Przeglądaj źródła

* first try to get cpupara working for x86_64

peter 21 lat temu
rodzic
commit
8c5b0f7d82

+ 6 - 3
compiler/arm/cgcpu.pas

@@ -92,7 +92,7 @@ unit cgcpu;
         procedure g_save_standard_registers(list : taasmoutput);override;
         procedure g_restore_standard_registers(list : taasmoutput);override;
         procedure g_save_all_registers(list : taasmoutput);override;
-        procedure g_restore_all_registers(list : taasmoutput;accused,acchiused:boolean);override;
+        procedure g_restore_all_registers(list : taasmoutput;const funcretparaloc:tparalocation);override;
 
         procedure a_jmp_cond(list : taasmoutput;cond : TOpCmp;l: tasmlabel);
         procedure fixref(list : taasmoutput;var ref : treference);
@@ -1138,7 +1138,7 @@ unit cgcpu;
       end;
 
 
-    procedure tcgarm.g_restore_all_registers(list : taasmoutput;accused,acchiused:boolean);
+    procedure tcgarm.g_restore_all_registers(list : taasmoutput;const funcretparaloc:tparalocation);
       begin
         { we support only ARM standard calling conventions so this procedure has no use on the ARM }
       end;
@@ -1273,7 +1273,10 @@ begin
 end.
 {
   $Log$
-  Revision 1.43  2004-01-29 17:09:32  florian
+  Revision 1.44  2004-02-04 22:01:13  peter
+    * first try to get cpupara working for x86_64
+
+  Revision 1.43  2004/01/29 17:09:32  florian
     * handling of floating point references fixed
 
   Revision 1.42  2004/01/28 15:36:47  florian

+ 7 - 6
compiler/cgobj.pas

@@ -440,7 +440,7 @@ unit cgobj;
           }
           procedure g_restore_standard_registers(list:Taasmoutput);virtual;abstract;
           procedure g_save_all_registers(list : taasmoutput);virtual;abstract;
-          procedure g_restore_all_registers(list : taasmoutput;accused,acchiused:boolean);virtual;abstract;
+          procedure g_restore_all_registers(list : taasmoutput;const funcretparaloc:tparalocation);virtual;abstract;
        end;
 
     {# @abstract(Abstract code generator for 64 Bit operations)
@@ -854,19 +854,17 @@ implementation
           LOC_CREGISTER,
           LOC_REGISTER:
             begin
+{$ifndef cpu64bit}
               if (locpara.size in [OS_S64,OS_64]) then
                 begin
-{$ifdef cpu64bit}
-                  ungetregister(list,locpara.register64);
-{$else cpu64bit}
                   getexplicitregister(list,locpara.registerlow);
                   getexplicitregister(list,locpara.registerhigh);
                   ungetregister(list,locpara.registerlow);
                   ungetregister(list,locpara.registerhigh);
-{$endif cpu64bit}
                   cg64.a_load64_reg_ref(list,locpara.register64,ref)
                 end
               else
+{$endif cpu64bit}
                 begin
                   getexplicitregister(list,locpara.register);
                   ungetregister(list,locpara.register);
@@ -2140,7 +2138,10 @@ finalization
 end.
 {
   $Log$
-  Revision 1.154  2004-02-03 22:32:53  peter
+  Revision 1.155  2004-02-04 22:01:13  peter
+    * first try to get cpupara working for x86_64
+
+  Revision 1.154  2004/02/03 22:32:53  peter
     * renamed xNNbittype to xNNinttype
     * renamed registers32 to registersint
     * replace some s32bit,u32bit with torddef([su]inttype).def.typ

+ 14 - 1
compiler/defutil.pas

@@ -163,6 +163,9 @@ interface
     {# Returns true, if def is a double type }
     function is_double(def : tdef) : boolean;
 
+    {# Returns true, if def is an extended type }
+    function is_extended(def : tdef) : boolean;
+
     {# Returns true, if def is a 64 bit integer type }
     function is_64bitint(def : tdef) : boolean;
 
@@ -234,6 +237,13 @@ implementation
       end;
 
 
+    function is_extended(def : tdef) : boolean;
+      begin
+        result:=(def.deftype=floatdef) and
+          (tfloatdef(def).typ=s80real);
+      end;
+
+
     function range_to_basetype(low,high:TConstExprInt):tbasetype;
       begin
         { generate a unsigned range if high<0 and low>=0 }
@@ -845,7 +855,10 @@ implementation
 end.
 {
   $Log$
-  Revision 1.9  2004-02-03 22:32:53  peter
+  Revision 1.10  2004-02-04 22:01:13  peter
+    * first try to get cpupara working for x86_64
+
+  Revision 1.9  2004/02/03 22:32:53  peter
     * renamed xNNbittype to xNNinttype
     * renamed registers32 to registersint
     * replace some s32bit,u32bit with torddef([su]inttype).def.typ

