https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86983
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |documentation
Status|UNCONFIRMED |NEW
Last reconfirmed| |2018-08-22
CC| |hubicka at gcc dot gnu.org,
| |jsm28 at gcc dot gnu.org
Ever confirmed|0 |1
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
case GIMPLE_GOTO:
t = gimple_goto_dest (stmt);
/* We will not inline a function which uses computed goto. The
addresses of its local labels, which may be tucked into
global storage, are of course not constant across
instantiations, which causes unexpected behavior. */
if (TREE_CODE (t) != LABEL_DECL)
{
inline_forbidden_reason
= G_("function %q+F can never be inlined "
"because it contains a computed goto");
*handled_ops_p = true;
return t;
}
we have to assume that the goto is to labels of the function or functions
inlined into it. We do not seem to compute whether any of those escape
(your testcase returns them though).
So the wording is correct and the Labels-as-Values documentation is
wrong.
Of course if we get enough analysis GCC might decide to inline, all the
wording always applies to a concrete implementation and is not to be
treated as language specification.
