[Git][ghc/ghc][wip/andreask/refactor_cmm_weights] Refactor linear reg alloc to remember past assignments.
Andreas Klebinger
gitlab at gitlab.haskell.org
Mon May 11 12:50:46 UTC 2020
Andreas Klebinger pushed to branch wip/andreask/refactor_cmm_weights at Glasgow Haskell Compiler / GHC
Commits:
04831c6d by Andreas Klebinger at 2020-05-11T14:48:30+02:00
Refactor linear reg alloc to remember past assignments.
When assigning registers we now first try registers we
assigned to in the past, instead of picking the "first"
one.
This is in extremely helpful when dealing with loops for
which variables are dead for part of the loop.
This is important for patterns like this:
foo = arg1
loop:
use(foo)
...
foo = getVal()
goto loop;
There we:
* assign foo to the register of arg1.
* use foo, it's dead after this use as it's overwritten after.
* do other things.
* look for a register to put foo in.
If we pick an arbitrary one it might differ from the register the
start of the loop expect's foo to be in.
To fix this we simply look for past register assignments for
the given variable. If we find one and the register is free we
use that register.
This reduces the need for fixup blocks which match the register
assignment between blocks. In the example above between the end
and the head of the loop.
This patch also moves branch weight estimation ahead of register
allocation and adds a flag to control it (cmm-static-pred).
* It means the linear allocator is more likely to assign the hotter
code paths first.
* If it assign these first we are:
+ Less likely to spill on the hot path.
+ Less likely to introduce fixup blocks on the hot path.
These two measure combined are surprisingly effective. Based on nofib
we get in the mean:
* -0.9% instructions executed
* -0.1% reads/writes
* -0.2% code size.
* -0.1% compiler allocations.
* -0.9% compile time.
* -0.8% runtime.
Most of the benefits are simply a result of removing redundant moves
and spills.
Reduced compiler allocations likely are the result of less code being
generated. (The added lookup is mostly non-allocating).
- - - - -
15 changed files:
- compiler/GHC/CmmToAsm.hs
- compiler/GHC/CmmToAsm/BlockLayout.hs
- compiler/GHC/CmmToAsm/CFG.hs
- compiler/GHC/CmmToAsm/Instr.hs
- compiler/GHC/CmmToAsm/Reg/Linear.hs
- compiler/GHC/CmmToAsm/Reg/Linear/Base.hs
- compiler/GHC/CmmToAsm/Reg/Linear/PPC.hs
- compiler/GHC/CmmToAsm/Reg/Linear/SPARC.hs
- compiler/GHC/CmmToAsm/Reg/Linear/State.hs
- compiler/GHC/CmmToAsm/Reg/Linear/X86.hs
- compiler/GHC/CmmToAsm/Reg/Linear/X86_64.hs
- compiler/GHC/Driver/Flags.hs
- compiler/GHC/Driver/Session.hs
- compiler/utils/Outputable.hs
- docs/users_guide/using-optimisation.rst
Changes:
=====================================
compiler/GHC/CmmToAsm.hs
=====================================
@@ -729,7 +729,7 @@ cmmNativeGen dflags this_mod modLoc ncgImpl us fileIds dbgMap cmm count
let optimizedCFG :: Maybe CFG
optimizedCFG =
- optimizeCFG (cfgWeightInfo dflags) cmm <$!> postShortCFG
+ optimizeCFG (gopt Opt_CmmStaticPred dflags) (cfgWeightInfo dflags) cmm <$!> postShortCFG
maybeDumpCfg dflags optimizedCFG "CFG Weights - Final" proc_name
=====================================
compiler/GHC/CmmToAsm/BlockLayout.hs
=====================================
@@ -636,22 +636,8 @@ sequenceChain :: forall a i. (Instruction i, Outputable i)
-> [GenBasicBlock i] -- ^ Blocks placed in sequence.
