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)
186 -> UnfWhen unSaturatedOk boringCxtOk -- Note [INLINE for small functions]
188 | top_bot -- See Note [Do not inline top-level bottoming functions]
192 -> UnfIfGoodArgs { ug_args = map (discount cased_bndrs) val_bndrs
193 , ug_size = iBox size
194 , ug_res = iBox scrut_discount }
197 = foldlBag (\acc (b',n) -> if bndr==b' then acc+n else acc)
200 (n_val_bndrs, guidance) }
203 Note [Computing the size of an expression]
204 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
205 The basic idea of sizeExpr is obvious enough: count nodes. But getting the
206 heuristics right has taken a long time. Here's the basic strategy:
208 * Variables, literals: 0
209 (Exception for string literals, see litSize.)
211 * Function applications (f e1 .. en): 1 + #value args
213 * Constructor applications: 1, regardless of #args
215 * Let(rec): 1 + size of components
230 Notice that 'x' counts 0, while (f x) counts 2. That's deliberate: there's
231 a function call to account for. Notice also that constructor applications
232 are very cheap, because exposing them to a caller is so valuable.
235 Note [Do not inline top-level bottoming functions]
236 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
237 The FloatOut pass has gone to some trouble to float out calls to 'error'
238 and similar friends. See Note [Bottoming floats] in SetLevels.
239 Do not re-inline them! But we *do* still inline if they are very small
240 (the uncondInline stuff).
243 Note [INLINE for small functions]
244 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
245 Consider {-# INLINE f #-}
248 Then f's RHS is no larger than its LHS, so we should inline it into
249 even the most boring context. In general, f the function is
250 sufficiently small that its body is as small as the call itself, the
251 inline unconditionally, regardless of how boring the context is.
255 * We inline *unconditionally* if inlined thing is smaller (using sizeExpr)
256 than the thing it's replacing. Notice that
257 (f x) --> (g 3) -- YES, unconditionally
258 (f x) --> x : [] -- YES, *even though* there are two
259 -- arguments to the cons
263 It's very important not to unconditionally replace a variable by
266 * We do this even if the thing isn't saturated, else we end up with the
270 doesn't inline. Even in a boring context, inlining without being
271 saturated will give a lambda instead of a PAP, and will be more
272 efficient at runtime.
274 * However, when the function's arity > 0, we do insist that it
275 has at least one value argument at the call site. Otherwise we find this:
278 If we inline f here we get
279 d = /\b. MkD (\x:b. x)
280 and then prepareRhs floats out the argument, abstracting the type
281 variables, so we end up with the original again!
285 uncondInline :: Arity -> Int -> Bool
286 -- Inline unconditionally if there no size increase
287 -- Size of call is arity (+1 for the function)
288 -- See Note [INLINE for small functions]
289 uncondInline arity size
290 | arity == 0 = size == 0
291 | otherwise = size <= arity + 1
296 sizeExpr :: FastInt -- Bomb out if it gets bigger than this
297 -> [Id] -- Arguments; we're interested in which of these
302 -- Note [Computing the size of an expression]
304 sizeExpr bOMB_OUT_SIZE top_args expr
307 size_up (Cast e _) = size_up e
308 size_up (Note _ e) = size_up e
309 size_up (Type _) = sizeZero -- Types cost nothing
310 size_up (Lit lit) = sizeN (litSize lit)
311 size_up (Var f) = size_up_call f [] -- Make sure we get constructor
312 -- discounts even on nullary constructors
314 size_up (App fun (Type _)) = size_up fun
315 size_up (App fun arg) = size_up arg `addSizeNSD`
316 size_up_app fun [arg]
318 size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
319 | otherwise = size_up e
321 size_up (Let (NonRec binder rhs) body)
322 = size_up rhs `addSizeNSD`
323 size_up body `addSizeN`
324 (if isUnLiftedType (idType binder) then 0 else 1)
325 -- For the allocation
326 -- If the binder has an unlifted type there is no allocation
328 size_up (Let (Rec pairs) body)
329 = foldr (addSizeNSD . size_up . snd)
330 (size_up body `addSizeN` length pairs) -- (length pairs) for the allocation
333 size_up (Case (Var v) _ _ alts)
334 | v `elem` top_args -- We are scrutinising an argument variable
335 = alts_size (foldr1 addAltSize alt_sizes)
336 (foldr1 maxSize alt_sizes)
337 -- Good to inline if an arg is scrutinised, because
338 -- that may eliminate allocation in the caller
339 -- And it eliminates the case itself
341 alt_sizes = map size_up_alt alts
343 -- alts_size tries to compute a good discount for
344 -- the case when we are scrutinising an argument variable
345 alts_size (SizeIs tot tot_disc tot_scrut) -- Size of all alternatives
346 (SizeIs max _ _) -- Size of biggest alternative
347 = SizeIs tot (unitBag (v, iBox (_ILIT(2) +# tot -# max)) `unionBags` tot_disc) tot_scrut
348 -- If the variable is known, we produce a discount that
349 -- will take us back to 'max', the size of the largest alternative
350 -- The 1+ is a little discount for reduced allocation in the caller
352 -- Notice though, that we return tot_disc, the total discount from
353 -- all branches. I think that's right.
