[PATCH] Make PRE PHI insertion use an already available PHI

[Richard Guenther]

On Thu, 21 Aug 2008, Daniel Berlin wrote:

> On Thu, Aug 21, 2008 at 10:14 AM, Richard Guenther <rguenther@suse.de> wrote:
> >
> > This makes us no longer rely on a second DOM pass for
> > gcc.dg/builtins-20.c.  The case in question is
> >
> > extern double cos(double);
> > void test2(double x, double y)
> > {
> >  if (cos(y<10 ? x : -y) != cos(y<10 ? x : y))
> >      link_error ();
> > }
> >
> > which we don't fold directly but have to deal with during optimization.
> > PRE figures out that all is equivalent but produces
> >
> > <bb 4>:
> >  # x_1 = PHI <x_5(D)(7), y_3(D)(3)>
> >  # D.1955_15 = PHI <D.1955_14(7), D.1955_7(3)>
> >  # prephitmp_22 = PHI <D.1955_14(7), D.1955_7(3)>
> >  D.1960_10 = prephitmp_22;
> >  if (D.1955_15 != D.1960_10)
> 
> Actually, this makes no sense to me.
> If you have a phi node already for the value, it should have detected
> it as a full redundancy, not a partial one.

It only gets a fully redundant value after PHI-translation, but we
do not look up the translated PHI nor do we seed the VN with that
result (so you could argue again that SCCVN should be better at
value-numbering PHIs).  The situation before PRE is:

test2 (double x, double y)
{
  double D.1960;
  double D.1955;

<bb 2>:
  if (y_3(D) >= 1.0e+1)
    goto <bb 3>;
  else
    goto <bb 7>;

<bb 3>:
  D.1955_7 = cos (y_3(D));

<bb 4>:
  # x_1 = PHI <x_5(D)(7), y_3(D)(3)>
  # D.1955_15 = PHI <D.1955_14(7), D.1955_7(3)>
  D.1960_10 = cos (x_1);
  if (D.1955_15 != D.1960_10)
    goto <bb 5>;
  else
    goto <bb 8>;

<bb 8>:
  goto <bb 6>;

<bb 5>:
  link_error ();

<bb 6>:
  return;

<bb 7>:
  D.1955_14 = cos (x_5(D));
  goto <bb 4>;

}

and cos (x_1) is available in both predecessors but of course computes
a different value in both, so we would insert a prephitmp to merge
those two values.  But this merge phi is available already.

Richard.