Writing a GCC Front End
GCC, the premier free software compiler suite, has undergone many changes in the last few years. One change in particular, the merging of the tree-ssa branch, has made it much simpler to write a new GCC front end.
GCC always has had two different internal representations, trees and RTL. RTL, the register transfer language, is the low-level representation that GCC uses when generating machine code. Traditionally, all optimizations were done in RTL. Trees are a higher-level representation; traditionally, they were less documented and less well known than RTL.
The tree-ssa Project, a long-term reworking of GCC internals spearheaded by Diego Novillo, changes all that. Now, trees are much better although still imperfectly documented, and many optimizations are done at the tree level. A side effect of this work on trees was the clear specification of a tree-based language called GENERIC. All GCC front ends generate GENERIC, which is later lowered to another tree-based representation called GIMPLE, and from there it goes to RTL.
What this means to you is that it is much, much simpler to write a new front end for GCC. In fact, it now is feasible to write a front end for GCC one without any knowledge of RTL whatsoever. This article provides a tour of how you would go about connecting your own compiler front end to GCC. The information in this article is specific to GCC 4.0, due to be released in 2005.
For our purposes, compilation is done in two phases, parsing and semantic analysis and then code generation. GCC handles the second phase for you, so the question is, what is the best way to implement phase one?
Traditional GCC front ends, such as the C and C++ front ends, generate trees during parsing. Front ends like these typically add their own tree codes for language-specific constructs. Then, after semantic analysis has completed, these trees are lowered to GENERIC by replacing high-level, language-specific trees with lower-level equivalents. One advantage of this approach is the language-specific trees usually are nearly GENERIC already. The lowering phase often can prevent too much garbage from generating.
The primary disadvantage of this approach is trees are typed dynamically. In theory, this might not seem so bad—many dynamically typed environments exist that can be used efficiently by developers, including Lisp and Python. However, these are complete environments, and GCC's heavily macro-ized C code doesn't confer the same advantages.
My preferred approach to writing a front end is to have a strongly typed, language-specific representation of the program, called an abstract syntax tree (AST). This is the approach used by the Ada front end and by gcjx, a rewrite of the front end for the Java programming language.
For instance, gcjx is written in C++ and has a class hierarchy that models the elements of the Java programming language. This code actually is independent of GCC and can be used for other purposes. In gcjx's case, the model can be lowered to GENERIC, but it also can be used to generate bytecode or JNI header files. In addition, it could be used for code introspection of various kinds; in practice, the front end is a reusable library.
This approach provides all the usual advantages of a strongly typed design, and in the GCC context, it results in a program that is easier to understand and debug. The relative independence of the resulting front end from the rest of GCC also is an advantage, because GCC changes rapidly and this loose coupling minimizes your exposure.
Potential disadvantages of this approach are the possibilities that your compiler might do more work than is strictly needed or use more memory. In practice, this doesn't seem to be too important.
Before we talk about some details of interfacing your front end to GCC, let's take a look at some of the documentation and source files you need to know. Because it hasn't been a priority in the GCC community to make it simpler to write front ends, some things you need to know are documented only in the source. The documentation references here refer to info pages and not URLs, because GCC 4.0 has not yet been released. Thus, the Web pages reflect earlier versions. Your best bet is to check out a copy of GCC from CVS and dig around in the source.
gcc/c.opt: describes command-line options used by the C family of front ends. More importantly, it describes the format of the .opt files. You'll be writing one of these.
gcc info page, node Spec Files (source file gcc/doc/invoke.texi): describes the spec minilanguage used by the GCC driver. You'll write some specs to tell GCC how to invoke your front end.
gccint info page, node Front End (source file gcc/doc/sourcebuild.texi): describes how to integrate your front end into the GCC build process.
gccint info page, node Tree SSA (source file gcc/doc/tree-ssa.texi): describes GENERIC.
