Re: [avr-gcc-list] Using current avr-gcc for target ATmega2561

[Martijn van Balen]

On 2005-10-03 17:26:10 Martijn van Balen
wrote:
>
> "What needs to be done (tasks)
>    (1) Align functions (and labels?) to 4-byte boundaries
so that gcc can continue to use 16-bit values to represent function pointers.
>    (2) Take into account, that these devices store three
bytes on the stack for each function call" 
>
> I can easily avoid (1) by using <128Kbyte flash from C. But I cannot
figure out what problem is solved by (2). 
> Indirect jumps can be done by IJMP/ICALL, no need to push addresses
yourself. Before diving into avr.md 
> and other files, could anybody point my in the right direction ?

>
To answer my own question, at least one problem is the addressing of function
parameters that were put on
the stack. I spend some time browsing
the gcc source code and finally found at least one
function in .../gcc/config/avr/avr.c
that needs knowledge of the size of the program counter:

int
initial_elimination_offset (int from,
int to)
{
  if (from == FRAME_POINTER_REGNUM
&& to == STACK_POINTER_REGNUM)
    return 0;
  else
    {
      int offset = frame_pointer_needed
? 2 : 0;

      offset += avr_regs_to_save
(NULL);
      return get_frame_size
() + 2 + 1 + offset;
    }
}

After a small experiment, I concluded
that replacing the last 2 with a 3 indeed solves
the problem with the function parameters.
So for now, I can use my patched compiler
to generate mega2561 code.

What I want to ask to the guru's is:
do you agree ? Or do I need to patch more to
get a useful compiler ?

Kind regards,
Martijn.

/*
 * Test program, generate assembly
with 'cc1 tst.c'
 *
 * p6 is passed through the stack
 *
 */
char f(long p1, long p2, long p3, long
p4, char p5, char p6)
{
  return p6;
}

main()
{
  f(42, 42, 42, 42, 42, 42);
}