Repository navigation
Description
Summary
Type::vector in cprover_bindings asserts its element type is numeric, so declaring any SIMD vector whose element is a pointer — Simd<*const T, N> / Simd<*mut T, N>, the portable_simd pointer-vector types — crashes the compiler:
thread 'rustc' panicked at cprover_bindings/src/goto_program/typ.rs:1283:9:
assertion failed: typ.is_numeric()
[Kani] current codegen item: declare_function: std::simd::Simd::<*const i32, 2>::splat
Reproducer
#![feature(portable_simd)]
use std::simd::Simd;
#[kani::proof]
fn check_ptr_simd() {
let v: Simd<*const i32, 2> = Simd::splat(core::ptr::null());
assert!(v.as_array().len() == 2);
}$ kani main.rs --only-codegen
thread 'rustc' panicked at cprover_bindings/src/goto_program/typ.rs:1283:9:
assertion failed: typ.is_numeric()Reproduced on 783e7f7 (current main) with nightly-2026-08-21. No autoharness needed — a plain #[kani::proof] is enough, and it fails at codegen.
Root cause
// cprover_bindings/src/goto_program/typ.rs:1281
pub fn vector(typ: Type, size: u64) -> Self {
assert!(typ.is_numeric());
Type::Vector { typ: Box::new(typ), size }
}with is_numeric() == is_floating_point() || is_integer(). A pointer element type is neither, so the assertion fires. portable_simd has supported pointer vectors (SimdConstPtr / SimdMutPtr) for a while, so this is a categorical gap rather than a corner case: every Simd<*const T, N> is unreachable for Kani today.
Relation to #2252
#2252 reports the same assertion (at the then-current line 1201, Kani 0.22.0) but from a different trigger — scatter_select_unchecked with SIMD masks and index vectors. Filing separately because the trigger and the fix surface here are different and much narrower: this one is purely "pointer is not a legal CBMC vector element type", with a 7-line reproducer and no gather/scatter involved. Happy to have it folded into #2252 if maintainers prefer.
Possible directions
- Model a pointer-element vector as a vector of the pointer-width integer, casting at element access. That keeps CBMC's numeric-element invariant and matches how the addresses are actually used (
SimdConstPtr::addr,wrapping_offset). - Or, if that is too invasive, turn the
assert!into a graceful unsupported-construct report so the user gets a verification failure naming the construct instead of a compiler crash. That alone would be a strict improvement — anassert!incprover_bindingssurfaces as "Kani unexpectedly panicked" with a request to file a bug, which is misleading for a known unsupported feature.
Why this matters for the standard library
Found while testing whether #4848 unblocks the autoharness jobs on model-checking/verify-rust-std#687 (the bump to nightly-2026-08-21). With #4848 applied, autoharness over library/ gets far enough to reach this, and core then fails codegen on:
declare_function: core::mem::transmute_copy::<core::core_simd::vector::Simd<usize, 2>, core::core_simd::vector::Simd<*const i32, 2>>
So this blocks the unfiltered autoharness path over the standard library (the Kani Metrics and Kani List jobs there, which pass no --include-pattern). The pattern-filtered autoharness verification jobs are unaffected, since no harness is generated for transmute_copy.
References
- Same assertion, different trigger: ICE: 'assertion failed: typ.is_numeric()', cprover_bindings/src/goto_program/typ.rs:1201:9 #2252
- Surfaced behind Fetch a candidate's MIR with the provider that serves its body kind #4848 / Autoharness ICEs on
#[rustc_comptime]intrinsics: optimized_mir called on a const-context def #4839 - Blocks the unfiltered autoharness path in Merge subtree update for toolchain nightly-2026-09-25 verify-rust-std#687
Activity
Model SIMD pointer lanes as integers of the same width (model-checkin…