[weihanglo]

What does this PR try to resolve?

Inspired by #17637.

Two things

• Move the build.rs embedded man pages generation logic to a xtask Rust file generation
• Drop the tgz compression and use direct string include_str! to embed those help texts. (as suggested in refactor(help): embed man pages without a tarball  #17640 (comment))

This shows that we can move build time code generation to xtask :)

Notes to reviewers

Binary size benchmark result (LLM assisted):

Binary	raw	gzip -9	xz -6	xz -9e
master	32,585,520	11,536,545	7,387,624	7,288,568
this PR	33,245,408	11,531,622	7,341,252	7,240,164
diff	+659,888	−4,923	−46,372	−48,404

so, about 644 KiB larger on disk, but about 4.8 KiB smaller gzipped and about 45-47 KB smaller xz-compressed. I think compressed size may matter more for downloads, unless we are going to expand our help man pages ten times more in the future.

BTW, in the future we should probably move some of our CI check to (snapshot) tests.
That is a pattern we kinda want to advocate.

---

🤖 LLM involvement: Help verify everything is the same and working well, as well as the binary size benchmark.

[rustbot]

r? @Muscraft

rustbot has assigned @Muscraft.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

Why was this reviewer chosen?

The reviewer was selected based on:

• Owners of files modified in this PR: @Muscraft, @epage, @weihanglo
• @Muscraft, @epage, @weihanglo expanded to Muscraft, epage, weihanglo
• Random selection from Muscraft, epage

[epage]

Alternatively, we stop compressing and include the text directly. Wasn't sure how reproducible the tgz would be or having a binary in the repo when I was planning to address this.

[weihanglo]

Under flate2's gz deterministic mode it should be fairly reproducible I think. Granted on Windows we need to deal with crlf stuff perhaps via gitattribute or normalization.

I can try digging a bit deeper and check the size difference with and without compression.

[weihanglo]

Under flate2's gz deterministic mode it should be fairly reproducible I think. Granted on Windows we need to deal with crlf stuff perhaps via gitattribute or normalization.

I can try digging a bit deeper and check the size difference with and without compression.

The result looks promising especially for the compressed size!

[epage]

If you feel the numbers look good, go ahead

View changes since this review

[weihanglo]

Those numbers in any direction don't matter that much. build.rs reduction is what I looks for.

Let's go.

@rustbot merge.