Cabal gets confused by benchmark stanza with a dependency that depends on the library itself

[RyanGlScott]

That's a bit of a confusing title, so hopefully this example makes it clear what I'm talking about. I want to run the benchmarks in vector-binary-instances, but Cabal gets mighty perplexed when trying to do so:

Preprocessing benchmark 'benchmarks' for vector-binary-instances-0.2.3.1...
[1 of 1] Compiling Main             ( benchmarks/Benchmarks.hs, dist/build/benchmarks/benchmarks-tmp/Main.o )

benchmarks/Benchmarks.hs:41:50:
    Overlapping instances for Binary (U.Vector Int)
      arising from a use of ‘decode’
    Matching instances:
      instance (U.Unbox a, Binary a) => Binary (U.Vector a)
        -- Defined in ‘vector-binary-instances-0.2.3.1@vecto_AfoM4V7e5UwKO3jo4zR1YO:Data.Vector.Binary’
      instance (U.Unbox a, Binary a) => Binary (U.Vector a)
        -- Defined in ‘Data.Vector.Binary’
    In the first argument of ‘nf’, namely ‘(decode :: V)’
    In the second argument of ‘($)’, namely ‘nf (decode :: V) bs’
    In the expression: bench "U.Vector Int" $ nf (decode :: V) bs

benchmarks/Benchmarks.hs:48:40:
    Overlapping instances for Binary (U.Vector Int)
      arising from a use of ‘encode’
    Matching instances:
      instance (U.Unbox a, Binary a) => Binary (U.Vector a)
        -- Defined in ‘vector-binary-instances-0.2.3.1@vecto_AfoM4V7e5UwKO3jo4zR1YO:Data.Vector.Binary’
      instance (U.Unbox a, Binary a) => Binary (U.Vector a)
        -- Defined in ‘Data.Vector.Binary’
    In the first argument of ‘nf’, namely ‘encode’
    In the second argument of ‘($)’, namely ‘nf encode vec1’
    In the expression: bench "U.Vector Int 3" $ nf encode vec1

What's confusing about the error message is that the "overlapping" instances both come from Data.Vector.Binary! I think Cabal might be getting tripped up over the fact that the benchmark depends on criterion, but criterion transitively depends on vector-binary-instances itself.

That being said, I don't know what the correct behavior is for this scenario. Should Cabal prevent you from even attempting to build the benchmark? Or is there some way to make this work?

[ezyang]

If you added the full log from Cabal it probably had a warning saying that the versions are inconsistent. Cabal doesn't enforce strong consistency so that you can manually configure a cycle like this (but it is definitely a "you get to keep all the pieces" type situation.) Enforcing consistency is the package manager's job.

You didn't mention it here, but Stack has a hack to allow cyclic dependencies like this (c.f. commercialhaskell/stack#220). So they purposely configured the benchmark suite with two different versions of the package. But since vector-binary-instances defines orphan instances, this is just not going to work; you really need only one vector-binary-instances in the plan. My expert recommendation is to have vector-binary-instances split out the implementation from the instances, and then in the test suite not use the instances to test (just call the implementation directly.)

For reference, cabal-install would reject this install plan. There's a PR to relax this #3290 but it is still buggy.

[RyanGlScott]

Oh, I see what went wrong. I had installed vector-binary-instances before running cabal configure --enable-benchmarks, so Cabal never warned me about the inconsistent versions. If I first create a sandbox and then try cabal install --enable-benchmarks, it errors.

Sorry for the noise! I'll close this as a duplicate of #1575, and try out your advice for fixing vector-binary-instances.