|Version 10 (modified by guest, 9 years ago) (diff)|
GHC Commentary: The GHC API
This section of the commentary describes everything between HscMain and the front-end; that is, the parts of GHC that coordinate the compilation of multiple modules.
The GHC API is rather stateful; the state of an interaction with GHC is stored in an abstract value of type GHC.Session. The only fundamental reason for this choice is that the Session models the state of the RTS's linker, which must be single-threaded.
Although the GHC API apparently supports multiple clients, because each can be interacting with a different Session, in fact it only supports one client that is actually executing code, because the RTS linker has a single global symbol table.
This part of the commentary is not a tutorial on using the GHC API: for that, see Using GHC as a Library. Here we are going to talk about the implementation.
A typical interaction with the GHC API goes something like the following:
- You probably want to wrap the whole program in defaultErrorHandler defaultDynFlags to get error messages
- Initialize the GHC top dir: init
- Create a new session: newSession
- Add some targets: setTargets, addTarget, guessTarget
- Perform Dependency Analysis: depanal
- Load (compile) the source files: load
The targets specify the source files or modules at the top of the dependency tree. For a Haskell program there is often just a single target Main.hs, but for a library the targets would consist of every visible module in the library.
The Target type is defined in compiler/main/HscTypes.lhs. Note that a Target includes not just the file or module name, but also optionally the complete source text of the module as a StringBuffer: this is to support an interactive development environment where the source file is being edited, and the in-memory copy of the source file is to be used in preference to the version on disk.
The dependency analysis phase determines all the Haskell source files that are to be compiled or loaded in the current session, by traversing the transitive dependencies of the targets. This process is called the downsweep because we are traversing the dependency tree downwards from the targets. (The upsweep, where we compile all these files happens in the opposite direction of course).
The downsweep function takes the targets and returns a list of ModSummary consisting of all the modules to be compiled/loaded.
The ModSummary type
A ModSummary (defined in compiler/main/HscTypes.h) contains various information about a module:
- Its Module, which includes the package that it belongs to
- Its ModLocation, which lists the pathnames of all the files associated with the module
- The modules that it imports
- The time it was last modified
- ... some other things
We collect ModSumary information for all the modules we are interested in during the downsweep, below. Extracting the information about the module name and the imports from a source file is the job of compiler/main/HeaderInfo.hs which partially parses the source file.
Converting a given module name into a ModSummary is done by summariseModule in compiler/main/GHC.hs. Similarly, if we have a filename rather than a module name, we generate a ModSummary using summariseFile.
Loading (compiling) the Modules
When the dependency analysis is complete, we can load these modules by calling GHC.load. The same interface is used regardless of whether we are loading modules into GHCi with the :load command, or compiling a program with ghc --make: we always end up calling GHC.load.
The process in principle is fairly simple:
- Visit each module in the dependency tree from the bottom up, invoking HscMain to compile it (the upsweep).
- Finally, link all the code together. In GHCi this involves loading all the object code into memory and linking it with the RTS linker, and then linking all the byte-code together. In --make mode this involves invoking the external linker to link the object code into a binary.
The process is made more tricky in practice for two reasons:
- We might not need to compile certain modules, if none of their dependencies have changed. GHC's recompilation checker? determines whether a module really needs to be compiled or not.
- In GHCi, we might just be reloading the program after making some changes, so we don't even want to re-link modules for which no dependencies have changed.