Diagnostic deduplication breaks with incr comp

[bjorn3]

I tried this code:

tests/ui/allocator/not-an-allocator.rs

#[global_allocator]
static A: usize = 0;
//~^ ERROR E0277
//~| ERROR E0277
//~| ERROR E0277
//~| ERROR E0277

fn main() {}

I expected to see this happen: Both without and with incr comp a single diagnostic.

Instead, this happened: Without incr comp a single diagnostic. With incr comp 4 duplicated diagnostics. With -Zdeduplicate-diagnostics=no in both cases 4 diagnostics are shown, but with incr comp it misses the message that the diagnostics were deduplicated.

Adding some debug prints shows that each diagnostic hashes to a different value with incr comp enabled, yet the debug print of the diagnostics are identical as is the DiagInner::keys() debug print.

Meta

rustc --version --verbose:

rustc 1.100.0-dev
binary: rustc
commit-hash: unknown
commit-date: unknown
host: x86_64-unknown-linux-gnu
release: 1.100.0-dev
LLVM version: 23.1.1

c999cef

[bjorn3]

Looks like the problem is that the representation of Span changes and the Hash impl of Span directly hashes this representation rather than the SpanData.

[nnethercote]

There are a bunch of inconsistencies in how Span and SpanData impl PartialEq/Eq/Hash/PartialOrd/Ord. Fixing those will fix this issue. I'll do this.

[nnethercote]

Looks like the problem is that the representation of Span changes and the Hash impl of Span directly hashes this representation rather than the SpanData.

I think the actual problem is the parent field. It's only used in incremental builds. It also doesn't get printed by the Debug impl. The proper fix will probably involve changing eq/hash to ignore that field, with the justification being that it's not really part of the spans identity the way lo/hi/ctxt are, that it's just some additional book-keeping.

[nnethercote]

Ignoring parent caused perf issues, so I will probably make a more targeted fix in dedup_hash. #163303 is a precursor that fixes a related problems with spans.

[nnethercote]

#163461 has the fix.