Hi. I've got the following testcase:

Code: Select all

#include <Eigen/Core>
#include <iostream>

template <typename T>
struct foo {
  foo(T t) : t(t) { }
  operator T () { return t; }
  T t;
};

int main() {
  foo<Eigen::Vector3d> f(Eigen::Vector3d(1.0, 1.0, 1.0));
  std::cout << (0.5 * f) << '\n';
}

It produces the following output:

Code: Select all

-1.55252e+231
-1.55252e+231
-1.55252e+231

That is clearly the wrong answer (I'm expecting the vector (0.5, 0.5, 0.5)^T). I can reproduce this problem with GCC 4.7.0 (Fedora 17) with Eigen 3.1.1, and with GCC 4.6.3 (Ubuntu 12.04) with Eigen 3.0.5. Making foo a non-template, replacing all T's for Eigen::Vector3, results in “error: no match for ‘operator*’ in ‘5.0e-1 * f’”. Since I have been unable to reproduce this problem without Eigen, I assume this must be something to do with Eigen or with my using Eigen.

So – does anyone have an idea what might be wrong here?

Thanks.

If you replace the last line with "std::cout << (0.5 * Eigen::Vector3d(f)) << '\n';" the problem goes away. The is something strange occurring in your implicit cast (implicit casts should be avoided IMHO).

Indeed, with an explicit conversion it works. There can't be much wrong going on in the implicit conversion function, as it's just a simple return of a copy of a vector. The problem must be in the interaction of this implicit conversion with the rest of Eigen. I'm still wondering what kind of interaction that is.

To be honest I'm surprised this does compile. It seems the cast operator is implicitly called only for a scalar multiple product, not for other operators involving a foo<> and a Vector3d. I also cannot reproduce the error with clang. Returning a const reference also fixes the issue.

I agree, this code probably shouldn't compile. I dug into it, and didn't really find a reason why it compiles at all. I tried explicitely calling op *, using Eigen::operator * (0.5, f), but that failed to compile. Yet, tracing through a debugger, I can clearly see Eigen::operator * (Scalar, Eigen::MatrixBase<Eigen::Matrix<…> >::StorageBaseType const&) called.

Unfortunately, I haven't been able to isolate those parts of Eigen that make this compile either.

It's strange you haven't reproduced the problem with Clang, though. I have been able to reproduce it with Clang 3.0 and Microsoft's Visual C++ 2010.

Also, from my debugging, I learned that what happens is roughly the following: operator T () gets called, and the vector gets copied. Then Eigen::op * (Scalar, Matrix) is called, which in turn calls Eigen::Matrix::operator * (Scalar). Then an object of type CwiseUnaryOp gets created which is supposed to contain the vector – however, the Matrix instance inside CwiseUnaryOp (m_xpr member) appears to be uninitialized – it also appears to have a different address from that copy that was created by calling operator T (). Then the product of 0.5 with this uninitialized matrix is computed and printed.