fix(context): persist resume_at in call_with_interpreter

[SipengXie2024]

Summary

call_with_interpreter creates a temporary EvmContext from the interpreter on each invocation. When JIT code suspends at a CALL or CREATE, it writes the resume point to ecx.resume_at. However, this value was never written back to the interpreter's bytecode PC before ecx was dropped.

On the next call, ResumeAt::load(pc, code) sees pc < code.len() (since PC was never updated) and returns 0, causing the JIT function to restart from the beginning instead of resuming at the correct point.

Impact

• Invisible for simple contracts without nested calls (e.g., the existing bench.rs benchmarks)
• Catastrophic performance degradation for real-world contracts with multiple CALL/CREATE operations

Uniswap V2 swap benchmark results:

Metric	Before fix	After fix
JIT execution	~921 µs	~29.5 µs
vs interpreter	17x slower	1.86x faster

Fix

Save ecx.resume_at before the borrow ends, then persist it via interpreter.bytecode.absolute_jump(resume_at) (+8 lines).

Root Cause Analysis

1. JIT suspends at CALL/CREATE → writes resume block address to ecx.resume_at
2. call_with_interpreter returns → ecx is dropped, resume_at is lost
3. Next invocation → EvmContext::from_interpreter_with_stack → ResumeAt::load(pc, code) → pc < code.len() → returns 0
4. JIT restarts from the beginning of the bytecode

Note: call() (low-level API used in bench.rs) is not affected because the caller manages EvmContext lifetime and ecx.resume_at persists across calls.