[Bug tree-optimization/91384] [7/8/9/10 Regression] Compare with negation is not eliminated

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91384

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Hmpf.  So the following was not optimized ever:

void foo (void);
void bar (void);

int
test (int a)
{
  int r = -a;

  if (a == 0)
    foo ();
  else
    bar ();

  return r;
}

Note this is also about canonicalization if you consider both jump forms
appearing in the source and the ability to thread them.

But yes, the CC re-using form is better (in case r is not dead otherwise).

On GIMPLE this shows we lack modeling of CCs.  I'm not sure doing

 _1 = __IFN_NEGATE_WITH_CCZ (a_2);
 r_3 = __real _1;
 _4 = __imag _1;
 if (_4 != 0)
   foo ();
 else
   bar ();

is the way to go though (that would be possible with current infrastructure).
A "nicer"/leaner way would be to generalize SSA projections so we do not
have to abuse COMPLEX and could allow

 _1 = __IFN_NEGATE_WITH_CC (a_2);
 r_3 = __value _1;
 if (__CCZ _1 != 0);
   foo ();
 else
   bar ();

but then doing this on GIMPLE is of course prone to CC spilling.

Note "fixing" the above testcase on GIMPLE is as hard as fixing it on RTL.