[Mesa-dev] i965 batch rollback: no more prepare() vs emit().

[Eric Anholt]

I did do some testing by making libdrm claim a smaller aperture size
and thus trigger the rollback code, which caught one bug.  Generally,
piglit doesn't trigger rollback at all, which is a bit concerning from
a testing coverage standpoint.

For people reviewing, I think the interesting part to review is mostly
making sure that my assertions about the incremental changes of moving
things from prepare() to emit() are true.  A particular risk would be
if I moved something from prepare() time (before all emit() calls) to
emit() time (after some emit() calls) without noticing the dependency
by those emits() in between.

No noticeable performance impact from the change, though I was hoping
to drop CPU usage a bit.  Oh well.