2 % (c) The AQUA Project, Glasgow University, 1993-1998
4 \section[SimplUtils]{The simplifier utilities}
9 mkLam, mkCase, prepareAlts, bindCaseBndr,
12 preInlineUnconditionally, postInlineUnconditionally,
13 activeInline, activeRule,
15 -- The continuation type
16 SimplCont(..), DupFlag(..), ArgInfo(..),
17 contIsDupable, contResultType, contIsTrivial, contArgs, dropArgs,
18 countValArgs, countArgs,
19 mkBoringStop, mkLazyArgStop, contIsRhsOrArg,
20 interestingCallContext, interestingArgContext,
22 interestingArg, mkArgInfo,
27 #include "HsVersions.h"
33 import qualified CoreSubst
40 import Var ( isCoVar )
43 import Type hiding( substTy )
44 import Coercion ( coercionKind )
46 import Unify ( dataConCannotMatch )
58 %************************************************************************
62 %************************************************************************
64 A SimplCont allows the simplifier to traverse the expression in a
65 zipper-like fashion. The SimplCont represents the rest of the expression,
66 "above" the point of interest.
68 You can also think of a SimplCont as an "evaluation context", using
69 that term in the way it is used for operational semantics. This is the
70 way I usually think of it, For example you'll often see a syntax for
71 evaluation context looking like
72 C ::= [] | C e | case C of alts | C `cast` co
73 That's the kind of thing we are doing here, and I use that syntax in
78 * A SimplCont describes a *strict* context (just like
79 evaluation contexts do). E.g. Just [] is not a SimplCont
81 * A SimplCont describes a context that *does not* bind
82 any variables. E.g. \x. [] is not a SimplCont
86 = Stop -- An empty context, or hole, []
87 CallCtxt -- True <=> There is something interesting about
88 -- the context, and hence the inliner
89 -- should be a bit keener (see interestingCallContext)
91 -- This is an argument of a function that has RULES
92 -- Inlining the call might allow the rule to fire
94 | CoerceIt -- C `cast` co
95 OutCoercion -- The coercion simplified
100 InExpr SimplEnv -- The argument and its static env
103 | Select -- case C of alts
105 InId [InAlt] SimplEnv -- The case binder, alts, and subst-env
108 -- The two strict forms have no DupFlag, because we never duplicate them
109 | StrictBind -- (\x* \xs. e) C
110 InId [InBndr] -- let x* = [] in e
111 InExpr SimplEnv -- is a special case
116 CallCtxt -- Whether *this* argument position is interesting
117 ArgInfo -- Whether the function at the head of e has rules, etc
118 SimplCont -- plus strictness flags for *further* args
122 ai_rules :: Bool, -- Function has rules (recursively)
123 -- => be keener to inline in all args
124 ai_strs :: [Bool], -- Strictness of arguments
125 -- Usually infinite, but if it is finite it guarantees
126 -- that the function diverges after being given
127 -- that number of args
128 ai_discs :: [Int] -- Discounts for arguments; non-zero => be keener to inline
132 instance Outputable SimplCont where
133 ppr (Stop interesting) = ptext (sLit "Stop") <> brackets (ppr interesting)
134 ppr (ApplyTo dup arg _ cont) = ((ptext (sLit "ApplyTo") <+> ppr dup <+> pprParendExpr arg)
135 {- $$ nest 2 (pprSimplEnv se) -}) $$ ppr cont
136 ppr (StrictBind b _ _ _ cont) = (ptext (sLit "StrictBind") <+> ppr b) $$ ppr cont
137 ppr (StrictArg f _ _ cont) = (ptext (sLit "StrictArg") <+> ppr f) $$ ppr cont
138 ppr (Select dup bndr alts _ cont) = (ptext (sLit "Select") <+> ppr dup <+> ppr bndr) $$
139 (nest 4 (ppr alts)) $$ ppr cont
140 ppr (CoerceIt co cont) = (ptext (sLit "CoerceIt") <+> ppr co) $$ ppr cont
142 data DupFlag = OkToDup | NoDup
144 instance Outputable DupFlag where
145 ppr OkToDup = ptext (sLit "ok")
146 ppr NoDup = ptext (sLit "nodup")
151 mkBoringStop :: SimplCont
152 mkBoringStop = Stop BoringCtxt
154 mkLazyArgStop :: CallCtxt -> SimplCont
155 mkLazyArgStop cci = Stop cci
158 contIsRhsOrArg :: SimplCont -> Bool
159 contIsRhsOrArg (Stop {}) = True
160 contIsRhsOrArg (StrictBind {}) = True
161 contIsRhsOrArg (StrictArg {}) = True
162 contIsRhsOrArg _ = False
165 contIsDupable :: SimplCont -> Bool
166 contIsDupable (Stop {}) = True
167 contIsDupable (ApplyTo OkToDup _ _ _) = True
168 contIsDupable (Select OkToDup _ _ _ _) = True
169 contIsDupable (CoerceIt _ cont) = contIsDupable cont
170 contIsDupable _ = False
173 contIsTrivial :: SimplCont -> Bool
174 contIsTrivial (Stop {}) = True
175 contIsTrivial (ApplyTo _ (Type _) _ cont) = contIsTrivial cont
176 contIsTrivial (CoerceIt _ cont) = contIsTrivial cont
177 contIsTrivial _ = False
180 contResultType :: SimplEnv -> OutType -> SimplCont -> OutType
181 contResultType env ty cont
184 subst_ty se ty = substTy (se `setInScope` env) ty
187 go (CoerceIt co cont) _ = go cont (snd (coercionKind co))
188 go (StrictBind _ bs body se cont) _ = go cont (subst_ty se (exprType (mkLams bs body)))
189 go (StrictArg fn _ _ cont) _ = go cont (funResultTy (exprType fn))
190 go (Select _ _ alts se cont) _ = go cont (subst_ty se (coreAltsType alts))
191 go (ApplyTo _ arg se cont) ty = go cont (apply_to_arg ty arg se)
193 apply_to_arg ty (Type ty_arg) se = applyTy ty (subst_ty se ty_arg)
194 apply_to_arg ty _ _ = funResultTy ty
197 countValArgs :: SimplCont -> Int
198 countValArgs (ApplyTo _ (Type _) _ cont) = countValArgs cont
199 countValArgs (ApplyTo _ _ _ cont) = 1 + countValArgs cont
202 countArgs :: SimplCont -> Int
203 countArgs (ApplyTo _ _ _ cont) = 1 + countArgs cont
206 contArgs :: SimplCont -> ([OutExpr], SimplCont)
207 -- Uses substitution to turn each arg into an OutExpr
208 contArgs cont = go [] cont
210 go args (ApplyTo _ arg se cont) = go (substExpr se arg : args) cont
211 go args cont = (reverse args, cont)
213 dropArgs :: Int -> SimplCont -> SimplCont
214 dropArgs 0 cont = cont
215 dropArgs n (ApplyTo _ _ _ cont) = dropArgs (n-1) cont
216 dropArgs n other = pprPanic "dropArgs" (ppr n <+> ppr other)
221 interestingArg :: OutExpr -> Bool
222 -- An argument is interesting if it has *some* structure
223 -- We are here trying to avoid unfolding a function that
224 -- is applied only to variables that have no unfolding
225 -- (i.e. they are probably lambda bound): f x y z
226 -- There is little point in inlining f here.
