ghcup: ghc toolchain installer

ghcup makes it easy to install specific versions of ghc on GNU/Linux, macOS (aka Darwin), FreeBSD and Windows and can also bootstrap a fresh Haskell developer environment from scratch. It follows the unix UNIX philosophy of do one thing and do it well.

Similar in scope to rustup, pyenv and jenv.

Table of Contents

• Installation
  • Supported platforms
  • Manual install
  • Vim integration
• Usage
  • Configuration
  • GPG verification
  • Manpages
  • Shell-completion
  • Compiling GHC from source
  • XDG support
  • Env variables
  • Installing custom bindists
  • Isolated Installs
  • CI
  • Tips and tricks
• Design goals
• How
• Known users
• Known problems
• FAQ

Installation

Most users should follow the instructions at https://www.haskell.org/ghcup/. Advanced users may want to perform a manual installation.

Supported platforms

This list may not be exhaustive and specifies support for bindists only.

Platform	Architecture	ghcup	GHC	cabal	HLS	stack
Windows 7	amd64	❔	✅	✅	✅	✅
Windows 10	amd64	✅	✅	✅	✅	✅
Windows Server 2016	amd64	✅	✅	✅	✅	✅
Windows Server 2019	amd64	✅	✅	✅	✅	✅
Windows Server 2022	amd64	✅	✅	✅	✅	✅
Windows WSL1	amd64	❌	❔	❔	❔	❔
Windows WSL2	amd64	✅	✅	✅	✅	✅
MacOS >=13	amd64	✅	✅	✅	✅	✅
MacOS <13	amd64	❌	❔	❔	❔	❔
MacOS	aarch64	✅	✅	✅	⚠️	❌
FreeBSD	amd64	✅	⚠️	✅	⚠️	❌
Linux generic	x86	✅	✅	✅	✅	✅
Linux generic	amd64	✅	✅	✅	✅	✅
Linux generic	aarch64	✅	⚠️	✅	⚠️	❌
Linux generic	armv7	✅	⚠️	✅	⚠️	❌

Windows 7

May or may not work, several issues:

• https://gitlab.haskell.org/haskell/ghcup-hs/-/issues/140
• https://gitlab.haskell.org/haskell/ghcup-hs/-/issues/197

WSL1

Unsupported. GHC may or may not work. Upgrade to WSL2.

MacOS <13

Not supported. Would require separate binaries, since >=13 binaries are incompatible. Please upgrade.

MacOS aarch64

HLS bindists are still experimental. Stack is theoretically supported, but has no binaries yet.

FreeBSD

Lacks some upstream bindists and may need compat libs, since most bindists are built on FreeBSD-12. HLS bindists are experimental.

Linux ARMv7/AARCH64

Lower availability of bindists. HLS only has experimental ones. Stack not supported currently.

Manual install

Download the binary for your platform at https://downloads.haskell.org/~ghcup/ and place it into your PATH anywhere.

Then adjust your PATH in ~/.bashrc (or similar, depending on your shell) like so:

export PATH="$HOME/.cabal/bin:$HOME/.ghcup/bin:$PATH"

Vim integration

See ghcup.vim.

Usage

See ghcup --help.

For the simple interactive TUI, run:

ghcup tui

For the full functionality via cli:

# list available ghc/cabal versions
ghcup list

# install the recommended GHC version
ghcup install ghc

# install a specific GHC version
ghcup install ghc 8.2.2

# set the currently "active" GHC version
ghcup set ghc 8.4.4

# install cabal-install
ghcup install cabal

# update ghcup itself
ghcup upgrade

GHCup works very well with cabal-install, which handles your haskell packages and can demand that a specific version of ghc is available, which ghcup can do.

Configuration

A configuration file can be put in ~/.ghcup/config.yaml. The default config file explaining all possible configurations can be found in this repo: config.yaml.

Partial configuration is fine. Command line options always override the config file settings.

GPG verification

GHCup supports verifying the GPG signature of the metadata file. The metadata file then contains SHA256 hashes of all downloads, so this is cryptographically secure.

