This paper summarizes my experience in implementing a compiler for a functional language. The language is ML(1) [Milner 84] and the compiler was first implemented in 1980 as a personal project when I was a postgraduate student at the University of Edinburgh(2). At the time, I was familiar with programming language semantics but knew very little about compiler technology; interpreters had been my main programming concern. Major influences in the design of this compiler have been [Steele 77] [Steele 78] and the implementation folklore for statically and dynamically scoped dialects of Lisp [Allen 78]. As a result, the internal structure of the compiler is fairly unorthodox, if compared for example with [Aho 78]. Anyway, a compiler for a language like ML has to be different. ML is interactive, statically scoped, strongly typed, polymorphic, and has first class higher-order functions, type inference and dynamic allocation. These features preclude many well-known implementation styles, particularly the ones used for Lisp (because of static scoping), the Algol family (because of functional values) and C (because of nested scoping and strong typing). The interaction of these features is what gives ML its “character”, and makes compilation challenging. The compiler has been recently partially converted to the new ML standard. The major points of interest which are discussed in this paper are: (a) the interactive interpreter-like usage; (b) the polymorphic type inference algorithm; (c) the compilation of pattern matching; (d) the optimization of the representation of user defined data types; (e) the compilation of functional closures, function application and variable access; (f) the intermediate abstract machine and its formal operational semantics; (g) modules and type-safe separate compilation.