Typos in comments [skip ci]
[ghc.git] / compiler / simplCore / Simplify.hs
1 {-
2 (c) The AQUA Project, Glasgow University, 1993-1998
3
4 \section[Simplify]{The main module of the simplifier}
5 -}
6
7 {-# LANGUAGE CPP #-}
8
9 module Simplify ( simplTopBinds, simplExpr, simplRules ) where
10
11 #include "HsVersions.h"
12
13 import DynFlags
14 import SimplMonad
15 import Type hiding ( substTy, substTyVar, extendTvSubst, extendCvSubst )
16 import SimplEnv
17 import SimplUtils
18 import FamInstEnv ( FamInstEnv )
19 import Literal ( litIsLifted ) --, mkMachInt ) -- temporalily commented out. See #8326
20 import Id
21 import MkId ( seqId, voidPrimId )
22 import MkCore ( mkImpossibleExpr, castBottomExpr )
23 import IdInfo
24 import Name ( Name, mkSystemVarName, isExternalName, getOccFS )
25 import Coercion hiding ( substCo, substCoVar )
26 import OptCoercion ( optCoercion )
27 import FamInstEnv ( topNormaliseType_maybe )
28 import DataCon ( DataCon, dataConWorkId, dataConRepStrictness
29 , isMarkedStrict, dataConRepArgTys ) --, dataConTyCon, dataConTag, fIRST_TAG )
30 --import TyCon ( isEnumerationTyCon ) -- temporalily commented out. See #8326
31 import CoreMonad ( Tick(..), SimplifierMode(..) )
32 import CoreSyn
33 import Demand ( StrictSig(..), dmdTypeDepth, isStrictDmd )
34 import PprCore ( pprCoreExpr )
35 import CoreUnfold
36 import CoreUtils
37 import CoreArity
38 --import PrimOp ( tagToEnumKey ) -- temporalily commented out. See #8326
39 import Rules ( mkRuleInfo, lookupRule, getRules )
40 import TysPrim ( voidPrimTy ) --, intPrimTy ) -- temporalily commented out. See #8326
41 import BasicTypes ( TopLevelFlag(..), isTopLevel, RecFlag(..) )
42 import MonadUtils ( foldlM, mapAccumLM, liftIO )
43 import Maybes ( orElse )
44 --import Unique ( hasKey ) -- temporalily commented out. See #8326
45 import Control.Monad
46 import Outputable
47 import FastString
48 import Pair
49 import Util
50 import ErrUtils
51
52 {-
53 The guts of the simplifier is in this module, but the driver loop for
54 the simplifier is in SimplCore.hs.
55
56
57 -----------------------------------------
58 *** IMPORTANT NOTE ***
59 -----------------------------------------
60 The simplifier used to guarantee that the output had no shadowing, but
61 it does not do so any more. (Actually, it never did!) The reason is
62 documented with simplifyArgs.
63
64
65 -----------------------------------------
66 *** IMPORTANT NOTE ***
67 -----------------------------------------
68 Many parts of the simplifier return a bunch of "floats" as well as an
69 expression. This is wrapped as a datatype SimplUtils.FloatsWith.
70
71 All "floats" are let-binds, not case-binds, but some non-rec lets may
72 be unlifted (with RHS ok-for-speculation).
73
74
75
76 -----------------------------------------
77 ORGANISATION OF FUNCTIONS
78 -----------------------------------------
79 simplTopBinds
80 - simplify all top-level binders
81 - for NonRec, call simplRecOrTopPair
82 - for Rec, call simplRecBind
83
84
85 ------------------------------
86 simplExpr (applied lambda) ==> simplNonRecBind
87 simplExpr (Let (NonRec ...) ..) ==> simplNonRecBind
88 simplExpr (Let (Rec ...) ..) ==> simplify binders; simplRecBind
89
90 ------------------------------
91 simplRecBind [binders already simplfied]
92 - use simplRecOrTopPair on each pair in turn
93
94 simplRecOrTopPair [binder already simplified]
95 Used for: recursive bindings (top level and nested)
96 top-level non-recursive bindings
97 Returns:
98 - check for PreInlineUnconditionally
99 - simplLazyBind
100
101 simplNonRecBind
102 Used for: non-top-level non-recursive bindings
103 beta reductions (which amount to the same thing)
104 Because it can deal with strict arts, it takes a
105 "thing-inside" and returns an expression
106
107 - check for PreInlineUnconditionally
108 - simplify binder, including its IdInfo
109 - if strict binding
110 simplStrictArg
111 mkAtomicArgs
112 completeNonRecX
113 else
114 simplLazyBind
115 addFloats
116
117 simplNonRecX: [given a *simplified* RHS, but an *unsimplified* binder]
118 Used for: binding case-binder and constr args in a known-constructor case
119 - check for PreInLineUnconditionally
120 - simplify binder
121 - completeNonRecX
122
123 ------------------------------
124 simplLazyBind: [binder already simplified, RHS not]
125 Used for: recursive bindings (top level and nested)
126 top-level non-recursive bindings
127 non-top-level, but *lazy* non-recursive bindings
128 [must not be strict or unboxed]
129 Returns floats + an augmented environment, not an expression
130 - substituteIdInfo and add result to in-scope
131 [so that rules are available in rec rhs]
132 - simplify rhs
133 - mkAtomicArgs
134 - float if exposes constructor or PAP
135 - completeBind
136
137
138 completeNonRecX: [binder and rhs both simplified]
139 - if the the thing needs case binding (unlifted and not ok-for-spec)
140 build a Case
141 else
142 completeBind
143 addFloats
144
145 completeBind: [given a simplified RHS]
146 [used for both rec and non-rec bindings, top level and not]
147 - try PostInlineUnconditionally
148 - add unfolding [this is the only place we add an unfolding]
149 - add arity
150
151
152
153 Right hand sides and arguments
154 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
155 In many ways we want to treat
156 (a) the right hand side of a let(rec), and
157 (b) a function argument
158 in the same way. But not always! In particular, we would
159 like to leave these arguments exactly as they are, so they
160 will match a RULE more easily.
161
162 f (g x, h x)
163 g (+ x)
164
165 It's harder to make the rule match if we ANF-ise the constructor,
166 or eta-expand the PAP:
167
168 f (let { a = g x; b = h x } in (a,b))
169 g (\y. + x y)
170
171 On the other hand if we see the let-defns
172
173 p = (g x, h x)
174 q = + x
175
176 then we *do* want to ANF-ise and eta-expand, so that p and q
177 can be safely inlined.
178
179 Even floating lets out is a bit dubious. For let RHS's we float lets
180 out if that exposes a value, so that the value can be inlined more vigorously.
181 For example
182
183 r = let x = e in (x,x)
184
185 Here, if we float the let out we'll expose a nice constructor. We did experiments
186 that showed this to be a generally good thing. But it was a bad thing to float
187 lets out unconditionally, because that meant they got allocated more often.
188
189 For function arguments, there's less reason to expose a constructor (it won't
190 get inlined). Just possibly it might make a rule match, but I'm pretty skeptical.
191 So for the moment we don't float lets out of function arguments either.
192
193
194 Eta expansion
195 ~~~~~~~~~~~~~~
196 For eta expansion, we want to catch things like
197
198 case e of (a,b) -> \x -> case a of (p,q) -> \y -> r
199
200 If the \x was on the RHS of a let, we'd eta expand to bring the two
201 lambdas together. And in general that's a good thing to do. Perhaps
202 we should eta expand wherever we find a (value) lambda? Then the eta
203 expansion at a let RHS can concentrate solely on the PAP case.
204
205
206 ************************************************************************
207 * *
208 \subsection{Bindings}
209 * *
210 ************************************************************************
211 -}
212
213 simplTopBinds :: SimplEnv -> [InBind] -> SimplM SimplEnv
214
215 simplTopBinds env0 binds0
216 = do { -- Put all the top-level binders into scope at the start
217 -- so that if a transformation rule has unexpectedly brought
218 -- anything into scope, then we don't get a complaint about that.
219 -- It's rather as if the top-level binders were imported.
220 -- See note [Glomming] in OccurAnal.
221 ; env1 <- simplRecBndrs env0 (bindersOfBinds binds0)
222 ; env2 <- simpl_binds env1 binds0
223 ; freeTick SimplifierDone
224 ; return env2 }
225 where
226 -- We need to track the zapped top-level binders, because
227 -- they should have their fragile IdInfo zapped (notably occurrence info)
228 -- That's why we run down binds and bndrs' simultaneously.
229 --
230 simpl_binds :: SimplEnv -> [InBind] -> SimplM SimplEnv
231 simpl_binds env [] = return env
232 simpl_binds env (bind:binds) = do { env' <- simpl_bind env bind
233 ; simpl_binds env' binds }
234
235 simpl_bind env (Rec pairs) = simplRecBind env TopLevel pairs
236 simpl_bind env (NonRec b r) = do { (env', b') <- addBndrRules env b (lookupRecBndr env b)
237 ; simplRecOrTopPair env' TopLevel NonRecursive b b' r }
238
239 {-
240 ************************************************************************
241 * *
242 \subsection{Lazy bindings}
243 * *
244 ************************************************************************
245
246 simplRecBind is used for
247 * recursive bindings only
248 -}
249
250 simplRecBind :: SimplEnv -> TopLevelFlag
251 -> [(InId, InExpr)]
252 -> SimplM SimplEnv
253 simplRecBind env0 top_lvl pairs0
254 = do { (env_with_info, triples) <- mapAccumLM add_rules env0 pairs0
255 ; env1 <- go (zapFloats env_with_info) triples
256 ; return (env0 `addRecFloats` env1) }
257 -- addFloats adds the floats from env1,
258 -- _and_ updates env0 with the in-scope set from env1
259 where
260 add_rules :: SimplEnv -> (InBndr,InExpr) -> SimplM (SimplEnv, (InBndr, OutBndr, InExpr))
261 -- Add the (substituted) rules to the binder
262 add_rules env (bndr, rhs)
263 = do { (env', bndr') <- addBndrRules env bndr (lookupRecBndr env bndr)
264 ; return (env', (bndr, bndr', rhs)) }
265
266 go env [] = return env
267
268 go env ((old_bndr, new_bndr, rhs) : pairs)
269 = do { env' <- simplRecOrTopPair env top_lvl Recursive old_bndr new_bndr rhs
270 ; go env' pairs }
271
272 {-
273 simplOrTopPair is used for
274 * recursive bindings (whether top level or not)
275 * top-level non-recursive bindings
276
277 It assumes the binder has already been simplified, but not its IdInfo.
278 -}
279
280 simplRecOrTopPair :: SimplEnv
281 -> TopLevelFlag -> RecFlag
282 -> InId -> OutBndr -> InExpr -- Binder and rhs
283 -> SimplM SimplEnv -- Returns an env that includes the binding
284
285 simplRecOrTopPair env top_lvl is_rec old_bndr new_bndr rhs
286 = do { dflags <- getDynFlags
287 ; trace_bind dflags $
288 if preInlineUnconditionally dflags env top_lvl old_bndr rhs
289 -- Check for unconditional inline
290 then do tick (PreInlineUnconditionally old_bndr)
291 return (extendIdSubst env old_bndr (mkContEx env rhs))
292 else simplLazyBind env top_lvl is_rec old_bndr new_bndr rhs env }
293 where
294 trace_bind dflags thing_inside
295 | not (dopt Opt_D_verbose_core2core dflags)
296 = thing_inside
297 | otherwise
298 = pprTrace "SimplBind" (ppr old_bndr) thing_inside
299 -- trace_bind emits a trace for each top-level binding, which
300 -- helps to locate the tracing for inlining and rule firing
301
302 {-
303 simplLazyBind is used for
304 * [simplRecOrTopPair] recursive bindings (whether top level or not)
305 * [simplRecOrTopPair] top-level non-recursive bindings
306 * [simplNonRecE] non-top-level *lazy* non-recursive bindings
307
308 Nota bene:
309 1. It assumes that the binder is *already* simplified,
310 and is in scope, and its IdInfo too, except unfolding
311
312 2. It assumes that the binder type is lifted.
313
314 3. It does not check for pre-inline-unconditionally;
315 that should have been done already.
316 -}
317
318 simplLazyBind :: SimplEnv
319 -> TopLevelFlag -> RecFlag
320 -> InId -> OutId -- Binder, both pre-and post simpl
321 -- The OutId has IdInfo, except arity, unfolding
322 -> InExpr -> SimplEnv -- The RHS and its environment
323 -> SimplM SimplEnv
324 -- Precondition: rhs obeys the let/app invariant
325 simplLazyBind env top_lvl is_rec bndr bndr1 rhs rhs_se
326 = -- pprTrace "simplLazyBind" ((ppr bndr <+> ppr bndr1) $$ ppr rhs $$ ppr (seIdSubst rhs_se)) $
327 do { let rhs_env = rhs_se `setInScope` env
328 (tvs, body) = case collectTyAndValBinders rhs of
329 (tvs, [], body)
330 | surely_not_lam body -> (tvs, body)
331 _ -> ([], rhs)
332
333 surely_not_lam (Lam {}) = False
334 surely_not_lam (Tick t e)
335 | not (tickishFloatable t) = surely_not_lam e
336 -- eta-reduction could float
337 surely_not_lam _ = True
338 -- Do not do the "abstract tyyvar" thing if there's
339 -- a lambda inside, because it defeats eta-reduction
340 -- f = /\a. \x. g a x
341 -- should eta-reduce.
342
343
344 ; (body_env, tvs') <- simplBinders rhs_env tvs
345 -- See Note [Floating and type abstraction] in SimplUtils
346
347 -- Simplify the RHS
348 ; let rhs_cont = mkRhsStop (substTy body_env (exprType body))
349 ; (body_env1, body1) <- simplExprF body_env body rhs_cont
350 -- ANF-ise a constructor or PAP rhs
351 ; (body_env2, body2) <- prepareRhs top_lvl body_env1 bndr1 body1
352
353 ; (env', rhs')
354 <- if not (doFloatFromRhs top_lvl is_rec False body2 body_env2)
355 then -- No floating, revert to body1
356 do { rhs' <- mkLam tvs' (wrapFloats body_env1 body1) rhs_cont
357 ; return (env, rhs') }
358
359 else if null tvs then -- Simple floating
360 do { tick LetFloatFromLet
361 ; return (addFloats env body_env2, body2) }
362
363 else -- Do type-abstraction first
364 do { tick LetFloatFromLet
365 ; (poly_binds, body3) <- abstractFloats tvs' body_env2 body2
366 ; rhs' <- mkLam tvs' body3 rhs_cont
367 ; env' <- foldlM (addPolyBind top_lvl) env poly_binds
368 ; return (env', rhs') }
369
370 ; completeBind env' top_lvl bndr bndr1 rhs' }
371
372 {-
373 A specialised variant of simplNonRec used when the RHS is already simplified,
374 notably in knownCon. It uses case-binding where necessary.
375 -}
376
377 simplNonRecX :: SimplEnv
378 -> InId -- Old binder
379 -> OutExpr -- Simplified RHS
380 -> SimplM SimplEnv
381 -- Precondition: rhs satisfies the let/app invariant
382 simplNonRecX env bndr new_rhs
383 | isDeadBinder bndr -- Not uncommon; e.g. case (a,b) of c { (p,q) -> p }
384 = return env -- Here c is dead, and we avoid creating
385 -- the binding c = (a,b)
386
387 | Coercion co <- new_rhs
388 = return (extendCvSubst env bndr co)
389
390 | otherwise
391 = do { (env', bndr') <- simplBinder env bndr
392 ; completeNonRecX NotTopLevel env' (isStrictId bndr) bndr bndr' new_rhs }
393 -- simplNonRecX is only used for NotTopLevel things
394
395 completeNonRecX :: TopLevelFlag -> SimplEnv
396 -> Bool
397 -> InId -- Old binder
398 -> OutId -- New binder
399 -> OutExpr -- Simplified RHS
400 -> SimplM SimplEnv
401 -- Precondition: rhs satisfies the let/app invariant
402 -- See Note [CoreSyn let/app invariant] in CoreSyn
403
404 completeNonRecX top_lvl env is_strict old_bndr new_bndr new_rhs
405 = do { (env1, rhs1) <- prepareRhs top_lvl (zapFloats env) new_bndr new_rhs
406 ; (env2, rhs2) <-
407 if doFloatFromRhs NotTopLevel NonRecursive is_strict rhs1 env1
408 then do { tick LetFloatFromLet
409 ; return (addFloats env env1, rhs1) } -- Add the floats to the main env
410 else return (env, wrapFloats env1 rhs1) -- Wrap the floats around the RHS
411 ; completeBind env2 NotTopLevel old_bndr new_bndr rhs2 }
412
413 {-
414 {- No, no, no! Do not try preInlineUnconditionally in completeNonRecX
415 Doing so risks exponential behaviour, because new_rhs has been simplified once already
416 In the cases described by the folowing commment, postInlineUnconditionally will
417 catch many of the relevant cases.
