The language is strongly typed, lexically scoped, and provides direct access to memory through load/stores. This is a low-level language without real data structures, lambdas/closures, nor memory management, but those can be written in this language.

The language is designed to expose the low-level flexibility of Wasm directly but in a friendly manner, which hides most of the complexity of the Wasm format.

It is intended to be used as for low-level Wasm programs, such as language runtimes (incl. memory allocators and gcs) for higher-level languages, or as a target for languages that want to compile directly to Wasm. It also helps explain how Wasm works by hiding some of the low-level details so the structure and semantics are clarified.

Introduction

Make WebAssembly Easy. WASE is a language, which tries to make WASM easy to write
directly. The language maps closely to WebAssembly, and compiles directly to Wasm
bytecode. As such, it can be seen as an alternative to the WAT text format,
but it is easier to use because of these things:

It uses more conventional syntax rather than S-expr
It has type inference
It handles all the index and table business for you

The language is strongly typed, lexically scoped, and provides direct access
to memory through load/stores. This is a low-level language without real data
structures, lambdas/closures, nor memory management, but those can be written
in this language.

The language is designed to expose the low-level flexibility of Wasm directly
but in a friendly manner, which hides most of the complexity of the Wasm format.

It is intended to be used as for low-level Wasm programs, such as language
runtimes (incl. memory allocators and gcs) for higher-level languages, or as
a target for languages that want to compile directly to Wasm. It also helps
explain how Wasm works by hiding some of the low-level details so the structure
and semantics are clarified.

Usage

To compile a tests/euler1.wase program to tests/euler1.wasm, clone this repository
and then use

bin\wase tests/euler1.wase

This requires a suitable Java Runtime in your path.

You can also consider to add wase/bin to your path.

You can also run it directly using java:

java -jar bin\wase.jar tests/euler1.-wase

Example

Here is a solution to the Project Euler challenge 1 (https://projecteuler.net/problem=1)
written in Wase:

// This includes Wasi imports and a printi32 helper
include lib/runtime;

// For Wasi, we have to export the memory
export memory 1;

// Iterative version using a loop
euler1Loop(limit : i32) -> i32 {
	var start = 1;
	var acc = 0;
	loop {
		// Breaks out of the function when start >= limit with the value acc
		break_if<1>(acc, start >= limit);
		if (start % 3 == 0 | start % 5 == 0) {
			acc := acc + start;
		};
		start := start + 1;
		// This is really continue and loops
		break<>();
	};
	// This is never reached, but we have to add this to get the right type
	acc
}

// A recursive implementation
foldRange(start : i32, end : i32, acc : i32) -> i32 {
	if (start <= end) {
		foldRange(start + 1, end, if (start % 3 == 0 | start % 5 == 0) {
			acc + start;
		} else acc)
	} else {
		acc;
	}
}

euler1(limit : i32) -> i32 {
	foldRange(1, limit - 1, 0);
}

// Wasi expects us to have a "_start" function exported
export _start() -> () {
	printi32(euler1Loop(1000)); // Correct: 233168
	printByte(10); // New line
	printi32(euler1(1000)); // Correct: 233168
	printByte(10); // New line
	{}
}

Compile with

bin\wase tests/euler1.wase

and run with wasmer (install)

wasmer tests/euler1.wasm
233168
233168

You can also run with wasm3, but it does not have as deep a stack, so the
recursion above causes a stack overflow unless you provide a bigger stack to
wasm3.

To compare Wase against Wat, check out the tests/ folder where the .wase
files have been compiled to .wasm and then decompiled to .wat.

Status

Beta. The compiler works, and the compiler can parse, type and compile
most instructions directly to WASM binaries that validate and run correctly.

One problem in daily use is that error messages are currently without positions.

This compiler is not validating. For example, you can use dynamic code in constant
contexts, which will be rejected by the Wasm runtime.