gcc/tree.def, gcc/tree.h: some attributes of trees don't seem to be documented, and reading these files can help. tree.def defines all the tree codes and is, in large part, explanatory comments. tree.h defines the tree node structures, the many accessor macros and declares functions that are useful in building trees of various types.
libcpp/include/line-map.h: line maps are used to represent source code locations in GCC. You may or may not use these in your front end—gcjx does not. Even if you do not use them, you need to build them when lowering to GENERIC, as information in line maps is used when generating debug information.
gcc/errors.h, gcc/diagnostic.h: defines the interface to GCC's error formatting functions, which you may choose to use.
gcc/gdbinit.in: defines some GDB commands that are handy when debugging GCC. For instance, the pt command prints a textual representation of a tree. The file .gdbinit also is made in the GCC build directory; if you debug there, the macros immediately are available.
gcc/langhooks.h: lang hooks are a mechanism GCC uses to allow front ends to control some aspects of GCC's behavior. Each front end must define its own copy of the langhooks structures; these structures consist largely of function pointers. GCC's middle and back ends call these functions to make language-specific decisions during compilation. The langhooks structures do change from time to time, but due to the way GCC expects front ends to initialize these structures, you largely are insulated from these changes at the source level. Some of these lang hooks are not optional, so your front end is going to implement them. Others are ad hoc additions for particular problems. For instance, the can_use_bit_fields_p hook was introduced solely to work around an optimization problem with the current gcj front end.
Currently GCC requires your front end to be visible at build time—there is no way to write a front end that is built separately and linked against an installed GCC. For this step, read through the appropriate section of the GCC manual to find out how to write the build infrastructure needed for your front end. Ordinarily, the simplest way is to copy another front end's files and modify them to suit.
Next, write two files to help integrate your front end into the GCC driver program. The lang-specs.h file describes your front end to the GCC driver. It tells the driver the file extensions that, when seen on the command line, should cause GCC to invoke your front end. It also gives the driver some instructions for what other programs must be run, such as whether the assembler should be run after your front end and how to pass or modify certain command-line options. It may take a while to write this file, as specs are their own strange language. However, examples in the other front ends can help.
The lang.opt file describes any command-line options specific to your front end. This is a plain-text file written in a straightforward format. Simple options, such as warning flags, can be put in lang.opt and do not require any other code on your part. Other arguments have to be handled by a lang hook you must write.
Next, implement the lang hooks needed to drive the compilation process. The important ones in this category are:
init_options: the first call made to your front end, before any option processing is done.
handle_option: called to handle a single command-line option.
post_options: called after all command-line processing has been done. This lang hook also is a convenient place to determine the name of the input file to parse.
init: called after post_options to initialize your front end.
finish: called after all compilation is done. You can use this to clean up after your front end, if necessary.
parse_file: a lang hook that does all the parsing, semantic analysis and code generation needed for the input file. It does all the actual work of compilation.
GCC needs your front end to do some initialization. Most of GCC is self-initializing, but in order to accommodate the needs of different front ends, it is possible to initialize some tree-related global variables in atypical ways. I recommend not trying to delve too deeply into this. It is simpler to define the standard tree nodes in the standard ways and to think up your own names for trees representing, say, the standard types in your language.
During initialization you want to call build_common_tree_nodes, set_sizetype and build_common_tree_nodes_2. set_sizetype is used to set the type of the internal equivalent of size_t; it is simplest to set this always to long_unsigned_type_node.
Other setup steps can be done in this phase. For instance, in the initialization code for gcjx, we build types representing various structures that we need to describe Java classes and methods.
Your parse_file lang hook calls your compiler to generate your internal data structures. Assuming this completes without errors, your front end now is ready to generate GENERIC trees from your AST. In gcjx, this is done by walking the AST for a class using a special visitor API. The GENERIC-specific implementation of this API incrementally builds trees representing the code and then hands this off to GCC.
All the details of generating trees are outside the scope of this article. Below are examples, however, showing three major tree types so you can see what each looks like.
