CodeGen.hs

-----------------------------------------------------------------------------
--
-- Generating machine code (instruction selection)
--
-- (c) The University of Glasgow 1996-2004
--
-----------------------------------------------------------------------------

-- This is a big module, but, if you pay attention to
-- (a) the sectioning, and (b) the type signatures, the
-- structure should not be too overwhelming.

module X86.CodeGen (
        cmmTopCodeGen,
        generateJumpTableForInstr,
        InstrBlock
)

where

#include "HsVersions.h"
#include "nativeGen/NCG.h"
#include "../includes/MachDeps.h"

-- NCG stuff:
import X86.Instr
import X86.Cond
import X86.Regs
import X86.RegInfo
import CodeGen.Platform
import CPrim
import Instruction
import PIC
import NCGMonad
import Size
import Reg
import Platform

-- Our intermediate code:
import BasicTypes
import BlockId
import Module           ( primPackageId )
import PprCmm           ()
import OldCmm
import OldPprCmm        ()
import CLabel

-- The rest:
import ForeignCall      ( CCallConv(..) )
import OrdList
import Outputable
import Unique
import FastString
import FastBool         ( isFastTrue )
import DynFlags
import Util

import Control.Monad
import Data.Bits
import Data.Int
import Data.Maybe
import Data.Word

is32BitPlatform :: NatM Bool
is32BitPlatform = do
    dflags <- getDynFlags
    return $ target32Bit (targetPlatform dflags)

sse2Enabled :: NatM Bool
sse2Enabled = do
  dflags <- getDynFlags
  case platformArch (targetPlatform dflags) of
      ArchX86_64 -> -- SSE2 is fixed on for x86_64.  It would be
                    -- possible to make it optional, but we'd need to
                    -- fix at least the foreign call code where the
                    -- calling convention specifies the use of xmm regs,
                    -- and possibly other places.
                    return True
      ArchX86    -> return (gopt Opt_SSE2 dflags || gopt Opt_SSE4_2 dflags)
      _          -> panic "sse2Enabled: Not an X86* arch"

sse4_2Enabled :: NatM Bool
sse4_2Enabled = do
  dflags <- getDynFlags
  return (gopt Opt_SSE4_2 dflags)

if_sse2 :: NatM a -> NatM a -> NatM a
if_sse2 sse2 x87 = do
  b <- sse2Enabled
  if b then sse2 else x87

cmmTopCodeGen
        :: RawCmmDecl
        -> NatM [NatCmmDecl (Alignment, CmmStatics) Instr]

cmmTopCodeGen (CmmProc info lab live (ListGraph blocks)) = do
  (nat_blocks,statics) <- mapAndUnzipM basicBlockCodeGen blocks
  picBaseMb <- getPicBaseMaybeNat
  dflags <- getDynFlags
  let proc = CmmProc info lab live (ListGraph $ concat nat_blocks)
      tops = proc : concat statics
      os   = platformOS $ targetPlatform dflags

  case picBaseMb of
      Just picBase -> initializePicBase_x86 ArchX86 os picBase tops
      Nothing -> return tops

cmmTopCodeGen (CmmData sec dat) = do
  return [CmmData sec (1, dat)]  -- no translation, we just use CmmStatic


basicBlockCodeGen
        :: CmmBasicBlock
        -> NatM ( [NatBasicBlock Instr]
                , [NatCmmDecl (Alignment, CmmStatics) Instr])

basicBlockCodeGen (BasicBlock id stmts) = do
  instrs <- stmtsToInstrs stmts
  -- code generation may introduce new basic block boundaries, which
  -- are indicated by the NEWBLOCK instruction.  We must split up the
  -- instruction stream into basic blocks again.  Also, we extract
  -- LDATAs here too.
  let
        (top,other_blocks,statics) = foldrOL mkBlocks ([],[],[]) instrs

        mkBlocks (NEWBLOCK id) (instrs,blocks,statics)
          = ([], BasicBlock id instrs : blocks, statics)
        mkBlocks (LDATA sec dat) (instrs,blocks,statics)
          = (instrs, blocks, CmmData sec dat:statics)
        mkBlocks instr (instrs,blocks,statics)
          = (instr:instrs, blocks, statics)
  return (BasicBlock id top : other_blocks, statics)


stmtsToInstrs :: [CmmStmt] -> NatM InstrBlock
stmtsToInstrs stmts
   = do instrss <- mapM stmtToInstrs stmts
        return (concatOL instrss)


stmtToInstrs :: CmmStmt -> NatM InstrBlock
stmtToInstrs stmt = do
  dflags <- getDynFlags
  is32Bit <- is32BitPlatform
  case stmt of
    CmmNop         -> return nilOL
    CmmComment s   -> return (unitOL (COMMENT s))

    CmmAssign reg src
      | isFloatType ty         -> assignReg_FltCode size reg src
      | is32Bit && isWord64 ty -> assignReg_I64Code      reg src
      | otherwise              -> assignReg_IntCode size reg src
        where ty = cmmRegType dflags reg
              size = cmmTypeSize ty

