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, simplGentlyForInlineRules,
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 simplGentlyForInlineRules :: SimplifierMode
426 simplGentlyForInlineRules = SimplGently { sm_rules = True, sm_inline = True }
427 -- Simplify as much as possible, subject to the usual "gentle" rules
430 Inlining is controlled partly by the SimplifierMode switch. This has two
433 SimplGently (a) Simplifying before specialiser/full laziness
434 (b) Simplifiying inside InlineRules
435 (c) Simplifying the LHS of a rule
436 (d) Simplifying a GHCi expression or Template
439 SimplPhase n _ Used at all other times
443 Gentle mode has a separate boolean flag to control
444 a) inlining (sm_inline flag)
445 b) rules (sm_rules flag)
446 A key invariant about Gentle mode is that it is treated as the EARLIEST
447 phase. Something is inlined if the sm_inline flag is on AND the thing
448 is inlinable in the earliest phase. This is important. Example
450 {-# INLINE [~1] g #-}
456 If we were to inline g into f's inlining, then an importing module would
458 f e --> g (g e) ---> RULE fires
459 because the InlineRule for f has had g inlined into it.
461 On the other hand, it is bad not to do ANY inlining into an
462 InlineRule, because then recursive knots in instance declarations
463 don't get unravelled.
465 However, *sometimes* SimplGently must do no call-site inlining at all.
466 Before full laziness we must be careful not to inline wrappers,
467 because doing so inhibits floating
468 e.g. ...(case f x of ...)...
469 ==> ...(case (case x of I# x# -> fw x#) of ...)...
470 ==> ...(case x of I# x# -> case fw x# of ...)...
471 and now the redex (f x) isn't floatable any more.
473 The no-inlining thing is also important for Template Haskell. You might be
474 compiling in one-shot mode with -O2; but when TH compiles a splice before
475 running it, we don't want to use -O2. Indeed, we don't want to inline
476 anything, because the byte-code interpreter might get confused about
477 unboxed tuples and suchlike.
479 Note [Simplifying gently inside InlineRules]
480 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
481 We don't do much simplification inside InlineRules (which come from
482 INLINE pragmas). It really is important to switch off inlinings
483 inside such expressions. Consider the following example
489 in ...g...g...g...g...g...
491 Now, if that's the ONLY occurrence of f, it will be inlined inside g,
492 and thence copied multiple times when g is inlined.
494 This function may be inlinined in other modules, so we don't want to
495 remove (by inlining) calls to functions that have specialisations, or
496 that may have transformation rules in an importing scope.
498 E.g. {-# INLINE f #-}
501 and suppose that g is strict *and* has specialisations. If we inline
502 g's wrapper, we deny f the chance of getting the specialised version
503 of g when f is inlined at some call site (perhaps in some other
506 It's also important not to inline a worker back into a wrapper.
508 wraper = inline_me (\x -> ...worker... )
509 Normally, the inline_me prevents the worker getting inlined into
510 the wrapper (initially, the worker's only call site!). But,
511 if the wrapper is sure to be called, the strictness analyser will
512 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
513 continuation. That's why the keep_inline predicate returns True for
514 ArgOf continuations. It shouldn't do any harm not to dissolve the
515 inline-me note under these circumstances.
517 Although we do very little simplification inside an InlineRule,
518 the RHS is simplified as normal. For example:
520 all xs = foldr (&&) True xs
521 any p = all . map p {-# INLINE any #-}
523 The RHS of 'any' will get optimised and deforested; but the InlineRule
524 will still mention the original RHS.
527 preInlineUnconditionally
528 ~~~~~~~~~~~~~~~~~~~~~~~~
529 @preInlineUnconditionally@ examines a bndr to see if it is used just
530 once in a completely safe way, so that it is safe to discard the
531 binding inline its RHS at the (unique) usage site, REGARDLESS of how
532 big the RHS might be. If this is the case we don't simplify the RHS
533 first, but just inline it un-simplified.
535 This is much better than first simplifying a perhaps-huge RHS and then
536 inlining and re-simplifying it. Indeed, it can be at least quadratically
545 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
546 This can happen with cascades of functions too:
553 THE MAIN INVARIANT is this:
555 ---- preInlineUnconditionally invariant -----
556 IF preInlineUnconditionally chooses to inline x = <rhs>
557 THEN doing the inlining should not change the occurrence
558 info for the free vars of <rhs>
559 ----------------------------------------------
561 For example, it's tempting to look at trivial binding like
563 and inline it unconditionally. But suppose x is used many times,
564 but this is the unique occurrence of y. Then inlining x would change
565 y's occurrence info, which breaks the invariant. It matters: y
566 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
569 Even RHSs labelled InlineMe aren't caught here, because there might be
570 no benefit from inlining at the call site.
572 [Sept 01] Don't unconditionally inline a top-level thing, because that
573 can simply make a static thing into something built dynamically. E.g.
577 [Remember that we treat \s as a one-shot lambda.] No point in
578 inlining x unless there is something interesting about the call site.
