#include "HsVersions.h"
#include "nativeGen/NCG.h"
-import MachInstrs
-import MachRegs
-import MachCodeGen
-import PprMach
-import RegAllocInfo
+
+#if alpha_TARGET_ARCH
+import Alpha.CodeGen
+import Alpha.Regs
+import Alpha.RegInfo
+import Alpha.Instr
+
+#elif i386_TARGET_ARCH || x86_64_TARGET_ARCH
+import X86.CodeGen
+import X86.Regs
+import X86.Instr
+import X86.Ppr
+
+#elif sparc_TARGET_ARCH
+import SPARC.CodeGen
+import SPARC.CodeGen.Expand
+import SPARC.Regs
+import SPARC.Instr
+import SPARC.Ppr
+import SPARC.ShortcutJump
+
+#elif powerpc_TARGET_ARCH
+import PPC.CodeGen
+import PPC.Cond
+import PPC.Regs
+import PPC.RegInfo
+import PPC.Instr
+import PPC.Ppr
+
+#else
+#error "AsmCodeGen: unknown architecture"
+
+#endif
+
+import RegAlloc.Liveness
+import qualified RegAlloc.Linear.Main as Linear
+
+import qualified GraphColor as Color
+import qualified RegAlloc.Graph.Main as Color
+import qualified RegAlloc.Graph.Stats as Color
+import qualified RegAlloc.Graph.TrivColorable as Color
+
+import TargetReg
+import Platform
+import Instruction
+import PIC
+import Reg
import NCGMonad
-import PositionIndependentCode
-import RegLiveness
-import RegCoalesce
-import qualified RegAllocLinear as Linear
-import qualified RegAllocColor as Color
-import qualified RegAllocStats as Color
-import qualified GraphColor as Color
+import BlockId
+import CgUtils ( fixStgRegisters )
import Cmm
import CmmOpt ( cmmMiniInline, cmmMachOpFold )
-import PprCmm ( pprStmt, pprCmms, pprCmm )
-import MachOp
+import PprCmm
import CLabel
-import State
import UniqFM
import Unique ( Unique, getUnique )
import UniqSupply
-import FastTypes
-import List ( groupBy, sortBy )
-import ErrUtils ( dumpIfSet_dyn )
import DynFlags
-import StaticFlags ( opt_Static, opt_PIC )
+import StaticFlags
import Util
-import Config ( cProjectVersion )
-import Module
+import Config
import Digraph
import qualified Pretty
+import BufWrite
import Outputable
import FastString
import UniqSet
import ErrUtils
+import Module
-- DEBUGGING ONLY
--import OrdList
import Data.List
-import Data.Int
-import Data.Word
-import Data.Bits
import Data.Maybe
-import GHC.Exts
import Control.Monad
+import System.IO
+import Distribution.System
{-
The native-code generator has machine-independent and
-- -----------------------------------------------------------------------------
-- Top-level of the native codegen
--- NB. We *lazilly* compile each block of code for space reasons.
-
--------------------
-nativeCodeGen :: DynFlags -> [RawCmm] -> UniqSupply -> IO Pretty.Doc
-nativeCodeGen dflags cmms us
+nativeCodeGen :: DynFlags -> Handle -> UniqSupply -> [RawCmm] -> IO ()
+nativeCodeGen dflags h us cmms
= do
- -- do native code generation on all these cmm things
- (us', result)
- <- mapAccumLM (cmmNativeGen dflags) us
- $ concat $ map add_split cmms
+ let split_cmms = concat $ map add_split cmms
+
+ -- BufHandle is a performance hack. We could hide it inside
+ -- Pretty if it weren't for the fact that we do lots of little
+ -- printDocs here (in order to do codegen in constant space).
