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,
23 mkInlineRule, mkWwInlineRule,
24 mkCompulsoryUnfolding,
26 couldBeSmallEnoughToInline,
27 certainlyWillInline, smallEnoughToInline,
29 callSiteInline, CallCtxt(..)
36 import PprCore () -- Instances
38 import CoreSubst ( emptySubst, substTy, extendIdSubst, extendTvSubst
39 , lookupIdSubst, substBndr, substBndrs, substRecBndrs )
46 import BasicTypes ( Arity )
47 import Type hiding( substTy, extendTvSubst )
58 %************************************************************************
60 \subsection{Making unfoldings}
62 %************************************************************************
65 mkTopUnfolding :: CoreExpr -> Unfolding
66 mkTopUnfolding expr = mkUnfolding True {- Top level -} expr
68 mkImplicitUnfolding :: CoreExpr -> Unfolding
69 -- For implicit Ids, do a tiny bit of optimising first
70 mkImplicitUnfolding expr
71 = CoreUnfolding (simpleOptExpr expr)
75 (calcUnfoldingGuidance opt_UF_CreationThreshold expr)
77 mkInlineRule :: CoreExpr -> Arity -> Unfolding
78 mkInlineRule expr arity
79 = InlineRule { uf_tmpl = simpleOptExpr expr,
80 uf_is_top = True, -- Conservative; this gets set more
81 -- accuately by the simplifier (slight hack)
82 -- in SimplEnv.substUnfolding
84 uf_is_value = exprIsHNF expr,
87 mkWwInlineRule :: CoreExpr -> Arity -> Id -> Unfolding
88 mkWwInlineRule expr arity wkr
89 = InlineRule { uf_tmpl = simpleOptExpr expr,
90 uf_is_top = True, -- Conservative; see mkInlineRule
92 uf_is_value = exprIsHNF expr,
93 uf_worker = Just wkr }
95 mkUnfolding :: Bool -> CoreExpr -> Unfolding
96 mkUnfolding top_lvl expr
97 = CoreUnfolding { uf_tmpl = occurAnalyseExpr expr,
99 uf_is_value = exprIsHNF expr,
100 uf_is_cheap = exprIsCheap expr,
101 uf_guidance = calcUnfoldingGuidance opt_UF_CreationThreshold expr }
102 -- Sometimes during simplification, there's a large let-bound thing
103 -- which has been substituted, and so is now dead; so 'expr' contains
104 -- two copies of the thing while the occurrence-analysed expression doesn't
105 -- Nevertheless, we don't occ-analyse before computing the size because the
106 -- size computation bales out after a while, whereas occurrence analysis does not.
108 -- This can occasionally mean that the guidance is very pessimistic;
109 -- it gets fixed up next round
111 mkCompulsoryUnfolding :: CoreExpr -> Unfolding
112 mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
113 = CompulsoryUnfolding (occurAnalyseExpr expr)
117 %************************************************************************
119 \subsection{The UnfoldingGuidance type}
121 %************************************************************************
124 calcUnfoldingGuidance
125 :: Int -- bomb out if size gets bigger than this
126 -> CoreExpr -- expression to look at
128 calcUnfoldingGuidance bOMB_OUT_SIZE expr
129 = case collectBinders expr of { (binders, body) ->
131 val_binders = filter isId binders
132 n_val_binders = length val_binders
134 case (sizeExpr (iUnbox bOMB_OUT_SIZE) val_binders body) of
135 TooBig -> UnfoldNever
136 SizeIs size cased_args scrut_discount
137 -> UnfoldIfGoodArgs { ug_arity = n_val_binders
138 , ug_args = map discount_for val_binders
139 , ug_size = iBox size
140 , ug_res = iBox scrut_discount }
142 discount_for b = foldlBag (\acc (b',n) -> if b==b' then acc+n else acc)