580 But watch out: if you aren't careful, some useful foldr/build fusion
581 can be lost (most notably in spectral/hartel/parstof) because the
582 foldr didn't see the build. Doing the dynamic allocation isn't a big
583 deal, in fact, but losing the fusion can be. But the right thing here
584 seems to be to do a callSiteInline based on the fact that there is
585 something interesting about the call site (it's strict). Hmm. That
588 Conclusion: inline top level things gaily until Phase 0 (the last
589 phase), at which point don't.
591 Note [pre/postInlineUnconditionally in gentle mode]
592 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
593 Even in gentle mode we want to do preInlineUnconditionally. The
594 reason is that too little clean-up happens if you don't inline
595 use-once things. Also a bit of inlining is *good* for full laziness;
596 it can expose constant sub-expressions. Example in
597 spectral/mandel/Mandel.hs, where the mandelset function gets a useful
598 let-float if you inline windowToViewport
600 However, as usual for Gentle mode, do not inline things that are
601 inactive in the intial stages. See Note [Gentle mode].
604 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
605 preInlineUnconditionally env top_lvl bndr rhs
607 | opt_SimplNoPreInlining = False
608 | otherwise = case idOccInfo bndr of
609 IAmDead -> True -- Happens in ((\x.1) v)
610 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
614 active = case phase of
615 SimplGently {} -> isEarlyActive act
616 -- See Note [pre/postInlineUnconditionally in gentle mode]
617 SimplPhase n _ -> isActive n act
618 act = idInlineActivation bndr
620 try_once in_lam int_cxt -- There's one textual occurrence
621 | not in_lam = isNotTopLevel top_lvl || early_phase
622 | otherwise = int_cxt && canInlineInLam rhs
624 -- Be very careful before inlining inside a lambda, becuase (a) we must not
625 -- invalidate occurrence information, and (b) we want to avoid pushing a
626 -- single allocation (here) into multiple allocations (inside lambda).
627 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
628 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
630 -- is_cheap = exprIsCheap rhs
631 -- ok = is_cheap && int_cxt
633 -- int_cxt The context isn't totally boring
634 -- E.g. let f = \ab.BIG in \y. map f xs
635 -- Don't want to substitute for f, because then we allocate
636 -- its closure every time the \y is called
637 -- But: let f = \ab.BIG in \y. map (f y) xs
638 -- Now we do want to substitute for f, even though it's not
639 -- saturated, because we're going to allocate a closure for
640 -- (f y) every time round the loop anyhow.
642 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
643 -- so substituting rhs inside a lambda doesn't change the occ info.
644 -- Sadly, not quite the same as exprIsHNF.
645 canInlineInLam (Lit _) = True
646 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
647 canInlineInLam (Note _ e) = canInlineInLam e
648 canInlineInLam _ = False
650 early_phase = case phase of
651 SimplPhase 0 _ -> False
653 -- If we don't have this early_phase test, consider
654 -- x = length [1,2,3]
655 -- The full laziness pass carefully floats all the cons cells to
656 -- top level, and preInlineUnconditionally floats them all back in.
657 -- Result is (a) static allocation replaced by dynamic allocation
658 -- (b) many simplifier iterations because this tickles
659 -- a related problem; only one inlining per pass
661 -- On the other hand, I have seen cases where top-level fusion is
662 -- lost if we don't inline top level thing (e.g. string constants)
663 -- Hence the test for phase zero (which is the phase for all the final
664 -- simplifications). Until phase zero we take no special notice of
665 -- top level things, but then we become more leery about inlining
670 postInlineUnconditionally
671 ~~~~~~~~~~~~~~~~~~~~~~~~~
672 @postInlineUnconditionally@ decides whether to unconditionally inline
673 a thing based on the form of its RHS; in particular if it has a
674 trivial RHS. If so, we can inline and discard the binding altogether.
676 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
677 only have *forward* references Hence, it's safe to discard the binding
679 NOTE: This isn't our last opportunity to inline. We're at the binding
680 site right now, and we'll get another opportunity when we get to the
683 Note that we do this unconditional inlining only for trival RHSs.
684 Don't inline even WHNFs inside lambdas; doing so may simply increase
685 allocation when the function is called. This isn't the last chance; see
688 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
689 Because we don't even want to inline them into the RHS of constructor
690 arguments. See NOTE above
692 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
693 it's best to inline it anyway. We often get a=E; b=a from desugaring,
694 with both a and b marked NOINLINE. But that seems incompatible with
695 our new view that inlining is like a RULE, so I'm sticking to the 'active'
699 postInlineUnconditionally
700 :: SimplEnv -> TopLevelFlag
701 -> OutId -- The binder (an InId would be fine too)
702 -> OccInfo -- From the InId
706 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
708 | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, don't inline
709 -- because it might be referred to "earlier"
710 | isExportedId bndr = False
711 | isInlineRule unfolding = False -- Note [InlineRule and postInlineUnconditionally]
712 | exprIsTrivial rhs = True
715 -- The point of examining occ_info here is that for *non-values*
716 -- that occur outside a lambda, the call-site inliner won't have
717 -- a chance (becuase it doesn't know that the thing
718 -- only occurs once). The pre-inliner won't have gotten
719 -- it either, if the thing occurs in more than one branch
720 -- So the main target is things like
723 -- True -> case x of ...
