https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118727
--- Comment #13 from Xi Ruoyao <xry111 at gcc dot gnu.org> --- For example for the original gcc.dg/pr108692.c: a.0_4 = (unsigned char) a_14; _5 = (int) a.0_4; b.1_6 = (unsigned char) b_16; _7 = (int) b.1_6; c_17 = _5 - _7; _8 = ABS_EXPR <c_17>; r_18 = _8 + r_23; becomes: patt_31 = .ABD (a.0_4, b.1_6); so when we get the ABD we already stripped the redundant (int) conversion here, note that the (unsigned char) conversion is really needed. And for my twisted test case: a.0_4 = (unsigned char) a_14; _5 = (unsigned int) a.0_4; _6 = (unsigned int) b_16; _7 = _5 - _6; c_17 = (int) _7; _8 = ABS_EXPR <c_17>; becomes: patt_38 = (unsigned short) a.0_4; patt_35 = (signed short) b_16; patt_36 = (signed short) patt_38; patt_33 = .ABD (patt_36, patt_35); To me this is already the "optimal" way. Correct me if I'm wrong, but anyway even if there's a redundant conversion it should be stripped in vect_recog_abd_pattern instead of vect_recog_sad_pattern.
