Bug 1358149 comment 55 indicates that transcendental function evaluation is having a significant effect on performance. The "Simple source without resampling (Stereo and positional)" performance benchmark differs from "Simple source without resampling (Stereo)" only in that it has a PannerNode. Comparison of the results implies about 20% of the time is attributable to the PannerNode.

Removing azimuth and elevation trigonometry from equalpower panning shows a 6-10% total time speed-up on WINNT 32- and 64-bit "Simple source without resampling (Stereo and positional)", which is a big chunk of the time attributable to the PannerNode. Linux numbers are slower, implying more other overheads, and speed-ups are around 3.5% to 5%.

The benchmark uses a fixed position. Performance improvements would be much more significant when these calculations are performed 128x as often for apps automating the position.