Virtual Slicer

The virtual slicer is a RUSTC_WRAPPER that stubs unreachable functions at the MIR level before LLVM sees them. It does not modify your source files or your Cargo.toml.

What "unreachable" means

Starting from the binary's main function, cargo-slicer traces the call graph across all workspace crates. Any function that cannot be reached from main is replaced with an abort() body. LLVM skips compilation, optimisation, and code emission for those functions entirely.

The analysis is conservative: trait impls, generics, async functions, closures, and any function called through a function pointer are always kept.

How it plugs into cargo

RUSTC_WRAPPER=cargo_slicer_dispatch  ← stable binary, < 1 ms startup
    │
    └─ local workspace crate?
           CARGO_SLICER_DRIVER=cargo-slicer-rustc  ← nightly driver, ~300 ms startup
               └─ BFS reachability analysis
               └─ MIR stub replacement
               └─ CGU filtering (skip codegen units with only stubs)

The dispatch binary keeps the nightly driver out of the fast path. Registry crates (which change rarely and are cached) never pay the 300 ms driver load.

Cross-crate pre-analysis

For accurate reachability across crate boundaries, run pre-analysis before the build:

cargo-slicer pre-analyze

This uses syn-based parsing to build a call graph across all workspace crates in seconds, writing results to .slicer-cache/. The driver reads these files at build time instead of re-analysing from scratch for every crate.

Without pre-analysis the slicer falls back to conservative per-crate analysis, which still works but produces fewer stubs.

Tuning

What cannot be stubbed

The slicer never stubs:

• Trait impl associated functions (vtable entries)
• Generic functions (monomorphised at the call site)
• async fn and closures
• unsafe fn (unless CARGO_SLICER_RELAX_UNSAFE=1)
• Any function reachable through a function pointer

These constraints are intentional. Stubbing them would either cause linker errors or produce incorrect binaries.

Upstream proposal

The virtual slicing logic has been extracted into a proposed rustc patch behind a -Z dead-fn-elimination flag. If accepted upstream, the install story becomes:

RUSTFLAGS="-Z dead-fn-elimination" cargo +nightly build --release

No extra binary, no nightly ABI compatibility shims.

Auto-delegation in cargo_slicer_dispatch

The dispatch wrapper detects whether the real rustc already implements the flag (rustc -Zhelp | grep dead-fn-elimination, cached per process). When present, dispatch transparently appends -Zdead-fn-elimination and execs rustc directly — bypassing the userspace driver, the .slicer-cache, and the per-nightly ABI shim. The userspace path is still used as a fallback when the flag is absent.

# Patched stable: dispatch detects the flag and delegates automatically
CARGO_SLICER_VIRTUAL=1 RUSTC_WRAPPER=$(which cargo_slicer_dispatch) \
  cargo +dfe-stage1 build --release

# Same command on unpatched nightly: dispatch falls back to the userspace driver
CARGO_SLICER_VIRTUAL=1 RUSTC_WRAPPER=$(which cargo_slicer_dispatch) \
  cargo +nightly build --release

Override knobs: CARGO_SLICER_NO_UPSTREAM_FLAG=1 forces the userspace path even when the flag is available; CARGO_SLICER_FORCE_UPSTREAM_FLAG=1 skips the probe and assumes the flag is supported.

Review status (2026-04)

The patch has been through a nine-point review by @petrochenkov (V1–V9 in the response document). After applying the patches:

• Scope is now explicit: binary crates only — library crate types early-return (V1).
• Seeds are reachable_set ∪ entry_fn ∪ address_taken ∪ vtable_methods rather than a re-implementation of reachable_set (V3, V5a, V5b, V8).
• The cross-crate call graph has been removed; cross-crate effects come through reachable_set seeding (V4).
• requires_monomorphization replaces blanket generics > 0, so lifetime-only generics are eligible (V6b).
• Vestigial #[test]/#[bench], unsafe, and signature-walk checks have been deleted — their safety is now provided structurally by the seed set (V7, V8a, V8b).
• Net effect on dead_fn_elim.rs: ~419 → ~340 lines.

Empirical answer to V9 ("will it converge to reachable_set?"): on the ASE 2026 corpus sweep, reported separately by crate kind per V10/V11. Binary subset (n=65): 59/65 built under both legs, zero slicer-only failures, median 1.38× speedup. Library subset (userspace tool only — in-tree flag is a no-op there today): 2,393/2,538 built, zero slicer-only failures, userspace median 1.50×. The patched stage1 oracle on rust-1.90.0 eliminated 904 functions on ripgrep with binary output identical to the baseline.

Full point-by-point response (V1–V11): vadim-response-results.md.