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.
46 -- The above warning supression flag is a temporary kludge.
47 -- While working on this module you are encouraged to remove it and fix
48 -- any warnings in the module. See
49 -- http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings
56 LevelledBind, LevelledExpr,
58 incMinorLvl, ltMajLvl, ltLvl, isTopLvl, isInlineCtxt
61 #include "HsVersions.h"
65 import DynFlags ( FloatOutSwitches(..) )
66 import CoreUtils ( exprType, exprIsTrivial, mkPiTypes )
67 import CoreFVs -- all of it
68 import CoreSubst ( Subst, emptySubst, extendInScope, extendIdSubst,
69 cloneIdBndr, cloneRecIdBndrs )
70 import Id ( Id, idType, mkSysLocal, isOneShotLambda,
72 idSpecialisation, idWorkerInfo, setIdInfo
74 import IdInfo ( workerExists, vanillaIdInfo, isEmptySpecInfo )
78 import Name ( getOccName )
79 import OccName ( occNameString )
80 import Type ( isUnLiftedType, Type )
81 import BasicTypes ( TopLevelFlag(..) )
83 import Util ( sortLe, isSingleton, count )
88 %************************************************************************
90 \subsection{Level numbers}
92 %************************************************************************
95 data Level = InlineCtxt -- A level that's used only for
96 -- the context parameter ctxt_lvl
97 | Level Int -- Level number of enclosing lambdas
98 Int -- Number of big-lambda and/or case expressions between
99 -- here and the nearest enclosing lambda
102 The {\em level number} on a (type-)lambda-bound variable is the
103 nesting depth of the (type-)lambda which binds it. The outermost lambda
104 has level 1, so (Level 0 0) means that the variable is bound outside any lambda.
106 On an expression, it's the maximum level number of its free
107 (type-)variables. On a let(rec)-bound variable, it's the level of its
108 RHS. On a case-bound variable, it's the number of enclosing lambdas.
110 Top-level variables: level~0. Those bound on the RHS of a top-level
111 definition but ``before'' a lambda; e.g., the \tr{x} in (levels shown
112 as ``subscripts'')...
114 a_0 = let b_? = ... in
115 x_1 = ... b ... in ...
118 The main function @lvlExpr@ carries a ``context level'' (@ctxt_lvl@).
119 That's meant to be the level number of the enclosing binder in the
120 final (floated) program. If the level number of a sub-expression is
121 less than that of the context, then it might be worth let-binding the
122 sub-expression so that it will indeed float.
124 If you can float to level @Level 0 0@ worth doing so because then your
125 allocation becomes static instead of dynamic. We always start with
131 @InlineCtxt@ very similar to @Level 0 0@, but is used for one purpose:
132 to say "don't float anything out of here". That's exactly what we
133 want for the body of an INLINE, where we don't want to float anything
134 out at all. See notes with lvlMFE below.
138 -- At one time I tried the effect of not float anything out of an InlineMe,
139 -- but it sometimes works badly. For example, consider PrelArr.done. It
140 -- has the form __inline (\d. e)
141 -- where e doesn't mention d. If we float this to
142 -- __inline (let x = e in \d. x)
143 -- things are bad. The inliner doesn't even inline it because it doesn't look
144 -- like a head-normal form. So it seems a lesser evil to let things float.
145 -- In SetLevels we do set the context to (Level 0 0) when we get to an InlineMe
146 -- which discourages floating out.
148 So the conclusion is: don't do any floating at all inside an InlineMe.
149 (In the above example, don't float the {x=e} out of the \d.)
