1 \section{Update Avoidance Analyser}
3 (c) Simon Marlow, Andre Santos 1992-1993
4 (c) The AQUA Project, Glasgow University, 1995-1996
6 %-----------------------------------------------------------------------------
7 \subsection{Module Interface}
11 module UpdAnal ( updateAnalyse ) where
13 #include "HsVersions.h"
15 import Prelude hiding ( lookup )
18 import MkId ( mkSysLocal )
19 import Id ( IdEnv, growIdEnv, addOneToIdEnv, combineIdEnvs, nullIdEnv,
20 unitIdEnv, mkIdEnv, rngIdEnv, lookupIdEnv,
22 getIdUpdateInfo, addIdUpdateInfo, idType,
26 import IdInfo ( UpdateInfo, UpdateSpec, mkUpdateInfo, updateInfoMaybe )
27 import Name ( isLocallyDefined )
28 import Type ( splitFunTys, splitSigmaTy )
30 import Unique ( getBuiltinUniques )
31 import SrcLoc ( noSrcLoc )
36 %-----------------------------------------------------------------------------
37 \subsection{Reverse application}
39 This is used instead of lazy pattern bindings to avoid space leaks.
46 %-----------------------------------------------------------------------------
49 List of closure references
53 x `notInRefs` y = not (x `elementOfUniqSet` y)
56 A closure value: environment of closures that are evaluated on entry,
57 a list of closures that are referenced from the result, and an
58 abstract value for the evaluated closure.
60 An IdEnv is used for the reference counts, as these environments are
61 combined often. A generic environment is used for the main environment
62 mapping closure names to values; as a common operation is extension of
63 this environment, this representation should be efficient.
66 -- partain: funny synonyms to cope w/ the fact
67 -- that IdEnvs know longer know what their keys are
68 -- (94/05) ToDo: improve
69 type IdEnvInt = IdEnv (Id, Int)
70 type IdEnvClosure = IdEnv (Id, Closure)
72 -- backward-compat functions
73 null_IdEnv :: IdEnv (Id, a)
74 null_IdEnv = nullIdEnv
76 unit_IdEnv :: Id -> a -> IdEnv (Id, a)
77 unit_IdEnv k v = unitIdEnv k (k, v)
79 mk_IdEnv :: [(Id, a)] -> IdEnv (Id, a)
80 mk_IdEnv pairs = mkIdEnv [ (k, (k,v)) | (k,v) <- pairs ]
82 grow_IdEnv :: IdEnv (Id, a) -> IdEnv (Id, a) -> IdEnv (Id, a)
83 grow_IdEnv env1 env2 = growIdEnv env1 env2
85 addOneTo_IdEnv :: IdEnv (Id, a) -> Id -> a -> IdEnv (Id, a)
86 addOneTo_IdEnv env k v = addOneToIdEnv env k (k, v)
88 combine_IdEnvs :: (a->a->a) -> IdEnv (Id, a) -> IdEnv (Id, a) -> IdEnv (Id, a)
89 combine_IdEnvs combiner env1 env2 = combineIdEnvs new_combiner env1 env2
91 new_combiner (id, x) (_, y) = (id, combiner x y)
93 dom_IdEnv :: IdEnv (Id, a) -> Refs
94 dom_IdEnv env = mkUniqSet [ i | (i,_) <- rngIdEnv env ]
96 lookup_IdEnv :: IdEnv (Id, a) -> Id -> Maybe a
97 lookup_IdEnv env key = case lookupIdEnv env key of
100 -- end backward compat stuff
102 type Closure = (IdEnvInt, Refs, AbFun)
104 type AbVal = IdEnvClosure -> Closure
105 newtype AbFun = Fun (Closure -> Closure)
107 -- partain: speeding-up stuff
109 type CaseBoundVars = IdSet
110 noCaseBound = emptyUniqSet
111 isCaseBound = elementOfUniqSet
112 x `notCaseBound` y = not (isCaseBound x y)
113 moreCaseBound :: CaseBoundVars -> [Id] -> CaseBoundVars
114 moreCaseBound old new = old `unionUniqSets` mkUniqSet new
119 %----------------------------------------------------------------------------
120 \subsection{Environment lookup}
122 If the requested value is not in the environment, we return an unknown
123 value. Lookup is designed to be partially applied to a variable, and
124 repeatedly applied to different environments after that.