+ bufh <- newBufHandle h
+ (imports, prof) <- cmmNativeGens dflags bufh us split_cmms [] [] 0
+ bFlush bufh
- let (native, imports, mColorStats, mLinearStats)
- = unzip4 result
+ let (native, colorStats, linearStats)
+ = unzip3 prof
+
+ -- dump native code
+ dumpIfSet_dyn dflags
+ Opt_D_dump_asm "Asm code"
+ (vcat $ map (docToSDoc . pprNatCmmTop) $ concat native)
-- dump global NCG stats for graph coloring allocator
- (case concat $ catMaybes mColorStats of
+ (case concat $ catMaybes colorStats of
[] -> return ()
stats -> do
-- build the global register conflict graph
dumpIfSet_dyn dflags
Opt_D_dump_asm_conflicts "Register conflict graph"
- $ Color.dotGraph Color.regDotColor trivColorable
+ $ Color.dotGraph
+ targetRegDotColor
+ (Color.trivColorable
+ targetVirtualRegSqueeze
+ targetRealRegSqueeze)
$ graphGlobal)
-- dump global NCG stats for linear allocator
- (case catMaybes mLinearStats of
+ (case concat $ catMaybes linearStats of
[] -> return ()
stats -> dumpSDoc dflags Opt_D_dump_asm_stats "NCG stats"
- $ Linear.pprStats (concat stats))
+ $ Linear.pprStats (concat native) stats)
+
+ -- write out the imports
+ Pretty.printDoc Pretty.LeftMode h
+ $ makeImportsDoc dflags (concat imports)
- return $ makeAsmDoc (concat native) (concat imports)
+ return ()
- where add_split (Cmm tops)
- | dopt Opt_SplitObjs dflags = split_marker : tops
- | otherwise = tops
+ where add_split (Cmm tops)
+ | dopt Opt_SplitObjs dflags = split_marker : tops
+ | otherwise = tops
- split_marker = CmmProc [] mkSplitMarkerLabel [] []
+ split_marker = CmmProc [] mkSplitMarkerLabel [] (ListGraph [])
+
+
+-- | Do native code generation on all these cmms.
+--
+cmmNativeGens :: DynFlags
+ -> BufHandle
+ -> UniqSupply
+ -> [RawCmmTop]
+ -> [[CLabel]]
+ -> [ ([NatCmmTop Instr],
+ Maybe [Color.RegAllocStats Instr],
+ Maybe [Linear.RegAllocStats]) ]
+ -> Int
+ -> IO ( [[CLabel]],
+ [([NatCmmTop Instr],
+ Maybe [Color.RegAllocStats Instr],
+ Maybe [Linear.RegAllocStats])] )
+
+cmmNativeGens _ _ _ [] impAcc profAcc _
+ = return (reverse impAcc, reverse profAcc)
+
+cmmNativeGens dflags h us (cmm : cmms) impAcc profAcc count
+ = do
+ (us', native, imports, colorStats, linearStats)
+ <- cmmNativeGen dflags us cmm count
+
+ Pretty.bufLeftRender h
+ $ {-# SCC "pprNativeCode" #-} Pretty.vcat $ map pprNatCmmTop native
+
+ -- carefully evaluate this strictly. Binding it with 'let'
+ -- and then using 'seq' doesn't work, because the let
+ -- apparently gets inlined first.
+ lsPprNative <- return $!
+ if dopt Opt_D_dump_asm dflags
+ || dopt Opt_D_dump_asm_stats dflags
+ then native
+ else []
+
+ count' <- return $! count + 1;
+
+ -- force evaulation all this stuff to avoid space leaks
+ seqString (showSDoc $ vcat $ map ppr imports) `seq` return ()
+
+ cmmNativeGens dflags h us' cmms
+ (imports : impAcc)
+ ((lsPprNative, colorStats, linearStats) : profAcc)
+ count'
+
+ where seqString [] = ()
+ seqString (x:xs) = x `seq` seqString xs `seq` ()
-- | Complete native code generation phase for a single top-level chunk of Cmm.
