Bug#716635: gcc-4.8: Miscompilation at optimization -O1

[Aurelien Jarno]

tag 716635 - help
tag 716635 + upstream
forwarded 716635 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59137 
retitle 716635 gcc-4.8: Miscompilation at optimization -O1 on mips/mipsel
thanks

On Wed, Jul 10, 2013 at 04:28:46PM -0400, Camm Maguire wrote:
> 
> Package: gcc-4.8
> Version: 4.8.1-5
> Severity: important
> 
> The following code
> 
> type_of(x0)==t_cons
> 
> from 
> 
> DEFUNO_NEW("LISTP",object,fLlistp,LISP
>    ,1,1,NONE,OO,OO,OO,OO,void,Llistp,(object x0),"")
> 
> {
>       /* 1 args */
> 
>       if (x0 == Cnil || type_of(x0) == t_cons)
>               x0 = Ct;
>       else
>               x0 = Cnil;
>       RETURN1(x0);
> }
> 
> where type_of(x0) is fully expanded as
> 
> ({register object _z=(object)(x0); ((((ufixnum)(_z))>=0x80000000) ?
> t_fixnum : ((!(_z)->d.e || (((ufixnum)((*(object *)(_z))))>=0x80000000))
> ? (_z==((object)&Cnil_body) ? t_symbol : t_cons) : _z->d.t));})
> 
> incorrectly gives true when compiled at O1, and correctly gives false
> when compiled at -O0 (only on mips).
> 
> The following is a gdb session followed by the full preprocessed source.
> 

Attached is a reduced testcase. The problem does appear at -O1, but not
at -O0 nor -O2. It is reproducible with GCC 4.7, 4.8 or snapshot, but
not with 4.6 or older. I have reported it upstream as issue #59137.

-- 
Aurelien Jarno                          GPG: 1024D/F1BCDB73
aurel...@aurel32.net                 http://www.aurel32.net

struct lispstruct
{
    int e;
    int t;
};

struct lispstruct Cnil_body;
struct lispstruct Ct_body;
int nvalues;

struct lispstruct * fLlistp (struct lispstruct *x0)
{
  if (x0 == &Cnil_body || (((unsigned long) x0 >= 0x80000000) ? 0 : (!x0->e ? (x0 != &Cnil_body) : x0->t)))
    x0 = &Ct_body;
  else
    x0 = &Cnil_body;
  nvalues = 1;
  return x0;
}

int main ()
{
  return fLlistp((struct lispstruct *) 0xa0000001) == &Cnil_body;
}