129 = \p -> case lookup_IdEnv p v of
131 Nothing -> unknownClosure
134 = const (case updateInfoMaybe (getIdUpdateInfo v) of
135 Nothing -> unknownClosure
136 Just spec -> convertUpdateSpec spec)
139 %-----------------------------------------------------------------------------
140 Represent a list of references as an ordered list.
143 mkRefs :: [Id] -> Refs
147 noRefs = emptyUniqSet
149 elemRefs = elementOfUniqSet
151 merge :: [Refs] -> Refs
152 merge xs = foldr merge2 emptyUniqSet xs
154 merge2 :: Refs -> Refs -> Refs
155 merge2 = unionUniqSets
158 %-----------------------------------------------------------------------------
159 \subsection{Some non-interesting values}
161 bottom will be used for abstract values that are not functions.
162 Hopefully its value will never be required!
166 bottom = panic "Internal: (Update Analyser) bottom"
169 noClosure is a value that is definitely not a function (i.e. primitive
170 values and constructor applications). unknownClosure is a value about
171 which we have no information at all. This should occur rarely, but
172 could happen when an id is imported and the exporting module was not
173 compiled with the update analyser.
176 noClosure, unknownClosure :: Closure
177 noClosure = (null_IdEnv, noRefs, bottom)
178 unknownClosure = (null_IdEnv, noRefs, dont_know noRefs)
181 dont_know is a black hole: it is something we know nothing about.
182 Applying dont_know to anything will generate a new dont_know that simply
183 contains more buried references.
186 dont_know :: Refs -> AbFun
188 = Fun (\(c,b,f) -> let b'' = dom_IdEnv c `merge2` b `merge2` b'
189 in (null_IdEnv, b'', dont_know b''))
192 -----------------------------------------------------------------------------
195 getrefs :: IdEnvClosure -> [AbVal] -> Refs -> Refs
196 getrefs p vs rest = foldr merge2 rest (getrefs' (map ($ p) vs))
199 getrefs' ((c,b,_):rs) = dom_IdEnv c : b : getrefs' rs
202 -----------------------------------------------------------------------------
204 udData is used when we are putting a list of closure references into a
205 data structure, or something else that we know nothing about.
208 udData :: [StgArg] -> CaseBoundVars -> AbVal
210 = \p -> (null_IdEnv, getrefs p local_ids noRefs, bottom)
211 where local_ids = [ lookup v | (StgVarArg v) <- vs, v `notCaseBound` cvs ]
214 %-----------------------------------------------------------------------------
215 \subsection{Analysing an atom}
218 udAtom :: CaseBoundVars -> StgArg -> AbVal
219 udAtom cvs (StgVarArg v)
220 | v `isCaseBound` cvs = const unknownClosure
221 | otherwise = lookup v
223 udAtom cvs _ = const noClosure
226 %-----------------------------------------------------------------------------
227 \subsection{Analysing an STG expression}
230 ud :: StgExpr -- Expression to be analysed
231 -> CaseBoundVars -- List of case-bound vars
232 -> IdEnvClosure -- Current environment
233 -> (StgExpr, AbVal) -- (New expression, abstract value)
235 ud e@(StgPrim _ vs _) cvs p = (e, udData vs cvs)
236 ud e@(StgCon _ vs _) cvs p = (e, udData vs cvs)
237 ud e@(StgSCC ty lab a) cvs p = ud a cvs p =: \(a', abval_a) ->
238 (StgSCC ty lab a', abval_a)
241 Here is application. The first thing to do is analyse the head, and
242 get an abstract function. Multiple applications are performed by using
243 a foldl with the function doApp. Closures are actually passed to the
244 abstract function iff the atom is a local variable.
