This is a nice example: for an arbitrary numerical integration method, one can choose a (xm, ym) and r small enough so that this will fail. For a smoothly differentiable function, one could find the peak numerically (especially after a meaningful transformation) and work around that, or use MCMC, but not for this one.