418 -- This happens; for example, the case_bndr during case of
419 -- known constructor: case (a,b) of x { (p,q) -> ... }
420 -- Here x isn't mentioned in the RHS, so we don't want to
421 -- create the (dead) let-binding let x = (a,b) in ...
422 --
423 -- Similarly, single occurrences can be inlined vigourously
424 -- e.g. case (f x, g y) of (a,b) -> ....
425 -- If a,b occur once we can avoid constructing the let binding for them.
426
427 Furthermore in the case-binding case preInlineUnconditionally risks extra thunks
428 -- Consider case I# (quotInt# x y) of
429 -- I# v -> let w = J# v in ...
430 -- If we gaily inline (quotInt# x y) for v, we end up building an
431 -- extra thunk:
432 -- let w = J# (quotInt# x y) in ...
433 -- because quotInt# can fail.
434
435 | preInlineUnconditionally env NotTopLevel bndr new_rhs
436 = thing_inside (extendIdSubst env bndr (DoneEx new_rhs))
437 -}
438
439 ----------------------------------
440 prepareRhs takes a putative RHS, checks whether it's a PAP or
441 constructor application and, if so, converts it to ANF, so that the
442 resulting thing can be inlined more easily. Thus
443 x = (f a, g b)
444 becomes
445 t1 = f a
446 t2 = g b
447 x = (t1,t2)
448
449 We also want to deal well cases like this
450 v = (f e1 `cast` co) e2
451 Here we want to make e1,e2 trivial and get
452 x1 = e1; x2 = e2; v = (f x1 `cast` co) v2
453 That's what the 'go' loop in prepareRhs does
454 -}
455
456 prepareRhs :: TopLevelFlag -> SimplEnv -> OutId -> OutExpr -> SimplM (SimplEnv, OutExpr)
457 -- Adds new floats to the env iff that allows us to return a good RHS
458 prepareRhs top_lvl env id (Cast rhs co) -- Note [Float coercions]
459 | Pair ty1 _ty2 <- coercionKind co -- Do *not* do this if rhs has an unlifted type
460 , not (isUnliftedType ty1) -- see Note [Float coercions (unlifted)]
461 = do { (env', rhs') <- makeTrivialWithInfo top_lvl env (getOccFS id) sanitised_info rhs
462 ; return (env', Cast rhs' co) }
463 where
464 sanitised_info = vanillaIdInfo `setStrictnessInfo` strictnessInfo info
465 `setDemandInfo` demandInfo info
466 info = idInfo id
467
468 prepareRhs top_lvl env0 id rhs0
469 = do { (_is_exp, env1, rhs1) <- go 0 env0 rhs0
470 ; return (env1, rhs1) }
471 where
472 go n_val_args env (Cast rhs co)
473 = do { (is_exp, env', rhs') <- go n_val_args env rhs
474 ; return (is_exp, env', Cast rhs' co) }
475 go n_val_args env (App fun (Type ty))
476 = do { (is_exp, env', rhs') <- go n_val_args env fun
477 ; return (is_exp, env', App rhs' (Type ty)) }
478 go n_val_args env (App fun arg)
479 = do { (is_exp, env', fun') <- go (n_val_args+1) env fun
480 ; case is_exp of
481 True -> do { (env'', arg') <- makeTrivial top_lvl env' (getOccFS id) arg
482 ; return (True, env'', App fun' arg') }
483 False -> return (False, env, App fun arg) }
484 go n_val_args env (Var fun)
485 = return (is_exp, env, Var fun)
486 where
487 is_exp = isExpandableApp fun n_val_args -- The fun a constructor or PAP
488 -- See Note [CONLIKE pragma] in BasicTypes
489 -- The definition of is_exp should match that in
490 -- OccurAnal.occAnalApp
491
492 go n_val_args env (Tick t rhs)
493 -- We want to be able to float bindings past this
494 -- tick. Non-scoping ticks don't care.
495 | tickishScoped t == NoScope
496 = do { (is_exp, env', rhs') <- go n_val_args env rhs
497 ; return (is_exp, env', Tick t rhs') }
498 -- On the other hand, for scoping ticks we need to be able to
499 -- copy them on the floats, which in turn is only allowed if
500 -- we can obtain non-counting ticks.
501 | not (tickishCounts t) || tickishCanSplit t
502 = do { (is_exp, env', rhs') <- go n_val_args (zapFloats env) rhs
503 ; let tickIt (id, expr) = (id, mkTick (mkNoCount t) expr)
504 floats' = seFloats $ env `addFloats` mapFloats env' tickIt
505 ; return (is_exp, env' { seFloats = floats' }, Tick t rhs') }
506
507 go _ env other
508 = return (False, env, other)
509
510 {-
511 Note [Float coercions]
512 ~~~~~~~~~~~~~~~~~~~~~~
513 When we find the binding
514 x = e `cast` co
515 we'd like to transform it to
516 x' = e
517 x = x `cast` co -- A trivial binding
518 There's a chance that e will be a constructor application or function, or something
519 like that, so moving the coercion to the usage site may well cancel the coercions
520 and lead to further optimisation. Example:
521
522 data family T a :: *
523 data instance T Int = T Int
524
525 foo :: Int -> Int -> Int
526 foo m n = ...
527 where
528 x = T m
529 go 0 = 0
530 go n = case x of { T m -> go (n-m) }
531 -- This case should optimise
532
533 Note [Preserve strictness when floating coercions]
534 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
535 In the Note [Float coercions] transformation, keep the strictness info.
536 Eg
537 f = e `cast` co -- f has strictness SSL
538 When we transform to
539 f' = e -- f' also has strictness SSL
540 f = f' `cast` co -- f still has strictness SSL
541
542 Its not wrong to drop it on the floor, but better to keep it.
543
544 Note [Float coercions (unlifted)]
545 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
546 BUT don't do [Float coercions] if 'e' has an unlifted type.
547 This *can* happen:
548
549 foo :: Int = (error (# Int,Int #) "urk")
550 `cast` CoUnsafe (# Int,Int #) Int
551
552 If do the makeTrivial thing to the error call, we'll get
553 foo = case error (# Int,Int #) "urk" of v -> v `cast` ...
554 But 'v' isn't in scope!
555
556 These strange casts can happen as a result of case-of-case
557 bar = case (case x of { T -> (# 2,3 #); F -> error "urk" }) of
558 (# p,q #) -> p+q
559 -}
560
561 makeTrivialArg :: SimplEnv -> ArgSpec -> SimplM (SimplEnv, ArgSpec)
562 makeTrivialArg env (ValArg e) = do
563 { (env', e') <- makeTrivial NotTopLevel env (fsLit "arg") e
564 ; return (env', ValArg e') }
565 makeTrivialArg env arg = return (env, arg) -- CastBy, TyArg
566
567 makeTrivial :: TopLevelFlag -> SimplEnv
568 -> FastString -- ^ a "friendly name" to build the new binder from
569 -> OutExpr -> SimplM (SimplEnv, OutExpr)
570 -- Binds the expression to a variable, if it's not trivial, returning the variable
571 makeTrivial top_lvl env context expr =
572 makeTrivialWithInfo top_lvl env context vanillaIdInfo expr
573
574 makeTrivialWithInfo :: TopLevelFlag -> SimplEnv
575 -> FastString
576 -- ^ a "friendly name" to build the new binder from
577 -> IdInfo -> OutExpr -> SimplM (SimplEnv, OutExpr)
578 -- Propagate strictness and demand info to the new binder
579 -- Note [Preserve strictness when floating coercions]
580 -- Returned SimplEnv has same substitution as incoming one
581 makeTrivialWithInfo top_lvl env context info expr
582 | exprIsTrivial expr -- Already trivial
583 || not (bindingOk top_lvl expr expr_ty) -- Cannot trivialise
584 -- See Note [Cannot trivialise]
585 = return (env, expr)
586 | otherwise -- See Note [Take care] below
587 = do { uniq <- getUniqueM
588 ; let name = mkSystemVarName uniq context
589 var = mkLocalIdOrCoVarWithInfo name expr_ty info
590 ; env' <- completeNonRecX top_lvl env False var var expr
591 ; expr' <- simplVar env' var
592 ; return (env', expr') }
593 -- The simplVar is needed becase we're constructing a new binding
594 -- a = rhs
595 -- And if rhs is of form (rhs1 |> co), then we might get
596 -- a1 = rhs1
597 -- a = a1 |> co
598 -- and now a's RHS is trivial and can be substituted out, and that
599 -- is what completeNonRecX will do
600 -- To put it another way, it's as if we'd simplified
601 -- let var = e in var
602 where
603 expr_ty = exprType expr
604
605 bindingOk :: TopLevelFlag -> CoreExpr -> Type -> Bool
606 -- True iff we can have a binding of this expression at this level
607 -- Precondition: the type is the type of the expression
608 bindingOk top_lvl _ expr_ty
609 | isTopLevel top_lvl = not (isUnliftedType expr_ty)
610 | otherwise = True
611
612 {-
613 Note [Cannot trivialise]
614 ~~~~~~~~~~~~~~~~~~~~~~~~
615 Consider tih
616 f :: Int -> Addr#
617
618 foo :: Bar
619 foo = Bar (f 3)
620
621 Then we can't ANF-ise foo, even though we'd like to, because
622 we can't make a top-level binding for the Addr# (f 3). And if
623 so we don't want to turn it into
624 foo = let x = f 3 in Bar x
625 because we'll just end up inlining x back, and that makes the
626 simplifier loop. Better not to ANF-ise it at all.
627
628 A case in point is literal strings (a MachStr is not regarded as
629 trivial):
630
631 foo = Ptr "blob"#
632
633 We don't want to ANF-ise this.
634
635 ************************************************************************
636 * *
637 \subsection{Completing a lazy binding}
638 * *
639 ************************************************************************
640
641 completeBind
642 * deals only with Ids, not TyVars
643 * takes an already-simplified binder and RHS
644 * is used for both recursive and non-recursive bindings
645 * is used for both top-level and non-top-level bindings
646
647 It does the following:
648 - tries discarding a dead binding
649 - tries PostInlineUnconditionally
650 - add unfolding [this is the only place we add an unfolding]
651 - add arity
652
653 It does *not* attempt to do let-to-case. Why? Because it is used for
654 - top-level bindings (when let-to-case is impossible)
655 - many situations where the "rhs" is known to be a WHNF
656 (so let-to-case is inappropriate).
657
658 Nor does it do the atomic-argument thing
659 -}
660
661 completeBind :: SimplEnv
662 -> TopLevelFlag -- Flag stuck into unfolding
663 -> InId -- Old binder
664 -> OutId -> OutExpr -- New binder and RHS
665 -> SimplM SimplEnv
666 -- completeBind may choose to do its work
667 -- * by extending the substitution (e.g. let x = y in ...)
668 -- * or by adding to the floats in the envt
669 --
670 -- Precondition: rhs obeys the let/app invariant
671 completeBind env top_lvl old_bndr new_bndr new_rhs
672 | isCoVar old_bndr
673 = case new_rhs of
674 Coercion co -> return (extendCvSubst env old_bndr co)
675 _ -> return (addNonRec env new_bndr new_rhs)
676
677 | otherwise
678 = ASSERT( isId new_bndr )
679 do { let old_info = idInfo old_bndr
680 old_unf = unfoldingInfo old_info
681 occ_info = occInfo old_info
682
683 -- Do eta-expansion on the RHS of the binding
684 -- See Note [Eta-expanding at let bindings] in SimplUtils
685 ; (new_arity, final_rhs) <- tryEtaExpandRhs env new_bndr new_rhs
686
687 -- Simplify the unfolding
688 ; new_unfolding <- simplLetUnfolding env top_lvl old_bndr final_rhs old_unf
689
690 ; dflags <- getDynFlags
691 ; if postInlineUnconditionally dflags env top_lvl new_bndr occ_info
692 final_rhs new_unfolding
693
694 -- Inline and discard the binding
695 then do { tick (PostInlineUnconditionally old_bndr)
696 ; return (extendIdSubst env old_bndr (DoneEx final_rhs)) }
697 -- Use the substitution to make quite, quite sure that the
698 -- substitution will happen, since we are going to discard the binding
699 else
700 do { let info1 = idInfo new_bndr `setArityInfo` new_arity
701
702 -- Unfolding info: Note [Setting the new unfolding]
703 info2 = info1 `setUnfoldingInfo` new_unfolding
704
705 -- Demand info: Note [Setting the demand info]
706 --
707 -- We also have to nuke demand info if for some reason
708 -- eta-expansion *reduces* the arity of the binding to less
709 -- than that of the strictness sig. This can happen: see Note [Arity decrease].
710 info3 | isEvaldUnfolding new_unfolding
711 || (case strictnessInfo info2 of
712 StrictSig dmd_ty -> new_arity < dmdTypeDepth dmd_ty)
713 = zapDemandInfo info2 `orElse` info2
714 | otherwise
715 = info2
716
717 final_id = new_bndr `setIdInfo` info3
718
719 ; -- pprTrace "Binding" (ppr final_id <+> ppr new_unfolding) $
720 return (addNonRec env final_id final_rhs) } }
721 -- The addNonRec adds it to the in-scope set too
722
723 ------------------------------
724 addPolyBind :: TopLevelFlag -> SimplEnv -> OutBind -> SimplM SimplEnv
725 -- Add a new binding to the environment, complete with its unfolding
726 -- but *do not* do postInlineUnconditionally, because we have already
727 -- processed some of the scope of the binding
728 -- We still want the unfolding though. Consider
729 -- let
730 -- x = /\a. let y = ... in Just y
731 -- in body
732 -- Then we float the y-binding out (via abstractFloats and addPolyBind)
733 -- but 'x' may well then be inlined in 'body' in which case we'd like the
734 -- opportunity to inline 'y' too.
735 --
736 -- INVARIANT: the arity is correct on the incoming binders
737
738 addPolyBind top_lvl env (NonRec poly_id rhs)
739 = do { unfolding <- simplLetUnfolding env top_lvl poly_id rhs noUnfolding
740 -- Assumes that poly_id did not have an INLINE prag
741 -- which is perhaps wrong. ToDo: think about this
742 ; let final_id = setIdInfo poly_id $
743 idInfo poly_id `setUnfoldingInfo` unfolding
744
745 ; return (addNonRec env final_id rhs) }
746
747 addPolyBind _ env bind@(Rec _)
748 = return (extendFloats env bind)
749 -- Hack: letrecs are more awkward, so we extend "by steam"
750 -- without adding unfoldings etc. At worst this leads to
751 -- more simplifier iterations
752
753 {- Note [Arity decrease]
754 ~~~~~~~~~~~~~~~~~~~~~~~~
755 Generally speaking the arity of a binding should not decrease. But it *can*
756 legitimately happen because of RULES. Eg
757 f = g Int
758 where g has arity 2, will have arity 2. But if there's a rewrite rule
759 g Int --> h
760 where h has arity 1, then f's arity will decrease. Here's a real-life example,
761 which is in the output of Specialise:
762
763 Rec {
764 $dm {Arity 2} = \d.\x. op d
765 {-# RULES forall d. $dm Int d = $s$dm #-}
766
767 dInt = MkD .... opInt ...
768 opInt {Arity 1} = $dm dInt
769
770 $s$dm {Arity 0} = \x. op dInt }
771
772 Here opInt has arity 1; but when we apply the rule its arity drops to 0.
773 That's why Specialise goes to a little trouble to pin the right arity
774 on specialised functions too.
775
776 Note [Setting the demand info]
777 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
778 If the unfolding is a value, the demand info may
779 go pear-shaped, so we nuke it. Example:
780 let x = (a,b) in
781 case x of (p,q) -> h p q x
782 Here x is certainly demanded. But after we've nuked
783 the case, we'll get just
784 let x = (a,b) in h a b x
785 and now x is not demanded (I'm assuming h is lazy)
786 This really happens. Similarly
787 let f = \x -> e in ...f..f...
788 After inlining f at some of its call sites the original binding may
789 (for example) be no longer strictly demanded.
790 The solution here is a bit ad hoc...
791
792
793 ************************************************************************
794 * *
795 \subsection[Simplify-simplExpr]{The main function: simplExpr}
796 * *
797 ************************************************************************
798
799 The reason for this OutExprStuff stuff is that we want to float *after*
800 simplifying a RHS, not before. If we do so naively we get quadratic
801 behaviour as things float out.
