-- -----------------------------------------------------------------------------
\begin{code}
+{-# OPTIONS -w #-}
+-- The above warning supression flag is a temporary kludge.
+-- While working on this module you are encouraged to remove it and fix
+-- any warnings in the module. See
+-- http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings
+-- for details
+
#include "nativeGen/NCG.h"
module MachRegs (
+ -- * Sizes
+ Size(..), intSize, floatSize, isFloatSize,
+ wordSize, cmmTypeSize, sizeToWidth,
+
-- * Immediate values
Imm(..), strImmLit, litToImm,
-- * The 'Reg' type
RegNo,
- Reg(..), isRealReg, isVirtualReg,
+ Reg(..), isRealReg, isVirtualReg, renameVirtualReg,
RegClass(..), regClass,
+ trivColorable,
getHiVRegFromLo,
mkVReg,
-- * Machine-dependent register-related stuff
allocatableRegs, argRegs, allArgRegs, callClobberedRegs,
+ allocatableRegsInClass,
freeReg,
spRel,
-- HACK: go for the max
#endif
-#include "MachRegs.h"
+#include "../includes/MachRegs.h"
import Cmm
-import MachOp ( MachRep(..) )
import CgUtils ( get_GlobalReg_addr )
-
import CLabel ( CLabel, mkMainCapabilityLabel )
import Pretty
import Outputable ( Outputable(..), pprPanic, panic )
import qualified Outputable
import Unique
+import UniqSet
import Constants
import FastTypes
+import FastBool
+import UniqFM
#if powerpc_TARGET_ARCH
import Data.Word ( Word8, Word16, Word32 )
#endif
-- -----------------------------------------------------------------------------
+-- Sizes on this architecture
+--
+-- A Size is usually a combination of width and class
+
+-- It looks very like the old MachRep, but it's now of purely local
+-- significance, here in the native code generator. You can change it
+-- without global consequences.
+--
+-- A major use is as an opcode qualifier; thus the opcode
+-- mov.l a b
+-- might be encoded
+-- MOV II32 a b
+-- where the Size field encodes the ".l" part.
+
+-- ToDo: it's not clear to me that we need separate signed-vs-unsigned sizes
+-- here. I've removed them from the x86 version, we'll see what happens --SDM
+
+-- ToDo: quite a few occurrences of Size could usefully be replaced by Width
+
+#if powerpc_TARGET_ARCH || i386_TARGET_ARCH || x86_64_TARGET_ARCH
+data Size -- For these three, the "size" also gives the int/float
+ -- distinction, because the instructions for int/float
+ -- differ only in their suffices
+ = II8 | II16 | II32 | II64 | FF32 | FF64 | FF80
+ deriving Eq
+
+intSize, floatSize :: Width -> Size
+intSize W8 = II8
+intSize W16 = II16
+intSize W32 = II32
+intSize W64 = II64
+intSize other = pprPanic "MachInstrs.intSize" (ppr other)
+
+floatSize W32 = FF32
+floatSize W64 = FF64
+floatSize other = pprPanic "MachInstrs.intSize" (ppr other)
+
+wordSize :: Size
+wordSize = intSize wordWidth
+
+sizeToWidth :: Size -> Width
+sizeToWidth II8 = W8
+sizeToWidth II16 = W16
+sizeToWidth II32 = W32
+sizeToWidth II64 = W64
+sizeToWidth FF32 = W32
+sizeToWidth FF64 = W64
+sizeToWidth _ = panic "MachInstrs.sizeToWidth"
+
+cmmTypeSize :: CmmType -> Size
+cmmTypeSize ty | isFloatType ty = floatSize (typeWidth ty)
+ | otherwise = intSize (typeWidth ty)
+
+isFloatSize :: Size -> Bool
+isFloatSize FF32 = True
+isFloatSize FF64 = True
+isFloatSize FF80 = True
+isFloatSize other = False
+#endif
+
+#if alpha_TARGET_ARCH
+data Size
+ = B -- byte
+ | Bu
+-- | W -- word (2 bytes): UNUSED
+-- | Wu -- : UNUSED
+ | L -- longword (4 bytes)
+ | Q -- quadword (8 bytes)
+-- | FF -- VAX F-style floating pt: UNUSED
+-- | GF -- VAX G-style floating pt: UNUSED
+-- | DF -- VAX D-style floating pt: UNUSED
+-- | SF -- IEEE single-precision floating pt: UNUSED
+ | TF -- IEEE double-precision floating pt
+ deriving Eq
+#endif
+