cmmNativeGen
:: DynFlags
-> UniqSupply
- -> RawCmmTop
- -> IO ( UniqSupply
- , ( [NatCmmTop]
- , [CLabel]
- , Maybe [Color.RegAllocStats]
- , Maybe [Linear.RegAllocStats]))
-
-cmmNativeGen dflags us cmm
+ -> RawCmmTop -- ^ the cmm to generate code for
+ -> Int -- ^ sequence number of this top thing
+ -> IO ( UniqSupply
+ , [NatCmmTop Instr] -- native code
+ , [CLabel] -- things imported by this cmm
+ , Maybe [Color.RegAllocStats Instr] -- stats for the coloring register allocator
+ , Maybe [Linear.RegAllocStats]) -- stats for the linear register allocators
+
+cmmNativeGen dflags us cmm count
= do
+
-- rewrite assignments to global regs
- let (fixed_cmm, usFix) =
- initUs us $ fixAssignsTop cmm
+ let fixed_cmm =
+ {-# SCC "fixStgRegisters" #-}
+ fixStgRegisters cmm
-- cmm to cmm optimisations
let (opt_cmm, imports) =
+ {-# SCC "cmmToCmm" #-}
cmmToCmm dflags fixed_cmm
dumpIfSet_dyn dflags
Opt_D_dump_opt_cmm "Optimised Cmm"
(pprCmm $ Cmm [opt_cmm])
-
-- generate native code from cmm
let ((native, lastMinuteImports), usGen) =
- initUs usFix $ genMachCode dflags opt_cmm
+ {-# SCC "genMachCode" #-}
+ initUs us $ genMachCode dflags opt_cmm
dumpIfSet_dyn dflags
Opt_D_dump_asm_native "Native code"
(vcat $ map (docToSDoc . pprNatCmmTop) native)
-
-- tag instructions with register liveness information
let (withLiveness, usLive) =
- initUs usGen $ mapUs regLiveness native
+ {-# SCC "regLiveness" #-}
+ initUs usGen
+ $ mapUs regLiveness
+ $ map natCmmTopToLive native
dumpIfSet_dyn dflags
Opt_D_dump_asm_liveness "Liveness annotations added"
(vcat $ map ppr withLiveness)
-
-- allocate registers
(alloced, usAlloc, ppr_raStatsColor, ppr_raStatsLinear) <-
- if dopt Opt_RegsGraph dflags
+ if ( dopt Opt_RegsGraph dflags
+ || dopt Opt_RegsIterative dflags)
then do
-- the regs usable for allocation
- let alloc_regs
+ let (alloc_regs :: UniqFM (UniqSet RealReg))
= foldr (\r -> plusUFM_C unionUniqSets
- $ unitUFM (regClass r) (unitUniqSet r))
+ $ unitUFM (targetClassOfRealReg r) (unitUniqSet r))
emptyUFM
- $ map RealReg allocatableRegs
+ $ allocatableRegs
- -- aggressively coalesce moves between virtual regs
- let (coalesced, usCoalesce)
- = initUs usLive $ regCoalesce withLiveness
-
- dumpIfSet_dyn dflags
- Opt_D_dump_asm_coalesce "Reg-Reg moves coalesced"
- (vcat $ map ppr coalesced)
-
- -- graph coloring register allocation
+ -- do the graph coloring register allocation
let ((alloced, regAllocStats), usAlloc)
- = initUs usCoalesce
- $ Color.regAlloc
+ = {-# SCC "RegAlloc" #-}
+ initUs usLive
+ $ Color.regAlloc
+ dflags
alloc_regs
(mkUniqSet [0..maxSpillSlots])
- coalesced
+ withLiveness
+ -- dump out what happened during register allocation
dumpIfSet_dyn dflags
Opt_D_dump_asm_regalloc "Registers allocated"
(vcat $ map (docToSDoc . pprNatCmmTop) alloced)
dumpIfSet_dyn dflags
Opt_D_dump_asm_regalloc_stages "Build/spill stages"
(vcat $ map (\(stage, stats)
- -> text "-- Stage " <> int stage
+ -> text "# --------------------------"
+ $$ text "# cmm " <> int count <> text " Stage " <> int stage
$$ ppr stats)
$ zip [0..] regAllocStats)
+ let mPprStats =
+ if dopt Opt_D_dump_asm_stats dflags
+ then Just regAllocStats else Nothing
+
+ -- force evaluation of the Maybe to avoid space leak
+ mPprStats `seq` return ()
+
