[Bug target/88892] [8/9 Regression] Double-to-float conversion uses wrong rounding mode when followed by memcpy

[segher at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88892

--- Comment #8 from Segher Boessenkool <segher at gcc dot gnu.org> ---
So we should use the current rounding mode for any float_trunc?  So we can use
float store instructions for it only with some fast-math option?