246 I've left the type signature for doApp in to make things a bit clearer.
249 ud e@(StgApp a atoms lvs) cvs p
252 abval_atoms = map (udAtom cvs) atoms
253 abval_a = udAtom cvs a
255 let doApp :: Closure -> AbVal -> Closure
256 doApp (c, b, Fun f) abval_atom =
257 abval_atom p =: \e@(_,_,_) ->
258 f e =: \(c', b', f') ->
259 (combine_IdEnvs (+) c' c, b', f')
260 in foldl doApp (abval_a p) abval_atoms
262 ud (StgCase expr lve lva uniq alts) cvs p
263 = ud expr cvs p =: \(expr', abval_selector) ->
264 udAlt alts p =: \(alts', abval_alts) ->
267 abval_selector p =: \(c, b, abfun_selector) ->
268 abval_alts p =: \(cs, bs, abfun_alts) ->
269 let bs' = b `merge2` bs in
270 (combine_IdEnvs (+) c cs, bs', dont_know bs')
272 (StgCase expr' lve lva uniq alts', abval_case)
277 -> (StgCaseAlts, AbVal)
279 udAlt (StgAlgAlts ty [alt] StgNoDefault) p
280 = udAlgAlt p alt =: \(alt', abval) ->
281 (StgAlgAlts ty [alt'] StgNoDefault, abval)
282 udAlt (StgAlgAlts ty [] def) p
283 = udDef def p =: \(def', abval) ->
284 (StgAlgAlts ty [] def', abval)
285 udAlt (StgAlgAlts ty alts def) p
286 = udManyAlts alts def udAlgAlt (StgAlgAlts ty) p
287 udAlt (StgPrimAlts ty [alt] StgNoDefault) p
288 = udPrimAlt p alt =: \(alt', abval) ->
289 (StgPrimAlts ty [alt'] StgNoDefault, abval)
290 udAlt (StgPrimAlts ty [] def) p
291 = udDef def p =: \(def', abval) ->
292 (StgPrimAlts ty [] def', abval)
293 udAlt (StgPrimAlts ty alts def) p
294 = udManyAlts alts def udPrimAlt (StgPrimAlts ty) p
297 = ud e cvs p =: \(e', v) -> ((l, e'), v)
299 udAlgAlt p (id, vs, use_mask, e)
300 = ud e (moreCaseBound cvs vs) p =: \(e', v) -> ((id, vs, use_mask, e'), v)
302 udDef :: StgCaseDefault
304 -> (StgCaseDefault, AbVal)
307 = (StgNoDefault, \p -> (null_IdEnv, noRefs, dont_know noRefs))
308 udDef (StgBindDefault v is_used expr) p
309 = ud expr (moreCaseBound cvs [v]) p =: \(expr', abval) ->
310 (StgBindDefault v is_used expr', abval)
312 udManyAlts alts def udalt stgalts p
313 = udDef def p =: \(def', abval_def) ->
314 unzip (map (udalt p) alts) =: \(alts', abvals_alts) ->
317 abval_def p =: \(cd, bd, _) ->
318 unzip3 (map ($ p) abvals_alts) =: \(cs, bs, _) ->
319 let bs' = merge (bd:bs) in
320 (foldr (combine_IdEnvs max) cd cs, bs', dont_know bs')
321 in (stgalts alts' def', abval_alts)
324 The heart of the analysis: here we decide whether to make a specific
325 closure updatable or not, based on the results of analysing the body.
