https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120927
--- Comment #9 from GCC Commits <cvs-commit at gcc dot gnu.org> --- The releases/gcc-15 branch has been updated by Richard Biener <rgue...@gcc.gnu.org>: https://gcc.gnu.org/g:b8599692a336b29851bdc5d8506a51d57521595c commit r15-9940-gb8599692a336b29851bdc5d8506a51d57521595c Author: Richard Biener <rguent...@suse.de> Date: Thu Jul 3 14:39:22 2025 +0200 tree-optimization/120927 - 510.parest_r segfault with masked epilog The following fixes bad alignment computaton for epilog vectorization when as in this case for 510.parest_r and masked epilog vectorization with AVX512 we end up choosing AVX to vectorize the main loop and masked AVX512 (sic!) to vectorize the epilog. In that case alignment analysis for the epilog tries to force alignment of the base to 64, but that cannot possibly help the epilog when the main loop had used a vector mode with smaller alignment requirement. There's another issue, that the check whether the step preserves alignment needs to consider possibly previously involved VFs (here, the main loops smaller VF) as well. These might not be the only case with problems for such a mode mix but at least there it seems wise to never use DR alignment forcing when analyzing an epilog. We get to chose this mode setup because the iteration over epilog modes doesn't prevent this, the maybe_ge (cached_vf_per_mode[0], first_vinfo_vf) skip is conditional on !supports_partial_vectors and it is also conditional on having a cached VF. Further nothing in vect_analyze_loop_1 rejects this setup - it might be conceivable that a target can do masking only for larger modes. There is a second reason we end up with this mode setup, which is that vect_need_peeling_or_partial_vectors_p says we do not need peeling or partial vectors when analyzing the main loop with AVX512 (if it would say so we'd have chosen a masked AVX512 epilog-only vectorization). It does that because it looks at LOOP_VINFO_COST_MODEL_THRESHOLD (which is not yet computed, so always zero at this point), and compares max_niter (5) against the VF (8), but not with equality as the comment says but with greater. This also needs looking at, PR120939. PR tree-optimization/120927 * tree-vect-data-refs.cc (vect_compute_data_ref_alignment): Do not force a DRs base alignment when analyzing an epilog loop. Check whether the step preserves alignment for all VFs possibly involved sofar. * gcc.dg/vect/vect-pr120927.c: New testcase. * gcc.dg/vect/vect-pr120927-2.c: Likewise. (cherry picked from commit 918f4517564c2cf7e5bb907428d5413742bee56f)
