A one-word review comment on my Eigen MR revealed that hand-rolling normalization with .norm() silently breaks on extreme inputs. Eigen already had the fix — I just wasn’t using it.
Eigen’s blocked Householder path only existed for left-side application. I added the right-side equivalent, upgrading M*Q from O(n) rank-1 updates to cache-friendly blocked matrix multiplies.
Eigen’s .inc plugin headers were read by Doxygen but never preprocessed as C++ — meaning macros like EIGEN_PARSED_BY_DOXYGEN were silently ignored. Here’s how I verified it in Docker and what the one-line fix looks like.
I submitted a Modified Gram-Schmidt QR decomposition to Eigen and a maintainer asked: why? Here’s the benchmark data that answered the question.
How a missing overflow check in Eigen’s TensorUInt128 division operator caused an infinite loop for any dividend above 2^127 — and the one-line fix that stopped it.
A performance optimization in Eigen’s fill path assumed all scalar types are equal. GCC’s libstdc++ disagreed — and AutoDiffScalar paid the price.
How a GCC 13 false-positive -Warray-bounds warning in Eigen’s TensorContraction led to a lesson about if constexpr, C++14 portability, and the right way to suppress compiler warnings in a large codebase.