724 -- False -> case x of ...
725 -- I'm not sure how important this is in practice
726 OneOcc in_lam _one_br int_cxt -- OneOcc => no code-duplication issue
727 -> smallEnoughToInline unfolding -- Small enough to dup
728 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
730 -- NB: Do NOT inline arbitrarily big things, even if one_br is True
731 -- Reason: doing so risks exponential behaviour. We simplify a big
732 -- expression, inline it, and simplify it again. But if the
733 -- very same thing happens in the big expression, we get
735 -- PRINCIPLE: when we've already simplified an expression once,
736 -- make sure that we only inline it if it's reasonably small.
738 && ((isNotTopLevel top_lvl && not in_lam) ||
739 -- But outside a lambda, we want to be reasonably aggressive
740 -- about inlining into multiple branches of case
741 -- e.g. let x = <non-value>
742 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
743 -- Inlining can be a big win if C3 is the hot-spot, even if
744 -- the uses in C1, C2 are not 'interesting'
745 -- An example that gets worse if you add int_cxt here is 'clausify'
747 (isCheapUnfolding unfolding && int_cxt))
748 -- isCheap => acceptable work duplication; in_lam may be true
749 -- int_cxt to prevent us inlining inside a lambda without some
750 -- good reason. See the notes on int_cxt in preInlineUnconditionally
752 IAmDead -> True -- This happens; for example, the case_bndr during case of
753 -- known constructor: case (a,b) of x { (p,q) -> ... }
754 -- Here x isn't mentioned in the RHS, so we don't want to
755 -- create the (dead) let-binding let x = (a,b) in ...
759 -- Here's an example that we don't handle well:
760 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
761 -- in \y. ....case f of {...} ....
762 -- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
764 -- - We can't preInlineUnconditionally because that woud invalidate
765 -- the occ info for b.
766 -- - We can't postInlineUnconditionally because the RHS is big, and
767 -- that risks exponential behaviour
768 -- - We can't call-site inline, because the rhs is big
772 active = case getMode env of
773 SimplGently {} -> isEarlyActive act
774 -- See Note [pre/postInlineUnconditionally in gentle mode]
775 SimplPhase n _ -> isActive n act
776 act = idInlineActivation bndr
778 activeInline :: SimplEnv -> OutId -> Bool
780 = case getMode env of
781 SimplGently { sm_inline = inlining_on }
782 -> inlining_on && isEarlyActive act
783 -- See Note [Gentle mode]
785 -- NB: we used to have a second exception, for data con wrappers.
786 -- On the grounds that we use gentle mode for rule LHSs, and
787 -- they match better when data con wrappers are inlined.
788 -- But that only really applies to the trivial wrappers (like (:)),
789 -- and they are now constructed as Compulsory unfoldings (in MkId)
790 -- so they'll happen anyway.
792 SimplPhase n _ -> isActive n act
794 act = idInlineActivation id
796 activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
797 -- Nothing => No rules at all
798 activeRule dflags env
799 | not (dopt Opt_EnableRewriteRules dflags)
800 = Nothing -- Rewriting is off
802 = case getMode env of
803 SimplGently { sm_rules = rules_on }
804 | rules_on -> Just isEarlyActive
805 | otherwise -> Nothing
806 -- Used to be Nothing (no rules in gentle mode)
807 -- Main motivation for changing is that I wanted
808 -- lift String ===> ...
809 -- to work in Template Haskell when simplifying
810 -- splices, so we get simpler code for literal strings
811 SimplPhase n _ -> Just (isActive n)
814 Note [InlineRule and postInlineUnconditionally]
815 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
816 Do not do postInlineUnconditionally if the Id has an InlineRule, otherwise
817 we lose the unfolding. Example
819 -- f has InlineRule with rhs (e |> co)
823 Then there's a danger we'll optimise to
828 and now postInlineUnconditionally, losing the InlineRule on f. Now f'
829 won't inline because 'e' is too big.
