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
43 import TcType ( tcSplitSigmaTy, tcSplitDFunHead )
45 import CoreSubst hiding( substTy )
46 import CoreFVs ( exprFreeVars )
47 import CoreArity ( manifestArity )
55 import BasicTypes ( Arity )
56 import TcType ( tcSplitDFunTy )
60 import VarEnv ( mkInScopeSet )
70 %************************************************************************
72 \subsection{Making unfoldings}
74 %************************************************************************
77 mkTopUnfolding :: Bool -> CoreExpr -> Unfolding
78 mkTopUnfolding is_bottoming expr
79 = mkUnfolding True {- Top level -} is_bottoming expr
81 mkImplicitUnfolding :: CoreExpr -> Unfolding
82 -- For implicit Ids, do a tiny bit of optimising first
83 mkImplicitUnfolding expr = mkTopUnfolding False (simpleOptExpr expr)
85 -- Note [Top-level flag on inline rules]
86 -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
87 -- Slight hack: note that mk_inline_rules conservatively sets the
88 -- top-level flag to True. It gets set more accurately by the simplifier
89 -- Simplify.simplUnfolding.
91 mkUnfolding :: Bool -> Bool -> CoreExpr -> Unfolding
92 mkUnfolding top_lvl is_bottoming expr
93 = CoreUnfolding { uf_tmpl = occurAnalyseExpr expr,
97 uf_is_value = exprIsHNF expr,
98 uf_is_conlike = exprIsConLike expr,
99 uf_expandable = exprIsExpandable expr,
100 uf_is_cheap = is_cheap,
101 uf_guidance = guidance }
103 is_cheap = exprIsCheap expr
104 (arity, guidance) = calcUnfoldingGuidance is_cheap (top_lvl && is_bottoming)
105 opt_UF_CreationThreshold expr
106 -- Sometimes during simplification, there's a large let-bound thing
107 -- which has been substituted, and so is now dead; so 'expr' contains
108 -- two copies of the thing while the occurrence-analysed expression doesn't
109 -- Nevertheless, we *don't* occ-analyse before computing the size because the
110 -- size computation bales out after a while, whereas occurrence analysis does not.
112 -- This can occasionally mean that the guidance is very pessimistic;
113 -- it gets fixed up next round. And it should be rare, because large
114 -- let-bound things that are dead are usually caught by preInlineUnconditionally
116 mkCoreUnfolding :: Bool -> UnfoldingSource -> CoreExpr
117 -> Arity -> UnfoldingGuidance -> Unfolding
118 -- Occurrence-analyses the expression before capturing it
119 mkCoreUnfolding top_lvl src expr arity guidance
120 = CoreUnfolding { uf_tmpl = occurAnalyseExpr expr,
124 uf_is_value = exprIsHNF expr,
125 uf_is_conlike = exprIsConLike expr,
126 uf_is_cheap = exprIsCheap expr,
127 uf_expandable = exprIsExpandable expr,
128 uf_guidance = guidance }
130 mkDFunUnfolding :: Type -> [CoreExpr] -> Unfolding
131 mkDFunUnfolding dfun_ty ops
132 = DFunUnfolding dfun_nargs data_con ops
134 (tvs, theta, head_ty) = tcSplitSigmaTy dfun_ty
135 -- NB: tcSplitSigmaTy: do not look through a newtype
136 -- when the dictionary type is a newtype
137 (cls, _) = tcSplitDFunHead head_ty
138 dfun_nargs = length tvs + length theta
139 data_con = classDataCon cls
141 mkWwInlineRule :: Id -> CoreExpr -> Arity -> Unfolding
142 mkWwInlineRule id expr arity
143 = mkCoreUnfolding True (InlineWrapper id)
144 (simpleOptExpr expr) arity
145 (UnfWhen unSaturatedOk boringCxtNotOk)
147 mkCompulsoryUnfolding :: CoreExpr -> Unfolding
148 mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
149 = mkCoreUnfolding True InlineCompulsory
150 expr 0 -- Arity of unfolding doesn't matter
151 (UnfWhen unSaturatedOk boringCxtOk)
153 mkInlineRule :: CoreExpr -> Maybe Arity -> Unfolding
154 mkInlineRule expr mb_arity
155 = mkCoreUnfolding True InlineRule -- Note [Top-level flag on inline rules]
157 (UnfWhen unsat_ok boring_ok)
159 expr' = simpleOptExpr expr
160 (unsat_ok, arity) = case mb_arity of
161 Nothing -> (unSaturatedOk, manifestArity expr')
162 Just ar -> (needSaturated, ar)
164 boring_ok = case calcUnfoldingGuidance True -- Treat as cheap
165 False -- But not bottoming
167 (_, UnfWhen _ boring_ok) -> boring_ok
168 _other -> boringCxtNotOk
169 -- See Note [INLINE for small functions]
173 %************************************************************************
175 \subsection{The UnfoldingGuidance type}
177 %************************************************************************
180 calcUnfoldingGuidance
181 :: Bool -- True <=> the rhs is cheap, or we want to treat it
182 -- as cheap (INLINE things)
183 -> Bool -- True <=> this is a top-level unfolding for a
184 -- diverging function; don't inline this
185 -> Int -- Bomb out if size gets bigger than this
186 -> CoreExpr -- Expression to look at
187 -> (Arity, UnfoldingGuidance)
188 calcUnfoldingGuidance expr_is_cheap top_bot bOMB_OUT_SIZE expr
189 = case collectBinders expr of { (bndrs, body) ->
191 val_bndrs = filter isId bndrs
192 n_val_bndrs = length val_bndrs
195 = case (sizeExpr (iUnbox bOMB_OUT_SIZE) val_bndrs body) of
197 SizeIs size cased_bndrs scrut_discount
198 | uncondInline n_val_bndrs (iBox size)
200 -> UnfWhen unSaturatedOk boringCxtOk -- Note [INLINE for small functions]
201 | top_bot -- See Note [Do not inline top-level bottoming functions]
205 -> UnfIfGoodArgs { ug_args = map (discount cased_bndrs) val_bndrs
206 , ug_size = iBox size
207 , ug_res = iBox scrut_discount }
210 = foldlBag (\acc (b',n) -> if bndr==b' then acc+n else acc)
213 (n_val_bndrs, guidance) }
216 Note [Computing the size of an expression]
217 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
218 The basic idea of sizeExpr is obvious enough: count nodes. But getting the
219 heuristics right has taken a long time. Here's the basic strategy:
221 * Variables, literals: 0
222 (Exception for string literals, see litSize.)
