Skip to content

mbrukman/porffor

 
 

Repository files navigation

Porffor  /ˈpɔrfɔr/  (poor-for)

A from-scratch experimental AOT optimizing JS/TS -> Wasm/C engine/compiler/runtime in JS. Not serious/intended for (real) use. (this is a straight forward, honest readme)
Age: ~6 months (very on and off)

Design

Porffor is a very unique JS engine, due many wildly different approaches. It is seriously limited, but what it can do, it does pretty well. Key differences:

  • 100% AOT compiled (no JIT)
  • No constant runtime/preluded code
  • Least Wasm imports possible (only I/O)

Porffor is primarily built from scratch, the only thing that is not is the parser (using Acorn). Binaryen/etc is not used, we make final wasm binaries ourself. You could imagine it as compiling a language which is a sub (some things unsupported) and super (new/custom apis) set of javascript. Not based on any particular spec version, focusing on function/working over spec compliance.

Limitations

  • No full object support yet
  • Little built-ins/prototype
  • No async/promise/await
  • No variables between scopes (except args and globals)
  • Literal callees only in calls (eg print() works, a = print; a() does not)
  • No eval() etc (since it is AOT)

Rhemyn

Rhemyn is Porffor's own regex engine; it compiles literal regex to Wasm bytecode AOT (remind you of anything?). It is quite basic and WIP. See its readme for more details.

2c

2c is Porffor's own Wasm -> C compiler, using generated Wasm bytecode and internal info to generate specific and efficient/fast C code. Little boilerplate/preluded code or required external files, just for CLI binaries (not like wasm2c very much).

Supported

See optimizations for opts implemented/supported.

Proposals

These include some early (stage 1/0) and/or dead (last commit years ago) proposals but I think they are pretty neat, so.

Math proposals (stage 1/0)

Language

  • Number literals
  • Declaring functions
  • Calling functions literal callees only
  • return
  • let/const/var basic declarations
  • Some basic integer operators (+-/*%)
  • Some basic integer bitwise operators (&|)
  • Equality operators (==, !=, etc)
  • GT/LT operators (>, <, >=, etc)
  • Some unary operators (!, +, -)
  • Logical operators (&&, ||)
  • Declaring multiple variables in one (let a, b = 0)
  • Global variables (var/none in top scope)
  • Functions returning 1 number
  • Bool literals as ints (not real type)
  • if and if ... else
  • Anonymous functions
  • Setting functions using vars (const foo = function() { ... })
  • Arrow functions
  • undefined/null as ints (hack)
  • Update expressions (a++, ++b, c--, etc)
  • for loops (for (let i = 0; i < N; i++), etc)
  • Basic objects (hack)
  • console.log (hack)
  • while loops
  • break and continue
  • Named export funcs
  • IIFE support
  • Assignment operators (+=, -=, >>=, &&=, etc)
  • Conditional/ternary operator (cond ? a : b)
  • Recursive functions
  • Bare returns (return)
  • throw (literals only)
  • Basic try { ... } catch { ... } (no error given)
  • Calling functions with non-matching arguments (eg f(a, b); f(0); f(1, 2, 3);)
  • typeof
  • Runtime errors for undeclared variables (ReferenceError), not functions (TypeError)
  • Array creation via [] (eg let arr = [ 1, 2, 3 ])
  • Array member access via arr[ind] (eg arr[0])
  • String literals ('hello world')
  • String member (char) access via str[ind] (eg str[0])
  • String concat (+) (eg 'a' + 'b')
  • Truthy/falsy (eg !'' == true)
  • String comparison (eg 'a' == 'a', 'a' != 'b')
  • Nullish coalescing operator (??)
  • for...of (arrays and strings)
  • Array member setting (arr[0] = 2, arr[0] += 2, etc)
  • Array constructor (Array(5), new Array(1, 2, 3))

