2 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
4 \section[FloatOut]{Float bindings outwards (towards the top level)}
6 ``Long-distance'' floating of bindings towards the top level.
9 module FloatOut ( floatOutwards ) where
11 #include "HsVersions.h"
14 import CoreUtils ( mkSCC )
16 import CmdLineOpts ( DynFlags, DynFlag(..), FloatOutSwitches(..) )
17 import ErrUtils ( dumpIfSet_dyn )
18 import CostCentre ( dupifyCC, CostCentre )
20 import CoreLint ( showPass, endPass )
21 import SetLevels ( Level(..), LevelledExpr, LevelledBind,
22 setLevels, ltMajLvl, ltLvl, isTopLvl )
23 import UniqSupply ( UniqSupply )
24 import List ( partition )
26 import Util ( notNull )
35 To float out sub-expressions that can thereby get outside
36 a non-one-shot value lambda, and hence may be shared.
39 To achieve this we may need to do two thing:
41 a) Let-bind the sub-expression:
43 f (g x) ==> let lvl = f (g x) in lvl
45 Now we can float the binding for 'lvl'.
47 b) More than that, we may need to abstract wrt a type variable
49 \x -> ... /\a -> let v = ...a... in ....
51 Here the binding for v mentions 'a' but not 'x'. So we
52 abstract wrt 'a', to give this binding for 'v':
57 Now the binding for vp can float out unimpeded.
58 I can't remember why this case seemed important enough to
59 deal with, but I certainly found cases where important floats
60 didn't happen if we did not abstract wrt tyvars.
62 With this in mind we can also achieve another goal: lambda lifting.
63 We can make an arbitrary (function) binding float to top level by
64 abstracting wrt *all* local variables, not just type variables, leaving
65 a binding that can be floated right to top level. Whether or not this
66 happens is controlled by a flag.
72 At the moment we never float a binding out to between two adjacent
76 \x y -> let t = x+x in ...
78 \x -> let t = x+x in \y -> ...
80 Reason: this is less efficient in the case where the original lambda
81 is never partially applied.
83 But there's a case I've seen where this might not be true. Consider:
89 elem' x (y:ys) = x==y || elem' x ys
91 It turns out that this generates a subexpression of the form
93 \deq x ys -> let eq = eqFromEqDict deq in ...
95 vwhich might usefully be separated to
97 \deq -> let eq = eqFromEqDict deq in \xy -> ...
99 Well, maybe. We don't do this at the moment.
102 type FloatBind = (Level, CoreBind)
103 type FloatBinds = [FloatBind]
106 %************************************************************************
108 \subsection[floatOutwards]{@floatOutwards@: let-floating interface function}
110 %************************************************************************
113 floatOutwards :: DynFlags
116 -> [CoreBind] -> IO [CoreBind]
118 floatOutwards dflags float_sws us pgm
120 showPass dflags float_msg ;
122 let { annotated_w_levels = setLevels float_sws pgm us ;
123 (fss, binds_s') = unzip (map floatTopBind annotated_w_levels)
126 dumpIfSet_dyn dflags Opt_D_verbose_core2core "Levels added:"
127 (vcat (map ppr annotated_w_levels));
129 let { (tlets, ntlets, lams) = get_stats (sum_stats fss) };
131 dumpIfSet_dyn dflags Opt_D_dump_simpl_stats "FloatOut stats:"
132 (hcat [ int tlets, ptext SLIT(" Lets floated to top level; "),
133 int ntlets, ptext SLIT(" Lets floated elsewhere; from "),
134 int lams, ptext SLIT(" Lambda groups")]);
136 endPass dflags float_msg Opt_D_verbose_core2core (concat binds_s')
137 {- no specific flag for dumping float-out -}
140 float_msg = showSDoc (text "Float out" <+> parens (sws float_sws))
141 sws (FloatOutSw lam const) = pp_not lam <+> text "lambdas" <> comma <+>
142 pp_not const <+> text "constants"
144 pp_not False = text "not"
146 floatTopBind bind@(NonRec _ _)
147 = case (floatBind bind) of { (fs, floats, bind') ->
148 (fs, floatsToBinds floats ++ [bind'])
151 floatTopBind bind@(Rec _)
152 = case (floatBind bind) of { (fs, floats, Rec pairs') ->
153 WARN( notNull floats, ppr bind $$ ppr floats )
154 (fs, [Rec (floatsToBindPairs floats ++ pairs')]) }
157 %************************************************************************
159 \subsection[FloatOut-Bind]{Floating in a binding (the business end)}
161 %************************************************************************
165 floatBind :: LevelledBind
166 -> (FloatStats, FloatBinds, CoreBind)
168 floatBind (NonRec (TB name level) rhs)
169 = case (floatRhs level rhs) of { (fs, rhs_floats, rhs') ->
170 (fs, rhs_floats, NonRec name rhs') }
172 floatBind bind@(Rec pairs)
173 = case (unzip3 (map do_pair pairs)) of { (fss, rhss_floats, new_pairs) ->
175 if not (isTopLvl bind_level) then
177 (sum_stats fss, concat rhss_floats, Rec new_pairs)
179 -- In a recursive binding, *destined for* the top level
180 -- (only), the rhs floats may contain references to the
181 -- bound things. For example
183 -- f = ...(let v = ...f... in b) ...
