a little JIT VM language for (probably/eventually) terminal UI apps.
Switch branches/tags
Nothing to show
Clone or download
MichaelRFairhurst Clear varsToPhi before an if statement's true blocks.
Suggested by a blog post reader -- thank you tinkerzf!

I'm not positive this is correct, since I haven't worked with SSA in a
,hile, that this is correct. However, varsToPhi is being cleared before
the false statements, so either that's broken or this line is indeed
missing.

I think its that this line is missing -- phi'ing too many vars should be
safe, and removed in optimizations/generate harmless code later. I don't
think there's a reason to phi vars changed in an outer block:

if (x) {
  x = true;
  if (y) {
    // no changes to x
  }
  // x shouldn't be phi'd here!
}
// x should only be phi'd here.

Adding this did not break any tests (not that my coverage is that
great). And I didn't add any since I don't think it would affect the
runtime, only the performance.
Latest commit 47a0ae7 Jul 14, 2018
Permalink
Failed to load latest commit information.
ast Void as a keyword (already a value/type) Jun 21, 2018
bench Omega change landing double support. Performance worse at the moment Apr 19, 2017
ssa Fix build for I guess newer C++ compiler? Jun 21, 2018
test Void as a keyword (already a value/type) Jun 21, 2018
ui_runtime Move cli ui code into own repo mostly to reduce clutter Mar 22, 2017
visitor Clear varsToPhi before an if statement's true blocks. Jul 15, 2018
.gitignore basic inlining, makefile improvements, debug disabling, and new instr… Mar 8, 2017
CONTRIBUTING.md Add contributing guideline Mar 23, 2017
LICENSE Initial commit Mar 1, 2017
README.md Void as a keyword (already a value/type) Jun 21, 2018
error_compiler.cpp Omega change landing double support. Performance worse at the moment Apr 19, 2017
error_compiler.h Omega change landing double support. Performance worse at the moment Apr 19, 2017
firstcompile.cpp Omega change landing double support. Performance worse at the moment Apr 19, 2017
firstcompile.h Delete remnants of location-resolved visitors, everything begins and … Apr 5, 2017
function_usage.h New benchmark which revealed some errors, tests and fixes for that an… Mar 26, 2017
functions.h Delete remnants of location-resolved visitors, everything begins and … Apr 5, 2017
heap.cpp Tests and stability, some expansion of std functions Mar 17, 2017
heap.h Undefined behavior cleanup in ssa typechecks, interpreter typecheck t… Mar 23, 2017
main.cpp Delete remnants of location-resolved visitors, everything begins and … Apr 5, 2017
makefile Fix build for I guess newer C++ compiler? Jun 21, 2018
optimize.cpp Omega change landing double support. Performance worse at the moment Apr 19, 2017
optimize.h Delete remnants of location-resolved visitors, everything begins and … Apr 5, 2017
parser.cpp Void as a keyword (already a value/type) Jun 21, 2018
parser.h if for / for if. Else if is no longer special Mar 24, 2017
runtime_error.cpp Use ssa as first code pass, begin double support Apr 5, 2017
runtime_error.h Use ssa as first code pass, begin double support Apr 5, 2017
scope.h Phis working maybe almost perfectly Mar 6, 2017
stack.cpp horrible first gc implementation Mar 13, 2017
stack.h horrible first gc implementation Mar 13, 2017
std.cpp Fix build for I guess newer C++ compiler? Jun 21, 2018
std.h Omega change landing double support. Performance worse at the moment Apr 19, 2017
tokenizer.cpp Void as a keyword (already a value/type) Jun 21, 2018
tokenizer.h Void as a keyword (already a value/type) Jun 21, 2018
type_info.h Delete remnants of location-resolved visitors, everything begins and … Apr 5, 2017
util.h horrible first gc implementation Mar 13, 2017
value.cpp Bugfixes/improvements to licm, added basic cse Mar 30, 2017
value.h Use ssa as first code pass, begin double support Apr 5, 2017

README.md

Vaiven

Vaiven is a little JIT programming language that I plan to eventually make into a little domain language for the terminal. Its codebase is currently under 15kloc, but it has an interpreter, a good-but-not-great compiler that profiles itself for type info that's used to create SSA-optimized code, and that only runs when the "hot" conditions are met.

Building:

git clone https://github.com/asmjit/asmjit.git
make

Tested with clang and gcc

Using

The current workflow of vaiven is to run over pipes, ie,

your_program | vvn

Frequently your_program will be a bash script. But if you have just a .vvn file to run, you can simply cat your_file.vvn | vvn.

