In previous implementation when converting a number to PropertyKey, we
always converted it to string and only when the conversion to
PropertyKey::Index was not possible we returned that string.
* Refactor vm calling convention to allow locals
* Move async generator object to the stack
* Apply review and some clean-up
* Add todo for storing fp in CallFrame
* Rename ASYNC_GENERATOR_OBJECT_LOCAL_INDEX to ASYNC_GENERATOR_OBJECT_REGISTER_INDEX
* `Context` independent `CodeBlock`s
This allows us to share `CodeBlock`s between `Context`, which makes
functions sharable between `Context`s.
* Apply review
* Move `FunctionKind` flag to CodeBlock
* Add `IS_ARROW` flag to codeblock
* Remove GetAsyncArrowFunction and GetFunctionAsync opcodes
* Remove GetGeneratorAsync opcode
* Remove direct conversion from `&str` to `JsValue`/`PropertyKey`.
* Allow unused static strings
* Introduce DHAT to benchmark data usage
* Create new release profile
* Fix docs
* Refactor `Context::run()` method
- Remove async generator close from run
- Remove promise capability from run
- Remove promise capability check from call_internal
- Remove generator creation from call_internal
- Move promise capability creation to separate opcode
* Inline ReThrow opcode in run
This PR implements `Hidden Classes`, I named them as `Shapes` (like Spidermonkey does), calling them maps like v8 seems confusing because there already is a JS builtin, likewise with `Hidden classes` since there are already classes in JS.
There are two types of shapes: `shared` shapes that create the transition tree, and are shared between objects, this is mainly intended for user defined objects this makes more sense because shapes can create transitions trees, doing that for the builtins seems wasteful (unless users wanted to creating an object with the same property names and the same property attributes in the same order... which seems unlikely). That's why I added `unique` shapes, only one object has it. This is similar to previous solution, but this architecture enables us to use inline caching.
There will probably be a performance hit until we implement inline caching.
There still a lot of work that needs to be done, on this:
- [x] Move Property Attributes to shape
- [x] Move Prototype to shape
- [x] ~~Move extensible flag to shape~~, On further evaluation this doesn't give any benefit (at least right now), since it isn't used by inline caching also adding one more transition.
- [x] Implement delete for unique shapes.
- [x] If the chain is too long we should probably convert it into a `unique` shape
- [x] Figure out threshold ~~(maybe more that 256 properties ?)~~ curently set to an arbitrary number (`1024`)
- [x] Implement shared property table between shared shapes
- [x] Add code Document
- [x] Varying size storage for properties (get+set = 2, data = 1)
- [x] Add shapes to more object:
- [x] ordinary object
- [x] Arrays
- [x] Functions
- [x] Other builtins
- [x] Add `shapes.md` doc explaining shapes in depth with mermaid diagrams :)
- [x] Add `$boa.shape` module
- [x] `$boa.shape.id(o)`
- [x] `$boa.shape.type(o)`
- [x] `$boa.shape.same(o1, o2)`
- [x] add doc to `boa_object.md`
This Pull Request changes the following:
- Creates a new `PromiseCapability` after every async function call instead of sharing the same capability for all calls of the same async function.
This Pull Request fixes some additional Annex B tests.
It changes the following:
- Fixes bugs related to parsing HTML closing comments (`-->`).
- Implements `RegExp::compile` behind the `annex-b` feature.
- Ignores the `legacy-regexp` feature flag, since it's still stage 3.
This should hopefully fix more async/futures issues related to resuming execution in the future, since we can leverage generator logic to handle this for us.
It changes the following:
- Refactors `GeneratorContext` to handle context preparation.
- Reuses the functionality of `GeneratorContext` in `Await`.
- Removes `EarlyReturnType` in favour of a single `r#await` bool flag in `CallFrame`.
Small steps towards ES5 conformance.
This PR changes the following:
- Implements HTML comments parsing (`<!--`, `-->`).
- Gates the functionality behind a new `annex-b` feature for `boa_parser`.
- Renames `strict_mode` to `strict` to be consistent with `Parser::set_strict`.
This Pull Request fixes#2317 and #1835, finally giving our engine proper realms 🥳.
It changes the following:
- Extracts the compile environment stack from `Realm` and into `Vm`.
- Adjusts the bytecompiler to accommodate this change.
- Adjusts `call/construct_internal` to accommodate this change. This also coincidentally fixed#2317, which I'm pretty happy about.
- Adjusts several APIs (`NativeJob`, `Realm`) and builtins (`eval`, initializers) to accommodate this change.
- Adjusts `JsNativeError`s to hold a reference to the Realm from which they were created. This only affects errors created within calls to function objects. Native calls don't need to set the realm because it's inherited by the next outer active function object. TLDR: `JsError` API stays the same, we just set the origin Realm of errors in `JsObject::call/construct_internal`.
This Pull Request fixes test [`assert-throws-same-realm.js`](eb44f67274/test/harness/assert-throws-same-realm.js).
It changes the following:
- Handles global variables through the global object, instead of the `context`.
- Adds an `active_function` field to the vm, which is used as the `NewTarget` when certain builtins aren't called with `new`.
- Adds a `realm_intrinsics` field to `Function`.
The fields like `[[PrivateElements]]` are hardly used but they occupy `24` bytes (on 64-bit arch.) the `ThinVec` type stores the `len` and `cap` right before the elements (like our `JsString` implementation) and only a pointer is kept (if not used it does not allocate!), was going to use this in #2723 , since it uses a `Vec` as a dense storage, but the PR is already too big.