802
803 To see why it's important to do it after, consider this (real) example:
804
805 let t = f x
806 in fst t
807 ==>
808 let t = let a = e1
809 b = e2
810 in (a,b)
811 in fst t
812 ==>
813 let a = e1
814 b = e2
815 t = (a,b)
816 in
817 a -- Can't inline a this round, cos it appears twice
818 ==>
819 e1
820
821 Each of the ==> steps is a round of simplification. We'd save a
822 whole round if we float first. This can cascade. Consider
823
824 let f = g d
825 in \x -> ...f...
826 ==>
827 let f = let d1 = ..d.. in \y -> e
828 in \x -> ...f...
829 ==>
830 let d1 = ..d..
831 in \x -> ...(\y ->e)...
832
833 Only in this second round can the \y be applied, and it
834 might do the same again.
835 -}
836
837 simplExpr :: SimplEnv -> CoreExpr -> SimplM CoreExpr
838 simplExpr env expr = simplExprC env expr (mkBoringStop expr_out_ty)
839 where
840 expr_out_ty :: OutType
841 expr_out_ty = substTy env (exprType expr)
842
843 simplExprC :: SimplEnv -> CoreExpr -> SimplCont -> SimplM CoreExpr
844 -- Simplify an expression, given a continuation
845 simplExprC env expr cont
846 = -- pprTrace "simplExprC" (ppr expr $$ ppr cont {- $$ ppr (seIdSubst env) -} $$ ppr (seFloats env) ) $
847 do { (env', expr') <- simplExprF (zapFloats env) expr cont
848 ; -- pprTrace "simplExprC ret" (ppr expr $$ ppr expr') $
849 -- pprTrace "simplExprC ret3" (ppr (seInScope env')) $
850 -- pprTrace "simplExprC ret4" (ppr (seFloats env')) $
851 return (wrapFloats env' expr') }
852
853 --------------------------------------------------
854 simplExprF :: SimplEnv -> InExpr -> SimplCont
855 -> SimplM (SimplEnv, OutExpr)
856
857 simplExprF env e cont
858 = {- pprTrace "simplExprF" (vcat
859 [ ppr e
860 , text "cont =" <+> ppr cont
861 , text "inscope =" <+> ppr (seInScope env)
862 , text "tvsubst =" <+> ppr (seTvSubst env)
863 , text "idsubst =" <+> ppr (seIdSubst env)
864 , text "cvsubst =" <+> ppr (seCvSubst env)
865 {- , ppr (seFloats env) -}
866 ]) $ -}
867 simplExprF1 env e cont
868
869 simplExprF1 :: SimplEnv -> InExpr -> SimplCont
870 -> SimplM (SimplEnv, OutExpr)
871 simplExprF1 env (Var v) cont = simplIdF env v cont
872 simplExprF1 env (Lit lit) cont = rebuild env (Lit lit) cont
873 simplExprF1 env (Tick t expr) cont = simplTick env t expr cont
874 simplExprF1 env (Cast body co) cont = simplCast env body co cont
875 simplExprF1 env (Coercion co) cont = simplCoercionF env co cont
876 simplExprF1 env (Type ty) cont = ASSERT( contIsRhsOrArg cont )
877 rebuild env (Type (substTy env ty)) cont
878
879 simplExprF1 env (App fun arg) cont
880 = simplExprF env fun $
881 case arg of
882 Type ty -> ApplyToTy { sc_arg_ty = substTy env ty
883 , sc_hole_ty = substTy env (exprType fun)
884 , sc_cont = cont }
885 _ -> ApplyToVal { sc_arg = arg, sc_env = env
886 , sc_dup = NoDup, sc_cont = cont }
887
888 simplExprF1 env expr@(Lam {}) cont
889 = simplLam env zapped_bndrs body cont
890 -- The main issue here is under-saturated lambdas
891 -- (\x1. \x2. e) arg1
892 -- Here x1 might have "occurs-once" occ-info, because occ-info
893 -- is computed assuming that a group of lambdas is applied
894 -- all at once. If there are too few args, we must zap the
895 -- occ-info, UNLESS the remaining binders are one-shot
896 where
897 (bndrs, body) = collectBinders expr
898 zapped_bndrs | need_to_zap = map zap bndrs
899 | otherwise = bndrs
900
901 need_to_zap = any zappable_bndr (drop n_args bndrs)
902 n_args = countArgs cont
903 -- NB: countArgs counts all the args (incl type args)
904 -- and likewise drop counts all binders (incl type lambdas)
905
906 zappable_bndr b = isId b && not (isOneShotBndr b)
907 zap b | isTyVar b = b
908 | otherwise = zapLamIdInfo b
909
910 simplExprF1 env (Case scrut bndr _ alts) cont
911 = simplExprF env scrut (Select { sc_dup = NoDup, sc_bndr = bndr
912 , sc_alts = alts
913 , sc_env = env, sc_cont = cont })
914
915 simplExprF1 env (Let (Rec pairs) body) cont
916 = do { env' <- simplRecBndrs env (map fst pairs)
917 -- NB: bndrs' don't have unfoldings or rules
918 -- We add them as we go down
919
920 ; env'' <- simplRecBind env' NotTopLevel pairs
921 ; simplExprF env'' body cont }
922
923 simplExprF1 env (Let (NonRec bndr rhs) body) cont
924 = simplNonRecE env bndr (rhs, env) ([], body) cont
925
926 ---------------------------------
927 simplType :: SimplEnv -> InType -> SimplM OutType
928 -- Kept monadic just so we can do the seqType
929 simplType env ty
930 = -- pprTrace "simplType" (ppr ty $$ ppr (seTvSubst env)) $
931 seqType new_ty `seq` return new_ty
932 where
933 new_ty = substTy env ty
934
935 ---------------------------------
936 simplCoercionF :: SimplEnv -> InCoercion -> SimplCont
937 -> SimplM (SimplEnv, OutExpr)
938 simplCoercionF env co cont
939 = do { co' <- simplCoercion env co
940 ; rebuild env (Coercion co') cont }
941
942 simplCoercion :: SimplEnv -> InCoercion -> SimplM OutCoercion
943 simplCoercion env co
944 = let opt_co = optCoercion (getTCvSubst env) co
945 in seqCo opt_co `seq` return opt_co
946
947 -----------------------------------
948 -- | Push a TickIt context outwards past applications and cases, as
949 -- long as this is a non-scoping tick, to let case and application
950 -- optimisations apply.
951
952 simplTick :: SimplEnv -> Tickish Id -> InExpr -> SimplCont
953 -> SimplM (SimplEnv, OutExpr)
954 simplTick env tickish expr cont
955 -- A scoped tick turns into a continuation, so that we can spot
956 -- (scc t (\x . e)) in simplLam and eliminate the scc. If we didn't do
957 -- it this way, then it would take two passes of the simplifier to
958 -- reduce ((scc t (\x . e)) e').
959 -- NB, don't do this with counting ticks, because if the expr is
960 -- bottom, then rebuildCall will discard the continuation.
961
962 -- XXX: we cannot do this, because the simplifier assumes that
963 -- the context can be pushed into a case with a single branch. e.g.
964 -- scc<f> case expensive of p -> e
965 -- becomes
966 -- case expensive of p -> scc<f> e
967 --
968 -- So I'm disabling this for now. It just means we will do more
969 -- simplifier iterations that necessary in some cases.
970
971 -- | tickishScoped tickish && not (tickishCounts tickish)
972 -- = simplExprF env expr (TickIt tickish cont)
973
974 -- For unscoped or soft-scoped ticks, we are allowed to float in new
975 -- cost, so we simply push the continuation inside the tick. This
976 -- has the effect of moving the tick to the outside of a case or
977 -- application context, allowing the normal case and application
978 -- optimisations to fire.
979 | tickish `tickishScopesLike` SoftScope
980 = do { (env', expr') <- simplExprF env expr cont
981 ; return (env', mkTick tickish expr')
982 }
983
984 -- Push tick inside if the context looks like this will allow us to
985 -- do a case-of-case - see Note [case-of-scc-of-case]
986 | Select {} <- cont, Just expr' <- push_tick_inside
987 = simplExprF env expr' cont
988
989 -- We don't want to move the tick, but we might still want to allow
990 -- floats to pass through with appropriate wrapping (or not, see
991 -- wrap_floats below)
992 --- | not (tickishCounts tickish) || tickishCanSplit tickish
993 -- = wrap_floats
994
995 | otherwise
996 = no_floating_past_tick
997
998 where
999
1000 -- Try to push tick inside a case, see Note [case-of-scc-of-case].
1001 push_tick_inside =
1002 case expr0 of
1003 Case scrut bndr ty alts
1004 -> Just $ Case (tickScrut scrut) bndr ty (map tickAlt alts)
1005 _other -> Nothing
1006 where (ticks, expr0) = stripTicksTop movable (Tick tickish expr)
1007 movable t = not (tickishCounts t) ||
1008 t `tickishScopesLike` NoScope ||
1009 tickishCanSplit t
1010 tickScrut e = foldr mkTick e ticks
1011 -- Alternatives get annotated with all ticks that scope in some way,
1012 -- but we don't want to count entries.
1013 tickAlt (c,bs,e) = (c,bs, foldr mkTick e ts_scope)
1014 ts_scope = map mkNoCount $
1015 filter (not . (`tickishScopesLike` NoScope)) ticks
1016
1017 no_floating_past_tick =
1018 do { let (inc,outc) = splitCont cont
1019 ; (env', expr') <- simplExprF (zapFloats env) expr inc
1020 ; let tickish' = simplTickish env tickish
1021 ; (env'', expr'') <- rebuild (zapFloats env')
1022 (wrapFloats env' expr')
1023 (TickIt tickish' outc)
1024 ; return (addFloats env env'', expr'')
1025 }
1026
1027 -- Alternative version that wraps outgoing floats with the tick. This
1028 -- results in ticks being duplicated, as we don't make any attempt to
1029 -- eliminate the tick if we re-inline the binding (because the tick
1030 -- semantics allows unrestricted inlining of HNFs), so I'm not doing
1031 -- this any more. FloatOut will catch any real opportunities for
1032 -- floating.
1033 --
1034 -- wrap_floats =
1035 -- do { let (inc,outc) = splitCont cont
1036 -- ; (env', expr') <- simplExprF (zapFloats env) expr inc
1037 -- ; let tickish' = simplTickish env tickish
1038 -- ; let wrap_float (b,rhs) = (zapIdStrictness (setIdArity b 0),
1039 -- mkTick (mkNoCount tickish') rhs)
1040 -- -- when wrapping a float with mkTick, we better zap the Id's
1041 -- -- strictness info and arity, because it might be wrong now.
1042 -- ; let env'' = addFloats env (mapFloats env' wrap_float)
1043 -- ; rebuild env'' expr' (TickIt tickish' outc)
1044 -- }
1045
1046
1047 simplTickish env tickish
1048 | Breakpoint n ids <- tickish
1049 = Breakpoint n (map (getDoneId . substId env) ids)
1050 | otherwise = tickish
1051
1052 -- Push type application and coercion inside a tick
1053 splitCont :: SimplCont -> (SimplCont, SimplCont)
1054 splitCont cont@(ApplyToTy { sc_cont = tail }) = (cont { sc_cont = inc }, outc)
1055 where (inc,outc) = splitCont tail
1056 splitCont (CastIt co c) = (CastIt co inc, outc)
1057 where (inc,outc) = splitCont c
1058 splitCont other = (mkBoringStop (contHoleType other), other)
1059
1060 getDoneId (DoneId id) = id
1061 getDoneId (DoneEx e) = getIdFromTrivialExpr e -- Note [substTickish] in CoreSubst
1062 getDoneId other = pprPanic "getDoneId" (ppr other)
1063
1064 -- Note [case-of-scc-of-case]
1065 -- It's pretty important to be able to transform case-of-case when
1066 -- there's an SCC in the way. For example, the following comes up
1067 -- in nofib/real/compress/Encode.hs:
1068 --
1069 -- case scctick<code_string.r1>
1070 -- case $wcode_string_r13s wild_XC w1_s137 w2_s138 l_aje
1071 -- of _ { (# ww1_s13f, ww2_s13g, ww3_s13h #) ->
1072 -- (ww1_s13f, ww2_s13g, ww3_s13h)
1073 -- }
1074 -- of _ { (ww_s12Y, ww1_s12Z, ww2_s130) ->
1075 -- tick<code_string.f1>
1076 -- (ww_s12Y,
1077 -- ww1_s12Z,
1078 -- PTTrees.PT
1079 -- @ GHC.Types.Char @ GHC.Types.Int wild2_Xj ww2_s130 r_ajf)
1080 -- }
1081 --
1082 -- We really want this case-of-case to fire, because then the 3-tuple
1083 -- will go away (indeed, the CPR optimisation is relying on this
1084 -- happening). But the scctick is in the way - we need to push it
1085 -- inside to expose the case-of-case. So we perform this
1086 -- transformation on the inner case:
1087 --
1088 -- scctick c (case e of { p1 -> e1; ...; pn -> en })
1089 -- ==>
1090 -- case (scctick c e) of { p1 -> scc c e1; ...; pn -> scc c en }
1091 --
1092 -- So we've moved a constant amount of work out of the scc to expose
1093 -- the case. We only do this when the continuation is interesting: in
1094 -- for now, it has to be another Case (maybe generalise this later).
1095
1096 {-
1097 ************************************************************************
1098 * *
1099 \subsection{The main rebuilder}
1100 * *
1101 ************************************************************************
1102 -}
1103
1104 rebuild :: SimplEnv -> OutExpr -> SimplCont -> SimplM (SimplEnv, OutExpr)
1105 -- At this point the substitution in the SimplEnv should be irrelevant
1106 -- only the in-scope set and floats should matter
1107 rebuild env expr cont
1108 = case cont of
1109 Stop {} -> return (env, expr)
1110 TickIt t cont -> rebuild env (mkTick t expr) cont
1111 CastIt co cont -> rebuild env (mkCast expr co) cont
1112 -- NB: mkCast implements the (Coercion co |> g) optimisation
1113
1114 Select { sc_bndr = bndr, sc_alts = alts, sc_env = se, sc_cont = cont }
1115 -> rebuildCase (se `setFloats` env) expr bndr alts cont
1116
1117 StrictArg info _ cont -> rebuildCall env (info `addValArgTo` expr) cont
1118 StrictBind b bs body se cont -> do { env' <- simplNonRecX (se `setFloats` env) b expr
1119 -- expr satisfies let/app since it started life
1120 -- in a call to simplNonRecE
1121 ; simplLam env' bs body cont }
1122
1123 ApplyToTy { sc_arg_ty = ty, sc_cont = cont}
1124 -> rebuild env (App expr (Type ty)) cont
1125 ApplyToVal { sc_arg = arg, sc_env = se, sc_dup = dup_flag, sc_cont = cont}
1126 -- See Note [Avoid redundant simplification]
1127 | isSimplified dup_flag -> rebuild env (App expr arg) cont
1128 | otherwise -> do { arg' <- simplExpr (se `setInScope` env) arg
1129 ; rebuild env (App expr arg') cont }
1130
1131
1132 {-
1133 ************************************************************************
1134 * *
1135 \subsection{Lambdas}
1136 * *
1137 ************************************************************************
1138 -}
1139
1140 simplCast :: SimplEnv -> InExpr -> Coercion -> SimplCont
1141 -> SimplM (SimplEnv, OutExpr)
1142 simplCast env body co0 cont0
1143 = do { co1 <- simplCoercion env co0
1144 ; cont1 <- addCoerce co1 cont0
1145 ; simplExprF env body cont1 }
1146 where
1147 addCoerce co cont = add_coerce co (coercionKind co) cont
1148
1149 add_coerce _co (Pair s1 k1) cont -- co :: ty~ty
1150 | s1 `eqType` k1 = return cont -- is a no-op
1151
1152 add_coerce co1 (Pair s1 _k2) (CastIt co2 cont)
1153 | (Pair _l1 t1) <- coercionKind co2
1154 -- e |> (g1 :: S1~L) |> (g2 :: L~T1)
1155 -- ==>
1156 -- e, if S1=T1
1157 -- e |> (g1 . g2 :: S1~T1) otherwise
1158 --
1159 -- For example, in the initial form of a worker
1160 -- we may find (coerce T (coerce S (\x.e))) y
1161 -- and we'd like it to simplify to e[y/x] in one round
1162 -- of simplification
1163 , s1 `eqType` t1 = return cont -- The coerces cancel out
1164 | otherwise = return (CastIt (mkTransCo co1 co2) cont)
1165
1166 add_coerce co (Pair s1s2 _t1t2) cont@(ApplyToTy { sc_arg_ty = arg_ty, sc_cont = tail })
1167 -- (f |> g) ty ---> (f ty) |> (g @ ty)
1168 -- This implements the PushT rule from the paper
1169 | isForAllTy s1s2
1170 = do { cont' <- addCoerce new_cast tail
1171 ; return (cont { sc_cont = cont' }) }
1172 where
1173 new_cast = mkInstCo co (mkNomReflCo arg_ty)
1174
1175 add_coerce co (Pair s1s2 t1t2) (ApplyToVal { sc_arg = arg, sc_env = arg_se
1176 , sc_dup = dup, sc_cont = cont })
1177 | isFunTy s1s2 -- This implements the Push rule from the paper
1178 , isFunTy t1t2 -- Check t1t2 to ensure 'arg' is a value arg
1179 -- (e |> (g :: s1s2 ~ t1->t2)) f
1180 -- ===>
1181 -- (e (f |> (arg g :: t1~s1))
1182 -- |> (res g :: s2->t2)
1183 --
1184 -- t1t2 must be a function type, t1->t2, because it's applied
1185 -- to something but s1s2 might conceivably not be
1186 --
1187 -- When we build the ApplyTo we can't mix the out-types
1188 -- with the InExpr in the argument, so we simply substitute
1189 -- to make it all consistent. It's a bit messy.
