Windows-rs: Optimizing docs generation

[jcotton42]

I had brought up the size of the generated rustdoc for windows-rs on the Rust Community Discord and rustdoc team member @jyn514 mentioned that adding the #![doc(html_no_source)] to the crate should suppress the HTML rendering of the crate source.

https://doc.rust-lang.org/stable/rustdoc/the-doc-attribute.html#html_no_source

[kennykerr]

Thanks!

[jyn514]

I can't build the docs on linux:

  = note: /usr/bin/ld: cannot find -loleaut32
          /usr/bin/ld: cannot find -lkernel32
          /usr/bin/ld: cannot find -loleaut32
          /usr/bin/ld: cannot find -lkernel32
          /usr/bin/ld: cannot find -lole32
          /usr/bin/ld: cannot find -lkernel32
          /usr/bin/ld: cannot find -lkernel32
          /usr/bin/ld: cannot find -lkernel32
          /usr/bin/ld: cannot find -loleaut32

I'd be interested how long rustdoc takes to run after you add html_no_source. Could someone generate some timing info for me?

$ rustup install nightly
$ time -v cargo +nightly rustdoc -- -Z time-passes

and then upload windows-PID.mm_profdata.
(time -v tells the total memory usage; if it's not on windows anything else that gives the max RSS works too)

[kennykerr]

Hey @jyn514 thanks for the help! @rylev may have some tips for how to build on Linux - I have no idea.

Note that you need to build a separate repo. The docs are too huge that I didn't want to include them inside the windows-rs repo. Instead, you can clone the https://github.com/microsoft/windows-docs-rs repo and build the crates/bindings package. Something like this:

C:\git\windows-docs-rs\crates\bindings>cargo doc --no-deps

This takes about 30 minutes and 20GB of memory. I updated the lib.rs file to look like this:

#![doc(html_no_source)]
::windows::include_bindings!();

Note that time -v is not available on Windows. I ran the following:

cargo +nightly rustdoc -- -Z time-passes

It took about 52 mins and it's peak working set (memory) is over 21GB.

I could not find a windows-PID.mm_profdata anywhere.

[jyn514]

Hmm, it should be there somewhere - do you see any file ending in .mm_profdata? It may be in the workspaces root, not the directory you ran cargo in.

If not, what version of rustdoc are you using? I think the format of the self-profile files changed a few months ago.

[kennykerr]

Not from what I can tell:

[kennykerr]

C:\git\windows-docs-rs>rustdoc --version
rustdoc 1.49.0 (e1884a8e3 2020-12-29)

[jyn514]

Oh shoot, I'm sorry, it should have been -Z self-profile, not -Z time-passes. It's ok, time-passes gave some useful info too. In particular it shows most of the time is spent on get_blanket_impls, which I've been meaning to improve for a while. It doesn't show where the memory usage is coming from, though, it already uses 18 GB before looking at impls.

You posted the default version of rustdoc - can you post the nightly version instead? rustdoc +nightly --version.

[kennykerr]

C:\git\windows-docs-rs>rustdoc +nightly --version
rustdoc 1.51.0-nightly (a4cbb44ae 2021-01-20)

OK, I'll rerun with self-profile.

[kennykerr]

Hmm, it's rather large. 😉

[kennykerr]

Not sure if it will work, but I've uploaded to OneDrive. Let me know if you can download it.

[jyn514]

That worked, thanks! Downloading it now, I'll let you know what I find :)

[jyn514]

By the way, you can analyze it yourself with measureme: https://github.com/rust-lang/measureme/

[jyn514]

A surprisingly long time is being spent not in rustdoc, but in the compiler itself:

In particular, resolve_crate and item_types_checking both take a lot longer than I would expect.

[kennykerr]

The Rust compiler certainly has a lot of trouble with the resulting 300M windows.rs containing the definitions for the entire Windows API. While it is a lot of code, the Rust compiler has a long way to go before it can compete with the like of the C++ compiler in terms of throughput.

[rylev]

For those following along at home, the summarize summary can be seen here.

[kennykerr]

I have applied the html_no_source update - thanks for the suggestion!

GitHub no longer refuses to accept some of the files but continues to warn about the huge search-index.js file:

remote: warning: File docs/doc/search-index.js is 57.14 MB; this is larger than GitHub's recommended maximum file size of 50.00 MB

The size of this file also seems to hurt page load and search performance in the browser.

[jyn514]

Yeah, it's a known issue: https://github.com/rust-lang/rust/issues/31387. I'll look into it after fixing https://github.com/rust-lang/rust/issues/81251.

[kennykerr]

Thanks again for all your help.

[jyn514]

FYI, what we've found is that the actual CPU time is not all that high (comparable to e.g. docs.rs), the problem is that compiling shoots up to 20 GB of memory usage which sets almost everyone's computer thrashing. @rylev compiled this on a computer with more than that amount of memory and it took about 2.5 minutes. https://rust-lang.zulipchat.com/#narrow/stream/247081-t-compiler.2Fperformance/topic/windows-rs.20perf/near/223954279

[kennykerr]

Was that on a Windows box? IO performance is very different between Windows and Linux. My dev box has 32GB of memory and it took around 30 minutes without much disk thrashing. It was definitely CPU bound and largely chugging away on a single thread. But maybe I missed something.

[jyn514]

Oh, that's really strange. Maybe @rylev measured the wrong thing? It definitely took him no more than 3 minutes:
bindings-18880.summarize.log

[rylev]

Sorry, I should have made this more clear to @jyn514 that I was testing a subset of the full windows API surface. So it's perfectly conceivable that it would take ~30 minutes on my machine as well. I also have 32 GB of RAM.

[rylev]

FYI: The changes I made in https://github.com/rust-lang/rust/pull/81419 and https://github.com/rust-lang/rust/pull/81414 reduce cargo build build times of the bindings in this repo from ~43m45s to ~32m45s! I've not measured the impact on cargo doc runs but I imagine it will also have a fairly large impact.

[rylev]

If anyone is interested what the timings look like for a smaller subset of the bindings, here they are.

[kennykerr]

That's fantastic. Looks like some more low-hanging fruit... 😄

[kennykerr]

Turns out its much slower than I first imagined. I was looking at the summary time reported by cargo but it's lying. Process Explorer reports that's its off by quite a bit. 😉

Also take a look at how much memory its using.

[kennykerr]

I've done all I can to optimize this from my end. Going to close this issue unless someone has a reason to keep it open.

[jyn514]

Hmm, I don't know if anyone has permissions to transfer this from microsoft/windows-rs to rust-lang/rust but if so that would be ideal.

[kennykerr]

I've asked someone at GitHub to take a look.

[kennykerr]

Unfortunately, GitHub doesn't support this. Happy to leave the issue open if you'd prefer.

[jyn514]

Hmm that's weird, there's definitely a "transfer issue" button on rustc-dev-guide for me. Maybe GitHub doesn't support transferring between organizations?

Anyway, I think most of the rustdoc side of this is tracked elsewhere so I'd be ok with closing this.

[kennykerr]

Yes, you can transfer within an organization - I do that regularly - but you cannot do so across organizations (at least according to my GitHub contact).