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 )
17 import Demand ( Demand(..), StrictSig(..), DmdType(..), DmdResult(..),
18 Demands(..), mkTopDmdType, isBotRes, returnsCPR, topSig, isAbsent
21 import BasicTypes ( RecFlag(..), isNonRec, isNeverActive,
22 Activation(..), InlinePragma(..),
23 inlinePragmaActivation, inlinePragmaRuleMatchInfo )
24 import VarEnv ( isEmptyVarEnv )
25 import Maybes ( orElse )
27 import Util ( lengthIs, notNull )
31 #include "HsVersions.h"
34 We take Core bindings whose binders have:
38 \item Strictness attached (by the front-end of the strictness
41 \item Constructed Product Result information attached by the CPR
46 and we return some ``plain'' bindings which have been
47 worker/wrapper-ified, meaning:
51 \item Functions have been split into workers and wrappers where
52 appropriate. If a function has both strictness and CPR properties
53 then only one worker/wrapper doing both transformations is produced;
55 \item Binders' @IdInfos@ have been updated to reflect the existence of
56 these workers/wrappers (this is where we get STRICTNESS and CPR pragma
57 info for exported values).
61 wwTopBinds :: UniqSupply -> [CoreBind] -> [CoreBind]
63 wwTopBinds us top_binds
65 top_binds' <- mapM wwBind top_binds
66 return (concat top_binds')
69 %************************************************************************
71 \subsection[wwBind-wwExpr]{@wwBind@ and @wwExpr@}
73 %************************************************************************
75 @wwBind@ works on a binding, trying each \tr{(binder, expr)} pair in
76 turn. Non-recursive case first, then recursive...
80 -> UniqSM [CoreBind] -- returns a WwBinding intermediate form;
81 -- the caller will convert to Expr/Binding,
84 wwBind (NonRec binder rhs) = do
86 new_pairs <- tryWW NonRecursive binder new_rhs
87 return [NonRec b e | (b,e) <- new_pairs]
88 -- Generated bindings must be non-recursive
89 -- because the original binding was.
92 = return . Rec <$> concatMapM do_one pairs
94 do_one (binder, rhs) = do new_rhs <- wwExpr rhs
95 tryWW Recursive binder new_rhs
98 @wwExpr@ basically just walks the tree, looking for appropriate
99 annotations that can be used. Remember it is @wwBind@ that does the
100 matching by looking for strict arguments of the correct type.
101 @wwExpr@ is a version that just returns the ``Plain'' Tree.
104 wwExpr :: CoreExpr -> UniqSM CoreExpr
106 wwExpr e@(Type {}) = return e
107 wwExpr e@(Lit {}) = return e
108 wwExpr e@(Var {}) = return e
110 wwExpr (Lam binder expr)
111 = Lam binder <$> wwExpr expr
114 = App <$> wwExpr f <*> wwExpr a
116 wwExpr (Note note expr)
117 = Note note <$> wwExpr expr
119 wwExpr (Cast expr co) = do
120 new_expr <- wwExpr expr
121 return (Cast new_expr co)
123 wwExpr (Let bind expr)
124 = mkLets <$> wwBind bind <*> wwExpr expr
126 wwExpr (Case expr binder ty alts) = do
127 new_expr <- wwExpr expr
128 new_alts <- mapM ww_alt alts
129 return (Case new_expr binder ty new_alts)
131 ww_alt (con, binders, rhs) = do
132 new_rhs <- wwExpr rhs
133 return (con, binders, new_rhs)
136 %************************************************************************
138 \subsection[tryWW]{@tryWW@: attempt a worker/wrapper pair}
140 %************************************************************************
142 @tryWW@ just accumulates arguments, converts strictness info from the
143 front-end into the proper form, then calls @mkWwBodies@ to do
146 The only reason this is monadised is for the unique supply.
148 Note [Don't w/w inline things (a)]
149 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
151 It's very important to refrain from w/w-ing an INLINE function (ie one
152 with an InlineRule) because the wrapper will then overwrite the
153 InlineRule unfolding.
155 Furthermore, if the programmer has marked something as INLINE,
156 we may lose by w/w'ing it.