832 %************************************************************************
836 %************************************************************************
839 mkLam :: SimplEnv -> [OutBndr] -> OutExpr -> SimplM OutExpr
840 -- mkLam tries three things
841 -- a) eta reduction, if that gives a trivial expression
842 -- b) eta expansion [only if there are some value lambdas]
847 = do { dflags <- getDOptsSmpl
848 ; mkLam' dflags bndrs body }
850 mkLam' :: DynFlags -> [OutBndr] -> OutExpr -> SimplM OutExpr
851 mkLam' dflags bndrs (Cast body co)
852 | not (any bad bndrs)
853 -- Note [Casts and lambdas]
854 = do { lam <- mkLam' dflags bndrs body
855 ; return (mkCoerce (mkPiTypes bndrs co) lam) }
857 co_vars = tyVarsOfType co
858 bad bndr = isCoVar bndr && bndr `elemVarSet` co_vars
860 mkLam' dflags bndrs body
861 | dopt Opt_DoEtaReduction dflags,
862 Just etad_lam <- tryEtaReduce bndrs body
863 = do { tick (EtaReduction (head bndrs))
866 | dopt Opt_DoLambdaEtaExpansion dflags,
867 not (inGentleMode env), -- In gentle mode don't eta-expansion
868 any isRuntimeVar bndrs -- because it can clutter up the code
869 -- with casts etc that may not be removed
870 = do { let body' = tryEtaExpansion dflags body
871 ; return (mkLams bndrs body') }
874 = return (mkLams bndrs body)
877 Note [Casts and lambdas]
878 ~~~~~~~~~~~~~~~~~~~~~~~~
880 (\x. (\y. e) `cast` g1) `cast` g2
881 There is a danger here that the two lambdas look separated, and the
882 full laziness pass might float an expression to between the two.
884 So this equation in mkLam' floats the g1 out, thus:
885 (\x. e `cast` g1) --> (\x.e) `cast` (tx -> g1)
888 In general, this floats casts outside lambdas, where (I hope) they
889 might meet and cancel with some other cast:
890 \x. e `cast` co ===> (\x. e) `cast` (tx -> co)
891 /\a. e `cast` co ===> (/\a. e) `cast` (/\a. co)
892 /\g. e `cast` co ===> (/\g. e) `cast` (/\g. co)
895 Notice that it works regardless of 'e'. Originally it worked only
896 if 'e' was itself a lambda, but in some cases that resulted in
897 fruitless iteration in the simplifier. A good example was when
898 compiling Text.ParserCombinators.ReadPrec, where we had a definition
899 like (\x. Get `cast` g)
900 where Get is a constructor with nonzero arity. Then mkLam eta-expanded
901 the Get, and the next iteration eta-reduced it, and then eta-expanded
904 Note also the side condition for the case of coercion binders.
905 It does not make sense to transform
906 /\g. e `cast` g ==> (/\g.e) `cast` (/\g.g)
907 because the latter is not well-kinded.
909 -- c) floating lets out through big lambdas
910 -- [only if all tyvar lambdas, and only if this lambda
911 -- is the RHS of a let]
913 {- Sept 01: I'm experimenting with getting the
914 full laziness pass to float out past big lambdsa
915 | all isTyVar bndrs, -- Only for big lambdas
916 contIsRhs cont -- Only try the rhs type-lambda floating
917 -- if this is indeed a right-hand side; otherwise
918 -- we end up floating the thing out, only for float-in
919 -- to float it right back in again!
920 = do (floats, body') <- tryRhsTyLam env bndrs body
921 return (floats, mkLams bndrs body')
925 %************************************************************************
929 %************************************************************************
931 Note [Eta reduction conditions]
932 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
933 We try for eta reduction here, but *only* if we get all the way to an
934 trivial expression. We don't want to remove extra lambdas unless we
935 are going to avoid allocating this thing altogether.
937 There are some particularly delicate points here:
939 * Eta reduction is not valid in general:
941 This matters, partly for old-fashioned correctness reasons but,
942 worse, getting it wrong can yield a seg fault. Consider
944 h y = case (case y of { True -> f `seq` True; False -> False }) of
945 True -> ...; False -> ...
947 If we (unsoundly) eta-reduce f to get f=f, the strictness analyser
948 says f=bottom, and replaces the (f `seq` True) with just
949 (f `cast` unsafe-co). BUT, as thing stand, 'f' got arity 1, and it
950 *keeps* arity 1 (perhaps also wrongly). So CorePrep eta-expands
951 the definition again, so that it does not termninate after all.
952 Result: seg-fault because the boolean case actually gets a function value.
955 So it's important to to the right thing.
957 * Note [Arity care]: we need to be careful if we just look at f's
958 arity. Currently (Dec07), f's arity is visible in its own RHS (see
959 Note [Arity robustness] in SimplEnv) so we must *not* trust the
960 arity when checking that 'f' is a value. Otherwise we will
965 Which might change a terminiating program (think (f `seq` e)) to a
966 non-terminating one. So we check for being a loop breaker first.
968 However for GlobalIds we can look at the arity; and for primops we
969 must, since they have no unfolding.
971 * Regardless of whether 'f' is a value, we always want to
972 reduce (/\a -> f a) to f
973 This came up in a RULE: foldr (build (/\a -> g a))
974 did not match foldr (build (/\b -> ...something complex...))
975 The type checker can insert these eta-expanded versions,
976 with both type and dictionary lambdas; hence the slightly
979 * Never *reduce* arity. For example
981 Then if h has arity 1 we don't want to eta-reduce because then
982 f's arity would decrease, and that is bad
984 These delicacies are why we don't use exprIsTrivial and exprIsHNF here.
988 tryEtaReduce :: [OutBndr] -> OutExpr -> Maybe OutExpr
989 tryEtaReduce bndrs body
990 = go (reverse bndrs) body
992 incoming_arity = count isId bndrs
994 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
995 go [] fun | ok_fun fun = Just fun -- Success!
