* do prefetching correctly

  * more prefetching inserted
  * unnecessary/wrong UpdateUsedRegs removed

git-svn-id: trunk@43878 -

compiler/aopt.pas

@@ -148,6 +148,7 @@ Unit aopt;
           p := BlockStart;
           While (P <> BlockEnd) Do
             Begin
+              prefetch(pointer(p.Next)^);
               Case p.typ Of
                 ait_Label:
                   begin
@@ -346,6 +347,7 @@ Unit aopt;
         p:=BlockStart;
         while p<>BlockEnd Do
           begin
+            prefetch(pointer(p.Next)^);
             if SchedulerPass1Cpu(p) then
               continue;
             p:=tai(p.next);

compiler/aoptbase.pas

@@ -195,7 +195,12 @@ unit aoptbase;
 {$endif cpudelayslot}
              ((Current.typ = ait_label) And
               labelCanBeSkipped(Tai_Label(Current)))) Do
-        Current := tai(Current.Next);
+        begin
+          { this won't help the current loop, but it helps when returning from GetNextInstruction
+            as the next entry is probably already in the cache }
+          prefetch(pointer(Current.Next)^);
+          Current := Tai(Current.Next);
+        end;
       If Assigned(Current) And
          (Current.typ = ait_Marker) And
          (Tai_Marker(Current).Kind = mark_NoPropInfoStart) Then
@@ -203,7 +208,12 @@ unit aoptbase;
           While Assigned(Current) And
                 ((Current.typ <> ait_Marker) Or
                  (Tai_Marker(Current).Kind <> mark_NoPropInfoEnd)) Do
-            Current := Tai(Current.Next);
+            begin
+              { this won't help the current loop, but it helps when returning from GetNextInstruction
+                as the next entry is probably already in the cache }
+              prefetch(pointer(Current.Next)^);
+              Current := Tai(Current.Next);
+            end;
         End;
     Until Not(Assigned(Current)) Or
           (Current.typ <> ait_Marker) Or

compiler/aoptobj.pas

@@ -1016,6 +1016,7 @@ Unit AoptObj;
                    ((p.typ = ait_marker) and
                     (tai_Marker(p).Kind in [mark_AsmBlockEnd,mark_NoLineInfoStart,mark_NoLineInfoEnd]))) do
                  begin
+                   prefetch(pointer(p.Next)^);
                    { Here's the optimise part }
                    if (p.typ in [ait_align, ait_label]) then
                      begin
@@ -1040,6 +1041,7 @@ Unit AoptObj;
             while assigned(p) and
                   (p.typ=ait_RegAlloc) Do
               begin
+                prefetch(pointer(p.Next)^);
                 case tai_regalloc(p).ratype of
                   ra_alloc :
                     Include(UsedRegs[getregtype(tai_regalloc(p).reg)].UsedRegs, getsupreg(tai_regalloc(p).reg));
@@ -1074,6 +1076,7 @@ Unit AoptObj;
             while assigned(p) and
                   (p.typ=ait_RegAlloc) Do
               begin
+                prefetch(pointer(p.Next)^);
                 case tai_regalloc(p).ratype of
                   ra_alloc :
                     Include(UsedRegs[getregtype(tai_regalloc(p).reg)].UsedRegs, getsupreg(tai_regalloc(p).reg));
@@ -1693,6 +1696,7 @@ Unit AoptObj;
         { Stop if hp is an instruction, for example }
         while (hp1 <> BlockEnd) and (hp1.typ in [ait_label,ait_align]) do
           begin
+            prefetch(pointer(hp1.Next)^);
             case hp1.typ of
               ait_label:
                 begin
@@ -2396,13 +2400,13 @@ Unit AoptObj;
         ClearUsedRegs;
         while (p <> BlockEnd) Do
           begin
-            UpdateUsedRegs(tai(p.next));
+            prefetch(pointer(p.Next)^);
             if PrePeepHoleOptsCpu(p) then
               continue;
             if assigned(p) then
               begin
-                UpdateUsedRegs(p);
                 p:=tai(p.next);
+                UpdateUsedRegs(p);
               end;
           end;
       end;
@@ -2425,7 +2429,7 @@ Unit AoptObj;
 
           while Assigned(p) and (p <> BlockEnd) Do
             begin
-              prefetch(p.Next);
+              prefetch(pointer(p.Next)^);
 
               { I'am not sure why this is done, UsedRegs should reflect the register usage before the instruction
                 If an instruction needs the information of this, it can easily create a TempUsedRegs (FK)
@@ -2439,7 +2443,6 @@ Unit AoptObj;
               { Handle jump optimizations first }
               if JumpOptsAvailable and DoJumpOptimizations(p, stoploop) then
                 begin
-                  UpdateUsedRegs(p);
                   if FirstInstruction then
                     { Update StartPoint, since the old p was removed;
                       don't set FirstInstruction to False though, as
@@ -2481,6 +2484,7 @@ Unit AoptObj;
         ClearUsedRegs;
         while (p <> BlockEnd) Do
           begin
+            prefetch(pointer(p.Next)^);
             if PeepHoleOptPass2Cpu(p) then
               continue;
             if assigned(p) then
@@ -2497,13 +2501,13 @@ Unit AoptObj;
         ClearUsedRegs;
         while (p <> BlockEnd) Do
           begin
-            UpdateUsedRegs(tai(p.next));
+            prefetch(pointer(p.Next)^);
             if PostPeepHoleOptsCpu(p) then
               continue;
             if assigned(p) then
               begin
-                UpdateUsedRegs(p);
                 p:=tai(p.next);
+                UpdateUsedRegs(p);
               end;
           end;
       end;

compiler/cclasses.pas

@@ -2190,7 +2190,7 @@ end;
         while assigned(NewNode) do
          begin
            Next:=NewNode.Next;
-           prefetch(next.next);
+           prefetch(pointer(Next)^);
            NewNode.Free;
            NewNode:=Next;
           end;