2 % (c) The AQUA Project, Glasgow University, 1993-1998
4 \section[SimplUtils]{The simplifier utilities}
11 preInlineUnconditionally, postInlineUnconditionally, activeInline, activeRule,
14 -- The continuation type
15 SimplCont(..), DupFlag(..), LetRhsFlag(..),
16 contIsDupable, contResultType,
17 countValArgs, countArgs, pushContArgs,
18 mkBoringStop, mkLazyArgStop, mkRhsStop, contIsRhs, contIsRhsOrArg,
19 getContArgs, interestingCallContext, interestingArgContext,
20 interestingArg, isStrictType
24 #include "HsVersions.h"
27 import DynFlags ( SimplifierSwitch(..), SimplifierMode(..),
28 DynFlags, DynFlag(..), dopt )
29 import StaticFlags ( opt_UF_UpdateInPlace, opt_SimplNoPreInlining,
32 import CoreFVs ( exprFreeVars )
33 import CoreUtils ( cheapEqExpr, exprType, exprIsTrivial,
34 etaExpand, exprEtaExpandArity, bindNonRec, mkCoerce,
35 findDefault, exprOkForSpeculation, exprIsHNF, mergeAlts,
38 import Literal ( mkStringLit )
39 import CoreUnfold ( smallEnoughToInline )
40 import MkId ( eRROR_ID, wrapNewTypeBody )
41 import Id ( Id, idType, isDataConWorkId, idOccInfo, isDictId,
42 isDeadBinder, idNewDemandInfo, isExportedId, mkSysLocal,
43 idUnfolding, idNewStrictness, idInlinePragma, idHasRules
45 import NewDemand ( isStrictDmd, isBotRes, splitStrictSig )
47 import Name ( mkSysTvName )
48 import Type ( Type, splitFunTys, dropForAlls, isStrictType,
49 splitTyConApp_maybe, tyConAppArgs, mkTyVarTys )
50 import Coercion ( isEqPredTy
52 import Coercion ( Coercion, mkUnsafeCoercion, coercionKind )
53 import TyCon ( tyConDataCons_maybe, isClosedNewTyCon )
54 import DataCon ( DataCon, dataConRepArity, dataConInstArgTys, dataConTyCon )
56 import BasicTypes ( TopLevelFlag(..), isNotTopLevel, OccInfo(..), isLoopBreaker, isOneOcc,
57 Activation, isAlwaysActive, isActive )
58 import Util ( lengthExceeds )
63 %************************************************************************
65 \subsection{The continuation data type}
67 %************************************************************************
70 data SimplCont -- Strict contexts
71 = Stop OutType -- Type of the result
73 Bool -- True <=> There is something interesting about
74 -- the context, and hence the inliner
75 -- should be a bit keener (see interestingCallContext)
77 -- (a) This is the RHS of a thunk whose type suggests
78 -- that update-in-place would be possible
79 -- (b) This is an argument of a function that has RULES
80 -- Inlining the call might allow the rule to fire
82 | CoerceIt OutCoercion -- The coercion simplified
86 CoreExpr -- The argument
87 (Maybe SimplEnv) -- (Just se) => the arg is un-simplified and this is its subst-env
88 -- Nothing => the arg is already simplified; don't repeatedly simplify it!
89 SimplCont -- and its environment
92 InId [InAlt] SimplEnv -- The case binder, alts, and subst-env
95 | ArgOf LetRhsFlag -- An arbitrary strict context: the argument
96 -- of a strict function, or a primitive-arg fn
98 -- No DupFlag, because we never duplicate it
99 OutType -- arg_ty: type of the argument itself
100 OutType -- cont_ty: the type of the expression being sought by the context
101 -- f (error "foo") ==> coerce t (error "foo")
103 -- We need to know the type t, to which to coerce.
105 (SimplEnv -> OutExpr -> SimplM FloatsWithExpr) -- What to do with the result
106 -- The result expression in the OutExprStuff has type cont_ty
108 data LetRhsFlag = AnArg -- It's just an argument not a let RHS
109 | AnRhs -- It's the RHS of a let (so please float lets out of big lambdas)
111 instance Outputable LetRhsFlag where
112 ppr AnArg = ptext SLIT("arg")
113 ppr AnRhs = ptext SLIT("rhs")
115 instance Outputable SimplCont where
116 ppr (Stop ty is_rhs _) = ptext SLIT("Stop") <> brackets (ppr is_rhs) <+> ppr ty
117 ppr (ApplyTo dup arg se cont) = (ptext SLIT("ApplyTo") <+> ppr dup <+> ppr arg) $$ ppr cont
118 ppr (ArgOf _ _ _ _) = ptext SLIT("ArgOf...")
