Fix error in Rust 2018 + no_core environment

[taiki-e]

Minimized reproduction: https://play.rust-lang.org/?version=nightly&mode=debug&edition=2018&gist=5b9f6c3026ec9d856699fa6dbd4361f0

This is a fix for the error that occurred in #58702.

r? @Centril

[Centril]

@petrochenkov Is this possibly due to hard-coding paths in HIR lowering? (can we stop doing that and instead introduce lang_items which are at least "principled hacks"? cc @eddyb)

[eddyb]

@Centril You'd need to move lang item collection to before HIR lowering. Doable, but AFAIK it currently is still happening after HIR lowering.

[Centril]

@Centril You'd need to move lang item collection to before HIR lowering. Doable, but AFAIK it currently is still happening after HIR lowering.

@eddyb What are the primary reasons why we haven't done so hitherto? perf? "no one has put in the work"? something else? Otherwise... It seems like something we should do then?

[eddyb]

It seems like something we should do then?

Yes. I think @durka poked at this a year or two ago, but, yeah, I agree we should try again.

[petrochenkov]

Should built-in macros also search traits like Default/Debug/etc as lang items?
I don't see much difference with the for case, all those items are not tied to the language so much as "real" lang items like e.g. Drop, they only need to be found somehow.

[petrochenkov]

Proc macros face a very similar problem when they need to find items from a certain crate, and we also have extern crate self as name when we are inside that certain crate.

Same approach can be used here, I believe - extern crate self as core + ::core::option/::core::iter/etc.

[petrochenkov]

The issue with this lowering-time desugaring is that the paths it generates not only need to be resolved, but they also should be "presentable" and look differently depending on no_std mode.

::std::option::Option should actually say std rather than core where it would be found as a lang item, that's why we have all this dance with injected_crate_name.

In this sense what this PR does is the correct solution.
I could perhaps be done a bit cleaner though, e.g. with crate_root being a &str rather than Option<&str> and initialized with std_inject::injected_crate_name().unwrap_or("crate").

[petrochenkov]

Nah, this is good as is.
crate::foo works on both editions unlike ::core::foo and ::crate doesn't work so PathRoot needs to be removed anyway, so it probably can't be made more uniform.

@bors r+

[bors]

📌 Commit 8033aefeef4c8f1625c26c1888d67d5498a79548 has been approved by petrochenkov

[eddyb]

@bors r- (see #59462 (comment))

[petrochenkov]

@bors r+ (#59462 (comment) is addressed)

[bors]

📌 Commit 362d243 has been approved by petrochenkov

[Centril]

Beta nominating for bootstrap compiler due to #58702 (comment).

[pnkfelix]

discussed at T-compiler meeting. Declining to beta-backport: The only client a backport would benefit would be rustc bootstraps (and not the stable channel), and we're already going to have nightly roll to beta next week anyway, so the risk/reward ratio is bad.