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

[fw at gcc dot gnu.org]

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

--- Comment #3 from Florian Weimer <fw at gcc dot gnu.org> ---
Started with r253210.  I don't think the new pattern is used in this case, so
maybe this is a pre-existing latent bug.