Re: [PATCH -v4 9/9] tracing: add function graph tracer support for MIPS

[Richard Sandiford <rdsandiford at googlemail dot com>]

Wu Zhangjin <wuzhangjin@gmail.com> writes:
> (Added linux-mips mailing list and the other people to the CC list.)
>
> On Thu, 2009-10-22 at 04:37 -0700, pajko wrote:
> [...]
>> > 
>> 
>> All this stuff can be avoided having PROFILE_BEFORE_PROLOGUE enabled in GCC
>> (gcc/config/mips/mips.h), like it is done one several other architectures.
>> Some of them even require it to be set for a working _mcount. 
>> Putting the call of _mcount before the function prologue should make no harm
>> (it's working for me), and this way RA can be hooked for function graph
>> tracing
>> before it is saved to stack in the function prologue. Thus there will be no
>> difference between leaf and non-leaf functions.
>
> Good idea! Seems PROFILE_BEFORE_PROLOGUE is commented by default in
> gcc/config/mips/mips.h of gcc 4.4:
>
> /* #define PROFILE_BEFORE_PROLOGUE */
>
> if we enable this macro, the handling will be the same to non-leaf and
> leaf function, so, David's patch to gcc is not need again.

Defining PROFILE_BEFORE_PROLOGUE isn't correct for abicalls code,
because "jal _mcount" is a macro that loads _mcount from the
GOT into $25.  We don't have access to $28 at the beginning of
the function, and we mustn't clobber the incoming value of $25.
So we could only make this change for non-abicalls code.

It's then a choice between (a) having new non-abicalls-specific
behaviour or (b) going with David's patch.  The advantage of
(a) is that the linux code is slightly simpler.  The disadvantage
is that it makes the _mcount interface differ between -mabicalls
and -mno-abicalls.  And IMO the disadvantage outweights the advantage.
If this new behaviour is useful for linux, it could easily be useful
for userspace too.  And with the new PLT support, non-shared abicalls
code is supposed to be link-compatible with non-abicalls code.

I think David's patch is the way to go.

Richard