Split mkInlineUnfolding into two functions
[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 CoreSubst ( pushCoTyArg, pushCoValArg )
39 --import PrimOp ( tagToEnumKey ) -- temporalily commented out. See #8326
40 import Rules ( mkRuleInfo, lookupRule, getRules )
41 import TysPrim ( voidPrimTy ) --, intPrimTy ) -- temporalily commented out. See #8326
42 import BasicTypes ( TopLevelFlag(..), isTopLevel, RecFlag(..) )
43 import MonadUtils ( foldlM, mapAccumLM, liftIO )
44 import Maybes ( orElse )
45 --import Unique ( hasKey ) -- temporalily commented out. See #8326
46 import Control.Monad
47 import Outputable
48 import FastString
49 import Pair
50 import Util
51 import ErrUtils
52
53 {-
54 The guts of the simplifier is in this module, but the driver loop for
55 the simplifier is in SimplCore.hs.
56
57
58 -----------------------------------------
59 *** IMPORTANT NOTE ***
60 -----------------------------------------
61 The simplifier used to guarantee that the output had no shadowing, but
62 it does not do so any more. (Actually, it never did!) The reason is
63 documented with simplifyArgs.
64
65
66 -----------------------------------------
67 *** IMPORTANT NOTE ***
68 -----------------------------------------
69 Many parts of the simplifier return a bunch of "floats" as well as an
70 expression. This is wrapped as a datatype SimplUtils.FloatsWith.
71
72 All "floats" are let-binds, not case-binds, but some non-rec lets may
73 be unlifted (with RHS ok-for-speculation).
74
75
76
77 -----------------------------------------
78 ORGANISATION OF FUNCTIONS
79 -----------------------------------------
80 simplTopBinds
81 - simplify all top-level binders
82 - for NonRec, call simplRecOrTopPair
83 - for Rec, call simplRecBind
84
85
86 ------------------------------
87 simplExpr (applied lambda) ==> simplNonRecBind
88 simplExpr (Let (NonRec ...) ..) ==> simplNonRecBind
89 simplExpr (Let (Rec ...) ..) ==> simplify binders; simplRecBind
90
91 ------------------------------
92 simplRecBind [binders already simplfied]
93 - use simplRecOrTopPair on each pair in turn
94
95 simplRecOrTopPair [binder already simplified]
96 Used for: recursive bindings (top level and nested)
97 top-level non-recursive bindings
98 Returns:
99 - check for PreInlineUnconditionally
100 - simplLazyBind
101
102 simplNonRecBind
103 Used for: non-top-level non-recursive bindings
104 beta reductions (which amount to the same thing)
105 Because it can deal with strict arts, it takes a
106 "thing-inside" and returns an expression
107
108 - check for PreInlineUnconditionally
109 - simplify binder, including its IdInfo
110 - if strict binding
111 simplStrictArg
112 mkAtomicArgs
113 completeNonRecX
114 else
115 simplLazyBind
116 addFloats
117
118 simplNonRecX: [given a *simplified* RHS, but an *unsimplified* binder]
119 Used for: binding case-binder and constr args in a known-constructor case
120 - check for PreInLineUnconditionally
121 - simplify binder
122 - completeNonRecX
123
124 ------------------------------
125 simplLazyBind: [binder already simplified, RHS not]
126 Used for: recursive bindings (top level and nested)
127 top-level non-recursive bindings
128 non-top-level, but *lazy* non-recursive bindings
129 [must not be strict or unboxed]
130 Returns floats + an augmented environment, not an expression
131 - substituteIdInfo and add result to in-scope
132 [so that rules are available in rec rhs]
133 - simplify rhs
134 - mkAtomicArgs
135 - float if exposes constructor or PAP
136 - completeBind
137
138
139 completeNonRecX: [binder and rhs both simplified]
140 - if the the thing needs case binding (unlifted and not ok-for-spec)
141 build a Case
142 else
143 completeBind
144 addFloats
145
146 completeBind: [given a simplified RHS]
147 [used for both rec and non-rec bindings, top level and not]
148 - try PostInlineUnconditionally
149 - add unfolding [this is the only place we add an unfolding]
150 - add arity
151
152
153
154 Right hand sides and arguments
155 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
156 In many ways we want to treat
157 (a) the right hand side of a let(rec), and
158 (b) a function argument
159 in the same way. But not always! In particular, we would
160 like to leave these arguments exactly as they are, so they
161 will match a RULE more easily.
162
163 f (g x, h x)
164 g (+ x)
165
166 It's harder to make the rule match if we ANF-ise the constructor,
167 or eta-expand the PAP:
168
169 f (let { a = g x; b = h x } in (a,b))
170 g (\y. + x y)
171
172 On the other hand if we see the let-defns
173
174 p = (g x, h x)
175 q = + x
176
177 then we *do* want to ANF-ise and eta-expand, so that p and q
178 can be safely inlined.
179
180 Even floating lets out is a bit dubious. For let RHS's we float lets
181 out if that exposes a value, so that the value can be inlined more vigorously.
182 For example
183
184 r = let x = e in (x,x)
185
186 Here, if we float the let out we'll expose a nice constructor. We did experiments
187 that showed this to be a generally good thing. But it was a bad thing to float
188 lets out unconditionally, because that meant they got allocated more often.
189
190 For function arguments, there's less reason to expose a constructor (it won't
191 get inlined). Just possibly it might make a rule match, but I'm pretty skeptical.
192 So for the moment we don't float lets out of function arguments either.
193
194
195 Eta expansion
196 ~~~~~~~~~~~~~~
197 For eta expansion, we want to catch things like
198
199 case e of (a,b) -> \x -> case a of (p,q) -> \y -> r
200
201 If the \x was on the RHS of a let, we'd eta expand to bring the two
202 lambdas together. And in general that's a good thing to do. Perhaps
203 we should eta expand wherever we find a (value) lambda? Then the eta
204 expansion at a let RHS can concentrate solely on the PAP case.
205
206
207 ************************************************************************
208 * *
209 \subsection{Bindings}
210 * *
211 ************************************************************************
212 -}
213
214 simplTopBinds :: SimplEnv -> [InBind] -> SimplM SimplEnv
215
216 simplTopBinds env0 binds0
217 = do { -- Put all the top-level binders into scope at the start
218 -- so that if a transformation rule has unexpectedly brought
219 -- anything into scope, then we don't get a complaint about that.
220 -- It's rather as if the top-level binders were imported.
221 -- See note [Glomming] in OccurAnal.
222 ; env1 <- simplRecBndrs env0 (bindersOfBinds binds0)
223 ; env2 <- simpl_binds env1 binds0
224 ; freeTick SimplifierDone
225 ; return env2 }
226 where
227 -- We need to track the zapped top-level binders, because
228 -- they should have their fragile IdInfo zapped (notably occurrence info)
229 -- That's why we run down binds and bndrs' simultaneously.
230 --
231 simpl_binds :: SimplEnv -> [InBind] -> SimplM SimplEnv
232 simpl_binds env [] = return env
233 simpl_binds env (bind:binds) = do { env' <- simpl_bind env bind
234 ; simpl_binds env' binds }
235
236 simpl_bind env (Rec pairs) = simplRecBind env TopLevel pairs
237 simpl_bind env (NonRec b r) = do { (env', b') <- addBndrRules env b (lookupRecBndr env b)
238 ; simplRecOrTopPair env' TopLevel NonRecursive b b' r }
239
240 {-
241 ************************************************************************
242 * *
243 \subsection{Lazy bindings}
244 * *
245 ************************************************************************
246
247 simplRecBind is used for
248 * recursive bindings only
249 -}
250
251 simplRecBind :: SimplEnv -> TopLevelFlag
252 -> [(InId, InExpr)]
253 -> SimplM SimplEnv
254 simplRecBind env0 top_lvl pairs0
255 = do { (env_with_info, triples) <- mapAccumLM add_rules env0 pairs0
256 ; env1 <- go (zapFloats env_with_info) triples
257 ; return (env0 `addRecFloats` env1) }
258 -- addFloats adds the floats from env1,
259 -- _and_ updates env0 with the in-scope set from env1
260 where
261 add_rules :: SimplEnv -> (InBndr,InExpr) -> SimplM (SimplEnv, (InBndr, OutBndr, InExpr))
262 -- Add the (substituted) rules to the binder
263 add_rules env (bndr, rhs)
264 = do { (env', bndr') <- addBndrRules env bndr (lookupRecBndr env bndr)
265 ; return (env', (bndr, bndr', rhs)) }
266
267 go env [] = return env
268
269 go env ((old_bndr, new_bndr, rhs) : pairs)
270 = do { env' <- simplRecOrTopPair env top_lvl Recursive old_bndr new_bndr rhs
271 ; go env' pairs }
272
273 {-
274 simplOrTopPair is used for
275 * recursive bindings (whether top level or not)
276 * top-level non-recursive bindings
277
278 It assumes the binder has already been simplified, but not its IdInfo.
279 -}
280
281 simplRecOrTopPair :: SimplEnv
282 -> TopLevelFlag -> RecFlag
283 -> InId -> OutBndr -> InExpr -- Binder and rhs
284 -> SimplM SimplEnv -- Returns an env that includes the binding
285
286 simplRecOrTopPair env top_lvl is_rec old_bndr new_bndr rhs
287 = do { dflags <- getDynFlags
288 ; trace_bind dflags $
289 if preInlineUnconditionally dflags env top_lvl old_bndr rhs
290 -- Check for unconditional inline
291 then do tick (PreInlineUnconditionally old_bndr)
292 return (extendIdSubst env old_bndr (mkContEx env rhs))
293 else simplLazyBind env top_lvl is_rec old_bndr new_bndr rhs env }
294 where
295 trace_bind dflags thing_inside
296 | not (dopt Opt_D_verbose_core2core dflags)
297 = thing_inside
298 | otherwise
299 = pprTrace "SimplBind" (ppr old_bndr) thing_inside
300 -- trace_bind emits a trace for each top-level binding, which
301 -- helps to locate the tracing for inlining and rule firing
302
303 {-
304 simplLazyBind is used for
305 * [simplRecOrTopPair] recursive bindings (whether top level or not)
306 * [simplRecOrTopPair] top-level non-recursive bindings
307 * [simplNonRecE] non-top-level *lazy* non-recursive bindings
308
309 Nota bene:
310 1. It assumes that the binder is *already* simplified,
311 and is in scope, and its IdInfo too, except unfolding
312
313 2. It assumes that the binder type is lifted.
314
315 3. It does not check for pre-inline-unconditionally;
316 that should have been done already.
317 -}
318
319 simplLazyBind :: SimplEnv
320 -> TopLevelFlag -> RecFlag
321 -> InId -> OutId -- Binder, both pre-and post simpl
322 -- The OutId has IdInfo, except arity, unfolding
323 -> InExpr -> SimplEnv -- The RHS and its environment
324 -> SimplM SimplEnv
325 -- Precondition: rhs obeys the let/app invariant
326 simplLazyBind env top_lvl is_rec bndr bndr1 rhs rhs_se
327 = -- pprTrace "simplLazyBind" ((ppr bndr <+> ppr bndr1) $$ ppr rhs $$ ppr (seIdSubst rhs_se)) $
328 do { let rhs_env = rhs_se `setInScope` env
329 (tvs, body) = case collectTyAndValBinders rhs of
330 (tvs, [], body)
331 | surely_not_lam body -> (tvs, body)
332 _ -> ([], rhs)
333
334 surely_not_lam (Lam {}) = False
335 surely_not_lam (Tick t e)
336 | not (tickishFloatable t) = surely_not_lam e
337 -- eta-reduction could float
338 surely_not_lam _ = True
339 -- Do not do the "abstract tyyvar" thing if there's
340 -- a lambda inside, because it defeats eta-reduction
341 -- f = /\a. \x. g a x
342 -- should eta-reduce.
343
344
345 ; (body_env, tvs') <- simplBinders rhs_env tvs
346 -- See Note [Floating and type abstraction] in SimplUtils
347
348 -- Simplify the RHS
349 ; let rhs_cont = mkRhsStop (substTy body_env (exprType body))
350 ; (body_env1, body1) <- simplExprF body_env body rhs_cont
351 -- ANF-ise a constructor or PAP rhs
352 ; (body_env2, body2) <- prepareRhs top_lvl body_env1 bndr1 body1
353
354 ; (env', rhs')
355 <- if not (doFloatFromRhs top_lvl is_rec False body2 body_env2)
356 then -- No floating, revert to body1
357 do { rhs' <- mkLam env tvs' (wrapFloats body_env1 body1) rhs_cont
358 ; return (env, rhs') }
359
360 else if null tvs then -- Simple floating
361 do { tick LetFloatFromLet
362 ; return (addFloats env body_env2, body2) }
363
364 else -- Do type-abstraction first
365 do { tick LetFloatFromLet
366 ; (poly_binds, body3) <- abstractFloats tvs' body_env2 body2
367 ; rhs' <- mkLam env tvs' body3 rhs_cont
368 ; env' <- foldlM (addPolyBind top_lvl) env poly_binds
369 ; return (env', rhs') }
370
371 ; completeBind env' top_lvl bndr bndr1 rhs' }
372
373 {-
374 A specialised variant of simplNonRec used when the RHS is already simplified,
375 notably in knownCon. It uses case-binding where necessary.
376 -}
377
378 simplNonRecX :: SimplEnv
379 -> InId -- Old binder
380 -> OutExpr -- Simplified RHS
381 -> SimplM SimplEnv
382 -- Precondition: rhs satisfies the let/app invariant
383 simplNonRecX env bndr new_rhs
384 | isDeadBinder bndr -- Not uncommon; e.g. case (a,b) of c { (p,q) -> p }
385 = return env -- Here c is dead, and we avoid creating
386 -- the binding c = (a,b)
387
388 | Coercion co <- new_rhs
389 = return (extendCvSubst env bndr co)
390
391 | otherwise
392 = do { (env', bndr') <- simplBinder env bndr
393 ; completeNonRecX NotTopLevel env' (isStrictId bndr) bndr bndr' new_rhs }
394 -- simplNonRecX is only used for NotTopLevel things
395
396 completeNonRecX :: TopLevelFlag -> SimplEnv
397 -> Bool
398 -> InId -- Old binder
399 -> OutId -- New binder
400 -> OutExpr -- Simplified RHS
401 -> SimplM SimplEnv
402 -- Precondition: rhs satisfies the let/app invariant
403 -- See Note [CoreSyn let/app invariant] in CoreSyn
404
405 completeNonRecX top_lvl env is_strict old_bndr new_bndr new_rhs
406 = do { (env1, rhs1) <- prepareRhs top_lvl (zapFloats env) new_bndr new_rhs
407 ; (env2, rhs2) <-
408 if doFloatFromRhs NotTopLevel NonRecursive is_strict rhs1 env1
409 then do { tick LetFloatFromLet
410 ; return (addFloats env env1, rhs1) } -- Add the floats to the main env
411 else return (env, wrapFloats env1 rhs1) -- Wrap the floats around the RHS
412 ; completeBind env2 NotTopLevel old_bndr new_bndr rhs2 }
413
414 {-
415 {- No, no, no! Do not try preInlineUnconditionally in completeNonRecX
416 Doing so risks exponential behaviour, because new_rhs has been simplified once already
417 In the cases described by the folowing commment, postInlineUnconditionally will
418 catch many of the relevant cases.
419 -- This happens; for example, the case_bndr during case of
420 -- known constructor: case (a,b) of x { (p,q) -> ... }
421 -- Here x isn't mentioned in the RHS, so we don't want to
422 -- create the (dead) let-binding let x = (a,b) in ...
423 --
424 -- Similarly, single occurrences can be inlined vigourously
425 -- e.g. case (f x, g y) of (a,b) -> ....
426 -- If a,b occur once we can avoid constructing the let binding for them.
427
428 Furthermore in the case-binding case preInlineUnconditionally risks extra thunks
429 -- Consider case I# (quotInt# x y) of
430 -- I# v -> let w = J# v in ...
431 -- If we gaily inline (quotInt# x y) for v, we end up building an
432 -- extra thunk:
433 -- let w = J# (quotInt# x y) in ...
434 -- because quotInt# can fail.
