2 ( ProcPointSet, Status(..)
3 , callProcPoints, minimalProcPointSet
4 , addProcPointProtocols, splitAtProcPoints, procPointAnalysis
8 import Prelude hiding (zip, unzip, last)
12 import Cmm hiding (blockId)
23 import MkZipCfgCmm hiding (CmmBlock, CmmGraph, CmmTopZ)
33 -- Compute a minimal set of proc points for a control-flow graph.
35 -- Determine a protocol for each proc point (which live variables will
36 -- be passed as arguments and which will be on the stack).
39 A proc point is a basic block that, after CPS transformation, will
40 start a new function. The entry block of the original function is a
41 proc point, as is the continuation of each function call.
42 A third kind of proc point arises if we want to avoid copying code.
43 Suppose we have code like the following:
46 if (...) { ..1..; call foo(); ..2..}
47 else { ..3..; call bar(); ..4..}
52 The statement 'x = y + z' can be reached from two different proc
53 points: the continuations of foo() and bar(). We would prefer not to
54 put a copy in each continuation; instead we would like 'x = y + z' to
55 be the start of a new procedure to which the continuations can jump:
58 if (...) { ..1..; push k_foo; jump foo_cps(); }
59 else { ..3..; push k_bar; jump bar_cps(); }
61 k_foo() { ..2..; jump k_join(y, z); }
62 k_bar() { ..4..; jump k_join(y, z); }
63 k_join(y, z) { x = y + z; return x; }
65 You might think then that a criterion to make a node a proc point is
66 that it is directly reached by two distinct proc points. (Note
67 [Direct reachability].) But this criterion is a bit too simple; for
68 example, 'return x' is also reached by two proc points, yet there is
69 no point in pulling it out of k_join. A good criterion would be to
70 say that a node should be made a proc point if it is reached by a set
71 of proc points that is different than its immediate dominator. NR
72 believes this criterion can be shown to produce a minimum set of proc
73 points, and given a dominator tree, the proc points can be chosen in
74 time linear in the number of blocks. Lacking a dominator analysis,
75 however, we turn instead to an iterative solution, starting with no
76 proc points and adding them according to these rules:
78 1. The entry block is a proc point.
79 2. The continuation of a call is a proc point.
80 3. A node is a proc point if it is directly reached by more proc
81 points than one of its predecessors.
83 Because we don't understand the problem very well, we apply rule 3 at
84 most once per iteration, then recompute the reachability information.
85 (See Note [No simple dataflow].) The choice of the new proc point is
86 arbitrary, and I don't know if the choice affects the final solution,
87 so I don't know if the number of proc points chosen is the
88 minimum---but the set will be minimal.
91 type ProcPointSet = BlockSet
94 = ReachedBy ProcPointSet -- set of proc points that directly reach the block
95 | ProcPoint -- this block is itself a proc point
97 instance Outputable Status where
99 | isEmptyBlockSet ps = text "<not-reached>"
100 | otherwise = text "reached by" <+>
101 (hsep $ punctuate comma $ map ppr $ blockSetToList ps)
102 ppr ProcPoint = text "<procpt>"
105 lattice :: DataflowLattice Status
106 lattice = DataflowLattice "direct proc-point reachability" unreached add_to False
107 where unreached = ReachedBy emptyBlockSet
108 add_to _ ProcPoint = noTx ProcPoint
109 add_to ProcPoint _ = aTx ProcPoint -- aTx because of previous case again
110 add_to (ReachedBy p) (ReachedBy p') =
111 let union = unionBlockSets p p'
112 in if sizeBlockSet union > sizeBlockSet p' then
113 aTx (ReachedBy union)
116 --------------------------------------------------
117 -- transfer equations
119 forward :: ForwardTransfers Middle Last Status
120 forward = ForwardTransfers first middle last exit
121 where first id ProcPoint = ReachedBy $ unitBlockSet id
124 last (LastCall _ (Just id) _ _ _) _ = LastOutFacts [(id, ProcPoint)]
125 last l x = LastOutFacts $ map (\id -> (id, x)) (succs l)
128 -- It is worth distinguishing two sets of proc points:
129 -- those that are induced by calls in the original graph
130 -- and those that are introduced because they're reachable from multiple proc points.
