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, mkInlineRule, mkWwInlineRule )
11 import CoreUtils ( exprType, exprIsHNF )
12 import CoreArity ( exprArity )
14 import Id ( idType, isOneShotLambda, idUnfolding,
15 setIdStrictness, mkWorkerId, setInlinePragma,
16 setInlineActivation, setIdUnfolding,
20 import Demand ( Demand(..), StrictSig(..), DmdType(..), DmdResult(..),
21 Demands(..), mkTopDmdType, isBotRes, returnsCPR, topSig, isAbsent
24 import BasicTypes ( RecFlag(..), isNonRec, isNeverActive,
25 Activation(..), InlinePragma(..),
26 inlinePragmaActivation, inlinePragmaRuleMatchInfo )
27 import VarEnv ( isEmptyVarEnv )
28 import Maybes ( orElse )
30 import Util ( lengthIs, notNull )
34 #include "HsVersions.h"
37 We take Core bindings whose binders have:
41 \item Strictness attached (by the front-end of the strictness
44 \item Constructed Product Result information attached by the CPR
49 and we return some ``plain'' bindings which have been
50 worker/wrapper-ified, meaning:
54 \item Functions have been split into workers and wrappers where
55 appropriate. If a function has both strictness and CPR properties
56 then only one worker/wrapper doing both transformations is produced;
58 \item Binders' @IdInfos@ have been updated to reflect the existence of
59 these workers/wrappers (this is where we get STRICTNESS and CPR pragma
60 info for exported values).
64 wwTopBinds :: UniqSupply -> [CoreBind] -> [CoreBind]
66 wwTopBinds us top_binds
68 top_binds' <- mapM wwBind top_binds
69 return (concat top_binds')
72 %************************************************************************
74 \subsection[wwBind-wwExpr]{@wwBind@ and @wwExpr@}
76 %************************************************************************
78 @wwBind@ works on a binding, trying each \tr{(binder, expr)} pair in
79 turn. Non-recursive case first, then recursive...
83 -> UniqSM [CoreBind] -- returns a WwBinding intermediate form;
84 -- the caller will convert to Expr/Binding,
87 wwBind (NonRec binder rhs) = do
89 new_pairs <- tryWW NonRecursive binder new_rhs
90 return [NonRec b e | (b,e) <- new_pairs]
91 -- Generated bindings must be non-recursive
92 -- because the original binding was.
95 = return . Rec <$> concatMapM do_one pairs
97 do_one (binder, rhs) = do new_rhs <- wwExpr rhs
98 tryWW Recursive binder new_rhs
101 @wwExpr@ basically just walks the tree, looking for appropriate
102 annotations that can be used. Remember it is @wwBind@ that does the
103 matching by looking for strict arguments of the correct type.
104 @wwExpr@ is a version that just returns the ``Plain'' Tree.
107 wwExpr :: CoreExpr -> UniqSM CoreExpr
109 wwExpr e@(Type {}) = return e
110 wwExpr e@(Lit {}) = return e
111 wwExpr e@(Var {}) = return e
113 wwExpr (Lam binder expr)
114 = Lam binder <$> wwExpr expr
117 = App <$> wwExpr f <*> wwExpr a
119 wwExpr (Note note expr)
120 = Note note <$> wwExpr expr
122 wwExpr (Cast expr co) = do
123 new_expr <- wwExpr expr
124 return (Cast new_expr co)
126 wwExpr (Let bind expr)
127 = mkLets <$> wwBind bind <*> wwExpr expr
129 wwExpr (Case expr binder ty alts) = do
130 new_expr <- wwExpr expr
131 new_alts <- mapM ww_alt alts
132 return (Case new_expr binder ty new_alts)
134 ww_alt (con, binders, rhs) = do
135 new_rhs <- wwExpr rhs
136 return (con, binders, new_rhs)
139 %************************************************************************
141 \subsection[tryWW]{@tryWW@: attempt a worker/wrapper pair}
143 %************************************************************************
145 @tryWW@ just accumulates arguments, converts strictness info from the
146 front-end into the proper form, then calls @mkWwBodies@ to do
149 The only reason this is monadised is for the unique supply.
151 Note [Don't w/w inline things (a)]
152 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
154 It's very important to refrain from w/w-ing an INLINE function (ie one
155 with an InlineRule) because the wrapper will then overwrite the
156 InlineRule unfolding.
158 Furthermore, if the programmer has marked something as INLINE,
159 we may lose by w/w'ing it.
