* also use sse2 instructions to convert int to real on i386

for (types convertable to) longint
  * fixed expectloc for x86_64 int_to_real if result is in mmreg

git-svn-id: trunk@9423 -

compiler/x86/nx86cnv.pas

@@ -61,7 +61,7 @@ implementation
       verbose,systems,globals,globtype,
       aasmbase,aasmtai,aasmdata,aasmcpu,
       symconst,symdef,
-      cgbase,cga,procinfo,pass_2,
+      cgbase,cga,procinfo,pass_1,pass_2,
       ncon,ncal,ncnv,
       cpubase,
       cgutils,cgobj,cgx86,ncgutil,
@@ -177,10 +177,21 @@ implementation
          current_procinfo.CurrFalseLabel:=oldFalseLabel;
        end;
 
+
     function tx86typeconvnode.first_int_to_real : tnode;
 
       begin
         first_int_to_real:=nil;
+        if (left.resultdef.size<4) then
+          begin
+            inserttypeconv(left,s32inttype);
+            firstpass(left)
+          end;
+
+        if use_sse(resultdef) and
+           (torddef(left.resultdef).ordtype = s32bit) then
+          expectloc:=LOC_MMREGISTER
+        else
           expectloc:=LOC_FPUREGISTER;
       end;
 
@@ -190,11 +201,52 @@ implementation
       var
          href : treference;
          l1,l2 : tasmlabel;
+         op: tasmop;
+         opsize: topsize;
          signtested : boolean;
       begin
-        if (left.location.loc in [LOC_SUBSETREG,LOC_CSUBSETREG,LOC_SUBSETREF,LOC_CSUBSETREF]) then
+        if not(left.location.loc in [LOC_REGISTER,LOC_CREGISTER,LOC_REFERENCE,LOC_CREFERENCE]) then
-          location_force_reg(current_asmdata.CurrAsmList,left.location,left.location.size,true);
+          location_force_reg(current_asmdata.CurrAsmList,left.location,left.location.size,false);
-
+        if use_sse(resultdef) and
+{$ifdef cpu64bit}
+           (torddef(left.resultdef).ordtype in [s32bit,s64bit]) then
+{$else cpu64bit}
+           (torddef(left.resultdef).ordtype=s32bit) then
+{$endif cpu64bit}
+          begin
+            location_reset(location,LOC_MMREGISTER,def_cgsize(resultdef));
+            location.register:=cg.getmmregister(current_asmdata.CurrAsmList,location.size);
+            case location.size of
+              OS_F32:
+                op:=A_CVTSI2SS;
+              OS_F64:
+                op:=A_CVTSI2SD;
+              else
+                internalerror(2007120902);
+            end;
+            case left.location.size of
+              OS_S32:
+                opsize:=S_L;
+              OS_S64:
+                opsize:=S_Q;
+              else
+                internalerror(2007120903);
+            end;
+            case left.location.loc of
+              LOC_REFERENCE,
+              LOC_CREFERENCE:
+                begin
+                  href:=left.location.reference;
+                  tcgx86(cg).make_simple_ref(current_asmdata.CurrAsmList,href);
+                  current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(op,opsize,href,location.register));
+                end;
+              LOC_REGISTER,
+              LOC_CREGISTER:
+                current_asmdata.CurrAsmList.concat(taicpu.op_reg_reg(op,opsize,left.location.register,location.register));
+            end;
+          end
+        else
+          begin
             location_reset(location,LOC_FPUREGISTER,def_cgsize(resultdef));
             if (left.location.loc=LOC_REGISTER) and (torddef(left.resultdef).ordtype=u64bit) then
               begin
@@ -266,21 +318,17 @@ implementation
               else
                 begin
                   if left.resultdef.size<4 then
-                begin
+                    internalerror(2007120901);
-                  tg.GetTemp(current_asmdata.CurrAsmList,4,tt_normal,href);
-                  location_freetemp(current_asmdata.CurrAsmList,left.location);
-                  cg.a_load_ref_ref(current_asmdata.CurrAsmList,left.location.size,OS_32,left.location.reference,href);
-                  left.location.reference:=href;
-                end;
                  href:=left.location.reference;
                  tcgx86(cg).make_simple_ref(current_asmdata.CurrAsmList,href);
                  current_asmdata.CurrAsmList.concat(taicpu.op_ref(A_FILD,S_IL,href));
                 end;
             end;
-        location_freetemp(current_asmdata.CurrAsmList,left.location);
             tcgx86(cg).inc_fpu_stack;
             location.register:=NR_ST;
           end;
+        location_freetemp(current_asmdata.CurrAsmList,left.location);
+      end;
 
