https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117195
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> --- Note this tree code (and optab) is useful only for reductions since it is left unspecified which of the 2x number of input lanes are accumulated to which output lane (even accumulating _all_ input lanes to only a single of the output lanes would be OK). Similar for DOT_PROD_EXPR and SAD_EXPR (lane_reducing_op_p). I do hope that at some point we kill those off in favor of optabs specifying how lanes are accumulated (even/odd, hi/lo). Investigation how the different ISAs behave here is required.
