Split Reg into vreg/hreg and add register pairs

[ghc-hetmet.git] / compiler / main / DynFlags.hs

diff --git a/compiler/main/DynFlags.hs b/compiler/main/DynFlags.hs
index 35949f7..3b027d8 100644 (file)
--- a/compiler/main/DynFlags.hs
+++ b/compiler/main/DynFlags.hs
@@ -64,12 +64,15 @@ module DynFlags (
 
 #include "HsVersions.h"
 
+#ifndef OMIT_NATIVE_CODEGEN
+import Platform
+#endif
 import Module
 import PackageConfig
 import PrelNames        ( mAIN, main_RDR_Unqual )
 import RdrName          ( RdrName, mkRdrUnqual )
 import OccName          ( mkVarOccFS )
-#ifdef i386_TARGET_ARCH
+#if defined(i386_TARGET_ARCH) || (!defined(mingw32_TARGET_OS) && !defined(darwin_TARGET_OS))
 import StaticFlags      ( opt_Static )
 #endif
 import StaticFlags      ( opt_PIC, WayName(..), v_Ways, v_Build_tag,
@@ -117,6 +120,7 @@ data DynFlag
    | Opt_D_dump_asm_regalloc_stages
    | Opt_D_dump_asm_conflicts
    | Opt_D_dump_asm_stats
+   | Opt_D_dump_asm_expanded
    | Opt_D_dump_cpranal
    | Opt_D_dump_deriv
    | Opt_D_dump_ds
@@ -220,6 +224,7 @@ data DynFlag
    | Opt_RelaxedPolyRec
    | Opt_StandaloneDeriving
    | Opt_DeriveDataTypeable
+   | Opt_DeriveFunctor
    | Opt_TypeSynonymInstances
    | Opt_FlexibleContexts
    | Opt_FlexibleInstances
@@ -338,6 +343,9 @@ data DynFlags = DynFlags {
   specConstrCount       :: Maybe Int,   -- ^ Max number of specialisations for any one function
   liberateCaseThreshold :: Maybe Int,   -- ^ Threshold for LiberateCase
 
+#ifndef OMIT_NATIVE_CODEGEN
+  targetPlatform       :: Platform,    -- ^ The platform we're compiling for. Used by the NCG.
+#endif
   stolen_x86_regs       :: Int,
   cmdlineHcIncludes     :: [String],    -- ^ @\-\#includes@
   importPaths           :: [FilePath],
@@ -583,6 +591,9 @@ defaultDynFlags =
         specConstrThreshold     = Just 200,
         specConstrCount         = Just 3,
         liberateCaseThreshold   = Just 200,
+#ifndef OMIT_NATIVE_CODEGEN
+        targetPlatform          = defaultTargetPlatform,
+#endif
         stolen_x86_regs         = 4,
         cmdlineHcIncludes       = [],
         importPaths             = ["."],
@@ -604,7 +615,7 @@ defaultDynFlags =
 
         outputFile              = Nothing,
         outputHi                = Nothing,
-        dynLibLoader            = Deployable,
+        dynLibLoader            = SystemDependent,
         dumpPrefix              = Nothing,
         dumpPrefixForce         = Nothing,
         includePaths            = [],
@@ -990,8 +1001,8 @@ pprFloatOutSwitches sw = pp_not (floatOutLambdas sw) <+> text "lambdas" <> comma
     pp_not False = text "not"
 
 -- | Switches that specify the minimum amount of floating out
-gentleFloatOutSwitches :: FloatOutSwitches
-gentleFloatOutSwitches = FloatOutSwitches False False
+-- gentleFloatOutSwitches :: FloatOutSwitches
+-- gentleFloatOutSwitches = FloatOutSwitches False False
 
 -- | Switches that do not specify floating out of lambdas, just of constants
 constantsOnlyFloatOutSwitches :: FloatOutSwitches
@@ -1093,7 +1104,14 @@ getCoreToDo dflags
         -- so that overloaded functions have all their dictionary lambdas manifest
         CoreDoSpecialising,
 