148 sizeExpr :: FastInt -- Bomb out if it gets bigger than this
149 -> [Id] -- Arguments; we're interested in which of these
154 sizeExpr bOMB_OUT_SIZE top_args expr
157 size_up (Type _) = sizeZero -- Types cost nothing
158 size_up (Var _) = sizeOne
159 size_up (Note _ body) = size_up body -- Notes cost nothing
160 size_up (Cast e _) = size_up e
161 size_up (App fun (Type _)) = size_up fun
162 size_up (App fun arg) = size_up_app fun [arg]
164 size_up (Lit lit) = sizeN (litSize lit)
166 size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
167 | otherwise = size_up e
169 size_up (Let (NonRec binder rhs) body)
170 = nukeScrutDiscount (size_up rhs) `addSize`
171 size_up body `addSizeN`
172 (if isUnLiftedType (idType binder) then 0 else 1)
173 -- For the allocation
174 -- If the binder has an unlifted type there is no allocation
176 size_up (Let (Rec pairs) body)
177 = nukeScrutDiscount rhs_size `addSize`
178 size_up body `addSizeN`
179 length pairs -- For the allocation
181 rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
183 size_up (Case (Var v) _ _ alts)
184 | v `elem` top_args -- We are scrutinising an argument variable
186 {- I'm nuking this special case; BUT see the comment with case alternatives.
188 (a) It's too eager. We don't want to inline a wrapper into a
189 context with no benefit.
190 E.g. \ x. f (x+x) no point in inlining (+) here!
192 (b) It's ineffective. Once g's wrapper is inlined, its case-expressions
193 aren't scrutinising arguments any more
197 [alt] -> size_up_alt alt `addSize` SizeIs (_ILIT(0)) (unitBag (v, 1)) (_ILIT(0))
198 -- We want to make wrapper-style evaluation look cheap, so that
199 -- when we inline a wrapper it doesn't make call site (much) bigger
200 -- Otherwise we get nasty phase ordering stuff:
203 -- If we inline g's wrapper, f looks big, and doesn't get inlined
204 -- into h; if we inline f first, while it looks small, then g's
205 -- wrapper will get inlined later anyway. To avoid this nasty
206 -- ordering difference, we make (case a of (x,y) -> ...),
207 -- *where a is one of the arguments* look free.
211 alts_size (foldr addSize sizeOne alt_sizes) -- The 1 is for the scrutinee
212 (foldr1 maxSize alt_sizes)
214 -- Good to inline if an arg is scrutinised, because
215 -- that may eliminate allocation in the caller
216 -- And it eliminates the case itself
219 alt_sizes = map size_up_alt alts
221 -- alts_size tries to compute a good discount for
222 -- the case when we are scrutinising an argument variable
223 alts_size (SizeIs tot _tot_disc _tot_scrut) -- Size of all alternatives
224 (SizeIs max max_disc max_scrut) -- Size of biggest alternative
225 = SizeIs tot (unitBag (v, iBox (_ILIT(1) +# tot -# max)) `unionBags` max_disc) max_scrut
226 -- If the variable is known, we produce a discount that
227 -- will take us back to 'max', the size of rh largest alternative
228 -- The 1+ is a little discount for reduced allocation in the caller
229 alts_size tot_size _ = tot_size
231 size_up (Case e _ _ alts) = nukeScrutDiscount (size_up e) `addSize`
232 foldr (addSize . size_up_alt) sizeZero alts
233 -- We don't charge for the case itself
234 -- It's a strict thing, and the price of the call
235 -- is paid by scrut. Also consider
236 -- case f x of DEFAULT -> e
237 -- This is just ';'! Don't charge for it.
