2 % (c) The AQUA Project, Glasgow University, 1994-1995
4 \section[LambdaLift]{A STG-code lambda lifter}
7 #include "HsVersions.h"
9 module LambdaLift ( liftProgram ) where
13 import Type ( mkForallTy, splitForalls, glueTyArgs,
14 Type, RhoType(..), TauType(..)
17 import Id ( mkSysLocal, idType, addIdArity, Id )
20 import SrcLoc ( mkUnknownSrcLoc, SrcLoc )
25 This is the lambda lifter. It turns lambda abstractions into
26 supercombinators on a selective basis:
28 * Let-no-escaped bindings are never lifted. That's one major reason
29 why the lambda lifter is done in STG.
31 * Non-recursive bindings whose RHS is a lambda abstractions are lifted,
32 provided all the occurrences of the bound variable is in a function
33 postition. In this example, f will be lifted:
41 $f p q r x = e -- Supercombinator
43 ..($f p q r a1)...($f p q r a2)...
45 NOTE that the original binding is eliminated.
47 But in this case, f won't be lifted:
54 Why? Because we have to heap-allocate a closure for f thus:
56 $f p q r x = e -- Supercombinator
61 ..(g f)...($f p q r a2)..
63 so it might as well be the original lambda abstraction.
65 We also do not lift if the function has an occurrence with no arguments, e.g.
71 as this form is more efficient than if we create a partial application
73 $f p q r x = e -- Supercombinator
77 * Recursive bindings *all* of whose RHSs are lambda abstractions are
79 - all the occurrences of all the binders are in a function position
80 - there aren't ``too many'' free variables.
82 Same reasoning as before for the function-position stuff. The ``too many
83 free variable'' part comes from considering the (potentially many)
84 recursive calls, which may now have lots of free vars.
87 * 2 might be already ``too many'' variables to abstract.
88 The problem is that the increase in the number of free variables
89 of closures refering to the lifted function (which is always # of
90 abstracted args - 1) may increase heap allocation a lot.
91 Expeiments are being done to check this...
92 * We do not lambda lift if the function has at least one occurrence
93 without any arguments. This caused lots of problems. Ex:
94 h = \ x -> ... let y = ...
95 in let let f = \x -> ...y...
99 h = \ x -> ... let y = ...
102 now f y is a partial application, so it will be updated, and this
106 --- NOT RELEVANT FOR STG ----
107 * All ``lone'' lambda abstractions are lifted. Notably this means lambda
109 - in a case alternative: case e of True -> (\x->b)
110 - in the body of a let: let x=e in (\y->b)
111 -----------------------------
113 %************************************************************************
115 \subsection[Lift-expressions]{The main function: liftExpr}
117 %************************************************************************
120 liftProgram :: UniqSupply -> [StgBinding] -> [StgBinding]
121 liftProgram us prog = concat (runLM Nothing us (mapLM liftTopBind prog))
124 liftTopBind :: StgBinding -> LiftM [StgBinding]
125 liftTopBind (StgNonRec id rhs)
126 = dontLiftRhs rhs `thenLM` \ (rhs', rhs_info) ->
127 returnLM (getScBinds rhs_info ++ [StgNonRec id rhs'])
129 liftTopBind (StgRec pairs)
130 = mapAndUnzipLM dontLiftRhs rhss `thenLM` \ (rhss', rhs_infos) ->
131 returnLM ([co_rec_ify (StgRec (ids `zip` rhss') :
132 getScBinds (unionLiftInfos rhs_infos))
135 (ids, rhss) = unzip pairs
141 -> LiftM (StgExpr, LiftInfo)
144 liftExpr expr@(StgCon con args lvs) = returnLM (expr, emptyLiftInfo)
145 liftExpr expr@(StgPrim op args lvs) = returnLM (expr, emptyLiftInfo)
147 liftExpr expr@(StgApp (StgLitArg lit) args lvs) = returnLM (expr, emptyLiftInfo)
148 liftExpr expr@(StgApp (StgVarArg v) args lvs)
149 = lookup v `thenLM` \ ~(sc, sc_args) -> -- NB the ~. We don't want to
150 -- poke these bindings too early!
151 returnLM (StgApp (StgVarArg sc) (map StgVarArg sc_args ++ args) lvs,
153 -- The lvs field is probably wrong, but we reconstruct it
154 -- anyway following lambda lifting
156 liftExpr (StgCase scrut lv1 lv2 uniq alts)
157 = liftExpr scrut `thenLM` \ (scrut', scrut_info) ->
158 lift_alts alts `thenLM` \ (alts', alts_info) ->
159 returnLM (StgCase scrut' lv1 lv2 uniq alts', scrut_info `unionLiftInfo` alts_info)
161 lift_alts (StgAlgAlts ty alg_alts deflt)
162 = mapAndUnzipLM lift_alg_alt alg_alts `thenLM` \ (alg_alts', alt_infos) ->
163 lift_deflt deflt `thenLM` \ (deflt', deflt_info) ->
164 returnLM (StgAlgAlts ty alg_alts' deflt', foldr unionLiftInfo deflt_info alt_infos)
166 lift_alts (StgPrimAlts ty prim_alts deflt)
167 = mapAndUnzipLM lift_prim_alt prim_alts `thenLM` \ (prim_alts', alt_infos) ->
168 lift_deflt deflt `thenLM` \ (deflt', deflt_info) ->
169 returnLM (StgPrimAlts ty prim_alts' deflt', foldr unionLiftInfo deflt_info alt_infos)
171 lift_alg_alt (con, args, use_mask, rhs)
172 = liftExpr rhs `thenLM` \ (rhs', rhs_info) ->
173 returnLM ((con, args, use_mask, rhs'), rhs_info)
175 lift_prim_alt (lit, rhs)
176 = liftExpr rhs `thenLM` \ (rhs', rhs_info) ->
177 returnLM ((lit, rhs'), rhs_info)
179 lift_deflt StgNoDefault = returnLM (StgNoDefault, emptyLiftInfo)
180 lift_deflt (StgBindDefault var used rhs)
181 = liftExpr rhs `thenLM` \ (rhs', rhs_info) ->
182 returnLM (StgBindDefault var used rhs', rhs_info)
185 Now the interesting cases. Let no escape isn't lifted. We turn it
186 back into a let, to play safe, because we have to redo that pass after
190 liftExpr (StgLetNoEscape _ _ (StgNonRec binder rhs) body)
191 = dontLiftRhs rhs `thenLM` \ (rhs', rhs_info) ->
192 liftExpr body `thenLM` \ (body', body_info) ->
193 returnLM (StgLet (StgNonRec binder rhs') body',
194 rhs_info `unionLiftInfo` body_info)
196 liftExpr (StgLetNoEscape _ _ (StgRec pairs) body)
197 = liftExpr body `thenLM` \ (body', body_info) ->
198 mapAndUnzipLM dontLiftRhs rhss `thenLM` \ (rhss', rhs_infos) ->
199 returnLM (StgLet (StgRec (binders `zipEqual` rhss')) body',
200 foldr unionLiftInfo body_info rhs_infos)
202 (binders,rhss) = unzip pairs
206 liftExpr (StgLet (StgNonRec binder rhs) body)
207 | not (isLiftable rhs)
208 = dontLiftRhs rhs `thenLM` \ (rhs', rhs_info) ->
209 liftExpr body `thenLM` \ (body', body_info) ->
210 returnLM (StgLet (StgNonRec binder rhs') body',
211 rhs_info `unionLiftInfo` body_info)
213 | otherwise -- It's a lambda
214 = -- Do the body of the let
215 fixLM (\ ~(sc_inline, _, _) ->
216 addScInlines [binder] [sc_inline] (
218 ) `thenLM` \ (body', body_info) ->
221 dontLiftRhs rhs `thenLM` \ (rhs', rhs_info) ->
223 -- All occurrences in function position, so lambda lift
224 getFinalFreeVars (rhsFreeVars rhs) `thenLM` \ final_free_vars ->
226 mkScPieces final_free_vars (binder,rhs') `thenLM` \ (sc_inline, sc_bind) ->
230 nonRecScBind rhs_info sc_bind `unionLiftInfo` body_info)
232 ) `thenLM` \ (_, expr', final_info) ->
234 returnLM (expr', final_info)
236 liftExpr (StgLet (StgRec pairs) body)
238 | not (all isLiftableRec rhss)
239 = liftExpr body `thenLM` \ (body', body_info) ->
240 mapAndUnzipLM dontLiftRhs rhss `thenLM` \ (rhss', rhs_infos) ->
241 returnLM (StgLet (StgRec (binders `zipEqual` rhss')) body',
242 foldr unionLiftInfo body_info rhs_infos)
244 | otherwise -- All rhss are liftable
245 = -- Do the body of the let
246 fixLM (\ ~(sc_inlines, _, _) ->
247 addScInlines binders sc_inlines (
249 liftExpr body `thenLM` \ (body', body_info) ->
250 mapAndUnzipLM dontLiftRhs rhss `thenLM` \ (rhss', rhs_infos) ->
252 -- Find the free vars of all the rhss,
253 -- excluding the binders themselves.
254 rhs_free_vars = unionManyUniqSets (map rhsFreeVars rhss)
258 rhs_info = unionLiftInfos rhs_infos
260 getFinalFreeVars rhs_free_vars `thenLM` \ final_free_vars ->
262 mapAndUnzipLM (mkScPieces final_free_vars) (binders `zip` rhss')
263 `thenLM` \ (sc_inlines, sc_pairs) ->
264 returnLM (sc_inlines,
266 recScBind rhs_info sc_pairs `unionLiftInfo` body_info)
268 )) `thenLM` \ (_, expr', final_info) ->
270 returnLM (expr', final_info)
272 (binders,rhss) = unzip pairs
276 liftExpr (StgSCC ty cc expr)
277 = liftExpr expr `thenLM` \ (expr2, expr_info) ->
278 returnLM (StgSCC ty cc expr2, expr_info)
281 A binding is liftable if it's a *function* (args not null) and never
282 occurs in an argument position.
285 isLiftable :: StgRhs -> Bool
287 isLiftable (StgRhsClosure _ (StgBinderInfo arg_occ _ _ _ unapplied_occ) fvs _ args _)
289 -- Experimental evidence suggests we should lift only if we will be
290 -- abstracting up to 4 fvs.
292 = if not (null args || -- Not a function
293 unapplied_occ || -- Has an occ with no args at all
294 arg_occ || -- Occurs in arg position
295 length fvs > 4 -- Too many free variables
297 then {-trace ("LL: " ++ show (length fvs))-} True
299 isLiftable other_rhs = False
301 isLiftableRec :: StgRhs -> Bool
303 -- this is just the same as for non-rec, except we only lift to
304 -- abstract up to 1 argument this avoids undoing Static Argument
305 -- Transformation work
307 {- Andre's longer comment about isLiftableRec: 1996/01:
309 A rec binding is "liftable" (according to our heuristics) if:
311 * all occurrences have arguments,
312 * does not occur in an argument position and
313 * has up to *2* free variables (including the rec binding variable
316 The point is: my experiments show that SAT is more important than LL.
317 Therefore if we still want to do LL, for *recursive* functions, we do
318 not want LL to undo what SAT did. We do this by avoiding LL recursive
319 functions that have more than 2 fvs, since if this recursive function
320 was created by SAT (we don't know!), it would have at least 3 fvs: one
321 for the rec binding itself and 2 more for the static arguments (note:
322 this matches with the choice of performing SAT to have at least 2
323 static arguments, if we change things there we should change things
327 isLiftableRec (StgRhsClosure _ (StgBinderInfo arg_occ _ _ _ unapplied_occ) fvs _ args _)
328 = if not (null args || -- Not a function
329 unapplied_occ || -- Has an occ with no args at all
330 arg_occ || -- Occurs in arg position
331 length fvs > 2 -- Too many free variables
333 then {-trace ("LLRec: " ++ show (length fvs))-} True
335 isLiftableRec other_rhs = False
337 rhsFreeVars :: StgRhs -> IdSet
338 rhsFreeVars (StgRhsClosure _ _ fvs _ _ _) = mkUniqSet fvs
339 rhsFreeVars other = panic "rhsFreeVars"
342 dontLiftRhs is like liftExpr, except that it does not lift a top-level
343 lambda abstraction. It is used for the right-hand sides of
344 definitions where we've decided *not* to lift: for example, top-level
345 ones or mutually-recursive ones where not all are lambdas.
348 dontLiftRhs :: StgRhs -> LiftM (StgRhs, LiftInfo)
350 dontLiftRhs rhs@(StgRhsCon cc v args) = returnLM (rhs, emptyLiftInfo)
352 dontLiftRhs (StgRhsClosure cc bi fvs upd args body)
353 = liftExpr body `thenLM` \ (body', body_info) ->
354 returnLM (StgRhsClosure cc bi fvs upd args body', body_info)
358 mkScPieces :: IdSet -- Extra args for the supercombinator
359 -> (Id, StgRhs) -- The processed RHS and original Id
360 -> LiftM ((Id,[Id]), -- Replace abstraction with this;
361 -- the set is its free vars
362 (Id,StgRhs)) -- Binding for supercombinator
364 mkScPieces extra_arg_set (id, StgRhsClosure cc bi _ upd args body)
365 = ASSERT( n_args > 0 )
366 -- Construct the rhs of the supercombinator, and its Id
367 -- this trace blackholes sometimes, don't use it
368 -- trace ("LL " ++ show (length (uniqSetToList extra_arg_set))) (
369 newSupercombinator sc_ty arity `thenLM` \ sc_id ->
371 returnLM ((sc_id, extra_args), (sc_id, sc_rhs))
375 extra_args = uniqSetToList extra_arg_set
376 arity = n_args + length extra_args
378 -- Construct the supercombinator type
379 type_of_original_id = idType id
380 extra_arg_tys = map idType extra_args
381 (tyvars, rest) = splitForalls type_of_original_id
382 sc_ty = mkForallTy tyvars (glueTyArgs extra_arg_tys rest)
384 sc_rhs = StgRhsClosure cc bi [] upd (extra_args ++ args) body
388 %************************************************************************
390 \subsection[Lift-monad]{The LiftM monad}
392 %************************************************************************
394 The monad is used only to distribute global stuff, and the unique supply.
397 type LiftM a = LiftFlags
399 -> (IdEnv -- Domain = candidates for lifting
400 (Id, -- The supercombinator
401 [Id]) -- Args to apply it to
406 type LiftFlags = Maybe Int -- No of fvs reqd to float recursive
407 -- binding; Nothing == infinity
410 runLM :: LiftFlags -> UniqSupply -> LiftM a -> a
411 runLM flags us m = m flags us nullIdEnv
413 thenLM :: LiftM a -> (a -> LiftM b) -> LiftM b
414 thenLM m k ci us idenv
415 = k (m ci us1 idenv) ci us2 idenv
417 (us1, us2) = splitUniqSupply us
419 returnLM :: a -> LiftM a
420 returnLM a ci us idenv = a
422 fixLM :: (a -> LiftM a) -> LiftM a
423 fixLM k ci us idenv = r
427 mapLM :: (a -> LiftM b) -> [a] -> LiftM [b]
428 mapLM f [] = returnLM []
429 mapLM f (a:as) = f a `thenLM` \ r ->
430 mapLM f as `thenLM` \ rs ->
433 mapAndUnzipLM :: (a -> LiftM (b,c)) -> [a] -> LiftM ([b],[c])
434 mapAndUnzipLM f [] = returnLM ([],[])
435 mapAndUnzipLM f (a:as) = f a `thenLM` \ (b,c) ->
436 mapAndUnzipLM f as `thenLM` \ (bs,cs) ->
437 returnLM (b:bs, c:cs)
441 newSupercombinator :: Type
445 newSupercombinator ty arity ci us idenv
446 = (mkSysLocal SLIT("sc") uniq ty mkUnknownSrcLoc) -- ToDo: improve location
448 -- ToDo: rm the addIdArity? Just let subsequent stg-saturation pass do it?
