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,
14 simplEnvForGHCi, simplEnvForRules, updModeForInlineRules,
16 -- The continuation type
17 SimplCont(..), DupFlag(..), ArgInfo(..),
18 contIsDupable, contResultType, contIsTrivial, contArgs, dropArgs,
19 countValArgs, countArgs,
20 mkBoringStop, mkRhsStop, mkLazyArgStop, contIsRhsOrArg,
21 interestingCallContext,
23 interestingArg, mkArgInfo,
28 #include "HsVersions.h"
34 import qualified CoreSubst
38 import CoreArity ( etaExpand, exprEtaExpandArity )
42 import Var ( isCoVar )
45 import Type hiding( substTy )
46 import Coercion ( coercionKind )
48 import Unify ( dataConCannotMatch )
60 %************************************************************************
64 %************************************************************************
66 A SimplCont allows the simplifier to traverse the expression in a
67 zipper-like fashion. The SimplCont represents the rest of the expression,
68 "above" the point of interest.
70 You can also think of a SimplCont as an "evaluation context", using
71 that term in the way it is used for operational semantics. This is the
72 way I usually think of it, For example you'll often see a syntax for
73 evaluation context looking like
74 C ::= [] | C e | case C of alts | C `cast` co
75 That's the kind of thing we are doing here, and I use that syntax in
80 * A SimplCont describes a *strict* context (just like
81 evaluation contexts do). E.g. Just [] is not a SimplCont
83 * A SimplCont describes a context that *does not* bind
84 any variables. E.g. \x. [] is not a SimplCont
88 = Stop -- An empty context, or hole, []
89 CallCtxt -- True <=> There is something interesting about
90 -- the context, and hence the inliner
91 -- should be a bit keener (see interestingCallContext)
93 -- This is an argument of a function that has RULES
94 -- Inlining the call might allow the rule to fire
96 | CoerceIt -- C `cast` co
97 OutCoercion -- The coercion simplified
102 InExpr SimplEnv -- The argument and its static env
105 | Select -- case C of alts
107 InId [InAlt] SimplEnv -- The case binder, alts, and subst-env
110 -- The two strict forms have no DupFlag, because we never duplicate them
111 | StrictBind -- (\x* \xs. e) C
112 InId [InBndr] -- let x* = [] in e
113 InExpr SimplEnv -- is a special case
117 OutExpr -- e; *always* of form (Var v `App1` e1 .. `App` en)
118 CallCtxt -- Whether *this* argument position is interesting
119 ArgInfo -- Whether the function at the head of e has rules, etc
120 SimplCont -- plus strictness flags for *further* args
124 ai_rules :: Bool, -- Function has rules (recursively)
125 -- => be keener to inline in all args
126 ai_strs :: [Bool], -- Strictness of arguments
127 -- Usually infinite, but if it is finite it guarantees
128 -- that the function diverges after being given
129 -- that number of args
130 ai_discs :: [Int] -- Discounts for arguments; non-zero => be keener to inline
134 instance Outputable SimplCont where
135 ppr (Stop interesting) = ptext (sLit "Stop") <> brackets (ppr interesting)
136 ppr (ApplyTo dup arg _ cont) = ((ptext (sLit "ApplyTo") <+> ppr dup <+> pprParendExpr arg)
137 {- $$ nest 2 (pprSimplEnv se) -}) $$ ppr cont
138 ppr (StrictBind b _ _ _ cont) = (ptext (sLit "StrictBind") <+> ppr b) $$ ppr cont
139 ppr (StrictArg f _ _ cont) = (ptext (sLit "StrictArg") <+> ppr f) $$ ppr cont
140 ppr (Select dup bndr alts _ cont) = (ptext (sLit "Select") <+> ppr dup <+> ppr bndr) $$
141 (nest 4 (ppr alts)) $$ ppr cont
142 ppr (CoerceIt co cont) = (ptext (sLit "CoerceIt") <+> ppr co) $$ ppr cont
144 data DupFlag = OkToDup | NoDup
146 instance Outputable DupFlag where
147 ppr OkToDup = ptext (sLit "ok")
148 ppr NoDup = ptext (sLit "nodup")
153 mkBoringStop :: SimplCont
154 mkBoringStop = Stop BoringCtxt
156 mkRhsStop :: SimplCont -- See Note [RHS of lets] in CoreUnfold
157 mkRhsStop = Stop (ArgCtxt False)
159 mkLazyArgStop :: CallCtxt -> SimplCont
160 mkLazyArgStop cci = Stop cci
163 contIsRhsOrArg :: SimplCont -> Bool
164 contIsRhsOrArg (Stop {}) = True
165 contIsRhsOrArg (StrictBind {}) = True
166 contIsRhsOrArg (StrictArg {}) = True
167 contIsRhsOrArg _ = False
170 contIsDupable :: SimplCont -> Bool
171 contIsDupable (Stop {}) = True
172 contIsDupable (ApplyTo OkToDup _ _ _) = True
173 contIsDupable (Select OkToDup _ _ _ _) = True
174 contIsDupable (CoerceIt _ cont) = contIsDupable cont
175 contIsDupable _ = False
178 contIsTrivial :: SimplCont -> Bool
179 contIsTrivial (Stop {}) = True
180 contIsTrivial (ApplyTo _ (Type _) _ cont) = contIsTrivial cont
181 contIsTrivial (CoerceIt _ cont) = contIsTrivial cont
182 contIsTrivial _ = False
185 contResultType :: SimplEnv -> OutType -> SimplCont -> OutType
186 contResultType env ty cont
189 subst_ty se ty = substTy (se `setInScope` env) ty
192 go (CoerceIt co cont) _ = go cont (snd (coercionKind co))
193 go (StrictBind _ bs body se cont) _ = go cont (subst_ty se (exprType (mkLams bs body)))
194 go (StrictArg fn _ _ cont) _ = go cont (funResultTy (exprType fn))
195 go (Select _ _ alts se cont) _ = go cont (subst_ty se (coreAltsType alts))
196 go (ApplyTo _ arg se cont) ty = go cont (apply_to_arg ty arg se)
198 apply_to_arg ty (Type ty_arg) se = applyTy ty (subst_ty se ty_arg)
199 apply_to_arg ty _ _ = funResultTy ty
202 countValArgs :: SimplCont -> Int
203 countValArgs (ApplyTo _ (Type _) _ cont) = countValArgs cont
204 countValArgs (ApplyTo _ _ _ cont) = 1 + countValArgs cont
207 countArgs :: SimplCont -> Int
208 countArgs (ApplyTo _ _ _ cont) = 1 + countArgs cont
211 contArgs :: SimplCont -> ([OutExpr], SimplCont)
212 -- Uses substitution to turn each arg into an OutExpr
213 contArgs cont = go [] cont
215 go args (ApplyTo _ arg se cont) = go (substExpr se arg : args) cont
216 go args cont = (reverse args, cont)
218 dropArgs :: Int -> SimplCont -> SimplCont
219 dropArgs 0 cont = cont
220 dropArgs n (ApplyTo _ _ _ cont) = dropArgs (n-1) cont
221 dropArgs n other = pprPanic "dropArgs" (ppr n <+> ppr other)
225 Note [Interesting call context]
226 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
227 We want to avoid inlining an expression where there can't possibly be
228 any gain, such as in an argument position. Hence, if the continuation
229 is interesting (eg. a case scrutinee, application etc.) then we
230 inline, otherwise we don't.
