If the computation of the composite fixed-point transform for RandR
overflows at any point, take the resulting floating point transform and
scale that back to fit in a fixed point matrix. This ensures that a matrix
will always be available, although perhaps at reduced precision. Someday we
should add floating point matrices to Render.
Signed-off-by: Keith Packard <keithp@keithp.com>
62fc98cb88