119 ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$
120 (nest 4 (ppr alts)) $$ ppr cont
121 ppr (CoerceIt co cont) = (ptext SLIT("CoerceIt") <+> ppr co) $$ ppr cont
123 data DupFlag = OkToDup | NoDup
125 instance Outputable DupFlag where
126 ppr OkToDup = ptext SLIT("ok")
127 ppr NoDup = ptext SLIT("nodup")
132 mkBoringStop :: OutType -> SimplCont
133 mkBoringStop ty = Stop ty AnArg False
135 mkLazyArgStop :: OutType -> Bool -> SimplCont
136 mkLazyArgStop ty has_rules = Stop ty AnArg (canUpdateInPlace ty || has_rules)
138 mkRhsStop :: OutType -> SimplCont
139 mkRhsStop ty = Stop ty AnRhs (canUpdateInPlace ty)
141 contIsRhs :: SimplCont -> Bool
142 contIsRhs (Stop _ AnRhs _) = True
143 contIsRhs (ArgOf AnRhs _ _ _) = True
144 contIsRhs other = False
146 contIsRhsOrArg (Stop _ _ _) = True
147 contIsRhsOrArg (ArgOf _ _ _ _) = True
148 contIsRhsOrArg other = False
151 contIsDupable :: SimplCont -> Bool
152 contIsDupable (Stop _ _ _) = True
153 contIsDupable (ApplyTo OkToDup _ _ _) = True
154 contIsDupable (Select OkToDup _ _ _ _) = True
155 contIsDupable (CoerceIt _ cont) = contIsDupable cont
156 contIsDupable other = False
159 discardableCont :: SimplCont -> Bool
160 discardableCont (Stop _ _ _) = False
161 discardableCont (CoerceIt _ cont) = discardableCont cont
162 discardableCont other = True
164 discardCont :: Type -- The type expected
165 -> SimplCont -- A continuation, expecting the previous type
166 -> SimplCont -- Replace the continuation with a suitable coerce
167 discardCont from_ty cont = case cont of
168 Stop to_ty is_rhs _ -> cont
169 other -> CoerceIt co (mkBoringStop to_ty)
171 co = mkUnsafeCoercion from_ty to_ty
172 to_ty = contResultType cont
175 contResultType :: SimplCont -> OutType
176 contResultType (Stop to_ty _ _) = to_ty
177 contResultType (ArgOf _ _ to_ty _) = to_ty
178 contResultType (ApplyTo _ _ _ cont) = contResultType cont
179 contResultType (CoerceIt _ cont) = contResultType cont
180 contResultType (Select _ _ _ _ cont) = contResultType cont
183 countValArgs :: SimplCont -> Int
184 countValArgs (ApplyTo _ (Type ty) se cont) = countValArgs cont
185 countValArgs (ApplyTo _ val_arg se cont) = 1 + countValArgs cont
186 countValArgs other = 0
188 countArgs :: SimplCont -> Int
189 countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont
193 pushContArgs ::[OutArg] -> SimplCont -> SimplCont
194 -- Pushes args with the specified environment
195 pushContArgs [] cont = cont
196 pushContArgs (arg : args) cont = ApplyTo NoDup arg Nothing (pushContArgs args cont)
201 getContArgs :: SwitchChecker
202 -> OutId -> SimplCont
203 -> ([(InExpr, Maybe SimplEnv, Bool)], -- Arguments; the Bool is true for strict args
204 SimplCont) -- Remaining continuation
205 -- getContArgs id k = (args, k', inl)
206 -- args are the leading ApplyTo items in k
207 -- (i.e. outermost comes first)
208 -- augmented with demand info from the functionn
209 getContArgs chkr fun orig_cont
211 -- Ignore strictness info if the no-case-of-case
212 -- flag is on. Strictness changes evaluation order
213 -- and that can change full laziness
214 stricts | switchIsOn chkr NoCaseOfCase = vanilla_stricts
215 | otherwise = computed_stricts
217 go [] stricts orig_cont
219 ----------------------------
222 go acc ss (ApplyTo _ arg@(Type _) se cont)
223 = go ((arg,se,False) : acc) ss cont
224 -- NB: don't bother to instantiate the function type
227 go acc (s:ss) (ApplyTo _ arg se cont)
228 = go ((arg,se,s) : acc) ss cont
230 -- We're run out of arguments, or else we've run out of demands
231 -- The latter only happens if the result is guaranteed bottom
232 -- This is the case for
233 -- * case (error "hello") of { ... }
234 -- * (error "Hello") arg
235 -- * f (error "Hello") where f is strict
237 -- Then, especially in the first of these cases, we'd like to discard
238 -- the continuation, leaving just the bottoming expression. But the
239 -- type might not be right, so we may have to add a coerce.
242 | null ss && discardableCont cont = (args, discardCont hole_ty cont)
243 | otherwise = (args, cont)
246 hole_ty = applyTypeToArgs (Var fun) (idType fun)
247 [substExpr_mb se arg | (arg,se,_) <- args]
248 substExpr_mb Nothing arg = arg
249 substExpr_mb (Just se) arg = substExpr se arg
251 ----------------------------
252 vanilla_stricts, computed_stricts :: [Bool]
253 vanilla_stricts = repeat False
254 computed_stricts = zipWith (||) fun_stricts arg_stricts
256 ----------------------------
257 (val_arg_tys, res_ty) = splitFunTys (dropForAlls (idType fun))
258 arg_stricts = map isStrictType val_arg_tys ++ repeat False
259 -- These argument types are used as a cheap and cheerful way to find
260 -- unboxed arguments, which must be strict. But it's an InType
261 -- and so there might be a type variable where we expect a function
262 -- type (the substitution hasn't happened yet). And we don't bother
263 -- doing the type applications for a polymorphic function.
264 -- Hence the splitFunTys*IgnoringForAlls*
266 ----------------------------
267 -- If fun_stricts is finite, it means the function returns bottom
268 -- after that number of value args have been consumed
269 -- Otherwise it's infinite, extended with False
271 = case splitStrictSig (idNewStrictness fun) of
272 (demands, result_info)
273 | not (demands `lengthExceeds` countValArgs orig_cont)
274 -> -- Enough args, use the strictness given.
275 -- For bottoming functions we used to pretend that the arg
276 -- is lazy, so that we don't treat the arg as an
277 -- interesting context. This avoids substituting
278 -- top-level bindings for (say) strings into
279 -- calls to error. But now we are more careful about
280 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
281 if isBotRes result_info then
282 map isStrictDmd demands -- Finite => result is bottom
284 map isStrictDmd demands ++ vanilla_stricts
286 other -> vanilla_stricts -- Not enough args, or no strictness
289 interestingArg :: OutExpr -> Bool
290 -- An argument is interesting if it has *some* structure
291 -- We are here trying to avoid unfolding a function that
292 -- is applied only to variables that have no unfolding
293 -- (i.e. they are probably lambda bound): f x y z
294 -- There is little point in inlining f here.
295 interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
296 -- Was: isValueUnfolding (idUnfolding v')
297 -- But that seems over-pessimistic
299 -- This accounts for an argument like
300 -- () or [], which is definitely interesting
301 interestingArg (Type _) = False
302 interestingArg (App fn (Type _)) = interestingArg fn
303 interestingArg (Note _ a) = interestingArg a
304 interestingArg other = True
305 -- Consider let x = 3 in f x
306 -- The substitution will contain (x -> ContEx 3), and we want to
307 -- to say that x is an interesting argument.
308 -- But consider also (\x. f x y) y
309 -- The substitution will contain (x -> ContEx y), and we want to say
310 -- that x is not interesting (assuming y has no unfolding)
313 Comment about interestingCallContext
314 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
315 We want to avoid inlining an expression where there can't possibly be
316 any gain, such as in an argument position. Hence, if the continuation
317 is interesting (eg. a case scrutinee, application etc.) then we
318 inline, otherwise we don't.
320 Previously some_benefit used to return True only if the variable was
321 applied to some value arguments. This didn't work:
323 let x = _coerce_ (T Int) Int (I# 3) in
324 case _coerce_ Int (T Int) x of
327 we want to inline x, but can't see that it's a constructor in a case
328 scrutinee position, and some_benefit is False.
332 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
334 .... case dMonadST _@_ x0 of (a,b,c) -> ....
