The new code (since implementing the mechanism of punting accelerated surfaces
to system memory, based on read operations of that surface) mistakenly
copies out the destination image to the temporary buffer in _all_ cases
when we really only need those pixels if we are actually reading the dest
pixels or doing a partial-write to them.

You can see the performance degredation if you run
src/share/test/java2d/J2DBench/J2DBench on a build prior to last week
(such as build 79) and a build from this week. Opaque image copies
run much slower now than they did before. Note that you must run with
-Dsun.java2d.noddraw on win32 or setenv J2D_NO_DGA on Solaris to see
this effect.