227 interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
228 -- Was: isValueUnfolding (idUnfolding v')
229 -- But that seems over-pessimistic
231 -- This accounts for an argument like
232 -- () or [], which is definitely interesting
233 interestingArg (Type _) = False
234 interestingArg (App fn (Type _)) = interestingArg fn
235 interestingArg (Note _ a) = interestingArg a
237 -- Idea (from Sam B); I'm not sure if it's a good idea, so commented out for now
238 -- interestingArg expr | isUnLiftedType (exprType expr)
239 -- -- Unlifted args are only ever interesting if we know what they are
244 interestingArg _ = True
245 -- Consider let x = 3 in f x
246 -- The substitution will contain (x -> ContEx 3), and we want to
247 -- to say that x is an interesting argument.
248 -- But consider also (\x. f x y) y
249 -- The substitution will contain (x -> ContEx y), and we want to say
250 -- that x is not interesting (assuming y has no unfolding)
254 Comment about interestingCallContext
255 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
256 We want to avoid inlining an expression where there can't possibly be
257 any gain, such as in an argument position. Hence, if the continuation
258 is interesting (eg. a case scrutinee, application etc.) then we
259 inline, otherwise we don't.
261 Previously some_benefit used to return True only if the variable was
262 applied to some value arguments. This didn't work:
264 let x = _coerce_ (T Int) Int (I# 3) in
265 case _coerce_ Int (T Int) x of
268 we want to inline x, but can't see that it's a constructor in a case
269 scrutinee position, and some_benefit is False.
273 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
275 .... case dMonadST _@_ x0 of (a,b,c) -> ....
277 we'd really like to inline dMonadST here, but we *don't* want to
278 inline if the case expression is just
280 case x of y { DEFAULT -> ... }
282 since we can just eliminate this case instead (x is in WHNF). Similar
283 applies when x is bound to a lambda expression. Hence
284 contIsInteresting looks for case expressions with just a single
289 interestingCallContext :: SimplCont -> CallCtxt
290 interestingCallContext cont
293 interesting (Select _ bndr _ _ _)
294 | isDeadBinder bndr = CaseCtxt
295 | otherwise = ArgCtxt False 2 -- If the binder is used, this
296 -- is like a strict let
298 interesting (ApplyTo _ arg _ cont)
299 | isTypeArg arg = interesting cont
300 | otherwise = ValAppCtxt -- Can happen if we have (f Int |> co) y
301 -- If f has an INLINE prag we need to give it some
302 -- motivation to inline. See Note [Cast then apply]
305 interesting (StrictArg _ cci _ _) = cci
306 interesting (StrictBind {}) = BoringCtxt
307 interesting (Stop cci) = cci
308 interesting (CoerceIt _ cont) = interesting cont
309 -- If this call is the arg of a strict function, the context
310 -- is a bit interesting. If we inline here, we may get useful
311 -- evaluation information to avoid repeated evals: e.g.
313 -- Here the contIsInteresting makes the '*' keener to inline,
314 -- which in turn exposes a constructor which makes the '+' inline.
315 -- Assuming that +,* aren't small enough to inline regardless.
317 -- It's also very important to inline in a strict context for things
320 -- Here, the context of (f x) is strict, and if f's unfolding is
321 -- a build it's *great* to inline it here. So we must ensure that
322 -- the context for (f x) is not totally uninteresting.
327 -> Int -- Number of value args
328 -> SimplCont -- Context of the cal
331 mkArgInfo fun n_val_args call_cont
332 | n_val_args < idArity fun -- Note [Unsaturated functions]
333 = ArgInfo { ai_rules = False
334 , ai_strs = vanilla_stricts
335 , ai_discs = vanilla_discounts }
337 = ArgInfo { ai_rules = interestingArgContext fun call_cont
338 , ai_strs = add_type_str (idType fun) arg_stricts
339 , ai_discs = arg_discounts }
341 vanilla_discounts, arg_discounts :: [Int]
342 vanilla_discounts = repeat 0
343 arg_discounts = case idUnfolding fun of
344 CoreUnfolding {uf_guidance = UnfoldIfGoodArgs {ug_args = discounts}}
345 -> discounts ++ vanilla_discounts
346 _ -> vanilla_discounts
348 vanilla_stricts, arg_stricts :: [Bool]
349 vanilla_stricts = repeat False
352 = case splitStrictSig (idNewStrictness fun) of
353 (demands, result_info)
354 | not (demands `lengthExceeds` n_val_args)
355 -> -- Enough args, use the strictness given.
356 -- For bottoming functions we used to pretend that the arg
357 -- is lazy, so that we don't treat the arg as an
358 -- interesting context. This avoids substituting
359 -- top-level bindings for (say) strings into
360 -- calls to error. But now we are more careful about
361 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
362 if isBotRes result_info then
363 map isStrictDmd demands -- Finite => result is bottom
365 map isStrictDmd demands ++ vanilla_stricts
367 -> WARN( True, text "More demands than arity" <+> ppr fun <+> ppr (idArity fun)
368 <+> ppr n_val_args <+> ppr demands )
369 vanilla_stricts -- Not enough args, or no strictness
371 add_type_str :: Type -> [Bool] -> [Bool]
372 -- If the function arg types are strict, record that in the 'strictness bits'
373 -- No need to instantiate because unboxed types (which dominate the strict
374 -- types) can't instantiate type variables.