First, obtain the gpg key:

gpg --batch --keyserver keys.openpgp.org --recv-keys 7784930957807690A66EBDBE3786C5262ECB4A3F

Then verify the gpg key in one of these ways:

1. find out where I live and visit me to do offline key signing
2. figure out my mobile phone number and call me to verify the fingerprint
3. more boring: contact me on Libera IRC (maerwald) and verify the fingerprint

Once you've verified the key, you have to figure out if you trust me.

If you trust me, then you can configure gpg in ~/.ghcup/config.yaml:

gpg-setting: GPGLax # GPGStrict | GPGLax | GPGNone

In GPGStrict mode, ghcup will fail if verification fails. In GPGLax mode it will just print a warning. You can also pass the mode via ghcup --gpg <strict|lax|none>.

Manpages

For man pages to work you need man-db as your man provider, then issue man ghc. Manpages only work for the currently set ghc. MANPATH may be required to be unset.

Shell-completion

Shell completions are in scripts/shell-completions directory of this repository.

For bash: install shell-completions/bash as e.g. /etc/bash_completion.d/ghcup (depending on distro) and make sure your bashrc sources the startup script (/usr/share/bash-completion/bash_completion on some distros).

Compiling GHC from source

Compiling from source is supported for both source tarballs and arbitrary git refs. See ghcup compile ghc --help for a list of all available options.

If you need to overwrite the existing build.mk, check the default files in data/build_mk, copy them somewhere, adjust them and pass --config path/to/build.mk to ghcup compile ghc. Common build.mk options are explained here.

Make sure your system meets all the prerequisites.

Cross support

ghcup can compile and install a cross GHC for any target. However, this requires that the build host has a complete cross toolchain and various libraries installed for the target platform.

Consult the GHC documentation on the prerequisites. For distributions with non-standard locations of cross toolchain and libraries, this may need some tweaking of build.mk or configure args. See ghcup compile ghc --help for further information.

XDG support

To enable XDG style directories, set the environment variable GHCUP_USE_XDG_DIRS to anything.

Then you can control the locations via XDG environment variables as such:

• XDG_DATA_HOME: GHCs will be unpacked in ghcup/ghc subdir (default: ~/.local/share)
• XDG_CACHE_HOME: logs and download files will be stored in ghcup subdir (default: ~/.cache)
• XDG_BIN_HOME: binaries end up here (default: ~/.local/bin)
• XDG_CONFIG_HOME: the config file is stored in ghcup subdir as config.yaml (default: ~/.config)

Note that ghcup makes some assumptions about structure of files in XDG_BIN_HOME. So if you have other tools installing e.g. stack/cabal/ghc into it, this will likely clash. In that case consider disabling XDG support.

Env variables

This is the complete list of env variables that change GHCup behavior:

• GHCUP_USE_XDG_DIRS: see XDG support above
• TMPDIR: where ghcup does the work (unpacking, building, ...)
• GHCUP_INSTALL_BASE_PREFIX: the base of ghcup (default: $HOME)
• GHCUP_CURL_OPTS: additional options that can be passed to curl
• GHCUP_WGET_OPTS: additional options that can be passed to wget
• GHCUP_GPG_OPTS: additional options that can be passed to gpg
• GHCUP_SKIP_UPDATE_CHECK: Skip the (possibly annoying) update check when you run a command
• CC/LD etc.: full environment is passed to the build system when compiling GHC via GHCup

Installing custom bindists

There are a couple of good use cases to install custom bindists:

1. manually built bindists (e.g. with patches)

• example: ghcup install ghc -u 'file:///home/mearwald/tmp/ghc-eff-patches/ghc-8.10.2-x86_64-deb10-linux.tar.xz' 8.10.2-eff

2. GHC head CI bindists

• example: ghcup install ghc -u 'https://gitlab.haskell.org/api/v4/projects/1/jobs/artifacts/master/raw/ghc-x86_64-fedora27-linux.tar.xz?job=validate-x86_64-linux-fedora27' head

3. DWARF bindists

• example: ghcup install ghc -u 'https://downloads.haskell.org/~ghc/8.10.2/ghc-8.10.2-x86_64-deb10-linux-dwarf.tar.xz' 8.10.2-dwarf

Since the version parser is pretty lax, 8.10.2-eff and head are both valid versions and produce the binaries ghc-8.10.2-eff and ghc-head respectively. GHCup always needs to know which version the bindist corresponds to (this is not automatically detected).