    CmmStore addr src
      | isFloatType ty         -> assignMem_FltCode size addr src
      | is32Bit && isWord64 ty -> assignMem_I64Code      addr src
      | otherwise              -> assignMem_IntCode size addr src
        where ty = cmmExprType dflags src
              size = cmmTypeSize ty

    CmmCall target result_regs args _
       -> genCCall is32Bit target result_regs args

    CmmBranch id          -> genBranch id
    CmmCondBranch arg id  -> genCondJump id arg
    CmmSwitch arg ids     -> do dflags <- getDynFlags
                                genSwitch dflags arg ids
    CmmJump arg gregs     -> do dflags <- getDynFlags
                                genJump arg (jumpRegs dflags gregs)
    CmmReturn             ->
      panic "stmtToInstrs: return statement should have been cps'd away"


jumpRegs :: DynFlags -> [GlobalReg] -> [Reg]
jumpRegs dflags gregs = [ RegReal r | Just r <- map (globalRegMaybe platform) gregs ]
    where platform = targetPlatform dflags

--------------------------------------------------------------------------------
-- | 'InstrBlock's are the insn sequences generated by the insn selectors.
--      They are really trees of insns to facilitate fast appending, where a
--      left-to-right traversal yields the insns in the correct order.
--
type InstrBlock
        = OrdList Instr


-- | Condition codes passed up the tree.
--
data CondCode
        = CondCode Bool Cond InstrBlock


-- | a.k.a "Register64"
--      Reg is the lower 32-bit temporary which contains the result.
--      Use getHiVRegFromLo to find the other VRegUnique.
--
--      Rules of this simplified insn selection game are therefore that
--      the returned Reg may be modified
--
data ChildCode64
   = ChildCode64
        InstrBlock
        Reg


-- | Register's passed up the tree.  If the stix code forces the register
--      to live in a pre-decided machine register, it comes out as @Fixed@;
--      otherwise, it comes out as @Any@, and the parent can decide which
--      register to put it in.
--
data Register
        = Fixed Size Reg InstrBlock
        | Any   Size (Reg -> InstrBlock)


swizzleRegisterRep :: Register -> Size -> Register
swizzleRegisterRep (Fixed _ reg code) size = Fixed size reg code
swizzleRegisterRep (Any _ codefn)     size = Any   size codefn


-- | Grab the Reg for a CmmReg
getRegisterReg :: Platform -> Bool -> CmmReg -> Reg

getRegisterReg _ use_sse2 (CmmLocal (LocalReg u pk))
  = let sz = cmmTypeSize pk in
    if isFloatSize sz && not use_sse2
       then RegVirtual (mkVirtualReg u FF80)
       else RegVirtual (mkVirtualReg u sz)

getRegisterReg platform _ (CmmGlobal mid)
  = case globalRegMaybe platform mid of
        Just reg -> RegReal $ reg
        Nothing  -> pprPanic "getRegisterReg-memory" (ppr $ CmmGlobal mid)
        -- By this stage, the only MagicIds remaining should be the
        -- ones which map to a real machine register on this
        -- platform.  Hence ...


-- | Memory addressing modes passed up the tree.
data Amode
        = Amode AddrMode InstrBlock

{-
Now, given a tree (the argument to an CmmLoad) that references memory,
produce a suitable addressing mode.

A Rule of the Game (tm) for Amodes: use of the addr bit must
immediately follow use of the code part, since the code part puts
values in registers which the addr then refers to.  So you can't put
anything in between, lest it overwrite some of those registers.  If
you need to do some other computation between the code part and use of
the addr bit, first store the effective address from the amode in a
temporary, then do the other computation, and then use the temporary:

    code
    LEA amode, tmp
    ... other computation ...
    ... (tmp) ...
-}


-- | Check whether an integer will fit in 32 bits.
--      A CmmInt is intended to be truncated to the appropriate
--      number of bits, so here we truncate it to Int64.  This is
--      important because e.g. -1 as a CmmInt might be either
--      -1 or 18446744073709551615.
--
is32BitInteger :: Integer -> Bool
is32BitInteger i = i64 <= 0x7fffffff && i64 >= -0x80000000
  where i64 = fromIntegral i :: Int64


-- | Convert a BlockId to some CmmStatic data
jumpTableEntry :: DynFlags -> Maybe BlockId -> CmmStatic
jumpTableEntry dflags Nothing = CmmStaticLit (CmmInt 0 (wordWidth dflags))
jumpTableEntry _ (Just blockid) = CmmStaticLit (CmmLabel blockLabel)
    where blockLabel = mkAsmTempLabel (getUnique blockid)


-- -----------------------------------------------------------------------------
-- General things for putting together code sequences

-- Expand CmmRegOff.  ToDo: should we do it this way around, or convert
-- CmmExprs into CmmRegOff?
mangleIndexTree :: DynFlags -> CmmReg -> Int -> CmmExpr
mangleIndexTree dflags reg off
  = CmmMachOp (MO_Add width) [CmmReg reg, CmmLit (CmmInt (fromIntegral off) width)]
  where width = typeWidth (cmmRegType dflags reg)

