2 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
4 \section[SCCfinal]{Modify and collect code generation for final STG program}
6 This is now a sort-of-normal STG-to-STG pass (WDP 94/06), run by stg2stg.
8 * Traverses the STG program collecting the cost centres. These are
9 required to declare the cost centres at the start of code
12 Note: because of cross-module unfolding, some of these cost centres
13 may be from other modules. But will still have to give them
14 "extern" declarations.
16 * Puts on CAF cost-centres if the user has asked for individual CAF
19 * Ditto for individual DICT cost-centres.
21 * Boxes top-level inherited functions passed as arguments.
23 * "Distributes" given cost-centres to all as-yet-unmarked RHSs.
26 module SCCfinal ( stgMassageForProfiling ) where
28 #include "HsVersions.h"
32 import CmdLineOpts ( opt_AutoSccsOnIndividualCafs )
33 import CostCentre -- lots of things
34 import Id ( Id, mkSysLocal, idType, idName )
35 import Module ( Module )
36 import UniqSupply ( uniqFromSupply, splitUniqSupply, UniqSupply )
37 import Unique ( Unique )
38 import Type ( splitForAllTys, splitTyConApp_maybe )
39 import TyCon ( isFunTyCon )
42 import Name ( isLocallyDefined )
43 import Util ( removeDups )
46 infixr 9 `thenMM`, `thenMM_`
50 type CollectedCCs = ([CostCentre], -- locally defined ones
51 [CostCentre], -- ones needing "extern" decls
52 [CostCentreStack]) -- singleton stacks (for CAFs)
54 stgMassageForProfiling
55 :: Module -- module name
56 -> UniqSupply -- unique supply
57 -> [StgBinding] -- input
58 -> (CollectedCCs, [StgBinding])
60 stgMassageForProfiling mod_name us stg_binds
62 ((local_ccs, extern_ccs, cc_stacks),
64 = initMM mod_name us (do_top_bindings stg_binds)
66 (fixed_ccs, fixed_cc_stacks)
67 = if opt_AutoSccsOnIndividualCafs
68 then ([],[]) -- don't need "all CAFs" CC
69 -- (for Prelude, we use PreludeCC)
70 else ([all_cafs_cc], [all_cafs_ccs])
72 local_ccs_no_dups = fst (removeDups cmpCostCentre local_ccs)
73 extern_ccs_no_dups = fst (removeDups cmpCostCentre extern_ccs)
75 ((fixed_ccs ++ local_ccs_no_dups,
77 fixed_cc_stacks ++ cc_stacks), stg_binds2)
80 all_cafs_cc = mkAllCafsCC mod_name
81 all_cafs_ccs = mkSingletonCCS all_cafs_cc
84 do_top_bindings :: [StgBinding] -> MassageM [StgBinding]
86 do_top_bindings [] = returnMM []
88 do_top_bindings (StgNonRec b rhs : bs)
89 = do_top_rhs b rhs `thenMM` \ rhs' ->
91 do_top_bindings bs `thenMM` \bs' ->
92 returnMM (StgNonRec b rhs' : bs')
95 do_top_bindings (StgRec pairs : bs)
96 = addTopLevelIshIds binders (
97 mapMM do_pair pairs `thenMM` \ pairs2 ->
98 do_top_bindings bs `thenMM` \ bs' ->
99 returnMM (StgRec pairs2 : bs')
102 binders = map fst pairs
104 = do_top_rhs b rhs `thenMM` \ rhs2 ->
108 do_top_rhs :: Id -> StgRhs -> MassageM StgRhs
110 do_top_rhs binder (StgRhsClosure _ bi srt fv u [] (StgSCC cc (StgConApp con args)))
111 | not (isSccCountCostCentre cc)
112 -- Trivial _scc_ around nothing but static data
113 -- Eliminate _scc_ ... and turn into StgRhsCon
114 = returnMM (StgRhsCon dontCareCCS con args)
116 {- Can't do this one with cost-centre stacks: --SDM
117 do_top_rhs binder (StgRhsClosure no_cc bi srt fv u [] (StgSCC ty cc expr))
118 | (noCCSAttached no_cc || currentOrSubsumedCCS no_cc)
119 && not (isSccCountCostCentre cc)
120 -- Top level CAF without a cost centre attached
121 -- Attach and collect cc of trivial _scc_ in body
122 = collectCC cc `thenMM_`
123 set_prevailing_cc cc (do_expr expr) `thenMM` \ expr' ->
124 returnMM (StgRhsClosure cc bi srt fv u [] expr')
127 do_top_rhs binder (StgRhsClosure no_cc bi srt fv u [] body)
128 | noCCSAttached no_cc || currentOrSubsumedCCS no_cc
129 -- Top level CAF without a cost centre attached
130 -- Attach CAF cc (collect if individual CAF ccs)
131 = (if opt_AutoSccsOnIndividualCafs
132 then let cc = mkAutoCC binder mod_name CafCC
133 ccs = mkSingletonCCS cc
135 collectCC cc `thenMM_`
136 collectCCS ccs `thenMM_`
139 returnMM all_cafs_ccs) `thenMM` \ caf_ccs ->
140 set_prevailing_cc caf_ccs (do_expr body) `thenMM` \ body' ->
141 returnMM (StgRhsClosure caf_ccs bi srt fv u [] body')
143 do_top_rhs binder (StgRhsClosure cc bi srt fv u [] body)
144 -- Top level CAF with cost centre attached
145 -- Should this be a CAF cc ??? Does this ever occur ???