232 Previously some_benefit used to return True only if the variable was
233 applied to some value arguments. This didn't work:
235 let x = _coerce_ (T Int) Int (I# 3) in
236 case _coerce_ Int (T Int) x of
239 we want to inline x, but can't see that it's a constructor in a case
240 scrutinee position, and some_benefit is False.
244 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
246 .... case dMonadST _@_ x0 of (a,b,c) -> ....
248 we'd really like to inline dMonadST here, but we *don't* want to
249 inline if the case expression is just
251 case x of y { DEFAULT -> ... }
253 since we can just eliminate this case instead (x is in WHNF). Similar
254 applies when x is bound to a lambda expression. Hence
255 contIsInteresting looks for case expressions with just a single
260 interestingCallContext :: SimplCont -> CallCtxt
261 -- See Note [Interesting call context]
262 interestingCallContext cont
265 interesting (Select _ bndr _ _ _)
266 | isDeadBinder bndr = CaseCtxt
267 | otherwise = ArgCtxt False -- If the binder is used, this
268 -- is like a strict let
269 -- See Note [RHS of lets] in CoreUnfold
271 interesting (ApplyTo _ arg _ cont)
272 | isTypeArg arg = interesting cont
273 | otherwise = ValAppCtxt -- Can happen if we have (f Int |> co) y
274 -- If f has an INLINE prag we need to give it some
275 -- motivation to inline. See Note [Cast then apply]
278 interesting (StrictArg _ cci _ _) = cci
279 interesting (StrictBind {}) = BoringCtxt
280 interesting (Stop cci) = cci
281 interesting (CoerceIt _ cont) = interesting cont
282 -- If this call is the arg of a strict function, the context
283 -- is a bit interesting. If we inline here, we may get useful
284 -- evaluation information to avoid repeated evals: e.g.
286 -- Here the contIsInteresting makes the '*' keener to inline,
287 -- which in turn exposes a constructor which makes the '+' inline.
288 -- Assuming that +,* aren't small enough to inline regardless.
290 -- It's also very important to inline in a strict context for things
293 -- Here, the context of (f x) is strict, and if f's unfolding is
294 -- a build it's *great* to inline it here. So we must ensure that
295 -- the context for (f x) is not totally uninteresting.
300 -> [CoreRule] -- Rules for function
301 -> Int -- Number of value args
302 -> SimplCont -- Context of the call
305 mkArgInfo fun rules n_val_args call_cont
306 | n_val_args < idArity fun -- Note [Unsaturated functions]
307 = ArgInfo { ai_rules = False
308 , ai_strs = vanilla_stricts
309 , ai_discs = vanilla_discounts }
311 = ArgInfo { ai_rules = interestingArgContext rules call_cont
312 , ai_strs = add_type_str (idType fun) arg_stricts
313 , ai_discs = arg_discounts }
315 vanilla_discounts, arg_discounts :: [Int]
316 vanilla_discounts = repeat 0
317 arg_discounts = case idUnfolding fun of
318 CoreUnfolding {uf_guidance = UnfoldIfGoodArgs {ug_args = discounts}}
319 -> discounts ++ vanilla_discounts
320 _ -> vanilla_discounts
322 vanilla_stricts, arg_stricts :: [Bool]
323 vanilla_stricts = repeat False
326 = case splitStrictSig (idNewStrictness fun) of
327 (demands, result_info)
328 | not (demands `lengthExceeds` n_val_args)
329 -> -- Enough args, use the strictness given.
330 -- For bottoming functions we used to pretend that the arg
331 -- is lazy, so that we don't treat the arg as an
332 -- interesting context. This avoids substituting
333 -- top-level bindings for (say) strings into
334 -- calls to error. But now we are more careful about
335 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
336 if isBotRes result_info then
337 map isStrictDmd demands -- Finite => result is bottom
339 map isStrictDmd demands ++ vanilla_stricts
341 -> WARN( True, text "More demands than arity" <+> ppr fun <+> ppr (idArity fun)
342 <+> ppr n_val_args <+> ppr demands )
343 vanilla_stricts -- Not enough args, or no strictness
345 add_type_str :: Type -> [Bool] -> [Bool]
346 -- If the function arg types are strict, record that in the 'strictness bits'
347 -- No need to instantiate because unboxed types (which dominate the strict
348 -- types) can't instantiate type variables.
349 -- add_type_str is done repeatedly (for each call); might be better
350 -- once-for-all in the function
351 -- But beware primops/datacons with no strictness
352 add_type_str _ [] = []
353 add_type_str fun_ty strs -- Look through foralls
354 | Just (_, fun_ty') <- splitForAllTy_maybe fun_ty -- Includes coercions
355 = add_type_str fun_ty' strs
356 add_type_str fun_ty (str:strs) -- Add strict-type info
357 | Just (arg_ty, fun_ty') <- splitFunTy_maybe fun_ty
358 = (str || isStrictType arg_ty) : add_type_str fun_ty' strs
362 {- Note [Unsaturated functions]
363 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
364 Consider (test eyeball/inline4)
367 where f has arity 2. Then we do not want to inline 'x', because
368 it'll just be floated out again. Even if f has lots of discounts
369 on its first argument -- it must be saturated for these to kick in
372 interestingArgContext :: [CoreRule] -> SimplCont -> Bool
373 -- If the argument has form (f x y), where x,y are boring,
374 -- and f is marked INLINE, then we don't want to inline f.
375 -- But if the context of the argument is
377 -- where g has rules, then we *do* want to inline f, in case it
378 -- exposes a rule that might fire. Similarly, if the context is
380 -- where h has rules, then we do want to inline f; hence the
381 -- call_cont argument to interestingArgContext
383 -- The ai-rules flag makes this happen; if it's
384 -- set, the inliner gets just enough keener to inline f
385 -- regardless of how boring f's arguments are, if it's marked INLINE
387 -- The alternative would be to *always* inline an INLINE function,
388 -- regardless of how boring its context is; but that seems overkill
389 -- For example, it'd mean that wrapper functions were always inlined
390 interestingArgContext rules call_cont
391 = notNull rules || enclosing_fn_has_rules
393 enclosing_fn_has_rules = go call_cont
395 go (Select {}) = False
396 go (ApplyTo {}) = False
397 go (StrictArg _ cci _ _) = interesting cci
398 go (StrictBind {}) = False -- ??