435
436 | preInlineUnconditionally env NotTopLevel bndr new_rhs
437 = thing_inside (extendIdSubst env bndr (DoneEx new_rhs))
438 -}
439
440 ----------------------------------
441 prepareRhs takes a putative RHS, checks whether it's a PAP or
442 constructor application and, if so, converts it to ANF, so that the
443 resulting thing can be inlined more easily. Thus
444 x = (f a, g b)
445 becomes
446 t1 = f a
447 t2 = g b
448 x = (t1,t2)
449
450 We also want to deal well cases like this
451 v = (f e1 `cast` co) e2
452 Here we want to make e1,e2 trivial and get
453 x1 = e1; x2 = e2; v = (f x1 `cast` co) v2
454 That's what the 'go' loop in prepareRhs does
455 -}
456
457 prepareRhs :: TopLevelFlag -> SimplEnv -> OutId -> OutExpr -> SimplM (SimplEnv, OutExpr)
458 -- Adds new floats to the env iff that allows us to return a good RHS
459 prepareRhs top_lvl env id (Cast rhs co) -- Note [Float coercions]
460 | Pair ty1 _ty2 <- coercionKind co -- Do *not* do this if rhs has an unlifted type
461 , not (isUnliftedType ty1) -- see Note [Float coercions (unlifted)]
462 = do { (env', rhs') <- makeTrivialWithInfo top_lvl env (getOccFS id) sanitised_info rhs
463 ; return (env', Cast rhs' co) }
464 where
465 sanitised_info = vanillaIdInfo `setStrictnessInfo` strictnessInfo info
466 `setDemandInfo` demandInfo info
467 info = idInfo id
468
469 prepareRhs top_lvl env0 id rhs0
470 = do { (_is_exp, env1, rhs1) <- go 0 env0 rhs0
471 ; return (env1, rhs1) }
472 where
473 go n_val_args env (Cast rhs co)
474 = do { (is_exp, env', rhs') <- go n_val_args env rhs
475 ; return (is_exp, env', Cast rhs' co) }
476 go n_val_args env (App fun (Type ty))
477 = do { (is_exp, env', rhs') <- go n_val_args env fun
478 ; return (is_exp, env', App rhs' (Type ty)) }
479 go n_val_args env (App fun arg)
480 = do { (is_exp, env', fun') <- go (n_val_args+1) env fun
481 ; case is_exp of
482 True -> do { (env'', arg') <- makeTrivial top_lvl env' (getOccFS id) arg
483 ; return (True, env'', App fun' arg') }
484 False -> return (False, env, App fun arg) }
485 go n_val_args env (Var fun)
486 = return (is_exp, env, Var fun)
487 where
488 is_exp = isExpandableApp fun n_val_args -- The fun a constructor or PAP
489 -- See Note [CONLIKE pragma] in BasicTypes
490 -- The definition of is_exp should match that in
491 -- OccurAnal.occAnalApp
492
493 go n_val_args env (Tick t rhs)
494 -- We want to be able to float bindings past this
495 -- tick. Non-scoping ticks don't care.
496 | tickishScoped t == NoScope
497 = do { (is_exp, env', rhs') <- go n_val_args env rhs
498 ; return (is_exp, env', Tick t rhs') }
499 -- On the other hand, for scoping ticks we need to be able to
500 -- copy them on the floats, which in turn is only allowed if
501 -- we can obtain non-counting ticks.
502 | not (tickishCounts t) || tickishCanSplit t
503 = do { (is_exp, env', rhs') <- go n_val_args (zapFloats env) rhs
504 ; let tickIt (id, expr) = (id, mkTick (mkNoCount t) expr)
505 floats' = seFloats $ env `addFloats` mapFloats env' tickIt
506 ; return (is_exp, env' { seFloats = floats' }, Tick t rhs') }
507
508 go _ env other
509 = return (False, env, other)
510
511 {-
512 Note [Float coercions]
513 ~~~~~~~~~~~~~~~~~~~~~~
514 When we find the binding
515 x = e `cast` co
516 we'd like to transform it to
517 x' = e
518 x = x `cast` co -- A trivial binding
519 There's a chance that e will be a constructor application or function, or something
520 like that, so moving the coercion to the usage site may well cancel the coercions
521 and lead to further optimisation. Example:
522
523 data family T a :: *
524 data instance T Int = T Int
525
526 foo :: Int -> Int -> Int
527 foo m n = ...
528 where
529 x = T m
530 go 0 = 0
531 go n = case x of { T m -> go (n-m) }
532 -- This case should optimise
533
534 Note [Preserve strictness when floating coercions]
535 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
536 In the Note [Float coercions] transformation, keep the strictness info.
537 Eg
538 f = e `cast` co -- f has strictness SSL
539 When we transform to
540 f' = e -- f' also has strictness SSL
541 f = f' `cast` co -- f still has strictness SSL
542
543 Its not wrong to drop it on the floor, but better to keep it.
544
545 Note [Float coercions (unlifted)]
546 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
547 BUT don't do [Float coercions] if 'e' has an unlifted type.
548 This *can* happen:
549
550 foo :: Int = (error (# Int,Int #) "urk")
551 `cast` CoUnsafe (# Int,Int #) Int
552
553 If do the makeTrivial thing to the error call, we'll get
554 foo = case error (# Int,Int #) "urk" of v -> v `cast` ...
555 But 'v' isn't in scope!
556
557 These strange casts can happen as a result of case-of-case
558 bar = case (case x of { T -> (# 2,3 #); F -> error "urk" }) of
559 (# p,q #) -> p+q
560 -}
561
562 makeTrivialArg :: SimplEnv -> ArgSpec -> SimplM (SimplEnv, ArgSpec)
563 makeTrivialArg env (ValArg e) = do
564 { (env', e') <- makeTrivial NotTopLevel env (fsLit "arg") e
565 ; return (env', ValArg e') }
566 makeTrivialArg env arg = return (env, arg) -- CastBy, TyArg
567
568 makeTrivial :: TopLevelFlag -> SimplEnv
569 -> FastString -- ^ a "friendly name" to build the new binder from
570 -> OutExpr -> SimplM (SimplEnv, OutExpr)
571 -- Binds the expression to a variable, if it's not trivial, returning the variable
572 makeTrivial top_lvl env context expr =
573 makeTrivialWithInfo top_lvl env context vanillaIdInfo expr
574
575 makeTrivialWithInfo :: TopLevelFlag -> SimplEnv
576 -> FastString
577 -- ^ a "friendly name" to build the new binder from
578 -> IdInfo -> OutExpr -> SimplM (SimplEnv, OutExpr)
579 -- Propagate strictness and demand info to the new binder
580 -- Note [Preserve strictness when floating coercions]
581 -- Returned SimplEnv has same substitution as incoming one
582 makeTrivialWithInfo top_lvl env context info expr
583 | exprIsTrivial expr -- Already trivial
584 || not (bindingOk top_lvl expr expr_ty) -- Cannot trivialise
585 -- See Note [Cannot trivialise]
586 = return (env, expr)
587 | otherwise -- See Note [Take care] below
588 = do { uniq <- getUniqueM
589 ; let name = mkSystemVarName uniq context
590 var = mkLocalIdOrCoVarWithInfo name expr_ty info
591 ; env' <- completeNonRecX top_lvl env False var var expr
592 ; expr' <- simplVar env' var
593 ; return (env', expr') }
594 -- The simplVar is needed because we're constructing a new binding
595 -- a = rhs
596 -- And if rhs is of form (rhs1 |> co), then we might get
597 -- a1 = rhs1
598 -- a = a1 |> co
599 -- and now a's RHS is trivial and can be substituted out, and that
600 -- is what completeNonRecX will do
601 -- To put it another way, it's as if we'd simplified
602 -- let var = e in var
603 where
604 expr_ty = exprType expr
605
606 bindingOk :: TopLevelFlag -> CoreExpr -> Type -> Bool
607 -- True iff we can have a binding of this expression at this level
608 -- Precondition: the type is the type of the expression
609 bindingOk top_lvl _ expr_ty
610 | isTopLevel top_lvl = not (isUnliftedType expr_ty)
611 | otherwise = True
612
613 {-
614 Note [Cannot trivialise]
615 ~~~~~~~~~~~~~~~~~~~~~~~~
616 Consider tih
617 f :: Int -> Addr#
618
619 foo :: Bar
620 foo = Bar (f 3)
621
622 Then we can't ANF-ise foo, even though we'd like to, because
623 we can't make a top-level binding for the Addr# (f 3). And if
624 so we don't want to turn it into
625 foo = let x = f 3 in Bar x
626 because we'll just end up inlining x back, and that makes the
627 simplifier loop. Better not to ANF-ise it at all.
628
629 A case in point is literal strings (a MachStr is not regarded as
630 trivial):
631
632 foo = Ptr "blob"#
633
634 We don't want to ANF-ise this.
635
636 ************************************************************************
637 * *
638 \subsection{Completing a lazy binding}
639 * *
640 ************************************************************************
641
642 completeBind
643 * deals only with Ids, not TyVars
644 * takes an already-simplified binder and RHS
645 * is used for both recursive and non-recursive bindings
646 * is used for both top-level and non-top-level bindings
647
648 It does the following:
649 - tries discarding a dead binding
650 - tries PostInlineUnconditionally
651 - add unfolding [this is the only place we add an unfolding]
652 - add arity
653
654 It does *not* attempt to do let-to-case. Why? Because it is used for
655 - top-level bindings (when let-to-case is impossible)
656 - many situations where the "rhs" is known to be a WHNF
657 (so let-to-case is inappropriate).
658
659 Nor does it do the atomic-argument thing
660 -}
661
662 completeBind :: SimplEnv
663 -> TopLevelFlag -- Flag stuck into unfolding
664 -> InId -- Old binder
665 -> OutId -> OutExpr -- New binder and RHS
666 -> SimplM SimplEnv
667 -- completeBind may choose to do its work
668 -- * by extending the substitution (e.g. let x = y in ...)
669 -- * or by adding to the floats in the envt
670 --
671 -- Precondition: rhs obeys the let/app invariant
672 completeBind env top_lvl old_bndr new_bndr new_rhs
673 | isCoVar old_bndr
674 = case new_rhs of
675 Coercion co -> return (extendCvSubst env old_bndr co)
676 _ -> return (addNonRec env new_bndr new_rhs)
677
678 | otherwise
679 = ASSERT( isId new_bndr )
680 do { let old_info = idInfo old_bndr
681 old_unf = unfoldingInfo old_info
682 occ_info = occInfo old_info
683
684 -- Do eta-expansion on the RHS of the binding
685 -- See Note [Eta-expanding at let bindings] in SimplUtils
686 ; (new_arity, final_rhs) <- tryEtaExpandRhs env new_bndr new_rhs
687
688 -- Simplify the unfolding
689 ; new_unfolding <- simplLetUnfolding env top_lvl old_bndr final_rhs old_unf
690
691 ; dflags <- getDynFlags
692 ; if postInlineUnconditionally dflags env top_lvl new_bndr occ_info
693 final_rhs new_unfolding
694
695 -- Inline and discard the binding
696 then do { tick (PostInlineUnconditionally old_bndr)
697 ; return (extendIdSubst env old_bndr (DoneEx final_rhs)) }
698 -- Use the substitution to make quite, quite sure that the
699 -- substitution will happen, since we are going to discard the binding
700 else
701 do { let info1 = idInfo new_bndr `setArityInfo` new_arity
702
703 -- Unfolding info: Note [Setting the new unfolding]
704 info2 = info1 `setUnfoldingInfo` new_unfolding
705
706 -- Demand info: Note [Setting the demand info]
707 --
708 -- We also have to nuke demand info if for some reason
709 -- eta-expansion *reduces* the arity of the binding to less
710 -- than that of the strictness sig. This can happen: see Note [Arity decrease].
711 info3 | isEvaldUnfolding new_unfolding
712 || (case strictnessInfo info2 of
713 StrictSig dmd_ty -> new_arity < dmdTypeDepth dmd_ty)
714 = zapDemandInfo info2 `orElse` info2
715 | otherwise
716 = info2
717
718 final_id = new_bndr `setIdInfo` info3
719
720 ; -- pprTrace "Binding" (ppr final_id <+> ppr new_unfolding) $
721 return (addNonRec env final_id final_rhs) } }
722 -- The addNonRec adds it to the in-scope set too
723
724 ------------------------------
725 addPolyBind :: TopLevelFlag -> SimplEnv -> OutBind -> SimplM SimplEnv
726 -- Add a new binding to the environment, complete with its unfolding
727 -- but *do not* do postInlineUnconditionally, because we have already
728 -- processed some of the scope of the binding
729 -- We still want the unfolding though. Consider
730 -- let
731 -- x = /\a. let y = ... in Just y
732 -- in body
733 -- Then we float the y-binding out (via abstractFloats and addPolyBind)
734 -- but 'x' may well then be inlined in 'body' in which case we'd like the
735 -- opportunity to inline 'y' too.
736 --
737 -- INVARIANT: the arity is correct on the incoming binders
738
739 addPolyBind top_lvl env (NonRec poly_id rhs)
740 = do { unfolding <- simplLetUnfolding env top_lvl poly_id rhs noUnfolding
741 -- Assumes that poly_id did not have an INLINE prag
742 -- which is perhaps wrong. ToDo: think about this
743 ; let final_id = setIdInfo poly_id $
744 idInfo poly_id `setUnfoldingInfo` unfolding
745
746 ; return (addNonRec env final_id rhs) }
747
748 addPolyBind _ env bind@(Rec _)
749 = return (extendFloats env bind)
750 -- Hack: letrecs are more awkward, so we extend "by steam"
751 -- without adding unfoldings etc. At worst this leads to
752 -- more simplifier iterations
753
754 {- Note [Arity decrease]
755 ~~~~~~~~~~~~~~~~~~~~~~~~
756 Generally speaking the arity of a binding should not decrease. But it *can*
757 legitimately happen because of RULES. Eg
758 f = g Int
759 where g has arity 2, will have arity 2. But if there's a rewrite rule
760 g Int --> h
761 where h has arity 1, then f's arity will decrease. Here's a real-life example,
762 which is in the output of Specialise:
763
764 Rec {
765 $dm {Arity 2} = \d.\x. op d
766 {-# RULES forall d. $dm Int d = $s$dm #-}
767
768 dInt = MkD .... opInt ...
769 opInt {Arity 1} = $dm dInt
770
771 $s$dm {Arity 0} = \x. op dInt }
772
773 Here opInt has arity 1; but when we apply the rule its arity drops to 0.
774 That's why Specialise goes to a little trouble to pin the right arity
775 on specialised functions too.
776
777 Note [Setting the demand info]
778 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
779 If the unfolding is a value, the demand info may
780 go pear-shaped, so we nuke it. Example:
781 let x = (a,b) in
782 case x of (p,q) -> h p q x
783 Here x is certainly demanded. But after we've nuked
784 the case, we'll get just
785 let x = (a,b) in h a b x
786 and now x is not demanded (I'm assuming h is lazy)
787 This really happens. Similarly
788 let f = \x -> e in ...f..f...
789 After inlining f at some of its call sites the original binding may
790 (for example) be no longer strictly demanded.
791 The solution here is a bit ad hoc...
792
793
794 ************************************************************************
795 * *
796 \subsection[Simplify-simplExpr]{The main function: simplExpr}
797 * *
798 ************************************************************************
799
800 The reason for this OutExprStuff stuff is that we want to float *after*
801 simplifying a RHS, not before. If we do so naively we get quadratic
802 behaviour as things float out.
803
804 To see why it's important to do it after, consider this (real) example:
805
806 let t = f x
807 in fst t
808 ==>
809 let t = let a = e1
810 b = e2
811 in (a,b)
812 in fst t
813 ==>
814 let a = e1
815 b = e2
816 t = (a,b)
817 in
818 a -- Can't inline a this round, cos it appears twice
819 ==>
820 e1
821
822 Each of the ==> steps is a round of simplification. We'd save a
823 whole round if we float first. This can cascade. Consider
824
825 let f = g d
826 in \x -> ...f...
827 ==>
828 let f = let d1 = ..d.. in \y -> e
829 in \x -> ...f...
830 ==>
831 let d1 = ..d..
832 in \x -> ...(\y ->e)...
833
834 Only in this second round can the \y be applied, and it
835 might do the same again.
