Hi there,

following up on a discussion we had at [1]...

Currently most of the recipes in core and meta-openembeeded are written like

recipe-a_1.0.bb
---
SUMMARY = "Foo"
LICENSE = "Bar"
LIC_FILES_CHKSUM = "file://LICENSE;md5=a4a2bbea1db029f21b3a328c7a059172"

DEPENDS = "some-dependency"
---

if you now think of the following:

local.conf
---
INHERIT += "magic"
---

magic.bbclass
---
DEPENDS = "magic-dependency"
---

then `magic-dependency` doesn't end up in recipe-a's DEPENDS chain.

While when you alter the recipe to

recipe-a_1.0.bb
---
SUMMARY = "Foo"
LICENSE = "Bar"
LIC_FILES_CHKSUM = "file://LICENSE;md5=a4a2bbea1db029f21b3a328c7a059172"

DEPENDS += "some-dependency"
---

or

recipe-a_1.0.bb
---
SUMMARY = "Foo"
LICENSE = "Bar"
LIC_FILES_CHKSUM = "file://LICENSE;md5=a4a2bbea1db029f21b3a328c7a059172"

DEPENDS:append = " some-dependency"
---

it works as expected.

Which bring me to my question:

Do we want to support these cases? even though it would mean that we might have to patch a lot of recipes in multiple layers

Currently I more and more come to the conclusion that append/prepend/+= should be the way to go, as otherwise globally inherited classes loose variable settings without us noticing (except for build failures ofc).

This issue spreads even further as not only DEPENDS is affected, but literally every hard assignment in a recipe that is also having a hard assignment in a globally inherited class.

TBF

the correct fix would be

magic.bbclass
---
DEPENDS:append = " magic-dependency"
---

but unfortunately sometimes this is out of our control (as we don't want to have bbclassappends for totally valid reasons)

Thoughts?

[1] https://lists.openembedded.org/g/openembedded-devel/topic/87562882#94227