355 alts_size tot_size _ = tot_size
357 size_up (Case e _ _ alts) = size_up e `addSizeNSD`
358 foldr (addAltSize . size_up_alt) sizeZero alts
359 -- We don't charge for the case itself
360 -- It's a strict thing, and the price of the call
361 -- is paid by scrut. Also consider
362 -- case f x of DEFAULT -> e
363 -- This is just ';'! Don't charge for it.
365 -- Moreover, we charge one per alternative.
368 -- size_up_app is used when there's ONE OR MORE value args
369 size_up_app (App fun arg) args
370 | isTypeArg arg = size_up_app fun args
371 | otherwise = size_up arg `addSizeNSD`
372 size_up_app fun (arg:args)
373 size_up_app (Var fun) args = size_up_call fun args
374 size_up_app other args = size_up other `addSizeN` length args
377 size_up_call :: Id -> [CoreExpr] -> ExprSize
378 size_up_call fun val_args
379 = case idDetails fun of
380 FCallId _ -> sizeN opt_UF_DearOp
381 DataConWorkId dc -> conSize dc (length val_args)
382 PrimOpId op -> primOpSize op (length val_args)
383 ClassOpId _ -> classOpSize top_args val_args
384 _ -> funSize top_args fun (length val_args)
387 size_up_alt (_con, _bndrs, rhs) = size_up rhs `addSizeN` 1
388 -- Don't charge for args, so that wrappers look cheap
389 -- (See comments about wrappers with Case)
391 -- IMPORATANT: *do* charge 1 for the alternative, else we
392 -- find that giant case nests are treated as practically free
393 -- A good example is Foreign.C.Error.errrnoToIOError
396 -- These addSize things have to be here because
397 -- I don't want to give them bOMB_OUT_SIZE as an argument
398 addSizeN TooBig _ = TooBig
399 addSizeN (SizeIs n xs d) m = mkSizeIs bOMB_OUT_SIZE (n +# iUnbox m) xs d
401 -- addAltSize is used to add the sizes of case alternatives
402 addAltSize TooBig _ = TooBig
403 addAltSize _ TooBig = TooBig
404 addAltSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
405 = mkSizeIs bOMB_OUT_SIZE (n1 +# n2)
407 (d1 +# d2) -- Note [addAltSize result discounts]
409 -- This variant ignores the result discount from its LEFT argument
410 -- It's used when the second argument isn't part of the result
411 addSizeNSD TooBig _ = TooBig
412 addSizeNSD _ TooBig = TooBig
413 addSizeNSD (SizeIs n1 xs _) (SizeIs n2 ys d2)
414 = mkSizeIs bOMB_OUT_SIZE (n1 +# n2)
420 -- | Finds a nominal size of a string literal.
421 litSize :: Literal -> Int
422 -- Used by CoreUnfold.sizeExpr
423 litSize (MachStr str) = 1 + ((lengthFS str + 3) `div` 4)
424 -- If size could be 0 then @f "x"@ might be too small
425 -- [Sept03: make literal strings a bit bigger to avoid fruitless
426 -- duplication of little strings]
427 litSize _other = 0 -- Must match size of nullary constructors
428 -- Key point: if x |-> 4, then x must inline unconditionally
429 -- (eg via case binding)
431 classOpSize :: [Id] -> [CoreExpr] -> ExprSize
432 -- See Note [Conlike is interesting]
435 classOpSize top_args (arg1 : other_args)
436 = SizeIs (iUnbox size) arg_discount (_ILIT(0))
438 size = 2 + length other_args
439 -- If the class op is scrutinising a lambda bound dictionary then
440 -- give it a discount, to encourage the inlining of this function
441 -- The actual discount is rather arbitrarily chosen
442 arg_discount = case arg1 of
443 Var dict | dict `elem` top_args
444 -> unitBag (dict, opt_UF_DictDiscount)
447 funSize :: [Id] -> Id -> Int -> ExprSize
448 -- Size for functions that are not constructors or primops
449 -- Note [Function applications]
450 funSize top_args fun n_val_args
451 | fun `hasKey` buildIdKey = buildSize
452 | fun `hasKey` augmentIdKey = augmentSize
453 | otherwise = SizeIs (iUnbox size) arg_discount (iUnbox res_discount)
455 some_val_args = n_val_args > 0
457 arg_discount | some_val_args && fun `elem` top_args
458 = unitBag (fun, opt_UF_FunAppDiscount)
459 | otherwise = emptyBag
460 -- If the function is an argument and is applied
461 -- to some values, give it an arg-discount
463 res_discount | idArity fun > n_val_args = opt_UF_FunAppDiscount
465 -- If the function is partially applied, show a result discount
467 size | some_val_args = 1 + n_val_args
469 -- The 1+ is for the function itself
470 -- Add 1 for each non-trivial arg;
471 -- the allocation cost, as in let(rec)
474 conSize :: DataCon -> Int -> ExprSize
475 conSize dc n_val_args
476 | n_val_args == 0 = SizeIs (_ILIT(0)) emptyBag (_ILIT(1)) -- Like variables
477 | isUnboxedTupleCon dc = SizeIs (_ILIT(0)) emptyBag (iUnbox n_val_args +# _ILIT(1))
478 | otherwise = SizeIs (_ILIT(1)) emptyBag (iUnbox n_val_args +# _ILIT(1))
479 -- Treat a constructors application as size 1, regardless of how
480 -- many arguments it has; we are keen to expose them
481 -- (and we charge separately for their args). We can't treat
482 -- them as size zero, else we find that (Just x) has size 0,
483 -- which is the same as a lone variable; and hence 'v' will
484 -- always be replaced by (Just x), where v is bound to Just x.
