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,
65 idSpecialisation, idWorkerInfo, setIdInfo
67 import IdInfo ( workerExists, vanillaIdInfo, isEmptySpecInfo )
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
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 = Level 0 0
154 iNLINE_CTXT = InlineCtxt
156 incMajorLvl :: Level -> Level
157 -- For InlineCtxt we ignore any inc's; we don't want
158 -- to do any floating at all; see notes above
159 incMajorLvl InlineCtxt = InlineCtxt
160 incMajorLvl (Level major minor) = Level (major+1) 0
162 incMinorLvl :: Level -> Level
163 incMinorLvl InlineCtxt = InlineCtxt
164 incMinorLvl (Level major minor) = Level major (minor+1)
166 maxLvl :: Level -> Level -> Level
167 maxLvl InlineCtxt l2 = l2
168 maxLvl l1 InlineCtxt = l1
169 maxLvl l1@(Level maj1 min1) l2@(Level maj2 min2)
170 | (maj1 > maj2) || (maj1 == maj2 && min1 > min2) = l1
173 ltLvl :: Level -> Level -> Bool
174 ltLvl any_lvl InlineCtxt = False
175 ltLvl InlineCtxt (Level _ _) = True
176 ltLvl (Level maj1 min1) (Level maj2 min2)
177 = (maj1 < maj2) || (maj1 == maj2 && min1 < min2)
179 ltMajLvl :: Level -> Level -> Bool
180 -- Tells if one level belongs to a difft *lambda* level to another
181 ltMajLvl any_lvl InlineCtxt = False
182 ltMajLvl InlineCtxt (Level maj2 _) = 0 < maj2
183 ltMajLvl (Level maj1 _) (Level maj2 _) = maj1 < maj2
185 isTopLvl :: Level -> Bool
186 isTopLvl (Level 0 0) = True
187 isTopLvl other = False
189 isInlineCtxt :: Level -> Bool
190 isInlineCtxt InlineCtxt = True
191 isInlineCtxt other = False
193 instance Outputable Level where
194 ppr InlineCtxt = text "<INLINE>"
195 ppr (Level maj min) = hcat [ char '<', int maj, char ',', int min, char '>' ]
197 instance Eq Level where
198 InlineCtxt == InlineCtxt = True
199 (Level maj1 min1) == (Level maj2 min2) = maj1==maj2 && min1==min2
204 %************************************************************************
206 \subsection{Main level-setting code}
208 %************************************************************************
211 setLevels :: FloatOutSwitches
216 setLevels float_lams binds us
217 = initLvl us (do_them binds)
219 -- "do_them"'s main business is to thread the monad along
220 -- It gives each top binding the same empty envt, because
221 -- things unbound in the envt have level number zero implicitly
222 do_them :: [CoreBind] -> LvlM [LevelledBind]
224 do_them [] = returnLvl []
226 = lvlTopBind init_env b `thenLvl` \ (lvld_bind, _) ->
227 do_them bs `thenLvl` \ lvld_binds ->
228 returnLvl (lvld_bind : lvld_binds)
230 init_env = initialEnv float_lams
232 lvlTopBind env (NonRec binder rhs)
233 = lvlBind TopLevel tOP_LEVEL env (AnnNonRec binder (freeVars rhs))
234 -- Rhs can have no free vars!
236 lvlTopBind env (Rec pairs)
237 = lvlBind TopLevel tOP_LEVEL env (AnnRec [(b,freeVars rhs) | (b,rhs) <- pairs])
240 %************************************************************************
242 \subsection{Setting expression levels}
244 %************************************************************************
247 lvlExpr :: Level -- ctxt_lvl: Level of enclosing expression
248 -> LevelEnv -- Level of in-scope names/tyvars
249 -> CoreExprWithFVs -- input expression
250 -> LvlM LevelledExpr -- Result expression
253 The @ctxt_lvl@ is, roughly, the level of the innermost enclosing
254 binder. Here's an example
256 v = \x -> ...\y -> let r = case (..x..) of
260 When looking at the rhs of @r@, @ctxt_lvl@ will be 1 because that's
261 the level of @r@, even though it's inside a level-2 @\y@. It's
262 important that @ctxt_lvl@ is 1 and not 2 in @r@'s rhs, because we
263 don't want @lvlExpr@ to turn the scrutinee of the @case@ into an MFE
264 --- because it isn't a *maximal* free expression.