996 go _ _ = Nothing -- Failure!
998 -- Note [Eta reduction conditions]
999 ok_fun (App fun (Type ty))
1000 | not (any (`elemVarSet` tyVarsOfType ty) bndrs)
1003 = not (fun_id `elem` bndrs)
1004 && (ok_fun_id fun_id || all ok_lam bndrs)
1007 ok_fun_id fun = fun_arity fun >= incoming_arity
1009 fun_arity fun -- See Note [Arity care]
1010 | isLocalId fun && isLoopBreaker (idOccInfo fun) = 0
1011 | otherwise = idArity fun
1013 ok_lam v = isTyVar v || isDictId v
1015 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
1019 %************************************************************************
1023 %************************************************************************
1027 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
1030 where (in both cases)
1032 * The xi can include type variables
1034 * The yi are all value variables
1036 * N is a NORMAL FORM (i.e. no redexes anywhere)
1037 wanting a suitable number of extra args.
1039 The biggest reason for doing this is for cases like
1045 Here we want to get the lambdas together. A good exmaple is the nofib
1046 program fibheaps, which gets 25% more allocation if you don't do this
1049 We may have to sandwich some coerces between the lambdas
1050 to make the types work. exprEtaExpandArity looks through coerces
1051 when computing arity; and etaExpand adds the coerces as necessary when
1052 actually computing the expansion.
1055 tryEtaExpansion :: DynFlags -> OutExpr -> OutExpr
1056 -- There is at least one runtime binder in the binders
1057 tryEtaExpansion dflags body
1058 = etaExpand fun_arity body
1060 fun_arity = exprEtaExpandArity dflags body
1064 %************************************************************************
1066 \subsection{Floating lets out of big lambdas}
1068 %************************************************************************
1070 Note [Floating and type abstraction]
1071 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1074 We'd like to float this to
1077 x = /\a. C (y1 a) (y2 a)
1078 for the usual reasons: we want to inline x rather vigorously.
1080 You may think that this kind of thing is rare. But in some programs it is
1081 common. For example, if you do closure conversion you might get:
1083 data a :-> b = forall e. (e -> a -> b) :$ e
1085 f_cc :: forall a. a :-> a
1086 f_cc = /\a. (\e. id a) :$ ()
1088 Now we really want to inline that f_cc thing so that the
1089 construction of the closure goes away.
1091 So I have elaborated simplLazyBind to understand right-hand sides that look
1095 and treat them specially. The real work is done in SimplUtils.abstractFloats,
1096 but there is quite a bit of plumbing in simplLazyBind as well.
1098 The same transformation is good when there are lets in the body:
1100 /\abc -> let(rec) x = e in b
1102 let(rec) x' = /\abc -> let x = x' a b c in e
1104 /\abc -> let x = x' a b c in b
1106 This is good because it can turn things like:
1108 let f = /\a -> letrec g = ... g ... in g
1110 letrec g' = /\a -> ... g' a ...
1112 let f = /\ a -> g' a
1114 which is better. In effect, it means that big lambdas don't impede
1117 This optimisation is CRUCIAL in eliminating the junk introduced by
1118 desugaring mutually recursive definitions. Don't eliminate it lightly!
1120 [May 1999] If we do this transformation *regardless* then we can
1121 end up with some pretty silly stuff. For example,
1124 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
1129 st = /\s -> ...[y1 s/x1, y2 s/x2]
1132 Unless the "..." is a WHNF there is really no point in doing this.
1133 Indeed it can make things worse. Suppose x1 is used strictly,
1136 x1* = case f y of { (a,b) -> e }
1138 If we abstract this wrt the tyvar we then can't do the case inline
1139 as we would normally do.
1141 That's why the whole transformation is part of the same process that
1142 floats let-bindings and constructor arguments out of RHSs. In particular,
1143 it is guarded by the doFloatFromRhs call in simplLazyBind.
1147 abstractFloats :: [OutTyVar] -> SimplEnv -> OutExpr -> SimplM ([OutBind], OutExpr)
1148 abstractFloats main_tvs body_env body
1149 = ASSERT( notNull body_floats )
1150 do { (subst, float_binds) <- mapAccumLM abstract empty_subst body_floats
1151 ; return (float_binds, CoreSubst.substExpr subst body) }
1153 main_tv_set = mkVarSet main_tvs
1154 body_floats = getFloats body_env
1155 empty_subst = CoreSubst.mkEmptySubst (seInScope body_env)
1157 abstract :: CoreSubst.Subst -> OutBind -> SimplM (CoreSubst.Subst, OutBind)
1158 abstract subst (NonRec id rhs)
1159 = do { (poly_id, poly_app) <- mk_poly tvs_here id
1160 ; let poly_rhs = mkLams tvs_here rhs'
1161 subst' = CoreSubst.extendIdSubst subst id poly_app
1162 ; return (subst', (NonRec poly_id poly_rhs)) }
1164 rhs' = CoreSubst.substExpr subst rhs
1165 tvs_here | any isCoVar main_tvs = main_tvs -- Note [Abstract over coercions]
1167 = varSetElems (main_tv_set `intersectVarSet` exprSomeFreeVars isTyVar rhs')
1169 -- Abstract only over the type variables free in the rhs
1170 -- wrt which the new binding is abstracted. But the naive
1171 -- approach of abstract wrt the tyvars free in the Id's type
1173 -- /\ a b -> let t :: (a,b) = (e1, e2)
1176 -- Here, b isn't free in x's type, but we must nevertheless
1177 -- abstract wrt b as well, because t's type mentions b.
