I'd like to, within a client program, use the GHC API from a
modified version of GHC. Its easy enough to get the GHC sources
and build and install the modified GHC, but installing the
modified GHC causes all older (Cabal) dependencies to break.
Consequently, I cannot use the GHC API as ghc -v says:
package ghc-7.0.3-... is unusable due to missing or recursive dependencies:...
Any clues on how to proceed?
Thanks!,
Ranjit.
You must build a custom version of GHC first, then install all the packages you need. That way, everything will be linked against your custom GHC, and you will be fine.
Related
I'm pretty new to Haskell, and I think I have a fundamental misunderstanding somewhere. When I'm in GHCi (using the ghci command), I can type import System.Random, and it works. I can then generate random numbers.
Next, I make a file called test.hs that contains nothing but one line: import System.Random. I then call the command ghc test.hs and get the following error message:
test.hs:1:1: error:
Could not find module ‘System.Random’
There are files missing in the ‘random-1.1’ package,
try running 'ghc-pkg check'.
Use -v to see a list of the files searched for.
|
1 | import System.Random
| ^^^^^^^^^^^^^^^^^^^^
However, if I go back to GHCi, I can type :load test.hs. This works, and allows me to generate random numbers.
When I run ghc-pkg check, I get only warnings about missing haddock interface files: https://pastebin.com/6a9f0nYZ. From what I understand, this isn't related to the current issue.
Also, when I run ghc-pkg list, random-1.1 is in the list, so random should be installed.
A couple of questions:
Why would GHC and GHCi have access to different imports? Why is the system set up that way? Maybe I just don't understand the relationship between GHC and GHCi.
According to the error message, there are "files missing." How can I figure out which files?
How can I make it so that I can compile Haskell files that use System.Random?
Edit: Both GHC and GHCi are the same version.
$ ghc --version
The Glorious Glasgow Haskell Compilation System, version 8.6.4
$ ghci --version
The Glorious Glasgow Haskell Compilation System, version 8.6.4
Edit: Both ghc and ghci are in /usr/bin/
$ which ghc
/usr/bin/ghc
$ which ghci
/usr/bin/ghci
Update: It looks like this is an Arch Linux peculiarity, not a corrupt package. I've updated my answer accordingly.
GHCi loads "dynamic" versions of modules. By default, GHC links with "static" versions of modules. In particular, when you run:
> import System.Random
under GHCi, it tries to access the file Random.dyn_hi to get interface information for the module. In contrast, when you compile a file with that import statement in it, GHC tries to access the file Random.hi.
You can verify this is the issue by running ghc-pkg field random import-dirs and peeking in the resulting directory. There should be a System subdirectory that usually has two files in it: System.hi and System.dyn_hi. If the former is missing, that's your problem.
Now, it looks like you're probably using Arch Linux. As documented on the Arch Haskell wiki page in the section "Problems with linking", the Arch Haskell community packages (including haskell-random) intentionally omit the static versions of interface files and libraries.
There are several of workarounds given there:
You can use dynamic linking when compiling with GHC. When using GHC directly, this just means passing the -dynamic flag. For Cabal-based projects, instructions are given on that page for modifying your ~/.cabal/config to use dynamic linking for all projects.
You can install the ghc-static and ghc-pristine packages and set up your path and/or Cabal to use the compiler in /usr/share/ghc-pristine/bin/ghc which will maintain its own separate package database that won't interfere with globally installed Haskell community packages, like haskell-random.
You can install ghc-static to get static versions of the base libraries and then run cabal install --force-reinstalls somepackage for all the non-base packages you need. Note that the Wiki notes that this can be tedious and complicated, since you have to manually determine all the package dependencies.
Now, it looks like you already had ghc-static installed, or when invoking GHC you would have also gotten an error about missing files in the base package. You ran cabal install --force-reinstalls random, though as #dfeuer notes, it might have been safer to run:
$ cabal install --force-reinstalls random-1.1
to ensure that the same version was reinstalled.
Anyway, this installed an additional copy of random in your user-specific package directory. If you run:
$ ghc-pkg list
you'll see that random-1.1 is listed both under the global database and user database:
/usr/lib/ghc-8.6.4/package.conf.d
...
random-1.1
...
/home/xxxx/.ghc/x86_64-linux-8.6.4/package.conf.d
random-1.1
and if you run:
$ ghc-pkg describe random
you'll see that it lists the two separate installed versions which is why you now get duplicate fields with ghc-pkg field random import-dirs.
There shouldn't be anything wrong with this. Your user database will take precedence over the global database, so your newly installed version of random will be used when you run GHCi or GHC.
Note that, if you change your mind and want to back out this reinstallation (and maybe try one of the other solutions suggested on the Wiki), you should be able to run:
$ ghc-pkg unregister --user random
Technically, this won't actually remove the package (as the compiled version will still be there under ~/.cabal/lib), but it should otherwise put things back the way they were.
