Places incorrect paths in extracted source maps for input that has relative sources

[rjgotten]

• Operating System: Windows 10
• Node Version: 10.15.3
• NPM Version: 6.9.0
• webpack Version: 4.31.0
• mini-css-extract-plugin Version: 0.7.0

Expected Behavior

When offered compiled CSS content backed by a source map where sources are relative URLs, extracts a .css file with a source map that has correct path information. That is: the sources are resolved against this.context of the compiled CSS content, producing the correct source locations.

Actual Behavior

When offered compiled CSS content backed by a source map where sources are relative URLs, extracts a css file with a source map that has incorrect path information. The sources seem to be output as-is to be interpreted by debuggers and dev-tools as relative to the extracted CSS file. This produces incorrect source locations.

Code

N/A

How Do We Reproduce?

Minimal repro case:

https://github.com/NetMatch/mini-css-extract-plugin-testcase

[rjgotten]

Somehow, I think this all ties back into the fact that Webpack itself wants disk-based - i.e. filesystem - absolute paths in source map sources whereas according to specification those are supposed to be URLs.

When said URLs are relative, they are relative to the generated file. Most external supplied source maps are relative, because that's the only way they can be portable.

---

Assuming Webpack is not going to budge and fix this mess, a work-around is needed where source map sources are eventually made compatible with Webpack's counter-spec status quo.

This is complicated for a few reasons, mostly because there's a lot of moving parts capable of influencing the source map for CSS: less-loader; sass-loader; postcss-loader; css-loader; mini-css-extract-plugin and possibly more.

Both postcss-loader and css-loader internally use PostCSS. This must receive proper (relative) URL sources in the map, or it produces bad paths in its output sources. Usually you ge especially bad results on Windows, where the path separator for filesystem paths is \ and not / making a further mess of things.

A custom loader module can be written to normalize any unfit source maps before they are offered to either of these loaders. However, specifically for css-loader there is no opportunity to rewrite its outmap maps into the disk-based absolute paths that core Webpack expects, as css-loader emits a JS module with the CSS source map encapsulated in it.

It and css-mini-extract-plugin following it in the loader chain, are opaque to end-users. I.e. the issue will need to be worked around from inside the plugin (and loader).

[alexander-akait]

PR welcome, maybe need fix in css loader first

[rjgotten]

@evilebottnawi

I'm now running a small experiment, using esprima and escodegen to parse the JS module output by css-loader; re-transform the relative URLs in its embedded CSS source-map to absolute disk paths; and then re-emit the corrected module.

Sofar, that seems to work well with the MiniCssExtractPlugin as well. If this holds up, I'll consider building up a pull request with a real fix. (There's some performance-related motivation there for me to provide an integrated fix. Because of course; using esprima for outside manipulation is not cheap. ;-) )

[alexander-akait]

Anyway feel free to put here any information and feedback

[alexander-akait]

Partial fixed for latest css-loader/mini-css-extract-plugin/webpack versions. Just note:

• webpack:./style/content.less should be webpack:///./style/content.less

[alexander-akait]

Now you will have "sources":["webpack:///./style/content.less","webpack:///./style/content.css"], it is hard to say css-loader what we should use content.less in from, because PostCSS doesn't support multiple values in the from option

[alexander-akait]

I want to move this issue to css-loader, because nothing to do here, we should found a way to say css-loader/postcss-loader using right source from source-map-loader

[alexander-akait]

@rjgotten Do you have ideas how we can achieve this?

[rjgotten]

Maybe just cop out? Introduce a rule that says the first candidate matching a given predicate test is passed on.
Or introduce a system which favors one extension type over the other?

It's situational, I think. So probably a user specified predicate/test regex or function is the better deal.

Then rewrite the source-map to only contain those matching sources, in situations where multiple are provided.

[alexander-akait]

@rjgotten We can't rewrite source maps, PostCSS does it, we just provide from/to and previously source map

[rjgotten]

Ah sorry, about the confusion there.

I didn't mean "rewrite the source-map" as in creating a chained source map, which indeed is what PostCSS should be doing.
I meant to rewrite the sources attribute of the source-map that is passed to PostCSS, so it only shows one source to PostCSS.

Thinking further though: why is the presence of multiple sources in the incoming map even an issue for the from option?
The PostCSS from and to options are the input and output CSS being processed.
And iirc the only connection from has to source-maps is that the file name provided there will be used as the content of a source-map's sources in the event that PostCSS starts a new source map - i.e. when it is not chaining off of an existing one. In case of an incoming source-map that already has multiple sources, that case obviously doesn't - or at least: shouldn't - apply.

[alexander-akait]

We can pass to from first source (map.sources[0]) if we have it in source map, so we can achieve this

[rjgotten]

Oh; right. Now I get what the problem is.
For compatibility's sake, you need to set a name for the source, because you're passing it in programmatically and not from disk even though you're chaining off of an existing source map. (Yeah; that took a moment to click with me... sorry.)

Sure; in that case just going with whatever is listed as the first entry in map.sources should work.
Heck, just passing in any arbitrary string should work, I guess, because it should never end up in the output source-map. The output source-map should be propagating the original sources from the input source-map, after all.

[alexander-akait]

Yep, I am worried that some of the tools may be generating incorrect invalid order of sources, but I think this is not our problem