This avoids triggering overflows in the filters, and avoids stray
test failures in the approximate functions on x86; due to rounding
differences, one implementation might overflow while another one
doesn't.
Signed-off-by: Martin Storsjö <martin@martin.st>