-- | The dual to getAnyReg: compute an expression into a register, but
--      we don't mind which one it is.
getSomeReg :: CmmExpr -> NatM (Reg, InstrBlock)
getSomeReg expr = do
  r <- getRegister expr
  case r of
    Any rep code -> do
        tmp <- getNewRegNat rep
        return (tmp, code tmp)
    Fixed _ reg code ->
        return (reg, code)


assignMem_I64Code :: CmmExpr -> CmmExpr -> NatM InstrBlock
assignMem_I64Code addrTree valueTree = do
  Amode addr addr_code <- getAmode addrTree
  ChildCode64 vcode rlo <- iselExpr64 valueTree
  let
        rhi = getHiVRegFromLo rlo

        -- Little-endian store
        mov_lo = MOV II32 (OpReg rlo) (OpAddr addr)
        mov_hi = MOV II32 (OpReg rhi) (OpAddr (fromJust (addrOffset addr 4)))
  return (vcode `appOL` addr_code `snocOL` mov_lo `snocOL` mov_hi)


assignReg_I64Code :: CmmReg  -> CmmExpr -> NatM InstrBlock
assignReg_I64Code (CmmLocal (LocalReg u_dst _)) valueTree = do
   ChildCode64 vcode r_src_lo <- iselExpr64 valueTree
   let
         r_dst_lo = RegVirtual $ mkVirtualReg u_dst II32
         r_dst_hi = getHiVRegFromLo r_dst_lo
         r_src_hi = getHiVRegFromLo r_src_lo
         mov_lo = MOV II32 (OpReg r_src_lo) (OpReg r_dst_lo)
         mov_hi = MOV II32 (OpReg r_src_hi) (OpReg r_dst_hi)
   return (
        vcode `snocOL` mov_lo `snocOL` mov_hi
     )

assignReg_I64Code _ _
   = panic "assignReg_I64Code(i386): invalid lvalue"


iselExpr64        :: CmmExpr -> NatM ChildCode64
iselExpr64 (CmmLit (CmmInt i _)) = do
  (rlo,rhi) <- getNewRegPairNat II32
  let
        r = fromIntegral (fromIntegral i :: Word32)
        q = fromIntegral (fromIntegral (i `shiftR` 32) :: Word32)
        code = toOL [
                MOV II32 (OpImm (ImmInteger r)) (OpReg rlo),
                MOV II32 (OpImm (ImmInteger q)) (OpReg rhi)
                ]
  return (ChildCode64 code rlo)

iselExpr64 (CmmLoad addrTree ty) | isWord64 ty = do
   Amode addr addr_code <- getAmode addrTree
   (rlo,rhi) <- getNewRegPairNat II32
   let
        mov_lo = MOV II32 (OpAddr addr) (OpReg rlo)
        mov_hi = MOV II32 (OpAddr (fromJust (addrOffset addr 4))) (OpReg rhi)
   return (
            ChildCode64 (addr_code `snocOL` mov_lo `snocOL` mov_hi)
                        rlo
     )

iselExpr64 (CmmReg (CmmLocal (LocalReg vu ty))) | isWord64 ty
   = return (ChildCode64 nilOL (RegVirtual $ mkVirtualReg vu II32))

-- we handle addition, but rather badly
iselExpr64 (CmmMachOp (MO_Add _) [e1, CmmLit (CmmInt i _)]) = do
   ChildCode64 code1 r1lo <- iselExpr64 e1
   (rlo,rhi) <- getNewRegPairNat II32
   let
        r = fromIntegral (fromIntegral i :: Word32)
        q = fromIntegral (fromIntegral (i `shiftR` 32) :: Word32)
        r1hi = getHiVRegFromLo r1lo
        code =  code1 `appOL`
                toOL [ MOV II32 (OpReg r1lo) (OpReg rlo),
                       ADD II32 (OpImm (ImmInteger r)) (OpReg rlo),
                       MOV II32 (OpReg r1hi) (OpReg rhi),
                       ADC II32 (OpImm (ImmInteger q)) (OpReg rhi) ]
   return (ChildCode64 code rlo)

iselExpr64 (CmmMachOp (MO_Add _) [e1,e2]) = do
   ChildCode64 code1 r1lo <- iselExpr64 e1
   ChildCode64 code2 r2lo <- iselExpr64 e2
   (rlo,rhi) <- getNewRegPairNat II32
   let
        r1hi = getHiVRegFromLo r1lo
        r2hi = getHiVRegFromLo r2lo
        code =  code1 `appOL`
                code2 `appOL`
                toOL [ MOV II32 (OpReg r1lo) (OpReg rlo),
                       ADD II32 (OpReg r2lo) (OpReg rlo),
                       MOV II32 (OpReg r1hi) (OpReg rhi),
                       ADC II32 (OpReg r2hi) (OpReg rhi) ]
   return (ChildCode64 code rlo)

iselExpr64 (CmmMachOp (MO_UU_Conv _ W64) [expr]) = do
     fn <- getAnyReg expr
     r_dst_lo <-  getNewRegNat II32
     let r_dst_hi = getHiVRegFromLo r_dst_lo
         code = fn r_dst_lo
     return (
             ChildCode64 (code `snocOL`
                          MOV II32 (OpImm (ImmInt 0)) (OpReg r_dst_hi))
                          r_dst_lo
            )

iselExpr64 expr
   = pprPanic "iselExpr64(i386)" (ppr expr)


