CmmContFlowOpt.hs 15.6 KB
Newer Older
1
{-# LANGUAGE GADTs #-}
2
{-# OPTIONS_GHC -fno-warn-incomplete-patterns #-}
3
module CmmContFlowOpt
4
    ( cmmCfgOpts
Simon Marlow's avatar
Simon Marlow committed
5
    , cmmCfgOptsProc
6
    , removeUnreachableBlocksProc
7
    , replaceLabels
8 9 10
    )
where

11
import Hoopl
12
import BlockId
13
import Cmm
14
import CmmUtils
15
import CmmSwitch (mapSwitchTargets)
16
import Maybes
17
import Panic
18
import Util
19

Ian Lynagh's avatar
Ian Lynagh committed
20
import Control.Monad
21
import Prelude hiding (succ, unzip, zip)
22

23 24 25 26 27 28 29 30 31 32

-- Note [What is shortcutting]
-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~
--
-- Consider this Cmm code:
--
-- L1: ...
--     goto L2;
-- L2: goto L3;
-- L3: ...
33
--
34 35 36
-- Here L2 is an empty block and contains only an unconditional branch
-- to L3. In this situation any block that jumps to L2 can jump
-- directly to L3:
37
--
38 39 40 41 42 43 44 45 46 47 48 49 50
-- L1: ...
--     goto L3;
-- L2: goto L3;
-- L3: ...
--
-- In this situation we say that we shortcut L2 to L3. One of
-- consequences of shortcutting is that some blocks of code may become
-- unreachable (in the example above this is true for L2).


-- Note [Control-flow optimisations]
-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
--
51
-- This optimisation does three things:
52 53 54 55 56 57 58 59 60 61 62
--
--   - If a block finishes in an unconditonal branch to another block
--     and that is the only jump to that block we concatenate the
--     destination block at the end of the current one.
--
--   - If a block finishes in a call whose continuation block is a
--     goto, then we can shortcut the destination, making the
--     continuation block the destination of the goto - but see Note
--     [Shortcut call returns].
--
--   - For any block that is not a call we try to shortcut the
63 64
--     destination(s). Additionally, if a block ends with a
--     conditional branch we try to invert the condition.
65 66 67 68 69 70 71 72 73 74 75 76 77
--
-- Blocks are processed using postorder DFS traversal. A side effect
-- of determining traversal order with a graph search is elimination
-- of any blocks that are unreachable.
--
-- Transformations are improved by working from the end of the graph
-- towards the beginning, because we may be able to perform many
-- shortcuts in one go.


-- Note [Shortcut call returns]
-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
--
78
-- We are going to maintain the "current" graph (LabelMap CmmBlock) as
79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133
-- we go, and also a mapping from BlockId to BlockId, representing
-- continuation labels that we have renamed.  This latter mapping is
-- important because we might shortcut a CmmCall continuation.  For
-- example:
--
--    Sp[0] = L
--    call g returns to L
--    L: goto M
--    M: ...
--
-- So when we shortcut the L block, we need to replace not only
-- the continuation of the call, but also references to L in the
-- code (e.g. the assignment Sp[0] = L):
--
--    Sp[0] = M
--    call g returns to M
--    M: ...
--
-- So we keep track of which labels we have renamed and apply the mapping
-- at the end with replaceLabels.


-- Note [Shortcut call returns and proc-points]
-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
--
-- Consider this code that you might get from a recursive
-- let-no-escape:
--
--       goto L1
--      L1:
--       if (Hp > HpLim) then L2 else L3
--      L2:
--       call stg_gc_noregs returns to L4
--      L4:
--       goto L1
--      L3:
--       ...
--       goto L1
--
-- Then the control-flow optimiser shortcuts L4.  But that turns L1
-- into the call-return proc point, and every iteration of the loop
-- has to shuffle variables to and from the stack.  So we must *not*
-- shortcut L4.
--
-- Moreover not shortcutting call returns is probably fine.  If L4 can
-- concat with its branch target then it will still do so.  And we
-- save some compile time because we don't have to traverse all the
-- code in replaceLabels.
--
-- However, we probably do want to do this if we are splitting proc
-- points, because L1 will be a proc-point anyway, so merging it with
-- L4 reduces the number of proc points.  Unfortunately recursive
-- let-no-escapes won't generate very good code with proc-point
-- splitting on - we should probably compile them to explicitly use
-- the native calling convention instead.
134

135 136
cmmCfgOpts :: Bool -> CmmGraph -> CmmGraph
cmmCfgOpts split g = fst (blockConcat split g)
137

