Re: PR 33089: Canonicalizing a + b != 0 as a != -b causes missed optimization

[Hans-Peter Nilsson <hp at bitrange dot com>]

On Fri, 17 Aug 2007, Rask Ingemann Lambertsen wrote:

>    We start out with (eq (plus X A) 0):

>    I want to hear some opinions on which form we should choose as the
> canonical one. I'm proposing two forms:
>
> 1) (eq (neg X) Y) because the change to simplify_comparison will be simpler.
> 2) (eq (plus X Y) 0) because that's the least surpricing form and the same
> as used for other comparisons, avoiding needless insn pattern duplication.
>
> The arm and i386 back ends implement 1).
> The c4x, i386, frv, rs6000, s390, score and sparc back ends implement 2).
> The c4x back end also implements the (eq (X (neg Y)) form.
>
>    I'm voting for 2).

Me-too, for the reasons you mentioned and so this case wouldn't
seem special anymore:

>    Note that when combining two instructions, of which one is a comparison
> of the result of a previous instruction against zero and that result is
> still needed after the comparison, combine knows not to mess with the
> expression. What's special in this case is that we are really only
> interested in the comparison result, so combine goes ahead and does the
> simplifications. Unfortunately, the result is an expression which targets
> aren't likely to have an instruction pattern for.

brgds, H-P