* Implemented stackframe optimization for ARM CPU.

git-svn-id: trunk@5374 -

compiler/arm/cgcpu.pas

@@ -1156,7 +1156,8 @@ unit cgcpu;
          firstfloatreg,lastfloatreg,
          r : byte;
          i : aint;
-        again : tasmlabel;
+         again : tasmlabel;
+         regs : tcpuregisterset;
       begin
         LocalSize:=align(LocalSize,4);
         if not(nostackframe) then
@@ -1171,19 +1172,28 @@ unit cgcpu;
                   lastfloatreg:=r;
                 end;
             a_reg_alloc(list,NR_STACK_POINTER_REG);
-            a_reg_alloc(list,NR_FRAME_POINTER_REG);
+            if current_procinfo.framepointer<>NR_STACK_POINTER_REG then
-            a_reg_alloc(list,NR_R12);
+              begin
+                a_reg_alloc(list,NR_FRAME_POINTER_REG);
+                a_reg_alloc(list,NR_R12);
 
-            list.concat(taicpu.op_reg_reg(A_MOV,NR_R12,NR_STACK_POINTER_REG));
+                list.concat(taicpu.op_reg_reg(A_MOV,NR_R12,NR_STACK_POINTER_REG));
+              end;
             { save int registers }
             reference_reset(ref);
             ref.index:=NR_STACK_POINTER_REG;
             ref.addressmode:=AM_PREINDEXED;
-            list.concat(setoppostfix(taicpu.op_ref_regset(A_STM,ref,
+            regs:=rg[R_INTREGISTER].used_in_proc-paramanager.get_volatile_registers_int(pocall_stdcall);
-              rg[R_INTREGISTER].used_in_proc-paramanager.get_volatile_registers_int(pocall_stdcall)+[RS_R11,RS_R12,RS_R14,RS_R15]),
+            if current_procinfo.framepointer<>NR_STACK_POINTER_REG then
-              PF_FD));
+              regs:=regs+[RS_R11,RS_R12,RS_R14,RS_R15]
+            else
+              if (regs<>[]) or (pi_do_call in current_procinfo.flags) then
+                include(regs,RS_R14);
+            if regs<>[] then
+              list.concat(setoppostfix(taicpu.op_ref_regset(A_STM,ref,regs),PF_FD));
 
-            list.concat(taicpu.op_reg_reg_const(A_SUB,NR_FRAME_POINTER_REG,NR_R12,4));
+            if current_procinfo.framepointer<>NR_STACK_POINTER_REG then
+              list.concat(taicpu.op_reg_reg_const(A_SUB,NR_FRAME_POINTER_REG,NR_R12,4));
 
             { allocate necessary stack size
               not necessary according to Yury Sidorov
@@ -1248,17 +1258,20 @@ unit cgcpu;
              end
             else
             }
-            if not(is_shifter_const(localsize,shift)) then
+            if LocalSize<>0 then
-              begin
+              if not(is_shifter_const(localsize,shift)) then
-                a_load_const_reg(list,OS_ADDR,LocalSize,NR_R12);
+                begin
-                list.concat(taicpu.op_reg_reg_reg(A_SUB,NR_STACK_POINTER_REG,NR_STACK_POINTER_REG,NR_R12));
+                  if current_procinfo.framepointer=NR_STACK_POINTER_REG then
-                a_reg_dealloc(list,NR_R12);
+                    a_reg_alloc(list,NR_R12);
-              end
+                  a_load_const_reg(list,OS_ADDR,LocalSize,NR_R12);
-            else
+                  list.concat(taicpu.op_reg_reg_reg(A_SUB,NR_STACK_POINTER_REG,NR_STACK_POINTER_REG,NR_R12));
-              begin
+                  a_reg_dealloc(list,NR_R12);
-                a_reg_dealloc(list,NR_R12);
+                end
-                list.concat(taicpu.op_reg_reg_const(A_SUB,NR_STACK_POINTER_REG,NR_STACK_POINTER_REG,LocalSize));
+              else
-              end;
+                begin
+                  a_reg_dealloc(list,NR_R12);
+                  list.concat(taicpu.op_reg_reg_const(A_SUB,NR_STACK_POINTER_REG,NR_STACK_POINTER_REG,LocalSize));
+                end;
 