Isolated installs

Ghcup also enables you to install a tool (GHC, Cabal, HLS, Stack) at an isolated location of your choosing. These installs, as the name suggests, are separate from your main installs and DO NOT conflict with them.

• No symlinks are made to these isolated installed tools, you'd have to manually point to them wherever you intend to use them.

• These installs, can also NOT be deleted from ghcup, you'd have to go and manually delete these.

You need to use the --isolate or -i flag followed by the directory path.

Examples:-

1. install an isolated GHC version at location /home/user/isolated_dir/ghc/

   • ghcup install ghc 8.10.5 --isolate /home/user/isolated_dir/ghc

2. isolated install Cabal at a location you desire

   • ghcup install cabal --isolate /home/username/my_isolated_dir/

3. do an isolated install with a custom bindist

   • ghcup install ghc --isolate /home/username/my_isolated_dir/ -u 'https://gitlab.haskell.org/api/v4/projects/1/jobs/artifacts/master/raw/ghc-x86_64-fedora27-linux.tar.xz?job=validate-x86_64-linux-fedora27' head

4. isolated install HLS

   • ghcup install hls --isolate /home/username/dir/hls/

5. you can even compile ghc to an isolated location.

   • ghcup compile ghc -j 4 -v 9.0.1 -b 8.10.5 -i /home/username/my/dir/ghc

   ---

CI

On windows, ghcup can be installed automatically on a CI runner like so:

Set-ExecutionPolicy Bypass -Scope Process -Force;[System.Net.ServicePointManager]::SecurityProtocol = [System.Net.ServicePointManager]::SecurityProtocol -bor 3072;Invoke-Command -ScriptBlock ([ScriptBlock]::Create((Invoke-WebRequest https://www.haskell.org/ghcup/sh/bootstrap-haskell.ps1 -UseBasicParsing))) -ArgumentList $false,$true,$true,$false,$false,$false,$false,"C:\"

On linux/darwin/freebsd, run the following on your runner:

curl --proto '=https' --tlsv1.2 -sSf https://get-ghcup.haskell.org | BOOTSTRAP_HASKELL_NONINTERACTIVE=1 BOOTSTRAP_HASKELL_MINIMAL=1 sh

This will just install ghcup and on windows additionally msys2.

Example github workflow

On github workflows you can use https://github.com/haskell/actions/

If you want to install ghcup manually though, here's an example config:

name: Haskell CI

on:
  push:
    branches: [ master ]
  pull_request:
    branches: [ master ]

jobs:
  build-cabal:

    runs-on: ${{ matrix.os }}
    strategy:
      fail-fast: false
      matrix:
        os: [ubuntu-latest, macOS-latest, windows-latest]
        ghc: ['8.10.7', '9.0.1']
        cabal: ['3.4.0.0']

    steps:
    - uses: actions/checkout@v2

    - if: matrix.os == 'windows-latest'
      name: Install ghcup on windows
      run: Set-ExecutionPolicy Bypass -Scope Process -Force;[System.Net.ServicePointManager]::SecurityProtocol = [System.Net.ServicePointManager]::SecurityProtocol -bor 3072;Invoke-Command -ScriptBlock ([ScriptBlock]::Create((Invoke-WebRequest https://www.haskell.org/ghcup/sh/bootstrap-haskell.ps1 -UseBasicParsing))) -ArgumentList $false,$true,$true,$false,$false,$false,$false,"C:\"

    - if: matrix.os == 'windows-latest'
      name: Add ghcup to PATH
      run: echo "/c/ghcup/bin" >> $GITHUB_PATH
      shell: bash

    - if: matrix.os != 'windows-latest'
      name: Install ghcup on non-windows
      run: curl --proto '=https' --tlsv1.2 -sSf https://get-ghcup.haskell.org | BOOTSTRAP_HASKELL_NONINTERACTIVE=1 BOOTSTRAP_HASKELL_MINIMAL=1 sh

