Re: A patch for linux 2.1.127

[Bernd Schmidt <crux at pool dot informatik dot rwth-aachen dot de>]

On Sun, 8 Nov 1998, Linus Torvalds wrote:
> On Sun, 8 Nov 1998, Richard Henderson wrote:
> > 
> > Unfortunately, those patches are way too massive to go back to the
> > 1.1 branch.  For that I don't really know what to do.
> 
> How about something simple like:
> 
> 	if (SMALL_REGISTER_SET)
> 		never_inline_functions_unless_the_user_asked_for_it();
> 
> which means that even with -O6 you would not inline functions unless they
> were marked inline. 
> 
> Note that this is not just a workaround for a bug.

It's not even a workaround for a bug.  Function inlining has nothing to
do with the problem.  The following testcase (gcc.dg/clobbers.c in the egcs
testsuite) demonstrates that:

int main ()
{
        int i;
        __asm__ ("movl $1,%0\n\txorl %%eax,%%eax" : "=r" (i) : : "eax");
        if (i != 1)
                abort ();
        __asm__ ("movl $1,%0\n\txorl %%ebx,%%ebx" : "=r" (i) : : "ebx");
        if (i != 1)
                abort ();
        __asm__ ("movl $1,%0\n\txorl %%ecx,%%ecx" : "=r" (i) : : "ecx");
        if (i != 1)
                abort ();
        __asm__ ("movl $1,%0\n\txorl %%edx,%%edx" : "=r" (i) : : "edx");
        if (i != 1)
                abort ();
        __asm__ ("movl $1,%0\n\txorl %%esi,%%esi" : "=r" (i) : : "esi");
        if (i != 1)
                abort ();
        __asm__ ("movl $1,%0\n\txorl %%edi,%%edi" : "=r" (i) : : "edi");
        if (i != 1)
                abort ();
        return 0;
}

It might be possible to construct a workaround for egcs-1.1, but it would most
likely be rather ugly, and it doesn't help people who use gcc-2.7.2 or
egcs-1.0.3.  The simplest workaround is to avoid using clobbers on the x86.

Bernd