Texture cache & memory writes

This is a conceptual question.
Suppose I have a texture reference bound to some linear pitch memory.
Furthermore, suppose I pass a pointer to this memory to a kernel.
In this kernel, I can read the memory through the texture reference and write to the memory through the raw pointer.
My question is can this work in a predictable (defined) way or is this just a bad thing to do? Does the texture cache become invalidated if a write occurs or could tex2D() return results inconsistent with the true data?

Nevermind, I answered this question by myself. The CUDA programming guide covers this in section