Re: [patch] (-fstrict-overflow) optimize counted loops on signed iv

["Richard Guenther" <richard dot guenther at gmail dot com>]

On 2/15/07, Zdenek Dvorak <rakdver@atrey.karlin.mff.cuni.cz> wrote:
> Hello,
>
> > But of course the following is the problem:
> >
> > Index: fold-const.c
> > ===================================================================
> > *** fold-const.c        (revision 121992)
> > --- fold-const.c        (working copy)
> > *************** tree_swap_operands_p (tree arg0, tree ar
> > *** 6668,6676 ****
> >        && (TREE_SIDE_EFFECTS (arg0) || TREE_SIDE_EFFECTS (arg1)))
> >      return 0;
> >
> > !   if (DECL_P (arg1))
> >      return 0;
> > !   if (DECL_P (arg0))
> >      return 1;
> >
> >    /* It is preferable to swap two SSA_NAME to ensure a canonical form
> > --- 6668,6676 ----
> >        && (TREE_SIDE_EFFECTS (arg0) || TREE_SIDE_EFFECTS (arg1)))
> >      return 0;
> >
> > !   if (DECL_P (arg1) || TREE_CODE (arg1) == SSA_NAME)
> >      return 0;
> > !   if (DECL_P (arg0) || TREE_CODE (arg0) == SSA_NAME)
> >      return 1;
> >
> >    /* It is preferable to swap two SSA_NAME to ensure a canonical form
> >
> > I'll bootstrap and test it.
>
> makes sense, although it won't fix the problem in general (e.g.,
> 2 * a < 2 * a + 4 still would not be folded).

That's true.  I have no strong opinion on whether we should handle this
by canonicalization or by making the transformation handle both cases.

Richard.