`x.py test linkchecker` builds rustc twice

[jyn514]

Found in https://rust-lang.zulipchat.com/#narrow/stream/242791-t-infra/topic/tool.20linkchecker.20needs.20stage.202.20compiler.20.3F: Currently, x.py doc uses --stage 0, but x.py linkchecker uses --stage 2. This causes enormous recompiles for no reason; the two should match.

[ehuss]

I believe x.py test src/tools/linkchecker defaults to runs at stage 1 like all other tests. Is there some config or something causing it to run at stage 2?

I'd be slightly concerned running it at stage 0, since some doc tools need to link with the compiler, they would be using an old compiler in that case. For quick checks, it should be OK, but like running any tests at stage 0, it is incomplete.

Also, I believe not all docs can be built at stage 0 (for example, if you have compiler-docs enabled), which could cause confusing errors. That should probably have some check to skip or error with compiler docs at stage 0.

[jyn514]

Hmm, this is a good point. @lzutao what do you think?

[jyn514]

I'd be slightly concerned running it at stage 0, since some doc tools need to link with the compiler, they would be using an old compiler in that case. For quick checks, it should be OK, but like running any tests at stage 0, it is incomplete.

In particular, they'd be missing recent fixes to intra-doc links which would fix broken links, yeah.

[tesuji]

I don't actually recall the stage 2 std or stage 2 compiler was built with linkchecker,
it actually took long enough that I don't want to run it again.

My config.toml:

ccache = true
ninja = true
targets = "AArch64;ARM;Mips;X86"
cargo = "/home/lzutao/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/bin/cargo"
rustc = "/home/lzutao/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/bin/rustc"
submodules = false
local-rebuild = true
codegen-units = 0
incremental = true
llvm-config = "/home/lzutao/.local/bin/llvm-config"

[ehuss]

After looking at this in more detail, I think the issue is that rustc is built twice. The second build is due to the error-index-generator, which links to rustc and rustdoc in order to reuse the markdown renderer. I'm not sure if there's a particularly good way to work around that. AFAIK, linking to rustc must use a stage1 build. The only alternative I can think of is using a different renderer (either mdbook, or something custom using pulldown-cmark?). I'm not sure that would be really worth it, but something to consider.

[Mark-Simulacrum]

You can definitely link to rustc libraries in any stage - not sure what you mean by stage 1 there.

In the long run I think mdbook is probably a good idea for the error index though.

[ehuss]

Yea, I was thinking about it more, and I was mistaken. I was thinking about ABI issues (with stage1-tools built with a stage1 compiler linking against stage0-rustc built with a stage0 compiler). But I guess it should use the same trick that rustdoc does to handle this (I think error-index-generator would need to replicate most of this logic to use stage-1).

A problem with mdbook is that it can't output to a single file. The URL https://doc.rust-lang.org/error-index.html might need to change to something like https://doc.rust-lang.org/error-index/index.html, which could be a little disruptive (at least needing a redirect). I'm also curious if it would be better to have each error on a separate page, or continue with one large page. (I'm uncertain about how slow a single page would be, considering mdbook does client-side syntax highlighting.) I'm also uncertain if anyone ever visits the error-index page at all (seems a bit silly to put effort into something nobody uses).

@Mark-Simulacrum do either of those approaches seem preferable (or neither)?

[jyn514]

I'm also uncertain if anyone ever visits the error-index page at all (seems a bit silly to put effort into something nobody uses).

Maybe a simpler solution would be to make the error-index opt-in instead of opt-out, so that it's only generated by CI? Then you wouldn't have to mess with it to get x.py test linkchecker --stage 0 working.

[Mark-Simulacrum]

I think the redirect shouldn't be too hard, I wouldn't see that as a problem. Distinct pages are a bit annoying, though I imagine that 99% of the time people are looking at just one error of the error index rather than all of them at the same time.

I think personally I'd prefer to make the error index generator not use rustdoc/rustc as a library. Maybe an easy way of doing that is to just roll the single file into rustdoc itself behind a forever unstable -Z flag or something, rather than trying to re-engineer a bunch of mdbook stuff.