2 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
6 ***************************
8 ***************************
10 1. We attach binding levels to Core bindings, in preparation for floating
11 outwards (@FloatOut@).
13 2. We also let-ify many expressions (notably case scrutinees), so they
14 will have a fighting chance of being floated sensible.
16 3. We clone the binders of any floatable let-binding, so that when it is
17 floated out it will be unique. (This used to be done by the simplifier
18 but the latter now only ensures that there's no shadowing; indeed, even
19 that may not be true.)
21 NOTE: this can't be done using the uniqAway idea, because the variable
22 must be unique in the whole program, not just its current scope,
23 because two variables in different scopes may float out to the
26 NOTE: Very tiresomely, we must apply this substitution to
27 the rules stored inside a variable too.
29 We do *not* clone top-level bindings, because some of them must not change,
30 but we *do* clone bindings that are heading for the top level
33 case x of wild { p -> ...wild... }
34 we substitute x for wild in the RHS of the case alternatives:
35 case x of wild { p -> ...x... }
36 This means that a sub-expression involving x is not "trapped" inside the RHS.
37 And it's not inconvenient because we already have a substitution.
39 Note that this is EXACTLY BACKWARDS from the what the simplifier does.
40 The simplifier tries to get rid of occurrences of x, in favour of wild,
41 in the hope that there will only be one remaining occurrence of x, namely
42 the scrutinee of the case, and we can inline it.
49 LevelledBind, LevelledExpr,
51 incMinorLvl, ltMajLvl, ltLvl, isTopLvl, isInlineCtxt
54 #include "HsVersions.h"
58 import DynFlags ( FloatOutSwitches(..) )
59 import CoreUtils ( exprType, exprIsTrivial, mkPiTypes )
60 import CoreFVs -- all of it
61 import CoreSubst ( Subst, emptySubst, extendInScope, extendIdSubst,
62 cloneIdBndr, cloneRecIdBndrs )
63 import Id ( Id, idType, mkSysLocal, isOneShotLambda,
64 zapDemandIdInfo, transferPolyIdInfo,
65 idSpecialisation, idWorkerInfo, setIdInfo
71 import Name ( getOccName )
72 import OccName ( occNameString )
73 import Type ( isUnLiftedType, Type )
74 import BasicTypes ( TopLevelFlag(..) )
76 import Util ( sortLe, isSingleton, count )
81 %************************************************************************
83 \subsection{Level numbers}
85 %************************************************************************
88 data Level = InlineCtxt -- A level that's used only for
89 -- the context parameter ctxt_lvl
90 | Level Int -- Level number of enclosing lambdas
91 Int -- Number of big-lambda and/or case expressions between
92 -- here and the nearest enclosing lambda
95 The {\em level number} on a (type-)lambda-bound variable is the
96 nesting depth of the (type-)lambda which binds it. The outermost lambda
97 has level 1, so (Level 0 0) means that the variable is bound outside any lambda.
99 On an expression, it's the maximum level number of its free
100 (type-)variables. On a let(rec)-bound variable, it's the level of its
101 RHS. On a case-bound variable, it's the number of enclosing lambdas.
103 Top-level variables: level~0. Those bound on the RHS of a top-level
104 definition but ``before'' a lambda; e.g., the \tr{x} in (levels shown
105 as ``subscripts'')...
107 a_0 = let b_? = ... in
108 x_1 = ... b ... in ...
111 The main function @lvlExpr@ carries a ``context level'' (@ctxt_lvl@).
112 That's meant to be the level number of the enclosing binder in the
113 final (floated) program. If the level number of a sub-expression is
114 less than that of the context, then it might be worth let-binding the
115 sub-expression so that it will indeed float.
117 If you can float to level @Level 0 0@ worth doing so because then your
118 allocation becomes static instead of dynamic. We always start with
122 Note [FloatOut inside INLINE]
123 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
124 @InlineCtxt@ very similar to @Level 0 0@, but is used for one purpose:
125 to say "don't float anything out of here". That's exactly what we
126 want for the body of an INLINE, where we don't want to float anything
127 out at all. See notes with lvlMFE below.
131 -- At one time I tried the effect of not float anything out of an InlineMe,
132 -- but it sometimes works badly. For example, consider PrelArr.done. It
133 -- has the form __inline (\d. e)
134 -- where e doesn't mention d. If we float this to
135 -- __inline (let x = e in \d. x)
136 -- things are bad. The inliner doesn't even inline it because it doesn't look
137 -- like a head-normal form. So it seems a lesser evil to let things float.
138 -- In SetLevels we do set the context to (Level 0 0) when we get to an InlineMe
139 -- which discourages floating out.
