GHC exposes its internal APIs to users through the built-in ghc package. It allows you to write programs that leverage GHC’s entire compilation driver, in order to analyze or compile Haskell code programmatically. Furthermore, GHC gives users the ability to load compiler plugins during compilation - modules which are allowed to view and change GHC’s internal intermediate representation, Core. Plugins are suitable for things like experimental optimizations or analysis, and offer a lower barrier of entry to compiler development for many common cases.
Furthermore, GHC offers a lightweight annotation mechanism that you can use to annotate your source code with metadata, which you can later inspect with either the compiler API or a compiler plugin.
Annotations are small pragmas that allow you to attach data to identifiers in source code, which are persisted when compiled. These pieces of data can then inspected and utilized when using GHC as a library or writing a compiler plugin.
Any expression that has both Typeable and Data instances may be attached to a top-level value binding using an ANN pragma. In particular, this means you can use ANN to annotate data constructors (e.g. Just) as well as normal values (e.g. take). By way of example, to annotate the function foo with the annotation Just "Hello" you would do this:
{-# ANN foo (Just "Hello") #-}
foo = ...
A number of restrictions apply to use of annotations:
The binder being annotated must be at the top level (i.e. no nested binders)
The binder being annotated must be declared in the current module
The expression you are annotating with must have a type with Typeable and Data instances
The Template Haskell staging restrictions apply to the expression being annotated with, so for example you cannot run a function from the module being compiled.
To be precise, the annotation {-# ANN x e #-} is well staged if and only if $(e) would be (disregarding the usual type restrictions of the splice syntax, and the usual restriction on splicing inside a splice - $([|1|]) is fine as an annotation, albeit redundant).
If you feel strongly that any of these restrictions are too onerous, please give the GHC team a shout.
However, apart from these restrictions, many things are allowed, including expressions which are not fully evaluated! Annotation expressions will be evaluated by the compiler just like Template Haskell splices are. So, this annotation is fine:
{-# ANN f SillyAnnotation { foo = (id 10) + $([| 20 |]), bar = 'f } #-}
f = ...
You can annotate types with the ANN pragma by using the type keyword. For example:
{-# ANN type Foo (Just "A `Maybe String' annotation") #-}
data Foo = ...
You can annotate modules with the ANN pragma by using the module keyword. For example:
{-# ANN module (Just "A `Maybe String' annotation") #-}
The ghc package exposes most of GHC’s frontend to users, and thus allows you to write programs that leverage it. This library is actually the same library used by GHC’s internal, frontend compilation driver, and thus allows you to write tools that programmatically compile source code and inspect it. Such functionality is useful in order to write things like IDE or refactoring tools. As a simple example, here’s a program which compiles a module, much like ghc itself does by default when invoked:
import GHC
import GHC.Paths ( libdir )
import DynFlags ( defaultLogAction )
main =
defaultErrorHandler defaultLogAction $ do
runGhc (Just libdir) $ do
dflags <- getSessionDynFlags
setSessionDynFlags dflags
target <- guessTarget "test_main.hs" Nothing
setTargets [target]
load LoadAllTargets
The argument to runGhc is a bit tricky. GHC needs this to find its libraries, so the argument must refer to the directory that is printed by ghc --print-libdir for the same version of GHC that the program is being compiled with. Above we therefore use the ghc-paths package which provides this for us.
Compiling it results in:
$ cat test_main.hs
main = putStrLn "hi"
$ ghc -package ghc simple_ghc_api.hs
[1 of 1] Compiling Main ( simple_ghc_api.hs, simple_ghc_api.o )
Linking simple_ghc_api ...
$ ./simple_ghc_api
$ ./test_main
hi
$
For more information on using the API, as well as more samples and references, please see this Haskell.org wiki page.
GHC has the ability to load compiler plugins at compile time. The feature is similar to the one provided by GCC, and allows users to write plugins that can adjust the behaviour of the constraint solver, inspect and modify the compilation pipeline, as well as transform and inspect GHC’s intermediate language, Core. Plugins are suitable for experimental analysis or optimization, and require no changes to GHC’s source code to use.
Plugins cannot optimize/inspect C–, nor can they implement things like parser/front-end modifications like GCC, apart from limited changes to the constraint solver. If you feel strongly that any of these restrictions are too onerous, please give the GHC team a shout.
Plugins can be specified on the command line with the -fplugin=⟨module⟩ option where ⟨module⟩ is a module in a registered package that exports a plugin. Arguments can be given to plugins with the -fplugin-opt=⟨module⟩:⟨args⟩ option.
