2 % (c) The GRASP/AQUA Project, Glasgow University, 1993-1998
4 \section[WorkWrap]{Worker/wrapper-generating back-end of strictness analyser}
7 module WorkWrap ( wwTopBinds, mkWrapper ) where
10 import CoreUnfold ( certainlyWillInline, mkInlineUnfolding, mkWwInlineRule )
11 import CoreUtils ( exprType, exprIsHNF )
12 import CoreArity ( exprArity )
20 import VarEnv ( isEmptyVarEnv )
21 import Maybes ( orElse )
23 import Util ( lengthIs, notNull )
27 #include "HsVersions.h"
30 We take Core bindings whose binders have:
34 \item Strictness attached (by the front-end of the strictness
37 \item Constructed Product Result information attached by the CPR
42 and we return some ``plain'' bindings which have been
43 worker/wrapper-ified, meaning:
47 \item Functions have been split into workers and wrappers where
48 appropriate. If a function has both strictness and CPR properties
49 then only one worker/wrapper doing both transformations is produced;
51 \item Binders' @IdInfos@ have been updated to reflect the existence of
52 these workers/wrappers (this is where we get STRICTNESS and CPR pragma
53 info for exported values).
57 wwTopBinds :: UniqSupply -> [CoreBind] -> [CoreBind]
59 wwTopBinds us top_binds
61 top_binds' <- mapM wwBind top_binds
62 return (concat top_binds')
65 %************************************************************************
67 \subsection[wwBind-wwExpr]{@wwBind@ and @wwExpr@}
69 %************************************************************************
71 @wwBind@ works on a binding, trying each \tr{(binder, expr)} pair in
72 turn. Non-recursive case first, then recursive...
76 -> UniqSM [CoreBind] -- returns a WwBinding intermediate form;
77 -- the caller will convert to Expr/Binding,
80 wwBind (NonRec binder rhs) = do
82 new_pairs <- tryWW NonRecursive binder new_rhs
83 return [NonRec b e | (b,e) <- new_pairs]
84 -- Generated bindings must be non-recursive
85 -- because the original binding was.
88 = return . Rec <$> concatMapM do_one pairs
90 do_one (binder, rhs) = do new_rhs <- wwExpr rhs
91 tryWW Recursive binder new_rhs
94 @wwExpr@ basically just walks the tree, looking for appropriate
95 annotations that can be used. Remember it is @wwBind@ that does the
96 matching by looking for strict arguments of the correct type.
97 @wwExpr@ is a version that just returns the ``Plain'' Tree.
100 wwExpr :: CoreExpr -> UniqSM CoreExpr
102 wwExpr e@(Type {}) = return e
103 wwExpr e@(Lit {}) = return e
104 wwExpr e@(Var {}) = return e
106 wwExpr (Lam binder expr)
107 = Lam binder <$> wwExpr expr
110 = App <$> wwExpr f <*> wwExpr a
112 wwExpr (Note note expr)
113 = Note note <$> wwExpr expr
115 wwExpr (Cast expr co) = do
116 new_expr <- wwExpr expr
117 return (Cast new_expr co)
119 wwExpr (Let bind expr)
120 = mkLets <$> wwBind bind <*> wwExpr expr
122 wwExpr (Case expr binder ty alts) = do
123 new_expr <- wwExpr expr
124 new_alts <- mapM ww_alt alts
125 return (Case new_expr binder ty new_alts)
127 ww_alt (con, binders, rhs) = do
128 new_rhs <- wwExpr rhs
129 return (con, binders, new_rhs)
132 %************************************************************************
134 \subsection[tryWW]{@tryWW@: attempt a worker/wrapper pair}
136 %************************************************************************
138 @tryWW@ just accumulates arguments, converts strictness info from the
139 front-end into the proper form, then calls @mkWwBodies@ to do
142 The only reason this is monadised is for the unique supply.
144 Note [Don't w/w INLINE things]
145 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
146 It's very important to refrain from w/w-ing an INLINE function (ie one
147 with an InlineRule) because the wrapper will then overwrite the
148 InlineRule unfolding.
150 Furthermore, if the programmer has marked something as INLINE,
151 we may lose by w/w'ing it.