1190 -- But it isn't a common case.
1191 --
1192 -- Example of use: Trac #995
1193 = do { (dup', arg_se', arg') <- simplArg env dup arg_se arg
1194 ; cont' <- addCoerce co2 cont
1195 ; return (ApplyToVal { sc_arg = mkCast arg' (mkSymCo co1)
1196 , sc_env = arg_se'
1197 , sc_dup = dup'
1198 , sc_cont = cont' }) }
1199 where
1200 -- we split coercion t1->t2 ~ s1->s2 into t1 ~ s1 and
1201 -- t2 ~ s2 with left and right on the curried form:
1202 -- (->) t1 t2 ~ (->) s1 s2
1203 [co1, co2] = decomposeCo 2 co
1204
1205 add_coerce co _ cont = return (CastIt co cont)
1206
1207 simplArg :: SimplEnv -> DupFlag -> StaticEnv -> CoreExpr
1208 -> SimplM (DupFlag, StaticEnv, OutExpr)
1209 simplArg env dup_flag arg_env arg
1210 | isSimplified dup_flag
1211 = return (dup_flag, arg_env, arg)
1212 | otherwise
1213 = do { arg' <- simplExpr (arg_env `setInScope` env) arg
1214 ; return (Simplified, zapSubstEnv arg_env, arg') }
1215
1216 {-
1217 ************************************************************************
1218 * *
1219 \subsection{Lambdas}
1220 * *
1221 ************************************************************************
1222
1223 Note [Zap unfolding when beta-reducing]
1224 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1225 Lambda-bound variables can have stable unfoldings, such as
1226 $j = \x. \b{Unf=Just x}. e
1227 See Note [Case binders and join points] below; the unfolding for lets
1228 us optimise e better. However when we beta-reduce it we want to
1229 revert to using the actual value, otherwise we can end up in the
1230 stupid situation of
1231 let x = blah in
1232 let b{Unf=Just x} = y
1233 in ...b...
1234 Here it'd be far better to drop the unfolding and use the actual RHS.
1235 -}
1236
1237 simplLam :: SimplEnv -> [InId] -> InExpr -> SimplCont
1238 -> SimplM (SimplEnv, OutExpr)
1239
1240 simplLam env [] body cont = simplExprF env body cont
1241
1242 -- Beta reduction
1243
1244 simplLam env (bndr:bndrs) body (ApplyToTy { sc_arg_ty = arg_ty, sc_cont = cont })
1245 = do { tick (BetaReduction bndr)
1246 ; simplLam (extendTvSubst env bndr arg_ty) bndrs body cont }
1247
1248 simplLam env (bndr:bndrs) body (ApplyToVal { sc_arg = arg, sc_env = arg_se
1249 , sc_cont = cont })
1250 = do { tick (BetaReduction bndr)
1251 ; simplNonRecE env' (zap_unfolding bndr) (arg, arg_se) (bndrs, body) cont }
1252 where
1253 env' | Coercion co <- arg
1254 = extendCvSubst env bndr co
1255 | otherwise
1256 = env
1257
1258 zap_unfolding bndr -- See Note [Zap unfolding when beta-reducing]
1259 | isId bndr, isStableUnfolding (realIdUnfolding bndr)
1260 = setIdUnfolding bndr NoUnfolding
1261 | otherwise = bndr
1262
1263 -- discard a non-counting tick on a lambda. This may change the
1264 -- cost attribution slightly (moving the allocation of the
1265 -- lambda elsewhere), but we don't care: optimisation changes
1266 -- cost attribution all the time.
1267 simplLam env bndrs body (TickIt tickish cont)
1268 | not (tickishCounts tickish)
1269 = simplLam env bndrs body cont
1270
1271 -- Not enough args, so there are real lambdas left to put in the result
1272 simplLam env bndrs body cont
1273 = do { (env', bndrs') <- simplLamBndrs env bndrs
1274 ; body' <- simplExpr env' body
1275 ; new_lam <- mkLam bndrs' body' cont
1276 ; rebuild env' new_lam cont }
1277
1278 simplLamBndrs :: SimplEnv -> [InBndr] -> SimplM (SimplEnv, [OutBndr])
1279 simplLamBndrs env bndrs = mapAccumLM simplLamBndr env bndrs
1280
1281 -------------
1282 simplLamBndr :: SimplEnv -> Var -> SimplM (SimplEnv, Var)
1283 -- Used for lambda binders. These sometimes have unfoldings added by
1284 -- the worker/wrapper pass that must be preserved, because they can't
1285 -- be reconstructed from context. For example:
1286 -- f x = case x of (a,b) -> fw a b x
1287 -- fw a b x{=(a,b)} = ...
1288 -- The "{=(a,b)}" is an unfolding we can't reconstruct otherwise.
1289 simplLamBndr env bndr
1290 | isId bndr && hasSomeUnfolding old_unf -- Special case
1291 = do { (env1, bndr1) <- simplBinder env bndr
1292 ; unf' <- simplUnfolding env1 NotTopLevel bndr old_unf
1293 ; let bndr2 = bndr1 `setIdUnfolding` unf'
1294 ; return (modifyInScope env1 bndr2, bndr2) }
1295
1296 | otherwise
1297 = simplBinder env bndr -- Normal case
1298 where
1299 old_unf = idUnfolding bndr
1300
1301 ------------------
1302 simplNonRecE :: SimplEnv
1303 -> InBndr -- The binder
1304 -> (InExpr, SimplEnv) -- Rhs of binding (or arg of lambda)
1305 -> ([InBndr], InExpr) -- Body of the let/lambda
1306 -- \xs.e
1307 -> SimplCont
1308 -> SimplM (SimplEnv, OutExpr)
1309
1310 -- simplNonRecE is used for
1311 -- * non-top-level non-recursive lets in expressions
1312 -- * beta reduction
1313 --
1314 -- It deals with strict bindings, via the StrictBind continuation,
1315 -- which may abort the whole process
1316 --
1317 -- Precondition: rhs satisfies the let/app invariant
1318 -- Note [CoreSyn let/app invariant] in CoreSyn
1319 --
1320 -- The "body" of the binding comes as a pair of ([InId],InExpr)
1321 -- representing a lambda; so we recurse back to simplLam
1322 -- Why? Because of the binder-occ-info-zapping done before
1323 -- the call to simplLam in simplExprF (Lam ...)
1324
1325 -- First deal with type applications and type lets
1326 -- (/\a. e) (Type ty) and (let a = Type ty in e)
1327 simplNonRecE env bndr (Type ty_arg, rhs_se) (bndrs, body) cont
1328 = ASSERT( isTyVar bndr )
1329 do { ty_arg' <- simplType (rhs_se `setInScope` env) ty_arg
1330 ; simplLam (extendTvSubst env bndr ty_arg') bndrs body cont }
1331
1332 simplNonRecE env bndr (rhs, rhs_se) (bndrs, body) cont
1333 = do dflags <- getDynFlags
1334 case () of
1335 _ | preInlineUnconditionally dflags env NotTopLevel bndr rhs
1336 -> do { tick (PreInlineUnconditionally bndr)
1337 ; -- pprTrace "preInlineUncond" (ppr bndr <+> ppr rhs) $
1338 simplLam (extendIdSubst env bndr (mkContEx rhs_se rhs)) bndrs body cont }
1339
1340 | isStrictId bndr -- Includes coercions
1341 -> simplExprF (rhs_se `setFloats` env) rhs
1342 (StrictBind bndr bndrs body env cont)
1343
1344 | otherwise
1345 -> ASSERT( not (isTyVar bndr) )
1346 do { (env1, bndr1) <- simplNonRecBndr env bndr
1347 ; (env2, bndr2) <- addBndrRules env1 bndr bndr1
1348 ; env3 <- simplLazyBind env2 NotTopLevel NonRecursive bndr bndr2 rhs rhs_se
1349 ; simplLam env3 bndrs body cont }
1350
1351 {-
1352 ************************************************************************
1353 * *
1354 Variables
1355 * *
1356 ************************************************************************
1357 -}
1358
1359 simplVar :: SimplEnv -> InVar -> SimplM OutExpr
1360 -- Look up an InVar in the environment
1361 simplVar env var
1362 | isTyVar var = return (Type (substTyVar env var))
1363 | isCoVar var = return (Coercion (substCoVar env var))
1364 | otherwise
1365 = case substId env var of
1366 DoneId var1 -> return (Var var1)
1367 DoneEx e -> return e
1368 ContEx tvs cvs ids e -> simplExpr (setSubstEnv env tvs cvs ids) e
1369
1370 simplIdF :: SimplEnv -> InId -> SimplCont -> SimplM (SimplEnv, OutExpr)
1371 simplIdF env var cont
1372 = case substId env var of
1373 DoneEx e -> simplExprF (zapSubstEnv env) e cont
1374 ContEx tvs cvs ids e -> simplExprF (setSubstEnv env tvs cvs ids) e cont
1375 DoneId var1 -> completeCall env var1 cont
1376 -- Note [zapSubstEnv]
1377 -- The template is already simplified, so don't re-substitute.
1378 -- This is VITAL. Consider
1379 -- let x = e in
1380 -- let y = \z -> ...x... in
1381 -- \ x -> ...y...
1382 -- We'll clone the inner \x, adding x->x' in the id_subst
1383 -- Then when we inline y, we must *not* replace x by x' in
1384 -- the inlined copy!!
1385
1386 ---------------------------------------------------------
1387 -- Dealing with a call site
1388
1389 completeCall :: SimplEnv -> OutId -> SimplCont -> SimplM (SimplEnv, OutExpr)
1390 completeCall env var cont
1391 = do { ------------- Try inlining ----------------
1392 dflags <- getDynFlags
1393 ; let (lone_variable, arg_infos, call_cont) = contArgs cont
1394 n_val_args = length arg_infos
1395 interesting_cont = interestingCallContext call_cont
1396 unfolding = activeUnfolding env var
1397 maybe_inline = callSiteInline dflags var unfolding
1398 lone_variable arg_infos interesting_cont
1399 ; case maybe_inline of {
1400 Just expr -- There is an inlining!
1401 -> do { checkedTick (UnfoldingDone var)
1402 ; dump_inline dflags expr cont
1403 ; simplExprF (zapSubstEnv env) expr cont }
1404
1405 ; Nothing -> do -- No inlining!
1406
1407 { rule_base <- getSimplRules
1408 ; let info = mkArgInfo var (getRules rule_base var) n_val_args call_cont
1409 ; rebuildCall env info cont
1410 }}}
1411 where
1412 dump_inline dflags unfolding cont
1413 | not (dopt Opt_D_dump_inlinings dflags) = return ()
1414 | not (dopt Opt_D_verbose_core2core dflags)
1415 = when (isExternalName (idName var)) $
1416 liftIO $ printOutputForUser dflags alwaysQualify $
1417 sep [text "Inlining done:", nest 4 (ppr var)]
1418 | otherwise
1419 = liftIO $ printOutputForUser dflags alwaysQualify $
1420 sep [text "Inlining done: " <> ppr var,
1421 nest 4 (vcat [text "Inlined fn: " <+> nest 2 (ppr unfolding),
1422 text "Cont: " <+> ppr cont])]
1423
1424 rebuildCall :: SimplEnv
1425 -> ArgInfo
1426 -> SimplCont
1427 -> SimplM (SimplEnv, OutExpr)
1428 rebuildCall env (ArgInfo { ai_fun = fun, ai_args = rev_args, ai_strs = [] }) cont
1429 -- When we run out of strictness args, it means
1430 -- that the call is definitely bottom; see SimplUtils.mkArgInfo
1431 -- Then we want to discard the entire strict continuation. E.g.
1432 -- * case (error "hello") of { ... }
1433 -- * (error "Hello") arg
1434 -- * f (error "Hello") where f is strict
1435 -- etc
1436 -- Then, especially in the first of these cases, we'd like to discard
1437 -- the continuation, leaving just the bottoming expression. But the
1438 -- type might not be right, so we may have to add a coerce.
1439 | not (contIsTrivial cont) -- Only do this if there is a non-trivial
1440 = return (env, castBottomExpr res cont_ty) -- contination to discard, else we do it
1441 where -- again and again!
1442 res = argInfoExpr fun rev_args
1443 cont_ty = contResultType cont
1444
1445 rebuildCall env info (CastIt co cont)
1446 = rebuildCall env (addCastTo info co) cont
1447
1448 rebuildCall env info (ApplyToTy { sc_arg_ty = arg_ty, sc_cont = cont })
1449 = rebuildCall env (info `addTyArgTo` arg_ty) cont
1450
1451 rebuildCall env info@(ArgInfo { ai_encl = encl_rules, ai_type = fun_ty
1452 , ai_strs = str:strs, ai_discs = disc:discs })
1453 (ApplyToVal { sc_arg = arg, sc_env = arg_se
1454 , sc_dup = dup_flag, sc_cont = cont })
1455 | isSimplified dup_flag -- See Note [Avoid redundant simplification]
1456 = rebuildCall env (addValArgTo info' arg) cont
1457
1458 | str -- Strict argument
1459 = -- pprTrace "Strict Arg" (ppr arg $$ ppr (seIdSubst env) $$ ppr (seInScope env)) $
1460 simplExprF (arg_se `setFloats` env) arg
1461 (StrictArg info' cci cont)
1462 -- Note [Shadowing]
1463
1464 | otherwise -- Lazy argument
1465 -- DO NOT float anything outside, hence simplExprC
1466 -- There is no benefit (unlike in a let-binding), and we'd
1467 -- have to be very careful about bogus strictness through
1468 -- floating a demanded let.
1469 = do { arg' <- simplExprC (arg_se `setInScope` env) arg
1470 (mkLazyArgStop (funArgTy fun_ty) cci)
1471 ; rebuildCall env (addValArgTo info' arg') cont }
1472 where
1473 info' = info { ai_strs = strs, ai_discs = discs }
1474 cci | encl_rules = RuleArgCtxt
1475 | disc > 0 = DiscArgCtxt -- Be keener here
1476 | otherwise = BoringCtxt -- Nothing interesting
1477
1478 rebuildCall env (ArgInfo { ai_fun = fun, ai_args = rev_args, ai_rules = rules }) cont
1479 | null rules
1480 = rebuild env (argInfoExpr fun rev_args) cont -- No rules, common case
1481
1482 | otherwise
1483 = do { -- We've accumulated a simplified call in <fun,rev_args>
1484 -- so try rewrite rules; see Note [RULEs apply to simplified arguments]
1485 -- See also Note [Rules for recursive functions]
1486 ; let env' = zapSubstEnv env -- See Note [zapSubstEnv];
1487 -- and NB that 'rev_args' are all fully simplified
1488 ; mb_rule <- tryRules env' rules fun (reverse rev_args) cont
1489 ; case mb_rule of {
1490 Just (rule_rhs, cont') -> simplExprF env' rule_rhs cont'
1491
1492 -- Rules don't match
1493 ; Nothing -> rebuild env (argInfoExpr fun rev_args) cont -- No rules
1494 } }
1495
1496 {-
1497 Note [RULES apply to simplified arguments]
1498 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1499 It's very desirable to try RULES once the arguments have been simplified, because
1500 doing so ensures that rule cascades work in one pass. Consider
1501 {-# RULES g (h x) = k x
1502 f (k x) = x #-}
1503 ...f (g (h x))...
1504 Then we want to rewrite (g (h x)) to (k x) and only then try f's rules. If
1505 we match f's rules against the un-simplified RHS, it won't match. This
1506 makes a particularly big difference when superclass selectors are involved:
1507 op ($p1 ($p2 (df d)))
1508 We want all this to unravel in one sweep.
1509
1510 Note [Avoid redundant simplification]
1511 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1512 Because RULES apply to simplified arguments, there's a danger of repeatedly
1513 simplifying already-simplified arguments. An important example is that of
1514 (>>=) d e1 e2
1515 Here e1, e2 are simplified before the rule is applied, but don't really
1516 participate in the rule firing. So we mark them as Simplified to avoid
1517 re-simplifying them.