375 -- add_type_str is done repeatedly (for each call); might be better
376 -- once-for-all in the function
377 -- But beware primops/datacons with no strictness
378 add_type_str _ [] = []
379 add_type_str fun_ty strs -- Look through foralls
380 | Just (_, fun_ty') <- splitForAllTy_maybe fun_ty -- Includes coercions
381 = add_type_str fun_ty' strs
382 add_type_str fun_ty (str:strs) -- Add strict-type info
383 | Just (arg_ty, fun_ty') <- splitFunTy_maybe fun_ty
384 = (str || isStrictType arg_ty) : add_type_str fun_ty' strs
388 {- Note [Unsaturated functions]
389 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
390 Consider (test eyeball/inline4)
393 where f has arity 2. Then we do not want to inline 'x', because
394 it'll just be floated out again. Even if f has lots of discounts
395 on its first argument -- it must be saturated for these to kick in
398 interestingArgContext :: Id -> SimplCont -> Bool
399 -- If the argument has form (f x y), where x,y are boring,
400 -- and f is marked INLINE, then we don't want to inline f.
401 -- But if the context of the argument is
403 -- where g has rules, then we *do* want to inline f, in case it
404 -- exposes a rule that might fire. Similarly, if the context is
406 -- where h has rules, then we do want to inline f; hence the
407 -- call_cont argument to interestingArgContext
409 -- The interesting_arg_ctxt flag makes this happen; if it's
410 -- set, the inliner gets just enough keener to inline f
411 -- regardless of how boring f's arguments are, if it's marked INLINE
413 -- The alternative would be to *always* inline an INLINE function,
414 -- regardless of how boring its context is; but that seems overkill
415 -- For example, it'd mean that wrapper functions were always inlined
416 interestingArgContext fn call_cont
417 = idHasRules fn || go call_cont
419 go (Select {}) = False
420 go (ApplyTo {}) = False
421 go (StrictArg _ cci _ _) = interesting cci
422 go (StrictBind {}) = False -- ??
423 go (CoerceIt _ c) = go c
424 go (Stop cci) = interesting cci
426 interesting (ArgCtxt rules _) = rules
427 interesting _ = False
432 %************************************************************************
434 \subsection{Decisions about inlining}
436 %************************************************************************
438 Inlining is controlled partly by the SimplifierMode switch. This has two
441 SimplGently (a) Simplifying before specialiser/full laziness
442 (b) Simplifiying inside INLINE pragma
443 (c) Simplifying the LHS of a rule
444 (d) Simplifying a GHCi expression or Template
447 SimplPhase n _ Used at all other times
449 The key thing about SimplGently is that it does no call-site inlining.
450 Before full laziness we must be careful not to inline wrappers,
451 because doing so inhibits floating
452 e.g. ...(case f x of ...)...
453 ==> ...(case (case x of I# x# -> fw x#) of ...)...
454 ==> ...(case x of I# x# -> case fw x# of ...)...
455 and now the redex (f x) isn't floatable any more.
457 The no-inlining thing is also important for Template Haskell. You might be
458 compiling in one-shot mode with -O2; but when TH compiles a splice before
459 running it, we don't want to use -O2. Indeed, we don't want to inline
460 anything, because the byte-code interpreter might get confused about
461 unboxed tuples and suchlike.
465 We don't simplify inside InlineRules (which come from INLINE pragmas).
466 It really is important to switch off inlinings inside such
467 expressions. Consider the following example
473 in ...g...g...g...g...g...
475 Now, if that's the ONLY occurrence of f, it will be inlined inside g,
476 and thence copied multiple times when g is inlined.
479 This function may be inlinined in other modules, so we
480 don't want to remove (by inlining) calls to functions that have
481 specialisations, or that may have transformation rules in an importing
484 E.g. {-# INLINE f #-}
487 and suppose that g is strict *and* has specialisations. If we inline
488 g's wrapper, we deny f the chance of getting the specialised version
489 of g when f is inlined at some call site (perhaps in some other
492 It's also important not to inline a worker back into a wrapper.
494 wraper = inline_me (\x -> ...worker... )
495 Normally, the inline_me prevents the worker getting inlined into
496 the wrapper (initially, the worker's only call site!). But,
497 if the wrapper is sure to be called, the strictness analyser will
498 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
499 continuation. That's why the keep_inline predicate returns True for
500 ArgOf continuations. It shouldn't do any harm not to dissolve the
501 inline-me note under these circumstances.
503 Note that the result is that we do very little simplification
506 all xs = foldr (&&) True xs
507 any p = all . map p {-# INLINE any #-}
509 Problem: any won't get deforested, and so if it's exported and the
510 importer doesn't use the inlining, (eg passes it as an arg) then we
511 won't get deforestation at all. We havn't solved this problem yet!
514 preInlineUnconditionally
515 ~~~~~~~~~~~~~~~~~~~~~~~~
516 @preInlineUnconditionally@ examines a bndr to see if it is used just
517 once in a completely safe way, so that it is safe to discard the
518 binding inline its RHS at the (unique) usage site, REGARDLESS of how
519 big the RHS might be. If this is the case we don't simplify the RHS
520 first, but just inline it un-simplified.
522 This is much better than first simplifying a perhaps-huge RHS and then
523 inlining and re-simplifying it. Indeed, it can be at least quadratically
532 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
533 This can happen with cascades of functions too:
540 THE MAIN INVARIANT is this:
542 ---- preInlineUnconditionally invariant -----
543 IF preInlineUnconditionally chooses to inline x = <rhs>
544 THEN doing the inlining should not change the occurrence
545 info for the free vars of <rhs>
546 ----------------------------------------------
548 For example, it's tempting to look at trivial binding like
550 and inline it unconditionally. But suppose x is used many times,
551 but this is the unique occurrence of y. Then inlining x would change
552 y's occurrence info, which breaks the invariant. It matters: y
553 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
556 Even RHSs labelled InlineMe aren't caught here, because there might be
557 no benefit from inlining at the call site.
559 [Sept 01] Don't unconditionally inline a top-level thing, because that
560 can simply make a static thing into something built dynamically. E.g.
564 [Remember that we treat \s as a one-shot lambda.] No point in
565 inlining x unless there is something interesting about the call site.
567 But watch out: if you aren't careful, some useful foldr/build fusion
568 can be lost (most notably in spectral/hartel/parstof) because the
569 foldr didn't see the build. Doing the dynamic allocation isn't a big
570 deal, in fact, but losing the fusion can be. But the right thing here
571 seems to be to do a callSiteInline based on the fact that there is
572 something interesting about the call site (it's strict). Hmm. That
575 Conclusion: inline top level things gaily until Phase 0 (the last
576 phase), at which point don't.
