+ enabled the rol/ror intrinsic on i8086

git-svn-id: trunk@35734 -

compiler/fpcdefs.inc

@@ -50,7 +50,7 @@
   {//$define SUPPORT_MMX}
   {$define cpumm}
   {$define fewintregisters}
-  {//$define cpurox}
+  {$define cpurox}
   {$define cpurefshaveindexreg}
   {$define SUPPORT_SAFECALL}
   {$define cpuneedsmulhelper}

compiler/i8086/cgcpu.pas

@@ -257,7 +257,7 @@ unit cgcpu;
         tmpreg: tregister;
         op1, op2: TAsmOp;
         ax_subreg: tregister;
-        hl_loop_start: tasmlabel;
+        hl_loop_start,hl_skip: tasmlabel;
         ai: taicpu;
         use_loop, use_186_fast_shift, use_8086_fast_shift,
           use_386_fast_shift: Boolean;
@@ -562,6 +562,61 @@ unit cgcpu;
                         end;
                     end;
                 end;
+              OP_ROL,OP_ROR:
+                begin
+                  a:=a and 31;
+                  if a=16 then
+                    list.Concat(taicpu.op_reg_reg(A_XCHG,S_W,reg,GetNextReg(reg)))
+                  else if ((a=1) and (op=OP_ROL)) or ((a=31) and (op=OP_ROR)) then
+                    begin
+                      list.Concat(taicpu.op_const_reg(A_SHL,S_W,1,GetNextReg(reg)));
+                      list.Concat(taicpu.op_const_reg(A_RCL,S_W,1,reg));
+                      list.Concat(taicpu.op_const_reg(A_ADC,S_W,0,GetNextReg(reg)));
+                    end
+                  else if ((a=15) and (op=OP_ROL)) or ((a=31) and (op=OP_ROR)) then
+                    begin
+                      list.Concat(taicpu.op_reg_reg(A_XCHG,S_W,reg,GetNextReg(reg)));
+
+                      list.Concat(taicpu.op_const_reg(A_SHR,S_W,1,reg));
+                      list.Concat(taicpu.op_const_reg(A_RCR,S_W,1,GetNextReg(reg)));
+
+                      current_asmdata.getjumplabel(hl_skip);
+
+                      ai:=Taicpu.Op_Sym(A_Jcc,S_NO,hl_skip);
+                      ai.SetCondition(C_NC);
+                      ai.is_jmp:=true;
+                      list.concat(ai);
+
+                      list.Concat(taicpu.op_const_reg(A_ADD,S_W,aint($8000),reg));
+
+                      a_label(list,hl_skip);
+                    end
+                  else if ((a=17) and (op=OP_ROL)) or ((a=31) and (op=OP_ROR)) then
+                    begin
+                      list.Concat(taicpu.op_reg_reg(A_XCHG,S_W,reg,GetNextReg(reg)));
+                      list.Concat(taicpu.op_const_reg(A_SHL,S_W,1,GetNextReg(reg)));
+                      list.Concat(taicpu.op_const_reg(A_RCL,S_W,1,reg));
+                      list.Concat(taicpu.op_const_reg(A_ADC,S_W,0,GetNextReg(reg)));
+                    end
+                  else if ((a=31) and (op=OP_ROL)) or ((a=1) and (op=OP_ROR)) then
+                    begin
+                      list.Concat(taicpu.op_const_reg(A_SHR,S_W,1,reg));
+                      list.Concat(taicpu.op_const_reg(A_RCR,S_W,1,GetNextReg(reg)));
+
+                      current_asmdata.getjumplabel(hl_skip);
+
+                      ai:=Taicpu.Op_Sym(A_Jcc,S_NO,hl_skip);
+                      ai.SetCondition(C_NC);
+                      ai.is_jmp:=true;
+                      list.concat(ai);
+
+                      list.Concat(taicpu.op_const_reg(A_ADD,S_W,aint($8000),reg));
+
+                      a_label(list,hl_skip);
+                    end
+                  else
+                    internalerror(2017040501);
+                end;
               else
                 begin
                   tmpreg:=getintregister(list,size);