151 One particular case is that of workers: we don't want to float the
152 call to the worker outside the wrapper, otherwise the worker might get
153 inlined into the floated expression, and an importing module won't see
157 type LevelledExpr = TaggedExpr Level
158 type LevelledBind = TaggedBind Level
160 tOP_LEVEL = Level 0 0
161 iNLINE_CTXT = InlineCtxt
163 incMajorLvl :: Level -> Level
164 -- For InlineCtxt we ignore any inc's; we don't want
165 -- to do any floating at all; see notes above
166 incMajorLvl InlineCtxt = InlineCtxt
167 incMajorLvl (Level major minor) = Level (major+1) 0
169 incMinorLvl :: Level -> Level
170 incMinorLvl InlineCtxt = InlineCtxt
171 incMinorLvl (Level major minor) = Level major (minor+1)
173 maxLvl :: Level -> Level -> Level
174 maxLvl InlineCtxt l2 = l2
175 maxLvl l1 InlineCtxt = l1
176 maxLvl l1@(Level maj1 min1) l2@(Level maj2 min2)
177 | (maj1 > maj2) || (maj1 == maj2 && min1 > min2) = l1
180 ltLvl :: Level -> Level -> Bool
181 ltLvl any_lvl InlineCtxt = False
182 ltLvl InlineCtxt (Level _ _) = True
183 ltLvl (Level maj1 min1) (Level maj2 min2)
184 = (maj1 < maj2) || (maj1 == maj2 && min1 < min2)
186 ltMajLvl :: Level -> Level -> Bool
187 -- Tells if one level belongs to a difft *lambda* level to another
188 ltMajLvl any_lvl InlineCtxt = False
189 ltMajLvl InlineCtxt (Level maj2 _) = 0 < maj2
190 ltMajLvl (Level maj1 _) (Level maj2 _) = maj1 < maj2
192 isTopLvl :: Level -> Bool
193 isTopLvl (Level 0 0) = True
194 isTopLvl other = False
196 isInlineCtxt :: Level -> Bool
197 isInlineCtxt InlineCtxt = True
198 isInlineCtxt other = False
200 instance Outputable Level where
201 ppr InlineCtxt = text "<INLINE>"
202 ppr (Level maj min) = hcat [ char '<', int maj, char ',', int min, char '>' ]
204 instance Eq Level where
205 InlineCtxt == InlineCtxt = True
206 (Level maj1 min1) == (Level maj2 min2) = maj1==maj2 && min1==min2
211 %************************************************************************
213 \subsection{Main level-setting code}
215 %************************************************************************
218 setLevels :: FloatOutSwitches
223 setLevels float_lams binds us
224 = initLvl us (do_them binds)
226 -- "do_them"'s main business is to thread the monad along
227 -- It gives each top binding the same empty envt, because
228 -- things unbound in the envt have level number zero implicitly
229 do_them :: [CoreBind] -> LvlM [LevelledBind]
231 do_them [] = returnLvl []
233 = lvlTopBind init_env b `thenLvl` \ (lvld_bind, _) ->
234 do_them bs `thenLvl` \ lvld_binds ->
235 returnLvl (lvld_bind : lvld_binds)
237 init_env = initialEnv float_lams
239 lvlTopBind env (NonRec binder rhs)
240 = lvlBind TopLevel tOP_LEVEL env (AnnNonRec binder (freeVars rhs))
241 -- Rhs can have no free vars!
243 lvlTopBind env (Rec pairs)
244 = lvlBind TopLevel tOP_LEVEL env (AnnRec [(b,freeVars rhs) | (b,rhs) <- pairs])
247 %************************************************************************
249 \subsection{Setting expression levels}
251 %************************************************************************
254 lvlExpr :: Level -- ctxt_lvl: Level of enclosing expression
255 -> LevelEnv -- Level of in-scope names/tyvars
256 -> CoreExprWithFVs -- input expression
257 -> LvlM LevelledExpr -- Result expression
260 The @ctxt_lvl@ is, roughly, the level of the innermost enclosing
261 binder. Here's an example
263 v = \x -> ...\y -> let r = case (..x..) of
267 When looking at the rhs of @r@, @ctxt_lvl@ will be 1 because that's
268 the level of @r@, even though it's inside a level-2 @\y@. It's
269 important that @ctxt_lvl@ is 1 and not 2 in @r@'s rhs, because we
270 don't want @lvlExpr@ to turn the scrutinee of the @case@ into an MFE
271 --- because it isn't a *maximal* free expression.