return ( alloced, usAlloc
- , if dopt Opt_D_dump_asm_stats dflags
- then Just regAllocStats else Nothing
+ , mPprStats
, Nothing)
else do
-- do linear register allocation
let ((alloced, regAllocStats), usAlloc)
- = initUs usLive
- $ liftM unzip
- $ mapUs Linear.regAlloc withLiveness
+ = {-# SCC "RegAlloc" #-}
+ initUs usLive
+ $ liftM unzip
+ $ mapUs Linear.regAlloc withLiveness
dumpIfSet_dyn dflags
Opt_D_dump_asm_regalloc "Registers allocated"
(vcat $ map (docToSDoc . pprNatCmmTop) alloced)
+ let mPprStats =
+ if dopt Opt_D_dump_asm_stats dflags
+ then Just (catMaybes regAllocStats) else Nothing
+
+ -- force evaluation of the Maybe to avoid space leak
+ mPprStats `seq` return ()
+
return ( alloced, usAlloc
, Nothing
- , if dopt Opt_D_dump_asm_stats dflags
- then Just (catMaybes regAllocStats) else Nothing)
+ , mPprStats)
---- shortcut branches
let shorted =
map sequenceTop shorted
---- x86fp_kludge
- let final_mach_code =
+ let kludged =
#if i386_TARGET_ARCH
{-# SCC "x86fp_kludge" #-}
map x86fp_kludge sequenced
sequenced
#endif
+ ---- expansion of SPARC synthetic instrs
+#if sparc_TARGET_ARCH
+ let expanded =
+ {-# SCC "sparc_expand" #-}
+ map expandTop kludged
+
+ dumpIfSet_dyn dflags
+ Opt_D_dump_asm_expanded "Synthetic instructions expanded"
+ (vcat $ map (docToSDoc . pprNatCmmTop) expanded)
+#else
+ let expanded =
+ kludged
+#endif
+
return ( usAlloc
- , ( final_mach_code
- , lastMinuteImports ++ imports
- , ppr_raStatsColor
- , ppr_raStatsLinear) )
+ , expanded
+ , lastMinuteImports ++ imports
+ , ppr_raStatsColor
+ , ppr_raStatsLinear)
#if i386_TARGET_ARCH
-x86fp_kludge :: NatCmmTop -> NatCmmTop
+x86fp_kludge :: NatCmmTop Instr -> NatCmmTop Instr
x86fp_kludge top@(CmmData _ _) = top
-x86fp_kludge top@(CmmProc info lbl params code) =
- CmmProc info lbl params (map bb_i386_insert_ffrees code)
- where
- bb_i386_insert_ffrees (BasicBlock id instrs) =
- BasicBlock id (i386_insert_ffrees instrs)
+x86fp_kludge (CmmProc info lbl params (ListGraph code)) =
+ CmmProc info lbl params (ListGraph $ i386_insert_ffrees code)
#endif
--- | Build assembler source file from native code and its imports.
+-- | Build a doc for all the imports.
--
-makeAsmDoc :: [NatCmmTop] -> [CLabel] -> Pretty.Doc
-makeAsmDoc native imports
- = Pretty.vcat (map pprNatCmmTop native)
- Pretty.$$ (Pretty.text "")
- Pretty.$$ dyld_stubs imports
+makeImportsDoc :: DynFlags -> [CLabel] -> Pretty.Doc
+makeImportsDoc dflags imports
+ = dyld_stubs imports
#if HAVE_SUBSECTIONS_VIA_SYMBOLS
-- On recent versions of Darwin, the linker supports
{- dyld_stubs imps = Pretty.vcat $ map pprDyldSymbolStub $
map head $ group $ sort imps-}
+ arch = platformArch $ targetPlatform dflags
+ os = platformOS $ targetPlatform dflags
+
-- (Hack) sometimes two Labels pretty-print the same, but have
-- different uniques; so we compare their text versions...
dyld_stubs imps
- | needImportedSymbols
+ | needImportedSymbols arch os
= Pretty.vcat $
- (pprGotDeclaration :) $
- map (pprImportedSymbol . fst . head) $
+ (pprGotDeclaration arch os :) $
+ map ( pprImportedSymbol arch os . fst . head) $
groupBy (\(_,a) (_,b) -> a == b) $
sortBy (\(_,a) (_,b) -> compare a b) $
map doPpr $
-- such that as many of the local jumps as possible turn into
-- fallthroughs.
