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,
22 interestingArg, mkArgInfo,
27 #include "HsVersions.h"
33 import qualified CoreSubst
37 import CoreArity ( etaExpand, exprEtaExpandArity )
41 import Var ( isCoVar )
44 import Type hiding( substTy )
45 import Coercion ( coercionKind )
47 import Unify ( dataConCannotMatch )
59 %************************************************************************
63 %************************************************************************
65 A SimplCont allows the simplifier to traverse the expression in a
66 zipper-like fashion. The SimplCont represents the rest of the expression,
67 "above" the point of interest.
69 You can also think of a SimplCont as an "evaluation context", using
70 that term in the way it is used for operational semantics. This is the
71 way I usually think of it, For example you'll often see a syntax for
72 evaluation context looking like
73 C ::= [] | C e | case C of alts | C `cast` co
74 That's the kind of thing we are doing here, and I use that syntax in
79 * A SimplCont describes a *strict* context (just like
80 evaluation contexts do). E.g. Just [] is not a SimplCont
82 * A SimplCont describes a context that *does not* bind
83 any variables. E.g. \x. [] is not a SimplCont
87 = Stop -- An empty context, or hole, []
88 CallCtxt -- True <=> There is something interesting about
89 -- the context, and hence the inliner
90 -- should be a bit keener (see interestingCallContext)
92 -- This is an argument of a function that has RULES
93 -- Inlining the call might allow the rule to fire
95 | CoerceIt -- C `cast` co
96 OutCoercion -- The coercion simplified
101 InExpr SimplEnv -- The argument and its static env
104 | Select -- case C of alts
106 InId [InAlt] SimplEnv -- The case binder, alts, and subst-env
109 -- The two strict forms have no DupFlag, because we never duplicate them
110 | StrictBind -- (\x* \xs. e) C
111 InId [InBndr] -- let x* = [] in e
112 InExpr SimplEnv -- is a special case
116 OutExpr -- e; *always* of form (Var v `App1` e1 .. `App` en)
117 CallCtxt -- Whether *this* argument position is interesting
118 ArgInfo -- Whether the function at the head of e has rules, etc
119 SimplCont -- plus strictness flags for *further* args
123 ai_rules :: Bool, -- Function has rules (recursively)
124 -- => be keener to inline in all args
125 ai_strs :: [Bool], -- Strictness of arguments
126 -- Usually infinite, but if it is finite it guarantees
127 -- that the function diverges after being given
128 -- that number of args
129 ai_discs :: [Int] -- Discounts for arguments; non-zero => be keener to inline
133 instance Outputable SimplCont where
134 ppr (Stop interesting) = ptext (sLit "Stop") <> brackets (ppr interesting)
135 ppr (ApplyTo dup arg _ cont) = ((ptext (sLit "ApplyTo") <+> ppr dup <+> pprParendExpr arg)
136 {- $$ nest 2 (pprSimplEnv se) -}) $$ ppr cont
137 ppr (StrictBind b _ _ _ cont) = (ptext (sLit "StrictBind") <+> ppr b) $$ ppr cont
138 ppr (StrictArg f _ _ cont) = (ptext (sLit "StrictArg") <+> ppr f) $$ ppr cont
139 ppr (Select dup bndr alts _ cont) = (ptext (sLit "Select") <+> ppr dup <+> ppr bndr) $$
140 (nest 4 (ppr alts)) $$ ppr cont
141 ppr (CoerceIt co cont) = (ptext (sLit "CoerceIt") <+> ppr co) $$ ppr cont
143 data DupFlag = OkToDup | NoDup
145 instance Outputable DupFlag where
146 ppr OkToDup = ptext (sLit "ok")
147 ppr NoDup = ptext (sLit "nodup")
152 mkBoringStop :: SimplCont
153 mkBoringStop = Stop BoringCtxt
155 mkLazyArgStop :: CallCtxt -> SimplCont
156 mkLazyArgStop cci = Stop cci
159 contIsRhsOrArg :: SimplCont -> Bool
160 contIsRhsOrArg (Stop {}) = True
161 contIsRhsOrArg (StrictBind {}) = True
162 contIsRhsOrArg (StrictArg {}) = True
163 contIsRhsOrArg _ = False
166 contIsDupable :: SimplCont -> Bool
167 contIsDupable (Stop {}) = True
168 contIsDupable (ApplyTo OkToDup _ _ _) = True
169 contIsDupable (Select OkToDup _ _ _ _) = True
170 contIsDupable (CoerceIt _ cont) = contIsDupable cont
171 contIsDupable _ = False
174 contIsTrivial :: SimplCont -> Bool
175 contIsTrivial (Stop {}) = True
176 contIsTrivial (ApplyTo _ (Type _) _ cont) = contIsTrivial cont
177 contIsTrivial (CoerceIt _ cont) = contIsTrivial cont
178 contIsTrivial _ = False
181 contResultType :: SimplEnv -> OutType -> SimplCont -> OutType
182 contResultType env ty cont
185 subst_ty se ty = substTy (se `setInScope` env) ty
188 go (CoerceIt co cont) _ = go cont (snd (coercionKind co))
189 go (StrictBind _ bs body se cont) _ = go cont (subst_ty se (exprType (mkLams bs body)))
190 go (StrictArg fn _ _ cont) _ = go cont (funResultTy (exprType fn))
191 go (Select _ _ alts se cont) _ = go cont (subst_ty se (coreAltsType alts))
192 go (ApplyTo _ arg se cont) ty = go cont (apply_to_arg ty arg se)
194 apply_to_arg ty (Type ty_arg) se = applyTy ty (subst_ty se ty_arg)
195 apply_to_arg ty _ _ = funResultTy ty
198 countValArgs :: SimplCont -> Int
199 countValArgs (ApplyTo _ (Type _) _ cont) = countValArgs cont
200 countValArgs (ApplyTo _ _ _ cont) = 1 + countValArgs cont
203 countArgs :: SimplCont -> Int
204 countArgs (ApplyTo _ _ _ cont) = 1 + countArgs cont
207 contArgs :: SimplCont -> ([OutExpr], SimplCont)
208 -- Uses substitution to turn each arg into an OutExpr
209 contArgs cont = go [] cont
211 go args (ApplyTo _ arg se cont) = go (substExpr se arg : args) cont
212 go args cont = (reverse args, cont)
214 dropArgs :: Int -> SimplCont -> SimplCont
215 dropArgs 0 cont = cont
216 dropArgs n (ApplyTo _ _ _ cont) = dropArgs (n-1) cont
217 dropArgs n other = pprPanic "dropArgs" (ppr n <+> ppr other)
221 Note [Interesting call context]
222 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
223 We want to avoid inlining an expression where there can't possibly be
224 any gain, such as in an argument position. Hence, if the continuation
225 is interesting (eg. a case scrutinee, application etc.) then we
226 inline, otherwise we don't.