266 If there were another lambda in @r@'s rhs, it would get level-2 as well.
269 lvlExpr _ _ (_, AnnType ty) = returnLvl (Type ty)
270 lvlExpr _ env (_, AnnVar v) = returnLvl (lookupVar env v)
271 lvlExpr _ env (_, AnnLit lit) = returnLvl (Lit lit)
273 lvlExpr ctxt_lvl env (_, AnnApp fun arg)
274 = lvl_fun fun `thenLvl` \ fun' ->
275 lvlMFE False ctxt_lvl env arg `thenLvl` \ arg' ->
276 returnLvl (App fun' arg')
279 lvl_fun (_, AnnCase _ _ _ _) = lvlMFE True ctxt_lvl env fun
280 lvl_fun other = lvlExpr ctxt_lvl env fun
281 -- We don't do MFE on partial applications generally,
282 -- but we do if the function is big and hairy, like a case
284 lvlExpr ctxt_lvl env (_, AnnNote InlineMe expr)
285 -- Don't float anything out of an InlineMe; hence the iNLINE_CTXT
286 = lvlExpr iNLINE_CTXT env expr `thenLvl` \ expr' ->
287 returnLvl (Note InlineMe expr')
289 lvlExpr ctxt_lvl env (_, AnnNote note expr)
290 = lvlExpr ctxt_lvl env expr `thenLvl` \ expr' ->
291 returnLvl (Note note expr')
293 lvlExpr ctxt_lvl env (_, AnnCast expr co)
294 = lvlExpr ctxt_lvl env expr `thenLvl` \ expr' ->
295 returnLvl (Cast expr' co)
297 -- We don't split adjacent lambdas. That is, given
299 -- we don't float to give
300 -- \x -> let v = x+y in \y -> (v,y)
301 -- Why not? Because partial applications are fairly rare, and splitting
302 -- lambdas makes them more expensive.
304 lvlExpr ctxt_lvl env expr@(_, AnnLam bndr rhs)
305 = lvlMFE True new_lvl new_env body `thenLvl` \ new_body ->
306 returnLvl (mkLams new_bndrs new_body)
308 (bndrs, body) = collectAnnBndrs expr
309 (new_lvl, new_bndrs) = lvlLamBndrs ctxt_lvl bndrs
310 new_env = extendLvlEnv env new_bndrs
311 -- At one time we called a special verion of collectBinders,
312 -- which ignored coercions, because we don't want to split
313 -- a lambda like this (\x -> coerce t (\s -> ...))
314 -- This used to happen quite a bit in state-transformer programs,
315 -- but not nearly so much now non-recursive newtypes are transparent.
316 -- [See SetLevels rev 1.50 for a version with this approach.]
318 lvlExpr ctxt_lvl env (_, AnnLet (AnnNonRec bndr rhs) body)
319 | isUnLiftedType (idType bndr)
320 -- Treat unlifted let-bindings (let x = b in e) just like (case b of x -> e)
321 -- That is, leave it exactly where it is
322 -- We used to float unlifted bindings too (e.g. to get a cheap primop
323 -- outside a lambda (to see how, look at lvlBind in rev 1.58)
324 -- but an unrelated change meant that these unlifed bindings
325 -- could get to the top level which is bad. And there's not much point;
326 -- unlifted bindings are always cheap, and so hardly worth floating.
