, the Id for a nullary constructor has
185 -- a trivial expression as its unfolding, and
186 -- we want to make sure that we always unfold it.
190 = case collectBinders expr of { (binders, body) ->
192 val_binders = filter isId binders
194 case (sizeExpr bOMB_OUT_SIZE val_binders body) of
196 TooBig -> UnfoldNever
198 SizeIs size cased_args scrut_discount
201 (map discount_for val_binders)
207 | is_fun_ty = num_cases * opt_UF_FunAppDiscount
208 | is_data_ty = num_cases * tyConFamilySize tycon * opt_UF_ScrutConDiscount
209 | otherwise = num_cases * opt_UF_PrimArgDiscount
211 num_cases = foldlBag (\n b' -> if b==b' then n+1 else n) 0 cased_args
212 -- Count occurrences of b in cased_args
214 is_fun_ty = maybeToBool (splitFunTy_maybe arg_ty)
215 (is_data_ty, tycon) = case (splitAlgTyConApp_maybe (idType b)) of
216 Nothing -> (False, panic "discount")
217 Just (tc,_,_) -> (True, tc)
222 sizeExpr :: Int -- Bomb out if it gets bigger than this
223 -> [Id] -- Arguments; we're interested in which of these
228 sizeExpr (I# bOMB_OUT_SIZE) args expr
231 size_up (Type t) = sizeZero -- Types cost nothing
232 size_up (Var v) = sizeOne
234 size_up (Note InlineMe _) = sizeTwo -- The idea is that this is one more
235 -- than the size of the "call" (i.e. 1)
236 -- We want to reply "no" to noSizeIncrease
237 -- for a bare reference (i.e. applied to no args)
238 -- to an INLINE thing
240 size_up (Note _ body) = size_up body -- Notes cost nothing
242 size_up (App fun (Type t)) = size_up fun
243 size_up (App fun arg) = size_up_app fun [arg]
245 size_up (Con con args) = foldr (addSize . size_up)
246 (size_up_con con args)
249 size_up (Lam b e) | isId b = size_up e `addSizeN` 1
250 | otherwise = size_up e
252 size_up (Let (NonRec binder rhs) body)
253 = nukeScrutDiscount (size_up rhs) `addSize`
254 size_up body `addSizeN`
255 (if isUnLiftedType (idType binder) then 0 else 1)
256 -- For the allocation
257 -- If the binder has an unlifted type there is no allocation
259 size_up (Let (Rec pairs) body)
260 = nukeScrutDiscount rhs_size `addSize`
261 size_up body `addSizeN`
262 length pairs -- For the allocation
264 rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
266 size_up (Case scrut _ alts)
267 = nukeScrutDiscount (size_up scrut) `addSize`
268 arg_discount scrut `addSize`
269 foldr (addSize . size_up_alt) sizeZero alts
271 -- Just charge for the alts that exist, not the ones that might exist
273 -- case (splitAlgTyConApp_maybe (coreExprType scrut)) of
275 -- Just (tc,_,_) -> tyConFamilySize tc
278 size_up_app (App fun arg) args = size_up_app fun (arg:args)
279 size_up_app fun args = foldr (addSize . size_up) (fun_discount fun) args
281 -- A function application with at least one value argument
282 -- so if the function is an argument give it an arg-discount
283 -- Also behave specially if the function is a build
284 fun_discount (Var fun) | idUnique fun == buildIdKey = buildSize
285 | idUnique fun == augmentIdKey = augmentSize
286 | fun `is_elem` args = scrutArg fun
287 fun_discount other = sizeZero
290 size_up_alt (con, bndrs, rhs) = size_up rhs
291 -- Don't charge for args, so that wrappers look cheap
294 size_up_con (Literal lit) args | isNoRepLit lit = sizeN opt_UF_NoRepLit
295 | otherwise = sizeOne
297 size_up_con (DataCon dc) args = conSizeN (valArgCount args)
299 size_up_con (PrimOp op) args = foldr addSize (sizeN op_cost) (map arg_discount args)
300 -- Give an arg-discount if a primop is applies to
301 -- one of the function's arguments
303 op_cost | primOpIsDupable op = opt_UF_CheapOp
304 | otherwise = opt_UF_DearOp
306 -- We want to record if we're case'ing, or applying, an argument
307 arg_discount (Var v) | v `is_elem` args = scrutArg v
308 arg_discount other = sizeZero
311 is_elem :: Id -> [Id] -> Bool
312 is_elem = isIn "size_up_scrut"
315 -- These addSize things have to be here because
316 -- I don't want to give them bOMB_OUT_SIZE as an argument
318 addSizeN TooBig _ = TooBig
319 addSizeN (SizeIs n xs d) (I# m)
320 | n_tot -# d <# bOMB_OUT_SIZE = SizeIs n_tot xs d
