[PATCH v2] arm-trusted-firmware-mvebu: stop cluttering Image Builder

[Tomasz Maciej Nowak]

W dniu 7.09.2022 o 22:57, Hauke Mehrtens pisze:
> On 9/7/22 19:23, Tomasz Maciej Nowak wrote:
>> W dniu 7.09.2022 o 00:15, Hauke Mehrtens pisze:
>>> On 8/31/22 17:03, Tomasz Maciej Nowak wrote:
>>>> From: Tomasz Maciej Nowak <tmn505 at gmail.com>
>>>>
>>>> All contents of staging_dir/image are included in Image Builder (IB) in
>>>> case some binary needs to be included in final image. But in case of
>>>> this package, all sources are stored there and those clutter the final
>>>> tarball of IB for no reason. Those sources are not used during image
>>>> creation and are just dead weight. To put it in perspective, the IB for
>>>> 21.02.0 is 158 MiB, 22.03.0-rc6 is 366 MiB and snapshot is over 620 MiB!
>>>> To fix it, put them in package build directory, so they won't end up
>>>> included in IB tarball. With that, the custom clean definition can be
>>>> removed, as the default one will take over.
>>>>
>>>> Signed-off-by: Tomasz Maciej Nowak <tmn505 at gmail.com>
>>>> Reviewed-by: Andre Heider <a.heider at gmail.com>
>>>> ---
>>>> v1 -> v2
>>>> - as pointed by Andre, we can use default clean definition
>>>> - add his review tag
>>>
>>> I think the build times are much higher with this change compared to before.
>>
>> Yes, that's correct, since all source is extracted/rebuilt for each variant.
>> The toolchain tarball is huge and it takes lot of time to extract. The second
>> culprit is cryptopp. It's rebuilt for each variant with one job.
>>
>>> I haven't found the time to do real measurements. I used this configuration to test:
>>> CONFIG_TARGET_mvebu=y
>>> CONFIG_TARGET_mvebu_cortexa53=y
>>> CONFIG_TARGET_mvebu_cortexa53_DEVICE_globalscale_espressobin=y
>>
>> I did measurements, before change:
>> time make -j4 package/arm-trusted-firmware-mvebu/compile
>> [ ... ]
>> real    13m17,046s
>> user    25m24,272s
>> sys     1m49,963s
>>
>> with the change:
>> time make -j4 package/arm-trusted-firmware-mvebu/compile
>> [ ... ]
>> real    28m26,566s
>> user    79m45,906s
>> sys     4m6,547s
>>
>> As You can see it's more than double. I also did test where $(BUILD_DIR)
>> was used:
>> time make -j4 package/arm-trusted-firmware-mvebu/compile
>> [ ... ]
>> real    13m19,462s
>> user    25m21,511s
>> sys     1m53,047s
>>
>> which looks same as in current state (ignore spinning rust hiccups). Should
>> I use $(BUILD_DIR) for the sources?
> 
> did you mean a common folder in $(BUILD_DIR) with your $(BUILD_DIR) approach?

Simply $(BUILD_DIR), which in our example translates to
"build_dir/target-aarch64_cortex-a53_musl" without any subdir. Naturally each
source is untarred to own subdir.

> I think it would be better to place the bootloader toolchain in some common place and not duplicated for each variant.

That's with current state using STAGING_DIR_IMAGE or with using BUILD_DIR,
so I'll use BUILD_DIR instead of PKG_BUILD_DIR.

> Is cryptopp also common for all variant and only used on the host?

Yes to both.

> Maybe we can make this all a host package?

That would make sense if we would untangle the whole TF-A build for mvebu,
which at current form, needs work, as we already discussed that [1]. Having
a second look at the build flow of it, I saw it's even more entangled than 
what I described in that topic. Anyway, lets keep it in current state, until
me/someone finds time to fix it.

1. http://lists.openwrt.org/pipermail/openwrt-devel/2020-October/031677.html

> Hauke

Regards, Tomek

-- 
TMN