jit-demo

jit-demo copied to clipboard

This is an interpreter in C for a mini Scheme/Racket language that support functions, numbers, if0, and a top-level table of definitions --- and a just-in-time (JIT) compiler to machine code that specializes frequently called closures.

There's no reader. Programs are constructed direct in AST form using a C API, as demonstrated in "test.c".

Continuations are explicit. A standard eval--continue interpreter is implemented in "eval.c".

Memory is managed with a two-space copying collector. The test suite allocates an AST before the GC is enabled so that all the test variables don't have to be registered as GC roots.

Numbers are encoded either as record or using a fixnum encoding where the low bit is set to indicate a number instead of a pointer. Configure with FIXNUM_ENCODING in "struct.h". The fixnum encoding is required for native-code JIT mode.

The evaluator supports an optional compilation pass that converts a local-variable reference to a De Bruijn index and substitutes a top-level reference with its value (which creates a cyclic structure for a recursive function). This compiler is already a kind of just-in-time compiler, because it relies on having values for top-level definitions.

A JIT compiler to machine code is supported when USE_JIT is defined to 1 and when linked with GNU lightning. To build in that mode, build with something like

make CPPFLAGS="-DUSE_JIT=1" LIBS="-llightning"

The JIT specializes closures by recompiling the closure body (taking into accunt actual values in the closure) after a closure is called SPECIALIZE_AFTER_COUNT times. This specialization pass inlines simple functions. Setting SPECIALIZE_AFTER_COUNT to -1 effectively disables specialization.

Naturally, the "test.c" examples include variants of fib. Change "main.c" to run test_Y_fib() or test_alt_fib() to try variants. Specialization does little for Y-combinator recursion, but it eliminates the abstraction overhead for the "alt" variant.