Fix truncate on amd64 NCG; fixes arith005.

[ghc-hetmet.git] / compiler / nativeGen / MachCodeGen.hs

diff --git a/compiler/nativeGen/MachCodeGen.hs b/compiler/nativeGen/MachCodeGen.hs
index 17ee624..3ff958d 100644 (file)
--- a/compiler/nativeGen/MachCodeGen.hs
+++ b/compiler/nativeGen/MachCodeGen.hs
@@ -49,6 +49,7 @@ import Control.Monad  ( mapAndUnzipM )
 import Data.Maybe      ( fromJust )
 import Data.Bits
 import Data.Word
+import Data.Int
 
 -- -----------------------------------------------------------------------------
 -- Top-level of the instruction selector
@@ -1828,15 +1829,18 @@ getAmode (CmmMachOp (MO_Add rep) [a@(CmmMachOp (MO_Shl _) _),
 getAmode (CmmMachOp (MO_Add rep) [x, CmmMachOp (MO_Shl _) 
                                        [y, CmmLit (CmmInt shift _)]])
   | shift == 0 || shift == 1 || shift == 2 || shift == 3
-  = do (x_reg, x_code) <- getNonClobberedReg x
-       -- x must be in a temp, because it has to stay live over y_code
-       -- we could compre x_reg and y_reg and do something better here...
-       (y_reg, y_code) <- getSomeReg y
-       let
-          code = x_code `appOL` y_code
-           base = case shift of 0 -> 1; 1 -> 2; 2 -> 4; 3 -> 8
-       return (Amode (AddrBaseIndex (EABaseReg x_reg) (EAIndex y_reg base) (ImmInt 0))
-               code)
+  = x86_complex_amode x y shift 0
+
+getAmode (CmmMachOp (MO_Add rep) 
+                [x, CmmMachOp (MO_Add _)
+                        [CmmMachOp (MO_Shl _) [y, CmmLit (CmmInt shift _)],
+                         CmmLit (CmmInt offset _)]])
+  | shift == 0 || shift == 1 || shift == 2 || shift == 3
+  && not (is64BitInteger offset)
+  = x86_complex_amode x y shift offset
+
+getAmode (CmmMachOp (MO_Add rep) [x,y])
+  = x86_complex_amode x y 0 0
 
 getAmode (CmmLit lit) | not (is64BitLit lit)
   = return (Amode (ImmAddr (litToImm lit) 0) nilOL)
@@ -1845,6 +1849,19 @@ getAmode expr = do
   (reg,code) <- getSomeReg expr
   return (Amode (AddrBaseIndex (EABaseReg reg) EAIndexNone (ImmInt 0)) code)
 
+
+x86_complex_amode :: CmmExpr -> CmmExpr -> Integer -> Integer -> NatM Amode
+x86_complex_amode base index shift offset
+  = do (x_reg, x_code) <- getNonClobberedReg base
+       -- x must be in a temp, because it has to stay live over y_code
+       -- we could compre x_reg and y_reg and do something better here...
+       (y_reg, y_code) <- getSomeReg index
+       let
+          code = x_code `appOL` y_code
+           base = case shift of 0 -> 1; 1 -> 2; 2 -> 4; 3 -> 8
+       return (Amode (AddrBaseIndex (EABaseReg x_reg) (EAIndex y_reg base) (ImmInt (fromIntegral offset)))
+               code)
+
 #endif /* i386_TARGET_ARCH || x86_64_TARGET_ARCH */
 
 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
@@ -2037,7 +2054,12 @@ is64BitLit x = False
 #endif
 
 is64BitInteger :: Integer -> Bool
-is64BitInteger i = i > 0x7fffffff || i < -0x80000000
+is64BitInteger i = i64 > 0x7fffffff || i64 < -0x80000000
+  where i64 = fromIntegral i :: Int64
+  -- 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.
 
 -- -----------------------------------------------------------------------------
 --  The 'CondCode' type:  Condition codes passed up the tree.
@@ -2359,6 +2381,25 @@ assignIntCode pk dst src
 #if i386_TARGET_ARCH || x86_64_TARGET_ARCH
 
 -- integer assignment to memory
+
+-- specific case of adding/subtracting an integer to a particular address.
+-- ToDo: catch other cases where we can use an operation directly on a memory 
+-- address.
+assignMem_IntCode pk addr (CmmMachOp op [CmmLoad addr2 _,
+                                                 CmmLit (CmmInt i _)])
+   | addr == addr2, pk /= I64 || not (is64BitInteger i),
+     Just instr <- check op
+   = do Amode amode code_addr <- getAmode addr
+        let code = code_addr `snocOL`
+                   instr pk (OpImm (ImmInt (fromIntegral i))) (OpAddr amode)
+        return code
+   where
+        check (MO_Add _) = Just ADD
+        check (MO_Sub _) = Just SUB
+        check _ = Nothing
+        -- ToDo: more?
+
+-- general case
 assignMem_IntCode pk addr src = do
     Amode addr code_addr <- getAmode addr
     (code_src, op_src)   <- get_op_RI src
@@ -4522,6 +4563,8 @@ remainderCode rep div x y = do
 -- -----------------------------------------------------------------------------
 --  Coercing to/from integer/floating-point...
 
+-- When going to integer, we truncate (round towards 0).
+
 -- @coerce(Int2FP|FP2Int)@ are more complicated integer/float
 -- conversions.  We have to store temporaries in memory to move
 -- between the integer and the floating point register sets.
@@ -4607,7 +4650,7 @@ coerceFP2Int from to x = do
 coerceFP2Int from to x = do
   (x_op, x_code) <- getOperand x  -- ToDo: could be a safe operand
   let
-        opc  = case from of F32 -> CVTSS2SI; F64 -> CVTSD2SI
+        opc  = case from of F32 -> CVTTSS2SIQ; F64 -> CVTTSD2SIQ
         code dst = x_code `snocOL` opc x_op dst
   -- in
   return (Any to code) -- works even if the destination rep is <I32