On Mon, 7 Nov 2016, Jakub Jelinek wrote:
> On Mon, Nov 07, 2016 at 01:17:25PM +0100, Richard Biener wrote:
> >
> > The following fixes phiopt to not introduce undefined behavior
> > in its abs replacement code in case we negate only positive values
> > in the original code.
> >
> > Bootstrapped and tested on x86_64-unknown-linux-gnu, applied to trunk.
> >
> > Richard.
> >
> > 2016-11-07 Richard Biener <rguent...@suse.de>
> >
> > PR tree-optimization/78228
> > * tree-ssa-phiopt.c (abs_replacement): Avoid introducing
> > undefined behavior.
> >
> > * gcc.dg/tree-ssa/phi-opt-15.c: New testcase.
> >
> > Index: gcc/tree-ssa-phiopt.c
> > ===================================================================
> > --- gcc/tree-ssa-phiopt.c (revision 241891)
> > +++ gcc/tree-ssa-phiopt.c (working copy)
> > @@ -1453,6 +1453,14 @@ abs_replacement (basic_block cond_bb, ba
> > else
> > negate = false;
> >
> > + /* If the code negates only iff positive then make sure to not
> > + introduce undefined behavior when negating or computing the absolute.
> > + ??? We could use range info if present to check for arg1 == INT_MIN.
> > */
>
> Perhaps just
>
> > + if (negate
> > + && (ANY_INTEGRAL_TYPE_P (TREE_TYPE (arg1))
> > + && ! TYPE_OVERFLOW_WRAPS (TREE_TYPE (arg1))))
> {
> wide_int minv = TYPE_MIN_VALUE (TYPE_DOMAIN (TREE_TYPE (arg1)));
> if (!expr_not_equal_to (arg1, minv))
> return false;
> }
> ?
rather wi::min_value (TREE_TYPE (arg1), SIGNED) I guess. Didn't know
of expr_not_equal_to, seems to be only used from i386.c at the moment.
We can improve things on trunk but I'd prefer to be safe on the
branch(es).
Richard.