224 * Function applications (f e1 .. en): 1 + #value args
226 * Constructor applications: 1, regardless of #args
228 * Let(rec): 1 + size of components
243 Notice that 'x' counts 0, while (f x) counts 2. That's deliberate: there's
244 a function call to account for. Notice also that constructor applications
245 are very cheap, because exposing them to a caller is so valuable.
248 Note [Do not inline top-level bottoming functions]
249 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
250 The FloatOut pass has gone to some trouble to float out calls to 'error'
251 and similar friends. See Note [Bottoming floats] in SetLevels.
252 Do not re-inline them! But we *do* still inline if they are very small
253 (the uncondInline stuff).
256 Note [INLINE for small functions]
257 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
258 Consider {-# INLINE f #-}
261 Then f's RHS is no larger than its LHS, so we should inline it into
262 even the most boring context. In general, f the function is
263 sufficiently small that its body is as small as the call itself, the
264 inline unconditionally, regardless of how boring the context is.
268 * We inline *unconditionally* if inlined thing is smaller (using sizeExpr)
269 than the thing it's replacing. Notice that
270 (f x) --> (g 3) -- YES, unconditionally
271 (f x) --> x : [] -- YES, *even though* there are two
272 -- arguments to the cons
276 It's very important not to unconditionally replace a variable by
279 * We do this even if the thing isn't saturated, else we end up with the
283 doesn't inline. Even in a boring context, inlining without being
284 saturated will give a lambda instead of a PAP, and will be more
285 efficient at runtime.
287 * However, when the function's arity > 0, we do insist that it
288 has at least one value argument at the call site. Otherwise we find this:
291 If we inline f here we get
292 d = /\b. MkD (\x:b. x)
293 and then prepareRhs floats out the argument, abstracting the type
294 variables, so we end up with the original again!
298 uncondInline :: Arity -> Int -> Bool
299 -- Inline unconditionally if there no size increase
300 -- Size of call is arity (+1 for the function)
301 -- See Note [INLINE for small functions]
302 uncondInline arity size
303 | arity == 0 = size == 0
304 | otherwise = size <= arity + 1
309 sizeExpr :: FastInt -- Bomb out if it gets bigger than this
310 -> [Id] -- Arguments; we're interested in which of these
315 -- Note [Computing the size of an expression]
317 sizeExpr bOMB_OUT_SIZE top_args expr
320 size_up (Cast e _) = size_up e
321 size_up (Note _ e) = size_up e
322 size_up (Type _) = sizeZero -- Types cost nothing
323 size_up (Lit lit) = sizeN (litSize lit)
324 size_up (Var f) = size_up_call f [] -- Make sure we get constructor
325 -- discounts even on nullary constructors
327 size_up (App fun (Type _)) = size_up fun
328 size_up (App fun arg) = size_up arg `addSizeNSD`
329 size_up_app fun [arg]
331 size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
332 | otherwise = size_up e
334 size_up (Let (NonRec binder rhs) body)
335 = size_up rhs `addSizeNSD`
336 size_up body `addSizeN`
337 (if isUnLiftedType (idType binder) then 0 else 1)
338 -- For the allocation
339 -- If the binder has an unlifted type there is no allocation
341 size_up (Let (Rec pairs) body)
342 = foldr (addSizeNSD . size_up . snd)
343 (size_up body `addSizeN` length pairs) -- (length pairs) for the allocation
346 size_up (Case (Var v) _ _ alts)
347 | v `elem` top_args -- We are scrutinising an argument variable
348 = alts_size (foldr1 addAltSize alt_sizes)
349 (foldr1 maxSize alt_sizes)
350 -- Good to inline if an arg is scrutinised, because
351 -- that may eliminate allocation in the caller
352 -- And it eliminates the case itself
354 alt_sizes = map size_up_alt alts
356 -- alts_size tries to compute a good discount for
357 -- the case when we are scrutinising an argument variable
358 alts_size (SizeIs tot tot_disc tot_scrut) -- Size of all alternatives
359 (SizeIs max _ _) -- Size of biggest alternative
360 = SizeIs tot (unitBag (v, iBox (_ILIT(2) +# tot -# max)) `unionBags` tot_disc) tot_scrut
361 -- If the variable is known, we produce a discount that
362 -- will take us back to 'max', the size of the largest alternative
363 -- The 1+ is a little discount for reduced allocation in the caller
365 -- Notice though, that we return tot_disc, the total discount from
366 -- all branches. I think that's right.
368 alts_size tot_size _ = tot_size
370 size_up (Case e _ _ alts) = size_up e `addSizeNSD`
371 foldr (addAltSize . size_up_alt) sizeZero alts
372 -- We don't charge for the case itself
373 -- It's a strict thing, and the price of the call
374 -- is paid by scrut. Also consider
375 -- case f x of DEFAULT -> e
376 -- This is just ';'! Don't charge for it.
378 -- Moreover, we charge one per alternative.