131 callProcPoints :: CmmGraph -> ProcPointSet
132 minimalProcPointSet :: ProcPointSet -> CmmGraph -> FuelMonad ProcPointSet
134 callProcPoints g = fold_blocks add (unitBlockSet (lg_entry g)) g
135 where add b set = case last $ unzip b of
136 LastOther (LastCall _ (Just k) _ _ _) -> extendBlockSet set k
139 minimalProcPointSet callProcPoints g = extendPPSet g (postorder_dfs g) callProcPoints
141 type PPFix = FuelMonad (ForwardFixedPoint Middle Last Status ())
143 procPointAnalysis :: ProcPointSet -> CmmGraph -> FuelMonad (BlockEnv Status)
144 procPointAnalysis procPoints g =
145 let addPP env id = extendBlockEnv env id ProcPoint
146 initProcPoints = foldl addPP emptyBlockEnv (blockSetToList procPoints)
147 in liftM zdfFpFacts $
148 (zdfSolveFrom initProcPoints "proc-point reachability" lattice
149 forward (fact_bot lattice) $ graphOfLGraph g :: PPFix)
151 extendPPSet :: CmmGraph -> [CmmBlock] -> ProcPointSet -> FuelMonad ProcPointSet
152 extendPPSet g blocks procPoints =
153 do env <- procPointAnalysis procPoints g
154 let add block pps = let id = blockId block
155 in case lookupBlockEnv env id of
156 Just ProcPoint -> extendBlockSet pps id
158 procPoints' = fold_blocks add emptyBlockSet g
159 newPoints = mapMaybe ppSuccessor blocks
160 newPoint = listToMaybe newPoints
161 ppSuccessor b@(Block bid _) =
162 let nreached id = case lookupBlockEnv env id `orElse`
163 pprPanic "no ppt" (ppr id <+> ppr b) of
165 ReachedBy ps -> sizeBlockSet ps
166 block_procpoints = nreached bid
167 -- | Looking for a successor of b that is reached by
168 -- more proc points than b and is not already a proc
169 -- point. If found, it can become a proc point.
170 newId succ_id = not (elemBlockSet succ_id procPoints') &&
171 nreached succ_id > block_procpoints
172 in listToMaybe $ filter newId $ succs b
175 [] -> return procPoints'
176 pps -> extendPPSet g blocks
177 (foldl extendBlockSet procPoints' pps)
179 case newPoint of Just id ->
180 if elemBlockSet id procPoints' then panic "added old proc pt"
181 else extendPPSet g blocks (extendBlockSet procPoints' id)
182 Nothing -> return procPoints'
185 ------------------------------------------------------------------------
186 -- Computing Proc-Point Protocols --
187 ------------------------------------------------------------------------
191 There is one major trick, discovered by Michael Adams, which is that
192 we want to choose protocols in a way that enables us to optimize away
193 some continuations. The optimization is very much like branch-chain
194 elimination, except that it involves passing results as well as
195 control. The idea is that if a call's continuation k does nothing but
196 CopyIn its results and then goto proc point P, the call's continuation
197 may be changed to P, *provided* P's protocol is identical to the
198 protocol for the CopyIn. We choose protocols to make this so.
200 Here's an explanatory example; we begin with the source code (lines
201 separate basic blocks):
209 Zipperization converts this code as follows:
212 call g() returns to k;
219 What we'd like to do is assign P the same CopyIn protocol as k, so we
223 call g() returns to P;
225 P: CopyIn(x, y); ..2..;
227 Of course, P may be the target of more than one continuation, and
228 different continuations may have different protocols. Michael Adams
229 implemented a voting mechanism, but he thinks a simple greedy
230 algorithm would be just as good, so that's what we do.
