2 % (c) The University of Glasgow 2006
3 % (c) The AQUA Project, Glasgow University, 1994-1998
8 Unfoldings (which can travel across module boundaries) are in Core
9 syntax (namely @CoreExpr@s).
11 The type @Unfolding@ sits ``above'' simply-Core-expressions
12 unfoldings, capturing ``higher-level'' things we know about a binding,
13 usually things that the simplifier found out (e.g., ``it's a
14 literal''). In the corner of a @CoreUnfolding@ unfolding, you will
15 find, unsurprisingly, a Core expression.
19 Unfolding, UnfoldingGuidance, -- Abstract types
21 noUnfolding, mkImplicitUnfolding,
22 mkTopUnfolding, mkUnfolding, mkCoreUnfolding,
23 mkInlineRule, mkWwInlineRule,
24 mkCompulsoryUnfolding, mkDFunUnfolding,
26 interestingArg, ArgSummary(..),
28 couldBeSmallEnoughToInline,
29 certainlyWillInline, smallEnoughToInline,
31 callSiteInline, CallCtxt(..),
37 #include "HsVersions.h"
42 import PprCore () -- Instances
44 import CoreSubst hiding( substTy )
45 import CoreFVs ( exprFreeVars )
53 import BasicTypes ( Arity )
54 import TcType ( tcSplitDFunTy )
58 import VarEnv ( mkInScopeSet )
68 %************************************************************************
70 \subsection{Making unfoldings}
72 %************************************************************************
75 mkTopUnfolding :: Bool -> CoreExpr -> Unfolding
76 mkTopUnfolding is_bottoming expr
77 = mkUnfolding True {- Top level -} is_bottoming expr
79 mkImplicitUnfolding :: CoreExpr -> Unfolding
80 -- For implicit Ids, do a tiny bit of optimising first
81 mkImplicitUnfolding expr = mkTopUnfolding False (simpleOptExpr expr)
83 -- Note [Top-level flag on inline rules]
84 -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
85 -- Slight hack: note that mk_inline_rules conservatively sets the
86 -- top-level flag to True. It gets set more accurately by the simplifier
87 -- Simplify.simplUnfolding.
89 mkUnfolding :: Bool -> Bool -> CoreExpr -> Unfolding
90 mkUnfolding top_lvl is_bottoming expr
91 = CoreUnfolding { uf_tmpl = occurAnalyseExpr expr,
95 uf_is_value = exprIsHNF expr,
96 uf_is_conlike = exprIsConLike expr,
97 uf_expandable = exprIsExpandable expr,
98 uf_is_cheap = is_cheap,
99 uf_guidance = guidance }
101 is_cheap = exprIsCheap expr
102 (arity, guidance) = calcUnfoldingGuidance is_cheap (top_lvl && is_bottoming)
103 opt_UF_CreationThreshold expr
104 -- Sometimes during simplification, there's a large let-bound thing
105 -- which has been substituted, and so is now dead; so 'expr' contains
106 -- two copies of the thing while the occurrence-analysed expression doesn't
107 -- Nevertheless, we *don't* occ-analyse before computing the size because the
108 -- size computation bales out after a while, whereas occurrence analysis does not.
110 -- This can occasionally mean that the guidance is very pessimistic;
111 -- it gets fixed up next round. And it should be rare, because large
112 -- let-bound things that are dead are usually caught by preInlineUnconditionally
114 mkCoreUnfolding :: Bool -> UnfoldingSource -> CoreExpr
115 -> Arity -> UnfoldingGuidance -> Unfolding
116 -- Occurrence-analyses the expression before capturing it
117 mkCoreUnfolding top_lvl src expr arity guidance
118 = CoreUnfolding { uf_tmpl = occurAnalyseExpr expr,
122 uf_is_value = exprIsHNF expr,
123 uf_is_conlike = exprIsConLike expr,
124 uf_is_cheap = exprIsCheap expr,
125 uf_expandable = exprIsExpandable expr,
126 uf_guidance = guidance }
128 mkDFunUnfolding :: DataCon -> [Id] -> Unfolding
129 mkDFunUnfolding con ops = DFunUnfolding con (map Var ops)
131 mkWwInlineRule :: Id -> CoreExpr -> Arity -> Unfolding
132 mkWwInlineRule id expr arity
133 = mkCoreUnfolding True (InlineWrapper id)
134 (simpleOptExpr expr) arity
135 (UnfWhen unSaturatedOk boringCxtNotOk)
137 mkCompulsoryUnfolding :: CoreExpr -> Unfolding
138 mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
139 = mkCoreUnfolding True InlineCompulsory
140 expr 0 -- Arity of unfolding doesn't matter
141 (UnfWhen unSaturatedOk boringCxtOk)
143 mkInlineRule :: Bool -> CoreExpr -> Arity -> Unfolding
144 mkInlineRule unsat_ok expr arity
145 = mkCoreUnfolding True InlineRule -- Note [Top-level flag on inline rules]
147 (UnfWhen unsat_ok boring_ok)
149 expr' = simpleOptExpr expr
150 boring_ok = case calcUnfoldingGuidance True -- Treat as cheap
151 False -- But not bottoming
153 (_, UnfWhen _ boring_ok) -> boring_ok
154 _other -> boringCxtNotOk
155 -- See Note [INLINE for small functions]
159 %************************************************************************
161 \subsection{The UnfoldingGuidance type}
163 %************************************************************************
166 calcUnfoldingGuidance
167 :: Bool -- True <=> the rhs is cheap, or we want to treat it
168 -- as cheap (INLINE things)
169 -> Bool -- True <=> this is a top-level unfolding for a
170 -- diverging function; don't inline this
171 -> Int -- Bomb out if size gets bigger than this
172 -> CoreExpr -- Expression to look at
173 -> (Arity, UnfoldingGuidance)
174 calcUnfoldingGuidance expr_is_cheap top_bot bOMB_OUT_SIZE expr
175 = case collectBinders expr of { (bndrs, body) ->
177 val_bndrs = filter isId bndrs
178 n_val_bndrs = length val_bndrs
181 = case (sizeExpr (iUnbox bOMB_OUT_SIZE) val_bndrs body) of
183 SizeIs size cased_bndrs scrut_discount
184 | uncondInline n_val_bndrs (iBox size) && expr_is_cheap
185 -> UnfWhen needSaturated boringCxtOk
187 | top_bot -- See Note [Do not inline top-level bottoming functions]
191 -> UnfIfGoodArgs { ug_args = map (discount cased_bndrs) val_bndrs
192 , ug_size = iBox size
193 , ug_res = iBox scrut_discount }
196 = foldlBag (\acc (b',n) -> if bndr==b' then acc+n else acc)
199 (n_val_bndrs, guidance) }
202 Note [Computing the size of an expression]
203 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
204 The basic idea of sizeExpr is obvious enough: count nodes. But getting the
205 heuristics right has taken a long time. Here's the basic strategy:
207 * Variables, literals: 0
208 (Exception for string literals, see litSize.)