486 -- However, unboxed tuples count as size zero
487 -- I found occasions where we had
488 -- f x y z = case op# x y z of { s -> (# s, () #) }
489 -- and f wasn't getting inlined
491 primOpSize :: PrimOp -> Int -> ExprSize
492 primOpSize op n_val_args
493 | not (primOpIsDupable op) = sizeN opt_UF_DearOp
494 | not (primOpOutOfLine op) = sizeN 1
495 -- Be very keen to inline simple primops.
496 -- We give a discount of 1 for each arg so that (op# x y z) costs 2.
497 -- We can't make it cost 1, else we'll inline let v = (op# x y z)
498 -- at every use of v, which is excessive.
500 -- A good example is:
501 -- let x = +# p q in C {x}
502 -- Even though x get's an occurrence of 'many', its RHS looks cheap,
503 -- and there's a good chance it'll get inlined back into C's RHS. Urgh!
505 | otherwise = sizeN n_val_args
508 buildSize :: ExprSize
509 buildSize = SizeIs (_ILIT(0)) emptyBag (_ILIT(4))
510 -- We really want to inline applications of build
511 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
512 -- Indeed, we should add a result_discount becuause build is
513 -- very like a constructor. We don't bother to check that the
514 -- build is saturated (it usually is). The "-2" discounts for the \c n,
515 -- The "4" is rather arbitrary.
517 augmentSize :: ExprSize
518 augmentSize = SizeIs (_ILIT(0)) emptyBag (_ILIT(4))
519 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
520 -- e plus ys. The -2 accounts for the \cn
522 -- When we return a lambda, give a discount if it's used (applied)
523 lamScrutDiscount :: ExprSize -> ExprSize
524 lamScrutDiscount (SizeIs n vs _) = SizeIs n vs (iUnbox opt_UF_FunAppDiscount)
525 lamScrutDiscount TooBig = TooBig
528 Note [addAltSize result discounts]
529 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
530 When adding the size of alternatives, we *add* the result discounts
531 too, rather than take the *maximum*. For a multi-branch case, this
532 gives a discount for each branch that returns a constructor, making us
533 keener to inline. I did try using 'max' instead, but it makes nofib
534 'rewrite' and 'puzzle' allocate significantly more, and didn't make
535 binary sizes shrink significantly either.
537 Note [Discounts and thresholds]
538 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
539 Constants for discounts and thesholds are defined in main/StaticFlags,
540 all of form opt_UF_xxxx. They are:
542 opt_UF_CreationThreshold (45)
543 At a definition site, if the unfolding is bigger than this, we
544 may discard it altogether
546 opt_UF_UseThreshold (6)
547 At a call site, if the unfolding, less discounts, is smaller than
548 this, then it's small enough inline
550 opt_UF_KeennessFactor (1.5)
551 Factor by which the discounts are multiplied before
552 subtracting from size
554 opt_UF_DictDiscount (1)
555 The discount for each occurrence of a dictionary argument
556 as an argument of a class method. Should be pretty small
557 else big functions may get inlined
559 opt_UF_FunAppDiscount (6)
560 Discount for a function argument that is applied. Quite
561 large, because if we inline we avoid the higher-order call.
564 The size of a foreign call or not-dupable PrimOp
567 Note [Function applications]
568 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
569 In a function application (f a b)
571 - If 'f' is an argument to the function being analysed,
572 and there's at least one value arg, record a FunAppDiscount for f
574 - If the application if a PAP (arity > 2 in this example)
575 record a *result* discount (because inlining
576 with "extra" args in the call may mean that we now
577 get a saturated application)
579 Code for manipulating sizes
582 data ExprSize = TooBig
583 | SizeIs FastInt -- Size found
584 (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
585 FastInt -- Size to subtract if result is scrutinised
586 -- by a case expression
588 instance Outputable ExprSize where
589 ppr TooBig = ptext (sLit "TooBig")
590 ppr (SizeIs a _ c) = brackets (int (iBox a) <+> int (iBox c))
592 -- subtract the discount before deciding whether to bale out. eg. we
593 -- want to inline a large constructor application into a selector:
594 -- tup = (a_1, ..., a_99)
595 -- x = case tup of ...
