proof-of-concept compiler engine using llvmlite

[cdetrio]

Supports just enough wasm instructions to run fibonacci.wasm.

(don't merge, this PR is only for reference).

$ python fib.py
calling visit on body: [['i32.const', 1], ['set_local', 1], ['block', [], [['get_local', 0], ['i32.const', 1], ['i32.lt_s'], ['br_if', 0], ['i32.const', 1], ['set_local', 3], ['i32.const', 0], ['set_local', 2], ['loop', [], [['get_local', 2], ['get_local', 3], ['i32.add'], ['set_local', 1], ['get_local', 3], ['set_local', 2], ['get_local', 1], ['set_local', 3], ['get_local', 0], ['i32.const', 4294967295], ['i32.add'], ['tee_local', 0], ['br_if', 0], ['end']]], ['end']]], ['get_local', 1], ['end']]
calling append_basic_block...
iterating over block_instrs...
got loop init block: %"loop.init" = loop.init:
iterating over loop block_instrs...
executing llvm code: ; ModuleID = "m_fibonacci_example"
target triple = "unknown-unknown-unknown"
target datalayout = ""

define i32 @"fn_fib"(i32 %".1")
{
entry:
  %".3" = alloca i32
  store i32 1, i32* %".3"
  br label %"block1"
block1:
  %".6" = alloca i32
  store i32 %".1", i32* %".6"
  %".8" = load i32, i32* %".6"
  %".9" = icmp slt i32 1, %".8"
  %".10" = alloca i32
  store i32 1, i32* %".10"
  %".12" = alloca i32
  store i32 0, i32* %".12"
  br label %"loop.init"
loop.init:
  %".15" = load i32, i32* %".12"
  %".16" = load i32, i32* %".10"
  %".17" = add i32 %".16", %".15"
  store i32 %".17", i32* %".3"
  %".19" = load i32, i32* %".10"
  store i32 %".19", i32* %".12"
  %".21" = load i32, i32* %".3"
  store i32 %".21", i32* %".10"
  %".23" = load i32, i32* %".6"
  %".24" = add i32 4294967295, %".23"
  store i32 %".24", i32* %".6"
  %".26" = icmp sgt i32 %".24", 0
  br i1 %".26", label %"loop.init.if", label %"loop.init.endif"
loop.init.if:
  br label %"loop.init"
loop.init.endif:
  br label %"contin-2"
contin-2:
  br label %"contin-1"
contin-1:
  %".31" = load i32, i32* %".3"
  ret i32 %".31"
}

verifying llvm module...
module verified!
c_fn_fib(9) = 55
0.029740958999999983
55

[poemm]

Wow! So impressed!

I ran it successfully.

It is impressive that you can write this barely touching spec_execution.py, and that compiler.py is almost standalone. I am curious whether you had to reference the spec when you wrote it, or if the code was enough.

It is good that you are starting with basic cases. I saw that they use the visitor pattern, and you visit by iterating over instructions in each block. I am curious if things will get complicated for some control flow cases.

I wish to help in this effort. I will read about llvmlite this weekend, and will think about whether there are any complicated cases.