sequenceChain _info _weights [] = []
sequenceChain _info _weights [x] = [x]
-sequenceChain info weights' blocks@((BasicBlock entry _):_) =
- let weights :: CFG
- weights = --pprTrace "cfg'" (pprEdgeWeights cfg')
- cfg'
- where
- (_, globalEdgeWeights) = {-# SCC mkGlobalWeights #-} mkGlobalWeights entry weights'
- cfg' = {-# SCC rewriteEdges #-}
- mapFoldlWithKey
- (\cfg from m ->
- mapFoldlWithKey
- (\cfg to w -> setEdgeWeight cfg (EdgeWeight w) from to )
- cfg m )
- weights'
- globalEdgeWeights
-
- directEdges :: [CfgEdge]
+sequenceChain info weights blocks@((BasicBlock entry _):_) =
+ let directEdges :: [CfgEdge]
directEdges = sortBy (flip compare) $ catMaybes . map relevantWeight $ (infoEdgeList weights)
where
relevantWeight :: CfgEdge -> Maybe CfgEdge
=====================================
compiler/GHC/CmmToAsm/CFG.hs
=====================================
@@ -670,11 +670,21 @@ findBackEdges root cfg =
typedEdges =
classifyEdges root getSuccs edges :: [((BlockId,BlockId),EdgeType)]
-
-optimizeCFG :: D.CfgWeights -> RawCmmDecl -> CFG -> CFG
-optimizeCFG _ (CmmData {}) cfg = cfg
-optimizeCFG weights (CmmProc info _lab _live graph) cfg =
- {-# SCC optimizeCFG #-}
+optimizeCFG :: Bool -> D.CfgWeights -> RawCmmDecl -> CFG -> CFG
+optimizeCFG _ _ (CmmData {}) cfg = cfg
+optimizeCFG doStaticPred weights proc@(CmmProc _info _lab _live graph) cfg =
+ (if doStaticPred then staticPredCfg (g_entry graph) else id) $
+ optHsPatterns weights proc $ cfg
+
+-- | Modify branch weights based on educated guess on
+-- patterns GHC tends to produce and how they affect
+-- performance.
+--
+-- Most importantly we penalize jumps across info tables.
+optHsPatterns :: D.CfgWeights -> RawCmmDecl -> CFG -> CFG
+optHsPatterns _ (CmmData {}) cfg = cfg
+optHsPatterns weights (CmmProc info _lab _live graph) cfg =
+ {-# SCC optHsPatterns #-}
-- pprTrace "Initial:" (pprEdgeWeights cfg) $
-- pprTrace "Initial:" (ppr $ mkGlobalWeights (g_entry graph) cfg) $
@@ -749,6 +759,21 @@ optimizeCFG weights (CmmProc info _lab _live graph) cfg =
| CmmSource { trans_cmmNode = CmmCondBranch {} } <- source = True
| otherwise = False
+-- |
+staticPredCfg :: BlockId -> CFG -> CFG
+staticPredCfg entry cfg = cfg'
+ where
+ (_, globalEdgeWeights) = {-# SCC mkGlobalWeights #-}
+ mkGlobalWeights entry cfg
+ cfg' = {-# SCC rewriteEdges #-}
+ mapFoldlWithKey
+ (\cfg from m ->
+ mapFoldlWithKey
+ (\cfg to w -> setEdgeWeight cfg (EdgeWeight w) from to )
+ cfg m )
+ cfg
+ globalEdgeWeights
+
-- | Determine loop membership of blocks based on SCC analysis
-- This is faster but only gives yes/no answers.
loopMembers :: HasDebugCallStack => CFG -> LabelMap Bool
@@ -922,6 +947,10 @@ revPostorderFrom cfg root =
-- reverse post order. Which is required for diamond control flow to work probably.
--
-- We also apply a few prediction heuristics (based on the same paper)
+--
+-- The returned result represents frequences.
+-- For blocks it's the expected number of executions and
+-- for edges is the number of traversals.
{-# NOINLINE mkGlobalWeights #-}
{-# SCC mkGlobalWeights #-}
=====================================
compiler/GHC/CmmToAsm/Instr.hs
=====================================
@@ -37,7 +37,10 @@ import GHC.CmmToAsm.Config
-- (for allocation purposes, anyway).