399 go (CoerceIt _ c) = go c
400 go (Stop cci) = interesting cci
402 interesting (ArgCtxt rules) = rules
403 interesting _ = False
408 %************************************************************************
410 \subsection{Decisions about inlining}
412 %************************************************************************
415 simplEnvForGHCi :: SimplEnv
416 simplEnvForGHCi = mkSimplEnv allOffSwitchChecker $
417 SimplGently { sm_rules = False, sm_inline = False }
418 -- Do not do any inlining, in case we expose some unboxed
419 -- tuple stuff that confuses the bytecode interpreter
421 simplEnvForRules :: SimplEnv
422 simplEnvForRules = mkSimplEnv allOffSwitchChecker $
423 SimplGently { sm_rules = True, sm_inline = False }
425 updModeForInlineRules :: SimplifierMode -> SimplifierMode
426 updModeForInlineRules mode
428 SimplGently {} -> mode -- Don't modify mode if we already gentle
429 SimplPhase {} -> SimplGently { sm_rules = True, sm_inline = True }
430 -- Simplify as much as possible, subject to the usual "gentle" rules
433 Inlining is controlled partly by the SimplifierMode switch. This has two
436 SimplGently (a) Simplifying before specialiser/full laziness
437 (b) Simplifiying inside InlineRules
438 (c) Simplifying the LHS of a rule
439 (d) Simplifying a GHCi expression or Template
442 SimplPhase n _ Used at all other times
446 Gentle mode has a separate boolean flag to control
447 a) inlining (sm_inline flag)
448 b) rules (sm_rules flag)
449 A key invariant about Gentle mode is that it is treated as the EARLIEST
450 phase. Something is inlined if the sm_inline flag is on AND the thing
451 is inlinable in the earliest phase. This is important. Example
453 {-# INLINE [~1] g #-}
459 If we were to inline g into f's inlining, then an importing module would
461 f e --> g (g e) ---> RULE fires
462 because the InlineRule for f has had g inlined into it.
464 On the other hand, it is bad not to do ANY inlining into an
465 InlineRule, because then recursive knots in instance declarations
466 don't get unravelled.
468 However, *sometimes* SimplGently must do no call-site inlining at all.
469 Before full laziness we must be careful not to inline wrappers,
470 because doing so inhibits floating
471 e.g. ...(case f x of ...)...
472 ==> ...(case (case x of I# x# -> fw x#) of ...)...
473 ==> ...(case x of I# x# -> case fw x# of ...)...
474 and now the redex (f x) isn't floatable any more.
476 The no-inlining thing is also important for Template Haskell. You might be
477 compiling in one-shot mode with -O2; but when TH compiles a splice before
478 running it, we don't want to use -O2. Indeed, we don't want to inline
479 anything, because the byte-code interpreter might get confused about
480 unboxed tuples and suchlike.
482 Note [RULEs enabled in SimplGently]
483 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
484 RULES are enabled when doing "gentle" simplification. Two reasons:
486 * We really want the class-op cancellation to happen:
487 op (df d1 d2) --> $cop3 d1 d2
488 because this breaks the mutual recursion between 'op' and 'df'
492 to work in Template Haskell when simplifying
493 splices, so we get simpler code for literal strings
495 Note [Simplifying gently inside InlineRules]
496 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
497 We don't do much simplification inside InlineRules (which come from
498 INLINE pragmas). It really is important to switch off inlinings
499 inside such expressions. Consider the following example
505 in ...g...g...g...g...g...
507 Now, if that's the ONLY occurrence of f, it will be inlined inside g,
508 and thence copied multiple times when g is inlined.
510 This function may be inlinined in other modules, so we don't want to
511 remove (by inlining) calls to functions that have specialisations, or
512 that may have transformation rules in an importing scope.
514 E.g. {-# INLINE f #-}
517 and suppose that g is strict *and* has specialisations. If we inline
518 g's wrapper, we deny f the chance of getting the specialised version
519 of g when f is inlined at some call site (perhaps in some other
522 It's also important not to inline a worker back into a wrapper.
524 wraper = inline_me (\x -> ...worker... )
525 Normally, the inline_me prevents the worker getting inlined into
526 the wrapper (initially, the worker's only call site!). But,
527 if the wrapper is sure to be called, the strictness analyser will
528 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
529 continuation. That's why the keep_inline predicate returns True for
530 ArgOf continuations. It shouldn't do any harm not to dissolve the
531 inline-me note under these circumstances.
533 Although we do very little simplification inside an InlineRule,
534 the RHS is simplified as normal. For example:
536 all xs = foldr (&&) True xs
537 any p = all . map p {-# INLINE any #-}
539 The RHS of 'any' will get optimised and deforested; but the InlineRule
540 will still mention the original RHS.
543 preInlineUnconditionally
544 ~~~~~~~~~~~~~~~~~~~~~~~~
545 @preInlineUnconditionally@ examines a bndr to see if it is used just
546 once in a completely safe way, so that it is safe to discard the
547 binding inline its RHS at the (unique) usage site, REGARDLESS of how
548 big the RHS might be. If this is the case we don't simplify the RHS
549 first, but just inline it un-simplified.
551 This is much better than first simplifying a perhaps-huge RHS and then
552 inlining and re-simplifying it. Indeed, it can be at least quadratically
561 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
562 This can happen with cascades of functions too:
569 THE MAIN INVARIANT is this:
571 ---- preInlineUnconditionally invariant -----
572 IF preInlineUnconditionally chooses to inline x = <rhs>
573 THEN doing the inlining should not change the occurrence
574 info for the free vars of <rhs>
575 ----------------------------------------------
577 For example, it's tempting to look at trivial binding like
579 and inline it unconditionally. But suppose x is used many times,
580 but this is the unique occurrence of y. Then inlining x would change
581 y's occurrence info, which breaks the invariant. It matters: y
582 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
585 Even RHSs labelled InlineMe aren't caught here, because there might be
586 no benefit from inlining at the call site.
588 [Sept 01] Don't unconditionally inline a top-level thing, because that
589 can simply make a static thing into something built dynamically. E.g.
593 [Remember that we treat \s as a one-shot lambda.] No point in
594 inlining x unless there is something interesting about the call site.
596 But watch out: if you aren't careful, some useful foldr/build fusion
597 can be lost (most notably in spectral/hartel/parstof) because the
598 foldr didn't see the build. Doing the dynamic allocation isn't a big
599 deal, in fact, but losing the fusion can be. But the right thing here
600 seems to be to do a callSiteInline based on the fact that there is
601 something interesting about the call site (it's strict). Hmm. That
604 Conclusion: inline top level things gaily until Phase 0 (the last
605 phase), at which point don't.