210 * Function applications (f e1 .. en): 1 + #value args
212 * Constructor applications: 1, regardless of #args
214 * Let(rec): 1 + size of components
229 Notice that 'x' counts 0, while (f x) counts 2. That's deliberate: there's
230 a function call to account for. Notice also that constructor applications
231 are very cheap, because exposing them to a caller is so valuable.
234 Note [Do not inline top-level bottoming functions]
235 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
236 The FloatOut pass has gone to some trouble to float out calls to 'error'
237 and similar friends. See Note [Bottoming floats] in SetLevels.
238 Do not re-inline them! But we *do* still inline if they are very small
239 (the uncondInline stuff).
242 Note [Unconditional inlining]
243 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
244 We inline *unconditionally* if inlined thing is smaller (using sizeExpr)
245 than the thing it's replacing. Notice that
246 (f x) --> (g 3) -- YES, unconditionally
247 (f x) --> x : [] -- YES, *even though* there are two
248 -- arguments to the cons
252 It's very important not to unconditionally replace a variable by
256 uncondInline :: Arity -> Int -> Bool
257 -- Inline unconditionally if there no size increase
258 -- Size of call is arity (+1 for the function)
259 -- See Note [Unconditional inlining]
260 uncondInline arity size
261 | arity == 0 = size == 0
262 | otherwise = size <= arity + 1
267 sizeExpr :: FastInt -- Bomb out if it gets bigger than this
268 -> [Id] -- Arguments; we're interested in which of these
273 -- Note [Computing the size of an expression]
275 sizeExpr bOMB_OUT_SIZE top_args expr
278 size_up (Cast e _) = size_up e
279 size_up (Note _ e) = size_up e
280 size_up (Type _) = sizeZero -- Types cost nothing
281 size_up (Lit lit) = sizeN (litSize lit)
282 size_up (Var f) = size_up_call f [] -- Make sure we get constructor
283 -- discounts even on nullary constructors
285 size_up (App fun (Type _)) = size_up fun
286 size_up (App fun arg) = size_up_app fun [arg]
287 `addSize` nukeScrutDiscount (size_up arg)
289 size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
290 | otherwise = size_up e
292 size_up (Let (NonRec binder rhs) body)
293 = nukeScrutDiscount (size_up rhs) `addSize`
294 size_up body `addSizeN`
295 (if isUnLiftedType (idType binder) then 0 else 1)
296 -- For the allocation
297 -- If the binder has an unlifted type there is no allocation
299 size_up (Let (Rec pairs) body)
300 = nukeScrutDiscount rhs_size `addSize`
301 size_up body `addSizeN`
302 length pairs -- For the allocation
304 rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
306 size_up (Case (Var v) _ _ alts)
307 | v `elem` top_args -- We are scrutinising an argument variable
308 = alts_size (foldr1 addSize alt_sizes) -- The 1 is for the case itself
309 (foldr1 maxSize alt_sizes)
310 -- Good to inline if an arg is scrutinised, because
311 -- that may eliminate allocation in the caller
312 -- And it eliminates the case itself
314 alt_sizes = map size_up_alt alts
316 -- alts_size tries to compute a good discount for
317 -- the case when we are scrutinising an argument variable
318 alts_size (SizeIs tot tot_disc _tot_scrut) -- Size of all alternatives
319 (SizeIs max _max_disc max_scrut) -- Size of biggest alternative
320 = SizeIs tot (unitBag (v, iBox (_ILIT(2) +# tot -# max)) `unionBags` tot_disc) max_scrut
321 -- If the variable is known, we produce a discount that
322 -- will take us back to 'max', the size of the largest alternative
323 -- The 1+ is a little discount for reduced allocation in the caller
325 -- Notice though, that we return tot_disc, the total discount from
326 -- all branches. I think that's right.
328 alts_size tot_size _ = tot_size
330 size_up (Case e _ _ alts) = foldr (addSize . size_up_alt)
331 (nukeScrutDiscount (size_up e))
333 -- We don't charge for the case itself
334 -- It's a strict thing, and the price of the call
335 -- is paid by scrut. Also consider
336 -- case f x of DEFAULT -> e
337 -- This is just ';'! Don't charge for it.
339 -- Moreover, we charge one per alternative.
