Sampling with constrained parameter

What should I do if I sample models with parameters having flat priors? Because MCHMC needs to know the gradient and in this case the gradient will go to infinity...