2 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
4 \section[CoreRules]{Transformation rules}
8 RuleBase, prepareLocalRuleBase, prepareOrphanRuleBase,
9 unionRuleBase, lookupRule, addRule, addIdSpecialisations,
10 ProtoCoreRule(..), pprProtoCoreRule, pprRuleBase,
14 #include "HsVersions.h"
16 import CoreSyn -- All of it
17 import OccurAnal ( occurAnalyseRule )
18 import BinderInfo ( markMany )
19 import CoreFVs ( exprFreeVars, idRuleVars, ruleRhsFreeVars, ruleSomeLhsFreeVars )
20 import CoreUnfold ( isCheapUnfolding, unfoldingTemplate )
21 import CoreUtils ( eqExpr )
22 import PprCore ( pprCoreRule )
23 import Subst ( Subst, InScopeSet, substBndr, lookupSubst, extendSubst,
24 mkSubst, substEnv, setSubstEnv, emptySubst, isInScope,
25 unBindSubst, bindSubstList, unBindSubstList, substInScope
27 import Id ( Id, idUnfolding, zapLamIdInfo,
28 idSpecialisation, setIdSpecialisation,
29 setIdNoDiscard, maybeModifyIdInfo, modifyIdInfo
31 import Name ( Name, isLocallyDefined )
32 import Var ( isTyVar, isId )
35 import Type ( mkTyVarTy, getTyVar_maybe )
36 import qualified Unify ( match )
37 import CmdLineOpts ( opt_D_dump_simpl, opt_D_verbose_core2core )
40 import ErrUtils ( dumpIfSet )
42 import Maybes ( maybeToBool )
43 import List ( partition )
44 import Util ( sortLt )
48 %************************************************************************
50 \subsection[specialisation-IdInfo]{Specialisation info about an @Id@}
52 %************************************************************************
54 A @CoreRule@ holds details of one rule for an @Id@, which
55 includes its specialisations.
57 For example, if a rule for @f@ contains the mapping:
59 forall a b d. [Type (List a), Type b, Var d] ===> f' a b
61 then when we find an application of f to matching types, we simply replace
62 it by the matching RHS:
64 f (List Int) Bool dict ===> f' Int Bool
66 All the stuff about how many dictionaries to discard, and what types
67 to apply the specialised function to, are handled by the fact that the
68 Rule contains a template for the result of the specialisation.
70 There is one more exciting case, which is dealt with in exactly the same
71 way. If the specialised value is unboxed then it is lifted at its
72 definition site and unlifted at its uses. For example:
74 pi :: forall a. Num a => a
76 might have a specialisation
78 [Int#] ===> (case pi' of Lift pi# -> pi#)
80 where pi' :: Lift Int# is the specialised version of pi.
83 %************************************************************************
87 %************************************************************************
90 matchRules :: InScopeSet -> [CoreRule] -> [CoreExpr] -> Maybe (RuleName, CoreExpr)
91 -- See comments on matchRule
92 matchRules in_scope [] args = Nothing
93 matchRules in_scope (rule:rules) args
94 = case matchRule in_scope rule args of
95 Just result -> Just result
96 Nothing -> matchRules in_scope rules args
99 matchRule :: InScopeSet -> CoreRule -> [CoreExpr] -> Maybe (RuleName, CoreExpr)
101 -- If (matchRule rule args) returns Just (name,rhs)
102 -- then (f args) matches the rule, and the corresponding
103 -- rewritten RHS is rhs
105 -- The bndrs and rhs is occurrence-analysed
110 -- forall f g x. map f (map g x) ==> map (f . g) x
112 -- CoreRule "map/map"
113 -- [f,g,x] -- tpl_vars
114 -- [f,map g x] -- tpl_args
115 -- map (f.g) x) -- rhs
117 -- Then the call: matchRule the_rule [e1,map e2 e3]
118 -- = Just ("map/map", (\f,g,x -> rhs) e1 e2 e3)
120 -- Any 'surplus' arguments in the input are simply put on the end
124 -- A1. No top-level variable is bound in the target
125 -- A2. No template variable is bound in the target
126 -- A3. No lambda bound template variable is free in any subexpression of the target
128 -- To see why A1 is necessary, consider matching
129 -- \x->f against \f->f
130 -- When we meet the lambdas we substitute [f/x] in the template (a no-op),
131 -- and then erroneously succeed in matching f against f.