579 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
580 preInlineUnconditionally env top_lvl bndr rhs
582 | opt_SimplNoPreInlining = False
583 | otherwise = case idOccInfo bndr of
584 IAmDead -> True -- Happens in ((\x.1) v)
585 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
589 active = case phase of
590 SimplGently -> isAlwaysActive prag
591 SimplPhase n _ -> isActive n prag
592 prag = idInlinePragma bndr
594 try_once in_lam int_cxt -- There's one textual occurrence
595 | not in_lam = isNotTopLevel top_lvl || early_phase
596 | otherwise = int_cxt && canInlineInLam rhs
598 -- Be very careful before inlining inside a lambda, becuase (a) we must not
599 -- invalidate occurrence information, and (b) we want to avoid pushing a
600 -- single allocation (here) into multiple allocations (inside lambda).
601 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
602 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
604 -- is_cheap = exprIsCheap rhs
605 -- ok = is_cheap && int_cxt
607 -- int_cxt The context isn't totally boring
608 -- E.g. let f = \ab.BIG in \y. map f xs
609 -- Don't want to substitute for f, because then we allocate
610 -- its closure every time the \y is called
611 -- But: let f = \ab.BIG in \y. map (f y) xs
612 -- Now we do want to substitute for f, even though it's not
613 -- saturated, because we're going to allocate a closure for
614 -- (f y) every time round the loop anyhow.
616 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
617 -- so substituting rhs inside a lambda doesn't change the occ info.
618 -- Sadly, not quite the same as exprIsHNF.
619 canInlineInLam (Lit _) = True
620 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
621 canInlineInLam (Note _ e) = canInlineInLam e
622 canInlineInLam _ = False
624 early_phase = case phase of
625 SimplPhase 0 _ -> False
627 -- If we don't have this early_phase test, consider
628 -- x = length [1,2,3]
629 -- The full laziness pass carefully floats all the cons cells to
630 -- top level, and preInlineUnconditionally floats them all back in.
631 -- Result is (a) static allocation replaced by dynamic allocation
632 -- (b) many simplifier iterations because this tickles
633 -- a related problem; only one inlining per pass
635 -- On the other hand, I have seen cases where top-level fusion is
636 -- lost if we don't inline top level thing (e.g. string constants)
637 -- Hence the test for phase zero (which is the phase for all the final
638 -- simplifications). Until phase zero we take no special notice of
639 -- top level things, but then we become more leery about inlining
644 postInlineUnconditionally
645 ~~~~~~~~~~~~~~~~~~~~~~~~~
646 @postInlineUnconditionally@ decides whether to unconditionally inline
647 a thing based on the form of its RHS; in particular if it has a
648 trivial RHS. If so, we can inline and discard the binding altogether.
650 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
651 only have *forward* references Hence, it's safe to discard the binding
653 NOTE: This isn't our last opportunity to inline. We're at the binding
654 site right now, and we'll get another opportunity when we get to the
657 Note that we do this unconditional inlining only for trival RHSs.
658 Don't inline even WHNFs inside lambdas; doing so may simply increase
659 allocation when the function is called. This isn't the last chance; see
662 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
663 Because we don't even want to inline them into the RHS of constructor
664 arguments. See NOTE above
666 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
667 it's best to inline it anyway. We often get a=E; b=a from desugaring,
668 with both a and b marked NOINLINE. But that seems incompatible with
669 our new view that inlining is like a RULE, so I'm sticking to the 'active'
673 postInlineUnconditionally
674 :: SimplEnv -> TopLevelFlag
675 -> InId -- The binder (an OutId would be fine too)
676 -> OccInfo -- From the InId
680 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
682 | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, don't inline
683 -- because it might be referred to "earlier"
684 | isExportedId bndr = False
685 | exprIsTrivial rhs = True
688 -- The point of examining occ_info here is that for *non-values*
689 -- that occur outside a lambda, the call-site inliner won't have
690 -- a chance (becuase it doesn't know that the thing
691 -- only occurs once). The pre-inliner won't have gotten
692 -- it either, if the thing occurs in more than one branch
693 -- So the main target is things like
696 -- True -> case x of ...
697 -- False -> case x of ...
698 -- I'm not sure how important this is in practice
699 OneOcc in_lam _one_br int_cxt -- OneOcc => no code-duplication issue
700 -> smallEnoughToInline unfolding -- Small enough to dup
701 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
703 -- NB: Do NOT inline arbitrarily big things, even if one_br is True
704 -- Reason: doing so risks exponential behaviour. We simplify a big
705 -- expression, inline it, and simplify it again. But if the
706 -- very same thing happens in the big expression, we get
708 -- PRINCIPLE: when we've already simplified an expression once,
709 -- make sure that we only inline it if it's reasonably small.
711 && ((isNotTopLevel top_lvl && not in_lam) ||
712 -- But outside a lambda, we want to be reasonably aggressive
713 -- about inlining into multiple branches of case
714 -- e.g. let x = <non-value>
715 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
716 -- Inlining can be a big win if C3 is the hot-spot, even if
717 -- the uses in C1, C2 are not 'interesting'
718 -- An example that gets worse if you add int_cxt here is 'clausify'
720 (isCheapUnfolding unfolding && int_cxt))
721 -- isCheap => acceptable work duplication; in_lam may be true
722 -- int_cxt to prevent us inlining inside a lambda without some
723 -- good reason. See the notes on int_cxt in preInlineUnconditionally
725 IAmDead -> True -- This happens; for example, the case_bndr during case of
726 -- known constructor: case (a,b) of x { (p,q) -> ... }
727 -- Here x isn't mentioned in the RHS, so we don't want to
728 -- create the (dead) let-binding let x = (a,b) in ...
732 -- Here's an example that we don't handle well:
733 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
734 -- in \y. ....case f of {...} ....
735 -- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
737 -- - We can't preInlineUnconditionally because that woud invalidate
738 -- the occ info for b.
739 -- - We can't postInlineUnconditionally because the RHS is big, and
740 -- that risks exponential behaviour
741 -- - We can't call-site inline, because the rhs is big
745 active = case getMode env of
746 SimplGently -> isAlwaysActive prag
747 SimplPhase n _ -> isActive n prag
748 prag = idInlinePragma bndr
750 activeInline :: SimplEnv -> OutId -> Bool
752 = case getMode env of
754 -- No inlining at all when doing gentle stuff,
755 -- except for local things that occur once (pre/postInlineUnconditionally)
756 -- The reason is that too little clean-up happens if you
757 -- don't inline use-once things. Also a bit of inlining is *good* for
758 -- full laziness; it can expose constant sub-expressions.
759 -- Example in spectral/mandel/Mandel.hs, where the mandelset
760 -- function gets a useful let-float if you inline windowToViewport
762 -- NB: we used to have a second exception, for data con wrappers.
763 -- On the grounds that we use gentle mode for rule LHSs, and
764 -- they match better when data con wrappers are inlined.