228 Previously some_benefit used to return True only if the variable was
229 applied to some value arguments. This didn't work:
231 let x = _coerce_ (T Int) Int (I# 3) in
232 case _coerce_ Int (T Int) x of
235 we want to inline x, but can't see that it's a constructor in a case
236 scrutinee position, and some_benefit is False.
240 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
242 .... case dMonadST _@_ x0 of (a,b,c) -> ....
244 we'd really like to inline dMonadST here, but we *don't* want to
245 inline if the case expression is just
247 case x of y { DEFAULT -> ... }
249 since we can just eliminate this case instead (x is in WHNF). Similar
250 applies when x is bound to a lambda expression. Hence
251 contIsInteresting looks for case expressions with just a single
256 interestingCallContext :: SimplCont -> CallCtxt
257 -- See Note [Interesting call context]
258 interestingCallContext cont
261 interesting (Select _ bndr _ _ _)
262 | isDeadBinder bndr = CaseCtxt
263 | otherwise = ArgCtxt False 2 -- If the binder is used, this
264 -- is like a strict let
266 interesting (ApplyTo _ arg _ cont)
267 | isTypeArg arg = interesting cont
268 | otherwise = ValAppCtxt -- Can happen if we have (f Int |> co) y
269 -- If f has an INLINE prag we need to give it some
270 -- motivation to inline. See Note [Cast then apply]
273 interesting (StrictArg _ cci _ _) = cci
274 interesting (StrictBind {}) = BoringCtxt
275 interesting (Stop cci) = cci
276 interesting (CoerceIt _ cont) = interesting cont
277 -- If this call is the arg of a strict function, the context
278 -- is a bit interesting. If we inline here, we may get useful
279 -- evaluation information to avoid repeated evals: e.g.
281 -- Here the contIsInteresting makes the '*' keener to inline,
282 -- which in turn exposes a constructor which makes the '+' inline.
283 -- Assuming that +,* aren't small enough to inline regardless.
285 -- It's also very important to inline in a strict context for things
288 -- Here, the context of (f x) is strict, and if f's unfolding is
289 -- a build it's *great* to inline it here. So we must ensure that
290 -- the context for (f x) is not totally uninteresting.
295 -> [CoreRule] -- Rules for function
296 -> Int -- Number of value args
297 -> SimplCont -- Context of the call
300 mkArgInfo fun rules n_val_args call_cont
301 | n_val_args < idArity fun -- Note [Unsaturated functions]
302 = ArgInfo { ai_rules = False
303 , ai_strs = vanilla_stricts
304 , ai_discs = vanilla_discounts }
306 = ArgInfo { ai_rules = interestingArgContext rules call_cont
307 , ai_strs = add_type_str (idType fun) arg_stricts
308 , ai_discs = arg_discounts }
310 vanilla_discounts, arg_discounts :: [Int]
311 vanilla_discounts = repeat 0
312 arg_discounts = case idUnfolding fun of
313 CoreUnfolding {uf_guidance = UnfoldIfGoodArgs {ug_args = discounts}}
314 -> discounts ++ vanilla_discounts
315 _ -> vanilla_discounts
317 vanilla_stricts, arg_stricts :: [Bool]
318 vanilla_stricts = repeat False
321 = case splitStrictSig (idNewStrictness fun) of
322 (demands, result_info)
323 | not (demands `lengthExceeds` n_val_args)
324 -> -- Enough args, use the strictness given.
325 -- For bottoming functions we used to pretend that the arg
326 -- is lazy, so that we don't treat the arg as an
327 -- interesting context. This avoids substituting
328 -- top-level bindings for (say) strings into
329 -- calls to error. But now we are more careful about
330 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
331 if isBotRes result_info then
332 map isStrictDmd demands -- Finite => result is bottom
334 map isStrictDmd demands ++ vanilla_stricts
336 -> WARN( True, text "More demands than arity" <+> ppr fun <+> ppr (idArity fun)
337 <+> ppr n_val_args <+> ppr demands )
338 vanilla_stricts -- Not enough args, or no strictness
340 add_type_str :: Type -> [Bool] -> [Bool]
341 -- If the function arg types are strict, record that in the 'strictness bits'
342 -- No need to instantiate because unboxed types (which dominate the strict
343 -- types) can't instantiate type variables.
344 -- add_type_str is done repeatedly (for each call); might be better
345 -- once-for-all in the function
346 -- But beware primops/datacons with no strictness
347 add_type_str _ [] = []
348 add_type_str fun_ty strs -- Look through foralls
349 | Just (_, fun_ty') <- splitForAllTy_maybe fun_ty -- Includes coercions
350 = add_type_str fun_ty' strs
351 add_type_str fun_ty (str:strs) -- Add strict-type info
352 | Just (arg_ty, fun_ty') <- splitFunTy_maybe fun_ty
353 = (str || isStrictType arg_ty) : add_type_str fun_ty' strs
357 {- Note [Unsaturated functions]
358 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
359 Consider (test eyeball/inline4)
362 where f has arity 2. Then we do not want to inline 'x', because
363 it'll just be floated out again. Even if f has lots of discounts
364 on its first argument -- it must be saturated for these to kick in
367 interestingArgContext :: [CoreRule] -> SimplCont -> Bool
368 -- If the argument has form (f x y), where x,y are boring,
369 -- and f is marked INLINE, then we don't want to inline f.
370 -- But if the context of the argument is
372 -- where g has rules, then we *do* want to inline f, in case it
373 -- exposes a rule that might fire. Similarly, if the context is
375 -- where h has rules, then we do want to inline f; hence the
376 -- call_cont argument to interestingArgContext
378 -- The ai-rules flag makes this happen; if it's
379 -- set, the inliner gets just enough keener to inline f
380 -- regardless of how boring f's arguments are, if it's marked INLINE
382 -- The alternative would be to *always* inline an INLINE function,
383 -- regardless of how boring its context is; but that seems overkill
384 -- For example, it'd mean that wrapper functions were always inlined
385 interestingArgContext rules call_cont
386 = notNull rules || enclosing_fn_has_rules
388 enclosing_fn_has_rules = go call_cont
390 go (Select {}) = False
391 go (ApplyTo {}) = False
392 go (StrictArg _ cci _ _) = interesting cci
393 go (StrictBind {}) = False -- ??
