2 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
4 %************************************************************************
6 \section[FloatIn]{Floating Inwards pass}
8 %************************************************************************
10 The main purpose of @floatInwards@ is floating into branches of a
11 case, so that we don't allocate things, save them on the stack, and
12 then discover that they aren't needed in the chosen branch.
15 module FloatIn ( floatInwards ) where
17 #include "HsVersions.h"
19 import CmdLineOpts ( DynFlags, DynFlag(..), dopt )
21 import CoreUtils ( exprIsValue, exprIsDupable )
22 import CoreLint ( showPass, endPass )
23 import CoreFVs ( CoreExprWithFVs, freeVars, freeVarsOf )
24 import Id ( isOneShotLambda )
25 import Var ( Id, idType, isTyVar )
26 import Type ( isUnLiftedType )
28 import Util ( zipEqual, zipWithEqual )
32 Top-level interface function, @floatInwards@. Note that we do not
33 actually float any bindings downwards from the top-level.
36 floatInwards :: DynFlags -> [CoreBind] -> IO [CoreBind]
38 floatInwards dflags binds
40 showPass dflags "Float inwards";
41 let { binds' = map fi_top_bind binds };
42 endPass dflags "Float inwards" Opt_D_verbose_core2core binds'
43 {- no specific flag for dumping float-in -}
47 fi_top_bind (NonRec binder rhs)
48 = NonRec binder (fiExpr [] (freeVars rhs))
49 fi_top_bind (Rec pairs)
50 = Rec [ (b, fiExpr [] (freeVars rhs)) | (b, rhs) <- pairs ]
53 %************************************************************************
55 \subsection{Mail from Andr\'e [edited]}
57 %************************************************************************
59 {\em Will wrote: What??? I thought the idea was to float as far
60 inwards as possible, no matter what. This is dropping all bindings
61 every time it sees a lambda of any kind. Help! }
63 You are assuming we DO DO full laziness AFTER floating inwards! We
64 have to [not float inside lambdas] if we don't.
66 If we indeed do full laziness after the floating inwards (we could
67 check the compilation flags for that) then I agree we could be more
68 aggressive and do float inwards past lambdas.
70 Actually we are not doing a proper full laziness (see below), which
71 was another reason for not floating inwards past a lambda.
73 This can easily be fixed. The problem is that we float lets outwards,
74 but there are a few expressions which are not let bound, like case
75 scrutinees and case alternatives. After floating inwards the
76 simplifier could decide to inline the let and the laziness would be
80 let a = expensive ==> \b -> case expensive of ...
81 in \ b -> case a of ...
86 to let bind the algebraic case scrutinees (done, I think) and
87 the case alternatives (except the ones with an
88 unboxed type)(not done, I think). This is best done in the
89 SetLevels.lhs module, which tags things with their level numbers.
91 do the full laziness pass (floating lets outwards).
93 simplify. The simplifier inlines the (trivial) lets that were
94 created but were not floated outwards.
97 With the fix I think Will's suggestion that we can gain even more from
98 strictness by floating inwards past lambdas makes sense.
100 We still gain even without going past lambdas, as things may be
101 strict in the (new) context of a branch (where it was floated to) or
104 let a = something case x of
105 in case x of alt1 -> case something of a -> a + a
106 alt1 -> a + a ==> alt2 -> b
109 let a = something let b = case something of a -> a + a
110 in let b = a + a ==> in (b,b)
113 Also, even if a is not found to be strict in the new context and is
114 still left as a let, if the branch is not taken (or b is not entered)
115 the closure for a is not built.
117 %************************************************************************
119 \subsection{Main floating-inwards code}
121 %************************************************************************
124 type FreeVarsSet = IdSet
126 type FloatingBinds = [(CoreBind, FreeVarsSet)]
127 -- In reverse dependency order (innermost bindiner first)
129 -- The FreeVarsSet is the free variables of the binding. In the case
130 -- of recursive bindings, the set doesn't include the bound
133 fiExpr :: FloatingBinds -- Binds we're trying to drop
134 -- as far "inwards" as possible
135 -> CoreExprWithFVs -- Input expr
136 -> CoreExpr -- Result
138 fiExpr to_drop (_, AnnVar v) = mkCoLets' to_drop (Var v)
140 fiExpr to_drop (_, AnnType ty) = ASSERT( null to_drop )
143 fiExpr to_drop (_, AnnLit lit) = Lit lit
146 Applications: we do float inside applications, mainly because we
147 need to get at all the arguments. The next simplifier run will
148 pull out any silly ones.