-sequenceTop :: NatCmmTop -> NatCmmTop
+sequenceTop
+ :: NatCmmTop Instr
+ -> NatCmmTop Instr
+
sequenceTop top@(CmmData _ _) = top
-sequenceTop (CmmProc info lbl params blocks) =
- CmmProc info lbl params (makeFarBranches $ sequenceBlocks blocks)
+sequenceTop (CmmProc info lbl params (ListGraph blocks)) =
+ CmmProc info lbl params (ListGraph $ makeFarBranches $ sequenceBlocks blocks)
-- The algorithm is very simple (and stupid): we make a graph out of
-- the blocks where there is an edge from one block to another iff the
-- output the block, then if it has an out edge, we move the
-- destination of the out edge to the front of the list, and continue.
-sequenceBlocks :: [NatBasicBlock] -> [NatBasicBlock]
+-- FYI, the classic layout for basic blocks uses postorder DFS; this
+-- algorithm is implemented in cmm/ZipCfg.hs (NR 6 Sep 2007).
+
+sequenceBlocks
+ :: Instruction instr
+ => [NatBasicBlock instr]
+ -> [NatBasicBlock instr]
+
sequenceBlocks [] = []
sequenceBlocks (entry:blocks) =
seqBlocks (mkNode entry : reverse (flattenSCCs (sccBlocks blocks)))
-- the first block is the entry point ==> it must remain at the start.
-sccBlocks :: [NatBasicBlock] -> [SCC (NatBasicBlock,Unique,[Unique])]
-sccBlocks blocks = stronglyConnCompR (map mkNode blocks)
-getOutEdges :: [Instr] -> [Unique]
-getOutEdges instrs = case jumpDests (last instrs) [] of
- [one] -> [getUnique one]
- _many -> []
- -- we're only interested in the last instruction of
- -- the block, and only if it has a single destination.
+sccBlocks
+ :: Instruction instr
+ => [NatBasicBlock instr]
+ -> [SCC ( NatBasicBlock instr
+ , Unique
+ , [Unique])]
+sccBlocks blocks = stronglyConnCompFromEdgedVerticesR (map mkNode blocks)
+
+-- we're only interested in the last instruction of
+-- the block, and only if it has a single destination.
+getOutEdges
+ :: Instruction instr
+ => [instr] -> [Unique]
+
+getOutEdges instrs
+ = case jumpDestsOfInstr (last instrs) of
+ [one] -> [getUnique one]
+ _many -> []
+
+mkNode :: (Instruction t)
+ => GenBasicBlock t
+ -> (GenBasicBlock t, Unique, [Unique])
mkNode block@(BasicBlock id instrs) = (block, getUnique id, getOutEdges instrs)
+seqBlocks :: (Eq t) => [(GenBasicBlock t1, t, [t])] -> [GenBasicBlock t1]
seqBlocks [] = []
seqBlocks ((block,_,[]) : rest)
= block : seqBlocks rest
-- fallthroughs within a loop.
seqBlocks _ = panic "AsmCodegen:seqBlocks"
-reorder id accum [] = (False, reverse accum)
+reorder :: (Eq a) => a -> [(t, a, t1)] -> [(t, a, t1)] -> (Bool, [(t, a, t1)])
+reorder _ accum [] = (False, reverse accum)
reorder id accum (b@(block,id',out) : rest)
| id == id' = (True, (block,id,out) : reverse accum ++ rest)
| otherwise = reorder id (b:accum) rest
-- Conditional branches on PowerPC are limited to +-32KB; if our Procs get too
-- big, we have to work around this limitation.
