For various reasons, we store the dispatch list for an object on a separate property of an `EventDispatcher` rather than dictating that all children of `EventDispatcher` use a specific object type. This is because `EventDispatcher` is a very general class with lots of object representations it needs to cover. So instead, we introduce a new object representation for a *property* and store it in a Ruffle private namespace that is as isolated from user code as alternate object representations are.
This also changes the underlying `DisplayObjectContainer` method to accept any type of range. Turns out enum trait objects aren't actually trait objects and don't need to worry about object safety!
`ChildContainer` is responsible for maintaining child lists for all display objects that can hold children. Currently, this is just `Button` and `MovieClip` since those are the only objects in AVM1 that can have children, but this will be extended to other objects in future commits.
The number of lists managed has also increased from two to three. The execution list is unchanged save for it's migration into the `ChildContainer` struct. The render list has been split in two to support AS3. Specifically, the render list is now a `Vec`. Render children are still rendered in order but they are now referenced by AS3 `id`s rather than depths. The old `BTreeMap` that served as a render list is now the depth list and serves to maintain compatibility with SWF tags and AVM1 code that refers to things on the timeline by depth.
This entirely abolishes the "global scope object" in AVM2. I even had to redefine several global object functions to work with the bottom of the scope stack, which seems to be where ASC likes to stick the script scope.
This is the same way that AVM1 actions run and it appears that frame scripts work exactly the same way. It fixes all outstanding bugs with movie clip navigation in AVM2 and allows me to remove a lot of weird workarounds I was writing for the old, incorrect behavior.
I'm also removing the "last run script frame" rule as `run_frame_internal` already had rules to prevent stopped clips from rerunning actions.
I originally added this with the anticipation that `impl` return syntax only allowed one trait plus OIBITs. This was prior experience in Rust but apparantly the compiler accepts this just fine, so I suppose my defensive coding practice was a bad/outdated idea.
This appears to work almost like it's own TObject method; you can run `Object.prototype.toLocaleString` on all sorts of things and it has separate behavior to what the class method for it might be. I have attempted to match Flash Player as best as I can.
The array being iterated is explicitly handed to all callbacks, and it is legal for the callback to mutate the array. Hence, we can't actually hold a `Ref` to the array storage when we call user code. Instead, we implement a custom `Iterator` which iterates over the object like user code would.
This actually can't be an `Iterator` impl due to limitations of the underlying trait. Hence, we have to `while let` instead of `for`.
This is for the sake of methods that want to change behavior based on if they're working with a number or some other kind of value. It should not be used otherwise.
This code also ensures that the prototypes of each system object are created in the appropriate `TObject` impl. This ensures that, for example, `new Array` hands you back an actual array.
* Implement `add`, with tests.
* Implement `add_i`.
There's no test, because for whatever reason, I can't figure out how to emit this from Animate CC 2020.
* avm2: Implement `bitand` with tests.
* Implement `bitnot` with tests.
* Implement `bitor` with tests.
* avm2: Implement `bitxor`
* avm2: Implement `declocal`, `declocal_i`, `decrement`, and `decrement_i`.
* tests: `swf_approx` tests should be allowed to print NaNs.
* avm2: Implement `divide`.
* avm2: Implement `inclocal`, `inclocal_i`, `increment`, and `increment_i`.
* avm2: Implement `lshift`.
* Implement `modulo`.
* avm2: Implement `multiply` and `multiply_i` (no tests for the latter)
* avm2: Implement `negate` and `negate_i` (no tests for the latter)
* avm2: Implement `rshift`
* avm2: Implement `subtract` and `subtract_i` (the latter without tests)
* avm2: Implement `urshift`.
This has some particularly annoying consequences for initialization order: notably, we can't actually create any ES4 classes using the standard machinery until after the three objects I just mentioned get created. Ergo, we have to create them through lower-level means, handing prototypes around, and then initialize AVM2's system prototypes list for it.
