[RFC 3/4] dtoc: add support for generate stuct udevice_id

[Simon Glass]

Hi Walter,

On Wed, 29 Jul 2020 at 10:00, Walter Lozano <walter.lozano at collabora.com> wrote:
>
> Hi Simon,
>
> On 28/7/20 23:42, Simon Glass wrote:
> > Hi Walter,
> >
> > On Sun, 26 Jul 2020 at 20:16, Walter Lozano <walter.lozano at collabora.com> wrote:
> >> Hi Simon,
> >>
> >> On 26/7/20 11:53, Simon Glass wrote:
> >>> Hi Walter,
> >>>
> >>> On Tue, 7 Jul 2020 at 08:08, Walter Lozano <walter.lozano at collabora.com> wrote:
> >>>> Hi Simon
> >>>>
> >>>> On 6/7/20 16:21, Simon Glass wrote:
> >>>>> Hi Walter,
> >>>>>
> >>>>> On Fri, 19 Jun 2020 at 15:12, Walter Lozano <walter.lozano at collabora.com> wrote:
> >>>>>> Based on several reports there is an increasing concern in the impact
> >>>>>> of adding additional features to drivers based on compatible strings.
> >>>>>> A good example of this situation is found in [1].
> >>>>>>
> >>>>>> In order to reduce this impact and as an initial step for further
> >>>>>> reduction, propose a new way to declare compatible strings, which allows
> >>>>>> to only include the useful ones.
> >>>>> What are the useful ones?
> >>>> The useful ones would be those that are used by the selected DTB by the
> >>>> current configuration. The idea of this patch is to declare all the
> >>>> possible compatible strings in a way that dtoc can generate code for
> >>>> only those which are going to be used, and in this way avoid lots of
> >>>> #ifdef like the ones shows in
> >>>>
> >>>> http://patchwork.ozlabs.org/project/uboot/patch/20200525202429.2146-1-agust@denx.de/
> >>>>
> >>>>
> >>>>>> The idea is to define compatible strings in a way to be easily parsed by
> >>>>>> dtoc, which will be responsible to build struct udevice_id [] based on
> >>>>>> the compatible strings present in the dtb.
> >>>>>>
> >>>>>> Additional features can be easily added, such as define constants
> >>>>>> depending on the presence of compatible strings, which allows to enable
> >>>>>> code blocks only in such cases without the need of adding additional
> >>>>>> configuration options.
> >>>>>>
> >>>>>> [1] http://patchwork.ozlabs.org/project/uboot/patch/20200525202429.2146-1-agust@denx.de/
> >>>>>>
> >>>>>> Signed-off-by: Walter Lozano <walter.lozano at collabora.com>
> >>>>>> ---
> >>>>>>     tools/dtoc/dtb_platdata.py | 32 ++++++++++++++++++++++++++++++++
> >>>>>>     1 file changed, 32 insertions(+)
> >>>>> I think dtoc should be able to parse the compatible strings as they
> >>>>> are today - e.g. see the tiny-dm stuff.
> >>>> Yes, I agree. My idea is that dtoc parses compatible strings as they are
> >>>> today but also in this new way. The reason for this is to allow dtoc to
> >>>> generate the code to include the useful compatible strings. Of course,
> >>>> this only makes sense if the idea of generating the compatible string
> >>>> associated  code is accepted.
> >>>>
> >>>> What do you think?
> >>> I think this is useful and better than using #ifdef in the source code
> >>> for this sort of thing. We need a way to specify the driver_data value
> >>> as well, right?
> >> Yes, I agree, it is better than #ifdef and c/ould give us some extra
> >> functionality.
> >>
> >> What doe you mean by driver_data value? Are you referring to the data
> >> field? like
> >>
> >> static struct esdhc_soc_data usdhc_imx7d_data = {
> >>           .flags = ESDHC_FLAG_USDHC | ESDHC_FLAG_STD_TUNING
> >>                           | ESDHC_FLAG_HAVE_CAP1 | ESDHC_FLAG_HS200
> >>                           | ESDHC_FLAG_HS400,
> >> };
> >>
> > Actually I was talking about the .data member in struct udevice_id.
> So my example is correct, as usdhc_imx7d_data is the value for .data in
> one case as shown bellow.
> >> If that is the case, I was thinking in defining a constant when specific
> >> compatible strings are enabled by dtoc, based in the above case
> >>
> >> #ifdef FSL_ESDHC_IMX_V2
> >> static struct esdhc_soc_data usdhc_imx7d_data = {
> >>           .flags = ESDHC_FLAG_USDHC | ESDHC_FLAG_STD_TUNING
> >>                           | ESDHC_FLAG_HAVE_CAP1 | ESDHC_FLAG_HS200
> >>                           | ESDHC_FLAG_HS400,
> >> };
> >> #endif
> >>
> >> COMPATIBLE(FSL_ESDHC, "fsl,imx7d-usdhc", &usdhc_imx7d_data, FSL_ESDHC_IMX_V2)
> >>
> >> So when dtoc parses COMPATIBLE and determines that compatible
> >> "fsl,imx7d-usdhc" should be added it also defines FSL_ESDHC_IMX_V2.
> > I think we can put that data in the dt-platdata.c file perhaps.
>
> I thought the same at the beginning, but then I changed my mind, because
>
> 1- in order to work dt-platdata.c will need to include several different
> .h, in this example, only for fsl_esdhc_imx to work, we will need to
> include fsl_esdhc_imx.h where all the flags are defined.

Yes I hit that problem with the tiny-dm experiment and ended up adding
a macro to specify the header.

Do you need FSL_ESDHC_IMX_V2? Is it just to avoid a warning about
usdhc_imx7d_data not being used? If so, we could use _maybe_unused

>
> 2- it case we use #define to avoid having to include several different
> .h probably the errors will be more difficult to catch/debug

Yes we would have to include the real header, not just copy bits out of it.
>
> What do you think?

I'm not sure overall. On the one hand I don't really like hiding C
code inside macros. On the other, it avoids the horrible manual
#ifdefs. So on balance I think your idea is the best approach. We can
always refine it later and it is easier to iterate on this sort of
thing if it is actually being used by some boards.


>
> >
> >> This is alsoAs I comment you in the tread about tiny-dm I think that we
> >> can save some space following your suggestions, and for instance implement
> >>
> >>
> >>> Re naming, perhaps DT_COMPAT() might be better than COMPATIBLE()? Or
> >>> even a name that indicates that it is optional, like DT_OPT_COMPAT() ?
> >>>
> >> I totally agree, naming is very important, and DT_COMPAT() is much better.
> >>
> >> What I don't fully understand is what are the cases for DT_OPT_COMPAT(),
> >> could you please clarify?
> > It's just an alternative name, with OPT meaning optional. But I think
> > we can leave out the OPT.
>
> Thanks for clarifying.

Regards,
SImon