https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118727
--- Comment #19 from GCC Commits <cvs-commit at gcc dot gnu.org> --- The master branch has been updated by Xi Ruoyao <xry...@gcc.gnu.org>: https://gcc.gnu.org/g:da88e7027a34a44de84f6d8d5a96d262c29080a7 commit r15-7370-gda88e7027a34a44de84f6d8d5a96d262c29080a7 Author: Xi Ruoyao <xry...@xry111.site> Date: Sun Feb 2 21:22:36 2025 +0800 vect: Fix wrong code with pr108692.c on targets with only non-widening ABD [PR118727] With things like // signed char a_14, a_16; 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; An ABD pattern will be recognized for _8: patt_31 = .ABD (a.0_4, b.1_6); It's still correct. But then when the SAD pattern is recognized: patt_29 = SAD_EXPR <a_14, b_16, r_23>; This is not correct. This only happens for targets with both uabd and sabd but not vec_widen_{s,u}abd, currently LoongArch is the only target affected. The problem is vect_look_through_possible_promotion will throw away a series of conversions if the effect is equivalent to a sign change and a promotion, but here the sign change is definitely relevant, and the promotion is also relevant for "mixed sign" cases like r += abs((unsigned int)(unsigned char) a - (signed int)(signed char) b (we need to promote to HImode as the difference can exceed the range of QImode). If there were any redundant promotion, it should have been stripped in vect_recog_abd_pattern (i.e. when patt_31 = .ABD (a.0_4, b.1_6) is recognized) instead of in vect_recog_sad_pattern, or we'd have a missed-optimization if the ABD output is not summerized. So anyway vect_recog_sad_pattern is just not a proper location to call vect_look_through_possible_promotion for the ABD inputs, remove the calls to fix the issue. gcc/ChangeLog: PR tree-optimization/118727 * tree-vect-patterns.cc (vect_recog_sad_pattern): Don't call vect_look_through_possible_promotion on ABD inputs. gcc/testsuite/ChangeLog: PR tree-optimization/118727 * gcc.dg/pr108692.c: Mention PR 118727 in the comment. * gcc.dg/pr118727.c: New test case.
