2 % (c) The AQUA Project, Glasgow University, 1993-1998
4 \section[SimplUtils]{The simplifier utilities}
9 mkLam, mkCase, prepareAlts, tryEtaExpand,
12 preInlineUnconditionally, postInlineUnconditionally,
13 activeUnfolding, activeUnfInRule, activeRule,
14 simplEnvForGHCi, simplEnvForRules, updModeForInlineRules,
16 -- The continuation type
17 SimplCont(..), DupFlag(..), ArgInfo(..),
19 contIsDupable, contResultType, contIsTrivial, contArgs, dropArgs,
20 pushSimplifiedArgs, countValArgs, countArgs, addArgTo,
21 mkBoringStop, mkRhsStop, mkLazyArgStop, contIsRhsOrArg,
22 interestingCallContext,
24 interestingArg, mkArgInfo,
29 #include "HsVersions.h"
32 import CoreMonad ( SimplifierMode(..), Tick(..) )
36 import qualified CoreSubst
47 import Type hiding( substTy )
48 import Coercion ( coercionKind )
50 import Unify ( dataConCannotMatch )
62 %************************************************************************
66 %************************************************************************
68 A SimplCont allows the simplifier to traverse the expression in a
69 zipper-like fashion. The SimplCont represents the rest of the expression,
70 "above" the point of interest.
72 You can also think of a SimplCont as an "evaluation context", using
73 that term in the way it is used for operational semantics. This is the
74 way I usually think of it, For example you'll often see a syntax for
75 evaluation context looking like
76 C ::= [] | C e | case C of alts | C `cast` co
77 That's the kind of thing we are doing here, and I use that syntax in
82 * A SimplCont describes a *strict* context (just like
83 evaluation contexts do). E.g. Just [] is not a SimplCont
85 * A SimplCont describes a context that *does not* bind
86 any variables. E.g. \x. [] is not a SimplCont
90 = Stop -- An empty context, or hole, []
91 CallCtxt -- True <=> There is something interesting about
92 -- the context, and hence the inliner
93 -- should be a bit keener (see interestingCallContext)
95 -- This is an argument of a function that has RULES
96 -- Inlining the call might allow the rule to fire
98 | CoerceIt -- C `cast` co
99 OutCoercion -- The coercion simplified
103 DupFlag -- See Note [DupFlag invariants]
104 InExpr StaticEnv -- The argument and its static env
107 | Select -- case C of alts
108 DupFlag -- See Note [DupFlag invariants]
109 InId [InAlt] StaticEnv -- The case binder, alts, and subst-env
112 -- The two strict forms have no DupFlag, because we never duplicate them
113 | StrictBind -- (\x* \xs. e) C
114 InId [InBndr] -- let x* = [] in e
115 InExpr StaticEnv -- is a special case
118 | StrictArg -- f e1 ..en C
119 ArgInfo -- Specifies f, e1..en, Whether f has rules, etc
120 -- plus strictness flags for *further* args
121 CallCtxt -- Whether *this* argument position is interesting
126 ai_fun :: Id, -- The function
127 ai_args :: [OutExpr], -- ...applied to these args (which are in *reverse* order)
128 ai_rules :: [CoreRule], -- Rules for this function
130 ai_encl :: Bool, -- Flag saying whether this function
131 -- or an enclosing one has rules (recursively)
132 -- True => be keener to inline in all args
134 ai_strs :: [Bool], -- Strictness of remaining arguments
135 -- Usually infinite, but if it is finite it guarantees
136 -- that the function diverges after being given
137 -- that number of args
138 ai_discs :: [Int] -- Discounts for remaining arguments; non-zero => be keener to inline
142 addArgTo :: ArgInfo -> OutExpr -> ArgInfo
143 addArgTo ai arg = ai { ai_args = arg : ai_args ai }
145 instance Outputable SimplCont where
146 ppr (Stop interesting) = ptext (sLit "Stop") <> brackets (ppr interesting)
147 ppr (ApplyTo dup arg _ cont) = ((ptext (sLit "ApplyTo") <+> ppr dup <+> pprParendExpr arg)
148 {- $$ nest 2 (pprSimplEnv se) -}) $$ ppr cont
149 ppr (StrictBind b _ _ _ cont) = (ptext (sLit "StrictBind") <+> ppr b) $$ ppr cont
150 ppr (StrictArg ai _ cont) = (ptext (sLit "StrictArg") <+> ppr (ai_fun ai)) $$ ppr cont
151 ppr (Select dup bndr alts se cont) = (ptext (sLit "Select") <+> ppr dup <+> ppr bndr) $$
152 (nest 2 $ vcat [ppr (seTvSubst se), ppr alts]) $$ ppr cont
153 ppr (CoerceIt co cont) = (ptext (sLit "CoerceIt") <+> ppr co) $$ ppr cont
155 data DupFlag = NoDup -- Unsimplified, might be big
156 | Simplified -- Simplified
157 | OkToDup -- Simplified and small
159 isSimplified :: DupFlag -> Bool
160 isSimplified NoDup = False
161 isSimplified _ = True -- Invariant: the subst-env is empty
163 instance Outputable DupFlag where
164 ppr OkToDup = ptext (sLit "ok")
165 ppr NoDup = ptext (sLit "nodup")
166 ppr Simplified = ptext (sLit "simpl")
169 Note [DupFlag invariants]
170 ~~~~~~~~~~~~~~~~~~~~~~~~~
171 In both (ApplyTo dup _ env k)
172 and (Select dup _ _ env k)
173 the following invariants hold
175 (a) if dup = OkToDup, then continuation k is also ok-to-dup
176 (b) if dup = OkToDup or Simplified, the subst-env is empty
177 (and and hence no need to re-simplify)
181 mkBoringStop :: SimplCont
182 mkBoringStop = Stop BoringCtxt
184 mkRhsStop :: SimplCont -- See Note [RHS of lets] in CoreUnfold
185 mkRhsStop = Stop (ArgCtxt False)
187 mkLazyArgStop :: CallCtxt -> SimplCont
188 mkLazyArgStop cci = Stop cci
191 contIsRhsOrArg :: SimplCont -> Bool
192 contIsRhsOrArg (Stop {}) = True
193 contIsRhsOrArg (StrictBind {}) = True
194 contIsRhsOrArg (StrictArg {}) = True
195 contIsRhsOrArg _ = False
198 contIsDupable :: SimplCont -> Bool
199 contIsDupable (Stop {}) = True
200 contIsDupable (ApplyTo OkToDup _ _ _) = True -- See Note [DupFlag invariants]
201 contIsDupable (Select OkToDup _ _ _ _) = True -- ...ditto...
