Personal tools

GHC/As a library

From HaskellWiki

Jump to: navigation, search

For instructions on the GHC API with GHC 6.8 or older please refer to GHC/As a library (up to 6.8)


1 Introduction

GHC's functionality can be useful for more things than just compiling Haskell programs. Important use cases are programs that analyse (and perhaps transform) Haskell code. Others include loading Haskell code dynamically in a GHCi-like manner. For this reason, a lot of GHC's features can be accessed by programs which import the ghc package.

The instructions on this page concern the API of GHC 6.10.1 and above. Please note that the GHC API is still in flux and may change quite significantly between major releases while we (the GHC team) provide new features or simplify certain aspects.

2 Getting Started

To use the GHC API you need GHC 6.10.1 or above and import the ghc package.

ghc -package ghc my_program.hs

In most cases you probably also want to use the ghc-paths package.

Most of the common functionality is provided by the GHC module, but occasionally you may have to import other modules. See the GHC's haddock documentation for a list of these modules. One good entry point into the docs is GHC.

3 A Simple Example

The following little program essentially does what ghc --make does.

import GHC
import GHC.Paths ( libdir )
import DynFlags
main = 
#if __GLASGOW_HASKELL__ > 704
    defaultErrorHandler defaultFatalMessager defaultFlushOut $ do
    defaultErrorHandler defaultLogAction $ do
      runGhc (Just libdir) $ do
        dflags <- getSessionDynFlags
        setSessionDynFlags dflags
        target <- guessTarget "test_main.hs" Nothing
        setTargets [target]
        load LoadAllTargets

The outermost function, defaultErrorHandler, sets up proper exception handlers and prints an error message and exits with exit code 1 if it encounters one of these exceptions.

Most of GHC's high-level API requires access to a current session. Therefore, these functions require to be called inside a monad that is an instance of the GhcMonad typeclass. Two default implementations of this typeclass are Ghc and GhcT. In the above example we used the Ghc monad since we don't need to track any extra state.

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.

4 Another example

Here we demonstrate calling parseModule, typecheckModule, desugarModule, getNamesInScope, and getModuleGraph. This works for haskell-platform, ghc-7.0.3 to ghc-7.6.x. It also demonstrates how to enable some extensions.

--invoke: ghci -package ghc A.hs
import GHC
import Outputable
import GHC.Paths ( libdir )
--GHC.Paths is available via cabal install ghc-paths
import DynFlags
targetFile = "B.hs"
main :: IO ()
main = do
   res <- example
#if __GLASGOW_HASKELL__ > 704
   str <- runGhc (Just libdir) $ do
      dflags <- getSessionDynFlags
      return $ showSDoc dflags $ ppr res
   putStrLn str
   putStrLn $ showSDoc ( ppr res )
example = 
#if __GLASGOW_HASKELL__ > 704
    defaultErrorHandler defaultFatalMessager defaultFlushOut $ do
    defaultErrorHandler defaultLogAction $ do
      runGhc (Just libdir) $ do
        dflags <- getSessionDynFlags
        let dflags' = foldl xopt_set dflags
                            [Opt_Cpp, Opt_ImplicitPrelude, Opt_MagicHash]
        setSessionDynFlags dflags'
        target <- guessTarget targetFile Nothing
        setTargets [target]
        load LoadAllTargets
        modSum <- getModSummary $ mkModuleName "B"
        p <- parseModule modSum
        t <- typecheckModule p
        d <- desugarModule t
        l <- loadModule d
        n <- getNamesInScope
        c <- return $ coreModule d
        g <- getModuleGraph
        mapM showModule g     
        return $ (parsedSource d,"/n-----/n",  typecheckedSource d)
module B where
main = print "Hello, World!"

5 Running interactive statements

Once you've loaded your module, you can run statements in that context, much as GHCi does, using the runStmt function.

Given the modSum ModSummary obtained earlier, you can dump the results of your statement, if the result type is a Show instance:

run :: GhcMonad m => ModSummary -> String -> m ()
run modSum expr = do
#if __GLASGOW_HASKELL__ < 704
  setContext [ms_mod modSum] []
#if __GLASGOW_HASKELL__ < 706
  setContext [IIModule $ ms_mod modSum]
  setContext [IIModule $ moduleName  $ ms_mod modSum]       
  rr<- runStmt expr RunToCompletion
  case rr of
    RunOk ns->
        let q=(qualName &&& qualModule) defaultUserStyle
        mapM_ (\n -> do
                mty <- lookupName n
                case mty of
                  Just (AnId aid) -> do
                      df <- getSessionDynFlags
                      t <- gtry $ obtainTermFromId maxBound True aid
                      evalDoc <- case t of
                          Right term -> showTerm term
                          Left  exn  -> return (text "*** Exception:" <+>
                                                  text (show (exn :: SomeException)))
                      liftIO $ putStrLn $ showSDocForUser df q evalDoc
                      return ()
                  _ -> return ()
                ) ns
    RunException e -> liftIO $ print e
    _ -> return ()

Note the call to setContext. Basically then runStmt return the names bound by the statement, which you then lookup via lookupName and obtainTermFromId and print via showTerm.

5.1 Session parameters for interactive evaluation

The DynFlags contain different flags that may be useful when working with the API, mainly to avoid generating the linker output. You need to be careful when you set them, though, since they can have long ranging effects. A common setting would be:

setSessionDynFlags oldFlgs  {hscTarget = HscInterpreted, ghcLink = LinkInMemory , ghcMode = CompManager }
  • HscInterpreted means you can still get TemplateHaskell support but you're not generating output code
  • LinkInMemory is required to be able to reload modules properly. Using NoLink will break that: reloading a module won't allow allow to run interactive statements on the new version of the module
  • CompManager ensures dependent modules are built and loaded if needed.

6 Further reading