--------------------------------------------------------------------------------
getRegister :: CmmExpr -> NatM Register
getRegister e = do dflags <- getDynFlags
                   is32Bit <- is32BitPlatform
                   getRegister' dflags is32Bit e

getRegister' :: DynFlags -> Bool -> CmmExpr -> NatM Register

getRegister' dflags is32Bit (CmmReg reg)
  = case reg of
        CmmGlobal PicBaseReg
         | is32Bit ->
            -- on x86_64, we have %rip for PicBaseReg, but it's not
            -- a full-featured register, it can only be used for
            -- rip-relative addressing.
            do reg' <- getPicBaseNat (archWordSize is32Bit)
               return (Fixed (archWordSize is32Bit) reg' nilOL)
        _ ->
            do use_sse2 <- sse2Enabled
               let
                 sz = cmmTypeSize (cmmRegType dflags reg)
                 size | not use_sse2 && isFloatSize sz = FF80
                      | otherwise                      = sz
               --
               let platform = targetPlatform dflags
               return (Fixed size (getRegisterReg platform use_sse2 reg) nilOL)


getRegister' dflags is32Bit (CmmRegOff r n)
  = getRegister' dflags is32Bit $ mangleIndexTree dflags r n

-- for 32-bit architectuers, support some 64 -> 32 bit conversions:
-- TO_W_(x), TO_W_(x >> 32)

getRegister' _ is32Bit (CmmMachOp (MO_UU_Conv W64 W32)
                     [CmmMachOp (MO_U_Shr W64) [x,CmmLit (CmmInt 32 _)]])
 | is32Bit = do
  ChildCode64 code rlo <- iselExpr64 x
  return $ Fixed II32 (getHiVRegFromLo rlo) code

getRegister' _ is32Bit (CmmMachOp (MO_SS_Conv W64 W32)
                     [CmmMachOp (MO_U_Shr W64) [x,CmmLit (CmmInt 32 _)]])
 | is32Bit = do
  ChildCode64 code rlo <- iselExpr64 x
  return $ Fixed II32 (getHiVRegFromLo rlo) code

getRegister' _ is32Bit (CmmMachOp (MO_UU_Conv W64 W32) [x])
 | is32Bit = do
  ChildCode64 code rlo <- iselExpr64 x
  return $ Fixed II32 rlo code

getRegister' _ is32Bit (CmmMachOp (MO_SS_Conv W64 W32) [x])
 | is32Bit = do
  ChildCode64 code rlo <- iselExpr64 x
  return $ Fixed II32 rlo code

getRegister' _ _ (CmmLit lit@(CmmFloat f w)) =
  if_sse2 float_const_sse2 float_const_x87
 where
  float_const_sse2
    | f == 0.0 = do
      let
          size = floatSize w
          code dst = unitOL  (XOR size (OpReg dst) (OpReg dst))
        -- I don't know why there are xorpd, xorps, and pxor instructions.
        -- They all appear to do the same thing --SDM
      return (Any size code)

   | otherwise = do
      Amode addr code <- memConstant (widthInBytes w) lit
      loadFloatAmode True w addr code

  float_const_x87 = case w of
    W64
      | f == 0.0 ->
        let code dst = unitOL (GLDZ dst)
        in  return (Any FF80 code)

      | f == 1.0 ->
        let code dst = unitOL (GLD1 dst)
        in  return (Any FF80 code)

    _otherwise -> do
      Amode addr code <- memConstant (widthInBytes w) lit
      loadFloatAmode False w addr code

-- catch simple cases of zero- or sign-extended load
getRegister' _ _ (CmmMachOp (MO_UU_Conv W8 W32) [CmmLoad addr _]) = do
  code <- intLoadCode (MOVZxL II8) addr
  return (Any II32 code)

getRegister' _ _ (CmmMachOp (MO_SS_Conv W8 W32) [CmmLoad addr _]) = do
  code <- intLoadCode (MOVSxL II8) addr
  return (Any II32 code)

getRegister' _ _ (CmmMachOp (MO_UU_Conv W16 W32) [CmmLoad addr _]) = do
  code <- intLoadCode (MOVZxL II16) addr
  return (Any II32 code)

getRegister' _ _ (CmmMachOp (MO_SS_Conv W16 W32) [CmmLoad addr _]) = do
  code <- intLoadCode (MOVSxL II16) addr
  return (Any II32 code)