765 -- But that only really applies to the trivial wrappers (like (:)),
766 -- and they are now constructed as Compulsory unfoldings (in MkId)
767 -- so they'll happen anyway.
769 SimplPhase n _ -> isActive n prag
771 prag = idInlinePragma id
773 activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
774 -- Nothing => No rules at all
775 activeRule dflags env
776 | not (dopt Opt_EnableRewriteRules dflags)
777 = Nothing -- Rewriting is off
779 = case getMode env of
780 SimplGently -> Just isAlwaysActive
781 -- Used to be Nothing (no rules in gentle mode)
782 -- Main motivation for changing is that I wanted
783 -- lift String ===> ...
784 -- to work in Template Haskell when simplifying
785 -- splices, so we get simpler code for literal strings
786 SimplPhase n _ -> Just (isActive n)
790 %************************************************************************
794 %************************************************************************
797 mkLam :: [OutBndr] -> OutExpr -> SimplM OutExpr
798 -- mkLam tries three things
799 -- a) eta reduction, if that gives a trivial expression
800 -- b) eta expansion [only if there are some value lambdas]
805 = do { dflags <- getDOptsSmpl
806 ; mkLam' dflags bndrs body }
808 mkLam' :: DynFlags -> [OutBndr] -> OutExpr -> SimplM OutExpr
809 mkLam' dflags bndrs (Cast body co)
810 | not (any bad bndrs)
811 -- Note [Casts and lambdas]
812 = do { lam <- mkLam' dflags bndrs body
813 ; return (mkCoerce (mkPiTypes bndrs co) lam) }
815 co_vars = tyVarsOfType co
816 bad bndr = isCoVar bndr && bndr `elemVarSet` co_vars
818 mkLam' dflags bndrs body
819 | dopt Opt_DoEtaReduction dflags,
820 Just etad_lam <- tryEtaReduce bndrs body
821 = do { tick (EtaReduction (head bndrs))
824 | dopt Opt_DoLambdaEtaExpansion dflags,
825 any isRuntimeVar bndrs
826 = do { body' <- tryEtaExpansion dflags body
827 ; return (mkLams bndrs body') }
830 = return (mkLams bndrs body)
833 Note [Casts and lambdas]
834 ~~~~~~~~~~~~~~~~~~~~~~~~
836 (\x. (\y. e) `cast` g1) `cast` g2
837 There is a danger here that the two lambdas look separated, and the
838 full laziness pass might float an expression to between the two.
840 So this equation in mkLam' floats the g1 out, thus:
841 (\x. e `cast` g1) --> (\x.e) `cast` (tx -> g1)
844 In general, this floats casts outside lambdas, where (I hope) they
845 might meet and cancel with some other cast:
846 \x. e `cast` co ===> (\x. e) `cast` (tx -> co)
847 /\a. e `cast` co ===> (/\a. e) `cast` (/\a. co)
848 /\g. e `cast` co ===> (/\g. e) `cast` (/\g. co)
851 Notice that it works regardless of 'e'. Originally it worked only
852 if 'e' was itself a lambda, but in some cases that resulted in
853 fruitless iteration in the simplifier. A good example was when
854 compiling Text.ParserCombinators.ReadPrec, where we had a definition
855 like (\x. Get `cast` g)
856 where Get is a constructor with nonzero arity. Then mkLam eta-expanded
857 the Get, and the next iteration eta-reduced it, and then eta-expanded
860 Note also the side condition for the case of coercion binders.
861 It does not make sense to transform
862 /\g. e `cast` g ==> (/\g.e) `cast` (/\g.g)
863 because the latter is not well-kinded.
865 -- c) floating lets out through big lambdas
866 -- [only if all tyvar lambdas, and only if this lambda
867 -- is the RHS of a let]
869 {- Sept 01: I'm experimenting with getting the
870 full laziness pass to float out past big lambdsa
871 | all isTyVar bndrs, -- Only for big lambdas
872 contIsRhs cont -- Only try the rhs type-lambda floating
873 -- if this is indeed a right-hand side; otherwise
874 -- we end up floating the thing out, only for float-in
875 -- to float it right back in again!
876 = do (floats, body') <- tryRhsTyLam env bndrs body
877 return (floats, mkLams bndrs body')
881 %************************************************************************
885 %************************************************************************
887 Note [Eta reduction conditions]
888 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
889 We try for eta reduction here, but *only* if we get all the way to an
890 trivial expression. We don't want to remove extra lambdas unless we
891 are going to avoid allocating this thing altogether.
893 There are some particularly delicate points here:
895 * Eta reduction is not valid in general:
897 This matters, partly for old-fashioned correctness reasons but,
898 worse, getting it wrong can yield a seg fault. Consider
900 h y = case (case y of { True -> f `seq` True; False -> False }) of
901 True -> ...; False -> ...
903 If we (unsoundly) eta-reduce f to get f=f, the strictness analyser
904 says f=bottom, and replaces the (f `seq` True) with just
905 (f `cast` unsafe-co). BUT, as thing stand, 'f' got arity 1, and it
906 *keeps* arity 1 (perhaps also wrongly). So CorePrep eta-expands
907 the definition again, so that it does not termninate after all.
908 Result: seg-fault because the boolean case actually gets a function value.
911 So it's important to to the right thing.
913 * Note [Arity care]: we need to be careful if we just look at f's
914 arity. Currently (Dec07), f's arity is visible in its own RHS (see
915 Note [Arity robustness] in SimplEnv) so we must *not* trust the
916 arity when checking that 'f' is a value. Otherwise we will
921 Which might change a terminiating program (think (f `seq` e)) to a
922 non-terminating one. So we check for being a loop breaker first.
924 However for GlobalIds we can look at the arity; and for primops we
925 must, since they have no unfolding.
927 * Regardless of whether 'f' is a value, we always want to
928 reduce (/\a -> f a) to f
929 This came up in a RULE: foldr (build (/\a -> g a))
930 did not match foldr (build (/\b -> ...something complex...))
931 The type checker can insert these eta-expanded versions,
932 with both type and dictionary lambdas; hence the slightly
935 * Never *reduce* arity. For example
937 Then if h has arity 1 we don't want to eta-reduce because then
938 f's arity would decrease, and that is bad
940 These delicacies are why we don't use exprIsTrivial and exprIsHNF here.
944 tryEtaReduce :: [OutBndr] -> OutExpr -> Maybe OutExpr
945 tryEtaReduce bndrs body
946 = go (reverse bndrs) body
948 incoming_arity = count isId bndrs
950 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
951 go [] fun | ok_fun fun = Just fun -- Success!
