Add a CI workflow for building against Eigen from source

[dfm]

This isn't really meant to be merged, but it's a follow up from #746 demonstrating the test failure with the current master branch of Eigen. This workflow seems to fail when run on macOS (but not on Ubuntu), and (as @hawkinsp suggested) I expect the issue is related to using clang instead of gcc.

One way to "fix" this test failure is to update this function:

nanobind/tests/test_eigen.cpp

Lines 172 to 179 in bff96e2

	using Matrix1d = Eigen::Matrix<double,1,1>;
	try {
	m.def(
	"default_arg", [](Matrix1d a, Matrix1d b) { return a + b; },
	"a"_a = Matrix1d::Zero(), "b"_a = Matrix1d::Zero());
	} catch (...) {
	// Ignore (NumPy not installed, etc.)
	}

as follows

-             "default_arg", [](Matrix1d a, Matrix1d b) { return a + b; },
+             "default_arg", [](Matrix1d a, Matrix1d b) -> Matrix1d { return a + b; },

but I think there's probably something more complicated happening here.

[wjakob]

That looks concerning since it affects such basic types. Let me add @WKarel as well.

[WKarel]

a + b should create an instance of a specialization of Eigen::CwiseBinaryOp for scalar sums and the given specialization of Eigen::Matrix. Upon return, this should call nanobind's specialization of type_caster for dense Eigen expressions, which should forward to the specialization of type_caster for Eigen::Array<double, 1, 1>.

What happens if you replace the trailing return type Eigen::Matrix<double,1,1> with Eigen::Array<double,1,1>?

I do not see why type_caster<dense expression>::from_cpp is a function template:

    template <typename T2>
    static handle from_cpp(T2 &&v, rv_policy policy, cleanup_list *cleanup) noexcept {
        return Caster::from_cpp(std::forward<T2>(v), policy, cleanup);
    }

Replacing it with

    static handle from_cpp(T&& v, rv_policy policy, cleanup_list* cleanup) noexcept {
      return Caster::from_cpp(std::move(v), policy, cleanup);
    }
    static handle from_cpp(const T& v, rv_policy policy, cleanup_list* cleanup) noexcept {
      return Caster::from_cpp(v, policy, cleanup);
    }

might actually fix the issue with CLang on MacOS.

[dfm]

Thanks for the suggestions, @WKarel!

What happens if you replace the trailing return type Eigen::Matrix<double,1,1> with Eigen::Array<double,1,1>?

I can confirm that updating the test as follows also fixes the error:

-             "default_arg", [](Matrix1d a, Matrix1d b) { return a + b; },
+             "default_arg", [](Matrix1d a, Matrix1d b) -> Eigen::Array<double, 1, 1> { return a + b; },

However, the error persists if Matrix1d in the test is redefined as Eigen::Array<double, 1, 1>, but the return type is omitted.

Replacing it with ... might actually fix the issue with CLang on MacOS.

I tried updating the following lines (I think this is what you were referring to):

nanobind/include/nanobind/eigen/dense.h

Lines 228 to 231 in 5ead8f0

	template <typename T2>
	static handle from_cpp(T2 &&v, rv_policy policy, cleanup_list *cleanup) noexcept {
	return Caster::from_cpp(std::forward<T2>(v), policy, cleanup);
	}

with the version you suggest, unfortunately that doesn't seem to fix the problem!

[WKarel]

I see. And yes, what you did is what I meant to give a try. Now I am out of ideas. So I would try to debug the problematic call, or temporarily introduce some debug output, e.g. starting with std::cerr << typeid(v).name(); just above

nanobind/include/nanobind/eigen/dense.h

Line 230 in 5ead8f0

return Caster::from_cpp(std::forward<T2>(v), policy, cleanup);

[wjakob]

@dfm: could you please provide the typeid() output as suggested by @WKarel?

[dfm]

@wjakob, @WKarel — I'm so sorry about disappearing here! I'll see if I can make any progress debugging this later today or tomorrow, but in the meantime, here's what I get for that typeid output when the test fails:

N5Eigen13CwiseBinaryOpINS_8internal13scalar_sum_opIddEEKNS_6MatrixIdLi1ELi1ELi0ELi1ELi1EEES6_EE

If that immediately suggests anything to try, let me know!

[wjakob]

Can you run that through c++filt or c++filt -t on your machine? Just run the command, enter that long string, and post what this outputs.

[dfm]

Absolutely! That prints:

Eigen::CwiseBinaryOp<Eigen::internal::scalar_sum_op<double, double>, Eigen::Matrix<double, 1, 1, 0, 1, 1> const, Eigen::Matrix<double, 1, 1, 0, 1, 1> const>

[WKarel]

That type meets my expectations. Does this test also fail in an unoptimized build?

[dfm]

Does this test also fail in an unoptimized build?

Good question and sorry for the slow response! It looks like switching to a Debug build fixes the error.