141 So the conclusion is: don't do any floating at all inside an InlineMe.
142 (In the above example, don't float the {x=e} out of the \d.)
144 One particular case is that of workers: we don't want to float the
145 call to the worker outside the wrapper, otherwise the worker might get
146 inlined into the floated expression, and an importing module won't see
150 type LevelledExpr = TaggedExpr Level
151 type LevelledBind = TaggedBind Level
153 tOP_LEVEL, iNLINE_CTXT :: Level
154 tOP_LEVEL = Level 0 0
155 iNLINE_CTXT = InlineCtxt
157 incMajorLvl :: Level -> Level
158 -- For InlineCtxt we ignore any inc's; we don't want
159 -- to do any floating at all; see notes above
160 incMajorLvl InlineCtxt = InlineCtxt
161 incMajorLvl (Level major _) = Level (major + 1) 0
163 incMinorLvl :: Level -> Level
164 incMinorLvl InlineCtxt = InlineCtxt
165 incMinorLvl (Level major minor) = Level major (minor+1)
167 maxLvl :: Level -> Level -> Level
168 maxLvl InlineCtxt l2 = l2
169 maxLvl l1 InlineCtxt = l1
170 maxLvl l1@(Level maj1 min1) l2@(Level maj2 min2)
171 | (maj1 > maj2) || (maj1 == maj2 && min1 > min2) = l1
174 ltLvl :: Level -> Level -> Bool
175 ltLvl _ InlineCtxt = False
176 ltLvl InlineCtxt (Level _ _) = True
177 ltLvl (Level maj1 min1) (Level maj2 min2)
178 = (maj1 < maj2) || (maj1 == maj2 && min1 < min2)
180 ltMajLvl :: Level -> Level -> Bool
181 -- Tells if one level belongs to a difft *lambda* level to another
182 ltMajLvl _ InlineCtxt = False
183 ltMajLvl InlineCtxt (Level maj2 _) = 0 < maj2
184 ltMajLvl (Level maj1 _) (Level maj2 _) = maj1 < maj2
186 isTopLvl :: Level -> Bool
187 isTopLvl (Level 0 0) = True
190 isInlineCtxt :: Level -> Bool
191 isInlineCtxt InlineCtxt = True
192 isInlineCtxt _ = False
194 instance Outputable Level where
195 ppr InlineCtxt = text "<INLINE>"
196 ppr (Level maj min) = hcat [ char '<', int maj, char ',', int min, char '>' ]
198 instance Eq Level where
199 InlineCtxt == InlineCtxt = True
200 (Level maj1 min1) == (Level maj2 min2) = maj1 == maj2 && min1 == min2
205 %************************************************************************
207 \subsection{Main level-setting code}
209 %************************************************************************
212 setLevels :: FloatOutSwitches
217 setLevels float_lams binds us
218 = initLvl us (do_them binds)
220 -- "do_them"'s main business is to thread the monad along
221 -- It gives each top binding the same empty envt, because
222 -- things unbound in the envt have level number zero implicitly
223 do_them :: [CoreBind] -> LvlM [LevelledBind]
225 do_them [] = return []
227 (lvld_bind, _) <- lvlTopBind init_env b
228 lvld_binds <- do_them bs
229 return (lvld_bind : lvld_binds)
231 init_env = initialEnv float_lams
233 lvlTopBind :: LevelEnv -> Bind Id -> LvlM (LevelledBind, LevelEnv)
234 lvlTopBind env (NonRec binder rhs)
235 = lvlBind TopLevel tOP_LEVEL env (AnnNonRec binder (freeVars rhs))
236 -- Rhs can have no free vars!
238 lvlTopBind env (Rec pairs)
239 = lvlBind TopLevel tOP_LEVEL env (AnnRec [(b,freeVars rhs) | (b,rhs) <- pairs])
242 %************************************************************************
244 \subsection{Setting expression levels}
246 %************************************************************************
249 lvlExpr :: Level -- ctxt_lvl: Level of enclosing expression
250 -> LevelEnv -- Level of in-scope names/tyvars
251 -> CoreExprWithFVs -- input expression
252 -> LvlM LevelledExpr -- Result expression
255 The @ctxt_lvl@ is, roughly, the level of the innermost enclosing
256 binder. Here's an example
258 v = \x -> ...\y -> let r = case (..x..) of
262 When looking at the rhs of @r@, @ctxt_lvl@ will be 1 because that's
263 the level of @r@, even though it's inside a level-2 @\y@. It's
264 important that @ctxt_lvl@ is 1 and not 2 in @r@'s rhs, because we
265 don't want @lvlExpr@ to turn the scrutinee of the @case@ into an MFE
266 --- because it isn't a *maximal* free expression.