1518
1519 Note [Shadowing]
1520 ~~~~~~~~~~~~~~~~
1521 This part of the simplifier may break the no-shadowing invariant
1522 Consider
1523 f (...(\a -> e)...) (case y of (a,b) -> e')
1524 where f is strict in its second arg
1525 If we simplify the innermost one first we get (...(\a -> e)...)
1526 Simplifying the second arg makes us float the case out, so we end up with
1527 case y of (a,b) -> f (...(\a -> e)...) e'
1528 So the output does not have the no-shadowing invariant. However, there is
1529 no danger of getting name-capture, because when the first arg was simplified
1530 we used an in-scope set that at least mentioned all the variables free in its
1531 static environment, and that is enough.
1532
1533 We can't just do innermost first, or we'd end up with a dual problem:
1534 case x of (a,b) -> f e (...(\a -> e')...)
1535
1536 I spent hours trying to recover the no-shadowing invariant, but I just could
1537 not think of an elegant way to do it. The simplifier is already knee-deep in
1538 continuations. We have to keep the right in-scope set around; AND we have
1539 to get the effect that finding (error "foo") in a strict arg position will
1540 discard the entire application and replace it with (error "foo"). Getting
1541 all this at once is TOO HARD!
1542
1543
1544 ************************************************************************
1545 * *
1546 Rewrite rules
1547 * *
1548 ************************************************************************
1549 -}
1550
1551 tryRules :: SimplEnv -> [CoreRule]
1552 -> Id -> [ArgSpec] -> SimplCont
1553 -> SimplM (Maybe (CoreExpr, SimplCont))
1554 -- The SimplEnv already has zapSubstEnv applied to it
1555
1556 tryRules env rules fn args call_cont
1557 | null rules
1558 = return Nothing
1559 {- Disabled until we fix #8326
1560 | fn `hasKey` tagToEnumKey -- See Note [Optimising tagToEnum#]
1561 , [_type_arg, val_arg] <- args
1562 , Select dup bndr ((_,[],rhs1) : rest_alts) se cont <- call_cont
1563 , isDeadBinder bndr
1564 = do { dflags <- getDynFlags
1565 ; let enum_to_tag :: CoreAlt -> CoreAlt
1566 -- Takes K -> e into tagK# -> e
1567 -- where tagK# is the tag of constructor K
1568 enum_to_tag (DataAlt con, [], rhs)
1569 = ASSERT( isEnumerationTyCon (dataConTyCon con) )
1570 (LitAlt tag, [], rhs)
1571 where
1572 tag = mkMachInt dflags (toInteger (dataConTag con - fIRST_TAG))
1573 enum_to_tag alt = pprPanic "tryRules: tagToEnum" (ppr alt)
1574
1575 new_alts = (DEFAULT, [], rhs1) : map enum_to_tag rest_alts
1576 new_bndr = setIdType bndr intPrimTy
1577 -- The binder is dead, but should have the right type
1578 ; return (Just (val_arg, Select dup new_bndr new_alts se cont)) }
1579 -}
1580 | otherwise
1581 = do { dflags <- getDynFlags
1582 ; case lookupRule dflags (getUnfoldingInRuleMatch env) (activeRule env)
1583 fn (argInfoAppArgs args) rules of {
1584 Nothing ->
1585 do { nodump dflags -- This ensures that an empty file is written
1586 ; return Nothing } ; -- No rule matches
1587 Just (rule, rule_rhs) ->
1588 do { checkedTick (RuleFired (ru_name rule))
1589 ; let cont' = pushSimplifiedArgs env
1590 (drop (ruleArity rule) args)
1591 call_cont
1592 -- (ruleArity rule) says how many args the rule consumed
1593 ; dump dflags rule rule_rhs
1594 ; return (Just (rule_rhs, cont')) }}}
1595 where
1596 dump dflags rule rule_rhs
1597 | dopt Opt_D_dump_rule_rewrites dflags
1598 = log_rule dflags Opt_D_dump_rule_rewrites "Rule fired" $ vcat
1599 [ text "Rule:" <+> ftext (ru_name rule)
1600 , text "Before:" <+> hang (ppr fn) 2 (sep (map ppr args))
1601 , text "After: " <+> pprCoreExpr rule_rhs
1602 , text "Cont: " <+> ppr call_cont ]
1603
1604 | dopt Opt_D_dump_rule_firings dflags
1605 = log_rule dflags Opt_D_dump_rule_firings "Rule fired:" $
1606 ftext (ru_name rule)
1607
1608 | otherwise
1609 = return ()
1610
1611 nodump dflags
1612 | dopt Opt_D_dump_rule_rewrites dflags
1613 = liftIO $ dumpSDoc dflags alwaysQualify Opt_D_dump_rule_rewrites "" empty
1614
1615 | dopt Opt_D_dump_rule_firings dflags
1616 = liftIO $ dumpSDoc dflags alwaysQualify Opt_D_dump_rule_firings "" empty
1617
1618 | otherwise
1619 = return ()
1620
1621 log_rule dflags flag hdr details
1622 = liftIO . dumpSDoc dflags alwaysQualify flag "" $
1623 sep [text hdr, nest 4 details]
1624
1625 {-
1626 Note [Optimising tagToEnum#]
1627 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1628 If we have an enumeration data type:
1629
1630 data Foo = A | B | C
1631
1632 Then we want to transform
1633
1634 case tagToEnum# x of ==> case x of
1635 A -> e1 DEFAULT -> e1
1636 B -> e2 1# -> e2
1637 C -> e3 2# -> e3
1638
1639 thereby getting rid of the tagToEnum# altogether. If there was a DEFAULT
1640 alternative we retain it (remember it comes first). If not the case must
1641 be exhaustive, and we reflect that in the transformed version by adding
1642 a DEFAULT. Otherwise Lint complains that the new case is not exhaustive.
1643 See #8317.
1644
1645 Note [Rules for recursive functions]
1646 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1647 You might think that we shouldn't apply rules for a loop breaker:
1648 doing so might give rise to an infinite loop, because a RULE is
1649 rather like an extra equation for the function:
1650 RULE: f (g x) y = x+y
1651 Eqn: f a y = a-y
1652
1653 But it's too drastic to disable rules for loop breakers.
1654 Even the foldr/build rule would be disabled, because foldr
1655 is recursive, and hence a loop breaker:
1656 foldr k z (build g) = g k z
1657 So it's up to the programmer: rules can cause divergence
1658
1659
1660 ************************************************************************
1661 * *
1662 Rebuilding a case expression
1663 * *
1664 ************************************************************************
1665
1666 Note [Case elimination]
1667 ~~~~~~~~~~~~~~~~~~~~~~~
1668 The case-elimination transformation discards redundant case expressions.
1669 Start with a simple situation:
1670
1671 case x# of ===> let y# = x# in e
1672 y# -> e
1673
1674 (when x#, y# are of primitive type, of course). We can't (in general)
1675 do this for algebraic cases, because we might turn bottom into
1676 non-bottom!
1677
1678 The code in SimplUtils.prepareAlts has the effect of generalise this
1679 idea to look for a case where we're scrutinising a variable, and we
1680 know that only the default case can match. For example:
1681
1682 case x of
1683 0# -> ...
1684 DEFAULT -> ...(case x of
1685 0# -> ...
1686 DEFAULT -> ...) ...
1687
1688 Here the inner case is first trimmed to have only one alternative, the
1689 DEFAULT, after which it's an instance of the previous case. This
1690 really only shows up in eliminating error-checking code.
1691
1692 Note that SimplUtils.mkCase combines identical RHSs. So
1693
1694 case e of ===> case e of DEFAULT -> r
1695 True -> r
1696 False -> r
1697
1698 Now again the case may be elminated by the CaseElim transformation.
1699 This includes things like (==# a# b#)::Bool so that we simplify
1700 case ==# a# b# of { True -> x; False -> x }
1701 to just
1702 x
1703 This particular example shows up in default methods for
1704 comparison operations (e.g. in (>=) for Int.Int32)
1705
1706 Note [Case elimination: lifted case]
1707 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1708 If a case over a lifted type has a single alternative, and is being used
1709 as a strict 'let' (all isDeadBinder bndrs), we may want to do this
1710 transformation:
1711
1712 case e of r ===> let r = e in ...r...
1713 _ -> ...r...
1714
1715 (a) 'e' is already evaluated (it may so if e is a variable)
1716 Specifically we check (exprIsHNF e). In this case
1717 we can just allocate the WHNF directly with a let.
1718 or
1719 (b) 'x' is not used at all and e is ok-for-speculation
1720 The ok-for-spec bit checks that we don't lose any
1721 exceptions or divergence.
1722
1723 NB: it'd be *sound* to switch from case to let if the
1724 scrutinee was not yet WHNF but was guaranteed to
1725 converge; but sticking with case means we won't build a
1726 thunk
1727
1728 or
1729 (c) 'x' is used strictly in the body, and 'e' is a variable
1730 Then we can just substitute 'e' for 'x' in the body.
1731 See Note [Eliminating redundant seqs]
1732
1733 For (b), the "not used at all" test is important. Consider
1734 case (case a ># b of { True -> (p,q); False -> (q,p) }) of
1735 r -> blah
1736 The scrutinee is ok-for-speculation (it looks inside cases), but we do
1737 not want to transform to
1738 let r = case a ># b of { True -> (p,q); False -> (q,p) }
1739 in blah
1740 because that builds an unnecessary thunk.
1741
1742 Note [Eliminating redundant seqs]
1743 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1744 If we have this:
1745 case x of r { _ -> ..r.. }
1746 where 'r' is used strictly in (..r..), the case is effectively a 'seq'
1747 on 'x', but since 'r' is used strictly anyway, we can safely transform to
1748 (...x...)
1749
1750 Note that this can change the error behaviour. For example, we might
1751 transform
1752 case x of { _ -> error "bad" }
1753 --> error "bad"
1754 which is might be puzzling if 'x' currently lambda-bound, but later gets
1755 let-bound to (error "good").
1756
1757 Nevertheless, the paper "A semantics for imprecise exceptions" allows
1758 this transformation. If you want to fix the evaluation order, use
1759 'pseq'. See Trac #8900 for an example where the loss of this
1760 transformation bit us in practice.
1761
1762 See also Note [Empty case alternatives] in CoreSyn.
1763
1764 Just for reference, the original code (added Jan 13) looked like this:
1765 || case_bndr_evald_next rhs
1766
1767 case_bndr_evald_next :: CoreExpr -> Bool
1768 -- See Note [Case binder next]
1769 case_bndr_evald_next (Var v) = v == case_bndr
1770 case_bndr_evald_next (Cast e _) = case_bndr_evald_next e
1771 case_bndr_evald_next (App e _) = case_bndr_evald_next e
1772 case_bndr_evald_next (Case e _ _ _) = case_bndr_evald_next e
1773 case_bndr_evald_next _ = False
1774
1775 (This came up when fixing Trac #7542. See also Note [Eta reduction of
1776 an eval'd function] in CoreUtils.)
1777
1778
1779 Note [Case elimination: unlifted case]
1780 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1781 Consider
1782 case a +# b of r -> ...r...
1783 Then we do case-elimination (to make a let) followed by inlining,
1784 to get
1785 .....(a +# b)....
1786 If we have
1787 case indexArray# a i of r -> ...r...
1788 we might like to do the same, and inline the (indexArray# a i).
1789 But indexArray# is not okForSpeculation, so we don't build a let
1790 in rebuildCase (lest it get floated *out*), so the inlining doesn't
1791 happen either.
1792
1793 This really isn't a big deal I think. The let can be
1794
1795
1796 Further notes about case elimination
1797 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1798 Consider: test :: Integer -> IO ()
1799 test = print
1800
1801 Turns out that this compiles to:
1802 Print.test
1803 = \ eta :: Integer
1804 eta1 :: Void# ->
1805 case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT ->
1806 case hPutStr stdout
1807 (PrelNum.jtos eta ($w[] @ Char))
1808 eta1
1809 of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }}
1810
1811 Notice the strange '<' which has no effect at all. This is a funny one.
1812 It started like this:
1813
1814 f x y = if x < 0 then jtos x
1815 else if y==0 then "" else jtos x
1816
1817 At a particular call site we have (f v 1). So we inline to get
1818
1819 if v < 0 then jtos x
1820 else if 1==0 then "" else jtos x
1821
1822 Now simplify the 1==0 conditional:
1823
1824 if v<0 then jtos v else jtos v
1825
1826 Now common-up the two branches of the case:
1827
1828 case (v<0) of DEFAULT -> jtos v
1829
1830 Why don't we drop the case? Because it's strict in v. It's technically
1831 wrong to drop even unnecessary evaluations, and in practice they
1832 may be a result of 'seq' so we *definitely* don't want to drop those.
1833 I don't really know how to improve this situation.
1834 -}
1835
1836 ---------------------------------------------------------
1837 -- Eliminate the case if possible
1838
1839 rebuildCase, reallyRebuildCase
1840 :: SimplEnv
1841 -> OutExpr -- Scrutinee
1842 -> InId -- Case binder
1843 -> [InAlt] -- Alternatives (inceasing order)
1844 -> SimplCont
1845 -> SimplM (SimplEnv, OutExpr)
1846
1847 --------------------------------------------------
1848 -- 1. Eliminate the case if there's a known constructor
1849 --------------------------------------------------
1850
1851 rebuildCase env scrut case_bndr alts cont
1852 | Lit lit <- scrut -- No need for same treatment as constructors
1853 -- because literals are inlined more vigorously
1854 , not (litIsLifted lit)
1855 = do { tick (KnownBranch case_bndr)
1856 ; case findAlt (LitAlt lit) alts of
1857 Nothing -> missingAlt env case_bndr alts cont
1858 Just (_, bs, rhs) -> simple_rhs bs rhs }
1859
1860 | Just (con, ty_args, other_args) <- exprIsConApp_maybe (getUnfoldingInRuleMatch env) scrut
1861 -- Works when the scrutinee is a variable with a known unfolding
1862 -- as well as when it's an explicit constructor application
1863 = do { tick (KnownBranch case_bndr)
1864 ; case findAlt (DataAlt con) alts of
1865 Nothing -> missingAlt env case_bndr alts cont
1866 Just (DEFAULT, bs, rhs) -> simple_rhs bs rhs
1867 Just (_, bs, rhs) -> knownCon env scrut con ty_args other_args
1868 case_bndr bs rhs cont
1869 }
1870 where
1871 simple_rhs bs rhs = ASSERT( null bs )
1872 do { env' <- simplNonRecX env case_bndr scrut
1873 -- scrut is a constructor application,
1874 -- hence satisfies let/app invariant
1875 ; simplExprF env' rhs cont }
1876
1877
1878 --------------------------------------------------
1879 -- 2. Eliminate the case if scrutinee is evaluated
1880 --------------------------------------------------
1881
1882 rebuildCase env scrut case_bndr alts@[(_, bndrs, rhs)] cont
1883 -- See if we can get rid of the case altogether
1884 -- See Note [Case elimination]
1885 -- mkCase made sure that if all the alternatives are equal,
1886 -- then there is now only one (DEFAULT) rhs
1887
1888 -- 2a. Dropping the case altogether, if
1889 -- a) it binds nothing (so it's really just a 'seq')
1890 -- b) evaluating the scrutinee has no side effects
1891 | is_plain_seq
1892 , exprOkForSideEffects scrut
1893 -- The entire case is dead, so we can drop it
1894 -- if the scrutinee converges without having imperative
1895 -- side effects or raising a Haskell exception
1896 -- See Note [PrimOp can_fail and has_side_effects] in PrimOp
1897 = simplExprF env rhs cont
1898
1899 -- 2b. Turn the case into a let, if
1900 -- a) it binds only the case-binder
1901 -- b) unlifted case: the scrutinee is ok-for-speculation
1902 -- lifted case: the scrutinee is in HNF (or will later be demanded)
1903 | all_dead_bndrs
1904 , if is_unlifted
1905 then exprOkForSpeculation scrut -- See Note [Case elimination: unlifted case]
1906 else exprIsHNF scrut -- See Note [Case elimination: lifted case]
1907 || scrut_is_demanded_var scrut
1908 = do { tick (CaseElim case_bndr)
1909 ; env' <- simplNonRecX env case_bndr scrut
1910 ; simplExprF env' rhs cont }
1911
1912 -- 2c. Try the seq rules if
1913 -- a) it binds only the case binder
1914 -- b) a rule for seq applies
1915 -- See Note [User-defined RULES for seq] in MkId
1916 | is_plain_seq
1917 = do { let scrut_ty = exprType scrut
1918 rhs_ty = substTy env (exprType rhs)
1919 out_args = [ TyArg { as_arg_ty = scrut_ty
1920 , as_hole_ty = seq_id_ty }
1921 , TyArg { as_arg_ty = rhs_ty
1922 , as_hole_ty = piResultTy seq_id_ty scrut_ty }
1923 , ValArg scrut]
1924 rule_cont = ApplyToVal { sc_dup = NoDup, sc_arg = rhs
1925 , sc_env = env, sc_cont = cont }
1926 env' = zapSubstEnv env
1927 -- Lazily evaluated, so we don't do most of this
1928
1929 ; rule_base <- getSimplRules
1930 ; mb_rule <- tryRules env' (getRules rule_base seqId) seqId out_args rule_cont
1931 ; case mb_rule of
1932 Just (rule_rhs, cont') -> simplExprF env' rule_rhs cont'
1933 Nothing -> reallyRebuildCase env scrut case_bndr alts cont }
1934 where
1935 is_unlifted = isUnliftedType (idType case_bndr)
1936 all_dead_bndrs = all isDeadBinder bndrs -- bndrs are [InId]
1937 is_plain_seq = all_dead_bndrs && isDeadBinder case_bndr -- Evaluation *only* for effect
1938 seq_id_ty = idType seqId
1939
1940 scrut_is_demanded_var :: CoreExpr -> Bool
1941 -- See Note [Eliminating redundant seqs]
1942 scrut_is_demanded_var (Cast s _) = scrut_is_demanded_var s
1943 scrut_is_demanded_var (Var _) = isStrictDmd (idDemandInfo case_bndr)
1944 scrut_is_demanded_var _ = False
1945
1946
1947 rebuildCase env scrut case_bndr alts cont
1948 = reallyRebuildCase env scrut case_bndr alts cont
1949
1950 --------------------------------------------------
1951 -- 3. Catch-all case
1952 --------------------------------------------------
1953
1954 reallyRebuildCase env scrut case_bndr alts cont
1955 = do { -- Prepare the continuation;
1956 -- The new subst_env is in place
1957 (env', dup_cont, nodup_cont) <- prepareCaseCont env alts cont
1958
1959 -- Simplify the alternatives
1960 ; (scrut', case_bndr', alts') <- simplAlts env' scrut case_bndr alts dup_cont
1961
1962 ; dflags <- getDynFlags
1963 ; let alts_ty' = contResultType dup_cont
1964 ; case_expr <- mkCase dflags scrut' case_bndr' alts_ty' alts'
1965
1966 -- Notice that rebuild gets the in-scope set from env', not alt_env
1967 -- (which in any case is only build in simplAlts)
1968 -- The case binder *not* scope over the whole returned case-expression
1969 ; rebuild env' case_expr nodup_cont }
1970
1971 {-
1972 simplCaseBinder checks whether the scrutinee is a variable, v. If so,
1973 try to eliminate uses of v in the RHSs in favour of case_bndr; that
1974 way, there's a chance that v will now only be used once, and hence
1975 inlined.