342 -- size_up_app is used when there's ONE OR MORE value args
343 size_up_app (App fun arg) args
344 | isTypeArg arg = size_up_app fun args
345 | otherwise = size_up_app fun (arg:args)
346 `addSize` nukeScrutDiscount (size_up arg)
347 size_up_app (Var fun) args = size_up_call fun args
348 size_up_app other args = size_up other `addSizeN` length args
351 size_up_call :: Id -> [CoreExpr] -> ExprSize
352 size_up_call fun val_args
353 = case idDetails fun of
354 FCallId _ -> sizeN opt_UF_DearOp
355 DataConWorkId dc -> conSize dc (length val_args)
356 PrimOpId op -> primOpSize op (length val_args)
357 ClassOpId _ -> classOpSize top_args val_args
358 _ -> funSize top_args fun (length val_args)
361 size_up_alt (_con, _bndrs, rhs) = size_up rhs `addSizeN` 1
362 -- Don't charge for args, so that wrappers look cheap
363 -- (See comments about wrappers with Case)
365 -- IMPORATANT: *do* charge 1 for the alternative, else we
366 -- find that giant case nests are treated as practically free
367 -- A good example is Foreign.C.Error.errrnoToIOError
370 -- These addSize things have to be here because
371 -- I don't want to give them bOMB_OUT_SIZE as an argument
372 addSizeN TooBig _ = TooBig
373 addSizeN (SizeIs n xs d) m = mkSizeIs bOMB_OUT_SIZE (n +# iUnbox m) xs d
375 addSize TooBig _ = TooBig
376 addSize _ TooBig = TooBig
377 addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
378 = mkSizeIs bOMB_OUT_SIZE (n1 +# n2) (xs `unionBags` ys) (d1 +# d2)
382 -- | Finds a nominal size of a string literal.
383 litSize :: Literal -> Int
384 -- Used by CoreUnfold.sizeExpr
385 litSize (MachStr str) = 1 + ((lengthFS str + 3) `div` 4)
386 -- If size could be 0 then @f "x"@ might be too small
387 -- [Sept03: make literal strings a bit bigger to avoid fruitless
388 -- duplication of little strings]
389 litSize _other = 0 -- Must match size of nullary constructors
390 -- Key point: if x |-> 4, then x must inline unconditionally
391 -- (eg via case binding)
393 classOpSize :: [Id] -> [CoreExpr] -> ExprSize
394 -- See Note [Conlike is interesting]
397 classOpSize top_args (arg1 : other_args)
398 = SizeIs (iUnbox size) arg_discount (_ILIT(0))
400 size = 2 + length other_args
401 -- If the class op is scrutinising a lambda bound dictionary then
402 -- give it a discount, to encourage the inlining of this function
403 -- The actual discount is rather arbitrarily chosen
404 arg_discount = case arg1 of
405 Var dict | dict `elem` top_args
406 -> unitBag (dict, opt_UF_DictDiscount)
409 funSize :: [Id] -> Id -> Int -> ExprSize
410 -- Size for functions that are not constructors or primops
411 -- Note [Function applications]
412 funSize top_args fun n_val_args
413 | fun `hasKey` buildIdKey = buildSize
414 | fun `hasKey` augmentIdKey = augmentSize
415 | otherwise = SizeIs (iUnbox size) arg_discount (iUnbox res_discount)
417 some_val_args = n_val_args > 0
419 arg_discount | some_val_args && fun `elem` top_args
420 = unitBag (fun, opt_UF_FunAppDiscount)
421 | otherwise = emptyBag
422 -- If the function is an argument and is applied
423 -- to some values, give it an arg-discount
425 res_discount | idArity fun > n_val_args = opt_UF_FunAppDiscount
427 -- If the function is partially applied, show a result discount
429 size | some_val_args = 1 + n_val_args
431 -- The 1+ is for the function itself
432 -- Add 1 for each non-trivial arg;
433 -- the allocation cost, as in let(rec)
436 conSize :: DataCon -> Int -> ExprSize
437 conSize dc n_val_args
438 | n_val_args == 0 = SizeIs (_ILIT(0)) emptyBag (_ILIT(1)) -- Like variables
439 | isUnboxedTupleCon dc = SizeIs (_ILIT(0)) emptyBag (iUnbox n_val_args +# _ILIT(1))
440 | otherwise = SizeIs (_ILIT(1)) emptyBag (iUnbox n_val_args +# _ILIT(1))
441 -- Treat a constructors application as size 1, regardless of how
442 -- many arguments it has; we are keen to expose them
443 -- (and we charge separately for their args). We can't treat
444 -- them as size zero, else we find that (Just x) has size 0,
445 -- which is the same as a lone variable; and hence 'v' will
446 -- always be replaced by (Just x), where v is bound to Just x.
448 -- However, unboxed tuples count as size zero
449 -- I found occasions where we had
450 -- f x y z = case op# x y z of { s -> (# s, () #) }
451 -- and f wasn't getting inlined
453 primOpSize :: PrimOp -> Int -> ExprSize
454 primOpSize op n_val_args
455 | not (primOpIsDupable op) = sizeN opt_UF_DearOp
456 | not (primOpOutOfLine op) = sizeN 1
457 -- Be very keen to inline simple primops.
458 -- We give a discount of 1 for each arg so that (op# x y z) costs 2.
459 -- We can't make it cost 1, else we'll inline let v = (op# x y z)
460 -- at every use of v, which is excessive.
462 -- A good example is:
463 -- let x = +# p q in C {x}
464 -- Even though x get's an occurrence of 'many', its RHS looks cheap,
465 -- and there's a good chance it'll get inlined back into C's RHS. Urgh!