268 If there were another lambda in @r@'s rhs, it would get level-2 as well.
271 lvlExpr _ _ ( _, AnnType ty) = return (Type ty)
272 lvlExpr _ env (_, AnnVar v) = return (lookupVar env v)
273 lvlExpr _ _ (_, AnnLit lit) = return (Lit lit)
275 lvlExpr ctxt_lvl env (_, AnnApp fun arg) = do
277 arg' <- lvlMFE False ctxt_lvl env arg
278 return (App fun' arg')
281 lvl_fun (_, AnnCase _ _ _ _) = lvlMFE True ctxt_lvl env fun
282 lvl_fun _ = lvlExpr ctxt_lvl env fun
283 -- We don't do MFE on partial applications generally,
284 -- but we do if the function is big and hairy, like a case
286 lvlExpr _ env (_, AnnNote InlineMe expr) = do
287 -- Don't float anything out of an InlineMe; hence the iNLINE_CTXT
288 expr' <- lvlExpr iNLINE_CTXT env expr
289 return (Note InlineMe expr')
291 lvlExpr ctxt_lvl env (_, AnnNote note expr) = do
292 expr' <- lvlExpr ctxt_lvl env expr
293 return (Note note expr')
295 lvlExpr ctxt_lvl env (_, AnnCast expr co) = do
296 expr' <- lvlExpr ctxt_lvl env expr
297 return (Cast expr' co)
299 -- We don't split adjacent lambdas. That is, given
301 -- we don't float to give
302 -- \x -> let v = x+y in \y -> (v,y)
303 -- Why not? Because partial applications are fairly rare, and splitting
304 -- lambdas makes them more expensive.
306 lvlExpr ctxt_lvl env expr@(_, AnnLam {}) = do
307 new_body <- lvlMFE True new_lvl new_env body
308 return (mkLams new_bndrs new_body)
310 (bndrs, body) = collectAnnBndrs expr
311 (new_lvl, new_bndrs) = lvlLamBndrs ctxt_lvl bndrs
312 new_env = extendLvlEnv env new_bndrs
313 -- At one time we called a special verion of collectBinders,
314 -- which ignored coercions, because we don't want to split
315 -- a lambda like this (\x -> coerce t (\s -> ...))
316 -- This used to happen quite a bit in state-transformer programs,
317 -- but not nearly so much now non-recursive newtypes are transparent.
318 -- [See SetLevels rev 1.50 for a version with this approach.]
320 lvlExpr ctxt_lvl env (_, AnnLet (AnnNonRec bndr rhs) body)
321 | isUnLiftedType (idType bndr) = do
322 -- Treat unlifted let-bindings (let x = b in e) just like (case b of x -> e)
323 -- That is, leave it exactly where it is
324 -- We used to float unlifted bindings too (e.g. to get a cheap primop
325 -- outside a lambda (to see how, look at lvlBind in rev 1.58)
326 -- but an unrelated change meant that these unlifed bindings
327 -- could get to the top level which is bad. And there's not much point;
328 -- unlifted bindings are always cheap, and so hardly worth floating.
329 rhs' <- lvlExpr ctxt_lvl env rhs
330 body' <- lvlExpr incd_lvl env' body
331 return (Let (NonRec bndr' rhs') body')
333 incd_lvl = incMinorLvl ctxt_lvl
334 bndr' = TB bndr incd_lvl
335 env' = extendLvlEnv env [bndr']
337 lvlExpr ctxt_lvl env (_, AnnLet bind body) = do
338 (bind', new_env) <- lvlBind NotTopLevel ctxt_lvl env bind
339 body' <- lvlExpr ctxt_lvl new_env body
340 return (Let bind' body')
342 lvlExpr ctxt_lvl env (_, AnnCase expr case_bndr ty alts) = do
343 expr' <- lvlMFE True ctxt_lvl env expr
344 let alts_env = extendCaseBndrLvlEnv env expr' case_bndr incd_lvl
345 alts' <- mapM (lvl_alt alts_env) alts
346 return (Case expr' (TB case_bndr incd_lvl) ty alts')
348 incd_lvl = incMinorLvl ctxt_lvl
350 lvl_alt alts_env (con, bs, rhs) = do
351 rhs' <- lvlMFE True incd_lvl new_env rhs
352 return (con, bs', rhs')
354 bs' = [ TB b incd_lvl | b <- bs ]
355 new_env = extendLvlEnv alts_env bs'
358 @lvlMFE@ is just like @lvlExpr@, except that it might let-bind
359 the expression, so that it can itself be floated.
