2 % (c) The AQUA Project, Glasgow University, 1993-1998
4 \section[SimplUtils]{The simplifier utilities}
8 mkLam, prepareAlts, mkCase,
11 preInlineUnconditionally, postInlineUnconditionally, activeInline, activeRule,
14 -- The continuation type
15 SimplCont(..), DupFlag(..), LetRhsFlag(..),
16 contIsDupable, contResultType,
17 countValArgs, countArgs, pushContArgs,
18 mkBoringStop, mkRhsStop, contIsRhs, contIsRhsOrArg,
19 getContArgs, interestingCallContext, interestingArg, isStrictType
23 #include "HsVersions.h"
26 import DynFlags ( SimplifierSwitch(..), SimplifierMode(..),
28 import StaticFlags ( opt_UF_UpdateInPlace, opt_SimplNoPreInlining,
32 import CoreFVs ( exprFreeVars )
33 import CoreUtils ( cheapEqExpr, exprType, exprIsTrivial, exprIsCheap,
34 etaExpand, exprEtaExpandArity, bindNonRec, mkCoerce2,
35 findDefault, exprOkForSpeculation, exprIsHNF
37 import CoreUnfold ( smallEnoughToInline )
38 import Id ( idType, isDataConWorkId, idOccInfo, isDictId, idArity,
39 mkSysLocal, isDeadBinder, idNewDemandInfo, isExportedId,
40 idUnfolding, idNewStrictness, idInlinePragma,
42 import NewDemand ( isStrictDmd, isBotRes, splitStrictSig )
44 import Type ( Type, splitFunTys, dropForAlls, isStrictType,
45 splitTyConApp_maybe, tyConAppArgs, mkTyVarTys
47 import Name ( mkSysTvName )
48 import TyCon ( tyConDataCons_maybe, isAlgTyCon, isNewTyCon )
49 import DataCon ( dataConRepArity, dataConTyVars, dataConArgTys, isVanillaDataCon )
50 import Var ( tyVarKind, mkTyVar )
52 import BasicTypes ( TopLevelFlag(..), isTopLevel, isNotTopLevel, OccInfo(..), isLoopBreaker, isOneOcc,
53 Activation, isAlwaysActive, isActive )
54 import Util ( lengthExceeds )
59 %************************************************************************
61 \subsection{The continuation data type}
63 %************************************************************************
66 data SimplCont -- Strict contexts
67 = Stop OutType -- Type of the result
69 Bool -- True <=> This is the RHS of a thunk whose type suggests
70 -- that update-in-place would be possible
71 -- (This makes the inliner a little keener.)
73 | CoerceIt OutType -- The To-type, simplified
76 | InlinePlease -- This continuation makes a function very
77 SimplCont -- keen to inline itelf
80 InExpr SimplEnv -- The argument, as yet unsimplified,
81 SimplCont -- and its environment
84 InId [InAlt] SimplEnv -- The case binder, alts, and subst-env
87 | ArgOf LetRhsFlag -- An arbitrary strict context: the argument
88 -- of a strict function, or a primitive-arg fn
90 -- No DupFlag because we never duplicate it
91 OutType -- arg_ty: type of the argument itself
92 OutType -- cont_ty: the type of the expression being sought by the context
93 -- f (error "foo") ==> coerce t (error "foo")
95 -- We need to know the type t, to which to coerce.
97 (SimplEnv -> OutExpr -> SimplM FloatsWithExpr) -- What to do with the result
98 -- The result expression in the OutExprStuff has type cont_ty
100 data LetRhsFlag = AnArg -- It's just an argument not a let RHS
101 | AnRhs -- It's the RHS of a let (so please float lets out of big lambdas)
103 instance Outputable LetRhsFlag where
104 ppr AnArg = ptext SLIT("arg")
105 ppr AnRhs = ptext SLIT("rhs")
107 instance Outputable SimplCont where
108 ppr (Stop _ is_rhs _) = ptext SLIT("Stop") <> brackets (ppr is_rhs)
109 ppr (ApplyTo dup arg se cont) = (ptext SLIT("ApplyTo") <+> ppr dup <+> ppr arg) $$ ppr cont
110 ppr (ArgOf _ _ _ _) = ptext SLIT("ArgOf...")
111 ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$
112 (nest 4 (ppr alts)) $$ ppr cont
113 ppr (CoerceIt ty cont) = (ptext SLIT("CoerceIt") <+> ppr ty) $$ ppr cont
114 ppr (InlinePlease cont) = ptext SLIT("InlinePlease") $$ ppr cont
116 data DupFlag = OkToDup | NoDup
118 instance Outputable DupFlag where
119 ppr OkToDup = ptext SLIT("ok")
120 ppr NoDup = ptext SLIT("nodup")
124 mkBoringStop, mkRhsStop :: OutType -> SimplCont
125 mkBoringStop ty = Stop ty AnArg (canUpdateInPlace ty)
126 mkRhsStop ty = Stop ty AnRhs (canUpdateInPlace ty)
128 contIsRhs :: SimplCont -> Bool
129 contIsRhs (Stop _ AnRhs _) = True
130 contIsRhs (ArgOf AnRhs _ _ _) = True
131 contIsRhs other = False
133 contIsRhsOrArg (Stop _ _ _) = True
134 contIsRhsOrArg (ArgOf _ _ _ _) = True
135 contIsRhsOrArg other = False
138 contIsDupable :: SimplCont -> Bool
139 contIsDupable (Stop _ _ _) = True
140 contIsDupable (ApplyTo OkToDup _ _ _) = True
141 contIsDupable (Select OkToDup _ _ _ _) = True
142 contIsDupable (CoerceIt _ cont) = contIsDupable cont
143 contIsDupable (InlinePlease cont) = contIsDupable cont
144 contIsDupable other = False
147 discardableCont :: SimplCont -> Bool
148 discardableCont (Stop _ _ _) = False
149 discardableCont (CoerceIt _ cont) = discardableCont cont
150 discardableCont (InlinePlease cont) = discardableCont cont
151 discardableCont other = True
153 discardCont :: SimplCont -- A continuation, expecting
154 -> SimplCont -- Replace the continuation with a suitable coerce
155 discardCont cont = case cont of
156 Stop to_ty is_rhs _ -> cont
157 other -> CoerceIt to_ty (mkBoringStop to_ty)
159 to_ty = contResultType cont
162 contResultType :: SimplCont -> OutType
163 contResultType (Stop to_ty _ _) = to_ty
164 contResultType (ArgOf _ _ to_ty _) = to_ty
165 contResultType (ApplyTo _ _ _ cont) = contResultType cont
166 contResultType (CoerceIt _ cont) = contResultType cont
167 contResultType (InlinePlease cont) = contResultType cont
168 contResultType (Select _ _ _ _ cont) = contResultType cont
171 countValArgs :: SimplCont -> Int
172 countValArgs (ApplyTo _ (Type ty) se cont) = countValArgs cont
173 countValArgs (ApplyTo _ val_arg se cont) = 1 + countValArgs cont
174 countValArgs other = 0
176 countArgs :: SimplCont -> Int
177 countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont
181 pushContArgs :: SimplEnv -> [OutArg] -> SimplCont -> SimplCont
182 -- Pushes args with the specified environment
183 pushContArgs env [] cont = cont
184 pushContArgs env (arg : args) cont = ApplyTo NoDup arg env (pushContArgs env args cont)
189 getContArgs :: SwitchChecker
190 -> OutId -> SimplCont
191 -> ([(InExpr, SimplEnv, Bool)], -- Arguments; the Bool is true for strict args
192 SimplCont, -- Remaining continuation
193 Bool) -- Whether we came across an InlineCall
194 -- getContArgs id k = (args, k', inl)