158 If the strictness analyser is run twice, this test also prevents
159 wrappers (which are INLINEd) from being re-done. (You can end up with
160 several liked-named Ids bouncing around at the same time---absolute
163 Notice that we refrain from w/w'ing an INLINE function even if it is
164 in a recursive group. It might not be the loop breaker. (We could
165 test for loop-breaker-hood, but I'm not sure that ever matters.)
167 Note [Don't w/w inline things (b)]
168 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
169 In general, we refrain from w/w-ing *small* functions, because they'll
170 inline anyway. But we must take care: it may look small now, but get
171 to be big later after other inling has happened. So we take the
172 precaution of adding an INLINE pragma to any such functions.
174 I made this change when I observed a big function at the end of
175 compilation with a useful strictness signature but no w-w. When
176 I measured it on nofib, it didn't make much difference; just a few
177 percent improved allocation on one benchmark (bspt/Euclid.space).
178 But nothing got worse.
180 Note [Wrapper activation]
181 ~~~~~~~~~~~~~~~~~~~~~~~~~
182 When should the wrapper inlining be active? It must not be active
183 earlier than the current Activation of the Id (eg it might have a
184 NOINLINE pragma). But in fact strictness analysis happens fairly
185 late in the pipeline, and we want to prioritise specialisations over
186 strictness. Eg if we have
188 f :: Num a => a -> Int -> a
189 f n 0 = n -- Strict in the Int, hence wrapper
190 f n x = f (n+n) (x-1)
193 g x = f x x -- Provokes a specialisation for f
201 Then we want the specialisation for 'f' to kick in before the wrapper does.
203 Now in fact the 'gentle' simplification pass encourages this, by
204 having rules on, but inlinings off. But that's kind of lucky. It seems
205 more robust to give the wrapper an Activation of (ActiveAfter 0),
206 so that it becomes active in an importing module at the same time that
207 it appears in the first place in the defining module.
211 -> Id -- The fn binder
212 -> CoreExpr -- The bound rhs; its innards
214 -> UniqSM [(Id, CoreExpr)] -- either *one* or *two* pairs;
215 -- if one, then no worker (only
216 -- the orig "wrapper" lives on);
217 -- if two, then a worker and a
219 tryWW is_rec fn_id rhs
220 | isNeverActive inline_act
221 -- No point in worker/wrappering if the thing is never inlined!
222 -- Because the no-inline prag will prevent the wrapper ever
223 -- being inlined at a call site.
225 -- Furthermore, don't even expose strictness info
226 = return [ (fn_id, rhs) ]
228 | is_thunk && worthSplittingThunk maybe_fn_dmd res_info
229 = ASSERT2( isNonRec is_rec, ppr new_fn_id ) -- The thunk must be non-recursive
230 checkSize new_fn_id rhs $
231 splitThunk new_fn_id rhs
233 | is_fun && worthSplittingFun wrap_dmds res_info
234 = checkSize new_fn_id rhs $
235 splitFun new_fn_id fn_info wrap_dmds res_info rhs
238 = return [ (new_fn_id, rhs) ]
241 fn_info = idInfo fn_id
242 maybe_fn_dmd = demandInfo fn_info
243 inline_act = inlinePragmaActivation (inlinePragInfo fn_info)
245 -- In practice it always will have a strictness
246 -- signature, even if it's a uninformative one
247 strict_sig = strictnessInfo fn_info `orElse` topSig
248 StrictSig (DmdType env wrap_dmds res_info) = strict_sig
250 -- new_fn_id has the DmdEnv zapped.