273 If there were another lambda in @r@'s rhs, it would get level-2 as well.
276 lvlExpr _ _ (_, AnnType ty) = returnLvl (Type ty)
277 lvlExpr _ env (_, AnnVar v) = returnLvl (lookupVar env v)
278 lvlExpr _ env (_, AnnLit lit) = returnLvl (Lit lit)
280 lvlExpr ctxt_lvl env (_, AnnApp fun arg)
281 = lvl_fun fun `thenLvl` \ fun' ->
282 lvlMFE False ctxt_lvl env arg `thenLvl` \ arg' ->
283 returnLvl (App fun' arg')
286 lvl_fun (_, AnnCase _ _ _ _) = lvlMFE True ctxt_lvl env fun
287 lvl_fun other = lvlExpr ctxt_lvl env fun
288 -- We don't do MFE on partial applications generally,
289 -- but we do if the function is big and hairy, like a case
291 lvlExpr ctxt_lvl env (_, AnnNote InlineMe expr)
292 -- Don't float anything out of an InlineMe; hence the iNLINE_CTXT
293 = lvlExpr iNLINE_CTXT env expr `thenLvl` \ expr' ->
294 returnLvl (Note InlineMe expr')
296 lvlExpr ctxt_lvl env (_, AnnNote note expr)
297 = lvlExpr ctxt_lvl env expr `thenLvl` \ expr' ->
298 returnLvl (Note note expr')
300 lvlExpr ctxt_lvl env (_, AnnCast expr co)
301 = lvlExpr ctxt_lvl env expr `thenLvl` \ expr' ->
302 returnLvl (Cast expr' co)
304 -- We don't split adjacent lambdas. That is, given
306 -- we don't float to give
307 -- \x -> let v = x+y in \y -> (v,y)
308 -- Why not? Because partial applications are fairly rare, and splitting
309 -- lambdas makes them more expensive.
311 lvlExpr ctxt_lvl env expr@(_, AnnLam bndr rhs)
312 = lvlMFE True new_lvl new_env body `thenLvl` \ new_body ->
313 returnLvl (mkLams new_bndrs new_body)
315 (bndrs, body) = collectAnnBndrs expr
316 (new_lvl, new_bndrs) = lvlLamBndrs ctxt_lvl bndrs
317 new_env = extendLvlEnv env new_bndrs
318 -- At one time we called a special verion of collectBinders,
319 -- which ignored coercions, because we don't want to split
320 -- a lambda like this (\x -> coerce t (\s -> ...))
321 -- This used to happen quite a bit in state-transformer programs,
322 -- but not nearly so much now non-recursive newtypes are transparent.
323 -- [See SetLevels rev 1.50 for a version with this approach.]
325 lvlExpr ctxt_lvl env (_, AnnLet (AnnNonRec bndr rhs) body)
326 | isUnLiftedType (idType bndr)
327 -- Treat unlifted let-bindings (let x = b in e) just like (case b of x -> e)
328 -- That is, leave it exactly where it is
329 -- We used to float unlifted bindings too (e.g. to get a cheap primop
330 -- outside a lambda (to see how, look at lvlBind in rev 1.58)
331 -- but an unrelated change meant that these unlifed bindings
332 -- could get to the top level which is bad. And there's not much point;
333 -- unlifted bindings are always cheap, and so hardly worth floating.
334 = lvlExpr ctxt_lvl env rhs `thenLvl` \ rhs' ->
335 lvlExpr incd_lvl env' body `thenLvl` \ body' ->
336 returnLvl (Let (NonRec bndr' rhs') body')
338 incd_lvl = incMinorLvl ctxt_lvl
339 bndr' = TB bndr incd_lvl
340 env' = extendLvlEnv env [bndr']
342 lvlExpr ctxt_lvl env (_, AnnLet bind body)
343 = lvlBind NotTopLevel ctxt_lvl env bind `thenLvl` \ (bind', new_env) ->
344 lvlExpr ctxt_lvl new_env body `thenLvl` \ body' ->
345 returnLvl (Let bind' body')
347 lvlExpr ctxt_lvl env (_, AnnCase expr case_bndr ty alts)
348 = lvlMFE True ctxt_lvl env expr `thenLvl` \ expr' ->
350 alts_env = extendCaseBndrLvlEnv env expr' case_bndr incd_lvl
352 mapLvl (lvl_alt alts_env) alts `thenLvl` \ alts' ->
353 returnLvl (Case expr' (TB case_bndr incd_lvl) ty alts')
355 incd_lvl = incMinorLvl ctxt_lvl
357 lvl_alt alts_env (con, bs, rhs)
358 = lvlMFE True incd_lvl new_env rhs `thenLvl` \ rhs' ->
359 returnLvl (con, bs', rhs')
361 bs' = [ TB b incd_lvl | b <- bs ]
362 new_env = extendLvlEnv alts_env bs'
365 @lvlMFE@ is just like @lvlExpr@, except that it might let-bind
366 the expression, so that it can itself be floated.
