feat(evm_interpreter): skip JUMPDEST dispatch when reached via JUMP/JUMPI

[0xVolosnikov]

What ❔

JUMP and JUMPI already validate that the target byte is a JUMPDEST opcode. The standalone JUMPDEST handler that the dispatcher runs on the next iteration is then redundant — its only effect is charging 1 gas + JUMPDEST_NATIVE_COST and returning Ok(()).

This PR inlines that work into JUMP / JUMPI only when building for the RISC-V proving target (cfg!(target_arch = "riscv32")).

• On host builds (forward / RPC / opcode-stats path), the dispatcher continues to run JUMPDEST normally, so EvmOpcodeStatsTracer, EvmOpcodesLogger, and any other forward-mode tracer that keys on per-opcode events keeps observing real JUMPDEST events with accurate gas/native deltas.
• On RISC-V builds (proving), JUMP/JUMPI fold the JUMPDEST gas + native charge into their own body and advance IP past the JUMPDEST byte. There is no live tracer on RISC-V that keys on JUMPDEST, so dropping the iteration is safe.

Mechanism

const INLINE_JUMPDEST: bool = cfg!(target_arch = "riscv32");

A compile-time const gates the optimization. The non-inlined branches are dead code on RISC-V (and vice versa), so the hot path stays lean on both targets.

• JUMP: when INLINE_JUMPDEST, fuses MID + JUMPDEST gas and JUMP_NATIVE + JUMPDEST_NATIVE native into a single spend_gas_and_native call (InvalidJump consumes all remaining gas anyway, so pre-charging is observationally identical). On host, charges only MID / JUMP_NATIVE as before and lands on dest.
• JUMPI: charges HIGH / JUMPI_NATIVE as before. On INLINE_JUMPDEST-and-taken-and-valid, a second spend_gas_and_native(JUMPDEST, JUMPDEST_NATIVE) charge is added (kept separate so the not-taken path does not pay JUMPDEST gas — the OOG boundary for marginal-gas frames must not shift). IP lands on dest + 1.
• JUMPDEST handler is preserved and runs as before on host. On RISC-V it only runs for fall-through cases (e.g. JUMPI condition false landing on a JUMPDEST byte).

Cycle-marker balance

cycle_marker keeps a host-side LABELS Vec and pairs it with marker CSR writes from the RISC-V binary; the two counts must match. Because host still dispatches JUMPDEST while RISC-V skips it, JUMP / JUMPI emit a synthetic cycle_marker::opcode_start!() / opcode_end!("JUMPDEST") pair on RISC-V to keep counts balanced.

Per-opcode cycle attribution between JUMP/JUMPI and JUMPDEST in the bench output gets scrambled because the synthetic pair nests inside the dispatcher's JUMP/JUMPI bracket — but the block-level process_block effective-cycle total is unaffected. That's the metric we measure for proving cost.

Why ❔

After the PUSH2..PUSH8 work, JUMPDEST showed as a top-tier total-cycle contributor in benchmark blocks despite costing only ~42 cycles per call, because of its sheer call count. Most of those JUMPDESTs are reached as jump targets where the validity check has already happened — the dispatch iteration is pure overhead (per-opcode STEP_NATIVE charge + dispatch match arm + cycle_marker pair).

Benchmark

bench_scripts/bench.sh compare, baseline = vv-push-small-specializations HEAD:

Block	Base Eff	Head Eff	Δ
block_19299001	208,011,772	207,739,099	−0.13% (−273K cycles)
block_22244135	134,167,657	132,764,589	−1.05% (−1.40M cycles)

Per-opcode (aggregated across both blocks, proving-side cycle_marker data):

Opcode	Count	Median cycles	Δ Median	Notes
JUMP	56,689	76 → 84	+10.5%	absorbs fused JUMPDEST charge
JUMPI	46,978	88 → 132	+29.4%	absorbs separate JUMPDEST charge
JUMPDEST	87,480	~42 → 2	−95%	synthetic-marker only on RISC-V; per-opcode attribution is the intentional artifact

Delegations (Blake / BigInt / Keccak): unchanged.

Cross-opcode noise

DUP1–DUP10 show ~+1.8% each (~150-200K cycles aggregate) consistent with the I-cache pressure already observed on this dispatch loop. Net per-block effective cycles are still solidly negative.

Is this a breaking change?

• Yes
• No

No protocol-visible behavior change. Total gas charged per JUMP+JUMPDEST sequence is 9 gas on both host and RISC-V builds. EVM state transitions are identical between modes. Failure semantics preserved (InvalidJump → consume all remaining gas, identical caller-visible outcome regardless of where the charge fails).

Checklist

• PR title corresponds to the body of PR.
• Tests for the changes have been added / updated.
  • cargo test -p evm_interpreter --features testing (13/13). tests/instances/evm rig (14/14). Existing tests exhaustively exercise JUMP/JUMPI/JUMPDEST through real EVM bytecode.
• Documentation comments have been added / updated.
• Code has been formatted.

Base branch note

Stacked on vv-push-small-specializations (PR #648). When that lands, this PR's base can be retargeted to custom-u256.

🤖 Generated with Claude Code