https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99728
--- Comment #14 from Richard Biener <rguenth at gcc dot gnu.org> --- With the hack and PR101293 fixed we're still left with Memory reference 13: *d_28(D).lam1.v Memory reference 14: VIEW_CONVERT_EXPR<vector(32) unsigned char>(*d_28(D).lam1) because the access types are incompatible (v4df vs v32qi). Possible to hack around but the initial hack creating the V_C_E has issues enough (though handling the aggregate copy in LIM will inevitably run into incomaptible types compared to other refs as well). How I love the fun of C++ and all the abstraction when using it ... :/
