extremlapin/fast_brainfuck.lua

High performance brainfuck transpiler/interpreter for Lua with FFI support. Very fast implementation with multiple optimization passes.

fast_brainfuck.lua's Issues

Ideas for improvement

After testing your transpiler i found a few limitations that you could remove:

1. programs that rely on wrapping cells don't work
2. moving the pointer to far left of cell 0 causes segmentation faults
3. dead code is not eliminated, e.g. a loop directly after another loop

Move to instructions CMP not lua code CMP in table

Support dynamic var folding with more than 1 assignation

Right now, it support for example :

	while data[i] ~= 0 do
		data[i] = data[i] - 1 (inc1)
		i = i + 9 (jmp1)
		data[i] = data[i] + 1 (inc2)
		i = i - 9 (jmp2, unused)
	end

To :

	data[i+jmp1] = data[i+jmp1] + (-(data[i]/inc1))*inc2
	data[i] = 0

---

	while data[i] ~= 0 do
		i = i - 1 (jmp1)
		data[i] = data[i] + 3 (inc1)
		i = i + 1 (jmp2)
		data[i] = data[i] - 1 (inc2)
	end

	data[i+jmp1] = data[i+jmp1] = (-(data[i]/inc2))*inc1
	data[i] = 0

jmp 1 = -jmp2

---

---

Would be nice to support :

			while data[i] ~= 0 do
				data[i] = data[i] - 1
				i = i - 1 (jmp1)
				data[i] = data[i] - 1
				i = i - 4 (jmp3)
				data[i] = data[i] + 1
				i = i + 5 (jmp3)
			end

jmp1 + jmp2 = -jmp3

Syntax level limit

Ran on OptimTease

chunk has too many syntax levels

Support more than 2^16 instructions

In LuaJIT, each function has an instruction limit of 2^16 because of how the bytecode register works.

I could do a final IR scan and find isolated loops and throw it in an external local function

I'll just leave this test program here

https://github.com/rdebath/Brainfuck/blob/broken-and-trivial/bitwidth.b

What is `Clanc C -Ofast`

I saw it in the benchmarks and I have never heard of it.