362 ~~~~~~~~~~~~~~~~~~~~~
363 We don't float unlifted MFEs, which potentially loses big opportunites.
366 where h :: Int -> Int# is expensive. We'd like to float the (h y) outside
367 the \x, but we don't because it's unboxed. Possible solution: box it.
370 lvlMFE :: Bool -- True <=> strict context [body of case or let]
371 -> Level -- Level of innermost enclosing lambda/tylam
372 -> LevelEnv -- Level of in-scope names/tyvars
373 -> CoreExprWithFVs -- input expression
374 -> LvlM LevelledExpr -- Result expression
376 lvlMFE _ _ _ (_, AnnType ty)
380 lvlMFE strict_ctxt ctxt_lvl env ann_expr@(fvs, _)
381 | isUnLiftedType ty -- Can't let-bind it; see Note [Unlifted MFEs]
382 || isInlineCtxt ctxt_lvl -- Don't float out of an __inline__ context
383 || exprIsTrivial expr -- Never float if it's trivial
384 || not good_destination
385 = -- Don't float it out
386 lvlExpr ctxt_lvl env ann_expr
388 | otherwise -- Float it out!
389 = do expr' <- lvlFloatRhs abs_vars dest_lvl env ann_expr
390 var <- newLvlVar "lvl" abs_vars ty
391 return (Let (NonRec (TB var dest_lvl) expr')
392 (mkVarApps (Var var) abs_vars))
394 expr = deAnnotate ann_expr
396 dest_lvl = destLevel env fvs (isFunction ann_expr)
397 abs_vars = abstractVars dest_lvl env fvs
399 -- A decision to float entails let-binding this thing, and we only do
400 -- that if we'll escape a value lambda, or will go to the top level.
402 | dest_lvl `ltMajLvl` ctxt_lvl -- Escapes a value lambda
404 -- OLD CODE: not (exprIsCheap expr) || isTopLvl dest_lvl
405 -- see Note [Escaping a value lambda]
407 | otherwise -- Does not escape a value lambda
408 = isTopLvl dest_lvl -- Only float if we are going to the top level
409 && floatConsts env -- and the floatConsts flag is on
410 && not strict_ctxt -- Don't float from a strict context
411 -- We are keen to float something to the top level, even if it does not
412 -- escape a lambda, because then it needs no allocation. But it's controlled
413 -- by a flag, because doing this too early loses opportunities for RULES
414 -- which (needless to say) are important in some nofib programs
415 -- (gcd is an example).
418 -- concat = /\ a -> foldr ..a.. (++) []
419 -- was getting turned into
420 -- concat = /\ a -> lvl a
421 -- lvl = /\ a -> foldr ..a.. (++) []
422 -- which is pretty stupid. Hence the strict_ctxt test
425 Note [Escaping a value lambda]
426 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
427 We want to float even cheap expressions out of value lambdas,
428 because that saves allocation. Consider
429 f = \x. .. (\y.e) ...
430 Then we'd like to avoid allocating the (\y.e) every time we call f,
431 (assuming e does not mention x).
433 An example where this really makes a difference is simplrun009.
435 Another reason it's good is because it makes SpecContr fire on functions.
437 f = \x. ....(f (\y.e))....
438 After floating we get
440 f = \x. ....(f lvl)...
441 and that is much easier for SpecConstr to generate a robust specialisation for.
443 The OLD CODE (given where this Note is referred to) prevents floating
444 of the example above, so I just don't understand the old code. I
445 don't understand the old comment either (which appears below). I
446 measured the effect on nofib of changing OLD CODE to 'True', and got
447 zeros everywhere, but a 4% win for 'puzzle'. Very small 0.5% loss for
448 'cse'; turns out to be because our arity analysis isn't good enough
449 yet (mentioned in Simon-nofib-notes).
452 Even if it escapes a value lambda, we only
453 float if it's not cheap (unless it'll get all the
454 way to the top). I've seen cases where we
455 float dozens of tiny free expressions, which cost
456 more to allocate than to evaluate.
457 NB: exprIsCheap is also true of bottom expressions, which
458 is good; we don't want to share them
460 It's only Really Bad to float a cheap expression out of a
461 strict context, because that builds a thunk that otherwise
462 would never be built. So another alternative would be to
464 || (strict_ctxt && not (exprIsBottom expr))
465 to the condition above. We should really try this out.