133 -- To see why A2 is needed consider matching
134 -- forall a. \b->b against \a->3
135 -- When we meet the lambdas we substitute [a/b] in the template, and then
136 -- erroneously succeed in matching what looks like the template variable 'a' against 3.
138 -- A3 is needed to validate the rule that says
141 -- (\x->E) matches (\x->F x)
144 matchRule in_scope rule@(BuiltinRule match_fn) args = match_fn args
146 matchRule in_scope rule@(Rule rn tpl_vars tpl_args rhs) args
147 = go tpl_args args emptySubst
148 -- We used to use the in_scope set, but I don't think that's necessary
149 -- After all, the result is going to be simplified again with that in_scope set
151 tpl_var_set = mkVarSet tpl_vars
153 -----------------------
155 go (tpl_arg:tpl_args) (arg:args) subst = match tpl_arg arg tpl_var_set (go tpl_args args) subst
157 -- Two easy ways to terminate
158 go [] [] subst = Just (rn, app_match subst (mkLams tpl_vars rhs) tpl_vars)
159 go [] args subst = Just (rn, app_match subst (mkLams tpl_vars rhs) tpl_vars `mkApps` args)
161 -- One tiresome way to terminate: check for excess unmatched
162 -- template arguments
163 go tpl_args [] subst = Nothing -- Failure
166 -----------------------
167 app_match subst fn vs = foldl go fn vs
169 senv = substEnv subst
170 go fn v = case lookupSubstEnv senv v of
171 Just (DoneEx ex) -> fn `App` ex
172 Just (DoneTy ty) -> fn `App` Type ty
173 -- Substitution should bind them all!
176 -----------------------
177 {- The code below tries to match even if there are more
178 template args than real args.
180 I now think this is probably a bad idea.
181 Should the template (map f xs) match (map g)? I think not.
182 For a start, in general eta expansion wastes work.
185 = case eta_complete tpl_args (mkVarSet leftovers) of
186 Just leftovers' -> Just (rn, mkLams done (mkLams leftovers' rhs),
187 mk_result_args subst done)
188 Nothing -> Nothing -- Failure
190 (done, leftovers) = partition (\v -> maybeToBool (lookupSubstEnv subst_env v))
191 (map zapOccInfo tpl_vars)
193 subst_env = substEnv subst
195 -----------------------
196 eta_complete [] vars = ASSERT( isEmptyVarSet vars )
198 eta_complete (Type ty:tpl_args) vars
199 = case getTyVar_maybe ty of
200 Just tv | tv `elemVarSet` vars
201 -> case eta_complete tpl_args (vars `delVarSet` tv) of
202 Just vars' -> Just (tv:vars')
206 eta_complete (Var v:tpl_args) vars
207 | v `elemVarSet` vars
208 = case eta_complete tpl_args (vars `delVarSet` v) of
209 Just vars' -> Just (v:vars')
212 eta_complete other vars = Nothing
216 zapOccInfo bndr | isTyVar bndr = bndr
217 | otherwise = zapLamIdInfo bndr
221 type Matcher result = VarSet -- Template variables
222 -> (Subst -> Maybe result) -- Continuation if success
223 -> Subst -> Maybe result -- Substitution so far -> result
224 -- The *SubstEnv* in these Substs apply to the TEMPLATE only
226 -- The *InScopeSet* in these Substs gives variables bound so far in the
227 -- target term. So when matching forall a. (\x. a x) against (\y. y y)
228 -- while processing the body of the lambdas, the in-scope set will be {y}.
