Re: [PATCH] Fix PR35607, another revenge of invariant addresses (ivopts)

[Zdenek Dvorak <rakdver at kam dot mff dot cuni dot cz>]

Hi,

> >  > > Yes.  address_invariant_p should be simply
> >  >
> >  > Like that one.
> >  >
> >  > Can we rely on TREE_INVARIANT on the base object of the reference
> >  > tree or shall we open-code that as well (hm, I wasn't able to
> >  > quickly come up with the correct thing here ;)).
> >  >
> >  > +   return TREE_INVARIANT (op);
> >
> >  return CONSTANT_CLASS_P (op) || DECL_P (op);
> 
> Actually recompute_tree_invariant_for_addr_expr uses the equivalent of
> 
>   if (DECL_P (op)
>       && (staticp (op)
>           || decl_function_context (op) == current_function_decl
>           || (TREE_CODE (op) == VAR_DECL
>               && DECL_THREAD_LOCAL_P (op))))
>     return true;
> 
>   return CONSTANT_CLASS_P (op);
> 
> so not all DECLs are invariant in the old sense.  Though I don't see
> why we shouldn't allow arbitrary DECLs here...

with code like

void foo(int n)
{
  int j, k;

  for (k = 1; k < n; k++)
    {
      int blah[k];

      for (j = 0; j < k; j++)
        {
	  int t;

	  bar (blah, &t)
	}
    }
}

&t probably is not invariant in the function.  So
decl_function_context (op) == current_function_decl does seem to make
sense.  Let's just copy the condition from
recompute_tree_invariant_for_addr_expr, that should be safe.

Zdenek