No penalty no tears: Least squares in high-dimensional linear models