2 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
4 \section[SpecConstr]{Specialise over constructors}
11 #include "HsVersions.h"
14 import CoreLint ( showPass, endPass )
15 import CoreUtils ( exprType, tcEqExpr, mkPiTypes )
16 import CoreFVs ( exprsFreeVars )
17 import CoreSubst ( Subst, mkSubst, substExpr )
18 import CoreTidy ( tidyRules )
19 import PprCore ( pprRules )
20 import WwLib ( mkWorkerArgs )
21 import DataCon ( dataConRepArity, isVanillaDataCon )
22 import Type ( tyConAppArgs, tyVarsOfTypes )
23 import Unify ( coreRefineTys )
24 import Id ( Id, idName, idType, isDataConWorkId_maybe,
25 mkUserLocal, mkSysLocal )
29 import Name ( nameOccName, nameSrcLoc )
30 import Rules ( addIdSpecialisations, mkLocalRule, rulesOfBinds )
31 import OccName ( mkSpecOcc )
32 import ErrUtils ( dumpIfSet_dyn )
33 import DynFlags ( DynFlags, DynFlag(..) )
34 import BasicTypes ( Activation(..) )
35 import Maybes ( orElse )
36 import Util ( mapAccumL, lengthAtLeast, notNull )
37 import List ( nubBy, partition )
43 -----------------------------------------------------
45 -----------------------------------------------------
50 drop n (x:xs) = drop (n-1) xs
52 After the first time round, we could pass n unboxed. This happens in
53 numerical code too. Here's what it looks like in Core:
55 drop n xs = case xs of
60 _ -> drop (I# (n# -# 1#)) xs
62 Notice that the recursive call has an explicit constructor as argument.
63 Noticing this, we can make a specialised version of drop
65 RULE: drop (I# n#) xs ==> drop' n# xs
67 drop' n# xs = let n = I# n# in ...orig RHS...
69 Now the simplifier will apply the specialisation in the rhs of drop', giving
71 drop' n# xs = case xs of
75 _ -> drop (n# -# 1#) xs
79 We'd also like to catch cases where a parameter is carried along unchanged,
80 but evaluated each time round the loop:
82 f i n = if i>0 || i>n then i else f (i*2) n
84 Here f isn't strict in n, but we'd like to avoid evaluating it each iteration.
85 In Core, by the time we've w/wd (f is strict in i) we get
87 f i# n = case i# ># 0 of
89 True -> case n of n' { I# n# ->
92 True -> f (i# *# 2#) n'
94 At the call to f, we see that the argument, n is know to be (I# n#),
95 and n is evaluated elsewhere in the body of f, so we can play the same
96 trick as above. However we don't want to do that if the boxed version
97 of n is needed (else we'd avoid the eval but pay more for re-boxing n).
98 So in this case we want that the *only* uses of n are in case statements.
103 * A self-recursive function. Ignore mutual recursion for now,
104 because it's less common, and the code is simpler for self-recursion.
108 a) At a recursive call, one or more parameters is an explicit
109 constructor application
111 That same parameter is scrutinised by a case somewhere in
112 the RHS of the function
116 b) At a recursive call, one or more parameters has an unfolding
117 that is an explicit constructor application
119 That same parameter is scrutinised by a case somewhere in
120 the RHS of the function
122 Those are the only uses of the parameter
125 There's a bit of a complication with type arguments. If the call
128 f p = ...f ((:) [a] x xs)...
130 then our specialised function look like
132 f_spec x xs = let p = (:) [a] x xs in ....as before....
134 This only makes sense if either
135 a) the type variable 'a' is in scope at the top of f, or
136 b) the type variable 'a' is an argument to f (and hence fs)
138 Actually, (a) may hold for value arguments too, in which case
139 we may not want to pass them. Supose 'x' is in scope at f's
140 defn, but xs is not. Then we'd like
142 f_spec xs = let p = (:) [a] x xs in ....as before....
144 Similarly (b) may hold too. If x is already an argument at the
145 call, no need to pass it again.
147 Finally, if 'a' is not in scope at the call site, we could abstract
148 it as we do the term variables:
150 f_spec a x xs = let p = (:) [a] x xs in ...as before...