Load the plugin in the given module. The module must be a member of a package registered in GHC’s package database.
Give arguments to a plugin module; module must be specified with -fplugin=⟨module⟩.
As an example, in order to load the plugin exported by Foo.Plugin in the package foo-ghc-plugin, and give it the parameter “baz”, we would invoke GHC like this:
$ ghc -fplugin Foo.Plugin -fplugin-opt Foo.Plugin:baz Test.hs
[1 of 1] Compiling Main ( Test.hs, Test.o )
Loading package ghc-prim ... linking ... done.
Loading package integer-gmp ... linking ... done.
Loading package base ... linking ... done.
Loading package ffi-1.0 ... linking ... done.
Loading package foo-ghc-plugin-0.1 ... linking ... done.
...
Linking Test ...
$
Alternatively, core plugins can be specified with Template Haskell.
addCorePlugin "Foo.Plugin"
This inserts the plugin as a core-to-core pass. Unlike -fplugin=(module), the plugin module can’t reside in the same package as the module calling Language.Haskell.TH.Syntax.addCorePlugin. This way, the implementation can expect the plugin to be built by the time it is needed.
Plugin modules live in a separate namespace from the user import namespace. By default, these two namespaces are the same; however, there are a few command line options which control specifically plugin packages:
This option causes the installed package ⟨pkg⟩ to be exposed for plugins, such as -fplugin=⟨module⟩. The package ⟨pkg⟩ can be specified in full with its version number (e.g. network-1.0) or the version number can be omitted if there is only one version of the package installed. If there are multiple versions of ⟨pkg⟩ installed and -hide-all-plugin-packages was not specified, then all other versions will become hidden. -plugin-package ⟨pkg⟩ supports thinning and renaming described in Thinning and renaming modules.
Unlike -package ⟨pkg⟩, this option does NOT cause package ⟨pkg⟩ to be linked into the resulting executable or shared object.
Exposes a package in the plugin namespace like -plugin-package ⟨pkg⟩, but the package is named by its installed package ID rather than by name. This is a more robust way to name packages, and can be used to select packages that would otherwise be shadowed. Cabal passes -plugin-package-id ⟨pkg-id⟩ flags to GHC. -plugin-package-id ⟨pkg-id⟩ supports thinning and renaming described in Thinning and renaming modules.
By default, all exposed packages in the normal, source import namespace are also available for plugins. This causes those packages to be hidden by default. If you use this flag, then any packages with plugins you require need to be explicitly exposed using -plugin-package ⟨pkg⟩ options.
At the moment, the only way to specify a dependency on a plugin in Cabal is to put it in build-depends (which uses the conventional -package-id ⟨unit-id⟩ flag); however, in the future there will be a separate field for specifying plugin dependencies specifically.
Plugins are modules that export at least a single identifier, plugin, of type GhcPlugins.Plugin. All plugins should import GhcPlugins as it defines the interface to the compilation pipeline.
A Plugin effectively holds a function which installs a compilation pass into the compiler pipeline. By default there is the empty plugin which does nothing, GhcPlugins.defaultPlugin, which you should override with record syntax to specify your installation function. Since the exact fields of the Plugin type are open to change, this is the best way to ensure your plugins will continue to work in the future with minimal interface impact.
Plugin exports a field, installCoreToDos which is a function of type [CommandLineOption] -> [CoreToDo] -> CoreM [CoreToDo]. A CommandLineOption is effectively just String, and a CoreToDo is basically a function of type Core -> Core. A CoreToDo gives your pass a name and runs it over every compiled module when you invoke GHC.
As a quick example, here is a simple plugin that just does nothing and just returns the original compilation pipeline, unmodified, and says ‘Hello’:
module DoNothing.Plugin (plugin) where
import GhcPlugins
plugin :: Plugin
plugin = defaultPlugin {
installCoreToDos = install
}
install :: [CommandLineOption] -> [CoreToDo] -> CoreM [CoreToDo]
install _ todo = do
putMsgS "Hello!"
return todo
Provided you compiled this plugin and registered it in a package (with cabal for instance,) you can then use it by just specifying -fplugin=DoNothing.Plugin on the command line, and during the compilation you should see GHC say ‘Hello’.
CoreToDo is effectively a data type that describes all the kinds of optimization passes GHC does on Core. There are passes for simplification, CSE, vectorisation, etc. There is a specific case for plugins, CoreDoPluginPass :: String -> PluginPass -> CoreToDo which should be what you always use when inserting your own pass into the pipeline. The first parameter is the name of the plugin, and the second is the pass you wish to insert.
