If we are out of memory, we can't try to allocate a string that could
fail as well. When Error is converted to String, this would result in an
endless OOM-throwing loop. Instead, pre-allocate the string on the VM,
and use it to construct the Error.
Note that as of this commit, the OOM string is still a DeprecatedString.
This is just preporatory for Error's conversion to String.
Depending on how this is invoked, the preprocessor may get confused when
pasting the VM parameter into this expression. For example, it trips up
on the JS REPL in cases such as:
TRY_OR_THROW_OOM(*g_vm, ...);
For example, consider cases where we want to propagate errors only in
specific instances:
auto result = read_data(); // something like ErrorOr<ByteBuffer>
if (result.is_error() && result.error().code() != EINTR)
continue;
auto bytes = TRY(result);
The TRY invocation will currently copy the byte buffer when the
expression (in this case, just a local variable) is stored into
_temporary_result.
This patch binds the expression to a reference to prevent such copies.
In less trival invocations (such as TRY(some_function()), this will
incur only temporary lifetime extensions, i.e. no functional change.
This is to allow using this macro in contexts that have defined `is`
already. For example, in ObjectConstructor, there is a native function
`is` which would trip up the compiler without this change.
This should solely be used to ignore completions from Heap::allocate in
currently-infallible contexts. It's mostly meant to let us both ignore
these errors and mark them with a FIXME in one go.
If a spec step hasn't been marked as fallible, but might throw due to
OOM, this is to make it clear that OOM is the only thing that may cause
a failure.
Currently, if you have a fallible function with an Optional return type
and try to return OptionalNone like this:
ThrowCompletionOr<Optional<SomeType>> { return OptionalNone {}; }
Then ThrowCompletionOr's m_value (whose type is an Optional<ValueType>)
will be set to empty, rather than containing an empty Optional. This is
due to the ThrowCompletionOr's WrappedValueType constructor.
This patch adds a constructor specifically for OptionalNone. If someone
attempts to construct a ThrowCompletionOr with OptionalNone, but the
ValueType is not an Optional, a compile error like the following will
occur:
ThrowCompletionOr<String> foo() { return OptionalNone {}; }
error: no matching constructor for initialization of 'AK::String'
Otherwise, it will fill ThrowCompletionOr's m_value with an empty
Optional.
If USING_AK_GLOBALLY is not defined, the name IsLvalueReference might
not be available in the global namespace. Follow the pattern established
in LibTest to fully qualify AK types in macros to avoid this problem.
DeprecatedFlyString relies heavily on DeprecatedString's StringImpl, so
let's rename it to A) match the name of DeprecatedString, B) write a new
FlyString class that is tied to String.
Move the macro to LibJS and change it to return a throw completion
instead of a WebIDL exception. This will let us use this macro within
LibJS to handle OOM conditions.
Default-constructing the m_value Completion made it have an undefined
JS value when not overridden in a constructor, such as the conditional
initialization in Optional(Optional<JS::Completion> const&).
See investigation by Tim here:
https://github.com/SerenityOS/serenity/pull/16498#discussion_r1049090456
Co-authored-by: Timothy Flynn <trflynn89@pm.me>
It was possible for the generic ThrowCompletionOr constructor to
"copy-construct" a JS Object when instantiating a ThrowCompletionOr
via e.g `return *object;`.
This happened because it chose the Object(Object& prototype) constructor
which will be removed in a subsequent commit. It was not easy to debug.
As a first step towards avoiding this in the future, the generic
ThrowCompletionOr constructor now takes the value as a const reference.
Since Completion has an enum for state we have plenty of space to add
an extra type which indicates on empty completion.
Since Optional<Completion> is used in every ThrowCompletionOr<...>
this saves quite some stack space as most function in LibJS return
types like that.
This saves 8 bytes for every Optional<Completion>.
In the ThrowCompletionOr constructors, the VERIFY statements are using
moved-from objects. We should not rely on those objects still being
valid after being moved.
In the end this is a nicer API than having separate has_{value,target}()
and having to check those first, and then making another Optional from
the unwrapped value:
completion.has_value() ? completion.value() : Optional<Value> {}
// ^^^^^^^^^^^^^^^^^^
// Implicit creation of non-empty Optional<Value>
This way we need to unwrap the optional ourselves, but can easily pass
it to something else as well.
This is in anticipation of the AST using completions :^)
It also needs to be able to take what the spec calls 'empty', which is
an Optional<Value> in this case (*not* an empty JS::Value). The common
use case is updating a completion with another completion, that may also
have an empty [[Value]] slot.
TL;DR: Instead of this:
return { TRY(my_object()) };
we can now do:
return TRY(my_object());
just like we mostly did for Value return types before ThrowCompletionOr.
The Completion constructor `VERIFY()`s that the passed argument is not
an empty Value, so normal_completion({}) would crash (although it's
currently not being used anywhere).
We want to pass an empty Optional<Value> instead.
We decided that we want to move away from throwing exceptions in AOs
and regular functions implicitly and then returning some
default-constructed value (usually an empty JS::Value) - this requires
remembering to check for an exception at the call site, which is
error-prone. It's also awkward for return values that cannot be
default-constructed, e.g. MarkedValueList.
Instead, the thrown value should be part of the function return value.
The solution to this is moving closer to the spec and using something
they call "completion records":
https://tc39.es/ecma262/#sec-completion-record-specification-type
This has various advantages:
- It becomes crystal clear whether some AO can fail or not, and errors
need to be handled and values unwrapped explicitly (for that reason
compatibility with the TRY() macro is already built-in, and a similar
TRY_OR_DISCARD() macro has been added specifically for use in LibJS,
while the majority of functions doesn't return ThrowCompletionOr yet)
- We no longer need to mix "valid" and "invalid" values of various types
for the success and exception outcomes (e.g. null/non-null AK::String,
empty/non-empty JS::Value)
- Subsequently it's no longer possible to accidentally use an exception
outcome return value as a success outcome return value (e.g. any AO
that returns a numeric type would return 0 even after throwing an
exception, at least before we started making use of Optional for that)
- Eventually the VM will no longer need to store an exception, and
temporarily clearing an exception (e.g. to call a function) becomes
obsolete - instead, completions will simply propagate up to the caller
outside of LibJS, which then can deal with it in any way
- Similar to throw we'll be able to implement the functionality of
break, continue, and return using completions, which will lead to
easier to understand code and fewer workarounds - the current
unwinding mechanism is not even remotely similar to the spec's
approach
The spec's NormalCompletion and ThrowCompletion AOs have been
implemented as simple wrappers around the JS::Completion constructor.
UpdateEmpty has been implemented as a JS::Completion method.
There's also a new VM::throw_completion<T>() helper, which basically
works like VM::throw_exception<T>() - it creates a T object (usually a
JS::Error), and returns it wrapped in a JS::Completion of Type::Throw.
Two temporary usage patterns have emerged:
1. Callee already returns ThrowCompletionOr, but caller doesn't:
auto foo = TRY_OR_DISCARD(bar());
2. Caller already returns ThrowCompletionOr, but callee doesn't:
auto foo = bar();
if (auto* exception = vm.exception())
return throw_completion(exception->value());
Eventually all error handling and unwrapping can be done with just TRY()
or possibly even operator? in the future :^)
Co-authored-by: Andreas Kling <kling@serenityos.org>