381 -- size_up_app is used when there's ONE OR MORE value args
382 size_up_app (App fun arg) args
383 | isTypeArg arg = size_up_app fun args
384 | otherwise = size_up arg `addSizeNSD`
385 size_up_app fun (arg:args)
386 size_up_app (Var fun) args = size_up_call fun args
387 size_up_app other args = size_up other `addSizeN` length args
390 size_up_call :: Id -> [CoreExpr] -> ExprSize
391 size_up_call fun val_args
392 = case idDetails fun of
393 FCallId _ -> sizeN opt_UF_DearOp
394 DataConWorkId dc -> conSize dc (length val_args)
395 PrimOpId op -> primOpSize op (length val_args)
396 ClassOpId _ -> classOpSize top_args val_args
397 _ -> funSize top_args fun (length val_args)
400 size_up_alt (_con, _bndrs, rhs) = size_up rhs `addSizeN` 1
401 -- Don't charge for args, so that wrappers look cheap
402 -- (See comments about wrappers with Case)
404 -- IMPORATANT: *do* charge 1 for the alternative, else we
405 -- find that giant case nests are treated as practically free
406 -- A good example is Foreign.C.Error.errrnoToIOError
409 -- These addSize things have to be here because
410 -- I don't want to give them bOMB_OUT_SIZE as an argument
411 addSizeN TooBig _ = TooBig
412 addSizeN (SizeIs n xs d) m = mkSizeIs bOMB_OUT_SIZE (n +# iUnbox m) xs d
414 -- addAltSize is used to add the sizes of case alternatives
415 addAltSize TooBig _ = TooBig
416 addAltSize _ TooBig = TooBig
417 addAltSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
418 = mkSizeIs bOMB_OUT_SIZE (n1 +# n2)
420 (d1 +# d2) -- Note [addAltSize result discounts]
422 -- This variant ignores the result discount from its LEFT argument
423 -- It's used when the second argument isn't part of the result
424 addSizeNSD TooBig _ = TooBig
425 addSizeNSD _ TooBig = TooBig
426 addSizeNSD (SizeIs n1 xs _) (SizeIs n2 ys d2)
427 = mkSizeIs bOMB_OUT_SIZE (n1 +# n2)
433 -- | Finds a nominal size of a string literal.
434 litSize :: Literal -> Int
435 -- Used by CoreUnfold.sizeExpr
436 litSize (MachStr str) = 1 + ((lengthFS str + 3) `div` 4)
437 -- If size could be 0 then @f "x"@ might be too small
438 -- [Sept03: make literal strings a bit bigger to avoid fruitless
439 -- duplication of little strings]
440 litSize _other = 0 -- Must match size of nullary constructors
441 -- Key point: if x |-> 4, then x must inline unconditionally
442 -- (eg via case binding)
444 classOpSize :: [Id] -> [CoreExpr] -> ExprSize
445 -- See Note [Conlike is interesting]
448 classOpSize top_args (arg1 : other_args)
449 = SizeIs (iUnbox size) arg_discount (_ILIT(0))
451 size = 2 + length other_args
452 -- If the class op is scrutinising a lambda bound dictionary then
453 -- give it a discount, to encourage the inlining of this function
454 -- The actual discount is rather arbitrarily chosen
455 arg_discount = case arg1 of
456 Var dict | dict `elem` top_args
457 -> unitBag (dict, opt_UF_DictDiscount)
460 funSize :: [Id] -> Id -> Int -> ExprSize
461 -- Size for functions that are not constructors or primops
462 -- Note [Function applications]
463 funSize top_args fun n_val_args
464 | fun `hasKey` buildIdKey = buildSize
465 | fun `hasKey` augmentIdKey = augmentSize
466 | otherwise = SizeIs (iUnbox size) arg_discount (iUnbox res_discount)
468 some_val_args = n_val_args > 0
470 arg_discount | some_val_args && fun `elem` top_args
471 = unitBag (fun, opt_UF_FunAppDiscount)
472 | otherwise = emptyBag
473 -- If the function is an argument and is applied
474 -- to some values, give it an arg-discount
476 res_discount | idArity fun > n_val_args = opt_UF_FunAppDiscount
478 -- If the function is partially applied, show a result discount
480 size | some_val_args = 1 + n_val_args
482 -- The 1+ is for the function itself
483 -- Add 1 for each non-trivial arg;
484 -- the allocation cost, as in let(rec)
487 conSize :: DataCon -> Int -> ExprSize
488 conSize dc n_val_args
489 | n_val_args == 0 = SizeIs (_ILIT(0)) emptyBag (_ILIT(1)) -- Like variables
491 -- See Note [Constructor size]
492 | isUnboxedTupleCon dc = SizeIs (_ILIT(0)) emptyBag (iUnbox n_val_args +# _ILIT(1))
494 -- See Note [Unboxed tuple result discount]
495 -- | isUnboxedTupleCon dc = SizeIs (_ILIT(0)) emptyBag (_ILIT(0))
497 -- See Note [Constructor size]
498 | otherwise = SizeIs (_ILIT(1)) emptyBag (iUnbox n_val_args +# _ILIT(1))
501 Note [Constructor size]
502 ~~~~~~~~~~~~~~~~~~~~~~~
503 Treat a constructors application as size 1, regardless of how many
504 arguments it has; we are keen to expose them (and we charge separately
505 for their args). We can't treat them as size zero, else we find that
506 (Just x) has size 0, which is the same as a lone variable; and hence
507 'v' will always be replaced by (Just x), where v is bound to Just x.
509 However, unboxed tuples count as size zero. I found occasions where we had
510 f x y z = case op# x y z of { s -> (# s, () #) }
511 and f wasn't getting inlined.
513 Note [Unboxed tuple result discount]
514 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
515 I tried giving unboxed tuples a *result discount* of zero (see the
516 commented-out line). Why? When returned as a result they do not
517 allocate, so maybe we don't want to charge so much for them If you
518 have a non-zero discount here, we find that workers often get inlined
519 back into wrappers, because it look like
520 f x = case $wf x of (# a,b #) -> (a,b)
521 and we are keener because of the case. However while this change
522 shrank binary sizes by 0.5% it also made spectral/boyer allocate 5%
523 more. All other changes were very small. So it's not a big deal but I
524 didn't adopt the idea.
527 primOpSize :: PrimOp -> Int -> ExprSize
528 primOpSize op n_val_args
529 | not (primOpIsDupable op) = sizeN opt_UF_DearOp
530 | not (primOpOutOfLine op) = sizeN 1
531 -- Be very keen to inline simple primops.