195 -- args are the leading ApplyTo items in k
196 -- (i.e. outermost comes first)
197 -- augmented with demand info from the functionn
198 getContArgs chkr fun orig_cont
200 -- Ignore strictness info if the no-case-of-case
201 -- flag is on. Strictness changes evaluation order
202 -- and that can change full laziness
203 stricts | switchIsOn chkr NoCaseOfCase = vanilla_stricts
204 | otherwise = computed_stricts
206 go [] stricts False orig_cont
208 ----------------------------
211 go acc ss inl (ApplyTo _ arg@(Type _) se cont)
212 = go ((arg,se,False) : acc) ss inl cont
213 -- NB: don't bother to instantiate the function type
216 go acc (s:ss) inl (ApplyTo _ arg se cont)
217 = go ((arg,se,s) : acc) ss inl cont
219 -- An Inline continuation
220 go acc ss inl (InlinePlease cont)
221 = go acc ss True cont
223 -- We're run out of arguments, or else we've run out of demands
224 -- The latter only happens if the result is guaranteed bottom
225 -- This is the case for
226 -- * case (error "hello") of { ... }
227 -- * (error "Hello") arg
228 -- * f (error "Hello") where f is strict
230 -- Then, especially in the first of these cases, we'd like to discard
231 -- the continuation, leaving just the bottoming expression. But the
232 -- type might not be right, so we may have to add a coerce.
234 | null ss && discardableCont cont = (reverse acc, discardCont cont, inl)
235 | otherwise = (reverse acc, cont, inl)
237 ----------------------------
238 vanilla_stricts, computed_stricts :: [Bool]
239 vanilla_stricts = repeat False
240 computed_stricts = zipWith (||) fun_stricts arg_stricts
242 ----------------------------
243 (val_arg_tys, _) = splitFunTys (dropForAlls (idType fun))
244 arg_stricts = map isStrictType val_arg_tys ++ repeat False
245 -- These argument types are used as a cheap and cheerful way to find
246 -- unboxed arguments, which must be strict. But it's an InType
247 -- and so there might be a type variable where we expect a function
248 -- type (the substitution hasn't happened yet). And we don't bother
249 -- doing the type applications for a polymorphic function.
250 -- Hence the splitFunTys*IgnoringForAlls*
252 ----------------------------
253 -- If fun_stricts is finite, it means the function returns bottom
254 -- after that number of value args have been consumed
255 -- Otherwise it's infinite, extended with False
257 = case splitStrictSig (idNewStrictness fun) of
258 (demands, result_info)
259 | not (demands `lengthExceeds` countValArgs orig_cont)
260 -> -- Enough args, use the strictness given.
261 -- For bottoming functions we used to pretend that the arg
262 -- is lazy, so that we don't treat the arg as an
263 -- interesting context. This avoids substituting
264 -- top-level bindings for (say) strings into
265 -- calls to error. But now we are more careful about
266 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
267 if isBotRes result_info then
268 map isStrictDmd demands -- Finite => result is bottom
270 map isStrictDmd demands ++ vanilla_stricts
272 other -> vanilla_stricts -- Not enough args, or no strictness
275 interestingArg :: OutExpr -> Bool
276 -- An argument is interesting if it has *some* structure
277 -- We are here trying to avoid unfolding a function that
278 -- is applied only to variables that have no unfolding
279 -- (i.e. they are probably lambda bound): f x y z
280 -- There is little point in inlining f here.
281 interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
282 -- Was: isValueUnfolding (idUnfolding v')
283 -- But that seems over-pessimistic
285 -- This accounts for an argument like
286 -- () or [], which is definitely interesting
287 interestingArg (Type _) = False
288 interestingArg (App fn (Type _)) = interestingArg fn
289 interestingArg (Note _ a) = interestingArg a
290 interestingArg other = True
291 -- Consider let x = 3 in f x
292 -- The substitution will contain (x -> ContEx 3), and we want to
293 -- to say that x is an interesting argument.
294 -- But consider also (\x. f x y) y
295 -- The substitution will contain (x -> ContEx y), and we want to say
296 -- that x is not interesting (assuming y has no unfolding)
299 Comment about interestingCallContext
300 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
301 We want to avoid inlining an expression where there can't possibly be
302 any gain, such as in an argument position. Hence, if the continuation
303 is interesting (eg. a case scrutinee, application etc.) then we
304 inline, otherwise we don't.
306 Previously some_benefit used to return True only if the variable was
307 applied to some value arguments. This didn't work:
309 let x = _coerce_ (T Int) Int (I# 3) in
310 case _coerce_ Int (T Int) x of
313 we want to inline x, but can't see that it's a constructor in a case
314 scrutinee position, and some_benefit is False.
318 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
320 .... case dMonadST _@_ x0 of (a,b,c) -> ....
322 we'd really like to inline dMonadST here, but we *don't* want to
323 inline if the case expression is just
325 case x of y { DEFAULT -> ... }
327 since we can just eliminate this case instead (x is in WHNF). Similar
328 applies when x is bound to a lambda expression. Hence
329 contIsInteresting looks for case expressions with just a single
333 interestingCallContext :: Bool -- False <=> no args at all
334 -> Bool -- False <=> no value args
336 -- The "lone-variable" case is important. I spent ages
337 -- messing about with unsatisfactory varaints, but this is nice.
338 -- The idea is that if a variable appear all alone
339 -- as an arg of lazy fn, or rhs Stop
340 -- as scrutinee of a case Select
341 -- as arg of a strict fn ArgOf
342 -- then we should not inline it (unless there is some other reason,
343 -- e.g. is is the sole occurrence). We achieve this by making
344 -- interestingCallContext return False for a lone variable.
346 -- Why? At least in the case-scrutinee situation, turning
347 -- let x = (a,b) in case x of y -> ...
349 -- let x = (a,b) in case (a,b) of y -> ...
351 -- let x = (a,b) in let y = (a,b) in ...
352 -- is bad if the binding for x will remain.
354 -- Another example: I discovered that strings
355 -- were getting inlined straight back into applications of 'error'
356 -- because the latter is strict.
358 -- f = \x -> ...(error s)...