607 Note [pre/postInlineUnconditionally in gentle mode]
608 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
609 Even in gentle mode we want to do preInlineUnconditionally. The
610 reason is that too little clean-up happens if you don't inline
611 use-once things. Also a bit of inlining is *good* for full laziness;
612 it can expose constant sub-expressions. Example in
613 spectral/mandel/Mandel.hs, where the mandelset function gets a useful
614 let-float if you inline windowToViewport
616 However, as usual for Gentle mode, do not inline things that are
617 inactive in the intial stages. See Note [Gentle mode].
620 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
621 preInlineUnconditionally env top_lvl bndr rhs
623 | opt_SimplNoPreInlining = False
624 | otherwise = case idOccInfo bndr of
625 IAmDead -> True -- Happens in ((\x.1) v)
626 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
630 active = case phase of
631 SimplGently {} -> isEarlyActive act
632 -- See Note [pre/postInlineUnconditionally in gentle mode]
633 SimplPhase n _ -> isActive n act
634 act = idInlineActivation bndr
636 try_once in_lam int_cxt -- There's one textual occurrence
637 | not in_lam = isNotTopLevel top_lvl || early_phase
638 | otherwise = int_cxt && canInlineInLam rhs
640 -- Be very careful before inlining inside a lambda, becuase (a) we must not
641 -- invalidate occurrence information, and (b) we want to avoid pushing a
642 -- single allocation (here) into multiple allocations (inside lambda).
643 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
644 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
646 -- is_cheap = exprIsCheap rhs
647 -- ok = is_cheap && int_cxt
649 -- int_cxt The context isn't totally boring
650 -- E.g. let f = \ab.BIG in \y. map f xs
651 -- Don't want to substitute for f, because then we allocate
652 -- its closure every time the \y is called
653 -- But: let f = \ab.BIG in \y. map (f y) xs
654 -- Now we do want to substitute for f, even though it's not
655 -- saturated, because we're going to allocate a closure for
656 -- (f y) every time round the loop anyhow.
658 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
659 -- so substituting rhs inside a lambda doesn't change the occ info.
660 -- Sadly, not quite the same as exprIsHNF.
661 canInlineInLam (Lit _) = True
662 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
663 canInlineInLam (Note _ e) = canInlineInLam e
664 canInlineInLam _ = False
666 early_phase = case phase of
667 SimplPhase 0 _ -> False
669 -- If we don't have this early_phase test, consider
670 -- x = length [1,2,3]
671 -- The full laziness pass carefully floats all the cons cells to
672 -- top level, and preInlineUnconditionally floats them all back in.
673 -- Result is (a) static allocation replaced by dynamic allocation
674 -- (b) many simplifier iterations because this tickles
675 -- a related problem; only one inlining per pass
677 -- On the other hand, I have seen cases where top-level fusion is
678 -- lost if we don't inline top level thing (e.g. string constants)
679 -- Hence the test for phase zero (which is the phase for all the final
680 -- simplifications). Until phase zero we take no special notice of
681 -- top level things, but then we become more leery about inlining
686 postInlineUnconditionally
687 ~~~~~~~~~~~~~~~~~~~~~~~~~
688 @postInlineUnconditionally@ decides whether to unconditionally inline
689 a thing based on the form of its RHS; in particular if it has a
690 trivial RHS. If so, we can inline and discard the binding altogether.
692 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
693 only have *forward* references Hence, it's safe to discard the binding
695 NOTE: This isn't our last opportunity to inline. We're at the binding
696 site right now, and we'll get another opportunity when we get to the
699 Note that we do this unconditional inlining only for trival RHSs.
700 Don't inline even WHNFs inside lambdas; doing so may simply increase
701 allocation when the function is called. This isn't the last chance; see
704 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
705 Because we don't even want to inline them into the RHS of constructor
706 arguments. See NOTE above
708 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
709 it's best to inline it anyway. We often get a=E; b=a from desugaring,
710 with both a and b marked NOINLINE. But that seems incompatible with
711 our new view that inlining is like a RULE, so I'm sticking to the 'active'
715 postInlineUnconditionally
716 :: SimplEnv -> TopLevelFlag
717 -> OutId -- The binder (an InId would be fine too)
718 -> OccInfo -- From the InId
722 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
724 | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, don't inline
725 -- because it might be referred to "earlier"
726 | isExportedId bndr = False
727 | isInlineRule unfolding = False -- Note [InlineRule and postInlineUnconditionally]
728 | exprIsTrivial rhs = True
731 -- The point of examining occ_info here is that for *non-values*
732 -- that occur outside a lambda, the call-site inliner won't have
733 -- a chance (becuase it doesn't know that the thing
734 -- only occurs once). The pre-inliner won't have gotten
735 -- it either, if the thing occurs in more than one branch
736 -- So the main target is things like
739 -- True -> case x of ...
740 -- False -> case x of ...
741 -- I'm not sure how important this is in practice
742 OneOcc in_lam _one_br int_cxt -- OneOcc => no code-duplication issue
743 -> smallEnoughToInline unfolding -- Small enough to dup
744 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
746 -- NB: Do NOT inline arbitrarily big things, even if one_br is True
747 -- Reason: doing so risks exponential behaviour. We simplify a big
748 -- expression, inline it, and simplify it again. But if the
749 -- very same thing happens in the big expression, we get
751 -- PRINCIPLE: when we've already simplified an expression once,
752 -- make sure that we only inline it if it's reasonably small.
754 && ((isNotTopLevel top_lvl && not in_lam) ||
755 -- But outside a lambda, we want to be reasonably aggressive
756 -- about inlining into multiple branches of case
757 -- e.g. let x = <non-value>
758 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
759 -- Inlining can be a big win if C3 is the hot-spot, even if
760 -- the uses in C1, C2 are not 'interesting'
761 -- An example that gets worse if you add int_cxt here is 'clausify'
763 (isCheapUnfolding unfolding && int_cxt))
764 -- isCheap => acceptable work duplication; in_lam may be true
765 -- int_cxt to prevent us inlining inside a lambda without some
766 -- good reason. See the notes on int_cxt in preInlineUnconditionally
768 IAmDead -> True -- This happens; for example, the case_bndr during case of
769 -- known constructor: case (a,b) of x { (p,q) -> ... }
770 -- Here x isn't mentioned in the RHS, so we don't want to
771 -- create the (dead) let-binding let x = (a,b) in ...
775 -- Here's an example that we don't handle well:
776 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
777 -- in \y. ....case f of {...} ....
778 -- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
780 -- - We can't preInlineUnconditionally because that woud invalidate
781 -- the occ info for b.