1178 -- Since t is floated too, we'd end up with the bogus:
1179 -- poly_t = /\ a b -> (e1, e2)
1180 -- poly_x = /\ a -> fst (poly_t a *b*)
1181 -- So for now we adopt the even more naive approach of
1182 -- abstracting wrt *all* the tyvars. We'll see if that
1183 -- gives rise to problems. SLPJ June 98
1185 abstract subst (Rec prs)
1186 = do { (poly_ids, poly_apps) <- mapAndUnzipM (mk_poly tvs_here) ids
1187 ; let subst' = CoreSubst.extendSubstList subst (ids `zip` poly_apps)
1188 poly_rhss = [mkLams tvs_here (CoreSubst.substExpr subst' rhs) | rhs <- rhss]
1189 ; return (subst', Rec (poly_ids `zip` poly_rhss)) }
1191 (ids,rhss) = unzip prs
1192 -- For a recursive group, it's a bit of a pain to work out the minimal
1193 -- set of tyvars over which to abstract:
1194 -- /\ a b c. let x = ...a... in
1195 -- letrec { p = ...x...q...
1196 -- q = .....p...b... } in
1198 -- Since 'x' is abstracted over 'a', the {p,q} group must be abstracted
1199 -- over 'a' (because x is replaced by (poly_x a)) as well as 'b'.
1200 -- Since it's a pain, we just use the whole set, which is always safe
1202 -- If you ever want to be more selective, remember this bizarre case too:
1204 -- Here, we must abstract 'x' over 'a'.
1207 mk_poly tvs_here var
1208 = do { uniq <- getUniqueM
1209 ; let poly_name = setNameUnique (idName var) uniq -- Keep same name
1210 poly_ty = mkForAllTys tvs_here (idType var) -- But new type of course
1211 poly_id = transferPolyIdInfo var tvs_here $ -- Note [transferPolyIdInfo] in Id.lhs
1212 mkLocalId poly_name poly_ty
1213 ; return (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tvs_here)) }
1214 -- In the olden days, it was crucial to copy the occInfo of the original var,
1215 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1216 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1217 -- at already simplified code, so it doesn't matter
1219 -- It's even right to retain single-occurrence or dead-var info:
1220 -- Suppose we started with /\a -> let x = E in B
1221 -- where x occurs once in B. Then we transform to:
1222 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1223 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1224 -- the occurrences of x' will be just the occurrences originally
1228 Note [Abstract over coercions]
1229 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1230 If a coercion variable (g :: a ~ Int) is free in the RHS, then so is the
1231 type variable a. Rather than sort this mess out, we simply bale out and abstract
1232 wrt all the type variables if any of them are coercion variables.
1235 Historical note: if you use let-bindings instead of a substitution, beware of this:
1237 -- Suppose we start with:
1239 -- x = /\ a -> let g = G in E
1241 -- Then we'll float to get
1243 -- x = let poly_g = /\ a -> G
1244 -- in /\ a -> let g = poly_g a in E
1246 -- But now the occurrence analyser will see just one occurrence
1247 -- of poly_g, not inside a lambda, so the simplifier will
1248 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1249 -- (I used to think that the "don't inline lone occurrences" stuff
1250 -- would stop this happening, but since it's the *only* occurrence,
1251 -- PreInlineUnconditionally kicks in first!)
1253 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1254 -- to appear many times. (NB: mkInlineMe eliminates
1255 -- such notes on trivial RHSs, so do it manually.)
1257 %************************************************************************
1261 %************************************************************************
1263 prepareAlts tries these things:
1265 1. If several alternatives are identical, merge them into
1266 a single DEFAULT alternative. I've occasionally seen this
1267 making a big difference:
1269 case e of =====> case e of
1270 C _ -> f x D v -> ....v....
1271 D v -> ....v.... DEFAULT -> f x
1274 The point is that we merge common RHSs, at least for the DEFAULT case.
1275 [One could do something more elaborate but I've never seen it needed.]