836 -}
837
838 simplExpr :: SimplEnv -> CoreExpr -> SimplM CoreExpr
839 simplExpr env expr = simplExprC env expr (mkBoringStop expr_out_ty)
840 where
841 expr_out_ty :: OutType
842 expr_out_ty = substTy env (exprType expr)
843
844 simplExprC :: SimplEnv -> CoreExpr -> SimplCont -> SimplM CoreExpr
845 -- Simplify an expression, given a continuation
846 simplExprC env expr cont
847 = -- pprTrace "simplExprC" (ppr expr $$ ppr cont {- $$ ppr (seIdSubst env) -} $$ ppr (seFloats env) ) $
848 do { (env', expr') <- simplExprF (zapFloats env) expr cont
849 ; -- pprTrace "simplExprC ret" (ppr expr $$ ppr expr') $
850 -- pprTrace "simplExprC ret3" (ppr (seInScope env')) $
851 -- pprTrace "simplExprC ret4" (ppr (seFloats env')) $
852 return (wrapFloats env' expr') }
853
854 --------------------------------------------------
855 simplExprF :: SimplEnv -> InExpr -> SimplCont
856 -> SimplM (SimplEnv, OutExpr)
857
858 simplExprF env e cont
859 = {- pprTrace "simplExprF" (vcat
860 [ ppr e
861 , text "cont =" <+> ppr cont
862 , text "inscope =" <+> ppr (seInScope env)
863 , text "tvsubst =" <+> ppr (seTvSubst env)
864 , text "idsubst =" <+> ppr (seIdSubst env)
865 , text "cvsubst =" <+> ppr (seCvSubst env)
866 {- , ppr (seFloats env) -}
867 ]) $ -}
868 simplExprF1 env e cont
869
870 simplExprF1 :: SimplEnv -> InExpr -> SimplCont
871 -> SimplM (SimplEnv, OutExpr)
872 simplExprF1 env (Var v) cont = simplIdF env v cont
873 simplExprF1 env (Lit lit) cont = rebuild env (Lit lit) cont
874 simplExprF1 env (Tick t expr) cont = simplTick env t expr cont
875 simplExprF1 env (Cast body co) cont = simplCast env body co cont
876 simplExprF1 env (Coercion co) cont = simplCoercionF env co cont
877 simplExprF1 env (Type ty) cont = ASSERT( contIsRhsOrArg cont )
878 rebuild env (Type (substTy env ty)) cont
879
880 simplExprF1 env (App fun arg) cont
881 = simplExprF env fun $
882 case arg of
883 Type ty -> ApplyToTy { sc_arg_ty = substTy env ty
884 , sc_hole_ty = substTy env (exprType fun)
885 , sc_cont = cont }
886 _ -> ApplyToVal { sc_arg = arg, sc_env = env
887 , sc_dup = NoDup, sc_cont = cont }
888
889 simplExprF1 env expr@(Lam {}) cont
890 = simplLam env zapped_bndrs body cont
891 -- The main issue here is under-saturated lambdas
892 -- (\x1. \x2. e) arg1
893 -- Here x1 might have "occurs-once" occ-info, because occ-info
894 -- is computed assuming that a group of lambdas is applied
895 -- all at once. If there are too few args, we must zap the
896 -- occ-info, UNLESS the remaining binders are one-shot
897 where
898 (bndrs, body) = collectBinders expr
899 zapped_bndrs | need_to_zap = map zap bndrs
900 | otherwise = bndrs
901
902 need_to_zap = any zappable_bndr (drop n_args bndrs)
903 n_args = countArgs cont
904 -- NB: countArgs counts all the args (incl type args)
905 -- and likewise drop counts all binders (incl type lambdas)
906
907 zappable_bndr b = isId b && not (isOneShotBndr b)
908 zap b | isTyVar b = b
909 | otherwise = zapLamIdInfo b
910
911 simplExprF1 env (Case scrut bndr _ alts) cont
912 = simplExprF env scrut (Select { sc_dup = NoDup, sc_bndr = bndr
913 , sc_alts = alts
914 , sc_env = env, sc_cont = cont })
915
916 simplExprF1 env (Let (Rec pairs) body) cont
917 = do { env' <- simplRecBndrs env (map fst pairs)
918 -- NB: bndrs' don't have unfoldings or rules
919 -- We add them as we go down
920
921 ; env'' <- simplRecBind env' NotTopLevel pairs
922 ; simplExprF env'' body cont }
923
924 simplExprF1 env (Let (NonRec bndr rhs) body) cont
925 = simplNonRecE env bndr (rhs, env) ([], body) cont
926
927 ---------------------------------
928 simplType :: SimplEnv -> InType -> SimplM OutType
929 -- Kept monadic just so we can do the seqType
930 simplType env ty
931 = -- pprTrace "simplType" (ppr ty $$ ppr (seTvSubst env)) $
932 seqType new_ty `seq` return new_ty
933 where
934 new_ty = substTy env ty
935
936 ---------------------------------
937 simplCoercionF :: SimplEnv -> InCoercion -> SimplCont
938 -> SimplM (SimplEnv, OutExpr)
939 simplCoercionF env co cont
940 = do { co' <- simplCoercion env co
941 ; rebuild env (Coercion co') cont }
942
943 simplCoercion :: SimplEnv -> InCoercion -> SimplM OutCoercion
944 simplCoercion env co
945 = let opt_co = optCoercion (getTCvSubst env) co
946 in seqCo opt_co `seq` return opt_co
947
948 -----------------------------------
949 -- | Push a TickIt context outwards past applications and cases, as
950 -- long as this is a non-scoping tick, to let case and application
951 -- optimisations apply.
952
953 simplTick :: SimplEnv -> Tickish Id -> InExpr -> SimplCont
954 -> SimplM (SimplEnv, OutExpr)
955 simplTick env tickish expr cont
956 -- A scoped tick turns into a continuation, so that we can spot
957 -- (scc t (\x . e)) in simplLam and eliminate the scc. If we didn't do
958 -- it this way, then it would take two passes of the simplifier to
959 -- reduce ((scc t (\x . e)) e').
960 -- NB, don't do this with counting ticks, because if the expr is
961 -- bottom, then rebuildCall will discard the continuation.
962
963 -- XXX: we cannot do this, because the simplifier assumes that
964 -- the context can be pushed into a case with a single branch. e.g.
965 -- scc<f> case expensive of p -> e
966 -- becomes
967 -- case expensive of p -> scc<f> e
968 --
969 -- So I'm disabling this for now. It just means we will do more
970 -- simplifier iterations that necessary in some cases.
971
972 -- | tickishScoped tickish && not (tickishCounts tickish)
973 -- = simplExprF env expr (TickIt tickish cont)
974
975 -- For unscoped or soft-scoped ticks, we are allowed to float in new
976 -- cost, so we simply push the continuation inside the tick. This
977 -- has the effect of moving the tick to the outside of a case or
978 -- application context, allowing the normal case and application
979 -- optimisations to fire.
980 | tickish `tickishScopesLike` SoftScope
981 = do { (env', expr') <- simplExprF env expr cont
982 ; return (env', mkTick tickish expr')
983 }
984
985 -- Push tick inside if the context looks like this will allow us to
986 -- do a case-of-case - see Note [case-of-scc-of-case]
987 | Select {} <- cont, Just expr' <- push_tick_inside
988 = simplExprF env expr' cont
989
990 -- We don't want to move the tick, but we might still want to allow
991 -- floats to pass through with appropriate wrapping (or not, see
992 -- wrap_floats below)
993 --- | not (tickishCounts tickish) || tickishCanSplit tickish
994 -- = wrap_floats
995
996 | otherwise
997 = no_floating_past_tick
998
999 where
1000
1001 -- Try to push tick inside a case, see Note [case-of-scc-of-case].
1002 push_tick_inside =
1003 case expr0 of
1004 Case scrut bndr ty alts
1005 -> Just $ Case (tickScrut scrut) bndr ty (map tickAlt alts)
1006 _other -> Nothing
1007 where (ticks, expr0) = stripTicksTop movable (Tick tickish expr)
1008 movable t = not (tickishCounts t) ||
1009 t `tickishScopesLike` NoScope ||
1010 tickishCanSplit t
1011 tickScrut e = foldr mkTick e ticks
1012 -- Alternatives get annotated with all ticks that scope in some way,
1013 -- but we don't want to count entries.
1014 tickAlt (c,bs,e) = (c,bs, foldr mkTick e ts_scope)
1015 ts_scope = map mkNoCount $
1016 filter (not . (`tickishScopesLike` NoScope)) ticks
1017
1018 no_floating_past_tick =
1019 do { let (inc,outc) = splitCont cont
1020 ; (env', expr') <- simplExprF (zapFloats env) expr inc
1021 ; let tickish' = simplTickish env tickish
1022 ; (env'', expr'') <- rebuild (zapFloats env')
1023 (wrapFloats env' expr')
1024 (TickIt tickish' outc)
1025 ; return (addFloats env env'', expr'')
1026 }
1027
1028 -- Alternative version that wraps outgoing floats with the tick. This
1029 -- results in ticks being duplicated, as we don't make any attempt to
1030 -- eliminate the tick if we re-inline the binding (because the tick
1031 -- semantics allows unrestricted inlining of HNFs), so I'm not doing
1032 -- this any more. FloatOut will catch any real opportunities for
1033 -- floating.
1034 --
1035 -- wrap_floats =
1036 -- do { let (inc,outc) = splitCont cont
1037 -- ; (env', expr') <- simplExprF (zapFloats env) expr inc
1038 -- ; let tickish' = simplTickish env tickish
1039 -- ; let wrap_float (b,rhs) = (zapIdStrictness (setIdArity b 0),
1040 -- mkTick (mkNoCount tickish') rhs)
1041 -- -- when wrapping a float with mkTick, we better zap the Id's
1042 -- -- strictness info and arity, because it might be wrong now.
1043 -- ; let env'' = addFloats env (mapFloats env' wrap_float)
1044 -- ; rebuild env'' expr' (TickIt tickish' outc)
1045 -- }
1046
1047
1048 simplTickish env tickish
1049 | Breakpoint n ids <- tickish
1050 = Breakpoint n (map (getDoneId . substId env) ids)
1051 | otherwise = tickish
1052
1053 -- Push type application and coercion inside a tick
1054 splitCont :: SimplCont -> (SimplCont, SimplCont)
1055 splitCont cont@(ApplyToTy { sc_cont = tail }) = (cont { sc_cont = inc }, outc)
1056 where (inc,outc) = splitCont tail
1057 splitCont (CastIt co c) = (CastIt co inc, outc)
1058 where (inc,outc) = splitCont c
1059 splitCont other = (mkBoringStop (contHoleType other), other)
1060
1061 getDoneId (DoneId id) = id
1062 getDoneId (DoneEx e) = getIdFromTrivialExpr e -- Note [substTickish] in CoreSubst
1063 getDoneId other = pprPanic "getDoneId" (ppr other)
1064
1065 -- Note [case-of-scc-of-case]
1066 -- It's pretty important to be able to transform case-of-case when
1067 -- there's an SCC in the way. For example, the following comes up
1068 -- in nofib/real/compress/Encode.hs:
1069 --
1070 -- case scctick<code_string.r1>
1071 -- case $wcode_string_r13s wild_XC w1_s137 w2_s138 l_aje
1072 -- of _ { (# ww1_s13f, ww2_s13g, ww3_s13h #) ->
1073 -- (ww1_s13f, ww2_s13g, ww3_s13h)
1074 -- }
1075 -- of _ { (ww_s12Y, ww1_s12Z, ww2_s130) ->
1076 -- tick<code_string.f1>
1077 -- (ww_s12Y,
1078 -- ww1_s12Z,
1079 -- PTTrees.PT
1080 -- @ GHC.Types.Char @ GHC.Types.Int wild2_Xj ww2_s130 r_ajf)
1081 -- }
1082 --
1083 -- We really want this case-of-case to fire, because then the 3-tuple
1084 -- will go away (indeed, the CPR optimisation is relying on this
1085 -- happening). But the scctick is in the way - we need to push it
1086 -- inside to expose the case-of-case. So we perform this
1087 -- transformation on the inner case:
1088 --
1089 -- scctick c (case e of { p1 -> e1; ...; pn -> en })
1090 -- ==>
1091 -- case (scctick c e) of { p1 -> scc c e1; ...; pn -> scc c en }
1092 --
1093 -- So we've moved a constant amount of work out of the scc to expose
1094 -- the case. We only do this when the continuation is interesting: in
1095 -- for now, it has to be another Case (maybe generalise this later).
1096
1097 {-
1098 ************************************************************************
1099 * *
1100 \subsection{The main rebuilder}
1101 * *
1102 ************************************************************************
1103 -}
1104
1105 rebuild :: SimplEnv -> OutExpr -> SimplCont -> SimplM (SimplEnv, OutExpr)
1106 -- At this point the substitution in the SimplEnv should be irrelevant
1107 -- only the in-scope set and floats should matter
1108 rebuild env expr cont
1109 = case cont of
1110 Stop {} -> return (env, expr)
1111 TickIt t cont -> rebuild env (mkTick t expr) cont
1112 CastIt co cont -> rebuild env (mkCast expr co) cont
1113 -- NB: mkCast implements the (Coercion co |> g) optimisation
1114
1115 Select { sc_bndr = bndr, sc_alts = alts, sc_env = se, sc_cont = cont }
1116 -> rebuildCase (se `setFloats` env) expr bndr alts cont
1117
1118 StrictArg info _ cont -> rebuildCall env (info `addValArgTo` expr) cont
1119 StrictBind b bs body se cont -> do { env' <- simplNonRecX (se `setFloats` env) b expr
1120 -- expr satisfies let/app since it started life
1121 -- in a call to simplNonRecE
1122 ; simplLam env' bs body cont }
1123
1124 ApplyToTy { sc_arg_ty = ty, sc_cont = cont}
1125 -> rebuild env (App expr (Type ty)) cont
1126 ApplyToVal { sc_arg = arg, sc_env = se, sc_dup = dup_flag, sc_cont = cont}
1127 -- See Note [Avoid redundant simplification]
1128 | isSimplified dup_flag -> rebuild env (App expr arg) cont
1129 | otherwise -> do { arg' <- simplExpr (se `setInScope` env) arg
1130 ; rebuild env (App expr arg') cont }
1131
1132
1133 {-
1134 ************************************************************************
1135 * *
1136 \subsection{Lambdas}
1137 * *
1138 ************************************************************************
1139 -}
1140
1141 simplCast :: SimplEnv -> InExpr -> Coercion -> SimplCont
1142 -> SimplM (SimplEnv, OutExpr)
1143 simplCast env body co0 cont0
1144 = do { co1 <- simplCoercion env co0
1145 ; cont1 <- addCoerce co1 cont0
1146 ; simplExprF env body cont1 }
1147 where
1148 addCoerce :: OutCoercion -> SimplCont -> SimplM SimplCont
1149 addCoerce co1 (CastIt co2 cont)
1150 = addCoerce (mkTransCo co1 co2) cont
1151
1152 addCoerce co cont@(ApplyToTy { sc_arg_ty = arg_ty, sc_cont = tail })
1153 | Just (arg_ty', co') <- pushCoTyArg co arg_ty
1154 = do { tail' <- addCoerce co' tail
1155 ; return (cont { sc_arg_ty = arg_ty', sc_cont = tail' }) }
1156
1157 addCoerce co (ApplyToVal { sc_arg = arg, sc_env = arg_se
1158 , sc_dup = dup, sc_cont = tail })
1159 | Just (co1, co2) <- pushCoValArg co
1160 = do { (dup', arg_se', arg') <- simplArg env dup arg_se arg
1161 -- When we build the ApplyTo we can't mix the OutCoercion
1162 -- 'co' with the InExpr 'arg', so we simplify
1163 -- to make it all consistent. It's a bit messy.
1164 -- But it isn't a common case.
1165 -- Example of use: Trac #995
1166 ; tail' <- addCoerce co2 tail
1167 ; return (ApplyToVal { sc_arg = mkCast arg' co1
1168 , sc_env = arg_se'
1169 , sc_dup = dup'
1170 , sc_cont = tail' }) }
1171
1172 addCoerce co cont
1173 | isReflexiveCo co = return cont
1174 | otherwise = return (CastIt co cont)
1175 -- It's worth checking isReflexiveCo.
1176 -- For example, in the initial form of a worker
1177 -- we may find (coerce T (coerce S (\x.e))) y
1178 -- and we'd like it to simplify to e[y/x] in one round
1179 -- of simplification
1180
1181 simplArg :: SimplEnv -> DupFlag -> StaticEnv -> CoreExpr
1182 -> SimplM (DupFlag, StaticEnv, OutExpr)
1183 simplArg env dup_flag arg_env arg
1184 | isSimplified dup_flag
1185 = return (dup_flag, arg_env, arg)
1186 | otherwise
1187 = do { arg' <- simplExpr (arg_env `setInScope` env) arg
1188 ; return (Simplified, zapSubstEnv arg_env, arg') }
1189
1190 {-
1191 ************************************************************************
1192 * *
1193 \subsection{Lambdas}
1194 * *
1195 ************************************************************************
1196
1197 Note [Zap unfolding when beta-reducing]
1198 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1199 Lambda-bound variables can have stable unfoldings, such as
1200 $j = \x. \b{Unf=Just x}. e
1201 See Note [Case binders and join points] below; the unfolding for lets
1202 us optimise e better. However when we beta-reduce it we want to
1203 revert to using the actual value, otherwise we can end up in the
1204 stupid situation of
1205 let x = blah in
1206 let b{Unf=Just x} = y
1207 in ...b...