532 -- We give a discount of 1 for each arg so that (op# x y z) costs 2.
533 -- We can't make it cost 1, else we'll inline let v = (op# x y z)
534 -- at every use of v, which is excessive.
536 -- A good example is:
537 -- let x = +# p q in C {x}
538 -- Even though x get's an occurrence of 'many', its RHS looks cheap,
539 -- and there's a good chance it'll get inlined back into C's RHS. Urgh!
541 | otherwise = sizeN n_val_args
544 buildSize :: ExprSize
545 buildSize = SizeIs (_ILIT(0)) emptyBag (_ILIT(4))
546 -- We really want to inline applications of build
547 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
548 -- Indeed, we should add a result_discount becuause build is
549 -- very like a constructor. We don't bother to check that the
550 -- build is saturated (it usually is). The "-2" discounts for the \c n,
551 -- The "4" is rather arbitrary.
553 augmentSize :: ExprSize
554 augmentSize = SizeIs (_ILIT(0)) emptyBag (_ILIT(4))
555 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
556 -- e plus ys. The -2 accounts for the \cn
558 -- When we return a lambda, give a discount if it's used (applied)
559 lamScrutDiscount :: ExprSize -> ExprSize
560 lamScrutDiscount (SizeIs n vs _) = SizeIs n vs (iUnbox opt_UF_FunAppDiscount)
561 lamScrutDiscount TooBig = TooBig
564 Note [addAltSize result discounts]
565 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
566 When adding the size of alternatives, we *add* the result discounts
567 too, rather than take the *maximum*. For a multi-branch case, this
568 gives a discount for each branch that returns a constructor, making us
569 keener to inline. I did try using 'max' instead, but it makes nofib
570 'rewrite' and 'puzzle' allocate significantly more, and didn't make
571 binary sizes shrink significantly either.
573 Note [Discounts and thresholds]
574 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
575 Constants for discounts and thesholds are defined in main/StaticFlags,
576 all of form opt_UF_xxxx. They are:
578 opt_UF_CreationThreshold (45)
579 At a definition site, if the unfolding is bigger than this, we
580 may discard it altogether
582 opt_UF_UseThreshold (6)
583 At a call site, if the unfolding, less discounts, is smaller than
584 this, then it's small enough inline
586 opt_UF_KeennessFactor (1.5)
587 Factor by which the discounts are multiplied before
588 subtracting from size
590 opt_UF_DictDiscount (1)
591 The discount for each occurrence of a dictionary argument
592 as an argument of a class method. Should be pretty small
593 else big functions may get inlined
595 opt_UF_FunAppDiscount (6)
596 Discount for a function argument that is applied. Quite
597 large, because if we inline we avoid the higher-order call.
600 The size of a foreign call or not-dupable PrimOp
603 Note [Function applications]
604 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
605 In a function application (f a b)
607 - If 'f' is an argument to the function being analysed,
608 and there's at least one value arg, record a FunAppDiscount for f
610 - If the application if a PAP (arity > 2 in this example)
611 record a *result* discount (because inlining
612 with "extra" args in the call may mean that we now
613 get a saturated application)
615 Code for manipulating sizes
618 data ExprSize = TooBig
619 | SizeIs FastInt -- Size found
620 (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
621 FastInt -- Size to subtract if result is scrutinised
622 -- by a case expression
624 instance Outputable ExprSize where
625 ppr TooBig = ptext (sLit "TooBig")
626 ppr (SizeIs a _ c) = brackets (int (iBox a) <+> int (iBox c))
628 -- subtract the discount before deciding whether to bale out. eg. we
629 -- want to inline a large constructor application into a selector:
630 -- tup = (a_1, ..., a_99)
631 -- x = case tup of ...
633 mkSizeIs :: FastInt -> FastInt -> Bag (Id, Int) -> FastInt -> ExprSize
634 mkSizeIs max n xs d | (n -# d) ># max = TooBig
635 | otherwise = SizeIs n xs d
637 maxSize :: ExprSize -> ExprSize -> ExprSize
638 maxSize TooBig _ = TooBig
639 maxSize _ TooBig = TooBig
640 maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
644 sizeN :: Int -> ExprSize
646 sizeZero = SizeIs (_ILIT(0)) emptyBag (_ILIT(0))
647 sizeN n = SizeIs (iUnbox n) emptyBag (_ILIT(0))
651 %************************************************************************
653 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
655 %************************************************************************
657 We use 'couldBeSmallEnoughToInline' to avoid exporting inlinings that
658 we ``couldn't possibly use'' on the other side. Can be overridden w/
659 flaggery. Just the same as smallEnoughToInline, except that it has no
663 couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
664 couldBeSmallEnoughToInline threshold rhs
665 = case sizeExpr (iUnbox threshold) [] body of
669 (_, body) = collectBinders rhs
672 smallEnoughToInline :: Unfolding -> Bool
673 smallEnoughToInline (CoreUnfolding {uf_guidance = UnfIfGoodArgs {ug_size = size}})
674 = size <= opt_UF_UseThreshold
675 smallEnoughToInline _
679 certainlyWillInline :: Unfolding -> Bool
680 -- Sees if the unfolding is pretty certain to inline
681 certainlyWillInline (CoreUnfolding { uf_is_cheap = is_cheap, uf_arity = n_vals, uf_guidance = guidance })
685 UnfIfGoodArgs { ug_size = size}
686 -> is_cheap && size - (n_vals +1) <= opt_UF_UseThreshold
688 certainlyWillInline _
692 %************************************************************************
694 \subsection{callSiteInline}
696 %************************************************************************
698 This is the key function. It decides whether to inline a variable at a call site
700 callSiteInline is used at call sites, so it is a bit more generous.
