|Version 22 (modified by simonmar, 6 years ago) (diff)|
GHC Commentary: Libraries
All GHC build trees contain a set of libraries, called the Boot Packages. These are the libraries that GHC's source code imports. Obviously you need the boot packages to build GHC at all.
The Boot Packages, along with the other subcomponents of the GHC build system, are in the file packages in a GHC tree. To get a list of them, you can run make show VALUE=PACKAGES in a configured GHC build tree. (This variable is set in $(TOP)/ghc.mk.)
You can see exactly which versions of what packages GHC depends on by looking in $(TOP)/compiler/ghc.cabal.in.
Boot packages can be classified in three different ways:
- Zero-boot/not zero-boot
- Installed/not installed
These distinctions are described in the following sub-sections.
Coupling to GHC
An important classification of the boot packages is as follows:
- INDEPENDENT: Independently maintained (e.g. time, haskeline)
- COUPLED: Tightly coupled to GHC, but used by others (base)
- SPECIFIC: Totally specific to GHC (e.g. template-haskell, DPH)
Most boot libraries are INDEPENDENT. INDEPENDENT libraries have a master repository somewhere separate from the GHC repositories. Whenever we release GHC, we ensure that the INDEPENDENT boot libraries that come with GHC are precisely sync'd with a particular released version of that library.
The current classification of packages is:
- SPECIFIC: ghc-prim, template-haskell, DPH
- COUPLED: base
- INDEPENDENT: all other packages
Since GHC's source code imports the boot packages, even the bootstrap compiler must have the boot packages available. (Or, more precisely, all the types and values that are imported must be available from some package in the bootstrap compiler; the exact set of packages does not need to be identical.)
For the most part we simply assume that the bootstrap compiler already has the boot packages installed. The Zero-boot Packages are a set of packages for which this assumption does not hold. For example, for certain fast-moving boot packages (eg Cabal), we don't want to rely on the user having installed a bang-up-to-date version of the package.
So we begin the entire build process by installing the zero-boot packages in the bootstrap compiler. (This installation is purely local to the build tree.)
As time goes on, a Zero-boot package may become an ordinary boot package, because the bootstrap compiler is expected to have (a sufficiently up to date) version of the package already.
The current Zero-boot packages are:
- Cabal: we frequently update Cabal and GHC in sync
- extensible-exceptions: this is a shim that provides an API to older versions of GHC that is compatible with what the current base package now exports. So, unusually, extensible-exceptions is a zero-boot package, but not a boot package.
- bin-package-db: a GHC-specific package that provides binary serialisation of the package database, use by ghc-pkg and GHC itself.
- binary (renamed to ghc-binary in the 6.12 branch): required by bin-package-db.
When we build a distribution of GHC, it includes at least some libraries, otherwise it would be utterly useless. Since GHC is part of the Haskell Platform, any library that is installed with GHC is necessarily part of the Haskell Platform, so we have to be a bit careful what we include.
Alas, since the ghc package (implementing the GHC API) is certainly an installed package, all the packages on which it depends must also be installed, and hence willy-nilly become part of the Haskell Platform. In practice that means that almost all the Boot Packages are installed. In some cases that is unfortunate. For example, we currently have a special version of the binary library, which we don't really expect Haskell users to use; in this case, we call it ghc-binary, and informally discourage its use.
Currently the Boot Packages that are not installed are haskelline, mtl, and terminfo; these are needed to build the GHC front-end, but not to build the ghc package.
Boot packages dependencies
- At the root of the hierarchy we have ghc-prim. As the name implies, this package contains the most primitive types and functions. It only contains a handful of modules, including GHC.Prim (which contains Int#, +#, etc) and GHC.Bool, containing the Bool datatype.
- Above ghc-prim is the integer-impl package, where impl is one of gmp and simple, which provides a definition of the Integer type (on top of the C gmp library, or in plain Haskell, respectively). Which functionality is provided in ghc-prim is mostly driven by what functionality the integer-impl packages need. By default integer-gmp is used; to use integer-simple define INTEGER_LIBRARY=integer-simple in mk/build.mk.
- Next is the base package. This contains a large number of modules, many of which are in one big cyclic import knot, mostly due to the Exception type.
- On top of base are a number of other, more specialised packages, whose purpose is generally clear from their name. If not, you can get more detail from the descriptions in their Cabal files. Currently these packages are are:
The haskell98, old-time and random packages are mostly only needed for Haskell 98 support, although dph currently uses random too.