debug-check (check 20): gdb round trip; findings 33-34 (gdb printers)

[zmaril]

Check 20 of docs/checks.md, the debugger round trip, as mirth-lab debug-check.

What it does

Each seed generates a program that builds known values. gdb, loaded with rust-gdb's pretty-printers, stops at a breakpoint and prints every local, plus *v for boxes and references. Each command runs as its own -ex, under a timeout. The output is parsed into a tree and compared structurally with the values the generator built, at -Copt-level 0, 1 and 2. At levels 1 and 2 an optimized-out value is accepted; a wrong value never is.

The values cover:

• integers of every width, and floats built from bit patterns (signed zeros, infinities, NaN payloads, subnormals)
• chars and strings with escapes and non-ASCII
• arrays, slices, Vec, VecDeque built to wrap around, and the four maps and sets
• Option and Result, including niche-optimized cases
• Box, Rc and Arc with known strong and weak counts, Weak, RefCell with live Ref/RefMut guards, Cell
• generated structs, tuple structs, unit structs, enums (with data, and fieldless with a repr and explicit discriminants), and unions
• references, PhantomData, OsString, Box<str>, Box<[T]>, Rc<str>, Arc<[T]>, PathBuf
• large vectors, and an Rc cycle, which must terminate

Results: 4,000 programs × 3 opt levels

Two findings, both in src/etc/gdb_providers.py. Both reproduce from 1.80.0 through nightly-2026-10-06, and no existing issue turned up for either.

• Finding 33:
  • The Vec, VecDeque and slice printers raise exceptions on zero-sized elements: "Cannot perform pointer math on incomplete type", and % cap where the stored capacity is 0.
  • BTreeMap and BTreeSet show every zero-sized key or value as (), whatever its type.
• Finding 34: the Ref/RefMut printer fails on every guard. It dereferences the NonNull field without unwrapping it, so neither the borrowed value nor the borrow count is shown. No gdb test prints a Ref.

Four more classes are gdb's own behaviour. They are labelled as expected, with the facts in docs/checks.md:

• gdb 15 can't read a 128-bit enum discriminant.
• An array of zero-sized elements prints as an address.
• A struct ending in a zero-sized field is taken for an unsized one, which breaks Box<[Arc<T>]>.
• An optimized-out pointer to an enum is read through at address 0.

Nothing else showed up: no other wrong value, no hang, no gdb crash.

Facts are in docs/hunt/gdb-printers.md, and the reductions in docs/hunt/tests/debug-check/. rustc/big-run.sh now runs this check too.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QXiEXbESemwqMLYKaWLDbT

[zmaril]

Merged into main together with #33–#38; renumbered at merge so finding numbers are unique: findings 33–34 are now 48–49 (docs/hunt.md).