-- catch simple cases of zero- or sign-extended load
getRegister' _ is32Bit (CmmMachOp (MO_UU_Conv W8 W64) [CmmLoad addr _])
 | not is32Bit = do
  code <- intLoadCode (MOVZxL II8) addr
  return (Any II64 code)

getRegister' _ is32Bit (CmmMachOp (MO_SS_Conv W8 W64) [CmmLoad addr _])
 | not is32Bit = do
  code <- intLoadCode (MOVSxL II8) addr
  return (Any II64 code)

getRegister' _ is32Bit (CmmMachOp (MO_UU_Conv W16 W64) [CmmLoad addr _])
 | not is32Bit = do
  code <- intLoadCode (MOVZxL II16) addr
  return (Any II64 code)

getRegister' _ is32Bit (CmmMachOp (MO_SS_Conv W16 W64) [CmmLoad addr _])
 | not is32Bit = do
  code <- intLoadCode (MOVSxL II16) addr
  return (Any II64 code)

getRegister' _ is32Bit (CmmMachOp (MO_UU_Conv W32 W64) [CmmLoad addr _])
 | not is32Bit = do
  code <- intLoadCode (MOV II32) addr -- 32-bit loads zero-extend
  return (Any II64 code)

getRegister' _ is32Bit (CmmMachOp (MO_SS_Conv W32 W64) [CmmLoad addr _])
 | not is32Bit = do
  code <- intLoadCode (MOVSxL II32) addr
  return (Any II64 code)

getRegister' _ is32Bit (CmmMachOp (MO_Add W64) [CmmReg (CmmGlobal PicBaseReg),
                                     CmmLit displacement])
 | not is32Bit = do
      return $ Any II64 (\dst -> unitOL $
        LEA II64 (OpAddr (ripRel (litToImm displacement))) (OpReg dst))

getRegister' dflags is32Bit (CmmMachOp mop [x]) = do -- unary MachOps
    sse2 <- sse2Enabled
    case mop of
      MO_F_Neg w
         | sse2      -> sse2NegCode w x
         | otherwise -> trivialUFCode FF80 (GNEG FF80) x

      MO_S_Neg w -> triv_ucode NEGI (intSize w)
      MO_Not w   -> triv_ucode NOT  (intSize w)

      -- Nop conversions
      MO_UU_Conv W32 W8  -> toI8Reg  W32 x
      MO_SS_Conv W32 W8  -> toI8Reg  W32 x
      MO_UU_Conv W16 W8  -> toI8Reg  W16 x
      MO_SS_Conv W16 W8  -> toI8Reg  W16 x
      MO_UU_Conv W32 W16 -> toI16Reg W32 x
      MO_SS_Conv W32 W16 -> toI16Reg W32 x

      MO_UU_Conv W64 W32 | not is32Bit -> conversionNop II64 x
      MO_SS_Conv W64 W32 | not is32Bit -> conversionNop II64 x
      MO_UU_Conv W64 W16 | not is32Bit -> toI16Reg W64 x
      MO_SS_Conv W64 W16 | not is32Bit -> toI16Reg W64 x
      MO_UU_Conv W64 W8  | not is32Bit -> toI8Reg  W64 x
      MO_SS_Conv W64 W8  | not is32Bit -> toI8Reg  W64 x

      MO_UU_Conv rep1 rep2 | rep1 == rep2 -> conversionNop (intSize rep1) x
      MO_SS_Conv rep1 rep2 | rep1 == rep2 -> conversionNop (intSize rep1) x

      -- widenings
      MO_UU_Conv W8  W32 -> integerExtend W8  W32 MOVZxL x
      MO_UU_Conv W16 W32 -> integerExtend W16 W32 MOVZxL x
      MO_UU_Conv W8  W16 -> integerExtend W8  W16 MOVZxL x

      MO_SS_Conv W8  W32 -> integerExtend W8  W32 MOVSxL x
      MO_SS_Conv W16 W32 -> integerExtend W16 W32 MOVSxL x
      MO_SS_Conv W8  W16 -> integerExtend W8  W16 MOVSxL x

      MO_UU_Conv W8  W64 | not is32Bit -> integerExtend W8  W64 MOVZxL x
      MO_UU_Conv W16 W64 | not is32Bit -> integerExtend W16 W64 MOVZxL x
      MO_UU_Conv W32 W64 | not is32Bit -> integerExtend W32 W64 MOVZxL x
      MO_SS_Conv W8  W64 | not is32Bit -> integerExtend W8  W64 MOVSxL x
      MO_SS_Conv W16 W64 | not is32Bit -> integerExtend W16 W64 MOVSxL x
      MO_SS_Conv W32 W64 | not is32Bit -> integerExtend W32 W64 MOVSxL x
        -- for 32-to-64 bit zero extension, amd64 uses an ordinary movl.
        -- However, we don't want the register allocator to throw it
        -- away as an unnecessary reg-to-reg move, so we keep it in
        -- the form of a movzl and print it as a movl later.

