Merged revision(s) 32020, 32036, 32536 from trunk (i386 PIC):

* Keep the GOT offset in a virtual register for i386 non-darwin platforms. 
  It fixes PIC code generation with GOT for i386 with enabled optimizations. Bugs #28667, #28668. 
  Prior the fix I have not been able to compile even RTL with -O2 due to not enough free registers, since EBX is reserved for GOT.

  It can be further optimized to teach register allocator to not spill the GOT register if possible.
  
........
* Fixed i386 case jump table when PIC with GOT and an external assembler is used.
........
* Fixed calling of external procs for i386 non-darwin targets when PIC is enabled. 
........

git-svn-id: branches/fixes_3_0@33439 -

compiler/aggas.pas

@@ -974,7 +974,7 @@ implementation
                        cpu_i386:
                          begin
                            AsmWrite(ait_const2str[aitconst_32bit]);
-                           AsmWrite(tai_const(hp).sym.name);
+                           AsmWrite(tai_const(hp).sym.name+'-_GLOBAL_OFFSET_TABLE_');
                          end;
                      else
                        InternalError(2014022602);

compiler/i386/cgcpu.pas

@@ -36,7 +36,6 @@ unit cgcpu;
     type
       tcg386 = class(tcgx86)
         procedure init_register_allocators;override;
-        procedure do_register_allocation(list:TAsmList;headertai:tai);override;
 
         { passing parameter using push instead of mov }
         procedure a_load_reg_cgpara(list : TAsmList;size : tcgsize;r : tregister;const cgpara : tcgpara);override;
@@ -85,29 +84,15 @@ unit cgcpu;
     procedure tcg386.init_register_allocators;
       begin
         inherited init_register_allocators;
-        if not(target_info.system in [system_i386_darwin,system_i386_iphonesim]) and
-           (cs_create_pic in current_settings.moduleswitches) then
-          rg[R_INTREGISTER]:=trgcpu.create(R_INTREGISTER,R_SUBWHOLE,[RS_EAX,RS_EDX,RS_ECX,RS_ESI,RS_EDI],first_int_imreg,[RS_EBP])
+        if (cs_useebp in current_settings.optimizerswitches) and assigned(current_procinfo) and (current_procinfo.framepointer<>NR_EBP) then
+          rg[R_INTREGISTER]:=trgcpu.create(R_INTREGISTER,R_SUBWHOLE,[RS_EAX,RS_EDX,RS_ECX,RS_EBX,RS_ESI,RS_EDI,RS_EBP],first_int_imreg,[])
         else
-          if (cs_useebp in current_settings.optimizerswitches) and assigned(current_procinfo) and (current_procinfo.framepointer<>NR_EBP) then
-            rg[R_INTREGISTER]:=trgcpu.create(R_INTREGISTER,R_SUBWHOLE,[RS_EAX,RS_EDX,RS_ECX,RS_EBX,RS_ESI,RS_EDI,RS_EBP],first_int_imreg,[])
-          else
-            rg[R_INTREGISTER]:=trgcpu.create(R_INTREGISTER,R_SUBWHOLE,[RS_EAX,RS_EDX,RS_ECX,RS_EBX,RS_ESI,RS_EDI],first_int_imreg,[RS_EBP]);
+          rg[R_INTREGISTER]:=trgcpu.create(R_INTREGISTER,R_SUBWHOLE,[RS_EAX,RS_EDX,RS_ECX,RS_EBX,RS_ESI,RS_EDI],first_int_imreg,[RS_EBP]);
         rg[R_MMXREGISTER]:=trgcpu.create(R_MMXREGISTER,R_SUBNONE,[RS_XMM0,RS_XMM1,RS_XMM2,RS_XMM3,RS_XMM4,RS_XMM5,RS_XMM6,RS_XMM7],first_mm_imreg,[]);
         rg[R_MMREGISTER]:=trgcpu.create(R_MMREGISTER,R_SUBWHOLE,[RS_XMM0,RS_XMM1,RS_XMM2,RS_XMM3,RS_XMM4,RS_XMM5,RS_XMM6,RS_XMM7],first_mm_imreg,[]);
         rgfpu:=Trgx86fpu.create;
       end;
 
