diff options
author | Jan Stolarek <jan.stolarek@p.lodz.pl> | 2013-10-18 17:43:55 +0200 |
---|---|---|
committer | Jan Stolarek <jan.stolarek@p.lodz.pl> | 2013-10-18 17:54:18 +0200 |
commit | 057bef6ef759bc0819d4ced291da92aa3feb445a (patch) | |
tree | ef27852e5c0fb0fb16ce03f9d3136043929fad49 /compiler/cmm/CmmContFlowOpt.hs | |
parent | b0545d81d2c838629498e84e144010870eaf2dfc (diff) | |
download | haskell-057bef6ef759bc0819d4ced291da92aa3feb445a.tar.gz |
Improve control flow optimisation algorithm
Fixes #8456. Previous version of control flow optimisations
did not update the list of block predecessors, leading to
unnecessary duplication of blocks in some cases. See Trac
and comments in the code for more details.
Diffstat (limited to 'compiler/cmm/CmmContFlowOpt.hs')
-rw-r--r-- | compiler/cmm/CmmContFlowOpt.hs | 435 |
1 files changed, 271 insertions, 164 deletions
diff --git a/compiler/cmm/CmmContFlowOpt.hs b/compiler/cmm/CmmContFlowOpt.hs index f39ab3c62b..7a008e2c0b 100644 --- a/compiler/cmm/CmmContFlowOpt.hs +++ b/compiler/cmm/CmmContFlowOpt.hs @@ -14,15 +14,127 @@ import BlockId import Cmm import CmmUtils import Maybes +import Panic import Control.Monad import Prelude hiding (succ, unzip, zip) ------------------------------------------------------------------------------ + +-- Note [What is shortcutting] +-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~ +-- +-- Consider this Cmm code: +-- +-- L1: ... +-- goto L2; +-- L2: goto L3; +-- L3: ... -- --- Control-flow optimisations +-- 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: -- ------------------------------------------------------------------------------ +-- 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] +-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +-- +-- This optimisation does four things: +-- +-- - 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 an unconditional branch, we may be able +-- to shortcut the destination block. +-- +-- - 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 block finishing in conditional branch we try to invert the +-- condition and shortcut destination of alternatives. +-- +-- - For any block that is not a call we try to shortcut the +-- destination(s). +-- +-- 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] +-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +-- +-- We are going to maintain the "current" graph (BlockEnv CmmBlock) as +-- 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. cmmCfgOpts :: Bool -> CmmGraph -> CmmGraph cmmCfgOpts split g = fst (blockConcat split g) @@ -43,65 +155,15 @@ cmmCfgOptsProc split (CmmProc info lbl live g) = CmmProc info' lbl live g' else info{ cit_lbl = infoTblLbl k' }) | otherwise = (k,info) - cmmCfgOptsProc _ top = top ------------------------------------------------------------------------------ --- --- Block concatenation --- ------------------------------------------------------------------------------ - --- This optimisation does three things: --- --- - If a block finishes with an unconditional branch, then we may --- be able to duplicate the block it points to and remove the --- branch. We do this if either --- a) the destination block is small (e.g. just another branch), or --- b) this is the only jump to this particular destination block. --- --- - If a block finishes in a call whose continuation block is a --- goto, then we can shortcut the destination, making the --- destination of the goto into the continuation. E.g. --- call g returns to L ==> call g returns to M --- L: goto M M: ...blah... --- M: ...blah... --- (but see Note [shortcut call returns]) --- --- - Remove any unreachable blocks from the graph. This is a side --- effect of starting with a postorder DFS traversal of the graph - --- Both 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. - --- We need to walk over the blocks from the end back to the --- beginning. We are going to maintain the "current" graph --- (BlockEnv CmmBlock) as 