161 If the strictness analyser is run twice, this test also prevents
162 wrappers (which are INLINEd) from being re-done. (You can end up with
163 several liked-named Ids bouncing around at the same time---absolute
166 Notice that we refrain from w/w'ing an INLINE function even if it is
167 in a recursive group. It might not be the loop breaker. (We could
168 test for loop-breaker-hood, but I'm not sure that ever matters.)
170 Note [Don't w/w inline things (b)]
171 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
172 In general, we refrain from w/w-ing *small* functions, because they'll
173 inline anyway. But we must take care: it may look small now, but get
174 to be big later after other inling has happened. So we take the
175 precaution of adding an INLINE pragma to any such functions.
177 I made this change when I observed a big function at the end of
178 compilation with a useful strictness signature but no w-w. When
179 I measured it on nofib, it didn't make much difference; just a few
180 percent improved allocation on one benchmark (bspt/Euclid.space).
181 But nothing got worse.
183 Note [Wrapper activation]
184 ~~~~~~~~~~~~~~~~~~~~~~~~~
185 When should the wrapper inlining be active? It must not be active
186 earlier than the current Activation of the Id (eg it might have a
187 NOINLINE pragma). But in fact strictness analysis happens fairly
188 late in the pipeline, and we want to prioritise specialisations over
189 strictness. Eg if we have
191 f :: Num a => a -> Int -> a
192 f n 0 = n -- Strict in the Int, hence wrapper
193 f n x = f (n+n) (x-1)
196 g x = f x x -- Provokes a specialisation for f
204 Then we want the specialisation for 'f' to kick in before the wrapper does.
206 Now in fact the 'gentle' simplification pass encourages this, by
207 having rules on, but inlinings off. But that's kind of lucky. It seems
208 more robust to give the wrapper an Activation of (ActiveAfter 0),
209 so that it becomes active in an importing module at the same time that
210 it appears in the first place in the defining module.
214 -> Id -- The fn binder
215 -> CoreExpr -- The bound rhs; its innards
217 -> UniqSM [(Id, CoreExpr)] -- either *one* or *two* pairs;
218 -- if one, then no worker (only
219 -- the orig "wrapper" lives on);
220 -- if two, then a worker and a
222 tryWW is_rec fn_id rhs
223 | isNeverActive inline_act
224 -- No point in worker/wrappering if the thing is never inlined!
225 -- Because the no-inline prag will prevent the wrapper ever
226 -- being inlined at a call site.
228 -- Furthermore, don't even expose strictness info
229 = return [ (fn_id, rhs) ]
231 | is_thunk && worthSplittingThunk maybe_fn_dmd res_info
232 = ASSERT2( isNonRec is_rec, ppr new_fn_id ) -- The thunk must be non-recursive
233 checkSize new_fn_id rhs $
234 splitThunk new_fn_id rhs
236 | is_fun && worthSplittingFun wrap_dmds res_info
237 = checkSize new_fn_id rhs $
238 splitFun new_fn_id fn_info wrap_dmds res_info rhs
241 = return [ (new_fn_id, rhs) ]
244 fn_info = idInfo fn_id
245 maybe_fn_dmd = demandInfo fn_info
246 inline_act = inlinePragmaActivation (inlinePragInfo fn_info)
248 -- In practice it always will have a strictness
249 -- signature, even if it's a uninformative one
250 strict_sig = strictnessInfo fn_info `orElse` topSig
251 StrictSig (DmdType env wrap_dmds res_info) = strict_sig
253 -- new_fn_id has the DmdEnv zapped.
