Is regs_ever_live valid in prologue? (SPARC quirk?)

["Antonio Cunei" <ack at cunei dot com>]

Hi, I am relatively new to GCC, and I need some advice,
or some advice regarding where to ask for advice.
Thank you for your suggestions! :)

The question, in a nutshell, is: is regs_ever_live[]
supposed to be valid during the prologue generation?

I am encountering, using the SPARC back end, cases
when some registers are marked as live when in fact
they are unused. Details are appended at the end of
this message.

Should I use a different method to extract the used
registers? Help! :)

---

I am also having troubles extracting, while in the
prologue, the machine mode of all the parameters
passed in registers. I used the following:

  for (arg=DECL_ARGUMENTS(current_function_decl);
                     arg!=0; arg=TREE_CHAIN(arg))
  {
    rtx parm=DECL_INCOMING_RTL(arg);
    int regn;

    if (REG_P(parm)) {
      enum machine_mode regmode = GET_MODE (parm);
      regn=REGNO(parm); // names used in the body

      // do something with regn and regmode
    }
  }

Would you use a different technique? This code doesn't
work 100% of the times, for me...


Thank you!!!! :)
 Toni
_______________________________________________________________________
 _____  ___  _  _ ___
|_   _!/ _ \| \| |_ _|      Antonio Cunei - University of Glasgow
  | | | (_) | .` || |   -   Room F152, Dept. of Computing Science
  !_!  \___/!_|\_!___|      17 Lilybank Gardens, Glasgow G12 8QQ


------------------------------------------------------------
Details of the anomaly:

compiling with -O3 the following, using the snapshot
20020422 (but I tried 3.0.4 as well)):

int correct(int a,int b)
{
 return a-b-a-b;
}

int incorrect(int a,int b)
{
 return b-a-b-a;
}


The results are as follows:

correct:
regs_ever_live:  %g1 %o0 %o1

        sub     %g0, %o1, %g1
        retl
        sub     %g1, %o1, %o0

Registers %o0 and %o1 are used for the two parameters,
so regs_ever_live is correct.
------
incorrect:
regs_ever_live:  %g1 %o0 %i1 <<<<<

        sub     %g0, %o0, %g1
        retl
        sub     %g1, %o0, %o0

%i1 appears in regs_ever_live instead of %o1 (a wrong
register window adjustment, maybe?)

------------------------------------------------------------