The db path i searched in order from left to right; --package-db appends on the right, whereas stack and cabal normally replace the second entry on the path. (The first is special; it *must* contain the base package with the runtime system. iirc this has caused stack issues in the past, when it needed to override some package other than base living in the first package db. It's also why cabal can't hide stuff installed in the global db.) So with --package-db, cabal will see the user package db before the sandbox --- when it needs to see the sandbox either before or in place of the user db.

On Friday, August 4, 2017, Emil Axelsson <78emil@gmail.com> wrote:
Ah, I see. I still don't understand what went wrong in this particular case, but at least it makes sense that there's a difference between `-package-db` and `cabal exec`.

I'll leave it at that.

Thanks!

/ Emil

Den 2017-08-04 kl. 18:45, skrev Brandon Allbery:
GHC_PACKAGE_PATH is the only one of those that it knows about. And the
difference is that --package-db appends to the package database path,
whereas GHC_PACKAGE_PATH sets the entire path. You would need an additional
option to override the *user* entry on the standard package db path with
that of the sandbox, to be compatible with how sandboxes work using only
command line options.

On Friday, August 4, 2017, Emil Axelsson <78emil@gmail.com> wrote:

I guess I get your point in the general case with various tools working
together. But in this case, only ghci is used, and it seems that it gets
some information from the variables `CABAL_SANDBOX_CONFIG`,
`CABAL_SANDBOX_PACKAGE_PATH` and `GHC_PACKAGE_PATH` (set by `cabal exec`)
that it doesn't get from the flag `-package-db`.

At least I find that a little strange.

/ Emil

Den 2017-08-04 kl. 16:58, skrev Brandon Allbery:

ghc including ghci does not know about sandboxes. At all. That is what
cabal/stack exec is for; it runs a command inside the sandbox.

There are sandboxing mechanisms that don't require this, but they require
your shell dotfiles to be set up in a way almost nobody does these days
(proper separation of environment variables; otherwise you get things like
the sandbox's package database, but the wrong $PATH). hsenv used to work
that way.

On Friday, August 4, 2017, Emil Axelsson <78emil@gmail.com
<javascript:_e(%7B%7D,'cvml','78emil@gmail.com');>> wrote:

Thanks for your reply! But I actually don't want `cabal repl`. The bigger
picture is explained here:

http://fun-discoveries.blogspot.com/2017/08/building-haskell
-projects-with-ghc.html

Towards the end of that post I suggest using `cabal exec` to integrate
with a Cabal sandbox; however, as Daniel Trstenjak pointed out, that
shouldn't be needed since `cabal-cargs` already sets the `-package-db`
flag
for GHC.

So my question is why `cabal exec` is needed (in this particular case)
even though `-package-db` is given?

I know `cabal exec` sets a few environment variables
(`CABAL_SANDBOX_CONFIG`, `CABAL_SANDBOX_PACKAGE_PATH` and
`GHC_PACKAGE_PATH`). I've checked that these are all set correctly, so
Cabal seems to be doing its job. But it's not clear why ghci gets
confused
when these variables are not set (and `-package-db` is given), but not
when
the variables are set.

Cheers

/ Emil

Den 2017-08-04 kl. 15:31, skrev nek0:

Hi Emil,

What you want is hidden behind the `cabal repl` command, which starts
ghci with the package-db of the sandbox.

Greetings,

nek0

On 4.8.2017 13:41, Emil Axelsson wrote:

Hi!

I have a small file Test.hs alone in a directory:

       {-# LANGUAGE DeriveGeneric #-}

       module Test where

       import Data.Hashable
       import Data.Scientific
       import GHC.Generics

       data Sc = Sc Scientific deriving (Generic)

       instance Hashable Sc

To be able to load this file, I set up a Cabal sandbox:

       $ ghc --numeric-version
       8.0.2

       $ cabal --numeric-version
       1.24.0.2

       $ cabal sandbox init
       ...

       $ cabal install hashable-1.2.6.0 scientific
       ...

(Note: not the latest version of hashable.)

Now, if I try to run GHCi and point it to the sandbox' package database
I get this error:

       $ ghci
-package-db=.cabal-sandbox/x86_64-linux-ghc-8.0.2-packages.conf.d
Test.hs GHCi, version 8.0.2: http://www.haskell.org/ghc/  :? for help
       [1 of 1] Compiling Test             ( Test.hs, interpreted )

       Test.hs:12:14: error:
           • No instance for (Hashable Scientific)
               arising from a use of
‘hashable-1.2.6.1:Data.Hashable.Class.$dmhashWithSalt’
           • In the expression:
               hashable-1.2.6.1:Data.Hashable.Class.$dmhashWithSalt @Sc
             In an equation for ‘hashWithSalt’:
                 hashWithSalt
                   = hashable-1.2.6.1:Data.Hashable
.Class.$dmhashWithSalt
@Sc
             In the instance declaration for ‘Hashable Sc’

Somehow it mixes in version 1.2.6.1 of hashable, even though this
package isn't installed (neither in the sandbox nor the global
database).

It turns out that wrapping the command in `cabal exec` fixes the
problem:

       $ cabal exec -- ghci
-package-db=.cabal-sandbox/x86_64-linux-ghc-8.0.2-packages.conf.d
Test.hs
       GHCi, version 8.0.2: http://www.haskell.org/ghc/  :? for help
       [1 of 1] Compiling Test             ( Test.hs, interpreted )
       Ok, modules loaded: Test.
       *Test>

Any idea what's going on?

/ Emil
_______________________________________________
Haskell-Cafe mailing list
To (un)subscribe, modify options or view archives go to:
http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe
Only members subscribed via the mailman list are allowed to post.

_______________________________________________

Haskell-Cafe mailing list
To (un)subscribe, modify options or view archives go to:
http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe
Only members subscribed via the mailman list are allowed to post.








--
brandon s allbery kf8nh                               sine nomine associates
allbery.b@gmail.com                                  ballbery@sinenomine.net
unix, openafs, kerberos, infrastructure, xmonad        http://sinenomine.net