229 -- That lets us do the occurs-check when matching 'a' against 'y'
231 match :: CoreExpr -- Template
232 -> CoreExpr -- Target
237 match (Var v1) e2 tpl_vars kont subst
238 = case lookupSubst subst v1 of
239 Nothing | v1 `elemVarSet` tpl_vars -- v1 is a template variable
240 -> if (any (`isInScope` subst) (varSetElems (exprFreeVars e2))) then
241 match_fail -- Occurs check failure
242 -- e.g. match forall a. (\x-> a x) against (\y. y y)
244 kont (extendSubst subst v1 (DoneEx e2))
247 | eqExpr (Var v1) e2 -> kont subst
248 -- v1 is not a template variable, so it must be a global constant
250 Just (DoneEx e2') | eqExpr e2' e2 -> kont subst
254 match (Lit lit1) (Lit lit2) tpl_vars kont subst
258 match (App f1 a1) (App f2 a2) tpl_vars kont subst
259 = match f1 f2 tpl_vars (match a1 a2 tpl_vars kont) subst
261 match (Lam x1 e1) (Lam x2 e2) tpl_vars kont subst
262 = bind [x1] [x2] (match e1 e2) tpl_vars kont subst
264 -- This rule does eta expansion
265 -- (\x.M) ~ N iff M ~ N x
267 match (Lam x1 e1) e2 tpl_vars kont subst
268 = bind [x1] [x1] (match e1 (App e2 (mkVarArg x1))) tpl_vars kont subst
270 -- Eta expansion the other way
271 -- M ~ (\y.N) iff \y.M y ~ \y.N
273 -- Remembering that by (A), y can't be free in M, we get this
274 match e1 (Lam x2 e2) tpl_vars kont subst
275 = bind [new_id] [x2] (match (App e1 (mkVarArg new_id)) e2) tpl_vars kont subst
277 new_id = uniqAway (substInScope subst) x2
278 -- This uniqAway is actually needed. Here's the example:
279 -- rule: foldr (mapFB (:) f) [] = mapList
280 -- target: foldr (\x. mapFB k f x) []
282 -- k = \x. mapFB ... x
283 -- The first \x is ok, but when we inline k, hoping it might
284 -- match (:) we find a second \x.
286 match (Case e1 x1 alts1) (Case e2 x2 alts2) tpl_vars kont subst
287 = match e1 e2 tpl_vars case_kont subst
289 case_kont subst = bind [x1] [x2] (match_alts alts1 (sortLt lt_alt alts2))
292 match (Type ty1) (Type ty2) tpl_vars kont subst
293 = match_ty ty1 ty2 tpl_vars kont subst
295 match (Note (Coerce to1 from1) e1) (Note (Coerce to2 from2) e2)
297 = (match_ty to1 to2 tpl_vars $
298 match_ty from1 from2 tpl_vars $
299 match e1 e2 tpl_vars kont) subst
302 {- I don't buy this let-rule any more
303 The let rule fails on matching
304 forall f,x,xs. f (x:xs)
306 f (let y = e in (y:[]))
307 because we just get x->y, which is bogus.
309 -- This is an interesting rule: we simply ignore lets in the
310 -- term being matched against! The unfolding inside it is (by assumption)
311 -- already inside any occurrences of the bound variables, so we'll expand
312 -- them when we encounter them. Meanwhile, we can't get false matches because
313 -- (also by assumption) the term being matched has no shadowing.
314 match e1 (Let bind e2) tpl_vars kont subst
315 = match e1 e2 tpl_vars kont subst
318 -- Here is another important rule: if the term being matched is a
319 -- variable, we expand it so long as its unfolding is a WHNF
320 -- (Its occurrence information is not necessarily up to date,
321 -- so we don't use it.)
322 match e1 (Var v2) tpl_vars kont subst
323 | isCheapUnfolding unfolding
324 = match e1 (unfoldingTemplate unfolding) tpl_vars kont subst
326 unfolding = idUnfolding v2
329 -- We can't cope with lets in the template
331 match e1 e2 tpl_vars kont subst = match_fail
334 ------------------------------------------
335 match_alts [] [] tpl_vars kont subst
337 match_alts ((c1,vs1,r1):alts1) ((c2,vs2,r2):alts2) tpl_vars kont subst
339 = bind vs1 vs2 (match r1 r2) tpl_vars
340 (match_alts alts1 alts2 tpl_vars kont)
342 match_alts alts1 alts2 tpl_vars kont subst = match_fail
344 lt_alt (con1, _, _) (con2, _, _) = con1 < con2
346 ----------------------------------------
347 bind :: [CoreBndr] -- Template binders
348 -> [CoreBndr] -- Target binders
351 -- This makes uses of assumption (A) above. For example,
353 -- Template: (\x.y) (y is free)
354 -- Target : (\y.y) (y is bound)
355 -- We rename x to y in the template... but then erroneously
356 -- match y against y. But this can't happen because of (A)
357 bind vs1 vs2 matcher tpl_vars kont subst
358 = WARN( not (all not_in_subst vs1), bug_msg )
359 matcher tpl_vars kont' subst'
361 kont' subst'' = kont (unBindSubstList subst'' vs1 vs2)
362 subst' = bindSubstList subst vs1 vs2
364 -- The unBindSubst relies on no shadowing in the template
365 not_in_subst v = not (maybeToBool (lookupSubst subst v))
366 bug_msg = sep [ppr vs1, ppr vs2]