1208 Here it'd be far better to drop the unfolding and use the actual RHS.
1209 -}
1210
1211 simplLam :: SimplEnv -> [InId] -> InExpr -> SimplCont
1212 -> SimplM (SimplEnv, OutExpr)
1213
1214 simplLam env [] body cont = simplExprF env body cont
1215
1216 -- Beta reduction
1217
1218 simplLam env (bndr:bndrs) body (ApplyToTy { sc_arg_ty = arg_ty, sc_cont = cont })
1219 = do { tick (BetaReduction bndr)
1220 ; simplLam (extendTvSubst env bndr arg_ty) bndrs body cont }
1221
1222 simplLam env (bndr:bndrs) body (ApplyToVal { sc_arg = arg, sc_env = arg_se
1223 , sc_cont = cont })
1224 = do { tick (BetaReduction bndr)
1225 ; simplNonRecE env' (zap_unfolding bndr) (arg, arg_se) (bndrs, body) cont }
1226 where
1227 env' | Coercion co <- arg
1228 = extendCvSubst env bndr co
1229 | otherwise
1230 = env
1231
1232 zap_unfolding bndr -- See Note [Zap unfolding when beta-reducing]
1233 | isId bndr, isStableUnfolding (realIdUnfolding bndr)
1234 = setIdUnfolding bndr NoUnfolding
1235 | otherwise = bndr
1236
1237 -- discard a non-counting tick on a lambda. This may change the
1238 -- cost attribution slightly (moving the allocation of the
1239 -- lambda elsewhere), but we don't care: optimisation changes
1240 -- cost attribution all the time.
1241 simplLam env bndrs body (TickIt tickish cont)
1242 | not (tickishCounts tickish)
1243 = simplLam env bndrs body cont
1244
1245 -- Not enough args, so there are real lambdas left to put in the result
1246 simplLam env bndrs body cont
1247 = do { (env', bndrs') <- simplLamBndrs env bndrs
1248 ; body' <- simplExpr env' body
1249 ; new_lam <- mkLam env bndrs' body' cont
1250 ; rebuild env' new_lam cont }
1251
1252 simplLamBndrs :: SimplEnv -> [InBndr] -> SimplM (SimplEnv, [OutBndr])
1253 simplLamBndrs env bndrs = mapAccumLM simplLamBndr env bndrs
1254
1255 -------------
1256 simplLamBndr :: SimplEnv -> Var -> SimplM (SimplEnv, Var)
1257 -- Used for lambda binders. These sometimes have unfoldings added by
1258 -- the worker/wrapper pass that must be preserved, because they can't
1259 -- be reconstructed from context. For example:
1260 -- f x = case x of (a,b) -> fw a b x
1261 -- fw a b x{=(a,b)} = ...
1262 -- The "{=(a,b)}" is an unfolding we can't reconstruct otherwise.
1263 simplLamBndr env bndr
1264 | isId bndr && hasSomeUnfolding old_unf -- Special case
1265 = do { (env1, bndr1) <- simplBinder env bndr
1266 ; unf' <- simplUnfolding env1 NotTopLevel bndr old_unf
1267 ; let bndr2 = bndr1 `setIdUnfolding` unf'
1268 ; return (modifyInScope env1 bndr2, bndr2) }
1269
1270 | otherwise
1271 = simplBinder env bndr -- Normal case
1272 where
1273 old_unf = idUnfolding bndr
1274
1275 ------------------
1276 simplNonRecE :: SimplEnv
1277 -> InBndr -- The binder
1278 -> (InExpr, SimplEnv) -- Rhs of binding (or arg of lambda)
1279 -> ([InBndr], InExpr) -- Body of the let/lambda
1280 -- \xs.e
1281 -> SimplCont
1282 -> SimplM (SimplEnv, OutExpr)
1283
1284 -- simplNonRecE is used for
1285 -- * non-top-level non-recursive lets in expressions
1286 -- * beta reduction
1287 --
1288 -- It deals with strict bindings, via the StrictBind continuation,
1289 -- which may abort the whole process
1290 --
1291 -- Precondition: rhs satisfies the let/app invariant
1292 -- Note [CoreSyn let/app invariant] in CoreSyn
1293 --
1294 -- The "body" of the binding comes as a pair of ([InId],InExpr)
1295 -- representing a lambda; so we recurse back to simplLam
1296 -- Why? Because of the binder-occ-info-zapping done before
1297 -- the call to simplLam in simplExprF (Lam ...)
1298
1299 -- First deal with type applications and type lets
1300 -- (/\a. e) (Type ty) and (let a = Type ty in e)
1301 simplNonRecE env bndr (Type ty_arg, rhs_se) (bndrs, body) cont
1302 = ASSERT( isTyVar bndr )
1303 do { ty_arg' <- simplType (rhs_se `setInScope` env) ty_arg
1304 ; simplLam (extendTvSubst env bndr ty_arg') bndrs body cont }
1305
1306 simplNonRecE env bndr (rhs, rhs_se) (bndrs, body) cont
1307 = do dflags <- getDynFlags
1308 case () of
1309 _ | preInlineUnconditionally dflags env NotTopLevel bndr rhs
1310 -> do { tick (PreInlineUnconditionally bndr)
1311 ; -- pprTrace "preInlineUncond" (ppr bndr <+> ppr rhs) $
1312 simplLam (extendIdSubst env bndr (mkContEx rhs_se rhs)) bndrs body cont }
1313
1314 | isStrictId bndr -- Includes coercions
1315 -> simplExprF (rhs_se `setFloats` env) rhs
1316 (StrictBind bndr bndrs body env cont)
1317
1318 | otherwise
1319 -> ASSERT( not (isTyVar bndr) )
1320 do { (env1, bndr1) <- simplNonRecBndr env bndr
1321 ; (env2, bndr2) <- addBndrRules env1 bndr bndr1
1322 ; env3 <- simplLazyBind env2 NotTopLevel NonRecursive bndr bndr2 rhs rhs_se
1323 ; simplLam env3 bndrs body cont }
1324
1325 {-
1326 ************************************************************************
1327 * *
1328 Variables
1329 * *
1330 ************************************************************************
1331 -}
1332
1333 simplVar :: SimplEnv -> InVar -> SimplM OutExpr
1334 -- Look up an InVar in the environment
1335 simplVar env var
1336 | isTyVar var = return (Type (substTyVar env var))
1337 | isCoVar var = return (Coercion (substCoVar env var))
1338 | otherwise
1339 = case substId env var of
1340 DoneId var1 -> return (Var var1)
1341 DoneEx e -> return e
1342 ContEx tvs cvs ids e -> simplExpr (setSubstEnv env tvs cvs ids) e
1343
1344 simplIdF :: SimplEnv -> InId -> SimplCont -> SimplM (SimplEnv, OutExpr)
1345 simplIdF env var cont
1346 = case substId env var of
1347 DoneEx e -> simplExprF (zapSubstEnv env) e cont
1348 ContEx tvs cvs ids e -> simplExprF (setSubstEnv env tvs cvs ids) e cont
1349 DoneId var1 -> completeCall env var1 cont
1350 -- Note [zapSubstEnv]
1351 -- The template is already simplified, so don't re-substitute.
1352 -- This is VITAL. Consider
1353 -- let x = e in
1354 -- let y = \z -> ...x... in
1355 -- \ x -> ...y...
1356 -- We'll clone the inner \x, adding x->x' in the id_subst
1357 -- Then when we inline y, we must *not* replace x by x' in
1358 -- the inlined copy!!
1359
1360 ---------------------------------------------------------
1361 -- Dealing with a call site
1362
1363 completeCall :: SimplEnv -> OutId -> SimplCont -> SimplM (SimplEnv, OutExpr)
1364 completeCall env var cont
1365 = do { ------------- Try inlining ----------------
1366 dflags <- getDynFlags
1367 ; let (lone_variable, arg_infos, call_cont) = contArgs cont
1368 n_val_args = length arg_infos
1369 interesting_cont = interestingCallContext call_cont
1370 unfolding = activeUnfolding env var
1371 maybe_inline = callSiteInline dflags var unfolding
1372 lone_variable arg_infos interesting_cont
1373 ; case maybe_inline of {
1374 Just expr -- There is an inlining!
1375 -> do { checkedTick (UnfoldingDone var)
1376 ; dump_inline dflags expr cont
1377 ; simplExprF (zapSubstEnv env) expr cont }
1378
1379 ; Nothing -> do -- No inlining!
1380
1381 { rule_base <- getSimplRules
1382 ; let info = mkArgInfo var (getRules rule_base var) n_val_args call_cont
1383 ; rebuildCall env info cont
1384 }}}
1385 where
1386 dump_inline dflags unfolding cont
1387 | not (dopt Opt_D_dump_inlinings dflags) = return ()
1388 | not (dopt Opt_D_verbose_core2core dflags)
1389 = when (isExternalName (idName var)) $
1390 liftIO $ printOutputForUser dflags alwaysQualify $
1391 sep [text "Inlining done:", nest 4 (ppr var)]
1392 | otherwise
1393 = liftIO $ printOutputForUser dflags alwaysQualify $
1394 sep [text "Inlining done: " <> ppr var,
1395 nest 4 (vcat [text "Inlined fn: " <+> nest 2 (ppr unfolding),
1396 text "Cont: " <+> ppr cont])]
1397
1398 rebuildCall :: SimplEnv
1399 -> ArgInfo
1400 -> SimplCont
1401 -> SimplM (SimplEnv, OutExpr)
1402 rebuildCall env (ArgInfo { ai_fun = fun, ai_args = rev_args, ai_strs = [] }) cont
1403 -- When we run out of strictness args, it means
1404 -- that the call is definitely bottom; see SimplUtils.mkArgInfo
1405 -- Then we want to discard the entire strict continuation. E.g.
1406 -- * case (error "hello") of { ... }
1407 -- * (error "Hello") arg
1408 -- * f (error "Hello") where f is strict
1409 -- etc
1410 -- Then, especially in the first of these cases, we'd like to discard
1411 -- the continuation, leaving just the bottoming expression. But the
1412 -- type might not be right, so we may have to add a coerce.
1413 | not (contIsTrivial cont) -- Only do this if there is a non-trivial
1414 = return (env, castBottomExpr res cont_ty) -- continuation to discard, else we do it
1415 where -- again and again!
1416 res = argInfoExpr fun rev_args
1417 cont_ty = contResultType cont
1418
1419 rebuildCall env info (CastIt co cont)
1420 = rebuildCall env (addCastTo info co) cont
1421
1422 rebuildCall env info (ApplyToTy { sc_arg_ty = arg_ty, sc_cont = cont })
1423 = rebuildCall env (info `addTyArgTo` arg_ty) cont
1424
1425 rebuildCall env info@(ArgInfo { ai_encl = encl_rules, ai_type = fun_ty
1426 , ai_strs = str:strs, ai_discs = disc:discs })
1427 (ApplyToVal { sc_arg = arg, sc_env = arg_se
1428 , sc_dup = dup_flag, sc_cont = cont })
1429 | isSimplified dup_flag -- See Note [Avoid redundant simplification]
1430 = rebuildCall env (addValArgTo info' arg) cont
1431
1432 | str -- Strict argument
1433 = -- pprTrace "Strict Arg" (ppr arg $$ ppr (seIdSubst env) $$ ppr (seInScope env)) $
1434 simplExprF (arg_se `setFloats` env) arg
1435 (StrictArg info' cci cont)
1436 -- Note [Shadowing]
1437
1438 | otherwise -- Lazy argument
1439 -- DO NOT float anything outside, hence simplExprC
1440 -- There is no benefit (unlike in a let-binding), and we'd
1441 -- have to be very careful about bogus strictness through
1442 -- floating a demanded let.
1443 = do { arg' <- simplExprC (arg_se `setInScope` env) arg
1444 (mkLazyArgStop (funArgTy fun_ty) cci)
1445 ; rebuildCall env (addValArgTo info' arg') cont }
1446 where
1447 info' = info { ai_strs = strs, ai_discs = discs }
1448 cci | encl_rules = RuleArgCtxt
1449 | disc > 0 = DiscArgCtxt -- Be keener here
1450 | otherwise = BoringCtxt -- Nothing interesting
1451
1452 rebuildCall env (ArgInfo { ai_fun = fun, ai_args = rev_args, ai_rules = rules }) cont
1453 | null rules
1454 = rebuild env (argInfoExpr fun rev_args) cont -- No rules, common case
1455
1456 | otherwise
1457 = do { -- We've accumulated a simplified call in <fun,rev_args>
1458 -- so try rewrite rules; see Note [RULEs apply to simplified arguments]
1459 -- See also Note [Rules for recursive functions]
1460 ; let env' = zapSubstEnv env -- See Note [zapSubstEnv];
1461 -- and NB that 'rev_args' are all fully simplified
1462 ; mb_rule <- tryRules env' rules fun (reverse rev_args) cont
1463 ; case mb_rule of {
1464 Just (rule_rhs, cont') -> simplExprF env' rule_rhs cont'
1465
1466 -- Rules don't match
1467 ; Nothing -> rebuild env (argInfoExpr fun rev_args) cont -- No rules
1468 } }
1469
1470 {-
1471 Note [RULES apply to simplified arguments]
1472 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1473 It's very desirable to try RULES once the arguments have been simplified, because
1474 doing so ensures that rule cascades work in one pass. Consider
1475 {-# RULES g (h x) = k x
1476 f (k x) = x #-}
1477 ...f (g (h x))...
1478 Then we want to rewrite (g (h x)) to (k x) and only then try f's rules. If
1479 we match f's rules against the un-simplified RHS, it won't match. This
1480 makes a particularly big difference when superclass selectors are involved:
1481 op ($p1 ($p2 (df d)))
1482 We want all this to unravel in one sweep.
1483
1484 Note [Avoid redundant simplification]
1485 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1486 Because RULES apply to simplified arguments, there's a danger of repeatedly
1487 simplifying already-simplified arguments. An important example is that of
1488 (>>=) d e1 e2
1489 Here e1, e2 are simplified before the rule is applied, but don't really
1490 participate in the rule firing. So we mark them as Simplified to avoid
1491 re-simplifying them.
1492
1493 Note [Shadowing]
1494 ~~~~~~~~~~~~~~~~
1495 This part of the simplifier may break the no-shadowing invariant
1496 Consider
1497 f (...(\a -> e)...) (case y of (a,b) -> e')
1498 where f is strict in its second arg
1499 If we simplify the innermost one first we get (...(\a -> e)...)
1500 Simplifying the second arg makes us float the case out, so we end up with
1501 case y of (a,b) -> f (...(\a -> e)...) e'
1502 So the output does not have the no-shadowing invariant. However, there is
1503 no danger of getting name-capture, because when the first arg was simplified
1504 we used an in-scope set that at least mentioned all the variables free in its
1505 static environment, and that is enough.
1506
1507 We can't just do innermost first, or we'd end up with a dual problem:
1508 case x of (a,b) -> f e (...(\a -> e')...)
1509
1510 I spent hours trying to recover the no-shadowing invariant, but I just could
1511 not think of an elegant way to do it. The simplifier is already knee-deep in
1512 continuations. We have to keep the right in-scope set around; AND we have
1513 to get the effect that finding (error "foo") in a strict arg position will
1514 discard the entire application and replace it with (error "foo"). Getting
1515 all this at once is TOO HARD!