1976
1977 Historical note: we use to do the "case binder swap" in the Simplifier
1978 so there were additional complications if the scrutinee was a variable.
1979 Now the binder-swap stuff is done in the occurrence analyer; see
1980 OccurAnal Note [Binder swap].
1981
1982 Note [knownCon occ info]
1983 ~~~~~~~~~~~~~~~~~~~~~~~~
1984 If the case binder is not dead, then neither are the pattern bound
1985 variables:
1986 case <any> of x { (a,b) ->
1987 case x of { (p,q) -> p } }
1988 Here (a,b) both look dead, but come alive after the inner case is eliminated.
1989 The point is that we bring into the envt a binding
1990 let x = (a,b)
1991 after the outer case, and that makes (a,b) alive. At least we do unless
1992 the case binder is guaranteed dead.
1993
1994 Note [Case alternative occ info]
1995 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1996 When we are simply reconstructing a case (the common case), we always
1997 zap the occurrence info on the binders in the alternatives. Even
1998 if the case binder is dead, the scrutinee is usually a variable, and *that*
1999 can bring the case-alternative binders back to life.
2000 See Note [Add unfolding for scrutinee]
2001
2002 Note [Improving seq]
2003 ~~~~~~~~~~~~~~~~~~~
2004 Consider
2005 type family F :: * -> *
2006 type instance F Int = Int
2007
2008 ... case e of x { DEFAULT -> rhs } ...
2009
2010 where x::F Int. Then we'd like to rewrite (F Int) to Int, getting
2011
2012 case e `cast` co of x'::Int
2013 I# x# -> let x = x' `cast` sym co
2014 in rhs
2015
2016 so that 'rhs' can take advantage of the form of x'.
2017
2018 Notice that Note [Case of cast] (in OccurAnal) may then apply to the result.
2019
2020 Nota Bene: We only do the [Improving seq] transformation if the
2021 case binder 'x' is actually used in the rhs; that is, if the case
2022 is *not* a *pure* seq.
2023 a) There is no point in adding the cast to a pure seq.
2024 b) There is a good reason not to: doing so would interfere
2025 with seq rules (Note [Built-in RULES for seq] in MkId).
2026 In particular, this [Improving seq] thing *adds* a cast
2027 while [Built-in RULES for seq] *removes* one, so they
2028 just flip-flop.
2029
2030 You might worry about
2031 case v of x { __DEFAULT ->
2032 ... case (v `cast` co) of y { I# -> ... }}
2033 This is a pure seq (since x is unused), so [Improving seq] won't happen.
2034 But it's ok: the simplifier will replace 'v' by 'x' in the rhs to get
2035 case v of x { __DEFAULT ->
2036 ... case (x `cast` co) of y { I# -> ... }}
2037 Now the outer case is not a pure seq, so [Improving seq] will happen,
2038 and then the inner case will disappear.
2039
2040 The need for [Improving seq] showed up in Roman's experiments. Example:
2041 foo :: F Int -> Int -> Int
2042 foo t n = t `seq` bar n
2043 where
2044 bar 0 = 0
2045 bar n = bar (n - case t of TI i -> i)
2046 Here we'd like to avoid repeated evaluating t inside the loop, by
2047 taking advantage of the `seq`.
2048
2049 At one point I did transformation in LiberateCase, but it's more
2050 robust here. (Otherwise, there's a danger that we'll simply drop the
2051 'seq' altogether, before LiberateCase gets to see it.)
2052 -}
2053
2054 simplAlts :: SimplEnv
2055 -> OutExpr
2056 -> InId -- Case binder
2057 -> [InAlt] -- Non-empty
2058 -> SimplCont
2059 -> SimplM (OutExpr, OutId, [OutAlt]) -- Includes the continuation
2060 -- Like simplExpr, this just returns the simplified alternatives;
2061 -- it does not return an environment
2062 -- The returned alternatives can be empty, none are possible
2063
2064 simplAlts env scrut case_bndr alts cont'
2065 = do { let env0 = zapFloats env
2066
2067 ; (env1, case_bndr1) <- simplBinder env0 case_bndr
2068
2069 ; fam_envs <- getFamEnvs
2070 ; (alt_env', scrut', case_bndr') <- improveSeq fam_envs env1 scrut
2071 case_bndr case_bndr1 alts
2072
2073 ; (imposs_deflt_cons, in_alts) <- prepareAlts scrut' case_bndr' alts
2074 -- NB: it's possible that the returned in_alts is empty: this is handled
2075 -- by the caller (rebuildCase) in the missingAlt function
2076
2077 ; alts' <- mapM (simplAlt alt_env' (Just scrut') imposs_deflt_cons case_bndr' cont') in_alts
2078 ; -- pprTrace "simplAlts" (ppr case_bndr $$ ppr alts_ty $$ ppr alts_ty' $$ ppr alts $$ ppr cont') $
2079 return (scrut', case_bndr', alts') }
2080
2081
2082 ------------------------------------
2083 improveSeq :: (FamInstEnv, FamInstEnv) -> SimplEnv
2084 -> OutExpr -> InId -> OutId -> [InAlt]
2085 -> SimplM (SimplEnv, OutExpr, OutId)
2086 -- Note [Improving seq]
2087 improveSeq fam_envs env scrut case_bndr case_bndr1 [(DEFAULT,_,_)]
2088 | not (isDeadBinder case_bndr) -- Not a pure seq! See Note [Improving seq]
2089 , Just (co, ty2) <- topNormaliseType_maybe fam_envs (idType case_bndr1)
2090 = do { case_bndr2 <- newId (fsLit "nt") ty2
2091 ; let rhs = DoneEx (Var case_bndr2 `Cast` mkSymCo co)
2092 env2 = extendIdSubst env case_bndr rhs
2093 ; return (env2, scrut `Cast` co, case_bndr2) }
2094
2095 improveSeq _ env scrut _ case_bndr1 _
2096 = return (env, scrut, case_bndr1)
2097
2098
2099 ------------------------------------
2100 simplAlt :: SimplEnv
2101 -> Maybe OutExpr -- The scrutinee
2102 -> [AltCon] -- These constructors can't be present when
2103 -- matching the DEFAULT alternative
2104 -> OutId -- The case binder
2105 -> SimplCont
2106 -> InAlt
2107 -> SimplM OutAlt
2108
2109 simplAlt env _ imposs_deflt_cons case_bndr' cont' (DEFAULT, bndrs, rhs)
2110 = ASSERT( null bndrs )
2111 do { let env' = addBinderUnfolding env case_bndr'
2112 (mkOtherCon imposs_deflt_cons)
2113 -- Record the constructors that the case-binder *can't* be.
2114 ; rhs' <- simplExprC env' rhs cont'
2115 ; return (DEFAULT, [], rhs') }
2116
2117 simplAlt env scrut' _ case_bndr' cont' (LitAlt lit, bndrs, rhs)
2118 = ASSERT( null bndrs )
2119 do { env' <- addAltUnfoldings env scrut' case_bndr' (Lit lit)
2120 ; rhs' <- simplExprC env' rhs cont'
2121 ; return (LitAlt lit, [], rhs') }
2122
2123 simplAlt env scrut' _ case_bndr' cont' (DataAlt con, vs, rhs)
2124 = do { -- Deal with the pattern-bound variables
2125 -- Mark the ones that are in ! positions in the
2126 -- data constructor as certainly-evaluated.
2127 -- NB: simplLamBinders preserves this eval info
2128 ; let vs_with_evals = add_evals (dataConRepStrictness con)
2129 ; (env', vs') <- simplLamBndrs env vs_with_evals
2130
2131 -- Bind the case-binder to (con args)
2132 ; let inst_tys' = tyConAppArgs (idType case_bndr')
2133 con_app :: OutExpr
2134 con_app = mkConApp2 con inst_tys' vs'
2135
2136 ; env'' <- addAltUnfoldings env' scrut' case_bndr' con_app
2137 ; rhs' <- simplExprC env'' rhs cont'
2138 ; return (DataAlt con, vs', rhs') }
2139 where
2140 -- add_evals records the evaluated-ness of the bound variables of
2141 -- a case pattern. This is *important*. Consider
2142 -- data T = T !Int !Int
2143 --
2144 -- case x of { T a b -> T (a+1) b }
2145 --
2146 -- We really must record that b is already evaluated so that we don't
2147 -- go and re-evaluate it when constructing the result.
2148 -- See Note [Data-con worker strictness] in MkId.hs
2149 add_evals the_strs
2150 = go vs the_strs
2151 where
2152 go [] [] = []
2153 go (v:vs') strs | isTyVar v = v : go vs' strs
2154 go (v:vs') (str:strs)
2155 | isMarkedStrict str = eval v : go vs' strs
2156 | otherwise = zap v : go vs' strs
2157 go _ _ = pprPanic "cat_evals"
2158 (ppr con $$
2159 ppr vs $$
2160 ppr_with_length the_strs $$
2161 ppr_with_length (dataConRepArgTys con) $$
2162 ppr_with_length (dataConRepStrictness con))
2163 where
2164 ppr_with_length list
2165 = ppr list <+> parens (text "length =" <+> ppr (length list))
2166 -- NB: If this panic triggers, note that
2167 -- NoStrictnessMark doesn't print!
2168
2169 zap v = zapIdOccInfo v -- See Note [Case alternative occ info]
2170 eval v = zap v `setIdUnfolding` evaldUnfolding
2171
2172 addAltUnfoldings :: SimplEnv -> Maybe OutExpr -> OutId -> OutExpr -> SimplM SimplEnv
2173 addAltUnfoldings env scrut case_bndr con_app
2174 = do { dflags <- getDynFlags
2175 ; let con_app_unf = mkSimpleUnfolding dflags con_app
2176 env1 = addBinderUnfolding env case_bndr con_app_unf
2177
2178 -- See Note [Add unfolding for scrutinee]
2179 env2 = case scrut of
2180 Just (Var v) -> addBinderUnfolding env1 v con_app_unf
2181 Just (Cast (Var v) co) -> addBinderUnfolding env1 v $
2182 mkSimpleUnfolding dflags (Cast con_app (mkSymCo co))
2183 _ -> env1
2184
2185 ; traceSmpl "addAltUnf" (vcat [ppr case_bndr <+> ppr scrut, ppr con_app])
2186 ; return env2 }
2187
2188 addBinderUnfolding :: SimplEnv -> Id -> Unfolding -> SimplEnv
2189 addBinderUnfolding env bndr unf
2190 | debugIsOn, Just tmpl <- maybeUnfoldingTemplate unf
2191 = WARN( not (eqType (idType bndr) (exprType tmpl)),
2192 ppr bndr $$ ppr (idType bndr) $$ ppr tmpl $$ ppr (exprType tmpl) )
2193 modifyInScope env (bndr `setIdUnfolding` unf)
2194
2195 | otherwise
2196 = modifyInScope env (bndr `setIdUnfolding` unf)
2197
2198 zapBndrOccInfo :: Bool -> Id -> Id
2199 -- Consider case e of b { (a,b) -> ... }
2200 -- Then if we bind b to (a,b) in "...", and b is not dead,
2201 -- then we must zap the deadness info on a,b
2202 zapBndrOccInfo keep_occ_info pat_id
2203 | keep_occ_info = pat_id
2204 | otherwise = zapIdOccInfo pat_id
2205
2206 {-
2207 Note [Add unfolding for scrutinee]
2208 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2209 In general it's unlikely that a variable scrutinee will appear
2210 in the case alternatives case x of { ...x unlikely to appear... }
2211 because the binder-swap in OccAnal has got rid of all such occcurrences
2212 See Note [Binder swap] in OccAnal.
2213
2214 BUT it is still VERY IMPORTANT to add a suitable unfolding for a
2215 variable scrutinee, in simplAlt. Here's why
2216 case x of y
2217 (a,b) -> case b of c
2218 I# v -> ...(f y)...
2219 There is no occurrence of 'b' in the (...(f y)...). But y gets
2220 the unfolding (a,b), and *that* mentions b. If f has a RULE
2221 RULE f (p, I# q) = ...
2222 we want that rule to match, so we must extend the in-scope env with a
2223 suitable unfolding for 'y'. It's *essential* for rule matching; but
2224 it's also good for case-elimintation -- suppose that 'f' was inlined
2225 and did multi-level case analysis, then we'd solve it in one
2226 simplifier sweep instead of two.
2227
2228 Exactly the same issue arises in SpecConstr;
2229 see Note [Add scrutinee to ValueEnv too] in SpecConstr
2230
2231 HOWEVER, given
2232 case x of y { Just a -> r1; Nothing -> r2 }
2233 we do not want to add the unfolding x -> y to 'x', which might seem cool,
2234 since 'y' itself has different unfoldings in r1 and r2. Reason: if we
2235 did that, we'd have to zap y's deadness info and that is a very useful
2236 piece of information.
2237
2238 So instead we add the unfolding x -> Just a, and x -> Nothing in the
2239 respective RHSs.
2240
2241
2242 ************************************************************************
2243 * *
2244 \subsection{Known constructor}
2245 * *
2246 ************************************************************************
2247
2248 We are a bit careful with occurrence info. Here's an example
2249
2250 (\x* -> case x of (a*, b) -> f a) (h v, e)
2251
2252 where the * means "occurs once". This effectively becomes
2253 case (h v, e) of (a*, b) -> f a)
2254 and then
2255 let a* = h v; b = e in f a
2256 and then
2257 f (h v)
2258
2259 All this should happen in one sweep.
2260 -}
2261
2262 knownCon :: SimplEnv
2263 -> OutExpr -- The scrutinee
2264 -> DataCon -> [OutType] -> [OutExpr] -- The scrutinee (in pieces)
2265 -> InId -> [InBndr] -> InExpr -- The alternative
2266 -> SimplCont
2267 -> SimplM (SimplEnv, OutExpr)
2268
2269 knownCon env scrut dc dc_ty_args dc_args bndr bs rhs cont
2270 = do { env' <- bind_args env bs dc_args
2271 ; env'' <- bind_case_bndr env'
2272 ; simplExprF env'' rhs cont }
2273 where
2274 zap_occ = zapBndrOccInfo (isDeadBinder bndr) -- bndr is an InId
2275
2276 -- Ugh!
