https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105554
--- Comment #16 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #14)
> So, I have tried
> --- gcc/cgraphclones.cc.jj 2023-02-24 11:05:19.704595633 +0100
> +++ gcc/cgraphclones.cc 2023-03-16 19:12:30.452503051 +0100
> @@ -1094,6 +1094,15 @@ cgraph_node::create_version_clone_with_b
> || in_lto_p)
> new_version_node->unique_name = true;
>
> + if (target_attributes)
> + {
> + push_cfun (DECL_STRUCT_FUNCTION (new_decl));
> + for (tree arg = DECL_ARGUMENTS (new_decl); arg; arg = DECL_CHAIN
> (arg))
> + if (VECTOR_TYPE_P (TREE_TYPE (arg)))
> + SET_DECL_MODE (arg, TYPE_MODE (TREE_TYPE (arg)));
> + pop_cfun ();
> + }
> +
> /* Update the call_expr on the edges to call the new version node. */
> update_call_expr (new_version_node);
>
> but that really didn't help, the problem seems to be actually different.
>
> From what I can see, tree_function_versioning does:
> 6240 DECL_RESULT (new_decl) = DECL_RESULT (old_decl);
> 6241 DECL_ARGUMENTS (new_decl) = DECL_ARGUMENTS (old_decl);
> 6242 initialize_cfun (new_decl, old_decl,
> 6243 new_entry ? new_entry->count :
> old_entry_block->count);
> and initialize_cfun will call allocate_function, which does:
> 4845 /* Now that we have activated any function-specific attributes
> 4846 that might affect layout, particularly vector modes,
> relayout
> 4847 each of the parameters and the result. */
> 4848 relayout_decl (result);
> 4849 for (tree parm = DECL_ARGUMENTS (fndecl); parm;
> 4850 parm = DECL_CHAIN (parm))
> 4851 relayout_decl (parm);
> So, we temporarily set DECL_ARGUMENTS and DECL_RESULT of new_decl to
> arguments/result of old_decl and allocate_function called relayout_decl on
> those, later on we create new arguments (which supposedly have correct
> DECL_MODE). But we've changed the old DECL_RESULT/DECL_ARGUMENTS.
Ick - that's one of the worst place to do this ... what happens if we
postpone setting DECL_RESULT/DECL_ARGUMENTS to after initialize_cfun
(and remove the setting of them there)?
I suppose on its own this doesn't fix the original issue but possibly
clobbering the original function arg decls layouts looks wrong ...