1516
1517
1518 ************************************************************************
1519 * *
1520 Rewrite rules
1521 * *
1522 ************************************************************************
1523 -}
1524
1525 tryRules :: SimplEnv -> [CoreRule]
1526 -> Id -> [ArgSpec] -> SimplCont
1527 -> SimplM (Maybe (CoreExpr, SimplCont))
1528 -- The SimplEnv already has zapSubstEnv applied to it
1529
1530 tryRules env rules fn args call_cont
1531 | null rules
1532 = return Nothing
1533 {- Disabled until we fix #8326
1534 | fn `hasKey` tagToEnumKey -- See Note [Optimising tagToEnum#]
1535 , [_type_arg, val_arg] <- args
1536 , Select dup bndr ((_,[],rhs1) : rest_alts) se cont <- call_cont
1537 , isDeadBinder bndr
1538 = do { dflags <- getDynFlags
1539 ; let enum_to_tag :: CoreAlt -> CoreAlt
1540 -- Takes K -> e into tagK# -> e
1541 -- where tagK# is the tag of constructor K
1542 enum_to_tag (DataAlt con, [], rhs)
1543 = ASSERT( isEnumerationTyCon (dataConTyCon con) )
1544 (LitAlt tag, [], rhs)
1545 where
1546 tag = mkMachInt dflags (toInteger (dataConTag con - fIRST_TAG))
1547 enum_to_tag alt = pprPanic "tryRules: tagToEnum" (ppr alt)
1548
1549 new_alts = (DEFAULT, [], rhs1) : map enum_to_tag rest_alts
1550 new_bndr = setIdType bndr intPrimTy
1551 -- The binder is dead, but should have the right type
1552 ; return (Just (val_arg, Select dup new_bndr new_alts se cont)) }
1553 -}
1554 | otherwise
1555 = do { dflags <- getDynFlags
1556 ; case lookupRule dflags (getUnfoldingInRuleMatch env) (activeRule env)
1557 fn (argInfoAppArgs args) rules of {
1558 Nothing ->
1559 do { nodump dflags -- This ensures that an empty file is written
1560 ; return Nothing } ; -- No rule matches
1561 Just (rule, rule_rhs) ->
1562 do { checkedTick (RuleFired (ru_name rule))
1563 ; let cont' = pushSimplifiedArgs env
1564 (drop (ruleArity rule) args)
1565 call_cont
1566 -- (ruleArity rule) says how many args the rule consumed
1567 ; dump dflags rule rule_rhs
1568 ; return (Just (rule_rhs, cont')) }}}
1569 where
1570 dump dflags rule rule_rhs
1571 | dopt Opt_D_dump_rule_rewrites dflags
1572 = log_rule dflags Opt_D_dump_rule_rewrites "Rule fired" $ vcat
1573 [ text "Rule:" <+> ftext (ru_name rule)
1574 , text "Before:" <+> hang (ppr fn) 2 (sep (map ppr args))
1575 , text "After: " <+> pprCoreExpr rule_rhs
1576 , text "Cont: " <+> ppr call_cont ]
1577
1578 | dopt Opt_D_dump_rule_firings dflags
1579 = log_rule dflags Opt_D_dump_rule_firings "Rule fired:" $
1580 ftext (ru_name rule)
1581
1582 | otherwise
1583 = return ()
1584
1585 nodump dflags
1586 | dopt Opt_D_dump_rule_rewrites dflags
1587 = liftIO $ dumpSDoc dflags alwaysQualify Opt_D_dump_rule_rewrites "" empty
1588
1589 | dopt Opt_D_dump_rule_firings dflags
1590 = liftIO $ dumpSDoc dflags alwaysQualify Opt_D_dump_rule_firings "" empty
1591
1592 | otherwise
1593 = return ()
1594
1595 log_rule dflags flag hdr details
1596 = liftIO . dumpSDoc dflags alwaysQualify flag "" $
1597 sep [text hdr, nest 4 details]
1598
1599 {-
1600 Note [Optimising tagToEnum#]
1601 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1602 If we have an enumeration data type:
1603
1604 data Foo = A | B | C
1605
1606 Then we want to transform
1607
1608 case tagToEnum# x of ==> case x of
1609 A -> e1 DEFAULT -> e1
1610 B -> e2 1# -> e2
1611 C -> e3 2# -> e3
1612
1613 thereby getting rid of the tagToEnum# altogether. If there was a DEFAULT
1614 alternative we retain it (remember it comes first). If not the case must
1615 be exhaustive, and we reflect that in the transformed version by adding
1616 a DEFAULT. Otherwise Lint complains that the new case is not exhaustive.
1617 See #8317.
1618
1619 Note [Rules for recursive functions]
1620 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1621 You might think that we shouldn't apply rules for a loop breaker:
1622 doing so might give rise to an infinite loop, because a RULE is
1623 rather like an extra equation for the function:
1624 RULE: f (g x) y = x+y
1625 Eqn: f a y = a-y
1626
1627 But it's too drastic to disable rules for loop breakers.
1628 Even the foldr/build rule would be disabled, because foldr
1629 is recursive, and hence a loop breaker:
1630 foldr k z (build g) = g k z
1631 So it's up to the programmer: rules can cause divergence
1632
1633
1634 ************************************************************************
1635 * *
1636 Rebuilding a case expression
1637 * *
1638 ************************************************************************
1639
1640 Note [Case elimination]
1641 ~~~~~~~~~~~~~~~~~~~~~~~
1642 The case-elimination transformation discards redundant case expressions.
1643 Start with a simple situation:
1644
1645 case x# of ===> let y# = x# in e
1646 y# -> e
1647
1648 (when x#, y# are of primitive type, of course). We can't (in general)
1649 do this for algebraic cases, because we might turn bottom into
1650 non-bottom!
1651
1652 The code in SimplUtils.prepareAlts has the effect of generalise this
1653 idea to look for a case where we're scrutinising a variable, and we
1654 know that only the default case can match. For example:
1655
1656 case x of
1657 0# -> ...
1658 DEFAULT -> ...(case x of
1659 0# -> ...
1660 DEFAULT -> ...) ...
1661
1662 Here the inner case is first trimmed to have only one alternative, the
1663 DEFAULT, after which it's an instance of the previous case. This
1664 really only shows up in eliminating error-checking code.
1665
1666 Note that SimplUtils.mkCase combines identical RHSs. So
1667
1668 case e of ===> case e of DEFAULT -> r
1669 True -> r
1670 False -> r
1671
1672 Now again the case may be elminated by the CaseElim transformation.
1673 This includes things like (==# a# b#)::Bool so that we simplify
1674 case ==# a# b# of { True -> x; False -> x }
1675 to just
1676 x
1677 This particular example shows up in default methods for
1678 comparison operations (e.g. in (>=) for Int.Int32)
1679
1680 Note [Case elimination: lifted case]
1681 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1682 If a case over a lifted type has a single alternative, and is being used
1683 as a strict 'let' (all isDeadBinder bndrs), we may want to do this
1684 transformation:
1685
1686 case e of r ===> let r = e in ...r...
1687 _ -> ...r...
1688
1689 (a) 'e' is already evaluated (it may so if e is a variable)
1690 Specifically we check (exprIsHNF e). In this case
1691 we can just allocate the WHNF directly with a let.
1692 or
1693 (b) 'x' is not used at all and e is ok-for-speculation
1694 The ok-for-spec bit checks that we don't lose any
1695 exceptions or divergence.
1696
1697 NB: it'd be *sound* to switch from case to let if the
1698 scrutinee was not yet WHNF but was guaranteed to
1699 converge; but sticking with case means we won't build a
1700 thunk
1701
1702 or
1703 (c) 'x' is used strictly in the body, and 'e' is a variable
1704 Then we can just substitute 'e' for 'x' in the body.
1705 See Note [Eliminating redundant seqs]
1706
1707 For (b), the "not used at all" test is important. Consider
1708 case (case a ># b of { True -> (p,q); False -> (q,p) }) of
1709 r -> blah
1710 The scrutinee is ok-for-speculation (it looks inside cases), but we do
1711 not want to transform to
1712 let r = case a ># b of { True -> (p,q); False -> (q,p) }
1713 in blah
1714 because that builds an unnecessary thunk.
1715
1716 Note [Eliminating redundant seqs]
1717 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1718 If we have this:
1719 case x of r { _ -> ..r.. }
1720 where 'r' is used strictly in (..r..), the case is effectively a 'seq'
1721 on 'x', but since 'r' is used strictly anyway, we can safely transform to
1722 (...x...)
1723
1724 Note that this can change the error behaviour. For example, we might
1725 transform
1726 case x of { _ -> error "bad" }
1727 --> error "bad"
1728 which is might be puzzling if 'x' currently lambda-bound, but later gets
1729 let-bound to (error "good").
1730
1731 Nevertheless, the paper "A semantics for imprecise exceptions" allows
1732 this transformation. If you want to fix the evaluation order, use
1733 'pseq'. See Trac #8900 for an example where the loss of this
1734 transformation bit us in practice.
1735
1736 See also Note [Empty case alternatives] in CoreSyn.
1737
1738 Just for reference, the original code (added Jan 13) looked like this:
1739 || case_bndr_evald_next rhs
1740
1741 case_bndr_evald_next :: CoreExpr -> Bool
1742 -- See Note [Case binder next]
1743 case_bndr_evald_next (Var v) = v == case_bndr
1744 case_bndr_evald_next (Cast e _) = case_bndr_evald_next e
1745 case_bndr_evald_next (App e _) = case_bndr_evald_next e
1746 case_bndr_evald_next (Case e _ _ _) = case_bndr_evald_next e
1747 case_bndr_evald_next _ = False
1748
1749 (This came up when fixing Trac #7542. See also Note [Eta reduction of
1750 an eval'd function] in CoreUtils.)
1751
1752
1753 Note [Case elimination: unlifted case]
1754 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1755 Consider
1756 case a +# b of r -> ...r...
1757 Then we do case-elimination (to make a let) followed by inlining,
1758 to get
1759 .....(a +# b)....
1760 If we have
1761 case indexArray# a i of r -> ...r...
1762 we might like to do the same, and inline the (indexArray# a i).
1763 But indexArray# is not okForSpeculation, so we don't build a let
1764 in rebuildCase (lest it get floated *out*), so the inlining doesn't
1765 happen either.
1766
1767 This really isn't a big deal I think. The let can be
1768
1769
1770 Further notes about case elimination
1771 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1772 Consider: test :: Integer -> IO ()
1773 test = print
1774
1775 Turns out that this compiles to:
1776 Print.test
1777 = \ eta :: Integer
1778 eta1 :: Void# ->
1779 case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT ->
1780 case hPutStr stdout
1781 (PrelNum.jtos eta ($w[] @ Char))
1782 eta1
1783 of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }}
1784
1785 Notice the strange '<' which has no effect at all. This is a funny one.
1786 It started like this:
1787
1788 f x y = if x < 0 then jtos x
1789 else if y==0 then "" else jtos x
1790
1791 At a particular call site we have (f v 1). So we inline to get
1792
1793 if v < 0 then jtos x
1794 else if 1==0 then "" else jtos x
1795
1796 Now simplify the 1==0 conditional:
1797
1798 if v<0 then jtos v else jtos v
1799
1800 Now common-up the two branches of the case:
1801
1802 case (v<0) of DEFAULT -> jtos v
1803
1804 Why don't we drop the case? Because it's strict in v. It's technically
1805 wrong to drop even unnecessary evaluations, and in practice they
1806 may be a result of 'seq' so we *definitely* don't want to drop those.
1807 I don't really know how to improve this situation.
1808 -}
1809
1810 ---------------------------------------------------------
1811 -- Eliminate the case if possible
1812
1813 rebuildCase, reallyRebuildCase
1814 :: SimplEnv
1815 -> OutExpr -- Scrutinee
1816 -> InId -- Case binder
1817 -> [InAlt] -- Alternatives (inceasing order)
1818 -> SimplCont
1819 -> SimplM (SimplEnv, OutExpr)
1820
1821 --------------------------------------------------
1822 -- 1. Eliminate the case if there's a known constructor
1823 --------------------------------------------------
1824
1825 rebuildCase env scrut case_bndr alts cont
1826 | Lit lit <- scrut -- No need for same treatment as constructors
1827 -- because literals are inlined more vigorously
1828 , not (litIsLifted lit)
1829 = do { tick (KnownBranch case_bndr)
1830 ; case findAlt (LitAlt lit) alts of
1831 Nothing -> missingAlt env case_bndr alts cont
1832 Just (_, bs, rhs) -> simple_rhs bs rhs }
1833
1834 | Just (con, ty_args, other_args) <- exprIsConApp_maybe (getUnfoldingInRuleMatch env) scrut
1835 -- Works when the scrutinee is a variable with a known unfolding
1836 -- as well as when it's an explicit constructor application
1837 = do { tick (KnownBranch case_bndr)
1838 ; case findAlt (DataAlt con) alts of
1839 Nothing -> missingAlt env case_bndr alts cont
1840 Just (DEFAULT, bs, rhs) -> simple_rhs bs rhs
1841 Just (_, bs, rhs) -> knownCon env scrut con ty_args other_args
1842 case_bndr bs rhs cont
1843 }
1844 where
1845 simple_rhs bs rhs = ASSERT( null bs )
1846 do { env' <- simplNonRecX env case_bndr scrut
1847 -- scrut is a constructor application,
1848 -- hence satisfies let/app invariant
1849 ; simplExprF env' rhs cont }
1850
1851
1852 --------------------------------------------------
1853 -- 2. Eliminate the case if scrutinee is evaluated
1854 --------------------------------------------------
1855
1856 rebuildCase env scrut case_bndr alts@[(_, bndrs, rhs)] cont
1857 -- See if we can get rid of the case altogether
1858 -- See Note [Case elimination]
1859 -- mkCase made sure that if all the alternatives are equal,
1860 -- then there is now only one (DEFAULT) rhs
1861
1862 -- 2a. Dropping the case altogether, if
1863 -- a) it binds nothing (so it's really just a 'seq')
1864 -- b) evaluating the scrutinee has no side effects
1865 | is_plain_seq
1866 , exprOkForSideEffects scrut
1867 -- The entire case is dead, so we can drop it
1868 -- if the scrutinee converges without having imperative
1869 -- side effects or raising a Haskell exception
1870 -- See Note [PrimOp can_fail and has_side_effects] in PrimOp
1871 = simplExprF env rhs cont
1872
1873 -- 2b. Turn the case into a let, if
1874 -- a) it binds only the case-binder
1875 -- b) unlifted case: the scrutinee is ok-for-speculation
1876 -- lifted case: the scrutinee is in HNF (or will later be demanded)
1877 | all_dead_bndrs
1878 , if is_unlifted
1879 then exprOkForSpeculation scrut -- See Note [Case elimination: unlifted case]
1880 else exprIsHNF scrut -- See Note [Case elimination: lifted case]
1881 || scrut_is_demanded_var scrut
1882 = do { tick (CaseElim case_bndr)
1883 ; env' <- simplNonRecX env case_bndr scrut
1884 ; simplExprF env' rhs cont }
1885
1886 -- 2c. Try the seq rules if
1887 -- a) it binds only the case binder
1888 -- b) a rule for seq applies
1889 -- See Note [User-defined RULES for seq] in MkId
1890 | is_plain_seq
1891 = do { let scrut_ty = exprType scrut
1892 rhs_ty = substTy env (exprType rhs)
1893 out_args = [ TyArg { as_arg_ty = scrut_ty
1894 , as_hole_ty = seq_id_ty }
1895 , TyArg { as_arg_ty = rhs_ty
1896 , as_hole_ty = piResultTy seq_id_ty scrut_ty }
1897 , ValArg scrut]
1898 rule_cont = ApplyToVal { sc_dup = NoDup, sc_arg = rhs
1899 , sc_env = env, sc_cont = cont }
1900 env' = zapSubstEnv env
1901 -- Lazily evaluated, so we don't do most of this
1902
1903 ; rule_base <- getSimplRules
1904 ; mb_rule <- tryRules env' (getRules rule_base seqId) seqId out_args rule_cont
1905 ; case mb_rule of
1906 Just (rule_rhs, cont') -> simplExprF env' rule_rhs cont'
1907 Nothing -> reallyRebuildCase env scrut case_bndr alts cont }
1908 where
1909 is_unlifted = isUnliftedType (idType case_bndr)
1910 all_dead_bndrs = all isDeadBinder bndrs -- bndrs are [InId]
1911 is_plain_seq = all_dead_bndrs && isDeadBinder case_bndr -- Evaluation *only* for effect
1912 seq_id_ty = idType seqId
1913
1914 scrut_is_demanded_var :: CoreExpr -> Bool
1915 -- See Note [Eliminating redundant seqs]
1916 scrut_is_demanded_var (Cast s _) = scrut_is_demanded_var s
1917 scrut_is_demanded_var (Var _) = isStrictDmd (idDemandInfo case_bndr)
1918 scrut_is_demanded_var _ = False
1919
1920
1921 rebuildCase env scrut case_bndr alts cont
1922 = reallyRebuildCase env scrut case_bndr alts cont
1923
1924 --------------------------------------------------
1925 -- 3. Catch-all case
1926 --------------------------------------------------
1927
1928 reallyRebuildCase env scrut case_bndr alts cont
1929 = do { -- Prepare the continuation;
1930 -- The new subst_env is in place
1931 (env', dup_cont, nodup_cont) <- prepareCaseCont env alts cont
1932
1933 -- Simplify the alternatives
1934 ; (scrut', case_bndr', alts') <- simplAlts env' scrut case_bndr alts dup_cont
1935
1936 ; dflags <- getDynFlags
1937 ; let alts_ty' = contResultType dup_cont
1938 ; case_expr <- mkCase dflags scrut' case_bndr' alts_ty' alts'
1939
1940 -- Notice that rebuild gets the in-scope set from env', not alt_env
1941 -- (which in any case is only build in simplAlts)
1942 -- The case binder *not* scope over the whole returned case-expression
1943 ; rebuild env' case_expr nodup_cont }
1944
1945 {-
1946 simplCaseBinder checks whether the scrutinee is a variable, v. If so,
1947 try to eliminate uses of v in the RHSs in favour of case_bndr; that
1948 way, there's a chance that v will now only be used once, and hence
1949 inlined.