368 [NOTE: unlifted MFEs]
369 We don't float unlifted MFEs, which potentially loses big opportunites.
372 where h :: Int -> Int# is expensive. We'd like to float the (h y) outside
373 the \x, but we don't because it's unboxed. Possible solution: box it.
376 lvlMFE :: Bool -- True <=> strict context [body of case or let]
377 -> Level -- Level of innermost enclosing lambda/tylam
378 -> LevelEnv -- Level of in-scope names/tyvars
379 -> CoreExprWithFVs -- input expression
380 -> LvlM LevelledExpr -- Result expression
382 lvlMFE strict_ctxt ctxt_lvl env (_, AnnType ty)
383 = returnLvl (Type ty)
386 lvlMFE strict_ctxt ctxt_lvl env ann_expr@(fvs, _)
387 | isUnLiftedType ty -- Can't let-bind it; see [NOTE: unlifted MFEs]
388 || isInlineCtxt ctxt_lvl -- Don't float out of an __inline__ context
389 || exprIsTrivial expr -- Never float if it's trivial
390 || not good_destination
391 = -- Don't float it out
392 lvlExpr ctxt_lvl env ann_expr
394 | otherwise -- Float it out!
395 = lvlFloatRhs abs_vars dest_lvl env ann_expr `thenLvl` \ expr' ->
396 newLvlVar "lvl" abs_vars ty `thenLvl` \ var ->
397 returnLvl (Let (NonRec (TB var dest_lvl) expr')
398 (mkVarApps (Var var) abs_vars))
400 expr = deAnnotate ann_expr
402 dest_lvl = destLevel env fvs (isFunction ann_expr)
403 abs_vars = abstractVars dest_lvl env fvs
405 -- A decision to float entails let-binding this thing, and we only do
406 -- that if we'll escape a value lambda, or will go to the top level.
408 | dest_lvl `ltMajLvl` ctxt_lvl -- Escapes a value lambda
410 -- OLD CODE: not (exprIsCheap expr) || isTopLvl dest_lvl
411 -- see Note [Escaping a value lambda]
413 | otherwise -- Does not escape a value lambda
414 = isTopLvl dest_lvl -- Only float if we are going to the top level
415 && floatConsts env -- and the floatConsts flag is on
416 && not strict_ctxt -- Don't float from a strict context
417 -- We are keen to float something to the top level, even if it does not
418 -- escape a lambda, because then it needs no allocation. But it's controlled
419 -- by a flag, because doing this too early loses opportunities for RULES
420 -- which (needless to say) are important in some nofib programs
421 -- (gcd is an example).
424 -- concat = /\ a -> foldr ..a.. (++) []
425 -- was getting turned into
426 -- concat = /\ a -> lvl a
427 -- lvl = /\ a -> foldr ..a.. (++) []
428 -- which is pretty stupid. Hence the strict_ctxt test
431 Note [Escaping a value lambda]
432 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
433 We want to float even cheap expressions out of value lambdas,
434 because that saves allocation. Consider
435 f = \x. .. (\y.e) ...
436 Then we'd like to avoid allocating the (\y.e) every time we call f,
437 (assuming e does not mention x).
439 An example where this really makes a difference is simplrun009.
441 Another reason it's good is because it makes SpecContr fire on functions.
443 f = \x. ....(f (\y.e))....
444 After floating we get
446 f = \x. ....(f lvl)...
447 and that is much easier for SpecConstr to generate a robust specialisation for.
449 The OLD CODE (given where this Note is referred to) prevents floating
450 of the example above, so I just don't understand the old code. I
451 don't understand the old comment either (which appears below). I
452 measured the effect on nofib of changing OLD CODE to 'True', and got
453 zeros everywhere, but a 4% win for 'puzzle'. Very small 0.5% loss for
454 'cse'; turns out to be because our arity analysis isn't good enough
455 yet (mentioned in Simon-nofib-notes).
458 Even if it escapes a value lambda, we only
459 float if it's not cheap (unless it'll get all the
460 way to the top). I've seen cases where we
461 float dozens of tiny free expressions, which cost
462 more to allocate than to evaluate.