CoreM is a monad that all of the Core optimizations live and operate inside of.
A plugin’s installation function (install in the above example) takes a list of CoreToDos and returns a list of CoreToDo. Before GHC begins compiling modules, it enumerates all the needed plugins you tell it to load, and runs all of their installation functions, initially on a list of passes that GHC specifies itself. After doing this for every plugin, the final list of passes is given to the optimizer, and are run by simply going over the list in order.
You should be careful with your installation function, because the list of passes you give back isn’t questioned or double checked by GHC at the time of this writing. An installation function like the following:
install :: [CommandLineOption] -> [CoreToDo] -> CoreM [CoreToDo]
install _ _ = return []
is certainly valid, but also certainly not what anyone really wants.
In the last section we saw that besides a name, a CoreDoPluginPass takes a pass of type PluginPass. A PluginPass is a synonym for (ModGuts -> CoreM ModGuts). ModGuts is a type that represents the one module being compiled by GHC at any given time.
A ModGuts holds all of the module’s top level bindings which we can examine. These bindings are of type CoreBind and effectively represent the binding of a name to body of code. Top-level module bindings are part of a ModGuts in the field mg_binds. Implementing a pass that manipulates the top level bindings merely needs to iterate over this field, and return a new ModGuts with an updated mg_binds field. Because this is such a common case, there is a function provided named bindsOnlyPass which lifts a function of type ([CoreBind] -> CoreM [CoreBind]) to type (ModGuts -> CoreM ModGuts).
Continuing with our example from the last section, we can write a simple plugin that just prints out the name of all the non-recursive bindings in a module it compiles:
module SayNames.Plugin (plugin) where
import GhcPlugins
plugin :: Plugin
plugin = defaultPlugin {
installCoreToDos = install
}
install :: [CommandLineOption] -> [CoreToDo] -> CoreM [CoreToDo]
install _ todo = do
return (CoreDoPluginPass "Say name" pass : todo)
pass :: ModGuts -> CoreM ModGuts
pass guts = do dflags <- getDynFlags
bindsOnlyPass (mapM (printBind dflags)) guts
where printBind :: DynFlags -> CoreBind -> CoreM CoreBind
printBind dflags bndr@(NonRec b _) = do
putMsgS $ "Non-recursive binding named " ++ showSDoc dflags (ppr b)
return bndr
printBind _ bndr = return bndr
Previously we discussed annotation pragmas (Source annotations), which we mentioned could be used to give compiler plugins extra guidance or information. Annotations for a module can be retrieved by a plugin, but you must go through the modules ModGuts in order to get it. Because annotations can be arbitrary instances of Data and Typeable, you need to give a type annotation specifying the proper type of data to retrieve from the interface file, and you need to make sure the annotation type used by your users is the same one your plugin uses. For this reason, we advise distributing annotations as part of the package which also provides compiler plugins if possible.
To get the annotations of a single binder, you can use getAnnotations and specify the proper type. Here’s an example that will print out the name of any top-level non-recursive binding with the SomeAnn annotation:
{-# LANGUAGE DeriveDataTypeable #-}
module SayAnnNames.Plugin (plugin, SomeAnn(..)) where
import GhcPlugins
import Control.Monad (unless)
import Data.Data
data SomeAnn = SomeAnn deriving Data
plugin :: Plugin
plugin = defaultPlugin {
installCoreToDos = install
}
install :: [CommandLineOption] -> [CoreToDo] -> CoreM [CoreToDo]
install _ todo = do
return (CoreDoPluginPass "Say name" pass : todo)
pass :: ModGuts -> CoreM ModGuts
pass g = do
dflags <- getDynFlags
mapM_ (printAnn dflags g) (mg_binds g) >> return g
where printAnn :: DynFlags -> ModGuts -> CoreBind -> CoreM CoreBind
printAnn dflags guts bndr@(NonRec b _) = do
anns <- annotationsOn guts b :: CoreM [SomeAnn]
unless (null anns) $ putMsgS $ "Annotated binding found: " ++ showSDoc dflags (ppr b)
return bndr
printAnn _ _ bndr = return bndr
annotationsOn :: Data a => ModGuts -> CoreBndr -> CoreM [a]
annotationsOn guts bndr = do
anns <- getAnnotations deserializeWithData guts
return $ lookupWithDefaultUFM anns [] (varUnique bndr)
Please see the GHC API documentation for more about how to use internal APIs, etc.
