Our benchmarks of the new R300 register allocator last month showed some promising results in terms of performance improvements where shaders with lots of scalar values are used, as the register count is down significantly.
What Tom's been doing over the past month is supporting for-loops and fixing various outstanding bugs. It also hooks into the recently merged floating-point texture additions.
He mentions this work in this blog post. The improved R300 register allocator for now isn't yet merged into mainline Mesa but is living in his personal Git repository for those interested in testing it with R300 through R500 (Radeon X1000 series) graphics processors.