251 -- (a) it is never used again
252 -- (b) it wastes space
253 -- (c) it becomes incorrect as things are cloned, because
254 -- we don't push the substitution into it
255 new_fn_id | isEmptyVarEnv env = fn_id
256 | otherwise = fn_id `setIdStrictness`
257 StrictSig (mkTopDmdType wrap_dmds res_info)
259 is_fun = notNull wrap_dmds
260 is_thunk = not is_fun && not (exprIsHNF rhs)
262 ---------------------
263 checkSize :: Id -> CoreExpr
264 -> UniqSM [(Id,CoreExpr)] -> UniqSM [(Id,CoreExpr)]
265 -- See Note [Don't w/w inline things (a) and (b)]
266 checkSize fn_id rhs thing_inside
267 | isStableUnfolding unfolding -- For DFuns and INLINE things, leave their
268 = return [ (fn_id, rhs) ] -- unfolding unchanged; but still attach
269 -- strictness info to the Id
271 | certainlyWillInline unfolding
272 = return [ (fn_id `setIdUnfolding` inline_rule, rhs) ]
273 -- Note [Don't w/w inline things (b)]
275 | otherwise = thing_inside
277 unfolding = idUnfolding fn_id
278 inline_rule = mkInlineRule rhs Nothing
280 ---------------------
281 splitFun :: Id -> IdInfo -> [Demand] -> DmdResult -> Expr Var
282 -> UniqSM [(Id, CoreExpr)]
283 splitFun fn_id fn_info wrap_dmds res_info rhs
284 = WARN( not (wrap_dmds `lengthIs` arity), ppr fn_id <+> (ppr arity $$ ppr wrap_dmds $$ ppr res_info) )
286 -- The arity should match the signature
287 (work_demands, wrap_fn, work_fn) <- mkWwBodies fun_ty wrap_dmds res_info one_shots
288 ; work_uniq <- getUniqueM
290 work_rhs = work_fn rhs
291 work_id = mkWorkerId work_uniq fn_id (exprType work_rhs)
292 `setIdOccInfo` occInfo fn_info
293 -- Copy over occurrence info from parent
294 -- Notably whether it's a loop breaker
295 -- Doesn't matter much, since we will simplify next, but
296 -- seems right-er to do so
298 `setInlineActivation` (inlinePragmaActivation inl_prag)
299 -- Any inline activation (which sets when inlining is active)
300 -- on the original function is duplicated on the worker
301 -- It *matters* that the pragma stays on the wrapper
302 -- It seems sensible to have it on the worker too, although we
303 -- can't think of a compelling reason. (In ptic, INLINE things are
304 -- not w/wd). However, the RuleMatchInfo is not transferred since
305 -- it does not make sense for workers to be constructorlike.
307 `setIdStrictness` StrictSig (mkTopDmdType work_demands work_res_info)
308 -- Even though we may not be at top level,
309 -- it's ok to give it an empty DmdEnv
311 `setIdArity` (exprArity work_rhs)
312 -- Set the arity so that the Core Lint check that the
313 -- arity is consistent with the demand type goes through
315 wrap_rhs = wrap_fn work_id
316 wrap_prag = InlinePragma { inl_inline = True
318 , inl_act = ActiveAfter 0
319 , inl_rule = rule_match_info }
320 -- See Note [Wrapper activation]
321 -- The RuleMatchInfo is (and must be) unaffected
322 -- The inl_inline is bound to be False, else we would not be
325 wrap_id = fn_id `setIdUnfolding` mkWwInlineRule work_id wrap_rhs arity
326 `setInlinePragma` wrap_prag
327 `setIdOccInfo` NoOccInfo
328 -- Zap any loop-breaker-ness, to avoid bleating from Lint
329 -- about a loop breaker with an INLINE rule
331 ; return ([(work_id, work_rhs), (wrap_id, wrap_rhs)]) })
332 -- Worker first, because wrapper mentions it
333 -- mkWwBodies has already built a wrap_rhs with an INLINE pragma wrapped around it
335 fun_ty = idType fn_id
336 inl_prag = inlinePragInfo fn_info
337 rule_match_info = inlinePragmaRuleMatchInfo inl_prag
338 arity = arityInfo fn_info
339 -- The arity is set by the simplifier using exprEtaExpandArity
340 -- So it may be more than the number of top-level-visible lambdas
342 work_res_info | isBotRes res_info = BotRes -- Cpr stuff done by wrapper
345 one_shots = get_one_shots rhs
347 -- If the original function has one-shot arguments, it is important to
348 -- make the wrapper and worker have corresponding one-shot arguments too.