When we start adding more system prototypes, we'll also have to fill the extras with blank objects and then slot them in as we create them.
This was surprisingly tricky - due to the need to look up superclasses, class trait instantiation requires an active `Activation` and `UpdateContext`. We can't get those during VM instance creation, since the player needs the VM first before it can give it a context to work with. Ergo, we have to tear the global scope initialization in two. At the first possible moment, the player calls a new `load_player_globals` method that initializes all class traits in global scope.
I have no idea why this is necessary - I was in a context where what *should* have been a `NativeMethod<'gc>` was instead being interpreted as some different function type with all the same lifetimes, but with an extra `'gc` lifetime as well. Funneling this through a non-trait method bypasses whatever is going on with the trait solver, and then at that point the trait solver knows what to do. Consider this an extra level of conversion.
ECMA-262 3rd ed. doesn't mention anything about different number types, so the standard as-if rule applies. If we are going to distinguish number types, we have to treat them as if they were the same type, promoting to `f64` as necessary to facilitate the conversion. I took a cursory look at an ECMA-262 4th ed. draft and it appears to do the same, although it calls everything `GeneralNumber` and has some really confusing psuedo-Pascal syntax for some reason.
I am extremely glad AVM2 does not provide access to 64-bit integer types (for now, at least).
Namespaces as values adds a bunch of extra special cases to the coercion and equality rules that don't really belong there. Namespace itself just returns it's URI as a string, so we can just make `NamespaceObject` do that and then treat it the same way we treat boxed primitives.
These include:
* Name resolution in `newobject`
* All runtime & late-bound multinames
* `Object.hasOwnProperty`
* `Object.propertyIsEnumerable`
* `Object.setPropertyIsEnumerable`
So, I overlooked this reading the 1.45 documentation, but the first thing they did is completely change f64 conversions. Apparantly, what I was doing (and what JavaScript spec dictates) is actually considered UB in LLVM, and my ability to actually write a concise wrapping u32 conversion is actually a soundness hole in Rust. Ergo, I'm now emulating the wrapping and sign calculation, which makes this both passing it's tests again and free of soundness holes and UB.
I don't know why I'm doing this - tests are failing in CI but not locally, and I can only assume that the most obvious conversion is broken in some way on whatever other architecture GitHub Actions uses. This will explicitly mask the integer result as a u64, and then convert it down to u32. A not-broken compiler should treat this code identically.
AVM2 is based on ES4, which as far as I'm aware, does not distinguish between "primitive values" and "objects". Thus, it is expedient to interpret any statement requiring something to be an Object to mean "not null or undefined".
Since we internally represent register values with primitive types, it is important that the VM always coerces to object before doing any other sort of type checking. Hence, something like `as_object` is unhelpful as it accidentally enforces a primitive/object distinction that ES4 attempted to remove.
The test is also far more in-depth than the `if_eq`/`if_ne` tests, which use the same set of vectors as the strict-equality tests from a while ago. Interestingly, this test passed on first run
Implementation is limited to generating exceptions on `null` or `undefined`. I'm not sure if primitive values don't exist in AVM2 or if this is supposed to box them like ES3, so I have decided to handle neither at this time.
This code is slightly over/under-precise compared to AVM2. This is because we handle precision limiting in binary floats rather than as part of the float printing process. Flash Player may also be rounding differently than us. However, I'm pretty sure ECMA-262 allows us to slightly differ here.
The ECMA-262 documentation is awfully overwrought for something that boils down to "chop off the non-whole part, wrap to 32 bits, then reinterpret as signed". Bitwise operations are *hell* to describe mathematically, and such descriptions are even harder to understand.
For whatever reason, `pushbyte` appears to be processed as a *signed* byte, despite the clear wording of "*byte_value* is an unsigned byte" in avm2overview.pdf. I guess it's supposed to be manually converted and promoted in this manner.
