rustdoc: broken "src" link

[noamraph]

In http://doc.rust-lang.org/std/sync/struct.Arc.html
click "src". It links to http://doc.rust-lang.org/alloc/arc/struct.Arc.html?gotosrc=1343 which is forwarded to http://doc.rust-lang.org/alloc/arc/undefined

[Gankra]

@alexcrichton Is there any historical or technical reason for suffixing [src] link id's with a unique numeric code? Stripping that would solve this problem. Although this might be a symptom of a larger rustdoc bug that should ultimately be fixed.

[Gankra]

After some minor sleuthing and discussion with acrichto, this is what I know.

• The bug is caused by the reexported crate yielding the wrong magic number in the ?gotosrc=xxx argument.
• The ?gotosrc=xxx argument tells rustdoc's JS to find a link on the page with id #src-xxx, and open the link.
• The reason the suffix exists is to support multiple [src] URLs on the same page, which is desired for the future.
• See src/librustdoc/html/static/main.js#732 for the JS logic
• See src/librustdoc/html/static/main.js#1264 for where the gotosrc number is written

From the latter we can see that this number is extracted from a fairly deep hierarchy of objects. self is an Item as defined in this module, from which we are extracting a clean::Item, from which we are extracting a syntax::ast::DefId, from which we are extracting a NodeId, which is just a u32. It is unclear where in this process the incorrect id is set (or perhaps the incorrect Item/DefId).

Unfortunately I'm out of my depth, and have other pressing matters at hand. As a shim to get correct behaviour, we could simply remove this unique id business until we actually need to support multiple [src] links.

[lifthrasiir]

I've analyzed this issue and this seems to be a fundamental problem with using DefId in the documentation. In the other words...

We should never use DefIds in the documentation because THEY WOULD CHANGE!

For liballoc we used to pass --cfg jemalloc to rustc but not to rustdoc, so we were actually working with different ASTs. A fix on Makefiles or even #2206 may avoid this issue (as we are looking at the same AST then) but they are not a definite solution. I'll work on the patch which removes all uses of DefIds from generated docs.