In addition to Core plugins, GHC has experimental support for typechecker plugins, which allow the behaviour of the constraint solver to be modified. For example, they make it possible to interface the compiler to an SMT solver, in order to support a richer theory of type-level arithmetic expressions than the theory built into GHC (see Computing With Type-Level Naturals).
The Plugin type has a field tcPlugin of type [CommandLineOption] -> Maybe TcPlugin, where the TcPlugin type is defined thus:
data TcPlugin = forall s . TcPlugin
{ tcPluginInit :: TcPluginM s
, tcPluginSolve :: s -> TcPluginSolver
, tcPluginStop :: s -> TcPluginM ()
}
type TcPluginSolver = [Ct] -> [Ct] -> [Ct] -> TcPluginM TcPluginResult
data TcPluginResult = TcPluginContradiction [Ct] | TcPluginOk [(EvTerm,Ct)] [Ct]
(The details of this representation are subject to change as we gain more experience writing typechecker plugins. It should not be assumed to be stable between GHC releases.)
The basic idea is as follows:
Plugin code runs in the TcPluginM monad, which provides a restricted interface to GHC API functionality that is relevant for typechecker plugins, including IO and reading the environment. If you need functionality that is not exposed in the TcPluginM module, you can use unsafeTcPluginTcM :: TcM a -> TcPluginM a, but are encouraged to contact the GHC team to suggest additions to the interface. Note that TcPluginM can perform arbitrary IO via tcPluginIO :: IO a -> TcPluginM a, although some care must be taken with side effects (particularly in tcPluginSolve). In general, it is up to the plugin author to make sure that any IO they do is safe.
The key component of a typechecker plugin is a function of type TcPluginSolver, like this:
solve :: [Ct] -> [Ct] -> [Ct] -> TcPluginM TcPluginResult
solve givens deriveds wanteds = ...
This function will be invoked at two points in the constraint solving process: after simplification of given constraints, and after unflattening of wanted constraints. The two phases can be distinguished because the deriveds and wanteds will be empty in the first case. In each case, the plugin should either
If the plugin cannot make any progress, it should return TcPluginOk [] []. Otherwise, if there were any new constraints, the main constraint solver will be re-invoked to simplify them, then the plugin will be invoked again. The plugin is responsible for making sure that this process eventually terminates.
Plugins are provided with all available constraints (including equalities and typeclass constraints), but it is easy for them to discard those that are not relevant to their domain, because they need return only those constraints for which they have made progress (either by solving or contradicting them).
Constraints that have been solved by the plugin must be provided with evidence in the form of an EvTerm of the type of the constraint. This evidence is ignored for given and derived constraints, which GHC “solves” simply by discarding them; typically this is used when they are uninformative (e.g. reflexive equations). For wanted constraints, the evidence will form part of the Core term that is generated after typechecking, and can be checked by -dcore-lint. It is possible for the plugin to create equality axioms for use in evidence terms, but GHC does not check their consistency, and inconsistent axiom sets may lead to segfaults or other runtime misbehaviour.
A frontend plugin allows you to add new major modes to GHC. You may prefer this over a traditional program which calls the GHC API, as GHC manages a lot of parsing flags and administrative nonsense which can be difficult to manage manually. To load a frontend plugin exported by Foo.FrontendPlugin, we just invoke GHC with the --frontend ⟨module⟩ flag as follows:
$ ghc --frontend Foo.FrontendPlugin ...other options...
Frontend plugins, like compiler plugins, are exported by registered plugins. However, unlike compiler modules, frontend plugins are modules that export at least a single identifier frontendPlugin of type GhcPlugins.FrontendPlugin.
FrontendPlugin exports a field frontend, which is a function [String] -> [(String, Maybe Phase)] -> Ghc (). The first argument is a list of extra flags passed to the frontend with -ffrontend-opt; the second argument is the list of arguments, usually source files and module names to be compiled (the Phase indicates if an -x flag was set), and a frontend simply executes some operation in the Ghc monad (which, among other things, has a Session).
As a quick example, here is a frontend plugin that prints the arguments that were passed to it, and then exits.
module DoNothing.FrontendPlugin (frontendPlugin) where
import GhcPlugins
frontendPlugin :: FrontendPlugin
frontendPlugin = defaultFrontendPlugin {
frontend = doNothing
}
doNothing :: [String] -> [(String, Maybe Phase)] -> Ghc ()
doNothing flags args = do
liftIO $ print flags
liftIO $ print args
Provided you have compiled this plugin and registered it in a package, you can just use it by specifying --frontend DoNothing.FrontendPlugin on the command line to GHC.