2277 bind_args env' [] _ = return env'
2278
2279 bind_args env' (b:bs') (Type ty : args)
2280 = ASSERT( isTyVar b )
2281 bind_args (extendTvSubst env' b ty) bs' args
2282
2283 bind_args env' (b:bs') (Coercion co : args)
2284 = ASSERT( isCoVar b )
2285 bind_args (extendCvSubst env' b co) bs' args
2286
2287 bind_args env' (b:bs') (arg : args)
2288 = ASSERT( isId b )
2289 do { let b' = zap_occ b
2290 -- Note that the binder might be "dead", because it doesn't
2291 -- occur in the RHS; and simplNonRecX may therefore discard
2292 -- it via postInlineUnconditionally.
2293 -- Nevertheless we must keep it if the case-binder is alive,
2294 -- because it may be used in the con_app. See Note [knownCon occ info]
2295 ; env'' <- simplNonRecX env' b' arg -- arg satisfies let/app invariant
2296 ; bind_args env'' bs' args }
2297
2298 bind_args _ _ _ =
2299 pprPanic "bind_args" $ ppr dc $$ ppr bs $$ ppr dc_args $$
2300 text "scrut:" <+> ppr scrut
2301
2302 -- It's useful to bind bndr to scrut, rather than to a fresh
2303 -- binding x = Con arg1 .. argn
2304 -- because very often the scrut is a variable, so we avoid
2305 -- creating, and then subsequently eliminating, a let-binding
2306 -- BUT, if scrut is a not a variable, we must be careful
2307 -- about duplicating the arg redexes; in that case, make
2308 -- a new con-app from the args
2309 bind_case_bndr env
2310 | isDeadBinder bndr = return env
2311 | exprIsTrivial scrut = return (extendIdSubst env bndr (DoneEx scrut))
2312 | otherwise = do { dc_args <- mapM (simplVar env) bs
2313 -- dc_ty_args are aready OutTypes,
2314 -- but bs are InBndrs
2315 ; let con_app = Var (dataConWorkId dc)
2316 `mkTyApps` dc_ty_args
2317 `mkApps` dc_args
2318 ; simplNonRecX env bndr con_app }
2319
2320 -------------------
2321 missingAlt :: SimplEnv -> Id -> [InAlt] -> SimplCont -> SimplM (SimplEnv, OutExpr)
2322 -- This isn't strictly an error, although it is unusual.
2323 -- It's possible that the simplifier might "see" that
2324 -- an inner case has no accessible alternatives before
2325 -- it "sees" that the entire branch of an outer case is
2326 -- inaccessible. So we simply put an error case here instead.
2327 missingAlt env case_bndr _ cont
2328 = WARN( True, text "missingAlt" <+> ppr case_bndr )
2329 return (env, mkImpossibleExpr (contResultType cont))
2330
2331 {-
2332 ************************************************************************
2333 * *
2334 \subsection{Duplicating continuations}
2335 * *
2336 ************************************************************************
2337 -}
2338
2339 prepareCaseCont :: SimplEnv
2340 -> [InAlt] -> SimplCont
2341 -> SimplM (SimplEnv,
2342 SimplCont, -- Dupable part
2343 SimplCont) -- Non-dupable part
2344 -- We are considering
2345 -- K[case _ of { p1 -> r1; ...; pn -> rn }]
2346 -- where K is some enclosing continuation for the case
2347 -- Goal: split K into two pieces Kdup,Knodup so that
2348 -- a) Kdup can be duplicated
2349 -- b) Knodup[Kdup[e]] = K[e]
2350 -- The idea is that we'll transform thus:
2351 -- Knodup[ (case _ of { p1 -> Kdup[r1]; ...; pn -> Kdup[rn] }
2352 --
2353 -- We may also return some extra bindings in SimplEnv (that scope over
2354 -- the entire continuation)
2355 --
2356 -- When case-of-case is off, just make the entire continuation non-dupable
2357
2358 prepareCaseCont env alts cont
2359 | not (sm_case_case (getMode env)) = return (env, mkBoringStop (contHoleType cont), cont)
2360 | not (many_alts alts) = return (env, cont, mkBoringStop (contResultType cont))
2361 | otherwise = mkDupableCont env cont
2362 where
2363 many_alts :: [InAlt] -> Bool -- True iff strictly > 1 non-bottom alternative
2364 many_alts [] = False -- See Note [Bottom alternatives]
2365 many_alts [_] = False
2366 many_alts (alt:alts)
2367 | is_bot_alt alt = many_alts alts
2368 | otherwise = not (all is_bot_alt alts)
2369
2370 is_bot_alt (_,_,rhs) = exprIsBottom rhs
2371
2372 {-
2373 Note [Bottom alternatives]
2374 ~~~~~~~~~~~~~~~~~~~~~~~~~~
2375 When we have
2376 case (case x of { A -> error .. ; B -> e; C -> error ..)
2377 of alts
2378 then we can just duplicate those alts because the A and C cases
2379 will disappear immediately. This is more direct than creating
2380 join points and inlining them away; and in some cases we would
2381 not even create the join points (see Note [Single-alternative case])
2382 and we would keep the case-of-case which is silly. See Trac #4930.
2383 -}
2384
2385 mkDupableCont :: SimplEnv -> SimplCont
2386 -> SimplM (SimplEnv, SimplCont, SimplCont)
2387
2388 mkDupableCont env cont
2389 | contIsDupable cont
2390 = return (env, cont, mkBoringStop (contResultType cont))
2391
2392 mkDupableCont _ (Stop {}) = panic "mkDupableCont" -- Handled by previous eqn
2393
2394 mkDupableCont env (CastIt ty cont)
2395 = do { (env', dup, nodup) <- mkDupableCont env cont
2396 ; return (env', CastIt ty dup, nodup) }
2397
2398 -- Duplicating ticks for now, not sure if this is good or not
2399 mkDupableCont env cont@(TickIt{})
2400 = return (env, mkBoringStop (contHoleType cont), cont)
2401
2402 mkDupableCont env cont@(StrictBind {})
2403 = return (env, mkBoringStop (contHoleType cont), cont)
2404 -- See Note [Duplicating StrictBind]
2405
2406 mkDupableCont env (StrictArg info cci cont)
2407 -- See Note [Duplicating StrictArg]
2408 = do { (env', dup, nodup) <- mkDupableCont env cont
2409 ; (env'', args') <- mapAccumLM makeTrivialArg env' (ai_args info)
2410 ; return (env'', StrictArg (info { ai_args = args' }) cci dup, nodup) }
2411
2412 mkDupableCont env cont@(ApplyToTy { sc_cont = tail })
2413 = do { (env', dup_cont, nodup_cont) <- mkDupableCont env tail
2414 ; return (env', cont { sc_cont = dup_cont }, nodup_cont ) }
2415
2416 mkDupableCont env (ApplyToVal { sc_arg = arg, sc_dup = dup, sc_env = se, sc_cont = cont })
2417 = -- e.g. [...hole...] (...arg...)
2418 -- ==>
2419 -- let a = ...arg...
2420 -- in [...hole...] a
2421 do { (env', dup_cont, nodup_cont) <- mkDupableCont env cont
2422 ; (_, se', arg') <- simplArg env' dup se arg
2423 ; (env'', arg'') <- makeTrivial NotTopLevel env' (fsLit "karg") arg'
2424 ; let app_cont = ApplyToVal { sc_arg = arg'', sc_env = se'
2425 , sc_dup = OkToDup, sc_cont = dup_cont }
2426 ; return (env'', app_cont, nodup_cont) }
2427
2428 mkDupableCont env cont@(Select { sc_bndr = case_bndr, sc_alts = [(_, bs, _rhs)] })
2429 -- See Note [Single-alternative case]
2430 -- | not (exprIsDupable rhs && contIsDupable case_cont)
2431 -- | not (isDeadBinder case_bndr)
2432 | all isDeadBinder bs -- InIds
2433 && not (isUnliftedType (idType case_bndr))
2434 -- Note [Single-alternative-unlifted]
2435 = return (env, mkBoringStop (contHoleType cont), cont)
2436
2437 mkDupableCont env (Select { sc_bndr = case_bndr, sc_alts = alts
2438 , sc_env = se, sc_cont = cont })
2439 = -- e.g. (case [...hole...] of { pi -> ei })
2440 -- ===>
2441 -- let ji = \xij -> ei
2442 -- in case [...hole...] of { pi -> ji xij }
2443 do { tick (CaseOfCase case_bndr)
2444 ; (env', dup_cont, nodup_cont) <- prepareCaseCont env alts cont
2445 -- NB: We call prepareCaseCont here. If there is only one
2446 -- alternative, then dup_cont may be big, but that's ok
2447 -- because we push it into the single alternative, and then
2448 -- use mkDupableAlt to turn that simplified alternative into
2449 -- a join point if it's too big to duplicate.
2450 -- And this is important: see Note [Fusing case continuations]
2451
2452 ; let alt_env = se `setInScope` env'
2453
2454 ; (alt_env', case_bndr') <- simplBinder alt_env case_bndr
2455 ; alts' <- mapM (simplAlt alt_env' Nothing [] case_bndr' dup_cont) alts
2456 -- Safe to say that there are no handled-cons for the DEFAULT case
2457 -- NB: simplBinder does not zap deadness occ-info, so
2458 -- a dead case_bndr' will still advertise its deadness
2459 -- This is really important because in
2460 -- case e of b { (# p,q #) -> ... }
2461 -- b is always dead, and indeed we are not allowed to bind b to (# p,q #),
2462 -- which might happen if e was an explicit unboxed pair and b wasn't marked dead.
2463 -- In the new alts we build, we have the new case binder, so it must retain
2464 -- its deadness.
2465 -- NB: we don't use alt_env further; it has the substEnv for
2466 -- the alternatives, and we don't want that
2467
2468 ; (env'', alts'') <- mkDupableAlts env' case_bndr' alts'
2469 ; return (env'', -- Note [Duplicated env]
2470 Select { sc_dup = OkToDup
2471 , sc_bndr = case_bndr', sc_alts = alts''
2472 , sc_env = zapSubstEnv env''
2473 , sc_cont = mkBoringStop (contHoleType nodup_cont) },
2474 nodup_cont) }
2475
2476
2477 mkDupableAlts :: SimplEnv -> OutId -> [InAlt]
2478 -> SimplM (SimplEnv, [InAlt])
2479 -- Absorbs the continuation into the new alternatives
2480
2481 mkDupableAlts env case_bndr' the_alts
2482 = go env the_alts
2483 where
2484 go env0 [] = return (env0, [])
2485 go env0 (alt:alts)
2486 = do { (env1, alt') <- mkDupableAlt env0 case_bndr' alt
2487 ; (env2, alts') <- go env1 alts
2488 ; return (env2, alt' : alts' ) }
2489
2490 mkDupableAlt :: SimplEnv -> OutId -> (AltCon, [CoreBndr], CoreExpr)
2491 -> SimplM (SimplEnv, (AltCon, [CoreBndr], CoreExpr))
2492 mkDupableAlt env case_bndr (con, bndrs', rhs') = do
2493 dflags <- getDynFlags
2494 if exprIsDupable dflags rhs' -- Note [Small alternative rhs]
2495 then return (env, (con, bndrs', rhs'))
2496 else
2497 do { let rhs_ty' = exprType rhs'
2498 scrut_ty = idType case_bndr
2499 case_bndr_w_unf
2500 = case con of
2501 DEFAULT -> case_bndr
2502 DataAlt dc -> setIdUnfolding case_bndr unf
2503 where
2504 -- See Note [Case binders and join points]
2505 unf = mkInlineUnfolding Nothing rhs
2506 rhs = mkConApp2 dc (tyConAppArgs scrut_ty) bndrs'
2507
2508 LitAlt {} -> WARN( True, text "mkDupableAlt"
2509 <+> ppr case_bndr <+> ppr con )
2510 case_bndr
2511 -- The case binder is alive but trivial, so why has
2512 -- it not been substituted away?
2513
2514 used_bndrs' | isDeadBinder case_bndr = filter abstract_over bndrs'
2515 | otherwise = bndrs' ++ [case_bndr_w_unf]
2516
2517 abstract_over bndr
2518 | isTyVar bndr = True -- Abstract over all type variables just in case
2519 | otherwise = not (isDeadBinder bndr)
2520 -- The deadness info on the new Ids is preserved by simplBinders
2521
2522 ; (final_bndrs', final_args) -- Note [Join point abstraction]
2523 <- if (any isId used_bndrs')
2524 then return (used_bndrs', varsToCoreExprs used_bndrs')
2525 else do { rw_id <- newId (fsLit "w") voidPrimTy
2526 ; return ([setOneShotLambda rw_id], [Var voidPrimId]) }
2527
2528 ; join_bndr <- newId (fsLit "$j") (mkLamTypes final_bndrs' rhs_ty')
2529 -- Note [Funky mkLamTypes]
2530
2531 ; let -- We make the lambdas into one-shot-lambdas. The
2532 -- join point is sure to be applied at most once, and doing so
2533 -- prevents the body of the join point being floated out by
2534 -- the full laziness pass
2535 really_final_bndrs = map one_shot final_bndrs'
2536 one_shot v | isId v = setOneShotLambda v
2537 | otherwise = v
2538 join_rhs = mkLams really_final_bndrs rhs'
2539 join_arity = exprArity join_rhs
2540 join_call = mkApps (Var join_bndr) final_args
2541
2542 ; env' <- addPolyBind NotTopLevel env (NonRec (join_bndr `setIdArity` join_arity) join_rhs)
2543 ; return (env', (con, bndrs', join_call)) }
2544 -- See Note [Duplicated env]
2545
2546 {-
2547 Note [Fusing case continuations]
2548 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2549 It's important to fuse two successive case continuations when the
2550 first has one alternative. That's why we call prepareCaseCont here.
2551 Consider this, which arises from thunk splitting (see Note [Thunk
2552 splitting] in WorkWrap):
2553
2554 let
2555 x* = case (case v of {pn -> rn}) of
2556 I# a -> I# a
2557 in body
2558
2559 The simplifier will find
2560 (Var v) with continuation
2561 Select (pn -> rn) (
2562 Select [I# a -> I# a] (
2563 StrictBind body Stop
2564
2565 So we'll call mkDupableCont on
2566 Select [I# a -> I# a] (StrictBind body Stop)
2567 There is just one alternative in the first Select, so we want to
2568 simplify the rhs (I# a) with continuation (StricgtBind body Stop)
2569 Supposing that body is big, we end up with
2570 let $j a = <let x = I# a in body>
2571 in case v of { pn -> case rn of
2572 I# a -> $j a }
2573 This is just what we want because the rn produces a box that
2574 the case rn cancels with.
2575
2576 See Trac #4957 a fuller example.
2577
2578 Note [Case binders and join points]
2579 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2580 Consider this
2581 case (case .. ) of c {
2582 I# c# -> ....c....
2583
2584 If we make a join point with c but not c# we get
2585 $j = \c -> ....c....
2586
2587 But if later inlining scrutines the c, thus
2588
2589 $j = \c -> ... case c of { I# y -> ... } ...
2590
2591 we won't see that 'c' has already been scrutinised. This actually
2592 happens in the 'tabulate' function in wave4main, and makes a significant
2593 difference to allocation.
2594
2595 An alternative plan is this:
2596
2597 $j = \c# -> let c = I# c# in ...c....
2598
2599 but that is bad if 'c' is *not* later scrutinised.
2600
2601 So instead we do both: we pass 'c' and 'c#' , and record in c's inlining
2602 (a stable unfolding) that it's really I# c#, thus
2603
2604 $j = \c# -> \c[=I# c#] -> ...c....
2605
2606 Absence analysis may later discard 'c'.
2607
2608 NB: take great care when doing strictness analysis;
2609 see Note [Lamba-bound unfoldings] in DmdAnal.
2610
2611 Also note that we can still end up passing stuff that isn't used. Before
2612 strictness analysis we have
2613 let $j x y c{=(x,y)} = (h c, ...)
2614 in ...
2615 After strictness analysis we see that h is strict, we end up with
2616 let $j x y c{=(x,y)} = ($wh x y, ...)
2617 and c is unused.
2618
2619 Note [Duplicated env]
2620 ~~~~~~~~~~~~~~~~~~~~~
2621 Some of the alternatives are simplified, but have not been turned into a join point
2622 So they *must* have an zapped subst-env. So we can't use completeNonRecX to
2623 bind the join point, because it might to do PostInlineUnconditionally, and
2624 we'd lose that when zapping the subst-env. We could have a per-alt subst-env,
2625 but zapping it (as we do in mkDupableCont, the Select case) is safe, and
2626 at worst delays the join-point inlining.
