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