I'm trying to use nix-shell as a #! wrapper for runghc, as suggested in the manpage. But it cannot find the libraries. Given the following example, cut-down from the manpage
#! /usr/bin/env nix-shell
#! nix-shell -i runghc -p haskellPackages.HTTP
import Network.HTTP
main = return ()
I get the following result:
[~:0]$ ./nixshelltest
nixshelltest:4:1: error:
Failed to load interface for ‘Network.HTTP’
Use -v to see a list of the files searched for.
[~:1]$
to my mind, that's exactly what nix-shell -p is to avoid.
Am I doing something wrong, missing the point, or is there a bug? This is both on a nixOS 17.03 host, and also a host running nix 17.09 on top of Ubuntu.
Thanks,
The environment that you're using to run the script is missing a step. It's got a GHC and a HTTP package, but the compiler doesn't know about the package.
The way GHC and library packages work in nix might be a little "inside out" from what you're expecting. You actually need to install a compiler that "depends on" all of the libraries you want, rather than simply installing the compiler and the library separately. The reason is that GHC is designed to have library packages added by modifying the file tree where GHC is installed. On a mutable file system with only a single system GHC install you would just modify GHC whenever a library was installed, but nix can't. Nix has to be able to install a frozen GHC that never changes, and potentially many of them.
So what happens instead is that you install a tiny wrapper which depends on the both the underlying "raw" GHC install and all of the libraries that you want to use. The wrapper then acts like an install of GHC that had those libraries registered, without actually needing to duplicate an entire GHC install. When you just install a library package on its own it just sits there inert, without any GHC being able to find it just by it existing.
In fact, the script you've shown here doesn't actually specify that it should have a compiler installed at all; it just asks for the HTTP library. When I tried your script I got command not found: runghc. The runghc is only working on your system because it happened to already be in your path when you ran this (perhaps because you have GHC installed in your profile?), and that GHC wasn't installed with the HTTP package and so can't see it. The nix-shell adding just the library to the environment doesn't help.
What you need to do instead is use this line:
#! nix-shell -i runghc -p "ghc.withPackages (ps: [ ps.HTTP ])"
You're not installing either ghc or HTTP directly; instead the ghc.withPackages function computes a nix package that installs a GHC wrapper that knows about the HTTP Haskell package. Under the hood this depends on a "raw" GHC with no additional libraries, and also on the HTTP library and will cause those to be installed too.
If you use lots of different Haskell environments (possibly via nix-shell scripts like this that each need a different set of libraries), then you will end up with a unique withPackages wrapper installed on your system for each combination of libraries you ever use. But that's okay because the wrappers themselves are tiny, and nix is able to share and reuse the underlying GHCs and library packages between all of those environments.
I am trying to install Haskell Platform 2014.2.0.0 from source on Red Hat Enterprise Linux 6.5. I have a functional install of Haskell Platform 2012.4.0.0 and GHC 7.4.2 from two years ago, plus a recently-installed Haskell Platform 2013.2.0.0 and GHC 7.6.3 from JustHub.
I've built GHC 7.8.3 from source, but it keeps coming up with seven failures in the test suite. I have no idea if these test failures are innocuous or not. (The test failures are not relevant to my question, but they may become significant later.)
I unpack the source tarball of 2014.2.0.0, read the README. It says that the way to build this iteration of Haskell is with a shell script, which is invoked:
./platform.sh $PATH_TO_GHC_BINDIST_TARBALL
I don't have a GHC binary distribution tarball. So far as I am able to tell, there is no binary distribution tarball of GHC 7.8.3 for any version of Red Hat Enterprise Linux. I have a built GHC 7.8.3. How do I tell platform.sh -- or whatever is underneath it -- that there is no tarball, and it should just use what's in $PATH? Alternately, how do I pack up my existing install of GHC 7.8.3 so that platform.sh will accept it?
The built GHC does not have a 'cabal' command, so the cabal commands in platform.sh are falling back to $PATH, which I can configure to be either of the other installed versions (2013.2/7.6.3 or 2012.4/7.4.2). It doesn't seem to make a difference: neither one recognizes 'cabal --sandbox'. Both result in complaints that I should run 'cd hptool ; cabal install --only-dependencies', which I've done, repeatedly. platform.sh never gets past that point.
If I run the commands in platform.sh by hand, I get to 'cd hptool; cabal build', which errors out: "cabal-1.16.0.2: Run the 'configure' command first.". But there is no 'configure' command available in the hptool directory.
I'm now stuck. How do I build Haskell Platform 2014 on RHEL 6?
You need to use your GHC sources to make your own "bindist." Directions at https://ghc.haskell.org/trac/ghc/wiki/MakingReleases
I managed to get Haskell Platform installed, and functional. I ended up abandoning platform.sh and just hand-installed all the packages in the Haskell Platform tarball -- and their dependencies -- with manual cabal commands. Along with the broken platform.sh, I ran into many problems on the way.
The ones that I remember:
platform.sh will never succeed if you have the stock Haskell Platform 2013 or previous installed. It wants a cabal that recognizes the '--sandbox' option, and cabal 1.18 doesn't know that option. You must have a newer cabal installed than Haskell Platform 2013 provides. (GHC 7.4 or 7.6 appears to be fine, though.)
