Re: [discuss] AMD x86-64 GCC development

[Jan Hubicka <jh at suse dot cz>]

>     Richard, I'm confused by your statement.  AFAIR you're on the x86-64
>     lists since the beginning and should have read all the discussions
>     that were going on.  
> 
> I'm on a half dozen lists.  I try to follow as much of what's going on
> as I can, but I can't be sure I get everything correct.
> 
> This came up in a discussion today where I'd assumed that once the NDA
> ended, the GCC x86-84 development had moved, but I didn't see the
> corresponding messages, so I was confused.
> 
>     The gcc work for AMD x86-64 is done on a *public* CVS tree that has
>     been started from the gcc.gnu.org CVS tree at the end of september and
>     officially announced 
> 
> Why is this not being done in the main GCC tree?
> 
>     We've created separate CVS trees for (so far) GCC, Binutils, Glibc and
>     the Linux kernel.  When the x86-64 port to GNU/Linux is in a stable
>     state, the CVS tree will be merged with the mainstream tree of the
>     different projects (according to the rules set up by the projects),
>     e.g. with GCC's tree at gcc.gnu.org with submissions to
>     patches@gcc-gnu.org with proper ChangeLogs etc.
> 
> But isn't this going to be a lot more work than if the development
> were done from the beginning in the main GCC tree?
Problem is that unline completely new chip, the x86-64 backend is done
in the existing i386 backend and active development on it introduces
breakage to i386 backend too. Stability of the i386 is important for
the other developers. With current scheme of not having conditionized compilation
of machine description, the patterns useless for i386, but required for x86-64
increases size of the i386 compiler by considerable amount. Thats another purpose
why I don't want to polute 3.0 with unfinished x86-64 support.

In the past similar stuff was done for other backends as well and usually such
early development stages was done out of mainline, so thats why I am doing so
here as well. Also since I need approvals for the patches, I want to minimize
their amount and send code that is really tested.  Until recently this was
impossible for x86-64.

Concerning the generic code changes and infrastructure stuff, I am mergning
all important bits to the mailine now.

Since development moved faster than I've expected and 3.0 is going to be later,
I would probably try to fit beta of x86-64 in 3.0 too, in case merging will be
fast enought.

Honza