240 size_up_app (App fun arg) args
241 | isTypeArg arg = size_up_app fun args
242 | otherwise = size_up_app fun (arg:args)
243 size_up_app fun args = foldr (addSize . nukeScrutDiscount . size_up)
244 (size_up_fun fun args)
247 -- A function application with at least one value argument
248 -- so if the function is an argument give it an arg-discount
250 -- Also behave specially if the function is a build
252 -- Also if the function is a constant Id (constr or primop)
253 -- compute discounts specially
254 size_up_fun (Var fun) args
255 | fun `hasKey` buildIdKey = buildSize
256 | fun `hasKey` augmentIdKey = augmentSize
258 = case globalIdDetails fun of
259 DataConWorkId dc -> conSizeN dc (valArgCount args)
261 FCallId _ -> sizeN opt_UF_DearOp
262 PrimOpId op -> primOpSize op (valArgCount args)
263 -- foldr addSize (primOpSize op) (map arg_discount args)
264 -- At one time I tried giving an arg-discount if a primop
265 -- is applied to one of the function's arguments, but it's
266 -- not good. At the moment, any unlifted-type arg gets a
267 -- 'True' for 'yes I'm evald', so we collect the discount even
268 -- if we know nothing about it. And just having it in a primop
269 -- doesn't help at all if we don't know something more.
271 _ -> fun_discount fun `addSizeN`
272 (1 + length (filter (not . exprIsTrivial) args))
273 -- The 1+ is for the function itself
274 -- Add 1 for each non-trivial arg;
275 -- the allocation cost, as in let(rec)
276 -- Slight hack here: for constructors the args are almost always
277 -- trivial; and for primops they are almost always prim typed
278 -- We should really only count for non-prim-typed args in the
279 -- general case, but that seems too much like hard work
281 size_up_fun other _ = size_up other
284 size_up_alt (_con, _bndrs, rhs) = size_up rhs
285 -- Don't charge for args, so that wrappers look cheap
286 -- (See comments about wrappers with Case)
289 -- We want to record if we're case'ing, or applying, an argument
290 fun_discount v | v `elem` top_args = SizeIs (_ILIT(0)) (unitBag (v, opt_UF_FunAppDiscount)) (_ILIT(0))
291 fun_discount _ = sizeZero
294 -- These addSize things have to be here because
295 -- I don't want to give them bOMB_OUT_SIZE as an argument
297 addSizeN TooBig _ = TooBig
298 addSizeN (SizeIs n xs d) m = mkSizeIs bOMB_OUT_SIZE (n +# iUnbox m) xs d
300 addSize TooBig _ = TooBig
301 addSize _ TooBig = TooBig
302 addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
303 = mkSizeIs bOMB_OUT_SIZE (n1 +# n2) (xs `unionBags` ys) (d1 +# d2)
306 Code for manipulating sizes
309 data ExprSize = TooBig
310 | SizeIs FastInt -- Size found
311 (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
312 FastInt -- Size to subtract if result is scrutinised
313 -- by a case expression
315 -- subtract the discount before deciding whether to bale out. eg. we
316 -- want to inline a large constructor application into a selector:
317 -- tup = (a_1, ..., a_99)
318 -- x = case tup of ...
320 mkSizeIs :: FastInt -> FastInt -> Bag (Id, Int) -> FastInt -> ExprSize
321 mkSizeIs max n xs d | (n -# d) ># max = TooBig
322 | otherwise = SizeIs n xs d
324 maxSize :: ExprSize -> ExprSize -> ExprSize
325 maxSize TooBig _ = TooBig
326 maxSize _ TooBig = TooBig
327 maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
330 sizeZero, sizeOne :: ExprSize
331 sizeN :: Int -> ExprSize
332 conSizeN :: DataCon ->Int -> ExprSize
334 sizeZero = SizeIs (_ILIT(0)) emptyBag (_ILIT(0))
335 sizeOne = SizeIs (_ILIT(1)) emptyBag (_ILIT(0))
336 sizeN n = SizeIs (iUnbox n) emptyBag (_ILIT(0))
338 | isUnboxedTupleCon dc = SizeIs (_ILIT(0)) emptyBag (iUnbox n +# _ILIT(1))
339 | otherwise = SizeIs (_ILIT(1)) emptyBag (iUnbox n +# _ILIT(1))
340 -- Treat constructors as size 1; we are keen to expose them
341 -- (and we charge separately for their args). We can't treat
342 -- them as size zero, else we find that (iBox x) has size 1,
343 -- which is the same as a lone variable; and hence 'v' will
344 -- always be replaced by (iBox x), where v is bound to iBox x.
