* Some work to allow mmx instructions to be used for 32 byte sets

compiler/i386/n386add.pas

@@ -45,6 +45,7 @@ interface
           procedure second_addboolean;
           procedure second_addfloat;
           procedure second_addsmallset;
+          procedure second_addmmxset;
           procedure second_mul;
 {$ifdef SUPPORT_MMX}
           procedure second_addmmx;
@@ -753,10 +754,8 @@ interface
             end;
           lten,gten:
             begin
-              If (not(nf_swaped in flags) and
+              if (not(nf_swaped in flags) and (nodetype = lten)) or
-                  (nodetype = lten)) or
+                 ((nf_swaped in flags) and (nodetype = gten)) then
-                 ((nf_swaped in flags) and
-                  (nodetype = gten)) then
                 swapleftright;
               location_force_reg(exprasmlist,left.location,opsize_2_cgsize[opsize],true);
               emit_op_right_left(A_AND,opsize);
@@ -787,6 +786,87 @@ interface
         set_result_location(cmpop,true);
       end;
 
+{*****************************************************************************
+                                   addmmxset
+*****************************************************************************}
+
+    procedure ti386addnode.second_addmmxset;
+
+    var opsize : TOpSize;
+        op     : TAsmOp;
+        cmpop,
+        pushedfpu,
+        noswap : boolean;
+    begin
+      pass_left_and_right(pushedfpu);
+
+      cmpop:=false;
+      noswap:=false;
+      opsize:=S_L;
+      case nodetype of
+        addn:
+          begin
+            { are we adding set elements ? }
+            if right.nodetype=setelementn then
+              begin
+                { adding elements is not commutative }
+{                if nf_swaped in flags then
+                  swapleftright;}
+                { bts requires both elements to be registers }
+{                location_force_reg(exprasmlist,left.location,opsize_2_cgsize[opsize],false);
+                location_force_reg(exprasmlist,right.location,opsize_2_cgsize[opsize],true);
+                op:=A_BTS;
+                noswap:=true;}
+              end
+            else
+              op:=A_POR;
+          end;
+        symdifn :
+          op:=A_PXOR;
+        muln:
+          op:=A_PAND;
+        subn:
+          op:=A_PANDN;
+        equaln,
+        unequaln :
+          begin
+            op:=A_PCMPEQD;
+            cmpop:=true;
+          end;
+        lten,gten:
+          begin
+            if (not(nf_swaped in flags) and (nodetype = lten)) or
+               ((nf_swaped in flags) and (nodetype = gten)) then
+              swapleftright;
+            location_force_reg(exprasmlist,left.location,opsize_2_cgsize[opsize],true);
+            emit_op_right_left(A_AND,opsize);
+            op:=A_PCMPEQD;
+            cmpop:=true;
+            { warning: ugly hack, we need a JE so change the node to equaln }
+            nodetype:=equaln;
+          end;
+          xorn :
+            op:=A_PXOR;
+          orn :
+            op:=A_POR;
+          andn :
+            op:=A_PAND;
+          else
+            internalerror(2003042215);
+        end;
+        { left must be a register }
+        left_must_be_reg(opsize,noswap);
+{        emit_generic_code(op,opsize,true,extra_not,false);}
+        location_freetemp(exprasmlist,right.location);
+        location_release(exprasmlist,right.location);
+        if cmpop then
+         begin
+           location_freetemp(exprasmlist,left.location);
+           location_release(exprasmlist,left.location);
+         end;
+        set_result_location(cmpop,true);
+      end;
+
 
 {*****************************************************************************
                                 Add64bit
@@ -1321,8 +1401,7 @@ interface
          extra_not : boolean;
 
       begin
-         { to make it more readable, string and set (not smallset!) have their
+         { to make it more readable, string and set have their own procedures }
-           own procedures }
          case left.resulttype.def.deftype of
            orddef :
              begin
@@ -1347,11 +1426,18 @@ interface
              end;
            setdef :
              begin
-               { normalsets are already handled in pass1 }
+              {Normalsets are already handled in pass1 if mmx
-               if (tsetdef(left.resulttype.def).settype<>smallset) then
+               should not be used.}
-                internalerror(200109041);
+              if (tsetdef(left.resulttype.def).settype<>smallset) then
-               second_addsmallset;
+                begin
-               exit;
+                  if cs_mmx in aktlocalswitches then
+                    second_addmmxset
+                  else
+                    internalerror(200109041);
+                end
+              else
+                second_addsmallset;
+              exit;
              end;
            arraydef :
              begin
@@ -1495,7 +1581,10 @@ begin
 end.
 {
   $Log$
-  Revision 1.88  2003-12-06 01:15:23  florian
+  Revision 1.89  2003-12-21 11:28:41  daniel
+    * Some work to allow mmx instructions to be used for 32 byte sets
+
+  Revision 1.88  2003/12/06 01:15:23  florian
     * reverted Peter's alloctemp patch; hopefully properly
 