Built-ins

  • NaN and Infinity (f64 only)
  • isNaN() and isFinite() (f64 only)
  • Most of Number (MAX_VALUE, MIN_VALUE, MAX_SAFE_INTEGER, MIN_SAFE_INTEGER, POSITIVE_INFINITY, NEGATIVE_INFINITY, EPSILON, NaN, isNaN, isFinite, isInteger, isSafeInteger) (some f64 only)
  • Some Math funcs (Math.sqrt, Math.abs, Math.floor, Math.sign, Math.round, Math.trunc, Math.clz32, Math.fround, Math.random) (f64 only)
  • Basic globalThis support
  • Basic Boolean and Number
  • Basic eval for literals
  • Math.random() using self-made xorshift128+ PRNG
  • Some of performance (now())
  • Some of Array.prototype (at, push, pop, shift, fill)
  • Some of String.prototype (at, charAt, charCodeAt)

Custom

  • Supports i32, i64, and f64 for valtypes
  • Start of a SIMD api (docs needed)
  • Intrinsic functions (see below)
  • Inlining wasm via asm`...` "macro"

Todo

No particular order and no guarentees, just what could happen soon™

  • Arrays
    • More of Array prototype
    • Arrays/strings inside arrays
    • Destructuring
  • Objects
    • Basic object expressions (eg {}, { a: 0 })
  • Wasm
    • Basic Wasm engine (interpreter) in JS
  • More math operators (**, etc)
  • do { ... } while (...)
  • Rewrite console.log to work with strings/arrays
  • Exceptions
    • Rewrite to use actual strings (optional?)
    • try { } finally { }
    • Rethrowing inside catch
  • Optimizations
    • Rewrite local indexes per func for smallest local header and remove unused idxs
    • Smarter inline selection (snapshots?)
    • Remove const ifs (if (true), etc)
    • Experiment with byte strings?
  • Runtime
    • WASI target
    • Run precompiled Wasm file if given
  • Cool proposals
  • Posts
    • Inlining investigation
  • Self hosted testing?

Performance

For the things it supports most of the time, Porffor is blazingly fast compared to most interpreters, and common engines running without JIT. For those with JIT, it is not that much slower like a traditional interpreter would be; mostly the same or a bit faster/slower depending on what.

Screenshot of comparison chart

Optimizations

Mostly for reducing size. I do not really care about compiler perf/time as long as it is reasonable. We do not use/rely on external opt tools (wasm-opt, etc), instead doing optimization inside the compiler itself creating even smaller code sizes than wasm-opt itself can produce as we have more internal information.

Traditional opts

  • Inlining functions (WIP, limited)
  • Inline const math ops
  • Tail calls (behind flag -tail-call)

Wasm transforms

  • local.set, local.get -> local.tee
  • i32.const 0, i32.eq -> i32.eqz
  • i64.extend_i32_s, i32.wrap_i64 -> ``
  • f64.convert_i32_u, i32.trunc_sat_f64_s -> ``
  • return, end -> end
  • Change const, convert to const of converted valtype (eg f64.const, i32.trunc_sat_f64_s -> i32.const`)
  • Remove some redundant sets/gets
  • Remove unneeded single just used vars
  • Remove unneeded blocks (no brs inside)
  • Remove unused imports
  • Use data segments for initing arrays/strings
  • (Likely more not documented yet, todo)

Wasm module

  • Type cache/index (no repeated types)
  • No main func if empty (and other exports)
  • No tags if unused/optimized out

Test262

Porffor can run Test262 via some hacks/transforms which remove unsupported features whilst still doing the same asserts (eg simpler error messages using literals only). It currently passes >10% (see latest commit desc for latest and details). Use node test262 to test, it will also show a difference of overall results between the last commit and current results.

