In-compiler invariants (check-invariants.patch, rustc-verify13) and invariant-sweep; findings 56-57

[zmaril]

Step 2 of leaning on mirth more: invariants checked inside rustc on every compilation.

The patch

docs/hunt/check-invariants.patch is applied last on the verify-reuse stack and is built into ~/mirth-work/rustc-verify13.

• With RUSTC_CHECK_INVARIANTS set, each violation prints rustc-invariant: <property>: <details> and compilation continues.
• RUSTC_CHECK_INVARIANTS=selftest also lists which checks ran. For Check reuse and report untracked reads inside rustc; findings 6 and 7 #18 it lists which queries' results are checked and which aren't.

# (properties.md)	Check	What upstream had
18	A query result contains no inference variables. Checked on each provider's typed result, dispatched on the value's type (autoref specialization over TypeVisitable, EarlyBinder, references, Option/Result, canonical responses, typeck node types, borrowck hidden types, clauses, impl headers, layouts). 99 of the ~200 query kinds a small program runs are checked.	nothing
14	A compile that emitted no error has no error types: typeck results (tainted, node types), type_of, fn_sig.	nothing
24	Local symbols are distinct from upstream crates' exported symbols.	within a session only (fatal SymbolAlreadyDefined)
15	Each metadata table entry is written once.	nothing
9	LLVM's ABI size of a lowered type equals the layout size. rustc's expensive layout sanity checks also run in release.	debug builds only
1	Argument lists fit the generics (debug_assert_args_compatible and the alias variant); violations are reported instead of raising a bug.	debug builds only
7	Encoded spans: lo <= hi, and lo is inside its file.	debug_assert! only

rustc/regen-patches.sh regenerates the new patch alongside the others; its embedded Python is replaced by sed. Checked by regenerating the stack:

• "0 files differ";
• verify-reuse.patch is identical apart from the length of the abbreviated index hashes.

The sweep

mirth-lab invariant-sweep compiles each standalone UI test with the variable set (codegen for tests that build). An ICE that happens only with the variable set counts as a finding.

Results:

• UI tests: 18,624 tests (7,408 compile, 11,198 fail as expected, 17 ICE without the checks too, 1 timeout). 315 have findings, all property 15. There is nothing for 18, 14, 24, 9, 1 or 7.
• gate-mutate: 12,000 mutants with the variable set gave no new ICE signatures.

Findings (docs/hunt/metadata-written-twice.md, facts only):

• 56: the crate root's module children are encoded twice in every library with a public item (encode_def_ids calls encode_info_for_mod(CRATE_DEF_ID), then reaches the root again in its loop).
• 57: coroutine layouts are encoded twice (encode_mir records mir_coroutine_witnesses inside if encode_opt and again at the end of the loop).
• Cost: low, in bytes only. A compiler with both removed saves 2,164 bytes of std's 8.0 MB .rmeta, 1,384 of core's, and 226 of a 16.5 KB async library's. With the fix, the invariant reports nothing.
• Versions: both are present since at least 1.80.0. Nothing was found upstream.

Findings 56–57 may need renumbering if other branches merge first.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QXiEXbESemwqMLYKaWLDbT