138
cmmCfgOptsProc :: Bool -> CmmDecl -> CmmDecl
139
cmmCfgOptsProc split (CmmProc info lbl live g) = CmmProc info' lbl live g'
140
    where (g', env) = blockConcat split g
141 142
          info' = info{ info_tbls = new_info_tbls }
          new_info_tbls = mapFromList (map upd_info (mapToList (info_tbls info)))
Simon Marlow's avatar
Simon Marlow committed
143

144 145 146
          -- If we changed any labels, then we have to update the info tables
          -- too, except for the top-level info table because that might be
          -- referred to by other procs.
147 148
          upd_info (k,info)
             | Just k' <- mapLookup k env
149 150 151
             = (k', if k' == g_entry g'
                       then info
                       else info{ cit_lbl = infoTblLbl k' })
152 153
             | otherwise
             = (k,info)
154
cmmCfgOptsProc _ top = top
155

156

157
blockConcat :: Bool -> CmmGraph -> (CmmGraph, LabelMap BlockId)
158
blockConcat splitting_procs g@CmmGraph { g_entry = entry_id }
159
  = (replaceLabels shortcut_map $ ofBlockMap new_entry new_blocks, shortcut_map')
160
  where
161 162
     -- We might be able to shortcut the entry BlockId itself.
     -- Remember to update the shortcut_map, since we also have to
163 164
     -- update the info_tbls mapping now.
     (new_entry, shortcut_map')
165 166
       | Just entry_blk <- mapLookup entry_id new_blocks
       , Just dest      <- canShortcut entry_blk
167
       = (dest, mapInsert entry_id dest shortcut_map)
168
       | otherwise
169
       = (entry_id, shortcut_map)
170

171 172 173
     -- blocks is a list of blocks in DFS postorder, while blockmap is
     -- a map of blocks. We process each element from blocks and update
     -- blockmap accordingly
174
     blocks = postorderDfs g
175
     blockmap = foldr addBlock emptyBody blocks
176

177 178 179 180 181 182 183 184 185 186 187 188 189
     -- Accumulator contains three components:
     --  * map of blocks in a graph
     --  * map of shortcut labels. See Note [Shortcut call returns]
     --  * map containing number of predecessors for each block. We discard
     --    it after we process all blocks.
     (new_blocks, shortcut_map, _) =
           foldr maybe_concat (blockmap, mapEmpty, initialBackEdges) blocks

     -- Map of predecessors for initial graph. We increase number of
     -- predecessors for entry block by one to denote that it is
     -- target of a jump, even if no block in the current graph jumps
     -- to it.
     initialBackEdges = incPreds entry_id (predMap blocks)
190 191

     maybe_concat :: CmmBlock
192 193
                  -> (LabelMap CmmBlock, LabelMap BlockId, LabelMap Int)
                  -> (LabelMap CmmBlock, LabelMap BlockId, LabelMap Int)
194 195 196
     maybe_concat block (blocks, shortcut_map, backEdges)
        -- If:
        --   (1) current block ends with unconditional branch to b' and
197 198
        --   (2) it has exactly one predecessor (namely, current block)
        --
199 200 201 202 203
        -- Then:
        --   (1) append b' block at the end of current block
        --   (2) remove b' from the map of blocks
        --   (3) remove information about b' from predecessors map
        --
204 205
        -- Since we know that the block has only one predecessor we call
        -- mapDelete directly instead of calling decPreds.
206 207 208
        --
        -- Note that we always maintain an up-to-date list of predecessors, so
        -- we can ignore the contents of shortcut_map
209
        | CmmBranch b' <- last
210
        , hasOnePredecessor b'
211
        , Just blk' <- mapLookup b' blocks
212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228
        = let bid' = entryLabel blk'
          in ( mapDelete bid' $ mapInsert bid (splice head blk') blocks
             , shortcut_map
             , mapDelete b' backEdges )

        -- If:
        --   (1) we are splitting proc points (see Note
        --       [Shortcut call returns and proc-points]) and
        --   (2) current block is a CmmCall or CmmForeignCall with
        --       continuation b' and
        --   (3) we can shortcut that continuation to dest
        -- Then:
        --   (1) we change continuation to point to b'
        --   (2) create mapping from b' to dest
        --   (3) increase number of predecessors of dest by 1
        --   (4) decrease number of predecessors of b' by 1
        --
Gabor Greif's avatar
Gabor Greif committed
229
        -- Later we will use replaceLabels to substitute all occurrences of b'
230 231
        -- with dest.
        | splitting_procs
232
        , Just b'   <- callContinuation_maybe last
233
        , Just blk' <- mapLookup b' blocks
Simon Marlow's avatar
Simon Marlow committed
234
        , Just dest <- canShortcut blk'
235 236 237 238 239 240 241 242 243 244 245
        = ( mapInsert bid (blockJoinTail head (update_cont dest)) blocks
          , mapInsert b' dest shortcut_map
          , decPreds b' $ incPreds dest backEdges )