-        runWhen full_laziness (CoreDoFloatOutwards gentleFloatOutSwitches),
+        runWhen full_laziness (CoreDoFloatOutwards constantsOnlyFloatOutSwitches),
+               -- Was: gentleFloatOutSwitches  
+               -- I have no idea why, but not floating constants to top level is
+               -- very bad in some cases. 
+               -- Notably: p_ident in spectral/rewrite
+               --          Changing from "gentle" to "constantsOnly" improved
+               --          rewrite's allocation by 19%, and made  0.0% difference
+               --          to any other nofib benchmark
 
         CoreDoFloatInwards,
 
@@ -1343,6 +1361,8 @@ dynamic_flags = [
          Supported
   , Flag "ddump-asm-stats"         (setDumpFlag Opt_D_dump_asm_stats)
          Supported
+  , Flag "ddump-asm-expanded"      (setDumpFlag Opt_D_dump_asm_expanded)
+         Supported
   , Flag "ddump-cpranal"           (setDumpFlag Opt_D_dump_cpranal)
          Supported
   , Flag "ddump-deriv"             (setDumpFlag Opt_D_dump_deriv)
@@ -1771,6 +1791,7 @@ xFlags = [
   ( "UnboxedTuples",                    Opt_UnboxedTuples, const Supported ),
   ( "StandaloneDeriving",               Opt_StandaloneDeriving, const Supported ),
   ( "DeriveDataTypeable",               Opt_DeriveDataTypeable, const Supported ),
+  ( "DeriveFunctor",                    Opt_DeriveFunctor, const Supported ),
   ( "TypeSynonymInstances",             Opt_TypeSynonymInstances, const Supported ),
   ( "FlexibleContexts",                 Opt_FlexibleContexts, const Supported ),
   ( "FlexibleInstances",                Opt_FlexibleInstances, const Supported ),
@@ -1791,6 +1812,8 @@ impliedFlags
                                                      --      be completely rigid for GADTs
 
     , (Opt_TypeFamilies,        Opt_RelaxedPolyRec)  -- Trac #2944 gives a nice example
+    , (Opt_TypeFamilies,        Opt_KindSignatures)  -- Type families use kind signatures
+                                                    -- all over the place
 
     , (Opt_ScopedTypeVariables, Opt_RelaxedPolyRec)  -- Ditto for scoped type variables; see
                                                      --      Note [Scoped tyvars] in TcBinds
@@ -1809,6 +1832,7 @@ glasgowExtsFlags = [
            , Opt_TypeSynonymInstances
            , Opt_StandaloneDeriving
            , Opt_DeriveDataTypeable
+           , Opt_DeriveFunctor
            , Opt_FlexibleContexts
            , Opt_FlexibleInstances
            , Opt_ConstrainedClassMethods
@@ -2020,12 +2044,15 @@ setOptLevel n dflags
 --                                      sometimes
 --    -fdicts-cheap                     always inline dictionaries
 --    -fmax-simplifier-iterations20     this is necessary sometimes
+--    -fsimplifier-phases=3             we use an additional simplifier phase
+--                                      for fusion
 --    -fno-spec-constr-threshold        run SpecConstr even for big loops
 --    -fno-spec-constr-count            SpecConstr as much as possible
 --    -finline-enough-args              hack to prevent excessive inlining
 --
 setDPHOpt :: DynFlags -> DynFlags
 setDPHOpt dflags = setOptLevel 2 (dflags { maxSimplIterations  = 20
+                                         , simplPhases         = 3
                                          , specConstrThreshold = Nothing
                                          , specConstrCount     = Nothing
                                          })
@@ -2234,7 +2261,13 @@ machdepCCOpts _dflags
         = ( [], ["-fomit-frame-pointer", "-G0"] )
 
 #elif x86_64_TARGET_ARCH
-        = ( [], ["-fomit-frame-pointer",
+        = (
+#if darwin_TARGET_OS
+            ["-m64"],
+#else
+            [],
+#endif
+                ["-fomit-frame-pointer",
                  "-fno-asynchronous-unwind-tables",
                         -- the unwind tables are unnecessary for HC code,
                         -- and get in the way of -split-objs.  Another option
@@ -2284,7 +2317,7 @@ picCCOpts _dflags
     | otherwise
         = []
 #else
-    | opt_PIC
+    | opt_PIC || not opt_Static
         = ["-fPIC", "-U __PIC__", "-D__PIC__"]
     | otherwise
         = []