2 % (c) The GRASP/AQUA Project, Glasgow University, 1993-1998
4 \section[WwLib]{A library for the ``worker/wrapper'' back-end to the strictness analyser}
7 module WwLib ( mkWwBodies, mkWWstr, mkWorkerArgs ) where
9 #include "HsVersions.h"
12 import CoreUtils ( exprType )
13 import Id ( Id, idType, mkSysLocal, idNewDemandInfo, setIdNewDemandInfo,
14 isOneShotLambda, setOneShotLambda, setIdUnfolding,
17 import IdInfo ( vanillaIdInfo )
19 import NewDemand ( Demand(..), DmdResult(..), Demands(..) )
20 import MkId ( realWorldPrimId, voidArgId, mkRuntimeErrorApp, rUNTIME_ERROR_ID,
21 mkUnpackCase, mkProductBox )
22 import TysWiredIn ( tupleCon )
24 import Coercion ( mkSymCoercion, splitNewTypeRepCo_maybe )
25 import BasicTypes ( Boxity(..) )
26 import Var ( Var, isId )
27 import UniqSupply ( returnUs, thenUs, getUniquesUs, UniqSM )
29 import Util ( zipWithEqual, notNull )
31 import List ( zipWith4 )
35 %************************************************************************
37 \subsection[mkWrapperAndWorker]{@mkWrapperAndWorker@}
39 %************************************************************************
41 Here's an example. The original function is:
44 g :: forall a . Int -> [a] -> a
52 From this, we want to produce:
54 -- wrapper (an unfolding)
55 g :: forall a . Int -> [a] -> a
60 -- call the worker; don't forget the type args!
63 $wg :: forall a . Int# -> [a] -> a
65 $wg = /\ a -> \ x# ys ->
69 case x of -- note: body of g moved intact
74 Something we have to be careful about: Here's an example:
77 -- "f" strictness: U(P)U(P)
78 f (I# a) (I# b) = a +# b
80 g = f -- "g" strictness same as "f"
83 \tr{f} will get a worker all nice and friendly-like; that's good.
84 {\em But we don't want a worker for \tr{g}}, even though it has the
85 same strictness as \tr{f}. Doing so could break laziness, at best.
87 Consequently, we insist that the number of strictness-info items is
88 exactly the same as the number of lambda-bound arguments. (This is
89 probably slightly paranoid, but OK in practice.) If it isn't the
90 same, we ``revise'' the strictness info, so that we won't propagate
91 the unusable strictness-info into the interfaces.
94 %************************************************************************
96 \subsection{The worker wrapper core}
98 %************************************************************************
100 @mkWwBodies@ is called when doing the worker/wrapper split inside a module.
103 mkWwBodies :: Type -- Type of original function
104 -> [Demand] -- Strictness of original function
105 -> DmdResult -- Info about function result
106 -> [Bool] -- One-shot-ness of the function
107 -> UniqSM ([Demand], -- Demands for worker (value) args
108 Id -> CoreExpr, -- Wrapper body, lacking only the worker Id
109 CoreExpr -> CoreExpr) -- Worker body, lacking the original function rhs
111 -- wrap_fn_args E = \x y -> E
112 -- work_fn_args E = E x y
114 -- wrap_fn_str E = case x of { (a,b) ->
115 -- case a of { (a1,a2) ->
117 -- work_fn_str E = \a2 a2 b y ->
118 -- let a = (a1,a2) in
122 mkWwBodies fun_ty demands res_info one_shots
123 = mkWWargs fun_ty demands one_shots' `thenUs` \ (wrap_args, wrap_fn_args, work_fn_args, res_ty) ->
124 mkWWstr wrap_args `thenUs` \ (work_args, wrap_fn_str, work_fn_str) ->
126 (work_lam_args, work_call_args) = mkWorkerArgs work_args res_ty
128 -- Don't do CPR if the worker doesn't have any value arguments
129 -- Then the worker is just a constant, so we don't want to unbox it.
130 (if any isId work_args then
131 mkWWcpr res_ty res_info
133 returnUs (id, id, res_ty)
134 ) `thenUs` \ (wrap_fn_cpr, work_fn_cpr, _cpr_res_ty) ->
136 returnUs ([idNewDemandInfo v | v <- work_call_args, isId v],
137 Note InlineMe . wrap_fn_args . wrap_fn_cpr . wrap_fn_str . applyToVars work_call_args . Var,
138 mkLams work_lam_args. work_fn_str . work_fn_cpr . work_fn_args)
139 -- We use an INLINE unconditionally, even if the wrapper turns out to be
140 -- something trivial like
142 -- f = __inline__ (coerce T fw)
143 -- The point is to propagate the coerce to f's call sites, so even though
144 -- f's RHS is now trivial (size 1) we still want the __inline__ to prevent
145 -- fw from being inlined into f's RHS
147 one_shots' = one_shots ++ repeat False
151 %************************************************************************
153 \subsection{Making wrapper args}
155 %************************************************************************
157 During worker-wrapper stuff we may end up with an unlifted thing
158 which we want to let-bind without losing laziness. So we
159 add a void argument. E.g.