597 mkSizeIs :: FastInt -> FastInt -> Bag (Id, Int) -> FastInt -> ExprSize
598 mkSizeIs max n xs d | (n -# d) ># max = TooBig
599 | otherwise = SizeIs n xs d
601 maxSize :: ExprSize -> ExprSize -> ExprSize
602 maxSize TooBig _ = TooBig
603 maxSize _ TooBig = TooBig
604 maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
608 sizeN :: Int -> ExprSize
610 sizeZero = SizeIs (_ILIT(0)) emptyBag (_ILIT(0))
611 sizeN n = SizeIs (iUnbox n) emptyBag (_ILIT(0))
615 %************************************************************************
617 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
619 %************************************************************************
621 We use 'couldBeSmallEnoughToInline' to avoid exporting inlinings that
622 we ``couldn't possibly use'' on the other side. Can be overridden w/
623 flaggery. Just the same as smallEnoughToInline, except that it has no
627 couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
628 couldBeSmallEnoughToInline threshold rhs
629 = case calcUnfoldingGuidance False False threshold rhs of
630 (_, UnfNever) -> False
634 smallEnoughToInline :: Unfolding -> Bool
635 smallEnoughToInline (CoreUnfolding {uf_guidance = UnfIfGoodArgs {ug_size = size}})
636 = size <= opt_UF_UseThreshold
637 smallEnoughToInline _
641 certainlyWillInline :: Unfolding -> Bool
642 -- Sees if the unfolding is pretty certain to inline
643 certainlyWillInline (CoreUnfolding { uf_is_cheap = is_cheap, uf_arity = n_vals, uf_guidance = guidance })
647 UnfIfGoodArgs { ug_size = size}
648 -> is_cheap && size - (n_vals +1) <= opt_UF_UseThreshold
650 certainlyWillInline _
654 %************************************************************************
656 \subsection{callSiteInline}
658 %************************************************************************
660 This is the key function. It decides whether to inline a variable at a call site
662 callSiteInline is used at call sites, so it is a bit more generous.
663 It's a very important function that embodies lots of heuristics.
664 A non-WHNF can be inlined if it doesn't occur inside a lambda,
665 and occurs exactly once or
666 occurs once in each branch of a case and is small
668 If the thing is in WHNF, there's no danger of duplicating work,
669 so we can inline if it occurs once, or is small
671 NOTE: we don't want to inline top-level functions that always diverge.
672 It just makes the code bigger. Tt turns out that the convenient way to prevent
673 them inlining is to give them a NOINLINE pragma, which we do in
674 StrictAnal.addStrictnessInfoToTopId
677 callSiteInline :: DynFlags
679 -> Unfolding -- Its unfolding (if active)
680 -> Bool -- True if there are are no arguments at all (incl type args)
681 -> [ArgSummary] -- One for each value arg; True if it is interesting
682 -> CallCtxt -- True <=> continuation is interesting
683 -> Maybe CoreExpr -- Unfolding, if any
686 instance Outputable ArgSummary where
687 ppr TrivArg = ptext (sLit "TrivArg")
688 ppr NonTrivArg = ptext (sLit "NonTrivArg")
689 ppr ValueArg = ptext (sLit "ValueArg")
691 data CallCtxt = BoringCtxt
693 | ArgCtxt -- We are somewhere in the argument of a function
694 Bool -- True <=> we're somewhere in the RHS of function with rules
695 -- False <=> we *are* the argument of a function with non-zero
698 -- we *are* the RHS of a let Note [RHS of lets]
699 -- In both cases, be a little keener to inline
701 | ValAppCtxt -- We're applied to at least one value arg
702 -- This arises when we have ((f x |> co) y)
703 -- Then the (f x) has argument 'x' but in a ValAppCtxt
705 | CaseCtxt -- We're the scrutinee of a case
706 -- that decomposes its scrutinee
708 instance Outputable CallCtxt where
709 ppr BoringCtxt = ptext (sLit "BoringCtxt")
710 ppr (ArgCtxt rules) = ptext (sLit "ArgCtxt") <+> ppr rules
711 ppr CaseCtxt = ptext (sLit "CaseCtxt")
712 ppr ValAppCtxt = ptext (sLit "ValAppCtxt")
714 callSiteInline dflags id unfolding lone_variable arg_infos cont_info
715 = case unfolding of {
716 NoUnfolding -> Nothing ;
717 OtherCon _ -> Nothing ;
718 DFunUnfolding {} -> Nothing ; -- Never unfold a DFun
719 CoreUnfolding { uf_tmpl = unf_template, uf_is_top = is_top, uf_is_value = is_value,
720 uf_is_cheap = is_cheap, uf_arity = uf_arity, uf_guidance = guidance } ->
721 -- uf_arity will typically be equal to (idArity id),
722 -- but may be less for InlineRules
724 n_val_args = length arg_infos
725 saturated = n_val_args >= uf_arity
727 result | yes_or_no = Just unf_template
728 | otherwise = Nothing
730 interesting_args = any nonTriv arg_infos
731 -- NB: (any nonTriv arg_infos) looks at the
732 -- over-saturated args too which is "wrong";
733 -- but if over-saturated we inline anyway.
