JIT-EVM Experiments

Goal

Evaluate performance gains to expect when EVM were to compile hot contracts into machine code (rather than interpreting the EVM bytecode)

Todos

• Support all instructions
• Performance evaluation
• Error handling
• Gas accounting

Experiment (MacOS, see Ubuntu below)

Setup:

• LLVM 14 installed with brew in /opt/homebrew/opt/llvm

Run:

LLVM_SYS_140_PREFIX=/opt/homebrew/opt/llvm cargo run

What it does:

• Fibonacci sequence calculator hand-implemented in EVM bytecode
• Runs it in EVM interpreter
• Compiles it to machine code and executes it
• You compare that the outputs correspond to the 15th Fibonacci number (which is 377)

How to interpret the output:

• Modulo indexing convention, 15th Fibonacci number is 377 (says Google)
• Last line of demo output (this is from JIT VM): return value is 377
• Scrolling up past the LLVM IR: last state of interpreter has 377 in the 0th stack entry

Tests (MacOS, see Ubuntu below)

RUST_BACKTRACE=1 LLVM_SYS_140_PREFIX=/opt/homebrew/opt/llvm RUST_MIN_STACK=8388608 cargo test -- --nocapture

Ubuntu

Install:

sudo apt-get install llvm-14 llvm-14-dev libllvm14

Adjust paths above accordingly (leaving LLVM_SYS_140_PREFIX=/opt/homebrew/opt/llvm away entirely "works for me").

References

• https://github.com/TheDan64/inkwell
• https://www.mattkeeter.com/projects/elfjit/
• https://github.com/mkeeter/advent-of-code/blob/master/2018/day21-jit/src/main.rs
• https://github.com/bluealloy/revm
• ethereum/evmone#320
• https://github.com/axic/snailtracer
• https://doc.rust-lang.org/reference/type-layout.html