-makeFarBranches :: [NatBasicBlock] -> [NatBasicBlock]
+makeFarBranches
+ :: [NatBasicBlock Instr]
+ -> [NatBasicBlock Instr]
#if powerpc_TARGET_ARCH
makeFarBranches blocks
handleBlock addr (BasicBlock id instrs)
= BasicBlock id (zipWith makeFar [addr..] instrs)
- makeFar addr (BCC ALWAYS tgt) = BCC ALWAYS tgt
+ makeFar _ (BCC ALWAYS tgt) = BCC ALWAYS tgt
makeFar addr (BCC cond tgt)
| abs (addr - targetAddr) >= nearLimit
= BCCFAR cond tgt
| otherwise
= BCC cond tgt
where Just targetAddr = lookupUFM blockAddressMap tgt
- makeFar addr other = other
+ makeFar _ other = other
nearLimit = 7000 -- 8192 instructions are allowed; let's keep some
-- distance, as we have a few pseudo-insns that are
-- -----------------------------------------------------------------------------
-- Shortcut branches
-shortcutBranches :: DynFlags -> [NatCmmTop] -> [NatCmmTop]
+shortcutBranches
+ :: DynFlags
+ -> [NatCmmTop Instr]
+ -> [NatCmmTop Instr]
+
shortcutBranches dflags tops
| optLevel dflags < 1 = tops -- only with -O or higher
| otherwise = map (apply_mapping mapping) tops'
(tops', mappings) = mapAndUnzip build_mapping tops
mapping = foldr plusUFM emptyUFM mappings
+build_mapping :: GenCmmTop d t (ListGraph Instr)
+ -> (GenCmmTop d t (ListGraph Instr), UniqFM JumpDest)
build_mapping top@(CmmData _ _) = (top, emptyUFM)
-build_mapping (CmmProc info lbl params [])
- = (CmmProc info lbl params [], emptyUFM)
-build_mapping (CmmProc info lbl params (head:blocks))
- = (CmmProc info lbl params (head:others), mapping)
+build_mapping (CmmProc info lbl params (ListGraph []))
+ = (CmmProc info lbl params (ListGraph []), emptyUFM)
+build_mapping (CmmProc info lbl params (ListGraph (head:blocks)))
+ = (CmmProc info lbl params (ListGraph (head:others)), mapping)
-- drop the shorted blocks, but don't ever drop the first one,
-- because it is pointed to by a global label.
where
-- find all the blocks that just consist of a jump that can be
-- shorted.
- (shortcut_blocks, others) = partitionWith split blocks
- split (BasicBlock id [insn]) | Just dest <- canShortcut insn
- = Left (id,dest)
- split other = Right other
+ -- Don't completely eliminate loops here -- that can leave a dangling jump!
+ (_, shortcut_blocks, others) = foldl split (emptyBlockSet, [], []) blocks
+ split (s, shortcut_blocks, others) b@(BasicBlock id [insn])
+ | Just (DestBlockId dest) <- canShortcut insn,
+ (elemBlockSet dest s) || dest == id -- loop checks
+ = (s, shortcut_blocks, b : others)
+ split (s, shortcut_blocks, others) (BasicBlock id [insn])
+ | Just dest <- canShortcut insn
+ = (extendBlockSet s id, (id,dest) : shortcut_blocks, others)
+ split (s, shortcut_blocks, others) other = (s, shortcut_blocks, other : others)
+
-- build a mapping from BlockId to JumpDest for shorting branches
mapping = foldl add emptyUFM shortcut_blocks
add ufm (id,dest) = addToUFM ufm id dest
+apply_mapping :: UniqFM JumpDest
+ -> GenCmmTop CmmStatic h (ListGraph Instr)
+ -> GenCmmTop CmmStatic h (ListGraph Instr)
apply_mapping ufm (CmmData sec statics)
= CmmData sec (map (shortcutStatic (lookupUFM ufm)) statics)
-- we need to get the jump tables, so apply the mapping to the entries
-- of a CmmData too.
-apply_mapping ufm (CmmProc info lbl params blocks)
- = CmmProc info lbl params (map short_bb blocks)
+apply_mapping ufm (CmmProc info lbl params (ListGraph blocks))
+ = CmmProc info lbl params (ListGraph $ map short_bb blocks)
where
short_bb (BasicBlock id insns) = BasicBlock id $! map short_insn insns
short_insn i = shortcutJump (lookupUFM ufm) i
-- Switching between the two monads whilst carrying along the same
-- Unique supply breaks abstraction. Is that bad?
-genMachCode :: DynFlags -> RawCmmTop -> UniqSM ([NatCmmTop], [CLabel])
+genMachCode
+ :: DynFlags
+ -> RawCmmTop
+ -> UniqSM
+ ( [NatCmmTop Instr]
+ , [CLabel])
genMachCode dflags cmm_top
= do { initial_us <- getUs
; let initial_st = mkNatM_State initial_us 0 dflags
- (new_tops, final_st) = initNat initial_st (cmmTopCodeGen cmm_top)
+ (new_tops, final_st) = initNat initial_st (cmmTopCodeGen dflags cmm_top)
final_delta = natm_delta final_st
final_imports = natm_imports final_st
; if final_delta == 0
else pprPanic "genMachCode: nonzero final delta" (int final_delta)
}
--- -----------------------------------------------------------------------------
--- Fixup assignments to global registers so that they assign to
--- locations within the RegTable, if appropriate.