1950
1951 Historical note: we use to do the "case binder swap" in the Simplifier
1952 so there were additional complications if the scrutinee was a variable.
1953 Now the binder-swap stuff is done in the occurrence analyer; see
1954 OccurAnal Note [Binder swap].
1955
1956 Note [knownCon occ info]
1957 ~~~~~~~~~~~~~~~~~~~~~~~~
1958 If the case binder is not dead, then neither are the pattern bound
1959 variables:
1960 case <any> of x { (a,b) ->
1961 case x of { (p,q) -> p } }
1962 Here (a,b) both look dead, but come alive after the inner case is eliminated.
1963 The point is that we bring into the envt a binding
1964 let x = (a,b)
1965 after the outer case, and that makes (a,b) alive. At least we do unless
1966 the case binder is guaranteed dead.
1967
1968 Note [Case alternative occ info]
1969 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1970 When we are simply reconstructing a case (the common case), we always
1971 zap the occurrence info on the binders in the alternatives. Even
1972 if the case binder is dead, the scrutinee is usually a variable, and *that*
1973 can bring the case-alternative binders back to life.
1974 See Note [Add unfolding for scrutinee]
1975
1976 Note [Improving seq]
1977 ~~~~~~~~~~~~~~~~~~~
1978 Consider
1979 type family F :: * -> *
1980 type instance F Int = Int
1981
1982 ... case e of x { DEFAULT -> rhs } ...
1983
1984 where x::F Int. Then we'd like to rewrite (F Int) to Int, getting
1985
1986 case e `cast` co of x'::Int
1987 I# x# -> let x = x' `cast` sym co
1988 in rhs
1989
1990 so that 'rhs' can take advantage of the form of x'.
1991
1992 Notice that Note [Case of cast] (in OccurAnal) may then apply to the result.
1993
1994 Nota Bene: We only do the [Improving seq] transformation if the
1995 case binder 'x' is actually used in the rhs; that is, if the case
1996 is *not* a *pure* seq.
1997 a) There is no point in adding the cast to a pure seq.
1998 b) There is a good reason not to: doing so would interfere
1999 with seq rules (Note [Built-in RULES for seq] in MkId).
2000 In particular, this [Improving seq] thing *adds* a cast
2001 while [Built-in RULES for seq] *removes* one, so they
2002 just flip-flop.
2003
2004 You might worry about
2005 case v of x { __DEFAULT ->
2006 ... case (v `cast` co) of y { I# -> ... }}
2007 This is a pure seq (since x is unused), so [Improving seq] won't happen.
2008 But it's ok: the simplifier will replace 'v' by 'x' in the rhs to get
2009 case v of x { __DEFAULT ->
2010 ... case (x `cast` co) of y { I# -> ... }}
2011 Now the outer case is not a pure seq, so [Improving seq] will happen,
2012 and then the inner case will disappear.
2013
2014 The need for [Improving seq] showed up in Roman's experiments. Example:
2015 foo :: F Int -> Int -> Int
2016 foo t n = t `seq` bar n
2017 where
2018 bar 0 = 0
2019 bar n = bar (n - case t of TI i -> i)
2020 Here we'd like to avoid repeated evaluating t inside the loop, by
2021 taking advantage of the `seq`.
2022
2023 At one point I did transformation in LiberateCase, but it's more
2024 robust here. (Otherwise, there's a danger that we'll simply drop the
2025 'seq' altogether, before LiberateCase gets to see it.)
2026 -}
2027
2028 simplAlts :: SimplEnv
2029 -> OutExpr
2030 -> InId -- Case binder
2031 -> [InAlt] -- Non-empty
2032 -> SimplCont
2033 -> SimplM (OutExpr, OutId, [OutAlt]) -- Includes the continuation
2034 -- Like simplExpr, this just returns the simplified alternatives;
2035 -- it does not return an environment
2036 -- The returned alternatives can be empty, none are possible
2037
2038 simplAlts env scrut case_bndr alts cont'
2039 = do { let env0 = zapFloats env
2040
2041 ; (env1, case_bndr1) <- simplBinder env0 case_bndr
2042 ; let case_bndr2 = case_bndr1 `setIdUnfolding` evaldUnfolding
2043 env2 = modifyInScope env1 case_bndr2
2044 -- See Note [Case-binder evaluated-ness]
2045
2046 ; fam_envs <- getFamEnvs
2047 ; (alt_env', scrut', case_bndr') <- improveSeq fam_envs env2 scrut
2048 case_bndr case_bndr2 alts
2049
2050 ; (imposs_deflt_cons, in_alts) <- prepareAlts scrut' case_bndr' alts
2051 -- NB: it's possible that the returned in_alts is empty: this is handled
2052 -- by the caller (rebuildCase) in the missingAlt function
2053
2054 ; alts' <- mapM (simplAlt alt_env' (Just scrut') imposs_deflt_cons case_bndr' cont') in_alts
2055 ; -- pprTrace "simplAlts" (ppr case_bndr $$ ppr alts_ty $$ ppr alts_ty' $$ ppr alts $$ ppr cont') $
2056 return (scrut', case_bndr', alts') }
2057
2058
2059 ------------------------------------
2060 improveSeq :: (FamInstEnv, FamInstEnv) -> SimplEnv
2061 -> OutExpr -> InId -> OutId -> [InAlt]
2062 -> SimplM (SimplEnv, OutExpr, OutId)
2063 -- Note [Improving seq]
2064 improveSeq fam_envs env scrut case_bndr case_bndr1 [(DEFAULT,_,_)]
2065 | not (isDeadBinder case_bndr) -- Not a pure seq! See Note [Improving seq]
2066 , Just (co, ty2) <- topNormaliseType_maybe fam_envs (idType case_bndr1)
2067 = do { case_bndr2 <- newId (fsLit "nt") ty2
2068 ; let rhs = DoneEx (Var case_bndr2 `Cast` mkSymCo co)
2069 env2 = extendIdSubst env case_bndr rhs
2070 ; return (env2, scrut `Cast` co, case_bndr2) }
2071
2072 improveSeq _ env scrut _ case_bndr1 _
2073 = return (env, scrut, case_bndr1)
2074
2075
2076 ------------------------------------
2077 simplAlt :: SimplEnv
2078 -> Maybe OutExpr -- The scrutinee
2079 -> [AltCon] -- These constructors can't be present when
2080 -- matching the DEFAULT alternative
2081 -> OutId -- The case binder
2082 -> SimplCont
2083 -> InAlt
2084 -> SimplM OutAlt
2085
2086 simplAlt env _ imposs_deflt_cons case_bndr' cont' (DEFAULT, bndrs, rhs)
2087 = ASSERT( null bndrs )
2088 do { let env' = addBinderUnfolding env case_bndr'
2089 (mkOtherCon imposs_deflt_cons)
2090 -- Record the constructors that the case-binder *can't* be.
2091 ; rhs' <- simplExprC env' rhs cont'
2092 ; return (DEFAULT, [], rhs') }
2093
2094 simplAlt env scrut' _ case_bndr' cont' (LitAlt lit, bndrs, rhs)
2095 = ASSERT( null bndrs )
2096 do { env' <- addAltUnfoldings env scrut' case_bndr' (Lit lit)
2097 ; rhs' <- simplExprC env' rhs cont'
2098 ; return (LitAlt lit, [], rhs') }
2099
2100 simplAlt env scrut' _ case_bndr' cont' (DataAlt con, vs, rhs)
2101 = do { -- Deal with the pattern-bound variables
2102 -- Mark the ones that are in ! positions in the
2103 -- data constructor as certainly-evaluated.
2104 -- NB: simplLamBinders preserves this eval info
2105 ; let vs_with_evals = add_evals (dataConRepStrictness con)
2106 ; (env', vs') <- simplLamBndrs env vs_with_evals
2107
2108 -- Bind the case-binder to (con args)
2109 ; let inst_tys' = tyConAppArgs (idType case_bndr')
2110 con_app :: OutExpr
2111 con_app = mkConApp2 con inst_tys' vs'
2112
2113 ; env'' <- addAltUnfoldings env' scrut' case_bndr' con_app
2114 ; rhs' <- simplExprC env'' rhs cont'
2115 ; return (DataAlt con, vs', rhs') }
2116 where
2117 -- add_evals records the evaluated-ness of the bound variables of
2118 -- a case pattern. This is *important*. Consider
2119 -- data T = T !Int !Int
2120 --
2121 -- case x of { T a b -> T (a+1) b }
2122 --
2123 -- We really must record that b is already evaluated so that we don't
2124 -- go and re-evaluate it when constructing the result.
2125 -- See Note [Data-con worker strictness] in MkId.hs
2126 add_evals the_strs
2127 = go vs the_strs
2128 where
2129 go [] [] = []
2130 go (v:vs') strs | isTyVar v = v : go vs' strs
2131 go (v:vs') (str:strs)
2132 | isMarkedStrict str = eval v : go vs' strs
2133 | otherwise = zap v : go vs' strs
2134 go _ _ = pprPanic "cat_evals"
2135 (ppr con $$
2136 ppr vs $$
2137 ppr_with_length the_strs $$
2138 ppr_with_length (dataConRepArgTys con) $$
2139 ppr_with_length (dataConRepStrictness con))
2140 where
2141 ppr_with_length list
2142 = ppr list <+> parens (text "length =" <+> ppr (length list))
2143 -- NB: If this panic triggers, note that
2144 -- NoStrictnessMark doesn't print!
2145
2146 zap v = zapIdOccInfo v -- See Note [Case alternative occ info]
2147 eval v = zap v `setIdUnfolding` evaldUnfolding
2148
2149 addAltUnfoldings :: SimplEnv -> Maybe OutExpr -> OutId -> OutExpr -> SimplM SimplEnv
2150 addAltUnfoldings env scrut case_bndr con_app
2151 = do { dflags <- getDynFlags
2152 ; let con_app_unf = mkSimpleUnfolding dflags con_app
2153 env1 = addBinderUnfolding env case_bndr con_app_unf
2154
2155 -- See Note [Add unfolding for scrutinee]
2156 env2 = case scrut of
2157 Just (Var v) -> addBinderUnfolding env1 v con_app_unf
2158 Just (Cast (Var v) co) -> addBinderUnfolding env1 v $
2159 mkSimpleUnfolding dflags (Cast con_app (mkSymCo co))
2160 _ -> env1
2161
2162 ; traceSmpl "addAltUnf" (vcat [ppr case_bndr <+> ppr scrut, ppr con_app])
2163 ; return env2 }
2164
2165 addBinderUnfolding :: SimplEnv -> Id -> Unfolding -> SimplEnv
2166 addBinderUnfolding env bndr unf
2167 | debugIsOn, Just tmpl <- maybeUnfoldingTemplate unf
2168 = WARN( not (eqType (idType bndr) (exprType tmpl)),
2169 ppr bndr $$ ppr (idType bndr) $$ ppr tmpl $$ ppr (exprType tmpl) )
2170 modifyInScope env (bndr `setIdUnfolding` unf)
2171
2172 | otherwise
2173 = modifyInScope env (bndr `setIdUnfolding` unf)
2174
2175 zapBndrOccInfo :: Bool -> Id -> Id
2176 -- Consider case e of b { (a,b) -> ... }
2177 -- Then if we bind b to (a,b) in "...", and b is not dead,
2178 -- then we must zap the deadness info on a,b
2179 zapBndrOccInfo keep_occ_info pat_id
2180 | keep_occ_info = pat_id
2181 | otherwise = zapIdOccInfo pat_id
2182
2183 {- Note [Case binder evaluated-ness]
2184 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2185 We pin on a (OtherCon []) unfolding to the case-binder of a Case,
2186 even though it'll be over-ridden in every case alternative with a more
2187 informative unfolding. Why? Because suppose a later, less clever, pass
2188 simply replaces all occurrences of the case binder with the binder itself;
2189 then Lint may complain about the let/app invariant. Example
2190 case e of b { DEFAULT -> let v = reallyUnsafePtrEq# b y in ....
2191 ; K -> blah }
2192
2193 The let/app invariant requires that y is evaluated in the call to
2194 reallyUnsafePtrEq#, which it is. But we still want that to be true if we
2195 propagate binders to occurrences.
2196
2197 This showed up in Trac #13027.
2198
2199 Note [Add unfolding for scrutinee]
2200 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2201 In general it's unlikely that a variable scrutinee will appear
2202 in the case alternatives case x of { ...x unlikely to appear... }
2203 because the binder-swap in OccAnal has got rid of all such occcurrences
2204 See Note [Binder swap] in OccAnal.
2205
2206 BUT it is still VERY IMPORTANT to add a suitable unfolding for a
2207 variable scrutinee, in simplAlt. Here's why
2208 case x of y
2209 (a,b) -> case b of c
2210 I# v -> ...(f y)...
2211 There is no occurrence of 'b' in the (...(f y)...). But y gets
2212 the unfolding (a,b), and *that* mentions b. If f has a RULE
2213 RULE f (p, I# q) = ...
2214 we want that rule to match, so we must extend the in-scope env with a
2215 suitable unfolding for 'y'. It's *essential* for rule matching; but
2216 it's also good for case-elimintation -- suppose that 'f' was inlined
2217 and did multi-level case analysis, then we'd solve it in one
2218 simplifier sweep instead of two.
2219
2220 Exactly the same issue arises in SpecConstr;
2221 see Note [Add scrutinee to ValueEnv too] in SpecConstr
2222
2223 HOWEVER, given
2224 case x of y { Just a -> r1; Nothing -> r2 }
2225 we do not want to add the unfolding x -> y to 'x', which might seem cool,
2226 since 'y' itself has different unfoldings in r1 and r2. Reason: if we
2227 did that, we'd have to zap y's deadness info and that is a very useful
2228 piece of information.
2229
2230 So instead we add the unfolding x -> Just a, and x -> Nothing in the
2231 respective RHSs.
2232
2233
2234 ************************************************************************
2235 * *
2236 \subsection{Known constructor}
2237 * *
2238 ************************************************************************
2239
2240 We are a bit careful with occurrence info. Here's an example
2241
2242 (\x* -> case x of (a*, b) -> f a) (h v, e)
2243
2244 where the * means "occurs once". This effectively becomes
2245 case (h v, e) of (a*, b) -> f a)
2246 and then
2247 let a* = h v; b = e in f a
2248 and then
2249 f (h v)
2250
2251 All this should happen in one sweep.
2252 -}
2253
2254 knownCon :: SimplEnv
2255 -> OutExpr -- The scrutinee
2256 -> DataCon -> [OutType] -> [OutExpr] -- The scrutinee (in pieces)
2257 -> InId -> [InBndr] -> InExpr -- The alternative
2258 -> SimplCont
2259 -> SimplM (SimplEnv, OutExpr)
2260
2261 knownCon env scrut dc dc_ty_args dc_args bndr bs rhs cont
2262 = do { env' <- bind_args env bs dc_args
2263 ; env'' <- bind_case_bndr env'
2264 ; simplExprF env'' rhs cont }
2265 where
2266 zap_occ = zapBndrOccInfo (isDeadBinder bndr) -- bndr is an InId
2267
2268 -- Ugh!
2269 bind_args env' [] _ = return env'
2270
2271 bind_args env' (b:bs') (Type ty : args)
2272 = ASSERT( isTyVar b )
2273 bind_args (extendTvSubst env' b ty) bs' args
2274
2275 bind_args env' (b:bs') (Coercion co : args)
2276 = ASSERT( isCoVar b )
2277 bind_args (extendCvSubst env' b co) bs' args
2278
2279 bind_args env' (b:bs') (arg : args)
2280 = ASSERT( isId b )
2281 do { let b' = zap_occ b
2282 -- Note that the binder might be "dead", because it doesn't
2283 -- occur in the RHS; and simplNonRecX may therefore discard
2284 -- it via postInlineUnconditionally.
