Re: Subtle bug with GCC optimization?

[Gary Houston <ghouston at arglist dot com>]

> From: Mikael Djurfeldt <mdj@mdj.nada.kth.se>
> Date: 14 Aug 2000 00:38:29 +0200

> Marius Vollmer <mvo@zagadka.ping.de> writes:
> 
> > after changing something in scm_ithrow, Guile crashed for me when
> > there was no matching tag on the wind list.  It crashed only when
> > compiled with -O2 on i386.  I traced it to the following code:
> > 
> >     for (winds = dynwinds; SCM_NIMP (winds); winds = SCM_CDR (winds))
> >       {
> >         if (! SCM_CONSP (winds))
> > 	  abort ();
> > 
> >         dynpair = SCM_CAR (winds);
> >         ...
> >       }
> > 
> > which is compiled to this code fragment (see below for complete
> > listings):
> > 
> > 	movl %ecx,%ebx
> > 	andl $6,%ebx                     ;; SCM_NIMP
> > 	movl (%ecx),%edx                 ;; SCM_CAR  -> crash!
> > 	jne .L19
> > 	testb $1,%dl                     ;; SCM_CONSP
> > 	je .L22
> > 	call abort
> > 	.p2align 4,,7
> > 
> > As you can see, the code checks whether winds is NIMP and takes it CAR
> > regardless (probably because it is needed to check for CONSP).
> > 
> > Is GCC allowed to do that?
> 
> Wow!  8-o
> 
> No, I don't think so.  First I thought that if we have the SCM type
> defined as a void *, GCC might have the right to assume that it's
> valid.
> 
> But this is not the default.
> 
> My guess it that it is a GCC bug.

It looks like the same GCC bug we ran into a few months ago: try adding
-fno-gcse.

Search for gcse at:

http://sources.redhat.com/ml/bug-guile/

or search for ghouston at:

http://gcc.gnu.org/ml/gcc-bugs/