LLVMpipe leverages the Low-Level Virtual Machine (LLVM) for optimizations and it -- along with Gallium3D itself -- has been in development for some time.
In this blog post by Zack Rusin he details the work that he's done recently on Gallium3D's LLVMpipe with the geometry processing pipeline. Below is the best part.
Anyway, with the new geometry paths the improvements are quite substantial. For vertex processing dominated tests (like geartrain which went from 35fps to 110fps) the improvements are between 2x to about 6x, for cases which are dominated by fragment processing it's obviously a lot less (e.g. openarena went from about 25fps to about 32fps). All in all llvmpipe is looking real good.