368 ----------------------------------------
369 match_ty ty1 ty2 tpl_vars kont subst
370 = case Unify.match ty1 ty2 tpl_vars Just (substEnv subst) of
371 Nothing -> match_fail
372 Just senv' -> kont (setSubstEnv subst senv')
374 ----------------------------------------
375 matches [] [] tpl_vars kont subst
377 matches (e:es) (e':es') tpl_vars kont subst
378 = match e e' tpl_vars (matches es es' tpl_vars kont) subst
379 matches es es' tpl_vars kont subst
382 ----------------------------------------
383 mkVarArg :: CoreBndr -> CoreArg
384 mkVarArg v | isId v = Var v
385 | otherwise = Type (mkTyVarTy v)
388 %************************************************************************
390 \subsection{Adding a new rule}
392 %************************************************************************
395 addRule :: Id -> CoreRules -> CoreRule -> CoreRules
397 -- Insert the new rule just before a rule that is *less specific*
398 -- than the new one; or at the end if there isn't such a one.
399 -- In this way we make sure that when looking up, the first match
400 -- is the most specific.
402 -- We make no check for rules that unify without one dominating
403 -- the other. Arguably this would be a bug.
405 addRule id (Rules rules rhs_fvs) rule@(BuiltinRule _)
406 = Rules (rule:rules) rhs_fvs
407 -- Put it at the start for lack of anything better
409 addRule id (Rules rules rhs_fvs) rule
410 = Rules (insertRule rules new_rule) (rhs_fvs `unionVarSet` new_rhs_fvs)
412 new_rule = occurAnalyseRule rule
413 new_rhs_fvs = ruleRhsFreeVars new_rule `delVarSet` id
415 -- Don't include the Id in its own rhs free-var set.
416 -- Otherwise the occurrence analyser makes bindings recursive
417 -- that shoudn't be. E.g.
418 -- RULE: f (f x y) z ==> f x (f y z)
420 insertRule rules new_rule@(Rule _ tpl_vars tpl_args _)
423 tpl_var_set = mkVarSet tpl_vars
424 -- Actually we should probably include the free vars of tpl_args,
425 -- but I can't be bothered
428 go (rule:rules) | new_is_more_specific rule = (new_rule:rule:rules)
429 | otherwise = rule : go rules
431 new_is_more_specific rule = maybeToBool (matchRule tpl_var_set rule tpl_args)
433 addIdSpecialisations :: Id -> [([CoreBndr], [CoreExpr], CoreExpr)] -> Id
434 addIdSpecialisations id spec_stuff
435 = setIdSpecialisation id new_rules
437 rule_name = _PK_ ("SPEC " ++ showSDoc (ppr id))
438 new_rules = foldr add (idSpecialisation id) spec_stuff
439 add (vars, args, rhs) rules = addRule id rules (Rule rule_name vars args rhs)
443 %************************************************************************
445 \subsection{Preparing the rule base
447 %************************************************************************
452 Bool -- True <=> this rule was defined in this module,
453 Id -- What Id is it for
454 CoreRule -- The rule itself
457 pprProtoCoreRule (ProtoCoreRule _ fn rule) = pprCoreRule (ppr fn) rule
459 lookupRule :: InScopeSet -> Id -> [CoreExpr] -> Maybe (RuleName, CoreExpr)
460 lookupRule in_scope fn args
461 = case idSpecialisation fn of
462 Rules rules _ -> matchRules in_scope rules args
464 localRule :: ProtoCoreRule -> Bool
465 localRule (ProtoCoreRule local _ _) = local
467 orphanRule :: ProtoCoreRule -> Bool
468 -- An "orphan rule" is one that is defined in this
469 -- module, but for an *imported* function. We need
470 -- to track these separately when generating the interface file
471 orphanRule (ProtoCoreRule local fn _)
472 = local && not (isLocallyDefined fn)
476 %************************************************************************
478 \subsection{Getting the rules ready}
480 %************************************************************************
483 type RuleBase = (IdSet, -- Imported Ids that have rules attached
484 IdSet) -- Ids (whether local or imported) mentioned on
485 -- LHS of some rule; these should be black listed
487 unionRuleBase (rule_ids1, black_ids1) (rule_ids2, black_ids2)
488 = (plusUFM_C merge_rules rule_ids1 rule_ids2,
489 unionVarSet black_ids1 black_ids2)
491 merge_rules id1 id2 = let rules1 = idSpecialisation id1
492 rules2 = idSpecialisation id2
493 new_rules = foldl (addRule id1) rules1 (rulesRules rules2)
495 setIdSpecialisation id1 new_rules
497 pprRuleBase :: RuleBase -> SDoc
498 pprRuleBase (rules,_) = vcat [ pprCoreRule (ppr id) rs
499 | id <- varSetElems rules,
500 rs <- rulesRules $ idSpecialisation id ]
502 -- prepareLocalRuleBase takes the CoreBinds and rules defined in this module.