325 addSize TooBig _ = TooBig
326 addSize _ TooBig = TooBig
327 addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
328 | (n_tot -# d_tot) <# bOMB_OUT_SIZE = SizeIs n_tot xys d_tot
333 xys = xs `unionBags` ys
336 Code for manipulating sizes
340 data ExprSize = TooBig
341 | SizeIs Int# -- Size found
342 (Bag Id) -- Arguments cased herein
343 Int# -- Size to subtract if result is scrutinised
344 -- by a case expression
346 sizeZero = SizeIs 0# emptyBag 0#
347 sizeOne = SizeIs 1# emptyBag 0#
348 sizeTwo = SizeIs 2# emptyBag 0#
349 sizeN (I# n) = SizeIs n emptyBag 0#
350 conSizeN (I# n) = SizeIs 1# emptyBag (n +# 1#)
351 -- Treat constructors as size 1, that unfoldAlways responsds 'False'
352 -- when asked about 'x' when x is bound to (C 3#).
353 -- This avoids gratuitous 'ticks' when x itself appears as an
354 -- atomic constructor argument.
356 buildSize = SizeIs (-2#) emptyBag 4#
357 -- We really want to inline applications of build
358 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
359 -- Indeed, we should add a result_discount becuause build is
360 -- very like a constructor. We don't bother to check that the
361 -- build is saturated (it usually is). The "-2" discounts for the \c n,
362 -- The "4" is rather arbitrary.
364 augmentSize = SizeIs (-2#) emptyBag 4#
365 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
366 -- e plus ys. The -2 accounts for the \cn
368 scrutArg v = SizeIs 0# (unitBag v) 0#
370 nukeScrutDiscount (SizeIs n vs d) = SizeIs n vs 0#
371 nukeScrutDiscount TooBig = TooBig
375 %************************************************************************
377 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
379 %************************************************************************
381 We have very limited information about an unfolding expression: (1)~so
382 many type arguments and so many value arguments expected---for our
383 purposes here, we assume we've got those. (2)~A ``size'' or ``cost,''
384 a single integer. (3)~An ``argument info'' vector. For this, what we
385 have at the moment is a Boolean per argument position that says, ``I
386 will look with great favour on an explicit constructor in this
387 position.'' (4)~The ``discount'' to subtract if the expression
388 is being scrutinised.
390 Assuming we have enough type- and value arguments (if not, we give up
391 immediately), then we see if the ``discounted size'' is below some
392 (semi-arbitrary) threshold. It works like this: for every argument
393 position where we're looking for a constructor AND WE HAVE ONE in our
394 hands, we get a (again, semi-arbitrary) discount [proportion to the
395 number of constructors in the type being scrutinized].
397 If we're in the context of a scrutinee ( \tr{(case <expr > of A .. -> ...;.. )})
398 and the expression in question will evaluate to a constructor, we use
399 the computed discount size *for the result only* rather than
400 computing the argument discounts. Since we know the result of
401 the expression is going to be taken apart, discounting its size
402 is more accurate (see @sizeExpr@ above for how this discount size
405 We use this one to avoid exporting inlinings that we ``couldn't possibly
406 use'' on the other side. Can be overridden w/ flaggery.
407 Just the same as smallEnoughToInline, except that it has no actual arguments.
410 couldBeSmallEnoughToInline :: UnfoldingGuidance -> Bool
411 couldBeSmallEnoughToInline UnfoldNever = False
412 couldBeSmallEnoughToInline other = True
414 certainlySmallEnoughToInline :: UnfoldingGuidance -> Bool
415 certainlySmallEnoughToInline UnfoldNever = False
416 certainlySmallEnoughToInline UnfoldAlways = True
417 certainlySmallEnoughToInline (UnfoldIfGoodArgs _ _ size _) = size <= opt_UF_UseThreshold
420 @okToUnfoldInHifile@ is used when emitting unfolding info into an interface
421 file to determine whether an unfolding candidate really should be unfolded.
422 The predicate is needed to prevent @_casm_@s (+ lit-lits) from being emitted
423 into interface files.