Functions that need to assert Boolness without coercion should either:
1. Ensure their function declaration requires a Boolean. (We don't enforce type errors on ES4 typehints yet, but we should.)
2. Check the value type themselves and raise their own errors if necessary.
As it stands the only users of `as_bool` either needed to check the type themselves or use `coerce_to_bool`. Notably, `setPropertyIsEnumerable` doesn't appear to coerce *or* throw an error: it instead fails silently if you hand it a non-`Boolean` value.
Holding a `Ref` on a garbage-collected object inherently extends any borrow locks on that object. Since ABC files are references already, taking a `Ref` to them only helps to skip the refcount update. This is less useful than expected: in most situations, using `abc_ref` causes double-borrow panics. The few methods that can use it are going to be fragile in the face of future refactors, so I'm nipping the problem in the bud now.
For good measure, most of the other methods in `value` for retrieving pool primitives now also use `TranslationUnit` instead of `AbcFile`. This is the result of a handful of cascading changes throughout the project, and itself caused a few more.
Interface methods are specifically not allowed to be called: as a result, they don't get a method body. Existing code assumed a 1:1 relationship between methods and bodies, which causes spurious errors.
This is inspired by Dinnerbone's similar PR on the AVM1 side, where the Action half of that VM's `Executable` was reduced from 128 bytes to 16 by shoving it in a `Gc`. This won't be as dramatic but should still save some memory.
In fact, it should save a *lot* of memory in bytecode execution, where thanks to the previous commit's rebase, we now need to clone the current method once *for each instruction executed*. That is terrible, but should stop now.
This also results in a far reduced role for `ReturnValue`, since I also took the liberty of removing most of it's use. Furthermore, I also made it apply equally to native and AVM2 code, which ensures all native implementations of methods don't double-borrow.
In AVM1, `ReturnValue` was actually removed entirely, because it's not needed. I attempted to do the same, but the fact that we're currently embedding `ScriptObjectData` in native objects means that we need it for virtual properties. Otherwise, virtual property implementations will see locked objects, which is bad.
While some code that references pool multinames has zero as a valid index, we cannot validate exactly what the zero index is for a given index. Hence, callers instantiating multinames must check for zero and substitute the correct zero-value interpretation for their given type. If zero is an invalid value, it should ideally throw a different error than what's provided here.
This commit breaks the build: we still need to tell `Avm2` how to turn ABC traits into our own internal `Trait<'gc>`, `Class<'gc>`, and `Method<'gc>` types. We also need something to track which traits have already been instantiated, because `callstatic` would otherwise reinstantiate the trait in a different scope. (In fact, I think it *does* do exactly that right now...)
The intention is to completely replace all usage of `Avm2XYZEntry` with `Class`, `Trait`, and `Method`. This will allow runtime-provided global class traits to coexist with those provided by user code.
Inspired by Dinnerbone's PR doing the exact same thing to AVM1.
On AVM2 we have a bit of a subtle issue: the base implementation of `set_property_local` and `init_property_local` *must* return `ReturnValue`s to avoid double-borrows. Each implementation of `TObject` must resolve them before returning.
This function has vague documentation about enabling locale-specific formatting in subclasses. As far as I can tell, none of the objects I implemented so far do anything different than `toString`, so I just have it use the same `TObject` property I set up for `toString`.
We still reuse the `FunctionObject` machinery internally. If necessary, we may want to split this into a separate `ClassObject` if some internal `TObject` method needs replacing for classes.
In practice not many movies will care about this, because the `AS3` namespace is open by default. You could opt-out of that, and I suppose that was there for using existing ES3 code in AS3 projects. ES4 would have had a similar ES4 namespace, which "JavaScript 2.0" code would need to opt into. Of course, ES4/JS2 never happened, so we just have this weird historical quirk here.
