avoid future implementation mischief. (Maybe not. Sometimes, general
distrust of theory leads people to distrust simple reasoning over times
from CPUs trying as hard as possible to mask DRAM latency via pre-fetch.)
(cherry picked from commit 196e747df1)