I like the elegance of bitwise operations and use them sometimes in signal processing applications to avoid a performance hit, though I'm no expert. But I avoid them in any high level code because a) the compiler is likely going to optimize that anyway and b) to be as human-readable as possible, the code should clearly describe the objective rather than exploit the underlying mechanism. In other words if I want to multiply or divide then I prefer to write that than rely on the reader's knowledge of what bit-shifting does, and also so that the code in question would run properly on some hypothetical ternary computer or suchlike. As I'm mainly a hobby/ad-hoc programmer I can afford not to care about the probable overhead this imposes.
Great question, I hope this stimulates some discussion.