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

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

[Philippe Strauss]

(I missed the begin of the thread, so I do some splicing from the mailing list archiver, sorry about that)

I would be very interested in an typed inline ASM in ocaml native, I've just dicovered SIMD intrisics in gcc, well, discovered using it, and that simple approach with static typing, and the C language to do function and loops prologue is a breeze.

I've never have the gut to learn full assembly (hum, on the mac, the 16bytes alignment of everything, including stack, is repulsive), only looking at the code generated by gcc, and tuning my C before then.
the cc intrisics approach is a low barrier entry level, which is good in computer science nowadays.

Regards.

-

Hello,

As mentioned earlier we have a student working on an implementation of the 
Linear Scan Register Allocator [1] for ocamlopt (and thereby ocamlnat). It 
took some time, but now there's a first working patch which looks promising. 
This work is done by Marcell Fischbach as part of his diploma thesis. The 
idea is to use the linear scan algorithm to drive the register allocation in 
the native top-level ocamlnat at some point, as suggested by Fabrice Le 
Fessant [2].

Marcell is now working to implement a proof-of-concept of an inline assembler 
for ocamlnat on i386 based on code from Alain Frisch an Fabrice Le Fessant. 
The result will also be contributed once ready, and will be used to 
effectively compare ocamlnat and the byte-code ocaml top-level.

The linear scan implementation reuses as much of the existing ocamlopt 
functionality as possible, so additional maintenance overhead should be 
manageable. Comments and suggestions are welcome of course. Please keep 
Marcell CC'ed with any replies as he's not subscribed to the list.

greets,
Benedikt

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

[Benedikt Meurer]

On Aug 1, 2011, at 16:53 , Benedikt Meurer wrote:

> The linear scan implementation reuses as much of the existing ocamlopt functionality as possible, so additional maintenance overhead should be manageable. Comments and suggestions are welcome of course. Please keep Marcell CC'ed with any replies as he's not subscribed to the list.

Mantis ticket created: http://caml.inria.fr/mantis/view.php?id=5324