-    procedure tcg386.do_register_allocation(list:TAsmList;headertai:tai);
-      begin
-        if (pi_needs_got in current_procinfo.flags) then
-          begin
-            if getsupreg(current_procinfo.got) < first_int_imreg then
-              include(rg[R_INTREGISTER].used_in_proc,getsupreg(current_procinfo.got));
-          end;
-        inherited do_register_allocation(list,headertai);
-      end;
-
 
     procedure tcg386.a_load_reg_cgpara(list : TAsmList;size : tcgsize;r : tregister;const cgpara : tcgpara);
       var
@@ -314,13 +299,6 @@ unit cgcpu;
         end;
 
       begin
-        { Release PIC register }
-        if (cs_create_pic in current_settings.moduleswitches) and
-           (tf_pic_uses_got in target_info.flags) and
-           (pi_needs_got in current_procinfo.flags) and
-           not(target_info.system in systems_darwin) then
-          list.concat(tai_regalloc.dealloc(NR_PIC_OFFSET_REG,nil));
-
         { MMX needs to call EMMS }
         if assigned(rg[R_MMXREGISTER]) and
            (rg[R_MMXREGISTER].uses_registers) then
@@ -582,26 +560,49 @@ unit cgcpu;
 
     procedure tcg386.g_maybe_got_init(list: TAsmList);
       var
-        notdarwin: boolean;
+        i: longint;
+        tmpreg: TRegister;
       begin
         { allocate PIC register }
         if (cs_create_pic in current_settings.moduleswitches) and
            (tf_pic_uses_got in target_info.flags) and
            (pi_needs_got in current_procinfo.flags) then
           begin
-            notdarwin:=not(target_info.system in [system_i386_darwin,system_i386_iphonesim]);
-            { on darwin, the got register is virtual (and allocated earlier
-              already) }
-            if notdarwin then
-              { ecx could be used in leaf procedures that don't use ecx to pass
-                aparameter }
-              current_procinfo.got:=NR_EBX;
-            if notdarwin { needs testing before it can be enabled for non-darwin platforms
-                and
-               (current_settings.optimizecputype in [cpu_Pentium2,cpu_Pentium3,cpu_Pentium4]) } then
+            if not (target_info.system in [system_i386_darwin,system_i386_iphonesim]) then
               begin
-                current_module.requires_ebx_pic_helper:=true;
-                a_call_name_static(list,'fpc_geteipasebx');
+                { Use ECX as a temp register by default }
+                tmpreg:=NR_ECX;
+                { Allocate registers used for parameters to make sure they
+                  never allocated during this PIC init code }
+                for i:=0 to current_procinfo.procdef.paras.Count - 1 do
+                  with tparavarsym(current_procinfo.procdef.paras[i]).paraloc[calleeside].Location^ do
+                    if Loc in [LOC_REGISTER, LOC_CREGISTER] then begin
+                      a_reg_alloc(list, register);
+                      { If ECX is used for a parameter, use EBX as temp }
+                      if getsupreg(register) = RS_ECX then
+                        tmpreg:=NR_EBX;
+                    end;
+
+                if tmpreg = NR_EBX then
+                  begin
+                    { Mark EBX as used in the proc }
+                    include(rg[R_INTREGISTER].used_in_proc,RS_EBX);
+                    current_module.requires_ebx_pic_helper:=true;
+                    a_call_name_static(list,'fpc_geteipasebx');
+                  end
+                else
+                  begin
+                    current_module.requires_ecx_pic_helper:=true;
+                    a_call_name_static(list,'fpc_geteipasecx');
+                  end;
+                list.concat(taicpu.op_sym_ofs_reg(A_ADD,S_L,current_asmdata.RefAsmSymbol('_GLOBAL_OFFSET_TABLE_'),0,tmpreg));
+                list.concat(taicpu.op_reg_reg(A_MOV,S_L,tmpreg,current_procinfo.got));
+
+                { Deallocate parameter registers }
+                for i:=0 to current_procinfo.procdef.paras.Count - 1 do
+                  with tparavarsym(current_procinfo.procdef.paras[i]).paraloc[calleeside].Location^ do
+                    if Loc in [LOC_REGISTER, LOC_CREGISTER] then
+                      a_reg_dealloc(list, register);
               end
             else
               begin
@@ -613,11 +614,6 @@ unit cgcpu;
                 a_label(list,current_procinfo.CurrGotLabel);
                 list.concat(taicpu.op_reg(A_POP,S_L,current_procinfo.got))
               end;
-            if notdarwin then
-              begin
-                list.concat(taicpu.op_sym_ofs_reg(A_ADD,S_L,current_asmdata.RefAsmSymbol('_GLOBAL_OFFSET_TABLE_'),0,NR_PIC_OFFSET_REG));
-                list.concat(tai_regalloc.alloc(NR_PIC_OFFSET_REG,nil));
-              end;
           end;
       end;
 