952 go _ _ = Nothing -- Failure!
954 -- Note [Eta reduction conditions]
955 ok_fun (App fun (Type ty))
956 | not (any (`elemVarSet` tyVarsOfType ty) bndrs)
959 = not (fun_id `elem` bndrs)
960 && (ok_fun_id fun_id || all ok_lam bndrs)
963 ok_fun_id fun = fun_arity fun >= incoming_arity
965 fun_arity fun -- See Note [Arity care]
966 | isLocalId fun && isLoopBreaker (idOccInfo fun) = 0
967 | otherwise = idArity fun
969 ok_lam v = isTyVar v || isDictId v
971 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
975 %************************************************************************
979 %************************************************************************
983 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
986 where (in both cases)
988 * The xi can include type variables
990 * The yi are all value variables
992 * N is a NORMAL FORM (i.e. no redexes anywhere)
993 wanting a suitable number of extra args.
995 The biggest reason for doing this is for cases like
1001 Here we want to get the lambdas together. A good exmaple is the nofib
1002 program fibheaps, which gets 25% more allocation if you don't do this
1005 We may have to sandwich some coerces between the lambdas
1006 to make the types work. exprEtaExpandArity looks through coerces
1007 when computing arity; and etaExpand adds the coerces as necessary when
1008 actually computing the expansion.
1011 tryEtaExpansion :: DynFlags -> OutExpr -> SimplM OutExpr
1012 -- There is at least one runtime binder in the binders
1013 tryEtaExpansion dflags body = do
1015 return (etaExpand fun_arity us body (exprType body))
1017 fun_arity = exprEtaExpandArity dflags body
1021 %************************************************************************
1023 \subsection{Floating lets out of big lambdas}
1025 %************************************************************************
1027 Note [Floating and type abstraction]
1028 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1031 We'd like to float this to
1034 x = /\a. C (y1 a) (y2 a)
1035 for the usual reasons: we want to inline x rather vigorously.
1037 You may think that this kind of thing is rare. But in some programs it is
1038 common. For example, if you do closure conversion you might get:
1040 data a :-> b = forall e. (e -> a -> b) :$ e
1042 f_cc :: forall a. a :-> a
1043 f_cc = /\a. (\e. id a) :$ ()
1045 Now we really want to inline that f_cc thing so that the
1046 construction of the closure goes away.
1048 So I have elaborated simplLazyBind to understand right-hand sides that look
1052 and treat them specially. The real work is done in SimplUtils.abstractFloats,
1053 but there is quite a bit of plumbing in simplLazyBind as well.
1055 The same transformation is good when there are lets in the body:
1057 /\abc -> let(rec) x = e in b
1059 let(rec) x' = /\abc -> let x = x' a b c in e
1061 /\abc -> let x = x' a b c in b
1063 This is good because it can turn things like:
1065 let f = /\a -> letrec g = ... g ... in g
1067 letrec g' = /\a -> ... g' a ...
1069 let f = /\ a -> g' a
1071 which is better. In effect, it means that big lambdas don't impede
1074 This optimisation is CRUCIAL in eliminating the junk introduced by
1075 desugaring mutually recursive definitions. Don't eliminate it lightly!
1077 [May 1999] If we do this transformation *regardless* then we can
1078 end up with some pretty silly stuff. For example,
1081 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
1086 st = /\s -> ...[y1 s/x1, y2 s/x2]
1089 Unless the "..." is a WHNF there is really no point in doing this.
1090 Indeed it can make things worse. Suppose x1 is used strictly,
1093 x1* = case f y of { (a,b) -> e }
1095 If we abstract this wrt the tyvar we then can't do the case inline
1096 as we would normally do.
1098 That's why the whole transformation is part of the same process that
1099 floats let-bindings and constructor arguments out of RHSs. In particular,
1100 it is guarded by the doFloatFromRhs call in simplLazyBind.
1104 abstractFloats :: [OutTyVar] -> SimplEnv -> OutExpr -> SimplM ([OutBind], OutExpr)
1105 abstractFloats main_tvs body_env body
1106 = ASSERT( notNull body_floats )
1107 do { (subst, float_binds) <- mapAccumLM abstract empty_subst body_floats
1108 ; return (float_binds, CoreSubst.substExpr subst body) }
1110 main_tv_set = mkVarSet main_tvs
1111 body_floats = getFloats body_env
1112 empty_subst = CoreSubst.mkEmptySubst (seInScope body_env)
1114 abstract :: CoreSubst.Subst -> OutBind -> SimplM (CoreSubst.Subst, OutBind)
1115 abstract subst (NonRec id rhs)
1116 = do { (poly_id, poly_app) <- mk_poly tvs_here id
1117 ; let poly_rhs = mkLams tvs_here rhs'
1118 subst' = CoreSubst.extendIdSubst subst id poly_app
1119 ; return (subst', (NonRec poly_id poly_rhs)) }
1121 rhs' = CoreSubst.substExpr subst rhs
1122 tvs_here | any isCoVar main_tvs = main_tvs -- Note [Abstract over coercions]
1124 = varSetElems (main_tv_set `intersectVarSet` exprSomeFreeVars isTyVar rhs')
1126 -- Abstract only over the type variables free in the rhs
1127 -- wrt which the new binding is abstracted. But the naive
1128 -- approach of abstract wrt the tyvars free in the Id's type
1130 -- /\ a b -> let t :: (a,b) = (e1, e2)
1133 -- Here, b isn't free in x's type, but we must nevertheless
1134 -- abstract wrt b as well, because t's type mentions b.
1135 -- Since t is floated too, we'd end up with the bogus:
1136 -- poly_t = /\ a b -> (e1, e2)
1137 -- poly_x = /\ a -> fst (poly_t a *b*)
1138 -- So for now we adopt the even more naive approach of
1139 -- abstracting wrt *all* the tyvars. We'll see if that
1140 -- gives rise to problems. SLPJ June 98
1142 abstract subst (Rec prs)
1143 = do { (poly_ids, poly_apps) <- mapAndUnzipM (mk_poly tvs_here) ids
1144 ; let subst' = CoreSubst.extendSubstList subst (ids `zip` poly_apps)
1145 poly_rhss = [mkLams tvs_here (CoreSubst.substExpr subst' rhs) | rhs <- rhss]
1146 ; return (subst', Rec (poly_ids `zip` poly_rhss)) }
1148 (ids,rhss) = unzip prs
1149 -- For a recursive group, it's a bit of a pain to work out the minimal
1150 -- set of tyvars over which to abstract:
1151 -- /\ a b c. let x = ...a... in
1152 -- letrec { p = ...x...q...