        -- If:
        --   (1) a block does not end with a call
        -- Then:
        --   (1) if it ends with a conditional attempt to invert the
        --       conditional
        --   (2) attempt to shortcut all destination blocks
        --   (3) if new successors of a block are different from the old ones
246 247 248 249
        --       update the of predecessors accordingly
        --
        -- A special case of this is a situation when a block ends with an
        -- unconditional jump to a block that can be shortcut.
250
        | Nothing <- callContinuation_maybe last
251 252 253 254 255 256 257 258 259
        = let oldSuccs = successors last
              newSuccs = successors swapcond_last
          in ( mapInsert bid (blockJoinTail head swapcond_last) blocks
             , shortcut_map
             , if oldSuccs == newSuccs
               then backEdges
               else foldr incPreds (foldr decPreds backEdges oldSuccs) newSuccs )

        -- Otherwise don't do anything
260
        | otherwise
261
        = ( blocks, shortcut_map, backEdges )
262
        where
Simon Marlow's avatar
Simon Marlow committed
263 264
          (head, last) = blockSplitTail block
          bid = entryLabel block
265

266 267 268 269 270 271 272 273
          -- Changes continuation of a call to a specified label
          update_cont dest =
              case last of
                CmmCall{}        -> last { cml_cont = Just dest }
                CmmForeignCall{} -> last { succ = dest }
                _                -> panic "Can't shortcut continuation."

          -- Attempts to shortcut successors of last node
274
          shortcut_last = mapSuccessors shortcut last
275 276 277
            where
              shortcut l =
                 case mapLookup l blocks of
278
                   Just b | Just dest <- canShortcut b -> dest
279 280
                   _otherwise -> l

281 282 283 284 285
          -- For a conditional, we invert the conditional if that would make it
          -- more likely that the branch-not-taken case becomes a fallthrough.
          -- This helps the native codegen a little bit, and probably has no
          -- effect on LLVM.  It's convenient to do it here, where we have the
          -- information about predecessors.
286
          swapcond_last
287 288
            | CmmCondBranch cond t f l <- shortcut_last
            , likelyFalse l
289
            , numPreds f > 1
290
            , hasOnePredecessor t
291
            , Just cond' <- maybeInvertCmmExpr cond
292
            = CmmCondBranch cond' f t (invertLikeliness l)
293 294 295 296

            | otherwise
            = shortcut_last

297 298 299 300 301 302 303
          likelyFalse (Just False) = True
          likelyFalse Nothing      = True
          likelyFalse _            = False

          invertLikeliness (Just b)     = Just (not b)
          invertLikeliness Nothing      = Nothing

304 305 306 307 308 309 310 311 312 313 314 315 316
          -- Number of predecessors for a block
          numPreds bid = mapLookup bid backEdges `orElse` 0

          hasOnePredecessor b = numPreds b == 1

-- Functions for incrementing and decrementing number of predecessors. If
-- decrementing would set the predecessor count to 0, we remove entry from the
-- map.
-- Invariant: if a block has no predecessors it should be dropped from the
-- graph because it is unreachable. maybe_concat is constructed to maintain
-- that invariant, but calling replaceLabels may introduce unreachable blocks.
-- We rely on subsequent passes in the Cmm pipeline to remove unreachable
-- blocks.
317
incPreds, decPreds :: BlockId -> LabelMap Int -> LabelMap Int
318 319 320 321 322 323 324 325 326 327 328 329
incPreds bid edges = mapInsertWith (+) bid 1 edges
decPreds bid edges = case mapLookup bid edges of
                       Just preds | preds > 1 -> mapInsert bid (preds - 1) edges
                       Just _                 -> mapDelete bid edges
                       _                      -> edges


-- Checks if a block consists only of "goto dest". If it does than we return
-- "Just dest" label. See Note [What is shortcutting]
canShortcut :: CmmBlock -> Maybe BlockId
canShortcut block
    | (_, middle, CmmBranch dest) <- blockSplit block
Peter Wortmann's avatar
Peter Wortmann committed
330
    , all dont_care $ blockToList middle
331 332 333
    = Just dest
    | otherwise
    = Nothing
Peter Wortmann's avatar
Peter Wortmann committed
334 335 336
    where dont_care CmmComment{} = True
          dont_care CmmTick{}    = True
          dont_care _other       = False
337 338 339 340 341