782 -- - We can't postInlineUnconditionally because the RHS is big, and
783 -- that risks exponential behaviour
784 -- - We can't call-site inline, because the rhs is big
788 active = case getMode env of
789 SimplGently {} -> isEarlyActive act
790 -- See Note [pre/postInlineUnconditionally in gentle mode]
791 SimplPhase n _ -> isActive n act
792 act = idInlineActivation bndr
794 activeInline :: SimplEnv -> OutId -> Bool
796 | isNonRuleLoopBreaker (idOccInfo id) -- Things with an INLINE pragma may have
797 -- an unfolding *and* be a loop breaker
798 = False -- (maybe the knot is not yet untied)
800 = case getMode env of
801 SimplGently { sm_inline = inlining_on }
802 -> inlining_on && isEarlyActive act
803 -- See Note [Gentle mode]
805 -- NB: we used to have a second exception, for data con wrappers.
806 -- On the grounds that we use gentle mode for rule LHSs, and
807 -- they match better when data con wrappers are inlined.
808 -- But that only really applies to the trivial wrappers (like (:)),
809 -- and they are now constructed as Compulsory unfoldings (in MkId)
810 -- so they'll happen anyway.
812 SimplPhase n _ -> isActive n act
814 act = idInlineActivation id
816 activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
817 -- Nothing => No rules at all
818 activeRule dflags env
819 | not (dopt Opt_EnableRewriteRules dflags)
820 = Nothing -- Rewriting is off
822 = case getMode env of
823 SimplGently { sm_rules = rules_on }
824 | rules_on -> Just isEarlyActive -- Note [RULEs enabled in SimplGently]
825 | otherwise -> Nothing
826 SimplPhase n _ -> Just (isActive n)
829 Note [InlineRule and postInlineUnconditionally]
830 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
831 Do not do postInlineUnconditionally if the Id has an InlineRule, otherwise
832 we lose the unfolding. Example
834 -- f has InlineRule with rhs (e |> co)
838 Then there's a danger we'll optimise to
843 and now postInlineUnconditionally, losing the InlineRule on f. Now f'
844 won't inline because 'e' is too big.
847 %************************************************************************
851 %************************************************************************
854 mkLam :: SimplEnv -> [OutBndr] -> OutExpr -> SimplM OutExpr
855 -- mkLam tries three things
856 -- a) eta reduction, if that gives a trivial expression
857 -- b) eta expansion [only if there are some value lambdas]
862 = do { dflags <- getDOptsSmpl
863 ; mkLam' dflags bndrs body }
865 mkLam' :: DynFlags -> [OutBndr] -> OutExpr -> SimplM OutExpr
866 mkLam' dflags bndrs (Cast body co)
867 | not (any bad bndrs)
868 -- Note [Casts and lambdas]
869 = do { lam <- mkLam' dflags bndrs body
870 ; return (mkCoerce (mkPiTypes bndrs co) lam) }
872 co_vars = tyVarsOfType co
873 bad bndr = isCoVar bndr && bndr `elemVarSet` co_vars
875 mkLam' dflags bndrs body
876 | dopt Opt_DoEtaReduction dflags,
877 Just etad_lam <- tryEtaReduce bndrs body
878 = do { tick (EtaReduction (head bndrs))
881 | dopt Opt_DoLambdaEtaExpansion dflags,
882 not (inGentleMode env), -- In gentle mode don't eta-expansion
883 any isRuntimeVar bndrs -- because it can clutter up the code
884 -- with casts etc that may not be removed
885 = do { let body' = tryEtaExpansion dflags body
886 ; return (mkLams bndrs body') }
889 = return (mkLams bndrs body)
892 Note [Casts and lambdas]
893 ~~~~~~~~~~~~~~~~~~~~~~~~
895 (\x. (\y. e) `cast` g1) `cast` g2
896 There is a danger here that the two lambdas look separated, and the
897 full laziness pass might float an expression to between the two.
899 So this equation in mkLam' floats the g1 out, thus:
900 (\x. e `cast` g1) --> (\x.e) `cast` (tx -> g1)
903 In general, this floats casts outside lambdas, where (I hope) they
904 might meet and cancel with some other cast:
905 \x. e `cast` co ===> (\x. e) `cast` (tx -> co)
906 /\a. e `cast` co ===> (/\a. e) `cast` (/\a. co)
907 /\g. e `cast` co ===> (/\g. e) `cast` (/\g. co)
910 Notice that it works regardless of 'e'. Originally it worked only
911 if 'e' was itself a lambda, but in some cases that resulted in
912 fruitless iteration in the simplifier. A good example was when
913 compiling Text.ParserCombinators.ReadPrec, where we had a definition
914 like (\x. Get `cast` g)
915 where Get is a constructor with nonzero arity. Then mkLam eta-expanded
916 the Get, and the next iteration eta-reduced it, and then eta-expanded
919 Note also the side condition for the case of coercion binders.
920 It does not make sense to transform
921 /\g. e `cast` g ==> (/\g.e) `cast` (/\g.g)
922 because the latter is not well-kinded.
924 -- c) floating lets out through big lambdas
925 -- [only if all tyvar lambdas, and only if this lambda
926 -- is the RHS of a let]
928 {- Sept 01: I'm experimenting with getting the
929 full laziness pass to float out past big lambdsa
930 | all isTyVar bndrs, -- Only for big lambdas
931 contIsRhs cont -- Only try the rhs type-lambda floating
932 -- if this is indeed a right-hand side; otherwise
933 -- we end up floating the thing out, only for float-in
934 -- to float it right back in again!
935 = do (floats, body') <- tryRhsTyLam env bndrs body
936 return (floats, mkLams bndrs body')
940 %************************************************************************
944 %************************************************************************
946 Note [Eta reduction conditions]
947 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
948 We try for eta reduction here, but *only* if we get all the way to an
949 trivial expression. We don't want to remove extra lambdas unless we
950 are going to avoid allocating this thing altogether.
952 There are some particularly delicate points here:
954 * Eta reduction is not valid in general:
956 This matters, partly for old-fashioned correctness reasons but,
957 worse, getting it wrong can yield a seg fault. Consider
959 h y = case (case y of { True -> f `seq` True; False -> False }) of
960 True -> ...; False -> ...
962 If we (unsoundly) eta-reduce f to get f=f, the strictness analyser
963 says f=bottom, and replaces the (f `seq` True) with just
964 (f `cast` unsafe-co). BUT, as thing stand, 'f' got arity 1, and it
965 *keeps* arity 1 (perhaps also wrongly). So CorePrep eta-expands
966 the definition again, so that it does not termninate after all.
967 Result: seg-fault because the boolean case actually gets a function value.
970 So it's important to to the right thing.
972 * Note [Arity care]: we need to be careful if we just look at f's
973 arity. Currently (Dec07), f's arity is visible in its own RHS (see
974 Note [Arity robustness] in SimplEnv) so we must *not* trust the
975 arity when checking that 'f' is a value. Otherwise we will
980 Which might change a terminiating program (think (f `seq` e)) to a
981 non-terminating one. So we check for being a loop breaker first.
983 However for GlobalIds we can look at the arity; and for primops we
984 must, since they have no unfolding.