    - name: Install ghc/cabal
      run: |
        ghcup install ghc ${{ matrix.ghc }}
        ghcup install cabal ${{ matrix.cabal }}
      shell: bash

    - name: Update cabal index
      run: cabal update
      shell: bash

    - name: Build
      run: cabal build --enable-tests --enable-benchmarks
      shell: bash

    - name: Run tests
      run: cabal test
      shell: bash

Tips and tricks

with_ghc wrapper (e.g. for HLS)

Due to some HLS bugs it's necessary that the ghc in PATH is the one defined in cabal.project. With some simple shell functions, we can start our editor with the appropriate path prepended.

For bash, in e.g. ~/.bashrc define:

with_ghc() {
  local np=$(ghcup --offline whereis -d ghc $1 || { ghcup --cache install ghc $1 && ghcup whereis -d ghc $1 ;})
  if [ -e "${np}" ] ; then
    shift
    PATH="$np:$PATH" "$@"
  else
    >&2 echo "Cannot find or install GHC version $1"
    return 1
  fi
}

For fish shell, in e.g. ~/.config/fish/config.fish define:

function with_ghc
  set --local np (ghcup --offline whereis -d ghc $argv[1] ; or begin ghcup --cache install ghc $argv[1] ; and ghcup whereis -d ghc $argv[1] ; end)
  if test -e "$np"
    PATH="$np:$PATH" $argv[2..-1]
  else
    echo "Cannot find or install GHC version $argv[1]" 1>&2
    return 1
  end
end

Then start a new shell and issue:

# replace 'code' with your editor
with_ghc 8.10.5 code path/to/haskell/source

Cabal and HLS will now see 8.10.5 as the primary GHC, without the need to run ghcup set all the time when switching between projects.

Design goals

1. simplicity
2. non-interactive
3. portable (eh)
4. do one thing and do it well (UNIX philosophy)

Non-goals

1. invoking sudo, apt-get or any package manager
2. handling system packages
3. handling cabal projects
4. being a stack alternative

How

Installs a specified GHC version into ~/.ghcup/ghc/<ver>, and places ghc-<ver> symlinks in ~/.ghcup/bin/.

Optionally, an unversioned ghc link can point to a default version of your choice.

This uses precompiled GHC binaries that have been compiled on fedora/debian by upstream GHC.

Alternatively, you can also tell it to compile from source (note that this might fail due to missing requirements).

In addition this script can also install cabal-install.

Known users

• Github actions:
  • actions/virtual-environments
  • haskell/actions/setup
• mirrors:
  • sjtug
• tools:
  • vabal

Known problems

Custom ghc version names

When installing ghc bindists with custom version names as outlined in installing custom bindists, then cabal might be unable to find the correct ghc-pkg (also see #73) if you use cabal build --with-compiler=ghc-foo. Instead, point it to the full path, such as: cabal build --with-compiler=$HOME/.ghcup/ghc/<version-name>/bin/ghc or set that GHC version as the current one via: ghcup set ghc <version-name>.

This problem doesn't exist for regularly installed GHC versions.

Limited distributions supported

Currently only GNU/Linux distributions compatible with the upstream GHC binaries are supported.

Precompiled binaries

Since this uses precompiled binaries you may run into several problems.

Missing libtinfo (ncurses)

You may run into problems with ncurses and missing libtinfo, in case your distribution doesn't use the legacy way of building ncurses and has no compatibility symlinks in place.

Ask your distributor on how to solve this or try to compile from source via ghcup compile <version>.

Libnuma required

This was a bug in the build system of some GHC versions that lead to unconditionally enabled libnuma support. To mitigate this you might have to install the libnuma package of your distribution. See here for a discussion.

Compilation

Although this script can compile GHC for you, it's just a very thin wrapper around the build system. It makes no effort in trying to figure out whether you have the correct toolchain and the correct dependencies. Refer to the official docs on how to prepare your environment for building GHC.