152 So the grand plan is:
154 * abstract the call site to a constructor-only pattern
155 e.g. C x (D (f p) (g q)) ==> C s1 (D s2 s3)
157 * Find the free variables of the abstracted pattern
159 * Pass these variables, less any that are in scope at
163 NOTICE that we only abstract over variables that are not in scope,
164 so we're in no danger of shadowing variables used in "higher up"
168 %************************************************************************
170 \subsection{Top level wrapper stuff}
172 %************************************************************************
175 specConstrProgram :: DynFlags -> UniqSupply -> [CoreBind] -> IO [CoreBind]
176 specConstrProgram dflags us binds
178 showPass dflags "SpecConstr"
180 let (binds', _) = initUs us (go emptyScEnv binds)
182 endPass dflags "SpecConstr" Opt_D_dump_spec binds'
184 dumpIfSet_dyn dflags Opt_D_dump_rules "Top-level specialisations"
185 (pprRules (tidyRules emptyTidyEnv (rulesOfBinds binds')))
189 go env [] = returnUs []
190 go env (bind:binds) = scBind env bind `thenUs` \ (env', _, bind') ->
191 go env' binds `thenUs` \ binds' ->
192 returnUs (bind' : binds')
196 %************************************************************************
198 \subsection{Environment: goes downwards}
200 %************************************************************************
203 data ScEnv = SCE { scope :: VarEnv HowBound,
204 -- Binds all non-top-level variables in scope
209 type ConstrEnv = IdEnv ConValue
210 data ConValue = CV AltCon [CoreArg]
211 -- Variables known to be bound to a constructor
212 -- in a particular case alternative
214 refineConstrEnv :: Subst -> ConstrEnv -> ConstrEnv
215 -- The substitution is a type substitution only
216 refineConstrEnv subst env = mapVarEnv refine_con_value env
218 refine_con_value (CV con args) = CV con (map (substExpr subst) args)
220 emptyScEnv = SCE { scope = emptyVarEnv, cons = emptyVarEnv }
222 data HowBound = RecFun -- These are the recursive functions for which
223 -- we seek interesting call patterns
225 | RecArg -- These are those functions' arguments; we are
226 -- interested to see if those arguments are scrutinised
228 | Other -- We track all others so we know what's in scope
229 -- This is used in spec_one to check what needs to be
230 -- passed as a parameter and what is in scope at the
231 -- function definition site
233 instance Outputable HowBound where
234 ppr RecFun = text "RecFun"
235 ppr RecArg = text "RecArg"
236 ppr Other = text "Other"
238 lookupScopeEnv env v = lookupVarEnv (scope env) v
240 extendBndrs env bndrs = env { scope = extendVarEnvList (scope env) [(b,Other) | b <- bndrs] }
241 extendBndr env bndr = env { scope = extendVarEnv (scope env) bndr Other }
246 -- we want to bind b, and perhaps scrut too, to (C x y)
247 extendCaseBndrs :: ScEnv -> Id -> CoreExpr -> AltCon -> [Var] -> ScEnv
248 extendCaseBndrs env case_bndr scrut DEFAULT alt_bndrs
249 = extendBndrs env (case_bndr : alt_bndrs)
251 extendCaseBndrs env case_bndr scrut con@(LitAlt lit) alt_bndrs
252 = ASSERT( null alt_bndrs ) extendAlt env case_bndr scrut (CV con []) []
254 extendCaseBndrs env case_bndr scrut con@(DataAlt data_con) alt_bndrs
255 | isVanillaDataCon data_con
256 = extendAlt env case_bndr scrut (CV con vanilla_args) alt_bndrs
259 = extendAlt env1 case_bndr scrut (CV con gadt_args) alt_bndrs
261 vanilla_args = map Type (tyConAppArgs (idType case_bndr)) ++
262 map varToCoreExpr alt_bndrs
264 gadt_args = map (substExpr subst . varToCoreExpr) alt_bndrs
266 (alt_tvs, _) = span isTyVar alt_bndrs
267 Just (tv_subst, is_local) = coreRefineTys data_con alt_tvs (idType case_bndr)
268 subst = mkSubst in_scope tv_subst emptyVarEnv -- No Id substitition
