[project @ 2004-08-13 13:04:50 by simonmar]

[ghc-hetmet.git] / ghc / compiler / codeGen / CodeGen.lhs

diff --git a/ghc/compiler/codeGen/CodeGen.lhs b/ghc/compiler/codeGen/CodeGen.lhs
index 95926aa..d7f2f70 100644 (file)
--- a/ghc/compiler/codeGen/CodeGen.lhs
+++ b/ghc/compiler/codeGen/CodeGen.lhs
@@ -19,122 +19,248 @@ module CodeGen ( codeGen ) where
 
 #include "HsVersions.h"
 
 
 #include "HsVersions.h"
 

-import StgSyn
-import CgMonad
-import AbsCSyn
-import CLabel          ( CLabel, mkSRTLabel, mkClosureLabel )
+import DriverState     ( v_Build_tag, v_MainModIs )

 
 

-import PprAbsC         ( dumpRealC )
-import AbsCUtils       ( mkAbstractCs, mkAbsCStmts, flattenAbsC )
-import CgBindery       ( CgIdInfo, addBindC, addBindsC )
+-- Kludge (??) so that CgExpr is reached via at least one non-SOURCE
+-- import.  Before, that wasn't the case, and CM therefore didn't 
+-- bother to compile it.
+import CgExpr           ( {-NOTHING!-} )       -- DO NOT DELETE THIS IMPORT
+import CgProf
+import CgMonad
+import CgBindery       ( CgIdInfo, addBindC, addBindsC, getCgIdInfo,
+                         cgIdInfoId )

 import CgClosure       ( cgTopRhsClosure )
 import CgClosure       ( cgTopRhsClosure )

-import CgCon           ( cgTopRhsCon )
-import CgConTbls       ( genStaticConBits )
-import ClosureInfo     ( mkClosureLFInfo )
-import CmdLineOpts     ( opt_SccProfilingOn, opt_EnsureSplittableC, 
-                         opt_D_dump_absC,    opt_SccGroup
-                       )
-import CostCentre       ( CostCentre, CostCentreStack )
-import FiniteMap       ( FiniteMap )
-import Id               ( Id, idName )
-import Module           ( Module, moduleString, ModuleName, moduleNameString )
-import PrimRep         ( getPrimRepSize, PrimRep(..) )
-import Type             ( Type )
-import TyCon            ( TyCon, isDataTyCon )
-import Class           ( Class, classTyCon )
-import BasicTypes      ( TopLevelFlag(..) )
-import UniqSupply      ( mkSplitUniqSupply )
-import ErrUtils                ( dumpIfSet )
-import Util
-import Panic           ( assertPanic )
+import CgCon           ( cgTopRhsCon, cgTyCon )
+import CgUtils         ( cmmRegOffW, emitRODataLits, cmmNeWord, emitRtsCall )
+
+import CLabel          ( mkSRTLabel, mkClosureLabel, moduleRegdLabel,
+                         mkPlainModuleInitLabel, mkModuleInitLabel )
+import Cmm
+import CmmUtils                ( zeroCLit, mkIntCLit, mkLblExpr )
+import PprCmm          ( pprCmms )
+import MachOp          ( wordRep, MachHint(..) )
+
+import StgSyn
+import PrelNames       ( gHC_PRIM, rOOT_MAIN, mAIN_Name, pREL_TOP_HANDLER )
+import CmdLineOpts     ( DynFlags, DynFlag(..), opt_EnsureSplittableC,
+                         opt_SccProfilingOn )
+
+import HscTypes                ( ForeignStubs(..), TypeEnv, typeEnvTyCons )
+import CostCentre       ( CollectedCCs )
+import Id               ( Id, idName, setIdName )
+import Name            ( nameSrcLoc, nameOccName, nameUnique, isInternalName, mkExternalName )
+import OccName         ( mkLocalOcc )
+import TyCon            ( isDataTyCon )
+import Module          ( Module, mkModuleName )
+import ErrUtils                ( dumpIfSet_dyn, showPass )
+import Panic           ( assertPanic, trace )
+import qualified Module ( moduleName )
+
+#ifdef DEBUG
+import Outputable
+#endif
+
+import DATA_IOREF      ( readIORef )