468 %************************************************************************
470 \subsection{Bindings}
472 %************************************************************************
474 The binding stuff works for top level too.
477 lvlBind :: TopLevelFlag -- Used solely to decide whether to clone
478 -> Level -- Context level; might be Top even for bindings nested in the RHS
479 -- of a top level binding
482 -> LvlM (LevelledBind, LevelEnv)
484 lvlBind top_lvl ctxt_lvl env (AnnNonRec bndr rhs@(rhs_fvs,_))
485 | isTyVar bndr -- Don't do anything for TyVar binders
486 -- (simplifier gets rid of them pronto)
487 || isInlineCtxt ctxt_lvl -- Don't do anything inside InlineMe
488 = do rhs' <- lvlExpr ctxt_lvl env rhs
489 return (NonRec (TB bndr ctxt_lvl) rhs', env)
492 = do -- No type abstraction; clone existing binder
493 rhs' <- lvlExpr dest_lvl env rhs
494 (env', bndr') <- cloneVar top_lvl env bndr ctxt_lvl dest_lvl
495 return (NonRec (TB bndr' dest_lvl) rhs', env')
498 = do -- Yes, type abstraction; create a new binder, extend substitution, etc
499 rhs' <- lvlFloatRhs abs_vars dest_lvl env rhs
500 (env', [bndr']) <- newPolyBndrs dest_lvl env abs_vars [bndr]
501 return (NonRec (TB bndr' dest_lvl) rhs', env')
504 bind_fvs = rhs_fvs `unionVarSet` idFreeVars bndr
505 abs_vars = abstractVars dest_lvl env bind_fvs
506 dest_lvl = destLevel env bind_fvs (isFunction rhs)
511 lvlBind top_lvl ctxt_lvl env (AnnRec pairs)
512 | isInlineCtxt ctxt_lvl -- Don't do anything inside InlineMe
513 = do rhss' <- mapM (lvlExpr ctxt_lvl env) rhss
514 return (Rec ([TB b ctxt_lvl | b <- bndrs] `zip` rhss'), env)
517 = do (new_env, new_bndrs) <- cloneRecVars top_lvl env bndrs ctxt_lvl dest_lvl
518 new_rhss <- mapM (lvlExpr ctxt_lvl new_env) rhss
519 return (Rec ([TB b dest_lvl | b <- new_bndrs] `zip` new_rhss), new_env)
521 | isSingleton pairs && count isIdVar abs_vars > 1
522 = do -- Special case for self recursion where there are
523 -- several variables carried around: build a local loop:
524 -- poly_f = \abs_vars. \lam_vars . letrec f = \lam_vars. rhs in f lam_vars
525 -- This just makes the closures a bit smaller. If we don't do
526 -- this, allocation rises significantly on some programs
528 -- We could elaborate it for the case where there are several
529 -- mutually functions, but it's quite a bit more complicated
531 -- This all seems a bit ad hoc -- sigh
533 (bndr,rhs) = head pairs
534 (rhs_lvl, abs_vars_w_lvls) = lvlLamBndrs dest_lvl abs_vars
535 rhs_env = extendLvlEnv env abs_vars_w_lvls
536 (rhs_env', new_bndr) <- cloneVar NotTopLevel rhs_env bndr rhs_lvl rhs_lvl
538 (lam_bndrs, rhs_body) = collectAnnBndrs rhs
539 (body_lvl, new_lam_bndrs) = lvlLamBndrs rhs_lvl lam_bndrs
540 body_env = extendLvlEnv rhs_env' new_lam_bndrs
541 new_rhs_body <- lvlExpr body_lvl body_env rhs_body
542 (poly_env, [poly_bndr]) <- newPolyBndrs dest_lvl env abs_vars [bndr]
543 return (Rec [(TB poly_bndr dest_lvl,
544 mkLams abs_vars_w_lvls $
545 mkLams new_lam_bndrs $
546 Let (Rec [(TB new_bndr rhs_lvl, mkLams new_lam_bndrs new_rhs_body)])
547 (mkVarApps (Var new_bndr) lam_bndrs))],
550 | otherwise = do -- Non-null abs_vars
551 (new_env, new_bndrs) <- newPolyBndrs dest_lvl env abs_vars bndrs
552 new_rhss <- mapM (lvlFloatRhs abs_vars dest_lvl new_env) rhss
553 return (Rec ([TB b dest_lvl | b <- new_bndrs] `zip` new_rhss), new_env)
556 (bndrs,rhss) = unzip pairs
558 -- Finding the free vars of the binding group is annoying
559 bind_fvs = (unionVarSets [ idFreeVars bndr `unionVarSet` rhs_fvs
560 | (bndr, (rhs_fvs,_)) <- pairs])
564 dest_lvl = destLevel env bind_fvs (all isFunction rhss)
565 abs_vars = abstractVars dest_lvl env bind_fvs
567 ----------------------------------------------------
568 -- Three help functons for the type-abstraction case
570 lvlFloatRhs :: [CoreBndr] -> Level -> LevelEnv -> CoreExprWithFVs
571 -> UniqSM (Expr (TaggedBndr Level))