701 It's a very important function that embodies lots of heuristics.
702 A non-WHNF can be inlined if it doesn't occur inside a lambda,
703 and occurs exactly once or
704 occurs once in each branch of a case and is small
706 If the thing is in WHNF, there's no danger of duplicating work,
707 so we can inline if it occurs once, or is small
709 NOTE: we don't want to inline top-level functions that always diverge.
710 It just makes the code bigger. Tt turns out that the convenient way to prevent
711 them inlining is to give them a NOINLINE pragma, which we do in
712 StrictAnal.addStrictnessInfoToTopId
715 callSiteInline :: DynFlags
717 -> Unfolding -- Its unfolding (if active)
718 -> Bool -- True if there are are no arguments at all (incl type args)
719 -> [ArgSummary] -- One for each value arg; True if it is interesting
720 -> CallCtxt -- True <=> continuation is interesting
721 -> Maybe CoreExpr -- Unfolding, if any
724 instance Outputable ArgSummary where
725 ppr TrivArg = ptext (sLit "TrivArg")
726 ppr NonTrivArg = ptext (sLit "NonTrivArg")
727 ppr ValueArg = ptext (sLit "ValueArg")
729 data CallCtxt = BoringCtxt
731 | ArgCtxt -- We are somewhere in the argument of a function
732 Bool -- True <=> we're somewhere in the RHS of function with rules
733 -- False <=> we *are* the argument of a function with non-zero
736 -- we *are* the RHS of a let Note [RHS of lets]
737 -- In both cases, be a little keener to inline
739 | ValAppCtxt -- We're applied to at least one value arg
740 -- This arises when we have ((f x |> co) y)
741 -- Then the (f x) has argument 'x' but in a ValAppCtxt
743 | CaseCtxt -- We're the scrutinee of a case
744 -- that decomposes its scrutinee
746 instance Outputable CallCtxt where
747 ppr BoringCtxt = ptext (sLit "BoringCtxt")
748 ppr (ArgCtxt rules) = ptext (sLit "ArgCtxt") <+> ppr rules
749 ppr CaseCtxt = ptext (sLit "CaseCtxt")
750 ppr ValAppCtxt = ptext (sLit "ValAppCtxt")
752 callSiteInline dflags id unfolding lone_variable arg_infos cont_info
753 = case unfolding of {
754 NoUnfolding -> Nothing ;
755 OtherCon _ -> Nothing ;
756 DFunUnfolding {} -> Nothing ; -- Never unfold a DFun
757 CoreUnfolding { uf_tmpl = unf_template, uf_is_top = is_top,
758 uf_is_cheap = is_cheap, uf_arity = uf_arity, uf_guidance = guidance } ->
759 -- uf_arity will typically be equal to (idArity id),
760 -- but may be less for InlineRules
762 n_val_args = length arg_infos
763 saturated = n_val_args >= uf_arity
765 result | yes_or_no = Just unf_template
766 | otherwise = Nothing
768 interesting_args = any nonTriv arg_infos
769 -- NB: (any nonTriv arg_infos) looks at the
770 -- over-saturated args too which is "wrong";
771 -- but if over-saturated we inline anyway.
773 -- some_benefit is used when the RHS is small enough
774 -- and the call has enough (or too many) value
775 -- arguments (ie n_val_args >= arity). But there must
776 -- be *something* interesting about some argument, or the
777 -- result context, to make it worth inlining
779 | not saturated = interesting_args -- Under-saturated
780 -- Note [Unsaturated applications]
781 | n_val_args > uf_arity = True -- Over-saturated
782 | otherwise = interesting_args -- Saturated
783 || interesting_saturated_call
785 interesting_saturated_call
787 BoringCtxt -> not is_top && uf_arity > 0 -- Note [Nested functions]
788 CaseCtxt -> not (lone_variable && is_cheap) -- Note [Lone variables]
789 ArgCtxt {} -> uf_arity > 0 -- Note [Inlining in ArgCtxt]
790 ValAppCtxt -> True -- Note [Cast then apply]
792 (yes_or_no, extra_doc)
794 UnfNever -> (False, empty)
796 UnfWhen unsat_ok boring_ok
797 -> (enough_args && (boring_ok || some_benefit), empty )
798 where -- See Note [INLINE for small functions]
799 enough_args = saturated || (unsat_ok && n_val_args > 0)
801 UnfIfGoodArgs { ug_args = arg_discounts, ug_res = res_discount, ug_size = size }
802 -> ( is_cheap && some_benefit && small_enough
803 , (text "discounted size =" <+> int discounted_size) )
805 discounted_size = size - discount
806 small_enough = discounted_size <= opt_UF_UseThreshold
807 discount = computeDiscount uf_arity arg_discounts
808 res_discount arg_infos cont_info
811 if (dopt Opt_D_dump_inlinings dflags && dopt Opt_D_verbose_core2core dflags) then
812 pprTrace ("Considering inlining: " ++ showSDoc (ppr id))
813 (vcat [text "arg infos" <+> ppr arg_infos,
814 text "uf arity" <+> ppr uf_arity,
815 text "interesting continuation" <+> ppr cont_info,
816 text "some_benefit" <+> ppr some_benefit,
817 text "is cheap:" <+> ppr is_cheap,
818 text "guidance" <+> ppr guidance,
820 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO"])
829 Be a tiny bit keener to inline in the RHS of a let, because that might
830 lead to good thing later
832 g y = let x = f y in ...(case x of (a,b,c) -> ...) ...
833 We'd inline 'f' if the call was in a case context, and it kind-of-is,
834 only we can't see it. So we treat the RHS of a let as not-totally-boring.
836 Note [Unsaturated applications]
837 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
838 When a call is not saturated, we *still* inline if one of the
839 arguments has interesting structure. That's sometimes very important.