327 = lvlExpr ctxt_lvl env rhs `thenLvl` \ rhs' ->
328 lvlExpr incd_lvl env' body `thenLvl` \ body' ->
329 returnLvl (Let (NonRec bndr' rhs') body')
331 incd_lvl = incMinorLvl ctxt_lvl
332 bndr' = TB bndr incd_lvl
333 env' = extendLvlEnv env [bndr']
335 lvlExpr ctxt_lvl env (_, AnnLet bind body)
336 = lvlBind NotTopLevel ctxt_lvl env bind `thenLvl` \ (bind', new_env) ->
337 lvlExpr ctxt_lvl new_env body `thenLvl` \ body' ->
338 returnLvl (Let bind' body')
340 lvlExpr ctxt_lvl env (_, AnnCase expr case_bndr ty alts)
341 = lvlMFE True ctxt_lvl env expr `thenLvl` \ expr' ->
343 alts_env = extendCaseBndrLvlEnv env expr' case_bndr incd_lvl
345 mapLvl (lvl_alt alts_env) alts `thenLvl` \ alts' ->
346 returnLvl (Case expr' (TB case_bndr incd_lvl) ty alts')
348 incd_lvl = incMinorLvl ctxt_lvl
350 lvl_alt alts_env (con, bs, rhs)
351 = lvlMFE True incd_lvl new_env rhs `thenLvl` \ rhs' ->
352 returnLvl (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.
361 [NOTE: unlifted MFEs]
362 We don't float unlifted MFEs, which potentially loses big opportunites.
365 where h :: Int -> Int# is expensive. We'd like to float the (h y) outside
366 the \x, but we don't because it's unboxed. Possible solution: box it.
369 lvlMFE :: Bool -- True <=> strict context [body of case or let]
370 -> Level -- Level of innermost enclosing lambda/tylam
371 -> LevelEnv -- Level of in-scope names/tyvars
372 -> CoreExprWithFVs -- input expression
373 -> LvlM LevelledExpr -- Result expression
375 lvlMFE strict_ctxt ctxt_lvl env (_, AnnType ty)
376 = returnLvl (Type ty)
379 lvlMFE strict_ctxt ctxt_lvl env ann_expr@(fvs, _)
380 | isUnLiftedType ty -- Can't let-bind it; see [NOTE: unlifted MFEs]
381 || isInlineCtxt ctxt_lvl -- Don't float out of an __inline__ context
382 || exprIsTrivial expr -- Never float if it's trivial
383 || not good_destination
384 = -- Don't float it out
385 lvlExpr ctxt_lvl env ann_expr
387 | otherwise -- Float it out!
388 = lvlFloatRhs abs_vars dest_lvl env ann_expr `thenLvl` \ expr' ->
389 newLvlVar "lvl" abs_vars ty `thenLvl` \ var ->
390 returnLvl (Let (NonRec (TB var dest_lvl) expr')
391 (mkVarApps (Var var) abs_vars))
393 expr = deAnnotate ann_expr
395 dest_lvl = destLevel env fvs (isFunction ann_expr)
396 abs_vars = abstractVars dest_lvl env fvs
398 -- A decision to float entails let-binding this thing, and we only do
399 -- that if we'll escape a value lambda, or will go to the top level.
401 | dest_lvl `ltMajLvl` ctxt_lvl -- Escapes a value lambda
403 -- OLD CODE: not (exprIsCheap expr) || isTopLvl dest_lvl
404 -- see Note [Escaping a value lambda]
406 | otherwise -- Does not escape a value lambda
407 = isTopLvl dest_lvl -- Only float if we are going to the top level
408 && floatConsts env -- and the floatConsts flag is on
409 && not strict_ctxt -- Don't float from a strict context
410 -- We are keen to float something to the top level, even if it does not
411 -- escape a lambda, because then it needs no allocation. But it's controlled
412 -- by a flag, because doing this too early loses opportunities for RULES
413 -- which (needless to say) are important in some nofib programs
414 -- (gcd is an example).
417 -- concat = /\ a -> foldr ..a.. (++) []
418 -- was getting turned into
419 -- concat = /\ a -> lvl a
420 -- lvl = /\ a -> foldr ..a.. (++) []
421 -- which is pretty stupid. Hence the strict_ctxt test
424 Note [Escaping a value lambda]
425 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
426 We want to float even cheap expressions out of value lambdas,
427 because that saves allocation. Consider
428 f = \x. .. (\y.e) ...
429 Then we'd like to avoid allocating the (\y.e) every time we call f,
430 (assuming e does not mention x).
432 An example where this really makes a difference is simplrun009.
434 Another reason it's good is because it makes SpecContr fire on functions.
436 f = \x. ....(f (\y.e))....
437 After floating we get
439 f = \x. ....(f lvl)...
440 and that is much easier for SpecConstr to generate a robust specialisation for.
442 The OLD CODE (given where this Note is referred to) prevents floating
443 of the example above, so I just don't understand the old code. I
444 don't understand the old comment either (which appears below). I
445 measured the effect on nofib of changing OLD CODE to 'True', and got
446 zeros everywhere, but a 4% win for 'puzzle'. Very small 0.5% loss for
447 'cse'; turns out to be because our arity analysis isn't good enough
448 yet (mentioned in Simon-nofib-notes).
451 Even if it escapes a value lambda, we only
452 float if it's not cheap (unless it'll get all the
453 way to the top). I've seen cases where we
454 float dozens of tiny free expressions, which cost
455 more to allocate than to evaluate.
456 NB: exprIsCheap is also true of bottom expressions, which
457 is good; we don't want to share them
459 It's only Really Bad to float a cheap expression out of a
460 strict context, because that builds a thunk that otherwise
461 would never be built. So another alternative would be to
463 || (strict_ctxt && not (exprIsBottom expr))
464 to the condition above. We should really try this out.
467 %************************************************************************
469 \subsection{Bindings}
471 %************************************************************************
473 The binding stuff works for top level too.
476 lvlBind :: TopLevelFlag -- Used solely to decide whether to clone
477 -> Level -- Context level; might be Top even for bindings nested in the RHS
478 -- of a top level binding
481 -> LvlM (LevelledBind, LevelEnv)
483 lvlBind top_lvl ctxt_lvl env (AnnNonRec bndr rhs@(rhs_fvs,_))
484 | isInlineCtxt ctxt_lvl -- Don't do anything inside InlineMe
485 = lvlExpr ctxt_lvl env rhs `thenLvl` \ rhs' ->
486 returnLvl (NonRec (TB bndr ctxt_lvl) rhs', env)
489 = -- No type abstraction; clone existing binder
490 lvlExpr dest_lvl env rhs `thenLvl` \ rhs' ->
491 cloneVar top_lvl env bndr ctxt_lvl dest_lvl `thenLvl` \ (env', bndr') ->
492 returnLvl (NonRec (TB bndr' dest_lvl) rhs', env')
495 = -- Yes, type abstraction; create a new binder, extend substitution, etc
496 lvlFloatRhs abs_vars dest_lvl env rhs `thenLvl` \ rhs' ->
497 newPolyBndrs dest_lvl env abs_vars [bndr] `thenLvl` \ (env', [bndr']) ->
498 returnLvl (NonRec (TB bndr' dest_lvl) rhs', env')
501 bind_fvs = rhs_fvs `unionVarSet` idFreeVars bndr
502 abs_vars = abstractVars dest_lvl env bind_fvs
503 dest_lvl = destLevel env bind_fvs (isFunction rhs)
508 lvlBind top_lvl ctxt_lvl env (AnnRec pairs)
509 | isInlineCtxt ctxt_lvl -- Don't do anything inside InlineMe
510 = mapLvl (lvlExpr ctxt_lvl env) rhss `thenLvl` \ rhss' ->
511 returnLvl (Rec ([TB b ctxt_lvl | b <- bndrs] `zip` rhss'), env)