328 ud (StgLet binds body) cvs p
329 = udBinding binds cvs p =: \(binds', vs, abval1, abval2) ->
330 abval1 p =: \(cs, p') ->
331 grow_IdEnv p p' =: \p ->
332 ud body cvs p =: \(body', abval_body) ->
333 abval_body p =: \(c, b, abfun) ->
334 tag b (combine_IdEnvs (+) cs c) binds' =: \tagged_binds ->
337 = abval2 p =: \(c1, p') ->
338 abval_body (grow_IdEnv p p') =: \(c2, b, abfun) ->
339 (combine_IdEnvs (+) c1 c2, b, abfun)
341 (StgLet tagged_binds body', abval)
344 %-----------------------------------------------------------------------------
345 \subsection{Analysing bindings}
347 For recursive sets of bindings we perform one iteration of a fixed
348 point algorithm, using (dont_know fv) as a safe approximation to the
349 real fixed point, where fv are the (mappings in the environment of
350 the) free variables of the function.
352 We'll return two new environments, one with the new closures in and
353 one without. There's no point in carrying around closures when their
354 respective bindings have already been analysed.
356 We don't need to find anything out about closures with arguments,
357 constructor closures etc.
360 udBinding :: StgBinding
365 IdEnvClosure -> (IdEnvInt, IdEnvClosure),
366 IdEnvClosure -> (IdEnvInt, IdEnvClosure))
368 udBinding (StgNonRec v rhs) cvs p
369 = udRhs rhs cvs p =: \(rhs', abval) ->
370 abval p =: \(c, b, abfun) ->
373 abval p =: \(c, b, abfun) ->
374 (c, unit_IdEnv v (a, b, abfun))
376 StgRhsClosure _ _ _ Updatable [] _ -> unit_IdEnv v 1
378 in (StgNonRec v rhs', [v], abval_rhs a, abval_rhs null_IdEnv)
380 udBinding (StgRec ve) cvs p
381 = (StgRec ve', [], abval_rhs, abval_rhs)
383 (vs, ve', abvals) = unzip3 (map udBind ve)
384 fv = (map lookup . filter (`notCaseBound` cvs) . concat . map collectfv) ve
388 p' = grow_IdEnv (mk_IdEnv (vs `zip` (repeat closure))) p
389 closure = (null_IdEnv, fv', dont_know fv')
390 fv' = getrefs p fv vs'
391 (cs, ps) = unzip (doRec vs abvals)
394 doRec (v:vs) (abval:as)
395 = abval p' =: \(c,b,abfun) ->
396 (c, (v,(null_IdEnv, b, abfun))) : doRec vs as
399 (foldr (combine_IdEnvs (+)) null_IdEnv cs, mk_IdEnv ps)
402 = udRhs rhs cvs p =: \(rhs', abval) ->
405 collectfv (_, StgRhsClosure _ _ fv _ _ _) = fv
406 collectfv (_, StgRhsCon _ con args) = [ v | (StgVarArg v) <- args ]
409 %-----------------------------------------------------------------------------
410 \subsection{Analysing Right-Hand Sides}
413 udRhs e@(StgRhsCon _ _ vs) cvs p = (e, udData vs cvs)
415 udRhs (StgRhsClosure cc bi fv u [] body) cvs p
416 = ud body cvs p =: \(body', abval_body) ->
417 (StgRhsClosure cc bi fv u [] body', abval_body)
420 Here is the code for closures with arguments. A closure has a number
421 of arguments, which correspond to a set of nested lambda expressions.
422 We build up the analysis using foldr with the function doLam to
423 analyse each lambda expression.
426 udRhs (StgRhsClosure cc bi fv u args body) cvs p
427 = ud body cvs p =: \(body', abval_body) ->
429 fv' = map lookup (filter (`notCaseBound` cvs) fv)
431 foldr doLam (\b -> abval_body) args (getrefs p fv' noRefs) p
433 (StgRhsClosure cc bi fv u args body', abval_rhs)
436 doLam :: Id -> (Refs -> AbVal) -> Refs -> AbVal
440 let b'' = dom_IdEnv c' `merge2` b' `merge2` b in
441 f b'' (addOneTo_IdEnv p i x)))
444 %-----------------------------------------------------------------------------
445 \subsection{Adjusting Update flags}
447 The closure is tagged single entry iff it is used at most once, it is
448 not referenced from inside a data structure or function, and it has no
449 arguments (closures with arguments are re-entrant).