986 * Regardless of whether 'f' is a value, we always want to
987 reduce (/\a -> f a) to f
988 This came up in a RULE: foldr (build (/\a -> g a))
989 did not match foldr (build (/\b -> ...something complex...))
990 The type checker can insert these eta-expanded versions,
991 with both type and dictionary lambdas; hence the slightly
994 * Never *reduce* arity. For example
996 Then if h has arity 1 we don't want to eta-reduce because then
997 f's arity would decrease, and that is bad
999 These delicacies are why we don't use exprIsTrivial and exprIsHNF here.
1003 tryEtaReduce :: [OutBndr] -> OutExpr -> Maybe OutExpr
1004 tryEtaReduce bndrs body
1005 = go (reverse bndrs) body
1007 incoming_arity = count isId bndrs
1009 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
1010 go [] fun | ok_fun fun = Just fun -- Success!
1011 go _ _ = Nothing -- Failure!
1013 -- Note [Eta reduction conditions]
1014 ok_fun (App fun (Type ty))
1015 | not (any (`elemVarSet` tyVarsOfType ty) bndrs)
1018 = not (fun_id `elem` bndrs)
1019 && (ok_fun_id fun_id || all ok_lam bndrs)
1022 ok_fun_id fun = fun_arity fun >= incoming_arity
1024 fun_arity fun -- See Note [Arity care]
1025 | isLocalId fun && isLoopBreaker (idOccInfo fun) = 0
1026 | otherwise = idArity fun
1028 ok_lam v = isTyVar v || isDictId v
1030 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
1034 %************************************************************************
1038 %************************************************************************
1042 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
1045 where (in both cases)
1047 * The xi can include type variables
1049 * The yi are all value variables
1051 * N is a NORMAL FORM (i.e. no redexes anywhere)
1052 wanting a suitable number of extra args.
1054 The biggest reason for doing this is for cases like
1060 Here we want to get the lambdas together. A good exmaple is the nofib
1061 program fibheaps, which gets 25% more allocation if you don't do this
1064 We may have to sandwich some coerces between the lambdas
1065 to make the types work. exprEtaExpandArity looks through coerces
1066 when computing arity; and etaExpand adds the coerces as necessary when
1067 actually computing the expansion.
1070 tryEtaExpansion :: DynFlags -> OutExpr -> OutExpr
1071 -- There is at least one runtime binder in the binders
1072 tryEtaExpansion dflags body
1073 = etaExpand fun_arity body
1075 fun_arity = exprEtaExpandArity dflags body
1079 %************************************************************************
1081 \subsection{Floating lets out of big lambdas}
1083 %************************************************************************
1085 Note [Floating and type abstraction]
1086 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1089 We'd like to float this to
1092 x = /\a. C (y1 a) (y2 a)
1093 for the usual reasons: we want to inline x rather vigorously.
1095 You may think that this kind of thing is rare. But in some programs it is
1096 common. For example, if you do closure conversion you might get:
1098 data a :-> b = forall e. (e -> a -> b) :$ e
1100 f_cc :: forall a. a :-> a
1101 f_cc = /\a. (\e. id a) :$ ()
1103 Now we really want to inline that f_cc thing so that the
1104 construction of the closure goes away.
1106 So I have elaborated simplLazyBind to understand right-hand sides that look
1110 and treat them specially. The real work is done in SimplUtils.abstractFloats,
1111 but there is quite a bit of plumbing in simplLazyBind as well.
1113 The same transformation is good when there are lets in the body:
1115 /\abc -> let(rec) x = e in b
1117 let(rec) x' = /\abc -> let x = x' a b c in e
1119 /\abc -> let x = x' a b c in b
1121 This is good because it can turn things like:
1123 let f = /\a -> letrec g = ... g ... in g
1125 letrec g' = /\a -> ... g' a ...
1127 let f = /\ a -> g' a
1129 which is better. In effect, it means that big lambdas don't impede
1132 This optimisation is CRUCIAL in eliminating the junk introduced by
1133 desugaring mutually recursive definitions. Don't eliminate it lightly!
1135 [May 1999] If we do this transformation *regardless* then we can
1136 end up with some pretty silly stuff. For example,
1139 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
1144 st = /\s -> ...[y1 s/x1, y2 s/x2]
1147 Unless the "..." is a WHNF there is really no point in doing this.
1148 Indeed it can make things worse. Suppose x1 is used strictly,
1151 x1* = case f y of { (a,b) -> e }
1153 If we abstract this wrt the tyvar we then can't do the case inline
1154 as we would normally do.
1156 That's why the whole transformation is part of the same process that
1157 floats let-bindings and constructor arguments out of RHSs. In particular,
1158 it is guarded by the doFloatFromRhs call in simplLazyBind.
1162 abstractFloats :: [OutTyVar] -> SimplEnv -> OutExpr -> SimplM ([OutBind], OutExpr)
1163 abstractFloats main_tvs body_env body
1164 = ASSERT( notNull body_floats )
1165 do { (subst, float_binds) <- mapAccumLM abstract empty_subst body_floats
1166 ; return (float_binds, CoreSubst.substExpr subst body) }
1168 main_tv_set = mkVarSet main_tvs
1169 body_floats = getFloats body_env
1170 empty_subst = CoreSubst.mkEmptySubst (seInScope body_env)
1172 abstract :: CoreSubst.Subst -> OutBind -> SimplM (CoreSubst.Subst, OutBind)
1173 abstract subst (NonRec id rhs)
1174 = do { (poly_id, poly_app) <- mk_poly tvs_here id
1175 ; let poly_rhs = mkLams tvs_here rhs'
1176 subst' = CoreSubst.extendIdSubst subst id poly_app
1177 ; return (subst', (NonRec poly_id poly_rhs)) }
1179 rhs' = CoreSubst.substExpr subst rhs
1180 tvs_here | any isCoVar main_tvs = main_tvs -- Note [Abstract over coercions]
1182 = varSetElems (main_tv_set `intersectVarSet` exprSomeFreeVars isTyVar rhs')
1184 -- Abstract only over the type variables free in the rhs
1185 -- wrt which the new binding is abstracted. But the naive
1186 -- approach of abstract wrt the tyvars free in the Id's type
1188 -- /\ a b -> let t :: (a,b) = (e1, e2)
1191 -- Here, b isn't free in x's type, but we must nevertheless
1192 -- abstract wrt b as well, because t's type mentions b.