346 -- However, unboxed tuples count as size zero
347 -- I found occasions where we had
348 -- f x y z = case op# x y z of { s -> (# s, () #) }
349 -- and f wasn't getting inlined
351 primOpSize :: PrimOp -> Int -> ExprSize
353 | not (primOpIsDupable op) = sizeN opt_UF_DearOp
354 | not (primOpOutOfLine op) = sizeN (2 - n_args)
355 -- Be very keen to inline simple primops.
356 -- We give a discount of 1 for each arg so that (op# x y z) costs 2.
357 -- We can't make it cost 1, else we'll inline let v = (op# x y z)
358 -- at every use of v, which is excessive.
360 -- A good example is:
361 -- let x = +# p q in C {x}
362 -- Even though x get's an occurrence of 'many', its RHS looks cheap,
363 -- and there's a good chance it'll get inlined back into C's RHS. Urgh!
364 | otherwise = sizeOne
366 buildSize :: ExprSize
367 buildSize = SizeIs (_ILIT(-2)) emptyBag (_ILIT(4))
368 -- We really want to inline applications of build
369 -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
370 -- Indeed, we should add a result_discount becuause build is
371 -- very like a constructor. We don't bother to check that the
372 -- build is saturated (it usually is). The "-2" discounts for the \c n,
373 -- The "4" is rather arbitrary.
375 augmentSize :: ExprSize
376 augmentSize = SizeIs (_ILIT(-2)) emptyBag (_ILIT(4))
377 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
378 -- e plus ys. The -2 accounts for the \cn
380 nukeScrutDiscount :: ExprSize -> ExprSize
381 nukeScrutDiscount (SizeIs n vs _) = SizeIs n vs (_ILIT(0))
382 nukeScrutDiscount TooBig = TooBig
384 -- When we return a lambda, give a discount if it's used (applied)
385 lamScrutDiscount :: ExprSize -> ExprSize
386 lamScrutDiscount (SizeIs n vs _) = case opt_UF_FunAppDiscount of { d -> SizeIs n vs (iUnbox d) }
387 lamScrutDiscount TooBig = TooBig
391 %************************************************************************
393 \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
395 %************************************************************************
397 We have very limited information about an unfolding expression: (1)~so
398 many type arguments and so many value arguments expected---for our
399 purposes here, we assume we've got those. (2)~A ``size'' or ``cost,''
400 a single integer. (3)~An ``argument info'' vector. For this, what we
401 have at the moment is a Boolean per argument position that says, ``I
402 will look with great favour on an explicit constructor in this
403 position.'' (4)~The ``discount'' to subtract if the expression
404 is being scrutinised.
406 Assuming we have enough type- and value arguments (if not, we give up
407 immediately), then we see if the ``discounted size'' is below some
408 (semi-arbitrary) threshold. It works like this: for every argument
409 position where we're looking for a constructor AND WE HAVE ONE in our
410 hands, we get a (again, semi-arbitrary) discount [proportion to the
411 number of constructors in the type being scrutinized].
413 If we're in the context of a scrutinee ( \tr{(case <expr > of A .. -> ...;.. )})
414 and the expression in question will evaluate to a constructor, we use
415 the computed discount size *for the result only* rather than
416 computing the argument discounts. Since we know the result of
417 the expression is going to be taken apart, discounting its size
418 is more accurate (see @sizeExpr@ above for how this discount size
421 We use this one to avoid exporting inlinings that we ``couldn't possibly
422 use'' on the other side. Can be overridden w/ flaggery.