463 NB: exprIsCheap is also true of bottom expressions, which
464 is good; we don't want to share them
466 It's only Really Bad to float a cheap expression out of a
467 strict context, because that builds a thunk that otherwise
468 would never be built. So another alternative would be to
470 || (strict_ctxt && not (exprIsBottom expr))
471 to the condition above. We should really try this out.
474 %************************************************************************
476 \subsection{Bindings}
478 %************************************************************************
480 The binding stuff works for top level too.
483 lvlBind :: TopLevelFlag -- Used solely to decide whether to clone
484 -> Level -- Context level; might be Top even for bindings nested in the RHS
485 -- of a top level binding
488 -> LvlM (LevelledBind, LevelEnv)
490 lvlBind top_lvl ctxt_lvl env (AnnNonRec bndr rhs@(rhs_fvs,_))
491 | isInlineCtxt ctxt_lvl -- Don't do anything inside InlineMe
492 = lvlExpr ctxt_lvl env rhs `thenLvl` \ rhs' ->
493 returnLvl (NonRec (TB bndr ctxt_lvl) rhs', env)
496 = -- No type abstraction; clone existing binder
497 lvlExpr dest_lvl env rhs `thenLvl` \ rhs' ->
498 cloneVar top_lvl env bndr ctxt_lvl dest_lvl `thenLvl` \ (env', bndr') ->
499 returnLvl (NonRec (TB bndr' dest_lvl) rhs', env')
502 = -- Yes, type abstraction; create a new binder, extend substitution, etc
503 lvlFloatRhs abs_vars dest_lvl env rhs `thenLvl` \ rhs' ->
504 newPolyBndrs dest_lvl env abs_vars [bndr] `thenLvl` \ (env', [bndr']) ->
505 returnLvl (NonRec (TB bndr' dest_lvl) rhs', env')
508 bind_fvs = rhs_fvs `unionVarSet` idFreeVars bndr
509 abs_vars = abstractVars dest_lvl env bind_fvs
510 dest_lvl = destLevel env bind_fvs (isFunction rhs)
515 lvlBind top_lvl ctxt_lvl env (AnnRec pairs)
516 | isInlineCtxt ctxt_lvl -- Don't do anything inside InlineMe
517 = mapLvl (lvlExpr ctxt_lvl env) rhss `thenLvl` \ rhss' ->
518 returnLvl (Rec ([TB b ctxt_lvl | b <- bndrs] `zip` rhss'), env)
521 = cloneRecVars top_lvl env bndrs ctxt_lvl dest_lvl `thenLvl` \ (new_env, new_bndrs) ->
522 mapLvl (lvlExpr ctxt_lvl new_env) rhss `thenLvl` \ new_rhss ->
523 returnLvl (Rec ([TB b dest_lvl | b <- new_bndrs] `zip` new_rhss), new_env)
525 | isSingleton pairs && count isId abs_vars > 1
526 = -- Special case for self recursion where there are
527 -- several variables carried around: build a local loop:
528 -- poly_f = \abs_vars. \lam_vars . letrec f = \lam_vars. rhs in f lam_vars
529 -- This just makes the closures a bit smaller. If we don't do
530 -- this, allocation rises significantly on some programs
532 -- We could elaborate it for the case where there are several
533 -- mutually functions, but it's quite a bit more complicated
535 -- This all seems a bit ad hoc -- sigh
537 (bndr,rhs) = head pairs
538 (rhs_lvl, abs_vars_w_lvls) = lvlLamBndrs dest_lvl abs_vars
539 rhs_env = extendLvlEnv env abs_vars_w_lvls
541 cloneVar NotTopLevel rhs_env bndr rhs_lvl rhs_lvl `thenLvl` \ (rhs_env', new_bndr) ->
543 (lam_bndrs, rhs_body) = collectAnnBndrs rhs
544 (body_lvl, new_lam_bndrs) = lvlLamBndrs rhs_lvl lam_bndrs
545 body_env = extendLvlEnv rhs_env' new_lam_bndrs
547 lvlExpr body_lvl body_env rhs_body `thenLvl` \ new_rhs_body ->
