Octane-Raytrace on Solaris/SPARC is showing a very large regression in the range of 25%.

The 25% regression is not always present. Part of the issue is that when we make multiple runs, the results shift between different modes. One of those modes is the original performance level. This may have started in b24, but that's not completely clear. However, the model nature of the results is causing this 25% shift in b25. If we understand whats causing the modality, we should be able to get performance back to the original performance level.

Note: the modality condition doesn't appear on x64 hardware.