202 contIsDupable (CoerceIt _ cont) = contIsDupable cont
203 contIsDupable _ = False
206 contIsTrivial :: SimplCont -> Bool
207 contIsTrivial (Stop {}) = True
208 contIsTrivial (ApplyTo _ (Type _) _ cont) = contIsTrivial cont
209 contIsTrivial (CoerceIt _ cont) = contIsTrivial cont
210 contIsTrivial _ = False
213 contResultType :: SimplEnv -> OutType -> SimplCont -> OutType
214 contResultType env ty cont
217 subst_ty se ty = substTy (se `setInScope` env) ty
220 go (CoerceIt co cont) _ = go cont (snd (coercionKind co))
221 go (StrictBind _ bs body se cont) _ = go cont (subst_ty se (exprType (mkLams bs body)))
222 go (StrictArg ai _ cont) _ = go cont (funResultTy (argInfoResultTy ai))
223 go (Select _ _ alts se cont) _ = go cont (subst_ty se (coreAltsType alts))
224 go (ApplyTo _ arg se cont) ty = go cont (apply_to_arg ty arg se)
226 apply_to_arg ty (Type ty_arg) se = applyTy ty (subst_ty se ty_arg)
227 apply_to_arg ty _ _ = funResultTy ty
229 argInfoResultTy :: ArgInfo -> OutType
230 argInfoResultTy (ArgInfo { ai_fun = fun, ai_args = args })
231 = foldr (\arg fn_ty -> applyTypeToArg fn_ty arg) (idType fun) args
234 countValArgs :: SimplCont -> Int
235 countValArgs (ApplyTo _ (Type _) _ cont) = countValArgs cont
236 countValArgs (ApplyTo _ _ _ cont) = 1 + countValArgs cont
239 countArgs :: SimplCont -> Int
240 countArgs (ApplyTo _ _ _ cont) = 1 + countArgs cont
243 contArgs :: SimplCont -> (Bool, [ArgSummary], SimplCont)
244 -- Uses substitution to turn each arg into an OutExpr
245 contArgs cont@(ApplyTo {})
246 = case go [] cont of { (args, cont') -> (False, args, cont') }
248 go args (ApplyTo _ arg se cont)
249 | isTypeArg arg = go args cont
250 | otherwise = go (is_interesting arg se : args) cont
251 go args cont = (reverse args, cont)
253 is_interesting arg se = interestingArg (substExpr (text "contArgs") se arg)
254 -- Do *not* use short-cutting substitution here
255 -- because we want to get as much IdInfo as possible
257 contArgs cont = (True, [], cont)
259 pushSimplifiedArgs :: SimplEnv -> [CoreExpr] -> SimplCont -> SimplCont
260 pushSimplifiedArgs _env [] cont = cont
261 pushSimplifiedArgs env (arg:args) cont = ApplyTo Simplified arg env (pushSimplifiedArgs env args cont)
262 -- The env has an empty SubstEnv
264 dropArgs :: Int -> SimplCont -> SimplCont
265 dropArgs 0 cont = cont
266 dropArgs n (ApplyTo _ _ _ cont) = dropArgs (n-1) cont
267 dropArgs n other = pprPanic "dropArgs" (ppr n <+> ppr other)
271 Note [Interesting call context]
272 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
273 We want to avoid inlining an expression where there can't possibly be
274 any gain, such as in an argument position. Hence, if the continuation
275 is interesting (eg. a case scrutinee, application etc.) then we
276 inline, otherwise we don't.
278 Previously some_benefit used to return True only if the variable was
279 applied to some value arguments. This didn't work:
281 let x = _coerce_ (T Int) Int (I# 3) in
282 case _coerce_ Int (T Int) x of
285 we want to inline x, but can't see that it's a constructor in a case
286 scrutinee position, and some_benefit is False.
290 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
292 .... case dMonadST _@_ x0 of (a,b,c) -> ....
294 we'd really like to inline dMonadST here, but we *don't* want to
295 inline if the case expression is just
297 case x of y { DEFAULT -> ... }
299 since we can just eliminate this case instead (x is in WHNF). Similar
300 applies when x is bound to a lambda expression. Hence
301 contIsInteresting looks for case expressions with just a single
306 interestingCallContext :: SimplCont -> CallCtxt
307 -- See Note [Interesting call context]
308 interestingCallContext cont
311 interesting (Select _ bndr _ _ _)
312 | isDeadBinder bndr = CaseCtxt
313 | otherwise = ArgCtxt False -- If the binder is used, this
314 -- is like a strict let
315 -- See Note [RHS of lets] in CoreUnfold
317 interesting (ApplyTo _ arg _ cont)
318 | isTypeArg arg = interesting cont
319 | otherwise = ValAppCtxt -- Can happen if we have (f Int |> co) y
320 -- If f has an INLINE prag we need to give it some
321 -- motivation to inline. See Note [Cast then apply]
324 interesting (StrictArg _ cci _) = cci
325 interesting (StrictBind {}) = BoringCtxt
326 interesting (Stop cci) = cci
327 interesting (CoerceIt _ cont) = interesting cont
328 -- If this call is the arg of a strict function, the context
329 -- is a bit interesting. If we inline here, we may get useful
330 -- evaluation information to avoid repeated evals: e.g.
332 -- Here the contIsInteresting makes the '*' keener to inline,
333 -- which in turn exposes a constructor which makes the '+' inline.
334 -- Assuming that +,* aren't small enough to inline regardless.
336 -- It's also very important to inline in a strict context for things
339 -- Here, the context of (f x) is strict, and if f's unfolding is
340 -- a build it's *great* to inline it here. So we must ensure that
341 -- the context for (f x) is not totally uninteresting.
346 -> [CoreRule] -- Rules for function
347 -> Int -- Number of value args
348 -> SimplCont -- Context of the call
351 mkArgInfo fun rules n_val_args call_cont
352 | n_val_args < idArity fun -- Note [Unsaturated functions]
353 = ArgInfo { ai_fun = fun, ai_args = [], ai_rules = rules
355 , ai_strs = vanilla_stricts
356 , ai_discs = vanilla_discounts }
358 = ArgInfo { ai_fun = fun, ai_args = [], ai_rules = rules
359 , ai_encl = interestingArgContext rules call_cont
360 , ai_strs = add_type_str (idType fun) arg_stricts
361 , ai_discs = arg_discounts }
363 vanilla_discounts, arg_discounts :: [Int]
364 vanilla_discounts = repeat 0
365 arg_discounts = case idUnfolding fun of
366 CoreUnfolding {uf_guidance = UnfIfGoodArgs {ug_args = discounts}}
367 -> discounts ++ vanilla_discounts
368 _ -> vanilla_discounts
370 vanilla_stricts, arg_stricts :: [Bool]
371 vanilla_stricts = repeat False
374 = case splitStrictSig (idStrictness fun) of
375 (demands, result_info)
376 | not (demands `lengthExceeds` n_val_args)
377 -> -- Enough args, use the strictness given.
378 -- For bottoming functions we used to pretend that the arg
379 -- is lazy, so that we don't treat the arg as an
380 -- interesting context. This avoids substituting
381 -- top-level bindings for (say) strings into
382 -- calls to error. But now we are more careful about
383 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
384 if isBotRes result_info then
385 map isStrictDmd demands -- Finite => result is bottom
387 map isStrictDmd demands ++ vanilla_stricts
389 -> WARN( True, text "More demands than arity" <+> ppr fun <+> ppr (idArity fun)
390 <+> ppr n_val_args <+> ppr demands )
391 vanilla_stricts -- Not enough args, or no strictness
393 add_type_str :: Type -> [Bool] -> [Bool]
394 -- If the function arg types are strict, record that in the 'strictness bits'
395 -- No need to instantiate because unboxed types (which dominate the strict
396 -- types) can't instantiate type variables.
397 -- add_type_str is done repeatedly (for each call); might be better
398 -- once-for-all in the function
399 -- But beware primops/datacons with no strictness
400 add_type_str _ [] = []
401 add_type_str fun_ty strs -- Look through foralls
402 | Just (_, fun_ty') <- splitForAllTy_maybe fun_ty -- Includes coercions
403 = add_type_str fun_ty' strs
404 add_type_str fun_ty (str:strs) -- Add strict-type info
405 | Just (arg_ty, fun_ty') <- splitFunTy_maybe fun_ty
406 = (str || isStrictType arg_ty) : add_type_str fun_ty' strs
410 {- Note [Unsaturated functions]
411 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
412 Consider (test eyeball/inline4)
415 where f has arity 2. Then we do not want to inline 'x', because
416 it'll just be floated out again. Even if f has lots of discounts
417 on its first argument -- it must be saturated for these to kick in
420 interestingArgContext :: [CoreRule] -> SimplCont -> Bool
421 -- If the argument has form (f x y), where x,y are boring,
422 -- and f is marked INLINE, then we don't want to inline f.