I had an existing .cabal and .ghc directory, which had incompatible builds and/or versions of various packages. I deleted both directories numerous times while testing things.
cabal install --global behaves rather differently from the default cabal install --user. .cabal contained something useful after I did 'cabal install cabal-install'. It took two or three tries at that to figure out where the new cabal binary went.
ghc and cabal pick up new libraries in .ghc and .cabal, but not new binaries.
Neither GHC nor cabal install default to --enable-shared, except when something wants it. I had to go rebuild everything before that -- all the way back to GHC 7.8.3 itself -- with --enable-shared once that happened.
haddock is ridiculously tightly bound to the version of GHC it was built with. I had to rebuild it to get --enable-documentation to work for anything built with GHC 7.8.3.
The test packages and the text package are so tightly integrated that they have circular dependencies if you try to do 'cabal install text --enable-tests'. Even after installing the most recent version of text and the test packages, cabal still wouldn't run the text test suite, so I gave up and installed text without testing it.
My default environment includes 'LC_ALL=C'. This trips a known bug in cabal -- apparently in all versions -- that breaks some package builds. To work around it, I had to shift it to 'LC_ALL=en_US.utf8'. I have no idea if the packages affected will work if you have LC_ALL, or any of the other locale variables such as LANG or LC_<anything else>, set to C.
cabal install --global is terribly inconsistent about where packages get stored. We split out individual packages into their own subdirectories, and then build a symlink tree in a known place out of all those subdirectories. So ghc is in its own subdirectory of /usr/sup/ghc-7.8.3; Haskell Platform is in another subdirectory, /usr/sup/haskell-platform-2014.2.0.0. I consistently used --prefix=/usr/sup/haskell-platform-2014.2.0.0 on every 'cabal install' command, but even then, some libraries ended up in /usr/sup/ghc-7.8.3.
Both GHC and Haskell Platform have a dictionary of what's built and where it is -- perhaps as a workaround for the install location inconsistency -- in /usr/sup/ghc-7.8.3/lib/ghc-7.8.3/package.conf.d/package.cache. If that package dictionary isn't world-readable, ghc breaks. What it should do is look at the actual file structure to find things. Given that ghc breaks if the dictionary isn't available, the file shouldn't be called 'cache', as a cache miss shouldn't cause catastrophic failure. Perhaps rename it to 'package-mandatory-dictionary'?
Ultimately, it all installed, but I have to wonder about the damage I did from so much banging my head on the wall.
again...
I'm messing around with profiling in GHC after recently starting to learn Haskell. I'm trying to use profiling to see how different implementations of functions vary in performance, for example using list comprehension instead of map. I'm trying to compile with the -prof flag but im getting the following output:
david#david-LinuxMint ~/Desktop/Sandbox/Haskell/a $ ghc --make Filt -prof
Filt.hs:1:1:
Could not find module `Prelude'
Perhaps you haven't installed the profiling libraries for package `base'?
Use -v to see a list of the files searched for
The program compiles fine with the -prof flag omitted. Anybody have any ideas where im going wrong? I've tried to find something on SO/internet but my Google-Fu is failing me? Cheers.
This is for Linux, Debian in particular, but I believe the OP is using Mint which I think might be a Debian variant.
You probably got base from the OS packaging system (Debian: APT), so you'll also want to get the profiling version of base from that same system. On Debian the hackage package "foo" is in the Debian package "libghc-foo-dev"[1]. The profiling version of "libghc-blah-dev" is in the Debian package "libghc-blah-prof".
Now, base is a little special. Those packages exist, but are "virtual"; they are listed in relationships, but are actually installed by (one or more) other real packages. In Debian virtual libghc-base-dev is provided by real package ghc and virtual libghc-base-dev is provided by real package ghc-prof.
So, install ghc-prof (or the Mint equivalent) and you'll be good, or at least move on the the next error.
Eventually, you'll end up downloading and building packages directly from hackage. You'll have to change your cabal settings to build the profiling versions of those libraries.
[1] If another Haskell compiler gets (back) into Debian, the "ghc" in "libghc" will change to a string suitable for that compiler.
I have successfully built PortFusion with the brand new 64-bit GHC 7.6.1 Release Candidate 1 for Windows.
Using freshly downloaded native 64-bit mingw binaries from http://www.drangon.org/mingw, the network package was as easy to install (after a bunch of non-relevant small fixes) as a simple
CC=mingw64/bin/gcc cabal install
There is also an LLVM toolchain package on the same website.
Now I wonder how I could tell GHC to use specific LLVM binaries during compilation.
Would it be something as simple and similar to above as:
#v??v
LLVM=????????? ghc -W -O2 -fllvm -optlo-O3 --make src/Main.hs
LLVM=????????? cabal install PortFusion -f llvm #¹
#^??^
¹ relevant line in PortFusion.cabal defining the llvm flag
or completely different?
For the record, the answer to the question in the title is
ghc -pgmlo opt_cmd -pgmlc llc_cmd -fllvm ...
opt_cmd and llc_cmd can be either command names (that will be looked up in PATH) or full pathnames.
You may be able to get GHC to use certain binaries by altering where they are defined in the PATH environment variable. Earlier takes precedence. Presumably, System PATH is also higher precedence than User PATH.