Skip to content

Slice modifies havoc reports an uninitialized read on uninitialized output with -Z uninit-checks #5181

Open

Description

@Tiramisu-Cake

When a function whose modifies target is a raw slice is replaced by its contract (stub_verified), the havoc of that slice fails an initialization check if the destination memory is not yet initialized. The same function body verifies, and the same contract with element-wise (thin pointer) targets also verifies.

use std::alloc::{alloc, dealloc, Layout};

// Safety: p points to a live, exclusively owned allocation of at least two bytes.
#[kani::modifies(std::ptr::slice_from_raw_parts_mut(p, 2))]
unsafe fn initialize(p: *mut u8) {
    p.add(0).write(1);
    p.add(1).write(2);
}

#[kani::proof_for_contract(initialize)]
fn check_body() {
    let layout = Layout::array::<u8>(2).unwrap();
    unsafe {
        let p = alloc(layout);
        if p.is_null() { return; }
        initialize(p);
        dealloc(p, layout);
    }
}

#[kani::proof]
#[kani::stub_verified(initialize)]
fn check_slice_havoc() {
    let layout = Layout::array::<u8>(2).unwrap();
    unsafe {
        let p = alloc(layout);
        if p.is_null() { return; }
        initialize(p);
        dealloc(p, layout);
    }
}
kani -Z function-contracts -Z uninit-checks -Z stubbing repro.rs --harness check_body --exact
kani -Z function-contracts -Z uninit-checks -Z stubbing repro.rs --harness check_slice_havoc --exact

Kani 0.68.0 at 0aac16ef62da9e9d5645782d458593e5d78143ba, nightly-2026-10-23, CBMC 6.11.0, x86_64 Linux.

Expected: both harnesses succeed. initialize only writes to the destination, and the modifies clause allows arbitrary new bytes there; nothing reads the old contents.

Actual: check_body succeeds. check_slice_havoc fails with exactly one failed check:

kani::internal::write_any_slice::<u8>.safety_check.1
Undefined Behavior: Reading from an uninitialized pointer of type `*mut [u8]`

Controls, all on the same revision and flags:

Variant check_body check_slice_havoc
uninitialized output, slice target (above) SUCCESS FAILURE (above)
output written with zeros before the call SUCCESS SUCCESS
uninitialized output, thin targets #[kani::modifies(p, p.wrapping_add(1))] SUCCESS SUCCESS

A separate harness that actually reads an uninitialized byte from the same kind of allocation (read_volatile) fails as expected with Reading from an uninitialized pointer of type *const u8, so the check itself works.

The slice havoc at kani_core/src/lib.rs:715–716 is (*slice).fill_with(T::any). That creates a &mut [T] over the existing elements and assigns to them, which reads them as initialized (and for types with destructors would also drop the old values). The thin-pointer havoc at lines 723–724 uses ptr::write, which is why the thin-target variant passes. Writing each element with ptr::write in the slice havoc as well would match the semantics of modifies for uninitialized output.

Possibly related but different: #3682 / #4749 (resolution of a doubled slice type before verification).

Disclosure: this issue was prepared with the help of AI assistants.

Activity

  1. added theissue type on Nov 17, 2026
  2. added
    [F] Spurious FailureIssues that cause Kani verification to fail despite the code being correct.
    on Nov 17, 2026
  3. added this to the Contracts milestone on Nov 17, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    [F] Spurious FailureIssues that cause Kani verification to fail despite the code being correct.

    Type

    Projects

    No projects

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions