1 \section{Update Avoidance Analyser}
3 (c) Simon Marlow, Andre Santos 1992-1993
4 (c) The AQUA Project, Glasgow University, 1995-1998
6 %-----------------------------------------------------------------------------
7 \subsection{Module Interface}
11 module UpdAnal ( updateAnalyse ) where
13 #include "HsVersions.h"
15 import Prelude hiding ( lookup )
20 import Id ( mkSysLocal,
21 idUpdateInfo, setIdUpdateInfo, idType,
25 import IdInfo ( UpdateInfo, UpdateSpec, mkUpdateInfo, updateInfoMaybe )
26 import Name ( isLocallyDefined )
27 import Type ( splitFunTys, splitSigmaTy )
28 import Unique ( getBuiltinUniques )
29 import Panic ( panic )
33 %-----------------------------------------------------------------------------
34 \subsection{Reverse application}
36 This is used instead of lazy pattern bindings to avoid space leaks.
43 %-----------------------------------------------------------------------------
46 List of closure references
50 x `notInRefs` y = not (x `elemVarSet` y)
53 A closure value: environment of closures that are evaluated on entry,
54 a list of closures that are referenced from the result, and an
55 abstract value for the evaluated closure.
57 An IdEnv is used for the reference counts, as these environments are
58 combined often. A generic environment is used for the main environment
59 mapping closure names to values; as a common operation is extension of
60 this environment, this representation should be efficient.
63 -- partain: funny synonyms to cope w/ the fact
64 -- that IdEnvs know longer know what their keys are
65 -- (94/05) ToDo: improve
66 type IdEnvInt = IdEnv (Id, Int)
67 type IdEnvClosure = IdEnv (Id, Closure)
69 -- backward-compat functions
70 null_IdEnv :: IdEnv (Id, a)
71 null_IdEnv = emptyVarEnv
73 unit_IdEnv :: Id -> a -> IdEnv (Id, a)
74 unit_IdEnv k v = unitVarEnv k (k, v)
76 mk_IdEnv :: [(Id, a)] -> IdEnv (Id, a)
77 mk_IdEnv pairs = mkVarEnv [ (k, (k,v)) | (k,v) <- pairs ]
79 grow_IdEnv :: IdEnv (Id, a) -> IdEnv (Id, a) -> IdEnv (Id, a)
80 grow_IdEnv env1 env2 = plusVarEnv env1 env2
82 addOneTo_IdEnv :: IdEnv (Id, a) -> Id -> a -> IdEnv (Id, a)
83 addOneTo_IdEnv env k v = extendVarEnv env k (k, v)
85 combine_IdEnvs :: (a->a->a) -> IdEnv (Id, a) -> IdEnv (Id, a) -> IdEnv (Id, a)
86 combine_IdEnvs combiner env1 env2 = plusVarEnv_C new_combiner env1 env2
88 new_combiner (id, x) (_, y) = (id, combiner x y)
90 dom_IdEnv :: IdEnv (Id, a) -> Refs
91 dom_IdEnv env = mkVarSet [ i | (i,_) <- rngVarEnv env ]
93 lookup_IdEnv :: IdEnv (Id, a) -> Id -> Maybe a
94 lookup_IdEnv env key = case lookupVarEnv env key of
97 -- end backward compat stuff
99 type Closure = (IdEnvInt, Refs, AbFun)
101 type AbVal = IdEnvClosure -> Closure
102 newtype AbFun = Fun (Closure -> Closure)
104 -- partain: speeding-up stuff
106 type CaseBoundVars = IdSet
107 noCaseBound = emptyVarSet
108 isCaseBound = elemVarSet
109 x `notCaseBound` y = not (isCaseBound x y)
110 moreCaseBound :: CaseBoundVars -> [Id] -> CaseBoundVars
111 moreCaseBound old new = old `unionVarSet` mkVarSet new
116 %----------------------------------------------------------------------------
117 \subsection{Environment lookup}
119 If the requested value is not in the environment, we return an unknown
120 value. Lookup is designed to be partially applied to a variable, and
121 repeatedly applied to different environments after that.