423 -- But if the context of the argument is
425 -- where g has rules, then we *do* want to inline f, in case it
426 -- exposes a rule that might fire. Similarly, if the context is
428 -- where h has rules, then we do want to inline f; hence the
429 -- call_cont argument to interestingArgContext
431 -- The ai-rules flag makes this happen; if it's
432 -- set, the inliner gets just enough keener to inline f
433 -- regardless of how boring f's arguments are, if it's marked INLINE
435 -- The alternative would be to *always* inline an INLINE function,
436 -- regardless of how boring its context is; but that seems overkill
437 -- For example, it'd mean that wrapper functions were always inlined
438 interestingArgContext rules call_cont
439 = notNull rules || enclosing_fn_has_rules
441 enclosing_fn_has_rules = go call_cont
443 go (Select {}) = False
444 go (ApplyTo {}) = False
445 go (StrictArg _ cci _) = interesting cci
446 go (StrictBind {}) = False -- ??
447 go (CoerceIt _ c) = go c
448 go (Stop cci) = interesting cci
450 interesting (ArgCtxt rules) = rules
451 interesting _ = False
455 %************************************************************************
459 %************************************************************************
461 Inlining is controlled partly by the SimplifierMode switch. This has two
464 SimplGently (a) Simplifying before specialiser/full laziness
465 (b) Simplifiying inside InlineRules
466 (c) Simplifying the LHS of a rule
467 (d) Simplifying a GHCi expression or Template
470 SimplPhase n _ Used at all other times
474 Gentle mode has a separate boolean flag to control
475 a) inlining (sm_inline flag)
476 b) rules (sm_rules flag)
477 A key invariant about Gentle mode is that it is treated as the EARLIEST
481 simplEnvForGHCi :: SimplEnv
482 simplEnvForGHCi = mkSimplEnv allOffSwitchChecker $
483 SimplGently { sm_rules = True, sm_inline = False }
484 -- Do not do any inlining, in case we expose some unboxed
485 -- tuple stuff that confuses the bytecode interpreter
487 simplEnvForRules :: SimplEnv
488 simplEnvForRules = mkSimplEnv allOffSwitchChecker $
489 SimplGently { sm_rules = True, sm_inline = False }
491 updModeForInlineRules :: Activation -> SimplifierMode -> SimplifierMode
492 -- See Note [Simplifying inside InlineRules]
493 updModeForInlineRules _inline_rule_act _current_mode
494 = SimplGently { sm_rules = True, sm_inline = True }
497 Note [Inlining in gentle mode]
498 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
499 Something is inlined if
500 (i) the sm_inline flag is on, AND
501 (ii) the thing has an INLINE pragma, AND
502 (iii) the thing is inlinable in the earliest phase.
504 Example of why (iii) is important:
505 {-# INLINE [~1] g #-}
511 If we were to inline g into f's inlining, then an importing module would
513 f e --> g (g e) ---> RULE fires
514 because the InlineRule for f has had g inlined into it.
516 On the other hand, it is bad not to do ANY inlining into an
517 InlineRule, because then recursive knots in instance declarations
518 don't get unravelled.
520 However, *sometimes* SimplGently must do no call-site inlining at all
521 (hence sm_inline = False). Before full laziness we must be careful
522 not to inline wrappers, because doing so inhibits floating
523 e.g. ...(case f x of ...)...
524 ==> ...(case (case x of I# x# -> fw x#) of ...)...
525 ==> ...(case x of I# x# -> case fw x# of ...)...
526 and now the redex (f x) isn't floatable any more.
528 The no-inlining thing is also important for Template Haskell. You might be
529 compiling in one-shot mode with -O2; but when TH compiles a splice before
530 running it, we don't want to use -O2. Indeed, we don't want to inline
531 anything, because the byte-code interpreter might get confused about
532 unboxed tuples and suchlike.
534 Note [RULEs enabled in SimplGently]
535 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
536 RULES are enabled when doing "gentle" simplification. Two reasons:
538 * We really want the class-op cancellation to happen:
539 op (df d1 d2) --> $cop3 d1 d2
540 because this breaks the mutual recursion between 'op' and 'df'
544 to work in Template Haskell when simplifying
545 splices, so we get simpler code for literal strings
547 But watch out: list fusion can prevent floating. So use phase control
548 to switch off those rules until after floating.
550 Currently (Oct10) I think that sm_rules is always True, so we
553 Note [Simplifying inside InlineRules]
554 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
555 We must take care with simplification inside InlineRules (which come from
558 First, consider the following example
563 in ...g...g...g...g...g...
564 Now, if that's the ONLY occurrence of f, it might be inlined inside g,
565 and thence copied multiple times when g is inlined. HENCE we treat
566 any occurrence in an InlineRule as a multiple occurrence, not a single
567 one; see OccurAnal.addRuleUsage.
569 Second, we do want *do* to some modest rules/inlining stuff in InlineRules,
570 partly to eliminate senseless crap, and partly to break the recursive knots
571 generated by instance declarations. To keep things simple, we always set
572 the phase to 'gentle' when processing InlineRules.
574 A note about wrappers
575 ~~~~~~~~~~~~~~~~~~~~~
576 It's also important not to inline a worker back into a wrapper.
578 wraper = inline_me (\x -> ...worker... )
579 Normally, the inline_me prevents the worker getting inlined into
580 the wrapper (initially, the worker's only call site!). But,
581 if the wrapper is sure to be called, the strictness analyser will
582 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
586 activeUnfolding :: SimplEnv -> IdUnfoldingFun
588 = case getMode env of
589 SimplGently { sm_inline = False } -> active_unfolding_minimal
590 SimplGently { sm_inline = True } -> active_unfolding_gentle
591 SimplPhase n _ -> active_unfolding n
593 activeUnfInRule :: SimplEnv -> IdUnfoldingFun
594 -- When matching in RULE, we want to "look through" an unfolding
595 -- (to see a constructor) if *rules* are on, even if *inlinings*
596 -- are not. A notable example is DFuns, which really we want to
597 -- match in rules like (op dfun) in gentle mode. Another example
598 -- is 'otherwise' which we want exprIsConApp_maybe to be able to
601 = case getMode env of
602 SimplGently { sm_rules = False } -> active_unfolding_minimal
603 SimplGently { sm_rules = True } -> active_unfolding_early
604 SimplPhase n _ -> active_unfolding n
606 active_unfolding_early id
607 | isEarlyActive (idInlineActivation id) = idUnfolding id
608 | otherwise = idUnfolding id
610 active_unfolding_minimal :: IdUnfoldingFun
611 -- Compuslory unfoldings only
612 -- Ignore SimplGently, because we want to inline regardless;
613 -- the Id has no top-level binding at all
615 -- NB: we used to have a second exception, for data con wrappers.
616 -- On the grounds that we use gentle mode for rule LHSs, and
617 -- they match better when data con wrappers are inlined.
618 -- But that only really applies to the trivial wrappers (like (:)),
619 -- and they are now constructed as Compulsory unfoldings (in MkId)
620 -- so they'll happen anyway.
621 active_unfolding_minimal id
622 | isCompulsoryUnfolding unf = unf
623 | otherwise = NoUnfolding
627 active_unfolding_gentle :: IdUnfoldingFun
628 -- Anything that is early-active
629 -- See Note [Gentle mode]
630 active_unfolding_gentle id
631 | isStableUnfolding unf
632 , isEarlyActive (idInlineActivation id) = unf
633 -- NB: wrappers are not early-active
634 | otherwise = NoUnfolding
637 -- idUnfolding checks for loop-breakers
638 -- Things with an INLINE pragma may have
639 -- an unfolding *and* be a loop breaker
640 -- (maybe the knot is not yet untied)
642 active_unfolding :: CompilerPhase -> IdUnfoldingFun
643 active_unfolding n id
644 | isActive n (idInlineActivation id) = idUnfolding id
645 | otherwise = NoUnfolding
647 activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
648 -- Nothing => No rules at all
649 activeRule dflags env
650 | not (dopt Opt_EnableRewriteRules dflags)
651 = Nothing -- Rewriting is off
653 = case getMode env of
654 SimplGently { sm_rules = rules_on }
655 | rules_on -> Just isEarlyActive -- Note [RULEs enabled in SimplGently]
656 | otherwise -> Nothing
657 SimplPhase n _ -> Just (isActive n)
660 --------------------------------------------------------------
662 Preserved just for now (Oct 10)
663 --------------------------------------------------------------
665 OK, so suppose we have
666 {-# INLINE <act> f #-}
668 meaning "inline f in phases p where activation <act>(p) holds".