 \end{code}
 
 \begin{code}
 \end{code}
 
 \begin{code}

+codeGen :: DynFlags
+       -> Module
+       -> TypeEnv
+       -> ForeignStubs
+       -> [Module]             -- directly-imported modules
+       -> CollectedCCs         -- (Local/global) cost-centres needing declaring/registering.
+       -> [(StgBinding,[(Id,[Id])])]   -- Bindings to convert, with SRTs
+       -> IO [Cmm]             -- Output
+
+codeGen dflags this_mod type_env foreign_stubs imported_mods 
+       cost_centre_info stg_binds
+  = do 
+  { showPass dflags "CodeGen"
+  ; way <- readIORef v_Build_tag
+  ; mb_main_mod <- readIORef v_MainModIs
+
+  ; let     tycons     = typeEnvTyCons type_env
+           data_tycons = filter isDataTyCon tycons
+
+-- Why?
+--   ; mapM_ (\x -> seq x (return ())) data_tycons
+
+  ; code_stuff <- initC this_mod $ do 
+                       { cmm_binds  <- mapM (getCmm . cgTopBinding) stg_binds
+                       ; cmm_tycons <- mapM cgTyCon data_tycons
+                       ; cmm_init   <- getCmm (mkModuleInit way cost_centre_info 
+                                                    this_mod mb_main_mod
+                                                    foreign_stubs imported_mods)
+                       ; return (cmm_binds ++ concat cmm_tycons ++ [cmm_init]) 
+                       }
+               -- Put datatype_stuff after code_stuff, because the
+               -- datatype closure table (for enumeration types) to
+               -- (say) PrelBase_True_closure, which is defined in
+               -- code_stuff
+
+  ; dumpIfSet_dyn dflags Opt_D_dump_cmm "Cmm" (pprCmms code_stuff)
+
+  ; return code_stuff }
+\end{code}
+
+%************************************************************************
+%*                                                                     *
+\subsection[codegen-init]{Module initialisation code}
+%*                                                                     *
+%************************************************************************
+
+/* -----------------------------------------------------------------------------
+   Module initialisation
+
+   The module initialisation code looks like this, roughly:
+
+       FN(__stginit_Foo) {
+         JMP_(__stginit_Foo_1_p)
+       }

 
 

+       FN(__stginit_Foo_1_p) {
+       ...
+       }

 
 

-codeGen :: Module              -- Module name
-       -> [ModuleName]         -- Import names
-       -> ([CostCentre],       -- Local cost-centres needing declaring/registering
-           [CostCentre],       -- "extern" cost-centres needing declaring
-           [CostCentreStack])  -- Pre-defined "singleton" cost centre stacks
-       -> [TyCon] -> [Class]   -- Local tycons and classes
-       -> [(StgBinding,[Id])]  -- Bindings to convert, with SRTs
-       -> IO AbstractC         -- Output
-
-codeGen mod_name imported_modules cost_centre_info
-       tycons classes stg_binds
-  = mkSplitUniqSupply 'f'      >>= \ fl_uniqs  -> -- absC flattener
-    let
-       datatype_stuff    = genStaticConBits cinfo data_tycons
-       code_stuff        = initC cinfo (cgTopBindings maybe_split stg_binds)
-       cost_centre_stuff = mkCostCentreStuff mod_name imported_modules cost_centre_info
-
-       abstractC = mkAbstractCs [ cost_centre_stuff, 
-                                  datatype_stuff,
-                                  code_stuff ]
-
-       flat_abstractC = flattenAbsC fl_uniqs abstractC
-    in
-    dumpIfSet opt_D_dump_absC "Abstract C" (dumpRealC abstractC)       >>
-    return flat_abstractC
+   We have one version of the init code with a module version and the
+   'way' attached to it.  The version number helps to catch cases
+   where modules are not compiled in dependency order before being
+   linked: if a module has been compiled since any modules which depend on
+   it, then the latter modules will refer to a different version in their
+   init blocks and a link error will ensue.