Join the flow9 discord, which also hosts discussions on Wase:

https://discord.gg/9gGJu6KU

The Wase Programming Language

Type system

Wase supports the types of Wasm:

i32, i64, f32, f64, v128, func, extern

At compile time, however, the compiler uses the real types for functions to ensure type
checking is complete. The syntax for functions is like this:

// A function that takes an i32 and a f64, and returns a i32
foo(i32, f64) -> i32 { 42 }

// Does not take any arguments, does not return anything
foo() -> () { }

// Multi-values is supported with this syntax: This function returns both an i32 and a f64
foo(i32) -> (i32, f64) { 
	[ 42, 3.141 ]
}

As a special type, there is also auto, where the type will be inferred by
the compiler:

// The return type of this function is inferred to be "i32"
bar() -> auto {
	1;
}

This also works for locals:

bar() -> auto {
	a : auto = 1;
	b = 2.0; // This is implicitly auto, and thus inferred
	a
}

The type inference is based on Hindley-Milner style unification, so it should be robust.

You can use explicit type annotations to verify types:

foo() -> auto {
	bar : i32
}

There are no implicit type conversions.

The compiler tracks whether variables and globals are mutable or not:

foo() {
	a = 1;
	a := 2; // Not allowed.

	var b = 1;
	b := 2; // This is allowed, since b is "var"
}

c : mutable i32 = 1;
bar() {
	c := 2; // This is allowed
}

Top-level Syntax

At the top-level, we have syntax for the different kinds of sections in
Wasm. This includes globals, functions, imports, tables and data.

The compiler will reorder the top-level, so imports, tables, memory and data
come first in the original order, and then after that, globals and functions
in their original order, in accordance with Wasm requirements.

Functions can call each other in mutual recursion, but globals can only refer
to globals that occur before themselves.

Global syntax

The grammar for globals is like this:

// Constant global
<id> : <type> = <expr>;

// Mutable global
<id> : mutable <type> = <expr>;

// Exported global
export <id> : <type> = <expr>;

// Exported mutable global
export <id> : mutable <type> = <expr>;

Some examples:

pi : f64 = 3.14159265359;
counter : mutable i32 = 0;
// Exports this global to the host with the name "secret"
export secret : i32 = 0xdeadbeaf;
// Exports this mutable global with the name "changes"
export changes : mutable f64 = 1.023;
// Exports this global to the host with the name "external"
export "external" internal : i32 = 1;

Function syntax

// Functions
foo(arg : i32) -> i32 {
	a : i32 = 1;
	a + arg
}

// This function is exported to the host using the name "foo"
export foo(arg : i32) -> i32 {
	arg
}

// Exports this function using the "_start" name to the host, in compliance with Wasi
export "_start" start() -> () {
	...
}

main() -> () {
	// This is the initial function automatically called otherwise
	...
}

If the program contains a function called main, it is marked as the
function that starts the program.

Include

Wase code can be split into multiple files, and included using this syntax at the
top-level:

include lib/runtime;

The compiler will read the file wase/lib/runtime.wase, parse it, and splice the
result into that place in the code. If a file is included more than once (also
transitively), it is only included the first occurrence.

The program is only type checked after includes are resolved.

Expressions

The body of globals and functions are expressions. There are no
statements, but only expressions. The syntax is close to the
typescript family:

h32 : i32 = 1;	// Int constants are i32
h32 : i32 = 0xDEADBEEF; // Hex notation

h64 : i64 = 1w; // w suffix for i64
h64 : i64 = 0x12345678DEADBEEFw; // Hex notation with suffix w for i64

fl32 : f32 = 0xDEADBEEFn; // Binary representation

fl64 : f64 = 10.0; // Floats are f64 by default
fl64 : f64 = 0x2345678DEADBEEFh; // Binary representation

// Let-binding have scope
a : i32 = 1;
<scope> // "a" is defined in this scope

var v : i32 = 1;
v := 2; // set local or global, when mutable

// Arithmetic, all signed.
1 + 2 / 3 * 4 % 5

// f64 arithmetic
23.4 + 2.3 * 3.141

// Bitwise operations
1 & 3 | 5 ^ 7

// Comparisons are signed. Use instructions for unsigned comparisons
1 < 5 | 5 >= 2 & a == 2.0 & lt_u<>(b, c) 

// Tuples, aka multi-values
[1, 2.0, 45]

// Tuple-let: This expands into multiple let-bindings with names like "t.0" and "t.1"
t : (i32, f64) = [1, 3.141];

