This fixes a rare case where NaNs could exist inside Cycles. When certain invalid meshes were passed in, Cycles would try too normalize a zero length normal during its setup stage. While it does check against division by zero, it still returns a zero length normal and passes it on to the path tracing kernel. The kernel then operates under the assumption that normals are valid, and in the case of such a zero length normal, would eventually create NaNs that propagate through and result in black pixels.
As far as I know you should only be able to get a zero length normal with zero area faces, which you can't hit. But of course with float rounding issues it's not so simple, and this workaround for these rare cases seems ok.