423 Just the same as smallEnoughToInline, except that it has no actual arguments.
426 couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
427 couldBeSmallEnoughToInline threshold rhs = case calcUnfoldingGuidance threshold rhs of
431 certainlyWillInline :: Unfolding -> Bool
432 -- Sees if the unfolding is pretty certain to inline
433 certainlyWillInline (CompulsoryUnfolding {}) = True
434 certainlyWillInline (InlineRule {}) = True
435 certainlyWillInline (CoreUnfolding
436 { uf_is_cheap = is_cheap
437 , uf_guidance = UnfoldIfGoodArgs {ug_arity = n_vals, ug_size = size}})
438 = is_cheap && size - (n_vals +1) <= opt_UF_UseThreshold
439 certainlyWillInline _
442 smallEnoughToInline :: Unfolding -> Bool
443 smallEnoughToInline (CoreUnfolding {uf_guidance = UnfoldIfGoodArgs {ug_size = size}})
444 = size <= opt_UF_UseThreshold
445 smallEnoughToInline _
449 %************************************************************************
451 \subsection{callSiteInline}
453 %************************************************************************
455 This is the key function. It decides whether to inline a variable at a call site
457 callSiteInline is used at call sites, so it is a bit more generous.
458 It's a very important function that embodies lots of heuristics.
459 A non-WHNF can be inlined if it doesn't occur inside a lambda,
460 and occurs exactly once or
461 occurs once in each branch of a case and is small
463 If the thing is in WHNF, there's no danger of duplicating work,
464 so we can inline if it occurs once, or is small
466 NOTE: we don't want to inline top-level functions that always diverge.
467 It just makes the code bigger. Tt turns out that the convenient way to prevent
468 them inlining is to give them a NOINLINE pragma, which we do in
469 StrictAnal.addStrictnessInfoToTopId
472 callSiteInline :: DynFlags
473 -> Bool -- True <=> the Id can be inlined
475 -> Bool -- True if there are are no arguments at all (incl type args)
476 -> [Bool] -- One for each value arg; True if it is interesting
477 -> CallCtxt -- True <=> continuation is interesting
478 -> Maybe CoreExpr -- Unfolding, if any
481 data CallCtxt = BoringCtxt
483 | ArgCtxt Bool -- We're somewhere in the RHS of function with rules
484 -- => be keener to inline
485 Int -- We *are* the argument of a function with this arg discount
486 -- => be keener to inline
487 -- INVARIANT: ArgCtxt False 0 ==> BoringCtxt
489 | ValAppCtxt -- We're applied to at least one value arg
490 -- This arises when we have ((f x |> co) y)
491 -- Then the (f x) has argument 'x' but in a ValAppCtxt
493 | CaseCtxt -- We're the scrutinee of a case
494 -- that decomposes its scrutinee
496 instance Outputable CallCtxt where
497 ppr BoringCtxt = ptext (sLit "BoringCtxt")
498 ppr (ArgCtxt _ _) = ptext (sLit "ArgCtxt")
499 ppr CaseCtxt = ptext (sLit "CaseCtxt")
500 ppr ValAppCtxt = ptext (sLit "ValAppCtxt")
502 callSiteInline dflags active_inline id lone_variable arg_infos cont_info
504 n_val_args = length arg_infos
506 case idUnfolding id of {
507 NoUnfolding -> Nothing ;
508 OtherCon _ -> Nothing ;
510 CompulsoryUnfolding unf_template -> Just unf_template ;
511 -- CompulsoryUnfolding => there is no top-level binding
512 -- for these things, so we must inline it.
513 -- Only a couple of primop-like things have
514 -- compulsory unfoldings (see MkId.lhs).