336 we'd really like to inline dMonadST here, but we *don't* want to
337 inline if the case expression is just
339 case x of y { DEFAULT -> ... }
341 since we can just eliminate this case instead (x is in WHNF). Similar
342 applies when x is bound to a lambda expression. Hence
343 contIsInteresting looks for case expressions with just a single
347 interestingCallContext :: Bool -- False <=> no args at all
348 -> Bool -- False <=> no value args
350 -- The "lone-variable" case is important. I spent ages
351 -- messing about with unsatisfactory varaints, but this is nice.
352 -- The idea is that if a variable appear all alone
353 -- as an arg of lazy fn, or rhs Stop
354 -- as scrutinee of a case Select
355 -- as arg of a strict fn ArgOf
356 -- then we should not inline it (unless there is some other reason,
357 -- e.g. is is the sole occurrence). We achieve this by making
358 -- interestingCallContext return False for a lone variable.
360 -- Why? At least in the case-scrutinee situation, turning
361 -- let x = (a,b) in case x of y -> ...
363 -- let x = (a,b) in case (a,b) of y -> ...
365 -- let x = (a,b) in let y = (a,b) in ...
366 -- is bad if the binding for x will remain.
368 -- Another example: I discovered that strings
369 -- were getting inlined straight back into applications of 'error'
370 -- because the latter is strict.
372 -- f = \x -> ...(error s)...
374 -- Fundamentally such contexts should not ecourage inlining because
375 -- the context can ``see'' the unfolding of the variable (e.g. case or a RULE)
376 -- so there's no gain.
378 -- However, even a type application or coercion isn't a lone variable.
380 -- case $fMonadST @ RealWorld of { :DMonad a b c -> c }
381 -- We had better inline that sucker! The case won't see through it.
383 -- For now, I'm treating treating a variable applied to types
384 -- in a *lazy* context "lone". The motivating example was
386 -- g = /\a. \y. h (f a)
387 -- There's no advantage in inlining f here, and perhaps
388 -- a significant disadvantage. Hence some_val_args in the Stop case
390 interestingCallContext some_args some_val_args cont
393 interesting (Select {}) = some_args
394 interesting (ApplyTo {}) = True -- Can happen if we have (coerce t (f x)) y
395 -- Perhaps True is a bit over-keen, but I've
396 -- seen (coerce f) x, where f has an INLINE prag,
397 -- So we have to give some motivaiton for inlining it
398 interesting (ArgOf {}) = some_val_args
399 interesting (Stop ty _ interesting) = some_val_args && interesting
400 interesting (CoerceIt _ cont) = interesting cont
401 -- If this call is the arg of a strict function, the context
402 -- is a bit interesting. If we inline here, we may get useful
403 -- evaluation information to avoid repeated evals: e.g.
405 -- Here the contIsInteresting makes the '*' keener to inline,
406 -- which in turn exposes a constructor which makes the '+' inline.
407 -- Assuming that +,* aren't small enough to inline regardless.
409 -- It's also very important to inline in a strict context for things
412 -- Here, the context of (f x) is strict, and if f's unfolding is
413 -- a build it's *great* to inline it here. So we must ensure that
414 -- the context for (f x) is not totally uninteresting.
418 interestingArgContext :: Id -> SimplCont -> Bool
419 -- If the argument has form (f x y), where x,y are boring,
420 -- and f is marked INLINE, then we don't want to inline f.
421 -- But if the context of the argument is
423 -- where g has rules, then we *do* want to inline f, in case it
424 -- exposes a rule that might fire. Similarly, if the context is
426 -- where h has rules, then we do want to inline f.
427 -- The interesting_arg_ctxt flag makes this happen; if it's
428 -- set, the inliner gets just enough keener to inline f
429 -- regardless of how boring f's arguments are, if it's marked INLINE
431 -- The alternative would be to *always* inline an INLINE function,
432 -- regardless of how boring its context is; but that seems overkill
433 -- For example, it'd mean that wrapper functions were always inlined
434 interestingArgContext fn cont
435 = idHasRules fn || go cont
437 go (Select {}) = False
438 go (ApplyTo {}) = False
440 go (CoerceIt _ c) = go c
441 go (Stop _ _ interesting) = interesting
444 canUpdateInPlace :: Type -> Bool
445 -- Consider let x = <wurble> in ...
446 -- If <wurble> returns an explicit constructor, we might be able
447 -- to do update in place. So we treat even a thunk RHS context
448 -- as interesting if update in place is possible. We approximate
449 -- this by seeing if the type has a single constructor with a
450 -- small arity. But arity zero isn't good -- we share the single copy
451 -- for that case, so no point in sharing.
454 | not opt_UF_UpdateInPlace = False
456 = case splitTyConApp_maybe ty of
458 Just (tycon, _) -> case tyConDataCons_maybe tycon of
459 Just [dc] -> arity == 1 || arity == 2
461 arity = dataConRepArity dc
467 %************************************************************************
469 \subsection{Decisions about inlining}
471 %************************************************************************
473 Inlining is controlled partly by the SimplifierMode switch. This has two
476 SimplGently (a) Simplifying before specialiser/full laziness
477 (b) Simplifiying inside INLINE pragma
478 (c) Simplifying the LHS of a rule
479 (d) Simplifying a GHCi expression or Template
482 SimplPhase n Used at all other times
484 The key thing about SimplGently is that it does no call-site inlining.
485 Before full laziness we must be careful not to inline wrappers,
486 because doing so inhibits floating
487 e.g. ...(case f x of ...)...
488 ==> ...(case (case x of I# x# -> fw x#) of ...)...
489 ==> ...(case x of I# x# -> case fw x# of ...)...
490 and now the redex (f x) isn't floatable any more.
492 The no-inlining thing is also important for Template Haskell. You might be
493 compiling in one-shot mode with -O2; but when TH compiles a splice before
494 running it, we don't want to use -O2. Indeed, we don't want to inline
495 anything, because the byte-code interpreter might get confused about
496 unboxed tuples and suchlike.
500 SimplGently is also used as the mode to simplify inside an InlineMe note.
503 inlineMode :: SimplifierMode
504 inlineMode = SimplGently
507 It really is important to switch off inlinings inside such
508 expressions. Consider the following example
514 in ...g...g...g...g...g...
516 Now, if that's the ONLY occurrence of f, it will be inlined inside g,
517 and thence copied multiple times when g is inlined.