 
 

+   The 'way' suffix helps to catch cases where modules compiled in different
+   ways are linked together (eg. profiled and non-profiled).
+
+   We provide a plain, unadorned, version of the module init code
+   which just jumps to the version with the label and way attached.  The
+   reason for this is that when using foreign exports, the caller of
+   startupHaskell() must supply the name of the init function for the "top"
+   module in the program, and we don't want to require that this name
+   has the version and way info appended to it.
+   -------------------------------------------------------------------------- */
+
+We initialise the module tree by keeping a work-stack, 
+       * pointed to by Sp
+       * that grows downward
+       * Sp points to the last occupied slot
+
+
+\begin{code}
+mkModuleInit 
+       :: String               -- the "way"
+       -> CollectedCCs         -- cost centre info
+       -> Module
+       -> Maybe String         -- Just m ==> we have flag: -main-is Foo.baz 
+       -> ForeignStubs
+       -> [Module]
+       -> Code
+mkModuleInit way cost_centre_info this_mod mb_main_mod foreign_stubs imported_mods
+  = do {       
+
+       -- Allocate the static boolean that records if this
+       -- module has been registered already
+       ; emitData Data [CmmDataLabel moduleRegdLabel, 
+                        CmmStaticLit zeroCLit]
+
+       ; emitSimpleProc real_init_lbl $ do
+           {   -- The return-code pops the work stack by 
+               -- incrementing Sp, and then jumpd to the popped item
+             ret_blk <- forkLabelledCode $ stmtsC
+                       [ CmmAssign spReg (cmmRegOffW spReg 1)
+                       , CmmJump (CmmLoad (cmmRegOffW spReg (-1)) wordRep) [] ]
+
+           ; init_blk <- forkLabelledCode $ do
+                           { mod_init_code; stmtC (CmmBranch ret_blk) }
+                       
+           ; stmtC (CmmCondBranch (cmmNeWord (CmmLit zeroCLit) mod_reg_val)
+                       ret_blk)
+           ; stmtC (CmmBranch init_blk)            
+           }
+
+
+           -- Make the "plain" procedure jump to the "real" init procedure
+       ; emitSimpleProc plain_init_lbl jump_to_init
+
+       -- When compiling the module in which the 'main' function lives,
+       -- (that is, Module.moduleName this_mod == main_mod_name)
+       -- we inject an extra stg_init procedure for stg_init_ZCMain, for the 
+       -- RTS to invoke.  We must consult the -main-is flag in case the
+       -- user specified a different function to Main.main
+       ; whenC (Module.moduleName this_mod == main_mod_name)
+               (emitSimpleProc plain_main_init_lbl jump_to_init)
+    }

