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

[Benedikt Meurer <benedikt.meurer@googlemail.com>]

On Aug 1, 2011, at 17:04 , Gabriel Scherer wrote:

> This work is meant to make a compromise between generated code quality
> and compilation speed to have good performances in rapid
> prototyping/development scenario.
> 
> Do you have more precise measurements on
> - the relative costs of the successive transformations during native
> compilation (including external linking etc.)? Which proportion of
> time is currently used for register allocation?

Right now most of the time in ocamlnat is spent in waiting for the assembler and linker to finish and the runtime linker to load the generated library file. However there are no precise timing results yet.

Marcell did a rough test with ocamlopt last week, building the test suite with both graph coloring and linear scan on a 2009 iMac (Core 2 Duo). The overall time spent in the register allocator dropped from 28s to 8s.

> - the performance cost of this new allocator in the generated code? I
> suppose the results may vary between different architectures (eg. x86
> is probably more sensitive to good allocation decisions than x86_64).

Same here... not yet. From what I've seen, the generated amd64 code is really close to the graph coloring code (isomorphic up to register renaming in most cases). Dunno for i386 yet.

Benedikt