Control flow

// Function call
foo(1, 2) 

if (a) b else c;
if (a) b;

// Sequence
{
	a();	// An implicit drop is added here
	b();	// An implicit drop is added here
	[1, 2];  // Two implicit drops are added here
	c
}

// Return from the function.
foo() -> i32 {
	if (early) {
		return 1;	 // Has to match function return type
	}
	code;
}
bar() -> () {
	if (early) {
		// Matches no result from function return type
		return;
	}
	code
}

Blocks

Wasm uses blocks for control flow. Each function introduces a block.
The block and loop constructs also do that, as well as the then
and else branches of if. The break and break_if instructions
refer to this stack of blocks. A break with level 0 breaks out of the
inner most block, while break<1>() breaks out of the next level. The number defines what parent block to break to.

block {
	code;
	// This breaks to the end of the block
	break_if<>(earlyStop);
	code;
}

loop {
	code;
	// In loops, the break goes to the top of the loop,
	// so this is an infinite loop
	break<>();
}

foo() -> () {
	loop {
		code;
		// This breaks out of the function
		break_if<1>(stopCondition);
		// This is really continue
		break<>();
	}
}

foo () -> f64 {
	loop {
		// Here, the break returns 3.141 from the function
		// since block one level up is the function, which
		// returns f64
		break_if<1>(3.141, earlyStop)
	}
}

// Here is a simple do-while loop, which prints from 1 to 10 and then F,
// but not E.
block {
	var i = 1;
	loop {
		printi32(i);
		printByte(10);
		i := i + 1;
		// This breaks out of the upper loop
		break_if<1>(i > 10);
		// This is really continue inside this loop
		break<>();
	};
	// Never reached "E"
	printByte(69);
	{}
};
// This is "F"
printByte(70);

TODO: illustrations (blocks with colors? with colorful brackets / font)

Indirect calls

Wase helps make indirect calls easier to work with. Consider a set of functions like this:

foo() -> i32 { 42 }
bar() -> i32 { 666	}
add1(v : i32) -> i32 { v + 1 }
sub1(v : i32) -> i32 { v - 1 }

Now it is possible to treat these functions as pointers, and call them indirectly using
call_indirect using the fn<id> construct:

call_indirect<>(fn<bar>());  // 666
call_indirect<>(fn<foo>());  // 42

call_indirect<>(41, fn<add1>()); // 42
call_indirect<>(43, fn<sub1>()); // 43

Behind the scenes, Wase will collect all fn<id> in the program, construct a element table for
them, and resolve them into i32 as required by call_indirect. If you treat these functions
as first order, the type of fn<bar> is i32. Also notice that the type of the indirect function is
inferred, so you might have to add a type annotation to resolve what the resulting type is:

myfn : i32 = if (cond) fn<bar>() else fn<foo>();
// We add a type annotation to "explain" that this call is for functions of type () -> i32 
call_indirect(myfn) : i32;

Low level instructions

The low-level instructions in Wase have the form

instruction<pars>(args)

where pars are parameters for the operation decided at compile time,
while args are arguments to the instruction put on the stack.

The instruction has the same name as the corresponding Wasm WAT format.

Load/store

Loads from memory are written like this:

load<>(index);

The width of the load is inferred from the use of the value.

// This is i32 load, since a is i32
a : i23 = load<>(0);

// This is f32 load, since f is f32
f : f32 = load<>(32):

Stores are written like this:

store<>(index, value);

The width of the store is inferred from the type of the value.

// This is f64.store, because 2.0 is f64
store<>(32, 2.0);

Loads and stores also exist in versions that work with smaller bit-widths:

// Loads a byte from the given memory address, sign extending it into a i32 or i64
signedByteValue : i32 = load8_s<>(0);
signedByteValue64 : i64 = load8_s<>(0);

// Loads an unsigned byte from the given memory address into a i32 or i64
byteValue : i32 = load8_u<>(0);
byteValue64 : i64 = load8_u<>(0);

sword : i32 = load16_s<>(0);
sword64 : i64 = load16_s<>(0);
uword : i32 = load16_u<>(0);
uword64 : i64 = load16_u<>(0);

sint32 : i64 = load32_s<>(0);
uint32 : i64 = load32_u<>(0);