2285 -- Nevertheless we must keep it if the case-binder is alive,
2286 -- because it may be used in the con_app. See Note [knownCon occ info]
2287 ; env'' <- simplNonRecX env' b' arg -- arg satisfies let/app invariant
2288 ; bind_args env'' bs' args }
2289
2290 bind_args _ _ _ =
2291 pprPanic "bind_args" $ ppr dc $$ ppr bs $$ ppr dc_args $$
2292 text "scrut:" <+> ppr scrut
2293
2294 -- It's useful to bind bndr to scrut, rather than to a fresh
2295 -- binding x = Con arg1 .. argn
2296 -- because very often the scrut is a variable, so we avoid
2297 -- creating, and then subsequently eliminating, a let-binding
2298 -- BUT, if scrut is a not a variable, we must be careful
2299 -- about duplicating the arg redexes; in that case, make
2300 -- a new con-app from the args
2301 bind_case_bndr env
2302 | isDeadBinder bndr = return env
2303 | exprIsTrivial scrut = return (extendIdSubst env bndr (DoneEx scrut))
2304 | otherwise = do { dc_args <- mapM (simplVar env) bs
2305 -- dc_ty_args are aready OutTypes,
2306 -- but bs are InBndrs
2307 ; let con_app = Var (dataConWorkId dc)
2308 `mkTyApps` dc_ty_args
2309 `mkApps` dc_args
2310 ; simplNonRecX env bndr con_app }
2311
2312 -------------------
2313 missingAlt :: SimplEnv -> Id -> [InAlt] -> SimplCont -> SimplM (SimplEnv, OutExpr)
2314 -- This isn't strictly an error, although it is unusual.
2315 -- It's possible that the simplifier might "see" that
2316 -- an inner case has no accessible alternatives before
2317 -- it "sees" that the entire branch of an outer case is
2318 -- inaccessible. So we simply put an error case here instead.
2319 missingAlt env case_bndr _ cont
2320 = WARN( True, text "missingAlt" <+> ppr case_bndr )
2321 return (env, mkImpossibleExpr (contResultType cont))
2322
2323 {-
2324 ************************************************************************
2325 * *
2326 \subsection{Duplicating continuations}
2327 * *
2328 ************************************************************************
2329 -}
2330
2331 prepareCaseCont :: SimplEnv
2332 -> [InAlt] -> SimplCont
2333 -> SimplM (SimplEnv,
2334 SimplCont, -- Dupable part
2335 SimplCont) -- Non-dupable part
2336 -- We are considering
2337 -- K[case _ of { p1 -> r1; ...; pn -> rn }]
2338 -- where K is some enclosing continuation for the case
2339 -- Goal: split K into two pieces Kdup,Knodup so that
2340 -- a) Kdup can be duplicated
2341 -- b) Knodup[Kdup[e]] = K[e]
2342 -- The idea is that we'll transform thus:
2343 -- Knodup[ (case _ of { p1 -> Kdup[r1]; ...; pn -> Kdup[rn] }
2344 --
2345 -- We may also return some extra bindings in SimplEnv (that scope over
2346 -- the entire continuation)
2347 --
2348 -- When case-of-case is off, just make the entire continuation non-dupable
2349
2350 prepareCaseCont env alts cont
2351 | not (sm_case_case (getMode env)) = return (env, mkBoringStop (contHoleType cont), cont)
2352 | not (many_alts alts) = return (env, cont, mkBoringStop (contResultType cont))
2353 | otherwise = mkDupableCont env cont
2354 where
2355 many_alts :: [InAlt] -> Bool -- True iff strictly > 1 non-bottom alternative
2356 many_alts [] = False -- See Note [Bottom alternatives]
2357 many_alts [_] = False
2358 many_alts (alt:alts)
2359 | is_bot_alt alt = many_alts alts
2360 | otherwise = not (all is_bot_alt alts)
2361
2362 is_bot_alt (_,_,rhs) = exprIsBottom rhs
2363
2364 {-
2365 Note [Bottom alternatives]
2366 ~~~~~~~~~~~~~~~~~~~~~~~~~~
2367 When we have
2368 case (case x of { A -> error .. ; B -> e; C -> error ..)
2369 of alts
2370 then we can just duplicate those alts because the A and C cases
2371 will disappear immediately. This is more direct than creating
2372 join points and inlining them away; and in some cases we would
2373 not even create the join points (see Note [Single-alternative case])
2374 and we would keep the case-of-case which is silly. See Trac #4930.
2375 -}
2376
2377 mkDupableCont :: SimplEnv -> SimplCont
2378 -> SimplM (SimplEnv, SimplCont, SimplCont)
2379
2380 mkDupableCont env cont
2381 | contIsDupable cont
2382 = return (env, cont, mkBoringStop (contResultType cont))
2383
2384 mkDupableCont _ (Stop {}) = panic "mkDupableCont" -- Handled by previous eqn
2385
2386 mkDupableCont env (CastIt ty cont)
2387 = do { (env', dup, nodup) <- mkDupableCont env cont
2388 ; return (env', CastIt ty dup, nodup) }
2389
2390 -- Duplicating ticks for now, not sure if this is good or not
2391 mkDupableCont env cont@(TickIt{})
2392 = return (env, mkBoringStop (contHoleType cont), cont)
2393
2394 mkDupableCont env cont@(StrictBind {})
2395 = return (env, mkBoringStop (contHoleType cont), cont)
2396 -- See Note [Duplicating StrictBind]
2397
2398 mkDupableCont env (StrictArg info cci cont)
2399 -- See Note [Duplicating StrictArg]
2400 = do { (env', dup, nodup) <- mkDupableCont env cont
2401 ; (env'', args') <- mapAccumLM makeTrivialArg env' (ai_args info)
2402 ; return (env'', StrictArg (info { ai_args = args' }) cci dup, nodup) }
2403
2404 mkDupableCont env cont@(ApplyToTy { sc_cont = tail })
2405 = do { (env', dup_cont, nodup_cont) <- mkDupableCont env tail
2406 ; return (env', cont { sc_cont = dup_cont }, nodup_cont ) }
2407
2408 mkDupableCont env (ApplyToVal { sc_arg = arg, sc_dup = dup, sc_env = se, sc_cont = cont })
2409 = -- e.g. [...hole...] (...arg...)
2410 -- ==>
2411 -- let a = ...arg...
2412 -- in [...hole...] a
2413 do { (env', dup_cont, nodup_cont) <- mkDupableCont env cont
2414 ; (_, se', arg') <- simplArg env' dup se arg
2415 ; (env'', arg'') <- makeTrivial NotTopLevel env' (fsLit "karg") arg'
2416 ; let app_cont = ApplyToVal { sc_arg = arg'', sc_env = se'
2417 , sc_dup = OkToDup, sc_cont = dup_cont }
2418 ; return (env'', app_cont, nodup_cont) }
2419
2420 mkDupableCont env cont@(Select { sc_bndr = case_bndr, sc_alts = [(_, bs, _rhs)] })
2421 -- See Note [Single-alternative case]
2422 -- | not (exprIsDupable rhs && contIsDupable case_cont)
2423 -- | not (isDeadBinder case_bndr)
2424 | all isDeadBinder bs -- InIds
2425 && not (isUnliftedType (idType case_bndr))
2426 -- Note [Single-alternative-unlifted]
2427 = return (env, mkBoringStop (contHoleType cont), cont)
2428
2429 mkDupableCont env (Select { sc_bndr = case_bndr, sc_alts = alts
2430 , sc_env = se, sc_cont = cont })
2431 = -- e.g. (case [...hole...] of { pi -> ei })
2432 -- ===>
2433 -- let ji = \xij -> ei
2434 -- in case [...hole...] of { pi -> ji xij }
2435 do { tick (CaseOfCase case_bndr)
2436 ; (env', dup_cont, nodup_cont) <- prepareCaseCont env alts cont
2437 -- NB: We call prepareCaseCont here. If there is only one
2438 -- alternative, then dup_cont may be big, but that's ok
2439 -- because we push it into the single alternative, and then
2440 -- use mkDupableAlt to turn that simplified alternative into
2441 -- a join point if it's too big to duplicate.
2442 -- And this is important: see Note [Fusing case continuations]
2443
2444 ; let alt_env = se `setInScope` env'
2445
2446 ; (alt_env', case_bndr') <- simplBinder alt_env case_bndr
2447 ; alts' <- mapM (simplAlt alt_env' Nothing [] case_bndr' dup_cont) alts
2448 -- Safe to say that there are no handled-cons for the DEFAULT case
2449 -- NB: simplBinder does not zap deadness occ-info, so
2450 -- a dead case_bndr' will still advertise its deadness
2451 -- This is really important because in
2452 -- case e of b { (# p,q #) -> ... }
2453 -- b is always dead, and indeed we are not allowed to bind b to (# p,q #),
2454 -- which might happen if e was an explicit unboxed pair and b wasn't marked dead.
2455 -- In the new alts we build, we have the new case binder, so it must retain
2456 -- its deadness.
2457 -- NB: we don't use alt_env further; it has the substEnv for
2458 -- the alternatives, and we don't want that
2459
2460 ; (env'', alts'') <- mkDupableAlts env' case_bndr' alts'
2461 ; return (env'', -- Note [Duplicated env]
2462 Select { sc_dup = OkToDup
2463 , sc_bndr = case_bndr', sc_alts = alts''
2464 , sc_env = zapSubstEnv env''
2465 , sc_cont = mkBoringStop (contHoleType nodup_cont) },
2466 nodup_cont) }
2467
2468
2469 mkDupableAlts :: SimplEnv -> OutId -> [InAlt]
2470 -> SimplM (SimplEnv, [InAlt])
2471 -- Absorbs the continuation into the new alternatives
2472
2473 mkDupableAlts env case_bndr' the_alts
2474 = go env the_alts
2475 where
2476 go env0 [] = return (env0, [])
2477 go env0 (alt:alts)
2478 = do { (env1, alt') <- mkDupableAlt env0 case_bndr' alt
2479 ; (env2, alts') <- go env1 alts
2480 ; return (env2, alt' : alts' ) }
2481
2482 mkDupableAlt :: SimplEnv -> OutId -> (AltCon, [CoreBndr], CoreExpr)
2483 -> SimplM (SimplEnv, (AltCon, [CoreBndr], CoreExpr))
2484 mkDupableAlt env case_bndr (con, bndrs', rhs') = do
2485 dflags <- getDynFlags
2486 if exprIsDupable dflags rhs' -- Note [Small alternative rhs]
2487 then return (env, (con, bndrs', rhs'))
2488 else
2489 do { let rhs_ty' = exprType rhs'
2490 scrut_ty = idType case_bndr
2491 case_bndr_w_unf
2492 = case con of
2493 DEFAULT -> case_bndr
2494 DataAlt dc -> setIdUnfolding case_bndr unf
2495 where
2496 -- See Note [Case binders and join points]
2497 unf = mkInlineUnfolding rhs
2498 rhs = mkConApp2 dc (tyConAppArgs scrut_ty) bndrs'
2499
2500 LitAlt {} -> WARN( True, text "mkDupableAlt"
2501 <+> ppr case_bndr <+> ppr con )
2502 case_bndr
2503 -- The case binder is alive but trivial, so why has
2504 -- it not been substituted away?
2505
2506 used_bndrs' | isDeadBinder case_bndr = filter abstract_over bndrs'
2507 | otherwise = bndrs' ++ [case_bndr_w_unf]
2508
2509 abstract_over bndr
2510 | isTyVar bndr = True -- Abstract over all type variables just in case
2511 | otherwise = not (isDeadBinder bndr)
2512 -- The deadness info on the new Ids is preserved by simplBinders
2513
2514 ; (final_bndrs', final_args) -- Note [Join point abstraction]
2515 <- if (any isId used_bndrs')
2516 then return (used_bndrs', varsToCoreExprs used_bndrs')
2517 else do { rw_id <- newId (fsLit "w") voidPrimTy
2518 ; return ([setOneShotLambda rw_id], [Var voidPrimId]) }
2519
2520 ; join_bndr <- newId (fsLit "$j") (mkLamTypes final_bndrs' rhs_ty')
2521 -- Note [Funky mkLamTypes]
2522
2523 ; let -- We make the lambdas into one-shot-lambdas. The
2524 -- join point is sure to be applied at most once, and doing so
2525 -- prevents the body of the join point being floated out by
2526 -- the full laziness pass
2527 really_final_bndrs = map one_shot final_bndrs'
2528 one_shot v | isId v = setOneShotLambda v
2529 | otherwise = v
2530 join_rhs = mkLams really_final_bndrs rhs'
2531 join_arity = exprArity join_rhs
2532 join_call = mkApps (Var join_bndr) final_args
2533
2534 ; env' <- addPolyBind NotTopLevel env (NonRec (join_bndr `setIdArity` join_arity) join_rhs)
2535 ; return (env', (con, bndrs', join_call)) }
2536 -- See Note [Duplicated env]
2537
2538 {-
2539 Note [Fusing case continuations]
2540 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2541 It's important to fuse two successive case continuations when the
2542 first has one alternative. That's why we call prepareCaseCont here.
2543 Consider this, which arises from thunk splitting (see Note [Thunk
2544 splitting] in WorkWrap):
2545
2546 let
2547 x* = case (case v of {pn -> rn}) of
2548 I# a -> I# a
2549 in body
2550
2551 The simplifier will find
2552 (Var v) with continuation
2553 Select (pn -> rn) (
2554 Select [I# a -> I# a] (
2555 StrictBind body Stop
2556
2557 So we'll call mkDupableCont on
2558 Select [I# a -> I# a] (StrictBind body Stop)
2559 There is just one alternative in the first Select, so we want to
2560 simplify the rhs (I# a) with continuation (StricgtBind body Stop)
2561 Supposing that body is big, we end up with
2562 let $j a = <let x = I# a in body>
2563 in case v of { pn -> case rn of
2564 I# a -> $j a }
2565 This is just what we want because the rn produces a box that
2566 the case rn cancels with.
2567
2568 See Trac #4957 a fuller example.
2569
2570 Note [Case binders and join points]
2571 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2572 Consider this
2573 case (case .. ) of c {
2574 I# c# -> ....c....
2575
2576 If we make a join point with c but not c# we get
2577 $j = \c -> ....c....
2578
2579 But if later inlining scrutines the c, thus
2580
2581 $j = \c -> ... case c of { I# y -> ... } ...
2582
2583 we won't see that 'c' has already been scrutinised. This actually
2584 happens in the 'tabulate' function in wave4main, and makes a significant
2585 difference to allocation.
2586
2587 An alternative plan is this:
2588
2589 $j = \c# -> let c = I# c# in ...c....
2590
2591 but that is bad if 'c' is *not* later scrutinised.
2592
2593 So instead we do both: we pass 'c' and 'c#' , and record in c's inlining
2594 (a stable unfolding) that it's really I# c#, thus
2595
2596 $j = \c# -> \c[=I# c#] -> ...c....
2597
2598 Absence analysis may later discard 'c'.
2599
2600 NB: take great care when doing strictness analysis;
2601 see Note [Lamba-bound unfoldings] in DmdAnal.
2602
2603 Also note that we can still end up passing stuff that isn't used. Before
2604 strictness analysis we have
2605 let $j x y c{=(x,y)} = (h c, ...)
2606 in ...
2607 After strictness analysis we see that h is strict, we end up with
2608 let $j x y c{=(x,y)} = ($wh x y, ...)
2609 and c is unused.
2610
2611 Note [Duplicated env]
2612 ~~~~~~~~~~~~~~~~~~~~~
2613 Some of the alternatives are simplified, but have not been turned into a join point
2614 So they *must* have an zapped subst-env. So we can't use completeNonRecX to
2615 bind the join point, because it might to do PostInlineUnconditionally, and
2616 we'd lose that when zapping the subst-env. We could have a per-alt subst-env,
2617 but zapping it (as we do in mkDupableCont, the Select case) is safe, and
2618 at worst delays the join-point inlining.
2619
2620 Note [Small alternative rhs]
2621 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2622 It is worth checking for a small RHS because otherwise we
2623 get extra let bindings that may cause an extra iteration of the simplifier to
2624 inline back in place. Quite often the rhs is just a variable or constructor.