394 go (CoerceIt _ c) = go c
395 go (Stop cci) = interesting cci
397 interesting (ArgCtxt rules _) = rules
398 interesting _ = False
403 %************************************************************************
405 \subsection{Decisions about inlining}
407 %************************************************************************
409 Inlining is controlled partly by the SimplifierMode switch. This has two
412 SimplGently (a) Simplifying before specialiser/full laziness
413 (b) Simplifiying inside InlineRules
414 (c) Simplifying the LHS of a rule
415 (d) Simplifying a GHCi expression or Template
418 SimplPhase n _ Used at all other times
420 The key thing about SimplGently is that it does no call-site inlining.
421 Before full laziness we must be careful not to inline wrappers,
422 because doing so inhibits floating
423 e.g. ...(case f x of ...)...
424 ==> ...(case (case x of I# x# -> fw x#) of ...)...
425 ==> ...(case x of I# x# -> case fw x# of ...)...
426 and now the redex (f x) isn't floatable any more.
428 The no-inlining thing is also important for Template Haskell. You might be
429 compiling in one-shot mode with -O2; but when TH compiles a splice before
430 running it, we don't want to use -O2. Indeed, we don't want to inline
431 anything, because the byte-code interpreter might get confused about
432 unboxed tuples and suchlike.
434 Note [Simplifying gently inside InlineRules]
435 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
436 We don't do much simplification inside InlineRules (which come from
437 INLINE pragmas). It really is important to switch off inlinings
438 inside such expressions. Consider the following example
444 in ...g...g...g...g...g...
446 Now, if that's the ONLY occurrence of f, it will be inlined inside g,
447 and thence copied multiple times when g is inlined.
449 This function may be inlinined in other modules, so we don't want to
450 remove (by inlining) calls to functions that have specialisations, or
451 that may have transformation rules in an importing scope.
453 E.g. {-# INLINE f #-}
456 and suppose that g is strict *and* has specialisations. If we inline
457 g's wrapper, we deny f the chance of getting the specialised version
458 of g when f is inlined at some call site (perhaps in some other
461 It's also important not to inline a worker back into a wrapper.
463 wraper = inline_me (\x -> ...worker... )
464 Normally, the inline_me prevents the worker getting inlined into
465 the wrapper (initially, the worker's only call site!). But,
466 if the wrapper is sure to be called, the strictness analyser will
467 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
468 continuation. That's why the keep_inline predicate returns True for
469 ArgOf continuations. It shouldn't do any harm not to dissolve the
470 inline-me note under these circumstances.
472 Although we do very little simplification inside an InlineRule,
473 the RHS is simplified as normal. For example:
475 all xs = foldr (&&) True xs
476 any p = all . map p {-# INLINE any #-}
478 The RHS of 'any' will get optimised and deforested; but the InlineRule
479 will still mention the original RHS.
482 preInlineUnconditionally
483 ~~~~~~~~~~~~~~~~~~~~~~~~
484 @preInlineUnconditionally@ examines a bndr to see if it is used just
485 once in a completely safe way, so that it is safe to discard the
486 binding inline its RHS at the (unique) usage site, REGARDLESS of how
487 big the RHS might be. If this is the case we don't simplify the RHS
488 first, but just inline it un-simplified.
490 This is much better than first simplifying a perhaps-huge RHS and then
491 inlining and re-simplifying it. Indeed, it can be at least quadratically
500 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
501 This can happen with cascades of functions too:
508 THE MAIN INVARIANT is this:
510 ---- preInlineUnconditionally invariant -----
511 IF preInlineUnconditionally chooses to inline x = <rhs>
512 THEN doing the inlining should not change the occurrence
513 info for the free vars of <rhs>
514 ----------------------------------------------
516 For example, it's tempting to look at trivial binding like
518 and inline it unconditionally. But suppose x is used many times,
519 but this is the unique occurrence of y. Then inlining x would change
520 y's occurrence info, which breaks the invariant. It matters: y
521 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
524 Even RHSs labelled InlineMe aren't caught here, because there might be
525 no benefit from inlining at the call site.
527 [Sept 01] Don't unconditionally inline a top-level thing, because that
528 can simply make a static thing into something built dynamically. E.g.
532 [Remember that we treat \s as a one-shot lambda.] No point in
533 inlining x unless there is something interesting about the call site.
535 But watch out: if you aren't careful, some useful foldr/build fusion
536 can be lost (most notably in spectral/hartel/parstof) because the
537 foldr didn't see the build. Doing the dynamic allocation isn't a big
538 deal, in fact, but losing the fusion can be. But the right thing here
539 seems to be to do a callSiteInline based on the fact that there is
540 something interesting about the call site (it's strict). Hmm. That
543 Conclusion: inline top level things gaily until Phase 0 (the last
544 phase), at which point don't.
547 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
548 preInlineUnconditionally env top_lvl bndr rhs
550 | opt_SimplNoPreInlining = False
551 | otherwise = case idOccInfo bndr of
552 IAmDead -> True -- Happens in ((\x.1) v)
553 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
557 active = case phase of
558 SimplGently -> isEarlyActive act
559 SimplPhase n _ -> isActive n act
560 act = idInlineActivation bndr
562 try_once in_lam int_cxt -- There's one textual occurrence
563 | not in_lam = isNotTopLevel top_lvl || early_phase
564 | otherwise = int_cxt && canInlineInLam rhs
566 -- Be very careful before inlining inside a lambda, becuase (a) we must not
567 -- invalidate occurrence information, and (b) we want to avoid pushing a
568 -- single allocation (here) into multiple allocations (inside lambda).
569 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
570 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
572 -- is_cheap = exprIsCheap rhs
573 -- ok = is_cheap && int_cxt
575 -- int_cxt The context isn't totally boring
576 -- E.g. let f = \ab.BIG in \y. map f xs
577 -- Don't want to substitute for f, because then we allocate
578 -- its closure every time the \y is called
579 -- But: let f = \ab.BIG in \y. map (f y) xs
580 -- Now we do want to substitute for f, even though it's not
581 -- saturated, because we're going to allocate a closure for
582 -- (f y) every time round the loop anyhow.
584 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
585 -- so substituting rhs inside a lambda doesn't change the occ info.
586 -- Sadly, not quite the same as exprIsHNF.
