vafinitcache_examples

Init cache-slot typing (Enhancement-326)

A mir::Value is a bare u32 index. sim_back's build_init_itern inserts init-function values into collapse_implicit (it stores val_map[&val]), but build_init_cache tested main-function values against that same set. Comparing indices across two independent namespaces does not fail loudly — it silently succeeds whenever the two counters happen to collide.

On a collision the slot's recorded hir::Type was wrong: an f64 cache slot was stamped Type::Bool, which lowers to i8. The store side then emitted trunc double .. to i8, and the noise loader read the slot back as a raw i8 straight into fmul i8 %x, double %y. The assertions build's LLVM verifier rejected that IR; the shipped release carried it into LLVM and died with EXC_BAD_ACCESS inside DoubleAPFloat::multiply — a SIGSEGV, no diagnostic.

Because the trigger is a numeric coincidence between two value counters, the original 917-byte fuzz reproducer resisted line-by-line minimization: deleting even a dead statement shifted an index and made the crash vanish. That is why initcache.va keeps two apparently pointless filler statements.

Ingredients (each verified essential by ablation)

1. a noise source whose power is an op-independent, non-parameter expression — it must land in a cache slot (EvalOutput::Cache); a bare parameter takes EvalOutput::Param and a literal takes EvalOutput::Const, both of which dodge the bug;
2. an analog operator that creates an implicit equation (idt, idtmod, or a variable-assigned ddt) — this produces the Bool-typed collapse flag;
3. that operator under a parameter-only, non-constant condition, so the flag is a runtime phi[FALSE,TRUE] rather than a folded constant;
4. the noise live in a contribution.

casex, hypot, cosh, nested loops and $temperature — all present in the original fuzz output — were ablated away as irrelevant.

Files

• initcache.va — the minimized crash reproducer (250 bytes). Compiling it with the pre-fix binary is a SIGSEGV.
• initcache_noise.va — the same ingredients with the noise left observable, so the value that flowed through the mis-typed slot can be checked numerically.
• verify_vafinitcache.py — 4 checks: it compiles, the module is valid IR, it simulates to a finite operating point, and the cache-slot noise power reads back as a real (onoise_total matches the closed form sqrt(P·R²·BW) to 0.02 %, which it cannot if the power is read as an i8).

Run

python3 verify_vafinitcache.py