360 -- Fundamentally such contexts should not ecourage inlining because
361 -- the context can ``see'' the unfolding of the variable (e.g. case or a RULE)
362 -- so there's no gain.
364 -- However, even a type application or coercion isn't a lone variable.
366 -- case $fMonadST @ RealWorld of { :DMonad a b c -> c }
367 -- We had better inline that sucker! The case won't see through it.
369 -- For now, I'm treating treating a variable applied to types
370 -- in a *lazy* context "lone". The motivating example was
372 -- g = /\a. \y. h (f a)
373 -- There's no advantage in inlining f here, and perhaps
374 -- a significant disadvantage. Hence some_val_args in the Stop case
376 interestingCallContext some_args some_val_args cont
379 interesting (InlinePlease _) = True
380 interesting (Select _ _ _ _ _) = some_args
381 interesting (ApplyTo _ _ _ _) = True -- Can happen if we have (coerce t (f x)) y
382 -- Perhaps True is a bit over-keen, but I've
383 -- seen (coerce f) x, where f has an INLINE prag,
384 -- So we have to give some motivaiton for inlining it
385 interesting (ArgOf _ _ _ _) = some_val_args
386 interesting (Stop ty _ upd_in_place) = some_val_args && upd_in_place
387 interesting (CoerceIt _ cont) = interesting cont
388 -- If this call is the arg of a strict function, the context
389 -- is a bit interesting. If we inline here, we may get useful
390 -- evaluation information to avoid repeated evals: e.g.
392 -- Here the contIsInteresting makes the '*' keener to inline,
393 -- which in turn exposes a constructor which makes the '+' inline.
394 -- Assuming that +,* aren't small enough to inline regardless.
396 -- It's also very important to inline in a strict context for things
399 -- Here, the context of (f x) is strict, and if f's unfolding is
400 -- a build it's *great* to inline it here. So we must ensure that
401 -- the context for (f x) is not totally uninteresting.
405 canUpdateInPlace :: Type -> Bool
406 -- Consider let x = <wurble> in ...
407 -- If <wurble> returns an explicit constructor, we might be able
408 -- to do update in place. So we treat even a thunk RHS context
409 -- as interesting if update in place is possible. We approximate
410 -- this by seeing if the type has a single constructor with a
411 -- small arity. But arity zero isn't good -- we share the single copy
412 -- for that case, so no point in sharing.
415 | not opt_UF_UpdateInPlace = False
417 = case splitTyConApp_maybe ty of
419 Just (tycon, _) -> case tyConDataCons_maybe tycon of
420 Just [dc] -> arity == 1 || arity == 2
422 arity = dataConRepArity dc
428 %************************************************************************
430 \subsection{Decisions about inlining}
432 %************************************************************************
434 Inlining is controlled partly by the SimplifierMode switch. This has two
437 SimplGently (a) Simplifying before specialiser/full laziness
438 (b) Simplifiying inside INLINE pragma
439 (c) Simplifying the LHS of a rule
440 (d) Simplifying a GHCi expression or Template
443 SimplPhase n Used at all other times
445 The key thing about SimplGently is that it does no call-site inlining.
446 Before full laziness we must be careful not to inline wrappers,
447 because doing so inhibits floating
448 e.g. ...(case f x of ...)...
449 ==> ...(case (case x of I# x# -> fw x#) of ...)...
450 ==> ...(case x of I# x# -> case fw x# of ...)...
451 and now the redex (f x) isn't floatable any more.
453 The no-inling thing is also important for Template Haskell. You might be
454 compiling in one-shot mode with -O2; but when TH compiles a splice before
455 running it, we don't want to use -O2. Indeed, we don't want to inline
456 anything, because the byte-code interpreter might get confused about
457 unboxed tuples and suchlike.
461 SimplGently is also used as the mode to simplify inside an InlineMe note.
464 inlineMode :: SimplifierMode
465 inlineMode = SimplGently
468 It really is important to switch off inlinings inside such
469 expressions. Consider the following example
475 in ...g...g...g...g...g...
477 Now, if that's the ONLY occurrence of f, it will be inlined inside g,
478 and thence copied multiple times when g is inlined.
481 This function may be inlinined in other modules, so we
482 don't want to remove (by inlining) calls to functions that have
483 specialisations, or that may have transformation rules in an importing
486 E.g. {-# INLINE f #-}
489 and suppose that g is strict *and* has specialisations. If we inline
490 g's wrapper, we deny f the chance of getting the specialised version
491 of g when f is inlined at some call site (perhaps in some other
494 It's also important not to inline a worker back into a wrapper.
496 wraper = inline_me (\x -> ...worker... )
497 Normally, the inline_me prevents the worker getting inlined into
498 the wrapper (initially, the worker's only call site!). But,
499 if the wrapper is sure to be called, the strictness analyser will
500 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
501 continuation. That's why the keep_inline predicate returns True for
502 ArgOf continuations. It shouldn't do any harm not to dissolve the
503 inline-me note under these circumstances.
505 Note that the result is that we do very little simplification
508 all xs = foldr (&&) True xs
509 any p = all . map p {-# INLINE any #-}
511 Problem: any won't get deforested, and so if it's exported and the
512 importer doesn't use the inlining, (eg passes it as an arg) then we
513 won't get deforestation at all. We havn't solved this problem yet!
516 preInlineUnconditionally
517 ~~~~~~~~~~~~~~~~~~~~~~~~
518 @preInlineUnconditionally@ examines a bndr to see if it is used just
519 once in a completely safe way, so that it is safe to discard the
520 binding inline its RHS at the (unique) usage site, REGARDLESS of how
521 big the RHS might be. If this is the case we don't simplify the RHS
522 first, but just inline it un-simplified.
524 This is much better than first simplifying a perhaps-huge RHS and then
525 inlining and re-simplifying it. Indeed, it can be at least quadratically
534 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
535 This can happen with cascades of functions too:
542 THE MAIN INVARIANT is this:
544 ---- preInlineUnconditionally invariant -----
545 IF preInlineUnconditionally chooses to inline x = <rhs>
546 THEN doing the inlining should not change the occurrence
547 info for the free vars of <rhs>
548 ----------------------------------------------
550 For example, it's tempting to look at trivial binding like
552 and inline it unconditionally. But suppose x is used many times,
553 but this is the unique occurrence of y. Then inlining x would change
554 y's occurrence info, which breaks the invariant. It matters: y
555 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
558 Evne RHSs labelled InlineMe aren't caught here, because there might be
559 no benefit from inlining at the call site.
561 [Sept 01] Don't unconditionally inline a top-level thing, because that
562 can simply make a static thing into something built dynamically. E.g.
566 [Remember that we treat \s as a one-shot lambda.] No point in
567 inlining x unless there is something interesting about the call site.