185 -- might get floated to
190 -- and hence we must (pessimistically) make all the floats recursive
191 -- with the top binding. Later dependency analysis will unravel it.
193 -- Can't happen on nested bindings because floatRhs will dump
194 -- the bindings in the RHS (partitionByMajorLevel treats top specially)
196 Rec (new_pairs ++ floatsToBindPairs (concat rhss_floats)))
199 bind_level = getBindLevel bind
201 do_pair (TB name level, rhs)
202 = case (floatRhs level rhs) of { (fs, rhs_floats, rhs') ->
203 (fs, rhs_floats, (name, rhs'))
207 %************************************************************************
209 \subsection[FloatOut-Expr]{Floating in expressions}
211 %************************************************************************
217 -> (FloatStats, FloatBinds, CoreExpr)
220 = case (floatExpr lvl arg) of { (fsa, floats, arg') ->
221 case (partitionByMajorLevel lvl floats) of { (floats', heres) ->
222 -- Dump bindings that aren't going to escape from a lambda
223 -- This is to avoid floating the x binding out of
224 -- f (let x = e in b)
225 -- unnecessarily. It even causes a bug to do so if we have
226 -- y = writeArr# a n (let x = e in b)
227 -- because the y binding is an expr-ok-for-speculation one.
228 -- [SLPJ Dec 01: I don't understand this last comment;
229 -- writeArr# is not ok-for-spec because of its side effect]
230 (fsa, floats', install heres arg') }}
232 floatExpr _ (Var v) = (zeroStats, [], Var v)
233 floatExpr _ (Type ty) = (zeroStats, [], Type ty)
234 floatExpr _ (Lit lit) = (zeroStats, [], Lit lit)
236 floatExpr lvl (App e a)
237 = case (floatExpr lvl e) of { (fse, floats_e, e') ->
238 case (floatRhs lvl a) of { (fsa, floats_a, a') ->
239 (fse `add_stats` fsa, floats_e ++ floats_a, App e' a') }}
241 floatExpr lvl lam@(Lam _ _)
243 (bndrs_w_lvls, body) = collectBinders lam
244 bndrs = [b | TB b _ <- bndrs_w_lvls]
245 lvls = [l | TB b l <- bndrs_w_lvls]
247 -- For the all-tyvar case we are prepared to pull
248 -- the lets out, to implement the float-out-of-big-lambda
249 -- transform; but otherwise we only float bindings that are
250 -- going to escape a value lambda.
251 -- In particular, for one-shot lambdas we don't float things
252 -- out; we get no saving by so doing.
253 partition_fn | all isTyVar bndrs = partitionByLevel
254 | otherwise = partitionByMajorLevel
256 case (floatExpr (last lvls) body) of { (fs, floats, body') ->
258 -- Dump any bindings which absolutely cannot go any further
259 case (partition_fn (head lvls) floats) of { (floats', heres) ->
261 (add_to_stats fs floats', floats', mkLams bndrs (install heres body'))
264 floatExpr lvl (Note note@(SCC cc) expr)
265 = case (floatExpr lvl expr) of { (fs, floating_defns, expr') ->
267 -- Annotate bindings floated outwards past an scc expression
268 -- with the cc. We mark that cc as "duplicated", though.
