In a wireless ad hoc network with multihop transmissions and interference-limited link rates, can we balance power control in the physical layer and congestion control in the transport layer to enhance the overall network performance, while maintaining the stability, robustness, and architectural modularity of the network? We present a distributive power control algorithm that couples with the original TCP protocols to increase the end-to-end throughput and energy efficiency of the network. Under the rigorous framework of nonlinearly constrained optimization, we prove the convergence of this coupled system to the global optimum of joint power control and congestion control, for both synchronized and asynchronous implementations. The rate of convergence is geometric and a desirable modularity between the transport and physical layers is maintained. In particular, when the congestion control mechanism is TCP Vegas, that a simple utilization in the physical layer of the router buffer occupancy information suffices to achieve the joint optimum of this cross layer design. Both analytic results and simulations illustrate other desirable properties of the proposed algorithm, including robustness to channel outage and to path loss estimation errors, and flexibility in trading-off performance optimality for implementation simplicity.