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
35 import Module ( Module )
36 import UniqSupply ( uniqFromSupply, splitUniqSupply, UniqSupply )
37 import Unique ( Unique )
39 import ListSetOps ( removeDups )
42 infixr 9 `thenMM`, `thenMM_`
46 type CollectedCCs = ([CostCentre], -- locally defined ones
47 [CostCentre], -- ones needing "extern" decls
48 [CostCentreStack]) -- singleton stacks (for CAFs)
50 stgMassageForProfiling
51 :: Module -- module name
52 -> UniqSupply -- unique supply
53 -> [StgBinding] -- input
54 -> (CollectedCCs, [StgBinding])
56 stgMassageForProfiling mod_name us stg_binds
58 ((local_ccs, extern_ccs, cc_stacks),
60 = initMM mod_name us (do_top_bindings stg_binds)
62 (fixed_ccs, fixed_cc_stacks)
63 = if opt_AutoSccsOnIndividualCafs
64 then ([],[]) -- don't need "all CAFs" CC
65 -- (for Prelude, we use PreludeCC)
66 else ([all_cafs_cc], [all_cafs_ccs])
68 local_ccs_no_dups = fst (removeDups cmpCostCentre local_ccs)
69 extern_ccs_no_dups = fst (removeDups cmpCostCentre extern_ccs)
71 ((fixed_ccs ++ local_ccs_no_dups,
73 fixed_cc_stacks ++ cc_stacks), stg_binds2)
76 all_cafs_cc = mkAllCafsCC mod_name
77 all_cafs_ccs = mkSingletonCCS all_cafs_cc
80 do_top_bindings :: [StgBinding] -> MassageM [StgBinding]
82 do_top_bindings [] = returnMM []
84 do_top_bindings (StgNonRec b rhs : bs)
85 = do_top_rhs b rhs `thenMM` \ rhs' ->
87 do_top_bindings bs `thenMM` \bs' ->
88 returnMM (StgNonRec b rhs' : bs')
91 do_top_bindings (StgRec pairs : bs)
92 = addTopLevelIshIds binders (
93 mapMM do_pair pairs `thenMM` \ pairs2 ->
94 do_top_bindings bs `thenMM` \ bs' ->
95 returnMM (StgRec pairs2 : bs')
98 binders = map fst pairs
100 = do_top_rhs b rhs `thenMM` \ rhs2 ->
104 do_top_rhs :: Id -> StgRhs -> MassageM StgRhs
106 do_top_rhs binder (StgRhsClosure _ bi srt fv u [] (StgSCC cc (StgConApp con args)))
107 | not (isSccCountCostCentre cc) && not (isDllConApp con args)
108 -- Trivial _scc_ around nothing but static data
109 -- Eliminate _scc_ ... and turn into StgRhsCon
111 -- isDllConApp checks for LitLit args too
112 = returnMM (StgRhsCon dontCareCCS con args)
114 {- Can't do this one with cost-centre stacks: --SDM
115 do_top_rhs binder (StgRhsClosure no_cc bi srt fv u [] (StgSCC ty cc expr))
116 | (noCCSAttached no_cc || currentOrSubsumedCCS no_cc)
117 && not (isSccCountCostCentre cc)
118 -- Top level CAF without a cost centre attached
119 -- Attach and collect cc of trivial _scc_ in body
120 = collectCC cc `thenMM_`
121 set_prevailing_cc cc (do_expr expr) `thenMM` \ expr' ->
122 returnMM (StgRhsClosure cc bi srt fv u [] expr')
125 do_top_rhs binder (StgRhsClosure no_cc bi srt fv u [] body)
126 | noCCSAttached no_cc || currentOrSubsumedCCS no_cc
127 -- Top level CAF without a cost centre attached
128 -- Attach CAF cc (collect if individual CAF ccs)
129 = (if opt_AutoSccsOnIndividualCafs
130 then let cc = mkAutoCC binder mod_name CafCC
131 ccs = mkSingletonCCS cc
133 collectCC cc `thenMM_`
134 collectCCS ccs `thenMM_`
137 returnMM all_cafs_ccs) `thenMM` \ caf_ccs ->
138 set_prevailing_cc caf_ccs (do_expr body) `thenMM` \ body' ->
139 returnMM (StgRhsClosure caf_ccs bi srt fv u [] body')
141 do_top_rhs binder (StgRhsClosure cc bi srt fv u [] body)
142 -- Top level CAF with cost centre attached
143 -- Should this be a CAF cc ??? Does this ever occur ???