669 Then what inlinings/rules can we apply to the copy of <rhs> captured in
670 f's InlineRule? Our model is that literally <rhs> is substituted for
671 f when it is inlined. So our conservative plan (implemented by
672 updModeForInlineRules) is this:
674 -------------------------------------------------------------
675 When simplifying the RHS of an InlineRule,
676 If the InlineRule becomes active in phase p, then
677 if the current phase is *earlier than* p,
678 make no inlinings or rules active when simplifying the RHS
680 set the phase to p when simplifying the RHS
682 -- Treat Gentle as phase "infinity"
683 -- If current_phase `earlier than` inline_rule_start_phase
686 -- if current_phase `same phase` inline_rule_start_phase
687 -- then current_phase (keep gentle flags)
688 -- else inline_rule_start_phase
689 -------------------------------------------------------------
693 a) Rules/inlinings that *cease* being active before p will
694 not apply to the InlineRule rhs, consistent with it being
695 inlined in its *original* form in phase p.
697 b) Rules/inlinings that only become active *after* p will
698 not apply to the InlineRule rhs, again to be consistent with
699 inlining the *original* rhs in phase p.
705 {-# NOINLINE [1] g #-}
708 {-# RULE h g = ... #-}
709 Here we must not inline g into f's RHS, even when we get to phase 0,
710 because when f is later inlined into some other module we want the
718 and suppose that there are auto-generated specialisations and a strictness
719 wrapper for g. The specialisations get activation AlwaysActive, and the
720 strictness wrapper get activation (ActiveAfter 0). So the strictness
721 wrepper fails the test and won't be inlined into f's InlineRule. That
722 means f can inline, expose the specialised call to g, so the specialisation
725 --------------------------------------------------------------
727 --------------------------------------------------------------
730 %************************************************************************
732 preInlineUnconditionally
734 %************************************************************************
736 preInlineUnconditionally
737 ~~~~~~~~~~~~~~~~~~~~~~~~
738 @preInlineUnconditionally@ examines a bndr to see if it is used just
739 once in a completely safe way, so that it is safe to discard the
740 binding inline its RHS at the (unique) usage site, REGARDLESS of how
741 big the RHS might be. If this is the case we don't simplify the RHS
742 first, but just inline it un-simplified.
744 This is much better than first simplifying a perhaps-huge RHS and then
745 inlining and re-simplifying it. Indeed, it can be at least quadratically
754 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
755 This can happen with cascades of functions too:
762 THE MAIN INVARIANT is this:
764 ---- preInlineUnconditionally invariant -----
765 IF preInlineUnconditionally chooses to inline x = <rhs>
766 THEN doing the inlining should not change the occurrence
767 info for the free vars of <rhs>
768 ----------------------------------------------
770 For example, it's tempting to look at trivial binding like
772 and inline it unconditionally. But suppose x is used many times,
773 but this is the unique occurrence of y. Then inlining x would change
774 y's occurrence info, which breaks the invariant. It matters: y
775 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
778 Even RHSs labelled InlineMe aren't caught here, because there might be
779 no benefit from inlining at the call site.
781 [Sept 01] Don't unconditionally inline a top-level thing, because that
782 can simply make a static thing into something built dynamically. E.g.
786 [Remember that we treat \s as a one-shot lambda.] No point in
787 inlining x unless there is something interesting about the call site.
789 But watch out: if you aren't careful, some useful foldr/build fusion
790 can be lost (most notably in spectral/hartel/parstof) because the
791 foldr didn't see the build. Doing the dynamic allocation isn't a big
792 deal, in fact, but losing the fusion can be. But the right thing here
793 seems to be to do a callSiteInline based on the fact that there is
794 something interesting about the call site (it's strict). Hmm. That
797 Conclusion: inline top level things gaily until Phase 0 (the last
798 phase), at which point don't.
800 Note [pre/postInlineUnconditionally in gentle mode]
801 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
802 Even in gentle mode we want to do preInlineUnconditionally. The
803 reason is that too little clean-up happens if you don't inline
804 use-once things. Also a bit of inlining is *good* for full laziness;
805 it can expose constant sub-expressions. Example in
806 spectral/mandel/Mandel.hs, where the mandelset function gets a useful
807 let-float if you inline windowToViewport
809 However, as usual for Gentle mode, do not inline things that are
810 inactive in the intial stages. See Note [Gentle mode].
812 Note [InlineRule and preInlineUnconditionally]
813 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
814 Surprisingly, do not pre-inline-unconditionally Ids with INLINE pragmas!
824 ...fInt...fInt...fInt...
826 Here f occurs just once, in the RHS of f1. But if we inline it there
827 we'll lose the opportunity to inline at each of fInt's call sites.
828 The INLINE pragma will only inline when the application is saturated
829 for exactly this reason; and we don't want PreInlineUnconditionally
830 to second-guess it. A live example is Trac #3736.
831 c.f. Note [InlineRule and postInlineUnconditionally]
833 Note [Top-level botomming Ids]
834 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
835 Don't inline top-level Ids that are bottoming, even if they are used just
836 once, because FloatOut has gone to some trouble to extract them out.
837 Inlining them won't make the program run faster!
840 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
841 preInlineUnconditionally env top_lvl bndr rhs
843 | isStableUnfolding (idUnfolding bndr) = False -- Note [InlineRule and preInlineUnconditionally]
844 | isTopLevel top_lvl && isBottomingId bndr = False -- Note [Top-level bottoming Ids]
845 | opt_SimplNoPreInlining = False
846 | otherwise = case idOccInfo bndr of
847 IAmDead -> True -- Happens in ((\x.1) v)
848 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
852 active = case phase of
853 SimplGently {} -> isEarlyActive act
854 -- See Note [pre/postInlineUnconditionally in gentle mode]
855 SimplPhase n _ -> isActive n act
856 act = idInlineActivation bndr
857 try_once in_lam int_cxt -- There's one textual occurrence
858 | not in_lam = isNotTopLevel top_lvl || early_phase
859 | otherwise = int_cxt && canInlineInLam rhs
861 -- Be very careful before inlining inside a lambda, because (a) we must not
862 -- invalidate occurrence information, and (b) we want to avoid pushing a
863 -- single allocation (here) into multiple allocations (inside lambda).
864 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
865 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
867 -- is_cheap = exprIsCheap rhs
868 -- ok = is_cheap && int_cxt
870 -- int_cxt The context isn't totally boring
871 -- E.g. let f = \ab.BIG in \y. map f xs
872 -- Don't want to substitute for f, because then we allocate
873 -- its closure every time the \y is called
874 -- But: let f = \ab.BIG in \y. map (f y) xs
875 -- Now we do want to substitute for f, even though it's not
876 -- saturated, because we're going to allocate a closure for
877 -- (f y) every time round the loop anyhow.
879 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
880 -- so substituting rhs inside a lambda doesn't change the occ info.
881 -- Sadly, not quite the same as exprIsHNF.
882 canInlineInLam (Lit _) = True
883 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
884 canInlineInLam (Note _ e) = canInlineInLam e
885 canInlineInLam _ = False
887 early_phase = case phase of
888 SimplPhase 0 _ -> False
890 -- If we don't have this early_phase test, consider
891 -- x = length [1,2,3]
892 -- The full laziness pass carefully floats all the cons cells to
893 -- top level, and preInlineUnconditionally floats them all back in.