             if firstfloatreg<>RS_NO then
               begin
@@ -1287,6 +1300,8 @@ unit cgcpu;
          firstfloatreg,lastfloatreg,
          r : byte;
          shift : byte;
+         regs : tcpuregisterset;
+         LocalSize : longint;
       begin
         if not(nostackframe) then
           begin
@@ -1320,7 +1335,37 @@ unit cgcpu;
               end;
 
             if (current_procinfo.framepointer=NR_STACK_POINTER_REG) then
-              list.concat(taicpu.op_reg_reg(A_MOV,NR_R15,NR_R14))
+              begin
+                LocalSize:=current_procinfo.calc_stackframe_size;
+                if LocalSize<>0 then
+                  if not(is_shifter_const(LocalSize,shift)) then
+                    begin
+                      a_reg_alloc(list,NR_R12);
+                      a_load_const_reg(list,OS_ADDR,LocalSize,NR_R12);
+                      list.concat(taicpu.op_reg_reg_reg(A_ADD,NR_STACK_POINTER_REG,NR_STACK_POINTER_REG,NR_R12));
+                      a_reg_dealloc(list,NR_R12);
+                    end
+                  else
+                    begin
+                      list.concat(taicpu.op_reg_reg_const(A_ADD,NR_STACK_POINTER_REG,NR_STACK_POINTER_REG,LocalSize));
+                    end;
+                    
+                regs:=rg[R_INTREGISTER].used_in_proc-paramanager.get_volatile_registers_int(pocall_stdcall);
+                if (pi_do_call in current_procinfo.flags) or (regs<>[]) then
+                  begin
+                    exclude(regs,RS_R14);
+                    include(regs,RS_R15);
+                  end;
+                if regs=[] then
+                  list.concat(taicpu.op_reg_reg(A_MOV,NR_R15,NR_R14))
+                else
+                  begin
+                    reference_reset(ref);
+                    ref.index:=NR_STACK_POINTER_REG;
+                    ref.addressmode:=AM_PREINDEXED;
+                    list.concat(setoppostfix(taicpu.op_ref_regset(A_LDM,ref,regs),PF_FD));
+                  end;
+              end
             else
               begin
                 { restore int registers and return }

compiler/arm/cpuinfo.pas

@@ -88,7 +88,7 @@ Const
                                  genericlevel3optimizerswitches-
                                  { no need to write info about those }
                                  [cs_opt_level1,cs_opt_level2,cs_opt_level3]+
-                                 [cs_opt_regvar,cs_opt_loopunroll,cs_opt_tailrecursion];
+                                 [cs_opt_regvar,cs_opt_loopunroll,cs_opt_tailrecursion,cs_opt_stackframe];
 
    level1optimizerswitches = genericlevel1optimizerswitches;
    level2optimizerswitches = genericlevel2optimizerswitches + level1optimizerswitches + [cs_opt_regvar,cs_opt_stackframe,cs_opt_tailrecursion];

compiler/arm/cpupi.pas

@@ -62,7 +62,8 @@ unit cpupi;
           is especially a problem when taking the address of a local. For now,
           this extra memory should hurt less than generating all local contants with offsets
           >256 as non shifter constants }
-        tg.setfirsttemp(-12-28);
+        if tg.direction = -1 then
+          tg.setfirsttemp(-12-28);
       end;
 
 

compiler/psub.pas

@@ -745,7 +745,7 @@ implementation
             { set the start offset to the start of the temp area in the stack }
             tg:=ttgobj.create;
 
-{$ifdef x86}
+{$if defined(x86) or defined(arm)}
             { try to strip the stack frame }
             { set the framepointer to esp if:
               - no assembler directive, those are handled elsewhere
@@ -761,7 +761,11 @@ implementation
                not(po_assembler in procdef.procoptions) and
                ((flags*[pi_has_assembler_block,pi_uses_exceptions,pi_is_assembler,
                        pi_needs_implicit_finally,pi_has_implicit_finally,pi_has_stackparameter,
-                       pi_needs_stackframe])=[]) then
+                       pi_needs_stackframe])=[])
+               {$ifdef arm}
+               and ((cs_fp_emulation in current_settings.moduleswitches) or not (pi_uses_fpu in flags))
+               {$endif arm}
+             then
                begin
                  { we need the parameter info here to determine if the procedure gets
                    parameters on the stack
@@ -777,8 +781,7 @@ implementation
                      tg.direction:=1;
                    end;
                end;
-{$endif x86}
+{$endif}
-
             { Create register allocator }
             cg.init_register_allocators;