467 | otherwise = sizeN n_val_args
470 buildSize :: ExprSize
471 buildSize = SizeIs (_ILIT(0)) emptyBag (_ILIT(4))
472 -- We really want to inline applications of build
473 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
474 -- Indeed, we should add a result_discount becuause build is
475 -- very like a constructor. We don't bother to check that the
476 -- build is saturated (it usually is). The "-2" discounts for the \c n,
477 -- The "4" is rather arbitrary.
479 augmentSize :: ExprSize
480 augmentSize = SizeIs (_ILIT(0)) emptyBag (_ILIT(4))
481 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
482 -- e plus ys. The -2 accounts for the \cn
484 nukeScrutDiscount :: ExprSize -> ExprSize
485 nukeScrutDiscount (SizeIs n vs _) = SizeIs n vs (_ILIT(0))
486 nukeScrutDiscount TooBig = TooBig
488 -- When we return a lambda, give a discount if it's used (applied)
489 lamScrutDiscount :: ExprSize -> ExprSize
490 lamScrutDiscount (SizeIs n vs _) = SizeIs n vs (iUnbox opt_UF_FunAppDiscount)
491 lamScrutDiscount TooBig = TooBig
494 Note [Discounts and thresholds]
495 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
496 Constants for discounts and thesholds are defined in main/StaticFlags,
497 all of form opt_UF_xxxx. They are:
499 opt_UF_CreationThreshold (45)
500 At a definition site, if the unfolding is bigger than this, we
501 may discard it altogether
503 opt_UF_UseThreshold (6)
504 At a call site, if the unfolding, less discounts, is smaller than
505 this, then it's small enough inline
507 opt_UF_KeennessFactor (1.5)
508 Factor by which the discounts are multiplied before
509 subtracting from size
511 opt_UF_DictDiscount (1)
512 The discount for each occurrence of a dictionary argument
513 as an argument of a class method. Should be pretty small
514 else big functions may get inlined
516 opt_UF_FunAppDiscount (6)
517 Discount for a function argument that is applied. Quite
518 large, because if we inline we avoid the higher-order call.
521 The size of a foreign call or not-dupable PrimOp
524 Note [Function applications]
525 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
526 In a function application (f a b)
528 - If 'f' is an argument to the function being analysed,
529 and there's at least one value arg, record a FunAppDiscount for f
531 - If the application if a PAP (arity > 2 in this example)
532 record a *result* discount (because inlining
533 with "extra" args in the call may mean that we now
534 get a saturated application)
536 Code for manipulating sizes
539 data ExprSize = TooBig
540 | SizeIs FastInt -- Size found
541 (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
542 FastInt -- Size to subtract if result is scrutinised
543 -- by a case expression
545 instance Outputable ExprSize where
546 ppr TooBig = ptext (sLit "TooBig")
547 ppr (SizeIs a _ c) = brackets (int (iBox a) <+> int (iBox c))
549 -- subtract the discount before deciding whether to bale out. eg. we
550 -- want to inline a large constructor application into a selector:
551 -- tup = (a_1, ..., a_99)
552 -- x = case tup of ...
554 mkSizeIs :: FastInt -> FastInt -> Bag (Id, Int) -> FastInt -> ExprSize
555 mkSizeIs max n xs d | (n -# d) ># max = TooBig
556 | otherwise = SizeIs n xs d
558 maxSize :: ExprSize -> ExprSize -> ExprSize
559 maxSize TooBig _ = TooBig
560 maxSize _ TooBig = TooBig
561 maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
565 sizeN :: Int -> ExprSize
567 sizeZero = SizeIs (_ILIT(0)) emptyBag (_ILIT(0))
568 sizeN n = SizeIs (iUnbox n) emptyBag (_ILIT(0))
572 %************************************************************************
574 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
576 %************************************************************************
578 We use 'couldBeSmallEnoughToInline' to avoid exporting inlinings that
579 we ``couldn't possibly use'' on the other side. Can be overridden w/
580 flaggery. Just the same as smallEnoughToInline, except that it has no
584 couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
585 couldBeSmallEnoughToInline threshold rhs
586 = case calcUnfoldingGuidance False False threshold rhs of
587 (_, UnfNever) -> False
591 smallEnoughToInline :: Unfolding -> Bool
592 smallEnoughToInline (CoreUnfolding {uf_guidance = UnfIfGoodArgs {ug_size = size}})
593 = size <= opt_UF_UseThreshold
594 smallEnoughToInline _
598 certainlyWillInline :: Unfolding -> Bool
599 -- Sees if the unfolding is pretty certain to inline
600 certainlyWillInline (CoreUnfolding { uf_is_cheap = is_cheap, uf_arity = n_vals, uf_guidance = guidance })
604 UnfIfGoodArgs { ug_size = size}
605 -> is_cheap && size - (n_vals +1) <= opt_UF_UseThreshold
607 certainlyWillInline _
611 %************************************************************************
613 \subsection{callSiteInline}
615 %************************************************************************
617 This is the key function. It decides whether to inline a variable at a call site
619 callSiteInline is used at call sites, so it is a bit more generous.
