Hi Tom,
Am 2019-12-06 00:58, schrieb Tom Rini:
On Fri, Dec 06, 2019 at 12:27:39AM +0100, Michael Walle wrote:
Hi Joe, Hi Tom,
Am 2019-12-05 16:55, schrieb Joe Hershberger:
> Hi Michael,
>
> On Fri, Oct 25, 2019 at 7:28 PM Michael Walle <mich...@walle.cc> wrote:
> >
> > Provide functions to read and write the Atheros debug registers.
> >
> > Signed-off-by: Michael Walle <mich...@walle.cc>
>
> This series is adding too much size to several of the boards' SPL it
> seems.
>
> https://travis-ci.org/jhershbe/u-boot/builds/620804934
>
> Please address this and resend.
So first of all, this was the old series. There was a v2 series, but
unfortunately, I've forgot to add the mailing to the recipients, so it
never ended up in the patchwork system. sorry :(
I've resend the v2 series here:
https://patchwork.ozlabs.org/project/uboot/list/?series=146771
Now coming to the real problem here. The sizes, or like some boards
handle
the SPL stuff. Btw. I could not reproduce it on the
am335x_boneblack_vboot
board with a gcc-8. I've seen the travis ci job uses the gcc-7 so this
also
depends on the gcc. gcc-8 seems to produce smaller code, because on
the
am335x_evm the overflow was only by some 100 bytes.
So taking the am335x_evm board for example. It has the following
options
set:
CONFIG_DM_ETH=y
CONFIG_SPL_NET_SUPPORT=y
CONFIG_PHY_ATHEROS=y (this one is set in the config.h!)
So adding a new binding for the phy obviously increases the code size.
But
the hard question is, how could that be fixed. IMHO the board has
wrong
settings. I really don't know how that could be "fixed" other than not
applying this series. Well, we could make the additions conditional
and
introduce a new Kconfig setting, but that is a relly ugly hack and
won't
last long, would it? Doh!
So, the gcc-7 from kernel.org is the min required and must work
toolchain.
Ok.
Maybe once gcc-9 is mature enough for people to have made
stand-alone toolchains for it we'll move up to that but gcc-8 for
everyone ends up being too hard. For the boneblack_vboot config, we
could just drop SPL networking, it's not super critical to that
particular example. But am335x_evm is the kitchen-sink EVM and it is
used and supported there.
That said, looking over the u-boot-spl.map, it looks like nfs stuff
doesn't get discarded for some reason, I'm going to look in to that.
Well but that is just some kind of band aid. These two boards were just
an example. There were also other boards which failed. I doubt the DM
stuff is getting smaller (I know there is some discussion how to
improve the situation). But there seems to be no way to disable the
ethernet device model for the SPL, correct? And unfortunately, this
also pulls the PHY library (with potential DM bindings).
-michael