503 -- It attaches those rules that are for local Ids to their binders, and
504 -- returns the remainder attached to Ids in an IdSet. It also returns
505 -- Ids mentioned on LHS of some rule; these should be blacklisted.
507 -- The rule Ids and LHS Ids are black-listed; that is, they aren't inlined
508 -- so that the opportunity to apply the rule isn't lost too soon
510 prepareLocalRuleBase :: [CoreBind] -> [ProtoCoreRule] -> ([CoreBind], RuleBase)
511 prepareLocalRuleBase binds local_rules
512 = (map zap_bind binds, (imported_id_rule_ids, rule_lhs_fvs))
514 (rule_ids, rule_lhs_fvs) = foldr add_rule (emptyVarSet, emptyVarSet) local_rules
515 imported_id_rule_ids = filterVarSet (not . isLocallyDefined) rule_ids
517 -- rule_fvs is the set of all variables mentioned in this module's rules
518 rule_fvs = foldVarSet (unionVarSet . idRuleVars) rule_lhs_fvs rule_ids
520 -- Attach the rules for each locally-defined Id to that Id.
521 -- - This makes the rules easier to look up
522 -- - It means that transformation rules and specialisations for
523 -- locally defined Ids are handled uniformly
524 -- - It keeps alive things that are referred to only from a rule
525 -- (the occurrence analyser knows about rules attached to Ids)
526 -- - It makes sure that, when we apply a rule, the free vars
527 -- of the RHS are more likely to be in scope
529 -- The LHS and RHS Ids are marked 'no-discard'.
530 -- This means that the binding won't be discarded EVEN if the binding
531 -- ends up being trivial (v = w) -- the simplifier would usually just
532 -- substitute w for v throughout, but we don't apply the substitution to
533 -- the rules (maybe we should?), so this substitution would make the rule
535 zap_bind (NonRec b r) = NonRec (zap_bndr b) r
536 zap_bind (Rec prs) = Rec [(zap_bndr b, r) | (b,r) <- prs]
538 zap_bndr bndr = case lookupVarSet rule_ids bndr of
539 Just bndr' -> setIdNoDiscard bndr'
540 Nothing | bndr `elemVarSet` rule_fvs -> setIdNoDiscard bndr
543 add_rule (ProtoCoreRule _ id rule)
544 (rule_id_set, rule_fvs)
545 = (rule_id_set `extendVarSet` new_id,
546 rule_fvs `unionVarSet` extendVarSet lhs_fvs id)
548 new_id = case lookupVarSet rule_id_set id of
549 Just id' -> addRuleToId id' rule
550 Nothing -> addRuleToId id rule
551 lhs_fvs = ruleSomeLhsFreeVars isId rule
552 -- Find *all* the free Ids of the LHS, not just
553 -- locally defined ones!!
555 addRuleToId id rule = setIdSpecialisation id (addRule id (idSpecialisation id) rule)
557 -- prepareOrphanRuleBase does exactly the same as prepareLocalRuleBase, except that
558 -- it assumes that none of the rules can be attached to local Ids.
560 prepareOrphanRuleBase :: [ProtoCoreRule] -> RuleBase
561 prepareOrphanRuleBase imported_rules
562 = foldr add_rule (emptyVarSet, emptyVarSet) imported_rules