While it may seem pointless, it the rouding mode is set to round towards
minus infinity then acos(-1) calculates atan2(sqrt(1.0 - 1.0), -1) the subtract
generates -0.0 which sqrt() preserves.
atan2(-0, -1) is -pi, but acos(-1) is expected to be +pi.
This might 'fix' the test failures seen in some environments, but they
are not failing due to an obvously incorrent x87 rounding mode.