572 lvlFloatRhs abs_vars dest_lvl env rhs = do
573 rhs' <- lvlExpr rhs_lvl rhs_env rhs
574 return (mkLams abs_vars_w_lvls rhs')
576 (rhs_lvl, abs_vars_w_lvls) = lvlLamBndrs dest_lvl abs_vars
577 rhs_env = extendLvlEnv env abs_vars_w_lvls
581 %************************************************************************
583 \subsection{Deciding floatability}
585 %************************************************************************
588 lvlLamBndrs :: Level -> [CoreBndr] -> (Level, [TaggedBndr Level])
589 -- Compute the levels for the binders of a lambda group
590 -- The binders returned are exactly the same as the ones passed,
591 -- but they are now paired with a level
595 lvlLamBndrs lvl bndrs
596 = go (incMinorLvl lvl)
597 False -- Havn't bumped major level in this group
600 go old_lvl bumped_major rev_lvld_bndrs (bndr:bndrs)
601 | isIdVar bndr && -- Go to the next major level if this is a value binder,
602 not bumped_major && -- and we havn't already gone to the next level (one jump per group)
603 not (isOneShotLambda bndr) -- and it isn't a one-shot lambda
604 = go new_lvl True (TB bndr new_lvl : rev_lvld_bndrs) bndrs
607 = go old_lvl bumped_major (TB bndr old_lvl : rev_lvld_bndrs) bndrs
610 new_lvl = incMajorLvl old_lvl
612 go old_lvl _ rev_lvld_bndrs []
613 = (old_lvl, reverse rev_lvld_bndrs)
614 -- a lambda like this (\x -> coerce t (\s -> ...))
615 -- This happens quite a bit in state-transformer programs
619 -- Destintion level is the max Id level of the expression
620 -- (We'll abstract the type variables, if any.)
621 destLevel :: LevelEnv -> VarSet -> Bool -> Level
622 destLevel env fvs is_function
624 && is_function = tOP_LEVEL -- Send functions to top level; see
625 -- the comments with isFunction
626 | otherwise = maxIdLevel env fvs
628 isFunction :: CoreExprWithFVs -> Bool
629 -- The idea here is that we want to float *functions* to
630 -- the top level. This saves no work, but
631 -- (a) it can make the host function body a lot smaller,
632 -- and hence inlinable.
633 -- (b) it can also save allocation when the function is recursive:
634 -- h = \x -> letrec f = \y -> ...f...y...x...
637 -- f = \x y -> ...(f x)...y...x...
639 -- No allocation for f now.
640 -- We may only want to do this if there are sufficiently few free
641 -- variables. We certainly only want to do it for values, and not for
642 -- constructors. So the simple thing is just to look for lambdas
643 isFunction (_, AnnLam b e) | isIdVar b = True
644 | otherwise = isFunction e
645 isFunction (_, AnnNote _ e) = isFunction e
650 %************************************************************************
652 \subsection{Free-To-Level Monad}
654 %************************************************************************
657 type LevelEnv = (FloatOutSwitches,
658 VarEnv Level, -- Domain is *post-cloned* TyVars and Ids
659 Subst, -- Domain is pre-cloned Ids; tracks the in-scope set
660 -- so that subtitution is capture-avoiding
661 IdEnv ([Var], LevelledExpr)) -- Domain is pre-cloned Ids
662 -- We clone let-bound variables so that they are still
663 -- distinct when floated out; hence the SubstEnv/IdEnv.
664 -- (see point 3 of the module overview comment).
665 -- We also use these envs when making a variable polymorphic
666 -- because we want to float it out past a big lambda.
668 -- The Subst and IdEnv always implement the same mapping, but the
669 -- Subst maps to CoreExpr and the IdEnv to LevelledExpr
670 -- Since the range is always a variable or type application,
671 -- there is never any difference between the two, but sadly
672 -- the types differ. The SubstEnv is used when substituting in
673 -- a variable's IdInfo; the IdEnv when we find a Var.
675 -- In addition the IdEnv records a list of tyvars free in the
676 -- type application, just so we don't have to call freeVars on
677 -- the type application repeatedly.