894 -- Result is (a) static allocation replaced by dynamic allocation
895 -- (b) many simplifier iterations because this tickles
896 -- a related problem; only one inlining per pass
898 -- On the other hand, I have seen cases where top-level fusion is
899 -- lost if we don't inline top level thing (e.g. string constants)
900 -- Hence the test for phase zero (which is the phase for all the final
901 -- simplifications). Until phase zero we take no special notice of
902 -- top level things, but then we become more leery about inlining
907 %************************************************************************
909 postInlineUnconditionally
911 %************************************************************************
913 postInlineUnconditionally
914 ~~~~~~~~~~~~~~~~~~~~~~~~~
915 @postInlineUnconditionally@ decides whether to unconditionally inline
916 a thing based on the form of its RHS; in particular if it has a
917 trivial RHS. If so, we can inline and discard the binding altogether.
919 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
920 only have *forward* references Hence, it's safe to discard the binding
922 NOTE: This isn't our last opportunity to inline. We're at the binding
923 site right now, and we'll get another opportunity when we get to the
926 Note that we do this unconditional inlining only for trival RHSs.
927 Don't inline even WHNFs inside lambdas; doing so may simply increase
928 allocation when the function is called. This isn't the last chance; see
931 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
932 Because we don't even want to inline them into the RHS of constructor
933 arguments. See NOTE above
935 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
936 it's best to inline it anyway. We often get a=E; b=a from desugaring,
937 with both a and b marked NOINLINE. But that seems incompatible with
938 our new view that inlining is like a RULE, so I'm sticking to the 'active'
942 postInlineUnconditionally
943 :: SimplEnv -> TopLevelFlag
944 -> OutId -- The binder (an InId would be fine too)
945 -> OccInfo -- From the InId
949 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
951 | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, don't inline
952 -- because it might be referred to "earlier"
953 | isExportedId bndr = False
954 | isStableUnfolding unfolding = False -- Note [InlineRule and postInlineUnconditionally]
955 | exprIsTrivial rhs = True
956 | isTopLevel top_lvl = False -- Note [Top level and postInlineUnconditionally]
959 -- The point of examining occ_info here is that for *non-values*
960 -- that occur outside a lambda, the call-site inliner won't have
961 -- a chance (becuase it doesn't know that the thing
962 -- only occurs once). The pre-inliner won't have gotten
963 -- it either, if the thing occurs in more than one branch
964 -- So the main target is things like
967 -- True -> case x of ...
968 -- False -> case x of ...
969 -- This is very important in practice; e.g. wheel-seive1 doubles
970 -- in allocation if you miss this out
971 OneOcc in_lam _one_br int_cxt -- OneOcc => no code-duplication issue
972 -> smallEnoughToInline unfolding -- Small enough to dup
973 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
975 -- NB: Do NOT inline arbitrarily big things, even if one_br is True
976 -- Reason: doing so risks exponential behaviour. We simplify a big
977 -- expression, inline it, and simplify it again. But if the
978 -- very same thing happens in the big expression, we get
980 -- PRINCIPLE: when we've already simplified an expression once,
981 -- make sure that we only inline it if it's reasonably small.
984 -- Outside a lambda, we want to be reasonably aggressive
985 -- about inlining into multiple branches of case
986 -- e.g. let x = <non-value>
987 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
988 -- Inlining can be a big win if C3 is the hot-spot, even if
989 -- the uses in C1, C2 are not 'interesting'
990 -- An example that gets worse if you add int_cxt here is 'clausify'
992 (isCheapUnfolding unfolding && int_cxt))
993 -- isCheap => acceptable work duplication; in_lam may be true
994 -- int_cxt to prevent us inlining inside a lambda without some
995 -- good reason. See the notes on int_cxt in preInlineUnconditionally
997 IAmDead -> True -- This happens; for example, the case_bndr during case of
998 -- known constructor: case (a,b) of x { (p,q) -> ... }
999 -- Here x isn't mentioned in the RHS, so we don't want to
1000 -- create the (dead) let-binding let x = (a,b) in ...
1004 -- Here's an example that we don't handle well:
1005 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
1006 -- in \y. ....case f of {...} ....
1007 -- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
1009 -- - We can't preInlineUnconditionally because that woud invalidate
1010 -- the occ info for b.
1011 -- - We can't postInlineUnconditionally because the RHS is big, and
1012 -- that risks exponential behaviour
1013 -- - We can't call-site inline, because the rhs is big
1017 active = case getMode env of
1018 SimplGently {} -> isEarlyActive act
1019 -- See Note [pre/postInlineUnconditionally in gentle mode]
1020 SimplPhase n _ -> isActive n act
1021 act = idInlineActivation bndr
1024 Note [Top level and postInlineUnconditionally]
1025 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1026 We don't do postInlineUnconditionally for top-level things (exept ones that
1028 * There is no point, because the main goal is to get rid of local
1029 bindings used in multiple case branches.
1030 * Doing so will inline top-level error expressions that have been
1031 carefully floated out by FloatOut. More generally, it might
1032 replace static allocation with dynamic.
1034 Note [InlineRule and postInlineUnconditionally]
1035 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1036 Do not do postInlineUnconditionally if the Id has an InlineRule, otherwise
1037 we lose the unfolding. Example
1039 -- f has InlineRule with rhs (e |> co)
1043 Then there's a danger we'll optimise to
1048 and now postInlineUnconditionally, losing the InlineRule on f. Now f'
1049 won't inline because 'e' is too big.
1051 c.f. Note [InlineRule and preInlineUnconditionally]
1054 %************************************************************************
1058 %************************************************************************
1061 mkLam :: SimplEnv -> [OutBndr] -> OutExpr -> SimplM OutExpr
1062 -- mkLam tries three things
1063 -- a) eta reduction, if that gives a trivial expression
1064 -- b) eta expansion [only if there are some value lambdas]
1068 mkLam _env bndrs body
1069 = do { dflags <- getDOptsSmpl
1070 ; mkLam' dflags bndrs body }
1072 mkLam' :: DynFlags -> [OutBndr] -> OutExpr -> SimplM OutExpr
1073 mkLam' dflags bndrs (Cast body co)
1074 | not (any bad bndrs)
1075 -- Note [Casts and lambdas]
1076 = do { lam <- mkLam' dflags bndrs body
1077 ; return (mkCoerce (mkPiTypes bndrs co) lam) }
1079 co_vars = tyVarsOfType co
1080 bad bndr = isCoVar bndr && bndr `elemVarSet` co_vars
1082 mkLam' dflags bndrs body@(Lam {})
1083 = mkLam' dflags (bndrs ++ bndrs1) body1
1085 (bndrs1, body1) = collectBinders body
1087 mkLam' dflags bndrs body
1088 | dopt Opt_DoEtaReduction dflags
1089 , Just etad_lam <- tryEtaReduce bndrs body
1090 = do { tick (EtaReduction (head bndrs))
1094 = return (mkLams bndrs body)
1098 Note [Casts and lambdas]
1099 ~~~~~~~~~~~~~~~~~~~~~~~~
1101 (\x. (\y. e) `cast` g1) `cast` g2
1102 There is a danger here that the two lambdas look separated, and the
1103 full laziness pass might float an expression to between the two.