520 This function may be inlinined in other modules, so we
521 don't want to remove (by inlining) calls to functions that have
522 specialisations, or that may have transformation rules in an importing
525 E.g. {-# INLINE f #-}
528 and suppose that g is strict *and* has specialisations. If we inline
529 g's wrapper, we deny f the chance of getting the specialised version
530 of g when f is inlined at some call site (perhaps in some other
533 It's also important not to inline a worker back into a wrapper.
535 wraper = inline_me (\x -> ...worker... )
536 Normally, the inline_me prevents the worker getting inlined into
537 the wrapper (initially, the worker's only call site!). But,
538 if the wrapper is sure to be called, the strictness analyser will
539 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
540 continuation. That's why the keep_inline predicate returns True for
541 ArgOf continuations. It shouldn't do any harm not to dissolve the
542 inline-me note under these circumstances.
544 Note that the result is that we do very little simplification
547 all xs = foldr (&&) True xs
548 any p = all . map p {-# INLINE any #-}
550 Problem: any won't get deforested, and so if it's exported and the
551 importer doesn't use the inlining, (eg passes it as an arg) then we
552 won't get deforestation at all. We havn't solved this problem yet!
555 preInlineUnconditionally
556 ~~~~~~~~~~~~~~~~~~~~~~~~
557 @preInlineUnconditionally@ examines a bndr to see if it is used just
558 once in a completely safe way, so that it is safe to discard the
559 binding inline its RHS at the (unique) usage site, REGARDLESS of how
560 big the RHS might be. If this is the case we don't simplify the RHS
561 first, but just inline it un-simplified.
563 This is much better than first simplifying a perhaps-huge RHS and then
564 inlining and re-simplifying it. Indeed, it can be at least quadratically
573 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
574 This can happen with cascades of functions too:
581 THE MAIN INVARIANT is this:
583 ---- preInlineUnconditionally invariant -----
584 IF preInlineUnconditionally chooses to inline x = <rhs>
585 THEN doing the inlining should not change the occurrence
586 info for the free vars of <rhs>
587 ----------------------------------------------
589 For example, it's tempting to look at trivial binding like
591 and inline it unconditionally. But suppose x is used many times,
592 but this is the unique occurrence of y. Then inlining x would change
593 y's occurrence info, which breaks the invariant. It matters: y
594 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
597 Evne RHSs labelled InlineMe aren't caught here, because there might be
598 no benefit from inlining at the call site.
600 [Sept 01] Don't unconditionally inline a top-level thing, because that
601 can simply make a static thing into something built dynamically. E.g.
605 [Remember that we treat \s as a one-shot lambda.] No point in
606 inlining x unless there is something interesting about the call site.
608 But watch out: if you aren't careful, some useful foldr/build fusion
609 can be lost (most notably in spectral/hartel/parstof) because the
610 foldr didn't see the build. Doing the dynamic allocation isn't a big
611 deal, in fact, but losing the fusion can be. But the right thing here
612 seems to be to do a callSiteInline based on the fact that there is
613 something interesting about the call site (it's strict). Hmm. That
616 Conclusion: inline top level things gaily until Phase 0 (the last
617 phase), at which point don't.
620 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
621 preInlineUnconditionally env top_lvl bndr rhs
623 | opt_SimplNoPreInlining = False
624 | otherwise = case idOccInfo bndr of
625 IAmDead -> True -- Happens in ((\x.1) v)
626 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
630 active = case phase of
631 SimplGently -> isAlwaysActive prag
632 SimplPhase n -> isActive n prag
633 prag = idInlinePragma bndr
635 try_once in_lam int_cxt -- There's one textual occurrence
636 | not in_lam = isNotTopLevel top_lvl || early_phase
637 | otherwise = int_cxt && canInlineInLam rhs
639 -- Be very careful before inlining inside a lambda, becuase (a) we must not
640 -- invalidate occurrence information, and (b) we want to avoid pushing a
641 -- single allocation (here) into multiple allocations (inside lambda).
642 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
643 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
645 -- is_cheap = exprIsCheap rhs
646 -- ok = is_cheap && int_cxt
648 -- int_cxt The context isn't totally boring
649 -- E.g. let f = \ab.BIG in \y. map f xs
650 -- Don't want to substitute for f, because then we allocate
651 -- its closure every time the \y is called
652 -- But: let f = \ab.BIG in \y. map (f y) xs
653 -- Now we do want to substitute for f, even though it's not
654 -- saturated, because we're going to allocate a closure for
655 -- (f y) every time round the loop anyhow.
657 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
658 -- so substituting rhs inside a lambda doesn't change the occ info.
659 -- Sadly, not quite the same as exprIsHNF.
660 canInlineInLam (Lit l) = True
661 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
662 canInlineInLam (Note _ e) = canInlineInLam e
663 canInlineInLam _ = False
665 early_phase = case phase of
666 SimplPhase 0 -> False
668 -- If we don't have this early_phase test, consider
669 -- x = length [1,2,3]
670 -- The full laziness pass carefully floats all the cons cells to
671 -- top level, and preInlineUnconditionally floats them all back in.
672 -- Result is (a) static allocation replaced by dynamic allocation
673 -- (b) many simplifier iterations because this tickles
674 -- a related problem; only one inlining per pass
676 -- On the other hand, I have seen cases where top-level fusion is
677 -- lost if we don't inline top level thing (e.g. string constants)
678 -- Hence the test for phase zero (which is the phase for all the final
679 -- simplifications). Until phase zero we take no special notice of
680 -- top level things, but then we become more leery about inlining
685 postInlineUnconditionally
686 ~~~~~~~~~~~~~~~~~~~~~~~~~
687 @postInlineUnconditionally@ decides whether to unconditionally inline
688 a thing based on the form of its RHS; in particular if it has a
689 trivial RHS. If so, we can inline and discard the binding altogether.
691 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
692 only have *forward* references Hence, it's safe to discard the binding
694 NOTE: This isn't our last opportunity to inline. We're at the binding
695 site right now, and we'll get another opportunity when we get to the
698 Note that we do this unconditional inlining only for trival RHSs.
699 Don't inline even WHNFs inside lambdas; doing so may simply increase
700 allocation when the function is called. This isn't the last chance; see
703 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
704 Because we don't even want to inline them into the RHS of constructor
705 arguments. See NOTE above
707 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
708 it's best to inline it anyway. We often get a=E; b=a from desugaring,
709 with both a and b marked NOINLINE. But that seems incompatible with
710 our new view that inlining is like a RULE, so I'm sticking to the 'active'
714 postInlineUnconditionally
715 :: SimplEnv -> TopLevelFlag
716 -> InId -- The binder (an OutId would be fine too)
717 -> OccInfo -- From the InId
721 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
723 | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, dont' inline
724 -- because it might be referred to "earlier"
725 | isExportedId bndr = False
726 | exprIsTrivial rhs = True
729 -- The point of examining occ_info here is that for *non-values*
730 -- that occur outside a lambda, the call-site inliner won't have
731 -- a chance (becuase it doesn't know that the thing
732 -- only occurs once). The pre-inliner won't have gotten
733 -- it either, if the thing occurs in more than one branch
734 -- So the main target is things like
737 -- True -> case x of ...