587 canInlineInLam (Lit _) = True
588 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
589 canInlineInLam (Note _ e) = canInlineInLam e
590 canInlineInLam _ = False
592 early_phase = case phase of
593 SimplPhase 0 _ -> False
595 -- If we don't have this early_phase test, consider
596 -- x = length [1,2,3]
597 -- The full laziness pass carefully floats all the cons cells to
598 -- top level, and preInlineUnconditionally floats them all back in.
599 -- Result is (a) static allocation replaced by dynamic allocation
600 -- (b) many simplifier iterations because this tickles
601 -- a related problem; only one inlining per pass
603 -- On the other hand, I have seen cases where top-level fusion is
604 -- lost if we don't inline top level thing (e.g. string constants)
605 -- Hence the test for phase zero (which is the phase for all the final
606 -- simplifications). Until phase zero we take no special notice of
607 -- top level things, but then we become more leery about inlining
612 postInlineUnconditionally
613 ~~~~~~~~~~~~~~~~~~~~~~~~~
614 @postInlineUnconditionally@ decides whether to unconditionally inline
615 a thing based on the form of its RHS; in particular if it has a
616 trivial RHS. If so, we can inline and discard the binding altogether.
618 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
619 only have *forward* references Hence, it's safe to discard the binding
621 NOTE: This isn't our last opportunity to inline. We're at the binding
622 site right now, and we'll get another opportunity when we get to the
625 Note that we do this unconditional inlining only for trival RHSs.
626 Don't inline even WHNFs inside lambdas; doing so may simply increase
627 allocation when the function is called. This isn't the last chance; see
630 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
631 Because we don't even want to inline them into the RHS of constructor
632 arguments. See NOTE above
634 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
635 it's best to inline it anyway. We often get a=E; b=a from desugaring,
636 with both a and b marked NOINLINE. But that seems incompatible with
637 our new view that inlining is like a RULE, so I'm sticking to the 'active'
641 postInlineUnconditionally
642 :: SimplEnv -> TopLevelFlag
643 -> OutId -- The binder (an InId would be fine too)
644 -> OccInfo -- From the InId
648 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
650 | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, don't inline
651 -- because it might be referred to "earlier"
652 | isExportedId bndr = False
653 | isInlineRule unfolding = False -- Note [InlineRule and postInlineUnconditionally]
654 | exprIsTrivial rhs = True
657 -- The point of examining occ_info here is that for *non-values*
658 -- that occur outside a lambda, the call-site inliner won't have
659 -- a chance (becuase it doesn't know that the thing
660 -- only occurs once). The pre-inliner won't have gotten
661 -- it either, if the thing occurs in more than one branch
662 -- So the main target is things like
665 -- True -> case x of ...
666 -- False -> case x of ...
667 -- I'm not sure how important this is in practice
668 OneOcc in_lam _one_br int_cxt -- OneOcc => no code-duplication issue
669 -> smallEnoughToInline unfolding -- Small enough to dup
670 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
672 -- NB: Do NOT inline arbitrarily big things, even if one_br is True
673 -- Reason: doing so risks exponential behaviour. We simplify a big
674 -- expression, inline it, and simplify it again. But if the
675 -- very same thing happens in the big expression, we get
677 -- PRINCIPLE: when we've already simplified an expression once,
678 -- make sure that we only inline it if it's reasonably small.
680 && ((isNotTopLevel top_lvl && not in_lam) ||
681 -- But outside a lambda, we want to be reasonably aggressive
682 -- about inlining into multiple branches of case
683 -- e.g. let x = <non-value>
684 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
685 -- Inlining can be a big win if C3 is the hot-spot, even if
686 -- the uses in C1, C2 are not 'interesting'
687 -- An example that gets worse if you add int_cxt here is 'clausify'
689 (isCheapUnfolding unfolding && int_cxt))
690 -- isCheap => acceptable work duplication; in_lam may be true
691 -- int_cxt to prevent us inlining inside a lambda without some
692 -- good reason. See the notes on int_cxt in preInlineUnconditionally
694 IAmDead -> True -- This happens; for example, the case_bndr during case of
695 -- known constructor: case (a,b) of x { (p,q) -> ... }
696 -- Here x isn't mentioned in the RHS, so we don't want to
697 -- create the (dead) let-binding let x = (a,b) in ...
701 -- Here's an example that we don't handle well:
702 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
703 -- in \y. ....case f of {...} ....
704 -- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
706 -- - We can't preInlineUnconditionally because that woud invalidate
707 -- the occ info for b.
708 -- - We can't postInlineUnconditionally because the RHS is big, and
709 -- that risks exponential behaviour
710 -- - We can't call-site inline, because the rhs is big
714 active = case getMode env of
715 SimplGently -> isAlwaysActive act
716 SimplPhase n _ -> isActive n act
717 act = idInlineActivation bndr
719 activeInline :: SimplEnv -> OutId -> Bool
721 = case getMode env of
723 -- No inlining at all when doing gentle stuff,
724 -- except for local things that occur once (pre/postInlineUnconditionally)
725 -- The reason is that too little clean-up happens if you
726 -- don't inline use-once things. Also a bit of inlining is *good* for
727 -- full laziness; it can expose constant sub-expressions.
728 -- Example in spectral/mandel/Mandel.hs, where the mandelset
729 -- function gets a useful let-float if you inline windowToViewport
731 -- NB: we used to have a second exception, for data con wrappers.
732 -- On the grounds that we use gentle mode for rule LHSs, and
733 -- they match better when data con wrappers are inlined.
734 -- But that only really applies to the trivial wrappers (like (:)),
735 -- and they are now constructed as Compulsory unfoldings (in MkId)
736 -- so they'll happen anyway.
738 SimplPhase n _ -> isActive n act
740 act = idInlineActivation id
742 activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
743 -- Nothing => No rules at all
744 activeRule dflags env
745 | not (dopt Opt_EnableRewriteRules dflags)
746 = Nothing -- Rewriting is off
748 = case getMode env of
749 SimplGently -> Just isAlwaysActive
750 -- Used to be Nothing (no rules in gentle mode)
751 -- Main motivation for changing is that I wanted
752 -- lift String ===> ...
753 -- to work in Template Haskell when simplifying
754 -- splices, so we get simpler code for literal strings
755 SimplPhase n _ -> Just (isActive n)
758 Note [InlineRule and postInlineUnconditionally]
759 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
760 Do not do postInlineUnconditionally if the Id has an InlineRule, otherwise
761 we lose the unfolding. Example
763 -- f has InlineRule with rhs (e |> co)
767 Then there's a danger we'll optimise to
772 and now postInlineUnconditionally, losing the InlineRule on f. Now f'
773 won't inline because 'e' is too big.