452 tag :: Refs -> IdEnvInt -> StgBinding -> StgBinding
454 tag b c r@(StgNonRec v (StgRhsClosure cc bi fv Updatable [] body))
455 = if (v `notInRefs` b) && (lookupc c v <= 1)
456 then -- trace "One!" (
457 StgNonRec v (StgRhsClosure cc bi fv SingleEntry [] body)
460 tag b c other = other
462 lookupc c v = case lookup_IdEnv c v of
467 %-----------------------------------------------------------------------------
468 \subsection{Top Level analysis}
470 Should we tag top level closures? This could have good implications
471 for CAFs (i.e. they could be made non-updateable if only used once,
472 thus preventing a space leak).
475 updateAnalyse :: [StgBinding] -> [StgBinding] {- Exported -}
477 = udProgram bs null_IdEnv
479 udProgram :: [StgBinding] -> IdEnvClosure -> [StgBinding]
482 = udBinding d noCaseBound p =: \(d', vs, _, abval_bind) ->
483 abval_bind p =: \(_, p') ->
484 grow_IdEnv p p' =: \p'' ->
485 attachUpdateInfoToBinds d' p'' =: \d'' ->
486 d'' : udProgram ds p''
489 %-----------------------------------------------------------------------------
490 \subsection{Exporting Update Information}
492 Convert the exported representation of a function's update function
493 into a real Closure value.
496 convertUpdateSpec :: UpdateSpec -> Closure
497 convertUpdateSpec = mkClosure null_IdEnv noRefs noRefs
499 mkClosure :: IdEnvInt -> Refs -> Refs -> UpdateSpec -> Closure
501 mkClosure c b b' [] = (c, b', dont_know b')
502 mkClosure c b b' (0 : ns) = (null_IdEnv, b, Fun (\ _ -> mkClosure c b b' ns))
503 mkClosure c b b' (1 : ns) = (null_IdEnv, b, Fun (\ (c',b'',f) ->
505 (combine_IdEnvs (+) c c')
506 (dom_IdEnv c' `merge2` b'' `merge2` b)
509 mkClosure c b b' (n : ns) = (null_IdEnv, b, Fun (\ (c',b'',f) ->
511 (dom_IdEnv c' `merge2` b'' `merge2` b)
512 (dom_IdEnv c' `merge2` b'' `merge2` b')
516 Convert a Closure into a representation that can be placed in a .hi file.
519 mkUpdateSpec :: Id -> Closure -> UpdateSpec
520 mkUpdateSpec v f = {- removeSuperfluous2s -} (map countUses ids)
522 (c,b,_) = foldl doApp f ids
523 ids = map mkid (getBuiltinUniques arity)
524 mkid u = mkSysLocal SLIT("upd") u noType noSrcLoc
525 countUses u = if u `elemRefs` b then 2 else min (lookupc c u) 2
526 noType = panic "UpdAnal: no type!"
529 = f (unit_IdEnv i 1, noRefs, dont_know noRefs) =: \(c',b',f') ->
530 (combine_IdEnvs (+) c' c, b', f')
532 (_,dict_tys,tau_ty) = (splitSigmaTy . idType) v
533 (reg_arg_tys, _) = splitFunTys tau_ty
534 arity = length dict_tys + length reg_arg_tys
537 removeSuperfluous2s = reverse . dropWhile (> 1) . reverse
539 %-----------------------------------------------------------------------------
540 \subsection{Attaching the update information to top-level bindings}
542 This is so that the information can later be retrieved for printing
543 out in the .hi file. This is not an ideal solution, however it will
547 attachUpdateInfoToBinds b p
549 StgNonRec v rhs -> StgNonRec (attachOne v) rhs
550 StgRec bs -> StgRec [ (attachOne v, rhs) | (v, rhs) <- bs ]
553 | externallyVisibleId v
554 = let c = lookup v p in
556 (mkUpdateInfo (mkUpdateSpec v c))
560 %-----------------------------------------------------------------------------