Current CPUs like Core duo or i7 are quite redundant, they can perform the same operations in various ways. Such ways use different amounts of electricity. Even stripped down CPUs like Atom have some redundancy. In very big data centres, like Google ones, the limiting factor is often the amount of electricity used by the very large number of computers present. Mobile phones, ebook readers, and similar gadgets often have powerful 32 bit CPUs with many MB of RAM, but for them it's important to maximize the number of hours they can run with a battery charge. So in theory it can be added a "-Oe" compilation argument to LLVM that tells the compiler to compile the code trying to minimize the amount of electricity the running program will use on the specified CPU. For example in -Oe mode the compiler can do the opposite of what I've written in bug 6209, and use the FP stack to store a single float value, saving the power needed to store an move around 128 bits in a SSE register.
Almost all computer literature shows that the best way to save power (on mainstream cpus that don't have super crazy power features) is to execute code faster so that the chip can enter a low power state sooner. Faster = lower power.