On Fri, Oct 03, 2014 at 03:38:49PM +0100, Rafael J. Wysocki wrote:

On Friday, October 03, 2014 02:58:26 PM Mark Rutland wrote:

quoted

On Fri, Oct 03, 2014 at 03:03:51AM +0100, Rafael J. Wysocki wrote:

quoted

On Thursday, October 02, 2014 04:36:54 PM Mika Westerberg wrote:

quoted

On Thu, Oct 02, 2014 at 02:46:30PM +0200, Arnd Bergmann wrote:

quoted

On Thursday 02 October 2014 15:15:08 Mika Westerberg wrote:

[cut]

quoted

Putting everything to a single package results this:

		Package () { "pwms", Package () {"led-red", ^PWM0, 0, 10, "led-green", ^PWM0, 1, 10 }}

But I think the below looks better:

I agree.

quoted

		Package () { "pwms", Package () {^PWM0, 0, 10, ^PWM0, 1, 10 }}
		Package () { "pwm-names", Package () {"led-red", "led-green"}}

and it is trivial to match with the corresponding DT fragment.

quoted

	}

vs.

	pwm-slave {
		pwms = <&pwm0 0 10>, <&pwm1 1 20>;
		pwm-names = "led-red", "led-green";
	};

I don't have strong feelings which way it should be. The current
implementation limits references so that you can have only integer
arguments, like {ref0, int, int, ref1, int} but if people think it is
better to allow strings there as well, it can be changed.

I would like to get comments from Darren and Rafael about this, though.

In my opinion there needs to be a "canonical" representation of the
binding that people always can expect to work.  It seems reasonable to
use the one exactly matching the DT representation for that.

I don't follow. The two forms would share the same high-level accessors,
but the binary representation is already different. Why should we choose
the inferior layout given they are already different binary formats?

Well, why is it inferior in the first place?  It represents the same information
and I'm not sure why the binary formats matter here?

Because people get the format wrong regardless of documentation. The
format:

Package () {
	Package () { ^ref1, data, data },
	Package () { ^ref2, data },
	Package () { ^ref3, data, data, data },
}

Is superior to the format:

Package () { ^ref1, data, data, ^ref2, data, ^ref3, data, data, data }

Because in the former you have delimiters that can be used to verify
each tuple. Imagine someone misses a data element for one of these
tuples. In the former layout you can detect this easily while in the
latter you cannot.

Additionally, the former can represent variadic phandle/reference + args
formats, which the latter cannot. The DT pinctrl bindings look the way
they do because we can't represent variadic args in DT due to a lack of
delimiters.

You have the ability to embed structure in the binary format. Throwing
this away because it doesn't quite match DT does not to me sounds like
the right tradeoff.

If I'm to create a _DSD with that information and have a DT template, it
surely is easier to copy it exactly than to figure out how to resolve it
to represent something I can actually put in there. 

Sure, you can put something together fast. However, that doesn't make it
necessarily better.

If you want to make reusing DT templates easier, why not just embed a
DTB and be done with it?

Thanks,
Mark.