It changes the following:
- Shrink object from `328` to `288` bytes (40 bytes reduction)
- Add the `thin_vec` lightweight crate (single file)
This Pull Request fixes#2673.
It changes the following:
- Add early errors to dynamic function constructors.
- Add tests that check for syntax errors when `super` is passed to the function constructor.
This Pull Request closes#1907.
It changes the following:
- Implement several early errors relating to labels, `break` and `continue` in the parser.
- Implement an early error for invalid cover grammar of object literals in the parser.
- Remove all remaining syntax errors from the bytecompiler.
When a rust string literal is given for a property key, boa checks if it can be parsed as an index and converts the string into a utf16 slice. This PR rewrites each hard-coded property key as a utf16 slice so that we can bypass those conversions at runtime.
This improves QuickJS benchmark score 5% on average.
Richards: 35.4 -> 37.0
DeltaBlue: 35.0 -> 38.1
Crypto: 57.6 -> 59.6
RayTrace: 137 -> 146
EarleyBoyer: 131 -> 138
Splay: 98.3 -> 104
NavierStokes: 10.2 -> 10.2
Small (ish?) step towards having proper realm records
This PR changes the following:
- Moves `Intrinsics` to `Realm`.
- Cleans up the initialization logic of our intrinsics to not depend on `Context`, unblocking things like #2314.
- Adds hooks to initialize the global object and the global this per the corresponding [`InitializeHostDefinedRealm ( )`](https://tc39.es/ecma262/#sec-initializehostdefinedrealm) hook. Though, this is currently broken because the vm uses `GlobalPropertyMap` instead of the `JsObject` API to initialize global properties.
Slightly related to #2411 since we need an API to pass module files, but more useful for #1760, #1313 and other error reporting issues.
It changes the following:
- Introduces a new `Source` API to store the path of a provided file or `None` if the source is a plain string.
- Improves the display of `boa_tester` to show the path of the tests being run. This also enables hyperlinks to directly jump to the tested file from the VS terminal.
- Adjusts the repo to this change.
Hopefully, this will improve our error display in the future.
The section about `Symbol` on the [specification](https://tc39.es/ecma262/#sec-ecmascript-language-types-symbol-type) says:
> The Symbol type is the set of all non-String values that may be used as the key of an Object property ([6.1.7](https://tc39.es/ecma262/#sec-object-type)).
Each possible Symbol value is unique and immutable.
Our previous implementation of `JsSymbol` used `Rc` and a thread local `Cell<usize>`. However, this meant that two different symbols in two different threads could share the same hash, making symbols not unique.
Also, the [GlobalSymbolRegistry](https://tc39.es/ecma262/#table-globalsymbolregistry-record-fields) is meant to be shared by all realms, including realms that are not in the same thread as the main one; this forces us to replace our current thread local global symbol registry with a thread-safe one that uses `DashMap` for concurrent access. However, the global symbol registry uses `JsString`s as keys and values, which forces us to either use `Vec<u16>` instead (wasteful and needs to allocate to convert to `JsString` on each access) or make `JsString` thread-safe with an atomic counter. For this reason, I implemented the second option.
This PR changes the following:
- Makes `JsSymbol` thread-safe by using Arc instead of Rc, and making `SYMBOL_HASH_COUNT` an `AtomicU64`.
- ~~Makes `JsString` thread-safe by using `AtomicUsize` instead of `Cell<usize>` for its ref count.~~ EDIT: Talked with @jasonwilliams and we decided to use `Box<[u16]>` for the global registry instead, because this won't penalize common usage of `JsString`, which is used a LOT more than `JsSymbol`.
- Makes the `GLOBAL_SYMBOL_REGISTRY` truly global, using `DashMap` as our global map that is shared by all threads.
- Replaces some thread locals with thread-safe alternatives, such as static arrays and static indices.
- Various improvements to all related code for this.
This Pull Request fixes various bugs related to classes.
The biggest changes are:
- Changed private names to be unique across multiple classes.
- Changed private name resolution to work via a visitor after a class is parsed. The way class early errors are defined makes it impossible to perform private name resolution while parsing.
- Added function names to class methods.
- Added class name binding to method function environments.
- Separated opcodes for `static` and non-`static` class method definitions to make the above operations possible.
There are still some bugs and further issues with classes but this is already a lot.
This PR is a complete redesign of our current native functions and closures API.
I was a bit dissatisfied with our previous design (even though I created it 😆), because it had a lot of superfluous traits, a forced usage of `Gc<GcCell<T>>` and an overly restrictive `NativeObject` bound. This redesign, on the other hand, simplifies a lot our public API, with a simple `NativeCallable` struct that has several constructors for each type of required native function.
This new design doesn't require wrapping every capture type with `Gc<GcCell<T>>`, relaxes the trait requirement to `Trace + 'static` for captures, can be reused in both `JsObject` functions and (soonish) host defined functions, and is (in my opinion) a bit cleaner than the previous iteration. It also offers an `unsafe` API as an escape hatch for users that want to pass non-Copy closures which don't capture traceable types.
Would ask for bikeshedding about the names though, because I don't know if `NativeCallable` is the most precise name for this. Same about the constructor names; I added the `from` prefix to all of them because it's the "standard" practice, but seeing the API doesn't have any other method aside from `call`, it may be better to just remove the prefix altogether.
Let me know what you think :)