It's been discussed in a bunch of other threads, but the 10,000 foot view is "it's too early to tell". There are probably a half dozen different IRs that are or will be on the table in the next year.
TGSI is probably the safest bet, particularly if Francisco's additions for compute are accepted and work out generally OK. LLVM IR is up there as well, but is a more natural fit for (scalar) compute than for (short vector) graphics, particularly when dealing with short vector or vliw hardware.
The original Gallium3D proposal envisioned using TGSI as the IR but then converting it to LLVM IR for optimization by LLVM before generating hardware instructions. That made sense to everyone but got a bit awkward when most of the OpenCL front ends *above* TGSI ended up using LLVM IR as well.