      MO_FF_Conv W32 W64
        | sse2      -> coerceFP2FP W64 x
        | otherwise -> conversionNop FF80 x

      MO_FF_Conv W64 W32 -> coerceFP2FP W32 x

      MO_FS_Conv from to -> coerceFP2Int from to x
      MO_SF_Conv from to -> coerceInt2FP from to x

      _other -> pprPanic "getRegister" (pprMachOp mop)
   where
        triv_ucode :: (Size -> Operand -> Instr) -> Size -> NatM Register
        triv_ucode instr size = trivialUCode size (instr size) x

        -- signed or unsigned extension.
        integerExtend :: Width -> Width
                      -> (Size -> Operand -> Operand -> Instr)
                      -> CmmExpr -> NatM Register
        integerExtend from to instr expr = do
            (reg,e_code) <- if from == W8 then getByteReg expr
                                          else getSomeReg expr
            let
                code dst =
                  e_code `snocOL`
                  instr (intSize from) (OpReg reg) (OpReg dst)
            return (Any (intSize to) code)

        toI8Reg :: Width -> CmmExpr -> NatM Register
        toI8Reg new_rep expr
            = do codefn <- getAnyReg expr
                 return (Any (intSize new_rep) codefn)
                -- HACK: use getAnyReg to get a byte-addressable register.
                -- If the source was a Fixed register, this will add the
                -- mov instruction to put it into the desired destination.
                -- We're assuming that the destination won't be a fixed
                -- non-byte-addressable register; it won't be, because all
                -- fixed registers are word-sized.

        toI16Reg = toI8Reg -- for now

        conversionNop :: Size -> CmmExpr -> NatM Register
        conversionNop new_size expr
            = do e_code <- getRegister' dflags is32Bit expr
                 return (swizzleRegisterRep e_code new_size)


getRegister' _ is32Bit (CmmMachOp mop [x, y]) = do -- dyadic MachOps
  sse2 <- sse2Enabled
  case mop of
      MO_F_Eq _ -> condFltReg is32Bit EQQ x y
      MO_F_Ne _ -> condFltReg is32Bit NE  x y
      MO_F_Gt _ -> condFltReg is32Bit GTT x y
      MO_F_Ge _ -> condFltReg is32Bit GE  x y
      MO_F_Lt _ -> condFltReg is32Bit LTT x y
      MO_F_Le _ -> condFltReg is32Bit LE  x y

      MO_Eq _   -> condIntReg EQQ x y
      MO_Ne _   -> condIntReg NE  x y

      MO_S_Gt _ -> condIntReg GTT x y
      MO_S_Ge _ -> condIntReg GE  x y
      MO_S_Lt _ -> condIntReg LTT x y
      MO_S_Le _ -> condIntReg LE  x y

      MO_U_Gt _ -> condIntReg GU  x y
      MO_U_Ge _ -> condIntReg GEU x y
      MO_U_Lt _ -> condIntReg LU  x y
      MO_U_Le _ -> condIntReg LEU x y

      MO_F_Add w  | sse2      -> trivialFCode_sse2 w ADD  x y
                  | otherwise -> trivialFCode_x87    GADD x y
      MO_F_Sub w  | sse2      -> trivialFCode_sse2 w SUB  x y
                  | otherwise -> trivialFCode_x87    GSUB x y
      MO_F_Quot w | sse2      -> trivialFCode_sse2 w FDIV x y
                  | otherwise -> trivialFCode_x87    GDIV x y
      MO_F_Mul w  | sse2      -> trivialFCode_sse2 w MUL x y
                  | otherwise -> trivialFCode_x87    GMUL x y

      MO_Add rep -> add_code rep x y
      MO_Sub rep -> sub_code rep x y

      MO_S_Quot rep -> div_code rep True  True  x y
      MO_S_Rem  rep -> div_code rep True  False x y
      MO_U_Quot rep -> div_code rep False True  x y
      MO_U_Rem  rep -> div_code rep False False x y

      MO_S_MulMayOflo rep -> imulMayOflo rep x y

      MO_Mul rep -> triv_op rep IMUL
      MO_And rep -> triv_op rep AND
      MO_Or  rep -> triv_op rep OR
      MO_Xor rep -> triv_op rep XOR

        {- Shift ops on x86s have constraints on their source, it
           either has to be Imm, CL or 1
            => trivialCode is not restrictive enough (sigh.)
        -}
      MO_Shl rep   -> shift_code rep SHL x y {-False-}
      MO_U_Shr rep -> shift_code rep SHR x y {-False-}
      MO_S_Shr rep -> shift_code rep SAR x y {-False-}

      _other -> pprPanic "getRegister(x86) - binary CmmMachOp (1)" (pprMachOp mop)
  where
    --------------------
    triv_op width instr = trivialCode width op (Just op) x y
                        where op   = instr (intSize width)

    imulMayOflo :: Width -> CmmExpr -> CmmExpr -> NatM Register
    imulMayOflo rep a b = do
         (a_reg, a_code) <- getNonClobberedReg a
         b_code <- getAnyReg b
         let
             shift_amt  = case rep of
                           W32 -> 31
                           W64 -> 63
                           _ -> panic "shift_amt"

             size = intSize rep
             code = a_code `appOL` b_code eax `appOL`
                        toOL [
                           IMUL2 size (OpReg a_reg),   -- result in %edx:%eax
                           SAR size (OpImm (ImmInt shift_amt)) (OpReg eax),
                                -- sign extend lower part
                           SUB size (OpReg edx) (OpReg eax)
                                -- compare against upper
                           -- eax==0 if high part == sign extended low part
                        ]
         return (Fixed size eax code)

