llvmpipe: avoid most 64 bit math in rasterization

llvmpipe: avoid most 64 bit math in rasterization

The trick here is to recognize that in the c + n * dcdx calculations,
not only can the lower FIXED_ORDER bits not change (as the dcdx values
have those all zero) but that this means the sign bit of the calculations
cannot be different as well, that is
sign(c + n*dcdx) == sign((c >> FIXED_ORDER) + n*(dcdx >> FIXED_ORDER)).
That shaves off more than enough bits to never require 64bit masks.
A shifted plane c value could still easily exceed 32 bits, however since we
throw out planes which are trivial accept even before binning (and similarly
don't even get to see tris for which there was a trivial reject plane)) this
is never a problem.
The idea isnt't all that revolutionary, in fact something similar was tried
ages ago (9773722c2b09d5f0615a47cecf4347859474dc56) back when the values were
only 32 bit anyway. I believe now it didn't quite work then because the
adjustment needed for testing trivial reject / partial masks wasn't handled
correctly.
This still keeps the separate 32/64 bit paths for now, as the 32 bit one still
looks minimally simpler (and also because if we'd pass in dcdx/dcdy/eo unscaled
from setup which would be a good reason to ditch the 32 bit path, we'd need to
change the special-purpose rasterization functions for small tris).

This passes piglit triangle-rasterization (-fbo -auto -max_size
-subpixelbits 8) and triangle-rasterization-overdraw (with some hacks
to make it work correctly with large sizes) easily (full piglit as
well of course, but most tests wouldn't use triangles large enough to
be affected, that is tris with a bounding box over 128x128).
The profiler says indeed time spent in rast_tri functions is reduced
substantially, BUT of course only if the tris are large. I measured a 3%
improvement in mesa gloss demo when supersized to twice the screen size...

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>