[haberman]

Blog entry describing my 100-line x86-64 Brainfuck JIT (built on DynASM): http://blog.reverberate.org/2012/12/hello-jit-world-joy-of-s...

(Direct link to the code: https://github.com/haberman/jitdemo/blob/master/jit3.dasc)

[Scaevolus]

I've been looking for a DynASM tutorial for ages, and wanted to use it to make a fast BF interpreter-- this is great!

I added some optimizations using Flex to identify patterns, and got a 3.8x speedup on the Mandelbrot benchmark -- https://github.com/rmmh/beefit

[haberman]

Very nice! I wouldn't have thought to use flex to recognize optimizable patterns. :) It crashes for me on OS X but I'll have to try it on Linux.

You may have tried this already, but where you have:

    mov al, byte [PTR]

Usually you want to write this instead to avoid a partial register stall (also in case there's junk sitting in the register).

    movzx  eax, byte [PTR]

[Scaevolus]

Hm, the OSX crash is probably because they have a different ABI from Linux -- Google "osx ia32 abi" for details-- I'm probably violating stack alignment.

Partial register stalls are where you write part of the register, then read the full register:

    mov al, byte [PTR]
    add ebx, eax

But I am emitting this:

    mov al, byte [PTR]
    add byte [PTR+3], al

So there's no stall.

I got jit4 about as far as is reasonable for a single-pass compiler, but I intend to implement a proper BF->IR->ASM compiler to implement some more complex multi-pass optimizations I have in mind.

[haberman]

Yes, I have encountered the problem before that OS X crashes on incorrect stack alignment even where other platforms tolerate it.

Good point about the stall -- I always thought that partial register stall happened at the point that you do the partial write, since the logical contents of the register now depend on its previous contents. I didn't realize that the dependency logic was sophisticated enough to allow the partial read without depending on the entire register's value.