[lno] [RFC] if-conversion and auto vectorizer

[Richard Henderson]

On Mon, Mar 15, 2004 at 10:01:29PM -0500, Diego Novillo wrote:
> > Well, *consist* of GIMPLE scalars.  According to is_gimple_condexpr
> > we can have "x rel-op y" as well.
> > 
> Yes.  Point being, that such a predicate can't have side effects.  Or
> you only want to accept boolean scalars there?

Not at all.

One thing that might should be checked before marking various 
variables for ssa-renaming, however, is that you get

	if (x_1 < y_1)
	  x_2 = a, y_2 = b;
	x_3 = PHI(x_1, x_2)
	y_3 = PHI(y_1, y_2)
 ==>
	x_3 = (x_1 < y_1 ? a : x_1)
	y_3 = (x_1 < y_1 ? b : y_1)

Note that x_1 outlives the creation of x_3, which is new to 
this transformed code.  This mere fact might argue for the
creation of a boolean temporary for the comparison.


r~