126 = \p -> case lookup_IdEnv p v of
128 Nothing -> unknownClosure
131 = const (case updateInfoMaybe (idUpdateInfo v) of
132 Nothing -> unknownClosure
133 Just spec -> convertUpdateSpec spec)
136 %-----------------------------------------------------------------------------
137 Represent a list of references as an ordered list.
140 mkRefs :: [Id] -> Refs
146 elemRefs = elemVarSet
148 merge :: [Refs] -> Refs
149 merge xs = foldr merge2 emptyVarSet xs
151 merge2 :: Refs -> Refs -> Refs
155 %-----------------------------------------------------------------------------
156 \subsection{Some non-interesting values}
158 bottom will be used for abstract values that are not functions.
159 Hopefully its value will never be required!
163 bottom = panic "Internal: (Update Analyser) bottom"
166 noClosure is a value that is definitely not a function (i.e. primitive
167 values and constructor applications). unknownClosure is a value about
168 which we have no information at all. This should occur rarely, but
169 could happen when an id is imported and the exporting module was not
170 compiled with the update analyser.
173 noClosure, unknownClosure :: Closure
174 noClosure = (null_IdEnv, noRefs, bottom)
175 unknownClosure = (null_IdEnv, noRefs, dont_know noRefs)
178 dont_know is a black hole: it is something we know nothing about.
179 Applying dont_know to anything will generate a new dont_know that simply
180 contains more buried references.
183 dont_know :: Refs -> AbFun
185 = Fun (\(c,b,f) -> let b'' = dom_IdEnv c `merge2` b `merge2` b'
186 in (null_IdEnv, b'', dont_know b''))
189 -----------------------------------------------------------------------------
192 getrefs :: IdEnvClosure -> [AbVal] -> Refs -> Refs
193 getrefs p vs rest = foldr merge2 rest (getrefs' (map ($ p) vs))
196 getrefs' ((c,b,_):rs) = dom_IdEnv c : b : getrefs' rs
199 -----------------------------------------------------------------------------
201 udData is used when we are putting a list of closure references into a
202 data structure, or something else that we know nothing about.
205 udData :: [StgArg] -> CaseBoundVars -> AbVal
207 = \p -> (null_IdEnv, getrefs p local_ids noRefs, bottom)
208 where local_ids = [ lookup v | StgVarArg v <- vs, v `notCaseBound` cvs ]
211 %-----------------------------------------------------------------------------
212 \subsection{Analysing an atom}
215 udVar :: CaseBoundVars -> Id -> AbVal
216 udVar cvs v | v `isCaseBound` cvs = const unknownClosure
217 | otherwise = lookup v
219 udAtom :: CaseBoundVars -> StgArg -> AbVal
220 udAtom cvs (StgVarArg v) = udVar cvs v
221 udAtom cvs _ = const noClosure
224 %-----------------------------------------------------------------------------
225 \subsection{Analysing an STG expression}
228 ud :: StgExpr -- Expression to be analysed
229 -> CaseBoundVars -- List of case-bound vars
230 -> IdEnvClosure -- Current environment
231 -> (StgExpr, AbVal) -- (New expression, abstract value)
233 ud e@(StgLit _) cvs p = (e, udData [] cvs)
234 ud e@(StgConApp _ vs) cvs p = (e, udData vs cvs)
235 ud e@(StgPrimApp _ vs _) cvs p = (e, udData vs cvs)
236 ud e@(StgSCC lab a) cvs p = ud a cvs p =: \(a', abval_a) ->
237 (StgSCC lab a', abval_a)
240 Here is application. The first thing to do is analyse the head, and
241 get an abstract function. Multiple applications are performed by using
242 a foldl with the function doApp. Closures are actually passed to the
243 abstract function iff the atom is a local variable.