The reason for this is that, in AVM2, `toString` and `valueOf` are not defined on the classes or prototypes of `Function` or `Class`. Instead, they use the `Object.prototype` versions of those functions. Ergo, string and primitive coercion are inherent object methods (the ones that get `[[DoubleSquareBrackets]]` in the ECMA standards). In Ruffle, our equivalent to `[[DoubleSquareBrackets]]` methods are methods on the `TObject` trait, so we're adding them there.
This mechanism will make implementing boxed value types (ala AVM1's `BoxedObject`) easier, too.
We also add some reasonable defaults for `ScriptObject` and `FunctionObject` which will appear on objects, functions, and classes.
Private names now return `false`, and we run any names through trait lookups. This also means any namespace resolution can fail now, in case we need to throw a `VerifyError`.
I have... significant reservations with the way object enumeration happens in AVM2. For comparison, AVM1 enumeration works like this: You enumerate the entire object at once, producing a list of property names, which are then pushed onto the stack after a sentinel value. This is a properly abstract way to handle property enumeration.
In AVM2, they completely replaced this with index-based enumeration. What this means is that you hand the object an index and it gives you back a name or value. There's also an instruction that will give you the next index in the object.
The only advantage I can think of is that it results in less stack manipulation if you want to bail out of iteration early. You just jump out of your loop and kill the registers you don't care about. The disadvantage is that it locks the object representation down pretty hard. They also screwed up the definition of `hasnext`, and thus the VM is stuck enumerating properties from 1. This is because `hasnext` and `hasnext2` increment the index value before checking the object. Code generated by Animate 2020 (which I suspect to be the final version of that software that generates AVM2 code) initializes the index at hero, and then does `hasnext2`, hence we have to start from one.
I actually cheated a little and added a separate `Vec` for storing enumerant names. I strongly suspect that Adobe's implementation has objects be inherently slot-oriented, and named properties are just hashmap lookups to slots. This would allow enumerating the slots to get names out of the object.
Our already odd `super` handling throws up another subtlety regarding bound recievers. Since we have to construct an instance of a parent class in order to get traits on it, we also have to make sure that we initialize traits with the correct reciever. I'll demonstrate here:
```let mut base = base_proto.construct(avm, context, &[])?;
let name = base.resolve_multiname(&multiname).unwrap();
let value = base.get_property(object, &name, avm, context)?.resolve(avm, context)?```
In this case, if `name` is the name of a method, getter, or setter trait, then `get_property` will instantiate that trait on `base` but bound to `reciever`. This is correct behavior for this case, but more generally, trait instantiation is permenant and therefore there's potential for confusing shenanigans if you `get_property` with the wrong reciever.
To be very clear, `reciever` should *always* be the same object that is getting `get_property` et. all called on it. In the event that you need to instantiate traits with a different `reciever`, you should construct a one-off object and retrieve prototypes from that.
Previously, we were treating ES4 classes like syntactic sugar over a prototype chain (like ES6 classes); e.g. each declared trait was set in the given prototype and then property look-ups happened as normal.
This already caused problems with virtual properties, which could be partially-defined in subclasses and required careful checks to make sure we stopped checking the prototype chain on the *correct* half of the property.
However, this is a hint of a larger problem, which is that ES4 classes don't actually define anything on the prototype chain. Instead, the instance itself constructs class properties and methods on itself. This allows things like methods automatically binding `this`, which isn't included in this commit but will be implemented really soon.
The prototype chain still exists even on pure ES4 classes, due to the need for backwards compatibility with ES3 code. Object, for example, still defines it's methods as prototype methods and thus there needs to be a prototype chain to reach them. I actually could have gotten away with using the prototype chain if AS3 *hadn't* retained this "legacy" detail of ES3 allowing this class/prototype distinction to leak out into upcoming tests.
We still actually use the prototype chain for one other thing: trait resolution. When we look for a trait to install onto an object, we pull traits from the prototype chain using a special set of `TObject` methods. This happens in opposite order from normal prototype lookups so that subclassing and verification can proceed correctly.