   where
   where

-    data_tycons = filter isDataTyCon (tycons ++ map classTyCon classes)
-                       -- Generate info tables  for the data constrs arising
-                       -- from class decls as well
-
-    maybe_split = if opt_EnsureSplittableC 
-                 then CSplitMarker 
-                 else AbsCNop
-    cinfo       = MkCompInfo mod_name
+    plain_init_lbl = mkPlainModuleInitLabel this_mod
+    real_init_lbl  = mkModuleInitLabel this_mod way
+    plain_main_init_lbl = mkPlainModuleInitLabel rOOT_MAIN
+
+    jump_to_init = stmtC (CmmJump (mkLblExpr real_init_lbl) [])
+
+    mod_reg_val = CmmLoad (mkLblExpr moduleRegdLabel) wordRep
+
+    main_mod_name = case mb_main_mod of
+                       Just mod_name -> mkModuleName mod_name
+                       Nothing       -> mAIN_Name
+
+    -- Main refers to GHC.TopHandler.runIO, so make sure we call the
+    -- init function for GHC.TopHandler.
+    extra_imported_mods
+       | Module.moduleName this_mod == main_mod_name = [pREL_TOP_HANDLER]
+       | otherwise                                   = []
+
+    mod_init_code = do
+       {       -- Set mod_reg to 1 to record that we've been here
+         stmtC (CmmStore (mkLblExpr moduleRegdLabel) (CmmLit (mkIntCLit 1)))
+
+               -- Now do local stuff
+       ; registerForeignExports foreign_stubs
+       ; initCostCentres cost_centre_info
+       ; mapCs (registerModuleImport way) (imported_mods++extra_imported_mods)
+       } 
+
+
+-----------------------
+registerModuleImport :: String -> Module -> Code
+registerModuleImport way mod 
+  | mod == gHC_PRIM
+  = nopC 
+  | otherwise  -- Push the init procedure onto the work stack
+  = stmtsC [ CmmAssign spReg (cmmRegOffW spReg (-1))
+          , CmmStore (CmmReg spReg) (mkLblExpr (mkModuleInitLabel mod way)) ]
+
+-----------------------
+registerForeignExports :: ForeignStubs -> Code
+registerForeignExports NoStubs 
+  = nopC
+registerForeignExports (ForeignStubs _ _ _ fe_bndrs)
+  = mapM_ mk_export_register fe_bndrs
+  where
+       mk_export_register bndr
+         = emitRtsCall SLIT("getStablePtr") 
+               [ (CmmLit (CmmLabel (mkClosureLabel (idName bndr))), PtrHint) ]

 \end{code}
 
 \end{code}
 

-Cost-centre profiling:
-Besides the usual stuff, we must produce:

 
 

-* Declarations for the cost-centres defined in this module;
-* Code to participate in "registering" all the cost-centres
-  in the program (done at startup time when the pgm is run).

 
 

-(The local cost-centres involved in this are passed
-into the code-generator, as are the imported-modules' names.)
+Cost-centre profiling: Besides the usual stuff, we must produce
+declarations for the cost-centres defined in this module;

 
 

-\begin{code}
-mkCostCentreStuff mod_name import_names (local_CCs, extern_CCs, singleton_CCSs)
-  | not opt_SccProfilingOn = AbsCNop
-  | otherwise = mkAbstractCs (
-                   map (CCostCentreDecl True)   local_CCs ++
-                   map (CCostCentreDecl False)  extern_CCs ++
-                   map CCostCentreStackDecl     singleton_CCSs ++
-                   mkCcRegister local_CCs singleton_CCSs import_names
-               )
+(The local cost-centres involved in this are passed into the
+code-generator.)

 
 

-  where
-    mkCcRegister ccs cc_stacks import_names
-      = let
-           register_ccs     = mkAbstractCs (map mk_register ccs)
-           register_imports
-             = foldr (mkAbsCStmts . mk_import_register) AbsCNop import_names
-           register_cc_stacks = mkAbstractCs (map mk_register_ccs cc_stacks)
-       in
-       [
-           CCallProfCCMacro SLIT("START_REGISTER_CCS") 
-              [ CLitLit (_PK_ ("_reg" ++ moduleString mod_name)) AddrRep],
-           register_ccs,
-           register_cc_stacks,
-           register_imports,
-           CCallProfCCMacro SLIT("END_REGISTER_CCS") []
-       ]
-      where
-       mk_register cc
-         = CCallProfCCMacro SLIT("REGISTER_CC") [mkCCostCentre cc]
-
-       mk_register_ccs ccs
-         = CCallProfCCMacro SLIT("REGISTER_CCS") [mkCCostCentreStack ccs]
-
-       mk_import_register import_name
-         = CCallProfCCMacro SLIT("REGISTER_IMPORT") 
-             [CLitLit (_PK_ ("_reg" ++ moduleNameString import_name)) AddrRep]
+\begin{code}
+initCostCentres :: CollectedCCs -> Code
+-- Emit the declarations, and return code to register them
+initCostCentres (local_CCs, ___extern_CCs, singleton_CCSs)
+  | not opt_SccProfilingOn = nopC
+  | otherwise
+  = do { mapM_ emitCostCentreDecl       local_CCs
+       ; mapM_ emitCostCentreStackDecl  singleton_CCSs
+       ; mapM_ emitRegisterCC           local_CCs
+       ; mapM_ emitRegisterCCS          singleton_CCSs
+       }