1153 -- q = .....p...b... } in
1155 -- Since 'x' is abstracted over 'a', the {p,q} group must be abstracted
1156 -- over 'a' (because x is replaced by (poly_x a)) as well as 'b'.
1157 -- Since it's a pain, we just use the whole set, which is always safe
1159 -- If you ever want to be more selective, remember this bizarre case too:
1161 -- Here, we must abstract 'x' over 'a'.
1164 mk_poly tvs_here var
1165 = do { uniq <- getUniqueM
1166 ; let poly_name = setNameUnique (idName var) uniq -- Keep same name
1167 poly_ty = mkForAllTys tvs_here (idType var) -- But new type of course
1168 poly_id = transferPolyIdInfo var $ -- Note [transferPolyIdInfo] in Id.lhs
1169 mkLocalId poly_name poly_ty
1170 ; return (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tvs_here)) }
1171 -- In the olden days, it was crucial to copy the occInfo of the original var,
1172 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1173 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1174 -- at already simplified code, so it doesn't matter
1176 -- It's even right to retain single-occurrence or dead-var info:
1177 -- Suppose we started with /\a -> let x = E in B
1178 -- where x occurs once in B. Then we transform to:
1179 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1180 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1181 -- the occurrences of x' will be just the occurrences originally
1185 Note [Abstract over coercions]
1186 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1187 If a coercion variable (g :: a ~ Int) is free in the RHS, then so is the
1188 type variable a. Rather than sort this mess out, we simply bale out and abstract
1189 wrt all the type variables if any of them are coercion variables.
1192 Historical note: if you use let-bindings instead of a substitution, beware of this:
1194 -- Suppose we start with:
1196 -- x = /\ a -> let g = G in E
1198 -- Then we'll float to get
1200 -- x = let poly_g = /\ a -> G
1201 -- in /\ a -> let g = poly_g a in E
1203 -- But now the occurrence analyser will see just one occurrence
1204 -- of poly_g, not inside a lambda, so the simplifier will
1205 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1206 -- (I used to think that the "don't inline lone occurrences" stuff
1207 -- would stop this happening, but since it's the *only* occurrence,
1208 -- PreInlineUnconditionally kicks in first!)
1210 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1211 -- to appear many times. (NB: mkInlineMe eliminates
1212 -- such notes on trivial RHSs, so do it manually.)
1214 %************************************************************************
1218 %************************************************************************
1220 prepareAlts tries these things:
1222 1. If several alternatives are identical, merge them into
1223 a single DEFAULT alternative. I've occasionally seen this
1224 making a big difference:
1226 case e of =====> case e of
1227 C _ -> f x D v -> ....v....
1228 D v -> ....v.... DEFAULT -> f x
1231 The point is that we merge common RHSs, at least for the DEFAULT case.
1232 [One could do something more elaborate but I've never seen it needed.]
1233 To avoid an expensive test, we just merge branches equal to the *first*
1234 alternative; this picks up the common cases
1235 a) all branches equal
1236 b) some branches equal to the DEFAULT (which occurs first)
1239 case e of b { ==> case e of b {
1240 p1 -> rhs1 p1 -> rhs1
1242 pm -> rhsm pm -> rhsm
1243 _ -> case b of b' { pn -> let b'=b in rhsn
1245 ... po -> let b'=b in rhso
1246 po -> rhso _ -> let b'=b in rhsd
1250 which merges two cases in one case when -- the default alternative of
1251 the outer case scrutises the same variable as the outer case This
1252 transformation is called Case Merging. It avoids that the same
1253 variable is scrutinised multiple times.
1256 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
1262 where @is@ was something like
1264 p `is` n = p /= (-1) && p == n
1266 This gave rise to a horrible sequence of cases
1273 and similarly in cascade for all the join points!
1276 ~~~~~~~~~~~~~~~~~~~~
1277 We do this *here*, looking at un-simplified alternatives, because we
1278 have to check that r doesn't mention the variables bound by the
1279 pattern in each alternative, so the binder-info is rather useful.
1282 prepareAlts :: SimplEnv -> OutExpr -> OutId -> [InAlt] -> SimplM ([AltCon], [InAlt])
1283 prepareAlts env scrut case_bndr' alts
1284 = do { dflags <- getDOptsSmpl
1285 ; alts <- combineIdenticalAlts case_bndr' alts
1287 ; let (alts_wo_default, maybe_deflt) = findDefault alts
1288 alt_cons = [con | (con,_,_) <- alts_wo_default]
1289 imposs_deflt_cons = nub (imposs_cons ++ alt_cons)
1290 -- "imposs_deflt_cons" are handled
1291 -- EITHER by the context,
1292 -- OR by a non-DEFAULT branch in this case expression.
1294 ; default_alts <- prepareDefault dflags env case_bndr' mb_tc_app
1295 imposs_deflt_cons maybe_deflt
1297 ; let trimmed_alts = filterOut impossible_alt alts_wo_default
1298 merged_alts = mergeAlts trimmed_alts default_alts
1299 -- We need the mergeAlts in case the new default_alt
1300 -- has turned into a constructor alternative.
1301 -- The merge keeps the inner DEFAULT at the front, if there is one
1302 -- and interleaves the alternatives in the right order
1304 ; return (imposs_deflt_cons, merged_alts) }
1306 mb_tc_app = splitTyConApp_maybe (idType case_bndr')
1307 Just (_, inst_tys) = mb_tc_app
1309 imposs_cons = case scrut of
1310 Var v -> otherCons (idUnfolding v)
1313 impossible_alt :: CoreAlt -> Bool
1314 impossible_alt (con, _, _) | con `elem` imposs_cons = True
1315 impossible_alt (DataAlt con, _, _) = dataConCannotMatch inst_tys con
1316 impossible_alt _ = False
1319 --------------------------------------------------
1320 -- 1. Merge identical branches
1321 --------------------------------------------------
1322 combineIdenticalAlts :: OutId -> [InAlt] -> SimplM [InAlt]
1324 combineIdenticalAlts case_bndr ((_con1,bndrs1,rhs1) : con_alts)
1325 | all isDeadBinder bndrs1, -- Remember the default
1326 length filtered_alts < length con_alts -- alternative comes first
1327 -- Also Note [Dead binders]
1328 = do { tick (AltMerge case_bndr)
1329 ; return ((DEFAULT, [], rhs1) : filtered_alts) }
1331 filtered_alts = filter keep con_alts
1332 keep (_con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1334 combineIdenticalAlts _ alts = return alts
1336 -------------------------------------------------------------------------
1337 -- Prepare the default alternative
1338 -------------------------------------------------------------------------
1339 prepareDefault :: DynFlags
1341 -> OutId -- Case binder; need just for its type. Note that as an
1342 -- OutId, it has maximum information; this is important.