// Stores the byte "32" at address 0
store8<>(0, 32)

// The same, except it comes from an i64
int64 : i64 = big;
store8<>(0, big); // Only picks the lower 8 bits of "big"

// Stores the lower 16 bits of the value at the given address
store16<>(address, value);

// Stores the lower 32 bits of the value at the given address
store32<>(address, value);

You can also define the offset and alignment explicitly:

load<offset>(index)
load<offset, align>(index)
store<offset>(index, value)
store<offset, alignment>(index, value)

The alignment is expressed as what power of 2 to use:

v : i32 = load8_u<0, 0>(i) // alignment at 1 byte
v : i32 = load16_u<0, 1>(i) // alignment at 2 bytes
v : i32 = load32_u<0, 2>(i) // alignment at 4 bytes
v : i64 = load_u<0, 3>(i) // alignment at 8 bytes

TODO: illustrations (offset, alignment)

Instructions

4 instructions plus all v128 SIMD instructions are not implemented yet:

table.init and elem.drop
memory.init and data.drop

Control instructions

| Wasm | Wase | Comments |
|-|-|-|
| block | block { exp }| Type is inferred
| loop | loop { exp } | Type is inferred
| if | if (cond) exp | Type is inferred
| ifelse | if (cond) exp else exp | Type is inferred
| unreachable | unreachable<>() |
| nop | nop<>() | No operation
| br | break<>() or break<int>() or break<>(val)or break<int>(val) | Default break level is 0. If there is a val, that is what we return with this break
| br_if | break_if<int>(cond) or break_if<>(cond) or break_if<int>(val, cond) or break_if<>(val, cond) | Default break level is 0. If there is a val, that is what the break returns
| br_table | br_table<l0, l1, l2, ..., ln, default>(index) or br_table<l0, l1, l2, ..., ln, default>(val, index) | Breaks the number of levels as indexed by the index. If a val is given, that is what the break returns. See tests/break_table.wase for details.
| return | return or return exp |
| call | fn(args) |
| call_indirect | call_indirect<>(args, fn<id>()) or call_indirect<>(fn<id>()) | The fn<id>() construct will convert the function id to an index into an automatically constructed element table of function index pointers.

Reference Instructions

Parametric Instructions

| Wasm | Wase | Comments |
|-|-|-|
| drop | drop<> or implicit in sequence {1;2} | The explicit drop<>() variant causes stack errors, since Wase is designed to be stack-safe.
| select | select<>(then, else, cond) | This is an eager if, where both then and else are always evaluated, but only one chosen based on the condition. This is branch-less so can be more efficient than normal if. (Automatically chooses the ref instruction version based on the type.)

Variable Instructions

Table Instructions

table.init and elem.drop not implemented yet. Requires elements first.

Memory Instructions

memory.init and data.drop not implemented yet. Requires data indexing.

Numeric Instructions

Vector Instructions

Memory instructions

Extract/replace lane instructions

Swizzle/splat instrucions

Relational i8x16 instructions

Relational i16x8 instructions

Relational i32x4 instructions

Relational i64x2 instructions

Relational f32x4 instructions

Relational f64x2 instructions

v128 instructions

i8x16 instructions

i16x8 instructions

i32x4 instructions

i64x2 instructions

f32x4 instructions

f64x2 instructions

Trunc/convert instructions

Wasm Sections

Imports of globals and functions

Use this syntax to import a function from the host:

// Function import from host
import println : (i32) -> void = console.log;

Notice imports have to be the first thing in the program.

The same works for globals:

// Global import
import id : i32 = module.name;
import id : mutable i32 = module.name;

Memory

You have to explicitly define how much memory is available for the runtime.
To reserve memory, use syntax like this:

export? memory <min> <max>?;

or in case it is imported from the host:

import memory <min> <max>? = module.name;

Examples:

// Reserves one page of 64k
memory 1;

// Reserves 64k at first, but maximum 1mb
memory 1 4;

// Reserves 128k and exports this memory under the name "memory" to the host
// This is the form that Wasi likes
export memory 2;

// Reserves and exports memory under the name "mymem"
export "mymem" memory 1 4;

// 64k Memory import from the host
import memory 1 = module.name;

Data

You can place constant data in the output file using syntax like this:

// Strings are placed as UTF8 but with the length first
// Hello will be bound to the address this data gets
data hello = "utf8 string is very comfortable";

// We can have a sequence of data. Each int is a byte.
data bytes = 1, 2, 3, "text", 3.0;

// Moving the data into offset 32 of the memory
data moved = "Hello, world!" offset 32;

The result is that this data is copied into memory on startup.