   Revision 1.87  2003/12/03 23:13:20  peter

compiler/nadd.pas

@@ -1644,27 +1644,35 @@ implementation
          else if (ld.deftype=setdef) then
            begin
              if tsetdef(ld).settype=smallset then
-              begin
+               begin
                  if nodetype in [ltn,lten,gtn,gten,equaln,unequaln] then
-                  expectloc:=LOC_FLAGS
+                   expectloc:=LOC_FLAGS
                  else
-                  expectloc:=LOC_REGISTER;
+                   expectloc:=LOC_REGISTER;
                  { are we adding set elements ? }
                  if right.nodetype=setelementn then
                    calcregisters(self,2,0,0)
                  else
                    calcregisters(self,1,0,0);
-              end
+               end
              else
-              begin
+             {$ifdef i386}
-                 result := first_addset;
+               if cs_mmx in aktlocalswitches then
-                 if assigned(result) then
+                 begin
-                   exit;
+                   expectloc:=LOC_MMXREGISTER;
-                 expectloc:=LOC_CREFERENCE;
+                   calcregisters(self,0,0,4);
-                 calcregisters(self,0,0,0);
+                 end
-                 { here we call SET... }
+               else
-                 include(current_procinfo.flags,pi_do_call);
+             {$endif}
-              end;
+                 begin
+                   result := first_addset;
+                   if assigned(result) then
+                     exit;
+                   expectloc:=LOC_CREFERENCE;
+                   calcregisters(self,0,0,0);
+                   { here we call SET... }
+                   include(current_procinfo.flags,pi_do_call);
+                 end;
            end
 
          { compare pchar by addresses like BP/Delphi }
@@ -1874,7 +1882,10 @@ begin
 end.
 {
   $Log$
-  Revision 1.100  2003-12-09 21:17:04  jonas
+  Revision 1.101  2003-12-21 11:28:41  daniel
+    * Some work to allow mmx instructions to be used for 32 byte sets
+
+  Revision 1.100  2003/12/09 21:17:04  jonas
     + support for evaluating qword constant expressions (both arguments have
       to be a qword, constants have to be explicitly typecasted to qword)
 

compiler/ncgadd.pas

@@ -52,6 +52,9 @@ interface
           procedure second_addfloat;virtual;abstract;
           procedure second_addboolean;virtual;
           procedure second_addsmallset;virtual;
+        {$ifdef i386}
+          procedure second_addmmxset;virtual;abstract;
+        {$endif}
           procedure second_add64bit;virtual;
           procedure second_addordinal;virtual;
           procedure second_cmpfloat;virtual;abstract;
@@ -709,10 +712,19 @@ interface
             end;
           setdef :
             begin
-              { normalsets are already handled in pass1 }
+              {Normalsets are already handled in pass1 if mmx
+               should not be used.}
               if (tsetdef(left.resulttype.def).settype<>smallset) then
-                internalerror(200109041);
+                begin
-              second_opsmallset;
+                {$ifdef i386}
+                  if cs_mmx in aktlocalswitches then
+                    second_opmmxset
+                  else
+                {$endif}
+                    internalerror(200109041);
+                end
+              else
+                second_opsmallset;
             end;
           arraydef :
             begin
@@ -735,7 +747,10 @@ begin
 end.
 {
   $Log$
-  Revision 1.22  2003-10-17 01:22:08  florian
+  Revision 1.23  2003-12-21 11:28:41  daniel
+    * Some work to allow mmx instructions to be used for 32 byte sets
+
+  Revision 1.22  2003/10/17 01:22:08  florian
     * compilation of the powerpc compiler fixed
 
   Revision 1.21  2003/10/10 17:48:13  peter