840 A good example is the Ord instance for Bool in Base:
843 $fOrdBool =GHC.Classes.D:Ord
848 $cmin_ajX [Occ=LoopBreaker] :: Bool -> Bool -> Bool
849 $cmin_ajX = GHC.Classes.$dmmin @ Bool $fOrdBool
852 But the defn of GHC.Classes.$dmmin is:
854 $dmmin :: forall a. GHC.Classes.Ord a => a -> a -> a
855 {- Arity: 3, HasNoCafRefs, Strictness: SLL,
856 Unfolding: (\ @ a $dOrd :: GHC.Classes.Ord a x :: a y :: a ->
857 case @ a GHC.Classes.<= @ a $dOrd x y of wild {
858 GHC.Bool.False -> y GHC.Bool.True -> x }) -}
860 We *really* want to inline $dmmin, even though it has arity 3, in
861 order to unravel the recursion.
864 Note [Things to watch]
865 ~~~~~~~~~~~~~~~~~~~~~~
866 * { y = I# 3; x = y `cast` co; ...case (x `cast` co) of ... }
867 Assume x is exported, so not inlined unconditionally.
868 Then we want x to inline unconditionally; no reason for it
869 not to, and doing so avoids an indirection.
871 * { x = I# 3; ....f x.... }
872 Make sure that x does not inline unconditionally!
873 Lest we get extra allocation.
875 Note [Inlining an InlineRule]
876 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
877 An InlineRules is used for
878 (a) programmer INLINE pragmas
879 (b) inlinings from worker/wrapper
881 For (a) the RHS may be large, and our contract is that we *only* inline
882 when the function is applied to all the arguments on the LHS of the
883 source-code defn. (The uf_arity in the rule.)
885 However for worker/wrapper it may be worth inlining even if the
886 arity is not satisfied (as we do in the CoreUnfolding case) so we don't
890 Note [Nested functions]
891 ~~~~~~~~~~~~~~~~~~~~~~~
892 If a function has a nested defn we also record some-benefit, on the
893 grounds that we are often able to eliminate the binding, and hence the
894 allocation, for the function altogether; this is good for join points.
895 But this only makes sense for *functions*; inlining a constructor
896 doesn't help allocation unless the result is scrutinised. UNLESS the
897 constructor occurs just once, albeit possibly in multiple case
898 branches. Then inlining it doesn't increase allocation, but it does
899 increase the chance that the constructor won't be allocated at all in
900 the branches that don't use it.
902 Note [Cast then apply]
903 ~~~~~~~~~~~~~~~~~~~~~~
905 myIndex = __inline_me ( (/\a. <blah>) |> co )
906 co :: (forall a. a -> a) ~ (forall a. T a)
907 ... /\a.\x. case ((myIndex a) |> sym co) x of { ... } ...
909 We need to inline myIndex to unravel this; but the actual call (myIndex a) has
910 no value arguments. The ValAppCtxt gives it enough incentive to inline.
912 Note [Inlining in ArgCtxt]
913 ~~~~~~~~~~~~~~~~~~~~~~~~~~
914 The condition (arity > 0) here is very important, because otherwise
915 we end up inlining top-level stuff into useless places; eg
918 This can make a very big difference: it adds 16% to nofib 'integer' allocs,
921 At one stage I replaced this condition by 'True' (leading to the above
922 slow-down). The motivation was test eyeball/inline1.hs; but that seems
925 NOTE: arguably, we should inline in ArgCtxt only if the result of the
926 call is at least CONLIKE. At least for the cases where we use ArgCtxt
927 for the RHS of a 'let', we only profit from the inlining if we get a
928 CONLIKE thing (modulo lets).
930 Note [Lone variables] See also Note [Interaction of exprIsCheap and lone variables]
931 ~~~~~~~~~~~~~~~~~~~~~ which appears below
932 The "lone-variable" case is important. I spent ages messing about
933 with unsatisfactory varaints, but this is nice. The idea is that if a
934 variable appears all alone
936 as an arg of lazy fn, or rhs BoringCtxt
937 as scrutinee of a case CaseCtxt
938 as arg of a fn ArgCtxt
940 it is bound to a cheap expression
942 then we should not inline it (unless there is some other reason,
943 e.g. is is the sole occurrence). That is what is happening at
944 the use of 'lone_variable' in 'interesting_saturated_call'.
946 Why? At least in the case-scrutinee situation, turning
947 let x = (a,b) in case x of y -> ...
949 let x = (a,b) in case (a,b) of y -> ...
951 let x = (a,b) in let y = (a,b) in ...
952 is bad if the binding for x will remain.
954 Another example: I discovered that strings
955 were getting inlined straight back into applications of 'error'
956 because the latter is strict.
958 f = \x -> ...(error s)...
960 Fundamentally such contexts should not encourage inlining because the
961 context can ``see'' the unfolding of the variable (e.g. case or a
962 RULE) so there's no gain. If the thing is bound to a value.
967 foo = _inline_ (\n. [n])
968 bar = _inline_ (foo 20)
969 baz = \n. case bar of { (m:_) -> m + n }
970 Here we really want to inline 'bar' so that we can inline 'foo'
971 and the whole thing unravels as it should obviously do. This is
972 important: in the NDP project, 'bar' generates a closure data
973 structure rather than a list.
975 So the non-inlining of lone_variables should only apply if the
976 unfolding is regarded as cheap; because that is when exprIsConApp_maybe
977 looks through the unfolding. Hence the "&& is_cheap" in the
980 * Even a type application or coercion isn't a lone variable.
982 case $fMonadST @ RealWorld of { :DMonad a b c -> c }
983 We had better inline that sucker! The case won't see through it.
985 For now, I'm treating treating a variable applied to types
986 in a *lazy* context "lone". The motivating example was
989 There's no advantage in inlining f here, and perhaps
990 a significant disadvantage. Hence some_val_args in the Stop case
992 Note [Interaction of exprIsCheap and lone variables]
993 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
994 The lone-variable test says "don't inline if a case expression
995 scrutines a lone variable whose unfolding is cheap". It's very
996 important that, under these circumstances, exprIsConApp_maybe
997 can spot a constructor application. So, for example, we don't
1000 to be cheap, and that's good because exprIsConApp_maybe doesn't
1001 think that expression is a constructor application.
1003 I used to test is_value rather than is_cheap, which was utterly
1004 wrong, because the above expression responds True to exprIsHNF.