269 in_scope = mkInScopeSet (tyVarsOfTypes (varEnvElts tv_subst))
271 env1 | is_local = env
272 | otherwise = env { cons = refineConstrEnv subst (cons env) }
276 extendAlt :: ScEnv -> Id -> CoreExpr -> ConValue -> [Var] -> ScEnv
277 extendAlt env case_bndr scrut val alt_bndrs
279 env1 = SCE { scope = extendVarEnvList (scope env) [(b,Other) | b <- case_bndr : alt_bndrs],
280 cons = extendVarEnv (cons env) case_bndr val }
283 Var v -> -- Bind the scrutinee in the ConstrEnv if it's a variable
284 -- Also forget if the scrutinee is a RecArg, because we're
285 -- now in the branch of a case, and we don't want to
286 -- record a non-scrutinee use of v if we have
287 -- case v of { (a,b) -> ...(f v)... }
288 SCE { scope = extendVarEnv (scope env1) v Other,
289 cons = extendVarEnv (cons env1) v val }
292 -- When we encounter a recursive function binding
294 -- we want to extend the scope env with bindings
295 -- that record that f is a RecFn and x,y are RecArgs
296 extendRecBndr env fn bndrs
297 = env { scope = scope env `extendVarEnvList`
298 ((fn,RecFun): [(bndr,RecArg) | bndr <- bndrs]) }
302 %************************************************************************
304 \subsection{Usage information: flows upwards}
306 %************************************************************************
311 calls :: !(IdEnv ([Call])), -- Calls
312 -- The functions are a subset of the
313 -- RecFuns in the ScEnv
315 occs :: !(IdEnv ArgOcc) -- Information on argument occurrences
316 } -- The variables are a subset of the
317 -- RecArg in the ScEnv
319 type Call = (ConstrEnv, [CoreArg])
320 -- The arguments of the call, together with the
321 -- env giving the constructor bindings at the call site
323 nullUsage = SCU { calls = emptyVarEnv, occs = emptyVarEnv }
325 combineUsage u1 u2 = SCU { calls = plusVarEnv_C (++) (calls u1) (calls u2),
326 occs = plusVarEnv_C combineOcc (occs u1) (occs u2) }
328 combineUsages [] = nullUsage
329 combineUsages us = foldr1 combineUsage us
331 data ArgOcc = CaseScrut
335 instance Outputable ArgOcc where
336 ppr CaseScrut = ptext SLIT("case-scrut")
337 ppr OtherOcc = ptext SLIT("other-occ")
338 ppr Both = ptext SLIT("case-scrut and other")
340 combineOcc CaseScrut CaseScrut = CaseScrut
341 combineOcc OtherOcc OtherOcc = OtherOcc
342 combineOcc _ _ = Both
346 %************************************************************************
348 \subsection{The main recursive function}
350 %************************************************************************
352 The main recursive function gathers up usage information, and
353 creates specialised versions of functions.
356 scExpr :: ScEnv -> CoreExpr -> UniqSM (ScUsage, CoreExpr)
357 -- The unique supply is needed when we invent
358 -- a new name for the specialised function and its args
360 scExpr env e@(Type t) = returnUs (nullUsage, e)
361 scExpr env e@(Lit l) = returnUs (nullUsage, e)
362 scExpr env e@(Var v) = returnUs (varUsage env v OtherOcc, e)
363 scExpr env (Note n e) = scExpr env e `thenUs` \ (usg,e') ->
364 returnUs (usg, Note n e')
365 scExpr env (Lam b e) = scExpr (extendBndr env b) e `thenUs` \ (usg,e') ->
366 returnUs (usg, Lam b e')
368 scExpr env (Case scrut b ty alts)
369 = sc_scrut scrut `thenUs` \ (scrut_usg, scrut') ->
370 mapAndUnzipUs sc_alt alts `thenUs` \ (alts_usgs, alts') ->
371 returnUs (combineUsages alts_usgs `combineUsage` scrut_usg,
372 Case scrut' b ty alts')
374 sc_scrut e@(Var v) = returnUs (varUsage env v CaseScrut, e)
375 sc_scrut e = scExpr env e
377 sc_alt (con,bs,rhs) = scExpr env1 rhs `thenUs` \ (usg,rhs') ->
378 returnUs (usg, (con,bs,rhs'))
380 env1 = extendCaseBndrs env b scrut con bs
382 scExpr env (Let bind body)
383 = scBind env bind `thenUs` \ (env', bind_usg, bind') ->