--
data RegUsage
- = RU [Reg] [Reg]
+ = RU {
+ reads :: [Reg],
+ writes :: [Reg]
+ }
-- | No regs read or written to.
noUsage :: RegUsage
=====================================
compiler/GHC/CmmToAsm/Reg/Linear.hs
=====================================
@@ -1,4 +1,5 @@
{-# LANGUAGE BangPatterns, CPP, ScopedTypeVariables #-}
+{-# LANGUAGE ConstraintKinds #-}
{-# OPTIONS_GHC -Wno-incomplete-uni-patterns #-}
@@ -137,6 +138,7 @@ import GHC.Platform
import Data.Maybe
import Data.List
import Control.Monad
+import Control.Applicative
-- -----------------------------------------------------------------------------
-- Top level of the register allocator
@@ -229,8 +231,12 @@ linearRegAlloc config entry_ids block_live sccs
go f = linearRegAlloc' config f entry_ids block_live sccs
platform = ncgPlatform config
+-- |
+type OutputableRegConstraint freeRegs instr =
+ (FR freeRegs, Outputable freeRegs, Outputable instr, Instruction instr)
+
linearRegAlloc'
- :: (FR freeRegs, Outputable instr, Instruction instr)
+ :: OutputableRegConstraint freeRegs instr
=> NCGConfig
-> freeRegs
-> [BlockId] -- ^ entry points
@@ -246,7 +252,7 @@ linearRegAlloc' config initFreeRegs entry_ids block_live sccs
return (blocks, stats, getStackUse stack)
-linearRA_SCCs :: (FR freeRegs, Instruction instr, Outputable instr)
+linearRA_SCCs :: OutputableRegConstraint freeRegs instr
=> [BlockId]
-> BlockMap RegSet
-> [NatBasicBlock instr]
@@ -281,7 +287,7 @@ linearRA_SCCs entry_ids block_live blocksAcc (CyclicSCC blocks : sccs)
more sanity checking to guard against this eventuality.
-}
-process :: (FR freeRegs, Instruction instr, Outputable instr)
+process :: OutputableRegConstraint freeRegs instr
=> [BlockId]
-> BlockMap RegSet
-> [GenBasicBlock (LiveInstr instr)]
@@ -325,15 +331,21 @@ process entry_ids block_live (b@(BasicBlock id _) : blocks)
-- | Do register allocation on this basic block
--
processBlock
- :: (FR freeRegs, Outputable instr, Instruction instr)
+ :: OutputableRegConstraint freeRegs instr
=> BlockMap RegSet -- ^ live regs on entry to each basic block
-> LiveBasicBlock instr -- ^ block to do register allocation on
-> RegM freeRegs [NatBasicBlock instr] -- ^ block with registers allocated
processBlock block_live (BasicBlock id instrs)
- = do initBlock id block_live
+ = do -- pprTraceM "processBlock" $ text "" $$ ppr (BasicBlock id instrs)
+ initBlock id block_live
+
+ -- assig <- getBlockAssigR
+ -- pprTraceM "assignment" $ ppr assig
+
(instrs', fixups)
<- linearRA block_live [] [] id instrs
+ -- pprTraceM "blockResult" $ ppr (instrs', fixups)
return $ BasicBlock id instrs' : fixups
@@ -369,7 +381,7 @@ initBlock id block_live
-- | Do allocation for a sequence of instructions.
linearRA
- :: (FR freeRegs, Outputable instr, Instruction instr)
+ :: OutputableRegConstraint freeRegs instr
=> BlockMap RegSet -- ^ map of what vregs are live on entry to each block.
-> [instr] -- ^ accumulator for instructions already processed.
-> [NatBasicBlock instr] -- ^ accumulator for blocks of fixup code.
@@ -396,7 +408,7 @@ linearRA block_live accInstr accFixups id (instr:instrs)
-- | Do allocation for a single instruction.
raInsn
- :: (FR freeRegs, Outputable instr, Instruction instr)
+ :: OutputableRegConstraint freeRegs instr
=> BlockMap RegSet -- ^ map of what vregs are love on entry to each block.
-> [instr] -- ^ accumulator for instructions already processed.
