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"
16 import DynFlags ( DynFlags, DynFlag(..), FloatOutSwitches(..) )
17 import ErrUtils ( dumpIfSet_dyn )
18 import CostCentre ( dupifyCC, CostCentre )
19 import Id ( Id, idType )
20 import Type ( isUnLiftedType )
21 import CoreLint ( showPass, endPass )
22 import SetLevels ( Level(..), LevelledExpr, LevelledBind,
23 setLevels, ltMajLvl, ltLvl, isTopLvl )
24 import UniqSupply ( UniqSupply )
25 import List ( partition )
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) -- INVARIANT: a FloatBind is always lifted
103 type FloatBinds = [FloatBind]
106 %************************************************************************
108 \subsection[floatOutwards]{@floatOutwards@: let-floating interface function}
110 %************************************************************************
113 floatOutwards :: FloatOutSwitches
116 -> [CoreBind] -> IO [CoreBind]
118 floatOutwards float_sws dflags 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"
147 = case (floatBind bind) of { (fs, floats) ->
148 (fs, floatsToBinds floats)
152 %************************************************************************
154 \subsection[FloatOut-Bind]{Floating in a binding (the business end)}
156 %************************************************************************
160 floatBind :: LevelledBind -> (FloatStats, FloatBinds)
162 floatBind (NonRec (TB name level) rhs)
163 = case (floatRhs level rhs) of { (fs, rhs_floats, rhs') ->
164 (fs, rhs_floats ++ [(level, NonRec name rhs')]) }
166 floatBind bind@(Rec pairs)
167 = case (unzip3 (map do_pair pairs)) of { (fss, rhss_floats, new_pairs) ->
168 let rhs_floats = concat rhss_floats in
170 if not (isTopLvl bind_dest_lvl) then
171 -- Find which bindings float out at least one lambda beyond this one
172 -- These ones can't mention the binders, because they couldn't
173 -- be escaping a major level if so.
174 -- The ones that are not going further can join the letrec;
175 -- they may not be mutually recursive but the occurrence analyser will
177 case (partitionByMajorLevel bind_dest_lvl rhs_floats) of { (floats', heres) ->
178 (sum_stats fss, floats' ++ [(bind_dest_lvl, Rec (floatsToBindPairs heres ++ new_pairs))]) }
180 -- In a recursive binding, *destined for* the top level
181 -- (only), the rhs floats may contain references to the
182 -- bound things. For example
183 -- f = ...(let v = ...f... in b) ...
184 -- might get floated to
187 -- and hence we must (pessimistically) make all the floats recursive
188 -- with the top binding. Later dependency analysis will unravel it.
190 -- This can only happen for bindings destined for the top level,
191 -- because only then will partitionByMajorLevel allow through a binding
192 -- that only differs in its minor level
193 (sum_stats fss, [(bind_dest_lvl, Rec (new_pairs ++ floatsToBindPairs rhs_floats))])
196 bind_dest_lvl = getBindLevel bind
198 do_pair (TB name level, rhs)
199 = case (floatRhs level rhs) of { (fs, rhs_floats, rhs') ->
200 (fs, rhs_floats, (name, rhs'))
204 %************************************************************************
206 \subsection[FloatOut-Expr]{Floating in expressions}
208 %************************************************************************
211 floatExpr, floatRhs, floatCaseAlt
214 -> (FloatStats, FloatBinds, CoreExpr)
216 floatCaseAlt lvl arg -- Used rec rhss, and case-alternative rhss
217 = case (floatExpr lvl arg) of { (fsa, floats, arg') ->
218 case (partitionByMajorLevel lvl floats) of { (floats', heres) ->
219 -- Dump bindings that aren't going to escape from a lambda;
220 -- in particular, we must dump the ones that are bound by
221 -- the rec or case alternative
222 (fsa, floats', install heres arg') }}
224 floatRhs lvl arg -- Used for nested non-rec rhss, and fn args
225 -- See Note [Floating out of RHS]
226 = case (floatExpr lvl arg) of { (fsa, floats, arg') ->
227 if exprIsCheap arg' then
230 case (partitionByMajorLevel lvl floats) of { (floats', heres) ->
231 (fsa, floats', install heres arg') }}
233 -- Note [Floating out of RHSs]
234 -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~
235 -- Dump bindings that aren't going to escape from a lambda
236 -- This isn't a scoping issue (the binder isn't in scope in the RHS
237 -- of a non-rec binding)
238 -- Rather, it is to avoid floating the x binding out of
239 -- f (let x = e in b)
240 -- unnecessarily. But we first test for values or trival rhss,
241 -- because (in particular) we don't want to insert new bindings between
242 -- the "=" and the "\". E.g.