569 But watch out: if you aren't careful, some useful foldr/build fusion
570 can be lost (most notably in spectral/hartel/parstof) because the
571 foldr didn't see the build. Doing the dynamic allocation isn't a big
572 deal, in fact, but losing the fusion can be. But the right thing here
573 seems to be to do a callSiteInline based on the fact that there is
574 something interesting about the call site (it's strict). Hmm. That
577 Conclusion: inline top level things gaily until Phase 0 (the last
578 phase), at which point don't.
581 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
582 preInlineUnconditionally env top_lvl bndr rhs
584 | opt_SimplNoPreInlining = False
585 | otherwise = case idOccInfo bndr of
586 IAmDead -> True -- Happens in ((\x.1) v)
587 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
591 active = case phase of
592 SimplGently -> isAlwaysActive prag
593 SimplPhase n -> isActive n prag
594 prag = idInlinePragma bndr
596 try_once in_lam int_cxt -- There's one textual occurrence
597 | not in_lam = isNotTopLevel top_lvl || early_phase
598 | otherwise = int_cxt && canInlineInLam rhs
600 -- Be very careful before inlining inside a lambda, becuase (a) we must not
601 -- invalidate occurrence information, and (b) we want to avoid pushing a
602 -- single allocation (here) into multiple allocations (inside lambda).
603 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
604 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
606 -- is_cheap = exprIsCheap rhs
607 -- ok = is_cheap && int_cxt
609 -- int_cxt The context isn't totally boring
610 -- E.g. let f = \ab.BIG in \y. map f xs
611 -- Don't want to substitute for f, because then we allocate
612 -- its closure every time the \y is called
613 -- But: let f = \ab.BIG in \y. map (f y) xs
614 -- Now we do want to substitute for f, even though it's not
615 -- saturated, because we're going to allocate a closure for
616 -- (f y) every time round the loop anyhow.
618 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
619 -- so substituting rhs inside a lambda doesn't change the occ info.
620 -- Sadly, not quite the same as exprIsHNF.
621 canInlineInLam (Lit l) = True
622 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
623 canInlineInLam (Note _ e) = canInlineInLam e
624 canInlineInLam _ = False
626 early_phase = case phase of
627 SimplPhase 0 -> False
629 -- If we don't have this early_phase test, consider
630 -- x = length [1,2,3]
631 -- The full laziness pass carefully floats all the cons cells to
632 -- top level, and preInlineUnconditionally floats them all back in.
633 -- Result is (a) static allocation replaced by dynamic allocation
634 -- (b) many simplifier iterations because this tickles
635 -- a related problem; only one inlining per pass
637 -- On the other hand, I have seen cases where top-level fusion is
638 -- lost if we don't inline top level thing (e.g. string constants)
639 -- Hence the test for phase zero (which is the phase for all the final
640 -- simplifications). Until phase zero we take no special notice of
641 -- top level things, but then we become more leery about inlining
646 postInlineUnconditionally
647 ~~~~~~~~~~~~~~~~~~~~~~~~~
648 @postInlineUnconditionally@ decides whether to unconditionally inline
649 a thing based on the form of its RHS; in particular if it has a
650 trivial RHS. If so, we can inline and discard the binding altogether.
652 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
653 only have *forward* references Hence, it's safe to discard the binding
655 NOTE: This isn't our last opportunity to inline. We're at the binding
656 site right now, and we'll get another opportunity when we get to the
659 Note that we do this unconditional inlining only for trival RHSs.
660 Don't inline even WHNFs inside lambdas; doing so may simply increase
661 allocation when the function is called. This isn't the last chance; see
664 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
665 Because we don't even want to inline them into the RHS of constructor
666 arguments. See NOTE above
668 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
669 it's best to inline it anyway. We often get a=E; b=a from desugaring,
670 with both a and b marked NOINLINE. But that seems incompatible with
671 our new view that inlining is like a RULE, so I'm sticking to the 'active'
675 postInlineUnconditionally :: SimplEnv -> TopLevelFlag -> OutId -> OccInfo -> OutExpr -> Unfolding -> Bool
676 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
678 | isLoopBreaker occ_info = False
679 | isExportedId bndr = False
680 | exprIsTrivial rhs = True
683 OneOcc in_lam one_br int_cxt
684 -> (one_br || smallEnoughToInline unfolding) -- Small enough to dup
685 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
687 -- NB: Do we want to inline arbitrarily big things becuase
688 -- one_br is True? that can lead to inline cascades. But
689 -- preInlineUnconditionlly has dealt with all the common cases
690 -- so perhaps it's worth the risk. Here's an example
691 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
693 -- We can't preInlineUnconditionally because that woud invalidate
694 -- the occ info for b. Yet f is used just once, and duplicating
695 -- the case work is fine (exprIsCheap).
697 && ((isNotTopLevel top_lvl && not in_lam) ||
698 -- But outside a lambda, we want to be reasonably aggressive
699 -- about inlining into multiple branches of case
700 -- e.g. let x = <non-value>
701 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
702 -- Inlining can be a big win if C3 is the hot-spot, even if
703 -- the uses in C1, C2 are not 'interesting'
704 -- An example that gets worse if you add int_cxt here is 'clausify'
706 (isCheapUnfolding unfolding && int_cxt))
707 -- isCheap => acceptable work duplication; in_lam may be true
708 -- int_cxt to prevent us inlining inside a lambda without some
709 -- good reason. See the notes on int_cxt in preInlineUnconditionally
712 -- The point here is that for *non-values* that occur
713 -- outside a lambda, the call-site inliner won't have
714 -- a chance (becuase it doesn't know that the thing
715 -- only occurs once). The pre-inliner won't have gotten
716 -- it either, if the thing occurs in more than one branch
717 -- So the main target is things like
720 -- True -> case x of ...
721 -- False -> case x of ...
722 -- I'm not sure how important this is in practice
724 active = case getMode env of
725 SimplGently -> isAlwaysActive prag
726 SimplPhase n -> isActive n prag
727 prag = idInlinePragma bndr
729 activeInline :: SimplEnv -> OutId -> OccInfo -> Bool
730 activeInline env id occ
731 = case getMode env of
732 SimplGently -> isOneOcc occ && isAlwaysActive prag
733 -- No inlining at all when doing gentle stuff,
734 -- except for local things that occur once
735 -- The reason is that too little clean-up happens if you
736 -- don't inline use-once things. Also a bit of inlining is *good* for
737 -- full laziness; it can expose constant sub-expressions.
738 -- Example in spectral/mandel/Mandel.hs, where the mandelset
739 -- function gets a useful let-float if you inline windowToViewport
741 -- NB: we used to have a second exception, for data con wrappers.
742 -- On the grounds that we use gentle mode for rule LHSs, and
743 -- they match better when data con wrappers are inlined.
744 -- But that only really applies to the trivial wrappers (like (:)),
745 -- and they are now constructed as Compulsory unfoldings (in MkId)
746 -- so they'll happen anyway.