-> BlockId -- ^ the id of the current block, for debugging
@@ -476,7 +488,7 @@ isInReg src assig | Just (InReg _) <- lookupUFM assig src = True
| otherwise = False
-genRaInsn :: (FR freeRegs, Instruction instr, Outputable instr)
+genRaInsn :: OutputableRegConstraint freeRegs instr
=> BlockMap RegSet
-> [instr]
-> BlockId
@@ -486,6 +498,7 @@ genRaInsn :: (FR freeRegs, Instruction instr, Outputable instr)
-> RegM freeRegs ([instr], [NatBasicBlock instr])
genRaInsn block_live new_instrs block_id instr r_dying w_dying = do
+-- pprTraceM "genRaInsn" $ ppr (block_id, instr)
platform <- getPlatform
case regUsageOfInstr platform instr of { RU read written ->
do
@@ -525,10 +538,12 @@ genRaInsn block_live new_instrs block_id instr r_dying w_dying = do
(fixup_blocks, adjusted_instr)
<- joinToTargets block_live block_id instr
+-- when (not $ null fixup_blocks) $ pprTraceM "genRA:FixBlocks" $ ppr fixup_blocks
+
-- Debugging - show places where the reg alloc inserted
-- assignment fixup blocks.
- -- when (not $ null fixup_blocks) $
- -- pprTrace "fixup_blocks" (ppr fixup_blocks) (return ())
+-- when (not $ null fixup_blocks) $
+-- pprTrace "fixup_blocks" (ppr fixup_blocks) (return ())
-- (e) Delete all register assignments for temps which are read
-- (only) and die here. Update the free register list.
@@ -737,7 +752,7 @@ data SpillLoc = ReadMem StackSlot -- reading from register only in memory
-- the list of free registers and free stack slots.
allocateRegsAndSpill
- :: (FR freeRegs, Outputable instr, Instruction instr)
+ :: forall freeRegs instr. (FR freeRegs, Outputable instr, Instruction instr)
=> Bool -- True <=> reading (load up spilled regs)
-> [VirtualReg] -- don't push these out
-> [instr] -- spill insns
@@ -749,7 +764,8 @@ allocateRegsAndSpill _ _ spills alloc []
= return (spills, reverse alloc)
allocateRegsAndSpill reading keep spills alloc (r:rs)
- = do assig <- getAssigR
+ = do assig <- getAssigR :: RegM freeRegs (RegMap Loc)
+ -- pprTraceM "allocateRegsAndSpill:assig" (ppr (r:rs) $$ ppr assig)
let doSpill = allocRegsAndSpill_spill reading keep spills alloc r rs assig
case lookupUFM assig r of
-- case (1a): already in a register
@@ -779,6 +795,22 @@ allocateRegsAndSpill reading keep spills alloc (r:rs)
| otherwise -> doSpill WriteNew
+-- | Given a virtual reg find a preferred real register.
+-- Note: I tried returning a list of registers, but that
+-- turned out to barely matter but added a few tenths of
+-- a percent to compile time.
+findPrefRealReg :: forall freeRegs u. Uniquable u
+ => u -> RegM freeRegs (Maybe RealReg)
+findPrefRealReg vreg = do
+ bassig <- getBlockAssigR :: RegM freeRegs (BlockMap (freeRegs,RegMap Loc))
+ return $ foldr (findVirtRegAssig) Nothing bassig
+ where
+ findVirtRegAssig :: (freeRegs,RegMap Loc) -> Maybe RealReg -> Maybe RealReg
+ findVirtRegAssig assig z =
+ z <|> case lookupUFM (snd assig) vreg of
+ Just (InReg real_reg) -> Just real_reg
+ Just (InBoth real_reg _) -> Just real_reg
+ _ -> z
-- reading is redundant with reason, but we keep it around because it's
-- convenient and it maintains the recursive structure of the allocator. -- EZY
@@ -795,18 +827,26 @@ allocRegsAndSpill_spill :: (FR freeRegs, Instruction instr, Outputable instr)
allocRegsAndSpill_spill reading keep spills alloc r rs assig spill_loc
= do platform <- getPlatform
freeRegs <- getFreeRegsR
- let freeRegs_thisClass = frGetFreeRegs platform (classOfVirtualReg r) freeRegs
+ let freeRegs_thisClass = frGetFreeRegs platform (classOfVirtualReg r) freeRegs :: [RealReg]
- case freeRegs_thisClass of
+ -- Can we put the variable into a register it already was?