425 The reason for inlining expressions containing _casm_s into interface files
426 is that these fragments of C are likely to mention functions/#defines that
427 will be out-of-scope when inlined into another module. This is not an
428 unfixable problem for the user (just need to -#include the approp. header
429 file), but turning it off seems to the simplest thing to do.
432 okToUnfoldInHiFile :: CoreExpr -> Bool
433 okToUnfoldInHiFile e = opt_UnfoldCasms || go e
435 -- Race over an expression looking for CCalls..
437 go (Con (Literal lit) _) = not (isLitLitLit lit)
438 go (Con (PrimOp op) args) = okToUnfoldPrimOp op && all go args
439 go (Con con args) = True -- con args are always atomic
440 go (App fun arg) = go fun && go arg
441 go (Lam _ body) = go body
442 go (Let binds body) = and (map go (body :rhssOfBind binds))
443 go (Case scrut bndr alts) = and (map go (scrut:rhssOfAlts alts))
444 go (Note _ body) = go body
447 -- ok to unfold a PrimOp as long as it's not a _casm_
448 okToUnfoldPrimOp (CCallOp _ is_casm _ _) = not is_casm
449 okToUnfoldPrimOp _ = True
453 %************************************************************************
455 \subsection{callSiteInline}
457 %************************************************************************
459 This is the key function. It decides whether to inline a variable at a call site
461 callSiteInline is used at call sites, so it is a bit more generous.
462 It's a very important function that embodies lots of heuristics.
463 A non-WHNF can be inlined if it doesn't occur inside a lambda,
464 and occurs exactly once or
465 occurs once in each branch of a case and is small
467 If the thing is in WHNF, there's no danger of duplicating work,
468 so we can inline if it occurs once, or is small
471 callSiteInline :: Bool -- True <=> the Id is black listed
472 -> Bool -- 'inline' note at call site
474 -> [Bool] -- One for each value arg; True if it is interesting
475 -> Bool -- True <=> continuation is interesting
476 -> Maybe CoreExpr -- Unfolding, if any
479 callSiteInline black_listed inline_call id arg_infos interesting_cont
480 = case getIdUnfolding id of {
481 NoUnfolding -> Nothing ;
482 OtherCon _ -> Nothing ;
483 CoreUnfolding unf_template is_cheap _ guidance ->
486 result | yes_or_no = Just unf_template
487 | otherwise = Nothing
489 inline_prag = getInlinePragma id
490 n_val_args = length arg_infos
494 IAmDead -> pprTrace "callSiteInline: dead" (ppr id) False
495 IMustNotBeINLINEd -> False
496 IAmALoopBreaker -> False
497 IMustBeINLINEd -> True -- Overrides absolutely everything, including the black list
498 ICanSafelyBeINLINEd in_lam one_br -> consider in_lam True one_br
499 NoInlinePragInfo -> consider InsideLam False False
501 consider in_lam once once_in_one_branch
502 | black_listed = False
504 | once_in_one_branch -- Be very keen to inline something if this is its unique occurrence; that
505 -- gives a good chance of eliminating the original binding for the thing.
506 -- The only time we hold back is when substituting inside a lambda;
507 -- then if the context is totally uninteresting (not applied, not scrutinised)
508 -- there is no point in substituting because it might just increase allocation.
509 = WARN( case in_lam of { NotInsideLam -> True; other -> False },
510 text "callSiteInline:oneOcc" <+> ppr id )
511 -- If it has one occurrence, not inside a lambda, PreInlineUnconditionally
512 -- should have zapped it already
513 is_cheap && (not (null arg_infos) || interesting_cont)
515 | otherwise -- Occurs (textually) more than once, so look at its size
519 UnfoldIfGoodArgs n_vals_wanted arg_discounts size res_discount
520 | enough_args && size <= (n_vals_wanted + 1)
522 -- Size of call is n_vals_wanted (+1 for the function)
525 InsideLam -> is_cheap
527 | not (or arg_infos || really_interesting_cont || once)
528 -- If it occurs more than once, there must be something interesting
529 -- about some argument, or the result, to make it worth inlining
530 -- We also drop this case if the thing occurs once, although perhaps in
531 -- several branches. In this case we are keener about inlining in the hope
532 -- that we'll be able to drop the allocation for the function altogether.