`super` somehow became even harder to implement: we now actually construct the parent class so we can get traits from it, which is going to complicate method binding as mentioned above.
The previous system for handling setters would execute the setter and then return a value to indicate whether or not the caller needed to resolve a stack frame. However, no caller of `Property.set` actually did this. Ergo, errors in setters and getters would not resolve up the stack at the correct time.
This problem also exists in AVM1 but is far less noticable as AVM1 only has two very uncommon runtime errors and very few movies use `throw`.
Normally, `set_property` only affects the object it was called on, which makes sense: otherwise, we couldn't override values that originate from a class prototype without accidentally monkey-patching the prototype. However, virtual setters only exist in prototypes and need to be accessible from child objects.
The solution to this is to have a specific method to check if a virtual setter exists. Virtual setters are then resolved through the prototype chain. If no virtual setter exists, then the reciever object is handed the value.
Note that we always use the `reciever` object rather than `self` so that `setsuper` can work correctly. In `setsuper`, we resolve the base class, and then set properties on it with the actual object in question as it's reciever. If a virtual setter is called, it will get the actual object it should be manipulating; and otherwise, prototypes will not be modified or consulted.
This required the reintroduction of dedicated reciever parameters to `Object.get_property_local` and `Object.set_property`, which I had removed from the AVM1 code I copied it from. It turns out being able to change the reciever was actually necessary in order to make super set/get work.
The previous system primarily relied on `Executable` to automatically start and continue a super chain. This works, but only for class hierarchies without *override gaps* - methods that override another method not defined by the direct superclass of the method. In that case, the override method would be called twice as the `base_class` was moved up one prototype at a time, which is wrong.
The new system relies on the call site to accurately report the prototype from which the current method was retrieved from. Super calls then start the resolution process *from the superclass of this prototype*, to ensure that the already-called method is skipped.
It should be noted that the proper `base_class` for things like `callmethod`, `callstatic`, `call`, `get`/`set` methods, and other call opcodes that don't use property look-up are best-effort guesses that may need to be amended later with better tests.
To facilitate `base_proto` resolution, a new `Object` method has been added. It's similar to `get_property`, but instead returns the closest prototype that can resolve the given `QName`, rather than the actual property's `ReturnValue`. Call operations use this to resolve the `base_proto`, and then resolve the method being called in `base_proto`. The existing `exec_super` method was removed and a `base_proto` method added to `exec` and `call`.
This works primarily by retaining the current superclass prototype in the activation object and then using it to retrieve the super method.
For constructors, we implement the `constructor` property, which is probably not the correct way to do this.
Also, implement a method table that method traits can optionally add themselves to.
Also also, add the ability to invoke a method without a `this` object. This required a non-trivial refactoring of the activation machinery, and changes to the signature of `NativeFunction`, and all native AVM2 functions.
Notably, this also removes `new_closure_scope` as it is not needed. AVM1 does not capture `with` scopes in closures, but AVM2 (as well as modern ECMAScript) does.
We already have a menagerie of `install_*` functions for adding static properties to a an object; and we don't have to support any kind of asinine nonsense liks `ASSetPropFlags` here. Ergo, we don't need this.
In AVM1, these are necessary because `ActionGetVariable` et. all directly interface with the scope chain. In AVM2, you `findpropstrict` up the scope chain, which gives you a normal object that you can interact with as you like. Ergo, the scope chain doesn't need set/get property methods.
All constant pools in an ABC file are actually numbered starting from one; there's an implicit 0 entry not stored in the file that the runtime is expected to retrieve when pulling constants from the pool.
The AVM2/ABC spec only mentions this in passing.
This allows the AVM to declare classes, which necessitated some refactoring to avoid double-borrows or having to do something "magic" that would dodge virtual properties.
I'm writing all this code assuming that classes and traits are syntactic sugar around ES3-style prototype chains on function objects. Hence, `FunctionObject` is still our workhorse object type for implementing typing.
David Wendt