 \end{code}
 
 %************************************************************************
 \end{code}
 
 %************************************************************************

@@ -143,7 +269,7 @@ mkCostCentreStuff mod_name import_names (local_CCs, extern_CCs, singleton_CCSs)
 %*                                                                     *
 %************************************************************************
 
 %*                                                                     *
 %************************************************************************
 

-@cgTopBindings@ is only used for top-level bindings, since they need
+@cgTopBinding@ is only used for top-level bindings, since they need

 to be allocated statically (not in the heap) and need to be labelled.
 No unboxed bindings can happen at top level.
 
 to be allocated statically (not in the heap) and need to be labelled.
 No unboxed bindings can happen at top level.
 

@@ -154,55 +280,83 @@ style, with the increasing static environment being plumbed as a state
 variable.
 
 \begin{code}
 variable.
 
 \begin{code}

-cgTopBindings :: AbstractC -> [(StgBinding,[Id])] -> Code
-
-cgTopBindings split bindings = mapCs (cgTopBinding split) bindings
-
-cgTopBinding :: AbstractC -> (StgBinding,[Id]) -> Code
-
-cgTopBinding split ((StgNonRec name rhs), srt)
-  = absC split                 `thenC`
-    absC (mkSRT srt_label srt)         `thenC`
-    setSRTLabel srt_label (
-    cgTopRhs name rhs          `thenFC` \ (name, info) ->
-    addBindC name info
-    )
-  where
-    srt_label = mkSRTLabel (idName name)
-
-cgTopBinding split ((StgRec pairs@((name,rhs):_)), srt)
-  = absC split                 `thenC`
-    absC (mkSRT srt_label srt)         `thenC`
-    setSRTLabel srt_label (
-    fixC (\ new_binds -> addBindsC new_binds   `thenC`
-                        mapFCs ( \ (b,e) -> cgTopRhs b e ) pairs
-    )                    `thenFC` \ new_binds ->
-    addBindsC new_binds
-    )
+cgTopBinding :: (StgBinding,[(Id,[Id])]) -> Code
+cgTopBinding (StgNonRec id rhs, srts)
+  = do { id' <- maybeExternaliseId id
+       ; mapM_ (mkSRT [id']) srts
+       ; (id,info) <- cgTopRhs id' rhs
+       ; addBindC id info      -- Add the *un-externalised* Id to the envt,
+                               -- so we find it when we look up occurrences
+       }
+
+cgTopBinding (StgRec pairs, srts)
+  = do { let (bndrs, rhss) = unzip pairs
+       ; bndrs' <- mapFCs maybeExternaliseId bndrs
+       ; let pairs' = zip bndrs' rhss
+       ; mapM_ (mkSRT bndrs')  srts
+       ; new_binds <- fixC (\ new_binds -> do 
+               { addBindsC new_binds
+               ; mapFCs ( \ (b,e) -> cgTopRhs b e ) pairs' })
+       ; nopC }
+
+mkSRT :: [Id] -> (Id,[Id]) -> Code
+mkSRT these (id,[])  = nopC
+mkSRT these (id,ids)
+  = do { ids <- mapFCs remap ids
+       ; id  <- remap id
+       ; emitRODataLits (mkSRTLabel (idName id)) 
+                      (map (CmmLabel . mkClosureLabel . idName) ids)
+       }