234 data Protocol = Protocol Convention CmmFormals Area
236 instance Outputable Protocol where
237 ppr (Protocol c fs a) = text "Protocol" <+> ppr c <+> ppr fs <+> ppr a
239 -- | Function 'optimize_calls' chooses protocols only for those proc
240 -- points that are relevant to the optimization explained above.
241 -- The others are assigned by 'add_unassigned', which is not yet clever.
243 addProcPointProtocols :: ProcPointSet -> ProcPointSet -> CmmGraph -> FuelMonad CmmGraph
244 addProcPointProtocols callPPs procPoints g =
245 do liveness <- cmmLivenessZ g
246 (protos, g') <- optimize_calls liveness g
247 blocks'' <- add_CopyOuts protos procPoints g'
248 return $ LGraph (lg_entry g) blocks''
249 where optimize_calls liveness g = -- see Note [Separate Adams optimization]
250 do let (protos, blocks') =
251 fold_blocks maybe_add_call (init_protocols, emptyBlockEnv) g
252 protos' = add_unassigned liveness procPoints protos
253 blocks <- add_CopyIns callPPs protos' blocks'
254 let g' = LGraph (lg_entry g) (mkBlockEnv (map withKey (concat blocks)))
255 withKey b@(Block bid _) = (bid, b)
256 return (protos', runTx removeUnreachableBlocksZ g')
257 maybe_add_call :: CmmBlock -> (BlockEnv Protocol, BlockEnv CmmBlock)
258 -> (BlockEnv Protocol, BlockEnv CmmBlock)
259 -- ^ If the block is a call whose continuation goes to a proc point
260 -- whose protocol either matches the continuation's or is not yet set,
261 -- redirect the call (cf 'newblock') and set the protocol if necessary
262 maybe_add_call block (protos, blocks) =
263 case goto_end $ unzip block of
264 (h, LastOther (LastCall tgt (Just k) args res s))
265 | Just proto <- lookupBlockEnv protos k,
266 Just pee <- branchesToProcPoint k
267 -> let newblock = zipht h (tailOfLast (LastCall tgt (Just pee)
269 changed_blocks = insertBlock newblock blocks
270 unchanged_blocks = insertBlock block blocks
271 in case lookupBlockEnv protos pee of
272 Nothing -> (extendBlockEnv protos pee proto,changed_blocks)
274 if proto == proto' then (protos, changed_blocks)
275 else (protos, unchanged_blocks)
276 _ -> (protos, insertBlock block blocks)
278 branchesToProcPoint :: BlockId -> Maybe BlockId
279 -- ^ Tells whether the named block is just a branch to a proc point
280 branchesToProcPoint id =
281 let (Block _ t) = lookupBlockEnv (lg_blocks g) id `orElse`
282 panic "branch out of graph"
284 ZLast (LastOther (LastBranch pee))
285 | elemBlockSet pee procPoints -> Just pee
287 init_protocols = fold_blocks maybe_add_proto emptyBlockEnv g
288 maybe_add_proto :: CmmBlock -> BlockEnv Protocol -> BlockEnv Protocol
289 --maybe_add_proto (Block id (ZTail (CopyIn c _ fs _srt) _)) env =
290 -- extendBlockEnv env id (Protocol c fs $ toArea id fs)
291 maybe_add_proto _ env = env
292 -- JD: Is this proto stuff even necessary, now that we have
293 -- common blockification?
295 -- | For now, following a suggestion by Ben Lippmeier, we pass all
296 -- live variables as arguments, hoping that a clever register
297 -- allocator might help.