    --------------------
    shift_code :: Width
               -> (Size -> Operand -> Operand -> Instr)
               -> CmmExpr
               -> CmmExpr
               -> NatM Register

    {- Case1: shift length as immediate -}
    shift_code width instr x (CmmLit lit) = do
          x_code <- getAnyReg x
          let
               size = intSize width
               code dst
                  = x_code dst `snocOL`
                    instr size (OpImm (litToImm lit)) (OpReg dst)
          return (Any size code)

    {- Case2: shift length is complex (non-immediate)
      * y must go in %ecx.
      * we cannot do y first *and* put its result in %ecx, because
        %ecx might be clobbered by x.
      * if we do y second, then x cannot be
        in a clobbered reg.  Also, we cannot clobber x's reg
        with the instruction itself.
      * so we can either:
        - do y first, put its result in a fresh tmp, then copy it to %ecx later
        - do y second and put its result into %ecx.  x gets placed in a fresh
          tmp.  This is likely to be better, becuase the reg alloc can
          eliminate this reg->reg move here (it won't eliminate the other one,
          because the move is into the fixed %ecx).
    -}
    shift_code width instr x y{-amount-} = do
        x_code <- getAnyReg x
        let size = intSize width
        tmp <- getNewRegNat size
        y_code <- getAnyReg y
        let
           code = x_code tmp `appOL`
                  y_code ecx `snocOL`
                  instr size (OpReg ecx) (OpReg tmp)
        return (Fixed size tmp code)

    --------------------
    add_code :: Width -> CmmExpr -> CmmExpr -> NatM Register
    add_code rep x (CmmLit (CmmInt y _))
        | is32BitInteger y = add_int rep x y
    add_code rep x y = trivialCode rep (ADD size) (Just (ADD size)) x y
      where size = intSize rep

    --------------------
    sub_code :: Width -> CmmExpr -> CmmExpr -> NatM Register
    sub_code rep x (CmmLit (CmmInt y _))
        | is32BitInteger (-y) = add_int rep x (-y)
    sub_code rep x y = trivialCode rep (SUB (intSize rep)) Nothing x y

    -- our three-operand add instruction:
    add_int width x y = do
        (x_reg, x_code) <- getSomeReg x
        let
            size = intSize width
            imm = ImmInt (fromInteger y)
            code dst
               = x_code `snocOL`
                 LEA size
                        (OpAddr (AddrBaseIndex (EABaseReg x_reg) EAIndexNone imm))
                        (OpReg dst)
        --
        return (Any size code)

    ----------------------
    div_code width signed quotient x y = do
           (y_op, y_code) <- getRegOrMem y -- cannot be clobbered
           x_code <- getAnyReg x
           let
             size = intSize width
             widen | signed    = CLTD size
                   | otherwise = XOR size (OpReg edx) (OpReg edx)

             instr | signed    = IDIV
                   | otherwise = DIV

             code = y_code `appOL`
                    x_code eax `appOL`
                    toOL [widen, instr size y_op]

             result | quotient  = eax
                    | otherwise = edx

           return (Fixed size result code)


getRegister' _ _ (CmmLoad mem pk)
  | isFloatType pk
  = do
    Amode addr mem_code <- getAmode mem
    use_sse2 <- sse2Enabled
    loadFloatAmode use_sse2 (typeWidth pk) addr mem_code

getRegister' _ is32Bit (CmmLoad mem pk)
  | is32Bit && not (isWord64 pk)
  = do
    code <- intLoadCode instr mem
    return (Any size code)
  where
    width = typeWidth pk
    size = intSize width
    instr = case width of
                W8     -> MOVZxL II8
                _other -> MOV size
        -- We always zero-extend 8-bit loads, if we
        -- can't think of anything better.  This is because
        -- we can't guarantee access to an 8-bit variant of every register
        -- (esi and edi don't have 8-bit variants), so to make things
        -- simpler we do our 8-bit arithmetic with full 32-bit registers.

