Elodin Cranelift

Elodin Cranelift Backend

libs/cranelift-mlir/ compiles StableHLO MLIR to native code via Cranelift JIT — the default CPU backend for Elodin simulations. Deep internals live in ARCHITECTURE.md; profiling is in PERFORMANCE.md.

Default backend

Cranelift is the default (backend="cranelift" in WorldBuilder.run / .build). Override per-run:

ELODIN_BACKEND=jax-cpu python examples/<name>/main.py run   # XLA native, reference for correctness
ELODIN_BACKEND=jax-gpu python examples/<name>/main.py run   # XLA CUDA

Or in Python:

w.run(system, backend="jax-cpu")

Where things live

Everyday commands

Always run inside nix develop.

cargo test -p cranelift-mlir                    # all unit + golden tests
cargo test -p cranelift-mlir --test ops         # per-op only
cargo clippy -p cranelift-mlir -- -Dwarnings
cargo fmt -p cranelift-mlir -- --check

# Full simulation regression (requires `just install` first):
just install
source .venv/bin/activate
ELODIN_BACKEND=cranelift bash scripts/ci/regress.sh --all                                 # every example
ELODIN_BACKEND=cranelift bash scripts/ci/regress.sh ball examples/ball/main.py            # one example
ELODIN_BACKEND=cranelift bash scripts/ci/regress.sh --update ball examples/ball/main.py   # re-baseline after verifying correctness

Baselines live in scripts/ci/baseline/; tolerances in scripts/ci/baseline/tolerances.json.

Shared Large Constants

Large stablehlo.constant dense<"0x..."> blobs over 1 MB are transparently interned by libs/cranelift-mlir:

• Parser: src/parser.rs decodes large hex constants as raw bytes and interns them through src/const_cache.rs.
• Cache: content-addressed by element type, byte length, and raw little-endian bytes under ${ELODIN_CACHE_DIR:-~/.cache/elodin/const-cache}.
• IR: ConstantValue::DenseExternal is cheap to clone and carries the cached mapping handle.
• Lowering: src/lower.rs registers each mapping with JITBuilder::symbol and lowers it through declare_data(Linkage::Import), so constants are not copied into the JIT arena.
• Lifetime: CompiledModule must retain the Arc<CachedConst> handles for every imported constant. The JIT symbol stores raw pointers, so dropping the parsed IR module before live execution would otherwise leave dangling mmap pointers.
• Concurrent publishing: cache writes must never replace an already-published file. Use a temp file plus no-clobber publish so simultaneous first writers converge on one inode; otherwise one process can map a now-deleted inode and lose page-cache sharing.
• Campaigns: elodin monte-carlo run ... pins one campaign ELODIN_CACHE_DIR across all workers. Add --memory-probe only for memory-proof runs; it writes memory.json with shared-constant PSS evidence without a bespoke runner.

Relevant tests:

cargo test -p cranelift-mlir --test large_constant_cache
cargo test -p cranelift-mlir --test imported_data_arena -- --ignored --nocapture
cargo test -p cranelift-mlir large_external_constant_8d_dynamic_slice

Customer-style checkpoint repro:

ELODIN_CRANELIFT_DEBUG_DIR=/tmp/dbg \
  cargo test -p cranelift-mlir --test checkpoint_test --release verify_checkpoint -- --ignored --nocapture

# To hold two verifier processes for mmap/PSS sampling:
ELODIN_CACHE_DIR=/tmp/elodin-const-cache-proof \
ELODIN_CRANELIFT_DEBUG_DIR=/tmp/dbg \
ELODIN_CRANELIFT_CHECKPOINT_HOLD_AFTER_TICK_MS=120000 \
  cargo test -p cranelift-mlir --test checkpoint_test --release verify_checkpoint -- --ignored --nocapture

Adding a new op

1. IR: add Instruction variant in src/ir.rs
2. Parser: add arm in parse_op() in src/parser.rs
3. Lowering (scalar): add match arm in lower_instruction() in src/lower.rs
4. Lowering (pointer-ABI): add match arm in lower_instruction_mem() in src/lower.rs
5. Runtime (if N-D): add tensor_<op>_f64 in src/tensor_rt.rs, register in TensorRtIds
6. Tests: golden-value tests in tests/ops.rs exercising both run_mlir and run_mlir_mem
7. cargo test -p cranelift-mlir

The dual-ABI / SIMD / tensor-runtime design constraints each step has to satisfy are documented in ARCHITECTURE.md.

Adding a new simulation example

1. Dump MLIR: ELODIN_BACKEND=cranelift ELODIN_CRANELIFT_DEBUG_DIR=/tmp/dbg python examples/<name>/main.py run
2. Catalog ops: python3 libs/cranelift-mlir/scripts/catalog_ops.py /tmp/dbg/stablehlo.mlir (safe on multi-GB files — do NOT grep them)
3. Copy the MLIR into testdata/, create an e2e test
4. Implement any missing ops (see "Adding a new op")
5. Regression: ELODIN_BACKEND=cranelift bash scripts/ci/regress.sh <name> examples/<name>/main.py

Debugging

Simulation produces wrong values

The tick checkpoint diagnostic tool is the workhorse. Full usage and MLIR-bisection workflow in ARCHITECTURE.md — Checkpoint Diagnostic Tool section. Quick commands:

# Capture XLA reference + Cranelift outputs for every tick input:
ELODIN_BACKEND=cranelift ELODIN_CRANELIFT_DEBUG_DIR=/tmp/ckpt \
  bash scripts/ci/regress.sh <example> examples/<example>/main.py

# Compare, per output, element-by-element:
ELODIN_CRANELIFT_DEBUG_DIR=/tmp/ckpt \
  cargo test -p cranelift-mlir --test checkpoint_test --release -- --ignored --nocapture

Once the diverging output is identified, reduce it to a minimal tests/ops.rs reproducer before fixing.

Simulation crashes (segfault / abort)

• Try --release first: some JIT paths trip ptr::copy_nonoverlapping debug-mode UB checks.
• Stack overflow on complex sims: the checkpoint test pins a 64 MB thread stack; mirror this if reproducing outside the test.
• NULL pointer in JIT code is almost always a cross-ABI marshaling bug.

Compiler says "unsupported instruction" or "unsupported custom_call target"

Follow "Adding a new op".

Environment variables

Only two matter day-to-day:

• ELODIN_BACKEND — cranelift (default), jax-cpu, jax-gpu.
• ELODIN_CRANELIFT_DEBUG_DIR=<dir> — single flag for every diagnostic (profile probes, op-category sampling, tick waveform, instr/fold reports, inliner and slot-pool traces, MLIR dump, first-tick XLA reference checkpoint). Files land flat under <dir>. Zero overhead when unset.

Full outputs and reading guide: PERFORMANCE.md.

Further reading

• libs/cranelift-mlir/ARCHITECTURE.md — compilation pipeline, dual ABI, SIMD (LaneRepr), JIT memory layout, tensor runtime, LAPACK via faer, gather patterns, while-loop scoping, checkpoint tool, testing strategy, opportunities.
• libs/cranelift-mlir/PERFORMANCE.md — ELODIN_CRANELIFT_DEBUG_DIR outputs, profile report fields, diff + waveform scripts, Tracy workflow.
• libs/cranelift-mlir/README.md — crate-level quick start.