+#if sparc_TARGET_ARCH /* || powerpc_TARGET_ARCH */
+data Size
+ = B -- byte (signed)
+ | Bu -- byte (unsigned)
+ | H -- halfword (signed, 2 bytes)
+ | Hu -- halfword (unsigned, 2 bytes)
+ | W -- word (4 bytes)
+ | F -- IEEE single-precision floating pt
+ | DF -- IEEE single-precision floating pt
+ deriving Eq
+#endif
+
+-- -----------------------------------------------------------------------------
-- Immediates
data Imm
litToImm :: CmmLit -> Imm
litToImm (CmmInt i _) = ImmInteger i
-litToImm (CmmFloat f F32) = ImmFloat f
-litToImm (CmmFloat f F64) = ImmDouble f
+litToImm (CmmFloat f W32) = ImmFloat f
+litToImm (CmmFloat f W64) = ImmDouble f
litToImm (CmmLabel l) = ImmCLbl l
litToImm (CmmLabelOff l off) = ImmIndex l off
litToImm (CmmLabelDiffOff l1 l2 off)
fits16Bits :: Integral a => a -> Bool
fits16Bits x = x >= -32768 && x < 32768
-makeImmediate :: Integral a => MachRep -> Bool -> a -> Maybe Imm
-
+makeImmediate :: Integral a => Width -> Bool -> a -> Maybe Imm
makeImmediate rep signed x = fmap ImmInt (toI16 rep signed)
where
- narrow I32 False = fromIntegral (fromIntegral x :: Word32)
- narrow I16 False = fromIntegral (fromIntegral x :: Word16)
- narrow I8 False = fromIntegral (fromIntegral x :: Word8)
- narrow I32 True = fromIntegral (fromIntegral x :: Int32)
- narrow I16 True = fromIntegral (fromIntegral x :: Int16)
- narrow I8 True = fromIntegral (fromIntegral x :: Int8)
+ narrow W32 False = fromIntegral (fromIntegral x :: Word32)
+ narrow W16 False = fromIntegral (fromIntegral x :: Word16)
+ narrow W8 False = fromIntegral (fromIntegral x :: Word8)
+ narrow W32 True = fromIntegral (fromIntegral x :: Int32)
+ narrow W16 True = fromIntegral (fromIntegral x :: Int16)
+ narrow W8 True = fromIntegral (fromIntegral x :: Int8)
narrowed = narrow rep signed
- toI16 I32 True
+ toI16 W32 True
| narrowed >= -32768 && narrowed < 32768 = Just narrowed
| otherwise = Nothing
- toI16 I32 False
+ toI16 W32 False
| narrowed >= 0 && narrowed < 65536 = Just narrowed
| otherwise = Nothing
toI16 _ _ = Just narrowed
| RcDouble
deriving Eq
+instance Uniquable RegClass where
+ getUnique RcInteger = mkUnique 'L' 0
+ getUnique RcFloat = mkUnique 'L' 1
+ getUnique RcDouble = mkUnique 'L' 2
+
type RegNo = Int
data Reg
unRealReg (RealReg i) = i
unRealReg vreg = pprPanic "unRealReg on VirtualReg" (ppr vreg)
-mkVReg :: Unique -> MachRep -> Reg
-mkVReg u rep
- = case rep of
+mkVReg :: Unique -> Size -> Reg
+mkVReg u size
+ | not (isFloatSize size) = VirtualRegI u
+ | otherwise
+ = case size of
#if sparc_TARGET_ARCH
- F32 -> VirtualRegF u
+ FF32 -> VirtualRegF u
#else
- F32 -> VirtualRegD u
+ FF32 -> VirtualRegD u
#endif
- F64 -> VirtualRegD u
- other -> VirtualRegI u
+ FF64 -> VirtualRegD u
+ _other -> panic "mkVReg"
isVirtualReg :: Reg -> Bool
isVirtualReg (RealReg _) = False
isRealReg :: Reg -> Bool
isRealReg = not . isVirtualReg
+renameVirtualReg :: Unique -> Reg -> Reg
+renameVirtualReg u r
+ = case r of
+ RealReg _ -> error "renameVirtualReg: can't change unique on a real reg"
+ VirtualRegI _ -> VirtualRegI u
+ VirtualRegHi _ -> VirtualRegHi u
+ VirtualRegF _ -> VirtualRegF u
+ VirtualRegD _ -> VirtualRegD u
+
instance Show Reg where
show (RealReg i) = showReg i
show (VirtualRegI u) = "%vI_" ++ show u
show (VirtualRegF u) = "%vF_" ++ show u
show (VirtualRegD u) = "%vD_" ++ show u
+instance Outputable RegClass where
+ ppr RcInteger = Outputable.text "I"
+ ppr RcFloat = Outputable.text "F"
+ ppr RcDouble = Outputable.text "D"
+
instance Outputable Reg where
ppr r = Outputable.text (show r)
+
+
+-- trivColorable function for the graph coloring allocator
+-- This gets hammered by scanGraph during register allocation,
+-- so needs to be fairly efficient.
