Use `Symbol` even more

[nnethercote]

These patches simplify the code a bit (fewer conversions) and also speed things up a bit (fewer with_interner calls).

r? @petrochenkov

[nnethercote]

I realize that this will need rebasing once #60740 lands.

[nnethercote]

@bors try

[bors]

⌛ Trying commit 67b969e with merge 8e849d8...

[bors]

☀️ Try build successful - checks-travis
Build commit: 8e849d8

[nnethercote]

@rust-timer build 8e849d8

[rust-timer]

Success: Queued 8e849d8 with parent 80e7cde, comparison URL.

[rust-timer]

Finished benchmarking try commit 8e849d8

[petrochenkov]

Any data structure using InternedString may use it for a reason, due to stable hashing.
IIRC, that's exactly the case for generic parameters, but I'll leave this to @michaelwoerister to review.

The (Local)InternedString::intern change LGTM.

r? @michaelwoerister

[michaelwoerister]

@petrochenkov, do we still use gensym? IIRC, we pushed for using InternedString in all cases where it was really only the raw string data that we care about. With "gensymmed" Symbols it's possible to distinguish between two identifiers with the same string contents, be there's no way to implement HashStable for that.

[petrochenkov]

Gensyms are still used.

Git blame says that InternedString in generic parameters was introduced in #49695 to fix #48923.
It looks like there's no specific regression test for that issue, so nothing in the test suite regressed from this PR.

[nnethercote]

It looks like there's no specific regression test for that issue

Sigh. My view is that Symbol is the fundamental type, and InternedString and LocalInternedString are alternatives that should only be used when necessary. (Using Symbol as much as possible is also good for performance.) I've been using the test suite as the guide for what conversions are possible, but if the test suite isn't comprehensive that makes things difficult :(

[nnethercote]

More generally: the difference between Symbol and InternedString is that the latter does PartialOrd/Ord and Hash on the chars, rather than the index.

• For PartialOrd/Ord I imagine the difference might affect some error messages.
• For Hash I am surprised that it is important, because if you need a stable hash you should use HashStable anyway, and Symbol could easily implement HashStable on the chars.

[petrochenkov]

Apparently the issue in #48923 was caused by query infrastructure using both 1) StableHash and 2) Hash/PartialEq.
If two things have the same StableHash, but different Hash/PartialEq then things break (that's exactly what happens with gensyms).

I don't know whether it's possible to use only StableHash and some different equality comparison function or trait instead of Hash and PartialEq in the query infrastructure.

[nnethercote]

Finally, I had to work out the difference between Symbol and InternedString myself, because there was no explanation in the code. (There is a comment about it now because I added one.) And none of the places that use InternedString have any explanation why they use it instead of Symbol, as far as I can tell. It is frustrating.

[michaelwoerister]

Yes, IIRC, I was not able to come up with a regression test for the fix in #49695. Adding an assertion that would catch the error early was the best I was able to do at the time:

rust/src/librustc/ty/query/plumbing.rs

Lines 526 to 535 in fbb6275

	// If the following assertion triggers, it can have two reasons:
	// 1. Something is wrong with DepNode creation, either here or
	// in DepGraph::try_mark_green()
	// 2. Two distinct query keys get mapped to the same DepNode
	// (see for example #48923)
	assert!(!self.dep_graph.dep_node_exists(&dep_node),
	"Forcing query with already existing DepNode.\n\
	- query-key: {:?}\n\
	- dep-node: {:?}",
	key, dep_node);

[nnethercote]

I have pushed new code that reverts the InternedString-to-Symbol change for ParamTy and related types. Some (but not all) of the performance benefit is lost as a result.

[nnethercote]

@bors try

[bors]

⌛ Trying commit 2175494 with merge ac67417...

[bors]

☀️ Try build successful - checks-travis
Build commit: ac67417

[nnethercote]

@rust-timer build ac67417

[rust-timer]

Success: Queued ac67417 with parent 4f53b5c, comparison URL.

[rust-timer]

Finished benchmarking try commit ac67417: comparison url

[nnethercote]

Still a clear perf improvement with the new code. It would be good to land this.

[petrochenkov]

r? @petrochenkov @bors r+

[bors]

📌 Commit 2175494 has been approved by petrochenkov

[nnethercote]

I rebased.

@bors r=petrochenkov

[bors]

📌 Commit c06cdbe has been approved by petrochenkov

[bors]

⌛ Testing commit c06cdbe with merge caef1e8...

[bors]

☀️ Test successful - checks-travis, status-appveyor
Approved by: petrochenkov
Pushing caef1e8 to master...