Vaiven is currently set to work as a REPL, so you can also just run vvn and experiment there.

Interpreting

Anything top-level in the vaiven program is evaluated and (currently) printed.

  1 + 1
  Int : 2

  "foo" + "bar"
  foobar

This is all run in an interpreter, which assumes the code never runs more than once, making it the fastest execution of unknown code, once. You can define loops in the interpreter, but you should not. You should put those in a function where they will be compiled.

Ints

Standard math operations +, *, /, -, >, <, etc. work on ints, doubles, and a combination of the two. Division always produces a double, otherwise an int is only produces when both operands are ints. Binary operations are not yet added though, nor is modulo.

Strings

Strings are much like javascript. You can not yet use double or single quotes, but you can concatenate them with +, and they are immutable.

You can also call toString on any value to, for instance, turn an int into a string.

Functions

Functions can be defined with or without parameters. Parameters are not typed. Like all control constructs in vaiven, it ends with "end".

Calling a function (such as print in this example) looks just like c syntax.

  fn foo
    print("hello!")
  end

Or, with a parameter,

  fn greet of name is
    print("hello " + name + "!")
  end

With multiple parameters, both definition and invocation can be comma separated:

fn twoArgs of x, y is
  ...
end

twoArgs(1, 2)

or newline separated:

fn twoArgs of
  x
  y
is
  ...
end

twoArgs(
  1
  2
)

(or both).

Return

Functions will automatically return the final statement, so you don't need to use a keyword to return a value in a function

  fn square of x is
    x * x
  end

This will not work yet, however, with if and for statements. So for early return you can use the keyword ret.

  fn square of x is
    ret x * x
  end

In this case, these do the same thing.

Currently, you cannot simply ret, you must ret a value.

Method call syntax

Methods don't actually exist in vaiven, which is common to dynamic languages. However, most dynamic languages maintain the feel of methods by using anonymous function members, which is an approach not taken by vaiven.

Instead vaiven emulates methods via a special syntax. Any function can be called on a value as if it were a method, and that value becomes the first parameter.

toString(1) // "1"
1.toString() // equivalent
cmp(1, 2) // false
1.cmp(2) // false

Note that this does not allow for polymorphism, as the function is a global. You can use tagged objects to emulate polymorphic behavior as needed, which keeps the language small.

Variables

Variables are scoped to their control structures and/or to functions, and/or globally. Global variables are only accessable from outside functions, and variables cannot share a name, even in subscopes (shadowing not allowed).

  var x = 4

Currently all variables must be initialized when declared.

If statements

If statements do not require parenthesis or brackets in vaiven.

  if true
    print("true")
  end

You can also use else like so:

  if true
    print("true")
  else
    print("false")
  end

Else if

You can also use else if as you would expect, and it does not have a special keyword.

  if hasShirt
    print("has shirt")
  else if hasShoes
    print("has shoes")
  else
    print("no service")
  end

be careful that this is not the same:

  if hasShirt
    print("has shirt")
  else
  if hasShoes
    print("has shoes")
  else
    print("no service")
  end

This actually requires one more end to finish the statement. While vaiven makes semicolons optional, it does not have whitespace significance. See the section on semicolons for why this example is treated as if hasShoes in within rather than after the else statement.

One line control structures

Control structures can be put on a single line, but you must use the do keyword and end. (The requirement for end may be relaxed in the future)

  if true do print("true") end
  if true do print("true") else do print("false") end

  if false
    ...
  else if foo do print("foo") end
  else do print("false") end

having to use do after else may also be relaxed in the future.

You can see the section on semicolons for why.

For

There is only one type of loop so far in vaiven, and that's a for condition loop. It simply runs the following block as long as the condition is true.

  for x < 10
    x = x + 1
  end

Like if and else if, this can be put on one line using do and end.

  for x < 10 do x = x + 1 end

Lists

Lists work just like javascript. You can create a list with literal-based syntax, assign to indexes and retrieve from indexes.

  var list = [1, 2, 3]
  list[0] // 1
  list[0] = 2
  list[0] // 2

You can also ask for indexes that don't exist in a list, and you'll get back void.

  list[5] // void

And you can assign anywhere in a list, regardless of its current length. If needed, the list will resize and fill in the empty spaces with void.

  var list = [];
  list[2] = 2;
  list // [void, void, 2]

You will get a type error if you attempt to use a non-integer as an index.