-
--- Note that we currently don't fixup reads here: they're done by
--- the generic optimiser below, to avoid having two separate passes
--- over the Cmm.
-
-fixAssignsTop :: RawCmmTop -> UniqSM RawCmmTop
-fixAssignsTop top@(CmmData _ _) = returnUs top
-fixAssignsTop (CmmProc info lbl params blocks) =
- mapUs fixAssignsBlock blocks `thenUs` \ blocks' ->
- returnUs (CmmProc info lbl params blocks')
-
-fixAssignsBlock :: CmmBasicBlock -> UniqSM CmmBasicBlock
-fixAssignsBlock (BasicBlock id stmts) =
- fixAssigns stmts `thenUs` \ stmts' ->
- returnUs (BasicBlock id stmts')
-
-fixAssigns :: [CmmStmt] -> UniqSM [CmmStmt]
-fixAssigns stmts =
- mapUs fixAssign stmts `thenUs` \ stmtss ->
- returnUs (concat stmtss)
-
-fixAssign :: CmmStmt -> UniqSM [CmmStmt]
-fixAssign (CmmAssign (CmmGlobal reg) src)
- | Left realreg <- reg_or_addr
- = returnUs [CmmAssign (CmmGlobal reg) src]
- | Right baseRegAddr <- reg_or_addr
- = returnUs [CmmStore baseRegAddr src]
- -- Replace register leaves with appropriate StixTrees for
- -- the given target. GlobalRegs which map to a reg on this
- -- arch are left unchanged. Assigning to BaseReg is always
- -- illegal, so we check for that.
- where
- reg_or_addr = get_GlobalReg_reg_or_addr reg
-
-fixAssign other_stmt = returnUs [other_stmt]
-- -----------------------------------------------------------------------------
-- Generic Cmm optimiser
(a) Constant folding
(b) Simple inlining: a temporary which is assigned to and then
used, once, can be shorted.
- (c) Replacement of references to GlobalRegs which do not have
- machine registers by the appropriate memory load (eg.
- Hp ==> *(BaseReg + 34) ).
- (d) Position independent code and dynamic linking
+ (c) Position independent code and dynamic linking
(i) introduce the appropriate indirections
and position independent refs
(ii) compile a list of imported symbols
cmmToCmm :: DynFlags -> RawCmmTop -> (RawCmmTop, [CLabel])
cmmToCmm _ top@(CmmData _ _) = (top, [])
-cmmToCmm dflags (CmmProc info lbl params blocks) = runCmmOpt dflags $ do
+cmmToCmm dflags (CmmProc info lbl params (ListGraph blocks)) = runCmmOpt dflags $ do
blocks' <- mapM cmmBlockConFold (cmmMiniInline blocks)
- return $ CmmProc info lbl params blocks'
+ return $ CmmProc info lbl params (ListGraph blocks')
newtype CmmOptM a = CmmOptM (([CLabel], DynFlags) -> (# a, [CLabel] #))
CmmOptM g' -> g' (imports', dflags)
addImportCmmOpt :: CLabel -> CmmOptM ()
-addImportCmmOpt lbl = CmmOptM $ \(imports, dflags) -> (# (), lbl:imports #)
+addImportCmmOpt lbl = CmmOptM $ \(imports, _dflags) -> (# (), lbl:imports #)
getDynFlagsCmmOpt :: CmmOptM DynFlags
getDynFlagsCmmOpt = CmmOptM $ \(imports, dflags) -> (# dflags, imports #)
stmts' <- mapM cmmStmtConFold stmts
return $ BasicBlock id stmts'
+cmmStmtConFold :: CmmStmt -> CmmOptM CmmStmt
cmmStmtConFold stmt
= case stmt of
CmmAssign reg src
e' <- cmmExprConFold CallReference e
return $ CmmCallee e' conv
other -> return other
- args' <- mapM (\(arg, hint) -> do
+ args' <- mapM (\(CmmHinted arg hint) -> do
arg' <- cmmExprConFold DataReference arg
- return (arg', hint)) args
+ return (CmmHinted arg' hint)) args
