On Sun, Mar 29, 2009 at 05:52:43PM +0200, Manuel Lauss wrote:
> From: Manuel Lauss <firstname.lastname@example.org>
> Date: Sun, 29 Mar 2009 17:52:43 +0200
> To: Kevin Hickey <email@example.com>
> Cc: Sergei Shtylyov <firstname.lastname@example.org>,
> Linux-MIPS <email@example.com>
> Subject: Re: [PATCH 0/3] Alchemy: platform updates
> Content-Type: text/plain; charset=US-ASCII
> On Sun, 29 Mar 2009 10:27:46 -0500
> Kevin Hickey <firstname.lastname@example.org> wrote:
> > On Sun, 2009-03-29 at 17:03 +0400, Sergei Shtylyov wrote:
> > > Single kernel binary? If it's at all possible, I am all for it.
> > On some level, I agree but not at the expense of a larger kernel or
> > longer boot times. Maybe I'm just not following how your implementation
> > works but it seems to me that runtime checks will add to boot time.
> > More importantly it adds to the kernel memory footprint as the tables of
> > constants for multiple CPUs will have to be compiled in. If I'm
> > designing a board with an Au1250 in it, I don't care about the interrupt
> > numbers for Au1100 or Au1500. This problem compounds when we introduce
> > Au1300 - several of its subsystems (like the interrupt controller) are
> > new requiring not only a new table of constants but a new object as
> > well. In the desktop space I can understand this approach, but in the
> > embedded space it seems like an unnecessary resource burden.
> > Please enlighten me :)
> You're right, from a single-cpu-board POV it doesn't make sense.
> However if you have a few boards which mostly differ in the Alchemy
> chip used (and not much else difference in board support code), I find
> this to be highly beneficial. If I can have a single binary for the
> folks testing these boards, all the better!
> Yes, increased binary size is to be expected, but I don't expect it to
> be in the megabyte range.
> I'm primarily doing this for company-internal purposes; I just thought
> I'd share the final result, maybe someone else might find it useful.
In the past the Alchemy code has frequently suffered from small silly
bugs that were affecting only a part of the Alchemy platforms. If
code was more generic there would be the additional benefit of improved
test coverage without more testing.
I frequently see people still optimizing code as if the problem was
still squeezing a program into a 2716 EPROM for an 8085. Trading a
few bytes or microseconds of startup time for sanity is a really good