1105 So this equation in mkLam' floats the g1 out, thus:
1106 (\x. e `cast` g1) --> (\x.e) `cast` (tx -> g1)
1109 In general, this floats casts outside lambdas, where (I hope) they
1110 might meet and cancel with some other cast:
1111 \x. e `cast` co ===> (\x. e) `cast` (tx -> co)
1112 /\a. e `cast` co ===> (/\a. e) `cast` (/\a. co)
1113 /\g. e `cast` co ===> (/\g. e) `cast` (/\g. co)
1114 (if not (g `in` co))
1116 Notice that it works regardless of 'e'. Originally it worked only
1117 if 'e' was itself a lambda, but in some cases that resulted in
1118 fruitless iteration in the simplifier. A good example was when
1119 compiling Text.ParserCombinators.ReadPrec, where we had a definition
1120 like (\x. Get `cast` g)
1121 where Get is a constructor with nonzero arity. Then mkLam eta-expanded
1122 the Get, and the next iteration eta-reduced it, and then eta-expanded
1125 Note also the side condition for the case of coercion binders.
1126 It does not make sense to transform
1127 /\g. e `cast` g ==> (/\g.e) `cast` (/\g.g)
1128 because the latter is not well-kinded.
1130 %************************************************************************
1134 %************************************************************************
1137 tryEtaExpand :: SimplEnv -> OutId -> OutExpr -> SimplM (Arity, OutExpr)
1138 -- See Note [Eta-expanding at let bindings]
1139 tryEtaExpand env bndr rhs
1140 = do { dflags <- getDOptsSmpl
1141 ; (new_arity, new_rhs) <- try_expand dflags
1143 ; WARN( new_arity < old_arity || new_arity < _dmd_arity,
1144 (ptext (sLit "Arity decrease:") <+> (ppr bndr <+> ppr old_arity
1145 <+> ppr new_arity <+> ppr _dmd_arity) $$ ppr new_rhs) )
1146 -- Note [Arity decrease]
1147 return (new_arity, new_rhs) }
1150 | dopt Opt_DoLambdaEtaExpansion dflags
1151 , not (exprIsTrivial rhs)
1152 , not (inGentleMode env) -- In gentle mode don't eta-expansion
1153 -- because it can clutter up the code
1154 -- with casts etc that may not be removed
1155 , let new_arity = exprEtaExpandArity dflags rhs
1156 , new_arity > old_arity
1157 = do { tick (EtaExpansion bndr)
1158 ; return (new_arity, etaExpand new_arity rhs) }
1160 = return (exprArity rhs, rhs)
1162 old_arity = idArity bndr
1163 _dmd_arity = length $ fst $ splitStrictSig $ idStrictness bndr
1166 Note [Eta-expanding at let bindings]
1167 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1168 We now eta expand at let-bindings, which is where the payoff
1171 One useful consequence is this example:
1172 genMap :: C a => ...
1173 {-# INLINE genMap #-}
1177 {-# INLINE myMap #-}
1180 Notice that 'genMap' should only inline if applied to two arguments.
1181 In the InlineRule for myMap we'll have the unfolding
1182 (\d -> genMap Int (..d..))
1183 We do not want to eta-expand to
1184 (\d f xs -> genMap Int (..d..) f xs)
1185 because then 'genMap' will inline, and it really shouldn't: at least
1186 as far as the programmer is concerned, it's not applied to two
1190 %************************************************************************
1192 \subsection{Floating lets out of big lambdas}
1194 %************************************************************************
1196 Note [Floating and type abstraction]
1197 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1200 We'd like to float this to
1203 x = /\a. C (y1 a) (y2 a)
1204 for the usual reasons: we want to inline x rather vigorously.
1206 You may think that this kind of thing is rare. But in some programs it is
1207 common. For example, if you do closure conversion you might get:
1209 data a :-> b = forall e. (e -> a -> b) :$ e
1211 f_cc :: forall a. a :-> a
1212 f_cc = /\a. (\e. id a) :$ ()
1214 Now we really want to inline that f_cc thing so that the
1215 construction of the closure goes away.
1217 So I have elaborated simplLazyBind to understand right-hand sides that look
1221 and treat them specially. The real work is done in SimplUtils.abstractFloats,
1222 but there is quite a bit of plumbing in simplLazyBind as well.
1224 The same transformation is good when there are lets in the body:
1226 /\abc -> let(rec) x = e in b
1228 let(rec) x' = /\abc -> let x = x' a b c in e
1230 /\abc -> let x = x' a b c in b
1232 This is good because it can turn things like:
1234 let f = /\a -> letrec g = ... g ... in g
1236 letrec g' = /\a -> ... g' a ...
1238 let f = /\ a -> g' a
1240 which is better. In effect, it means that big lambdas don't impede
1243 This optimisation is CRUCIAL in eliminating the junk introduced by
1244 desugaring mutually recursive definitions. Don't eliminate it lightly!
1246 [May 1999] If we do this transformation *regardless* then we can
1247 end up with some pretty silly stuff. For example,
1250 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
1255 st = /\s -> ...[y1 s/x1, y2 s/x2]
1258 Unless the "..." is a WHNF there is really no point in doing this.
1259 Indeed it can make things worse. Suppose x1 is used strictly,
1262 x1* = case f y of { (a,b) -> e }
1264 If we abstract this wrt the tyvar we then can't do the case inline
1265 as we would normally do.
1267 That's why the whole transformation is part of the same process that
1268 floats let-bindings and constructor arguments out of RHSs. In particular,
1269 it is guarded by the doFloatFromRhs call in simplLazyBind.
1273 abstractFloats :: [OutTyVar] -> SimplEnv -> OutExpr -> SimplM ([OutBind], OutExpr)
1274 abstractFloats main_tvs body_env body
1275 = ASSERT( notNull body_floats )
1276 do { (subst, float_binds) <- mapAccumLM abstract empty_subst body_floats
1277 ; return (float_binds, CoreSubst.substExpr (text "abstract_floats1") subst body) }
1279 main_tv_set = mkVarSet main_tvs
1280 body_floats = getFloats body_env
1281 empty_subst = CoreSubst.mkEmptySubst (seInScope body_env)
1283 abstract :: CoreSubst.Subst -> OutBind -> SimplM (CoreSubst.Subst, OutBind)
1284 abstract subst (NonRec id rhs)
1285 = do { (poly_id, poly_app) <- mk_poly tvs_here id
1286 ; let poly_rhs = mkLams tvs_here rhs'
1287 subst' = CoreSubst.extendIdSubst subst id poly_app
1288 ; return (subst', (NonRec poly_id poly_rhs)) }
1290 rhs' = CoreSubst.substExpr (text "abstract_floats2") subst rhs
1291 tvs_here | any isCoVar main_tvs = main_tvs -- Note [Abstract over coercions]
1293 = varSetElems (main_tv_set `intersectVarSet` exprSomeFreeVars isTyCoVar rhs')
1295 -- Abstract only over the type variables free in the rhs
1296 -- wrt which the new binding is abstracted. But the naive
1297 -- approach of abstract wrt the tyvars free in the Id's type
1299 -- /\ a b -> let t :: (a,b) = (e1, e2)
1302 -- Here, b isn't free in x's type, but we must nevertheless
1303 -- abstract wrt b as well, because t's type mentions b.
1304 -- Since t is floated too, we'd end up with the bogus:
1305 -- poly_t = /\ a b -> (e1, e2)
1306 -- poly_x = /\ a -> fst (poly_t a *b*)
1307 -- So for now we adopt the even more naive approach of
1308 -- abstracting wrt *all* the tyvars. We'll see if that
1309 -- gives rise to problems. SLPJ June 98
1311 abstract subst (Rec prs)
1312 = do { (poly_ids, poly_apps) <- mapAndUnzipM (mk_poly tvs_here) ids
1313 ; let subst' = CoreSubst.extendSubstList subst (ids `zip` poly_apps)
1314 poly_rhss = [mkLams tvs_here (CoreSubst.substExpr (text "abstract_floats3") subst' rhs)
1316 ; return (subst', Rec (poly_ids `zip` poly_rhss)) }
1318 (ids,rhss) = unzip prs
1319 -- For a recursive group, it's a bit of a pain to work out the minimal
1320 -- set of tyvars over which to abstract:
1321 -- /\ a b c. let x = ...a... in
1322 -- letrec { p = ...x...q...
