From investigation on the large picture macrobenchmark, the vast majority of time is spent in Display List alloc and realloc. This is somewhat expected, as the test itself is attempting to stress the recording format by drawing a large number of operations. Still, we could probably make this case a bit faster by growing the display list buffer a bit more quickly. If I'm reading dl_builder.cc correctly, we're only growing by 4096 bytes at a time, which is probably a bit too slow.

We shouldn't double, at least not without end as that is a good way to run out of memory. But using the currently allocated size as a hueristic to choose multiple page sizes would be good.