just check cannot pass inside the recommended nix develop shell: the probe pins llc by SHA-256, nixpkgs rewrites it #1211

Activity

1. nihalpasham commented on Sep 3, 2026

   

   nihalpasham

   on Sep 3, 2026

   Last edited by nihalpasham

   Collaborator

   More actions

   Confirmed on main: the probe hashes the whole llc file (probe.rs validate_backend_identity), and the rustup nightly-2026-08-28 llc here is your 2e5bc688 byte for byte. The bytes are the wrong identity, and not only for Nix.

   • That llc is a 281 KB driver. The NVPTX lowering lives in libLLVM.so.23.1-rust-1.100.0-nightly (208 MB), which nothing hashes. So the SHA records which binary produced the evidence; the real check is already the emitted PTX matching the catalog's expected instruction on every probe.
   • patchelf (the Nix step that rewrites where a binary finds its loader and libraries) leaves the code untouched. Same compiler, new hash.
   • Plain rustup on aarch64 Linux (the flake lists it) hits the same wall: same LLVM, different bytes.
   • Version string alone is too coarse: every nightly in the 1.100.0 cycle prints 23.1.0-rust-1.100.0-nightly, and Rust bumps its LLVM inside a cycle.

   enforced identity                  rustup x86_64  nix   aarch64  tells nightlies apart
   llc file SHA-256 (today)           ok             fail  fail     unreliably (driver bytes only)
   llc --version alone                ok             ok    ok       no
   llc --version + rustc commit-hash  ok             ok    ok       yes
   

   My lean, as one PR:

   • llvm_nvptx evidence gains backend_toolchain_commit = rustc -vV commit-hash; the resolver requires it for that kind (missing = loud failure, no silent downgrade). backend_sha256 stays recorded and printed as provenance.
   • validate_backend_identity checks version + commit in the default run (the one that validates committed evidence; --llc comparison runs stay unchecked as today). rust_toolchain_llc already parses rustc -vV, so this is one more line read.
   • tests: same commit + different bytes passes; different commit fails; different version fails; evidence without the field fails at resolve time.
   • before/after: the probe's first output line under rustup and under nix develop. catalog.json changes, so re-run bash scripts/check-book-catalog-stamp.sh after the final commit, and update the two comments that describe the gate (Justfile check-intrinsics, unit-tests.yml).
   • also fine: a structured [rust_toolchain] commit_hash in intrinsics/upstream.lock (the resolver already checks evidence against that file) compared by the probe, one field in one file, no catalog.json churn. Pick whichever you prefer and say so in the PR.
   • no Nix CI lane, per fix(nix): advance the rust-overlay pin to the toolchain's nightly, and gate the flake in CI #1210. Criteria: https://discord.com/channels/1515530041767759993/1515561621156663487/1540970771323551775

   I do not run Nix, so the Nix-side pass is yours to show. Of your three open asks with me, this one goes first (#1189 and #1213 follow, one series at a time). Do you want to send this PR?

   👍React with 👍1Reacted by Sam
2. almightychang commented on Sep 3, 2026

   

   almightychang

   on Sep 3, 2026

   ContributorAuthor

   More actions

   Yes. Sent as #1227, taking the upstream.lock placement: a required [rust_toolchain] commit_hash compared by the probe, evidence and catalog.json untouched (backend_sha256 stays as printed provenance). The lock already pins what the evidence was built from and its [llvm] provenance line already named this commit in prose; rust-toolchain.toml pins one channel for the whole repository and the probe runs every record under it, so the commit is a repository-level fact rather than a per-evidence field. The PR body has the before/after first lines under rustup and nix develop, and just check run as one command inside nix develop passes.
3. nihalpasham

   closed this as completedin #1227on Sep 6, 2026
4. nihalpasham

   added a commit that references this issue on Sep 6, 2026

   fix(intrinsics): pin the probe toolchain by rustc commit, not llc byt…

   ...

   This issue will close once commit ef48b17 is merged into the 'main' branch.

   ef48b17