620 It's a very important function that embodies lots of heuristics.
621 A non-WHNF can be inlined if it doesn't occur inside a lambda,
622 and occurs exactly once or
623 occurs once in each branch of a case and is small
625 If the thing is in WHNF, there's no danger of duplicating work,
626 so we can inline if it occurs once, or is small
628 NOTE: we don't want to inline top-level functions that always diverge.
629 It just makes the code bigger. Tt turns out that the convenient way to prevent
630 them inlining is to give them a NOINLINE pragma, which we do in
631 StrictAnal.addStrictnessInfoToTopId
634 callSiteInline :: DynFlags
636 -> Unfolding -- Its unfolding (if active)
637 -> Bool -- True if there are are no arguments at all (incl type args)
638 -> [ArgSummary] -- One for each value arg; True if it is interesting
639 -> CallCtxt -- True <=> continuation is interesting
640 -> Maybe CoreExpr -- Unfolding, if any
643 instance Outputable ArgSummary where
644 ppr TrivArg = ptext (sLit "TrivArg")
645 ppr NonTrivArg = ptext (sLit "NonTrivArg")
646 ppr ValueArg = ptext (sLit "ValueArg")
648 data CallCtxt = BoringCtxt
650 | ArgCtxt -- We are somewhere in the argument of a function
651 Bool -- True <=> we're somewhere in the RHS of function with rules
652 -- False <=> we *are* the argument of a function with non-zero
655 -- we *are* the RHS of a let Note [RHS of lets]
656 -- In both cases, be a little keener to inline
658 | ValAppCtxt -- We're applied to at least one value arg
659 -- This arises when we have ((f x |> co) y)
660 -- Then the (f x) has argument 'x' but in a ValAppCtxt
662 | CaseCtxt -- We're the scrutinee of a case
663 -- that decomposes its scrutinee
665 instance Outputable CallCtxt where
666 ppr BoringCtxt = ptext (sLit "BoringCtxt")
667 ppr (ArgCtxt rules) = ptext (sLit "ArgCtxt") <+> ppr rules
668 ppr CaseCtxt = ptext (sLit "CaseCtxt")
669 ppr ValAppCtxt = ptext (sLit "ValAppCtxt")
671 callSiteInline dflags id unfolding lone_variable arg_infos cont_info
672 = case unfolding of {
673 NoUnfolding -> Nothing ;
674 OtherCon _ -> Nothing ;
675 DFunUnfolding {} -> Nothing ; -- Never unfold a DFun
676 CoreUnfolding { uf_tmpl = unf_template, uf_is_top = is_top, uf_is_value = is_value,
677 uf_is_cheap = is_cheap, uf_arity = uf_arity, uf_guidance = guidance } ->
678 -- uf_arity will typically be equal to (idArity id),
679 -- but may be less for InlineRules
681 n_val_args = length arg_infos
682 saturated = n_val_args >= uf_arity
684 result | yes_or_no = Just unf_template
685 | otherwise = Nothing
687 interesting_args = any nonTriv arg_infos
688 -- NB: (any nonTriv arg_infos) looks at the
689 -- over-saturated args too which is "wrong";
690 -- but if over-saturated we inline anyway.
692 -- some_benefit is used when the RHS is small enough
693 -- and the call has enough (or too many) value
694 -- arguments (ie n_val_args >= arity). But there must
695 -- be *something* interesting about some argument, or the
696 -- result context, to make it worth inlining
698 | not saturated = interesting_args -- Under-saturated
699 -- Note [Unsaturated applications]
700 | n_val_args > uf_arity = True -- Over-saturated
701 | otherwise = interesting_args -- Saturated
702 || interesting_saturated_call
704 interesting_saturated_call
706 BoringCtxt -> not is_top && uf_arity > 0 -- Note [Nested functions]
707 CaseCtxt -> not (lone_variable && is_value) -- Note [Lone variables]
708 ArgCtxt {} -> uf_arity > 0 -- Note [Inlining in ArgCtxt]
709 ValAppCtxt -> True -- Note [Cast then apply]
711 (yes_or_no, extra_doc)
713 UnfNever -> (False, empty)
715 UnfWhen unsat_ok boring_ok -> ( (unsat_ok || saturated)
716 && (boring_ok || some_benefit)
718 -- For the boring_ok part see Note [INLINE for small functions]
720 UnfIfGoodArgs { ug_args = arg_discounts, ug_res = res_discount, ug_size = size }
721 -> ( is_cheap && some_benefit && small_enough
722 , (text "discounted size =" <+> int discounted_size) )
724 discounted_size = size - discount
725 small_enough = discounted_size <= opt_UF_UseThreshold
726 discount = computeDiscount uf_arity arg_discounts
727 res_discount arg_infos cont_info
730 if dopt Opt_D_dump_inlinings dflags then
731 pprTrace ("Considering inlining: " ++ showSDoc (ppr id))
732 (vcat [text "arg infos" <+> ppr arg_infos,
733 text "uf arity" <+> ppr uf_arity,
734 text "interesting continuation" <+> ppr cont_info,
735 text "some_benefit" <+> ppr some_benefit,
736 text "is value:" <+> ppr is_value,
737 text "is cheap:" <+> ppr is_cheap,
738 text "guidance" <+> ppr guidance,
740 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO"])
749 Be a tiny bit keener to inline in the RHS of a let, because that might
750 lead to good thing later
752 g y = let x = f y in ...(case x of (a,b,c) -> ...) ...
753 We'd inline 'f' if the call was in a case context, and it kind-of-is,
754 only we can't see it. So we treat the RHS of a let as not-totally-boring.
756 Note [Unsaturated applications]
757 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
758 When a call is not saturated, we *still* inline if one of the
759 arguments has interesting structure. That's sometimes very important.
760 A good example is the Ord instance for Bool in Base:
763 $fOrdBool =GHC.Classes.D:Ord
768 $cmin_ajX [Occ=LoopBreaker] :: Bool -> Bool -> Bool
769 $cmin_ajX = GHC.Classes.$dmmin @ Bool $fOrdBool
772 But the defn of GHC.Classes.$dmmin is:
774 $dmmin :: forall a. GHC.Classes.Ord a => a -> a -> a
775 {- Arity: 3, HasNoCafRefs, Strictness: SLL,
776 Unfolding: (\ @ a $dOrd :: GHC.Classes.Ord a x :: a y :: a ->
777 case @ a GHC.Classes.<= @ a $dOrd x y of wild {
778 GHC.Bool.False -> y GHC.Bool.True -> x }) -}
780 We *really* want to inline $dmmin, even though it has arity 3, in
781 order to unravel the recursion.