299 add_unassigned :: BlockEnv CmmLive -> ProcPointSet -> BlockEnv Protocol ->
301 add_unassigned = pass_live_vars_as_args
303 pass_live_vars_as_args :: BlockEnv CmmLive -> ProcPointSet ->
304 BlockEnv Protocol -> BlockEnv Protocol
305 pass_live_vars_as_args _liveness procPoints protos = protos'
306 where protos' = foldBlockSet addLiveVars protos procPoints
307 addLiveVars :: BlockId -> BlockEnv Protocol -> BlockEnv Protocol
308 addLiveVars id protos =
309 case lookupBlockEnv protos id of
311 Nothing -> let live = emptyRegSet
312 --lookupBlockEnv _liveness id `orElse`
313 --panic ("no liveness at block " ++ show id)
314 formals = uniqSetToList live
315 prot = Protocol Private formals $ CallArea $ Young id
316 in extendBlockEnv protos id prot
319 -- | Add copy-in instructions to each proc point that did not arise from a call
320 -- instruction. (Proc-points that arise from calls already have their copy-in instructions.)
322 add_CopyIns :: ProcPointSet -> BlockEnv Protocol -> BlockEnv CmmBlock ->
323 FuelMonad [[CmmBlock]]
324 add_CopyIns callPPs protos blocks =
325 liftUniq $ mapM maybe_insert_CopyIns (blockEnvToList blocks)
326 where maybe_insert_CopyIns (_, b@(Block id t))
327 | not $ elemBlockSet id callPPs
328 = case lookupBlockEnv protos id of
329 Just (Protocol c fs _area) ->
330 do LGraph _ blocks <-
331 lgraphOfAGraph (mkLabel id <*> copyInSlot c fs <*> mkZTail t)
332 return (map snd $ blockEnvToList blocks)
333 Nothing -> return [b]
334 | otherwise = return [b]
336 -- | Add a CopyOut node before each procpoint.
337 -- If the predecessor is a call, then the copy outs should already be done by the callee.
338 -- Note: If we need to add copy-out instructions, they may require stack space,
339 -- so we accumulate a map from the successors to the necessary stack space,
340 -- then update the successors after we have finished inserting the copy-outs.
342 add_CopyOuts :: BlockEnv Protocol -> ProcPointSet -> CmmGraph ->
343 FuelMonad (BlockEnv CmmBlock)
344 add_CopyOuts protos procPoints g = fold_blocks mb_copy_out (return emptyBlockEnv) g
345 where mb_copy_out :: CmmBlock -> FuelMonad (BlockEnv CmmBlock) ->
346 FuelMonad (BlockEnv CmmBlock)
347 mb_copy_out b@(Block bid _) z | bid == lg_entry g = skip b z
349 case last $ unzip b of
350 LastOther (LastCall _ _ _ _ _) -> skip b z -- copy out done by callee
352 copy_out b z = fold_succs trySucc b init >>= finish
353 where init = z >>= (\bmap -> return (b, bmap))
355 if elemBlockSet succId procPoints then
356 case lookupBlockEnv protos succId of
358 Just (Protocol c fs _area) -> insert z succId $ copyOutSlot c fs
362 (b, bs) <- insertBetween b m succId
363 -- pprTrace "insert for succ" (ppr succId <> ppr m) $ do
364 return $ (b, foldl (flip insertBlock) bmap bs)
365 finish (b@(Block bid _), bmap) =
366 return $ (extendBlockEnv bmap bid b)
367 skip b@(Block bid _) bs =
368 bs >>= (\bmap -> return (extendBlockEnv bmap bid b))
370 -- At this point, we have found a set of procpoints, each of which should be
371 -- the entry point of a procedure.
372 -- Now, we create the procedure for each proc point,
373 -- which requires that we:
374 -- 1. build a map from proc points to the blocks reachable from the proc point
375 -- 2. turn each branch to a proc point into a jump
376 -- 3. turn calls and returns into jumps
377 -- 4. build info tables for the procedures -- and update the info table for
378 -- the SRTs in the entry procedure as well.
379 -- Input invariant: A block should only be reachable from a single ProcPoint.