1193 -- Since t is floated too, we'd end up with the bogus:
1194 -- poly_t = /\ a b -> (e1, e2)
1195 -- poly_x = /\ a -> fst (poly_t a *b*)
1196 -- So for now we adopt the even more naive approach of
1197 -- abstracting wrt *all* the tyvars. We'll see if that
1198 -- gives rise to problems. SLPJ June 98
1200 abstract subst (Rec prs)
1201 = do { (poly_ids, poly_apps) <- mapAndUnzipM (mk_poly tvs_here) ids
1202 ; let subst' = CoreSubst.extendSubstList subst (ids `zip` poly_apps)
1203 poly_rhss = [mkLams tvs_here (CoreSubst.substExpr subst' rhs) | rhs <- rhss]
1204 ; return (subst', Rec (poly_ids `zip` poly_rhss)) }
1206 (ids,rhss) = unzip prs
1207 -- For a recursive group, it's a bit of a pain to work out the minimal
1208 -- set of tyvars over which to abstract:
1209 -- /\ a b c. let x = ...a... in
1210 -- letrec { p = ...x...q...
1211 -- q = .....p...b... } in
1213 -- Since 'x' is abstracted over 'a', the {p,q} group must be abstracted
1214 -- over 'a' (because x is replaced by (poly_x a)) as well as 'b'.
1215 -- Since it's a pain, we just use the whole set, which is always safe
1217 -- If you ever want to be more selective, remember this bizarre case too:
1219 -- Here, we must abstract 'x' over 'a'.
1222 mk_poly tvs_here var
1223 = do { uniq <- getUniqueM
1224 ; let poly_name = setNameUnique (idName var) uniq -- Keep same name
1225 poly_ty = mkForAllTys tvs_here (idType var) -- But new type of course
1226 poly_id = transferPolyIdInfo var tvs_here $ -- Note [transferPolyIdInfo] in Id.lhs
1227 mkLocalId poly_name poly_ty
1228 ; return (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tvs_here)) }
1229 -- In the olden days, it was crucial to copy the occInfo of the original var,
1230 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1231 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1232 -- at already simplified code, so it doesn't matter
1234 -- It's even right to retain single-occurrence or dead-var info:
1235 -- Suppose we started with /\a -> let x = E in B
1236 -- where x occurs once in B. Then we transform to:
1237 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1238 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1239 -- the occurrences of x' will be just the occurrences originally
1243 Note [Abstract over coercions]
1244 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1245 If a coercion variable (g :: a ~ Int) is free in the RHS, then so is the
1246 type variable a. Rather than sort this mess out, we simply bale out and abstract
1247 wrt all the type variables if any of them are coercion variables.
1250 Historical note: if you use let-bindings instead of a substitution, beware of this:
1252 -- Suppose we start with:
1254 -- x = /\ a -> let g = G in E
1256 -- Then we'll float to get
1258 -- x = let poly_g = /\ a -> G
1259 -- in /\ a -> let g = poly_g a in E
1261 -- But now the occurrence analyser will see just one occurrence
1262 -- of poly_g, not inside a lambda, so the simplifier will
1263 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1264 -- (I used to think that the "don't inline lone occurrences" stuff
1265 -- would stop this happening, but since it's the *only* occurrence,
1266 -- PreInlineUnconditionally kicks in first!)
1268 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1269 -- to appear many times. (NB: mkInlineMe eliminates
1270 -- such notes on trivial RHSs, so do it manually.)
1272 %************************************************************************
1276 %************************************************************************
1278 prepareAlts tries these things:
1280 1. If several alternatives are identical, merge them into
1281 a single DEFAULT alternative. I've occasionally seen this
1282 making a big difference:
1284 case e of =====> case e of
1285 C _ -> f x D v -> ....v....
1286 D v -> ....v.... DEFAULT -> f x
1289 The point is that we merge common RHSs, at least for the DEFAULT case.
1290 [One could do something more elaborate but I've never seen it needed.]
1291 To avoid an expensive test, we just merge branches equal to the *first*
1292 alternative; this picks up the common cases
1293 a) all branches equal
1294 b) some branches equal to the DEFAULT (which occurs first)
1297 case e of b { ==> case e of b {
1298 p1 -> rhs1 p1 -> rhs1
1300 pm -> rhsm pm -> rhsm
1301 _ -> case b of b' { pn -> let b'=b in rhsn
1303 ... po -> let b'=b in rhso
1304 po -> rhso _ -> let b'=b in rhsd
1308 which merges two cases in one case when -- the default alternative of
1309 the outer case scrutises the same variable as the outer case This
1310 transformation is called Case Merging. It avoids that the same
1311 variable is scrutinised multiple times.
1314 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
1320 where @is@ was something like
1322 p `is` n = p /= (-1) && p == n
1324 This gave rise to a horrible sequence of cases
1331 and similarly in cascade for all the join points!
1334 ~~~~~~~~~~~~~~~~~~~~
1335 We do this *here*, looking at un-simplified alternatives, because we
1336 have to check that r doesn't mention the variables bound by the
1337 pattern in each alternative, so the binder-info is rather useful.
1340 prepareAlts :: SimplEnv -> OutExpr -> OutId -> [InAlt] -> SimplM ([AltCon], [InAlt])
1341 prepareAlts env scrut case_bndr' alts
1342 = do { dflags <- getDOptsSmpl
1343 ; alts <- combineIdenticalAlts case_bndr' alts
1345 ; let (alts_wo_default, maybe_deflt) = findDefault alts
1346 alt_cons = [con | (con,_,_) <- alts_wo_default]
1347 imposs_deflt_cons = nub (imposs_cons ++ alt_cons)
1348 -- "imposs_deflt_cons" are handled
1349 -- EITHER by the context,
1350 -- OR by a non-DEFAULT branch in this case expression.
1352 ; default_alts <- prepareDefault dflags env case_bndr' mb_tc_app
1353 imposs_deflt_cons maybe_deflt
1355 ; let trimmed_alts = filterOut impossible_alt alts_wo_default
1356 merged_alts = mergeAlts trimmed_alts default_alts
1357 -- We need the mergeAlts in case the new default_alt
1358 -- has turned into a constructor alternative.
1359 -- The merge keeps the inner DEFAULT at the front, if there is one
1360 -- and interleaves the alternatives in the right order
1362 ; return (imposs_deflt_cons, merged_alts) }
1364 mb_tc_app = splitTyConApp_maybe (idType case_bndr')
1365 Just (_, inst_tys) = mb_tc_app
1367 imposs_cons = case scrut of
1368 Var v -> otherCons (idUnfolding v)
1371 impossible_alt :: CoreAlt -> Bool
1372 impossible_alt (con, _, _) | con `elem` imposs_cons = True
1373 impossible_alt (DataAlt con, _, _) = dataConCannotMatch inst_tys con
1374 impossible_alt _ = False
1377 --------------------------------------------------
1378 -- 1. Merge identical branches
1379 --------------------------------------------------
1380 combineIdenticalAlts :: OutId -> [InAlt] -> SimplM [InAlt]
1382 combineIdenticalAlts case_bndr ((_con1,bndrs1,rhs1) : con_alts)
1383 | all isDeadBinder bndrs1, -- Remember the default
1384 length filtered_alts < length con_alts -- alternative comes first
1385 -- Also Note [Dead binders]
1386 = do { tick (AltMerge case_bndr)
1387 ; return ((DEFAULT, [], rhs1) : filtered_alts) }
1389 filtered_alts = filter keep con_alts
1390 keep (_con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1392 combineIdenticalAlts _ alts = return alts
1394 -------------------------------------------------------------------------
1395 -- Prepare the default alternative
1396 -------------------------------------------------------------------------
1397 prepareDefault :: DynFlags
1399 -> OutId -- Case binder; need just for its type. Note that as an
1400 -- OutId, it has maximum information; this is important.