515 -- We don't allow them to be inactive
517 InlineRule { uf_tmpl = unf_template, uf_arity = arity, uf_is_top = is_top
518 , uf_is_value = is_value, uf_worker = mb_worker }
519 -> let yes_or_no | not active_inline = False
520 | n_val_args < arity = yes_unsat -- Not enough value args
521 | n_val_args == arity = yes_exact -- Exactly saturated
522 | otherwise = True -- Over-saturated
523 result | yes_or_no = Just unf_template
524 | otherwise = Nothing
526 -- See Note [Inlining an InlineRule]
527 is_wrapper = isJust mb_worker
528 yes_unsat | is_wrapper = or arg_infos
531 yes_exact = or arg_infos || interesting_saturated_call
532 interesting_saturated_call
534 BoringCtxt -> not is_top -- Note [Nested functions]
535 CaseCtxt -> not lone_variable || not is_value -- Note [Lone variables]
536 ArgCtxt {} -> arity > 0 -- Note [Inlining in ArgCtxt]
537 ValAppCtxt -> True -- Note [Cast then apply]
539 if dopt Opt_D_dump_inlinings dflags then
540 pprTrace ("Considering InlineRule for: " ++ showSDoc (ppr id))
541 (vcat [text "active:" <+> ppr active_inline,
542 text "arg infos" <+> ppr arg_infos,
543 text "interesting call" <+> ppr interesting_saturated_call,
544 text "is value:" <+> ppr is_value,
545 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO"])
549 CoreUnfolding { uf_tmpl = unf_template, uf_is_top = is_top, uf_is_value = is_value,
550 uf_is_cheap = is_cheap, uf_guidance = guidance } ->
553 result | yes_or_no = Just unf_template
554 | otherwise = Nothing
556 yes_or_no = active_inline && is_cheap && consider_safe
557 -- We consider even the once-in-one-branch
558 -- occurrences, because they won't all have been
559 -- caught by preInlineUnconditionally. In particular,
560 -- if the occurrence is once inside a lambda, and the
561 -- rhs is cheap but not a manifest lambda, then
562 -- pre-inline will not have inlined it for fear of
563 -- invalidating the occurrence info in the rhs.
566 -- consider_safe decides whether it's a good idea to
567 -- inline something, given that there's no
568 -- work-duplication issue (the caller checks that).
571 UnfoldIfGoodArgs { ug_arity = n_vals_wanted, ug_args = arg_discounts
572 , ug_res = res_discount, ug_size = size }
573 | enough_args && size <= (n_vals_wanted + 1)
574 -- Inline unconditionally if there no size increase
575 -- Size of call is n_vals_wanted (+1 for the function)
579 -> some_benefit && small_enough && inline_enough_args
582 enough_args = n_val_args >= n_vals_wanted
584 not (dopt Opt_InlineIfEnoughArgs dflags) || enough_args
587 some_benefit = or arg_infos || really_interesting_cont
588 -- There must be something interesting
589 -- about some argument, or the result
590 -- context, to make it worth inlining
592 really_interesting_cont
593 | n_val_args < n_vals_wanted = False -- Too few args
594 | n_val_args == n_vals_wanted = interesting_saturated_call
595 | otherwise = True -- Extra args
596 -- really_interesting_cont tells if the result of the
597 -- call is in an interesting context.
599 interesting_saturated_call
601 BoringCtxt -> not is_top && n_vals_wanted > 0 -- Note [Nested functions]
602 CaseCtxt -> not lone_variable || not is_value -- Note [Lone variables]
603 ArgCtxt {} -> n_vals_wanted > 0 -- Note [Inlining in ArgCtxt]
604 ValAppCtxt -> True -- Note [Cast then apply]
606 small_enough = (size - discount) <= opt_UF_UseThreshold
607 discount = computeDiscount n_vals_wanted arg_discounts
608 res_discount' arg_infos
609 res_discount' = case cont_info of
611 CaseCtxt -> res_discount
612 _other -> 4 `min` res_discount
613 -- res_discount can be very large when a function returns
614 -- construtors; but we only want to invoke that large discount
615 -- when there's a case continuation.
616 -- Otherwise we, rather arbitrarily, threshold it. Yuk.
