Make `just verify` actually invoke Why3 + alt-ergo + z3 against every
Pearlite VC, and fix the mutant + miri pipelines along the way.
Verification crate
- Drop the const-generic CAP types for EventQueue / TimerTable.
Mirror with std::vec::Vec, whose creusot-std specs carry the
proof burden. Capacity behavior is exercised by the runtime's
own unit tests + correspondence tests.
- View impls are #[logic(open)] over a pub field so contracts on
push/pop/etc. can see the link to the inner Vec's @ view.
- 14/14 VCs discharge: EventQueue {new, push, pop, is_empty, len,
clear, Default}; TimerTable {new, len, is_empty, Default};
Entry derives.
- `cargo creusot prove --no-cache` always runs the SMT solvers;
`time` shows ~7s CPU on ~1.2s wall-clock = real parallel work.
Tooling
- verification/mise.toml pins nightly-2025-11-13 + rust-src,
rustc-dev, llvm-tools-preview, rustfmt; declares opam, python,
uv, and the project-local .venv. Inherits root mise.toml's just
+ cargo-* tools.
- just verify clones creusot v0.9.0 to ~/.local/share/hsmc-verify/
and runs the canonical ./INSTALL — that's the only way the
bundled prelude-generator + creusot-rustc actually build.
- just verify --no-cache forces SMT re-runs every time.
- Drop bespoke pin/install of cargo-creusot from mise.toml; the
cargo backend goes through cargo-binstall which can't pin a
multi-bin git workspace at a tag.
Mutants — was missing every JournalSink mutant
- Recipe runs with --features tokio,journal so journal_* and det_*
tests (gated on both) actually execute.
- Add clear_journal_empties_the_journal test in journal_smoke.rs
to kill the JournalSink::clear with-() mutant.
- Remove dead JournalSink::len + is_empty (called nowhere; only
existed as mutation-bait).
Result: 0 missed, 42 caught, 211 unviable (proc-macro emitted
code that mutants can't standalone-compile).
Miri — was hitting the trybuild glob isolation issue
- --exclude hsmc-macros (trybuild ui test isn't UB-relevant)
- --exclude hsmc-verification (different pinned toolchain)
- 66 tests pass under Miri, 0 UB.
Gitignore
- /verif/, /why3find.json, /.why3find/, **/.venv/ — all
regenerated by `just verify`. Repo stays clean of per-machine
prover timings + venv artifacts.
Misc
- Fix root-state targetability: `on(Trig) => <ChartName>;` resolves
to root id 0 in both ir.rs::resolve_transitions and
validate.rs's separate name_map.
- `set shell` and `set script-interpreter` route every justfile
recipe through `mise exec`, so all subshells see the right
Rust + opam + cargo-* tools regardless of how `just` was
invoked. No shebangs anywhere.
123 lines
3.3 KiB
Rust
123 lines
3.3 KiB
Rust
//! Correspondence test: the verified [`EventQueue`] mirror behaves
|
|
//! identically to the real runtime [`hsmc::__private::EventQueue`]
|
|
//! for the operations we've verified.
|
|
//!
|
|
//! The mirror is unbounded (uses `Vec<E>`) where the runtime is
|
|
//! capacity-bounded (uses `heapless::Deque<E, N>`). This test ignores
|
|
//! capacity exhaustion — that path is exercised by the runtime's own
|
|
//! unit tests in `hsmc/src/lib.rs::private_internals`. What this test
|
|
//! pins down is the FIFO discipline + length tracking + clear, which
|
|
//! ARE proven by Creusot for the mirror.
|
|
|
|
use hsmc::__private::EventQueue as RealQueue;
|
|
use hsmc_verification::event_queue::EventQueue as MirrorQueue;
|
|
|
|
#[derive(Debug, Clone, Copy)]
|
|
enum Op {
|
|
Push(u32),
|
|
Pop,
|
|
Clear,
|
|
IsEmpty,
|
|
Len,
|
|
}
|
|
|
|
/// Drive both queues with the same op sequence and assert observable
|
|
/// behavior matches at every step.
|
|
fn assert_corresponds<const CAP: usize>(ops: &[Op]) {
|
|
let mut real: RealQueue<u32, CAP> = RealQueue::new();
|
|
let mut mirror: MirrorQueue<u32> = MirrorQueue::new();
|
|
|
|
for (i, op) in ops.iter().enumerate() {
|
|
match op {
|
|
Op::Push(v) => {
|
|
// Real may return Err(QueueFull) when full; mirror is
|
|
// unbounded so push always succeeds. Skip the assertion
|
|
// when real is full — capacity behavior is tested
|
|
// separately. Otherwise both succeed; lengths track.
|
|
let r = real.push(*v);
|
|
if r.is_ok() {
|
|
mirror.push(*v);
|
|
}
|
|
}
|
|
Op::Pop => {
|
|
let r = real.pop();
|
|
let m = mirror.pop();
|
|
assert_eq!(r, m, "step {i} Pop: real {r:?} vs mirror {m:?}");
|
|
}
|
|
Op::Clear => {
|
|
real.clear();
|
|
mirror.clear();
|
|
}
|
|
Op::IsEmpty => {
|
|
assert_eq!(
|
|
real.is_empty(),
|
|
mirror.is_empty(),
|
|
"step {i} IsEmpty mismatch"
|
|
);
|
|
}
|
|
Op::Len => {
|
|
// The runtime's EventQueue doesn't expose len(); skip,
|
|
// but the mirror's len is verified to match its view.
|
|
let _ = mirror.len();
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
#[test]
|
|
fn corr_basic_push_pop() {
|
|
assert_corresponds::<8>(&[
|
|
Op::IsEmpty,
|
|
Op::Push(1),
|
|
Op::Push(2),
|
|
Op::Push(3),
|
|
Op::Len,
|
|
Op::IsEmpty,
|
|
Op::Pop,
|
|
Op::Pop,
|
|
Op::Pop,
|
|
Op::IsEmpty,
|
|
Op::Pop,
|
|
]);
|
|
}
|
|
|
|
#[test]
|
|
fn corr_clear_empties() {
|
|
assert_corresponds::<8>(&[
|
|
Op::Push(1),
|
|
Op::Push(2),
|
|
Op::Push(3),
|
|
Op::Clear,
|
|
Op::IsEmpty,
|
|
Op::Pop,
|
|
Op::Push(99),
|
|
Op::Pop,
|
|
]);
|
|
}
|
|
|
|
#[test]
|
|
fn corr_alternating() {
|
|
let mut ops = Vec::new();
|
|
for i in 0..50u32 {
|
|
ops.push(Op::Push(i));
|
|
ops.push(Op::Push(i + 100));
|
|
ops.push(Op::Pop);
|
|
}
|
|
assert_corresponds::<8>(&ops);
|
|
}
|
|
|
|
#[test]
|
|
fn corr_fill_drain_repeat() {
|
|
let mut ops = Vec::new();
|
|
for round in 0..5 {
|
|
for i in 0..8u32 {
|
|
ops.push(Op::Push(round * 100 + i));
|
|
}
|
|
for _ in 0..8 {
|
|
ops.push(Op::Pop);
|
|
}
|
|
}
|
|
ops.push(Op::IsEmpty);
|
|
assert_corresponds::<8>(&ops);
|
|
}
|