https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118727
--- Comment #9 from Xi Ruoyao <xry111 at gcc dot gnu.org> --- (In reply to Tamar Christina from comment #8) > That change was made in g:aec90c8bf30cbd66e4febae2c78622dc217f3918, but no > real explanation as to why. > > patt_40 = (signed char) a.0_4; > patt_41 = SAD_EXPR <patt_40, b_16, r_23>; > > would be ok if it was > > patt_41 = SAD_EXPR <a.0_4, patt_b_16, r_23>; > > where a.0_4 is > > a.0_4 = (unsigned char) a_14; > patt_b_16 = (unsigned char) b_16 > > and the point of the vect_look_through_possible_promotion is to increase > your VF by removing any unneeded intermediate promotions. > > On AArch64 we produce the right code, so the question is why does this > diverge from what you get. > > It seems to me that it's not the promotion that is wrong, but rather > > /* Get the inputs to the SAD_EXPR in the appropriate types. */ > tree sad_oprnd[2]; > vect_convert_inputs (vinfo, stmt_vinfo, 2, sad_oprnd, half_type, > unprom, half_vectype); > > is not doing the right thing for you. This call is what's responsible for > getting the inputs in the same type. If I read it correctly it just gets the inputs into half_type here. So if half_type is wrong the behavior will be wrong. And "half_type = unprom[0].type;" so it's from vect_look_through_possible_promotion.