compiler/nadd.pas

@@ -1047,8 +1047,13 @@ implementation
                    tshlshrnode(left).left.isequal(tshlshrnode(right).left) and
                    not(might_have_sideeffects(tshlshrnode(left).left)) then
                    begin
-                     if tordconstnode(tshlshrnode(left).right).value=
+                     if (tordconstnode(tshlshrnode(left).right).value=
-                       tshlshrnode(left).left.resultdef.size*8-tordconstnode(tshlshrnode(right).right).value then
+                       tshlshrnode(left).left.resultdef.size*8-tordconstnode(tshlshrnode(right).right).value)
+{$ifdef i8086}
+                         and (not(torddef(left.resultdef).ordtype in [s32bit,u32bit]) or
+                           (tordconstnode(tshlshrnode(left).right).value.svalue in [1,15,16,17,31]))
+{$endif i8086}
+                        then
                        begin
                          result:=cinlinenode.create(in_ror_x_y,false,
                            ccallparanode.create(tshlshrnode(left).right,
@@ -1057,8 +1062,13 @@ implementation
                          tshlshrnode(left).right:=nil;
                          exit;
                        end
-                     else if tordconstnode(tshlshrnode(right).right).value=
+                     else if (tordconstnode(tshlshrnode(right).right).value=
-                       tshlshrnode(left).left.resultdef.size*8-tordconstnode(tshlshrnode(left).right).value then
+                       tshlshrnode(left).left.resultdef.size*8-tordconstnode(tshlshrnode(left).right).value)
+{$ifdef i8086}
+                         and (not(torddef(left.resultdef).ordtype in [s32bit,u32bit]) or
+                           (tordconstnode(tshlshrnode(right).right).value.svalue in [1,15,16,17,31]))
+{$endif i8086}
+                        then
                        begin
                          result:=cinlinenode.create(in_rol_x_y,false,
                            ccallparanode.create(tshlshrnode(right).right,
@@ -1076,8 +1086,13 @@ implementation
                    tshlshrnode(left).left.isequal(tshlshrnode(right).left) and
                    not(might_have_sideeffects(tshlshrnode(left).left)) then
                    begin
-                     if tordconstnode(tshlshrnode(left).right).value=
+                     if (tordconstnode(tshlshrnode(left).right).value=
-                       tshlshrnode(left).left.resultdef.size*8-tordconstnode(tshlshrnode(right).right).value then
+                       tshlshrnode(left).left.resultdef.size*8-tordconstnode(tshlshrnode(right).right).value)
+{$ifdef i8086}
+                         and (not(torddef(left.resultdef).ordtype in [s32bit,u32bit]) or
+                           (tordconstnode(tshlshrnode(left).right).value.svalue in [1,15,16,17,31]))
+{$endif i8086}
+                        then
                        begin
                          result:=cinlinenode.create(in_rol_x_y,false,
                            ccallparanode.create(tshlshrnode(left).right,
@@ -1086,8 +1101,13 @@ implementation
                          tshlshrnode(left).right:=nil;
                          exit;
                        end
-                     else if tordconstnode(tshlshrnode(right).right).value=
+                     else if (tordconstnode(tshlshrnode(right).right).value=
-                       tshlshrnode(left).left.resultdef.size*8-tordconstnode(tshlshrnode(left).right).value then
+                       tshlshrnode(left).left.resultdef.size*8-tordconstnode(tshlshrnode(left).right).value)
+{$ifdef i8086}
+                         and (not(torddef(left.resultdef).ordtype in [s32bit,u32bit]) or
+                           (tordconstnode(tshlshrnode(right).right).value.svalue in [1,15,16,17,31]))
+{$endif i8086}
+                        then
                        begin
                          result:=cinlinenode.create(in_ror_x_y,false,
                            ccallparanode.create(tshlshrnode(right).right,

rtl/inc/systemh.inc

@@ -904,10 +904,10 @@ function NtoLE(const AValue: QWord): QWord;{$ifdef SYSTEMINLINE}inline;{$endif}
 
 {$ifdef FPC_HAS_INTERNAL_ROX}
 
-{$if defined(cpux86_64) or defined(cpui386)}
+{$if defined(cpux86_64) or defined(cpui386) or defined(cpui8086)}
 {$define FPC_HAS_INTERNAL_ROX_BYTE}
 {$define FPC_HAS_INTERNAL_ROX_WORD}
-{$endif defined(cpux86_64) or defined(cpui386)}
+{$endif defined(cpux86_64) or defined(cpui386) or defined(cpui8086)}
 
 {$if defined(cpux86_64) or defined(cpui386) or defined(arm) or defined(powerpc) or defined(powerpc64) or defined(cpuaarch64)}
 {$define FPC_HAS_INTERNAL_ROX_DWORD}