735 -- some_benefit is used when the RHS is small enough
736 -- and the call has enough (or too many) value
737 -- arguments (ie n_val_args >= arity). But there must
738 -- be *something* interesting about some argument, or the
739 -- result context, to make it worth inlining
741 | not saturated = interesting_args -- Under-saturated
742 -- Note [Unsaturated applications]
743 | n_val_args > uf_arity = True -- Over-saturated
744 | otherwise = interesting_args -- Saturated
745 || interesting_saturated_call
747 interesting_saturated_call
749 BoringCtxt -> not is_top && uf_arity > 0 -- Note [Nested functions]
750 CaseCtxt -> not (lone_variable && is_value) -- Note [Lone variables]
751 ArgCtxt {} -> uf_arity > 0 -- Note [Inlining in ArgCtxt]
752 ValAppCtxt -> True -- Note [Cast then apply]
754 (yes_or_no, extra_doc)
756 UnfNever -> (False, empty)
758 UnfWhen unsat_ok boring_ok
759 -> (enough_args && (boring_ok || some_benefit), empty )
760 where -- See Note [INLINE for small functions]
761 enough_args = saturated || (unsat_ok && n_val_args > 0)
763 UnfIfGoodArgs { ug_args = arg_discounts, ug_res = res_discount, ug_size = size }
764 -> ( is_cheap && some_benefit && small_enough
765 , (text "discounted size =" <+> int discounted_size) )
767 discounted_size = size - discount
768 small_enough = discounted_size <= opt_UF_UseThreshold
769 discount = computeDiscount uf_arity arg_discounts
770 res_discount arg_infos cont_info
773 if (dopt Opt_D_dump_inlinings dflags && dopt Opt_D_verbose_core2core dflags) then
774 pprTrace ("Considering inlining: " ++ showSDoc (ppr id))
775 (vcat [text "arg infos" <+> ppr arg_infos,
776 text "uf arity" <+> ppr uf_arity,
777 text "interesting continuation" <+> ppr cont_info,
778 text "some_benefit" <+> ppr some_benefit,
779 text "is value:" <+> ppr is_value,
780 text "is cheap:" <+> ppr is_cheap,
781 text "guidance" <+> ppr guidance,
783 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO"])
792 Be a tiny bit keener to inline in the RHS of a let, because that might
793 lead to good thing later
795 g y = let x = f y in ...(case x of (a,b,c) -> ...) ...
796 We'd inline 'f' if the call was in a case context, and it kind-of-is,
797 only we can't see it. So we treat the RHS of a let as not-totally-boring.
799 Note [Unsaturated applications]
800 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
801 When a call is not saturated, we *still* inline if one of the
802 arguments has interesting structure. That's sometimes very important.
803 A good example is the Ord instance for Bool in Base:
806 $fOrdBool =GHC.Classes.D:Ord
811 $cmin_ajX [Occ=LoopBreaker] :: Bool -> Bool -> Bool
812 $cmin_ajX = GHC.Classes.$dmmin @ Bool $fOrdBool
815 But the defn of GHC.Classes.$dmmin is:
817 $dmmin :: forall a. GHC.Classes.Ord a => a -> a -> a
818 {- Arity: 3, HasNoCafRefs, Strictness: SLL,
819 Unfolding: (\ @ a $dOrd :: GHC.Classes.Ord a x :: a y :: a ->
820 case @ a GHC.Classes.<= @ a $dOrd x y of wild {
821 GHC.Bool.False -> y GHC.Bool.True -> x }) -}
823 We *really* want to inline $dmmin, even though it has arity 3, in
824 order to unravel the recursion.
827 Note [Things to watch]
828 ~~~~~~~~~~~~~~~~~~~~~~
829 * { y = I# 3; x = y `cast` co; ...case (x `cast` co) of ... }
830 Assume x is exported, so not inlined unconditionally.
831 Then we want x to inline unconditionally; no reason for it
832 not to, and doing so avoids an indirection.
834 * { x = I# 3; ....f x.... }
835 Make sure that x does not inline unconditionally!
836 Lest we get extra allocation.
838 Note [Inlining an InlineRule]
839 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
840 An InlineRules is used for
841 (a) programmer INLINE pragmas
842 (b) inlinings from worker/wrapper
844 For (a) the RHS may be large, and our contract is that we *only* inline
845 when the function is applied to all the arguments on the LHS of the
846 source-code defn. (The uf_arity in the rule.)
848 However for worker/wrapper it may be worth inlining even if the
849 arity is not satisfied (as we do in the CoreUnfolding case) so we don't
853 Note [Nested functions]
854 ~~~~~~~~~~~~~~~~~~~~~~~
855 If a function has a nested defn we also record some-benefit, on the
856 grounds that we are often able to eliminate the binding, and hence the
857 allocation, for the function altogether; this is good for join points.