+ 42 - 2
compiler/i386/cgcpu.pas

@@ -42,6 +42,8 @@ unit cgcpu;
       tcg386 = class(tcgx86)
         procedure init_register_allocators;override;
         class function reg_cgsize(const reg: tregister): tcgsize; override;
+        procedure g_save_all_registers(list : taasmoutput);override;
+        procedure g_restore_all_registers(list : taasmoutput;const funcretparaloc:tparalocation);override;
      end;
 
       tcg64f386 = class(tcg64f32)
@@ -57,7 +59,7 @@ unit cgcpu;
 
     uses
        globtype,globals,verbose,systems,cutils,
-       symdef,symsym,defutil,paramgr,
+       symdef,symsym,defutil,paramgr,procinfo,
        rgcpu,rgx86,tgobj;
 
 
@@ -111,6 +113,41 @@ unit cgcpu;
         result := opsize_2_cgsize[reg2opsize(reg)];
       end;}
 
+    procedure tcg386.g_save_all_registers(list : taasmoutput);
+      begin
+        list.concat(Taicpu.Op_none(A_PUSHA,S_L));
+        tg.GetTemp(list,POINTER_SIZE,tt_noreuse,current_procinfo.save_regs_ref);
+        a_load_reg_ref(list,OS_ADDR,OS_ADDR,NR_STACK_POINTER_REG,current_procinfo.save_regs_ref);
+      end;
+
+
+    procedure tcg386.g_restore_all_registers(list : taasmoutput;const funcretparaloc:tparalocation);
+      var
+        href : treference;
+      begin
+        a_load_ref_reg(list,OS_ADDR,OS_ADDR,current_procinfo.save_regs_ref,NR_STACK_POINTER_REG);
+        tg.UnGetTemp(list,current_procinfo.save_regs_ref);
+        if funcretparaloc.loc=LOC_REGISTER then
+          begin
+            if funcretparaloc.size in [OS_64,OS_S64] then
+              begin
+                reference_reset_base(href,NR_STACK_POINTER_REG,20);
+                a_load_reg_ref(list,OS_32,OS_32,NR_FUNCTION_RETURN64_HIGH_REG,href);
+                reference_reset_base(href,NR_STACK_POINTER_REG,28);
+                a_load_reg_ref(list,OS_32,OS_32,NR_FUNCTION_RETURN64_LOW_REG,href);
+              end
+            else
+              begin
+                reference_reset_base(href,NR_STACK_POINTER_REG,28);
+                a_load_reg_ref(list,OS_32,OS_32,NR_FUNCTION_RETURN_REG,href);
+              end;
+          end;
+        list.concat(Taicpu.Op_none(A_POPA,S_L));
+        { We add a NOP because of the 386DX CPU bugs with POPAD }
+        list.concat(taicpu.op_none(A_NOP,S_L));
+      end;
+
+
 
 { ************* 64bit operations ************ }
 
@@ -246,7 +283,10 @@ begin
 end.
 {
   $Log$
-  Revision 1.44  2004-01-14 23:39:05  florian
+  Revision 1.45  2004-02-04 22:01:13  peter
+    * first try to get cpupara working for x86_64
+
+  Revision 1.44  2004/01/14 23:39:05  florian
     * another bunch of x86-64 fixes mainly calling convention and
       assembler reader related
 

+ 9 - 12
compiler/ncgutil.pas

@@ -56,7 +56,7 @@ interface
     procedure gen_stackalloc_code(list:Taasmoutput);
     procedure gen_stackfree_code(list:Taasmoutput;usesacc,usesacchi:boolean);
     procedure gen_save_used_regs(list:TAAsmoutput);
-    procedure gen_restore_used_regs(list:TAAsmoutput;usesacc,usesacchi,usesfpu:boolean);
+    procedure gen_restore_used_regs(list:TAAsmoutput;const funcretparaloc:tparalocation);
     procedure gen_initialize_code(list:TAAsmoutput;inlined:boolean);
     procedure gen_finalize_code(list:TAAsmoutput;inlined:boolean);
     procedure gen_entry_code(list:TAAsmoutput);
@@ -1164,14 +1164,8 @@ implementation
             while assigned(hp) do
               begin
                 case tvarsym(hp.parasym).localloc.loc of
-                  LOC_REGISTER :
-                    begin
-                      gotregvarparas := true;
-                      { cg.a_load_param_reg will first allocate and then deallocate paraloc }
-                      { register (if the parameter resides in a register) and then allocate }
-                      { the regvar (which is currently not allocated)                       }
-                      cg.a_loadany_param_reg(list,hp.paraloc[calleeside],tvarsym(hp.parasym).localloc.register,nil);
-                    end;
+                  LOC_REGISTER,
+                  LOC_MMREGISTER,
                   LOC_FPUREGISTER:
                     begin
                       gotregvarparas := true;
@@ -1642,7 +1636,7 @@ implementation
       end;
 
