EthProofs worker: the optimized guest (in-guest JUMPDEST table, zkVM syscalls)

[eigmax]

Stacked on #23. This brings the optimized guest into the Ziren 2.0 EthProofs worker.

Why

The Fleet worker (#23 and its follow-ups) branched from the guest optimization line in April and still ships the older guest:

• Soundness. ClientExecutorInput.bytecodes carries revm Bytecode, including the host's JumpTable. The guest checks the code hash with hash_slow, which hashes only the code bytes. A prover can therefore mark a 0x5b inside PUSH data as a valid jump target, or a real JUMPDEST as invalid, under the correct code hash. The optimized guest takes the raw code (RawCode) and derives the jump table itself.
• Cost. The older guest lacks the zkVM syscalls for 256-bit arithmetic and the curves, the arena tries, and the revm 32-bit fast paths.

What is in it

Two commits on top of #23:

• Prepare consecutive EthProofs blocks in parallel: the worker change the current deployment runs.

• guest: the optimised guest line in the Ziren 2.0 EthProofs worker: the guest optimization line (feat/guest-syscalls, whose commits are public there) squashed into the worker, plus two further guest changes:

  • TrieDB and the post-state root pass reuse keccak(address) and keccak(slot) inside the zkVM.
  • The derived header takes the logs bloom that verify_receipts already compared with the receipts, instead of building every receipt bloom a second time. Optimism still accumulates it.

  Its message describes how the two lines were reconciled. In short:

  • bin/eth-proofs, the root manifest and the lock come from the worker. The git-locked Ziren 2.0 graph is kept.
  • The guest, mpt and the witness RPC database come from the optimization line.
  • The worker's storage rules are kept: untouched storage roots are preserved and empty accounts are removed (already in ArenaState::update). Omitted storage of an absent account, or of one with an authenticated empty root, reads as zero, and any other missing trie is an error; that rule is now ArenaState::storage, with its three tests.
  • Verified bytecode recovery is the optimization line's complete_account_code.

Results

Same blocks, identical committed block hashes, every proof verified:

Block	Deployed guest	This guest
26184581	504.6M cycles, 73.97 s	278.7M cycles, 53.73 s
26184582	557.1M cycles, 73.64 s	269.6M cycles, 53.40 s
26184583	89.0M cycles, 20.61 s	52.7M cycles, 18.01 s

The times are single-card prove times. Running as the production worker on 8 cards, the first 24 blocks after the switch averaged 7.97 s per block (median 7.82 s, p90 9.89 s). The last 100 blocks of the previous guest averaged 11.98 s (median 10.92 s, p90 16.51 s).

• Guest ELF SHA-256: 941629f6acb9a229d810a7cb32dbd0c5d57ad050bf1ad43b5edd7d2a21454e1b
• Program vk: 0x00e87eb3c2f0e827d29811dddb265bf03653e65acc08ca48e88536c83ed9de79
• cargo test -p mpt: 24 passed.

Further guest changes

Four more commits make the guest cheaper and close one more input hole:

• perf(guest): bind the chain to mainnet, gate the precompile tracker. The proof commits only the block hash, so the chain spec and the beneficiary override in the input came from the prover unchecked. The guest now accepts mainnet inputs only, with no custom beneficiary. The per-precompile cycle-tracker wrapper, which turned the precompile set into a map probed on every call, is now behind a precompile-tracking feature.
• perf(guest): walk the JUMPDEST map in assembly on the zkVM. The in-guest jump-table walk is hand-scheduled MIPS (jumpdest.s): six instructions per opcode instead of nine, with the branch delay slots filled and a read-ahead that stays inside the code's zero padding. It was checked against the Rust walk on every code of two mainnet blocks and on 2,000 random codes. Native builds keep the Rust walk.
• perf(guest): revm fork with cheaper dispatch and jumps. This moves to ziren-patches/revm patch-31.0.2-mips32-dispatch (interpreter: cheaper dispatch and jumps on 32-bit ziren-patches/revm#4): static gas charged on the low word of the remaining gas, JUMPDEST folded into the taken jump that lands on it, and PUSH2 + JUMP / JUMPI run as one instruction.
• deps: Ziren from perf/shamir-msm. This uses perf(zkvm-lib): Shamir's trick in the two-point MSM Ziren#562 (Shamir's trick in the zkVM lib's two-point MSM, which ECDSA recovery runs per transaction). The Ziren checkout the guest builds from is to be on that branch. The lockfile also keeps the precompile patches' main-branch zkm-lib, zkm-primitives and zkm-zkvm; those edges exist only for the zkVM target, where the guest's path patch replaces them.

Once revm#4 and Ziren#562 merge, the revm and Ziren dependencies go back to patch-31.0.2-mips32 and main.

Measured in the executor on blocks 26184948 / 26184949, with identical committed block hashes:

Change	Cycles	Main-trace area
Chain binding, precompile tracker gated	−0.19% / −0.28%	−0.11% / −0.16%
JUMPDEST walk in assembly	−1.39% / −1.79%	−0.80% / −1.00%
revm dispatch and jumps	−3.20% / −3.44%	−2.06% / −2.16%
Shamir MSM	−1.05% / −1.00%	−1.83% / −1.71%
All four, vs f30cf7d	402.97M → 379.91M (−5.72%) / 233.60M → 218.71M (−6.38%)	−4.72% / −4.95%

These four have not been proved on a GPU yet. On the current 5-card production, proving time fits 4.44 s + 33.7 ms per million cycles over the last 200 blocks, which puts the expected gain at about 4% of the mean block time (12.45 s → ~12.0 s). The program vk changes with the guest ELF.