776 %************************************************************************
780 %************************************************************************
783 mkLam :: SimplEnv -> [OutBndr] -> OutExpr -> SimplM OutExpr
784 -- mkLam tries three things
785 -- a) eta reduction, if that gives a trivial expression
786 -- b) eta expansion [only if there are some value lambdas]
791 = do { dflags <- getDOptsSmpl
792 ; mkLam' dflags bndrs body }
794 mkLam' :: DynFlags -> [OutBndr] -> OutExpr -> SimplM OutExpr
795 mkLam' dflags bndrs (Cast body co)
796 | not (any bad bndrs)
797 -- Note [Casts and lambdas]
798 = do { lam <- mkLam' dflags bndrs body
799 ; return (mkCoerce (mkPiTypes bndrs co) lam) }
801 co_vars = tyVarsOfType co
802 bad bndr = isCoVar bndr && bndr `elemVarSet` co_vars
804 mkLam' dflags bndrs body
805 | dopt Opt_DoEtaReduction dflags,
806 Just etad_lam <- tryEtaReduce bndrs body
807 = do { tick (EtaReduction (head bndrs))
810 | dopt Opt_DoLambdaEtaExpansion dflags,
811 not (inGentleMode env), -- In gentle mode don't eta-expansion
812 any isRuntimeVar bndrs -- because it can clutter up the code
813 -- with casts etc that may not be removed
814 = do { let body' = tryEtaExpansion dflags body
815 ; return (mkLams bndrs body') }
818 = return (mkLams bndrs body)
821 Note [Casts and lambdas]
822 ~~~~~~~~~~~~~~~~~~~~~~~~
824 (\x. (\y. e) `cast` g1) `cast` g2
825 There is a danger here that the two lambdas look separated, and the
826 full laziness pass might float an expression to between the two.
828 So this equation in mkLam' floats the g1 out, thus:
829 (\x. e `cast` g1) --> (\x.e) `cast` (tx -> g1)
832 In general, this floats casts outside lambdas, where (I hope) they
833 might meet and cancel with some other cast:
834 \x. e `cast` co ===> (\x. e) `cast` (tx -> co)
835 /\a. e `cast` co ===> (/\a. e) `cast` (/\a. co)
836 /\g. e `cast` co ===> (/\g. e) `cast` (/\g. co)
839 Notice that it works regardless of 'e'. Originally it worked only
840 if 'e' was itself a lambda, but in some cases that resulted in
841 fruitless iteration in the simplifier. A good example was when
842 compiling Text.ParserCombinators.ReadPrec, where we had a definition
843 like (\x. Get `cast` g)
844 where Get is a constructor with nonzero arity. Then mkLam eta-expanded
845 the Get, and the next iteration eta-reduced it, and then eta-expanded
848 Note also the side condition for the case of coercion binders.
849 It does not make sense to transform
850 /\g. e `cast` g ==> (/\g.e) `cast` (/\g.g)
851 because the latter is not well-kinded.
853 -- c) floating lets out through big lambdas
854 -- [only if all tyvar lambdas, and only if this lambda
855 -- is the RHS of a let]
857 {- Sept 01: I'm experimenting with getting the
858 full laziness pass to float out past big lambdsa
859 | all isTyVar bndrs, -- Only for big lambdas
860 contIsRhs cont -- Only try the rhs type-lambda floating
861 -- if this is indeed a right-hand side; otherwise
862 -- we end up floating the thing out, only for float-in
863 -- to float it right back in again!
864 = do (floats, body') <- tryRhsTyLam env bndrs body
865 return (floats, mkLams bndrs body')
869 %************************************************************************
873 %************************************************************************
875 Note [Eta reduction conditions]
876 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
877 We try for eta reduction here, but *only* if we get all the way to an
878 trivial expression. We don't want to remove extra lambdas unless we
879 are going to avoid allocating this thing altogether.
881 There are some particularly delicate points here:
883 * Eta reduction is not valid in general:
885 This matters, partly for old-fashioned correctness reasons but,
886 worse, getting it wrong can yield a seg fault. Consider
888 h y = case (case y of { True -> f `seq` True; False -> False }) of
889 True -> ...; False -> ...
891 If we (unsoundly) eta-reduce f to get f=f, the strictness analyser
892 says f=bottom, and replaces the (f `seq` True) with just
893 (f `cast` unsafe-co). BUT, as thing stand, 'f' got arity 1, and it
894 *keeps* arity 1 (perhaps also wrongly). So CorePrep eta-expands
895 the definition again, so that it does not termninate after all.
896 Result: seg-fault because the boolean case actually gets a function value.
899 So it's important to to the right thing.
901 * Note [Arity care]: we need to be careful if we just look at f's
902 arity. Currently (Dec07), f's arity is visible in its own RHS (see
903 Note [Arity robustness] in SimplEnv) so we must *not* trust the
904 arity when checking that 'f' is a value. Otherwise we will
909 Which might change a terminiating program (think (f `seq` e)) to a
910 non-terminating one. So we check for being a loop breaker first.
912 However for GlobalIds we can look at the arity; and for primops we
913 must, since they have no unfolding.
915 * Regardless of whether 'f' is a value, we always want to
916 reduce (/\a -> f a) to f
917 This came up in a RULE: foldr (build (/\a -> g a))
918 did not match foldr (build (/\b -> ...something complex...))
919 The type checker can insert these eta-expanded versions,
920 with both type and dictionary lambdas; hence the slightly
923 * Never *reduce* arity. For example
925 Then if h has arity 1 we don't want to eta-reduce because then
926 f's arity would decrease, and that is bad
928 These delicacies are why we don't use exprIsTrivial and exprIsHNF here.
932 tryEtaReduce :: [OutBndr] -> OutExpr -> Maybe OutExpr
933 tryEtaReduce bndrs body
934 = go (reverse bndrs) body
936 incoming_arity = count isId bndrs
938 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
939 go [] fun | ok_fun fun = Just fun -- Success!
940 go _ _ = Nothing -- Failure!
