Re: [avr-gcc-list] GCC guru question (progue/epilogue sequences)

[Larry Barello]

In my RTOS (www.barello.net/avrx) I do the same thing you do.  Since
GCC allocates up to 8  bytes of locals out of registers, kernel
functions have the restriction that there be no more than 8 bytes of
locals and you don't do anything that requires a frame (e.g. take the
address of a local).

The NAKED attribute is unfortunate in that it doesn't set up the
frame.  What is really needed is a "No Return" attribute so that
registers are not saved by GCC and the RTOS can handle context
switches without chewing up so much SRAM.

In assembler, where I control everything, AvrX runs comfortably in a
2313. With GCC AvrX needs an 8515 to do useful work since each context
needs, at a minimum, 35 bytes before any that GCC uses.

Note: as soon as you call a function (from your task or interrupt
handler) you get a valid frame and GCC pushes a bunch of registers
that are likely unused at the top (naked) level.

Cheers!

----- Original Message -----
From: "Marko Panger" <address@hidden>
To: <address@hidden>
Cc: "AVR Mailing List" <address@hidden>
Sent: Tuesday, July 09, 2002 2:23 PM
Subject: Re: [avr-gcc-list] GCC guru question (progue/epilogue
sequences)


HI again !

My INT handler looks like

RTOS_INTERRUPT(VECTOR_NUMBER)        /*naked attribute*/
{
    uint16_t var1;        /* this is a variable used by the user*/

    IntEnter();        /*here I save registers*/

    User code bla...bla;
    bla;
    bla;

    IntExit();        /* here I    swap context*/
}

As you can se I am doing a sort of pro/epi by my own. I should save
the entire context in IntEnter() because the user code could make an
another task ready to run, so the context is swapped in IntExit(). Now
the extra code generated by the compiler is redundant and more - it
wastes ram (PUSH, POP).

After IntEnter() Y and the SP are the same. My prologue sequence
should be:

in r28, SPL
in r29, SPH
sbiw r28, NUMBER OF BYTES REQUIRED BY USER LOCAL VARIABLES
out    SPL, R28
out    SPH, R29

----------------------------------------------------------------------
---------------
Anyway I have just solved the problem with a little trick:

RTOS_INTERRUPT(VECTOR_NUMBER)        /*naked attribute*/
{
    IntEnter();        /*here I save registers*/

    UserFunction();

    IntExit();        /* here I    swap context*/
}

void UserFunction(void);

Thanks for helping me !

Regards,
Marko P.



----- Original Message -----
  From: Marc R. Wetzel
  To: Marko Panger ; AVR Mailing List
  Sent: Tuesday, July 09, 2002 11:05 PM
  Subject: RE: [avr-gcc-list] GCC guru question (progue/epilogue
sequences)


  Did you had a look at ethernut? Something very useful is already
there. This project
  looks always for contributors.

  To your question:

  if an interrupt occurs your kernel has no chance to save any
registers. so the pro/epilogue is very helpful.
  your kernel will be interrupted and it is not clear -- when.

  Look at the ethernut-way: it uses a kind of wrapper for any possible
interrupt source available.
  and if you attach an real interrupt-handler to it, it will be
called.

  Hopefully I understand your question right...

  /Marc
    -----Original Message-----
    From: address@hidden
[mailto:address@hidden Behalf Of Marko Panger
    Sent: Tuesday, July 09, 2002 10:56 PM
    To: AVR Mailing List
    Subject: [avr-gcc-list] GCC guru question (progue/epilogue
sequences)


    Hi all !!

    I realy hope somebody could answer the following question. I will
try to explain the situation and my problem.

    So... I am writting a RTOS and I would like to have interrupt
handlers written in C language. If I define a function as a INTERRUPT
or SIGNAL the compiler generates the prologue/epilogue sequence
automaticaly. But I don't need this. Registers saving an restoring is
managed by my kernel.

    There is another possibility. I could decalre the interrupt
function with NAKED attribute. This would be fine for me since the
compiler doesn't generate the prologue/epilogue sequence. But here
comes the problem. What will happen if I decalare some local variables
wich are putted on the stack ?

    I was wondering if is possible to write my own pro/epi sequences ?
How will I now how many variables are putted onto the stack ?





avr-gcc-list at http://avr1.org