compiler/i386/cpupi.pas

@@ -97,8 +97,7 @@ unit cpupi;
 
     procedure ti386procinfo.allocate_got_register(list: tasmlist);
       begin
-        if (target_info.system in [system_i386_darwin,system_i386_iphonesim]) and
-           (cs_create_pic in current_settings.moduleswitches) then
+        if (cs_create_pic in current_settings.moduleswitches) then
           begin
             got := cg.getaddressregister(list);
           end;

compiler/i386/hlcgcpu.pas

@@ -40,6 +40,7 @@ interface
      protected
       procedure gen_loadfpu_loc_cgpara(list: TAsmList; size: tdef; const l: tlocation; const cgpara: tcgpara; locintsize: longint); override;
      public
+      function a_call_name(list: TAsmList; pd: tprocdef; const s: TSymStr; const paras: array of pcgpara; forceresdef: tdef; weak: boolean): tcgpara; override;
       procedure g_copyvaluepara_openarray(list: TAsmList; const ref: treference; const lenloc: tlocation; arrdef: tarraydef; destreg: tregister); override;
       procedure g_releasevaluepara_openarray(list: TAsmList; arrdef: tarraydef; const l: tlocation); override;
     end;
@@ -49,7 +50,10 @@ interface
 implementation
 
   uses
+    globals, procinfo,
     globtype,verbose,
+    fmodule,systems,
+    aasmbase,aasmtai,
     paramgr,
     cpubase,tgobj,cgobj,cgcpu;
 
@@ -170,6 +174,30 @@ implementation
     end;
 
 
+  function thlcgcpu.a_call_name(list: TAsmList; pd: tprocdef; const s: TSymStr; const paras: array of pcgpara; forceresdef: tdef; weak: boolean): tcgpara;
+  var
+    need_got_load: boolean;
+  begin
+    { Load GOT address to EBX before calling an external function.
+      It is needed because GOT stubs for external function calls
+      generated by a linker expect EBX as a GOT register. }
+    need_got_load:=not (target_info.system in systems_darwin) and
+                   (cs_create_pic in current_settings.moduleswitches) and
+                   (tf_pic_uses_got in target_info.flags) and
+                   (po_external in pd.procoptions);
+    if need_got_load then
+      begin
+        { Alloc EBX }
+        getcpuregister(list, NR_PIC_OFFSET_REG);
+        list.concat(taicpu.op_reg_reg(A_MOV,S_L,current_procinfo.got,NR_PIC_OFFSET_REG));
+      end;
+    Result:=inherited a_call_name(list, pd, s, paras, forceresdef, weak);
+    { Free EBX }
+    if need_got_load then
+      ungetcpuregister(list, NR_PIC_OFFSET_REG);
+  end;
+
+
   procedure thlcgcpu.g_copyvaluepara_openarray(list: TAsmList; const ref: treference; const lenloc: tlocation; arrdef: tarraydef; destreg: tregister);
     begin
       if paramanager.use_fixed_stack then

compiler/x86/aasmcpu.pas

@@ -1124,10 +1124,7 @@ implementation
 {$ifdef i386}
                      or (
                          (ref^.refaddr in [addr_pic]) and
-                         { allow any base for assembler blocks }
-                        ((assigned(current_procinfo) and
-                         (pi_has_assembler_block in current_procinfo.flags) and
-                         (ref^.base<>NR_NO)) or (ref^.base=NR_EBX))
+                         (ref^.base<>NR_NO)
                         )
 {$endif i386}
 {$ifdef x86_64}