1276 To avoid an expensive test, we just merge branches equal to the *first*
1277 alternative; this picks up the common cases
1278 a) all branches equal
1279 b) some branches equal to the DEFAULT (which occurs first)
1282 case e of b { ==> case e of b {
1283 p1 -> rhs1 p1 -> rhs1
1285 pm -> rhsm pm -> rhsm
1286 _ -> case b of b' { pn -> let b'=b in rhsn
1288 ... po -> let b'=b in rhso
1289 po -> rhso _ -> let b'=b in rhsd
1293 which merges two cases in one case when -- the default alternative of
1294 the outer case scrutises the same variable as the outer case This
1295 transformation is called Case Merging. It avoids that the same
1296 variable is scrutinised multiple times.
1299 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
1305 where @is@ was something like
1307 p `is` n = p /= (-1) && p == n
1309 This gave rise to a horrible sequence of cases
1316 and similarly in cascade for all the join points!
1319 ~~~~~~~~~~~~~~~~~~~~
1320 We do this *here*, looking at un-simplified alternatives, because we
1321 have to check that r doesn't mention the variables bound by the
1322 pattern in each alternative, so the binder-info is rather useful.
1325 prepareAlts :: SimplEnv -> OutExpr -> OutId -> [InAlt] -> SimplM ([AltCon], [InAlt])
1326 prepareAlts env scrut case_bndr' alts
1327 = do { dflags <- getDOptsSmpl
1328 ; alts <- combineIdenticalAlts case_bndr' alts
1330 ; let (alts_wo_default, maybe_deflt) = findDefault alts
1331 alt_cons = [con | (con,_,_) <- alts_wo_default]
1332 imposs_deflt_cons = nub (imposs_cons ++ alt_cons)
1333 -- "imposs_deflt_cons" are handled
1334 -- EITHER by the context,
1335 -- OR by a non-DEFAULT branch in this case expression.
1337 ; default_alts <- prepareDefault dflags env case_bndr' mb_tc_app
1338 imposs_deflt_cons maybe_deflt
1340 ; let trimmed_alts = filterOut impossible_alt alts_wo_default
1341 merged_alts = mergeAlts trimmed_alts default_alts
1342 -- We need the mergeAlts in case the new default_alt
1343 -- has turned into a constructor alternative.
1344 -- The merge keeps the inner DEFAULT at the front, if there is one
1345 -- and interleaves the alternatives in the right order
1347 ; return (imposs_deflt_cons, merged_alts) }
1349 mb_tc_app = splitTyConApp_maybe (idType case_bndr')
1350 Just (_, inst_tys) = mb_tc_app
1352 imposs_cons = case scrut of
1353 Var v -> otherCons (idUnfolding v)
1356 impossible_alt :: CoreAlt -> Bool
1357 impossible_alt (con, _, _) | con `elem` imposs_cons = True
1358 impossible_alt (DataAlt con, _, _) = dataConCannotMatch inst_tys con
1359 impossible_alt _ = False
1362 --------------------------------------------------
1363 -- 1. Merge identical branches
1364 --------------------------------------------------
1365 combineIdenticalAlts :: OutId -> [InAlt] -> SimplM [InAlt]
1367 combineIdenticalAlts case_bndr ((_con1,bndrs1,rhs1) : con_alts)
1368 | all isDeadBinder bndrs1, -- Remember the default
1369 length filtered_alts < length con_alts -- alternative comes first
1370 -- Also Note [Dead binders]
1371 = do { tick (AltMerge case_bndr)
1372 ; return ((DEFAULT, [], rhs1) : filtered_alts) }
1374 filtered_alts = filter keep con_alts
1375 keep (_con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1377 combineIdenticalAlts _ alts = return alts
1379 -------------------------------------------------------------------------
1380 -- Prepare the default alternative
1381 -------------------------------------------------------------------------
1382 prepareDefault :: DynFlags
1384 -> OutId -- Case binder; need just for its type. Note that as an
1385 -- OutId, it has maximum information; this is important.
1386 -- Test simpl013 is an example
1387 -> Maybe (TyCon, [Type]) -- Type of scrutinee, decomposed
1388 -> [AltCon] -- These cons can't happen when matching the default
1389 -> Maybe InExpr -- Rhs
1390 -> SimplM [InAlt] -- Still unsimplified
1391 -- We use a list because it's what mergeAlts expects,
1392 -- And becuase case-merging can cause many to show up
1394 ------- Merge nested cases ----------
1395 prepareDefault dflags env outer_bndr _bndr_ty imposs_cons (Just deflt_rhs)
1396 | dopt Opt_CaseMerge dflags
1397 , Case (Var inner_scrut_var) inner_bndr _ inner_alts <- deflt_rhs
1398 , DoneId inner_scrut_var' <- substId env inner_scrut_var
1399 -- Remember, inner_scrut_var is an InId, but outer_bndr is an OutId
1400 , inner_scrut_var' == outer_bndr
1401 -- NB: the substId means that if the outer scrutinee was a
1402 -- variable, and inner scrutinee is the same variable,
1403 -- then inner_scrut_var' will be outer_bndr
1404 -- via the magic of simplCaseBinder
1405 = do { tick (CaseMerge outer_bndr)
1407 ; let munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1408 ; return [(con, args, munge_rhs rhs) | (con, args, rhs) <- inner_alts,
1409 not (con `elem` imposs_cons) ]
1410 -- NB: filter out any imposs_cons. Example:
1413 -- DEFAULT -> case x of
1416 -- When we merge, we must ensure that e1 takes
1417 -- precedence over e2 as the value for A!