738 -- False -> case x of ...
739 -- I'm not sure how important this is in practice
740 OneOcc in_lam one_br int_cxt -- OneOcc => no work-duplication issue
741 -> smallEnoughToInline unfolding -- Small enough to dup
742 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
744 -- NB: Do NOT inline arbitrarily big things, even if one_br is True
745 -- Reason: doing so risks exponential behaviour. We simplify a big
746 -- expression, inline it, and simplify it again. But if the
747 -- very same thing happens in the big expression, we get
749 -- PRINCIPLE: when we've already simplified an expression once,
750 -- make sure that we only inline it if it's reasonably small.
752 && ((isNotTopLevel top_lvl && not in_lam) ||
753 -- But outside a lambda, we want to be reasonably aggressive
754 -- about inlining into multiple branches of case
755 -- e.g. let x = <non-value>
756 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
757 -- Inlining can be a big win if C3 is the hot-spot, even if
758 -- the uses in C1, C2 are not 'interesting'
759 -- An example that gets worse if you add int_cxt here is 'clausify'
761 (isCheapUnfolding unfolding && int_cxt))
762 -- isCheap => acceptable work duplication; in_lam may be true
763 -- int_cxt to prevent us inlining inside a lambda without some
764 -- good reason. See the notes on int_cxt in preInlineUnconditionally
766 IAmDead -> True -- This happens; for example, the case_bndr during case of
767 -- known constructor: case (a,b) of x { (p,q) -> ... }
768 -- Here x isn't mentioned in the RHS, so we don't want to
769 -- create the (dead) let-binding let x = (a,b) in ...
773 -- Here's an example that we don't handle well:
774 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
775 -- in \y. ....case f of {...} ....
776 -- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
778 -- * We can't preInlineUnconditionally because that woud invalidate
779 -- the occ info for b.
780 -- * We can't postInlineUnconditionally because the RHS is big, and
781 -- that risks exponential behaviour
782 -- * We can't call-site inline, because the rhs is big
786 active = case getMode env of
787 SimplGently -> isAlwaysActive prag
788 SimplPhase n -> isActive n prag
789 prag = idInlinePragma bndr
791 activeInline :: SimplEnv -> OutId -> Bool
793 = case getMode env of
795 -- No inlining at all when doing gentle stuff,
796 -- except for local things that occur once
797 -- The reason is that too little clean-up happens if you
798 -- don't inline use-once things. Also a bit of inlining is *good* for
799 -- full laziness; it can expose constant sub-expressions.
800 -- Example in spectral/mandel/Mandel.hs, where the mandelset
801 -- function gets a useful let-float if you inline windowToViewport
803 -- NB: we used to have a second exception, for data con wrappers.
804 -- On the grounds that we use gentle mode for rule LHSs, and
805 -- they match better when data con wrappers are inlined.
806 -- But that only really applies to the trivial wrappers (like (:)),
807 -- and they are now constructed as Compulsory unfoldings (in MkId)
808 -- so they'll happen anyway.
810 SimplPhase n -> isActive n prag
812 prag = idInlinePragma id
814 activeRule :: SimplEnv -> Maybe (Activation -> Bool)
815 -- Nothing => No rules at all
817 | opt_RulesOff = Nothing
819 = case getMode env of
820 SimplGently -> Just isAlwaysActive
821 -- Used to be Nothing (no rules in gentle mode)
822 -- Main motivation for changing is that I wanted
823 -- lift String ===> ...
824 -- to work in Template Haskell when simplifying
825 -- splices, so we get simpler code for literal strings
826 SimplPhase n -> Just (isActive n)
830 %************************************************************************
832 \subsection{Rebuilding a lambda}
834 %************************************************************************
837 mkLam :: SimplEnv -> [OutBinder] -> OutExpr -> SimplCont -> SimplM FloatsWithExpr
841 a) eta reduction, if that gives a trivial expression
842 b) eta expansion [only if there are some value lambdas]
843 c) floating lets out through big lambdas
844 [only if all tyvar lambdas, and only if this lambda
848 mkLam env bndrs body cont
849 = getDOptsSmpl `thenSmpl` \dflags ->
850 mkLam' dflags env bndrs body cont
852 mkLam' dflags env bndrs body cont
853 | dopt Opt_DoEtaReduction dflags,
854 Just etad_lam <- tryEtaReduce bndrs body
855 = tick (EtaReduction (head bndrs)) `thenSmpl_`
856 returnSmpl (emptyFloats env, etad_lam)
858 | dopt Opt_DoLambdaEtaExpansion dflags,
859 any isRuntimeVar bndrs
860 = tryEtaExpansion dflags body `thenSmpl` \ body' ->
861 returnSmpl (emptyFloats env, mkLams bndrs body')
863 {- Sept 01: I'm experimenting with getting the
864 full laziness pass to float out past big lambdsa
865 | all isTyVar bndrs, -- Only for big lambdas
866 contIsRhs cont -- Only try the rhs type-lambda floating
867 -- if this is indeed a right-hand side; otherwise
868 -- we end up floating the thing out, only for float-in
869 -- to float it right back in again!
870 = tryRhsTyLam env bndrs body `thenSmpl` \ (floats, body') ->
871 returnSmpl (floats, mkLams bndrs body')
875 = returnSmpl (emptyFloats env, mkLams bndrs body)
879 %************************************************************************
881 \subsection{Eta expansion and reduction}
883 %************************************************************************
885 We try for eta reduction here, but *only* if we get all the
886 way to an exprIsTrivial expression.
887 We don't want to remove extra lambdas unless we are going
888 to avoid allocating this thing altogether
891 tryEtaReduce :: [OutBinder] -> OutExpr -> Maybe OutExpr
892 tryEtaReduce bndrs body
893 -- We don't use CoreUtils.etaReduce, because we can be more
895 -- (a) we already have the binders
896 -- (b) we can do the triviality test before computing the free vars
897 = go (reverse bndrs) body
899 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
900 go [] fun | ok_fun fun = Just fun -- Success!
