Breaking Through the Normalization Barrier: A Self-Interpreter for F-omega, by Matt Brown and Jens Palsberg:
According to conventional wisdom, a self-interpreter for a strongly normalizing λ-calculus is impossible. We call this the normalization barrier. The normalization barrier stems from a theorem in computability theory that says that a total universal function for the total computable functions is impossible. In this paper we break through the normalization barrier and define a self-interpreter for System Fω, a strongly normalizing λ-calculus. After a careful analysis of the classical theorem, we show that static type checking in Fω can exclude the proof’s diagonalization gadget, leaving open the possibility for a self-interpreter. Along with the self-interpreter, we program four other operations in Fω, including a continuation-passing style transformation. Our operations rely on a new approach to program representation that may be useful in theorem provers and compilers.
I haven't gone through the whole paper, but their claims are compelling. They have created self-interpreters in System F, System Fω and System Fω+, which are all strongly normalizing typed languages. Previously, the only instance of this for a typed language was Girard's System U, which is not strongly normalizing. The key lynchpin appears in this paragraph on page 2:
Our result breaks through the normalization barrier. The conventional wisdom underlying the normalization barrier makes an implicit assumption that all representations will behave like their counterpart in the computability theorem, and therefore the theorem must apply to them as well. This assumption excludes other notions of representation, about which the theorem says nothing. Thus, our result does not contradict the theorem, but shows that the theorem is less far-reaching than previously thought.
Pretty cool if this isn't too complicated in any given language. Could let one move some previously non-typesafe runtime features, into type safe libraries.