Hadrian is the build system for the Glasgow Haskell Compiler. It is based on the Shake library. If you are curious about the rationale behind the project and the architecture of the build system you can find more details in this Haskell Symposium 2016 paper and this Haskell eXchange 2016 talk.
Put on the helmet and run the following commands from the root of the GHC tree:
./boot && ./configure
hadrian/build -j
or on Windows:
./boot && ./configure --enable-tarballs-autodownload
hadrian/build.bat -j
Here flag -j
enables parallelism and is optional. We will further refer to the
build script simply as build
.
-
On Windows, if you do not want to install MSYS, you can use the Stack-based build script (Stack provides a managed MSYS environment), as described in these instructions. If you don't mind installing MSYS yourself or already have it, you can use the Cabal-based build script.
-
Hadrian is written in Haskell and depends on
shake
(plus a few packages thatshake
depends on),mtl
,quickcheck
, and GHC core libraries. -
If you have never built GHC before, start with the preparation guide.
-
By default Hadrian will fetch dependencies from the internet. If this is a problem for your build system then see the bootstrap directory for instructions about how to build hadrian without an internet connection.
Once your first build is successful, simply run build
to rebuild after some
changes. Build results are placed into _build
by default.
There are many different ways to build a compiler, each way is called a flavour.
--flavour=FLAVOUR
: choose a build flavour. The following settings are currently supported:default
,quick
,quickest
,perf
,prof
,devel1
anddevel2
. As an example, thequickest
flavour adds-O0
flag to all GHC invocations and builds libraries only in thevanilla
way, which speeds up builds by 3-4x.
In addition to the overall build flavour there are also "flavour transformers" which can slightly modify the build settings for a flavour. Some common flavour transformers are:
no_profiled_libs
: Don't build profiled librarieslint
: Build with core lint enabled.- etc
A flavour transformer is appended to a normal flavour name using +
--flavour=default+lint
Build flavours and flavour transformers are documented here.
In addition to standard Shake flags (try --help
), the build system
currently supports several others:
-
--build-root=PATH
or-oPATH
: specify the directory in which you want to store all build products. By default Hadrian builds everything in the_build/
subdirectory of the GHC source tree. This option is useful for GHC developers who want to build GHC in different ways or at different commits, from the same source directory, and have the build products sit in different, isolated folders. -
--freeze1
: freeze Stage1 GHC, i.e. do not rebuild it even if some of its source files are out-of-date. This allows to significantly reduce the rebuild time when you are working on a feature that affects both Stage1 and Stage2 compilers, but may lead to incorrect build results. To unfreeze Stage1 GHC simply drop the--freeze1
flag and Hadrian will rebuild all out-of-date files. -
--freeze2
: just like--freeze1
but tell Hadrian to additionally freeze Stage2 GHC. -
--skip-depends
: skips rebuilding Haskell module dependency files. -
--bignum={native,gmp,check-gmp,ffi}
: choose which bignum implementation to use. The default isgmp
. -
--color
and--no-color
: choose whether to use colors when printing build progress info. By default, Hadrian tries to determine if the terminal supports colored output, and proceeds accordingly. -
--progress-info=STYLE
: choose how build progress info is printed. There are four settings:none
,brief
(one line per build command; this is the default setting),normal
(typically a box per build command), andunicorn
(whennormal
just won't do). -
-V
/--verbose
: run Hadrian in verbose mode. This makes commands print their stdout and produces slightly more output on a failure (including hadrian call stacks). -
-VV
: run hadrian in diagnostics mode: In particular this prints diagnostic messages by Shake oracles and full command lines for all commands. -
--lint
: run Shake Lint during the build to check that the build system is well formed. Note that the Lint check currently fails under certain circumstances, as discussed in this ticket.
Hadrian expressions are used extensively for specifying build settings. For an explanation of how they work, see the documentation.
The Make-based build system uses mk/build.mk
to specify user build settings.
Hadrian uses hadrian/UserSettings.hs
for the same purpose, see
documentation.
You can build a specific library or executable for a given stage by running
build stage<N>:<lib|exe>:<package name>
. Examples:
# Build the Stage1 GHC compiler, and place the binary to the directory
# _build/stage0/bin/ghc (because it is built by the Stage0 compiler).
build stage1:exe:ghc-bin
# Build the Stage2 GHC compiler, and place the binary to the directory
# _build/stage1/bin/ghc (because it is built by the Stage1 compiler).
build stage2:exe:ghc-bin
# Build the `ghc` library with the bootstrapping (Stage0) compiler, and register
# it in the package database stored in the directory _build/stage0/lib.
build stage0:lib:ghc
# Build the Cabal library with the Stage1 compiler and register it
# in the package database stored in the directory _build/stage1/lib.
# Build the `text` library with the Stage2 compiler and register it
# in the package database stored in the directory _build/stage2/lib.
build stage2:lib:text
# Build Haddock using the Stage1 compiler and place the binary into the
# directory _build/stage1/haddock.
build stage1:exe:haddock
The GHC
environment variable can be used to control which GHC is used to build hadrian. By
default the version of GHC on your path is used to build hadrian.
This can be a different version of GHC to the one you want to use as the boot compiler (which
is selected during ./configure).
GHC=$(which ghc-9.4.2) ./hadrian/build
-- hadrian is built using ghc-9.4.2
Running the ./hadrian/ghci
script will load the main compiler into
a ghci session. This is intended for fast development feedback, modules are only
typechecked so it isn't possible to run any functions in the repl.
./hadrian/ghci
You can also use this target with ghcid
.
ghcid --command="./hadrian/ghci"
The first time this command is run hadrian will need to compile a few dependencies which will take 1-2 minutes. Loading GHC into GHCi itself takes about 30 seconds and reloads after that take in the region of 1-5 seconds depending on which modules need to be recompiled.