679 -- The domain of the both envs is *pre-cloned* Ids, though
681 -- The domain of the VarEnv Level is the *post-cloned* Ids
683 initialEnv :: FloatOutSwitches -> LevelEnv
684 initialEnv float_lams = (float_lams, emptyVarEnv, emptySubst, emptyVarEnv)
686 floatLams :: LevelEnv -> Bool
687 floatLams (FloatOutSw float_lams _, _, _, _) = float_lams
689 floatConsts :: LevelEnv -> Bool
690 floatConsts (FloatOutSw _ float_consts, _, _, _) = float_consts
692 extendLvlEnv :: LevelEnv -> [TaggedBndr Level] -> LevelEnv
693 -- Used when *not* cloning
694 extendLvlEnv (float_lams, lvl_env, subst, id_env) prs
696 foldl add_lvl lvl_env prs,
697 foldl del_subst subst prs,
698 foldl del_id id_env prs)
700 add_lvl env (TB v l) = extendVarEnv env v l
701 del_subst env (TB v _) = extendInScope env v
702 del_id env (TB v _) = delVarEnv env v
703 -- We must remove any clone for this variable name in case of
704 -- shadowing. This bit me in the following case
705 -- (in nofib/real/gg/Spark.hs):
708 -- ... -> case e of wild {
709 -- ... -> ... wild ...
713 -- The inside occurrence of @wild@ was being replaced with @ds@,
714 -- incorrectly, because the SubstEnv was still lying around. Ouch!
717 -- extendCaseBndrLvlEnv adds the mapping case-bndr->scrut-var if it can
718 -- (see point 4 of the module overview comment)
719 extendCaseBndrLvlEnv :: LevelEnv -> Expr (TaggedBndr Level) -> Var -> Level
721 extendCaseBndrLvlEnv (float_lams, lvl_env, subst, id_env) (Var scrut_var) case_bndr lvl
723 extendVarEnv lvl_env case_bndr lvl,
724 extendIdSubst subst case_bndr (Var scrut_var),
725 extendVarEnv id_env case_bndr ([scrut_var], Var scrut_var))
727 extendCaseBndrLvlEnv env _scrut case_bndr lvl
728 = extendLvlEnv env [TB case_bndr lvl]
730 extendPolyLvlEnv :: Level -> LevelEnv -> [Var] -> [(Var, Var)] -> LevelEnv
731 extendPolyLvlEnv dest_lvl (float_lams, lvl_env, subst, id_env) abs_vars bndr_pairs
733 foldl add_lvl lvl_env bndr_pairs,
734 foldl add_subst subst bndr_pairs,
735 foldl add_id id_env bndr_pairs)
737 add_lvl env (_, v') = extendVarEnv env v' dest_lvl
738 add_subst env (v, v') = extendIdSubst env v (mkVarApps (Var v') abs_vars)
739 add_id env (v, v') = extendVarEnv env v ((v':abs_vars), mkVarApps (Var v') abs_vars)
741 extendCloneLvlEnv :: Level -> LevelEnv -> Subst -> [(Var, Var)] -> LevelEnv
742 extendCloneLvlEnv lvl (float_lams, lvl_env, _, id_env) new_subst bndr_pairs
744 foldl add_lvl lvl_env bndr_pairs,
746 foldl add_id id_env bndr_pairs)
748 add_lvl env (_, v') = extendVarEnv env v' lvl
749 add_id env (v, v') = extendVarEnv env v ([v'], Var v')
752 maxIdLevel :: LevelEnv -> VarSet -> Level
753 maxIdLevel (_, lvl_env,_,id_env) var_set
754 = foldVarSet max_in tOP_LEVEL var_set
756 max_in in_var lvl = foldr max_out lvl (case lookupVarEnv id_env in_var of
757 Just (abs_vars, _) -> abs_vars
761 | isIdVar out_var = case lookupVarEnv lvl_env out_var of
762 Just lvl' -> maxLvl lvl' lvl
764 | otherwise = lvl -- Ignore tyvars in *maxIdLevel*
766 lookupVar :: LevelEnv -> Id -> LevelledExpr
767 lookupVar (_, _, _, id_env) v = case lookupVarEnv id_env v of
768 Just (_, expr) -> expr
771 abstractVars :: Level -> LevelEnv -> VarSet -> [Var]
772 -- Find the variables in fvs, free vars of the target expresion,
773 -- whose level is greater than the destination level
774 -- These are the ones we are going to abstract out
775 abstractVars dest_lvl (_, lvl_env, _, id_env) fvs
776 = map zap $ uniq $ sortLe le
777 [var | fv <- varSetElems fvs
778 , var <- absVarsOf id_env fv
780 -- NB: it's important to call abstract_me only on the OutIds the
781 -- come from absVarsOf (not on fv, which is an InId)
783 -- Sort the variables so the true type variables come first;
784 -- the tyvars scope over Ids and coercion vars
785 v1 `le` v2 = case (is_tv v1, is_tv v2) of
786 (True, False) -> True
787 (False, True) -> False
788 _ -> v1 <= v2 -- Same family
790 is_tv v = isTyVar v && not (isCoVar v)
792 uniq :: [Var] -> [Var]
793 -- Remove adjacent duplicates; the sort will have brought them together
794 uniq (v1:v2:vs) | v1 == v2 = uniq (v2:vs)
795 | otherwise = v1 : uniq (v2:vs)
798 abstract_me v = case lookupVarEnv lvl_env v of
799 Just lvl -> dest_lvl `ltLvl` lvl
802 -- We are going to lambda-abstract, so nuke any IdInfo,
803 -- and add the tyvars of the Id (if necessary)
804 zap v | isIdVar v = WARN( workerExists (idWorkerInfo v) ||
805 not (isEmptySpecInfo (idSpecialisation v)),
806 text "absVarsOf: discarding info on" <+> ppr v )
807 setIdInfo v vanillaIdInfo
810 absVarsOf :: IdEnv ([Var], LevelledExpr) -> Var -> [Var]
811 -- If f is free in the expression, and f maps to poly_f a b c in the
812 -- current substitution, then we must report a b c as candidate type
815 -- Also, if x::a is an abstracted variable, then so is a; that is,
816 -- we must look in x's type
817 -- And similarly if x is a coercion variable.