617 -- But we want to aovid inlining large functions that return
618 -- constructors into contexts that are simply "interesting"
621 if dopt Opt_D_dump_inlinings dflags then
622 pprTrace ("Considering inlining: " ++ showSDoc (ppr id))
623 (vcat [text "active:" <+> ppr active_inline,
624 text "arg infos" <+> ppr arg_infos,
625 text "interesting continuation" <+> ppr cont_info,
626 text "is value:" <+> ppr is_value,
627 text "is cheap:" <+> ppr is_cheap,
628 text "guidance" <+> ppr guidance,
629 text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO"])
636 Note [Inlining an InlineRule]
637 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
638 An InlineRules is used for
639 (a) pogrammer INLINE pragmas
640 (b) inlinings from worker/wrapper
642 For (a) the RHS may be large, and our contract is that we *only* inline
643 when the function is applied to all the arguments on the LHS of the
644 source-code defn. (The uf_arity in the rule.)
646 However for worker/wrapper it may be worth inlining even if the
647 arity is not satisfied (as we do in the CoreUnfolding case) so we don't
651 Note [Nested functions]
652 ~~~~~~~~~~~~~~~~~~~~~~~
653 If a function has a nested defn we also record some-benefit, on the
654 grounds that we are often able to eliminate the binding, and hence the
655 allocation, for the function altogether; this is good for join points.
656 But this only makes sense for *functions*; inlining a constructor
657 doesn't help allocation unless the result is scrutinised. UNLESS the
658 constructor occurs just once, albeit possibly in multiple case
659 branches. Then inlining it doesn't increase allocation, but it does
660 increase the chance that the constructor won't be allocated at all in
661 the branches that don't use it.
663 Note [Cast then apply]
664 ~~~~~~~~~~~~~~~~~~~~~~
666 myIndex = __inline_me ( (/\a. <blah>) |> co )
667 co :: (forall a. a -> a) ~ (forall a. T a)
668 ... /\a.\x. case ((myIndex a) |> sym co) x of { ... } ...
670 We need to inline myIndex to unravel this; but the actual call (myIndex a) has
671 no value arguments. The ValAppCtxt gives it enough incentive to inline.
673 Note [Inlining in ArgCtxt]
674 ~~~~~~~~~~~~~~~~~~~~~~~~~~
675 The condition (n_vals_wanted > 0) here is very important, because otherwise
676 we end up inlining top-level stuff into useless places; eg
679 This can make a very big difference: it adds 16% to nofib 'integer' allocs,
682 At one stage I replaced this condition by 'True' (leading to the above
683 slow-down). The motivation was test eyeball/inline1.hs; but that seems
686 Note [Lone variables]
687 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
688 The "lone-variable" case is important. I spent ages messing about
689 with unsatisfactory varaints, but this is nice. The idea is that if a
690 variable appears all alone
691 as an arg of lazy fn, or rhs Stop
692 as scrutinee of a case Select
693 as arg of a strict fn ArgOf
695 it is bound to a value
696 then we should not inline it (unless there is some other reason,
697 e.g. is is the sole occurrence). That is what is happening at
698 the use of 'lone_variable' in 'interesting_saturated_call'.
700 Why? At least in the case-scrutinee situation, turning
701 let x = (a,b) in case x of y -> ...
703 let x = (a,b) in case (a,b) of y -> ...
705 let x = (a,b) in let y = (a,b) in ...
706 is bad if the binding for x will remain.
708 Another example: I discovered that strings
709 were getting inlined straight back into applications of 'error'
710 because the latter is strict.
712 f = \x -> ...(error s)...
714 Fundamentally such contexts should not encourage inlining because the
715 context can ``see'' the unfolding of the variable (e.g. case or a
716 RULE) so there's no gain. If the thing is bound to a value.
721 foo = _inline_ (\n. [n])
722 bar = _inline_ (foo 20)
723 baz = \n. case bar of { (m:_) -> m + n }
724 Here we really want to inline 'bar' so that we can inline 'foo'
725 and the whole thing unravels as it should obviously do. This is
726 important: in the NDP project, 'bar' generates a closure data
727 structure rather than a list.