245 I've left the type signature for doApp in to make things a bit clearer.
248 ud e@(StgApp a atoms) cvs p
251 abval_atoms = map (udAtom cvs) atoms
252 abval_a = udVar cvs a
254 let doApp :: Closure -> AbVal -> Closure
255 doApp (c, b, Fun f) abval_atom =
256 abval_atom p =: \e@(_,_,_) ->
257 f e =: \(c', b', f') ->
258 (combine_IdEnvs (+) c' c, b', f')
259 in foldl doApp (abval_a p) abval_atoms
261 ud (StgCase expr lve lva bndr srt alts) cvs p
262 = ud expr cvs p =: \(expr', abval_selector) ->
263 udAlt alts p =: \(alts', abval_alts) ->
266 abval_selector p =: \(c, b, abfun_selector) ->
267 abval_alts p =: \(cs, bs, abfun_alts) ->
268 let bs' = b `merge2` bs in
269 (combine_IdEnvs (+) c cs, bs', dont_know bs')
271 (StgCase expr' lve lva bndr srt alts', abval_case)
274 alts_cvs = moreCaseBound cvs [bndr]
278 -> (StgCaseAlts, AbVal)
280 udAlt (StgAlgAlts ty [alt] StgNoDefault) p
281 = udAlgAlt p alt =: \(alt', abval) ->
282 (StgAlgAlts ty [alt'] StgNoDefault, abval)
283 udAlt (StgAlgAlts ty [] def) p
284 = udDef def p =: \(def', abval) ->
285 (StgAlgAlts ty [] def', abval)
286 udAlt (StgAlgAlts ty alts def) p
287 = udManyAlts alts def udAlgAlt (StgAlgAlts ty) p
288 udAlt (StgPrimAlts ty [alt] StgNoDefault) p
289 = udPrimAlt p alt =: \(alt', abval) ->
290 (StgPrimAlts ty [alt'] StgNoDefault, abval)
291 udAlt (StgPrimAlts ty [] def) p
292 = udDef def p =: \(def', abval) ->
293 (StgPrimAlts ty [] def', abval)
294 udAlt (StgPrimAlts ty alts def) p
295 = udManyAlts alts def udPrimAlt (StgPrimAlts ty) p
298 = ud e alts_cvs p =: \(e', v) -> ((l, e'), v)
300 udAlgAlt p (id, vs, use_mask, e)
301 = ud e (moreCaseBound alts_cvs vs) p
302 =: \(e', v) -> ((id, vs, use_mask, e'), v)
304 udDef :: StgCaseDefault
306 -> (StgCaseDefault, AbVal)
309 = (StgNoDefault, \p -> (null_IdEnv, noRefs, dont_know noRefs))
310 udDef (StgBindDefault expr) p
311 = ud expr alts_cvs p =: \(expr', abval) ->
312 (StgBindDefault expr', abval)
314 udManyAlts alts def udalt stgalts p
315 = udDef def p =: \(def', abval_def) ->
316 unzip (map (udalt p) alts) =: \(alts', abvals_alts) ->
319 abval_def p =: \(cd, bd, _) ->
320 unzip3 (map ($ p) abvals_alts) =: \(cs, bs, _) ->
321 let bs' = merge (bd:bs) in
322 (foldr (combine_IdEnvs max) cd cs, bs', dont_know bs')
323 in (stgalts alts' def', abval_alts)
326 The heart of the analysis: here we decide whether to make a specific
327 closure updatable or not, based on the results of analysing the body.
330 ud (StgLet binds body) cvs p
331 = udBinding binds cvs p =: \(binds', vs, abval1, abval2) ->
332 abval1 p =: \(cs, p') ->
333 grow_IdEnv p p' =: \p ->
334 ud body cvs p =: \(body', abval_body) ->
335 abval_body p =: \(c, b, abfun) ->
336 tag b (combine_IdEnvs (+) cs c) binds' =: \tagged_binds ->
339 = abval2 p =: \(c1, p') ->
340 abval_body (grow_IdEnv p p') =: \(c2, b, abfun) ->
341 (combine_IdEnvs (+) c1 c2, b, abfun)
343 (StgLet tagged_binds body', abval)
346 %-----------------------------------------------------------------------------
347 \subsection{Analysing bindings}
349 For recursive sets of bindings we perform one iteration of a fixed
350 point algorithm, using (dont_know fv) as a safe approximation to the
351 real fixed point, where fv are the (mappings in the environment of
352 the) free variables of the function.
