[Bug target/127055] Missing vectorizations of round math functions

[mkretz at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127055

--- Comment #24 from Matthias Kretz (Vir) <mkretz at gcc dot gnu.org> ---
https://compiler-explorer.com/z/K3a374YcE shows that GCC does the same as
Clang:
copysign from argument to nextafter(.5, 0) -> add that to the argument -> trunc

The only difference is that Clang uses vroundps 11 and GCC uses vroundps 3. The
difference is the imm[3] bit which controls FE_INEXACT. The dependence on the
fenv here is only about signaling inexact. There's no dependency on the dynamic
rounding mode.