Skip to content

Fix SMT pointer type dependency discovery - #9159

Open
chaizhenhua wants to merge 1 commit into
diffblue:developfrom
chaizhenhua:fix/smt-pointer-type-dependencies-20260910-upstream
Open

chaizhenhua wants to merge 1 commit into
diffblue:developfrom
chaizhenhua:fix/smt-pointer-type-dependencies-20260910-upstream

Conversation

@chaizhenhua

Copy link
Copy Markdown

Recursive pointer types can make SMT conversion use a struct-tag alias
before registration and throw map::at. Keep pointer targets in the existing
type visitor, but discover their expression dependencies without declaring
unneeded pointee sorts. Visiting a tag in this mode does not suppress a later
by-value visit, and VLA size expressions retain their full dependencies.

This change only fixes pointer type dependency discovery. It makes no solver
performance claim and does not change response batching, binders, or unknown
expression handling.

Validation includes a failing baseline regression, 48 SMT unit cases, the
complete unit suite, cbmc-CORE, smt2_solver-CORE, seven existing pointer/VLA
regressions through Z3, and both recursive alias entry orders producing the
expected counterexample. Existing CORE exclusions remain unchanged.

  • Each commit message has a non-empty body explaining the change.
  • Added non-obvious types and procedures are documented.
  • User guide: not applicable; this restores existing pointer semantics.
  • Unit regressions are included in the bug-fix commit.
  • Performance data: not applicable; no performance claim.
  • Restricted to one bug fix.
  • No unrelated whitespace changes.

Pointers are encoded as bitvectors, but type discovery currently declares
their pointee datatypes. A recursive Leaf -> Internal pointer -> Leaf value
type can consequently use a struct-tag alias before it has been registered,
throwing map::at during SMT conversion.

Separate sort and expression dependencies in the existing type visitor.
Pointer targets only require expression discovery, including implicit VLA
sizes. Track visited tags together with their discovery mode so a pointer
visit cannot suppress a later by-value visit. Size expressions continue to
discover the full dependencies of their operands.

Add focused regressions for both recursive entry orders, mixed pointer/value
visits, nested VLA sizes, complex/state sorts, bitvector pointer arithmetic,
and recursive unions. The recursive pointer test fails on the unmodified
develop implementation and passes with this change.

Validation: clang-format 18 on changed lines; SMT unit tests (48 cases,
237 assertions); all unit tests (620 cases, 19494 assertions) and expected
failures; cbmc-CORE (1128 passed, 67 skipped); smt2_solver-CORE (42 passed,
2 skipped); seven existing pointer/VLA regressions using Z3; both recursive
alias entry orders produce the expected counterexample through Z3.
feliperodri pushed a commit to tautschnig/kani that referenced this pull request Nov 20, 2026
…ecking#5203)

## Summary

Lower core's native comparison and pattern assertion failures through
Kani's existing GOTO panic model. On the current PR base, this reduces
the GOTO program and reports failures at the user's assertion with a
fixed diagnostic. A CBMC SMT conversion failure in the Challenge 18
`ChunksExactMut::nth_back` harness on its pinned toolchain motivated the
change.

## Additional implementation notes

- Match `assert_failed`, `assert_failed_inner`, and
`assert_matches_failed` in the same core module as `panic_fmt`. Apply
the rewrite before reachability in GOTO, which provides the `PanicStub`
hook.
- Check the enabled user-stub map once per codegen unit. An active stub
for any of the three assertion entries disables their lowering for that
unit. This preserves an `assert_failed_inner` stub when the assertion
calls an outer entry first. The formatter helper guard also uses the
enabled stub map.
- `RustcIntrinsicsPass` prepends a static diagnostic and keeps the
original MIR operands in the rewritten call to `panic_stub`. The GOTO
hook reads operand 0 as the message and evaluates the other operands in
order for their safety checks before emitting the assertion property.
- Clear the original generic arguments when redirecting to the
non-generic `panic_stub`. This also applies to the existing rewrite of
generic panic functions such as `begin_panic::<M>`.
- With the default assertion configuration, `assert_ne!` and
`debug_assert_eq!` report the comparison expression. `core::assert_eq!`
and `assert_matches!` report the new fixed diagnostics at the assertion
lines.