151 fiExpr to_drop (_,AnnApp fun arg)
152 = mkCoLets' drop_here (App (fiExpr fun_drop fun) (fiExpr arg_drop arg))
154 [drop_here, fun_drop, arg_drop] = sepBindsByDropPoint False [freeVarsOf fun, freeVarsOf arg] to_drop
157 We are careful about lambdas:
159 * We must be careful about floating inside inside a value lambda.
160 That risks losing laziness.
161 The float-out pass might rescue us, but then again it might not.
163 * We must be careful about type lambdas too. At one time we did, and
164 there is no risk of duplicating work thereby, but we do need to be
165 careful. In particular, here is a bad case (it happened in the
168 in let f = /\t -> \a -> ...
170 let f = /\t -> let v = ... in \a -> ...
171 This is bad as now f is an updatable closure (update PAP)
174 So we treat lambda in groups, using the following rule:
176 Float inside a group of lambdas only if
177 they are all either type lambdas or one-shot lambdas.
179 Otherwise drop all the bindings outside the group.
182 fiExpr to_drop (_, AnnLam b body)
183 = case collect [b] body of
185 | all is_ok bndrs -> mkLams bndrs (fiExpr to_drop real_body)
186 | otherwise -> mkCoLets' to_drop (mkLams bndrs (fiExpr [] real_body))
188 collect bs (_, AnnLam b body) = collect (b:bs) body
189 collect bs body = (reverse bs, body)
191 is_ok bndr = isTyVar bndr || isOneShotLambda bndr
194 We don't float lets inwards past an SCC.
195 ToDo: keep info on current cc, and when passing
196 one, if it is not the same, annotate all lets in binds with current
197 cc, change current cc to the new one and float binds into expr.
200 fiExpr to_drop (_, AnnNote note@(SCC cc) expr)
201 = -- Wimp out for now
202 mkCoLets' to_drop (Note note (fiExpr [] expr))
204 fiExpr to_drop (_, AnnNote InlineCall expr)
205 = -- Wimp out for InlineCall; keep it close
206 -- the the call it annotates
207 mkCoLets' to_drop (Note InlineCall (fiExpr [] expr))
209 fiExpr to_drop (_, AnnNote InlineMe expr)
210 = -- Ditto... don't float anything into an INLINE expression
211 mkCoLets' to_drop (Note InlineMe (fiExpr [] expr))
213 fiExpr to_drop (_, AnnNote note@(Coerce _ _) expr)
214 = -- Just float in past coercion
215 Note note (fiExpr to_drop expr)
218 For @Lets@, the possible ``drop points'' for the \tr{to_drop}
219 bindings are: (a)~in the body, (b1)~in the RHS of a NonRec binding,
220 or~(b2), in each of the RHSs of the pairs of a @Rec@.
222 Note that we do {\em weird things} with this let's binding. Consider:
231 Look at the inner \tr{let}. As \tr{w} is used in both the bind and
232 body of the inner let, we could panic and leave \tr{w}'s binding where
233 it is. But \tr{v} is floatable further into the body of the inner let, and
234 {\em then} \tr{w} will also be only in the body of that inner let.
236 So: rather than drop \tr{w}'s binding here, we add it onto the list of
237 things to drop in the outer let's body, and let nature take its
241 fiExpr to_drop (_,AnnLet (AnnNonRec id rhs@(rhs_fvs, ann_rhs)) body)
242 = fiExpr new_to_drop body
244 body_fvs = freeVarsOf body
246 final_body_fvs | noFloatIntoRhs ann_rhs
247 || isUnLiftedType (idType id) = body_fvs `unionVarSet` rhs_fvs
248 | otherwise = body_fvs
249 -- See commments with letrec below
250 -- No point in floating in only to float straight out again
251 -- Ditto ok-for-speculation unlifted RHSs
253 [shared_binds, rhs_binds, body_binds] = sepBindsByDropPoint False [rhs_fvs, final_body_fvs] to_drop
255 new_to_drop = body_binds ++ -- the bindings used only in the body
256 [(NonRec id rhs', rhs_fvs')] ++ -- the new binding itself
257 shared_binds -- the bindings used both in rhs and body
259 -- Push rhs_binds into the right hand side of the binding
260 rhs' = fiExpr rhs_binds rhs
261 rhs_fvs' = rhs_fvs `unionVarSet` floatedBindsFVs rhs_binds
263 fiExpr to_drop (_,AnnLet (AnnRec bindings) body)
264 = fiExpr new_to_drop body
266 rhss = map snd bindings
268 rhss_fvs = map freeVarsOf rhss
269 body_fvs = freeVarsOf body
271 -- Add to body_fvs the free vars of any RHS that has
272 -- a lambda at the top. This has the effect of making it seem
273 -- that such things are used in the body as well, and hence prevents
274 -- them getting floated in. The big idea is to avoid turning:
277 -- letrec f = \z. ...x#...f...
280 -- letrec f = let x# = y# +# 1# in \z. ...x#...f... in ...
282 -- Because now we can't float the let out again, because a letrec
283 -- can't have unboxed bindings.
285 final_body_fvs = foldr (unionVarSet . get_extras) body_fvs rhss
286 get_extras (rhs_fvs, rhs) | noFloatIntoRhs rhs = rhs_fvs
287 | otherwise = emptyVarSet
289 (shared_binds:body_binds:rhss_binds) = sepBindsByDropPoint False (final_body_fvs:rhss_fvs) to_drop
291 new_to_drop = -- the bindings used only in the body
293 -- the new binding itself
294 [(Rec (fi_bind rhss_binds bindings), rhs_fvs')] ++
295 -- the bindings used both in rhs and body or in more than one rhs
298 rhs_fvs' = unionVarSet (unionVarSets rhss_fvs)
299 (unionVarSets (map floatedBindsFVs rhss_binds))
301 -- Push rhs_binds into the right hand side of the binding
302 fi_bind :: [FloatingBinds] -- one per "drop pt" conjured w/ fvs_of_rhss
303 -> [(Id, CoreExprWithFVs)]
306 fi_bind to_drops pairs
307 = [ (binder, fiExpr to_drop rhs)
308 | ((binder, rhs), to_drop) <- zipEqual "fi_bind" pairs to_drops ]
311 For @Case@, the possible ``drop points'' for the \tr{to_drop}
312 bindings are: (a)~inside the scrutinee, (b)~inside one of the
313 alternatives/default [default FVs always {\em first}!].
316 fiExpr to_drop (_, AnnCase scrut case_bndr alts)
317 = mkCoLets' drop_here1 $
318 mkCoLets' drop_here2 $
319 Case (fiExpr scrut_drops scrut) case_bndr
320 (zipWith fi_alt alts_drops_s alts)
322 -- Float into the scrut and alts-considered-together just like App
323 [drop_here1, scrut_drops, alts_drops] = sepBindsByDropPoint False [scrut_fvs, all_alts_fvs] to_drop
325 -- Float into the alts with the is_case flag set
326 (drop_here2 : alts_drops_s) = sepBindsByDropPoint True alts_fvs alts_drops
328 scrut_fvs = freeVarsOf scrut
329 alts_fvs = map alt_fvs alts
330 all_alts_fvs = unionVarSets alts_fvs
331 alt_fvs (con, args, rhs) = foldl delVarSet (freeVarsOf rhs) (case_bndr:args)
332 -- Delete case_bndr and args from free vars of rhs
333 -- to get free vars of alt
335 fi_alt to_drop (con, args, rhs) = (con, args, fiExpr to_drop rhs)
337 noFloatIntoRhs (AnnNote InlineMe _) = True
338 noFloatIntoRhs (AnnLam b _) = not (isId b && isOneShotLambda b)
339 -- IMPORTANT: don't say 'True' for a RHS with a one-shot lambda at the top.
340 -- This makes a big difference for things like
341 -- f x# = let x = I# x#
342 -- in let j = \() -> ...x...
343 -- in if <condition> then normal-path else j ()
344 -- If x is used only in the error case join point, j, we must float the
345 -- boxing constructor into it, else we box it every time which is very bad
348 noFloatIntoRhs rhs = exprIsValue (deAnnotate' rhs) -- We'd just float rigt back out again...