+--
+-- NOTE: This only works for arcitectures with just RcInteger and RcDouble
+-- (which are disjoint) ie. x86, x86_64 and ppc
+--
+
+-- BL 2007/09
+-- Doing a nice fold over the UniqSet makes trivColorable use
+-- 32% of total compile time and 42% of total alloc when compiling SHA1.lhs from darcs.
+{-
+trivColorable :: RegClass -> UniqSet Reg -> UniqSet Reg -> Bool
+trivColorable classN conflicts exclusions
+ = let
+
+ acc :: Reg -> (Int, Int) -> (Int, Int)
+ acc r (cd, cf)
+ = case regClass r of
+ RcInteger -> (cd+1, cf)
+ RcDouble -> (cd, cf+1)
+ _ -> panic "MachRegs.trivColorable: reg class not handled"
+
+ tmp = foldUniqSet acc (0, 0) conflicts
+ (countInt, countFloat) = foldUniqSet acc tmp exclusions
+
+ squeese = worst countInt classN RcInteger
+ + worst countFloat classN RcDouble
+
+ in squeese < allocatableRegsInClass classN
+
+-- | Worst case displacement
+-- node N of classN has n neighbors of class C.
+--
+-- We currently only have RcInteger and RcDouble, which don't conflict at all.
+-- This is a bit boring compared to what's in RegArchX86.
+--
+worst :: Int -> RegClass -> RegClass -> Int
+worst n classN classC
+ = case classN of
+ RcInteger
+ -> case classC of
+ RcInteger -> min n (allocatableRegsInClass RcInteger)
+ RcDouble -> 0
+
+ RcDouble
+ -> case classC of
+ RcDouble -> min n (allocatableRegsInClass RcDouble)
+ RcInteger -> 0
+-}
+
+
+-- The number of allocatable regs is hard coded here so we can do a fast comparision
+-- in trivColorable. It's ok if these numbers are _less_ than the actual number of
+-- free regs, but they can't be more or the register conflict graph won't color.
+--
+-- There is an allocatableRegsInClass :: RegClass -> Int, but doing the unboxing
+-- is too slow for us here.
+--
+-- Compare MachRegs.freeRegs and MachRegs.h to get these numbers.
+--
+#if i386_TARGET_ARCH
+#define ALLOCATABLE_REGS_INTEGER (_ILIT(3))
+#define ALLOCATABLE_REGS_DOUBLE (_ILIT(6))
+
+#elif x86_64_TARGET_ARCH
+#define ALLOCATABLE_REGS_INTEGER (_ILIT(5))
+#define ALLOCATABLE_REGS_DOUBLE (_ILIT(2))
+
+#elif powerpc_TARGET_ARCH
+#define ALLOCATABLE_REGS_INTEGER (_ILIT(16))
+#define ALLOCATABLE_REGS_DOUBLE (_ILIT(26))
+
+#else
+#error ToDo: define ALLOCATABLE_REGS_INTEGER and ALLOCATABLE_REGS_DOUBLE
+#endif
+
+{-# INLINE regClass #-}
+trivColorable :: RegClass -> UniqSet Reg -> UniqSet Reg -> Bool
+trivColorable classN conflicts exclusions
+ = {-# SCC "trivColorable" #-}
+ let
+ {-# INLINE isSqueesed #-}
+ isSqueesed cI cF ufm
+ = case ufm of
+ NodeUFM _ _ left right
+ -> case isSqueesed cI cF right of
+ (# s, cI', cF' #)
+ -> case s of
+ False -> isSqueesed cI' cF' left
+ True -> (# True, cI', cF' #)
+
+ LeafUFM _ reg
+ -> case regClass reg of
+ RcInteger
+ -> case cI +# _ILIT(1) of
+ cI' -> (# cI' >=# ALLOCATABLE_REGS_INTEGER, cI', cF #)
+
+ RcDouble
+ -> case cF +# _ILIT(1) of
+ cF' -> (# cF' >=# ALLOCATABLE_REGS_DOUBLE, cI, cF' #)
+
+ EmptyUFM
+ -> (# False, cI, cF #)
+
+ in case isSqueesed (_ILIT(0)) (_ILIT(0)) conflicts of
+ (# False, cI', cF' #)
+ -> case isSqueesed cI' cF' exclusions of
+ (# s, _, _ #) -> not s
+
+ (# True, _, _ #)
+ -> False
+
+
+
-- -----------------------------------------------------------------------------
-- Machine-specific register stuff
fake4 = RealReg 12
fake5 = RealReg 13
+
-- On x86, we might want to have an 8-bit RegClass, which would
-- contain just regs 1-4 (the others don't have 8-bit versions).
