##What? Brainfuck is a very minimal programming language which models a (limited) Turing Machine. It has a very small set of commmands which makes it an ideal target for a toy compiler project, like this one.

##Why? This was inspired by Erik's post, as well as my interest in building some form of a compiler. This work follows on from the assembler I started building for the early version of Notch's DCPU.

##How?

1. Parse the program
2. Generate code
   1. Pull blocks out of a look-up-table
   2. Construct fixup table
3. Link jumps and printf calls
4. mprotect the code buffer to make it executable
5. Jump to it

1. Build (with buildjit.sh)
2. Run with a brainfuck program in argv[1], e.g. ./jit '>+++++++++[<++++++++>-]<.>+++++++[<++++>-]<+.+++++++..+++.[-]>++++++++[<++++>-]<.>+++++++++++[<++++++++>-]<-.--------.+++.------.--------.[-]>++++++++[<++++>-]<+.[-]++++++++++.' (From Speedy's implementation over at helloworld.org)

##Issues

• Doesn't support input, yet.
• Handling of invalid commands (AKA comments) is brittle - spaces break it, for instance.
• Isn't doing anything particularly clever. Since the IR is a list with some metadata for loops it's not possible to do any fancy optimisations.
• The lookup-table codegen approach is klunky. Since the code blocks are fixed, smart things like register allocation aren't possible. I guess since block-level info is available it should be feasible to deduce where fragments that repeatedly load & store the same cell could be exchanged for ones that don't waste this time.
• It doesn't bounds-check the pointer, so it's vulnerable to memory corruption attacks.