Codebase

  • compiler: contains the compiler itself

    • builtins.js: all built-ins of the engine (spec, custom. vars, funcs)
    • codeGen.js: code (wasm) generation, ast -> wasm. The bulk of the effort
    • decompile.js: basic wasm decompiler for debug info
    • embedding.js: utils for embedding consts
    • encoding.js: utils for encoding things as bytes as wasm expects
    • expression.js: mapping most operators to an opcode (advanced are as built-ins eg f64_%)
    • index.js: doing all the compiler steps, takes code in, wasm out
    • opt.js: self-made wasm bytecode optimizer
    • parse.js: parser simply wrapping acorn
    • sections.js: assembles wasm ops and metadata into a wasm module/file
    • wasmSpec.js: "enums"/info from wasm spec
    • wrap.js: wrapper for compiler which instantiates and produces nice exports
  • runner: contains utils for running JS with the compiler

    • index.js: the main file, you probably want to use this
    • info.js: runs with extra info printed
    • repl.js: basic repl (uses node:repl)
  • rhemyn: contains Rhemyn - our regex engine (used by Porffor)

    • compile.js: compiles regex ast into wasm bytecode
    • parse.js: own regex parser
  • test: contains many test files for majority of supported features

  • test262: test262 runner and utils

Usecases

Basically none (other than giving people headaches). Potential ideas to come?

Usage

Basically nothing will work :). See files in test for examples.

  1. Clone repo
  2. npm install
  3. node test to run tests (some will fail)
  4. node runner path/to/code.js to run a file (or node runner to use wip repl)

You can also use Deno (deno run -A ... instead of node ...), or Bun (bun ... instead of node ...).

Options

  • -target=wasm|c|native (default: wasm) to set target output (native compiles c output to binary, see args below)
  • -target=c|native only:
    • -o=out.c|out.exe|out to set file to output c or binary
  • -target=native only:
    • -compiler=clang to set compiler binary (path/name) to use to compile
    • -cO=O3 to set compiler opt argument
  • -parser=acorn|@babel/parser|meriyah|hermes-parser (default: acorn) to set which parser to use
  • -parse-types to enable parsing type annotations/typescript. if -parser is unset, changes default to @babel/parser. does not type check
  • -opt-types to perform optimizations using type annotations as compiler hints. does not type check
  • -valtype=i32|i64|f64 (default: f64) to set valtype
  • -O0 to disable opt
  • -O1 (default) to enable basic opt (simplify insts, treeshake wasm imports)
  • -O2 to enable advanced opt (inlining). unstable
  • -O3 to enable advanceder opt (precompute const math). unstable
  • -no-run to not run wasm output, just compile
  • -opt-log to log some opts
  • -code-log to log some codegen (you probably want -funcs)
  • -regex-log to log some regex
  • -funcs to log funcs
  • -ast-log to log AST
  • -opt-funcs to log funcs after opt
  • -sections to log sections as hex
  • -opt-no-inline to not inline any funcs
  • -tail-call to enable tail calls (experimental + not widely implemented)
  • -compile-hints to enable V8 compilation hints (experimental + doesn't seem to do much?)

VSCode extension

There is a vscode extension in porffor-for-vscode which tweaks JS syntax highlighting to be nicer with porffor features (eg highlighting wasm inside of inline asm).

Isn't this the same as AssemblyScript/other Wasm langs?

No. they are not alike at all internally and have very different goals/ideals:

  • Porffor is made as a generic JS engine, not for Wasm stuff specifically
  • Porffor primarily consumes JS
  • Porffor is written in pure JS and compiles itself, not using Binaryen/etc
  • (Also I didn't know it existed when I started this, lol)

FAQ

1. Why the name?

purple in Welsh is porffor. Why purple?

  • No other JS engine is purple colored
  • Purple is pretty cool
  • Purple apparently represents "ambition", which is.. one word to describe this project
  • The hard to speak name is also the noise your brain makes in reaction to this idea!

2. Why at all?

Yes!

3. But what about spec compliance?

Lol, no. (sorry.)

About

A from-scratch experimental AOT JS -> Wasm engine in JS

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • JavaScript 84.5%
  • HTML 10.9%
  • TypeScript 4.6%