One kind of tree represents a type. Here is an example from gcjx of the Java char type:
tree type_jchar = make_node (CHAR_TYPE); TYPE_PRECISION (type_jchar) = 16; fixup_unsigned_type (type_jchar);
You can represent any type using trees. In particular, there are tree types representing records, unions, pointers and integers of various sizes.
Decl represents a declaration or, in other words, a name given to some object. For instance, a local variable in the source code is represented by a decl:
tree local = build_decl (VAR_DECL, get_identifier ("variable_name"), type_jchar);
There are decls representing various named objects in a program: translation units, functions, fields, variables, parameters, constants, labels and types. A type decl represents the declaration of the type, as opposed to the type itself.
Many kinds of expr trees are available that represent the various kinds of expressions in a program. These are similar to C expressions but are more general in some ways. For instance, trees do not distinguish between if statements and conditional expressions—both are represented by a COND_EXPR, with the only difference being that an if statement has void type. Here's how we can build an expression that adds our variable to itself:
tree addition = build2 (PLUS_EXPR, type_jchar, local, local);
Trees that represent statements are linked together using a special iterator API. Here is how we would chain together two statements, s1 and s2:
tree result = alloc_stmt_list (); tree_stmt_iterator out = tsi_start (result); tsi_link_after (&out, s1, TSI_CONTINUE_LINKING); tsi_link_after (&out, s2, TSI_CONTINUE_LINKING); // Now `result' holds the list of statements.
Other kinds of tree nodes exist; read tree.def and the manual for a more complete understanding. It also is possible for a front end to define its own tree codes; however, if you have your own AST, you should not need to do this.
The overall structure of the program you generate probably is going to resemble a translation unit decl, which would contain types, variables and functions.
Once you've built the trees representing a function, a global variable or a type for which you want to generate debugging information, you need to pass the tree to the appropriate function to handle the rest of compilation. Three such functions are available at present: rest_of_decl_compilation handles compilation for a decl node, cgraph_finalize_function handles compilation for a function and rest_of_type_compilation handles compilation for a type.
Although GCC has a fair number of internal consistency checks, it still is too easy to provoke crashes in code that are unrelated to your front end. In many cases, you can move up the stack, printing whatever trees are being manipulated, until you find some discrepancy caused by incorrect tree generation. This technique requires surprisingly little general GCC knowledge in order to effectively debug your code.
GCC has some handy debug functionality. In the debugger you can call the debug_tree function to print a tree. You also can use the -fdump-tree family of command-line options to print trees after various passes have been run.
My experience writing gcjx has been that lowering its strongly typed intermediate representation to trees is quite simple. The tree back end to gcjx, one back end among several, represents roughly 10% of the total code of the compiler. Although unfinished, it currently weighs in at 6,000 lines of code (raw wc -l count)—around the same size as the bytecode back end. One inference to draw from this is if you already have a compiler, the task of attaching it to GCC can be accomplished easily.
As trees are high-level, I haven't looked at any RTL while writing this front end. I haven't spent any time at all thinking about or dealing with processor-specific issues. Unless your language has some esoteric requirements, this ought to hold for you as well.
The statically typed AST in gcjx is easily reused. Currently, there are four back ends, and I expect to write more later. For instance, it would be simple to build a back end that writes a cross-reference representation of the program to a database. Or, it would be possible to write a back end that walks the AST searching for typical errors, akin to the FindBugs program. This consideration would be even more compelling for languages, which, unlike Java, don't already have a wealth of analysis tools.
The process of writing a front end certainly could be made even easier. For instance, there is no need to require lang-specs.h. Instead, a front end could install a description file that the GCC driver would read at startup. Similarly, lang.opt probably could be dispensed with. With more work, it even would be possible to compile front ends separately from the rest of GCC.
Resources for this article: /article/8138.
Tom Tromey has been involved with free software since 1991 and has worked on many different programs. He currently is employed as an engineer at Red Hat, working on GCJ.