349 -- Otherwise we spuriously float stuff out of case-expression join points,
350 -- which is very annoying.
351 get_one_shots :: Expr Var -> [Bool]
352 get_one_shots (Lam b e)
353 | isId b = isOneShotLambda b : get_one_shots e
354 | otherwise = get_one_shots e
355 get_one_shots (Note _ e) = get_one_shots e
356 get_one_shots _ = noOneShotInfo
361 Suppose x is used strictly (never mind whether it has the CPR
368 splitThunk transforms like this:
371 x* = case x-rhs of { I# a -> I# a }
374 Now simplifier will transform to
377 I# a -> let x* = I# a
380 which is what we want. Now suppose x-rhs is itself a case:
382 x-rhs = case e of { T -> I# a; F -> I# b }
384 The join point will abstract over a, rather than over (which is
385 what would have happened before) which is fine.
387 Notice that x certainly has the CPR property now!
389 In fact, splitThunk uses the function argument w/w splitting
390 function, so that if x's demand is deeper (say U(U(L,L),L))
391 then the splitting will go deeper too.
394 -- splitThunk converts the *non-recursive* binding
399 -- I# y -> let x = I# y in x }
400 -- See comments above. Is it not beautifully short?
402 splitThunk :: Var -> Expr Var -> UniqSM [(Var, Expr Var)]
403 splitThunk fn_id rhs = do
404 (_, wrap_fn, work_fn) <- mkWWstr [fn_id]
405 return [ (fn_id, Let (NonRec fn_id rhs) (wrap_fn (work_fn (Var fn_id)))) ]
409 %************************************************************************
411 \subsection{Functions over Demands}
413 %************************************************************************
416 worthSplittingFun :: [Demand] -> DmdResult -> Bool
417 -- True <=> the wrapper would not be an identity function
418 worthSplittingFun ds res
419 = any worth_it ds || returnsCPR res
420 -- worthSplitting returns False for an empty list of demands,
421 -- and hence do_strict_ww is False if arity is zero and there is no CPR
422 -- See Note [Worker-wrapper for bottoming functions]
424 worth_it Abs = True -- Absent arg
425 worth_it (Eval (Prod _)) = True -- Product arg to evaluate
428 worthSplittingThunk :: Maybe Demand -- Demand on the thunk
429 -> DmdResult -- CPR info for the thunk
431 worthSplittingThunk maybe_dmd res
432 = worth_it maybe_dmd || returnsCPR res
434 -- Split if the thing is unpacked
435 worth_it (Just (Eval (Prod ds))) = not (all isAbsent ds)
439 Note [Worker-wrapper for bottoming functions]
440 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
441 We used not to split if the result is bottom.
442 [Justification: there's no efficiency to be gained.]
444 But it's sometimes bad not to make a wrapper. Consider
445 fw = \x# -> let x = I# x# in case e of
449 The re-boxing code won't go away unless error_fn gets a wrapper too.
450 [We don't do reboxing now, but in general it's better to pass an
451 unboxed thing to f, and have it reboxed in the error cases....]
454 %************************************************************************
456 \subsection{The worker wrapper core}
458 %************************************************************************
460 @mkWrapper@ is called when importing a function. We have the type of
461 the function and the name of its worker, and we want to make its body (the wrapper).
464 mkWrapper :: Type -- Wrapper type
465 -> StrictSig -- Wrapper strictness info
466 -> UniqSM (Id -> CoreExpr) -- Wrapper body, missing worker Id
468 mkWrapper fun_ty (StrictSig (DmdType _ demands res_info)) = do
469 (_, wrap_fn, _) <- mkWwBodies fun_ty demands res_info noOneShotInfo
472 noOneShotInfo :: [Bool]
473 noOneShotInfo = repeat False