748 SimplPhase n -> isActive n prag
750 prag = idInlinePragma id
752 activeRule :: SimplEnv -> Maybe (Activation -> Bool)
753 -- Nothing => No rules at all
755 | opt_RulesOff = Nothing
757 = case getMode env of
758 SimplGently -> Just isAlwaysActive
759 -- Used to be Nothing (no rules in gentle mode)
760 -- Main motivation for changing is that I wanted
761 -- lift String ===> ...
762 -- to work in Template Haskell when simplifying
763 -- splices, so we get simpler code for literal strings
764 SimplPhase n -> Just (isActive n)
768 %************************************************************************
770 \subsection{Rebuilding a lambda}
772 %************************************************************************
775 mkLam :: SimplEnv -> [OutBinder] -> OutExpr -> SimplCont -> SimplM FloatsWithExpr
779 a) eta reduction, if that gives a trivial expression
780 b) eta expansion [only if there are some value lambdas]
781 c) floating lets out through big lambdas
782 [only if all tyvar lambdas, and only if this lambda
786 mkLam env bndrs body cont
787 = getDOptsSmpl `thenSmpl` \dflags ->
788 mkLam' dflags env bndrs body cont
790 mkLam' dflags env bndrs body cont
791 | dopt Opt_DoEtaReduction dflags,
792 Just etad_lam <- tryEtaReduce bndrs body
793 = tick (EtaReduction (head bndrs)) `thenSmpl_`
794 returnSmpl (emptyFloats env, etad_lam)
796 | dopt Opt_DoLambdaEtaExpansion dflags,
797 any isRuntimeVar bndrs
798 = tryEtaExpansion body `thenSmpl` \ body' ->
799 returnSmpl (emptyFloats env, mkLams bndrs body')
801 {- Sept 01: I'm experimenting with getting the
802 full laziness pass to float out past big lambdsa
803 | all isTyVar bndrs, -- Only for big lambdas
804 contIsRhs cont -- Only try the rhs type-lambda floating
805 -- if this is indeed a right-hand side; otherwise
806 -- we end up floating the thing out, only for float-in
807 -- to float it right back in again!
808 = tryRhsTyLam env bndrs body `thenSmpl` \ (floats, body') ->
809 returnSmpl (floats, mkLams bndrs body')
813 = returnSmpl (emptyFloats env, mkLams bndrs body)
817 %************************************************************************
819 \subsection{Eta expansion and reduction}
821 %************************************************************************
823 We try for eta reduction here, but *only* if we get all the
824 way to an exprIsTrivial expression.
825 We don't want to remove extra lambdas unless we are going
826 to avoid allocating this thing altogether
829 tryEtaReduce :: [OutBinder] -> OutExpr -> Maybe OutExpr
830 tryEtaReduce bndrs body
831 -- We don't use CoreUtils.etaReduce, because we can be more
833 -- (a) we already have the binders
834 -- (b) we can do the triviality test before computing the free vars
835 = go (reverse bndrs) body
837 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
838 go [] fun | ok_fun fun = Just fun -- Success!
839 go _ _ = Nothing -- Failure!
841 ok_fun fun = exprIsTrivial fun
842 && not (any (`elemVarSet` (exprFreeVars fun)) bndrs)
843 && (exprIsHNF fun || all ok_lam bndrs)
844 ok_lam v = isTyVar v || isDictId v
845 -- The exprIsHNF is because eta reduction is not
846 -- valid in general: \x. bot /= bot
847 -- So we need to be sure that the "fun" is a value.
849 -- However, we always want to reduce (/\a -> f a) to f
850 -- This came up in a RULE: foldr (build (/\a -> g a))
851 -- did not match foldr (build (/\b -> ...something complex...))
852 -- The type checker can insert these eta-expanded versions,
853 -- with both type and dictionary lambdas; hence the slightly
856 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
860 Try eta expansion for RHSs
863 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
866 where (in both cases)
868 * The xi can include type variables
870 * The yi are all value variables
872 * N is a NORMAL FORM (i.e. no redexes anywhere)
873 wanting a suitable number of extra args.
875 We may have to sandwich some coerces between the lambdas
876 to make the types work. exprEtaExpandArity looks through coerces
877 when computing arity; and etaExpand adds the coerces as necessary when
878 actually computing the expansion.
881 tryEtaExpansion :: OutExpr -> SimplM OutExpr
882 -- There is at least one runtime binder in the binders
884 = getUniquesSmpl `thenSmpl` \ us ->
885 returnSmpl (etaExpand fun_arity us body (exprType body))
887 fun_arity = exprEtaExpandArity body
891 %************************************************************************
893 \subsection{Floating lets out of big lambdas}
895 %************************************************************************
897 tryRhsTyLam tries this transformation, when the big lambda appears as
898 the RHS of a let(rec) binding:
900 /\abc -> let(rec) x = e in b
902 let(rec) x' = /\abc -> let x = x' a b c in e
904 /\abc -> let x = x' a b c in b
906 This is good because it can turn things like:
908 let f = /\a -> letrec g = ... g ... in g
910 letrec g' = /\a -> ... g' a ...
914 which is better. In effect, it means that big lambdas don't impede
917 This optimisation is CRUCIAL in eliminating the junk introduced by
918 desugaring mutually recursive definitions. Don't eliminate it lightly!
920 So far as the implementation is concerned:
922 Invariant: go F e = /\tvs -> F e
926 = Let x' = /\tvs -> F e
930 G = F . Let x = x' tvs
932 go F (Letrec xi=ei in b)
933 = Letrec {xi' = /\tvs -> G ei}
937 G = F . Let {xi = xi' tvs}
939 [May 1999] If we do this transformation *regardless* then we can
940 end up with some pretty silly stuff. For example,
943 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
948 st = /\s -> ...[y1 s/x1, y2 s/x2]
951 Unless the "..." is a WHNF there is really no point in doing this.
952 Indeed it can make things worse. Suppose x1 is used strictly,
955 x1* = case f y of { (a,b) -> e }
957 If we abstract this wrt the tyvar we then can't do the case inline
958 as we would normally do.
962 {- Trying to do this in full laziness
964 tryRhsTyLam :: SimplEnv -> [OutTyVar] -> OutExpr -> SimplM FloatsWithExpr
965 -- Call ensures that all the binders are type variables
967 tryRhsTyLam env tyvars body -- Only does something if there's a let
968 | not (all isTyVar tyvars)
969 || not (worth_it body) -- inside a type lambda,
970 = returnSmpl (emptyFloats env, body) -- and a WHNF inside that
973 = go env (\x -> x) body
976 worth_it e@(Let _ _) = whnf_in_middle e
979 whnf_in_middle (Let (NonRec x rhs) e) | isUnLiftedType (idType x) = False
980 whnf_in_middle (Let _ e) = whnf_in_middle e
981 whnf_in_middle e = exprIsCheap e
983 main_tyvar_set = mkVarSet tyvars
985 go env fn (Let bind@(NonRec var rhs) body)
987 = go env (fn . Let bind) body
989 go env fn (Let (NonRec var rhs) body)
990 = mk_poly tyvars_here var `thenSmpl` \ (var', rhs') ->
991 addAuxiliaryBind env (NonRec var' (mkLams tyvars_here (fn rhs))) $ \ env ->
992 go env (fn . Let (mk_silly_bind var rhs')) body
996 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprSomeFreeVars isTyVar rhs)
997 -- Abstract only over the type variables free in the rhs
998 -- wrt which the new binding is abstracted. But the naive
999 -- approach of abstract wrt the tyvars free in the Id's type
1001 -- /\ a b -> let t :: (a,b) = (e1, e2)
1004 -- Here, b isn't free in x's type, but we must nevertheless
1005 -- abstract wrt b as well, because t's type mentions b.