858 But this only makes sense for *functions*; inlining a constructor
859 doesn't help allocation unless the result is scrutinised. UNLESS the
860 constructor occurs just once, albeit possibly in multiple case
861 branches. Then inlining it doesn't increase allocation, but it does
862 increase the chance that the constructor won't be allocated at all in
863 the branches that don't use it.
865 Note [Cast then apply]
866 ~~~~~~~~~~~~~~~~~~~~~~
868 myIndex = __inline_me ( (/\a. <blah>) |> co )
869 co :: (forall a. a -> a) ~ (forall a. T a)
870 ... /\a.\x. case ((myIndex a) |> sym co) x of { ... } ...
872 We need to inline myIndex to unravel this; but the actual call (myIndex a) has
873 no value arguments. The ValAppCtxt gives it enough incentive to inline.
875 Note [Inlining in ArgCtxt]
876 ~~~~~~~~~~~~~~~~~~~~~~~~~~
877 The condition (arity > 0) here is very important, because otherwise
878 we end up inlining top-level stuff into useless places; eg
881 This can make a very big difference: it adds 16% to nofib 'integer' allocs,
884 At one stage I replaced this condition by 'True' (leading to the above
885 slow-down). The motivation was test eyeball/inline1.hs; but that seems
888 NOTE: arguably, we should inline in ArgCtxt only if the result of the
889 call is at least CONLIKE. At least for the cases where we use ArgCtxt
890 for the RHS of a 'let', we only profit from the inlining if we get a
891 CONLIKE thing (modulo lets).
893 Note [Lone variables]
894 ~~~~~~~~~~~~~~~~~~~~~
895 The "lone-variable" case is important. I spent ages messing about
896 with unsatisfactory varaints, but this is nice. The idea is that if a
897 variable appears all alone
899 as an arg of lazy fn, or rhs BoringCtxt
900 as scrutinee of a case CaseCtxt
901 as arg of a fn ArgCtxt
903 it is bound to a value
905 then we should not inline it (unless there is some other reason,
906 e.g. is is the sole occurrence). That is what is happening at
907 the use of 'lone_variable' in 'interesting_saturated_call'.
909 Why? At least in the case-scrutinee situation, turning
910 let x = (a,b) in case x of y -> ...
912 let x = (a,b) in case (a,b) of y -> ...
914 let x = (a,b) in let y = (a,b) in ...
915 is bad if the binding for x will remain.
917 Another example: I discovered that strings
918 were getting inlined straight back into applications of 'error'
919 because the latter is strict.
921 f = \x -> ...(error s)...
923 Fundamentally such contexts should not encourage inlining because the
924 context can ``see'' the unfolding of the variable (e.g. case or a
925 RULE) so there's no gain. If the thing is bound to a value.
930 foo = _inline_ (\n. [n])
931 bar = _inline_ (foo 20)
932 baz = \n. case bar of { (m:_) -> m + n }
933 Here we really want to inline 'bar' so that we can inline 'foo'
934 and the whole thing unravels as it should obviously do. This is
935 important: in the NDP project, 'bar' generates a closure data
936 structure rather than a list.
938 So the non-inlining of lone_variables should only apply if the
939 unfolding is regarded as cheap; because that is when exprIsConApp_maybe
940 looks through the unfolding. Hence the "&& is_cheap" in the
943 * Even a type application or coercion isn't a lone variable.
945 case $fMonadST @ RealWorld of { :DMonad a b c -> c }
946 We had better inline that sucker! The case won't see through it.
948 For now, I'm treating treating a variable applied to types
949 in a *lazy* context "lone". The motivating example was
952 There's no advantage in inlining f here, and perhaps
953 a significant disadvantage. Hence some_val_args in the Stop case
956 computeDiscount :: Int -> [Int] -> Int -> [ArgSummary] -> CallCtxt -> Int
957 computeDiscount n_vals_wanted arg_discounts res_discount arg_infos cont_info
958 -- We multiple the raw discounts (args_discount and result_discount)
959 -- ty opt_UnfoldingKeenessFactor because the former have to do with
960 -- *size* whereas the discounts imply that there's some extra
961 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
964 = 1 -- Discount of 1 because the result replaces the call
965 -- so we count 1 for the function itself
967 + length (take n_vals_wanted arg_infos)
968 -- Discount of (un-scaled) 1 for each arg supplied,
969 -- because the result replaces the call
971 + round (opt_UF_KeenessFactor *
972 fromIntegral (arg_discount + res_discount'))
974 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
976 mk_arg_discount _ TrivArg = 0
977 mk_arg_discount _ NonTrivArg = 1
978 mk_arg_discount discount ValueArg = discount
980 res_discount' = case cont_info of
982 CaseCtxt -> res_discount
983 _other -> 4 `min` res_discount
984 -- res_discount can be very large when a function returns
985 -- constructors; but we only want to invoke that large discount
986 -- when there's a case continuation.
987 -- Otherwise we, rather arbitrarily, threshold it. Yuk.
