Re: bad dead code detection ?

[Fergus Henderson <fjh at cs dot mu dot OZ dot AU>]

On 25-Jun-2003, David Jobet <david.jobet@free.fr> wrote:
> I try to write an exception handling library, and I'm forced to play with gcc
> inline assembly.

I don't see why.  For an example of an exception handling
library which does not use any inline assembly, see
<http://www.cs.mu.oz.au/~fjh/CXCPT/>.

> The trick is to register address of local labels in a global static table. Later
> when an exception is launched, I manage to jump to this label to do clean up.
> The problem is gcc detects the label cannot be reached and dead code elimination
> completely remove the exception handling code from the executable when compiled
> with -O3.
...
> I'm wondering if it's a bug, or if there is a way to prevent this.

It's not a bug... but there are ways to prevent it.

To prevent dead code elimination from eliminating a label,
you need to do two things:

1.  Make sure that the address of each label is taken *and appears to
    GCC to be used*, e.g. by storing it in a volatile global variable,

	extern void *volatile volatile_global;
	...
	volatile_global = &&label;

    or passing it to a volatile asm statement.

	__asm__ __volatile__("" : : "g"(&&label))

    The latter approach is better since it doesn't generate any code.

    Of course this use of &&label needs to occur in *reachable* code,
    e.g. at the start of the function.

2.  You also need to make sure that the function contains at least one
    reachable jump to an unknown-to-GCC location, e.g. by putting
    the following at the start of the function, immediately after the
    above-mentioned code to take and apparently use the addresses of labels.

    For example, add

	goto *identity_function(&&start);
	start:

    where identity_function() is a function whose body is not visible to GCC
    that just returns its argument.

    I suppose that

	volatile_global = &&start;
	goto *volatile_global;

    ought to have a similar effect too, although I haven't tested that.

> When compiled without -03, the code is not removed, but the label is
> moved from the *good* location to the return label.

See the thread "equality and label addresses" starting at
<http://gcc.gnu.org/ml/gcc/2003-01/msg01519.html>, which is
at least relevant, though I'm not sure it completely addresses
your concern here.
 
-- 
Fergus Henderson <fjh@cs.mu.oz.au>  |  "I have always known that the pursuit
The University of Melbourne         |  of excellence is a lethal habit"
WWW: <http://www.cs.mu.oz.au/~fjh>  |     -- the last words of T. S. Garp.