1006 This kind of thing can occur if you have
1009 foo = let x = e in (x,x)
1014 computeDiscount :: Int -> [Int] -> Int -> [ArgSummary] -> CallCtxt -> Int
1015 computeDiscount n_vals_wanted arg_discounts res_discount arg_infos cont_info
1016 -- We multiple the raw discounts (args_discount and result_discount)
1017 -- ty opt_UnfoldingKeenessFactor because the former have to do with
1018 -- *size* whereas the discounts imply that there's some extra
1019 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
1022 = 1 -- Discount of 1 because the result replaces the call
1023 -- so we count 1 for the function itself
1025 + length (take n_vals_wanted arg_infos)
1026 -- Discount of (un-scaled) 1 for each arg supplied,
1027 -- because the result replaces the call
1029 + round (opt_UF_KeenessFactor *
1030 fromIntegral (arg_discount + res_discount'))
1032 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
1034 mk_arg_discount _ TrivArg = 0
1035 mk_arg_discount _ NonTrivArg = 1
1036 mk_arg_discount discount ValueArg = discount
1038 res_discount' = case cont_info of
1040 CaseCtxt -> res_discount
1041 _other -> 4 `min` res_discount
1042 -- res_discount can be very large when a function returns
1043 -- constructors; but we only want to invoke that large discount
1044 -- when there's a case continuation.
1045 -- Otherwise we, rather arbitrarily, threshold it. Yuk.
1046 -- But we want to aovid inlining large functions that return
1047 -- constructors into contexts that are simply "interesting"
1050 %************************************************************************
1052 Interesting arguments
1054 %************************************************************************
1056 Note [Interesting arguments]
1057 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1058 An argument is interesting if it deserves a discount for unfoldings
1059 with a discount in that argument position. The idea is to avoid
1060 unfolding a function that is applied only to variables that have no
1061 unfolding (i.e. they are probably lambda bound): f x y z There is
1062 little point in inlining f here.
1064 Generally, *values* (like (C a b) and (\x.e)) deserve discounts. But
1065 we must look through lets, eg (let x = e in C a b), because the let will
1066 float, exposing the value, if we inline. That makes it different to
1069 Before 2009 we said it was interesting if the argument had *any* structure
1070 at all; i.e. (hasSomeUnfolding v). But does too much inlining; see Trac #3016.
1072 But we don't regard (f x y) as interesting, unless f is unsaturated.
1073 If it's saturated and f hasn't inlined, then it's probably not going
1076 Note [Conlike is interesting]
1077 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1079 f d = ...((*) d x y)...
1081 where df is con-like. Then we'd really like to inline 'f' so that the
1082 rule for (*) (df d) can fire. To do this
1083 a) we give a discount for being an argument of a class-op (eg (*) d)
1084 b) we say that a con-like argument (eg (df d)) is interesting
1087 data ArgSummary = TrivArg -- Nothing interesting
1088 | NonTrivArg -- Arg has structure
1089 | ValueArg -- Arg is a con-app or PAP
1090 -- ..or con-like. Note [Conlike is interesting]
1092 interestingArg :: CoreExpr -> ArgSummary
1093 -- See Note [Interesting arguments]
1094 interestingArg e = go e 0
1096 -- n is # value args to which the expression is applied
1097 go (Lit {}) _ = ValueArg
1099 | isConLikeId v = ValueArg -- Experimenting with 'conlike' rather that
1100 -- data constructors here
1101 | idArity v > n = ValueArg -- Catches (eg) primops with arity but no unfolding
1102 | n > 0 = NonTrivArg -- Saturated or unknown call
1103 | conlike_unfolding = ValueArg -- n==0; look for an interesting unfolding
1104 -- See Note [Conlike is interesting]
1105 | otherwise = TrivArg -- n==0, no useful unfolding
1107 conlike_unfolding = isConLikeUnfolding (idUnfolding v)
1109 go (Type _) _ = TrivArg
1110 go (App fn (Type _)) n = go fn n
1111 go (App fn _) n = go fn (n+1)
1112 go (Note _ a) n = go a n
1113 go (Cast e _) n = go e n
1115 | isTyCoVar v = go e n
1117 | otherwise = ValueArg
1118 go (Let _ e) n = case go e n of { ValueArg -> ValueArg; _ -> NonTrivArg }
1119 go (Case {}) _ = NonTrivArg
1121 nonTriv :: ArgSummary -> Bool
1122 nonTriv TrivArg = False
1126 %************************************************************************
1130 %************************************************************************
1132 Note [exprIsConApp_maybe]
1133 ~~~~~~~~~~~~~~~~~~~~~~~~~
1134 exprIsConApp_maybe is a very important function. There are two principal
1136 * case e of { .... }
1137 * cls_op e, where cls_op is a class operation
1139 In both cases you want to know if e is of form (C e1..en) where C is
1142 However e might not *look* as if
1145 -- | Returns @Just (dc, [t1..tk], [x1..xn])@ if the argument expression is
1146 -- a *saturated* constructor application of the form @dc t1..tk x1 .. xn@,
1147 -- where t1..tk are the *universally-qantified* type args of 'dc'
1148 exprIsConApp_maybe :: IdUnfoldingFun -> CoreExpr -> Maybe (DataCon, [Type], [CoreExpr])
1150 exprIsConApp_maybe id_unf (Note _ expr)
1151 = exprIsConApp_maybe id_unf expr
1152 -- We ignore all notes. For example,
1153 -- case _scc_ "foo" (C a b) of
1155 -- should be optimised away, but it will be only if we look
1156 -- through the SCC note.
1158 exprIsConApp_maybe id_unf (Cast expr co)
1159 = -- Here we do the KPush reduction rule as described in the FC paper
1160 -- The transformation applies iff we have
1161 -- (C e1 ... en) `cast` co
1162 -- where co :: (T t1 .. tn) ~ to_ty
1163 -- The left-hand one must be a T, because exprIsConApp returned True
1164 -- but the right-hand one might not be. (Though it usually will.)