384 scExpr env' body `thenUs` \ (body_usg, body') ->
385 returnUs (bind_usg `combineUsage` body_usg, Let bind' body')
387 scExpr env e@(App _ _)
389 (fn, args) = collectArgs e
391 mapAndUnzipUs (scExpr env) args `thenUs` \ (usgs, args') ->
393 arg_usg = combineUsages usgs
394 fn_usg | Var f <- fn,
395 Just RecFun <- lookupScopeEnv env f
396 = SCU { calls = unitVarEnv f [(cons env, args)],
401 returnUs (arg_usg `combineUsage` fn_usg, mkApps fn args')
402 -- Don't bother to look inside fn;
403 -- it's almost always a variable
405 ----------------------
406 scBind :: ScEnv -> CoreBind -> UniqSM (ScEnv, ScUsage, CoreBind)
407 scBind env (Rec [(fn,rhs)])
409 = scExpr env_fn_body body `thenUs` \ (usg, body') ->
411 SCU { calls = calls, occs = occs } = usg
413 specialise env fn bndrs body usg `thenUs` \ (rules, spec_prs) ->
414 returnUs (extendBndr env fn, -- For the body of the letrec, just
415 -- extend the env with Other to record
416 -- that it's in scope; no funny RecFun business
417 SCU { calls = calls `delVarEnv` fn, occs = occs `delVarEnvList` val_bndrs},
418 Rec ((fn `addIdSpecialisations` rules, mkLams bndrs body') : spec_prs))
420 (bndrs,body) = collectBinders rhs
421 val_bndrs = filter isId bndrs
422 env_fn_body = extendRecBndr env fn bndrs
425 = mapAndUnzipUs do_one prs `thenUs` \ (usgs, prs') ->
426 returnUs (extendBndrs env (map fst prs), combineUsages usgs, Rec prs')
428 do_one (bndr,rhs) = scExpr env rhs `thenUs` \ (usg, rhs') ->
429 returnUs (usg, (bndr,rhs'))
431 scBind env (NonRec bndr rhs)
432 = scExpr env rhs `thenUs` \ (usg, rhs') ->
433 returnUs (extendBndr env bndr, usg, NonRec bndr rhs')
435 ----------------------
437 | Just RecArg <- lookupScopeEnv env v = SCU { calls = emptyVarEnv,
438 occs = unitVarEnv v use }
439 | otherwise = nullUsage
443 %************************************************************************
445 \subsection{The specialiser}
447 %************************************************************************
452 -> [CoreBndr] -> CoreExpr -- Its RHS
453 -> ScUsage -- Info on usage
454 -> UniqSM ([CoreRule], -- Rules
455 [(Id,CoreExpr)]) -- Bindings
457 specialise env fn bndrs body (SCU {calls=calls, occs=occs})
458 = getUs `thenUs` \ us ->
460 all_calls = lookupVarEnv calls fn `orElse` []
462 good_calls :: [[CoreArg]]
464 | (con_env, call_args) <- all_calls,
465 call_args `lengthAtLeast` n_bndrs, -- App is saturated
466 let call = (bndrs `zip` call_args),
467 any (good_arg con_env occs) call, -- At least one arg is a constr app
468 let (_, pats) = argsToPats con_env us call_args
471 mapAndUnzipUs (spec_one env fn (mkLams bndrs body))
472 (nubBy same_call good_calls `zip` [1..])
474 n_bndrs = length bndrs
475 same_call as1 as2 = and (zipWith tcEqExpr as1 as2)
477 ---------------------
478 good_arg :: ConstrEnv -> IdEnv ArgOcc -> (CoreBndr, CoreArg) -> Bool
479 good_arg con_env arg_occs (bndr, arg)
480 = case is_con_app_maybe con_env arg of
481 Just _ -> bndr_usg_ok arg_occs bndr arg
484 bndr_usg_ok :: IdEnv ArgOcc -> Var -> CoreArg -> Bool
485 bndr_usg_ok arg_occs bndr arg
486 = case lookupVarEnv arg_occs bndr of
487 Just CaseScrut -> True -- Used only by case scrutiny
488 Just Both -> case arg of -- Used by case and elsewhere
489 App _ _ -> True -- so the arg should be an explicit con app
491 other -> False -- Not used, or used wonkily
494 ---------------------
497 -> CoreExpr -- Rhs of the original function
499 -> UniqSM (CoreRule, (Id,CoreExpr)) -- Rule and binding
501 -- spec_one creates a specialised copy of the function, together
502 -- with a rule for using it. I'm very proud of how short this
503 -- function is, considering what it does :-).
509 f = /\b \y::[(a,b)] -> ....f (b,c) ((:) (a,(b,c)) (x,v) (h w))...