 
-    procedure gen_restore_used_regs(list:TAAsmoutput;usesacc,usesacchi,usesfpu:boolean);
+    procedure gen_restore_used_regs(list:TAAsmoutput;const funcretparaloc:tparalocation);
       begin
         { Pure assembler routines need to save the registers themselves }
         if (po_assembler in current_procinfo.procdef.procoptions) then
@@ -1651,7 +1645,7 @@ implementation
         { for the save all registers we can simply use a pusha,popa which
           push edi,esi,ebp,esp(ignored),ebx,edx,ecx,eax }
         if (po_saveregisters in current_procinfo.procdef.procoptions) then
-          cg.g_restore_all_registers(list,usesacc,usesacchi)
+          cg.g_restore_all_registers(list,funcretparaloc)
         else
           if current_procinfo.procdef.proccalloption in savestdregs_pocalls then
             cg.g_restore_standard_registers(list);
@@ -2117,7 +2111,10 @@ implementation
 end.
 {
   $Log$
-  Revision 1.187  2004-02-03 22:32:54  peter
+  Revision 1.188  2004-02-04 22:01:13  peter
+    * first try to get cpupara working for x86_64
+
+  Revision 1.187  2004/02/03 22:32:54  peter
     * renamed xNNbittype to xNNinttype
     * renamed registers32 to registersint
     * replace some s32bit,u32bit with torddef([su]inttype).def.typ

+ 6 - 3
compiler/powerpc/cgcpu.pas

@@ -97,7 +97,7 @@ unit cgcpu;
         procedure g_save_standard_registers(list:Taasmoutput);override;
         procedure g_restore_standard_registers(list:Taasmoutput);override;
         procedure g_save_all_registers(list : taasmoutput);override;
-        procedure g_restore_all_registers(list : taasmoutput;accused,acchiused:boolean);override;
+        procedure g_restore_all_registers(list : taasmoutput;const funcretparaloc:tparalocation);override;
 
         procedure a_jmp_cond(list : taasmoutput;cond : TOpCmp;l: tasmlabel);
 
@@ -809,7 +809,7 @@ const
          {$warning FIX ME}
        end;
 
-     procedure tcgppc.g_restore_all_registers(list : taasmoutput;accused,acchiused:boolean);
+     procedure tcgppc.g_restore_all_registers(list : taasmoutput;const funcretparaloc:tparalocation);
        begin
          {$warning FIX ME}
        end;
@@ -2292,7 +2292,10 @@ begin
 end.
 {
   $Log$
-  Revision 1.157  2004-02-03 19:49:24  jonas
+  Revision 1.158  2004-02-04 22:01:13  peter
+    * first try to get cpupara working for x86_64
+
+  Revision 1.157  2004/02/03 19:49:24  jonas
     - removed mov "reg, reg" optimizations, as they are removed by the
       register allocator and may be necessary to indicate a register may not
       be reused before some point

+ 5 - 2
compiler/psub.pas

@@ -761,7 +761,7 @@ implementation
             gen_save_used_regs(templist);
             aktproccode.insertlistafter(headertai,templist);
             aktfilepos:=exitpos;
-            gen_restore_used_regs(aktproccode,usesacc,usesacchi,usesfpu);
+            gen_restore_used_regs(aktproccode,procdef.funcret_paraloc[calleeside]);
             { Add stack allocation code after header }
             aktfilepos:=entrypos;
             gen_stackalloc_code(templist);
@@ -1342,7 +1342,10 @@ implementation
 end.
 {
   $Log$
-  Revision 1.183  2004-02-03 22:32:54  peter
+  Revision 1.184  2004-02-04 22:01:13  peter
+    * first try to get cpupara working for x86_64
+
+  Revision 1.183  2004/02/03 22:32:54  peter
     * renamed xNNbittype to xNNinttype
     * renamed registers32 to registersint
     * replace some s32bit,u32bit with torddef([su]inttype).def.typ

