Titre : CamllVM - an LLVM based runtime for Ocaml

Auteur : Raphael Amiard (UPMC)

Abstract : 

There are two official ways of running Ocaml code : The first is using
the bytecode compiler and interpreter, the second is using the
ocamlopt compiler to generate native code directly.

There has been many discussions about using LLVM in the context of
OCaml code generation. The theoretical advantages of using LLVM for
code generations are amongst others, faster execution, better
portability, and shared optimizations amongst all users of LLVM.

The CamllVM project takes the approach of generating LLVM code from
the OCaml bytecode usually interpreted by ocamlrun. There are many
problems to tackle in such an approach, but the two main pain points
arise from two facts :

- OCaml bytecode was meant to be interpreted, not compiled. It is
  untyped, where LLVM intermediate representation is typed, and uses
  stack semantics where LLVM uses register based semantics.

- LLVM, despite aiming to be a general library apt to compile down all
  sorts of languages, is of course designed with C, C++, and
  Objective-C in mind, since it is the main languages it is used to
  compile. It thus presents a number of difficulties to us, mainly
  about handling exceptions and garbage collection.

In this talk, I will present the advantages and drawbacks of the
approach CamllVM has taken, what leads were explored, and what is
working. I will also summarize the main difficulties in compiling
OCaml to LLVM.