729 * Even a type application or coercion isn't a lone variable.
731 case $fMonadST @ RealWorld of { :DMonad a b c -> c }
732 We had better inline that sucker! The case won't see through it.
734 For now, I'm treating treating a variable applied to types
735 in a *lazy* context "lone". The motivating example was
738 There's no advantage in inlining f here, and perhaps
739 a significant disadvantage. Hence some_val_args in the Stop case
742 computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Int
743 computeDiscount n_vals_wanted arg_discounts result_discount arg_infos
744 -- We multiple the raw discounts (args_discount and result_discount)
745 -- ty opt_UnfoldingKeenessFactor because the former have to do with
746 -- *size* whereas the discounts imply that there's some extra
747 -- *efficiency* to be gained (e.g. beta reductions, case reductions)
750 -- we also discount 1 for each argument passed, because these will
751 -- reduce with the lambdas in the function (we count 1 for a lambda
753 = 1 + -- Discount of 1 because the result replaces the call
754 -- so we count 1 for the function itself
755 length (take n_vals_wanted arg_infos) +
756 -- Discount of 1 for each arg supplied, because the
757 -- result replaces the call
758 round (opt_UF_KeenessFactor *
759 fromIntegral (arg_discount + result_discount))
761 arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
763 mk_arg_discount discount is_evald | is_evald = discount
767 %************************************************************************
769 The Very Simple Optimiser
771 %************************************************************************
775 simpleOptExpr :: CoreExpr -> CoreExpr
776 -- Return an occur-analysed and slightly optimised expression
777 -- The optimisation is very straightforward: just
778 -- inline non-recursive bindings that are used only once,
779 -- or wheere the RHS is trivial
782 = go emptySubst (occurAnalyseExpr expr)
784 go subst (Var v) = lookupIdSubst subst v
785 go subst (App e1 e2) = App (go subst e1) (go subst e2)
786 go subst (Type ty) = Type (substTy subst ty)
787 go _ (Lit lit) = Lit lit
788 go subst (Note note e) = Note note (go subst e)
789 go subst (Cast e co) = Cast (go subst e) (substTy subst co)
790 go subst (Let bind body) = go_bind subst bind body
791 go subst (Lam bndr body) = Lam bndr' (go subst' body)
793 (subst', bndr') = substBndr subst bndr
795 go subst (Case e b ty as) = Case (go subst e) b'
797 (map (go_alt subst') as)
799 (subst', b') = substBndr subst b
802 ----------------------
803 go_alt subst (con, bndrs, rhs) = (con, bndrs', go subst' rhs)
805 (subst', bndrs') = substBndrs subst bndrs
807 ----------------------
808 go_bind subst (Rec prs) body = Let (Rec (bndrs' `zip` rhss'))
811 (bndrs, rhss) = unzip prs
812 (subst', bndrs') = substRecBndrs subst bndrs
813 rhss' = map (go subst') rhss
815 go_bind subst (NonRec b r) body = go_nonrec subst b (go subst r) body
817 ----------------------
818 go_nonrec subst b (Type ty') body
819 | isTyVar b = go (extendTvSubst subst b ty') body
820 -- let a::* = TYPE ty in <body>
821 go_nonrec subst b r' body
822 | isId b -- let x = e in <body>
823 , exprIsTrivial r' || safe_to_inline (idOccInfo b)
824 = go (extendIdSubst subst b r') body
825 go_nonrec subst b r' body
826 = Let (NonRec b' r') (go subst' body)
828 (subst', b') = substBndr subst b
830 ----------------------
831 -- Unconditionally safe to inline
832 safe_to_inline :: OccInfo -> Bool
833 safe_to_inline IAmDead = True
834 safe_to_inline (OneOcc in_lam one_br _) = not in_lam && one_br
835 safe_to_inline (IAmALoopBreaker {}) = False
836 safe_to_inline NoOccInfo = False