rfcs
rfcs copied to clipboard
Replacing `moduleName` in template meta
Historically, Ember compiles a moduleName
into the wire format of each template. The Ember Inspector displays this when navigating the hierarchy of components.
This isn't really viable going forward, for two reasons:
-
moduleName
as traditionally understood doesn't exist reliably anymore- it used to be the "runtime" name of a module, which was unambiguous because all packages were forced into one flat namespace. But we don't do that anymore. Under embroider (and even in classic builds when consuming v2 addons), we follow normal node_modules resolution, so packages are not a flat namespace.
- the mapping from publicly-visible name to internal-name can be complicated, based on the full feature set of package.json exports. For us to try to produce a traditional
moduleName
we would need to faithfully run that transformation, backwards.
-
The existence of Template Tag means that templates are not one-to-one with modules anymore. So identifying which module a component lives in is not enough to unambiguously locate it.
Since templates are javascript values (or to be precise: templates are a detail of components and components are javascript values), it's worth asking: why do we think we need to label them with their module names, when we don't do that with any other javascript values?
I would like to suggest that instead of trying to label components by their template moduleName, we can change the inspector to offer a "jump to definition" button on components.
- For class-based components,
window.inspect(theComponentInstance.constructor)
already works. - But for template-only components, there's no function present to point at.
- we could introduce one inside the wire format of the template
- or we could make
templateOnlyComponent()
capture an Exception and save its stack so that it would be possible to navigate back to the site at which the component was defined.
A related concern here is that AST transforms can see moduleName
and I know some rely on it. They need a deprecation path to stop doing that. It's going to troll them anyway as soon as people are using template tag.
glimmer vm is also probably passing it as the only template information. captureRenderTree items has a property ``template` which is the same as moduleName
Isn’t the scope closure good enough?
Sorry I should be more specific, I was responding to the need for a function to “point at” for jumping to the definition; scopeFn could serve that purpose rather than introducing one specifically for debug purposes only, right?
Copying my comment from https://github.com/emberjs/ember-inspector/issues/2425#issuecomment-1665157967 here:
Wee may still need moduleName
to support hot module replacement for components (in vite), at the moment we map changed component files to rendered components by moduleName
resolved from getComponentTemplate
or component itself. (https://github.com/lifeart/demo-ember-vite/blob/master/plugins/hot-reload.ts#L169)
Today in the tooling meeting we discussed this and are generally in favor of trying to eliminate moduleName
from <template>
tag by default.
I looked around it seems like most places already assume moduleName
could be undefined and generally have fallbacks like ?? (unknown module)
I propose that
- we add a
{ moduleName = undefined }
option totemplate()
- for
<template>
, depending on how we feel about using the "attributes space", either
or<template moduleName="...">...</template>
(to be consistent with #982)// moduleName ... <template>...</template>
- if it seems like it is important to have this for dev mode and/or backwards compatibility, we ship a babel plugin that adds back the above in a best-effort compatible way
I don't think HMR should be using moduleName
at all, it should just replace the imported value.
import Foo from "./foo";
import bar from "./bar";
<template>
<Foo @foo="..." />
{{bar "some" "args"}}
</template>
Should be transformed into something like
import { tracked } from "@glimmer/tracking";
import Foo from "./foo";
import bar from "./bar";
// TODO eliminate this when not import.meta.hot
const __imports__ = new (class {
@tracked Foo = Foo;
@tracked bar = bar;
})();
if (import.meta.hot) {
import.meta.hot.accept('./foo', (Foo) => __imports__.Foo = Foo);
import.meta.hot.accept('./bar', (bar) => __imports__.bar = bar);
}
<template>
<__imports__.Foo @foo="..." />
{{__imports__.bar "some" "args"}}
</template>
Then everything should just work.
what about just "name"?
export const Foo = <template>
</template>;
gets transpiled to:
// ...
export const Foo = template('...', { name: 'Foo', /* ... */ });
no user-land customization?
The point of the feature is for backwards compatibility anyway and I think we either try to preserve existing behavior as much as possible or not do it at all
if you want it name
-ed for some reason, just use a class:
class Foo { <template>...</template> }
That would naturally give you the .name
property and you can use the normal pragma to tell the minifier to leave it alone (or not)
it's just <template ...attribute space>
would require more syntax work for some editors / tools, so my preference would be to only add something to attribute space if it's going to be a future-facing feature, rather than a backcompat one
I think that space is spec-ed/reserved for content-tag
, so it may need to happen anyway (e.g. depending on #982), but I agree I wouldn't do it just for this. So, IMO if we are using it for other things, then we should use it for this, otherwise the anything is fine.
That being said it does have its issues: e.g. what if you want to make it conditional based on import.meta.DEBUG
. However, content-tag
probably has to support that in general anyway.