514 = cloneRecVars top_lvl env bndrs ctxt_lvl dest_lvl `thenLvl` \ (new_env, new_bndrs) ->
515 mapLvl (lvlExpr ctxt_lvl new_env) rhss `thenLvl` \ new_rhss ->
516 returnLvl (Rec ([TB b dest_lvl | b <- new_bndrs] `zip` new_rhss), new_env)
518 | isSingleton pairs && count isId abs_vars > 1
519 = -- Special case for self recursion where there are
520 -- several variables carried around: build a local loop:
521 -- poly_f = \abs_vars. \lam_vars . letrec f = \lam_vars. rhs in f lam_vars
522 -- This just makes the closures a bit smaller. If we don't do
523 -- this, allocation rises significantly on some programs
525 -- We could elaborate it for the case where there are several
526 -- mutually functions, but it's quite a bit more complicated
528 -- This all seems a bit ad hoc -- sigh
530 (bndr,rhs) = head pairs
531 (rhs_lvl, abs_vars_w_lvls) = lvlLamBndrs dest_lvl abs_vars
532 rhs_env = extendLvlEnv env abs_vars_w_lvls
534 cloneVar NotTopLevel rhs_env bndr rhs_lvl rhs_lvl `thenLvl` \ (rhs_env', new_bndr) ->
536 (lam_bndrs, rhs_body) = collectAnnBndrs rhs
537 (body_lvl, new_lam_bndrs) = lvlLamBndrs rhs_lvl lam_bndrs
538 body_env = extendLvlEnv rhs_env' new_lam_bndrs
540 lvlExpr body_lvl body_env rhs_body `thenLvl` \ new_rhs_body ->
541 newPolyBndrs dest_lvl env abs_vars [bndr] `thenLvl` \ (poly_env, [poly_bndr]) ->
542 returnLvl (Rec [(TB poly_bndr dest_lvl,
543 mkLams abs_vars_w_lvls $
544 mkLams new_lam_bndrs $
545 Let (Rec [(TB new_bndr rhs_lvl, mkLams new_lam_bndrs new_rhs_body)])
546 (mkVarApps (Var new_bndr) lam_bndrs))],
549 | otherwise -- Non-null abs_vars
550 = newPolyBndrs dest_lvl env abs_vars bndrs `thenLvl` \ (new_env, new_bndrs) ->
551 mapLvl (lvlFloatRhs abs_vars dest_lvl new_env) rhss `thenLvl` \ new_rhss ->
552 returnLvl (Rec ([TB b dest_lvl | b <- new_bndrs] `zip` new_rhss), new_env)
555 (bndrs,rhss) = unzip pairs
557 -- Finding the free vars of the binding group is annoying
558 bind_fvs = (unionVarSets [ idFreeVars bndr `unionVarSet` rhs_fvs
559 | (bndr, (rhs_fvs,_)) <- pairs])
563 dest_lvl = destLevel env bind_fvs (all isFunction rhss)
564 abs_vars = abstractVars dest_lvl env bind_fvs
566 ----------------------------------------------------
567 -- Three help functons for the type-abstraction case
569 lvlFloatRhs abs_vars dest_lvl env rhs
570 = lvlExpr rhs_lvl rhs_env rhs `thenLvl` \ rhs' ->
571 returnLvl (mkLams abs_vars_w_lvls rhs')
573 (rhs_lvl, abs_vars_w_lvls) = lvlLamBndrs dest_lvl abs_vars
574 rhs_env = extendLvlEnv env abs_vars_w_lvls
578 %************************************************************************
580 \subsection{Deciding floatability}
582 %************************************************************************
585 lvlLamBndrs :: Level -> [CoreBndr] -> (Level, [TaggedBndr Level])
586 -- Compute the levels for the binders of a lambda group
587 -- The binders returned are exactly the same as the ones passed,
588 -- but they are now paired with a level
592 lvlLamBndrs lvl bndrs
593 = go (incMinorLvl lvl)
594 False -- Havn't bumped major level in this group
597 go old_lvl bumped_major rev_lvld_bndrs (bndr:bndrs)
598 | isId bndr && -- Go to the next major level if this is a value binder,
599 not bumped_major && -- and we havn't already gone to the next level (one jump per group)
600 not (isOneShotLambda bndr) -- and it isn't a one-shot lambda
601 = go new_lvl True (TB bndr new_lvl : rev_lvld_bndrs) bndrs
604 = go old_lvl bumped_major (TB bndr old_lvl : rev_lvld_bndrs) bndrs
607 new_lvl = incMajorLvl old_lvl
609 go old_lvl _ rev_lvld_bndrs []
610 = (old_lvl, reverse rev_lvld_bndrs)
611 -- a lambda like this (\x -> coerce t (\s -> ...))
