rustdoc doesn't include all text that cross attribute type boundaries in the summary

[Nemo157]

I tried this code:

#[doc = " Some text"]
#[doc = " that should"]
/// be concatenated.
pub fn main() {
    println!("Hello, world!");
}

I expected to see this happen: The entire first paragraph is used as the summary for the function in the module docs: "Some text that should be concatenated"

Instead, this happened: Only the unsugared doc attributes are used: "Some text that should"

Meta

rustc --version --verbose:

rustc 1.54.0-nightly (4e3e6db01 2021-05-18)
binary: rustc
commit-hash: 4e3e6db011c5b482d2bef8ba02274657f93b5e0d
commit-date: 2021-05-18
host: x86_64-unknown-linux-gnu
release: 1.54.0-nightly
LLVM version: 12.0.1

[ben0x539]

This seems to be achieved here:

rust/src/librustdoc/clean/types.rs

Line 1060 in 1639a16

if new_frag.kind != ori.kind || new_frag.parent_module != ori.parent_module {

(probably from #80261?)

We're concatenating the text from doc attrs here, but only until we find one of a different kind. Seems intentional enough but I don't know what it's for.

To me naively it only makes sense to concatenate single-line doc comments to generate the summary; with either multiline doc comments or explicit attrs I would expect that where they end is a good place to end the summary.

[Nemo157]

My usecase is dynamically generating some words in the first sentence based on cfg. I don't see any reason that rustdoc should distinguish between attribute types here, it has markdown's concept of a paragraph to determine how much to look at for the summary, there's no need to limit how users can construct that first paragraph.

[ben0x539]

Right, that's a compelling use case for sure! My take was "separate attributes are probably intended as separate paragraphs", but I didn't consider that you'd dynamically construct the set of attributes. I'm curious if that was the original motivation for the kind comparison or if it's needed for another reason.