Re: call graph

[Carlo R. <ca...@al...>]

On Fri, Dec 19, 2003 at 04:18:16AM +0000, Philippe Elie wrote:
> Carlo I didn't understand your "log the top of stack in sample files",
> I don't see how it can work, can you explain it ?

When an interrupt occurs, you read
the memory contents of (%ebp) in order
to know what is the (return) address
of the caller - and that is the only
thing, currently, that is written to
the oprofile (along with the current
program counter (pc)).

Consider:

f3() { }
f2() { f3(); }
f1() { f2(); }

or

f3:
.LFB2:
        pushl   %ebp
.LCFI0:
        movl    %esp, %ebp
.LCFI1:
        popl    %ebp
        ret
.LFE2:


When an NMI occurs at the moment
that you are at .LFB2, then pc points
to f3 and (%ebp) points to f1 (this is
what causes us to think that f1 calls f3).
Same thing when you are at .LCFI0.

Therefore, you'd actually need the
contents of (%esp) too, "just in case".
Then later you can decide to insert
that as caller.

Now I see that you ALSO need ((%esp+1),
because that is what you need when you
are at .LCFI0.

Perhaps it is possible to do this in
the driver though.  The driver has to
get the stack trace, so it knows exactly
how it works on the current target.
In the above case (linux x86_32), it
could do the following:

- If the current pc points to an
  instruction that disassembles as
  'push %ebp', then add an escape
  sequence that adds the top-of-stack.

- If the current pc points to an
  instruction that disassembles as
  'movl %esp, %ebp', then add
  an escape sequence that adds
  the content of the stack one deeper
  (assuming therefore that there
  was one push since the last call).

The profile converter (ie opgprof)
can ignore these insertions when they
are not inside the prologue of a
function.

Disadvantage:

- driver needs to do two compares and
  two conditional jumps extra (but we
  can do branch prediction here: normally
  the pc will point to other code).
- This is architecture dependent, so
  its gonna be some work to implement.
  There will also be several ways that
  an epilogue can look like.

Advantage:
- because the pc normally points somewhere
  else, you don't need to add the escape
  sequences often.
- when there is no debug info, you can
  guess that every inserted escape indeed
  is inside a prologue and probably still
  be quite accurate.

-- 
Carlo Wood <ca...@al...>