   where
   where

-    srt_label = mkSRTLabel (idName name)
-
-mkSRT :: CLabel -> [Id] -> AbstractC
-mkSRT lbl []  = AbsCNop
-mkSRT lbl ids = CSRT lbl (map (mkClosureLabel . idName) ids)
+       -- Sigh, better map all the ids against the environment in 
+       -- case they've been externalised (see maybeExternaliseId below).
+    remap id = case filter (==id) these of
+               (id':_) -> returnFC id'
+               [] -> do { info <- getCgIdInfo id; return (cgIdInfoId info) }

 
 -- Urgh!  I tried moving the forkStatics call from the rhss of cgTopRhs
 -- to enclose the listFCs in cgTopBinding, but that tickled the
 -- statics "error" call in initC.  I DON'T UNDERSTAND WHY!
 
 cgTopRhs :: Id -> StgRhs -> FCode (Id, CgIdInfo)
 
 -- Urgh!  I tried moving the forkStatics call from the rhss of cgTopRhs
 -- to enclose the listFCs in cgTopBinding, but that tickled the
 -- statics "error" call in initC.  I DON'T UNDERSTAND WHY!
 
 cgTopRhs :: Id -> StgRhs -> FCode (Id, CgIdInfo)

-       -- the Id is passed along for setting up a binding...
+       -- The Id is passed along for setting up a binding...
+       -- It's already been externalised if necessary

 
 cgTopRhs bndr (StgRhsCon cc con args)
 
 cgTopRhs bndr (StgRhsCon cc con args)

-  = forkStatics (cgTopRhsCon bndr con args (all zero_size args))
+  = forkStatics (cgTopRhsCon bndr con args)
+
+cgTopRhs bndr (StgRhsClosure cc bi fvs upd_flag srt args body)
+  = ASSERT(null fvs)    -- There should be no free variables
+    setSRTLabel (mkSRTLabel (idName bndr)) $ 
+    forkStatics (cgTopRhsClosure bndr cc bi srt upd_flag args body)
+\end{code}
+
+
+%************************************************************************
+%*                                                                     *
+\subsection{Stuff to support splitting}
+%*                                                                     *
+%************************************************************************
+
+If we're splitting the object, we need to externalise all the top-level names
+(and then make sure we only use the externalised one in any C label we use
+which refers to this name).
+
+\begin{code}
+maybeExternaliseId :: Id -> FCode Id
+maybeExternaliseId id
+  | opt_EnsureSplittableC,     -- Externalise the name for -split-objs
+    isInternalName name = do { mod <- moduleName
+                            ; returnFC (setIdName id (externalise mod)) }
+  | otherwise          = returnFC id

   where
   where

-    zero_size atom = getPrimRepSize (getArgPrimRep atom) == 0
-
-cgTopRhs bndr (StgRhsClosure cc bi srt fvs upd_flag args body)
-  = ASSERT(null fvs) -- There should be no free variables
-    getSRTLabel `thenFC` \srt_label ->
-    let lf_info = 
-         mkClosureLFInfo bndr TopLevel [{-no fvs-}] upd_flag args srt_label srt
-    in
-    forkStatics (cgTopRhsClosure bndr cc bi args body lf_info)
+    externalise mod = mkExternalName uniq mod new_occ Nothing loc
+    name    = idName id
+    uniq    = nameUnique name
+    new_occ = mkLocalOcc uniq (nameOccName name)
+    loc     = nameSrcLoc name
+       -- We want to conjure up a name that can't clash with any
+       -- existing name.  So we generate
+       --      Mod_$L243foo
+       -- where 243 is the unique.

 \end{code}
 \end{code}