[Bug tree-optimization/99728] code pessimization when using wrapper classes around SIMD types

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99728

--- Comment #15 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:9f34b780b0461ec7b2b2defe96e44ab616ea2aa3

commit r12-2097-g9f34b780b0461ec7b2b2defe96e44ab616ea2aa3
Author: Richard Biener <rguenther@suse.de>
Date:   Wed Jul 7 11:41:03 2021 +0200

    tree-optimization/99728 - improve LIM for loops with aggregate copies

    This improves LIM by recording aggregate copies for disambiguation
    purposes instead of as UNANALYZABLE_MEM which will prevent any
    invariant or store motion across it.  This allows four of the six
    references in the loop of the testcase to be promoted.

    2021-07-07  Richard Biener  <rguenther@suse.de>

            PR tree-optimization/99728
            * tree-ssa-loop-im.c (gather_mem_refs_stmt): Record
            aggregate copies.
            (mem_refs_may_alias_p): Add assert we handled aggregate
            copies elsewhere.
            (sm_seq_valid_bb): Give up when running into aggregate copies.
            (ref_indep_loop_p): Handle aggregate copies as never
            being invariant themselves but allow other refs to be
            disambiguated against them.
            (can_sm_ref_p): Do not try to apply store-motion to aggregate
            copies.

            * g++.dg/opt/pr99728.C: New testcase.