2627
2628 Note [Small alternative rhs]
2629 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2630 It is worth checking for a small RHS because otherwise we
2631 get extra let bindings that may cause an extra iteration of the simplifier to
2632 inline back in place. Quite often the rhs is just a variable or constructor.
2633 The Ord instance of Maybe in PrelMaybe.hs, for example, took several extra
2634 iterations because the version with the let bindings looked big, and so wasn't
2635 inlined, but after the join points had been inlined it looked smaller, and so
2636 was inlined.
2637
2638 NB: we have to check the size of rhs', not rhs.
2639 Duplicating a small InAlt might invalidate occurrence information
2640 However, if it *is* dupable, we return the *un* simplified alternative,
2641 because otherwise we'd need to pair it up with an empty subst-env....
2642 but we only have one env shared between all the alts.
2643 (Remember we must zap the subst-env before re-simplifying something).
2644 Rather than do this we simply agree to re-simplify the original (small) thing later.
2645
2646 Note [Funky mkLamTypes]
2647 ~~~~~~~~~~~~~~~~~~~~~~
2648 Notice the funky mkLamTypes. If the contructor has existentials
2649 it's possible that the join point will be abstracted over
2650 type variables as well as term variables.
2651 Example: Suppose we have
2652 data T = forall t. C [t]
2653 Then faced with
2654 case (case e of ...) of
2655 C t xs::[t] -> rhs
2656 We get the join point
2657 let j :: forall t. [t] -> ...
2658 j = /\t \xs::[t] -> rhs
2659 in
2660 case (case e of ...) of
2661 C t xs::[t] -> j t xs
2662
2663 Note [Join point abstraction]
2664 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2665 Join points always have at least one value argument,
2666 for several reasons
2667
2668 * If we try to lift a primitive-typed something out
2669 for let-binding-purposes, we will *caseify* it (!),
2670 with potentially-disastrous strictness results. So
2671 instead we turn it into a function: \v -> e
2672 where v::Void#. The value passed to this function is void,
2673 which generates (almost) no code.
2674
2675 * CPR. We used to say "&& isUnliftedType rhs_ty'" here, but now
2676 we make the join point into a function whenever used_bndrs'
2677 is empty. This makes the join-point more CPR friendly.
2678 Consider: let j = if .. then I# 3 else I# 4
2679 in case .. of { A -> j; B -> j; C -> ... }
2680
2681 Now CPR doesn't w/w j because it's a thunk, so
2682 that means that the enclosing function can't w/w either,
2683 which is a lose. Here's the example that happened in practice:
2684 kgmod :: Int -> Int -> Int
2685 kgmod x y = if x > 0 && y < 0 || x < 0 && y > 0
2686 then 78
2687 else 5
2688
2689 * Let-no-escape. We want a join point to turn into a let-no-escape
2690 so that it is implemented as a jump, and one of the conditions
2691 for LNE is that it's not updatable. In CoreToStg, see
2692 Note [What is a non-escaping let]
2693
2694 * Floating. Since a join point will be entered once, no sharing is
2695 gained by floating out, but something might be lost by doing
2696 so because it might be allocated.
2697
2698 I have seen a case alternative like this:
2699 True -> \v -> ...
2700 It's a bit silly to add the realWorld dummy arg in this case, making
2701 $j = \s v -> ...
2702 True -> $j s
2703 (the \v alone is enough to make CPR happy) but I think it's rare
2704
2705 There's a slight infelicity here: we pass the overall
2706 case_bndr to all the join points if it's used in *any* RHS,
2707 because we don't know its usage in each RHS separately
2708
2709
2710 Note [Duplicating StrictArg]
2711 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2712 The original plan had (where E is a big argument)
2713 e.g. f E [..hole..]
2714 ==> let $j = \a -> f E a
2715 in $j [..hole..]
2716
2717 But this is terrible! Here's an example:
2718 && E (case x of { T -> F; F -> T })
2719 Now, && is strict so we end up simplifying the case with
2720
2721 an ArgOf continuation. If we let-bind it, we get
2722 let $j = \v -> && E v
2723 in simplExpr (case x of { T -> F; F -> T })
2724 (ArgOf (\r -> $j r)
2725 And after simplifying more we get
2726 let $j = \v -> && E v
2727 in case x of { T -> $j F; F -> $j T }
2728 Which is a Very Bad Thing
2729
2730 What we do now is this
2731 f E [..hole..]
2732 ==> let a = E
2733 in f a [..hole..]
2734 Now if the thing in the hole is a case expression (which is when
2735 we'll call mkDupableCont), we'll push the function call into the
2736 branches, which is what we want. Now RULES for f may fire, and
2737 call-pattern specialisation. Here's an example from Trac #3116
2738 go (n+1) (case l of
2739 1 -> bs'
2740 _ -> Chunk p fpc (o+1) (l-1) bs')
2741 If we can push the call for 'go' inside the case, we get
2742 call-pattern specialisation for 'go', which is *crucial* for
2743 this program.
2744
2745 Here is the (&&) example:
2746 && E (case x of { T -> F; F -> T })
2747 ==> let a = E in
2748 case x of { T -> && a F; F -> && a T }
2749 Much better!
2750
2751 Notice that
2752 * Arguments to f *after* the strict one are handled by
2753 the ApplyToVal case of mkDupableCont. Eg
2754 f [..hole..] E
2755
2756 * We can only do the let-binding of E because the function
2757 part of a StrictArg continuation is an explicit syntax
2758 tree. In earlier versions we represented it as a function
2759 (CoreExpr -> CoreEpxr) which we couldn't take apart.
2760
2761 Do *not* duplicate StrictBind and StritArg continuations. We gain
2762 nothing by propagating them into the expressions, and we do lose a
2763 lot.
2764
2765 The desire not to duplicate is the entire reason that
2766 mkDupableCont returns a pair of continuations.
2767
2768 Note [Duplicating StrictBind]
2769 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2770 Unlike StrictArg, there doesn't seem anything to gain from
2771 duplicating a StrictBind continuation, so we don't.
2772
2773
2774 Note [Single-alternative cases]
2775 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2776 This case is just like the ArgOf case. Here's an example:
2777 data T a = MkT !a
2778 ...(MkT (abs x))...
2779 Then we get
2780 case (case x of I# x' ->
2781 case x' <# 0# of
2782 True -> I# (negate# x')
2783 False -> I# x') of y {
2784 DEFAULT -> MkT y
2785 Because the (case x) has only one alternative, we'll transform to
2786 case x of I# x' ->
2787 case (case x' <# 0# of
2788 True -> I# (negate# x')
2789 False -> I# x') of y {
2790 DEFAULT -> MkT y
2791 But now we do *NOT* want to make a join point etc, giving
2792 case x of I# x' ->
2793 let $j = \y -> MkT y
2794 in case x' <# 0# of
2795 True -> $j (I# (negate# x'))
2796 False -> $j (I# x')
2797 In this case the $j will inline again, but suppose there was a big
2798 strict computation enclosing the orginal call to MkT. Then, it won't
2799 "see" the MkT any more, because it's big and won't get duplicated.
2800 And, what is worse, nothing was gained by the case-of-case transform.
2801
2802 So, in circumstances like these, we don't want to build join points
2803 and push the outer case into the branches of the inner one. Instead,
2804 don't duplicate the continuation.
2805
2806 When should we use this strategy? We should not use it on *every*
2807 single-alternative case:
2808 e.g. case (case ....) of (a,b) -> (# a,b #)
2809 Here we must push the outer case into the inner one!
2810 Other choices:
2811
2812 * Match [(DEFAULT,_,_)], but in the common case of Int,
2813 the alternative-filling-in code turned the outer case into
2814 case (...) of y { I# _ -> MkT y }
2815
2816 * Match on single alternative plus (not (isDeadBinder case_bndr))
2817 Rationale: pushing the case inwards won't eliminate the construction.
2818 But there's a risk of
2819 case (...) of y { (a,b) -> let z=(a,b) in ... }
2820 Now y looks dead, but it'll come alive again. Still, this
2821 seems like the best option at the moment.
2822
2823 * Match on single alternative plus (all (isDeadBinder bndrs))
2824 Rationale: this is essentially seq.
2825
2826 * Match when the rhs is *not* duplicable, and hence would lead to a
2827 join point. This catches the disaster-case above. We can test
2828 the *un-simplified* rhs, which is fine. It might get bigger or
2829 smaller after simplification; if it gets smaller, this case might
2830 fire next time round. NB also that we must test contIsDupable
2831 case_cont *too, because case_cont might be big!
2832
2833 HOWEVER: I found that this version doesn't work well, because
2834 we can get let x = case (...) of { small } in ...case x...
2835 When x is inlined into its full context, we find that it was a bad
2836 idea to have pushed the outer case inside the (...) case.
2837
2838 There is a cost to not doing case-of-case; see Trac #10626.
2839
2840 Note [Single-alternative-unlifted]
2841 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2842 Here's another single-alternative where we really want to do case-of-case:
2843
2844 data Mk1 = Mk1 Int# | Mk2 Int#
2845
2846 M1.f =
2847 \r [x_s74 y_s6X]
2848 case
2849 case y_s6X of tpl_s7m {
2850 M1.Mk1 ipv_s70 -> ipv_s70;
2851 M1.Mk2 ipv_s72 -> ipv_s72;
2852 }
2853 of
2854 wild_s7c
2855 { __DEFAULT ->
2856 case
2857 case x_s74 of tpl_s7n {
2858 M1.Mk1 ipv_s77 -> ipv_s77;
2859 M1.Mk2 ipv_s79 -> ipv_s79;
2860 }
2861 of
2862 wild1_s7b
2863 { __DEFAULT -> ==# [wild1_s7b wild_s7c];
2864 };
2865 };
2866
2867 So the outer case is doing *nothing at all*, other than serving as a
2868 join-point. In this case we really want to do case-of-case and decide
2869 whether to use a real join point or just duplicate the continuation:
2870
2871 let $j s7c = case x of
2872 Mk1 ipv77 -> (==) s7c ipv77
2873 Mk1 ipv79 -> (==) s7c ipv79
2874 in
2875 case y of
2876 Mk1 ipv70 -> $j ipv70
2877 Mk2 ipv72 -> $j ipv72
2878
2879 Hence: check whether the case binder's type is unlifted, because then
2880 the outer case is *not* a seq.
2881
2882 ************************************************************************
2883 * *
2884 Unfoldings
2885 * *
2886 ************************************************************************
2887 -}
2888
2889 simplLetUnfolding :: SimplEnv-> TopLevelFlag
2890 -> InId
2891 -> OutExpr
2892 -> Unfolding -> SimplM Unfolding
2893 simplLetUnfolding env top_lvl id new_rhs unf
2894 | isStableUnfolding unf
2895 = simplUnfolding env top_lvl id unf
2896 | otherwise
2897 = bottoming `seq` -- See Note [Force bottoming field]
2898 do { dflags <- getDynFlags
2899 ; return (mkUnfolding dflags InlineRhs (isTopLevel top_lvl) bottoming new_rhs) }
2900 -- We make an unfolding *even for loop-breakers*.
2901 -- Reason: (a) It might be useful to know that they are WHNF
2902 -- (b) In TidyPgm we currently assume that, if we want to
2903 -- expose the unfolding then indeed we *have* an unfolding
2904 -- to expose. (We could instead use the RHS, but currently
2905 -- we don't.) The simple thing is always to have one.
2906 where
2907 bottoming = isBottomingId id
2908
2909 simplUnfolding :: SimplEnv-> TopLevelFlag -> InId -> Unfolding -> SimplM Unfolding
2910 -- Note [Setting the new unfolding]
2911 simplUnfolding env top_lvl id unf
2912 = case unf of
2913 NoUnfolding -> return unf
2914 OtherCon {} -> return unf
2915
2916 DFunUnfolding { df_bndrs = bndrs, df_con = con, df_args = args }
2917 -> do { (env', bndrs') <- simplBinders rule_env bndrs
2918 ; args' <- mapM (simplExpr env') args
2919 ; return (mkDFunUnfolding bndrs' con args') }
2920
2921 CoreUnfolding { uf_tmpl = expr, uf_src = src, uf_guidance = guide }
2922 | isStableSource src
2923 -> do { expr' <- simplExpr rule_env expr
2924 ; case guide of
2925 UnfWhen { ug_arity = arity, ug_unsat_ok = sat_ok } -- Happens for INLINE things
2926 -> let guide' = UnfWhen { ug_arity = arity, ug_unsat_ok = sat_ok
2927 , ug_boring_ok = inlineBoringOk expr' }
2928 -- Refresh the boring-ok flag, in case expr'
2929 -- has got small. This happens, notably in the inlinings
2930 -- for dfuns for single-method classes; see
2931 -- Note [Single-method classes] in TcInstDcls.
2932 -- A test case is Trac #4138
2933 in return (mkCoreUnfolding src is_top_lvl expr' guide')
2934 -- See Note [Top-level flag on inline rules] in CoreUnfold
2935
2936 _other -- Happens for INLINABLE things
2937 -> bottoming `seq` -- See Note [Force bottoming field]
2938 do { dflags <- getDynFlags
2939 ; return (mkUnfolding dflags src is_top_lvl bottoming expr') } }
2940 -- If the guidance is UnfIfGoodArgs, this is an INLINABLE
2941 -- unfolding, and we need to make sure the guidance is kept up
2942 -- to date with respect to any changes in the unfolding.
2943
2944 | otherwise -> return noUnfolding -- Discard unstable unfoldings
2945 where
2946 bottoming = isBottomingId id
2947 is_top_lvl = isTopLevel top_lvl
2948 act = idInlineActivation id
2949 rule_env = updMode (updModeForStableUnfoldings act) env
2950 -- See Note [Simplifying inside stable unfoldings] in SimplUtils
2951
2952 {-
2953 Note [Force bottoming field]
2954 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2955 We need to force bottoming, or the new unfolding holds
2956 on to the old unfolding (which is part of the id).
2957
2958 Note [Setting the new unfolding]
2959 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2960 * If there's an INLINE pragma, we simplify the RHS gently. Maybe we
2961 should do nothing at all, but simplifying gently might get rid of
2962 more crap.
2963
2964 * If not, we make an unfolding from the new RHS. But *only* for
2965 non-loop-breakers. Making loop breakers not have an unfolding at all
2966 means that we can avoid tests in exprIsConApp, for example. This is
2967 important: if exprIsConApp says 'yes' for a recursive thing, then we
2968 can get into an infinite loop
2969
2970 If there's an stable unfolding on a loop breaker (which happens for
2971 INLINEABLE), we hang on to the inlining. It's pretty dodgy, but the
2972 user did say 'INLINE'. May need to revisit this choice.
2973
2974 ************************************************************************
2975 * *
2976 Rules
2977 * *
2978 ************************************************************************
2979
2980 Note [Rules in a letrec]
2981 ~~~~~~~~~~~~~~~~~~~~~~~~
2982 After creating fresh binders for the binders of a letrec, we
2983 substitute the RULES and add them back onto the binders; this is done
2984 *before* processing any of the RHSs. This is important. Manuel found
2985 cases where he really, really wanted a RULE for a recursive function
2986 to apply in that function's own right-hand side.
2987
2988 See Note [Loop breaking and RULES] in OccAnal.
2989 -}
2990
2991 addBndrRules :: SimplEnv -> InBndr -> OutBndr -> SimplM (SimplEnv, OutBndr)
2992 -- Rules are added back into the bin
2993 addBndrRules env in_id out_id
2994 | null old_rules
2995 = return (env, out_id)
2996 | otherwise
2997 = do { new_rules <- simplRules env (Just (idName out_id)) old_rules
2998 ; let final_id = out_id `setIdSpecialisation` mkRuleInfo new_rules
2999 ; return (modifyInScope env final_id, final_id) }
3000 where
3001 old_rules = ruleInfoRules (idSpecialisation in_id)
3002
3003 simplRules :: SimplEnv -> Maybe Name -> [CoreRule] -> SimplM [CoreRule]
3004 simplRules env mb_new_nm rules
3005 = mapM simpl_rule rules
3006 where
3007 simpl_rule rule@(BuiltinRule {})
3008 = return rule
3009
3010 simpl_rule rule@(Rule { ru_bndrs = bndrs, ru_args = args
3011 , ru_fn = fn_name, ru_rhs = rhs })
3012 = do { (env', bndrs') <- simplBinders env bndrs
3013 ; let rule_env = updMode updModeForRules env'
3014 ; args' <- mapM (simplExpr rule_env) args
3015 ; rhs' <- simplExpr rule_env rhs
3016 ; return (rule { ru_bndrs = bndrs'
3017 , ru_fn = mb_new_nm `orElse` fn_name
3018 , ru_args = args'
3019 , ru_rhs = rhs' }) }