Re: egcs-current: Major optimizer bug

[Jeffrey A Law <law at cygnus dot com>]

  In message <19980630114738.H30376@noris.de>you write:
  > egcs-2.91.45 19980628 (gcc2 ss-980609 experimental) severely miscompiles
  > the following, quite simple, code:
  > 
  > >>>>>>>>>>>>>>>>
  > int dn_skipname(unsigned char *ptr, unsigned char *eom) {
  >     unsigned char *saveptr = ptr;
  > 
  >     if (ns_name_skip(&ptr, eom) == -1)
  > 	        return (-1);
  >     return (ptr - saveptr);
  > }
  > >>>>>>>>>>>>>>>>
  > 
  > The resulting assembly output (i386) looks like the following. As you can
  > see, it will always return zero, which is not at all what's intended.  :-(
Thanks.  Initial analysis indicates that it's a problem with the
ADDRESSOF optimizations.  It looks like we've got an invalid REG_EQUIV
note for a load of the stack slot contining "ptr".  The REG_EQUIV
note causes later optimizations to optimize away the load of
the slot after the call to ns_name_skip.

I've added this test to the testsuite.

Thanks,

jeff