901 go _ _ = Nothing -- Failure!
903 ok_fun fun = exprIsTrivial fun
904 && not (any (`elemVarSet` (exprFreeVars fun)) bndrs)
905 && (exprIsHNF fun || all ok_lam bndrs)
906 ok_lam v = isTyVar v || isDictId v
907 -- The exprIsHNF is because eta reduction is not
908 -- valid in general: \x. bot /= bot
909 -- So we need to be sure that the "fun" is a value.
911 -- However, we always want to reduce (/\a -> f a) to f
912 -- This came up in a RULE: foldr (build (/\a -> g a))
913 -- did not match foldr (build (/\b -> ...something complex...))
914 -- The type checker can insert these eta-expanded versions,
915 -- with both type and dictionary lambdas; hence the slightly
918 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
922 Try eta expansion for RHSs
925 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
928 where (in both cases)
930 * The xi can include type variables
932 * The yi are all value variables
934 * N is a NORMAL FORM (i.e. no redexes anywhere)
935 wanting a suitable number of extra args.
937 We may have to sandwich some coerces between the lambdas
938 to make the types work. exprEtaExpandArity looks through coerces
939 when computing arity; and etaExpand adds the coerces as necessary when
940 actually computing the expansion.
943 tryEtaExpansion :: DynFlags -> OutExpr -> SimplM OutExpr
944 -- There is at least one runtime binder in the binders
945 tryEtaExpansion dflags body
946 = getUniquesSmpl `thenSmpl` \ us ->
947 returnSmpl (etaExpand fun_arity us body (exprType body))
949 fun_arity = exprEtaExpandArity dflags body
953 %************************************************************************
955 \subsection{Floating lets out of big lambdas}
957 %************************************************************************
959 tryRhsTyLam tries this transformation, when the big lambda appears as
960 the RHS of a let(rec) binding:
962 /\abc -> let(rec) x = e in b
964 let(rec) x' = /\abc -> let x = x' a b c in e
966 /\abc -> let x = x' a b c in b
968 This is good because it can turn things like:
970 let f = /\a -> letrec g = ... g ... in g
972 letrec g' = /\a -> ... g' a ...
976 which is better. In effect, it means that big lambdas don't impede
979 This optimisation is CRUCIAL in eliminating the junk introduced by
980 desugaring mutually recursive definitions. Don't eliminate it lightly!
982 So far as the implementation is concerned:
984 Invariant: go F e = /\tvs -> F e
988 = Let x' = /\tvs -> F e
992 G = F . Let x = x' tvs
994 go F (Letrec xi=ei in b)
995 = Letrec {xi' = /\tvs -> G ei}
999 G = F . Let {xi = xi' tvs}
1001 [May 1999] If we do this transformation *regardless* then we can
1002 end up with some pretty silly stuff. For example,
1005 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
1010 st = /\s -> ...[y1 s/x1, y2 s/x2]
1013 Unless the "..." is a WHNF there is really no point in doing this.
1014 Indeed it can make things worse. Suppose x1 is used strictly,
1017 x1* = case f y of { (a,b) -> e }
1019 If we abstract this wrt the tyvar we then can't do the case inline
1020 as we would normally do.
1024 {- Trying to do this in full laziness
1026 tryRhsTyLam :: SimplEnv -> [OutTyVar] -> OutExpr -> SimplM FloatsWithExpr
1027 -- Call ensures that all the binders are type variables
1029 tryRhsTyLam env tyvars body -- Only does something if there's a let
1030 | not (all isTyVar tyvars)
1031 || not (worth_it body) -- inside a type lambda,
1032 = returnSmpl (emptyFloats env, body) -- and a WHNF inside that
1035 = go env (\x -> x) body
1038 worth_it e@(Let _ _) = whnf_in_middle e
1041 whnf_in_middle (Let (NonRec x rhs) e) | isUnLiftedType (idType x) = False
1042 whnf_in_middle (Let _ e) = whnf_in_middle e
1043 whnf_in_middle e = exprIsCheap e
1045 main_tyvar_set = mkVarSet tyvars
1047 go env fn (Let bind@(NonRec var rhs) body)
1049 = go env (fn . Let bind) body
1051 go env fn (Let (NonRec var rhs) body)
1052 = mk_poly tyvars_here var `thenSmpl` \ (var', rhs') ->
1053 addAuxiliaryBind env (NonRec var' (mkLams tyvars_here (fn rhs))) $ \ env ->
1054 go env (fn . Let (mk_silly_bind var rhs')) body
1058 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprSomeFreeVars isTyVar rhs)
1059 -- Abstract only over the type variables free in the rhs
1060 -- wrt which the new binding is abstracted. But the naive
1061 -- approach of abstract wrt the tyvars free in the Id's type
1063 -- /\ a b -> let t :: (a,b) = (e1, e2)
1066 -- Here, b isn't free in x's type, but we must nevertheless
1067 -- abstract wrt b as well, because t's type mentions b.
1068 -- Since t is floated too, we'd end up with the bogus:
1069 -- poly_t = /\ a b -> (e1, e2)
1070 -- poly_x = /\ a -> fst (poly_t a *b*)
1071 -- So for now we adopt the even more naive approach of
1072 -- abstracting wrt *all* the tyvars. We'll see if that
1073 -- gives rise to problems. SLPJ June 98
1075 go env fn (Let (Rec prs) body)
1076 = mapAndUnzipSmpl (mk_poly tyvars_here) vars `thenSmpl` \ (vars', rhss') ->
1078 gn body = fn (foldr Let body (zipWith mk_silly_bind vars rhss'))
1079 pairs = vars' `zip` [mkLams tyvars_here (gn rhs) | rhs <- rhss]
1081 addAuxiliaryBind env (Rec pairs) $ \ env ->
1084 (vars,rhss) = unzip prs
1085 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprsSomeFreeVars isTyVar (map snd prs))
1086 -- See notes with tyvars_here above
1088 go env fn body = returnSmpl (emptyFloats env, fn body)
1090 mk_poly tyvars_here var
1091 = getUniqueSmpl `thenSmpl` \ uniq ->
1093 poly_name = setNameUnique (idName var) uniq -- Keep same name
1094 poly_ty = mkForAllTys tyvars_here (idType var) -- But new type of course
1095 poly_id = mkLocalId poly_name poly_ty
1097 -- In the olden days, it was crucial to copy the occInfo of the original var,
1098 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1099 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1100 -- at already simplified code, so it doesn't matter
1102 -- It's even right to retain single-occurrence or dead-var info:
1103 -- Suppose we started with /\a -> let x = E in B
1104 -- where x occurs once in B. Then we transform to:
1105 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1106 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1107 -- the occurrences of x' will be just the occurrences originally
1110 returnSmpl (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tyvars_here))
1112 mk_silly_bind var rhs = NonRec var (Note InlineMe rhs)
1113 -- Suppose we start with:
1115 -- x = /\ a -> let g = G in E
1117 -- Then we'll float to get
1119 -- x = let poly_g = /\ a -> G
1120 -- in /\ a -> let g = poly_g a in E
1122 -- But now the occurrence analyser will see just one occurrence
1123 -- of poly_g, not inside a lambda, so the simplifier will
1124 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1125 -- (I used to think that the "don't inline lone occurrences" stuff
1126 -- would stop this happening, but since it's the *only* occurrence,
1127 -- PreInlineUnconditionally kicks in first!)