Tables

Importing of tables is done like this:

// Table of functions of min size 1, named module.mytable in the host
import mytable : table<func>(1) = module.mytable;

// Table of externs of min size 5, max size 10, named module.myexterns in the host
import myexterns : table<extern>(5 10) = module.myexterns;

Standard library

There is a minimal standard library you get with include lib/runtime.
It contains these helpers:

// Prints a byte to stdout
printByte(i : i32) -> ()

// Print unsigned i32 as decimal
printu32(i : i32) -> ()

// Print signed i32 as decimal
printi32(i : i32) -> ()

// Prints unsigned as hexadecimal
printHex32(i : i32) -> ()

These require Wasi. There is a stub for the wasi interface in lib/wasi.wase
but that is very incomplete at the moment.

There is a start of a memory allocator in lib/malloc.wase. It has
a double-linked list of blocks, and a separate area for small allocations
(<16 bytes>). It will be probably be refactored to expose the different
management disciplines more directly.

Development

Wase is written in flow9. To develop on Wase itself, first install flow9:

https://github.com/area9innovation/flow9

We recommend to use VS Code with the flow9 extension.

To run the compiler from the command line, use

bin\wased tests/euler1.wase

This will compile the compiler, and use this to compile the euler1.wase file
to a euler1.wasm file. It is thus exactly the same as bin\wase except that
it uses the latest source of the compiler.

When you make changes to the compiler, it is recommended to make sure there are
no regressions by running all test cases and produce .wat output as well
using wasm2wat:

bin\wased test=1 wat=1

This requires that you install WABT first

https://github.com/WebAssembly/wabt

If there are regressions, the .wat and/or .wasm files in the tests
folder will have changes when you do git status.

You can compile a new wase.jar binary release used by bin\wase using:

flowc1 wase/wase.flow jar=bin/wase.jar

This requires a Java JDK like OpenJDK.

Implementation notes

For each instruction, we need to define three basic different things:

Syntax. Done in grammar.flow using Gringo.
Typing. Done in type.flow using DSL typing.
Compilation. Done in compile.flow using the Wase intermediate AST
Low-level compilation to bytecode is done in the flow9 repository in flow9/lib/formats/wasm/wasm_encode.flow

TODOS

There are a number of things, that would make Wase better:

Add the last instructions
Add SIMD instructions
Complete the Wasi interface
Encode 64-bit constants as S64, rather than U64.
Check all I32 encodings whether they are S32 or U32
Add syntax for function arguments to define whether it is mutable or not.
Right now, all function arguments are considered mutable.
Better parse errors with positions
We support a special "hole<>()" instruction, which does nothing.
The hole construct could in principle allow stack-like code:

[1, hole<>() + 2 ]

Right now, typing infer that type to (i32, i32), while it really is
i32. So the code above compiles correctly, and works at runtime, but
our type inference is not smart enough to know this.

Refactor the malloc to be separate allocation policies
Add syntax for passive data, which is not automatically copied into memory
until memory.init is called.
Add support for naming the data index for memory.init and data.drop

data id : 1, 23, ... ?
Capture the address and/or size of data segments to make use easier?
Document the implicit tables for ref.func
Add syntax for elements, which are pieces to initialize tables
Small, simple optimizations:
Detect a == 0 and make that eqz
Detect load<>(a + 4) and make that load<4>(a)
Detect const expressions and evaluate them
Detect if (a) const else const, and make that select
local.set x ;local.get x -> local.tee x
*2, *4, /2, /4 can be turned into shifts

Tags: language wasm assembly

Last modified 16 December 2024

Wase

A friendly, low-level language for Wasm.