1006 -- Since t is floated too, we'd end up with the bogus:
1007 -- poly_t = /\ a b -> (e1, e2)
1008 -- poly_x = /\ a -> fst (poly_t a *b*)
1009 -- So for now we adopt the even more naive approach of
1010 -- abstracting wrt *all* the tyvars. We'll see if that
1011 -- gives rise to problems. SLPJ June 98
1013 go env fn (Let (Rec prs) body)
1014 = mapAndUnzipSmpl (mk_poly tyvars_here) vars `thenSmpl` \ (vars', rhss') ->
1016 gn body = fn (foldr Let body (zipWith mk_silly_bind vars rhss'))
1017 pairs = vars' `zip` [mkLams tyvars_here (gn rhs) | rhs <- rhss]
1019 addAuxiliaryBind env (Rec pairs) $ \ env ->
1022 (vars,rhss) = unzip prs
1023 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprsSomeFreeVars isTyVar (map snd prs))
1024 -- See notes with tyvars_here above
1026 go env fn body = returnSmpl (emptyFloats env, fn body)
1028 mk_poly tyvars_here var
1029 = getUniqueSmpl `thenSmpl` \ uniq ->
1031 poly_name = setNameUnique (idName var) uniq -- Keep same name
1032 poly_ty = mkForAllTys tyvars_here (idType var) -- But new type of course
1033 poly_id = mkLocalId poly_name poly_ty
1035 -- In the olden days, it was crucial to copy the occInfo of the original var,
1036 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1037 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1038 -- at already simplified code, so it doesn't matter
1040 -- It's even right to retain single-occurrence or dead-var info:
1041 -- Suppose we started with /\a -> let x = E in B
1042 -- where x occurs once in B. Then we transform to:
1043 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1044 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1045 -- the occurrences of x' will be just the occurrences originally
1048 returnSmpl (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tyvars_here))
1050 mk_silly_bind var rhs = NonRec var (Note InlineMe rhs)
1051 -- Suppose we start with:
1053 -- x = /\ a -> let g = G in E
1055 -- Then we'll float to get
1057 -- x = let poly_g = /\ a -> G
1058 -- in /\ a -> let g = poly_g a in E
1060 -- But now the occurrence analyser will see just one occurrence
1061 -- of poly_g, not inside a lambda, so the simplifier will
1062 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1063 -- (I used to think that the "don't inline lone occurrences" stuff
1064 -- would stop this happening, but since it's the *only* occurrence,
1065 -- PreInlineUnconditionally kicks in first!)
1067 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1068 -- to appear many times. (NB: mkInlineMe eliminates
1069 -- such notes on trivial RHSs, so do it manually.)
1073 %************************************************************************
1075 \subsection{Case alternative filtering
1077 %************************************************************************
1079 prepareAlts does two things:
1081 1. Eliminate alternatives that cannot match, including the
1082 DEFAULT alternative.
1084 2. If the DEFAULT alternative can match only one possible constructor,
1085 then make that constructor explicit.
1087 case e of x { DEFAULT -> rhs }
1089 case e of x { (a,b) -> rhs }
1090 where the type is a single constructor type. This gives better code
1091 when rhs also scrutinises x or e.
1093 It's a good idea do do this stuff before simplifying the alternatives, to
1094 avoid simplifying alternatives we know can't happen, and to come up with
1095 the list of constructors that are handled, to put into the IdInfo of the
1096 case binder, for use when simplifying the alternatives.
1098 Eliminating the default alternative in (1) isn't so obvious, but it can
1101 data Colour = Red | Green | Blue
1110 DEFAULT -> [ case y of ... ]
1112 If we inline h into f, the default case of the inlined h can't happen.
1113 If we don't notice this, we may end up filtering out *all* the cases
1114 of the inner case y, which give us nowhere to go!
1118 prepareAlts :: OutExpr -- Scrutinee
1119 -> InId -- Case binder
1120 -> [InAlt] -- Increasing order
1121 -> SimplM ([InAlt], -- Better alternatives, still incresaing order
1122 [AltCon]) -- These cases are handled
1124 prepareAlts scrut case_bndr alts
1126 (alts_wo_default, maybe_deflt) = findDefault alts
1128 impossible_cons = case scrut of
1129 Var v -> otherCons (idUnfolding v)
1132 -- Filter out alternatives that can't possibly match
1133 better_alts | null impossible_cons = alts_wo_default
1134 | otherwise = [alt | alt@(con,_,_) <- alts_wo_default,
1135 not (con `elem` impossible_cons)]
1137 -- "handled_cons" are handled either by the context,
1138 -- or by a branch in this case expression
1139 -- (Don't add DEFAULT to the handled_cons!!)
1140 handled_cons = impossible_cons ++ [con | (con,_,_) <- better_alts]
1142 -- Filter out the default, if it can't happen,
1143 -- or replace it with "proper" alternative if there
1144 -- is only one constructor left
1145 prepareDefault case_bndr handled_cons maybe_deflt `thenSmpl` \ deflt_alt ->
1147 returnSmpl (mergeAlts better_alts deflt_alt, handled_cons)
1148 -- We need the mergeAlts in case the new default_alt
1149 -- has turned into a constructor alternative.
1151 prepareDefault case_bndr handled_cons (Just rhs)
1152 | Just (tycon, inst_tys) <- splitTyConApp_maybe (idType case_bndr),
1153 isAlgTyCon tycon, -- It's a data type, tuple, or unboxed tuples.
1154 not (isNewTyCon tycon), -- We can have a newtype, if we are just doing an eval:
1155 -- case x of { DEFAULT -> e }
1156 -- and we don't want to fill in a default for them!
1157 Just all_cons <- tyConDataCons_maybe tycon,
1158 not (null all_cons), -- This is a tricky corner case. If the data type has no constructors,
1159 -- which GHC allows, then the case expression will have at most a default
1160 -- alternative. We don't want to eliminate that alternative, because the
1161 -- invariant is that there's always one alternative. It's more convenient
1163 -- case x of { DEFAULT -> e }
1164 -- as it is, rather than transform it to
1165 -- error "case cant match"
1166 -- which would be quite legitmate. But it's a really obscure corner, and
1167 -- not worth wasting code on.