254 -- (a) it is never used again
255 -- (b) it wastes space
256 -- (c) it becomes incorrect as things are cloned, because
257 -- we don't push the substitution into it
258 new_fn_id | isEmptyVarEnv env = fn_id
259 | otherwise = fn_id `setIdStrictness`
260 StrictSig (mkTopDmdType wrap_dmds res_info)
262 is_fun = notNull wrap_dmds
263 is_thunk = not is_fun && not (exprIsHNF rhs)
265 ---------------------
266 checkSize :: Id -> CoreExpr
267 -> UniqSM [(Id,CoreExpr)] -> UniqSM [(Id,CoreExpr)]
268 -- See Note [Don't w/w inline things (a) and (b)]
269 checkSize fn_id rhs thing_inside
270 | isStableUnfolding unfolding -- For DFuns and INLINE things, leave their
271 = return [ (fn_id, rhs) ] -- unfolding unchanged; but still attach
272 -- strictness info to the Id
274 | certainlyWillInline unfolding
275 = return [ (fn_id `setIdUnfolding` inline_rule, rhs) ]
276 -- Note [Don't w/w inline things (b)]
278 | otherwise = thing_inside
280 unfolding = idUnfolding fn_id
281 inline_rule = mkInlineRule unSaturatedOk rhs (unfoldingArity unfolding)
283 ---------------------
284 splitFun :: Id -> IdInfo -> [Demand] -> DmdResult -> Expr Var
285 -> UniqSM [(Id, CoreExpr)]
286 splitFun fn_id fn_info wrap_dmds res_info rhs
287 = WARN( not (wrap_dmds `lengthIs` arity), ppr fn_id <+> (ppr arity $$ ppr wrap_dmds $$ ppr res_info) )
289 -- The arity should match the signature
290 (work_demands, wrap_fn, work_fn) <- mkWwBodies fun_ty wrap_dmds res_info one_shots
291 ; work_uniq <- getUniqueM
293 work_rhs = work_fn rhs
294 work_id = mkWorkerId work_uniq fn_id (exprType work_rhs)
295 `setInlineActivation` (inlinePragmaActivation inl_prag)
296 -- Any inline activation (which sets when inlining is active)
297 -- on the original function is duplicated on the worker
298 -- It *matters* that the pragma stays on the wrapper
299 -- It seems sensible to have it on the worker too, although we
300 -- can't think of a compelling reason. (In ptic, INLINE things are
301 -- not w/wd). However, the RuleMatchInfo is not transferred since
302 -- it does not make sense for workers to be constructorlike.
304 `setIdStrictness` StrictSig (mkTopDmdType work_demands work_res_info)
305 -- Even though we may not be at top level,
306 -- it's ok to give it an empty DmdEnv
308 `setIdArity` (exprArity work_rhs)
309 -- Set the arity so that the Core Lint check that the
310 -- arity is consistent with the demand type goes through
312 wrap_rhs = wrap_fn work_id
313 wrap_prag = InlinePragma { inl_inline = True
314 , inl_act = ActiveAfter 0
315 , inl_rule = rule_match_info }
317 wrap_id = fn_id `setIdUnfolding` mkWwInlineRule work_id wrap_rhs arity
318 `setInlinePragma` wrap_prag
319 -- See Note [Wrapper activation]
320 -- The RuleMatchInfo is (and must be) unaffected
321 -- The inl_inline is bound to be False, else we would not be
324 ; return ([(work_id, work_rhs), (wrap_id, wrap_rhs)]) })
325 -- Worker first, because wrapper mentions it
326 -- mkWwBodies has already built a wrap_rhs with an INLINE pragma wrapped around it
328 fun_ty = idType fn_id
329 inl_prag = inlinePragInfo fn_info
330 rule_match_info = inlinePragmaRuleMatchInfo inl_prag
331 arity = arityInfo fn_info
332 -- The arity is set by the simplifier using exprEtaExpandArity
333 -- So it may be more than the number of top-level-visible lambdas
335 work_res_info | isBotRes res_info = BotRes -- Cpr stuff done by wrapper
338 one_shots = get_one_shots rhs
340 -- If the original function has one-shot arguments, it is important to
341 -- make the wrapper and worker have corresponding one-shot arguments too.