612 -- This happens quite a bit in state-transformer programs
616 -- Destintion level is the max Id level of the expression
617 -- (We'll abstract the type variables, if any.)
618 destLevel :: LevelEnv -> VarSet -> Bool -> Level
619 destLevel env fvs is_function
621 && is_function = tOP_LEVEL -- Send functions to top level; see
622 -- the comments with isFunction
623 | otherwise = maxIdLevel env fvs
625 isFunction :: CoreExprWithFVs -> Bool
626 -- The idea here is that we want to float *functions* to
627 -- the top level. This saves no work, but
628 -- (a) it can make the host function body a lot smaller,
629 -- and hence inlinable.
630 -- (b) it can also save allocation when the function is recursive:
631 -- h = \x -> letrec f = \y -> ...f...y...x...
634 -- f = \x y -> ...(f x)...y...x...
636 -- No allocation for f now.
637 -- We may only want to do this if there are sufficiently few free
638 -- variables. We certainly only want to do it for values, and not for
639 -- constructors. So the simple thing is just to look for lambdas
640 isFunction (_, AnnLam b e) | isId b = True
641 | otherwise = isFunction e
642 isFunction (_, AnnNote n e) = isFunction e
643 isFunction other = False
647 %************************************************************************
649 \subsection{Free-To-Level Monad}
651 %************************************************************************
654 type LevelEnv = (FloatOutSwitches,
655 VarEnv Level, -- Domain is *post-cloned* TyVars and Ids
656 Subst, -- Domain is pre-cloned Ids; tracks the in-scope set
657 -- so that subtitution is capture-avoiding
658 IdEnv ([Var], LevelledExpr)) -- Domain is pre-cloned Ids
659 -- We clone let-bound variables so that they are still
660 -- distinct when floated out; hence the SubstEnv/IdEnv.
661 -- (see point 3 of the module overview comment).
662 -- We also use these envs when making a variable polymorphic
663 -- because we want to float it out past a big lambda.
665 -- The SubstEnv and IdEnv always implement the same mapping, but the
666 -- SubstEnv maps to CoreExpr and the IdEnv to LevelledExpr
667 -- Since the range is always a variable or type application,
668 -- there is never any difference between the two, but sadly
669 -- the types differ. The SubstEnv is used when substituting in
670 -- a variable's IdInfo; the IdEnv when we find a Var.
672 -- In addition the IdEnv records a list of tyvars free in the
673 -- type application, just so we don't have to call freeVars on
674 -- the type application repeatedly.
676 -- The domain of the both envs is *pre-cloned* Ids, though
678 -- The domain of the VarEnv Level is the *post-cloned* Ids
680 initialEnv :: FloatOutSwitches -> LevelEnv
681 initialEnv float_lams = (float_lams, emptyVarEnv, emptySubst, emptyVarEnv)
683 floatLams :: LevelEnv -> Bool
684 floatLams (FloatOutSw float_lams _, _, _, _) = float_lams
686 floatConsts :: LevelEnv -> Bool
687 floatConsts (FloatOutSw _ float_consts, _, _, _) = float_consts
689 extendLvlEnv :: LevelEnv -> [TaggedBndr Level] -> LevelEnv
690 -- Used when *not* cloning
691 extendLvlEnv (float_lams, lvl_env, subst, id_env) prs
693 foldl add_lvl lvl_env prs,
694 foldl del_subst subst prs,
695 foldl del_id id_env prs)
697 add_lvl env (TB v l) = extendVarEnv env v l
698 del_subst env (TB v _) = extendInScope env v
699 del_id env (TB v _) = delVarEnv env v
700 -- We must remove any clone for this variable name in case of
701 -- shadowing. This bit me in the following case
702 -- (in nofib/real/gg/Spark.hs):
705 -- ... -> case e of wild {
706 -- ... -> ... wild ...