Objects

Object literals are not yet supported, but you can make an object via the function object. It can then be used like an associative array, or using . member syntax.

  var o = object()
  o.foo = "bar"
  o.foo // "bar"
  o["foo"] // "bar"

  o["baz"] // void
  o.baz // void

  o[myVar] = "dynamic property"
  o[myVar] // "dynamic property"

Soon object literals will be added.

Type rules

Vaiven has some loosely defined type rules, and almost no type-coersion.

For the plus operator, when one side's static type can be inferred, you will get an error requesting that type on the other side. When neither is inferred, you will get a more general error. Only strings and ints and doubles can be used with +, and if either operand is a string, both must be.

[] + [] // Type Error: expected int or double or string
1 + object() // Type Error: expected int or double
1.0 + object() // Type Error: expected double
[] + 1 // Type Error: expected int or double
"" + true // Type Error: expected string
[] + false // Type Error: expected string
"" + 1 // Type Error: expected strint or double
1 + void // Type Error: expected int or double

Other math operators only support ints and doubles

"" * [] // Type Error: expected int or double
"" - true // Type Error: expected int or double
false / "" // Type Error: expected int or double
[] > object() // Type Error: expected int or double

Like +, [] is polymorphic but requires agreement between both sides. If the static type of one side can be inferred, you will get an error expecting the other operand to match. Otherwise, you will get a generic error. Lists can be subscripted with ints and objects can be subscripted with strings. Doubles can not be used to index arrays, though.

1[object()] // Type Error: expected list or object
true[false] // Type Error: expected list or object
1[""] // Type Error: expected object
""[1] // Type Error: expected list
1[""] // Type Error: expected object
object()[1] // Type Error: expected string
[1,2,3][00.4] // Type Error: expected int

Comparisons can safely cross types, and behave like === in js. However, ints and doubles are also comparable.

1 == true // false
0 == "0" // false
true == "true" // false
[] == "" // false
object() == "{}" // false

1 == 1.0 // true

Conditions (like in if, for) and ! must be bools.

!true // OK
if true // OK
if 1 // Type Error: expected bool
  !1 // Type Error: expected bool
end
if "" // Type Error: expected bool
  !"" // Type Error: expected bool
end
if [] // Type Error: expected bool
  ![] // Type Error: expected bool
end
if void // Type Error: expected bool
  !void // Type Error: expected bool
end
if object() // Type Error: expected bool
  !object() // Type Error: expected bool
end

Semicolons

Vaiven has been made to have optional semicolons for pretty syntax, with predictable effects and minimal whitespace significance. It has also been designed so that newlines are special, not else if.

You can write vaiven code with no special newline handling, by always using semicolons and do.

  if true do
    print("stuff");
  end

  for false do
    print("doesn't execute");
  end

This is how the language looked originally.

Now, in places where do or semicolon are expected, a newline will suffice. Therefore, the above code is exactly equivalent to

  if true
    print("stuff")
  end

  for false
    print("doesn't execute")
  end

and this is why the oneline versions of if/for/else look like the original grammar; do is the semicolon of control structors, and so both are optional. But one-liners require them.

To make if true do something; end a little cleaner, we also made end suffice in place of a semicolon, but leave it for the enclosing statement to consume.

And since commas are the semicolons of parameter lists, so we made those optional too.

  foo(
    1 + 2
    3 + 4
  )

Newlines can also fill in for is, and that's why the shorthand function syntax works:

fn foo
  print("foo")
end

Oneline functions with no arguments require is.

fn foo is print("foo") end

and once you start listing arguments, we need is to know when those arguments end.

Else if isn't special

I am of the opinion that else if should not be special, and so in vaiven, it is not.

You may have noticed, if do is the semicolon of statements, and else if is not special, then how does else if work without a do or a newline?

Well, else can be followed by any control structure. This allows the else to not require an end, which gives the appearance of creating a hybrid conrol structure. It is not unique to else if, as you can also do else for.

  if ...
    ...
  else for x
    ...
    x = x + 1
  end

and the same logic will applies to if and for and ret as well:

  if a ret 4

  if x for y
    ...
    y = y + 1
  end

  for z if q
    ...
  end

  for z if q ret a

So else if is not special. All control structures can be blended as desired!

Performance

Feel free to run bench/bench.sh for numbers specific to your system.

The one cross language benchmark we have, recursive fib, is barely faster than v8 though a bit slower than dart. But this is not to say that vaiven is twice as fast as v8. Its just to say that the right code is exactly as performant as you would expect from a high-quality JIT -- see optimization tips for details about what we don't do well, as its still a pretty long list.