+ 6 - 3
compiler/sparc/cgcpu.pas

@@ -79,7 +79,7 @@ interface
         procedure g_flags2reg(list:TAasmOutput;Size:TCgSize;const f:tresflags;reg:TRegister);override;
         procedure g_overflowCheck(List:TAasmOutput;const Loc:TLocation;def:TDef);override;
         procedure g_stackframe_entry(list:TAasmOutput;localsize:LongInt);override;
-        procedure g_restore_all_registers(list:TAasmOutput;accused,acchiused:boolean);override;
+        procedure g_restore_all_registers(list:TAasmOutput;const funcretparaloc:tparalocation);override;
         procedure g_restore_frame_pointer(list:TAasmOutput);override;
         procedure g_restore_standard_registers(list:taasmoutput);override;
         procedure g_return_from_proc(list:TAasmOutput;parasize:aword);override;
@@ -813,7 +813,7 @@ implementation
       end;
 
 
-    procedure TCgSparc.g_restore_all_registers(list:TaasmOutput;accused,acchiused:boolean);
+    procedure TCgSparc.g_restore_all_registers(list:TaasmOutput;const funcretparaloc:tparalocation);
       begin
         { The sparc port uses the sparc standard calling convetions so this function has no used }
       end;
@@ -1107,7 +1107,10 @@ begin
 end.
 {
   $Log$
-  Revision 1.77  2004-01-12 22:11:38  peter
+  Revision 1.78  2004-02-04 22:01:13  peter
+    * first try to get cpupara working for x86_64
+
+  Revision 1.77  2004/01/12 22:11:38  peter
     * use localalign info for alignment for locals and temps
     * sparc fpu flags branching added
     * moved powerpc copy_valye_openarray to generic

+ 4 - 33
compiler/x86/cgx86.pas

@@ -127,8 +127,6 @@ unit cgx86;
         procedure g_return_from_proc(list : taasmoutput;parasize : aword);override;
         procedure g_save_standard_registers(list:Taasmoutput);override;
         procedure g_restore_standard_registers(list:Taasmoutput);override;
-        procedure g_save_all_registers(list : taasmoutput);override;
-        procedure g_restore_all_registers(list : taasmoutput;accused,acchiused:boolean);override;
 
         procedure g_overflowcheck(list: taasmoutput; const l:tlocation;def:tdef);override;
 
@@ -1829,36 +1827,6 @@ unit cgx86;
       end;
 
 
-    procedure tcgx86.g_save_all_registers(list : taasmoutput);
-      begin
-        list.concat(Taicpu.Op_none(A_PUSHA,S_L));
-        tg.GetTemp(list,POINTER_SIZE,tt_noreuse,current_procinfo.save_regs_ref);
-        a_load_reg_ref(list,OS_ADDR,OS_ADDR,NR_ESP,current_procinfo.save_regs_ref);
-      end;
-
-
-    procedure tcgx86.g_restore_all_registers(list : taasmoutput;accused,acchiused:boolean);
-      var
-        href : treference;
-      begin
-        a_load_ref_reg(list,OS_ADDR,OS_ADDR,current_procinfo.save_regs_ref,NR_ESP);
-        tg.UnGetTemp(list,current_procinfo.save_regs_ref);
-        if acchiused then
-         begin
-           reference_reset_base(href,NR_ESP,20);
-           list.concat(Taicpu.Op_reg_ref(A_MOV,S_L,NR_EDX,href));
-         end;
-        if accused then
-         begin
-           reference_reset_base(href,NR_ESP,28);
-           list.concat(Taicpu.Op_reg_ref(A_MOV,S_L,NR_EAX,href));
-         end;
-        list.concat(Taicpu.Op_none(A_POPA,S_L));
-        { We add a NOP because of the 386DX CPU bugs with POPAD }
-        list.concat(taicpu.op_none(A_NOP,S_L));
-      end;
-
-
     { produces if necessary overflowcode }
     procedure tcgx86.g_overflowcheck(list: taasmoutput; const l:tlocation;def:tdef);
       var
@@ -1889,7 +1857,10 @@ unit cgx86;
 end.
 {
   $Log$
-  Revision 1.105  2004-02-04 19:22:27  peter
+  Revision 1.106  2004-02-04 22:01:13  peter
+    * first try to get cpupara working for x86_64
+
+  Revision 1.105  2004/02/04 19:22:27  peter
   *** empty log message ***
 