146 = pprPanic "SCCfinal: CAF with cc:" (ppr cc)
148 {- can't do this with cost-centre stacks: --SDM
149 do_top_rhs binder (StgRhsClosure _ bi srt fv u args (StgSCC cc expr))
150 | not (isSccCountCostCentre cc)
151 -- Top level function with trivial _scc_ in body
152 -- Attach and collect cc of trivial _scc_
153 = collectCC cc `thenMM_`
154 set_prevailing_cc cc (do_expr expr) `thenMM` \ expr' ->
155 returnMM (StgRhsClosure cc bi srt fv u args expr')
158 do_top_rhs binder (StgRhsClosure no_ccs bi srt fv u args body)
159 -- Top level function, probably subsumed
160 | noCCSAttached no_ccs
161 = set_lambda_cc (do_expr body) `thenMM` \ body' ->
162 returnMM (StgRhsClosure subsumedCCS bi srt fv u args body')
165 = pprPanic "SCCfinal: CAF with cc:" (ppr no_ccs)
167 do_top_rhs binder (StgRhsCon ccs con args)
168 -- Top-level (static) data is not counted in heap
169 -- profiles; nor do we set CCCS from it; so we
170 -- just slam in dontCareCostCentre
171 = returnMM (StgRhsCon dontCareCCS con args)
174 do_expr :: StgExpr -> MassageM StgExpr
176 do_expr (StgLit l) = returnMM (StgLit l)
178 do_expr (StgApp fn args)
179 = boxHigherOrderArgs (StgApp fn) args
181 do_expr (StgConApp con args)
182 = boxHigherOrderArgs (\args -> StgConApp con args) args
184 do_expr (StgPrimApp con args res_ty)
185 = boxHigherOrderArgs (\args -> StgPrimApp con args res_ty) args
187 do_expr (StgSCC cc expr) -- Ha, we found a cost centre!
188 = collectCC cc `thenMM_`
189 do_expr expr `thenMM` \ expr' ->
190 returnMM (StgSCC cc expr')
192 do_expr (StgCase expr fv1 fv2 bndr srt alts)
193 = do_expr expr `thenMM` \ expr' ->
194 do_alts alts `thenMM` \ alts' ->
195 returnMM (StgCase expr' fv1 fv2 bndr srt alts')
197 do_alts (StgAlgAlts ty alts def)
198 = mapMM do_alt alts `thenMM` \ alts' ->
199 do_deflt def `thenMM` \ def' ->
200 returnMM (StgAlgAlts ty alts' def')
202 do_alt (id, bs, use_mask, e)
203 = do_expr e `thenMM` \ e' ->
204 returnMM (id, bs, use_mask, e')
206 do_alts (StgPrimAlts ty alts def)
207 = mapMM do_alt alts `thenMM` \ alts' ->
208 do_deflt def `thenMM` \ def' ->
209 returnMM (StgPrimAlts ty alts' def')
212 = do_expr e `thenMM` \ e' ->
215 do_deflt StgNoDefault = returnMM StgNoDefault
216 do_deflt (StgBindDefault e)
217 = do_expr e `thenMM` \ e' ->
218 returnMM (StgBindDefault e')
221 = do_let b e `thenMM` \ (b,e) ->
222 returnMM (StgLet b e)
224 do_expr (StgLetNoEscape lvs1 lvs2 b e)
225 = do_let b e `thenMM` \ (b,e) ->
226 returnMM (StgLetNoEscape lvs1 lvs2 b e)
228 ----------------------------------
230 do_let (StgNonRec b rhs) e
231 = do_rhs rhs `thenMM` \ rhs' ->
233 do_expr e `thenMM` \ e' ->
234 returnMM (StgNonRec b rhs',e')
237 do_let (StgRec pairs) e
238 = addTopLevelIshIds binders (
239 mapMM do_pair pairs `thenMM` \ pairs' ->
240 do_expr e `thenMM` \ e' ->
241 returnMM (StgRec pairs', e')
244 binders = map fst pairs
246 = do_rhs rhs `thenMM` \ rhs2 ->
249 ----------------------------------
250 do_rhs :: StgRhs -> MassageM StgRhs
251 -- We play much the same game as we did in do_top_rhs above;
252 -- but we don't have to worry about cafs etc.