## Diagnostics

Rewritten comparison assertions report `assertion failed`. Rewritten
pattern assertions report `pattern assertion failed`. These fixed
diagnostics also apply when a literal custom message is supplied. The
base reports the runtime-formatting placeholder from `core::panicking`
for the same native assertions.

The model retains the caller's construction of `fmt::Arguments` and
evaluates the original operands before emitting the panic property and
ending the path. Formatting callbacks such as `Debug::fmt` are outside
the modeled panic path.

## Validation

Validated revision `dc85d0d` against upstream base `c25a57bc` with
`nightly-2026-11-18`, CBMC 6.11.0, and CVC5 1.3.0.

- The diagnostics regression observes the user's own messages for stubs
of `assert_failed` and `assert_failed_inner`. Separate `assert_eq!` and
`assert_matches!` harnesses check the inner stub with distinct expected
messages.
- The defensive inactive-stub regression uses raw compiler flags to
allow the annotation while leaving the stub pass disabled. The explicit
MIR regression checks that a null pointer dereference in a projected
operand is reported in a `should_panic` harness.
- The `Assert`, `Stubbing`, and `FmtDerive` groups report 59 passes with
two existing fixme cases ignored. All 53 expected-output cases selected
by `assert panic stub` and all 25 LLBC cases pass.

### Validation clarification

The standalone CVC5 harness passes on upstream base `c25a57bc` and does
not reproduce the original Challenge 18 SMT conversion failure there. On
that base the demonstrated benefits are a smaller GOTO program and
improved assertion diagnostics. `assert_failed_native.rs` and
`native-assert-projected-argument.rs` pass on both the base and the
revision as regression guards for assertion, panic, and safety-check
behavior.

The original Challenge 18 end-to-end result uses its project pin. The PR
base's nightly stops while compiling that project's `core`, before the
Challenge 18 harness can run.

### GOTO program size

Set `main_checkout` and `candidate_checkout` to the absolute paths of
the independently built base and PR checkouts, then run the script with
Bash. It runs the same fixture for both harnesses in fresh temporary
directories.

```sh
set -eu
KANI_MAIN="$main_checkout/scripts/kani"
KANI_CANDIDATE="$candidate_checkout/scripts/kani"
fixture="$candidate_checkout/tests/kani/Assert/assert_failed_native.rs"

for KANI in "$KANI_MAIN" "$KANI_CANDIDATE"; do
  for harness in check_failed_equality check_failed_pattern; do
    (
      run_dir="$(mktemp -d)"
      trap 'rm -rf -- "$run_dir"' EXIT
      cp "$fixture" "$run_dir/native.rs"
      cd "$run_dir"

      printf '%s %s\n' "$KANI" "$harness"
      "$KANI" native.rs -Z unstable-options --no-assert-overrides --exact \
        --harness "$harness" --only-codegen --keep-temps
      wc -c ./*.symtab.out
      python3 -c 'import json, sys; print(len(json.load(open(sys.argv[1]))))' ./*.pretty_name_map.json
    )
  done
done
```

| Harness | Base symbol table bytes | Revised symbol table bytes | Base
pretty-name entries | Revised pretty-name entries |
| --- | ---: | ---: | ---: | ---: |
| `check_failed_equality` | 2,210,210 | 24,478 | 7,424 | 160 |
| `check_failed_pattern` | 2,207,361 | 17,297 | 7,413 | 133 |

Extends the panic lowering from model-checking#4169. The pinned-toolchain SMT failure
is related to diffblue/cbmc#9159.

By submitting this pull request, I confirm that my contribution is made
under the terms of the Apache 2.0 and MIT licenses.

---------

Co-authored-by: Mingzhi Gao <gmz563527968@126.com>
Co-authored-by: v3risec <272978695+v3risec@users.noreply.github.com>

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant