To avoid compilation errors it is desirable to verify that a compiler is type correct—i.e., given well-typed source code, it always outputs well-typed target code. This can be done intrinsically by implementing it as a function in a dependently typed programming language, such as Agda. This function manipulates data types of well-typed source and target programs, and is therefore type correct by construction. A key challenge in implementing an intrinsically typed compiler is the representation of labels in bytecode. Because label names are global, bytecode typing appears to be inherently a non-compositional, whole-program property. The individual operations of the compiler do not preserve this property, which requires the programmer to reason about labels, which spoils the compiler definition with proof terms.

In this paper, we address this problem using a new nameless and co-contextual representation of typed global label binding, which is compositional. Our key idea is to use linearity to ensure that all labels are defined exactly once. To write concise compilers that manipulate programs in our representation, we develop a linear, dependently typed, shallowly embedded language in Agda, based on separation logic. We show that this language enables the concise specification and implementation of intrinsically typed operations on bytecode, culminating in an intrinsically typed compiler for a language with structured control-flow.