 begin
   ctypeconvnode:=tx86typeconvnode

compiler/x86_64/nx64cnv.pas

@@ -26,7 +26,7 @@ unit nx64cnv;
 interface
 
     uses
-      node,ncgcnv,defutil,defcmp,
+      node,ncgcnv,defutil,defcmp,pass_1,
       nx86cnv;
 
     type
@@ -41,6 +41,7 @@ interface
          { procedure second_chararray_to_string;override; }
          { procedure second_char_to_string;override; }
          { function first_int_to_real: tnode; override; }
+         function first_int_to_real : tnode;override;
          procedure second_int_to_real;override;
          { procedure second_real_to_real;override; }
          { procedure second_cord_to_pointer;override; }
@@ -68,6 +69,22 @@ implementation
       tgobj;
 
 
+    function tx8664typeconvnode.first_int_to_real : tnode;
+      begin
+        result:=nil;
+        if use_sse(resultdef) and
+           (torddef(left.resultdef).ordtype=u32bit) then
+          begin
+            inserttypeconv(left,s64inttype);
+            firstpass(left);
+          end
+        else
+          result:=inherited first_int_to_real;
+       if use_sse(resultdef) then
+         expectloc:=LOC_MMREGISTER;
+      end;
+
+
     procedure tx8664typeconvnode.second_int_to_real;
       var
          href : treference;
@@ -76,11 +93,6 @@ implementation
       begin
         if use_sse(resultdef) then
           begin
-            { We can only directly convert s32bit and s64bit,u64bit values, for other
-              values convert first to s64bit }
-            if not(torddef(left.resultdef).ordtype in [s32bit,s64bit,u64bit]) then
-              location_force_reg(current_asmdata.CurrAsmList,left.location,OS_S64,false);
-
             if is_double(resultdef) then
               op:=A_CVTSI2SD
             else if is_single(resultdef) then
@@ -89,7 +101,7 @@ implementation
               internalerror(200506061);
 
             location_reset(location,LOC_MMREGISTER,def_cgsize(resultdef));
-            location.register:=cg.getmmregister(current_asmdata.CurrAsmList,def_cgsize(resultdef));
+            location.register:=cg.getmmregister(current_asmdata.CurrAsmList,location.size);
 
             case torddef(left.resultdef).ordtype of
               u64bit:
@@ -148,24 +160,7 @@ implementation
                    cg.a_label(current_asmdata.CurrAsmList,l2);
                 end
               else
-                begin
+                inherited second_int_to_real;
-                  if not(left.location.loc in [LOC_REGISTER,LOC_CREGISTER,LOC_REFERENCE,LOC_CREFERENCE]) then
-                    location_force_reg(current_asmdata.CurrAsmList,left.location,left.location.size,false);
-                  case left.location.loc of
-                    LOC_CREFERENCE,
-                    LOC_REFERENCE :
-                      begin
-                        href:=left.location.reference;
-                        tcgx86(cg).make_simple_ref(current_asmdata.CurrAsmList,href);
-                        current_asmdata.CurrAsmList.concat(Taicpu.op_ref_reg(op,tcgsize2opsize[left.location.size],href,location.register));
-                      end;
-                    LOC_CREGISTER,
-                    LOC_REGISTER :
-                      current_asmdata.CurrAsmList.concat(Taicpu.op_reg_reg(op,tcgsize2opsize[left.location.size],left.location.register,location.register));
-                    else
-                      internalerror(200506072);
-                  end;
-                end;
             end;
           end
         else