243 -- f = \x -> let <bind> in <body>
245 -- f = let <bind> in \x -> <body>
246 -- (a) The simplifier will immediately float it further out, so we may
247 -- as well do so right now; in general, keeping rhss as manifest
249 -- (b) If a float-in pass follows immediately, it might add yet more
250 -- bindings just after the '='. And some of them might (correctly)
251 -- be strict even though the 'let f' is lazy, because f, being a value,
252 -- gets its demand-info zapped by the simplifier.
254 -- We use exprIsCheap because that is also what's used by the simplifier
255 -- to decide whether to float a let out of a let
257 floatExpr _ (Var v) = (zeroStats, [], Var v)
258 floatExpr _ (Type ty) = (zeroStats, [], Type ty)
259 floatExpr _ (Lit lit) = (zeroStats, [], Lit lit)
261 floatExpr lvl (App e a)
262 = case (floatExpr lvl e) of { (fse, floats_e, e') ->
263 case (floatRhs lvl a) of { (fsa, floats_a, a') ->
264 (fse `add_stats` fsa, floats_e ++ floats_a, App e' a') }}
266 floatExpr lvl lam@(Lam _ _)
268 (bndrs_w_lvls, body) = collectBinders lam
269 bndrs = [b | TB b _ <- bndrs_w_lvls]
270 lvls = [l | TB b l <- bndrs_w_lvls]
272 -- For the all-tyvar case we are prepared to pull
273 -- the lets out, to implement the float-out-of-big-lambda
274 -- transform; but otherwise we only float bindings that are
275 -- going to escape a value lambda.
276 -- In particular, for one-shot lambdas we don't float things
277 -- out; we get no saving by so doing.
278 partition_fn | all isTyVar bndrs = partitionByLevel
279 | otherwise = partitionByMajorLevel
281 case (floatExpr (last lvls) body) of { (fs, floats, body') ->
283 -- Dump any bindings which absolutely cannot go any further
284 case (partition_fn (head lvls) floats) of { (floats', heres) ->
286 (add_to_stats fs floats', floats', mkLams bndrs (install heres body'))
289 floatExpr lvl (Note note@(SCC cc) expr)
290 = case (floatExpr lvl expr) of { (fs, floating_defns, expr') ->
292 -- Annotate bindings floated outwards past an scc expression
293 -- with the cc. We mark that cc as "duplicated", though.
295 annotated_defns = annotate (dupifyCC cc) floating_defns
297 (fs, annotated_defns, Note note expr') }
299 annotate :: CostCentre -> FloatBinds -> FloatBinds
301 annotate dupd_cc defn_groups
302 = [ (level, ann_bind floater) | (level, floater) <- defn_groups ]
304 ann_bind (NonRec binder rhs)
305 = NonRec binder (mkSCC dupd_cc rhs)
308 = Rec [(binder, mkSCC dupd_cc rhs) | (binder, rhs) <- pairs]
310 floatExpr lvl (Note InlineMe expr) -- Other than SCCs
311 = case floatExpr InlineCtxt expr of { (fs, floating_defns, expr') ->
312 -- There can be some floating_defns, arising from
313 -- ordinary lets that were there all the time. It seems
314 -- more efficient to test once here than to avoid putting
315 -- them into floating_defns (which would mean testing for
316 -- inlineCtxt at every let)
317 (fs, [], Note InlineMe (install floating_defns expr')) } -- See notes in SetLevels
319 floatExpr lvl (Note note expr) -- Other than SCCs
320 = case (floatExpr lvl expr) of { (fs, floating_defns, expr') ->
321 (fs, floating_defns, Note note expr') }
323 floatExpr lvl (Cast expr co)
324 = case (floatExpr lvl expr) of { (fs, floating_defns, expr') ->
325 (fs, floating_defns, Cast expr' co) }
327 floatExpr lvl (Let (NonRec (TB bndr bndr_lvl) rhs) body)
328 | isUnLiftedType (idType bndr) -- Treat unlifted lets just like a case
329 -- I.e. floatExpr for rhs, floatCaseAlt for body
330 = case floatExpr lvl rhs of { (fs, rhs_floats, rhs') ->