380 splitAtProcPoints :: CLabel -> ProcPointSet-> ProcPointSet -> BlockEnv Status ->
381 CmmTopZ -> FuelMonad [CmmTopZ]
382 splitAtProcPoints entry_label callPPs procPoints procMap
383 (CmmProc (CmmInfo gc upd_fr info_tbl) top_l top_args
384 (stackInfo, g@(LGraph entry blocks))) =
385 do -- Build a map from procpoints to the blocks they reach
386 let addBlock b@(Block bid _) graphEnv =
387 case lookupBlockEnv procMap bid of
388 Just ProcPoint -> add graphEnv bid bid b
389 Just (ReachedBy set) ->
390 case blockSetToList set of
392 [id] -> add graphEnv id bid b
393 _ -> panic "Each block should be reachable from only one ProcPoint"
394 Nothing -> pprPanic "block not reached by a proc point?" (ppr bid)
395 add graphEnv procId bid b = extendBlockEnv graphEnv procId graph'
396 where graph = lookupBlockEnv graphEnv procId `orElse` emptyBlockEnv
397 graph' = extendBlockEnv graph bid b
398 graphEnv <- return $ fold_blocks addBlock emptyBlockEnv g
399 -- Build a map from proc point BlockId to labels for their new procedures
400 -- Due to common blockification, we may overestimate the set of procpoints.
401 let add_label map pp = return $ addToFM map pp lbl
402 where lbl = if pp == entry then entry_label else blockLbl pp
403 procLabels <- foldM add_label emptyFM
404 (filter (elemBlockEnv blocks) (blockSetToList procPoints))
405 -- For each procpoint, we need to know the SP offset on entry.
406 -- If the procpoint is:
407 -- - continuation of a call, the SP offset is in the call
408 -- - otherwise, 0 -- no overflow for passing those variables
409 let add_sp_off b env =
410 case last (unzip b) of
411 LastOther (LastCall {cml_cont = Just succ, cml_ret_args = off,
412 cml_ret_off = updfr_off}) ->
413 extendBlockEnv env succ (off, updfr_off)
415 spEntryMap = fold_blocks add_sp_off (mkBlockEnv [(entry, stackInfo)]) g
416 getStackInfo id = lookupBlockEnv spEntryMap id `orElse` (0, Nothing)
417 -- In each new graph, add blocks jumping off to the new procedures,
418 -- and replace branches to procpoints with branches to the jump-off blocks
419 let add_jump_block (env, bs) (pp, l) =
420 do bid <- liftM mkBlockId getUniqueM
421 let b = Block bid (ZLast (LastOther jump))
422 (argSpace, _) = getStackInfo pp
423 jump = LastCall (CmmLit (CmmLabel l')) Nothing argSpace 0 Nothing
424 l' = if elemBlockSet pp callPPs then entryLblToInfoLbl l else l
425 return (extendBlockEnv env pp bid, b : bs)
426 add_jumps (newGraphEnv) (ppId, blockEnv) =
427 do let needed_jumps = -- find which procpoints we currently branch to
428 foldBlockEnv' add_if_branch_to_pp [] blockEnv
429 add_if_branch_to_pp block rst =
430 case last (unzip block) of
431 LastOther (LastBranch id) -> add_if_pp id rst
432 LastOther (LastCondBranch _ ti fi) ->
433 add_if_pp ti (add_if_pp fi rst)
434 LastOther (LastSwitch _ tbl) -> foldr add_if_pp rst (catMaybes tbl)
436 add_if_pp id rst = case lookupFM procLabels id of
437 Just x -> (id, x) : rst
439 (jumpEnv, jumpBlocks) <-
440 foldM add_jump_block (emptyBlockEnv, []) needed_jumps
441 -- update the entry block
442 let b = expectJust "block in env" $ lookupBlockEnv blockEnv ppId
443 off = getStackInfo ppId
444 blockEnv' = extendBlockEnv blockEnv ppId b
445 -- replace branches to procpoints with branches to jumps
446 LGraph _ blockEnv'' = replaceBranches jumpEnv $ LGraph ppId blockEnv'
447 -- add the jump blocks to the graph
448 blockEnv''' = foldl (flip insertBlock) blockEnv'' jumpBlocks
449 let g' = (off, LGraph ppId blockEnv''')
450 -- pprTrace "g' pre jumps" (ppr g') $ do
451 return (extendBlockEnv newGraphEnv ppId g')