   Revision 1.104  2004/02/03 19:46:48  jonas

+ 10 - 3
compiler/x86/cpubase.pas

@@ -256,6 +256,9 @@ uses
       tparalocation = record
          size : TCGSize;
          loc  : TCGLoc;
+         { Location type of registerhigh, for x86_64 this can
+           be different from loc when pushing structures of 16 bytes }
+         lochigh : TCGLoc;
          alignment : byte;
          case TCGLoc of
             LOC_REFERENCE : (reference : tparareference);
@@ -265,9 +268,10 @@ uses
                 1 : (register,registerhigh : tregister);
                 { overlay a registerlow }
                 2 : (registerlow : tregister);
+{$ifndef cpu64bit}
                 { overlay a 64 Bit register type }
-                3 : (reg64 : tregister64);
-                4 : (register64 : tregister64);
+                3 : (register64 : tregister64);
+{$endif cpu64bit}
               );
             { it's only for better handling }
             LOC_MMXREGISTER,LOC_CMMXREGISTER : (
@@ -535,7 +539,10 @@ implementation
 end.
 {
   $Log$
-  Revision 1.38  2004-01-30 13:42:03  florian
+  Revision 1.39  2004-02-04 22:01:13  peter
+    * first try to get cpupara working for x86_64
+
+  Revision 1.38  2004/01/30 13:42:03  florian
     * fixed more alignment issues
 
   Revision 1.37  2004/01/15 14:01:32  florian

+ 18 - 1
compiler/x86_64/cgcpu.pas

@@ -38,6 +38,8 @@ unit cgcpu;
       tcgx86_64 = class(tcgx86)
         procedure init_register_allocators;override;
         class function reg_cgsize(const reg: tregister): tcgsize; override;
+        procedure g_save_all_registers(list : taasmoutput);override;
+        procedure g_restore_all_registers(list : taasmoutput;const funcretparaloc:tparalocation);override;
         procedure g_concatcopy(list : taasmoutput;const source,dest : treference;len : aword; delsource,loadref : boolean);override;
       end;
 
@@ -89,6 +91,18 @@ unit cgcpu;
       end;
 
 
+    procedure tcgx86_64.g_save_all_registers(list : taasmoutput);
+      begin
+        {$warning todo tcgx86_64.g_save_all_registers}
+      end;
+
+
+    procedure tcgx86_64.g_restore_all_registers(list : taasmoutput;const funcretparaloc:tparalocation);
+      begin
+        {$warning todo tcgx86_64.g_restore_all_registers}
+      end;
+
+
     procedure tcgx86_64.g_concatcopy(list : taasmoutput;const source,dest : treference;len : aword; delsource,loadref : boolean);
       var
          ecxpushed : boolean;
@@ -220,7 +234,10 @@ begin
 end.
 {
   $Log$
-  Revision 1.9  2004-01-14 23:39:05  florian
+  Revision 1.10  2004-02-04 22:01:13  peter
+    * first try to get cpupara working for x86_64
+
+  Revision 1.9  2004/01/14 23:39:05  florian
     * another bunch of x86-64 fixes mainly calling convention and
       assembler reader related
 

+ 206 - 96
compiler/x86_64/cpupara.pas

@@ -30,95 +30,127 @@ unit cpupara;
 
     uses
       globtype,
-      cpubase,
+      cpubase,cgbase,
       symconst,symbase,symtype,symdef,
       paramgr;
 
     type
-       { Returns the location for the nr-st 32 Bit int parameter
-         if every parameter before is an 32 Bit int parameter as well
-         and if the calling conventions for the helper routines of the
-         rtl are used.
-       }
        tx86_64paramanager = class(tparamanager)
+       private
+          procedure create_funcret_paraloc_info(p : tabstractprocdef; side: tcallercallee);
+       public
           function getintparaloc(calloption : tproccalloption; nr : longint) : tparalocation;override;
+          function get_volatile_registers_int(calloption : tproccalloption):tcpuregisterset;override;
+          function get_volatile_registers_mm(calloption : tproccalloption):tcpuregisterset;override;
+          function get_volatile_registers_fpu(calloption : tproccalloption):tcpuregisterset;override;
           function create_paraloc_info(p : tabstractprocdef; side: tcallercallee):longint;override;
        end;
 
   implementation
 
     uses
-       verbose,
-       cpuinfo,cgbase,systems,
+       cutils,verbose,
+       cpuinfo,systems,
        defutil;
 
     const
       paraintsupregs : array[0..5] of tsuperregister = (RS_RDI,RS_RSI,RS_RDX,RS_RCX,RS_R8,RS_R9);
       parammsupregs : array[0..7] of tsuperregister = (RS_XMM0,RS_XMM1,RS_XMM2,RS_XMM3,RS_XMM4,RS_XMM5,RS_XMM6,RS_XMM7);
 
