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

[David Daney <ddaney@caviumnetworks.com>]

Steven Rostedt wrote:
> On Wed, 2009-10-21 at 09:37 -0700, David Daney wrote:
> 
>> There is no deterministic way to identify MIPS function prologs.  This 
>> is especially true for leaf functions, but also for functions with 
>> multiple return sites.
>>
>> For certain GCC versions there may be a set of command line options that 
>> would give good results, but in general it is not possible.  Attempts at 
>> fast backtrace generation using code inspection are not reliable and 
>> will invariably result in faults and panics when they fail.
> 
> Thanks for the update.
> 
> We can easily protect against panics, since we do fault protection
> within the code (although currently it will panic on fault, but we can
> fix that ;-). We can limit the search to a couple of 100 instructions,
> as well as fail on first panic.
> 
> But are you sure that when compiled with -pg, that GCC does not give a
> reliable prologue. Things are different when GCC is compiled with -pg,
> it may indeed always have something that we can flag.
> 
> We could also add other tests, like the subtraction of the stack too.
> 

I have not used -pg, so I don't know for sure, I think all it does is 
add the calls to _mcount.  Someone could investigate 
-fno-omit-frame-pointer, with that you may be able to use:

    move    s8,sp

To identify function prologs, but it would still be ad hoc, as modern 
versions of GCC will reorder instructions in the prolog for better 
scheduling.

David Daney