144 = pprPanic "SCCfinal: CAF with cc:" (ppr cc)
146 {- can't do this with cost-centre stacks: --SDM
147 do_top_rhs binder (StgRhsClosure _ bi srt fv u args (StgSCC cc expr))
148 | not (isSccCountCostCentre cc)
149 -- Top level function with trivial _scc_ in body
150 -- Attach and collect cc of trivial _scc_
151 = collectCC cc `thenMM_`
152 set_prevailing_cc cc (do_expr expr) `thenMM` \ expr' ->
153 returnMM (StgRhsClosure cc bi srt fv u args expr')
156 do_top_rhs binder (StgRhsClosure no_ccs bi srt fv u args body)
157 -- Top level function, probably subsumed
158 | noCCSAttached no_ccs
159 = set_lambda_cc (do_expr body) `thenMM` \ body' ->
160 returnMM (StgRhsClosure subsumedCCS bi srt fv u args body')
163 = pprPanic "SCCfinal: CAF with cc:" (ppr no_ccs)
165 do_top_rhs binder (StgRhsCon ccs con args)
166 -- Top-level (static) data is not counted in heap
167 -- profiles; nor do we set CCCS from it; so we
168 -- just slam in dontCareCostCentre
169 = returnMM (StgRhsCon dontCareCCS con args)
172 do_expr :: StgExpr -> MassageM StgExpr
174 do_expr (StgLit l) = returnMM (StgLit l)
176 do_expr (StgApp fn args)
177 = boxHigherOrderArgs (StgApp fn) args
179 do_expr (StgConApp con args)
180 = boxHigherOrderArgs (\args -> StgConApp con args) args
182 do_expr (StgPrimApp con args res_ty)
183 = boxHigherOrderArgs (\args -> StgPrimApp con args res_ty) args
185 do_expr (StgSCC cc expr) -- Ha, we found a cost centre!
186 = collectCC cc `thenMM_`
187 do_expr expr `thenMM` \ expr' ->
188 returnMM (StgSCC cc expr')
190 do_expr (StgCase expr fv1 fv2 bndr srt alts)
191 = do_expr expr `thenMM` \ expr' ->
192 do_alts alts `thenMM` \ alts' ->
193 returnMM (StgCase expr' fv1 fv2 bndr srt alts')
195 do_alts (StgAlgAlts ty alts def)
196 = mapMM do_alt alts `thenMM` \ alts' ->
197 do_deflt def `thenMM` \ def' ->
198 returnMM (StgAlgAlts ty alts' def')
200 do_alt (id, bs, use_mask, e)
201 = do_expr e `thenMM` \ e' ->
202 returnMM (id, bs, use_mask, e')
204 do_alts (StgPrimAlts ty alts def)
205 = mapMM do_alt alts `thenMM` \ alts' ->
206 do_deflt def `thenMM` \ def' ->
207 returnMM (StgPrimAlts ty alts' def')
210 = do_expr e `thenMM` \ e' ->
213 do_deflt StgNoDefault = returnMM StgNoDefault
214 do_deflt (StgBindDefault e)
215 = do_expr e `thenMM` \ e' ->
216 returnMM (StgBindDefault e')
219 = do_let b e `thenMM` \ (b,e) ->
220 returnMM (StgLet b e)
222 do_expr (StgLetNoEscape lvs1 lvs2 b e)
223 = do_let b e `thenMM` \ (b,e) ->
224 returnMM (StgLetNoEscape lvs1 lvs2 b e)
226 ----------------------------------
228 do_let (StgNonRec b rhs) e
229 = do_rhs rhs `thenMM` \ rhs' ->
231 do_expr e `thenMM` \ e' ->
232 returnMM (StgNonRec b rhs',e')
235 do_let (StgRec pairs) e
236 = addTopLevelIshIds binders (
237 mapMM do_pair pairs `thenMM` \ pairs' ->
238 do_expr e `thenMM` \ e' ->
239 returnMM (StgRec pairs', e')
242 binders = map fst pairs
244 = do_rhs rhs `thenMM` \ rhs2 ->
247 ----------------------------------
248 do_rhs :: StgRhs -> MassageM StgRhs
249 -- We play much the same game as we did in do_top_rhs above;
250 -- but we don't have to worry about cafs etc.