537 NotInsideLam -> small_enough
538 InsideLam -> is_cheap && small_enough
541 enough_args = n_val_args >= n_vals_wanted
542 really_interesting_cont | n_val_args < n_vals_wanted = False -- Too few args
543 | n_val_args == n_vals_wanted = interesting_cont
544 | otherwise = True -- Extra args
545 -- This rather elaborate defn for really_interesting_cont is important
546 -- Consider an I# = INLINE (\x -> I# {x})
547 -- The unfolding guidance deems it to have size 2, and no arguments.
548 -- So in an application (I# y) we must take the extra arg 'y' as
549 -- evidence of an interesting context!
551 small_enough = (size - discount) <= opt_UF_UseThreshold
552 discount = computeDiscount n_vals_wanted arg_discounts res_discount
553 arg_infos really_interesting_cont
558 if opt_D_dump_inlinings then
559 pprTrace "Considering inlining"
560 (ppr id <+> vcat [text "black listed" <+> ppr black_listed,
561 text "inline prag:" <+> ppr inline_prag,
562 text "arg infos" <+> ppr arg_infos,
563 text "interesting continuation" <+> ppr interesting_cont,
564 text "is cheap" <+> ppr is_cheap,
565 text "guidance" <+> ppr guidance,
566 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO",
568 text "Unfolding =" <+> pprCoreExpr unf_template
576 computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Bool -> Int
577 computeDiscount n_vals_wanted arg_discounts res_discount arg_infos result_used
578 -- We multiple the raw discounts (args_discount and result_discount)
579 -- ty opt_UnfoldingKeenessFactor because the former have to do with
580 -- *size* whereas the discounts imply that there's some extra
581 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
584 -- we also discount 1 for each argument passed, because these will
585 -- reduce with the lambdas in the function (we count 1 for a lambda
587 = length (take n_vals_wanted arg_infos) +
588 -- Discount of 1 for each arg supplied, because the
589 -- result replaces the call
590 round (opt_UF_KeenessFactor *
591 fromInt (arg_discount + result_discount))
593 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
595 mk_arg_discount discount is_evald | is_evald = discount
598 -- Don't give a result discount unless there are enough args
599 result_discount | result_used = res_discount -- Over-applied, or case scrut
604 %************************************************************************
606 \subsection{Black-listing}
608 %************************************************************************
610 Inlining is controlled by the "Inline phase" number, which is set
611 by the per-simplification-pass '-finline-phase' flag.
613 For optimisation we use phase 1,2 and nothing (i.e. no -finline-phase flag)
614 in that order. The meanings of these are determined by the @blackListed@ function
618 blackListed :: IdSet -- Used in transformation rules
619 -> Maybe Int -- Inline phase
620 -> Id -> Bool -- True <=> blacklisted
622 -- The blackListed function sees whether a variable should *not* be
623 -- inlined because of the inline phase we are in. This is the sole
624 -- place that the inline phase number is looked at.
626 -- Phase 0: used for 'no imported inlinings please'
627 -- This prevents wrappers getting inlined which in turn is bad for full laziness
628 blackListed rule_vars (Just 0)
629 = \v -> not (isLocallyDefined v)
631 -- Phase 1: don't inline any rule-y things or things with specialisations
632 blackListed rule_vars (Just 1)
633 = \v -> let v_uniq = idUnique v
634 in v `elemVarSet` rule_vars
635 || not (isEmptyCoreRules (getIdSpecialisation v))
636 || v_uniq == runSTRepIdKey
638 -- Phase 2: allow build/augment to inline, and specialisations
639 blackListed rule_vars (Just 2)
640 = \v -> let v_uniq = idUnique v
641 in (v `elemVarSet` rule_vars && not (v_uniq == buildIdKey ||
642 v_uniq == augmentIdKey))
643 || v_uniq == runSTRepIdKey
645 -- Otherwise just go for it
646 blackListed rule_vars phase
651 SLPJ 95/04: Why @runST@ must be inlined very late:
655 (a, s') = newArray# 100 [] s
656 (_, s'') = fill_in_array_or_something a x s'
660 If we inline @runST@, we'll get:
663 (a, s') = newArray# 100 [] realWorld#{-NB-}
664 (_, s'') = fill_in_array_or_something a x s'
668 And now the @newArray#@ binding can be floated to become a CAF, which
669 is totally and utterly wrong:
672 (a, s') = newArray# 100 [] realWorld#{-NB-} -- YIKES!!!
675 let (_, s'') = fill_in_array_or_something a x s' in
678 All calls to @f@ will share a {\em single} array!
680 Yet we do want to inline runST sometime, so we can avoid
681 needless code. Solution: black list it until the last moment.