-    function getparaloc(p : tdef) : tcgloc;
+    procedure getvalueparaloc(p : tdef;var loc_lo,loc_hi:tcgloc);
+      begin
+        loc_lo:=LOC_INVALID;
+        loc_hi:=LOC_INVALID;
+        case p.deftype of
+           orddef:
+             begin
+               loc_lo:=LOC_REGISTER;
+               {$warning TODO 128bit also needs lochigh}
+             end;
+           floatdef:
+             begin
+               case tfloatdef(p).typ of
+                  s80real:
+                    loc_lo:=LOC_REFERENCE;
+                  s32real,
+                  s64real,
+                  s64comp,
+                  s64currency :
+                    loc_lo:=LOC_MMREGISTER;
+                  s128real:
+                    begin
+                      loc_lo:=LOC_MMREGISTER;
+                      loc_hi:=LOC_MMREGISTER;
+                      {$warning TODO float 128bit needs SSEUP lochigh}
+                    end;
+               end;
+             end;
+           recorddef:
+             begin
+               if p.size<=16 then
+                 begin
+                   {$warning TODO location depends on the fields}
+                   loc_lo:=LOC_REFERENCE;
+                 end
+               else
+                 loc_lo:=LOC_REFERENCE;
+             end;
+           objectdef:
+             begin
+               if is_object(p) then
+                 loc_lo:=LOC_REFERENCE
+               else
+                 loc_lo:=LOC_REGISTER;
+             end;
+           arraydef:
+             loc_lo:=LOC_REFERENCE;
+           variantdef:
+             loc_lo:=LOC_REFERENCE;
+           stringdef:
+             if is_shortstring(p) or is_longstring(p) then
+               loc_lo:=LOC_REFERENCE
+             else
+               loc_lo:=LOC_REGISTER;
+           setdef:
+             if is_smallset(p) then
+               loc_lo:=LOC_REGISTER
+             else
+               loc_lo:=LOC_REFERENCE;
+           procvardef:
+             begin
+               { This is a record < 16 bytes }
+               if (po_methodpointer in tprocvardef(p).procoptions) then
+                 begin
+                   loc_lo:=LOC_REGISTER;
+                   loc_hi:=LOC_REGISTER;
+                 end
+               else
+                 loc_lo:=LOC_REGISTER;
+             end;
+           else
+             begin
+               { default for pointers,enums,etc }
+               loc_lo:=LOC_REGISTER;
+             end;
+        end;
+      end;
+
 
+    function tx86_64paramanager.get_volatile_registers_int(calloption : tproccalloption):tcpuregisterset;
       begin
-         { Later, the LOC_REFERENCE is in most cases changed into LOC_REGISTER
-           if push_addr_param for the def is true
-         }
-         // !!!!! Fix aggregate types
-         case p.deftype of
-            orddef:
-              getparaloc:=LOC_REGISTER;
-            floatdef:
-              case tfloatdef(p).typ of
-                 s80real:
-                   getparaloc:=LOC_REFERENCE;
-                 s32real,
-                 s64real,
-                 s64comp,
-                 s64currency,
-                 s128real:
-                   getparaloc:=LOC_MMREGISTER;
-              end;
-            enumdef:
-              getparaloc:=LOC_REGISTER;
-            pointerdef:
-              getparaloc:=LOC_REGISTER;
-            formaldef:
-              getparaloc:=LOC_REGISTER;
-            classrefdef:
-              getparaloc:=LOC_REGISTER;
-            recorddef:
-              getparaloc:=LOC_REFERENCE;
-            objectdef:
-              if is_object(p) then
-                getparaloc:=LOC_REFERENCE
-              else
-                getparaloc:=LOC_REGISTER;
-            stringdef:
-              if is_shortstring(p) or is_longstring(p) then
-                getparaloc:=LOC_REFERENCE
-              else
-                getparaloc:=LOC_REGISTER;
-            procvardef:
-              if (po_methodpointer in tprocvardef(p).procoptions) then
-                getparaloc:=LOC_REFERENCE
-              else
-                getparaloc:=LOC_REGISTER;
-            filedef:
-              getparaloc:=LOC_REGISTER;
-            arraydef:
-              getparaloc:=LOC_REFERENCE;
-            setdef:
-              if is_smallset(p) then
-                getparaloc:=LOC_REGISTER
-              else
-                getparaloc:=LOC_REFERENCE;
-            variantdef:
-              getparaloc:=LOC_REFERENCE;
-            { avoid problems with errornous definitions }
-            errordef:
-              getparaloc:=LOC_REGISTER;
-            else
-              internalerror(2002071001);
-         end;
+        result:=[RS_RAX,RS_RCX,RS_RDX,RS_RSI,RS_RDI,RS_R8,RS_R9,RS_R10,RS_R11];
+      end;
+
+
+    function tx86_64paramanager.get_volatile_registers_mm(calloption : tproccalloption):tcpuregisterset;
+      begin
+        result:=[RS_XMM0..RS_XMM15];
+      end;
+
+
+    function tx86_64paramanager.get_volatile_registers_fpu(calloption : tproccalloption):tcpuregisterset;
+      begin
+        result:=[RS_ST0..RS_ST7];
       end;
 
