diff --git a/compiler/i386/aoptcpu.pas b/compiler/i386/aoptcpu.pas
index 07ba868da1..1f28e33efa 100644
--- a/compiler/i386/aoptcpu.pas
+++ b/compiler/i386/aoptcpu.pas
@@ -40,7 +40,6 @@ unit aoptcpu;
         procedure PeepHoleOptPass1; override;
         procedure PeepHoleOptPass2; override;
         procedure PostPeepHoleOpts; override;
-        function DoFpuLoadStoreOpt(var p : tai) : boolean;
       end;
 
     Var
@@ -58,56 +57,6 @@ unit aoptcpu;
       { units we should get rid off: }
       symsym,symconst;
 
-    function TCPUAsmoptimizer.DoFpuLoadStoreOpt(var p: tai): boolean;
-    { returns true if a "continue" should be done after this optimization }
-    var hp1, hp2: tai;
-    begin
-      DoFpuLoadStoreOpt := false;
-      if (taicpu(p).oper[0]^.typ = top_ref) and
-         getNextInstruction(p, hp1) and
-         (hp1.typ = ait_instruction) and
-         (((taicpu(hp1).opcode = A_FLD) and
-           (taicpu(p).opcode = A_FSTP)) or
-          ((taicpu(p).opcode = A_FISTP) and
-           (taicpu(hp1).opcode = A_FILD))) and
-         (taicpu(hp1).oper[0]^.typ = top_ref) and
-         (taicpu(hp1).opsize = taicpu(p).opsize) and
-         RefsEqual(taicpu(p).oper[0]^.ref^, taicpu(hp1).oper[0]^.ref^) then
-        begin
-          { replacing fstp f;fld f by fst f is only valid for extended because of rounding }
-          if (taicpu(p).opsize=S_FX) and
-             getNextInstruction(hp1, hp2) and
-             (hp2.typ = ait_instruction) and
-             IsExitCode(hp2) and
-             (taicpu(p).oper[0]^.ref^.base = current_procinfo.FramePointer) and
-             not(assigned(current_procinfo.procdef.funcretsym) and
-                 (taicpu(p).oper[0]^.ref^.offset < tabstractnormalvarsym(current_procinfo.procdef.funcretsym).localloc.reference.offset)) and
-             (taicpu(p).oper[0]^.ref^.index = NR_NO) then
-            begin
-              asml.remove(p);
-              asml.remove(hp1);
-              p.free;
-              hp1.free;
-              p := hp2;
-              removeLastDeallocForFuncRes(p);
-              doFPULoadStoreOpt := true;
-            end
-          (* can't be done because the store operation rounds
-          else
-            { fst can't store an extended value! }
-            if (taicpu(p).opsize <> S_FX) and
-               (taicpu(p).opsize <> S_IQ) then
-              begin
-                if (taicpu(p).opcode = A_FSTP) then
-                  taicpu(p).opcode := A_FST
-                else taicpu(p).opcode := A_FIST;
-                asml.remove(hp1);
-                hp1.free;
-              end
-          *)
-        end;
-    end;
-
 
   { converts a TChange variable to a TRegister }
   function tch2reg(ch: tinschange): tsuperregister;
@@ -577,7 +526,7 @@ begin
                             end
                     end;
                   A_FSTP,A_FISTP:
-                    if doFpuLoadStoreOpt(p) then
+                    if OptPass1FSTP(p) then
                       continue;
                   A_LEA:
                     begin
@@ -776,7 +725,7 @@ begin
                 if OptPass2Jcc(p) then
                   continue;
               A_FSTP,A_FISTP:
-                if DoFpuLoadStoreOpt(p) then
+                if OptPass1FSTP(p) then
                   continue;
               A_IMUL:
                 if OptPass2Imul(p) then
diff --git a/compiler/x86/aoptx86.pas b/compiler/x86/aoptx86.pas
index aa1a30d3c4..6afa3de88c 100644
--- a/compiler/x86/aoptx86.pas
+++ b/compiler/x86/aoptx86.pas
@@ -71,6 +71,7 @@ unit aoptx86;
         function OptPass1Sub(var p : tai) : boolean;
         function OptPass1SHLSAL(var p : tai) : boolean;
         function OptPass1SETcc(var p: tai): boolean;
+        function OptPass1FSTP(var p: tai): boolean;
 
         function OptPass2MOV(var p : tai) : boolean;
         function OptPass2Imul(var p : tai) : boolean;
@@ -2448,6 +2449,58 @@ unit aoptx86;
       end;
 
 
+    function TX86AsmOptimizer.OptPass1FSTP(var p: tai): boolean;
+      { returns true if a "continue" should be done after this optimization }
+      var
+        hp1, hp2: tai;
+      begin
+        Result := false;
+        if MatchOpType(taicpu(p),top_ref) and
+           GetNextInstruction(p, hp1) and
+           (hp1.typ = ait_instruction) and
+           (((taicpu(hp1).opcode = A_FLD) and
+             (taicpu(p).opcode = A_FSTP)) or
+            ((taicpu(p).opcode = A_FISTP) and
+             (taicpu(hp1).opcode = A_FILD))) and
+           MatchOpType(taicpu(hp1),top_ref) and
+           (taicpu(hp1).opsize = taicpu(p).opsize) and
+           RefsEqual(taicpu(p).oper[0]^.ref^, taicpu(hp1).oper[0]^.ref^) then
+          begin
+            { replacing fstp f;fld f by fst f is only valid for extended because of rounding }
+            if (taicpu(p).opsize=S_FX) and
+               GetNextInstruction(hp1, hp2) and
+               (hp2.typ = ait_instruction) and
+               IsExitCode(hp2) and
+               (taicpu(p).oper[0]^.ref^.base = current_procinfo.FramePointer) and
+               not(assigned(current_procinfo.procdef.funcretsym) and
+                   (taicpu(p).oper[0]^.ref^.offset < tabstractnormalvarsym(current_procinfo.procdef.funcretsym).localloc.reference.offset)) and
+               (taicpu(p).oper[0]^.ref^.index = NR_NO) then
+              begin
+                asml.remove(p);
+                asml.remove(hp1);
+                p.free;
+                hp1.free;
+                p := hp2;
+                RemoveLastDeallocForFuncRes(p);
+                Result := true;
+              end
+            (* can't be done because the store operation rounds
+            else
+              { fst can't store an extended value! }
+              if (taicpu(p).opsize <> S_FX) and
+                 (taicpu(p).opsize <> S_IQ) then
+                begin
+                  if (taicpu(p).opcode = A_FSTP) then
+                    taicpu(p).opcode := A_FST
+                  else taicpu(p).opcode := A_FIST;
+                  asml.remove(hp1);
+                  hp1.free;
+                end
+            *)
+          end;
+      end;
+
+
     function TX86AsmOptimizer.OptPass2MOV(var p : tai) : boolean;
       var
        hp1,hp2: tai;
diff --git a/compiler/x86_64/aoptcpu.pas b/compiler/x86_64/aoptcpu.pas
index 818d3c2353..960144abe4 100644
--- a/compiler/x86_64/aoptcpu.pas
+++ b/compiler/x86_64/aoptcpu.pas
@@ -118,6 +118,8 @@ uses
                 result:=OptPass1SHLSAL(p);
               A_SETcc:
                 result:=OptPass1SETcc(p);
+              A_FSTP,A_FISTP:
+                result:=OptPass1FSTP(p);
             end;
           end;
         end;