161 f = /\a -> \x y z -> E::Int# -- E does not mention x,y,z
163 fw = /\ a -> \void -> E
164 f = /\ a -> \x y z -> fw realworld
166 We use the state-token type which generates no code.
169 mkWorkerArgs :: [Var]
170 -> Type -- Type of body
171 -> ([Var], -- Lambda bound args
172 [Var]) -- Args at call site
173 mkWorkerArgs args res_ty
174 | any isId args || not (isUnLiftedType res_ty)
177 = (args ++ [voidArgId], args ++ [realWorldPrimId])
181 %************************************************************************
183 \subsection{Coercion stuff}
185 %************************************************************************
188 We really want to "look through" coerces.
189 Reason: I've seen this situation:
191 let f = coerce T (\s -> E)
197 If only we w/w'd f, we'd get
198 let f = coerce T (\s -> fw s)
202 Now we'll inline f to get
210 Now we'll see that fw has arity 1, and will arity expand
211 the \x to get what we want.
214 -- mkWWargs is driven off the function type and arity.
215 -- It chomps bites off foralls, arrows, newtypes
216 -- and keeps repeating that until it's satisfied the supplied arity
220 -> [Bool] -- True for a one-shot arg; ** may be infinite **
221 -> UniqSM ([Var], -- Wrapper args
222 CoreExpr -> CoreExpr, -- Wrapper fn
223 CoreExpr -> CoreExpr, -- Worker fn
224 Type) -- Type of wrapper body
226 mkWWargs fun_ty demands one_shots
227 | Just (rep_ty, co) <- splitNewTypeRepCo_maybe fun_ty
228 -- The newtype case is for when the function has
229 -- a recursive newtype after the arrow (rare)
230 -- We check for arity >= 0 to avoid looping in the case
231 -- of a function whose type is, in effect, infinite
232 -- [Arity is driven by looking at the term, not just the type.]
234 -- It's also important when we have a function returning (say) a pair
235 -- wrapped in a recursive newtype, at least if CPR analysis can look
236 -- through such newtypes, which it probably can since they are
238 = mkWWargs rep_ty demands one_shots `thenUs` \ (wrap_args, wrap_fn_args, work_fn_args, res_ty) ->
240 \ e -> Cast (wrap_fn_args e) (mkSymCoercion co),
241 \ e -> work_fn_args (Cast e co),
244 = getUniquesUs `thenUs` \ wrap_uniqs ->
246 (tyvars, tau) = splitForAllTys fun_ty
247 (arg_tys, body_ty) = splitFunTys tau
249 n_demands = length demands
250 n_arg_tys = length arg_tys
251 n_args = n_demands `min` n_arg_tys
253 new_fun_ty = mkFunTys (drop n_demands arg_tys) body_ty
254 new_demands = drop n_arg_tys demands
255 new_one_shots = drop n_args one_shots
257 val_args = zipWith4 mk_wrap_arg wrap_uniqs arg_tys demands one_shots
258 wrap_args = tyvars ++ val_args
260 {- ASSERT( notNull tyvars || notNull arg_tys ) -}
261 if (null tyvars) && (null arg_tys) then
262 pprTrace "mkWWargs" (ppr fun_ty $$ ppr demands)
263 returnUs ([], id, id, fun_ty)
268 new_one_shots `thenUs` \ (more_wrap_args, wrap_fn_args, work_fn_args, res_ty) ->
270 returnUs (wrap_args ++ more_wrap_args,
271 mkLams wrap_args . wrap_fn_args,
272 work_fn_args . applyToVars wrap_args,
276 = returnUs ([], id, id, fun_ty)
279 applyToVars :: [Var] -> CoreExpr -> CoreExpr
280 applyToVars vars fn = mkVarApps fn vars
282 mk_wrap_arg :: Unique -> Type -> NewDemand.Demand -> Bool -> Id
283 mk_wrap_arg uniq ty dmd one_shot
284 = set_one_shot one_shot (setIdNewDemandInfo (mkSysLocal FSLIT("w") uniq ty) dmd)
286 set_one_shot True id = setOneShotLambda id
287 set_one_shot False id = id
291 %************************************************************************
293 \subsection{Strictness stuff}
295 %************************************************************************
298 mkWWstr :: [Var] -- Wrapper args; have their demand info on them
299 -- *Includes type variables*
300 -> UniqSM ([Var], -- Worker args
301 CoreExpr -> CoreExpr, -- Wrapper body, lacking the worker call
302 -- and without its lambdas
303 -- This fn adds the unboxing
305 CoreExpr -> CoreExpr) -- Worker body, lacking the original body of the function,
306 -- and lacking its lambdas.
