Re: [PATCH] MIPS: Add option to pass return address location to _mcount. Was: [PATCH -v4 4/9] tracing: add static function tracer support for MIPS

[Wu Zhangjin <wuzhangjin at gmail dot com>]

Hi,

On Sat, 2009-10-24 at 10:12 +0100, Richard Sandiford wrote:
> Thanks for the patch.
[...]
> > How about this patch, I think it does what you suggest.
> >
> > When we pass -pg -mmcount-raloc, the address of the return address 
> > relative to sp is passed in $12 to _mcount.  If the return address is 
> > not saved, $12 will be zero.  I think this will work as registers are 
> > never saved with an offset of zero.  $12 is a temporary register that is 
> > not part of the ABI.
> 
> Hmm, well, the suggestion was to pass a pointer rather than an offset,
> but both you and Wu Zhangjin seem to prefer the offset.  Is there a
> reason for that?  I suggested a pointer because
> 
>   (a) they're more C-like
>   (b) they're just as quick and easy to compute
>   (c) MIPS doesn't have indexed addresses (well, except for a few
>       special cases) so the callee is going to have to compute the
>       pointer at some stage anyway
> 

Agree with you.

if not with -fno-omit-frame-pointer, we also need to calculate the frame
pointer, and then plus it with the offset. with pointer, we can get it
directly, but it may need a more instruction(lui..., addiu...) for
loading the pointer. of course, at last, the pointer will save more time
for us :-)

so, David, could you please use pointer instead? and then I will test it
asap(cloning the latest gcc currently). thanks!

> (It sounds from Wu Zhangjin's reply like he'd alread suggested the
> offset thing before I sent my message.  If so, sorry for not using
> that earlier message as context.)
> 

It doesn't matter, Seems at that time, you were not added in the CC
list, but added by David Daney later.

> > +  if (TARGET_RALOC)
> > +    {
> > +      /* If TARGET_RALOC load $12 with the offset of the ra save
> > +	 location.  */
> > +      if (mips_raloc_in_delay_slot_p())
> > +	emit_small_ra_offset = 1;
> > +      else
> > +	{
> > +	  if (Pmode == DImode)
> > +	    fprintf (file, "\tdli\t%s,%d\t\t# offset of ra\n", reg_names[12],
> > +		     cfun->machine->frame.ra_fp_offset);
> > +	  else
> > +	    fprintf (file, "\tli\t%s,%d\t\t# offset of ra\n", reg_names[12],
> > +		     cfun->machine->frame.ra_fp_offset);
> > +	}
> > +    }
> 
> We shouldn't need to do the delay slot dance.  With either the pointer
> ((D)LA) or offset ((D)LI) approach, the macro won't use $at, so we can
> insert the new code immediately before the jump, leaving the assembler
> to fill the delay slot.  This is simpler and should mean that the delay
> slot gets filled more often in the multi-insn-macro cases.
> 
> Looks good otherwise, but I'd be interested in other suggestions for
> the option name.  I kept misreading "raloc" as a typo for "reloc".
> 

The same misreading to me, what about -mmcount-ra-loc? add one "-", or
-mcount-ra-location?

BTW: Just made dynamic function tracer for MIPS support module tracing
with the help of -mlong-calls. after some more tests, I will send it as
-v5 revision later. hope the -v6 revision work with this new feature of
gcc from David Daney.

Regards,
	Wu Zhangjin