270 annotated_defns = annotate (dupifyCC cc) floating_defns
272 (fs, annotated_defns, Note note expr') }
274 annotate :: CostCentre -> FloatBinds -> FloatBinds
276 annotate dupd_cc defn_groups
277 = [ (level, ann_bind floater) | (level, floater) <- defn_groups ]
279 ann_bind (NonRec binder rhs)
280 = NonRec binder (mkSCC dupd_cc rhs)
283 = Rec [(binder, mkSCC dupd_cc rhs) | (binder, rhs) <- pairs]
285 floatExpr lvl (Note InlineMe expr) -- Other than SCCs
286 = case floatExpr InlineCtxt expr of { (fs, floating_defns, expr') ->
287 -- There can be some floating_defns, arising from
288 -- ordinary lets that were there all the time. It seems
289 -- more efficient to test once here than to avoid putting
290 -- them into floating_defns (which would mean testing for
291 -- inlineCtxt at every let)
292 (fs, [], Note InlineMe (install floating_defns expr')) } -- See notes in SetLevels
294 floatExpr lvl (Note note expr) -- Other than SCCs
295 = case (floatExpr lvl expr) of { (fs, floating_defns, expr') ->
296 (fs, floating_defns, Note note expr') }
298 floatExpr lvl (Let bind body)
299 = case (floatBind bind) of { (fsb, rhs_floats, bind') ->
300 case (floatExpr lvl body) of { (fse, body_floats, body') ->
301 -- if isInlineCtxt lvl then -- No floating inside an InlineMe
302 -- ASSERT( null rhs_floats && null body_floats )
303 -- (add_stats fsb fse, [], Let bind' body')
306 rhs_floats ++ [(bind_lvl, bind')] ++ body_floats,
310 bind_lvl = getBindLevel bind
312 floatExpr lvl (Case scrut (TB case_bndr case_lvl) alts)
313 = case floatExpr lvl scrut of { (fse, fde, scrut') ->
314 case floatList float_alt alts of { (fsa, fda, alts') ->
315 (add_stats fse fsa, fda ++ fde, Case scrut' case_bndr alts')
318 -- Use floatRhs for the alternatives, so that we
319 -- don't gratuitiously float bindings out of the RHSs
320 float_alt (con, bs, rhs)
321 = case (floatRhs case_lvl rhs) of { (fs, rhs_floats, rhs') ->
322 (fs, rhs_floats, (con, [b | TB b _ <- bs], rhs')) }
325 floatList :: (a -> (FloatStats, FloatBinds, b)) -> [a] -> (FloatStats, FloatBinds, [b])
326 floatList f [] = (zeroStats, [], [])
327 floatList f (a:as) = case f a of { (fs_a, binds_a, b) ->
328 case floatList f as of { (fs_as, binds_as, bs) ->
329 (fs_a `add_stats` fs_as, binds_a ++ binds_as, b:bs) }}
332 %************************************************************************
334 \subsection{Utility bits for floating stats}
336 %************************************************************************
338 I didn't implement this with unboxed numbers. I don't want to be too
339 strict in this stuff, as it is rarely turned on. (WDP 95/09)
343 = FlS Int -- Number of top-floats * lambda groups they've been past
344 Int -- Number of non-top-floats * lambda groups they've been past
345 Int -- Number of lambda (groups) seen
347 get_stats (FlS a b c) = (a, b, c)
349 zeroStats = FlS 0 0 0
351 sum_stats xs = foldr add_stats zeroStats xs
353 add_stats (FlS a1 b1 c1) (FlS a2 b2 c2)
354 = FlS (a1 + a2) (b1 + b2) (c1 + c2)
356 add_to_stats (FlS a b c) floats
357 = FlS (a + length top_floats) (b + length other_floats) (c + 1)
359 (top_floats, other_floats) = partition to_very_top floats
361 to_very_top (my_lvl, _) = isTopLvl my_lvl
365 %************************************************************************
367 \subsection{Utility bits for floating}
369 %************************************************************************
372 getBindLevel (NonRec (TB _ lvl) _) = lvl
373 getBindLevel (Rec (((TB _ lvl), _) : _)) = lvl
377 partitionByMajorLevel, partitionByLevel
378 :: Level -- Partitioning level
380 -> FloatBinds -- Defns to be divided into 2 piles...
382 -> (FloatBinds, -- Defns with level strictly < partition level,
383 FloatBinds) -- The rest
386 partitionByMajorLevel ctxt_lvl defns
387 = partition float_further defns
389 -- Float it if we escape a value lambda, or if we get to the top level
390 float_further (my_lvl, bind) = my_lvl `ltMajLvl` ctxt_lvl || isTopLvl my_lvl
391 -- The isTopLvl part says that if we can get to the top level, say "yes" anyway
397 -- which is as it should be
399 partitionByLevel ctxt_lvl defns
400 = partition float_further defns
402 float_further (my_lvl, _) = my_lvl `ltLvl` ctxt_lvl
406 floatsToBinds :: FloatBinds -> [CoreBind]
407 floatsToBinds floats = map snd floats
409 floatsToBindPairs :: FloatBinds -> [(Id,CoreExpr)]
411 floatsToBindPairs floats = concat (map mk_pairs floats)
413 mk_pairs (_, Rec pairs) = pairs
414 mk_pairs (_, NonRec binder rhs) = [(binder,rhs)]
416 install :: FloatBinds -> CoreExpr -> CoreExpr
418 install defn_groups expr
419 = foldr install_group expr defn_groups
421 install_group (_, defns) body = Let defns body