Re: [PATCH v9 1/7] typeinfo: Introduce KCFI typeinfo mangling API

[Kees Cook <kees@kernel.org>]

On Fri, Dec 12, 2025 at 05:29:43PM -0800, Andrew Pinski wrote:
> On Fri, Dec 12, 2025 at 5:24 PM Kees Cook <kees@kernel.org> wrote:
> >
> > On Fri, Dec 12, 2025 at 03:07:51PM -0800, Andrew Pinski wrote:
> > > On Tue, Dec 9, 2025 at 6:27 PM Kees Cook <kees@kernel.org> wrote:
> > > [...]
> > > > +uint32_t
> > > > +typeinfo_get_hash (tree type)
> > > > +{
> > > > +  gcc_assert (type != NULL_TREE);
> > > > +  uint32_t hash_state = 2166136261U; /* FNV-1a 32-bit offset basis.  */
> > > > +
> > > > +  mangle_type (type, nullptr, &hash_state);
> > > > +  return hash_state;
> > > > +}
> > >
> > > It might make sense to do a few self_test here instead of (in addition
> > > to) having a testcase.
> > > This way the testing is done earlier.
> >
> > I tried to do this back in v2 and could not make it work. See [1] for
> > more details on what I found, but basically I don't have access to the
> > parser itself in the selftests, so I couldn't build end-to-end testing
> > of arbitrary C (testing node types isn't really a sufficient test in my
> > view). I could to basic type tests, but it would be redundant to what
> > I ended up with in dg, and I didn't want to split up the testing.
> 
> I am asking for self-tests of the API and not depending on the C
> parser. So generating some FUNCTION_TYPE types and then getting back
> the hash and/or name.
> 
> > (testing node types isn't really a sufficient test in my view)
> It might not be a fully sufficient test in itself; it does not mean it
> is NOT a useful test to have. This is why I mentioned in addition to.
> Testing the API outside of the end-to-end testing is a good thing and
> can find bugs early on.  GCC does not have enough self-tests really
> and folks don't use it as much as we should.

Okay, yeah, if doing this kind of "manual" construction is viewed as
meaningful, I'll give it a shot to augment the existing tests.

-Kees

-- 
Kees Cook