It's essential to `oxc_traverse`'s safety scheme that the user cannot create a `TraverseAncestry`, because they could then substitute it for the one stored in `TraverseCtx`, and cause a buffer underrun when an ancestor gets popped off stack which should never be empty - but it is because user has sneakily swapped it for another one.
Not being able to create a `TraverseAncestry` also requires that user cannot obtain an owned `TraverseCtx` either, because you can obtain an owned `TraverseAncestry` from an owned `TraverseCtx`.
Therefore, it's unsound for `TraverseCtx::new` to be public.
However, it is useful in minifier to be able to re-use the same `TraverseCtx` over and over, which requires having an owned `TraverseCtx`.
To support this use case, introduce `ReusableTraverseCtx`. It is an opaque wrapper around `TraverseCtx`, which prevents accessing the `TraverseCtx` inside it. It's safe for user to own a `ReusableTraverseCtx`, because there's nothing they can do with it except for using it to traverse via `traverse_mut_with_ctx`, which ensures the safety invariants are upheld.
At some point, we'll hopefully be able to reduce the number of passes in the minifier, and so remove the need for `ReusableTraverseCtx`.But in the meantime, this keeps `Traverse`'s API safe from unsound abuse.
Note: Strictly speaking, there is still room to abuse the API and produce UB by initiating a 2nd traversal of a different AST in an `Traverse` visitor, and then `mem::swap` the 2 x `&mut TraverseCtx`s. But this is a completely bizarre thing to do, and would basically require you to write malicious code specifically designed to cause UB, so it's not a real risk in practice. We'd need branded lifetimes to close that hole too.
So this PR doesn't 100% ensure safety in a formal sense, but it at least makes it very hard to trigger UB *by accident*, which was the risk before.
This definitely needs some cleanup by somebody more familiar with Rust,
but I think it contains most of the important pieces. I'll leave a
review with the different pieces that I know need cleanup.
Ref #6966.
Small optimization. `transform_call_expression_impl` does not require the `Expression::CallExpression`, only the `CallExpression` itself. So pass that into the function, instead of having to unwrap it (which is an unnecessary branch).
Apologies @Boshen for the stupidly large diff.
I've done the following:
- `RuleEnum`'s members are now prefixed with the plugin name. e.g. `NoDebugger` has become `EslintNoDebugger`
- updated tester.rs to accept the rule's NAME, CATEGORY to allow us to test rules with the same name (diff category)
- updates `declare_all_lint_rules` satisfy the first above change.
Script to update fixtures for class properties transform was failing to transform fixtures where output is `output.mjs` (instead of `output.js`). This PR fixes that.
Transform private property accesses in static prop initializers. e.g.:
Input:
```js
class C {
static #x = 123;
static y = this.#x;
}
```
Transformed:
```js
class C {}
var _x = { _: 123 };
babelHelpers.defineProperty(C, "y", babelHelpers.assertClassBrand(C, C, _x)._);
```
`this.#x` has been transformed to `babelHelpers.assertClassBrand(C, C, _x)._`.
- Remove `format!` macro
- Rename string related macros, `&'static`: `text!` and `'a`: `dynamic_text!`
- Apply `wrap!` macro instead of manually using `enter|leave_node`
- Introduce `/ir` directory and move `Doc`, `impl Display`, `DocBuilder`, etc.
I'm not yet determined how to, how deep to use macro, but this is first stepping stone... 🥌
Add a no-op placeholder method for transforming private fields as `AssignmentPattern`s. e.g. `[object.#prop] = []`.
Implementation will be added later.
Fix edge case in class static properties transform. When transforming `super()` in class constructor, stop searching when hit a `TSModuleBlock`. `TSModuleBlock` is essentially a function, so `super()` there is invalid.
