Optimization of nested procedures and functions
lazarus / 64-FPC development

----------------------------------------------------------------------
lagprogramming · 2022-07-16 10:17

According to the documentation:
QuoteWhen a routine is declared within the scope of a procedure or
function, it is said to be nested. In this case, an additional
invisible parameter is passed to the nested routine. This additional
parameter is the frame pointer address of the parent routine. This
permits the nested routine to access the local variables and
parameters of the calling routine.
In the situation when the nested routine doesn't access any parameter
or local variable of the calling routine, would there be a problem if
the produced code would be as if the routine was not nested? The frame
pointer parameter occupies space with no benefit and it looks like it
might block a useful CPU register(look at variable "f" in the
following nested_yes and nested_no functions).

Here is an example where function1 and function2 do the same thing
except for the fact that one of them calls a nested routine and the
other one doesn't:

Code: Pascal  [Select][+][-]window.onload = function(){var x1 =
document.getElementById("main_content_section"); if (x1) { var x =
document.getElementsByClassName("geshi");for (var i = 0; i program
Project1; function nested_no(a,b,c,d,e,f:sizeint):sizeint;begin  if
a+b+c+d+e+f>0 then result:=-1 else result:=0;end; function
function1(a,b,c,d,e,f:sizeint):sizeint;begin 
result:=nested_no(a,b,a,b,a,b);end; function
function2(a,b,c,d,e,f:sizeint):sizeint;  function
nested_yes(a,b,c,d,e,f:sizeint):sizeint;  begin    if a+b+c+d+e+f>0
then result:=-1 else result:=0;  end;begin  result:=nested_yes(a,b,a,b
,a,b);end; beginwriteln(function1(round(random),round(random),round(ra
ndom),round(random),round(random),round(random)));writeln(function2(ro
und(random),round(random),round(random),round(random),round(random),ro
und(random)));end.

I attach FPC's produced code(look at variable "f" in the following
nested_yes and nested_no functions)

----------------------------------------------------------------------
marcov · 2022-07-16 10:56

If you mark nested_no as "inline", the assembler looks like

Code: [Select]pushl     %ebp
.Lc9:
.Lc10:
# Var a located in register eax
# Var b located in register edx
# Var c located in register ecx
# Var d located in stack [ebp+16]
# Var e located in stack [ebp+12]
# Var f located in stack [ebp+8]
.Ll6:
# [10] result:=nested_no(a,b,a,b,a,b);
        leal    (%eax,%edx),%ecx
        addl    %eax,%ecx
        addl    %edx,%ecx
        addl    %ecx,%eax
        addl    %edx,%eax
        testl   %eax,%eax
        jng     .Lj11
        movl    $-1,%eax
        jmp     .Lj12
        .p2align 4,,10
        .p2align 3
.Lj11:
        xorl    %eax,%eax
.Lj12:
# Var $result located in register eax
.Lc11:
.Lc12:
.Ll7:
# [11] end;
        popl    %ebp
        ret     $12


AND for function2:
.Ll10:
# [19] result:=nested_yes(a,b,a,b,a,b);
        leal    (%eax,%edx),%ecx
        addl    %eax,%ecx
        addl    %edx,%ecx
        addl    %ecx,%eax
        addl    %edx,%eax
        testl   %eax,%eax
        jng     .Lj18
        movl    $-1,%eax
        jmp     .Lj19
        .p2align 4,,10
        .p2align 3
.Lj18:
        xorl    %eax,%eax



Which doesn't call nested_no. So mark the nested functions you want to
inline with the inline directive like this:

Code: Pascal  [Select][+][-]window.onload = function(){var x1 =
document.getElementById("main_content_section"); if (x1) { var x =
document.getElementsByClassName("geshi");for (var i = 0; i function
nested_no(a,b,c,d,e,f:sizeint):sizeint; inline; // and  function
nested_yes(a,b,c,d,e,f:sizeint):sizeint; inline;
Note that inlining very large functions does not always produce a
better result. But

----------------------------------------------------------------------
PascalDragon · 2022-07-16 14:34

Quote from: lagprogramming on July 16, 2022, 12:17:20 pmIn the
situation when the nested routine doesn't access any parameter or
local variable of the calling routine, would there be a problem if the
produced code would be as if the routine was not nested? The frame
pointer parameter occupies space with no benefit and it looks like it
might block a useful CPU register(look at variable "f" in the
following nested_yes and nested_no functions).
FPC main contains the UnusedPara optimization which removes the parent
framepointer parameter in your example.