On Tue, 3 May 2016, Jan Hubicka wrote:
> Hi,
> the code path handling the case where callee is missing return statement but
> calle
> statement has LHS is broken in tree-inline since anonymous SSA_NAMEs was
> introduced.
> This code is not not used because all those inlines are disabled by
> gimple_check_call_matching_types, but since I would like to drop the checks
> it seems
> good idea to fix the code path rather than dropping it. This copies what
> Jakub's
> code does when redirecting to unreachable.
>
> Bootstrapped/regtesetd x86_64-linux, OK?
LGTM. [we should eventually simply allow "anonymous" default defs
which would simply be not registered with the var->default-def map
but just be marked SSA_NAME_IS_DEFAULT_DEF and having a GIMPLE_NOP def.
Not sure where having this would break today, maybe some checking bits
in verify-ssa are not happy with such default defs]
Thanks,
Richard.
> Honza
>
> * tree-inline.c (expand_call_inline): Fix path inlining function
> with no return statement.
> Index: tree-inline.c
> ===================================================================
> --- tree-inline.c (revision 235839)
> +++ tree-inline.c (working copy)
> @@ -4708,7 +4708,7 @@ expand_call_inline (basic_block bb, gimp
> {
> tree name = gimple_call_lhs (stmt);
> tree var = SSA_NAME_VAR (name);
> - tree def = ssa_default_def (cfun, var);
> + tree def = var ? ssa_default_def (cfun, var) : NULL;
>
> if (def)
> {
> @@ -4719,6 +4719,11 @@ expand_call_inline (basic_block bb, gimp
> }
> else
> {
> + if (!var)
> + {
> + tree var = create_tmp_reg_fn (cfun, TREE_TYPE (name), NULL);
> + SET_SSA_NAME_VAR_OR_IDENTIFIER (name, var);
> + }
> /* Otherwise make this variable undefined. */
> gsi_remove (&stmt_gsi, true);
> set_ssa_default_def (cfun, var, name);