307 -- This fn does the reboxing
309 = returnUs ([], nop_fn, nop_fn)
312 = mkWWstr_one arg `thenUs` \ (args1, wrap_fn1, work_fn1) ->
313 mkWWstr args `thenUs` \ (args2, wrap_fn2, work_fn2) ->
314 returnUs (args1 ++ args2, wrap_fn1 . wrap_fn2, work_fn1 . work_fn2)
316 ----------------------
317 -- mkWWstr_one wrap_arg = (work_args, wrap_fn, work_fn)
318 -- * wrap_fn assumes wrap_arg is in scope,
319 -- brings into scope work_args (via cases)
320 -- * work_fn assumes work_args are in scope, a
321 -- brings into scope wrap_arg (via lets)
322 mkWWstr_one :: Var -> UniqSM ([Var], CoreExpr -> CoreExpr, CoreExpr -> CoreExpr)
325 = returnUs ([arg], nop_fn, nop_fn)
328 = case idNewDemandInfo arg of
330 -- Absent case. We don't deal with absence for unlifted types,
331 -- though, because it's not so easy to manufacture a placeholder
332 -- We'll see if this turns out to be a problem
333 Abs | not (isUnLiftedType (idType arg)) ->
334 returnUs ([], nop_fn, mk_absent_let arg)
338 | Just (_arg_tycon, _tycon_arg_tys, data_con, inst_con_arg_tys)
339 <- deepSplitProductType_maybe (idType arg)
340 -> getUniquesUs `thenUs` \ uniqs ->
342 unpk_args = zipWith mk_ww_local uniqs inst_con_arg_tys
343 unpk_args_w_ds = zipWithEqual "mkWWstr" set_worker_arg_info unpk_args cs
344 unbox_fn = mkUnpackCase (sanitiseCaseBndr arg) (Var arg) unpk_args data_con
345 rebox_fn = Let (NonRec arg con_app)
346 con_app = mkProductBox unpk_args (idType arg)
348 mkWWstr unpk_args_w_ds `thenUs` \ (worker_args, wrap_fn, work_fn) ->
349 returnUs (worker_args, unbox_fn . wrap_fn, work_fn . rebox_fn)
350 -- Don't pass the arg, rebox instead
352 -- `seq` demand; evaluate in wrapper in the hope
353 -- of dropping seqs in the worker
356 arg_w_unf = arg `setIdUnfolding` evaldUnfolding
357 -- Tell the worker arg that it's sure to be evaluated
358 -- so that internal seqs can be dropped
360 returnUs ([arg_w_unf], mk_seq_case arg, nop_fn)
361 -- Pass the arg, anyway, even if it is in theory discarded
364 -- x gets a (Eval (Poly Abs)) demand, but if we fail to pass it to the worker
365 -- we ABSOLUTELY MUST record that x is evaluated in the wrapper.
367 -- f x y = x `seq` fw y
368 -- fw y = let x{Evald} = error "oops" in (x `seq` y)
369 -- If we don't pin on the "Evald" flag, the seq doesn't disappear, and
370 -- we end up evaluating the absent thunk.
371 -- But the Evald flag is pretty weird, and I worry that it might disappear
372 -- during simplification, so for now I've just nuked this whole case
375 _other_demand -> returnUs ([arg], nop_fn, nop_fn)
378 -- If the wrapper argument is a one-shot lambda, then
379 -- so should (all) the corresponding worker arguments be
380 -- This bites when we do w/w on a case join point
381 set_worker_arg_info worker_arg demand = set_one_shot (setIdNewDemandInfo worker_arg demand)
383 set_one_shot | isOneShotLambda arg = setOneShotLambda
384 | otherwise = \x -> x
386 ----------------------
387 nop_fn :: CoreExpr -> CoreExpr
392 %************************************************************************
394 \subsection{CPR stuff}
396 %************************************************************************
399 @mkWWcpr@ takes the worker/wrapper pair produced from the strictness
400 info and adds in the CPR transformation. The worker returns an
401 unboxed tuple containing non-CPR components. The wrapper takes this
402 tuple and re-produces the correct structured output.
404 The non-CPR results appear ordered in the unboxed tuple as if by a
405 left-to-right traversal of the result structure.