1323 -- q = .....p...b... } in
1325 -- Since 'x' is abstracted over 'a', the {p,q} group must be abstracted
1326 -- over 'a' (because x is replaced by (poly_x a)) as well as 'b'.
1327 -- Since it's a pain, we just use the whole set, which is always safe
1329 -- If you ever want to be more selective, remember this bizarre case too:
1331 -- Here, we must abstract 'x' over 'a'.
1334 mk_poly tvs_here var
1335 = do { uniq <- getUniqueM
1336 ; let poly_name = setNameUnique (idName var) uniq -- Keep same name
1337 poly_ty = mkForAllTys tvs_here (idType var) -- But new type of course
1338 poly_id = transferPolyIdInfo var tvs_here $ -- Note [transferPolyIdInfo] in Id.lhs
1339 mkLocalId poly_name poly_ty
1340 ; return (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tvs_here)) }
1341 -- In the olden days, it was crucial to copy the occInfo of the original var,
1342 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1343 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1344 -- at already simplified code, so it doesn't matter
1346 -- It's even right to retain single-occurrence or dead-var info:
1347 -- Suppose we started with /\a -> let x = E in B
1348 -- where x occurs once in B. Then we transform to:
1349 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1350 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1351 -- the occurrences of x' will be just the occurrences originally
1355 Note [Abstract over coercions]
1356 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1357 If a coercion variable (g :: a ~ Int) is free in the RHS, then so is the
1358 type variable a. Rather than sort this mess out, we simply bale out and abstract
1359 wrt all the type variables if any of them are coercion variables.
1362 Historical note: if you use let-bindings instead of a substitution, beware of this:
1364 -- Suppose we start with:
1366 -- x = /\ a -> let g = G in E
1368 -- Then we'll float to get
1370 -- x = let poly_g = /\ a -> G
1371 -- in /\ a -> let g = poly_g a in E
1373 -- But now the occurrence analyser will see just one occurrence
1374 -- of poly_g, not inside a lambda, so the simplifier will
1375 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1376 -- (I used to think that the "don't inline lone occurrences" stuff
1377 -- would stop this happening, but since it's the *only* occurrence,
1378 -- PreInlineUnconditionally kicks in first!)
1380 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1381 -- to appear many times. (NB: mkInlineMe eliminates
1382 -- such notes on trivial RHSs, so do it manually.)
1384 %************************************************************************
1388 %************************************************************************
1390 prepareAlts tries these things:
1392 1. Eliminate alternatives that cannot match, including the
1393 DEFAULT alternative.
1395 2. If the DEFAULT alternative can match only one possible constructor,
1396 then make that constructor explicit.
1398 case e of x { DEFAULT -> rhs }
1400 case e of x { (a,b) -> rhs }
1401 where the type is a single constructor type. This gives better code
1402 when rhs also scrutinises x or e.
1404 3. Returns a list of the constructors that cannot holds in the
1405 DEFAULT alternative (if there is one)
1407 Here "cannot match" includes knowledge from GADTs
1409 It's a good idea do do this stuff before simplifying the alternatives, to
1410 avoid simplifying alternatives we know can't happen, and to come up with
1411 the list of constructors that are handled, to put into the IdInfo of the
1412 case binder, for use when simplifying the alternatives.
1414 Eliminating the default alternative in (1) isn't so obvious, but it can
1417 data Colour = Red | Green | Blue
1426 DEFAULT -> [ case y of ... ]
1428 If we inline h into f, the default case of the inlined h can't happen.
1429 If we don't notice this, we may end up filtering out *all* the cases
1430 of the inner case y, which give us nowhere to go!
1433 prepareAlts :: OutExpr -> OutId -> [InAlt] -> SimplM ([AltCon], [InAlt])
1434 prepareAlts scrut case_bndr' alts
1435 = do { let (alts_wo_default, maybe_deflt) = findDefault alts
1436 alt_cons = [con | (con,_,_) <- alts_wo_default]
1437 imposs_deflt_cons = nub (imposs_cons ++ alt_cons)
1438 -- "imposs_deflt_cons" are handled
1439 -- EITHER by the context,
1440 -- OR by a non-DEFAULT branch in this case expression.
1442 ; default_alts <- prepareDefault case_bndr' mb_tc_app
1443 imposs_deflt_cons maybe_deflt
1445 ; let trimmed_alts = filterOut impossible_alt alts_wo_default
1446 merged_alts = mergeAlts trimmed_alts default_alts
1447 -- We need the mergeAlts in case the new default_alt
1448 -- has turned into a constructor alternative.
1449 -- The merge keeps the inner DEFAULT at the front, if there is one
1450 -- and interleaves the alternatives in the right order
1452 ; return (imposs_deflt_cons, merged_alts) }
1454 mb_tc_app = splitTyConApp_maybe (idType case_bndr')
1455 Just (_, inst_tys) = mb_tc_app
1457 imposs_cons = case scrut of
1458 Var v -> otherCons (idUnfolding v)
1461 impossible_alt :: CoreAlt -> Bool
1462 impossible_alt (con, _, _) | con `elem` imposs_cons = True
1463 impossible_alt (DataAlt con, _, _) = dataConCannotMatch inst_tys con
1464 impossible_alt _ = False
1467 prepareDefault :: OutId -- Case binder; need just for its type. Note that as an
1468 -- OutId, it has maximum information; this is important.
1469 -- Test simpl013 is an example
1470 -> Maybe (TyCon, [Type]) -- Type of scrutinee, decomposed
1471 -> [AltCon] -- These cons can't happen when matching the default
1472 -> Maybe InExpr -- Rhs
1473 -> SimplM [InAlt] -- Still unsimplified
1474 -- We use a list because it's what mergeAlts expects,
1476 --------- Fill in known constructor -----------
1477 prepareDefault case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
1478 | -- This branch handles the case where we are
1479 -- scrutinisng an algebraic data type
1480 isAlgTyCon tycon -- It's a data type, tuple, or unboxed tuples.
1481 , not (isNewTyCon tycon) -- We can have a newtype, if we are just doing an eval:
1482 -- case x of { DEFAULT -> e }
1483 -- and we don't want to fill in a default for them!
1484 , Just all_cons <- tyConDataCons_maybe tycon
1485 , not (null all_cons)
1486 -- This is a tricky corner case. If the data type has no constructors,
1487 -- which GHC allows, then the case expression will have at most a default
1488 -- alternative. We don't want to eliminate that alternative, because the
1489 -- invariant is that there's always one alternative. It's more convenient
1491 -- case x of { DEFAULT -> e }
1492 -- as it is, rather than transform it to
1493 -- error "case cant match"
1494 -- which would be quite legitmate. But it's a really obscure corner, and
1495 -- not worth wasting code on.