352 %************************************************************************
354 \subsection{@sepBindsByDropPoint@}
356 %************************************************************************
358 This is the crucial function. The idea is: We have a wad of bindings
359 that we'd like to distribute inside a collection of {\em drop points};
360 insides the alternatives of a \tr{case} would be one example of some
361 drop points; the RHS and body of a non-recursive \tr{let} binding
362 would be another (2-element) collection.
364 So: We're given a list of sets-of-free-variables, one per drop point,
365 and a list of floating-inwards bindings. If a binding can go into
366 only one drop point (without suddenly making something out-of-scope),
367 in it goes. If a binding is used inside {\em multiple} drop points,
368 then it has to go in a you-must-drop-it-above-all-these-drop-points
371 We have to maintain the order on these drop-point-related lists.
375 :: Bool -- True <=> is case expression
376 -> [FreeVarsSet] -- One set of FVs per drop point
377 -> FloatingBinds -- Candidate floaters
378 -> [FloatingBinds] -- FIRST one is bindings which must not be floated
379 -- inside any drop point; the rest correspond
380 -- one-to-one with the input list of FV sets
382 -- Every input floater is returned somewhere in the result;
383 -- none are dropped, not even ones which don't seem to be
384 -- free in *any* of the drop-point fvs. Why? Because, for example,
385 -- a binding (let x = E in B) might have a specialised version of
386 -- x (say x') stored inside x, but x' isn't free in E or B.
388 type DropBox = (FreeVarsSet, FloatingBinds)
390 sepBindsByDropPoint is_case drop_pts []
391 = [] : [[] | p <- drop_pts] -- cut to the chase scene; it happens
393 sepBindsByDropPoint is_case drop_pts floaters
394 = go floaters (map (\fvs -> (fvs, [])) (emptyVarSet : drop_pts))
396 go :: FloatingBinds -> [DropBox] -> [FloatingBinds]
397 -- The *first* one in the argument list is the drop_here set
398 -- The FloatingBinds in the lists are in the reverse of
399 -- the normal FloatingBinds order; that is, they are the right way round!
401 go [] drop_boxes = map (reverse . snd) drop_boxes
403 go (bind_w_fvs@(bind, bind_fvs) : binds) drop_boxes@(here_box : fork_boxes)
406 -- "here" means the group of bindings dropped at the top of the fork
408 (used_here : used_in_flags) = [ any (`elemVarSet` fvs) (bindersOf bind)
409 | (fvs, drops) <- drop_boxes]
411 drop_here = used_here || not can_push
413 -- For case expressions we duplicate the binding if it is
414 -- reasonably small, and if it is not used in all the RHSs
415 -- This is good for situations like
420 -- E -> ...not mentioning x...
422 n_alts = length used_in_flags
423 n_used_alts = length [() | True <- used_in_flags]
425 can_push = n_used_alts == 1 -- Used in just one branch
426 || (is_case && -- We are looking at case alternatives
427 n_used_alts > 1 && -- It's used in more than one
428 n_used_alts < n_alts && -- ...but not all
429 bindIsDupable bind) -- and we can duplicate the binding
431 new_boxes | drop_here = (insert here_box : fork_boxes)
432 | otherwise = (here_box : new_fork_boxes)
434 new_fork_boxes = zipWithEqual "FloatIn.sepBinds" insert_maybe fork_boxes used_in_flags
436 insert :: DropBox -> DropBox
437 insert (fvs,drops) = (fvs `unionVarSet` bind_fvs, bind_w_fvs:drops)
439 insert_maybe box True = insert box
440 insert_maybe box False = box
443 floatedBindsFVs :: FloatingBinds -> FreeVarsSet
444 floatedBindsFVs binds = unionVarSets (map snd binds)
446 mkCoLets' :: FloatingBinds -> CoreExpr -> CoreExpr
447 mkCoLets' to_drop e = foldl (flip (Let . fst)) e to_drop
448 -- Remember to_drop is in *reverse* dependency order
450 bindIsDupable (Rec prs) = all (exprIsDupable . snd) prs
451 bindIsDupable (NonRec b r) = exprIsDupable r