Conditionals in project files, processed in order or not?

[phadej]

If I have

packages:   indexed-traversable
packages:   indexed-traversable-instances
if impl(ghc <9.12)
  tests: False
  packages: indexed-traversable-benchmarks

tests:      True
benchmarks: True

it seems that with GHC-9.10 (i.e. when conditional is true), the tests are turned off, even lexically tests: True comes later.

This is consistent with how conditionals work in .cabal files (conditional sections processed as if they appear last), but it's undocumented in https://cabal.readthedocs.io/en/3.12/cabal-project-description-file.html#conditionals-and-imports so I report this as a bug.

[Vekhir]

if impl(ghc <9.12)

should to be false with GHC 9.12, shouldn't it?
Given the rest of the description, I'd expect if impl(ghc >=9.12) is meant.

[phadej]

Edited: with GHC-9.10

[phadej]

Can someone comment what's the behavior. Are conditionals processed interspersed with top-level fields or are contents reordered?

[ffaf1]

When reimplementing cabal check, I recall that CondTree and friends had this tricky behaviour.

cabal/Cabal-syntax/src/Distribution/Types/CondTree.hs

Lines 57 to 62 in 54d364d

	data CondTree v c a = CondNode
	{ condTreeData :: a
	, condTreeConstraints :: c
	, condTreeComponents :: [CondBranch v c a]
	}
	deriving (Show, Eq, Data, Generic, Functor, Foldable, Traversable)

Check the type and it is clear that conditionals are not “interspersed”. The same CondTree is used for projects (see ProjectConfigSkeleton).

So yes, I would say it is like in .cabal files.

[phadej]

So yes, I would say it is like in .cabal files.

I'll welcome the documentation calling that out. Please don't ask people to look into how things are implemented to figure out how they work.

I only found out about this because I was aware this might be the case. Someone might get surprised by this behavior. Especially when they "quickly" and temporarily add a non conditional override at the end of project file when doing something else, the behavior might be still be surprising, but if it's in the docs they will know that it's expected.

EDIT: changing this behavior will be breaking change, and there are no better way to find out breaking changes when you find yourself updating i.e. changing the documentation (and tests).