2625 The Ord instance of Maybe in PrelMaybe.hs, for example, took several extra
2626 iterations because the version with the let bindings looked big, and so wasn't
2627 inlined, but after the join points had been inlined it looked smaller, and so
2628 was inlined.
2629
2630 NB: we have to check the size of rhs', not rhs.
2631 Duplicating a small InAlt might invalidate occurrence information
2632 However, if it *is* dupable, we return the *un* simplified alternative,
2633 because otherwise we'd need to pair it up with an empty subst-env....
2634 but we only have one env shared between all the alts.
2635 (Remember we must zap the subst-env before re-simplifying something).
2636 Rather than do this we simply agree to re-simplify the original (small) thing later.
2637
2638 Note [Funky mkLamTypes]
2639 ~~~~~~~~~~~~~~~~~~~~~~
2640 Notice the funky mkLamTypes. If the constructor has existentials
2641 it's possible that the join point will be abstracted over
2642 type variables as well as term variables.
2643 Example: Suppose we have
2644 data T = forall t. C [t]
2645 Then faced with
2646 case (case e of ...) of
2647 C t xs::[t] -> rhs
2648 We get the join point
2649 let j :: forall t. [t] -> ...
2650 j = /\t \xs::[t] -> rhs
2651 in
2652 case (case e of ...) of
2653 C t xs::[t] -> j t xs
2654
2655 Note [Join point abstraction]
2656 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2657 Join points always have at least one value argument,
2658 for several reasons
2659
2660 * If we try to lift a primitive-typed something out
2661 for let-binding-purposes, we will *caseify* it (!),
2662 with potentially-disastrous strictness results. So
2663 instead we turn it into a function: \v -> e
2664 where v::Void#. The value passed to this function is void,
2665 which generates (almost) no code.
2666
2667 * CPR. We used to say "&& isUnliftedType rhs_ty'" here, but now
2668 we make the join point into a function whenever used_bndrs'
2669 is empty. This makes the join-point more CPR friendly.
2670 Consider: let j = if .. then I# 3 else I# 4
2671 in case .. of { A -> j; B -> j; C -> ... }
2672
2673 Now CPR doesn't w/w j because it's a thunk, so
2674 that means that the enclosing function can't w/w either,
2675 which is a lose. Here's the example that happened in practice:
2676 kgmod :: Int -> Int -> Int
2677 kgmod x y = if x > 0 && y < 0 || x < 0 && y > 0
2678 then 78
2679 else 5
2680
2681 * Let-no-escape. We want a join point to turn into a let-no-escape
2682 so that it is implemented as a jump, and one of the conditions
2683 for LNE is that it's not updatable. In CoreToStg, see
2684 Note [What is a non-escaping let]
2685
2686 * Floating. Since a join point will be entered once, no sharing is
2687 gained by floating out, but something might be lost by doing
2688 so because it might be allocated.
2689
2690 I have seen a case alternative like this:
2691 True -> \v -> ...
2692 It's a bit silly to add the realWorld dummy arg in this case, making
2693 $j = \s v -> ...
2694 True -> $j s
2695 (the \v alone is enough to make CPR happy) but I think it's rare
2696
2697 There's a slight infelicity here: we pass the overall
2698 case_bndr to all the join points if it's used in *any* RHS,
2699 because we don't know its usage in each RHS separately
2700
2701
2702 Note [Duplicating StrictArg]
2703 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2704 The original plan had (where E is a big argument)
2705 e.g. f E [..hole..]
2706 ==> let $j = \a -> f E a
2707 in $j [..hole..]
2708
2709 But this is terrible! Here's an example:
2710 && E (case x of { T -> F; F -> T })
2711 Now, && is strict so we end up simplifying the case with
2712
2713 an ArgOf continuation. If we let-bind it, we get
2714 let $j = \v -> && E v
2715 in simplExpr (case x of { T -> F; F -> T })
2716 (ArgOf (\r -> $j r)
2717 And after simplifying more we get
2718 let $j = \v -> && E v
2719 in case x of { T -> $j F; F -> $j T }
2720 Which is a Very Bad Thing
2721
2722 What we do now is this
2723 f E [..hole..]
2724 ==> let a = E
2725 in f a [..hole..]
2726 Now if the thing in the hole is a case expression (which is when
2727 we'll call mkDupableCont), we'll push the function call into the
2728 branches, which is what we want. Now RULES for f may fire, and
2729 call-pattern specialisation. Here's an example from Trac #3116
2730 go (n+1) (case l of
2731 1 -> bs'
2732 _ -> Chunk p fpc (o+1) (l-1) bs')
2733 If we can push the call for 'go' inside the case, we get
2734 call-pattern specialisation for 'go', which is *crucial* for
2735 this program.
2736
2737 Here is the (&&) example:
2738 && E (case x of { T -> F; F -> T })
2739 ==> let a = E in
2740 case x of { T -> && a F; F -> && a T }
2741 Much better!
2742
2743 Notice that
2744 * Arguments to f *after* the strict one are handled by
2745 the ApplyToVal case of mkDupableCont. Eg
2746 f [..hole..] E
2747
2748 * We can only do the let-binding of E because the function
2749 part of a StrictArg continuation is an explicit syntax
2750 tree. In earlier versions we represented it as a function
2751 (CoreExpr -> CoreEpxr) which we couldn't take apart.
2752
2753 Do *not* duplicate StrictBind and StritArg continuations. We gain
2754 nothing by propagating them into the expressions, and we do lose a
2755 lot.
2756
2757 The desire not to duplicate is the entire reason that
2758 mkDupableCont returns a pair of continuations.
2759
2760 Note [Duplicating StrictBind]
2761 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2762 Unlike StrictArg, there doesn't seem anything to gain from
2763 duplicating a StrictBind continuation, so we don't.
2764
2765
2766 Note [Single-alternative cases]
2767 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2768 This case is just like the ArgOf case. Here's an example:
2769 data T a = MkT !a
2770 ...(MkT (abs x))...
2771 Then we get
2772 case (case x of I# x' ->
2773 case x' <# 0# of
2774 True -> I# (negate# x')
2775 False -> I# x') of y {
2776 DEFAULT -> MkT y
2777 Because the (case x) has only one alternative, we'll transform to
2778 case x of I# x' ->
2779 case (case x' <# 0# of
2780 True -> I# (negate# x')
2781 False -> I# x') of y {
2782 DEFAULT -> MkT y
2783 But now we do *NOT* want to make a join point etc, giving
2784 case x of I# x' ->
2785 let $j = \y -> MkT y
2786 in case x' <# 0# of
2787 True -> $j (I# (negate# x'))
2788 False -> $j (I# x')
2789 In this case the $j will inline again, but suppose there was a big
2790 strict computation enclosing the orginal call to MkT. Then, it won't
2791 "see" the MkT any more, because it's big and won't get duplicated.
2792 And, what is worse, nothing was gained by the case-of-case transform.
2793
2794 So, in circumstances like these, we don't want to build join points
2795 and push the outer case into the branches of the inner one. Instead,
2796 don't duplicate the continuation.
2797
2798 When should we use this strategy? We should not use it on *every*
2799 single-alternative case:
2800 e.g. case (case ....) of (a,b) -> (# a,b #)
2801 Here we must push the outer case into the inner one!
2802 Other choices:
2803
2804 * Match [(DEFAULT,_,_)], but in the common case of Int,
2805 the alternative-filling-in code turned the outer case into
2806 case (...) of y { I# _ -> MkT y }
2807
2808 * Match on single alternative plus (not (isDeadBinder case_bndr))
2809 Rationale: pushing the case inwards won't eliminate the construction.
2810 But there's a risk of
2811 case (...) of y { (a,b) -> let z=(a,b) in ... }
2812 Now y looks dead, but it'll come alive again. Still, this
2813 seems like the best option at the moment.
2814
2815 * Match on single alternative plus (all (isDeadBinder bndrs))
2816 Rationale: this is essentially seq.
2817
2818 * Match when the rhs is *not* duplicable, and hence would lead to a
2819 join point. This catches the disaster-case above. We can test
2820 the *un-simplified* rhs, which is fine. It might get bigger or
2821 smaller after simplification; if it gets smaller, this case might
2822 fire next time round. NB also that we must test contIsDupable
2823 case_cont *too, because case_cont might be big!
2824
2825 HOWEVER: I found that this version doesn't work well, because
2826 we can get let x = case (...) of { small } in ...case x...
2827 When x is inlined into its full context, we find that it was a bad
2828 idea to have pushed the outer case inside the (...) case.
2829
2830 There is a cost to not doing case-of-case; see Trac #10626.
2831
2832 Note [Single-alternative-unlifted]
2833 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2834 Here's another single-alternative where we really want to do case-of-case:
2835
2836 data Mk1 = Mk1 Int# | Mk2 Int#
2837
2838 M1.f =
2839 \r [x_s74 y_s6X]
2840 case
2841 case y_s6X of tpl_s7m {
2842 M1.Mk1 ipv_s70 -> ipv_s70;
2843 M1.Mk2 ipv_s72 -> ipv_s72;
2844 }
2845 of
2846 wild_s7c
2847 { __DEFAULT ->
2848 case
2849 case x_s74 of tpl_s7n {
2850 M1.Mk1 ipv_s77 -> ipv_s77;
2851 M1.Mk2 ipv_s79 -> ipv_s79;
2852 }
2853 of
2854 wild1_s7b
2855 { __DEFAULT -> ==# [wild1_s7b wild_s7c];
2856 };
2857 };
2858
2859 So the outer case is doing *nothing at all*, other than serving as a
2860 join-point. In this case we really want to do case-of-case and decide
2861 whether to use a real join point or just duplicate the continuation:
2862
2863 let $j s7c = case x of
2864 Mk1 ipv77 -> (==) s7c ipv77
2865 Mk1 ipv79 -> (==) s7c ipv79
2866 in
2867 case y of
2868 Mk1 ipv70 -> $j ipv70
2869 Mk2 ipv72 -> $j ipv72
2870
2871 Hence: check whether the case binder's type is unlifted, because then
2872 the outer case is *not* a seq.
2873
2874 ************************************************************************
2875 * *
2876 Unfoldings
2877 * *
2878 ************************************************************************
2879 -}
2880
2881 simplLetUnfolding :: SimplEnv-> TopLevelFlag
2882 -> InId
2883 -> OutExpr
2884 -> Unfolding -> SimplM Unfolding
2885 simplLetUnfolding env top_lvl id new_rhs unf
2886 | isStableUnfolding unf
2887 = simplUnfolding env top_lvl id unf
2888 | otherwise
2889 = is_bottoming `seq` -- See Note [Force bottoming field]
2890 do { dflags <- getDynFlags
2891 ; return (mkUnfolding dflags InlineRhs is_top_lvl is_bottoming new_rhs) }
2892 -- We make an unfolding *even for loop-breakers*.
2893 -- Reason: (a) It might be useful to know that they are WHNF
2894 -- (b) In TidyPgm we currently assume that, if we want to
2895 -- expose the unfolding then indeed we *have* an unfolding
2896 -- to expose. (We could instead use the RHS, but currently
2897 -- we don't.) The simple thing is always to have one.
2898 where
2899 is_top_lvl = isTopLevel top_lvl
2900 is_bottoming = isBottomingId id
2901
2902 simplUnfolding :: SimplEnv-> TopLevelFlag -> InId -> Unfolding -> SimplM Unfolding
2903 -- Note [Setting the new unfolding]
2904 simplUnfolding env top_lvl id unf
2905 = case unf of
2906 NoUnfolding -> return unf
2907 BootUnfolding -> return unf
2908 OtherCon {} -> return unf
2909
2910 DFunUnfolding { df_bndrs = bndrs, df_con = con, df_args = args }
2911 -> do { (env', bndrs') <- simplBinders rule_env bndrs
2912 ; args' <- mapM (simplExpr env') args
2913 ; return (mkDFunUnfolding bndrs' con args') }
2914
2915 CoreUnfolding { uf_tmpl = expr, uf_src = src, uf_guidance = guide }
2916 | isStableSource src
2917 -> do { expr' <- simplExpr rule_env expr
2918 ; case guide of
2919 UnfWhen { ug_arity = arity, ug_unsat_ok = sat_ok } -- Happens for INLINE things
2920 -> let guide' = UnfWhen { ug_arity = arity, ug_unsat_ok = sat_ok
2921 , ug_boring_ok = inlineBoringOk expr' }
2922 -- Refresh the boring-ok flag, in case expr'
2923 -- has got small. This happens, notably in the inlinings
2924 -- for dfuns for single-method classes; see
2925 -- Note [Single-method classes] in TcInstDcls.
2926 -- A test case is Trac #4138
2927 in return (mkCoreUnfolding src is_top_lvl expr' guide')
2928 -- See Note [Top-level flag on inline rules] in CoreUnfold
2929
2930 _other -- Happens for INLINABLE things
2931 -> is_bottoming `seq` -- See Note [Force bottoming field]
2932 do { dflags <- getDynFlags
2933 ; return (mkUnfolding dflags src is_top_lvl is_bottoming expr') } }
2934 -- If the guidance is UnfIfGoodArgs, this is an INLINABLE
2935 -- unfolding, and we need to make sure the guidance is kept up
2936 -- to date with respect to any changes in the unfolding.
2937
2938 | otherwise -> return noUnfolding -- Discard unstable unfoldings
2939 where
2940 is_top_lvl = isTopLevel top_lvl
2941 is_bottoming = isBottomingId id
2942 act = idInlineActivation id
2943 rule_env = updMode (updModeForStableUnfoldings act) env
2944 -- See Note [Simplifying inside stable unfoldings] in SimplUtils
2945
2946 {-
2947 Note [Force bottoming field]
2948 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2949 We need to force bottoming, or the new unfolding holds
2950 on to the old unfolding (which is part of the id).
2951
2952 Note [Setting the new unfolding]
2953 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2954 * If there's an INLINE pragma, we simplify the RHS gently. Maybe we
2955 should do nothing at all, but simplifying gently might get rid of
2956 more crap.
2957
2958 * If not, we make an unfolding from the new RHS. But *only* for
2959 non-loop-breakers. Making loop breakers not have an unfolding at all
2960 means that we can avoid tests in exprIsConApp, for example. This is
2961 important: if exprIsConApp says 'yes' for a recursive thing, then we
2962 can get into an infinite loop
2963
2964 If there's an stable unfolding on a loop breaker (which happens for
2965 INLINABLE), we hang on to the inlining. It's pretty dodgy, but the
2966 user did say 'INLINE'. May need to revisit this choice.
2967
2968 ************************************************************************
2969 * *
2970 Rules
2971 * *
2972 ************************************************************************
2973
2974 Note [Rules in a letrec]
2975 ~~~~~~~~~~~~~~~~~~~~~~~~
2976 After creating fresh binders for the binders of a letrec, we
2977 substitute the RULES and add them back onto the binders; this is done
2978 *before* processing any of the RHSs. This is important. Manuel found
2979 cases where he really, really wanted a RULE for a recursive function
2980 to apply in that function's own right-hand side.
2981
2982 See Note [Loop breaking and RULES] in OccAnal.
2983 -}
2984
2985 addBndrRules :: SimplEnv -> InBndr -> OutBndr -> SimplM (SimplEnv, OutBndr)
2986 -- Rules are added back into the bin
2987 addBndrRules env in_id out_id
2988 | null old_rules
2989 = return (env, out_id)
2990 | otherwise
2991 = do { new_rules <- simplRules env (Just (idName out_id)) old_rules
2992 ; let final_id = out_id `setIdSpecialisation` mkRuleInfo new_rules
2993 ; return (modifyInScope env final_id, final_id) }
2994 where
2995 old_rules = ruleInfoRules (idSpecialisation in_id)
2996
2997 simplRules :: SimplEnv -> Maybe Name -> [CoreRule] -> SimplM [CoreRule]
2998 simplRules env mb_new_nm rules
2999 = mapM simpl_rule rules
3000 where
3001 simpl_rule rule@(BuiltinRule {})
3002 = return rule
3003
3004 simpl_rule rule@(Rule { ru_bndrs = bndrs, ru_args = args
3005 , ru_fn = fn_name, ru_rhs = rhs })
3006 = do { (env', bndrs') <- simplBinders env bndrs
3007 ; let rule_env = updMode updModeForRules env'
3008 ; args' <- mapM (simplExpr rule_env) args
3009 ; rhs' <- simplExpr rule_env rhs
3010 ; return (rule { ru_bndrs = bndrs'
3011 , ru_fn = mb_new_nm `orElse` fn_name
3012 , ru_args = args'
3013 , ru_rhs = rhs' }) }