942 -- Note [Eta reduction conditions]
943 ok_fun (App fun (Type ty))
944 | not (any (`elemVarSet` tyVarsOfType ty) bndrs)
947 = not (fun_id `elem` bndrs)
948 && (ok_fun_id fun_id || all ok_lam bndrs)
951 ok_fun_id fun = fun_arity fun >= incoming_arity
953 fun_arity fun -- See Note [Arity care]
954 | isLocalId fun && isLoopBreaker (idOccInfo fun) = 0
955 | otherwise = idArity fun
957 ok_lam v = isTyVar v || isDictId v
959 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
963 %************************************************************************
967 %************************************************************************
971 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
974 where (in both cases)
976 * The xi can include type variables
978 * The yi are all value variables
980 * N is a NORMAL FORM (i.e. no redexes anywhere)
981 wanting a suitable number of extra args.
983 The biggest reason for doing this is for cases like
989 Here we want to get the lambdas together. A good exmaple is the nofib
990 program fibheaps, which gets 25% more allocation if you don't do this
993 We may have to sandwich some coerces between the lambdas
994 to make the types work. exprEtaExpandArity looks through coerces
995 when computing arity; and etaExpand adds the coerces as necessary when
996 actually computing the expansion.
999 tryEtaExpansion :: DynFlags -> OutExpr -> OutExpr
1000 -- There is at least one runtime binder in the binders
1001 tryEtaExpansion dflags body
1002 = etaExpand fun_arity body
1004 fun_arity = exprEtaExpandArity dflags body
1008 %************************************************************************
1010 \subsection{Floating lets out of big lambdas}
1012 %************************************************************************
1014 Note [Floating and type abstraction]
1015 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1018 We'd like to float this to
1021 x = /\a. C (y1 a) (y2 a)
1022 for the usual reasons: we want to inline x rather vigorously.
1024 You may think that this kind of thing is rare. But in some programs it is
1025 common. For example, if you do closure conversion you might get:
1027 data a :-> b = forall e. (e -> a -> b) :$ e
1029 f_cc :: forall a. a :-> a
1030 f_cc = /\a. (\e. id a) :$ ()
1032 Now we really want to inline that f_cc thing so that the
1033 construction of the closure goes away.
1035 So I have elaborated simplLazyBind to understand right-hand sides that look
1039 and treat them specially. The real work is done in SimplUtils.abstractFloats,
1040 but there is quite a bit of plumbing in simplLazyBind as well.
1042 The same transformation is good when there are lets in the body:
1044 /\abc -> let(rec) x = e in b
1046 let(rec) x' = /\abc -> let x = x' a b c in e
1048 /\abc -> let x = x' a b c in b
1050 This is good because it can turn things like:
1052 let f = /\a -> letrec g = ... g ... in g
1054 letrec g' = /\a -> ... g' a ...
1056 let f = /\ a -> g' a
1058 which is better. In effect, it means that big lambdas don't impede
1061 This optimisation is CRUCIAL in eliminating the junk introduced by
1062 desugaring mutually recursive definitions. Don't eliminate it lightly!
1064 [May 1999] If we do this transformation *regardless* then we can
1065 end up with some pretty silly stuff. For example,
1068 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
1073 st = /\s -> ...[y1 s/x1, y2 s/x2]
1076 Unless the "..." is a WHNF there is really no point in doing this.
1077 Indeed it can make things worse. Suppose x1 is used strictly,
1080 x1* = case f y of { (a,b) -> e }
1082 If we abstract this wrt the tyvar we then can't do the case inline
1083 as we would normally do.
1085 That's why the whole transformation is part of the same process that
1086 floats let-bindings and constructor arguments out of RHSs. In particular,
1087 it is guarded by the doFloatFromRhs call in simplLazyBind.
1091 abstractFloats :: [OutTyVar] -> SimplEnv -> OutExpr -> SimplM ([OutBind], OutExpr)
1092 abstractFloats main_tvs body_env body
1093 = ASSERT( notNull body_floats )
1094 do { (subst, float_binds) <- mapAccumLM abstract empty_subst body_floats
1095 ; return (float_binds, CoreSubst.substExpr subst body) }
1097 main_tv_set = mkVarSet main_tvs
1098 body_floats = getFloats body_env
1099 empty_subst = CoreSubst.mkEmptySubst (seInScope body_env)
1101 abstract :: CoreSubst.Subst -> OutBind -> SimplM (CoreSubst.Subst, OutBind)
1102 abstract subst (NonRec id rhs)
1103 = do { (poly_id, poly_app) <- mk_poly tvs_here id
1104 ; let poly_rhs = mkLams tvs_here rhs'
1105 subst' = CoreSubst.extendIdSubst subst id poly_app
1106 ; return (subst', (NonRec poly_id poly_rhs)) }
1108 rhs' = CoreSubst.substExpr subst rhs
1109 tvs_here | any isCoVar main_tvs = main_tvs -- Note [Abstract over coercions]
1111 = varSetElems (main_tv_set `intersectVarSet` exprSomeFreeVars isTyVar rhs')
1113 -- Abstract only over the type variables free in the rhs
1114 -- wrt which the new binding is abstracted. But the naive
1115 -- approach of abstract wrt the tyvars free in the Id's type
1117 -- /\ a b -> let t :: (a,b) = (e1, e2)
1120 -- Here, b isn't free in x's type, but we must nevertheless
1121 -- abstract wrt b as well, because t's type mentions b.
1122 -- Since t is floated too, we'd end up with the bogus:
1123 -- poly_t = /\ a b -> (e1, e2)
1124 -- poly_x = /\ a -> fst (poly_t a *b*)
1125 -- So for now we adopt the even more naive approach of
1126 -- abstracting wrt *all* the tyvars. We'll see if that
1127 -- gives rise to problems. SLPJ June 98
1129 abstract subst (Rec prs)
1130 = do { (poly_ids, poly_apps) <- mapAndUnzipM (mk_poly tvs_here) ids
1131 ; let subst' = CoreSubst.extendSubstList subst (ids `zip` poly_apps)
1132 poly_rhss = [mkLams tvs_here (CoreSubst.substExpr subst' rhs) | rhs <- rhss]
1133 ; return (subst', Rec (poly_ids `zip` poly_rhss)) }
1135 (ids,rhss) = unzip prs
1136 -- For a recursive group, it's a bit of a pain to work out the minimal
1137 -- set of tyvars over which to abstract:
1138 -- /\ a b c. let x = ...a... in
1139 -- letrec { p = ...x...q...
1140 -- q = .....p...b... } in
1142 -- Since 'x' is abstracted over 'a', the {p,q} group must be abstracted
1143 -- over 'a' (because x is replaced by (poly_x a)) as well as 'b'.
