Re: [Caml-list] Linear Scan Register Allocator for ocamlopt/ocamlnat

[Erik de Castro Lopo <mle+ocaml@mega-nerd.com>]

Pierre-Alexandre Voye wrote:

> I have a stupid question : I wonder if it would not be a bad idea that Ocaml
> output C code and let gcc do its work, so compile code with good
> performances in a lot of architecture ? Gcc is able to do autovectorization
> (SSE, MMX, Larabee in the futur, etc...), very specific processor
> optimization, etc...
> But maybe it's a stupid idea ?

I have some experience in thie area. I work on the DDC compiler [0]
a compiler for a strict by default (optionally lazy) evaluation
dialect of Haskell.

When I joined the project the compiler had a working compile via C
backend, to which I added an LLVM backend [1].

Executables compiled via the LLVM backend (even without exploring
any of the LLVM optimisation passes) were faster than the same
executables compiled via C (gcc -O2). I suspect this is because
the generated C code was nothing like the C code people write and
the GCC is only good at optimising idiomatic C code.

I highly recommend LLVM as a compiler backend.

HTH,
Erik

[0] http://disciple.ouroborus.net/
[1] http://www.mega-nerd.com/erikd/Blog/CodeHacking/DDC/index.html

-- 
----------------------------------------------------------------------
Erik de Castro Lopo
http://www.mega-nerd.com/