Comment # 3 on bug 1127564 from Richard Biener

Hmm, this is a strange case of excess-precision and resulting differences
(x87 doubles have 80bit internal precision).  Usually the solution is
-fexcess-precision=standard but somehow that doesn't work for the
multiplication-and-conversion-to-integer sequence.  The "older" -ffloat-store
workaround works fine though.

Filed upstream.