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 ) 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 )
18 import DataCon ( splitProductType_maybe, splitProductType )
19 import NewDemand ( Demand(..), Keepity(..), DmdResult(..) )
20 import DmdAnal ( both )
21 import PrelInfo ( realWorldPrimId, aBSENT_ERROR_ID, eRROR_CSTRING_ID )
22 import TysPrim ( realWorldStatePrimTy )
23 import TysWiredIn ( tupleCon )
24 import Type ( Type, isUnLiftedType, mkFunTys,
25 splitForAllTys, splitFunTys, splitNewType_maybe, isAlgType
27 import Literal ( Literal(MachStr) )
28 import BasicTypes ( Boxity(..) )
29 import Var ( Var, isId )
30 import UniqSupply ( returnUs, thenUs, getUniqueUs, getUniquesUs, UniqSM )
31 import Util ( zipWithEqual )
33 import List ( zipWith4 )
37 %************************************************************************
39 \subsection[mkWrapperAndWorker]{@mkWrapperAndWorker@}
41 %************************************************************************
43 Here's an example. The original function is:
46 g :: forall a . Int -> [a] -> a
54 From this, we want to produce:
56 -- wrapper (an unfolding)
57 g :: forall a . Int -> [a] -> a
62 -- call the worker; don't forget the type args!
65 $wg :: forall a . Int# -> [a] -> a
67 $wg = /\ a -> \ x# ys ->
71 case x of -- note: body of g moved intact
76 Something we have to be careful about: Here's an example:
79 -- "f" strictness: U(P)U(P)
80 f (I# a) (I# b) = a +# b
82 g = f -- "g" strictness same as "f"
85 \tr{f} will get a worker all nice and friendly-like; that's good.
86 {\em But we don't want a worker for \tr{g}}, even though it has the
87 same strictness as \tr{f}. Doing so could break laziness, at best.
89 Consequently, we insist that the number of strictness-info items is
90 exactly the same as the number of lambda-bound arguments. (This is
91 probably slightly paranoid, but OK in practice.) If it isn't the
92 same, we ``revise'' the strictness info, so that we won't propagate
93 the unusable strictness-info into the interfaces.
96 %************************************************************************
98 \subsection{The worker wrapper core}
100 %************************************************************************
102 @mkWwBodies@ is called when doing the worker/wrapper split inside a module.
105 mkWwBodies :: Type -- Type of original function
106 -> [Demand] -- Strictness of original function
107 -> DmdResult -- Info about function result
108 -> [Bool] -- One-shot-ness of the function
109 -> UniqSM ([Demand], -- Demands for worker (value) args
110 Id -> CoreExpr, -- Wrapper body, lacking only the worker Id
111 CoreExpr -> CoreExpr) -- Worker body, lacking the original function rhs
113 -- wrap_fn_args E = \x y -> E
114 -- work_fn_args E = E x y
116 -- wrap_fn_str E = case x of { (a,b) ->
117 -- case a of { (a1,a2) ->
119 -- work_fn_str E = \a2 a2 b y ->
120 -- let a = (a1,a2) in
124 mkWwBodies fun_ty demands res_info one_shots
125 = mkWWargs fun_ty demands one_shots' `thenUs` \ (wrap_args, wrap_fn_args, work_fn_args, res_ty) ->
126 mkWWcpr res_ty res_info `thenUs` \ (wrap_fn_cpr, work_fn_cpr, cpr_res_ty) ->
127 mkWWstr cpr_res_ty wrap_args `thenUs` \ (work_dmds, wrap_fn_str, work_fn_str) ->
130 Note InlineMe . wrap_fn_args . wrap_fn_cpr . wrap_fn_str . Var,
131 work_fn_str . work_fn_cpr . work_fn_args)
132 -- We use an INLINE unconditionally, even if the wrapper turns out to be
133 -- something trivial like
135 -- f = __inline__ (coerce T fw)
136 -- The point is to propagate the coerce to f's call sites, so even though
137 -- f's RHS is now trivial (size 1) we still want the __inline__ to prevent
138 -- fw from being inlined into f's RHS
140 one_shots' = one_shots ++ repeat False
144 %************************************************************************
146 \subsection{Coercion stuff}
148 %************************************************************************
151 We really want to "look through" coerces.
152 Reason: I've seen this situation:
154 let f = coerce T (\s -> E)
160 If only we w/w'd f, we'd get
161 let f = coerce T (\s -> fw s)
165 Now we'll inline f to get
173 Now we'll see that fw has arity 1, and will arity expand
174 the \x to get what we want.
177 -- mkWWargs is driven off the function type and arity.