1401 -- Test simpl013 is an example
1402 -> Maybe (TyCon, [Type]) -- Type of scrutinee, decomposed
1403 -> [AltCon] -- These cons can't happen when matching the default
1404 -> Maybe InExpr -- Rhs
1405 -> SimplM [InAlt] -- Still unsimplified
1406 -- We use a list because it's what mergeAlts expects,
1407 -- And becuase case-merging can cause many to show up
1409 ------- Merge nested cases ----------
1410 prepareDefault dflags env outer_bndr _bndr_ty imposs_cons (Just deflt_rhs)
1411 | dopt Opt_CaseMerge dflags
1412 , Case (Var inner_scrut_var) inner_bndr _ inner_alts <- deflt_rhs
1413 , DoneId inner_scrut_var' <- substId env inner_scrut_var
1414 -- Remember, inner_scrut_var is an InId, but outer_bndr is an OutId
1415 , inner_scrut_var' == outer_bndr
1416 -- NB: the substId means that if the outer scrutinee was a
1417 -- variable, and inner scrutinee is the same variable,
1418 -- then inner_scrut_var' will be outer_bndr
1419 -- via the magic of simplCaseBinder
1420 = do { tick (CaseMerge outer_bndr)
1422 ; let munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1423 ; return [(con, args, munge_rhs rhs) | (con, args, rhs) <- inner_alts,
1424 not (con `elem` imposs_cons) ]
1425 -- NB: filter out any imposs_cons. Example:
1428 -- DEFAULT -> case x of
1431 -- When we merge, we must ensure that e1 takes
1432 -- precedence over e2 as the value for A!
1434 -- Warning: don't call prepareAlts recursively!
1435 -- Firstly, there's no point, because inner alts have already had
1436 -- mkCase applied to them, so they won't have a case in their default
1437 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1438 -- in munge_rhs may put a case into the DEFAULT branch!
1441 --------- Fill in known constructor -----------
1442 prepareDefault _ _ case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
1443 | -- This branch handles the case where we are
1444 -- scrutinisng an algebraic data type
1445 isAlgTyCon tycon -- It's a data type, tuple, or unboxed tuples.
1446 , not (isNewTyCon tycon) -- We can have a newtype, if we are just doing an eval:
1447 -- case x of { DEFAULT -> e }
1448 -- and we don't want to fill in a default for them!
1449 , Just all_cons <- tyConDataCons_maybe tycon
1450 , not (null all_cons) -- This is a tricky corner case. If the data type has no constructors,
1451 -- which GHC allows, then the case expression will have at most a default
1452 -- alternative. We don't want to eliminate that alternative, because the
1453 -- invariant is that there's always one alternative. It's more convenient
1455 -- case x of { DEFAULT -> e }
1456 -- as it is, rather than transform it to
1457 -- error "case cant match"
1458 -- which would be quite legitmate. But it's a really obscure corner, and
1459 -- not worth wasting code on.
1460 , let imposs_data_cons = [con | DataAlt con <- imposs_cons] -- We now know it's a data type
1461 impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
1462 = case filterOut impossible all_cons of
1463 [] -> return [] -- Eliminate the default alternative
1464 -- altogether if it can't match
1466 [con] -> -- It matches exactly one constructor, so fill it in
1467 do { tick (FillInCaseDefault case_bndr)
1469 ; let (ex_tvs, co_tvs, arg_ids) =
1470 dataConRepInstPat us con inst_tys
1471 ; return [(DataAlt con, ex_tvs ++ co_tvs ++ arg_ids, deflt_rhs)] }
1473 _ -> return [(DEFAULT, [], deflt_rhs)]
1475 | debugIsOn, isAlgTyCon tycon, not (isOpenTyCon tycon), null (tyConDataCons tycon)
1476 -- This can legitimately happen for type families, so don't report that
1477 = pprTrace "prepareDefault" (ppr case_bndr <+> ppr tycon)
1478 $ return [(DEFAULT, [], deflt_rhs)]
1480 --------- Catch-all cases -----------
1481 prepareDefault _dflags _env _case_bndr _bndr_ty _imposs_cons (Just deflt_rhs)
1482 = return [(DEFAULT, [], deflt_rhs)]
1484 prepareDefault _dflags _env _case_bndr _bndr_ty _imposs_cons Nothing
1485 = return [] -- No default branch
1490 =================================================================================
1492 mkCase tries these things
1494 1. Eliminate the case altogether if possible
1502 and similar friends.
1506 mkCase :: OutExpr -> OutId -> [OutAlt] -- Increasing order
1509 --------------------------------------------------
1511 --------------------------------------------------
1513 mkCase scrut case_bndr alts -- Identity case
1514 | all identity_alt alts
1515 = do tick (CaseIdentity case_bndr)
1516 return (re_cast scrut)
1518 identity_alt (con, args, rhs) = check_eq con args (de_cast rhs)
1520 check_eq DEFAULT _ (Var v) = v == case_bndr
1521 check_eq (LitAlt lit') _ (Lit lit) = lit == lit'
1522 check_eq (DataAlt con) args rhs = rhs `cheapEqExpr` mkConApp con (arg_tys ++ varsToCoreExprs args)
1523 || rhs `cheapEqExpr` Var case_bndr
1524 check_eq _ _ _ = False
1526 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1529 -- case e of x { _ -> x `cast` c }
1530 -- And we definitely want to eliminate this case, to give
1532 -- So we throw away the cast from the RHS, and reconstruct
1533 -- it at the other end. All the RHS casts must be the same
1534 -- if (all identity_alt alts) holds.
1536 -- Don't worry about nested casts, because the simplifier combines them
1537 de_cast (Cast e _) = e
1540 re_cast scrut = case head alts of
1541 (_,_,Cast _ co) -> Cast scrut co
1546 --------------------------------------------------
1548 --------------------------------------------------
1549 mkCase scrut bndr alts = return (Case scrut bndr (coreAltsType alts) alts)
1553 When adding auxiliary bindings for the case binder, it's worth checking if
1554 its dead, because it often is, and occasionally these mkCase transformations
1555 cascade rather nicely.
1558 bindCaseBndr :: Id -> CoreExpr -> CoreExpr -> CoreExpr
1559 bindCaseBndr bndr rhs body
1560 | isDeadBinder bndr = body
1561 | otherwise = bindNonRec bndr rhs body