**Description:**
Following up https://github.com/swc-project/swc/pull/7422, `swc_core::plugin` itself is not exposed to `__plugin_transform_host`, so still not able to access the inner plugin proxy.
**Description:**
One of the oversight around design of `TransformExecutor` is
encapsulating plugin module logic. It has access to the cache and do its
own loading & storing. This means consumer of plugin runner have tricky
challenge to control its caching system. First, there is no way to
escape how swc_plugin_runner controls cache and cannot synchronize into
their own, also depends on the usecases cannot control the features they
want to opt in: for example, there's no way one interface uses in-memory
cache, and another uses filesystem since it is compile time configured
singleton.
PR revisits overall design of TransformExecutor: now it accepts a tratir
`PluginModuleBytes`, which abstracts any kind of bytes we are dealing
with, such as raw file slice or serialized `wasmer::Module`. Cache
instantiation and managing is now bubbled up to the application level
(`swc` in here), so if someone wants non-singleton caching or integrate
into their own caching system it can be customized.
Lastly, deprecated `memory_cache` feature and only exposes
`filesystem_cache`. Cache implementation uses in-memory is always
available, and can opt in filesystem cache where it's supported.
**BREAKING CHANGE:**
This is clearly breaking changes for the consumers of swc_core. for the
@swc/core, this PR takes care of necessary changes. I'll work on
next-swc changes later once we have new @swc/core version with this
changes.
**Description:**
This issue is more severe than I originally thought. It raises not in
array indexing, but in function calls and property mutation. We should
treat all function arguments as potentially be property mutated,
otherwise following example
```js
class A {
a = 1
toString() {
return this.a
}
}
const a = new A()
function foo(x) {
x.a++
}
const b = a + 1
foo(a)
console.log(b)
```
would be error(It should log 2, but logs 3 after compress).
As the result, massive regressions is unavoidable, since some of these
optimizations may indeed cause error. Part of them can be mitigated with
following optimization -- allow inline of ident even if its original
value is mutated. Consider
```js
export function foo(x) {
const y = x
x.a = 1
y.b = 2
}
```
If x is a primitive value, all mutations to its properties are ignored;
if x is a object, then y refers to the same object no matter what
mutation is performed.
And there's still room for more, currently following code
```js
export function foo(x) {
const y = Math.floor(x);
g(y);
}
```
But I'd rather do it in a separate PR.
**Related issue:**
- Closes#7402.