1144 -- Since it's a pain, we just use the whole set, which is always safe
1146 -- If you ever want to be more selective, remember this bizarre case too:
1148 -- Here, we must abstract 'x' over 'a'.
1151 mk_poly tvs_here var
1152 = do { uniq <- getUniqueM
1153 ; let poly_name = setNameUnique (idName var) uniq -- Keep same name
1154 poly_ty = mkForAllTys tvs_here (idType var) -- But new type of course
1155 poly_id = transferPolyIdInfo var tvs_here $ -- Note [transferPolyIdInfo] in Id.lhs
1156 mkLocalId poly_name poly_ty
1157 ; return (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tvs_here)) }
1158 -- In the olden days, it was crucial to copy the occInfo of the original var,
1159 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1160 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1161 -- at already simplified code, so it doesn't matter
1163 -- It's even right to retain single-occurrence or dead-var info:
1164 -- Suppose we started with /\a -> let x = E in B
1165 -- where x occurs once in B. Then we transform to:
1166 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1167 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1168 -- the occurrences of x' will be just the occurrences originally
1172 Note [Abstract over coercions]
1173 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1174 If a coercion variable (g :: a ~ Int) is free in the RHS, then so is the
1175 type variable a. Rather than sort this mess out, we simply bale out and abstract
1176 wrt all the type variables if any of them are coercion variables.
1179 Historical note: if you use let-bindings instead of a substitution, beware of this:
1181 -- Suppose we start with:
1183 -- x = /\ a -> let g = G in E
1185 -- Then we'll float to get
1187 -- x = let poly_g = /\ a -> G
1188 -- in /\ a -> let g = poly_g a in E
1190 -- But now the occurrence analyser will see just one occurrence
1191 -- of poly_g, not inside a lambda, so the simplifier will
1192 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1193 -- (I used to think that the "don't inline lone occurrences" stuff
1194 -- would stop this happening, but since it's the *only* occurrence,
1195 -- PreInlineUnconditionally kicks in first!)
1197 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1198 -- to appear many times. (NB: mkInlineMe eliminates
1199 -- such notes on trivial RHSs, so do it manually.)
1201 %************************************************************************
1205 %************************************************************************
1207 prepareAlts tries these things:
1209 1. If several alternatives are identical, merge them into
1210 a single DEFAULT alternative. I've occasionally seen this
1211 making a big difference:
1213 case e of =====> case e of
1214 C _ -> f x D v -> ....v....
1215 D v -> ....v.... DEFAULT -> f x
1218 The point is that we merge common RHSs, at least for the DEFAULT case.
1219 [One could do something more elaborate but I've never seen it needed.]
1220 To avoid an expensive test, we just merge branches equal to the *first*
1221 alternative; this picks up the common cases
1222 a) all branches equal
1223 b) some branches equal to the DEFAULT (which occurs first)
1226 case e of b { ==> case e of b {
1227 p1 -> rhs1 p1 -> rhs1
1229 pm -> rhsm pm -> rhsm
1230 _ -> case b of b' { pn -> let b'=b in rhsn
1232 ... po -> let b'=b in rhso
1233 po -> rhso _ -> let b'=b in rhsd
1237 which merges two cases in one case when -- the default alternative of
1238 the outer case scrutises the same variable as the outer case This
1239 transformation is called Case Merging. It avoids that the same
1240 variable is scrutinised multiple times.
1243 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
1249 where @is@ was something like
1251 p `is` n = p /= (-1) && p == n
1253 This gave rise to a horrible sequence of cases
1260 and similarly in cascade for all the join points!
1263 ~~~~~~~~~~~~~~~~~~~~
1264 We do this *here*, looking at un-simplified alternatives, because we
1265 have to check that r doesn't mention the variables bound by the
1266 pattern in each alternative, so the binder-info is rather useful.
1269 prepareAlts :: SimplEnv -> OutExpr -> OutId -> [InAlt] -> SimplM ([AltCon], [InAlt])
1270 prepareAlts env scrut case_bndr' alts
1271 = do { dflags <- getDOptsSmpl
1272 ; alts <- combineIdenticalAlts case_bndr' alts
1274 ; let (alts_wo_default, maybe_deflt) = findDefault alts
1275 alt_cons = [con | (con,_,_) <- alts_wo_default]
1276 imposs_deflt_cons = nub (imposs_cons ++ alt_cons)
1277 -- "imposs_deflt_cons" are handled
1278 -- EITHER by the context,
1279 -- OR by a non-DEFAULT branch in this case expression.
1281 ; default_alts <- prepareDefault dflags env case_bndr' mb_tc_app
1282 imposs_deflt_cons maybe_deflt
1284 ; let trimmed_alts = filterOut impossible_alt alts_wo_default
1285 merged_alts = mergeAlts trimmed_alts default_alts
1286 -- We need the mergeAlts in case the new default_alt
1287 -- has turned into a constructor alternative.
1288 -- The merge keeps the inner DEFAULT at the front, if there is one
1289 -- and interleaves the alternatives in the right order
1291 ; return (imposs_deflt_cons, merged_alts) }
1293 mb_tc_app = splitTyConApp_maybe (idType case_bndr')
1294 Just (_, inst_tys) = mb_tc_app
1296 imposs_cons = case scrut of
1297 Var v -> otherCons (idUnfolding v)
1300 impossible_alt :: CoreAlt -> Bool
1301 impossible_alt (con, _, _) | con `elem` imposs_cons = True
1302 impossible_alt (DataAlt con, _, _) = dataConCannotMatch inst_tys con
1303 impossible_alt _ = False
1306 --------------------------------------------------
1307 -- 1. Merge identical branches
1308 --------------------------------------------------
1309 combineIdenticalAlts :: OutId -> [InAlt] -> SimplM [InAlt]
1311 combineIdenticalAlts case_bndr ((_con1,bndrs1,rhs1) : con_alts)
1312 | all isDeadBinder bndrs1, -- Remember the default
1313 length filtered_alts < length con_alts -- alternative comes first
1314 -- Also Note [Dead binders]
1315 = do { tick (AltMerge case_bndr)
1316 ; return ((DEFAULT, [], rhs1) : filtered_alts) }
1318 filtered_alts = filter keep con_alts
1319 keep (_con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1321 combineIdenticalAlts _ alts = return alts
1323 -------------------------------------------------------------------------
1324 -- Prepare the default alternative
1325 -------------------------------------------------------------------------
1326 prepareDefault :: DynFlags
1328 -> OutId -- Case binder; need just for its type. Note that as an
1329 -- OutId, it has maximum information; this is important.