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). So we keep track of --- which labels we have renamed and apply the mapping at the end --- with replaceLabels. - blockConcat :: Bool -> CmmGraph -> (CmmGraph, BlockEnv BlockId) blockConcat splitting_procs g@CmmGraph { g_entry = entry_id } = (replaceLabels shortcut_map $ ofBlockMap new_entry new_blocks, shortcut_map') where - -- we might be able to shortcut the entry BlockId itself. - -- remember to update the shortcut_map', since we also have to + -- We might be able to shortcut the entry BlockId itself. + -- Remember to update the shortcut_map, since we also have to -- update the info_tbls mapping now. (new_entry, shortcut_map') | Just entry_blk <- mapLookup entry_id new_blocks @@ -110,146 +172,196 @@ blockConcat splitting_procs g@CmmGraph { g_entry = entry_id } | otherwise = (entry_id, shortcut_map) + -- 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 blocks = postorderDfs g blockmap = foldr addBlock emptyBody blocks - -- the initial blockmap is constructed from the postorderDfs result, - -- so that we automatically throw away unreachable blocks. - (new_blocks, shortcut_map) = - foldr maybe_concat (blockmap, mapEmpty) blocks + -- 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) maybe_concat :: CmmBlock - -> (BlockEnv CmmBlock, BlockEnv BlockId) - -> (BlockEnv CmmBlock, BlockEnv BlockId) - maybe_concat block (blocks, shortcut_map) + -> (BlockEnv CmmBlock, BlockEnv BlockId, BlockEnv Int) + -> (BlockEnv CmmBlock, BlockEnv BlockId, BlockEnv Int) + maybe_concat block (blocks, shortcut_map, backEdges) + -- If: + -- (1) current block ends with unconditional branch to b' and + -- (2) it has exactly one predecessor (namely, current block) and + -- (3) we have not mapped any other label to b' + -- (see Note [Shortcut call returns]). + -- 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 + -- + -- This guard must be first so that we always eliminate blocks that have + -- only one predecessor. If we had a target block that is both + -- shorcutable and has only one predecessor and attempted to shortcut it + -- first we would make that block unreachable but would not remove it + -- from the graph. | CmmBranch b' <- last , Just blk' <- mapLookup b' blocks - , shouldConcatWith b' blk' - = (mapInsert bid (splice head blk') blocks, shortcut_map) - - -- calls: if we can shortcut the continuation label, then - -- we must *also* remember to substitute for the label in the - -- code, because we will push it somewhere. - | splitting_procs -- Note [shortcut call returns] + , hasOnePredecessor b' + , hasNotBeenMappedTo b' shortcut_map + = let bid' = entryLabel blk' + in ( mapDelete bid' $ mapInsert bid (splice head blk') blocks + , shortcut_map + , mapDelete b' backEdges ) + + -- If: + -- (1) current block ends with unconditional branch to b' and + -- (2) we can shortcut block b' + -- Then: + -- (1) concatenate b' at the end of current block, effectively + -- changing target of uncondtional jump from b' to dest + -- (2) increase number of predecessors of dest by 1 + -- (3) decrease number of predecessors of b' by 1 + | CmmBranch b' <- last + , Just blk' <- mapLookup b' blocks + , Just dest <- canShortcut blk' + = ( mapInsert bid (splice head blk') blocks, shortcut_map, + decPreds b' $ incPreds dest 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 + -- + -- Later we will use replaceLabels to substitute all occurences of b' + -- with dest. + | splitting_procs , Just b' <- callContinuation_maybe last , Just blk' <- mapLookup b' blocks , Just dest <- canShortcut blk' - = (blocks, mapInsert b' dest shortcut_map) - -- replaceLabels will substitute dest for b' everywhere, later - - -- non-calls: see if we can shortcut any of the successors, - -- and check whether we should invert the conditional + = ( 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 + -- we update the of predecessors accordingly | Nothing <- callContinuation_maybe last - = ( mapInsert bid (blockJoinTail head