255 do_rhs (StgRhsClosure closure_cc bi srt fv u [] (StgSCC ty cc (StgCon (DataCon con) args _)))
256 | not (isSccCountCostCentre cc)
257 = collectCC cc `thenMM_`
258 returnMM (StgRhsCon cc con args)
262 do_rhs (StgRhsClosure _ bi srt fv u args (StgSCC ty cc expr))
263 | not (isSccCountCostCentre cc)
264 = collectCC cc `thenMM_`
265 set_prevailing_cc cc (do_expr expr) `thenMM` \ expr' ->
266 returnMM (StgRhsClosure cc bi srt fv u args expr')
269 do_rhs (StgRhsClosure cc bi srt fv u [] body)
270 = do_expr body `thenMM` \ body' ->
271 returnMM (StgRhsClosure currentCCS bi srt fv u [] body')
273 do_rhs (StgRhsClosure cc bi srt fv u args body)
274 = set_lambda_cc (do_expr body) `thenMM` \ body' ->
275 get_prevailing_cc `thenMM` \ prev_ccs ->
276 returnMM (StgRhsClosure currentCCS bi srt fv u args body')
278 do_rhs (StgRhsCon cc con args)
279 = returnMM (StgRhsCon currentCCS con args)
282 %************************************************************************
284 \subsection{Boxing higher-order args}
286 %************************************************************************
288 Boxing is *turned off* at the moment, until we can figure out how to
289 do it properly in general.
293 :: ([StgArg] -> StgExpr)
294 -- An application lacking its arguments
295 -> [StgArg] -- arguments which we might box
298 #ifndef PROF_DO_BOXING
299 boxHigherOrderArgs almost_expr args
300 = returnMM (almost_expr args)
302 boxHigherOrderArgs almost_expr args
303 = getTopLevelIshIds `thenMM` \ ids ->
304 mapAccumMM (do_arg ids) [] args `thenMM` \ (let_bindings, new_args) ->
305 returnMM (foldr (mk_stg_let currentCCS) (almost_expr new_args) let_bindings)
309 do_arg ids bindings arg@(StgVarArg old_var)
310 | (not (isLocallyDefined old_var) || elemVarSet old_var ids)
311 && isFunType var_type
312 = -- make a trivial let-binding for the top-level function
313 getUniqueMM `thenMM` \ uniq ->
315 new_var = mkSysLocal SLIT("sf") uniq var_type
317 returnMM ( (new_var, old_var) : bindings, StgVarArg new_var )
319 var_type = idType old_var
321 do_arg ids bindings arg = returnMM (bindings, arg)
324 mk_stg_let :: CostCentreStack -> (Id, Id) -> StgExpr -> StgExpr
326 mk_stg_let cc (new_var, old_var) body
328 rhs_body = StgApp old_var [{-args-}]
329 rhs_closure = StgRhsClosure cc stgArgOcc NoSRT [{-fvs-}] ReEntrant [{-args-}] rhs_body
331 StgLet (StgNonRec new_var rhs_closure) body
333 bOGUS_LVs = emptyUniqSet -- easier to print than: panic "mk_stg_let: LVs"
336 = case splitForAllTys var_type of
337 (_, ty) -> case splitTyConApp_maybe ty of
338 Just (tycon,_) | isFunTyCon tycon -> True
343 %************************************************************************
345 \subsection{Boring monad stuff for this}
347 %************************************************************************
351 = Module -- module name
352 -> CostCentreStack -- prevailing CostCentre
353 -- if none, subsumedCosts at top-level
354 -- useCurrentCostCentre at nested levels
356 -> VarSet -- toplevel-ish Ids for boxing
358 -> (CollectedCCs, result)