You can further speed up the script by passing -j
as an argument. This will
have the effect of passing -j
to both hadrian and ghci so they will both
build in parallel.
./hadrian/ghci -j8
You can also use the experimental multi-repl if you are booting with GHC-9.4 or later.
./hadrian/ghci-multi
To run GHC testsuite, use build test
. See
doc/testsuite.md to learn about all associated command line
flags, as well as about the equivalents of the features that the Make build
system offers.
NOTE: The only build flavours which are expected to pass the testsuite are those
tested in CI. If you use an untested flavour such as "Quick" then you run the
risk that not all tests will pass. In particular you can rely on the validate
and perf
flavours being tested but no others.
build selftest
runs tests of the build system. The current test coverage
is close to zero (see #197).
There are two targets which runs the lint commands used by CI:
lint:base
, runs hlint on thebase
library.lint:compiler
, runs hlint on theghc
library.
It's useful to know that you can combine multiple targets in build command. For example, you can tell hadrian to build the compiler and also run the linters:
./hadrian/build stage2:exe:ghc-bin lint:compiler
-
build clean
removes all build artefacts. -
build distclean
additionally remove the mingw tarballs and fs* files created byconfigure
. -
build -B
forces Shake to rerun all rules, even if the previous build results are still up-to-date.
GHC is a self-hosted compiler and consequently the build proceeds in several stages:
- The build begins with a user-provided installation of GHC called the
stage0 (or bootstrap) compiler which is used (via the
build.*.sh
scripts) to build Hadrian. - Hadrian uses the stage0 compiler to build a stage1 compiler (somewhat
confusingly found in
_build/stage0/bin/ghc
), linking against the stage0 compiler's core libraries (e.g.base
). - The stage1 compiler is used to build new core libraries (found in
_build/stage1/lib
). - The stage1 compiler is used to build a stage2 compiler (found in
_build/stage1/bin/ghc
), linking against these new core libraries. - Optionally (see the Building Stage3 section below) the
stage2 compiler can be used to build a stage3 compiler (found in
build/stage2/bin/ghc
) as a further smoke-test.
Note that the stage directories in the _build
directory can be thought of as
named after the stage that was used to build the artifacts in each directory.
These stages can be summarized graphically:
To build all GHC documentation, run build docs
. This includes
- Haddock documentation for all libraries
- The user guide (PDF and HTML)
- The man page
In order to only build haddock document there is the build docs-haddock
target.
In order to build the haddock documentation for just one package use the docs:<pkg>
command,
for example docs:base
will just build the documentation for base
.
Alternatively,
you can use the --docs
CLI flag to selectively disable some or
all of the documentation targets:
--docs=none
: don't build any docs--docs=no-haddocks
: don't build haddocks--docs=no-sphinx
: don't build any user manual or manpage--docs=no-sphinx-html
: don't build HTML versions of manuals--docs=no-sphinx-pdfs
: don't build PDF versions of manuals--docs=no-sphinx-man
: don't build the manpage
You can pass several --docs=...
flags, Hadrian will combine
their effects.
To build haddock documentation for upload to hackage you need to pass the --haddock-base-url
flag,
by default this will choose a url suitable for uploading to hackage but you might also want to pass something like
http://127.0.0.1:8080/package/%pkg%/docs
for testing upload locally on a local hackage server.
To build a GHC source distribution tarball, run build source-dist
.
To build a GHC binary distribution, run build binary-dist
. The resulting
tarball contains just enough to support the
$ ./configure [--prefix=PATH] && make install
workflow, for now.
Note: On windows you need to use the reloc-binary-dist
target.
If you require a relocatable binary distribution (for example on Windows), then you
can build the reloc-binary-dist
target.
You can get Hadrian to build and install a binary distribution in one go with the following command:
$ build install --prefix=/some/absolute/path
This builds everything that would be shipped in a bindist, without creating
the archive, and just runs ./configure --prefix=PATH
and make install
to get GHC installed installed at /some/absolute/path
.
It is possible to define a build flavour that builds a Stage3 compiler, which is a compiler built using Stage2. This is useful for cross-compilation. Detailed instructions can be found in the corresponding part of the user settings manual.
The tool-args
target is designed to allow hadrian to be integrated into other
tooling which uses the GHC API.
tool-args
prints out a list of flags which hadrian will use to compile
a module in the compiler
directory. Using these flags you can then set up
a GHC API session with the correct environment to load a module into your own
GHC session. This is how haskell-ide-engine
is able to support hadrian.
> ./hadrian/build tool-args
-hide-all-packages -no-user-package-db -package-db _build/stage0/lib/packag...
The ./hadrian/ghci
script is implemented using this target.
If you need help in debugging Hadrian, read the wiki and Shake's debugging tutorial. If nothing helps, don't hesitate to create a GHC issue.
The new build system still lacks many important features:
- Dynamic linking on Windows is not supported #343.
You can find a cheatsheet-style document that shows the Hadrian equivalents of commands that GHC users are used to run with the Make build system here.
The list of people who helped create Hadrian is long, and we hope that it will soon become even longer! The project was first developed in a separate GitHub repository, where you can find the list of original contributors. They had to stare at Makefiles for months, so give them all a round of applause. Simon Peyton Jones and Simon Marlow helped with deciphering these Makefiles, and Neil Mitchell patiently explained how to translate Makefiles to (much nicer) Shakefiles.
The initial development of Hadrian was funded by Newcastle University, EPSRC and Microsoft Research. Other organisations that contributed at various stages of the project are Haskell.Org and Google (both through supporting summer student projects), as well as Well-Typed.