153 If the strictness analyser is run twice, this test also prevents
154 wrappers (which are INLINEd) from being re-done. (You can end up with
155 several liked-named Ids bouncing around at the same time---absolute
158 Notice that we refrain from w/w'ing an INLINE function even if it is
159 in a recursive group. It might not be the loop breaker. (We could
160 test for loop-breaker-hood, but I'm not sure that ever matters.)
162 Note [Don't w/w INLINABLE things]
163 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
167 then in principle we might get a more efficient loop by w/w'ing f.
168 But that would make a new unfolding which would overwrite the old
169 one. So we leave INLINABLE things alone too.
171 This is a slight infelicity really, because it means that adding
172 an INLINABLE pragma could make a program a bit less efficient,
173 because you lose the worker/wrapper stuff. But I don't see a way
176 Note [Don't w/w inline small non-loop-breaker things]
177 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
178 In general, we refrain from w/w-ing *small* functions, which are not
179 loop breakers, because they'll inline anyway. But we must take care:
180 it may look small now, but get to be big later after other inlining
181 has happened. So we take the precaution of adding an INLINE pragma to
184 I made this change when I observed a big function at the end of
185 compilation with a useful strictness signature but no w-w. (It was
186 small during demand analysis, we refrained from w/w, and then got big
187 when something was inlined in its rhs.) When I measured it on nofib,
188 it didn't make much difference; just a few percent improved allocation
189 on one benchmark (bspt/Euclid.space). But nothing got worse.
191 There is an infelicity though. We may get something like
194 g x = case gw x of r -> I# r
196 f {- InlineStable, Template = g val -}
197 f = case gw x of r -> I# r
199 The code for f duplicates that for g, without any real benefit. It
200 won't really be executed, because calls to f will go via the inlining.
202 Note [Wrapper activation]
203 ~~~~~~~~~~~~~~~~~~~~~~~~~
204 When should the wrapper inlining be active? It must not be active
205 earlier than the current Activation of the Id (eg it might have a
206 NOINLINE pragma). But in fact strictness analysis happens fairly
207 late in the pipeline, and we want to prioritise specialisations over
208 strictness. Eg if we have
210 f :: Num a => a -> Int -> a
211 f n 0 = n -- Strict in the Int, hence wrapper
212 f n x = f (n+n) (x-1)
215 g x = f x x -- Provokes a specialisation for f
223 Then we want the specialisation for 'f' to kick in before the wrapper does.
225 Now in fact the 'gentle' simplification pass encourages this, by
226 having rules on, but inlinings off. But that's kind of lucky. It seems
227 more robust to give the wrapper an Activation of (ActiveAfter 0),
228 so that it becomes active in an importing module at the same time that
229 it appears in the first place in the defining module.
233 -> Id -- The fn binder
234 -> CoreExpr -- The bound rhs; its innards
236 -> UniqSM [(Id, CoreExpr)] -- either *one* or *two* pairs;
237 -- if one, then no worker (only
238 -- the orig "wrapper" lives on);
239 -- if two, then a worker and a
241 tryWW is_rec fn_id rhs
242 | isNeverActive inline_act
243 -- No point in worker/wrappering if the thing is never inlined!
244 -- Because the no-inline prag will prevent the wrapper ever
245 -- being inlined at a call site.
247 -- Furthermore, don't even expose strictness info
248 = return [ (fn_id, rhs) ]
250 | is_thunk && worthSplittingThunk maybe_fn_dmd res_info
251 -- See Note [Thunk splitting]
252 = ASSERT2( isNonRec is_rec, ppr new_fn_id ) -- The thunk must be non-recursive
253 checkSize new_fn_id rhs $
254 splitThunk new_fn_id rhs
256 | is_fun && worthSplittingFun wrap_dmds res_info
257 = checkSize new_fn_id rhs $
258 splitFun new_fn_id fn_info wrap_dmds res_info rhs
261 = return [ (new_fn_id, rhs) ]
264 fn_info = idInfo fn_id
265 maybe_fn_dmd = demandInfo fn_info
266 inline_act = inlinePragmaActivation (inlinePragInfo fn_info)
268 -- In practice it always will have a strictness
269 -- signature, even if it's a uninformative one
270 strict_sig = strictnessInfo fn_info `orElse` topSig
271 StrictSig (DmdType env wrap_dmds res_info) = strict_sig
273 -- new_fn_id has the DmdEnv zapped.