360 -- the initMM function also returns the final CollectedCCs
362 initMM :: Module -- module name, which we may consult
367 initMM mod_name init_us m = m mod_name noCCS init_us emptyVarSet ([],[],[])
369 thenMM :: MassageM a -> (a -> MassageM b) -> MassageM b
370 thenMM_ :: MassageM a -> (MassageM b) -> MassageM b
372 thenMM expr cont mod scope_cc us ids ccs
373 = case splitUniqSupply us of { (s1, s2) ->
374 case (expr mod scope_cc s1 ids ccs) of { (ccs2, result) ->
375 cont result mod scope_cc s2 ids ccs2 }}
377 thenMM_ expr cont mod scope_cc us ids ccs
378 = case splitUniqSupply us of { (s1, s2) ->
379 case (expr mod scope_cc s1 ids ccs) of { (ccs2, _) ->
380 cont mod scope_cc s2 ids ccs2 }}
382 returnMM :: a -> MassageM a
383 returnMM result mod scope_cc us ids ccs = (ccs, result)
386 nopMM mod scope_cc us ids ccs = (ccs, ())
388 mapMM :: (a -> MassageM b) -> [a] -> MassageM [b]
389 mapMM f [] = returnMM []
391 = f m `thenMM` \ r ->
392 mapMM f ms `thenMM` \ rs ->
395 mapAccumMM :: (acc -> x -> MassageM (acc, y)) -> acc -> [x] -> MassageM (acc, [y])
396 mapAccumMM f b [] = returnMM (b, [])
397 mapAccumMM f b (m:ms)
398 = f b m `thenMM` \ (b2, r) ->
399 mapAccumMM f b2 ms `thenMM` \ (b3, rs) ->
402 getUniqueMM :: MassageM Unique
403 getUniqueMM mod scope_cc us ids ccs = (ccs, uniqFromSupply us)
405 addTopLevelIshId :: Id -> MassageM a -> MassageM a
406 addTopLevelIshId id scope mod scope_cc us ids ccs
407 | isCurrentCCS scope_cc = scope mod scope_cc us ids ccs
408 | otherwise = scope mod scope_cc us (extendVarSet ids id) ccs
410 addTopLevelIshIds :: [Id] -> MassageM a -> MassageM a
411 addTopLevelIshIds [] cont = cont
412 addTopLevelIshIds (id:ids) cont
413 = addTopLevelIshId id (addTopLevelIshIds ids cont)
415 getTopLevelIshIds :: MassageM VarSet
416 getTopLevelIshIds mod scope_cc us ids ccs = (ccs, ids)
419 The prevailing CCS is used to tell whether we're in a top-levelish
420 position, where top-levelish is defined as "not inside a lambda".
421 Prevailing CCs used to be used for something much more complicated,
425 set_lambda_cc :: MassageM a -> MassageM a
426 set_lambda_cc action mod scope_cc us ids ccs
427 = action mod currentCCS us ids ccs
429 set_prevailing_cc :: CostCentreStack -> MassageM a -> MassageM a
430 set_prevailing_cc cc_to_set_to action mod scope_cc us ids ccs
431 = action mod cc_to_set_to us ids ccs
433 get_prevailing_cc :: MassageM CostCentreStack
434 get_prevailing_cc mod scope_cc us ids ccs = (ccs, scope_cc)
438 collectCC :: CostCentre -> MassageM ()
440 collectCC cc mod_name scope_cc us ids (local_ccs, extern_ccs, ccss)
441 = ASSERT(not (noCCAttached cc))
442 if (cc `ccFromThisModule` mod_name) then
443 ((cc : local_ccs, extern_ccs, ccss), ())
444 else -- must declare it "extern"
445 ((local_ccs, cc : extern_ccs, ccss), ())
447 collectCCS :: CostCentreStack -> MassageM ()
449 collectCCS ccs mod_name scope_cc us ids (local_ccs, extern_ccs, ccss)
450 = ASSERT(not (noCCSAttached ccs))
451 ((local_ccs, extern_ccs, ccs : ccss), ())