Migrate to op2. Making a few decisions to get this across the line:
- Empty slices, no matter where the come from, are null pointers. The v8
bugs (https://bugs.chromium.org/p/v8/issues/detail?id=13489) and
(https://bugs.chromium.org/p/v8/issues/detail?id=13488) make passing
around zero-length slice pointers too dangerous as they might be
uninitialized or null data.
- Offsets and lengths are `#[number] isize` and `#[number] usize`
respectively -- 53 bits should be enough for anyone
- Pointers are bigints. This is a u64 in the fastcall world, and can
accept Integer/Int32/Number/BigInt v8 types in the slow world.
Few improvements to FFI types:
1. Export `PointerObject` for convenience. It's fairly commonly used in
library code and thus should be exported.
2. Fix various comments around `PointerValue` and `UnsafePointer` and
expand upon them to better reflect reality.
3. Instead of using a `Record<"value", type>[T]` for determining the
type of an FFI symbol parameter use direct `T extends "value" ? type :
never` comparison.
The last part enables smuggling extra information into the parameter and
return value string declarations at the type level. eg. Instead of just
`"u8"` the parameter can be `"u8" & { [brand]: T }` for some `T extends
number`. That `T` can then be extracted from the parameter to form the
TypeScript function's parameter or return value type. Essentially, this
enables type-safe FFI!
The foremost use-cases for this are enums and pointer safety. These are
implemented in the second commit which should enable, in a backwards
compatible way, for pointer parameters to declare what sort of pointer
they mean, functions to declare what the API definition of the native
function is, and for numbers to declare what Enum they stand for (if
any).
Attempts to fix the thread_safe_callback flakiness. It's unclear what
the flake is about, the exit code is apparently `C0000005` or
`ACCESS_VIOLATION`, pointing to an issue with memory access. My only
guess is that maybe dropping the `Option<extern "C" fn ()>` is somehow
checking the validity of the function pointer and since the function has
been dropped, the pointer is no longer valid and sometimes points to
memory that should not be accessed.
So now the will explicitly drop the functions before they get
deallocated. If this doesn't fix the flake then something beyond my
understanding is wrong.
This patch makes `NativeType` to `libffi::middle::Type` conversion
failliable and w.t disallows struct with empty fields. libffi does not
handle "empty" struct because they don't exist in C (or Rust).
Fixes #17481
Adds support for passing and returning structs as buffers to FFI. This does not implement fastapi support for structs. Needed for certain system APIs such as AppKit on macOS.
Currently, slow call path will always create a dangling pointer to
replace a null pointer when called with eg. a `new Uint8Array()`
parameter, which V8 initialises as a null pointer backed buffer.
However, the fast call path will never change the pointer value and will
thus expose a null pointer. Thus, it's possible that the pointer value
that a native call sees coming from Deno changes between two sequential
invocations of the same function with the exact same parameters.
Since null pointers can be quite important, and `Uint8Array` is the
chosen fast path for Deno FFI `"buffer"` parameters, I think it is
fairly important that the null pointer be properly exposed to the native
code. Thus this PR.
### `*mut c_void`
While here, I also changed the type of our pointer values to `*mut
c_void`. This is mainly due to JS buffers always being `*mut`, and
because we offer a way to turn a pointer into a JS `ArrayBuffer`
(`op_ffi_get_buf`) which is read-write. I'm not exactly sure which way
we should really go here, we have pointers that are definitely mut but
we also cannot assume all of our pointers are. So, do we go with the
maxima or the minima?
### `optimisedCall(new Uint8Array())`
V8 seems to have a bug where calling an optimised function with a newly
created empty `Uint8Array` (no argument or 0) will not see the data
pointer being null but instead it's some stable pointer, perhaps
pointing to some internal null-backing-store. The pointer value is also
an odd (not even) number, so it might specifically be a tagged pointer.
This will probably be an issue for some users, if they try to use eg.
`method(cstr("something"), new Uint8Array())` as a way to do a fast call
to `method` with a null pointer as the second parameter.
If instead of a `new Uint8Array()` the user instead uses some `const
NULL = new Uint8Array()` where the `NULL` buffer has been passed to a
slow call previously, then the fast call will properly see a null
pointer.
I'll take this up with some V8 engineers to see if this couldn't be
fixed.
Potential fix for type-code mismatch in FFI buffer types. The code
supports ArrayBuffers, but types only reflect TypedArray support.
There's also an existing type for this sort of stuff: `BufferSource`.
(Although, it uses `ArrayBufferView` which doesn't actually connect with
the TypedArray interfaces specifically, but it's just a type inheritance
difference and nothing more.)
Makes `op_ffi_ptr_of` fast. One of the tests changed from printing
`false` to `true` as the fast `&[u8]` slice path creates the slice with
a null pointer. Thus the `op_ffi_ptr_of` will now return a null pointer
value whereas previously it returned a dangling pointer value.
This PR makes pointer read methods of `Deno.UnsafePointerView` Fast API
compliant, with the exception of `getCString` which cannot be made fast
with current V8 Fast API.
Fixes #15136
Currently `UnsafeCallback` class' `ref()` and `unref()` methods rely on
the `event_loop_middleware` implementation in core. If even a single
`UnsafeCallback` is ref'ed, then the FFI event loop middleware will
always return `true` to signify that there may still be more work for
the event loop to do.
The middleware handling in core does not wait a moment to check again,
but will instead synchronously directly re-poll the event loop and
middlewares for more work. This becomes a live-loop.
This PR introduces a `Future` implementation for the `CallbackInfo`
struct that acts as the intermediary data storage between an
`UnsafeCallback` and the `libffi` C callback. Ref'ing a callback now
means calling an async op that binds to the `CallbackInfo` Future and
only resolves once the callback is unref'ed. The `libffi` C callback
will call the waker of this Future when it fires to make sure that the
main thread wakes up to receive the callback.
Welcome to better optimised op calls! Currently opSync is called with parameters of every type and count. This most definitely makes the call megamorphic. Additionally, it seems that spread params leads to V8 not being able to optimise the calls quite as well (apparently Fast Calls cannot be used with spread params).
Monomorphising op calls should lead to some improved performance. Now that unwrapping of sync ops results is done on Rust side, this is pretty simple:
```
opSync("op_foo", param1, param2);
// -> turns to
ops.op_foo(param1, param2);
```
This means sync op calls are now just directly calling the native binding function. When V8 Fast API Calls are enabled, this will enable those to be called on the optimised path.
Monomorphising async ops likely requires using callbacks and is left as an exercise to the reader.
