We give a polynomial-time algorithm for detecting very long cycles in dense regular graphs. Specifically, we show that, given $α\in (0,1)$, there exists a $c=c(α)$ such that the following holds: there is a polynomial-time algorithm that, given a $D$-regular graph $G$ on $n$ vertices with $D\geq αn$, determines whether $G$ contains a cycle on at least $n - c$ vertices. The problem becomes NP-complete if we drop either the density or the regularity condition. The algorithm combines tools from extremal graph theory and spectral partitioning as well as some further algorithmic ingredients.