[Bug tree-optimization/57742] memset(malloc(n),0,n) -> calloc(n,1)

["glisse at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57742

--- Comment #10 from Marc Glisse <glisse at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #9)
> (In reply to Marc Glisse from comment #7)
> > Also, for this testcase:
> > void* f(int n,double*d){
> >   int* p=__builtin_malloc(n);
> >   ++*d;
> >   __builtin_memset(p,0,n);
> >   return p;
> > }
> > I actually get a callback for the store in *d, which gcc believes might
> > alias :-(
> 
> Yeah well, either because of pass placement or because of points-to
> analysis being not context sensitive.

forwprop is run 4 times, it would be really unlucky if all 4 were badly placed.
I am surprised that in FRE/PRE stmt_may_clobber_ref_p_1 returns false for p and
q in the following code (while called from forwprop in the above testcase it
returns true). The main difference I can see is that the SSA_NAME_PTR_INFO of p
has vars_contains_global=1 when I test my code and 0 when I test the following
in FRE.

int g(int*a,int n,double*q){
  int*p=__builtin_malloc(n);
  p[2]=3;
  *q=5.;
  return p[2];
}


> Exact pattern matching of the CFG involved might be the easiest, plus
> manually implementing walk_aliased_vdefs by simply walking the use-def
> chain of the virtual operands from the memset operation to the malloc

That was also my conclusion, and it is a bit disappointing, since the
transformation shouldn't mind if for instance there is an unrelated loop
between malloc and memset. Better than nothing I guess :-/

> and checking stmt_may_clobber_ref_p_1 on the ao_ref_init_from_ptr_and_size
> ref.

Well, that won't help since stmt_may_clobber_ref_p_1 doesn't notice that
unrelated stores are unrelated, I first need to work on that.