https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127055
--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Matthias Kretz (Vir) from comment #3)
> Right, for round I understand. But why is using the packed vround/vrndscale
> instruction (rather than the scalar one) gated on -fno-trapping-math? There
> is no observable difference between e.g.
>
> "test_v2df_ceil(double __vector(2))":
> vrndscalesd xmm1, xmm0, xmm0, 10
> vunpckhpd xmm0, xmm0, xmm0
> vrndscalesd xmm0, xmm0, xmm0, 10
> vunpcklpd xmm0, xmm1, xmm0
> ret
>
> and
>
> "test_v2df_ceil(double __vector(2))":
> vrndscalepd xmm0, xmm0, 10
> ret
t.C:15:188: note: ==> examining statement: _5 = __builtin_ceil (_4);
t.C:15:188: missed: function is not vectorizable.
t.C:15:177: missed: not vectorized: relevant stmt not supported: _5 =
__builtin_ceil (_4);
it's vectorized with -fno-trapping-math to
vect__5.116_9 = .CEIL (x_2(D));
and
vrndscalepd $10, %xmm0, %xmm0
ret