342 -- Otherwise we spuriously float stuff out of case-expression join points,
343 -- which is very annoying.
344 get_one_shots :: Expr Var -> [Bool]
345 get_one_shots (Lam b e)
346 | isId b = isOneShotLambda b : get_one_shots e
347 | otherwise = get_one_shots e
348 get_one_shots (Note _ e) = get_one_shots e
349 get_one_shots _ = noOneShotInfo
354 Suppose x is used strictly (never mind whether it has the CPR
361 splitThunk transforms like this:
364 x* = case x-rhs of { I# a -> I# a }
367 Now simplifier will transform to
370 I# a -> let x* = I# a
373 which is what we want. Now suppose x-rhs is itself a case:
375 x-rhs = case e of { T -> I# a; F -> I# b }
377 The join point will abstract over a, rather than over (which is
378 what would have happened before) which is fine.
380 Notice that x certainly has the CPR property now!
382 In fact, splitThunk uses the function argument w/w splitting
383 function, so that if x's demand is deeper (say U(U(L,L),L))
384 then the splitting will go deeper too.
387 -- splitThunk converts the *non-recursive* binding
392 -- I# y -> let x = I# y in x }
393 -- See comments above. Is it not beautifully short?
395 splitThunk :: Var -> Expr Var -> UniqSM [(Var, Expr Var)]
396 splitThunk fn_id rhs = do
397 (_, wrap_fn, work_fn) <- mkWWstr [fn_id]
398 return [ (fn_id, Let (NonRec fn_id rhs) (wrap_fn (work_fn (Var fn_id)))) ]
402 %************************************************************************
404 \subsection{Functions over Demands}
406 %************************************************************************
409 worthSplittingFun :: [Demand] -> DmdResult -> Bool
410 -- True <=> the wrapper would not be an identity function
411 worthSplittingFun ds res
412 = any worth_it ds || returnsCPR res
413 -- worthSplitting returns False for an empty list of demands,
414 -- and hence do_strict_ww is False if arity is zero and there is no CPR
415 -- See Note [Worker-wrapper for bottoming functions]
417 worth_it Abs = True -- Absent arg
418 worth_it (Eval (Prod _)) = True -- Product arg to evaluate
421 worthSplittingThunk :: Maybe Demand -- Demand on the thunk
422 -> DmdResult -- CPR info for the thunk
424 worthSplittingThunk maybe_dmd res
425 = worth_it maybe_dmd || returnsCPR res
427 -- Split if the thing is unpacked
428 worth_it (Just (Eval (Prod ds))) = not (all isAbsent ds)
432 Note [Worker-wrapper for bottoming functions]
433 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
434 We used not to split if the result is bottom.
435 [Justification: there's no efficiency to be gained.]
437 But it's sometimes bad not to make a wrapper. Consider
438 fw = \x# -> let x = I# x# in case e of
442 The re-boxing code won't go away unless error_fn gets a wrapper too.
443 [We don't do reboxing now, but in general it's better to pass an
444 unboxed thing to f, and have it reboxed in the error cases....]
447 %************************************************************************
449 \subsection{The worker wrapper core}
451 %************************************************************************
453 @mkWrapper@ is called when importing a function. We have the type of
454 the function and the name of its worker, and we want to make its body (the wrapper).
457 mkWrapper :: Type -- Wrapper type
458 -> StrictSig -- Wrapper strictness info
459 -> UniqSM (Id -> CoreExpr) -- Wrapper body, missing worker Id
461 mkWrapper fun_ty (StrictSig (DmdType _ demands res_info)) = do
462 (_, wrap_fn, _) <- mkWwBodies fun_ty demands res_info noOneShotInfo
465 noOneShotInfo :: [Bool]
466 noOneShotInfo = repeat False