988 -- But we want to aovid inlining large functions that return
989 -- constructors into contexts that are simply "interesting"
992 %************************************************************************
994 Interesting arguments
996 %************************************************************************
998 Note [Interesting arguments]
999 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1000 An argument is interesting if it deserves a discount for unfoldings
1001 with a discount in that argument position. The idea is to avoid
1002 unfolding a function that is applied only to variables that have no
1003 unfolding (i.e. they are probably lambda bound): f x y z There is
1004 little point in inlining f here.
1006 Generally, *values* (like (C a b) and (\x.e)) deserve discounts. But
1007 we must look through lets, eg (let x = e in C a b), because the let will
1008 float, exposing the value, if we inline. That makes it different to
1011 Before 2009 we said it was interesting if the argument had *any* structure
1012 at all; i.e. (hasSomeUnfolding v). But does too much inlining; see Trac #3016.
1014 But we don't regard (f x y) as interesting, unless f is unsaturated.
1015 If it's saturated and f hasn't inlined, then it's probably not going
1018 Note [Conlike is interesting]
1019 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1021 f d = ...((*) d x y)...
1023 where df is con-like. Then we'd really like to inline 'f' so that the
1024 rule for (*) (df d) can fire. To do this
1025 a) we give a discount for being an argument of a class-op (eg (*) d)
1026 b) we say that a con-like argument (eg (df d)) is interesting
1029 data ArgSummary = TrivArg -- Nothing interesting
1030 | NonTrivArg -- Arg has structure
1031 | ValueArg -- Arg is a con-app or PAP
1032 -- ..or con-like. Note [Conlike is interesting]
1034 interestingArg :: CoreExpr -> ArgSummary
1035 -- See Note [Interesting arguments]
1036 interestingArg e = go e 0
1038 -- n is # value args to which the expression is applied
1039 go (Lit {}) _ = ValueArg
1041 | isConLikeId v = ValueArg -- Experimenting with 'conlike' rather that
1042 -- data constructors here
1043 | idArity v > n = ValueArg -- Catches (eg) primops with arity but no unfolding
1044 | n > 0 = NonTrivArg -- Saturated or unknown call
1045 | conlike_unfolding = ValueArg -- n==0; look for an interesting unfolding
1046 -- See Note [Conlike is interesting]
1047 | otherwise = TrivArg -- n==0, no useful unfolding
1049 conlike_unfolding = isConLikeUnfolding (idUnfolding v)
1051 go (Type _) _ = TrivArg
1052 go (App fn (Type _)) n = go fn n
1053 go (App fn _) n = go fn (n+1)
1054 go (Note _ a) n = go a n
1055 go (Cast e _) n = go e n
1057 | isTyVar v = go e n
1059 | otherwise = ValueArg
1060 go (Let _ e) n = case go e n of { ValueArg -> ValueArg; _ -> NonTrivArg }
1061 go (Case {}) _ = NonTrivArg
1063 nonTriv :: ArgSummary -> Bool
1064 nonTriv TrivArg = False
1068 %************************************************************************
1072 %************************************************************************
1074 Note [exprIsConApp_maybe]
1075 ~~~~~~~~~~~~~~~~~~~~~~~~~
1076 exprIsConApp_maybe is a very important function. There are two principal
1078 * case e of { .... }
1079 * cls_op e, where cls_op is a class operation
1081 In both cases you want to know if e is of form (C e1..en) where C is
1084 However e might not *look* as if
1087 -- | Returns @Just (dc, [t1..tk], [x1..xn])@ if the argument expression is
1088 -- a *saturated* constructor application of the form @dc t1..tk x1 .. xn@,
1089 -- where t1..tk are the *universally-qantified* type args of 'dc'
1090 exprIsConApp_maybe :: IdUnfoldingFun -> CoreExpr -> Maybe (DataCon, [Type], [CoreExpr])
1092 exprIsConApp_maybe id_unf (Note _ expr)
1093 = exprIsConApp_maybe id_unf expr
1094 -- We ignore all notes. For example,
1095 -- case _scc_ "foo" (C a b) of
1097 -- should be optimised away, but it will be only if we look
1098 -- through the SCC note.
1100 exprIsConApp_maybe id_unf (Cast expr co)
1101 = -- Here we do the KPush reduction rule as described in the FC paper
1102 -- The transformation applies iff we have
1103 -- (C e1 ... en) `cast` co
1104 -- where co :: (T t1 .. tn) ~ to_ty
1105 -- The left-hand one must be a T, because exprIsConApp returned True
1106 -- but the right-hand one might not be. (Though it usually will.)