409 mkWWcpr :: Type -- function body type
410 -> DmdResult -- CPR analysis results
411 -> UniqSM (CoreExpr -> CoreExpr, -- New wrapper
412 CoreExpr -> CoreExpr, -- New worker
413 Type) -- Type of worker's body
415 mkWWcpr body_ty RetCPR
416 | not (isClosedAlgType body_ty)
418 text "mkWWcpr: non-algebraic or open body type" <+> ppr body_ty )
419 returnUs (id, id, body_ty)
421 | n_con_args == 1 && isUnLiftedType con_arg_ty1
422 -- Special case when there is a single result of unlifted type
424 -- Wrapper: case (..call worker..) of x -> C x
425 -- Worker: case ( ..body.. ) of C x -> x
426 = getUniquesUs `thenUs` \ (work_uniq : arg_uniq : _) ->
428 work_wild = mk_ww_local work_uniq body_ty
429 arg = mk_ww_local arg_uniq con_arg_ty1
430 con_app = mkProductBox [arg] body_ty
432 returnUs (\ wkr_call -> Case wkr_call (arg) (exprType con_app) [(DEFAULT, [], con_app)],
433 \ body -> workerCase (work_wild) body [arg] data_con (Var arg),
436 | otherwise -- The general case
437 -- Wrapper: case (..call worker..) of (# a, b #) -> C a b
438 -- Worker: case ( ...body... ) of C a b -> (# a, b #)
439 = getUniquesUs `thenUs` \ uniqs ->
441 (wrap_wild : work_wild : args) = zipWith mk_ww_local uniqs (ubx_tup_ty : body_ty : con_arg_tys)
442 arg_vars = map Var args
443 ubx_tup_con = tupleCon Unboxed n_con_args
444 ubx_tup_ty = exprType ubx_tup_app
445 ubx_tup_app = mkConApp ubx_tup_con (map Type con_arg_tys ++ arg_vars)
446 con_app = mkProductBox args body_ty
448 returnUs (\ wkr_call -> Case wkr_call (wrap_wild) (exprType con_app) [(DataAlt ubx_tup_con, args, con_app)],
449 \ body -> workerCase (work_wild) body args data_con ubx_tup_app,
452 (_arg_tycon, _tycon_arg_tys, data_con, con_arg_tys) = deepSplitProductType "mkWWcpr" body_ty
453 n_con_args = length con_arg_tys
454 con_arg_ty1 = head con_arg_tys
456 mkWWcpr body_ty _other -- No CPR info
457 = returnUs (id, id, body_ty)
459 -- If the original function looked like
460 -- f = \ x -> _scc_ "foo" E
462 -- then we want the CPR'd worker to look like
463 -- \ x -> _scc_ "foo" (case E of I# x -> x)
464 -- and definitely not
465 -- \ x -> case (_scc_ "foo" E) of I# x -> x)
467 -- This transform doesn't move work or allocation
468 -- from one cost centre to another
469 workerCase :: Id -> CoreExpr -> [Id] -> DataCon -> CoreExpr -> CoreExpr
470 workerCase bndr (Note (SCC cc) e) args con body = Note (SCC cc) (mkUnpackCase bndr e args con body)
471 workerCase bndr e args con body = mkUnpackCase bndr e args con body
475 %************************************************************************
477 \subsection{Utilities}
479 %************************************************************************
483 mk_absent_let :: Id -> CoreExpr -> CoreExpr
484 mk_absent_let arg body
485 | not (isUnLiftedType arg_ty)
486 = Let (NonRec arg abs_rhs) body
488 = panic "WwLib: haven't done mk_absent_let for primitives yet"
491 abs_rhs = mkRuntimeErrorApp rUNTIME_ERROR_ID arg_ty msg
492 msg = "Oops! Entered absent arg " ++ showSDocDebug (ppr arg <+> ppr (idType arg))
494 mk_seq_case :: Id -> CoreExpr -> CoreExpr
495 mk_seq_case arg body = Case (Var arg) (sanitiseCaseBndr arg) (exprType body) [(DEFAULT, [], body)]
497 sanitiseCaseBndr :: Id -> Id
498 -- The argument we are scrutinising has the right type to be
499 -- a case binder, so it's convenient to re-use it for that purpose.
500 -- But we *must* throw away all its IdInfo. In particular, the argument
501 -- will have demand info on it, and that demand info may be incorrect for
502 -- the case binder. e.g. case ww_arg of ww_arg { I# x -> ... }
503 -- Quite likely ww_arg isn't used in '...'. The case may get discarded
504 -- if the case binder says "I'm demanded". This happened in a situation
505 -- like (x+y) `seq` ....
506 sanitiseCaseBndr id = id `setIdInfo` vanillaIdInfo
508 mk_ww_local :: Unique -> Type -> Id
509 mk_ww_local uniq ty = mkSysLocal FSLIT("ww") uniq ty