1168 let handled_data_cons = [data_con | DataAlt data_con <- handled_cons],
1169 let missing_cons = [con | con <- all_cons,
1170 not (con `elem` handled_data_cons)]
1171 = case missing_cons of
1172 [] -> returnSmpl [] -- Eliminate the default alternative
1173 -- if it can't match
1175 [con] -> -- It matches exactly one constructor, so fill it in
1176 tick (FillInCaseDefault case_bndr) `thenSmpl_`
1177 mk_args con inst_tys `thenSmpl` \ args ->
1178 returnSmpl [(DataAlt con, args, rhs)]
1180 two_or_more -> returnSmpl [(DEFAULT, [], rhs)]
1183 = returnSmpl [(DEFAULT, [], rhs)]
1185 prepareDefault case_bndr handled_cons Nothing
1188 mk_args missing_con inst_tys
1189 = mk_tv_bndrs missing_con inst_tys `thenSmpl` \ (tv_bndrs, inst_tys') ->
1190 getUniquesSmpl `thenSmpl` \ id_uniqs ->
1191 let arg_tys = dataConArgTys missing_con inst_tys'
1192 arg_ids = zipWith (mkSysLocal FSLIT("a")) id_uniqs arg_tys
1194 returnSmpl (tv_bndrs ++ arg_ids)
1196 mk_tv_bndrs missing_con inst_tys
1197 | isVanillaDataCon missing_con
1198 = returnSmpl ([], inst_tys)
1200 = getUniquesSmpl `thenSmpl` \ tv_uniqs ->
1201 let new_tvs = zipWith mk tv_uniqs (dataConTyVars missing_con)
1202 mk uniq tv = mkTyVar (mkSysTvName uniq FSLIT("t")) (tyVarKind tv)
1204 returnSmpl (new_tvs, mkTyVarTys new_tvs)
1208 %************************************************************************
1210 \subsection{Case absorption and identity-case elimination}
1212 %************************************************************************
1214 mkCase puts a case expression back together, trying various transformations first.
1217 mkCase :: OutExpr -> OutId -> OutType
1218 -> [OutAlt] -- Increasing order
1221 mkCase scrut case_bndr ty alts
1222 = getDOptsSmpl `thenSmpl` \dflags ->
1223 mkAlts dflags scrut case_bndr alts `thenSmpl` \ better_alts ->
1224 mkCase1 scrut case_bndr ty better_alts
1228 mkAlts tries these things:
1230 1. If several alternatives are identical, merge them into
1231 a single DEFAULT alternative. I've occasionally seen this
1232 making a big difference:
1234 case e of =====> case e of
1235 C _ -> f x D v -> ....v....
1236 D v -> ....v.... DEFAULT -> f x
1239 The point is that we merge common RHSs, at least for the DEFAULT case.
1240 [One could do something more elaborate but I've never seen it needed.]
1241 To avoid an expensive test, we just merge branches equal to the *first*
1242 alternative; this picks up the common cases
1243 a) all branches equal
1244 b) some branches equal to the DEFAULT (which occurs first)
1247 case e of b { ==> case e of b {
1248 p1 -> rhs1 p1 -> rhs1
1250 pm -> rhsm pm -> rhsm
1251 _ -> case b of b' { pn -> let b'=b in rhsn
1253 ... po -> let b'=b in rhso
1254 po -> rhso _ -> let b'=b in rhsd
1258 which merges two cases in one case when -- the default alternative of
1259 the outer case scrutises the same variable as the outer case This
1260 transformation is called Case Merging. It avoids that the same
1261 variable is scrutinised multiple times.
1264 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
1270 where @is@ was something like
1272 p `is` n = p /= (-1) && p == n
1274 This gave rise to a horrible sequence of cases
1281 and similarly in cascade for all the join points!
1286 --------------------------------------------------
1287 -- 1. Merge identical branches
1288 --------------------------------------------------
1289 mkAlts dflags scrut case_bndr alts@((con1,bndrs1,rhs1) : con_alts)
1290 | all isDeadBinder bndrs1, -- Remember the default
1291 length filtered_alts < length con_alts -- alternative comes first
1292 = tick (AltMerge case_bndr) `thenSmpl_`
1293 returnSmpl better_alts
1295 filtered_alts = filter keep con_alts
1296 keep (con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1297 better_alts = (DEFAULT, [], rhs1) : filtered_alts
1300 --------------------------------------------------
1301 -- 2. Merge nested cases
1302 --------------------------------------------------
1304 mkAlts dflags scrut outer_bndr outer_alts
1305 | dopt Opt_CaseMerge dflags,
1306 (outer_alts_without_deflt, maybe_outer_deflt) <- findDefault outer_alts,
1307 Just (Case (Var scrut_var) inner_bndr _ inner_alts) <- maybe_outer_deflt,
1308 scruting_same_var scrut_var
1310 munged_inner_alts = [(con, args, munge_rhs rhs) | (con, args, rhs) <- inner_alts]
1311 munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1313 new_alts = mergeAlts outer_alts_without_deflt munged_inner_alts
1314 -- The merge keeps the inner DEFAULT at the front, if there is one
1315 -- and eliminates any inner_alts that are shadowed by the outer_alts
1317 tick (CaseMerge outer_bndr) `thenSmpl_`
1319 -- Warning: don't call mkAlts recursively!
1320 -- Firstly, there's no point, because inner alts have already had
1321 -- mkCase applied to them, so they won't have a case in their default
1322 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1323 -- in munge_rhs may put a case into the DEFAULT branch!
1325 -- We are scrutinising the same variable if it's
1326 -- the outer case-binder, or if the outer case scrutinises a variable
1327 -- (and it's the same). Testing both allows us not to replace the
1328 -- outer scrut-var with the outer case-binder (Simplify.simplCaseBinder).
1329 scruting_same_var = case scrut of
1330 Var outer_scrut -> \ v -> v == outer_bndr || v == outer_scrut
1331 other -> \ v -> v == outer_bndr
1333 ------------------------------------------------
1335 ------------------------------------------------
1337 mkAlts dflags scrut case_bndr other_alts = returnSmpl other_alts
1340 ---------------------------------
1341 mergeAlts :: [OutAlt] -> [OutAlt] -> [OutAlt]
1342 -- Merge preserving order; alternatives in the first arg
1343 -- shadow ones in the second
1344 mergeAlts [] as2 = as2
1345 mergeAlts as1 [] = as1
1346 mergeAlts (a1:as1) (a2:as2)
1347 = case a1 `cmpAlt` a2 of
1348 LT -> a1 : mergeAlts as1 (a2:as2)
1349 EQ -> a1 : mergeAlts as1 as2 -- Discard a2
1350 GT -> a2 : mergeAlts (a1:as1) as2
1355 =================================================================================
1357 mkCase1 tries these things
1359 1. Eliminate the case altogether if possible
1367 and similar friends.
