floating-point fix for x86_64

For 32-bit float primtives like sinFloat#, we currently call the
double versions of the C library functions (sin(), cos() etc.).  It
seems more correct to call the float versions (sinf(), cosf() etc.).
This makes a difference on x86_64, I'm not entirely sure why, but this
way at least generates more consistent results and avoids extra
promotion/demotion instructions when calling these primitives.

diff --git a/ghc/compiler/cmm/PprC.hs b/ghc/compiler/cmm/PprC.hs
index d085ce0..6ce2df5 100644 (file)
--- a/ghc/compiler/cmm/PprC.hs
+++ b/ghc/compiler/cmm/PprC.hs
@@ -533,19 +533,19 @@ pprCallishMachOp_for_C mop
         MO_F64_Log  -> ptext SLIT("log")
         MO_F64_Exp  -> ptext SLIT("exp")
         MO_F64_Sqrt -> ptext SLIT("sqrt")
-        MO_F32_Pwr  -> ptext SLIT("pow")
-        MO_F32_Sin  -> ptext SLIT("sin")
-        MO_F32_Cos  -> ptext SLIT("cos")
-        MO_F32_Tan  -> ptext SLIT("tan")
-        MO_F32_Sinh -> ptext SLIT("sinh")
-        MO_F32_Cosh -> ptext SLIT("cosh")
-        MO_F32_Tanh -> ptext SLIT("tanh")
-        MO_F32_Asin -> ptext SLIT("asin")
-        MO_F32_Acos -> ptext SLIT("acos")
-        MO_F32_Atan -> ptext SLIT("atan")
-        MO_F32_Log  -> ptext SLIT("log")
-        MO_F32_Exp  -> ptext SLIT("exp")
-        MO_F32_Sqrt -> ptext SLIT("sqrt")
+        MO_F32_Pwr  -> ptext SLIT("powf")
+        MO_F32_Sin  -> ptext SLIT("sinf")
+        MO_F32_Cos  -> ptext SLIT("cosf")
+        MO_F32_Tan  -> ptext SLIT("tanf")
+        MO_F32_Sinh -> ptext SLIT("sinhf")
+        MO_F32_Cosh -> ptext SLIT("coshf")
+        MO_F32_Tanh -> ptext SLIT("tanhf")
+        MO_F32_Asin -> ptext SLIT("asinf")
+        MO_F32_Acos -> ptext SLIT("acosf")
+        MO_F32_Atan -> ptext SLIT("atanf")
+        MO_F32_Log  -> ptext SLIT("logf")
+        MO_F32_Exp  -> ptext SLIT("expf")
+        MO_F32_Sqrt -> ptext SLIT("sqrtf")
 
 -- ---------------------------------------------------------------------
 -- Useful #defines

diff --git a/ghc/compiler/nativeGen/MachCodeGen.hs b/ghc/compiler/nativeGen/MachCodeGen.hs
index 2d1e6aa..32dad13 100644 (file)
--- a/ghc/compiler/nativeGen/MachCodeGen.hs
+++ b/ghc/compiler/nativeGen/MachCodeGen.hs
@@ -3083,27 +3083,32 @@ outOfLineFloatOp mop res args vols
           code2 <- stmtToInstrs (CmmAssign res (demote (CmmReg tmp)))
           return (code1 `appOL` code2)
   where
+#if i386_TARGET_ARCH
         promote (x,hint) = (CmmMachOp (MO_S_Conv F32 F64) [x], hint)
         demote  x = CmmMachOp (MO_S_Conv F64 F32) [x]
+#else
+        promote (x,hint) = (x,hint)
+        demote  x = x
+#endif
 
        lbl = mkForeignLabel fn Nothing True
 
        fn = case mop of
-             MO_F32_Sqrt  -> FSLIT("sqrt")
-             MO_F32_Sin   -> FSLIT("sin")
-             MO_F32_Cos   -> FSLIT("cos")
-             MO_F32_Tan   -> FSLIT("tan")
-             MO_F32_Exp   -> FSLIT("exp")
-             MO_F32_Log   -> FSLIT("log")
-
-             MO_F32_Asin  -> FSLIT("asin")
-             MO_F32_Acos  -> FSLIT("acos")
-             MO_F32_Atan  -> FSLIT("atan")
-
-             MO_F32_Sinh  -> FSLIT("sinh")
-             MO_F32_Cosh  -> FSLIT("cosh")
-             MO_F32_Tanh  -> FSLIT("tanh")
-             MO_F32_Pwr   -> FSLIT("pow")
+             MO_F32_Sqrt  -> FSLIT("sqrtf")
+             MO_F32_Sin   -> FSLIT("sinf")
+             MO_F32_Cos   -> FSLIT("cosf")
+             MO_F32_Tan   -> FSLIT("tanf")
+             MO_F32_Exp   -> FSLIT("expf")
+             MO_F32_Log   -> FSLIT("logf")
+
+             MO_F32_Asin  -> FSLIT("asinf")
+             MO_F32_Acos  -> FSLIT("acosf")
+             MO_F32_Atan  -> FSLIT("atanf")
+
+             MO_F32_Sinh  -> FSLIT("sinhf")
+             MO_F32_Cosh  -> FSLIT("coshf")
+             MO_F32_Tanh  -> FSLIT("tanhf")
+             MO_F32_Pwr   -> FSLIT("powf")
 
              MO_F64_Sqrt  -> FSLIT("sqrt")
              MO_F64_Sin   -> FSLIT("sin")