return $ CmmCall target' regs args' srt returns
CmmCondBranch test dest
CmmComment (mkFastString ("deleted: " ++
showSDoc (pprStmt stmt)))
- CmmLit (CmmInt n _) -> CmmBranch dest
- other -> CmmCondBranch test' dest
+ CmmLit (CmmInt _ _) -> CmmBranch dest
+ _other -> CmmCondBranch test' dest
CmmSwitch expr ids
-> do expr' <- cmmExprConFold DataReference expr
-> return other
+cmmExprConFold :: ReferenceKind -> CmmExpr -> CmmOptM CmmExpr
cmmExprConFold referenceKind expr
= case expr of
CmmLoad addr rep
-> do
dflags <- getDynFlagsCmmOpt
dynRef <- cmmMakeDynamicReference dflags addImportCmmOpt referenceKind lbl
- return $ cmmMachOpFold (MO_Add wordRep) [
+ return $ cmmMachOpFold (MO_Add wordWidth) [
dynRef,
- (CmmLit $ CmmInt (fromIntegral off) wordRep)
+ (CmmLit $ CmmInt (fromIntegral off) wordWidth)
]
-#if powerpc_TARGET_ARCH
- -- On powerpc (non-PIC), it's easier to jump directly to a label than
- -- to use the register table, so we replace these registers
- -- with the corresponding labels:
+ -- On powerpc (non-PIC), it's easier to jump directly to a label than
+ -- to use the register table, so we replace these registers
+ -- with the corresponding labels:
+ CmmReg (CmmGlobal EagerBlackholeInfo)
+ | cTargetArch == PPC && not opt_PIC
+ -> cmmExprConFold referenceKind $
+ CmmLit (CmmLabel (mkCmmCodeLabel rtsPackageId (fsLit "__stg_EAGER_BLACKHOLE_info")))
CmmReg (CmmGlobal GCEnter1)
- | not opt_PIC
+ | cTargetArch == PPC && not opt_PIC
-> cmmExprConFold referenceKind $
- CmmLit (CmmLabel (mkRtsCodeLabel SLIT( "__stg_gc_enter_1")))
+ CmmLit (CmmLabel (mkCmmCodeLabel rtsPackageId (fsLit "__stg_gc_enter_1")))
CmmReg (CmmGlobal GCFun)
- | not opt_PIC
+ | cTargetArch == PPC && not opt_PIC
-> cmmExprConFold referenceKind $
- CmmLit (CmmLabel (mkRtsCodeLabel SLIT( "__stg_gc_fun")))
-#endif
+ CmmLit (CmmLabel (mkCmmCodeLabel rtsPackageId (fsLit "__stg_gc_fun")))
- CmmReg (CmmGlobal mid)
- -- Replace register leaves with appropriate StixTrees for
- -- the given target. MagicIds which map to a reg on this
- -- arch are left unchanged. For the rest, BaseReg is taken
- -- to mean the address of the reg table in MainCapability,
- -- and for all others we generate an indirection to its
- -- location in the register table.
- -> case get_GlobalReg_reg_or_addr mid of
- Left realreg -> return expr
- Right baseRegAddr
- -> case mid of
- BaseReg -> cmmExprConFold DataReference baseRegAddr
- other -> cmmExprConFold DataReference
- (CmmLoad baseRegAddr (globalRegRep mid))
- -- eliminate zero offsets
- CmmRegOff reg 0
- -> cmmExprConFold referenceKind (CmmReg reg)
-
- CmmRegOff (CmmGlobal mid) offset
- -- RegOf leaves are just a shorthand form. If the reg maps
- -- to a real reg, we keep the shorthand, otherwise, we just
- -- expand it and defer to the above code.
- -> case get_GlobalReg_reg_or_addr mid of
- Left realreg -> return expr
- Right baseRegAddr
- -> cmmExprConFold DataReference (CmmMachOp (MO_Add wordRep) [
- CmmReg (CmmGlobal mid),
- CmmLit (CmmInt (fromIntegral offset)
- wordRep)])
other
-> return other
--- -----------------------------------------------------------------------------
--- Utils
-
-bind f x = x $! f
-
\end{code}