1129 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1130 -- to appear many times. (NB: mkInlineMe eliminates
1131 -- such notes on trivial RHSs, so do it manually.)
1135 %************************************************************************
1137 \subsection{Case absorption and identity-case elimination}
1139 %************************************************************************
1142 mkCase puts a case expression back together, trying various transformations first.
1145 mkCase :: OutExpr -> OutId -> OutType
1146 -> [OutAlt] -- Increasing order
1149 mkCase scrut case_bndr ty alts
1150 = getDOptsSmpl `thenSmpl` \dflags ->
1151 mkAlts dflags scrut case_bndr alts `thenSmpl` \ better_alts ->
1152 mkCase1 scrut case_bndr ty better_alts
1156 mkAlts tries these things:
1158 1. If several alternatives are identical, merge them into
1159 a single DEFAULT alternative. I've occasionally seen this
1160 making a big difference:
1162 case e of =====> case e of
1163 C _ -> f x D v -> ....v....
1164 D v -> ....v.... DEFAULT -> f x
1167 The point is that we merge common RHSs, at least for the DEFAULT case.
1168 [One could do something more elaborate but I've never seen it needed.]
1169 To avoid an expensive test, we just merge branches equal to the *first*
1170 alternative; this picks up the common cases
1171 a) all branches equal
1172 b) some branches equal to the DEFAULT (which occurs first)
1175 case e of b { ==> case e of b {
1176 p1 -> rhs1 p1 -> rhs1
1178 pm -> rhsm pm -> rhsm
1179 _ -> case b of b' { pn -> let b'=b in rhsn
1181 ... po -> let b'=b in rhso
1182 po -> rhso _ -> let b'=b in rhsd
1186 which merges two cases in one case when -- the default alternative of
1187 the outer case scrutises the same variable as the outer case This
1188 transformation is called Case Merging. It avoids that the same
1189 variable is scrutinised multiple times.
1192 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
1198 where @is@ was something like
1200 p `is` n = p /= (-1) && p == n
1202 This gave rise to a horrible sequence of cases
1209 and similarly in cascade for all the join points!
1214 --------------------------------------------------
1215 -- 1. Merge identical branches
1216 --------------------------------------------------
1217 mkAlts dflags scrut case_bndr alts@((con1,bndrs1,rhs1) : con_alts)
1218 | all isDeadBinder bndrs1, -- Remember the default
1219 length filtered_alts < length con_alts -- alternative comes first
1220 = tick (AltMerge case_bndr) `thenSmpl_`
1221 returnSmpl better_alts
1223 filtered_alts = filter keep con_alts
1224 keep (con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1225 better_alts = (DEFAULT, [], rhs1) : filtered_alts
1228 --------------------------------------------------
1229 -- 2. Merge nested cases
1230 --------------------------------------------------
1232 mkAlts dflags scrut outer_bndr outer_alts
1233 | dopt Opt_CaseMerge dflags,
1234 (outer_alts_without_deflt, maybe_outer_deflt) <- findDefault outer_alts,
1235 Just (Case (Var scrut_var) inner_bndr _ inner_alts) <- maybe_outer_deflt,
1236 scruting_same_var scrut_var
1238 munged_inner_alts = [(con, args, munge_rhs rhs) | (con, args, rhs) <- inner_alts]
1239 munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1241 new_alts = mergeAlts outer_alts_without_deflt munged_inner_alts
1242 -- The merge keeps the inner DEFAULT at the front, if there is one
1243 -- and eliminates any inner_alts that are shadowed by the outer_alts
1245 tick (CaseMerge outer_bndr) `thenSmpl_`
1247 -- Warning: don't call mkAlts recursively!
1248 -- Firstly, there's no point, because inner alts have already had
1249 -- mkCase applied to them, so they won't have a case in their default
1250 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1251 -- in munge_rhs may put a case into the DEFAULT branch!
1253 -- We are scrutinising the same variable if it's
1254 -- the outer case-binder, or if the outer case scrutinises a variable
1255 -- (and it's the same). Testing both allows us not to replace the
1256 -- outer scrut-var with the outer case-binder (Simplify.simplCaseBinder).
1257 scruting_same_var = case scrut of
1258 Var outer_scrut -> \ v -> v == outer_bndr || v == outer_scrut
1259 other -> \ v -> v == outer_bndr
1261 ------------------------------------------------
1263 ------------------------------------------------
1265 mkAlts dflags scrut case_bndr other_alts = returnSmpl other_alts
1270 =================================================================================
1272 mkCase1 tries these things
1274 1. Eliminate the case altogether if possible
1282 and similar friends.
1285 Start with a simple situation:
1287 case x# of ===> e[x#/y#]
1290 (when x#, y# are of primitive type, of course). We can't (in general)
1291 do this for algebraic cases, because we might turn bottom into
1294 Actually, we generalise this idea to look for a case where we're
1295 scrutinising a variable, and we know that only the default case can
1300 other -> ...(case x of
1304 Here the inner case can be eliminated. This really only shows up in
1305 eliminating error-checking code.
1307 We also make sure that we deal with this very common case:
1312 Here we are using the case as a strict let; if x is used only once
1313 then we want to inline it. We have to be careful that this doesn't
1314 make the program terminate when it would have diverged before, so we
1316 - x is used strictly, or
1317 - e is already evaluated (it may so if e is a variable)
1319 Lastly, we generalise the transformation to handle this:
1325 We only do this for very cheaply compared r's (constructors, literals
1326 and variables). If pedantic bottoms is on, we only do it when the
1327 scrutinee is a PrimOp which can't fail.