274 -- (a) it is never used again
275 -- (b) it wastes space
276 -- (c) it becomes incorrect as things are cloned, because
277 -- we don't push the substitution into it
278 new_fn_id | isEmptyVarEnv env = fn_id
279 | otherwise = fn_id `setIdStrictness`
280 StrictSig (mkTopDmdType wrap_dmds res_info)
282 is_fun = notNull wrap_dmds
283 is_thunk = not is_fun && not (exprIsHNF rhs)
285 ---------------------
286 checkSize :: Id -> CoreExpr
287 -> UniqSM [(Id,CoreExpr)] -> UniqSM [(Id,CoreExpr)]
288 checkSize fn_id rhs thing_inside
289 | isStableUnfolding (realIdUnfolding fn_id)
290 = return [ (fn_id, rhs) ]
291 -- See Note [Don't w/w INLINABLE things]
292 -- and Note [Don't w/w INLINABLABLE things]
293 -- NB: use realIdUnfolding because we want to see the unfolding
294 -- even if it's a loop breaker!
296 | certainlyWillInline (idUnfolding fn_id)
297 = return [ (fn_id `setIdUnfolding` inline_rule, rhs) ]
298 -- Note [Don't w/w inline small non-loop-breaker things]
299 -- NB: use idUnfolding because we don't want to apply
300 -- this criterion to a loop breaker!
302 | otherwise = thing_inside
304 inline_rule = mkInlineUnfolding Nothing rhs
306 ---------------------
307 splitFun :: Id -> IdInfo -> [Demand] -> DmdResult -> Expr Var
308 -> UniqSM [(Id, CoreExpr)]
309 splitFun fn_id fn_info wrap_dmds res_info rhs
310 = WARN( not (wrap_dmds `lengthIs` arity), ppr fn_id <+> (ppr arity $$ ppr wrap_dmds $$ ppr res_info) )
312 -- The arity should match the signature
313 (work_demands, wrap_fn, work_fn) <- mkWwBodies fun_ty wrap_dmds res_info one_shots
314 ; work_uniq <- getUniqueM
316 work_rhs = work_fn rhs
317 work_id = mkWorkerId work_uniq fn_id (exprType work_rhs)
318 `setIdOccInfo` occInfo fn_info
319 -- Copy over occurrence info from parent
320 -- Notably whether it's a loop breaker
321 -- Doesn't matter much, since we will simplify next, but
322 -- seems right-er to do so
324 `setInlineActivation` (inlinePragmaActivation inl_prag)
325 -- Any inline activation (which sets when inlining is active)
326 -- on the original function is duplicated on the worker
327 -- It *matters* that the pragma stays on the wrapper
328 -- It seems sensible to have it on the worker too, although we
329 -- can't think of a compelling reason. (In ptic, INLINE things are
330 -- not w/wd). However, the RuleMatchInfo is not transferred since
331 -- it does not make sense for workers to be constructorlike.
333 `setIdStrictness` StrictSig (mkTopDmdType work_demands work_res_info)
334 -- Even though we may not be at top level,
335 -- it's ok to give it an empty DmdEnv
337 `setIdArity` (exprArity work_rhs)
338 -- Set the arity so that the Core Lint check that the
339 -- arity is consistent with the demand type goes through
341 wrap_rhs = wrap_fn work_id
342 wrap_prag = InlinePragma { inl_inline = Inline
344 , inl_act = ActiveAfter 0
345 , inl_rule = rule_match_info }
346 -- See Note [Wrapper activation]
347 -- The RuleMatchInfo is (and must be) unaffected
348 -- The inl_inline is bound to be False, else we would not be
351 wrap_id = fn_id `setIdUnfolding` mkWwInlineRule work_id wrap_rhs arity
352 `setInlinePragma` wrap_prag
353 `setIdOccInfo` NoOccInfo
354 -- Zap any loop-breaker-ness, to avoid bleating from Lint
355 -- about a loop breaker with an INLINE rule
357 ; return ([(work_id, work_rhs), (wrap_id, wrap_rhs)]) })
358 -- Worker first, because wrapper mentions it
359 -- mkWwBodies has already built a wrap_rhs with an INLINE pragma wrapped around it
361 fun_ty = idType fn_id
362 inl_prag = inlinePragInfo fn_info
363 rule_match_info = inlinePragmaRuleMatchInfo inl_prag
364 arity = arityInfo fn_info
365 -- The arity is set by the simplifier using exprEtaExpandArity
366 -- So it may be more than the number of top-level-visible lambdas
368 work_res_info | isBotRes res_info = BotRes -- Cpr stuff done by wrapper
371 one_shots = get_one_shots rhs
373 -- If the original function has one-shot arguments, it is important to
374 -- make the wrapper and worker have corresponding one-shot arguments too.