1370 Start with a simple situation:
1372 case x# of ===> e[x#/y#]
1375 (when x#, y# are of primitive type, of course). We can't (in general)
1376 do this for algebraic cases, because we might turn bottom into
1379 Actually, we generalise this idea to look for a case where we're
1380 scrutinising a variable, and we know that only the default case can
1385 other -> ...(case x of
1389 Here the inner case can be eliminated. This really only shows up in
1390 eliminating error-checking code.
1392 We also make sure that we deal with this very common case:
1397 Here we are using the case as a strict let; if x is used only once
1398 then we want to inline it. We have to be careful that this doesn't
1399 make the program terminate when it would have diverged before, so we
1401 - x is used strictly, or
1402 - e is already evaluated (it may so if e is a variable)
1404 Lastly, we generalise the transformation to handle this:
1410 We only do this for very cheaply compared r's (constructors, literals
1411 and variables). If pedantic bottoms is on, we only do it when the
1412 scrutinee is a PrimOp which can't fail.
1414 We do it *here*, looking at un-simplified alternatives, because we
1415 have to check that r doesn't mention the variables bound by the
1416 pattern in each alternative, so the binder-info is rather useful.
1418 So the case-elimination algorithm is:
1420 1. Eliminate alternatives which can't match
1422 2. Check whether all the remaining alternatives
1423 (a) do not mention in their rhs any of the variables bound in their pattern
1424 and (b) have equal rhss
1426 3. Check we can safely ditch the case:
1427 * PedanticBottoms is off,
1428 or * the scrutinee is an already-evaluated variable
1429 or * the scrutinee is a primop which is ok for speculation
1430 -- ie we want to preserve divide-by-zero errors, and
1431 -- calls to error itself!
1433 or * [Prim cases] the scrutinee is a primitive variable
1435 or * [Alg cases] the scrutinee is a variable and
1436 either * the rhs is the same variable
1437 (eg case x of C a b -> x ===> x)
1438 or * there is only one alternative, the default alternative,
1439 and the binder is used strictly in its scope.
1440 [NB this is helped by the "use default binder where
1441 possible" transformation; see below.]
1444 If so, then we can replace the case with one of the rhss.
1446 Further notes about case elimination
1447 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1448 Consider: test :: Integer -> IO ()
1451 Turns out that this compiles to:
1454 eta1 :: State# RealWorld ->
1455 case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT ->
1457 (PrelNum.jtos eta ($w[] @ Char))
1459 of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }}
1461 Notice the strange '<' which has no effect at all. This is a funny one.
1462 It started like this:
1464 f x y = if x < 0 then jtos x
1465 else if y==0 then "" else jtos x
1467 At a particular call site we have (f v 1). So we inline to get
1469 if v < 0 then jtos x
1470 else if 1==0 then "" else jtos x
1472 Now simplify the 1==0 conditional:
1474 if v<0 then jtos v else jtos v
1476 Now common-up the two branches of the case:
1478 case (v<0) of DEFAULT -> jtos v
1480 Why don't we drop the case? Because it's strict in v. It's technically
1481 wrong to drop even unnecessary evaluations, and in practice they
1482 may be a result of 'seq' so we *definitely* don't want to drop those.
1483 I don't really know how to improve this situation.
1487 --------------------------------------------------
1488 -- 0. Check for empty alternatives
1489 --------------------------------------------------
1492 mkCase1 scrut case_bndr ty []
1493 = pprTrace "mkCase1: null alts" (ppr case_bndr <+> ppr scrut) $
1497 --------------------------------------------------
1498 -- 1. Eliminate the case altogether if poss
1499 --------------------------------------------------
1501 mkCase1 scrut case_bndr ty [(con,bndrs,rhs)]
1502 -- See if we can get rid of the case altogether
1503 -- See the extensive notes on case-elimination above
1504 -- mkCase made sure that if all the alternatives are equal,
1505 -- then there is now only one (DEFAULT) rhs
1506 | all isDeadBinder bndrs,
1508 -- Check that the scrutinee can be let-bound instead of case-bound
1509 exprOkForSpeculation scrut
1510 -- OK not to evaluate it
1511 -- This includes things like (==# a# b#)::Bool
1512 -- so that we simplify
1513 -- case ==# a# b# of { True -> x; False -> x }
1516 -- This particular example shows up in default methods for
1517 -- comparision operations (e.g. in (>=) for Int.Int32)
1518 || exprIsHNF scrut -- It's already evaluated
1519 || var_demanded_later scrut -- It'll be demanded later
1521 -- || not opt_SimplPedanticBottoms) -- Or we don't care!
1522 -- We used to allow improving termination by discarding cases, unless -fpedantic-bottoms was on,
1523 -- but that breaks badly for the dataToTag# primop, which relies on a case to evaluate
1524 -- its argument: case x of { y -> dataToTag# y }
1525 -- Here we must *not* discard the case, because dataToTag# just fetches the tag from
1526 -- the info pointer. So we'll be pedantic all the time, and see if that gives any
1528 -- Also we don't want to discard 'seq's
1529 = tick (CaseElim case_bndr) `thenSmpl_`
1530 returnSmpl (bindCaseBndr case_bndr scrut rhs)
1533 -- The case binder is going to be evaluated later,
1534 -- and the scrutinee is a simple variable
1535 var_demanded_later (Var v) = isStrictDmd (idNewDemandInfo case_bndr)
1536 var_demanded_later other = False
1539 --------------------------------------------------
1541 --------------------------------------------------
1543 mkCase1 scrut case_bndr ty alts -- Identity case
1544 | all identity_alt alts
1545 = tick (CaseIdentity case_bndr) `thenSmpl_`
1546 returnSmpl (re_note scrut)
1548 identity_alt (con, args, rhs) = de_note rhs `cheapEqExpr` identity_rhs con args
1550 identity_rhs (DataAlt con) args = mkConApp con (arg_tys ++ map varToCoreExpr args)
1551 identity_rhs (LitAlt lit) _ = Lit lit
1552 identity_rhs DEFAULT _ = Var case_bndr
1554 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1557 -- case coerce T e of x { _ -> coerce T' x }
1558 -- And we definitely want to eliminate this case!
1559 -- So we throw away notes from the RHS, and reconstruct
1560 -- (at least an approximation) at the other end
1561 de_note (Note _ e) = de_note e
1564 -- re_note wraps a coerce if it might be necessary
1565 re_note scrut = case head alts of
1566 (_,_,rhs1@(Note _ _)) -> mkCoerce2 (exprType rhs1) (idType case_bndr) scrut
1570 --------------------------------------------------
1572 --------------------------------------------------
1573 mkCase1 scrut bndr ty alts = returnSmpl (Case scrut bndr ty alts)
1577 When adding auxiliary bindings for the case binder, it's worth checking if
1578 its dead, because it often is, and occasionally these mkCase transformations
1579 cascade rather nicely.
1582 bindCaseBndr bndr rhs body
1583 | isDeadBinder bndr = body
1584 | otherwise = bindNonRec bndr rhs body