swapcond_last) blocks - , shortcut_map ) - + = 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 | otherwise - = (blocks, shortcut_map) + = ( blocks, shortcut_map, backEdges ) where (head, last) = blockSplitTail block bid = entryLabel block + -- 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 shortcut_last = mapSuccessors shortcut last where shortcut l = case mapLookup l blocks of - Just b | Just dest <- canShortcut b -> dest + Just b | Just dest <- canShortcut b -> dest _otherwise -> l - -- 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. - -- + -- 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. swapcond_last | CmmCondBranch cond t f <- shortcut_last , numPreds f > 1 - , numPreds t == 1 + , hasOnePredecessor t , Just cond' <- maybeInvertCmmExpr cond = CmmCondBranch cond' f t | otherwise = shortcut_last - - shouldConcatWith b block - | okToDuplicate block = True -- short enough to duplicate - | numPreds b == 1 = True -- only one predecessor: go for it - | otherwise = False - - numPreds bid = mapLookup bid backEdges `orElse` 0 - - canShortcut :: CmmBlock -> Maybe BlockId - canShortcut block - | (_, middle, CmmBranch dest) <- blockSplit block - , isEmptyBlock middle - = Just dest - | otherwise - = Nothing - - backEdges :: BlockEnv Int -- number of predecessors for each block - backEdges = mapInsertWith (+) entry_id 1 $ -- add 1 for the entry id - predMap blocks - - splice :: Block CmmNode C O -> CmmBlock -> CmmBlock - splice head rest = head `blockAppend` snd (blockSplitHead rest) - - + -- Number of predecessors for a block + numPreds bid = mapLookup bid backEdges `orElse` 0 + + hasOnePredecessor b = numPreds b == 1 + + hasNotBeenMappedTo :: BlockId -> BlockEnv BlockId -> Bool + hasNotBeenMappedTo b successor_map = mapMember b successor_map + +-- 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. +incPreds, decPreds :: BlockId -> BlockEnv Int -> BlockEnv Int +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 + , isEmptyBlock middle + = Just dest + | otherwise + = Nothing + + +-- 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 +splice head rest = head `blockAppend` snd (blockSplitHead rest) + + +-- If node is a call with continuation call return Just label of that +-- continuation. Otherwise return Nothing. 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 -okToDuplicate :: CmmBlock -> Bool -okToDuplicate block - = case blockSplit block of - (_, m, CmmBranch _) -> isEmptyBlock m - -- cheap and cheerful; we might expand this in the future to - -- e.g. spot blocks that represent a single instruction or two. - -- Be careful: a CmmCall can be more than one instruction, it - -- has a CmmExpr inside it. - _otherwise -> False - - -{- Note [shortcut call returns] - -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. --} - ------------------------------------------------------------------------- + -- Map over the CmmGraph, replacing each label with its mapping in the -- supplied BlockEnv. - replaceLabels :: BlockEnv BlockId -> CmmGraph -> CmmGraph replaceLabels env g | mapNull env = g @@ -275,19 +387,14 @@ replaceLabels env g mkCmmCondBranch :: CmmExpr -> Label -> Label -> CmmNode O C mkCmmCondBranch p t f = if t == f then CmmBranch t else CmmCondBranch p t f ----------------------------------------------------------------- --- Build a map from a block to its set of predecessors. Very useful. - +-- Build a map from a block to its set of predecessors. predMap :: [CmmBlock] -> BlockEnv Int -predMap blocks = foldr add_preds mapEmpty blocks -- find the back edges +predMap blocks = foldr add_preds mapEmpty blocks where add_preds block env = foldr add env (successors block) where add lbl env = mapInsertWith (+) lbl 1 env ------------------------------------------------------------------------------ --- -- Removing unreachable blocks - removeUnreachableBlocksProc :: CmmDecl -> CmmDecl removeUnreachableBlocksProc (CmmProc info lbl live g) = CmmProc info lbl live (removeUnreachableBlocks g) |