This is my hobby project to make a comprephensive, C99-compliant C compiler that supports precise garbage collection. My ultimate goal is to make it a drop-in replacement for GCC for projects using Boehm’s conservative GC.
I started developing the compiler May 2010, and as of July 2010, it can compile non-trivial programs, such as 8 queen problem solver. Still a lot of features are missing, like global variables, floating point numbers, declarations, miscellaneous operators, etc.
8cc is currently a two pass compiler. In the first pass, C source code is parsed with hand-written recursive-descent parser and operator-precedence parser, and compiled into IL (intermediate langauge). IL is a list of objects representing an instruction code and zero-or-more operands. Here is a partial list of instructions:
-
+ dst src0 src1 (addition)
-
ADDRESS dst src (pointer arithmetic “&”)
-
IF var then-block else-block continue-block
The operands of IF instruction are basic blocks, that don’t contain any branch instructions in the middle of it. All branch instructions jump to the beginning of a basic block. IF instruction, for example, is compiled to the machine codes tha jumps either then-block or else-block, depending on the value of variable “var”. “for” and “while” loops in C are implemented using IF operator. The program in IL consists of the entry basic block and all basic blocks reachable from the entry.
The point of the first pass is that flattens nested C expressions. For example, C expression “A + B + C” is compiled to something like “T0 = A + B” and “T1 = T0 + C” (the result of the expression is T1.) No AST would be created. This should contributes compilation speed and simplifies the code by eliminating extra pass to convert AST to IL.
In the second pass, ILs are converted to x86-64 machine instructions. At the moment, all variables are assigned to the stack, and registers are used only as scratch memory. For instance, “A = B + C” is compiled to the code that (1) loads the value of variable B in the stack to register RAX, (2) loads C to register R11, (3) addd them, and (4) stores the result in RAX to varaible A. It always copies values between the registers and the stack even if the value will be used just after the current instruction.
8cc directly outputs ELF object file, like TCC but unlike the most other ccompilers. No assembler is needed.
At last, GCC would be invoked to link an object file to make an executable.
I’ll implement all C99 features. After that, I’d go with copying GC. Because compiler knows precise type, runtime can do precise GC as long as compiler retains type information to a binary in the form that is accessible from runtime. Type of heap-allocated objects would be detected by recognizing the special form in source code, malloc(sizeof(type) * size). Once it’s done, I’d implement flow analysis and SSA conversion, and then implement efficient register allocation algorithm, such as Linear Scan Register Allocation to improve performance.