548 newPolyBndrs dest_lvl env abs_vars [bndr] `thenLvl` \ (poly_env, [poly_bndr]) ->
549 returnLvl (Rec [(TB poly_bndr dest_lvl,
550 mkLams abs_vars_w_lvls $
551 mkLams new_lam_bndrs $
552 Let (Rec [(TB new_bndr rhs_lvl, mkLams new_lam_bndrs new_rhs_body)])
553 (mkVarApps (Var new_bndr) lam_bndrs))],
556 | otherwise -- Non-null abs_vars
557 = newPolyBndrs dest_lvl env abs_vars bndrs `thenLvl` \ (new_env, new_bndrs) ->
558 mapLvl (lvlFloatRhs abs_vars dest_lvl new_env) rhss `thenLvl` \ new_rhss ->
559 returnLvl (Rec ([TB b dest_lvl | b <- new_bndrs] `zip` new_rhss), new_env)
562 (bndrs,rhss) = unzip pairs
564 -- Finding the free vars of the binding group is annoying
565 bind_fvs = (unionVarSets [ idFreeVars bndr `unionVarSet` rhs_fvs
566 | (bndr, (rhs_fvs,_)) <- pairs])
570 dest_lvl = destLevel env bind_fvs (all isFunction rhss)
571 abs_vars = abstractVars dest_lvl env bind_fvs
573 ----------------------------------------------------
574 -- Three help functons for the type-abstraction case
576 lvlFloatRhs abs_vars dest_lvl env rhs
577 = lvlExpr rhs_lvl rhs_env rhs `thenLvl` \ rhs' ->
578 returnLvl (mkLams abs_vars_w_lvls rhs')
580 (rhs_lvl, abs_vars_w_lvls) = lvlLamBndrs dest_lvl abs_vars
581 rhs_env = extendLvlEnv env abs_vars_w_lvls
585 %************************************************************************
587 \subsection{Deciding floatability}
589 %************************************************************************
592 lvlLamBndrs :: Level -> [CoreBndr] -> (Level, [TaggedBndr Level])
593 -- Compute the levels for the binders of a lambda group
594 -- The binders returned are exactly the same as the ones passed,
595 -- but they are now paired with a level
599 lvlLamBndrs lvl bndrs
600 = go (incMinorLvl lvl)
601 False -- Havn't bumped major level in this group
604 go old_lvl bumped_major rev_lvld_bndrs (bndr:bndrs)
605 | isId bndr && -- Go to the next major level if this is a value binder,
606 not bumped_major && -- and we havn't already gone to the next level (one jump per group)
607 not (isOneShotLambda bndr) -- and it isn't a one-shot lambda
608 = go new_lvl True (TB bndr new_lvl : rev_lvld_bndrs) bndrs
611 = go old_lvl bumped_major (TB bndr old_lvl : rev_lvld_bndrs) bndrs
614 new_lvl = incMajorLvl old_lvl
616 go old_lvl _ rev_lvld_bndrs []
617 = (old_lvl, reverse rev_lvld_bndrs)
618 -- a lambda like this (\x -> coerce t (\s -> ...))
619 -- This happens quite a bit in state-transformer programs
623 -- Destintion level is the max Id level of the expression
624 -- (We'll abstract the type variables, if any.)
625 destLevel :: LevelEnv -> VarSet -> Bool -> Level
626 destLevel env fvs is_function
628 && is_function = tOP_LEVEL -- Send functions to top level; see
629 -- the comments with isFunction
630 | otherwise = maxIdLevel env fvs
632 isFunction :: CoreExprWithFVs -> Bool
633 -- The idea here is that we want to float *functions* to
634 -- the top level. This saves no work, but
635 -- (a) it can make the host function body a lot smaller,
636 -- and hence inlinable.
637 -- (b) it can also save allocation when the function is recursive:
638 -- h = \x -> letrec f = \y -> ...f...y...x...
641 -- f = \x y -> ...(f x)...y...x...
643 -- No allocation for f now.