375 -- Otherwise we spuriously float stuff out of case-expression join points,
376 -- which is very annoying.
377 get_one_shots :: Expr Var -> [Bool]
378 get_one_shots (Lam b e)
379 | isId b = isOneShotLambda b : get_one_shots e
380 | otherwise = get_one_shots e
381 get_one_shots (Note _ e) = get_one_shots e
382 get_one_shots _ = noOneShotInfo
385 Note [Thunk splitting]
386 ~~~~~~~~~~~~~~~~~~~~~~
387 Suppose x is used strictly (never mind whether it has the CPR
394 splitThunk transforms like this:
397 x* = case x-rhs of { I# a -> I# a }
400 Now simplifier will transform to
403 I# a -> let x* = I# a
406 which is what we want. Now suppose x-rhs is itself a case:
408 x-rhs = case e of { T -> I# a; F -> I# b }
410 The join point will abstract over a, rather than over (which is
411 what would have happened before) which is fine.
413 Notice that x certainly has the CPR property now!
415 In fact, splitThunk uses the function argument w/w splitting
416 function, so that if x's demand is deeper (say U(U(L,L),L))
417 then the splitting will go deeper too.
420 -- See Note [Thunk splitting]
421 -- splitThunk converts the *non-recursive* binding
426 -- I# y -> let x = I# y in x }
427 -- See comments above. Is it not beautifully short?
429 splitThunk :: Var -> Expr Var -> UniqSM [(Var, Expr Var)]
430 splitThunk fn_id rhs = do
431 (_, wrap_fn, work_fn) <- mkWWstr [fn_id]
432 return [ (fn_id, Let (NonRec fn_id rhs) (wrap_fn (work_fn (Var fn_id)))) ]
436 %************************************************************************
438 \subsection{Functions over Demands}
440 %************************************************************************
443 worthSplittingFun :: [Demand] -> DmdResult -> Bool
444 -- True <=> the wrapper would not be an identity function
445 worthSplittingFun ds res
446 = any worth_it ds || returnsCPR res
447 -- worthSplitting returns False for an empty list of demands,
448 -- and hence do_strict_ww is False if arity is zero and there is no CPR
449 -- See Note [Worker-wrapper for bottoming functions]
451 worth_it Abs = True -- Absent arg
452 worth_it (Eval (Prod _)) = True -- Product arg to evaluate
455 worthSplittingThunk :: Maybe Demand -- Demand on the thunk
456 -> DmdResult -- CPR info for the thunk
458 worthSplittingThunk maybe_dmd res
459 = worth_it maybe_dmd || returnsCPR res
461 -- Split if the thing is unpacked
462 worth_it (Just (Eval (Prod ds))) = not (all isAbsent ds)
466 Note [Worker-wrapper for bottoming functions]
467 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
468 We used not to split if the result is bottom.
469 [Justification: there's no efficiency to be gained.]
471 But it's sometimes bad not to make a wrapper. Consider
472 fw = \x# -> let x = I# x# in case e of
476 The re-boxing code won't go away unless error_fn gets a wrapper too.
477 [We don't do reboxing now, but in general it's better to pass an
478 unboxed thing to f, and have it reboxed in the error cases....]
481 %************************************************************************
483 \subsection{The worker wrapper core}
485 %************************************************************************
487 @mkWrapper@ is called when importing a function. We have the type of
488 the function and the name of its worker, and we want to make its body (the wrapper).
491 mkWrapper :: Type -- Wrapper type
492 -> StrictSig -- Wrapper strictness info
493 -> UniqSM (Id -> CoreExpr) -- Wrapper body, missing worker Id
495 mkWrapper fun_ty (StrictSig (DmdType _ demands res_info)) = do
496 (_, wrap_fn, _) <- mkWwBodies fun_ty demands res_info noOneShotInfo
499 noOneShotInfo :: [Bool]
500 noOneShotInfo = repeat False