1330 -- Test simpl013 is an example
1331 -> Maybe (TyCon, [Type]) -- Type of scrutinee, decomposed
1332 -> [AltCon] -- These cons can't happen when matching the default
1333 -> Maybe InExpr -- Rhs
1334 -> SimplM [InAlt] -- Still unsimplified
1335 -- We use a list because it's what mergeAlts expects,
1336 -- And becuase case-merging can cause many to show up
1338 ------- Merge nested cases ----------
1339 prepareDefault dflags env outer_bndr _bndr_ty imposs_cons (Just deflt_rhs)
1340 | dopt Opt_CaseMerge dflags
1341 , Case (Var inner_scrut_var) inner_bndr _ inner_alts <- deflt_rhs
1342 , DoneId inner_scrut_var' <- substId env inner_scrut_var
1343 -- Remember, inner_scrut_var is an InId, but outer_bndr is an OutId
1344 , inner_scrut_var' == outer_bndr
1345 -- NB: the substId means that if the outer scrutinee was a
1346 -- variable, and inner scrutinee is the same variable,
1347 -- then inner_scrut_var' will be outer_bndr
1348 -- via the magic of simplCaseBinder
1349 = do { tick (CaseMerge outer_bndr)
1351 ; let munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1352 ; return [(con, args, munge_rhs rhs) | (con, args, rhs) <- inner_alts,
1353 not (con `elem` imposs_cons) ]
1354 -- NB: filter out any imposs_cons. Example:
1357 -- DEFAULT -> case x of
1360 -- When we merge, we must ensure that e1 takes
1361 -- precedence over e2 as the value for A!
1363 -- Warning: don't call prepareAlts recursively!
1364 -- Firstly, there's no point, because inner alts have already had
1365 -- mkCase applied to them, so they won't have a case in their default
1366 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1367 -- in munge_rhs may put a case into the DEFAULT branch!
1370 --------- Fill in known constructor -----------
1371 prepareDefault _ _ case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
1372 | -- This branch handles the case where we are
1373 -- scrutinisng an algebraic data type
1374 isAlgTyCon tycon -- It's a data type, tuple, or unboxed tuples.
1375 , not (isNewTyCon tycon) -- We can have a newtype, if we are just doing an eval:
1376 -- case x of { DEFAULT -> e }
1377 -- and we don't want to fill in a default for them!
1378 , Just all_cons <- tyConDataCons_maybe tycon
1379 , not (null all_cons) -- This is a tricky corner case. If the data type has no constructors,
1380 -- which GHC allows, then the case expression will have at most a default
1381 -- alternative. We don't want to eliminate that alternative, because the
1382 -- invariant is that there's always one alternative. It's more convenient
1384 -- case x of { DEFAULT -> e }
1385 -- as it is, rather than transform it to
1386 -- error "case cant match"
1387 -- which would be quite legitmate. But it's a really obscure corner, and
1388 -- not worth wasting code on.
1389 , let imposs_data_cons = [con | DataAlt con <- imposs_cons] -- We now know it's a data type
1390 impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
1391 = case filterOut impossible all_cons of
1392 [] -> return [] -- Eliminate the default alternative
1393 -- altogether if it can't match
1395 [con] -> -- It matches exactly one constructor, so fill it in
1396 do { tick (FillInCaseDefault case_bndr)
1398 ; let (ex_tvs, co_tvs, arg_ids) =
1399 dataConRepInstPat us con inst_tys
1400 ; return [(DataAlt con, ex_tvs ++ co_tvs ++ arg_ids, deflt_rhs)] }
1402 _ -> return [(DEFAULT, [], deflt_rhs)]
1404 | debugIsOn, isAlgTyCon tycon, not (isOpenTyCon tycon), null (tyConDataCons tycon)
1405 -- This can legitimately happen for type families, so don't report that
1406 = pprTrace "prepareDefault" (ppr case_bndr <+> ppr tycon)
1407 $ return [(DEFAULT, [], deflt_rhs)]
1409 --------- Catch-all cases -----------
1410 prepareDefault _dflags _env _case_bndr _bndr_ty _imposs_cons (Just deflt_rhs)
1411 = return [(DEFAULT, [], deflt_rhs)]
1413 prepareDefault _dflags _env _case_bndr _bndr_ty _imposs_cons Nothing
1414 = return [] -- No default branch
1419 =================================================================================
1421 mkCase tries these things
1423 1. Eliminate the case altogether if possible
1431 and similar friends.
1435 mkCase :: OutExpr -> OutId -> [OutAlt] -- Increasing order
1438 --------------------------------------------------
1440 --------------------------------------------------
1442 mkCase scrut case_bndr alts -- Identity case
1443 | all identity_alt alts
1444 = do tick (CaseIdentity case_bndr)
1445 return (re_cast scrut)
1447 identity_alt (con, args, rhs) = check_eq con args (de_cast rhs)
1449 check_eq DEFAULT _ (Var v) = v == case_bndr
1450 check_eq (LitAlt lit') _ (Lit lit) = lit == lit'
1451 check_eq (DataAlt con) args rhs = rhs `cheapEqExpr` mkConApp con (arg_tys ++ varsToCoreExprs args)
1452 || rhs `cheapEqExpr` Var case_bndr
1453 check_eq _ _ _ = False
1455 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1458 -- case e of x { _ -> x `cast` c }
1459 -- And we definitely want to eliminate this case, to give
1461 -- So we throw away the cast from the RHS, and reconstruct
1462 -- it at the other end. All the RHS casts must be the same
1463 -- if (all identity_alt alts) holds.
1465 -- Don't worry about nested casts, because the simplifier combines them
1466 de_cast (Cast e _) = e
1469 re_cast scrut = case head alts of
1470 (_,_,Cast _ co) -> Cast scrut co
1475 --------------------------------------------------
1477 --------------------------------------------------
1478 mkCase scrut bndr alts = return (Case scrut bndr (coreAltsType alts) alts)
1482 When adding auxiliary bindings for the case binder, it's worth checking if
1483 its dead, because it often is, and occasionally these mkCase transformations
1484 cascade rather nicely.
1487 bindCaseBndr :: Id -> CoreExpr -> CoreExpr -> CoreExpr
1488 bindCaseBndr bndr rhs body
1489 | isDeadBinder bndr = body
1490 | otherwise = bindNonRec bndr rhs body