452 lookup :: Id -> LiftM (Id,[Id])
454 = case lookupIdEnv idenv v of
455 Just result -> result
458 addScInlines :: [Id] -> [(Id,[Id])] -> LiftM a -> LiftM a
459 addScInlines ids values m ci us idenv
462 idenv' = growIdEnvList idenv (ids `zip_lazy` values)
464 -- zip_lazy zips two things together but matches lazily on the
465 -- second argument. This is important, because the ids are know here,
466 -- but the things they are bound to are decided only later
468 zip_lazy (x:xs) ~(y:ys) = (x,y) : zip_lazy xs ys
471 -- The free vars reported by the free-var analyser will include
472 -- some ids, f, which are to be replaced by ($f a b c), where $f
473 -- is the supercombinator. Hence instead of f being a free var,
478 -- f a = ...y1..y2.....
485 -- Here the free vars of g are {f,z}; but f will be lambda-lifted
486 -- with free vars {y1,y2}, so the "real~ free vars of g are {y1,y2,z}.
488 getFinalFreeVars :: IdSet -> LiftM IdSet
490 getFinalFreeVars free_vars ci us idenv
491 = unionManyUniqSets (map munge_it (uniqSetToList free_vars))
493 munge_it :: Id -> IdSet -- Takes a free var and maps it to the "real"
495 munge_it id = case lookupIdEnv idenv id of
496 Just (_, args) -> mkUniqSet args
497 Nothing -> singletonUniqSet id
502 %************************************************************************
504 \subsection[Lift-info]{The LiftInfo type}
506 %************************************************************************
509 type LiftInfo = Bag StgBinding -- Float to top
511 emptyLiftInfo = emptyBag
513 unionLiftInfo :: LiftInfo -> LiftInfo -> LiftInfo
514 unionLiftInfo binds1 binds2 = binds1 `unionBags` binds2
516 unionLiftInfos :: [LiftInfo] -> LiftInfo
517 unionLiftInfos infos = foldr unionLiftInfo emptyLiftInfo infos
519 mkScInfo :: StgBinding -> LiftInfo
520 mkScInfo bind = unitBag bind
522 nonRecScBind :: LiftInfo -- From body of supercombinator
523 -> (Id, StgRhs) -- Supercombinator and its rhs
525 nonRecScBind binds (sc_id,sc_rhs) = binds `snocBag` (StgNonRec sc_id sc_rhs)
528 -- In the recursive case, all the SCs from the RHSs of the recursive group
529 -- are dealing with might potentially mention the new, recursive SCs.
530 -- So we flatten the whole lot into a single recursive group.
532 recScBind :: LiftInfo -- From body of supercombinator
533 -> [(Id,StgRhs)] -- Supercombinator rhs
536 recScBind binds pairs = unitBag (co_rec_ify (StgRec pairs : bagToList binds))
538 co_rec_ify :: [StgBinding] -> StgBinding
539 co_rec_ify binds = StgRec (concat (map f binds))
541 f (StgNonRec id rhs) = [(id,rhs)]
542 f (StgRec pairs) = pairs
545 getScBinds :: LiftInfo -> [StgBinding]
546 getScBinds binds = bagToList binds
548 looksLikeSATRhs [(f,StgRhsClosure _ _ _ _ ls _)] (StgApp (StgVarArg f') args _)
549 = (f == f') && (length args == length ls)
550 looksLikeSATRhs _ _ = False