Re: [RFA] Arm/Thumb tweak for generic_dummy_frames

[Michael Snyder <msnyder@redhat.com>]

Richard Earnshaw wrote:
> 
> msnyder@cygnus.com said:
> >  This is a corner case that Andrew missed when he did the transition
> > to generic dummy frames.
> 
> > 2002-05-21  Michael Snyder  <msnyder@redhat.com>
> 
> >       * arm-tdep.c (arm_frame_chain): Recognize dummy-frame as a
> >       special case that does not indicate a transition from arm
> >       to thumb or vice versa.
> 
> I can't (easily) work out from this what was wrong, and how you've fixed
> it.  Could you provide some more detailed analysis? 

Yeah, but not without drawing some pictures.   ;-)

OK, here's the context.  You're debugging in main, you put a breakpoint
at foo, and then you call foo, thusly:

  (gdb) print foo()

(Let's say that main and foo are both thumb).
You hit the breakpoint, and now your runtime stack looks something like
this:

  #0  foo
  #1  dummy frame created by gdb
  #2  main

At this point, GDB has already constructed a frame_info for foo, 
doing all the prolog analysis and computing the saved registers.

So now you say "backtrace".

So gdb callse arm_frame_chain to find and start analyzing frame #1.

The first thing arm_frame_chain does is ask whether this frame
(frame #0) is a dummy frame.  No, it's not, so we continue.

Then we compute the return address of frame #0, which for a 
generic dummy frame is always the entry point address ("_start"
or crt0 or whatever).

Next, we compare the return address to the current frame's address,
to see whether we've made a transition between arm code and
thumb code.  But this isn't a valid question if the caller_pc
is actually gdb's dummy frame, because the situation
(caller_pc == _start) is artificial.  _start may be arm code, 
and frame #0 may be thumb code, but we weren't actually called
from _start, so we don't want to act as if we were.

Now as to your second question --

> Why would a dummy frame never involve a transition between 
> ARM and Thumb state?

Well, that's not what this change is actually about -- this change
is about thumb-calling-thumb (with a bogus "arm" dummy frame in
the middle).  However, Andrew and I discussed what would happen
if you were in arm code and you tried to have GDB call a thumb
function (or vice versa), and decided that, well, we haven't
really handled that situation either before or after this change.
Someday we need to address it.

I guess I also should have mentioned that I ran testsuites 
before and after.  This change definitely fixes some failures,
and it doesn't introduce any new ones.