253 do_rhs (StgRhsClosure closure_cc bi srt fv u [] (StgSCC ty cc (StgCon (DataCon con) args _)))
254 | not (isSccCountCostCentre cc)
255 = collectCC cc `thenMM_`
256 returnMM (StgRhsCon cc con args)
260 do_rhs (StgRhsClosure _ bi srt fv u args (StgSCC ty cc expr))
261 | not (isSccCountCostCentre cc)
262 = collectCC cc `thenMM_`
263 set_prevailing_cc cc (do_expr expr) `thenMM` \ expr' ->
264 returnMM (StgRhsClosure cc bi srt fv u args expr')
267 do_rhs (StgRhsClosure cc bi srt fv u [] body)
268 = do_expr body `thenMM` \ body' ->
269 returnMM (StgRhsClosure currentCCS bi srt fv u [] body')
271 do_rhs (StgRhsClosure cc bi srt fv u args body)
272 = set_lambda_cc (do_expr body) `thenMM` \ body' ->
273 get_prevailing_cc `thenMM` \ prev_ccs ->
274 returnMM (StgRhsClosure currentCCS bi srt fv u args body')
276 do_rhs (StgRhsCon cc con args)
277 = returnMM (StgRhsCon currentCCS con args)
280 %************************************************************************
282 \subsection{Boxing higher-order args}
284 %************************************************************************
286 Boxing is *turned off* at the moment, until we can figure out how to
287 do it properly in general.
291 :: ([StgArg] -> StgExpr)
292 -- An application lacking its arguments
293 -> [StgArg] -- arguments which we might box
296 #ifndef PROF_DO_BOXING
297 boxHigherOrderArgs almost_expr args
298 = returnMM (almost_expr args)
300 boxHigherOrderArgs almost_expr args
301 = getTopLevelIshIds `thenMM` \ ids ->
302 mapAccumMM (do_arg ids) [] args `thenMM` \ (let_bindings, new_args) ->
303 returnMM (foldr (mk_stg_let currentCCS) (almost_expr new_args) let_bindings)
307 do_arg ids bindings arg@(StgVarArg old_var)
308 | (not (isLocallyDefined old_var) || elemVarSet old_var ids)
309 && isFunType var_type
310 = -- make a trivial let-binding for the top-level function
311 getUniqueMM `thenMM` \ uniq ->
313 new_var = mkSysLocal SLIT("sf") uniq var_type
315 returnMM ( (new_var, old_var) : bindings, StgVarArg new_var )
317 var_type = idType old_var
319 do_arg ids bindings arg = returnMM (bindings, arg)
322 mk_stg_let :: CostCentreStack -> (Id, Id) -> StgExpr -> StgExpr
324 mk_stg_let cc (new_var, old_var) body
326 rhs_body = StgApp old_var [{-args-}]
327 rhs_closure = StgRhsClosure cc stgArgOcc NoSRT [{-fvs-}] ReEntrant [{-args-}] rhs_body
329 StgLet (StgNonRec new_var rhs_closure) body
331 bOGUS_LVs = emptyUniqSet -- easier to print than: panic "mk_stg_let: LVs"
334 = case splitForAllTys var_type of
335 (_, ty) -> case splitTyConApp_maybe ty of
336 Just (tycon,_) | isFunTyCon tycon -> True
341 %************************************************************************
343 \subsection{Boring monad stuff for this}
345 %************************************************************************
349 = Module -- module name
350 -> CostCentreStack -- prevailing CostCentre
351 -- if none, subsumedCosts at top-level
352 -- useCurrentCostCentre at nested levels
354 -> VarSet -- toplevel-ish Ids for boxing
356 -> (CollectedCCs, result)
