This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: From beginner: What is the workload of writing a RTL expanderfor a new IR?
- From: Diego Novillo <dnovillo at redhat dot com>
- To: Chunhua Liao <chunhualiao at hotmail dot com>
- Cc: "gcc at gcc dot gnu dot org" <gcc at gcc dot gnu dot org>
- Date: Thu, 18 Dec 2003 09:32:44 -0500
- Subject: Re: From beginner: What is the workload of writing a RTL expanderfor a new IR?
- Organization: Red Hat Canada
- References: <Law11-F639EzsiGPwi30004c45a@hotmail.com>
On Wed, 2003-12-17 at 21:40, Chunhua Liao wrote:
> I notice that there are several projects aiming at improve the
> optimization infrastructure of GCC. And it seems all of them involve
> introducing new IR(s) between the original tree representation and RTL, thus
> new RTL expanders are necessary for them.
> My questions are:
> 1. Is RTL a complete representation for whole translation-unit(a complete
> input sourcefile contains) which can be used for code generator without the
> help from tree representation?
>
No.
> If yes, why did I found some tree references even at the very back end?
>
Precisely for that reason.
> 2. What is the workload(in terms of n persons m months) of writing such new
> RTL expander from a new IR? My current understanding is that it is
> non-trival, since GCC calls several levels of functions to expand a function
> in tree to a sequence of RTL instruction nodes and most of them contain
> parameters in tree type. We may need rewrite them with the parameters of
> new IR type?
>
I'd say more than year. Assuming that you have a good working knowledge
of GCC internals to begin with.
Diego.