In general, compiled code is up to 100x faster than interpreted code. And usually the hot code is about 2x faster than that, though it can be even higher.

For instance, the recursion based test currently does a better job reducing polymorphic + overhead, and finishes in 0.000 seconds (too low to measure) vs about 20 second for the equivalent interpreted code on my system.

Optimizations

Vaiven currently performs the following optimizations on your code:

  • Instruction combining (x * 1 == x, x * 4 / 2 == x * 2, ...)
  • Constant propagation (3 * 4 == 12...)
  • Dead code elimination (including branches which can't be taken)
  • Type profiling guided optimizations via guarded specializations
  • Typecheck/boxing/unboxing elimination
  • Function inlining
  • Recursive function inlining (essentially akin to loop unrolling)
  • Tail call recursion (including jumping past guards)
  • Basic jump threading (not range-based yet, just jumps to jumps ignoring conditions)
  • Basic loop invariant code motion (can get mixed up by phis)
  • Basic common subexpression elimination

and generates otherwise clean asm

  • boolean values are massaged into jmpcc calls
  • argument counts are known statically and compiled with the right behavior hardcoded
  • all tagged values are 64 bit using an inverse punboxing system
  • some bounds checking elimination for instance in list initialization

And vaiven uses asmjit's register allocation with basic coalescing added on top, which is a very good though not great (yet!) register allocator.

Under the hood: Optimization tips

Firstly, anything that will be executed more than once should be put in a function. Anything executed only once should not. The interpreter is about 2x faster than compilation and execution.

However, for anything that runs multiple times (ie has loops, recursion, etc), the compilation is much much much faster than the interpreter. The interpreter is deliberately using the slowest interpreting pattern out there, a visitor. This is because it is faster than compiling to bytecode before running, but only for one run.

After a certain count of times, function will self optimize using argument type data. This is tracked on previous executions in a bitmap of, has x been an int, string, etc.

Currently, when an argument has been multiple types, no optimization is done against that parameter. It is therefore critical that in the warm-up stage, you call functions monomorphically where that is how they will run.

On the other hand, after optimization occurs, code violating the profiler's data will run even slower. Its not as slow as full blown dynamic deoptimization, its simply a guard check in the function header, but its still slower than the original code. So if you have a function which may accept strings or ints, and only feed it ints during the warm up phase, it will be slower when it gets strings later.

Vaiven also will not yet throw away optimized code when it seems to be missed frequently, so be careful that you don't teach it bad tricks.

At the very least, even if a function is completely polymorphic and has no optimizability, its still best to optimize it as that. Some optimizations (such as inlining) are still possible, and the data collection will be stripped out as well.

Since vaiven doesn't yet have true dynamic deoptimization, only guarded optimizations, be wary of polymorphic code in the middle of your function. Things like fetching from an object or an array will not be profiled and optimized, since we won't have any fall back strategy for when we don't get what we expected.

For the same reason, slow loops cannot optimize themselves mid-loop like v8 and other JITs. Its best to warm up a loop ahead of time, perhaps with an argument for how many iterations should run, before running a long slow loop.

Inlining is also slightly controllable, in a few ways. If you see something is being inlined which shouldn't, you can add a bunch of dead code to trick the inlining into fear of code explosion. On the flip side, you can add arguments which are unused, since those arguments represent a chance at optimization. With the right compilation flags, you can track the compilers inlining decisions to make use of these tricks.

To Do

Core language:

  • object literals
  • unary -/+
  • binary operators?
  • &&, || with short circuiting
  • stack traces
  • imports
  • void return
  • for over lists (and objects?)
  • larger std library
  • falsy values (at the very least void and empty string?)
  • escapes inside strings
  • check type of object at runtime
  • perhaps more type coercions, like toString() before property access or parseInt before array access

Performance:

  • Bounds check elimination
  • Inlining native functions
  • track "constant" lists and objects
  • Precise GC and/or better allocator for heap
  • Custom string/list/object implementation for inlinability
  • fuzz test for security/stability
  • dynamic deoptimization/on-stack optimization
  • profile jumps to reduce branching
  • profile jumps to make better inlining decisions
  • profile hash lookups to make for faster property access
  • better benchmarks
  • align jumps

Domain language

Once vaiven runs as a TUI framework, it needs these things. Currently, vaiven is just a little VM that doesn't need any domain-related aspects.

  • Choose a strategy: fork the core language? Plugin model? Neither?
  • fully spec language
  • ...

Legal

This is not an official google product. Please see LICENSE for more.