710 -- The inside occurrence of @wild@ was being replaced with @ds@,
711 -- incorrectly, because the SubstEnv was still lying around. Ouch!
714 -- extendCaseBndrLvlEnv adds the mapping case-bndr->scrut-var if it can
715 -- (see point 4 of the module overview comment)
716 extendCaseBndrLvlEnv (float_lams, lvl_env, subst, id_env) (Var scrut_var) case_bndr lvl
718 extendVarEnv lvl_env case_bndr lvl,
719 extendIdSubst subst case_bndr (Var scrut_var),
720 extendVarEnv id_env case_bndr ([scrut_var], Var scrut_var))
722 extendCaseBndrLvlEnv env scrut case_bndr lvl
723 = extendLvlEnv env [TB case_bndr lvl]
725 extendPolyLvlEnv dest_lvl (float_lams, lvl_env, subst, id_env) abs_vars bndr_pairs
727 foldl add_lvl lvl_env bndr_pairs,
728 foldl add_subst subst bndr_pairs,
729 foldl add_id id_env bndr_pairs)
731 add_lvl env (v,v') = extendVarEnv env v' dest_lvl
732 add_subst env (v,v') = extendIdSubst env v (mkVarApps (Var v') abs_vars)
733 add_id env (v,v') = extendVarEnv env v ((v':abs_vars), mkVarApps (Var v') abs_vars)
735 extendCloneLvlEnv lvl (float_lams, lvl_env, _, id_env) new_subst bndr_pairs
737 foldl add_lvl lvl_env bndr_pairs,
739 foldl add_id id_env bndr_pairs)
741 add_lvl env (v,v') = extendVarEnv env v' lvl
742 add_id env (v,v') = extendVarEnv env v ([v'], Var v')
745 maxIdLevel :: LevelEnv -> VarSet -> Level
746 maxIdLevel (_, lvl_env,_,id_env) var_set
747 = foldVarSet max_in tOP_LEVEL var_set
749 max_in in_var lvl = foldr max_out lvl (case lookupVarEnv id_env in_var of
750 Just (abs_vars, _) -> abs_vars
754 | isId out_var = case lookupVarEnv lvl_env out_var of
755 Just lvl' -> maxLvl lvl' lvl
757 | otherwise = lvl -- Ignore tyvars in *maxIdLevel*
759 lookupVar :: LevelEnv -> Id -> LevelledExpr
760 lookupVar (_, _, _, id_env) v = case lookupVarEnv id_env v of
761 Just (_, expr) -> expr
764 abstractVars :: Level -> LevelEnv -> VarSet -> [Var]
765 -- Find the variables in fvs, free vars of the target expresion,
766 -- whose level is greater than the destination level
767 -- These are the ones we are going to abstract out
768 abstractVars dest_lvl env fvs
769 = uniq (sortLe le [var | fv <- varSetElems fvs, var <- absVarsOf dest_lvl env fv])
771 -- Sort the variables so we don't get
772 -- mixed-up tyvars and Ids; it's just messy
773 v1 `le` v2 = case (isId v1, isId v2) of
774 (True, False) -> False
775 (False, True) -> True
776 other -> v1 <= v2 -- Same family
778 uniq :: [Var] -> [Var]
779 -- Remove adjacent duplicates; the sort will have brought them together
780 uniq (v1:v2:vs) | v1 == v2 = uniq (v2:vs)
781 | otherwise = v1 : uniq (v2:vs)
784 absVarsOf :: Level -> LevelEnv -> Var -> [Var]
785 -- If f is free in the expression, and f maps to poly_f a b c in the
786 -- current substitution, then we must report a b c as candidate type