-- However, we can get away without this at the moment because the
then regNames !! n
else "%unknown_x86_real_reg_" ++ show n
+
#endif
{-
= let isFree i = isFastTrue (freeReg i)
in filter isFree allMachRegNos
+
+-- | The number of regs in each class.
+-- We go via top level CAFs to ensure that we're not recomputing
+-- the length of these lists each time the fn is called.
+allocatableRegsInClass :: RegClass -> Int
+allocatableRegsInClass cls
+ = case cls of
+ RcInteger -> allocatableRegsInteger
+ RcDouble -> allocatableRegsDouble
+
+allocatableRegsInteger
+ = length $ filter (\r -> regClass r == RcInteger)
+ $ map RealReg allocatableRegs
+
+allocatableRegsDouble
+ = length $ filter (\r -> regClass r == RcDouble)
+ $ map RealReg allocatableRegs
+
+
-- these are the regs which we cannot assume stay alive over a
-- C call.
callClobberedRegs :: [Reg]
globalRegMaybe BaseReg = Just (RealReg REG_Base)
#endif
#ifdef REG_R1
-globalRegMaybe (VanillaReg 1) = Just (RealReg REG_R1)
+globalRegMaybe (VanillaReg 1 _) = Just (RealReg REG_R1)
#endif
#ifdef REG_R2
-globalRegMaybe (VanillaReg 2) = Just (RealReg REG_R2)
+globalRegMaybe (VanillaReg 2 _) = Just (RealReg REG_R2)
#endif
#ifdef REG_R3
-globalRegMaybe (VanillaReg 3) = Just (RealReg REG_R3)
+globalRegMaybe (VanillaReg 3 _) = Just (RealReg REG_R3)
#endif
#ifdef REG_R4
-globalRegMaybe (VanillaReg 4) = Just (RealReg REG_R4)
+globalRegMaybe (VanillaReg 4 _) = Just (RealReg REG_R4)
#endif
#ifdef REG_R5
-globalRegMaybe (VanillaReg 5) = Just (RealReg REG_R5)
+globalRegMaybe (VanillaReg 5 _) = Just (RealReg REG_R5)
#endif
#ifdef REG_R6
-globalRegMaybe (VanillaReg 6) = Just (RealReg REG_R6)
+globalRegMaybe (VanillaReg 6 _) = Just (RealReg REG_R6)
#endif
#ifdef REG_R7
-globalRegMaybe (VanillaReg 7) = Just (RealReg REG_R7)
+globalRegMaybe (VanillaReg 7 _) = Just (RealReg REG_R7)
#endif
#ifdef REG_R8
-globalRegMaybe (VanillaReg 8) = Just (RealReg REG_R8)
+globalRegMaybe (VanillaReg 8 _) = Just (RealReg REG_R8)
#endif
#ifdef REG_R9
-globalRegMaybe (VanillaReg 9) = Just (RealReg REG_R9)
+globalRegMaybe (VanillaReg 9 _) = Just (RealReg REG_R9)
#endif
#ifdef REG_R10
-globalRegMaybe (VanillaReg 10) = Just (RealReg REG_R10)
+globalRegMaybe (VanillaReg 10 _) = Just (RealReg REG_R10)
#endif
#ifdef REG_F1
globalRegMaybe (FloatReg 1) = Just (RealReg REG_F1)
projects / ghc-hetmet.git / blobdiff