 
@@ -142,6 +174,34 @@ unit cpupara;
       end;
 
 
+    procedure tx86_64paramanager.create_funcret_paraloc_info(p : tabstractprocdef; side: tcallercallee);
+      var
+        paraloc : tparalocation;
+      begin
+        { Function return }
+        fillchar(paraloc,sizeof(tparalocation),0);
+        paraloc.size:=def_cgsize(p.rettype.def);
+        { Return in FPU register? }
+        if is_extended(p.rettype.def) then
+          begin
+            paraloc.loc:=LOC_FPUREGISTER;
+            paraloc.register:=NR_FPU_RESULT_REG;
+          end
+        else
+         { Return in register? }
+         if not ret_in_param(p.rettype.def,p.proccalloption) then
+          begin
+            paraloc.loc:=LOC_REGISTER;
+            paraloc.register:=NR_FUNCTION_RETURN_REG;
+          end
+        else
+          begin
+            paraloc.loc:=LOC_REFERENCE;
+          end;
+        p.funcret_paraloc[side]:=paraloc;
+      end;
+
+
     function tx86_64paramanager.create_paraloc_info(p : tabstractprocdef; side: tcallercallee):longint;
       var
         hp : tparaitem;
@@ -163,61 +223,108 @@ unit cpupara;
           begin
             pushaddr:=push_addr_param(hp.paratyp,hp.paratype.def,p.proccalloption);
             if pushaddr then
-              paraloc.size:=OS_ADDR
+              begin
+                paraloc.size:=OS_ADDR;
+                paraloc.loc:=LOC_REGISTER;
+                paraloc.lochigh:=LOC_INVALID;
+              end
             else
-              paraloc.size:=def_cgsize(hp.paratype.def);
+              begin
+                paraloc.size:=def_cgsize(hp.paratype.def);
+                getvalueparaloc(hp.paratype.def,paraloc.loc,paraloc.lochigh);
+              end;
             paraloc.alignment:=paraalign;
-            if (intparareg<=high(paraintsupregs)) and
-               not(
-                   ((hp.paratype.def.deftype in [floatdef,recorddef,arraydef]) and
-                    (not pushaddr))
-                  ) then
+            { Location low }
+            if (paraloc.loc=LOC_REGISTER) and
+               (intparareg<=high(paraintsupregs)) then
               begin
-                paraloc.loc:=LOC_REGISTER;
-                if paraloc.size=OS_NO then
+                if (paraloc.size=OS_NO) or (paraloc.lochigh<>LOC_INVALID) then
                   subreg:=R_SUBWHOLE
                 else
                   subreg:=cgsize2subreg(paraloc.size);
-                paraloc.alignment:=paraalign;
                 paraloc.register:=newreg(R_INTREGISTER,paraintsupregs[intparareg],subreg);
                 inc(intparareg);
               end
-            else if (mmparareg<=high(parammsupregs)) then
+            else
+             if (paraloc.loc=LOC_MMREGISTER) and
+                (mmparareg<=high(parammsupregs)) then
               begin
+                paraloc.register:=newreg(R_MMREGISTER,parammsupregs[mmparareg],R_SUBNONE);
+                inc(mmparareg);
               end
             else
               begin
                 paraloc.loc:=LOC_REFERENCE;
+                paraloc.lochigh:=LOC_INVALID;
                 if side=callerside then
                   paraloc.reference.index:=NR_STACK_POINTER_REG
                 else
                   paraloc.reference.index:=NR_FRAME_POINTER_REG;
                 l:=push_size(hp.paratyp,hp.paratype.def,p.proccalloption);
-                // varalign:=size_2_align(l);
+                varalign:=size_2_align(l);
                 paraloc.reference.offset:=parasize;
-                // varalign:=used_align(varalign,paraalign,paraalign);
-                // parasize:=align(parasize+l,varalign);
+                varalign:=used_align(varalign,paraalign,paraalign);
+                parasize:=align(parasize+l,varalign);
+              end;
+            { Location High if required }
+            if (paraloc.lochigh<>LOC_INVALID) then
+              begin
+                if (paraloc.lochigh=LOC_REGISTER) and
+                   (intparareg<=high(paraintsupregs)) then
+                  begin
+                    paraloc.registerhigh:=newreg(R_INTREGISTER,paraintsupregs[intparareg],R_SUBWHOLE);
+                    inc(intparareg);
+                  end
+                else
+                 if (paraloc.lochigh=LOC_MMREGISTER) and
+                    (mmparareg<=high(parammsupregs)) then
+                  begin
+                    paraloc.registerhigh:=newreg(R_MMREGISTER,parammsupregs[mmparareg],R_SUBNONE);
+                    inc(mmparareg);
+                  end
+                else
+                  begin
+                    { Release when location low has already registers
+                      assigned }
+                    if paraloc.loc=LOC_REGISTER then
+                      dec(intparareg);
+                    if paraloc.loc=LOC_MMREGISTER then
+                      dec(mmparareg);
+                    { Overwrite with LOC_REFERENCE }
+                    paraloc.loc:=LOC_REFERENCE;
+                    paraloc.lochigh:=LOC_INVALID;
+                    fillchar(paraloc.reference,sizeof(paraloc.reference),0);
+                    if side=callerside then
+                      paraloc.reference.index:=NR_STACK_POINTER_REG
+                    else
+                      paraloc.reference.index:=NR_FRAME_POINTER_REG;
+                    l:=push_size(hp.paratyp,hp.paratype.def,p.proccalloption);
+                    varalign:=size_2_align(l);
+                    paraloc.reference.offset:=parasize;
+                    varalign:=used_align(varalign,paraalign,paraalign);
+                    parasize:=align(parasize+l,varalign);
+                  end;
               end;
             hp.paraloc[side]:=paraloc;
             hp:=tparaitem(hp.next);
           end;
-        { Register parameters are assigned from left-to-right, adapt offset
-          for calleeside to be reversed }
-        hp:=tparaitem(p.para.first);
-        while assigned(hp) do
+        { Register parameters are assigned from left-to-right, but the
+          offsets on the stack are right-to-left. There is no need
+          to reverse the offset, only adapt the calleeside with the
+          start offset of the first param on the stack }
+        if side=calleeside then
           begin
-            if (hp.paraloc[side].loc=LOC_REFERENCE) then
+            hp:=tparaitem(p.para.first);
+            while assigned(hp) do
               begin
-                l:=push_size(hp.paratyp,hp.paratype.def,p.proccalloption);
-                // varalign:=used_align(size_2_align(l),paraalign,paraalign);
-                // l:=align(l,varalign);
-                hp.paraloc[side].reference.offset:=parasize-hp.paraloc[side].reference.offset-l;
-                if side=calleeside then
+                if (hp.paraloc[side].loc=LOC_REFERENCE) then
                   inc(hp.paraloc[side].reference.offset,target_info.first_parm_offset);
+                hp:=tparaitem(hp.next);
               end;
-            hp:=tparaitem(hp.next);
           end;
-        { We need to return the size allocated }
+        { Create Function result paraloc }
+        create_funcret_paraloc_info(p,side);
+        { We need to return the size allocated on the stack }
         result:=parasize;
       end;
 