784 Note [INLINE for small functions]
785 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
786 Consider {-# INLINE f #-}
789 Then f's RHS is no larger than its LHS, so we should inline it
790 into even the most boring context. (We do so if there is no INLINE
794 Note [Things to watch]
795 ~~~~~~~~~~~~~~~~~~~~~~
796 * { y = I# 3; x = y `cast` co; ...case (x `cast` co) of ... }
797 Assume x is exported, so not inlined unconditionally.
798 Then we want x to inline unconditionally; no reason for it
799 not to, and doing so avoids an indirection.
801 * { x = I# 3; ....f x.... }
802 Make sure that x does not inline unconditionally!
803 Lest we get extra allocation.
805 Note [Inlining an InlineRule]
806 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
807 An InlineRules is used for
808 (a) programmer INLINE pragmas
809 (b) inlinings from worker/wrapper
811 For (a) the RHS may be large, and our contract is that we *only* inline
812 when the function is applied to all the arguments on the LHS of the
813 source-code defn. (The uf_arity in the rule.)
815 However for worker/wrapper it may be worth inlining even if the
816 arity is not satisfied (as we do in the CoreUnfolding case) so we don't
820 Note [Nested functions]
821 ~~~~~~~~~~~~~~~~~~~~~~~
822 If a function has a nested defn we also record some-benefit, on the
823 grounds that we are often able to eliminate the binding, and hence the
824 allocation, for the function altogether; this is good for join points.
825 But this only makes sense for *functions*; inlining a constructor
826 doesn't help allocation unless the result is scrutinised. UNLESS the
827 constructor occurs just once, albeit possibly in multiple case
828 branches. Then inlining it doesn't increase allocation, but it does
829 increase the chance that the constructor won't be allocated at all in
830 the branches that don't use it.
832 Note [Cast then apply]
833 ~~~~~~~~~~~~~~~~~~~~~~
835 myIndex = __inline_me ( (/\a. <blah>) |> co )
836 co :: (forall a. a -> a) ~ (forall a. T a)
837 ... /\a.\x. case ((myIndex a) |> sym co) x of { ... } ...
839 We need to inline myIndex to unravel this; but the actual call (myIndex a) has
840 no value arguments. The ValAppCtxt gives it enough incentive to inline.
842 Note [Inlining in ArgCtxt]
843 ~~~~~~~~~~~~~~~~~~~~~~~~~~
844 The condition (arity > 0) here is very important, because otherwise
845 we end up inlining top-level stuff into useless places; eg
848 This can make a very big difference: it adds 16% to nofib 'integer' allocs,
851 At one stage I replaced this condition by 'True' (leading to the above
852 slow-down). The motivation was test eyeball/inline1.hs; but that seems
855 NOTE: arguably, we should inline in ArgCtxt only if the result of the
856 call is at least CONLIKE. At least for the cases where we use ArgCtxt
857 for the RHS of a 'let', we only profit from the inlining if we get a
858 CONLIKE thing (modulo lets).
860 Note [Lone variables]
861 ~~~~~~~~~~~~~~~~~~~~~
862 The "lone-variable" case is important. I spent ages messing about
863 with unsatisfactory varaints, but this is nice. The idea is that if a
864 variable appears all alone
866 as an arg of lazy fn, or rhs BoringCtxt
867 as scrutinee of a case CaseCtxt
868 as arg of a fn ArgCtxt
870 it is bound to a value
872 then we should not inline it (unless there is some other reason,
873 e.g. is is the sole occurrence). That is what is happening at
874 the use of 'lone_variable' in 'interesting_saturated_call'.
876 Why? At least in the case-scrutinee situation, turning
877 let x = (a,b) in case x of y -> ...
879 let x = (a,b) in case (a,b) of y -> ...
881 let x = (a,b) in let y = (a,b) in ...
882 is bad if the binding for x will remain.
884 Another example: I discovered that strings
885 were getting inlined straight back into applications of 'error'
886 because the latter is strict.
888 f = \x -> ...(error s)...
890 Fundamentally such contexts should not encourage inlining because the
891 context can ``see'' the unfolding of the variable (e.g. case or a
892 RULE) so there's no gain. If the thing is bound to a value.
897 foo = _inline_ (\n. [n])
898 bar = _inline_ (foo 20)
899 baz = \n. case bar of { (m:_) -> m + n }
900 Here we really want to inline 'bar' so that we can inline 'foo'
901 and the whole thing unravels as it should obviously do. This is
902 important: in the NDP project, 'bar' generates a closure data
903 structure rather than a list.
905 So the non-inlining of lone_variables should only apply if the
906 unfolding is regarded as cheap; because that is when exprIsConApp_maybe
907 looks through the unfolding. Hence the "&& is_cheap" in the
910 * Even a type application or coercion isn't a lone variable.