178 -- It chomps bites off foralls, arrows, newtypes
179 -- and keeps repeating that until it's satisfied the supplied arity
183 -> [Bool] -- True for a one-shot arg; ** may be infinite **
184 -> UniqSM ([Var], -- Wrapper args
185 CoreExpr -> CoreExpr, -- Wrapper fn
186 CoreExpr -> CoreExpr, -- Worker fn
187 Type) -- Type of wrapper body
189 mkWWargs fun_ty demands one_shots
190 | Just rep_ty <- splitNewType_maybe fun_ty
191 -- The newtype case is for when the function has
192 -- a recursive newtype after the arrow (rare)
193 -- We check for arity >= 0 to avoid looping in the case
194 -- of a function whose type is, in effect, infinite
195 -- [Arity is driven by looking at the term, not just the type.]
197 -- It's also important when we have a function returning (say) a pair
198 -- wrapped in a recursive newtype, at least if CPR analysis can look
199 -- through such newtypes, which it probably can since they are
201 = mkWWargs rep_ty demands one_shots `thenUs` \ (wrap_args, wrap_fn_args, work_fn_args, res_ty) ->
203 Note (Coerce fun_ty rep_ty) . wrap_fn_args,
204 work_fn_args . Note (Coerce rep_ty fun_ty),
208 = getUniquesUs `thenUs` \ wrap_uniqs ->
210 (tyvars, tau) = splitForAllTys fun_ty
211 (arg_tys, body_ty) = splitFunTys tau
213 n_demands = length demands
214 n_arg_tys = length arg_tys
215 n_args = n_demands `min` n_arg_tys
217 new_fun_ty = mkFunTys (drop n_demands arg_tys) body_ty
218 new_demands = drop n_arg_tys demands
219 new_one_shots = drop n_args one_shots
221 val_args = zipWith4 mk_wrap_arg wrap_uniqs arg_tys demands one_shots
222 wrap_args = tyvars ++ val_args
224 {- ASSERT( not (null tyvars) || not (null arg_tys) ) -}
225 if (null tyvars) && (null arg_tys) then
226 pprTrace "mkWWargs" (ppr fun_ty $$ ppr demands)
227 returnUs ([], id, id, fun_ty)
232 new_one_shots `thenUs` \ (more_wrap_args, wrap_fn_args, work_fn_args, res_ty) ->
234 returnUs (wrap_args ++ more_wrap_args,
235 mkLams wrap_args . wrap_fn_args,
236 work_fn_args . applyToVars wrap_args,
240 = returnUs ([], id, id, fun_ty)
243 applyToVars :: [Var] -> CoreExpr -> CoreExpr
244 applyToVars vars fn = mkVarApps fn vars
246 mk_wrap_arg uniq ty dmd one_shot
247 = set_one_shot one_shot (setIdNewDemandInfo (mkSysLocal SLIT("w") uniq ty) dmd)
249 set_one_shot True id = setOneShotLambda id
250 set_one_shot False id = id
254 %************************************************************************
256 \subsection{Strictness stuff}
258 %************************************************************************
261 mkWWstr :: Type -- Result type
262 -> [Var] -- Wrapper args; have their demand info on them
263 -- *Includes type variables*
264 -> UniqSM ([Demand], -- Demand on worker (value) args
265 CoreExpr -> CoreExpr, -- Wrapper body, lacking the worker call
266 -- and without its lambdas
267 -- This fn adds the unboxing, and makes the
268 -- call passing the unboxed things
270 CoreExpr -> CoreExpr) -- Worker body, lacking the original body of the function,
271 -- but *with* lambdas
273 mkWWstr res_ty wrap_args
274 = mk_ww_str_s wrap_args `thenUs` \ (work_args, take_apart, put_together) ->
276 work_dmds = [idNewDemandInfo v | v <- work_args, isId v]
277 apply_to args fn = mkVarApps fn args
279 if not (null work_dmds && isUnLiftedType res_ty) then
280 returnUs ( work_dmds,
281 take_apart . applyToVars work_args,
282 mkLams work_args . put_together)
284 -- Horrid special case. If the worker would have no arguments, and the
285 -- function returns a primitive type value, that would make the worker into
286 -- an unboxed value. We box it by passing a dummy void argument, thus:
288 -- f = /\abc. \xyz. fw abc void
289 -- fw = /\abc. \v. body
291 -- We use the state-token type which generates no code
292 getUniqueUs `thenUs` \ void_arg_uniq ->
294 void_arg = mk_ww_local void_arg_uniq realWorldStatePrimTy
297 take_apart . applyToVars [realWorldPrimId] . apply_to work_args,