1108 case exprIsConApp_maybe id_unf expr of {
1109 Nothing -> Nothing ;
1110 Just (dc, _dc_univ_args, dc_args) ->
1112 let (_from_ty, to_ty) = coercionKind co
1113 dc_tc = dataConTyCon dc
1115 case splitTyConApp_maybe to_ty of {
1116 Nothing -> Nothing ;
1117 Just (to_tc, to_tc_arg_tys)
1118 | dc_tc /= to_tc -> Nothing
1119 -- These two Nothing cases are possible; we might see
1120 -- (C x y) `cast` (g :: T a ~ S [a]),
1121 -- where S is a type function. In fact, exprIsConApp
1122 -- will probably not be called in such circumstances,
1123 -- but there't nothing wrong with it
1127 tc_arity = tyConArity dc_tc
1128 dc_univ_tyvars = dataConUnivTyVars dc
1129 dc_ex_tyvars = dataConExTyVars dc
1130 arg_tys = dataConRepArgTys dc
1132 dc_eqs :: [(Type,Type)] -- All equalities from the DataCon
1133 dc_eqs = [(mkTyVarTy tv, ty) | (tv,ty) <- dataConEqSpec dc] ++
1134 [getEqPredTys eq_pred | eq_pred <- dataConEqTheta dc]
1136 (ex_args, rest1) = splitAtList dc_ex_tyvars dc_args
1137 (co_args, val_args) = splitAtList dc_eqs rest1
1139 -- Make the "theta" from Fig 3 of the paper
1140 gammas = decomposeCo tc_arity co
1141 theta = zipOpenTvSubst (dc_univ_tyvars ++ dc_ex_tyvars)
1142 (gammas ++ stripTypeArgs ex_args)
1144 -- Cast the existential coercion arguments
1145 cast_co (ty1, ty2) (Type co)
1146 = Type $ mkSymCoercion (substTy theta ty1)
1147 `mkTransCoercion` co
1148 `mkTransCoercion` (substTy theta ty2)
1149 cast_co _ other_arg = pprPanic "cast_co" (ppr other_arg)
1150 new_co_args = zipWith cast_co dc_eqs co_args
1152 -- Cast the value arguments (which include dictionaries)
1153 new_val_args = zipWith cast_arg arg_tys val_args
1154 cast_arg arg_ty arg = mkCoerce (substTy theta arg_ty) arg
1157 let dump_doc = vcat [ppr dc, ppr dc_univ_tyvars, ppr dc_ex_tyvars,
1158 ppr arg_tys, ppr dc_args, ppr _dc_univ_args,
1159 ppr ex_args, ppr val_args]
1161 ASSERT2( coreEqType _from_ty (mkTyConApp dc_tc _dc_univ_args), dump_doc )
1162 ASSERT2( all isTypeArg (ex_args ++ co_args), dump_doc )
1163 ASSERT2( equalLength val_args arg_tys, dump_doc )
1166 Just (dc, to_tc_arg_tys, ex_args ++ new_co_args ++ new_val_args)
1169 exprIsConApp_maybe id_unf expr
1172 analyse (App fun arg) args = analyse fun (arg:args)
1173 analyse fun@(Lam {}) args = beta fun [] args
1175 analyse (Var fun) args
1176 | Just con <- isDataConWorkId_maybe fun
1178 , let (univ_ty_args, rest_args) = splitAtList (dataConUnivTyVars con) args
1179 = Just (con, stripTypeArgs univ_ty_args, rest_args)
1181 -- Look through dictionary functions; see Note [Unfolding DFuns]
1182 | DFunUnfolding con ops <- unfolding
1184 , let (dfun_tvs, _cls, dfun_res_tys) = tcSplitDFunTy (idType fun)
1185 subst = zipOpenTvSubst dfun_tvs (stripTypeArgs (takeList dfun_tvs args))
1186 = Just (con, substTys subst dfun_res_tys,
1187 [mkApps op args | op <- ops])
1189 -- Look through unfoldings, but only cheap ones, because
1190 -- we are effectively duplicating the unfolding
1191 | Just rhs <- expandUnfolding_maybe unfolding
1192 = -- pprTrace "expanding" (ppr fun $$ ppr rhs) $
1195 is_saturated = count isValArg args == idArity fun
1196 unfolding = id_unf fun
1198 analyse _ _ = Nothing
1201 beta (Lam v body) pairs (arg : args)
1203 = beta body ((v,arg):pairs) args
1205 beta (Lam {}) _ _ -- Un-saturated, or not a type lambda
1209 = case analyse (substExpr (text "subst-expr-is-con-app") subst fun) args of
1210 Nothing -> -- pprTrace "Bale out! exprIsConApp_maybe" doc $
1212 Just ans -> -- pprTrace "Woo-hoo! exprIsConApp_maybe" doc $
1215 subst = mkOpenSubst (mkInScopeSet (exprFreeVars fun)) pairs
1216 -- doc = vcat [ppr fun, ppr expr, ppr pairs, ppr args]
1219 stripTypeArgs :: [CoreExpr] -> [Type]
1220 stripTypeArgs args = ASSERT2( all isTypeArg args, ppr args )
1221 [ty | Type ty <- args]
1224 Note [Unfolding DFuns]
1225 ~~~~~~~~~~~~~~~~~~~~~~
1228 df :: forall a b. (Eq a, Eq b) -> Eq (a,b)
1229 df a b d_a d_b = MkEqD (a,b) ($c1 a b d_a d_b)
1232 So to split it up we just need to apply the ops $c1, $c2 etc
1233 to the very same args as the dfun. It takes a little more work
1234 to compute the type arguments to the dictionary constructor.