1419 -- Warning: don't call prepareAlts recursively!
1420 -- Firstly, there's no point, because inner alts have already had
1421 -- mkCase applied to them, so they won't have a case in their default
1422 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1423 -- in munge_rhs may put a case into the DEFAULT branch!
1426 --------- Fill in known constructor -----------
1427 prepareDefault _ _ case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
1428 | -- This branch handles the case where we are
1429 -- scrutinisng an algebraic data type
1430 isAlgTyCon tycon -- It's a data type, tuple, or unboxed tuples.
1431 , not (isNewTyCon tycon) -- We can have a newtype, if we are just doing an eval:
1432 -- case x of { DEFAULT -> e }
1433 -- and we don't want to fill in a default for them!
1434 , Just all_cons <- tyConDataCons_maybe tycon
1435 , not (null all_cons) -- This is a tricky corner case. If the data type has no constructors,
1436 -- which GHC allows, then the case expression will have at most a default
1437 -- alternative. We don't want to eliminate that alternative, because the
1438 -- invariant is that there's always one alternative. It's more convenient
1440 -- case x of { DEFAULT -> e }
1441 -- as it is, rather than transform it to
1442 -- error "case cant match"
1443 -- which would be quite legitmate. But it's a really obscure corner, and
1444 -- not worth wasting code on.
1445 , let imposs_data_cons = [con | DataAlt con <- imposs_cons] -- We now know it's a data type
1446 impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
1447 = case filterOut impossible all_cons of
1448 [] -> return [] -- Eliminate the default alternative
1449 -- altogether if it can't match
1451 [con] -> -- It matches exactly one constructor, so fill it in
1452 do { tick (FillInCaseDefault case_bndr)
1454 ; let (ex_tvs, co_tvs, arg_ids) =
1455 dataConRepInstPat us con inst_tys
1456 ; return [(DataAlt con, ex_tvs ++ co_tvs ++ arg_ids, deflt_rhs)] }
1458 _ -> return [(DEFAULT, [], deflt_rhs)]
1460 | debugIsOn, isAlgTyCon tycon, not (isOpenTyCon tycon), null (tyConDataCons tycon)
1461 -- This can legitimately happen for type families, so don't report that
1462 = pprTrace "prepareDefault" (ppr case_bndr <+> ppr tycon)
1463 $ return [(DEFAULT, [], deflt_rhs)]
1465 --------- Catch-all cases -----------
1466 prepareDefault _dflags _env _case_bndr _bndr_ty _imposs_cons (Just deflt_rhs)
1467 = return [(DEFAULT, [], deflt_rhs)]
1469 prepareDefault _dflags _env _case_bndr _bndr_ty _imposs_cons Nothing
1470 = return [] -- No default branch
1475 =================================================================================
1477 mkCase tries these things
1479 1. Eliminate the case altogether if possible
1487 and similar friends.
1491 mkCase :: OutExpr -> OutId -> [OutAlt] -- Increasing order
1494 --------------------------------------------------
1496 --------------------------------------------------
1498 mkCase scrut case_bndr alts -- Identity case
1499 | all identity_alt alts
1500 = do tick (CaseIdentity case_bndr)
1501 return (re_cast scrut)
1503 identity_alt (con, args, rhs) = check_eq con args (de_cast rhs)
1505 check_eq DEFAULT _ (Var v) = v == case_bndr
1506 check_eq (LitAlt lit') _ (Lit lit) = lit == lit'
1507 check_eq (DataAlt con) args rhs = rhs `cheapEqExpr` mkConApp con (arg_tys ++ varsToCoreExprs args)
1508 || rhs `cheapEqExpr` Var case_bndr
1509 check_eq _ _ _ = False
1511 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1514 -- case e of x { _ -> x `cast` c }
1515 -- And we definitely want to eliminate this case, to give
1517 -- So we throw away the cast from the RHS, and reconstruct
1518 -- it at the other end. All the RHS casts must be the same
1519 -- if (all identity_alt alts) holds.
1521 -- Don't worry about nested casts, because the simplifier combines them
1522 de_cast (Cast e _) = e
1525 re_cast scrut = case head alts of
1526 (_,_,Cast _ co) -> Cast scrut co
1531 --------------------------------------------------
1533 --------------------------------------------------
1534 mkCase scrut bndr alts = return (Case scrut bndr (coreAltsType alts) alts)
1538 When adding auxiliary bindings for the case binder, it's worth checking if
1539 its dead, because it often is, and occasionally these mkCase transformations
1540 cascade rather nicely.
1543 bindCaseBndr :: Id -> CoreExpr -> CoreExpr -> CoreExpr
1544 bindCaseBndr bndr rhs body
1545 | isDeadBinder bndr = body
1546 | otherwise = bindNonRec bndr rhs body