644 -- We may only want to do this if there are sufficiently few free
645 -- variables. We certainly only want to do it for values, and not for
646 -- constructors. So the simple thing is just to look for lambdas
647 isFunction (_, AnnLam b e) | isId b = True
648 | otherwise = isFunction e
649 isFunction (_, AnnNote n e) = isFunction e
650 isFunction other = False
654 %************************************************************************
656 \subsection{Free-To-Level Monad}
658 %************************************************************************
661 type LevelEnv = (FloatOutSwitches,
662 VarEnv Level, -- Domain is *post-cloned* TyVars and Ids
663 Subst, -- Domain is pre-cloned Ids; tracks the in-scope set
664 -- so that subtitution is capture-avoiding
665 IdEnv ([Var], LevelledExpr)) -- Domain is pre-cloned Ids
666 -- We clone let-bound variables so that they are still
667 -- distinct when floated out; hence the SubstEnv/IdEnv.
668 -- (see point 3 of the module overview comment).
669 -- We also use these envs when making a variable polymorphic
670 -- because we want to float it out past a big lambda.
672 -- The SubstEnv and IdEnv always implement the same mapping, but the
673 -- SubstEnv maps to CoreExpr and the IdEnv to LevelledExpr
674 -- Since the range is always a variable or type application,
675 -- there is never any difference between the two, but sadly
676 -- the types differ. The SubstEnv is used when substituting in
677 -- a variable's IdInfo; the IdEnv when we find a Var.
679 -- In addition the IdEnv records a list of tyvars free in the
680 -- type application, just so we don't have to call freeVars on
681 -- the type application repeatedly.
683 -- The domain of the both envs is *pre-cloned* Ids, though
685 -- The domain of the VarEnv Level is the *post-cloned* Ids
687 initialEnv :: FloatOutSwitches -> LevelEnv
688 initialEnv float_lams = (float_lams, emptyVarEnv, emptySubst, emptyVarEnv)
690 floatLams :: LevelEnv -> Bool
691 floatLams (FloatOutSw float_lams _, _, _, _) = float_lams
693 floatConsts :: LevelEnv -> Bool
694 floatConsts (FloatOutSw _ float_consts, _, _, _) = float_consts
696 extendLvlEnv :: LevelEnv -> [TaggedBndr Level] -> LevelEnv
697 -- Used when *not* cloning
698 extendLvlEnv (float_lams, lvl_env, subst, id_env) prs
700 foldl add_lvl lvl_env prs,
701 foldl del_subst subst prs,
702 foldl del_id id_env prs)
704 add_lvl env (TB v l) = extendVarEnv env v l
705 del_subst env (TB v _) = extendInScope env v
706 del_id env (TB v _) = delVarEnv env v
707 -- We must remove any clone for this variable name in case of
708 -- shadowing. This bit me in the following case
709 -- (in nofib/real/gg/Spark.hs):
712 -- ... -> case e of wild {
713 -- ... -> ... wild ...
717 -- The inside occurrence of @wild@ was being replaced with @ds@,
718 -- incorrectly, because the SubstEnv was still lying around. Ouch!
721 -- extendCaseBndrLvlEnv adds the mapping case-bndr->scrut-var if it can
722 -- (see point 4 of the module overview comment)
723 extendCaseBndrLvlEnv (float_lams, lvl_env, subst, id_env) (Var scrut_var) case_bndr lvl
725 extendVarEnv lvl_env case_bndr lvl,
726 extendIdSubst subst case_bndr (Var scrut_var),
727 extendVarEnv id_env case_bndr ([scrut_var], Var scrut_var))
729 extendCaseBndrLvlEnv env scrut case_bndr lvl
730 = extendLvlEnv env [TB case_bndr lvl]
732 extendPolyLvlEnv dest_lvl (float_lams, lvl_env, subst, id_env) abs_vars bndr_pairs
734 foldl add_lvl lvl_env bndr_pairs,
735 foldl add_subst subst bndr_pairs,
736 foldl add_id id_env bndr_pairs)
738 add_lvl env (v,v') = extendVarEnv env v' dest_lvl
739 add_subst env (v,v') = extendIdSubst env v (mkVarApps (Var v') abs_vars)
