https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126771

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
So eliding the conversion during pattern recog is difficult, IIRC with aarch64
SVE the mask mode for V2SI is different from that for V2DI, and
vectorizable_condition requires a match in lanes of the condition and the
result (and the result and the true/false operands being the same).
It shouldn't be an issue for BB vectorization, but for loop vectorization
the conversion is the more natural place for "unpacking".

Reply via email to