298 mkLams work_args . Lam void_arg . put_together)
300 ----------------------
303 ----------------------
305 = returnUs ([], nop_fn, nop_fn)
307 mk_ww_str_s (arg : args)
308 = mk_ww_str arg `thenUs` \ (args1, wrap_fn1, work_fn1) ->
309 mk_ww_str_s args `thenUs` \ (args2, wrap_fn2, work_fn2) ->
310 returnUs (args1 ++ args2, wrap_fn1 . wrap_fn2, work_fn1 . work_fn2)
313 ----------------------
316 = returnUs ([arg], nop_fn, nop_fn)
319 = case idNewDemandInfo arg of
321 -- Absent case. We don't deal with absence for unlifted types,
322 -- though, because it's not so easy to manufacture a placeholder
323 -- We'll see if this turns out to be a problem
324 Abs | not (isUnLiftedType (idType arg)) ->
325 returnUs ([], nop_fn, mk_absent_let arg)
330 arg_w_unf = arg `setIdUnfolding` mkOtherCon []
331 -- Tell the worker arg that it's sure to be evaluated
332 -- so that internal seqs can be dropped
334 returnUs ([arg_w_unf], mk_seq_case arg, nop_fn)
335 -- Pass the arg, anyway, even if it is in theory discarded
338 -- x gets a (Seq Drop []) demand, but if we fail to pass it to the worker
339 -- we ABSOLUTELY MUST record that x is evaluated in the wrapper.
341 -- f x y = x `seq` fw y
342 -- fw y = let x{Evald} = error "oops" in (x `seq` y)
343 -- If we don't pin on the "Evald" flag, the seq doesn't disappear, and
344 -- we end up evaluating the absent thunk.
345 -- But the Evald flag is pretty wierd, and I worry that it might disappear
346 -- during simplification, so for now I've just nuked this whole case
350 | Just (arg_tycon, tycon_arg_tys, data_con, inst_con_arg_tys)
351 <- splitProductType_maybe (idType arg)
352 -> getUniquesUs `thenUs` \ uniqs ->
354 unpk_args = zipWith mk_ww_local uniqs inst_con_arg_tys
355 unpk_args_w_ds = zipWithEqual "mk_ww_str" set_worker_arg_info unpk_args cs'
356 unbox_fn = mk_unpk_case arg unpk_args data_con arg_tycon
357 rebox_fn = Let (NonRec arg con_app)
358 con_app = mkConApp data_con (map Type tycon_arg_tys ++ map Var unpk_args)
361 Keep -> map (DmdAnal.both Lazy) cs -- Careful! Now we don't pass
362 -- the box, we must pass all the
363 -- components. In effect
367 mk_ww_str_s unpk_args_w_ds `thenUs` \ (worker_args, wrap_fn, work_fn) ->
370 -- Keep -> returnUs (arg : worker_args, unbox_fn . wrap_fn, work_fn)
371 -- -- Pass the arg, no need to rebox
372 -- Drop -> returnUs (worker_args, unbox_fn . wrap_fn, work_fn . rebox_fn)
373 -- -- Don't pass the arg, rebox instead
374 -- I used to be clever here, but consider
376 -- f n (x:xs) = f (n+x) xs
377 -- Here n gets (Seq Keep [L]), but it's BAD BAD BAD to pass both n and n#
378 -- Needs more thought, but the simple thing to do is to accept the reboxing
379 -- stuff if there are any non-absent arguments (and that case is dealt with above):
381 returnUs (worker_args, unbox_fn . wrap_fn, work_fn . rebox_fn)
382 -- Don't pass the arg, rebox instead
385 WARN( True, ppr arg )
386 returnUs ([arg], nop_fn, nop_fn)
389 other_demand -> returnUs ([arg], nop_fn, nop_fn)
392 -- If the wrapper argument is a one-shot lambda, then
393 -- so should (all) the corresponding worker arguments be
394 -- This bites when we do w/w on a case join point
395 set_worker_arg_info worker_arg demand = set_one_shot (setIdNewDemandInfo worker_arg demand)
397 set_one_shot | isOneShotLambda arg = setOneShotLambda
398 | otherwise = \x -> x
402 %************************************************************************
404 \subsection{CPR stuff}
406 %************************************************************************
409 @mkWWcpr@ takes the worker/wrapper pair produced from the strictness
410 info and adds in the CPR transformation. The worker returns an
411 unboxed tuple containing non-CPR components. The wrapper takes this
412 tuple and re-produces the correct structured output.
414 The non-CPR results appear ordered in the unboxed tuple as if by a
415 left-to-right traversal of the result structure.