+ pref_reg <- findPrefRealReg r
+ case freeRegs_thisClass of
-- case (2): we have a free register
- (my_reg : _) ->
- do spills' <- loadTemp r spill_loc my_reg spills
+ (first_free : _) ->
+ do let final_reg
+ | Just reg <- pref_reg
+ , reg `elem` freeRegs_thisClass
+ = reg
+ | otherwise
+ = first_free
+ spills' <- loadTemp r spill_loc final_reg spills
- setAssigR (addToUFM assig r $! newLocation spill_loc my_reg)
- setFreeRegsR $ frAllocateReg platform my_reg freeRegs
+ setAssigR (addToUFM assig r $! newLocation spill_loc final_reg)
+ setFreeRegsR $ frAllocateReg platform final_reg freeRegs
- allocateRegsAndSpill reading keep spills' (my_reg : alloc) rs
+ allocateRegsAndSpill reading keep spills' (final_reg : alloc) rs
-- case (3): we need to push something out to free up a register
@@ -814,7 +854,8 @@ allocRegsAndSpill_spill reading keep spills alloc r rs assig spill_loc
do let inRegOrBoth (InReg _) = True
inRegOrBoth (InBoth _ _) = True
inRegOrBoth _ = False
- let candidates' =
+ let candidates' :: UniqFM Loc
+ candidates' =
flip delListFromUFM keep $
filterUFM inRegOrBoth $
assig
=====================================
compiler/GHC/CmmToAsm/Reg/Linear/Base.hs
=====================================
@@ -30,6 +30,7 @@ import GHC.Types.Unique.FM
import GHC.Types.Unique.Supply
import GHC.Cmm.BlockId
+data ReadingOrWriting = Reading | Writing deriving (Eq,Ord)
-- | Used to store the register assignment on entry to a basic block.
-- We use this to handle join points, where multiple branch instructions
@@ -138,6 +139,8 @@ data RA_State freeRegs
, ra_config :: !NCGConfig
-- | (from,fixup,to) : We inserted fixup code between from and to
- , ra_fixups :: [(BlockId,BlockId,BlockId)] }
+ , ra_fixups :: [(BlockId,BlockId,BlockId)]
+
+ }
=====================================
compiler/GHC/CmmToAsm/Reg/Linear/PPC.hs
=====================================
@@ -1,3 +1,5 @@
+{-# LANGUAGE GeneralizedNewtypeDeriving #-}
+
-- | Free regs map for PowerPC
module GHC.CmmToAsm.Reg.Linear.PPC where
@@ -27,6 +29,9 @@ import Data.Bits
data FreeRegs = FreeRegs !Word32 !Word32
deriving( Show ) -- The Show is used in an ASSERT
+instance Outputable FreeRegs where
+ ppr = text . show
+
noFreeRegs :: FreeRegs
noFreeRegs = FreeRegs 0 0
=====================================
compiler/GHC/CmmToAsm/Reg/Linear/SPARC.hs
=====================================
@@ -1,4 +1,5 @@
{-# LANGUAGE CPP #-}
+{-# LANGUAGE GeneralizedNewtypeDeriving #-}
-- | Free regs map for SPARC
module GHC.CmmToAsm.Reg.Linear.SPARC where
@@ -38,6 +39,9 @@ data FreeRegs
instance Show FreeRegs where
show = showFreeRegs
+instance Outputable FreeRegs where
+ ppr = text . showFreeRegs
+
-- | A reg map where no regs are free to be allocated.