819 | isIdVar v = [av2 | av1 <- lookup_avs v
820 , av2 <- add_tyvars av1]
821 | isCoVar v = add_tyvars v
825 lookup_avs v = case lookupVarEnv id_env v of
826 Just (abs_vars, _) -> abs_vars
829 add_tyvars v = v : varSetElems (varTypeTyVars v)
833 type LvlM result = UniqSM result
835 initLvl :: UniqSupply -> UniqSM a -> a
841 newPolyBndrs :: Level -> LevelEnv -> [Var] -> [Id] -> UniqSM (LevelEnv, [Id])
842 newPolyBndrs dest_lvl env abs_vars bndrs = do
844 let new_bndrs = zipWith mk_poly_bndr bndrs uniqs
845 return (extendPolyLvlEnv dest_lvl env abs_vars (bndrs `zip` new_bndrs), new_bndrs)
847 mk_poly_bndr bndr uniq = transferPolyIdInfo bndr $ -- Note [transferPolyIdInfo] in Id.lhs
848 mkSysLocal (mkFastString str) uniq poly_ty
850 str = "poly_" ++ occNameString (getOccName bndr)
851 poly_ty = mkPiTypes abs_vars (idType bndr)
854 -> [CoreBndr] -> Type -- Abstract wrt these bndrs
856 newLvlVar str vars body_ty = do
858 return (mkSysLocal (mkFastString str) uniq (mkPiTypes vars body_ty))
860 -- The deeply tiresome thing is that we have to apply the substitution
861 -- to the rules inside each Id. Grr. But it matters.
863 cloneVar :: TopLevelFlag -> LevelEnv -> Id -> Level -> Level -> LvlM (LevelEnv, Id)
864 cloneVar TopLevel env v _ _
865 = return (env, v) -- Don't clone top level things
866 cloneVar NotTopLevel env@(_,_,subst,_) v ctxt_lvl dest_lvl
867 = ASSERT( isIdVar v ) do
868 us <- getUniqueSupplyM
870 (subst', v1) = cloneIdBndr subst us v
871 v2 = zap_demand ctxt_lvl dest_lvl v1
872 env' = extendCloneLvlEnv dest_lvl env subst' [(v,v2)]
875 cloneRecVars :: TopLevelFlag -> LevelEnv -> [Id] -> Level -> Level -> LvlM (LevelEnv, [Id])
876 cloneRecVars TopLevel env vs _ _
877 = return (env, vs) -- Don't clone top level things
878 cloneRecVars NotTopLevel env@(_,_,subst,_) vs ctxt_lvl dest_lvl
879 = ASSERT( all isIdVar vs ) do
880 us <- getUniqueSupplyM
882 (subst', vs1) = cloneRecIdBndrs subst us vs
883 vs2 = map (zap_demand ctxt_lvl dest_lvl) vs1
884 env' = extendCloneLvlEnv dest_lvl env subst' (vs `zip` vs2)
887 -- VERY IMPORTANT: we must zap the demand info
888 -- if the thing is going to float out past a lambda,
889 -- or if it's going to top level (where things can't be strict)
890 zap_demand :: Level -> Level -> Id -> Id
891 zap_demand dest_lvl ctxt_lvl id
892 | ctxt_lvl == dest_lvl,
893 not (isTopLvl dest_lvl) = id -- Stays, and not going to top level
894 | otherwise = zapDemandIdInfo id -- Floats out