354 We'll return two new environments, one with the new closures in and
355 one without. There's no point in carrying around closures when their
356 respective bindings have already been analysed.
358 We don't need to find anything out about closures with arguments,
359 constructor closures etc.
362 udBinding :: StgBinding
367 IdEnvClosure -> (IdEnvInt, IdEnvClosure),
368 IdEnvClosure -> (IdEnvInt, IdEnvClosure))
370 udBinding (StgNonRec v rhs) cvs p
371 = udRhs rhs cvs p =: \(rhs', abval) ->
372 abval p =: \(c, b, abfun) ->
375 abval p =: \(c, b, abfun) ->
376 (c, unit_IdEnv v (a, b, abfun))
378 StgRhsClosure _ _ _ _ Updatable [] _ -> unit_IdEnv v 1
380 in (StgNonRec v rhs', [v], abval_rhs a, abval_rhs null_IdEnv)
382 udBinding (StgRec ve) cvs p
383 = (StgRec ve', [], abval_rhs, abval_rhs)
385 (vs, ve', abvals) = unzip3 (map udBind ve)
386 fv = (map lookup . filter (`notCaseBound` cvs) . concat . map collectfv) ve
390 p' = grow_IdEnv (mk_IdEnv (vs `zip` (repeat closure))) p
391 closure = (null_IdEnv, fv', dont_know fv')
392 fv' = getrefs p fv vs'
393 (cs, ps) = unzip (doRec vs abvals)
396 doRec (v:vs) (abval:as)
397 = abval p' =: \(c,b,abfun) ->
398 (c, (v,(null_IdEnv, b, abfun))) : doRec vs as
401 (foldr (combine_IdEnvs (+)) null_IdEnv cs, mk_IdEnv ps)
404 = udRhs rhs cvs p =: \(rhs', abval) ->
407 collectfv (_, StgRhsClosure _ _ _ fv _ _ _) = fv
408 collectfv (_, StgRhsCon _ con args) = [ v | StgVarArg v <- args ]
411 %-----------------------------------------------------------------------------
412 \subsection{Analysing Right-Hand Sides}
415 udRhs e@(StgRhsCon _ _ vs) cvs p = (e, udData vs cvs)
417 udRhs (StgRhsClosure cc bi srt fv u [] body) cvs p
418 = ud body cvs p =: \(body', abval_body) ->
419 (StgRhsClosure cc bi srt fv u [] body', abval_body)
422 Here is the code for closures with arguments. A closure has a number
423 of arguments, which correspond to a set of nested lambda expressions.
424 We build up the analysis using foldr with the function doLam to
425 analyse each lambda expression.
428 udRhs (StgRhsClosure cc bi srt fv u args body) cvs p
429 = ud body cvs p =: \(body', abval_body) ->
431 fv' = map lookup (filter (`notCaseBound` cvs) fv)
433 foldr doLam (\b -> abval_body) args (getrefs p fv' noRefs) p
435 (StgRhsClosure cc bi srt fv u args body', abval_rhs)
438 doLam :: Id -> (Refs -> AbVal) -> Refs -> AbVal
442 let b'' = dom_IdEnv c' `merge2` b' `merge2` b in
443 f b'' (addOneTo_IdEnv p i x)))
446 %-----------------------------------------------------------------------------
447 \subsection{Adjusting Update flags}
449 The closure is tagged single entry iff it is used at most once, it is
450 not referenced from inside a data structure or function, and it has no
451 arguments (closures with arguments are re-entrant).