-- Simpler memory load code on x86_64
getRegister' _ is32Bit (CmmLoad mem pk)
 | not is32Bit
  = do
    code <- intLoadCode (MOV size) mem
    return (Any size code)
  where size = intSize $ typeWidth pk

getRegister' _ is32Bit (CmmLit (CmmInt 0 width))
  = let
        size = intSize width

        -- x86_64: 32-bit xor is one byte shorter, and zero-extends to 64 bits
        size1 = if is32Bit then size
                           else case size of
                                II64 -> II32
                                _ -> size
        code dst
           = unitOL (XOR size1 (OpReg dst) (OpReg dst))
    in
        return (Any size code)

  -- optimisation for loading small literals on x86_64: take advantage
  -- of the automatic zero-extension from 32 to 64 bits, because the 32-bit
  -- instruction forms are shorter.
getRegister' dflags is32Bit (CmmLit lit)
  | not is32Bit, isWord64 (cmmLitType dflags lit), not (isBigLit lit)
  = let
        imm = litToImm lit
        code dst = unitOL (MOV II32 (OpImm imm) (OpReg dst))
    in
        return (Any II64 code)
  where
   isBigLit (CmmInt i _) = i < 0 || i > 0xffffffff
   isBigLit _ = False
        -- note1: not the same as (not.is32BitLit), because that checks for
        -- signed literals that fit in 32 bits, but we want unsigned
        -- literals here.
        -- note2: all labels are small, because we're assuming the
        -- small memory model (see gcc docs, -mcmodel=small).

getRegister' dflags _ (CmmLit lit)
  = do let size = cmmTypeSize (cmmLitType dflags lit)
           imm = litToImm lit
           code dst = unitOL (MOV size (OpImm imm) (OpReg dst))
       return (Any size code)

getRegister' _ _ other = pprPanic "getRegister(x86)" (ppr other)


intLoadCode :: (Operand -> Operand -> Instr) -> CmmExpr
   -> NatM (Reg -> InstrBlock)
intLoadCode instr mem = do
  Amode src mem_code <- getAmode mem
  return (\dst -> mem_code `snocOL` instr (OpAddr src) (OpReg dst))

-- Compute an expression into *any* register, adding the appropriate
-- move instruction if necessary.
getAnyReg :: CmmExpr -> NatM (Reg -> InstrBlock)
getAnyReg expr = do
  r <- getRegister expr
  anyReg r

anyReg :: Register -> NatM (Reg -> InstrBlock)
anyReg (Any _ code)          = return code
anyReg (Fixed rep reg fcode) = return (\dst -> fcode `snocOL` reg2reg rep reg dst)

-- A bit like getSomeReg, but we want a reg that can be byte-addressed.
-- Fixed registers might not be byte-addressable, so we make sure we've
-- got a temporary, inserting an extra reg copy if necessary.
getByteReg :: CmmExpr -> NatM (Reg, InstrBlock)
getByteReg expr = do
  is32Bit <- is32BitPlatform
  if is32Bit
      then do r <- getRegister expr
              case r of
                Any rep code -> do
                    tmp <- getNewRegNat rep
                    return (tmp, code tmp)
                Fixed rep reg code
                    | isVirtualReg reg -> return (reg,code)
                    | otherwise -> do
                        tmp <- getNewRegNat rep
                        return (tmp, code `snocOL` reg2reg rep reg tmp)
                    -- ToDo: could optimise slightly by checking for
                    -- byte-addressable real registers, but that will
                    -- happen very rarely if at all.
      else getSomeReg expr -- all regs are byte-addressable on x86_64

-- Another variant: this time we want the result in a register that cannot
-- be modified by code to evaluate an arbitrary expression.
getNonClobberedReg :: CmmExpr -> NatM (Reg, InstrBlock)
getNonClobberedReg expr = do
  dflags <- getDynFlags
  r <- getRegister expr
  case r of
    Any rep code -> do
        tmp <- getNewRegNat rep
        return (tmp, code tmp)
    Fixed rep reg code
        -- only certain regs can be clobbered
        | reg `elem` instrClobberedRegs (targetPlatform dflags)
        -> do
                tmp <- getNewRegNat rep
                return (tmp, code `snocOL` reg2reg rep reg tmp)
        | otherwise ->
                return (reg, code)

reg2reg :: Size -> Reg -> Reg -> Instr
reg2reg size src dst
  | size == FF80 = GMOV src dst
  | otherwise    = MOV size (OpReg src) (OpReg dst)


--------------------------------------------------------------------------------
getAmode :: CmmExpr -> NatM Amode
getAmode e = do is32Bit <- is32BitPlatform
                getAmode' is32Bit e

getAmode' :: Bool -> CmmExpr -> NatM Amode
getAmode' _ (CmmRegOff r n) = do dflags <- getDynFlags
                                 getAmode $ mangleIndexTree dflags r n

getAmode' is32Bit (CmmMachOp (MO_Add W64) [CmmReg (CmmGlobal PicBaseReg),
                                                  CmmLit displacement])
 | not is32Bit
    = return $ Amode (ripRel (litToImm displacement)) nilOL


-- This is all just ridiculous, since it carefully undoes
-- what mangleIndexTree has just done.
getAmode' is32Bit (CmmMachOp (MO_Sub _rep) [x, CmmLit lit@(CmmInt i _)])
  | is32BitLit is32Bit lit
  -- ASSERT(rep == II32)???
  = do (x_reg, x_code) <- getSomeReg x
       let off = ImmInt (-(fromInteger i))
       return (Amode (AddrBaseIndex (EABaseReg x_reg) EAIndexNone off) x_code)

getAmode' is32Bit (CmmMachOp (MO_Add _rep) [x, CmmLit lit])
  | is32BitLit is32Bit lit