788 absVarsOf dest_lvl (_, lvl_env, _, id_env) v
790 = [zap av2 | av1 <- lookup_avs v, av2 <- add_tyvars av1, abstract_me av2]
793 = if abstract_me v then [v] else []
796 abstract_me v = case lookupVarEnv lvl_env v of
797 Just lvl -> dest_lvl `ltLvl` lvl
800 lookup_avs v = case lookupVarEnv id_env v of
801 Just (abs_vars, _) -> abs_vars
804 add_tyvars v = v : varSetElems (varTypeTyVars v)
806 -- We are going to lambda-abstract, so nuke any IdInfo,
807 -- and add the tyvars of the Id (if necessary)
808 zap v | isId v = WARN( workerExists (idWorkerInfo v) ||
809 not (isEmptySpecInfo (idSpecialisation v)),
810 text "absVarsOf: discarding info on" <+> ppr v )
811 setIdInfo v vanillaIdInfo
816 type LvlM result = UniqSM result
825 newPolyBndrs dest_lvl env abs_vars bndrs
826 = getUniquesUs `thenLvl` \ uniqs ->
828 new_bndrs = zipWith mk_poly_bndr bndrs uniqs
830 returnLvl (extendPolyLvlEnv dest_lvl env abs_vars (bndrs `zip` new_bndrs), new_bndrs)
832 mk_poly_bndr bndr uniq = mkSysLocal (mkFastString str) uniq poly_ty
834 str = "poly_" ++ occNameString (getOccName bndr)
835 poly_ty = mkPiTypes abs_vars (idType bndr)
839 -> [CoreBndr] -> Type -- Abstract wrt these bndrs
841 newLvlVar str vars body_ty
842 = getUniqueUs `thenLvl` \ uniq ->
843 returnUs (mkSysLocal (mkFastString str) uniq (mkPiTypes vars body_ty))
845 -- The deeply tiresome thing is that we have to apply the substitution
846 -- to the rules inside each Id. Grr. But it matters.
848 cloneVar :: TopLevelFlag -> LevelEnv -> Id -> Level -> Level -> LvlM (LevelEnv, Id)
849 cloneVar TopLevel env v ctxt_lvl dest_lvl
850 = returnUs (env, v) -- Don't clone top level things
851 cloneVar NotTopLevel env@(_,_,subst,_) v ctxt_lvl dest_lvl
853 getUs `thenLvl` \ us ->
855 (subst', v1) = cloneIdBndr subst us v
856 v2 = zap_demand ctxt_lvl dest_lvl v1
857 env' = extendCloneLvlEnv dest_lvl env subst' [(v,v2)]
861 cloneRecVars :: TopLevelFlag -> LevelEnv -> [Id] -> Level -> Level -> LvlM (LevelEnv, [Id])
862 cloneRecVars TopLevel env vs ctxt_lvl dest_lvl
863 = returnUs (env, vs) -- Don't clone top level things
864 cloneRecVars NotTopLevel env@(_,_,subst,_) vs ctxt_lvl dest_lvl
865 = ASSERT( all isId vs )
866 getUs `thenLvl` \ us ->
868 (subst', vs1) = cloneRecIdBndrs subst us vs
869 vs2 = map (zap_demand ctxt_lvl dest_lvl) vs1
870 env' = extendCloneLvlEnv dest_lvl env subst' (vs `zip` vs2)
874 -- VERY IMPORTANT: we must zap the demand info
875 -- if the thing is going to float out past a lambda,
876 -- or if it's going to top level (where things can't be strict)
877 zap_demand dest_lvl ctxt_lvl id
878 | ctxt_lvl == dest_lvl,
879 not (isTopLvl dest_lvl) = id -- Stays, and not going to top level
880 | otherwise = zapDemandIdInfo id -- Floats out