noFreeRegs :: FreeRegs
noFreeRegs = FreeRegs 0 0 0
=====================================
compiler/GHC/CmmToAsm/Reg/Linear/State.hs
=====================================
@@ -1,4 +1,5 @@
{-# LANGUAGE CPP, PatternSynonyms, DeriveFunctor #-}
+{-# LANGUAGE ScopedTypeVariables #-}
#if !defined(GHC_LOADED_INTO_GHCI)
{-# LANGUAGE UnboxedTuples #-}
=====================================
compiler/GHC/CmmToAsm/Reg/Linear/X86.hs
=====================================
@@ -1,3 +1,4 @@
+{-# LANGUAGE GeneralizedNewtypeDeriving #-}
-- | Free regs map for i386
module GHC.CmmToAsm.Reg.Linear.X86 where
@@ -9,12 +10,13 @@ import GHC.Platform.Reg.Class
import GHC.Platform.Reg
import Panic
import GHC.Platform
+import Outputable
import Data.Word
import Data.Bits
newtype FreeRegs = FreeRegs Word32
- deriving Show
+ deriving (Show,Outputable)
noFreeRegs :: FreeRegs
noFreeRegs = FreeRegs 0
=====================================
compiler/GHC/CmmToAsm/Reg/Linear/X86_64.hs
=====================================
@@ -1,3 +1,4 @@
+{-# LANGUAGE GeneralizedNewtypeDeriving #-}
-- | Free regs map for x86_64
module GHC.CmmToAsm.Reg.Linear.X86_64 where
@@ -9,12 +10,13 @@ import GHC.Platform.Reg.Class
import GHC.Platform.Reg
import Panic
import GHC.Platform
+import Outputable
import Data.Word
import Data.Bits
newtype FreeRegs = FreeRegs Word64
- deriving Show
+ deriving (Show,Outputable)
noFreeRegs :: FreeRegs
noFreeRegs = FreeRegs 0
=====================================
compiler/GHC/Driver/Flags.hs
=====================================
@@ -181,6 +181,7 @@ data GeneralFlag
| Opt_LlvmFillUndefWithGarbage -- Testing for undef bugs (hidden flag)
| Opt_IrrefutableTuples
| Opt_CmmSink
+ | Opt_CmmStaticPred
| Opt_CmmElimCommonBlocks
| Opt_AsmShortcutting
| Opt_OmitYields
=====================================
compiler/GHC/Driver/Session.hs
=====================================
@@ -3523,6 +3523,7 @@ fFlagsDeps = [
flagSpec "case-folding" Opt_CaseFolding,
flagSpec "cmm-elim-common-blocks" Opt_CmmElimCommonBlocks,
flagSpec "cmm-sink" Opt_CmmSink,
+ flagSpec "cmm-static-pred" Opt_CmmStaticPred,
flagSpec "cse" Opt_CSE,
flagSpec "stg-cse" Opt_StgCSE,
flagSpec "stg-lift-lams" Opt_StgLiftLams,
@@ -4074,6 +4075,7 @@ optLevelFlags -- see Note [Documenting optimisation flags]
, ([1,2], Opt_CmmElimCommonBlocks)
, ([2], Opt_AsmShortcutting)
, ([1,2], Opt_CmmSink)
+ , ([1,2], Opt_CmmStaticPred)
, ([1,2], Opt_CSE)
, ([1,2], Opt_StgCSE)
, ([2], Opt_StgLiftLams)
=====================================
compiler/utils/Outputable.hs
=====================================
@@ -847,6 +847,9 @@ instance Outputable Word16 where
instance Outputable Word32 where
ppr n = integer $ fromIntegral n
+instance Outputable Word64 where
+ ppr n = integer $ fromIntegral n
+
instance Outputable Word where
ppr n = integer $ fromIntegral n
=====================================
docs/users_guide/using-optimisation.rst
=====================================
@@ -214,6 +214,19 @@ by saying ``-fno-wombat``.
to their usage sites. It also inlines simple expressions like
literals or registers.
+.. ghc-flag:: -fcmm-static-pred
+ :shortdesc: Enable static control flow prediction :ghc-flag:`-O`.
+ :type: dynamic
+ :reverse: -fno-cmm-static-pred
+ :category:
+
+ :default: off but enabled with :ghc-flag:`-O`.
+
+ This enables static control flow prediction on the final Cmm
+ code. If enabled GHC will apply certain heuristics to identify
+ loops and hot code paths. This information is then used by the
+ register allocation and code layout passes.
+
.. ghc-flag:: -fasm-shortcutting
:shortdesc: Enable shortcutting on assembly. Implied by :ghc-flag:`-O2`.
:type: dynamic
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/04831c6d1ace8eb951b1045eb9a526ed51bc4fad
--
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/04831c6d1ace8eb951b1045eb9a526ed51bc4fad
You're receiving this email because of your account on gitlab.haskell.org.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-commits/attachments/20200511/f4839a73/attachment-0001.html>
More information about the ghc-commits
mailing list