[PATCH] phiopt: Optimize (x <=> y) cmp z [PR94589]

[Jakub Jelinek]

On Wed, May 05, 2021 at 01:45:29PM +0200, Marc Glisse wrote:
> On Tue, 4 May 2021, Jakub Jelinek via Gcc-patches wrote:
> 
> > 2) the pr94589-2.C testcase should be matching just 12 times each, but runs
> > into operator>=(strong_ordering, unspecified) being defined as
> > (_M_value&1)==_M_value
> > rather than _M_value>=0.  When not honoring NaNs, the 2 case should be
> > unreachable and so (_M_value&1)==_M_value is then equivalent to _M_value>=0,
> > but is not a single use but two uses.  I'll need to pattern match that case
> > specially.
> 
> Somewhere in RTL (_M_value&1)==_M_value is turned into (_M_value&-2)==0,
> that could be worth doing already in GIMPLE.

Apparently it is
  /* Simplify eq/ne (and/ior x y) x/y) for targets with a BICS instruction or
     constant folding if x/y is a constant.  */
  if ((code == EQ || code == NE)
      && (op0code == AND || op0code == IOR)
      && !side_effects_p (op1)
      && op1 != CONST0_RTX (cmp_mode))
    {
      /* Both (eq/ne (and x y) x) and (eq/ne (ior x y) y) simplify to
         (eq/ne (and (not y) x) 0).  */
...
      /* Both (eq/ne (and x y) y) and (eq/ne (ior x y) x) simplify to
         (eq/ne (and (not x) y) 0).  */
Yes, doing that on GIMPLE for the case where the not argument is constant
would simplify the phiopt follow-up (it would be single imm use then).

	Jakub