331 case floatCaseAlt bndr_lvl body of { (fs, body_floats, body') ->
332 (fs, rhs_floats ++ body_floats, Let (NonRec bndr rhs') body') }}
334 floatExpr lvl (Let bind body)
335 = case (floatBind bind) of { (fsb, bind_floats) ->
336 case (floatExpr lvl body) of { (fse, body_floats, body') ->
338 bind_floats ++ body_floats,
341 floatExpr lvl (Case scrut (TB case_bndr case_lvl) ty alts)
342 = case floatExpr lvl scrut of { (fse, fde, scrut') ->
343 case floatList float_alt alts of { (fsa, fda, alts') ->
344 (add_stats fse fsa, fda ++ fde, Case scrut' case_bndr ty alts')
347 -- Use floatCaseAlt for the alternatives, so that we
348 -- don't gratuitiously float bindings out of the RHSs
349 float_alt (con, bs, rhs)
350 = case (floatCaseAlt case_lvl rhs) of { (fs, rhs_floats, rhs') ->
351 (fs, rhs_floats, (con, [b | TB b _ <- bs], rhs')) }
354 floatList :: (a -> (FloatStats, FloatBinds, b)) -> [a] -> (FloatStats, FloatBinds, [b])
355 floatList f [] = (zeroStats, [], [])
356 floatList f (a:as) = case f a of { (fs_a, binds_a, b) ->
357 case floatList f as of { (fs_as, binds_as, bs) ->
358 (fs_a `add_stats` fs_as, binds_a ++ binds_as, b:bs) }}
361 %************************************************************************
363 \subsection{Utility bits for floating stats}
365 %************************************************************************
367 I didn't implement this with unboxed numbers. I don't want to be too
368 strict in this stuff, as it is rarely turned on. (WDP 95/09)
372 = FlS Int -- Number of top-floats * lambda groups they've been past
373 Int -- Number of non-top-floats * lambda groups they've been past
374 Int -- Number of lambda (groups) seen
376 get_stats (FlS a b c) = (a, b, c)
378 zeroStats = FlS 0 0 0
380 sum_stats xs = foldr add_stats zeroStats xs
382 add_stats (FlS a1 b1 c1) (FlS a2 b2 c2)
383 = FlS (a1 + a2) (b1 + b2) (c1 + c2)
385 add_to_stats (FlS a b c) floats
386 = FlS (a + length top_floats) (b + length other_floats) (c + 1)
388 (top_floats, other_floats) = partition to_very_top floats
390 to_very_top (my_lvl, _) = isTopLvl my_lvl
394 %************************************************************************
396 \subsection{Utility bits for floating}
398 %************************************************************************
401 getBindLevel (NonRec (TB _ lvl) _) = lvl
402 getBindLevel (Rec (((TB _ lvl), _) : _)) = lvl
406 partitionByMajorLevel, partitionByLevel
407 :: Level -- Partitioning level
409 -> FloatBinds -- Defns to be divided into 2 piles...
411 -> (FloatBinds, -- Defns with level strictly < partition level,
412 FloatBinds) -- The rest
415 partitionByMajorLevel ctxt_lvl defns
416 = partition float_further defns
418 -- Float it if we escape a value lambda, or if we get to the top level
419 float_further (my_lvl, bind) = my_lvl `ltMajLvl` ctxt_lvl || isTopLvl my_lvl
420 -- The isTopLvl part says that if we can get to the top level, say "yes" anyway
426 -- which is as it should be
428 partitionByLevel ctxt_lvl defns
429 = partition float_further defns
431 float_further (my_lvl, _) = my_lvl `ltLvl` ctxt_lvl
435 floatsToBinds :: FloatBinds -> [CoreBind]
436 floatsToBinds floats = map snd floats
438 floatsToBindPairs :: FloatBinds -> [(Id,CoreExpr)]
440 floatsToBindPairs floats = concat (map mk_pairs floats)
442 mk_pairs (_, Rec pairs) = pairs
443 mk_pairs (_, NonRec binder rhs) = [(binder,rhs)]
445 install :: FloatBinds -> CoreExpr -> CoreExpr
447 install defn_groups expr
448 = foldr install_group expr defn_groups
450 install_group (_, defns) body = Let defns body