358 -- the initMM function also returns the final CollectedCCs
360 initMM :: Module -- module name, which we may consult
365 initMM mod_name init_us m = m mod_name noCCS init_us emptyVarSet ([],[],[])
367 thenMM :: MassageM a -> (a -> MassageM b) -> MassageM b
368 thenMM_ :: MassageM a -> (MassageM b) -> MassageM b
370 thenMM expr cont mod scope_cc us ids ccs
371 = case splitUniqSupply us of { (s1, s2) ->
372 case (expr mod scope_cc s1 ids ccs) of { (ccs2, result) ->
373 cont result mod scope_cc s2 ids ccs2 }}
375 thenMM_ expr cont mod scope_cc us ids ccs
376 = case splitUniqSupply us of { (s1, s2) ->
377 case (expr mod scope_cc s1 ids ccs) of { (ccs2, _) ->
378 cont mod scope_cc s2 ids ccs2 }}
380 returnMM :: a -> MassageM a
381 returnMM result mod scope_cc us ids ccs = (ccs, result)
384 nopMM mod scope_cc us ids ccs = (ccs, ())
386 mapMM :: (a -> MassageM b) -> [a] -> MassageM [b]
387 mapMM f [] = returnMM []
389 = f m `thenMM` \ r ->
390 mapMM f ms `thenMM` \ rs ->
393 mapAccumMM :: (acc -> x -> MassageM (acc, y)) -> acc -> [x] -> MassageM (acc, [y])
394 mapAccumMM f b [] = returnMM (b, [])
395 mapAccumMM f b (m:ms)
396 = f b m `thenMM` \ (b2, r) ->
397 mapAccumMM f b2 ms `thenMM` \ (b3, rs) ->
400 getUniqueMM :: MassageM Unique
401 getUniqueMM mod scope_cc us ids ccs = (ccs, uniqFromSupply us)
403 addTopLevelIshId :: Id -> MassageM a -> MassageM a
404 addTopLevelIshId id scope mod scope_cc us ids ccs
405 | isCurrentCCS scope_cc = scope mod scope_cc us ids ccs
406 | otherwise = scope mod scope_cc us (extendVarSet ids id) ccs
408 addTopLevelIshIds :: [Id] -> MassageM a -> MassageM a
409 addTopLevelIshIds [] cont = cont
410 addTopLevelIshIds (id:ids) cont
411 = addTopLevelIshId id (addTopLevelIshIds ids cont)
413 getTopLevelIshIds :: MassageM VarSet
414 getTopLevelIshIds mod scope_cc us ids ccs = (ccs, ids)
417 The prevailing CCS is used to tell whether we're in a top-levelish
418 position, where top-levelish is defined as "not inside a lambda".
419 Prevailing CCs used to be used for something much more complicated,
423 set_lambda_cc :: MassageM a -> MassageM a
424 set_lambda_cc action mod scope_cc us ids ccs
425 = action mod currentCCS us ids ccs
427 set_prevailing_cc :: CostCentreStack -> MassageM a -> MassageM a
428 set_prevailing_cc cc_to_set_to action mod scope_cc us ids ccs
429 = action mod cc_to_set_to us ids ccs
431 get_prevailing_cc :: MassageM CostCentreStack
432 get_prevailing_cc mod scope_cc us ids ccs = (ccs, scope_cc)
436 collectCC :: CostCentre -> MassageM ()
438 collectCC cc mod_name scope_cc us ids (local_ccs, extern_ccs, ccss)
439 = ASSERT(not (noCCAttached cc))
440 if (cc `ccFromThisModule` mod_name) then
441 ((cc : local_ccs, extern_ccs, ccss), ())
442 else -- must declare it "extern"
443 ((local_ccs, cc : extern_ccs, ccss), ())
445 collectCCS :: CostCentreStack -> MassageM ()
447 collectCCS ccs mod_name scope_cc us ids (local_ccs, extern_ccs, ccss)
448 = ASSERT(not (noCCSAttached ccs))
449 ((local_ccs, extern_ccs, ccs : ccss), ())