1496 , let imposs_data_cons = [con | DataAlt con <- imposs_cons] -- We now know it's a data type
1497 impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
1498 = case filterOut impossible all_cons of
1499 [] -> return [] -- Eliminate the default alternative
1500 -- altogether if it can't match
1502 [con] -> -- It matches exactly one constructor, so fill it in
1503 do { tick (FillInCaseDefault case_bndr)
1505 ; let (ex_tvs, co_tvs, arg_ids) =
1506 dataConRepInstPat us con inst_tys
1507 ; return [(DataAlt con, ex_tvs ++ co_tvs ++ arg_ids, deflt_rhs)] }
1509 _ -> return [(DEFAULT, [], deflt_rhs)]
1511 | debugIsOn, isAlgTyCon tycon
1512 , null (tyConDataCons tycon)
1513 , not (isFamilyTyCon tycon || isAbstractTyCon tycon)
1514 -- Check for no data constructors
1515 -- This can legitimately happen for abstract types and type families,
1516 -- so don't report that
1517 = pprTrace "prepareDefault" (ppr case_bndr <+> ppr tycon)
1518 $ return [(DEFAULT, [], deflt_rhs)]
1520 --------- Catch-all cases -----------
1521 prepareDefault _case_bndr _bndr_ty _imposs_cons (Just deflt_rhs)
1522 = return [(DEFAULT, [], deflt_rhs)]
1524 prepareDefault _case_bndr _bndr_ty _imposs_cons Nothing
1525 = return [] -- No default branch
1530 %************************************************************************
1534 %************************************************************************
1536 mkCase tries these things
1538 1. Merge Nested Cases
1540 case e of b { ==> case e of b {
1541 p1 -> rhs1 p1 -> rhs1
1543 pm -> rhsm pm -> rhsm
1544 _ -> case b of b' { pn -> let b'=b in rhsn
1546 ... po -> let b'=b in rhso
1547 po -> rhso _ -> let b'=b in rhsd
1551 which merges two cases in one case when -- the default alternative of
1552 the outer case scrutises the same variable as the outer case. This
1553 transformation is called Case Merging. It avoids that the same
1554 variable is scrutinised multiple times.
1556 2. Eliminate Identity Case
1562 and similar friends.
1564 3. Merge identical alternatives.
1565 If several alternatives are identical, merge them into
1566 a single DEFAULT alternative. I've occasionally seen this
1567 making a big difference:
1569 case e of =====> case e of
1570 C _ -> f x D v -> ....v....
1571 D v -> ....v.... DEFAULT -> f x
1574 The point is that we merge common RHSs, at least for the DEFAULT case.
1575 [One could do something more elaborate but I've never seen it needed.]
1576 To avoid an expensive test, we just merge branches equal to the *first*
1577 alternative; this picks up the common cases
1578 a) all branches equal
1579 b) some branches equal to the DEFAULT (which occurs first)
1581 The case where Merge Identical Alternatives transformation showed up
1582 was like this (base/Foreign/C/Err/Error.lhs):
1588 where @is@ was something like
1590 p `is` n = p /= (-1) && p == n
1592 This gave rise to a horrible sequence of cases
1599 and similarly in cascade for all the join points!
1603 mkCase, mkCase1, mkCase2
1606 -> [OutAlt] -- Alternatives in standard (increasing) order
1609 --------------------------------------------------
1610 -- 1. Merge Nested Cases
1611 --------------------------------------------------
1613 mkCase dflags scrut outer_bndr ((DEFAULT, _, deflt_rhs) : outer_alts)
1614 | dopt Opt_CaseMerge dflags
1615 , Case (Var inner_scrut_var) inner_bndr _ inner_alts <- deflt_rhs
1616 , inner_scrut_var == outer_bndr
1617 = do { tick (CaseMerge outer_bndr)
1619 ; let wrap_alt (con, args, rhs) = ASSERT( outer_bndr `notElem` args )
1620 (con, args, wrap_rhs rhs)
1621 -- Simplifier's no-shadowing invariant should ensure
1622 -- that outer_bndr is not shadowed by the inner patterns
1623 wrap_rhs rhs = Let (NonRec inner_bndr (Var outer_bndr)) rhs
1624 -- The let is OK even for unboxed binders,
1626 wrapped_alts | isDeadBinder inner_bndr = inner_alts
1627 | otherwise = map wrap_alt inner_alts
1629 merged_alts = mergeAlts outer_alts wrapped_alts
1630 -- NB: mergeAlts gives priority to the left
1633 -- DEFAULT -> case x of
1636 -- When we merge, we must ensure that e1 takes
1637 -- precedence over e2 as the value for A!
1639 ; mkCase1 dflags scrut outer_bndr merged_alts
1641 -- Warning: don't call mkCase recursively!
1642 -- Firstly, there's no point, because inner alts have already had
1643 -- mkCase applied to them, so they won't have a case in their default
1644 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1645 -- in munge_rhs may put a case into the DEFAULT branch!
1647 mkCase dflags scrut bndr alts = mkCase1 dflags scrut bndr alts
1649 --------------------------------------------------
1650 -- 2. Eliminate Identity Case
1651 --------------------------------------------------
1653 mkCase1 _dflags scrut case_bndr alts -- Identity case
1654 | all identity_alt alts
1655 = do { tick (CaseIdentity case_bndr)
1656 ; return (re_cast scrut) }
1658 identity_alt (con, args, rhs) = check_eq con args (de_cast rhs)
1660 check_eq DEFAULT _ (Var v) = v == case_bndr
1661 check_eq (LitAlt lit') _ (Lit lit) = lit == lit'
1662 check_eq (DataAlt con) args rhs = rhs `cheapEqExpr` mkConApp con (arg_tys ++ varsToCoreExprs args)
1663 || rhs `cheapEqExpr` Var case_bndr
1664 check_eq _ _ _ = False
1666 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1669 -- case e of x { _ -> x `cast` c }
1670 -- And we definitely want to eliminate this case, to give
1672 -- So we throw away the cast from the RHS, and reconstruct
1673 -- it at the other end. All the RHS casts must be the same
1674 -- if (all identity_alt alts) holds.
1676 -- Don't worry about nested casts, because the simplifier combines them
1677 de_cast (Cast e _) = e
1680 re_cast scrut = case head alts of
1681 (_,_,Cast _ co) -> Cast scrut co
1684 --------------------------------------------------
1685 -- 3. Merge Identical Alternatives
1686 --------------------------------------------------
1687 mkCase1 dflags scrut case_bndr ((_con1,bndrs1,rhs1) : con_alts)
1688 | all isDeadBinder bndrs1 -- Remember the default
1689 , length filtered_alts < length con_alts -- alternative comes first
1690 -- Also Note [Dead binders]
1691 = do { tick (AltMerge case_bndr)
1692 ; mkCase2 dflags scrut case_bndr alts' }
1694 alts' = (DEFAULT, [], rhs1) : filtered_alts
1695 filtered_alts = filter keep con_alts
1696 keep (_con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1698 mkCase1 dflags scrut bndr alts = mkCase2 dflags scrut bndr alts
1700 --------------------------------------------------
1702 --------------------------------------------------
1703 mkCase2 _dflags scrut bndr alts
1704 = return (Case scrut bndr (coreAltsType alts) alts)
1708 ~~~~~~~~~~~~~~~~~~~~
1709 Note that dead-ness is maintained by the simplifier, so that it is
1710 accurate after simplification as well as before.
1713 Note [Cascading case merge]
1714 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
1715 Case merging should cascade in one sweep, because it
1719 DEFAULT -> case a of b
1720 DEFAULT -> case b of c {
1727 DEFAULT -> case a of b
1728 DEFAULT -> let c = b in e
1729 A -> let c = b in ea
1734 DEFAULT -> let b = a in let c = b in e
1735 A -> let b = a in let c = b in ea
1736 B -> let b = a in eb
1740 However here's a tricky case that we still don't catch, and I don't
1741 see how to catch it in one pass:
1743 case x of c1 { I# a1 ->
1746 DEFAULT -> case x of c3 { I# a2 ->
1749 After occurrence analysis (and its binder-swap) we get this
1751 case x of c1 { I# a1 ->
1752 let x = c1 in -- Binder-swap addition
1755 DEFAULT -> case x of c3 { I# a2 ->
1758 When we simplify the inner case x, we'll see that
1759 x=c1=I# a1. So we'll bind a2 to a1, and get
1761 case x of c1 { I# a1 ->
1764 DEFAULT -> case a1 of ...
1766 This is corect, but we can't do a case merge in this sweep
1767 because c2 /= a1. Reason: the binding c1=I# a1 went inwards
1768 without getting changed to c1=I# c2.
1770 I don't think this is worth fixing, even if I knew how. It'll
1771 all come out in the next pass anyway.