[PATCH][match.pd] PR middle-end/66915 Restrict A - B -> A + (-B) to non-fixed-point types

[Richard Biener]

On Tue, 21 Jul 2015, Kyrill Tkachov wrote:

> 
> On 21/07/15 08:24, Richard Biener wrote:
> > On Mon, 20 Jul 2015, Kyrill Tkachov wrote:
> > 
> > > Hi all,
> > > 
> > > This patch fixes the PR in question which is a miscompilation of
> > > gcc.dg/fixed-point/unary.c on arm.
> > > It just restricts the A - B -> A + (-B) transformation when the type is
> > > fixed-point.
> > > 
> > > This fixes the testcase for me.
> > > Is this the right approach?
> > > 
> > > Bootstrap and test on arm and x86 running.
> > > 
> > > Ok if testing is clean?
> > Ok, but I think the fold-const.c code has the same issue, no:
> > 
> >        /* A - B -> A + (-B) if B is easily negatable.  */
> >        if (negate_expr_p (arg1)
> >            && !TYPE_OVERFLOW_SANITIZED (type)
> >            && ((FLOAT_TYPE_P (type)
> >                 /* Avoid this transformation if B is a positive REAL_CST.
> > */
> >                 && (TREE_CODE (arg1) != REAL_CST
> >                     ||  REAL_VALUE_NEGATIVE (TREE_REAL_CST (arg1))))
> >                || INTEGRAL_TYPE_P (type)))
> >          return fold_build2_loc (loc, PLUS_EXPR, type,
> >                              fold_convert_loc (loc, type, arg0),
> >                              fold_convert_loc (loc, type,
> >                                                negate_expr (arg1)));
> > 
> > ah, no.  The above only applies to float-type and integral-types.
> > 
> > Thus yes, your patch is ok.  Can you double-check the other pattern,
> > 
> > /* -(A + B) -> (-B) - A.  */
> > (simplify
> >   (negate (plus:c @0 negate_expr_p@1))
> >   (if (!HONOR_SIGN_DEPENDENT_ROUNDING (element_mode (type))
> >        && !HONOR_SIGNED_ZEROS (element_mode (type)))
> >    (minus (negate @1) @0)))
> > 
> > ?
> 
> Thanks, committed with r226028.
> I can add (FLOAT_TYPE_P (type) || INTEGRAL_TYPE_P (type)) to the condition.
> That would more closely mirror the original logic, right?
> That passes x86_64 bootstrap and aarch64 testing looks ok.

Yeah, that works for me, too.

Thanks,
Richard.