-- Concatenates two blocks. First one is assumed to be open on exit, the second
-- is assumed to be closed on entry (i.e. it has a label attached to it, which
-- the splice function removes by calling snd on result of blockSplitHead).
splice :: Block CmmNode C O -> CmmBlock -> CmmBlock
Peter Wortmann's avatar
Peter Wortmann committed
342 343 344 345
splice head rest = entry `blockJoinHead` code0 `blockAppend` code1
  where (CmmEntry lbl sc0, code0) = blockSplitHead head
        (CmmEntry _   sc1, code1) = blockSplitHead rest
        entry = CmmEntry lbl (combineTickScopes sc0 sc1)
346 347 348

-- If node is a call with continuation call return Just label of that
-- continuation. Otherwise return Nothing.
349 350 351 352 353
callContinuation_maybe :: CmmNode O C -> Maybe BlockId
callContinuation_maybe (CmmCall { cml_cont = Just b }) = Just b
callContinuation_maybe (CmmForeignCall { succ = b })   = Just b
callContinuation_maybe _ = Nothing

354

355
-- Map over the CmmGraph, replacing each label with its mapping in the
356 357
-- supplied LabelMap.
replaceLabels :: LabelMap BlockId -> CmmGraph -> CmmGraph
358
replaceLabels env g
Simon Marlow's avatar
Simon Marlow committed
359 360
  | mapNull env = g
  | otherwise   = replace_eid $ mapGraphNodes1 txnode g
361
   where
362
     replace_eid g = g {g_entry = lookup (g_entry g)}
363 364 365
     lookup id = mapLookup id env `orElse` id

     txnode :: CmmNode e x -> CmmNode e x
366 367 368 369 370 371 372 373 374 375
     txnode (CmmBranch bid) = CmmBranch (lookup bid)
     txnode (CmmCondBranch p t f l) =
       mkCmmCondBranch (exp p) (lookup t) (lookup f) l
     txnode (CmmSwitch e ids) =
       CmmSwitch (exp e) (mapSwitchTargets lookup ids)
     txnode (CmmCall t k rg a res r) =
       CmmCall (exp t) (liftM lookup k) rg a res r
     txnode fc@CmmForeignCall{} =
       fc{ args = map exp (args fc), succ = lookup (succ fc) }
     txnode other = mapExpDeep exp other
376 377

     exp :: CmmExpr -> CmmExpr
378
     exp (CmmLit (CmmBlock bid))                = CmmLit (CmmBlock (lookup bid))
Simon Marlow's avatar
Simon Marlow committed
379
     exp (CmmStackSlot (Young id) i) = CmmStackSlot (Young (lookup id)) i
380 381
     exp e                                      = e

382 383 384
mkCmmCondBranch :: CmmExpr -> Label -> Label -> Maybe Bool -> CmmNode O C
mkCmmCondBranch p t f l =
  if t == f then CmmBranch t else CmmCondBranch p t f l
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
385

386
-- Build a map from a block to its set of predecessors.
387
predMap :: [CmmBlock] -> LabelMap Int
388
predMap blocks = foldr add_preds mapEmpty blocks
Simon Marlow's avatar
Simon Marlow committed
389 390 391
  where
    add_preds block env = foldr add env (successors block)
      where add lbl env = mapInsertWith (+) lbl 1 env
392 393

-- Removing unreachable blocks
394
removeUnreachableBlocksProc :: CmmDecl -> CmmDecl
395
removeUnreachableBlocksProc proc@(CmmProc info lbl live g)
396
   | used_blocks `lengthLessThan` mapSize (toBlockMap g)
397
   = CmmProc info' lbl live g'
398 399 400 401 402
   | otherwise
   = proc
   where
     g'    = ofBlockList (g_entry g) used_blocks
     info' = info { info_tbls = keep_used (info_tbls info) }
403
             -- Remove any info_tbls for unreachable
404

405
     keep_used :: LabelMap CmmInfoTable -> LabelMap CmmInfoTable
406
     keep_used bs = mapFoldWithKey keep mapEmpty bs
407

408
     keep :: Label -> CmmInfoTable -> LabelMap CmmInfoTable -> LabelMap CmmInfoTable
409 410 411 412 413 414 415 416
     keep l i env | l `setMember` used_lbls = mapInsert l i env
                  | otherwise               = env

     used_blocks :: [CmmBlock]
     used_blocks = postorderDfs g

     used_lbls :: LabelSet
     used_lbls = foldr (setInsert . entryLabel) setEmpty used_blocks