1329 We do it *here*, looking at un-simplified alternatives, because we
1330 have to check that r doesn't mention the variables bound by the
1331 pattern in each alternative, so the binder-info is rather useful.
1333 So the case-elimination algorithm is:
1335 1. Eliminate alternatives which can't match
1337 2. Check whether all the remaining alternatives
1338 (a) do not mention in their rhs any of the variables bound in their pattern
1339 and (b) have equal rhss
1341 3. Check we can safely ditch the case:
1342 * PedanticBottoms is off,
1343 or * the scrutinee is an already-evaluated variable
1344 or * the scrutinee is a primop which is ok for speculation
1345 -- ie we want to preserve divide-by-zero errors, and
1346 -- calls to error itself!
1348 or * [Prim cases] the scrutinee is a primitive variable
1350 or * [Alg cases] the scrutinee is a variable and
1351 either * the rhs is the same variable
1352 (eg case x of C a b -> x ===> x)
1353 or * there is only one alternative, the default alternative,
1354 and the binder is used strictly in its scope.
1355 [NB this is helped by the "use default binder where
1356 possible" transformation; see below.]
1359 If so, then we can replace the case with one of the rhss.
1361 Further notes about case elimination
1362 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1363 Consider: test :: Integer -> IO ()
1366 Turns out that this compiles to:
1369 eta1 :: State# RealWorld ->
1370 case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT ->
1372 (PrelNum.jtos eta ($w[] @ Char))
1374 of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }}
1376 Notice the strange '<' which has no effect at all. This is a funny one.
1377 It started like this:
1379 f x y = if x < 0 then jtos x
1380 else if y==0 then "" else jtos x
1382 At a particular call site we have (f v 1). So we inline to get
1384 if v < 0 then jtos x
1385 else if 1==0 then "" else jtos x
1387 Now simplify the 1==0 conditional:
1389 if v<0 then jtos v else jtos v
1391 Now common-up the two branches of the case:
1393 case (v<0) of DEFAULT -> jtos v
1395 Why don't we drop the case? Because it's strict in v. It's technically
1396 wrong to drop even unnecessary evaluations, and in practice they
1397 may be a result of 'seq' so we *definitely* don't want to drop those.
1398 I don't really know how to improve this situation.
1402 --------------------------------------------------
1403 -- 0. Check for empty alternatives
1404 --------------------------------------------------
1406 -- This isn't strictly an error. It's possible that the simplifer might "see"
1407 -- that an inner case has no accessible alternatives before it "sees" that the
1408 -- entire branch of an outer case is inaccessible. So we simply
1409 -- put an error case here insteadd
1410 mkCase1 scrut case_bndr ty []
1411 = pprTrace "mkCase1: null alts" (ppr case_bndr <+> ppr scrut) $
1412 return (mkApps (Var eRROR_ID)
1413 [Type ty, Lit (mkStringLit "Impossible alternative")])
1415 --------------------------------------------------
1416 -- 1. Eliminate the case altogether if poss
1417 --------------------------------------------------
1419 mkCase1 scrut case_bndr ty [(con,bndrs,rhs)]
1420 -- See if we can get rid of the case altogether
1421 -- See the extensive notes on case-elimination above
1422 -- mkCase made sure that if all the alternatives are equal,
1423 -- then there is now only one (DEFAULT) rhs
1424 | all isDeadBinder bndrs,
1426 -- Check that the scrutinee can be let-bound instead of case-bound
1427 exprOkForSpeculation scrut
1428 -- OK not to evaluate it
1429 -- This includes things like (==# a# b#)::Bool
1430 -- so that we simplify
1431 -- case ==# a# b# of { True -> x; False -> x }
1434 -- This particular example shows up in default methods for
1435 -- comparision operations (e.g. in (>=) for Int.Int32)
1436 || exprIsHNF scrut -- It's already evaluated
1437 || var_demanded_later scrut -- It'll be demanded later
1439 -- || not opt_SimplPedanticBottoms) -- Or we don't care!
1440 -- We used to allow improving termination by discarding cases, unless -fpedantic-bottoms was on,
1441 -- but that breaks badly for the dataToTag# primop, which relies on a case to evaluate
1442 -- its argument: case x of { y -> dataToTag# y }
1443 -- Here we must *not* discard the case, because dataToTag# just fetches the tag from
1444 -- the info pointer. So we'll be pedantic all the time, and see if that gives any
1446 -- Also we don't want to discard 'seq's
1447 = tick (CaseElim case_bndr) `thenSmpl_`
1448 returnSmpl (bindCaseBndr case_bndr scrut rhs)
1451 -- The case binder is going to be evaluated later,
1452 -- and the scrutinee is a simple variable
1453 var_demanded_later (Var v) = isStrictDmd (idNewDemandInfo case_bndr)
1454 var_demanded_later other = False
1457 --------------------------------------------------
1459 --------------------------------------------------
1461 mkCase1 scrut case_bndr ty alts -- Identity case
1462 | all identity_alt alts
1463 = tick (CaseIdentity case_bndr) `thenSmpl_`
1464 returnSmpl (re_cast scrut)
1466 identity_alt (con, args, rhs) = de_cast rhs `cheapEqExpr` mk_id_rhs con args
1468 mk_id_rhs (DataAlt con) args = mkConApp con (arg_tys ++ varsToCoreExprs args)
1469 mk_id_rhs (LitAlt lit) _ = Lit lit
1470 mk_id_rhs DEFAULT _ = Var case_bndr
1472 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1475 -- case e of x { _ -> x `cast` c }
1476 -- And we definitely want to eliminate this case, to give
1478 -- So we throw away the cast from the RHS, and reconstruct
1479 -- it at the other end. All the RHS casts must be the same
1480 -- if (all identity_alt alts) holds.
1482 -- Don't worry about nested casts, because the simplifier combines them
1483 de_cast (Cast e _) = e
1486 re_cast scrut = case head alts of
1487 (_,_,Cast _ co) -> Cast scrut co
1492 --------------------------------------------------
1494 --------------------------------------------------
1495 mkCase1 scrut bndr ty alts = returnSmpl (Case scrut bndr ty alts)
1499 When adding auxiliary bindings for the case binder, it's worth checking if
1500 its dead, because it often is, and occasionally these mkCase transformations
1501 cascade rather nicely.
1504 bindCaseBndr bndr rhs body
1505 | isDeadBinder bndr = body
1506 | otherwise = bindNonRec bndr rhs body