452 graphEnv <- foldM add_jumps emptyBlockEnv $ blockEnvToList graphEnv
453 let to_proc (bid, g) | elemBlockSet bid callPPs =
455 CmmProc (CmmInfo gc upd_fr info_tbl) top_l top_args (replacePPIds g)
457 CmmProc emptyContInfoTable lbl [] (replacePPIds g)
458 where lbl = expectJust "pp label" $ lookupFM procLabels bid
460 CmmProc (CmmInfo Nothing Nothing CmmNonInfoTable) lbl [] (replacePPIds g)
461 where lbl = expectJust "pp label" $ lookupFM procLabels bid
462 -- References to procpoint IDs can now be replaced with the infotable's label
463 replacePPIds (x, g) = (x, map_nodes id (mapExpMiddle repl) (mapExpLast repl) g)
464 where repl e@(CmmLit (CmmBlock bid)) =
465 case lookupFM procLabels bid of
466 Just l -> CmmLit (CmmLabel (entryLblToInfoLbl l))
469 -- The C back end expects to see return continuations before the call sites.
470 -- Here, we sort them in reverse order -- it gets reversed later.
471 let (_, block_order) = foldl add_block_num (0::Int, emptyBlockEnv) (postorder_dfs g)
472 add_block_num (i, map) (Block bid _) = (i+1, extendBlockEnv map bid i)
473 sort_fn (bid, _) (bid', _) =
474 compare (expectJust "block_order" $ lookupBlockEnv block_order bid)
475 (expectJust "block_order" $ lookupBlockEnv block_order bid')
476 procs <- return $ map to_proc $ sortBy sort_fn $ blockEnvToList graphEnv
477 return -- pprTrace "procLabels" (ppr procLabels)
478 -- pprTrace "splitting graphs" (ppr procs)
480 splitAtProcPoints _ _ _ _ t@(CmmData _ _) = return [t]
482 ----------------------------------------------------------------
485 Note [Direct reachability]
487 Block B is directly reachable from proc point P iff control can flow
488 from P to B without passing through an intervening proc point.
491 ----------------------------------------------------------------
494 Note [No simple dataflow]
496 Sadly, it seems impossible to compute the proc points using a single
497 dataflow pass. One might attempt to use this simple lattice:
499 data Location = Unknown
500 | InProc BlockId -- node is in procedure headed by the named proc point
501 | ProcPoint -- node is itself a proc point
503 At a join, a node in two different blocks becomes a proc point.
504 The difficulty is that the change of information during iterative
505 computation may promote a node prematurely. Here's a program that
506 illustrates the difficulty:
515 L2: if (...) { g(); goto L1; }
519 The only proc-point needed (besides the entry) is L1. But in an
520 iterative analysis, consider what happens to L2. On the first pass
521 through, it rises from Unknown to 'InProc entry', but when L1 is
522 promoted to a proc point (because it's the successor of g()), L1's
523 successors will be promoted to 'InProc L1'. The problem hits when the
524 new fact 'InProc L1' flows into L2 which is already bound to 'InProc entry'.
525 The join operation makes it a proc point when in fact it needn't be,
526 because its immediate dominator L1 is already a proc point and there
527 are no other proc points that directly reach L2.
532 {- Note [Separate Adams optimization]
533 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
534 It may be worthwhile to attempt the Adams optimization by rewriting
535 the graph before the assignment of proc-point protocols. Here are a
538 g() returns to k; g() returns to L;
539 k: CopyIn c ress; goto L:
541 L: // no CopyIn node here L: CopyIn c ress;
544 And when c == c' and ress == ress', this also:
546 g() returns to k; g() returns to L;
547 k: CopyIn c ress; goto L:
549 L: CopyIn c' ress' L: CopyIn c' ress' ;
551 In both cases the goal is to eliminate k.