On a SNB i5-2500:

Speedups
========
         firefox-chalkboard  34284.16 -> 19637.40:  1.74x speedup
         swfdec-giant-steps    778.35 ->   665.37:  1.17x speedup
                  ocitysmap    485.64 ->   431.94:  1.12x speedup

Slowdowns
=========
           firefox-fishbowl  46878.98 -> 54407.14:  1.16x slowdown

That slow down is due to overhead of the increased number of calls to
pixman_image_composite32() (pixman_transform_point for analyzing the
source extents in particular) outweighing any advantage gained by
performing the rasterisation in a single pass and eliding gaps. The
solution that has been floated in the past is for an interface into
pixman to only perform the analysis once and then to return a kernel to
use for all spans.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>