419 mkWWcpr :: Type -- function body type
420 -> DmdResult -- CPR analysis results
421 -> UniqSM (CoreExpr -> CoreExpr, -- New wrapper
422 CoreExpr -> CoreExpr, -- New worker
423 Type) -- Type of worker's body
425 mkWWcpr body_ty RetCPR
426 | not (isAlgType body_ty)
427 = WARN( True, text "mkWWcpr: non-algebraic body type" <+> ppr body_ty )
428 returnUs (id, id, body_ty)
430 | n_con_args == 1 && isUnLiftedType con_arg_ty1
431 -- Special case when there is a single result of unlifted type
432 = getUniquesUs `thenUs` \ (work_uniq : arg_uniq : _) ->
434 work_wild = mk_ww_local work_uniq body_ty
435 arg = mk_ww_local arg_uniq con_arg_ty1
437 returnUs (\ wkr_call -> Case wkr_call arg [(DEFAULT, [], mkConApp data_con (map Type tycon_arg_tys ++ [Var arg]))],
438 \ body -> workerCase body work_wild [(DataAlt data_con, [arg], Var arg)],
441 | otherwise -- The general case
442 = getUniquesUs `thenUs` \ uniqs ->
444 (wrap_wild : work_wild : args) = zipWith mk_ww_local uniqs (ubx_tup_ty : body_ty : con_arg_tys)
445 arg_vars = map Var args
446 ubx_tup_con = tupleCon Unboxed n_con_args
447 ubx_tup_ty = exprType ubx_tup_app
448 ubx_tup_app = mkConApp ubx_tup_con (map Type con_arg_tys ++ arg_vars)
449 con_app = mkConApp data_con (map Type tycon_arg_tys ++ arg_vars)
451 returnUs (\ wkr_call -> Case wkr_call wrap_wild [(DataAlt ubx_tup_con, args, con_app)],
452 \ body -> workerCase body work_wild [(DataAlt data_con, args, ubx_tup_app)],
455 (_, tycon_arg_tys, data_con, con_arg_tys) = splitProductType "mkWWcpr" body_ty
456 n_con_args = length con_arg_tys
457 con_arg_ty1 = head con_arg_tys
459 mkWWcpr body_ty other -- No CPR info
460 = returnUs (id, id, body_ty)
462 -- If the original function looked like
463 -- f = \ x -> _scc_ "foo" E
465 -- then we want the CPR'd worker to look like
466 -- \ x -> _scc_ "foo" (case E of I# x -> x)
467 -- and definitely not
468 -- \ x -> case (_scc_ "foo" E) of I# x -> x)
470 -- This transform doesn't move work or allocation
471 -- from one cost centre to another
473 workerCase (Note (SCC cc) e) arg alts = Note (SCC cc) (Case e arg alts)
474 workerCase e arg alts = Case e arg alts
478 %************************************************************************
480 \subsection{Utilities}
482 %************************************************************************
486 mk_absent_let arg body
487 | not (isUnLiftedType arg_ty)
488 = Let (NonRec arg abs_rhs) body
490 = panic "WwLib: haven't done mk_absent_let for primitives yet"
493 -- abs_rhs = mkTyApps (Var aBSENT_ERROR_ID) [arg_ty]
494 abs_rhs = mkApps (Var eRROR_CSTRING_ID) [Type arg_ty, Lit (MachStr (_PK_ msg))]
495 msg = "Oops! Entered absent arg " ++ showSDocDebug (ppr arg <+> ppr (idType arg))
497 mk_unpk_case arg unpk_args boxing_con boxing_tycon body
500 (sanitiseCaseBndr arg)
501 [(DataAlt boxing_con, unpk_args, body)]
503 mk_seq_case arg body = Case (Var arg) (sanitiseCaseBndr arg) [(DEFAULT, [], body)]
505 sanitiseCaseBndr :: Id -> Id
506 -- The argument we are scrutinising has the right type to be
507 -- a case binder, so it's convenient to re-use it for that purpose.
508 -- But we *must* throw away all its IdInfo. In particular, the argument
509 -- will have demand info on it, and that demand info may be incorrect for
510 -- the case binder. e.g. case ww_arg of ww_arg { I# x -> ... }
511 -- Quite likely ww_arg isn't used in '...'. The case may get discarded
512 -- if the case binder says "I'm demanded". This happened in a situation
513 -- like (x+y) `seq` ....
514 sanitiseCaseBndr id = id `setIdInfo` vanillaIdInfo
516 mk_ww_local uniq ty = mkSysLocal SLIT("ww") uniq ty