912 case $fMonadST @ RealWorld of { :DMonad a b c -> c }
913 We had better inline that sucker! The case won't see through it.
915 For now, I'm treating treating a variable applied to types
916 in a *lazy* context "lone". The motivating example was
919 There's no advantage in inlining f here, and perhaps
920 a significant disadvantage. Hence some_val_args in the Stop case
923 computeDiscount :: Int -> [Int] -> Int -> [ArgSummary] -> CallCtxt -> Int
924 computeDiscount n_vals_wanted arg_discounts res_discount arg_infos cont_info
925 -- We multiple the raw discounts (args_discount and result_discount)
926 -- ty opt_UnfoldingKeenessFactor because the former have to do with
927 -- *size* whereas the discounts imply that there's some extra
928 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
931 = 1 -- Discount of 1 because the result replaces the call
932 -- so we count 1 for the function itself
934 + length (take n_vals_wanted arg_infos)
935 -- Discount of (un-scaled) 1 for each arg supplied,
936 -- because the result replaces the call
938 + round (opt_UF_KeenessFactor *
939 fromIntegral (arg_discount + res_discount'))
941 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
943 mk_arg_discount _ TrivArg = 0
944 mk_arg_discount _ NonTrivArg = 1
945 mk_arg_discount discount ValueArg = discount
947 res_discount' = case cont_info of
949 CaseCtxt -> res_discount
950 _other -> 4 `min` res_discount
951 -- res_discount can be very large when a function returns
952 -- constructors; but we only want to invoke that large discount
953 -- when there's a case continuation.
954 -- Otherwise we, rather arbitrarily, threshold it. Yuk.
955 -- But we want to aovid inlining large functions that return
956 -- constructors into contexts that are simply "interesting"
959 %************************************************************************
961 Interesting arguments
963 %************************************************************************
965 Note [Interesting arguments]
966 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
967 An argument is interesting if it deserves a discount for unfoldings
968 with a discount in that argument position. The idea is to avoid
969 unfolding a function that is applied only to variables that have no
970 unfolding (i.e. they are probably lambda bound): f x y z There is
971 little point in inlining f here.
973 Generally, *values* (like (C a b) and (\x.e)) deserve discounts. But
974 we must look through lets, eg (let x = e in C a b), because the let will
975 float, exposing the value, if we inline. That makes it different to
978 Before 2009 we said it was interesting if the argument had *any* structure
979 at all; i.e. (hasSomeUnfolding v). But does too much inlining; see Trac #3016.
981 But we don't regard (f x y) as interesting, unless f is unsaturated.
982 If it's saturated and f hasn't inlined, then it's probably not going
985 Note [Conlike is interesting]
986 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
988 f d = ...((*) d x y)...
990 where df is con-like. Then we'd really like to inline 'f' so that the
991 rule for (*) (df d) can fire. To do this
992 a) we give a discount for being an argument of a class-op (eg (*) d)
993 b) we say that a con-like argument (eg (df d)) is interesting
996 data ArgSummary = TrivArg -- Nothing interesting
997 | NonTrivArg -- Arg has structure
998 | ValueArg -- Arg is a con-app or PAP
999 -- ..or con-like. Note [Conlike is interesting]
1001 interestingArg :: CoreExpr -> ArgSummary
1002 -- See Note [Interesting arguments]
1003 interestingArg e = go e 0
1005 -- n is # value args to which the expression is applied
1006 go (Lit {}) _ = ValueArg
1008 | isConLikeId v = ValueArg -- Experimenting with 'conlike' rather that
1009 -- data constructors here
1010 | idArity v > n = ValueArg -- Catches (eg) primops with arity but no unfolding
1011 | n > 0 = NonTrivArg -- Saturated or unknown call
1012 | conlike_unfolding = ValueArg -- n==0; look for an interesting unfolding
1013 -- See Note [Conlike is interesting]
1014 | otherwise = TrivArg -- n==0, no useful unfolding
1016 conlike_unfolding = isConLikeUnfolding (idUnfolding v)
1018 go (Type _) _ = TrivArg
1019 go (App fn (Type _)) n = go fn n
1020 go (App fn _) n = go fn (n+1)
1021 go (Note _ a) n = go a n
1022 go (Cast e _) n = go e n
1024 | isTyVar v = go e n
1026 | otherwise = ValueArg
1027 go (Let _ e) n = case go e n of { ValueArg -> ValueArg; _ -> NonTrivArg }
1028 go (Case {}) _ = NonTrivArg
1030 nonTriv :: ArgSummary -> Bool
1031 nonTriv TrivArg = False
1035 %************************************************************************
1039 %************************************************************************
1041 Note [exprIsConApp_maybe]
1042 ~~~~~~~~~~~~~~~~~~~~~~~~~
1043 exprIsConApp_maybe is a very important function. There are two principal
1045 * case e of { .... }
1046 * cls_op e, where cls_op is a class operation
1048 In both cases you want to know if e is of form (C e1..en) where C is
1051 However e might not *look* as if
1054 -- | Returns @Just (dc, [t1..tk], [x1..xn])@ if the argument expression is
1055 -- a *saturated* constructor application of the form @dc t1..tk x1 .. xn@,
1056 -- where t1..tk are the *universally-qantified* type args of 'dc'
1057 exprIsConApp_maybe :: IdUnfoldingFun -> CoreExpr -> Maybe (DataCon, [Type], [CoreExpr])
1059 exprIsConApp_maybe id_unf (Note _ expr)
1060 = exprIsConApp_maybe id_unf expr
1061 -- We ignore all notes. For example,
1062 -- case _scc_ "foo" (C a b) of
1064 -- should be optimised away, but it will be only if we look
1065 -- through the SCC note.
1067 exprIsConApp_maybe id_unf (Cast expr co)
1068 = -- Here we do the KPush reduction rule as described in the FC paper
1069 -- The transformation applies iff we have
1070 -- (C e1 ... en) `cast` co
1071 -- where co :: (T t1 .. tn) ~ to_ty
1072 -- The left-hand one must be a T, because exprIsConApp returned True
1073 -- but the right-hand one might not be. (Though it usually will.)