510 [c::*, v::(b,c) are presumably bound by the (...) part]
512 f_spec = /\ b c \ v::(b,c) hw::[(a,(b,c))] ->
513 (...entire RHS of f...) (b,c) ((:) (a,(b,c)) (x,v) hw)
515 RULE: forall b::* c::*, -- Note, *not* forall a, x
519 f (b,c) ((:) (a,(b,c)) (x,v) hw) = f_spec b c v hw
522 spec_one env fn rhs (pats, rule_number)
523 = getUniqueUs `thenUs` \ spec_uniq ->
526 fn_loc = nameSrcLoc fn_name
527 spec_occ = mkSpecOcc (nameOccName fn_name)
528 pat_fvs = varSetElems (exprsFreeVars pats)
529 vars_to_bind = filter not_avail pat_fvs
530 not_avail v = not (v `elemVarEnv` scope env)
531 -- Put the type variables first; the type of a term
532 -- variable may mention a type variable
533 (tvs, ids) = partition isTyVar vars_to_bind
535 spec_body = mkApps rhs pats
536 body_ty = exprType spec_body
538 (spec_lam_args, spec_call_args) = mkWorkerArgs bndrs body_ty
539 -- Usual w/w hack to avoid generating
540 -- a spec_rhs of unlifted type and no args
542 rule_name = mkFastString ("SC:" ++ showSDoc (ppr fn <> int rule_number))
543 spec_rhs = mkLams spec_lam_args spec_body
544 spec_id = mkUserLocal spec_occ spec_uniq (mkPiTypes spec_lam_args body_ty) fn_loc
545 rule_rhs = mkVarApps (Var spec_id) spec_call_args
546 rule = mkLocalRule rule_name specConstrActivation fn_name bndrs pats rule_rhs
548 returnUs (rule, (spec_id, spec_rhs))
550 -- In which phase should the specialise-constructor rules be active?
551 -- Originally I made them always-active, but Manuel found that
552 -- this defeated some clever user-written rules. So Plan B
553 -- is to make them active only in Phase 0; after all, currently,
554 -- the specConstr transformation is only run after the simplifier
555 -- has reached Phase 0. In general one would want it to be
556 -- flag-controllable, but for now I'm leaving it baked in
558 specConstrActivation :: Activation
559 specConstrActivation = ActiveAfter 0 -- Baked in; see comments above
562 %************************************************************************
564 \subsection{Argument analysis}
566 %************************************************************************
568 This code deals with analysing call-site arguments to see whether
569 they are constructor applications.
572 -- argToPat takes an actual argument, and returns an abstracted
573 -- version, consisting of just the "constructor skeleton" of the
574 -- argument, with non-constructor sub-expression replaced by new
575 -- placeholder variables. For example:
576 -- C a (D (f x) (g y)) ==> C p1 (D p2 p3)
578 argToPat :: ConstrEnv -> UniqSupply -> CoreArg -> (UniqSupply, CoreExpr)
579 argToPat env us (Type ty)
583 | Just (CV dc args) <- is_con_app_maybe env arg
585 (us',args') = argsToPats env us args
587 (us', mk_con_app dc args')
589 argToPat env us (Var v) -- Don't uniqify existing vars,
590 = (us, Var v) -- so that we can spot when we pass them twice
593 = (us1, Var (mkSysLocal FSLIT("sc") (uniqFromSupply us2) (exprType arg)))
595 (us1,us2) = splitUniqSupply us
597 argsToPats :: ConstrEnv -> UniqSupply -> [CoreArg] -> (UniqSupply, [CoreExpr])
598 argsToPats env us args = mapAccumL (argToPat env) us args
603 is_con_app_maybe :: ConstrEnv -> CoreExpr -> Maybe ConValue
604 is_con_app_maybe env (Var v)
606 -- You might think we could look in the idUnfolding here
607 -- but that doesn't take account of which branch of a
608 -- case we are in, which is the whole point
610 is_con_app_maybe env (Lit lit)
611 = Just (CV (LitAlt lit) [])
613 is_con_app_maybe env expr
614 = case collectArgs expr of
615 (Var fun, args) | Just con <- isDataConWorkId_maybe fun,
616 args `lengthAtLeast` dataConRepArity con
617 -- Might be > because the arity excludes type args
618 -> Just (CV (DataAlt con) args)
622 mk_con_app :: AltCon -> [CoreArg] -> CoreExpr
623 mk_con_app (LitAlt lit) [] = Lit lit
624 mk_con_app (DataAlt con) args = mkConApp con args