Stack support

There may be a number of bugs when trying to make ghcup installed GHC versions work with stack, such as:

• https://gitlab.haskell.org/haskell/ghcup-hs/-/issues/188

Further, stack's upgrade procedure may break/confuse ghcup. There are a number of integration issues discussed here:

• https://gitlab.haskell.org/haskell/ghcup-hs/-/issues/153

Windows support

Windows support is in early stages. Since windows doesn't support symbolic links properly, ghcup uses a shimgen wrapper. It seems to work well, but there may be unknown issues with that approach.

Windows 7 and Powershell 2.0 aren't well supported at the moment, also see:

• https://gitlab.haskell.org/haskell/ghcup-hs/-/issues/140
• https://gitlab.haskell.org/haskell/ghcup-hs/-/issues/197

FAQ

Why reimplement stack?

GHCup is not a reimplementation of stack. The only common part is automatic installation of GHC, but even that differs in scope and design.

Why should I use ghcup over stack?

GHCup is not a replacement for stack. Instead, it supports installing and managing stack versions. It does the same for cabal, GHC and HLS. As such, It doesn't make a workflow choice for you.

Why should I let ghcup manage stack?

You don't need to. However, some users seem to prefer to have a central tool that manages cabal and stack at the same time. Additionally, it can allow better sharing of GHC installation across these tools. Also see:

• https://docs.haskellstack.org/en/stable/yaml_configuration/#system-ghc
• https://github.com/commercialhaskell/stack/pull/5585

Why does ghcup not use stack code?

Oddly, this question has been asked a couple of times. For the curious, here are a few reasons:

1. GHCup started as a shell script. At the time of rewriting it in Haskell, the authors didn't even know that stack exposes some of its installation API
2. Even if they did, it doesn't seem it would have satisfied their needs
   • it didn't support cabal installation, which was the main motivation behind GHCup back then
   • depending on a codebase as big as stack for a central part of one's application without having a short contribution pipeline would likely have caused stagnation or resulted in simply copy-pasting the relevant code in order to adjust it
   • it's not clear how GHCup would have been implemented with the provided API. It seems the codebases are fairly different. GHCup does a lot of symlink handling to expose a central bin/ directory that users can easily put in PATH, without having to worry about anything more. It also provides explicit removal functionality, GHC cross-compilation, a TUI, etc etc.
3. GHCup is built around unix principles and supposed to be simple.

Why not unify...

...stack and Cabal and do away with standalone installers

GHCup is not involved in such decisions. cabal-install and stack might have a sufficiently different user experience to warrant having a choice.

...installer implementations and have a common library

This sounds like an interesting goal. However, GHC installation isn't a hard engineering problem and the shared code wouldn't be too exciting. For such an effort to make sense, all involved parties would need to collaborate and have a short pipeline to get patches in.

It's true this would solve the integration problem, but following unix principles, we can do similar via hooks. Both cabal and stack can support installation hooks. These hooks can then call into ghcup or anything else, also see:

• https://github.com/haskell/cabal/issues/7394
• https://github.com/commercialhaskell/stack/pull/5585

...installers (like, all of it)

So far, there hasn't been an open discussion about this. Is this even a good idea? Sometimes projects converge eventually if their overlap is big enough, sometimes they don't.

While unification sounds like a simplification of the ecosystem, it also takes away choice. Take curl and wget as an example.

How bad do we need this?

Why not support windows?

Windows is supported since GHCup version 0.1.15.1.

Why the haskell reimplementation?

GHCup started as a portable posix shell script of maybe 50 LOC. GHC installation itself can be carried out in about ~3 lines of shell code (download, unpack , configure+make install). However, much convenient functionality has been added since, as well as ensuring that all operations are safe and correct. The shell script ended up with over 2k LOC, which was very hard to maintain.

The main concern when switching from a portable shell script to haskell was platform/architecture support. However, ghcup now re-uses GHCs CI infrastructure and as such is perfectly in sync with all platforms that GHC supports.

Is GHCup affiliated with the Haskell Foundation?

There has been some collaboration: Windows and Stack support were mainly requested by the Haskell Foundation and those seemed interesting features to add.

Other than that, GHCup is dedicated only to its users and is supported by haskell.org through hosting and CI infrastructure.