740 add_id env (v,v') = extendVarEnv env v ((v':abs_vars), mkVarApps (Var v') abs_vars)
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,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 | isId 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 env fvs
776 = uniq (sortLe le [var | fv <- varSetElems fvs, var <- absVarsOf dest_lvl env fv])
778 -- Sort the variables so we don't get
779 -- mixed-up tyvars and Ids; it's just messy
780 v1 `le` v2 = case (isId v1, isId v2) of
781 (True, False) -> False
782 (False, True) -> True
783 other -> v1 <= v2 -- Same family
785 uniq :: [Var] -> [Var]
786 -- Remove adjacent duplicates; the sort will have brought them together
787 uniq (v1:v2:vs) | v1 == v2 = uniq (v2:vs)
788 | otherwise = v1 : uniq (v2:vs)
791 absVarsOf :: Level -> LevelEnv -> Var -> [Var]
792 -- If f is free in the expression, and f maps to poly_f a b c in the
793 -- current substitution, then we must report a b c as candidate type
795 absVarsOf dest_lvl (_, lvl_env, _, id_env) v
797 = [zap av2 | av1 <- lookup_avs v, av2 <- add_tyvars av1, abstract_me av2]
800 = if abstract_me v then [v] else []
803 abstract_me v = case lookupVarEnv lvl_env v of
804 Just lvl -> dest_lvl `ltLvl` lvl
807 lookup_avs v = case lookupVarEnv id_env v of
808 Just (abs_vars, _) -> abs_vars
811 add_tyvars v = v : varSetElems (varTypeTyVars v)
813 -- We are going to lambda-abstract, so nuke any IdInfo,
814 -- and add the tyvars of the Id (if necessary)
815 zap v | isId v = WARN( workerExists (idWorkerInfo v) ||
816 not (isEmptySpecInfo (idSpecialisation v)),
817 text "absVarsOf: discarding info on" <+> ppr v )
818 setIdInfo v vanillaIdInfo
823 type LvlM result = UniqSM result
832 newPolyBndrs dest_lvl env abs_vars bndrs
833 = getUniquesUs `thenLvl` \ uniqs ->
835 new_bndrs = zipWith mk_poly_bndr bndrs uniqs
837 returnLvl (extendPolyLvlEnv dest_lvl env abs_vars (bndrs `zip` new_bndrs), new_bndrs)
839 mk_poly_bndr bndr uniq = mkSysLocal (mkFastString str) uniq poly_ty
841 str = "poly_" ++ occNameString (getOccName bndr)
842 poly_ty = mkPiTypes abs_vars (idType bndr)
846 -> [CoreBndr] -> Type -- Abstract wrt these bndrs
848 newLvlVar str vars body_ty
849 = getUniqueUs `thenLvl` \ uniq ->
850 returnUs (mkSysLocal (mkFastString str) uniq (mkPiTypes vars body_ty))
852 -- The deeply tiresome thing is that we have to apply the substitution
853 -- to the rules inside each Id. Grr. But it matters.
855 cloneVar :: TopLevelFlag -> LevelEnv -> Id -> Level -> Level -> LvlM (LevelEnv, Id)
856 cloneVar TopLevel env v ctxt_lvl dest_lvl
857 = returnUs (env, v) -- Don't clone top level things
858 cloneVar NotTopLevel env@(_,_,subst,_) v ctxt_lvl dest_lvl
860 getUs `thenLvl` \ us ->
862 (subst', v1) = cloneIdBndr subst us v
863 v2 = zap_demand ctxt_lvl dest_lvl v1
864 env' = extendCloneLvlEnv dest_lvl env subst' [(v,v2)]
868 cloneRecVars :: TopLevelFlag -> LevelEnv -> [Id] -> Level -> Level -> LvlM (LevelEnv, [Id])
869 cloneRecVars TopLevel env vs ctxt_lvl dest_lvl
870 = returnUs (env, vs) -- Don't clone top level things
871 cloneRecVars NotTopLevel env@(_,_,subst,_) vs ctxt_lvl dest_lvl
872 = ASSERT( all isId vs )
873 getUs `thenLvl` \ us ->
875 (subst', vs1) = cloneRecIdBndrs subst us vs
876 vs2 = map (zap_demand ctxt_lvl dest_lvl) vs1
877 env' = extendCloneLvlEnv dest_lvl env subst' (vs `zip` vs2)
881 -- VERY IMPORTANT: we must zap the demand info
882 -- if the thing is going to float out past a lambda,
883 -- or if it's going to top level (where things can't be strict)
884 zap_demand dest_lvl ctxt_lvl id
885 | ctxt_lvl == dest_lvl,
886 not (isTopLvl dest_lvl) = id -- Stays, and not going to top level
887 | otherwise = zapDemandIdInfo id -- Floats out