new-build: Repeating build fails

[BardurArantsson]

Repeating a new-build command twice fails on my project with the error message

improveInstallPlanWithUpToDatePackages: dep on non lib package
CallStack (from HasCallStack):
error, called at ./Distribution/Client/ProjectBuilding.hs:358:7 in
main:Distribution.Client.ProjectBuilding

the second time I run the build. The command I'm running is

cabal new-build cqrs-{core,example,memory,postgresql,testkit}

and the repository is current 'master' here: https://github.com/BardurArantsson/cqrs/tree/master

[BardurArantsson]

Oh, and the exact cabal-install version I'm running is commit b7b666e

[23Skidoo]

Any chance we can get a small test case?

[BardurArantsson]

I have absolutely no idea how I would go about making such a test case as I'm not familiar with any of the new-build stuff :).

Maybe someone who is can chime in on what I could try?

[23Skidoo]

OK, since your library is open source the issue should be reproducible.

[BardurArantsson]

Hopefully, yes.

[ezyang]

TBF, this is a lot of deps ;) (currently building...)

[ezyang]

I cannot repro. I am also on b7b666e

Do I have to edit some files?

[BardurArantsson]

No, all my files are exactly as in that repo.

I'm assuming you did the "repeat the command" bit? (Just making sure.)

If it doesn't reproduce for you then some global state must be "leakin in" somewhere... Hum.

OK, I'll try wiping everything I can and try to build again. Anything I should wipe besides ~/.cabal and ~/.ghc...?

[ezyang]

ezyang@sabre:~/Dev/labs/cqrs$ cabal new-build cqrs-{core,example,memory,postgresql,testkit}
In order, the following will be built (use -v for more details):
cqrs-example-0.11.0
Preprocessing executable 'cqrs-example' for cqrs-example-0.11.0...
ezyang@sabre:~/Dev/labs/cqrs$ cabal new-build cqrs-{core,example,memory,postgresql,testkit}
In order, the following will be built (use -v for more details):
cqrs-example-0.11.0
Preprocessing executable 'cqrs-example' for cqrs-example-0.11.0...

For some reason new-build is repeatedly building cqrs-example, which does seem like a bug.

Before you wipe everything, first try wiping dist-newstyle and trying again.

[BardurArantsson]

I did try dist-newstyle, but same result :( .

I'll try the "full wipe" tomorrow and report back.

[BardurArantsson]

OK, I've tried the full wipe of ~/.ghc and ~/.cabal. Issue remains.

Can you think of any other system-wide state that might influence this?

@ezyang : Just out of curiousity... did you do a completely clean build? (That is, with ~/.ghc and ~ /.cabal at least moved out of the way.)

[ezyang]

Actually, I've reproduced this, but on a different project! The problem goes away if you make a modification to the project in question.

[grayjay]

I just encountered this when building cabal with cabal new-build cabal. I was using commit 9fea99d.

[ezyang]

So, for #3644, I'm going to be rewriting this chunk of code. Maybe the problem will go away after I've done so.

[grayjay]

I ran git bisect and got commit 97cff5c.

[ezyang]

OK... my guess what is going on is that improveInstallPlanWithUpToDatePackages used to handle the case where there was no cached InstalledPackageInfo, but when Duncan refactored it, he made it partial in this case.

Unfortunately, we do have to handle this case, because an inplace executable may be improved such that we don't need to rebuild it. I'll have to analyze preexisting more carefully to understand what we need to do here.

The steps to reproduce should be to build a package with an executable but no library twice.

[ezyang]

So I guess there are two possible ways to fix this:

1. Add a new case to GenericPlanPackage to represent installed executables (which don't have an InstalledPackageInfo)
2. Don't bother updating the InstallPlan. It shouldn't matter if we have an "updated" entry in it, because no one can depend on it, and so we'll never actually query for it.

[ezyang]

This is now fixed.