454 tag :: Refs -> IdEnvInt -> StgBinding -> StgBinding
456 tag b c r@(StgNonRec v (StgRhsClosure cc bi srt fv Updatable [] body))
457 = if (v `notInRefs` b) && (lookupc c v <= 1)
458 then -- trace "One!" (
459 StgNonRec v (StgRhsClosure cc bi srt fv SingleEntry [] body)
462 tag b c other = other
464 lookupc c v = case lookup_IdEnv c v of
469 %-----------------------------------------------------------------------------
470 \subsection{Top Level analysis}
472 Should we tag top level closures? This could have good implications
473 for CAFs (i.e. they could be made non-updateable if only used once,
474 thus preventing a space leak).
477 updateAnalyse :: [StgBinding] -> [StgBinding] {- Exported -}
479 = udProgram bs null_IdEnv
481 udProgram :: [StgBinding] -> IdEnvClosure -> [StgBinding]
484 = udBinding d noCaseBound p =: \(d', vs, _, abval_bind) ->
485 abval_bind p =: \(_, p') ->
486 grow_IdEnv p p' =: \p'' ->
487 attachUpdateInfoToBinds d' p'' =: \d'' ->
488 d'' : udProgram ds p''
491 %-----------------------------------------------------------------------------
492 \subsection{Exporting Update Information}
494 Convert the exported representation of a function's update function
495 into a real Closure value.
498 convertUpdateSpec :: UpdateSpec -> Closure
499 convertUpdateSpec = mkClosure null_IdEnv noRefs noRefs
501 mkClosure :: IdEnvInt -> Refs -> Refs -> UpdateSpec -> Closure
503 mkClosure c b b' [] = (c, b', dont_know b')
504 mkClosure c b b' (0 : ns) = (null_IdEnv, b, Fun (\ _ -> mkClosure c b b' ns))
505 mkClosure c b b' (1 : ns) = (null_IdEnv, b, Fun (\ (c',b'',f) ->
507 (combine_IdEnvs (+) c c')
508 (dom_IdEnv c' `merge2` b'' `merge2` b)
511 mkClosure c b b' (n : ns) = (null_IdEnv, b, Fun (\ (c',b'',f) ->
513 (dom_IdEnv c' `merge2` b'' `merge2` b)
514 (dom_IdEnv c' `merge2` b'' `merge2` b')
518 Convert a Closure into a representation that can be placed in a .hi file.
521 mkUpdateSpec :: Id -> Closure -> UpdateSpec
522 mkUpdateSpec v f = {- removeSuperfluous2s -} (map countUses ids)
524 (c,b,_) = foldl doApp f ids
525 ids = map mkid (getBuiltinUniques arity)
526 mkid u = mkSysLocal SLIT("upd") u noType
527 countUses u = if u `elemRefs` b then 2 else min (lookupc c u) 2
528 noType = panic "UpdAnal: no type!"
531 = f (unit_IdEnv i 1, noRefs, dont_know noRefs) =: \(c',b',f') ->
532 (combine_IdEnvs (+) c' c, b', f')
534 (_,dict_tys,tau_ty) = (splitSigmaTy . idType) v
535 (reg_arg_tys, _) = splitFunTys tau_ty
536 arity = length dict_tys + length reg_arg_tys
539 removeSuperfluous2s = reverse . dropWhile (> 1) . reverse
541 %-----------------------------------------------------------------------------
542 \subsection{Attaching the update information to top-level bindings}
544 This is so that the information can later be retrieved for printing
545 out in the .hi file. This is not an ideal solution, however it will
549 attachUpdateInfoToBinds b p
551 StgNonRec v rhs -> StgNonRec (attachOne v) rhs
552 StgRec bs -> StgRec [ (attachOne v, rhs) | (v, rhs) <- bs ]
555 | externallyVisibleId v
556 = let c = lookup v p in
558 (mkUpdateInfo (mkUpdateSpec v c))
562 %-----------------------------------------------------------------------------