1075 case exprIsConApp_maybe id_unf expr of {
1076 Nothing -> Nothing ;
1077 Just (dc, _dc_univ_args, dc_args) ->
1079 let (_from_ty, to_ty) = coercionKind co
1080 dc_tc = dataConTyCon dc
1082 case splitTyConApp_maybe to_ty of {
1083 Nothing -> Nothing ;
1084 Just (to_tc, to_tc_arg_tys)
1085 | dc_tc /= to_tc -> Nothing
1086 -- These two Nothing cases are possible; we might see
1087 -- (C x y) `cast` (g :: T a ~ S [a]),
1088 -- where S is a type function. In fact, exprIsConApp
1089 -- will probably not be called in such circumstances,
1090 -- but there't nothing wrong with it
1094 tc_arity = tyConArity dc_tc
1095 dc_univ_tyvars = dataConUnivTyVars dc
1096 dc_ex_tyvars = dataConExTyVars dc
1097 arg_tys = dataConRepArgTys dc
1099 dc_eqs :: [(Type,Type)] -- All equalities from the DataCon
1100 dc_eqs = [(mkTyVarTy tv, ty) | (tv,ty) <- dataConEqSpec dc] ++
1101 [getEqPredTys eq_pred | eq_pred <- dataConEqTheta dc]
1103 (ex_args, rest1) = splitAtList dc_ex_tyvars dc_args
1104 (co_args, val_args) = splitAtList dc_eqs rest1
1106 -- Make the "theta" from Fig 3 of the paper
1107 gammas = decomposeCo tc_arity co
1108 theta = zipOpenTvSubst (dc_univ_tyvars ++ dc_ex_tyvars)
1109 (gammas ++ stripTypeArgs ex_args)
1111 -- Cast the existential coercion arguments
1112 cast_co (ty1, ty2) (Type co)
1113 = Type $ mkSymCoercion (substTy theta ty1)
1114 `mkTransCoercion` co
1115 `mkTransCoercion` (substTy theta ty2)
1116 cast_co _ other_arg = pprPanic "cast_co" (ppr other_arg)
1117 new_co_args = zipWith cast_co dc_eqs co_args
1119 -- Cast the value arguments (which include dictionaries)
1120 new_val_args = zipWith cast_arg arg_tys val_args
1121 cast_arg arg_ty arg = mkCoerce (substTy theta arg_ty) arg
1124 let dump_doc = vcat [ppr dc, ppr dc_univ_tyvars, ppr dc_ex_tyvars,
1125 ppr arg_tys, ppr dc_args, ppr _dc_univ_args,
1126 ppr ex_args, ppr val_args]
1128 ASSERT2( coreEqType _from_ty (mkTyConApp dc_tc _dc_univ_args), dump_doc )
1129 ASSERT2( all isTypeArg (ex_args ++ co_args), dump_doc )
1130 ASSERT2( equalLength val_args arg_tys, dump_doc )
1133 Just (dc, to_tc_arg_tys, ex_args ++ new_co_args ++ new_val_args)
1136 exprIsConApp_maybe id_unf expr
1139 analyse (App fun arg) args = analyse fun (arg:args)
1140 analyse fun@(Lam {}) args = beta fun [] args
1142 analyse (Var fun) args
1143 | Just con <- isDataConWorkId_maybe fun
1145 , let (univ_ty_args, rest_args) = splitAtList (dataConUnivTyVars con) args
1146 = Just (con, stripTypeArgs univ_ty_args, rest_args)
1148 -- Look through dictionary functions; see Note [Unfolding DFuns]
1149 | DFunUnfolding con ops <- unfolding
1151 , let (dfun_tvs, _cls, dfun_res_tys) = tcSplitDFunTy (idType fun)
1152 subst = zipOpenTvSubst dfun_tvs (stripTypeArgs (takeList dfun_tvs args))
1153 = Just (con, substTys subst dfun_res_tys,
1154 [mkApps op args | op <- ops])
1156 -- Look through unfoldings, but only cheap ones, because
1157 -- we are effectively duplicating the unfolding
1158 | CoreUnfolding { uf_expandable = expand_me, uf_tmpl = rhs } <- unfolding
1159 , expand_me = -- pprTrace "expanding" (ppr fun $$ ppr rhs) $
1162 is_saturated = count isValArg args == idArity fun
1163 unfolding = id_unf fun -- Does not look through loop breakers
1164 -- ToDo: we *may* look through variables that are NOINLINE
1165 -- in this phase, and that is really not right
1167 analyse _ _ = Nothing
1170 in_scope = mkInScopeSet (exprFreeVars expr)
1173 beta (Lam v body) pairs (arg : args)
1175 = beta body ((v,arg):pairs) args
1177 beta (Lam {}) _ _ -- Un-saturated, or not a type lambda
1181 = case analyse (substExpr subst fun) args of
1182 Nothing -> -- pprTrace "Bale out! exprIsConApp_maybe" doc $
1184 Just ans -> -- pprTrace "Woo-hoo! exprIsConApp_maybe" doc $
1187 subst = mkOpenSubst in_scope pairs
1188 -- doc = vcat [ppr fun, ppr expr, ppr pairs, ppr args]
1191 stripTypeArgs :: [CoreExpr] -> [Type]
1192 stripTypeArgs args = ASSERT2( all isTypeArg args, ppr args )
1193 [ty | Type ty <- args]
1196 Note [Unfolding DFuns]
1197 ~~~~~~~~~~~~~~~~~~~~~~
1200 df :: forall a b. (Eq a, Eq b) -> Eq (a,b)
1201 df a b d_a d_b = MkEqD (a,b) ($c1 a b d_a d_b)
1204 So to split it up we just need to apply the ops $c1, $c2 etc
1205 to the very same args as the dfun. It takes a little more work
1206 to compute the type arguments to the dictionary constructor.