Fix linear register allocator bug involving read/write to same registers.

[ghc-hetmet.git] / compiler / nativeGen / X86 / CodeGen.hs

diff --git a/compiler/nativeGen/X86/CodeGen.hs b/compiler/nativeGen/X86/CodeGen.hs
index 44311a4..86ecbf9 100644 (file)
--- a/compiler/nativeGen/X86/CodeGen.hs
+++ b/compiler/nativeGen/X86/CodeGen.hs
@@ -605,9 +605,7 @@ getRegister (CmmMachOp mop [x]) = do -- unary MachOps
         | sse2      -> coerceFP2FP W64 x
         | otherwise -> conversionNop FF80 x 
 
-      MO_FF_Conv W64 W32
-        | sse2      -> coerceFP2FP W32 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
@@ -1589,12 +1587,24 @@ genCCall target dest_regs args = do
             | otherwise
 #endif
             = concatOL push_codes
+       
+         -- Deallocate parameters after call for ccall;
+         -- but not for stdcall (callee does it)
+         --
+         -- We have to pop any stack padding we added
+         -- on Darwin even if we are doing stdcall, though (#5052)
+       pop_size | cconv /= StdCallConv = tot_arg_size
+                | otherwise
+#if darwin_TARGET_OS
+                 = arg_pad_size
+#else
+                 = 0
+#endif
+       
        call = callinsns `appOL`
                toOL (
-                       -- Deallocate parameters after call for ccall;
-                       -- but not for stdcall (callee does it)
-                  (if cconv == StdCallConv || tot_arg_size==0 then [] else 
-                  [ADD II32 (OpImm (ImmInt tot_arg_size)) (OpReg esp)])
+                  (if pop_size==0 then [] else 
+                  [ADD II32 (OpImm (ImmInt pop_size)) (OpReg esp)])
                   ++
                   [DELTA (delta + tot_arg_size)]
                )
@@ -2257,12 +2267,14 @@ coerceFP2Int from to x = if_sse2 coerceFP2Int_sse2 coerceFP2Int_x87
 --------------------------------------------------------------------------------
 coerceFP2FP :: Width -> CmmExpr -> NatM Register
 coerceFP2FP to x = do
+  use_sse2 <- sse2Enabled
   (x_reg, x_code) <- getSomeReg x
   let
-        opc  = case to of W32 -> CVTSD2SS; W64 -> CVTSS2SD
+        opc | use_sse2  = case to of W32 -> CVTSD2SS; W64 -> CVTSS2SD
+            | otherwise = GDTOF
         code dst = x_code `snocOL` opc x_reg dst
   -- in
-  return (Any (floatSize to) code)
+  return (Any (if use_sse2 then floatSize to else FF80) code)
 
 --------------------------------------------------------------------------------