1343 -- Test simpl013 is an example
1344 -> Maybe (TyCon, [Type]) -- Type of scrutinee, decomposed
1345 -> [AltCon] -- These cons can't happen when matching the default
1346 -> Maybe InExpr -- Rhs
1347 -> SimplM [InAlt] -- Still unsimplified
1348 -- We use a list because it's what mergeAlts expects,
1349 -- And becuase case-merging can cause many to show up
1351 ------- Merge nested cases ----------
1352 prepareDefault dflags env outer_bndr _bndr_ty imposs_cons (Just deflt_rhs)
1353 | dopt Opt_CaseMerge dflags
1354 , Case (Var inner_scrut_var) inner_bndr _ inner_alts <- deflt_rhs
1355 , DoneId inner_scrut_var' <- substId env inner_scrut_var
1356 -- Remember, inner_scrut_var is an InId, but outer_bndr is an OutId
1357 , inner_scrut_var' == outer_bndr
1358 -- NB: the substId means that if the outer scrutinee was a
1359 -- variable, and inner scrutinee is the same variable,
1360 -- then inner_scrut_var' will be outer_bndr
1361 -- via the magic of simplCaseBinder
1362 = do { tick (CaseMerge outer_bndr)
1364 ; let munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1365 ; return [(con, args, munge_rhs rhs) | (con, args, rhs) <- inner_alts,
1366 not (con `elem` imposs_cons) ]
1367 -- NB: filter out any imposs_cons. Example:
1370 -- DEFAULT -> case x of
1373 -- When we merge, we must ensure that e1 takes
1374 -- precedence over e2 as the value for A!
1376 -- Warning: don't call prepareAlts recursively!
1377 -- Firstly, there's no point, because inner alts have already had
1378 -- mkCase applied to them, so they won't have a case in their default
1379 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1380 -- in munge_rhs may put a case into the DEFAULT branch!
1383 --------- Fill in known constructor -----------
1384 prepareDefault _ _ case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
1385 | -- This branch handles the case where we are
1386 -- scrutinisng an algebraic data type
1387 isAlgTyCon tycon -- It's a data type, tuple, or unboxed tuples.
1388 , not (isNewTyCon tycon) -- We can have a newtype, if we are just doing an eval:
1389 -- case x of { DEFAULT -> e }
1390 -- and we don't want to fill in a default for them!
1391 , Just all_cons <- tyConDataCons_maybe tycon
1392 , not (null all_cons) -- This is a tricky corner case. If the data type has no constructors,
1393 -- which GHC allows, then the case expression will have at most a default
1394 -- alternative. We don't want to eliminate that alternative, because the
1395 -- invariant is that there's always one alternative. It's more convenient
1397 -- case x of { DEFAULT -> e }
1398 -- as it is, rather than transform it to
1399 -- error "case cant match"
1400 -- which would be quite legitmate. But it's a really obscure corner, and
1401 -- not worth wasting code on.
1402 , let imposs_data_cons = [con | DataAlt con <- imposs_cons] -- We now know it's a data type
1403 impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
1404 = case filterOut impossible all_cons of
1405 [] -> return [] -- Eliminate the default alternative
1406 -- altogether if it can't match
1408 [con] -> -- It matches exactly one constructor, so fill it in
1409 do { tick (FillInCaseDefault case_bndr)
1411 ; let (ex_tvs, co_tvs, arg_ids) =
1412 dataConRepInstPat us con inst_tys
1413 ; return [(DataAlt con, ex_tvs ++ co_tvs ++ arg_ids, deflt_rhs)] }
1415 _ -> return [(DEFAULT, [], deflt_rhs)]
1417 | debugIsOn, isAlgTyCon tycon, not (isOpenTyCon tycon), null (tyConDataCons tycon)
1418 -- This can legitimately happen for type families, so don't report that
1419 = pprTrace "prepareDefault" (ppr case_bndr <+> ppr tycon)
1420 $ return [(DEFAULT, [], deflt_rhs)]
1422 --------- Catch-all cases -----------
1423 prepareDefault _dflags _env _case_bndr _bndr_ty _imposs_cons (Just deflt_rhs)
1424 = return [(DEFAULT, [], deflt_rhs)]
1426 prepareDefault _dflags _env _case_bndr _bndr_ty _imposs_cons Nothing
1427 = return [] -- No default branch
1432 =================================================================================
1434 mkCase tries these things
1436 1. Eliminate the case altogether if possible
1444 and similar friends.
1448 mkCase :: OutExpr -> OutId -> [OutAlt] -- Increasing order
1451 --------------------------------------------------
1453 --------------------------------------------------
1455 mkCase scrut case_bndr alts -- Identity case
1456 | all identity_alt alts
1457 = do tick (CaseIdentity case_bndr)
1458 return (re_cast scrut)
1460 identity_alt (con, args, rhs) = check_eq con args (de_cast rhs)
1462 check_eq DEFAULT _ (Var v) = v == case_bndr
1463 check_eq (LitAlt lit') _ (Lit lit) = lit == lit'
1464 check_eq (DataAlt con) args rhs = rhs `cheapEqExpr` mkConApp con (arg_tys ++ varsToCoreExprs args)
1465 || rhs `cheapEqExpr` Var case_bndr
1466 check_eq _ _ _ = False
1468 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1471 -- case e of x { _ -> x `cast` c }
1472 -- And we definitely want to eliminate this case, to give
1474 -- So we throw away the cast from the RHS, and reconstruct
1475 -- it at the other end. All the RHS casts must be the same
1476 -- if (all identity_alt alts) holds.
1478 -- Don't worry about nested casts, because the simplifier combines them
1479 de_cast (Cast e _) = e
1482 re_cast scrut = case head alts of
1483 (_,_,Cast _ co) -> Cast scrut co
1488 --------------------------------------------------
1490 --------------------------------------------------
1491 mkCase scrut bndr alts = return (Case scrut bndr (coreAltsType alts) alts)
1495 When adding auxiliary bindings for the case binder, it's worth checking if
1496 its dead, because it often is, and occasionally these mkCase transformations
1497 cascade rather nicely.
1500 bindCaseBndr :: Id -> CoreExpr -> CoreExpr -> CoreExpr
1501 bindCaseBndr bndr rhs body
1502 | isDeadBinder bndr = body
1503 | otherwise = bindNonRec bndr rhs body