https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126998

--- Comment #6 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Richard Biener <[email protected]>:

https://gcc.gnu.org/g:408162bbcde30c7d6672a20f96aa6dc7a1c5fa58

commit r17-3694-g408162bbcde30c7d6672a20f96aa6dc7a1c5fa58
Author: Richard Biener <[email protected]>
Date:   Wed Aug 26 13:46:22 2026 +0200

    tree-optimization/126998 - avoid vect_get_num_copies on SLP inputs

    The following avoids using vect_get_num_copies on the SLP input
    from vectorizable_lane_reducing which is in this case used
    for costing.  The transform phase will simply look at the actual
    number of vector inputs code generated, but for costing we
    can use vect_get_num_copies_for_invariant which is exported with
    the patch.  It must be possible to compute the number from
    the number of vector outputs, the number of SLP lanes and the
    width of the elements involved, but given what the transform
    phase does this matches it exactly.

            PR tree-optimization/126998
            * tree-vectorizer.h (vect_get_num_copies_for_invariant):
            Declare.
            * tree-vect-slp.cc (vect_get_num_copies_for_invariant): Export.
            * tree-vect-loop.cc (vectorizable_lane_reducing): Use
            vect_get_num_copies_for_invariant on the SLP input.

            * gcc.dg/vect/vect-reduc-dot-pr126998.c: New testcase.

Reply via email to