An automatic way to reduce times of visiting useless AST nodes

[hyf0]

I think about this while reducing times of visiting useless AST nodes manually.

let’s say we have a macro called magic_visit_mut . We could use it like

struct OnlyCareModuleItem;

magic_visit_mut!(impl VisitMut for OnlyCareModuleItem {
  fn visit_module_item(&mut self, node: &mut ModuleItem) {
		// do something
  }
})

and the macro will expand as

struct OnlyCareModuleItem;

impl VisitMut for OnlyCareModuleItem {
  // --- noop visiting function automatically generated by  `magic_visit_mut`
  fn visit_mut_script(&mut self, node: &mut Script) {}
  // --- End

  fn visit_module_item(&mut self, node: &mut ModuleItem) {
    // do something
  }
}

algorithm

We will start with the node visiting function that is declared in magic_visit_mut. If a node could be visited, then all possible parents and children of it could be visited. After the analysis, we will generate noop visiting function for non-visited nodes.

And, we could add extra hints for each visitor to help with analysis. Such as #[magic_visit_mut(wont_visit_children)]

struct OnlyCareModuleItem;

magic_visit_mut!(impl VisitMut for OnlyCareModuleItem {
  #[magic_visit_mut(wont_visit_children)]
  fn visit_module_item(&mut self, node: &mut ModuleItem) {
		// do something
  }
})

the macro will expand as

struct OnlyCareModule;

impl VisitMut for OnlyCareModuleItem {
  // --- noop visiting function automatically generated by  `magic_visit_mut`
  fn visit_mut_script(&mut self, node: &mut Script) {}

  fn visit_mut_import_decl(&mut self, node: &mut Script) {}

  fn visit_mut_export_decl(&mut self, node: &mut Script) {}

  fn visit_mut_stmts(&mut self, node: &mut Script) {}

  // ....
  // --- End

  fn visit_module_item(&mut self, node: &mut ModuleItem) {
    // do something
  }
}

I’m not sure how much the performance could be improved, but I guess it wasn’t bad to raise an issue to discuss it.

[kdy1]

I thought rustc would optimize it to noop, but seems like it was not the case?

[kdy1]

Doesn't LTO work? There's no #[inline], so LTO is required to optimize it out.

[hyf0]

Not sure, I will look into it.

[kdy1]

Did it work?

[hyf0]

I haven't investigated it yet. But I found a visitor of SWC only takes 20+ms to visit a huge AST parsed from a 20mb+ javascript file, which makes me feel it's unnecessary to do such optimization.

[kdy1]

I see. Thank you for sharing!