Unwinding past main

Unwinding past main

[Randolph Chung]

I'm seeing a stack unwinding problem on hppa. Seems like this should be
applicable to other targets too -- how do you solve this problem?

Take a simple test program:
====
int foo(void) { return 1; }
int main(void) { return foo(); }
====

gcc does some tail call optimizations on this (when compiled with -O2),
so that the generated code looks like:

00010518 <foo>:
   10518:       e8 40 c0 00     bv r0(rp)
   1051c:       34 1c 00 02     ldi 1,ret0

00010520 <main>:
   10520:       6b c2 3f d9     stw rp,-14(,sp)
   10524:       37 de 00 80     ldo 40(sp),sp
   10528:       4b c2 3f 59     ldw -54(,sp),rp
   1052c:       e8 1f 1f cd     b,l 10518 <foo>,r0
   10530:       37 de 3f 81     ldo -40(sp),sp
   10534:       08 00 02 40     nop

doing a backtrace from foo results in an infinite loop :) I see some
code in gcc that tries to determine when we are in the main() function,
and stops the backtrace. But here, foo() returns directly to the caller
of main() (__libc_start_main on a glibc system), so we try to continue
to unwind through glibc and hit some asm bits that cannot be unwound
properly.

what is the "proper" way to handle this?

thanks,
randolph
-- 
Randolph Chung
Debian GNU/Linux Developer, hppa/ia64 ports
http://www.tausq.org/

Re: Unwinding past main

[Randolph Chung]

> doing a backtrace from foo results in an infinite loop :) I see some
> code in gcc that tries to determine when we are in the main() function,
          ^^^ should be "gdb"
> and stops the backtrace. But here, foo() returns directly to the caller

randolph
-- 
Randolph Chung
Debian GNU/Linux Developer, hppa/ia64 ports
http://www.tausq.org/

Re: Unwinding past main

[Joel Brobecker]

> I'm seeing a stack unwinding problem on hppa. Seems like this should be
> applicable to other targets too -- how do you solve this problem?

Funny you would see that. I am just looking at the same sort of
problem, but on IA64. I have an infinite loop when unwinding the
call stack from a thread. I don't understand all the conditions
yet (I am completely new to IA64, and it's quite complex for me).

So you are right in saying that this can happen on other targets too.
In my case, the "main" is the NPTL thread start routine, and we try
to unwind past it. In some cases, we are lucky, because we end up
with 2 identical frames which cause an error, but in some others,
the frame base changes slightly from frame to frame and we loop...

-- 
Joel

Re: Unwinding past main

[Randolph Chung]

> Funny you would see that. I am just looking at the same sort of
> problem, but on IA64. I have an infinite loop when unwinding the
> call stack from a thread. I don't understand all the conditions
> yet (I am completely new to IA64, and it's quite complex for me).

oh, it used to do this on hppa too... we unwind to the clone call and
then it gets all confused.

> So you are right in saying that this can happen on other targets too.
> In my case, the "main" is the NPTL thread start routine, and we try
> to unwind past it. In some cases, we are lucky, because we end up
> with 2 identical frames which cause an error, but in some others,
> the frame base changes slightly from frame to frame and we loop...

yup. that's what i see too in this case.

randolph
-- 
Randolph Chung
Debian GNU/Linux Developer, hppa/ia64 ports
http://www.tausq.org/