@@ -227,7 +334,10 @@ begin
 end.
 {
   $Log$
-  Revision 1.6  2004-01-14 23:39:05  florian
+  Revision 1.7  2004-02-04 22:01:13  peter
+    * first try to get cpupara working for x86_64
+
+  Revision 1.6  2004/01/14 23:39:05  florian
     * another bunch of x86-64 fixes mainly calling convention and
       assembler reader related
 

+ 18 - 3
compiler/x86_64/cpupi.pas

@@ -26,24 +26,39 @@ unit cpupi;
 
 {$i fpcdefs.inc}
 
-  interface
+interface
 
     uses
        psub,procinfo;
 
     type
        tx86_64procinfo = class(tcgprocinfo)
+         function calc_stackframe_size:longint;override;
        end;
 
 
-  implementation
+implementation
+
+    uses
+      cutils,
+      tgobj;
+
+
+    function tx86_64procinfo.calc_stackframe_size:longint;
+      begin
+        { RSP should be aligned on 16 bytes }
+        result:=Align(tg.direction*tg.lasttemp,16);
+      end;
 
 begin
    cprocinfo:=tx86_64procinfo;
 end.
 {
   $Log$
-  Revision 1.2  2003-12-24 00:33:10  florian
+  Revision 1.3  2004-02-04 22:01:13  peter
+    * first try to get cpupara working for x86_64
+
+  Revision 1.2  2003/12/24 00:33:10  florian
     * x86-64 compilation fixed
 
   Revision 1.1  2003/01/05 13:36:54  florian