1166 case exprIsConApp_maybe id_unf expr of {
1167 Nothing -> Nothing ;
1168 Just (dc, _dc_univ_args, dc_args) ->
1170 let (_from_ty, to_ty) = coercionKind co
1171 dc_tc = dataConTyCon dc
1173 case splitTyConApp_maybe to_ty of {
1174 Nothing -> Nothing ;
1175 Just (to_tc, to_tc_arg_tys)
1176 | dc_tc /= to_tc -> Nothing
1177 -- These two Nothing cases are possible; we might see
1178 -- (C x y) `cast` (g :: T a ~ S [a]),
1179 -- where S is a type function. In fact, exprIsConApp
1180 -- will probably not be called in such circumstances,
1181 -- but there't nothing wrong with it
1185 tc_arity = tyConArity dc_tc
1186 dc_univ_tyvars = dataConUnivTyVars dc
1187 dc_ex_tyvars = dataConExTyVars dc
1188 arg_tys = dataConRepArgTys dc
1190 dc_eqs :: [(Type,Type)] -- All equalities from the DataCon
1191 dc_eqs = [(mkTyVarTy tv, ty) | (tv,ty) <- dataConEqSpec dc] ++
1192 [getEqPredTys eq_pred | eq_pred <- dataConEqTheta dc]
1194 (ex_args, rest1) = splitAtList dc_ex_tyvars dc_args
1195 (co_args, val_args) = splitAtList dc_eqs rest1
1197 -- Make the "theta" from Fig 3 of the paper
1198 gammas = decomposeCo tc_arity co
1199 theta = zipOpenTvSubst (dc_univ_tyvars ++ dc_ex_tyvars)
1200 (gammas ++ stripTypeArgs ex_args)
1202 -- Cast the existential coercion arguments
1203 cast_co (ty1, ty2) (Type co)
1204 = Type $ mkSymCoercion (substTy theta ty1)
1205 `mkTransCoercion` co
1206 `mkTransCoercion` (substTy theta ty2)
1207 cast_co _ other_arg = pprPanic "cast_co" (ppr other_arg)
1208 new_co_args = zipWith cast_co dc_eqs co_args
1210 -- Cast the value arguments (which include dictionaries)
1211 new_val_args = zipWith cast_arg arg_tys val_args
1212 cast_arg arg_ty arg = mkCoerce (substTy theta arg_ty) arg
1215 let dump_doc = vcat [ppr dc, ppr dc_univ_tyvars, ppr dc_ex_tyvars,
1216 ppr arg_tys, ppr dc_args, ppr _dc_univ_args,
1217 ppr ex_args, ppr val_args]
1219 ASSERT2( coreEqType _from_ty (mkTyConApp dc_tc _dc_univ_args), dump_doc )
1220 ASSERT2( all isTypeArg (ex_args ++ co_args), dump_doc )
1221 ASSERT2( equalLength val_args arg_tys, dump_doc )
1224 Just (dc, to_tc_arg_tys, ex_args ++ new_co_args ++ new_val_args)
1227 exprIsConApp_maybe id_unf expr
1230 analyse (App fun arg) args = analyse fun (arg:args)
1231 analyse fun@(Lam {}) args = beta fun [] args
1233 analyse (Var fun) args
1234 | Just con <- isDataConWorkId_maybe fun
1235 , count isValArg args == idArity fun
1236 , let (univ_ty_args, rest_args) = splitAtList (dataConUnivTyVars con) args
1237 = Just (con, stripTypeArgs univ_ty_args, rest_args)
1239 -- Look through dictionary functions; see Note [Unfolding DFuns]
1240 | DFunUnfolding dfun_nargs con ops <- unfolding
1241 , let sat = length args == dfun_nargs -- See Note [DFun arity check]
1242 in if sat then True else
1243 pprTrace "Unsaturated dfun" (ppr fun <+> int dfun_nargs $$ ppr args) False
1244 , let (dfun_tvs, _cls, dfun_res_tys) = tcSplitDFunTy (idType fun)
1245 subst = zipOpenTvSubst dfun_tvs (stripTypeArgs (takeList dfun_tvs args))
1246 = Just (con, substTys subst dfun_res_tys,
1247 [mkApps op args | op <- ops])
1249 -- Look through unfoldings, but only cheap ones, because
1250 -- we are effectively duplicating the unfolding
1251 | Just rhs <- expandUnfolding_maybe unfolding
1252 = -- pprTrace "expanding" (ppr fun $$ ppr rhs) $
1255 unfolding = id_unf fun
1257 analyse _ _ = Nothing
1260 beta (Lam v body) pairs (arg : args)
1262 = beta body ((v,arg):pairs) args
1264 beta (Lam {}) _ _ -- Un-saturated, or not a type lambda
1268 = analyse (substExpr (text "subst-expr-is-con-app") subst fun) args
1270 subst = mkOpenSubst (mkInScopeSet (exprFreeVars fun)) pairs
1271 -- doc = vcat [ppr fun, ppr expr, ppr pairs, ppr args]
1274 stripTypeArgs :: [CoreExpr] -> [Type]
1275 stripTypeArgs args = ASSERT2( all isTypeArg args, ppr args )
1276 [ty | Type ty <- args]
1279 Note [Unfolding DFuns]
1280 ~~~~~~~~~~~~~~~~~~~~~~
1283 df :: forall a b. (Eq a, Eq b) -> Eq (a,b)
1284 df a b d_a d_b = MkEqD (a,b) ($c1 a b d_a d_b)
1287 So to split it up we just need to apply the ops $c1, $c2 etc
1288 to the very same args as the dfun. It takes a little more work
1289 to compute the type arguments to the dictionary constructor.
1291 Note [DFun arity check]
1292 ~~~~~~~~~~~~~~~~~~~~~~~
1293 Here we check that the total number of supplied arguments (inclding
1294 type args) matches what the dfun is expecting. This may be *less*
1295 than the ordinary arity of the dfun: see Note [DFun unfoldings] in CoreSyn