What is the root cause of the performance improvement from 0.19.0 to 0.20.0

[ImJeremyHe]

I mocked 10000 transactions to check the performance for revm. After bumping revm to version 0.20.0, I found there is a great improvement.
I have checked the release log, but still have no clue.
Can anyone help me find out what is the root cause? Thanks!

[rakita]

Initialization was not optimized earlier and precompile "Loading" was done in the not fastest way, So I optimized that part and it became faster.
This part:

revm/crates/revm/src/evm_impl.rs

Lines 167 to 181 in b19acaa

	let mut subroutine = SubRoutine::default();
	if env.cfg.perf_all_precompiles_have_balance {
	// load precompiles without asking db.
	let mut precompile_acc = Vec::new();
	for (add, _) in precompiles.as_slice() {
	precompile_acc.push(*add);
	}
	subroutine.load_precompiles_default(&precompile_acc);
	} else {
	let mut precompile_acc = Map::new();
	for (add, _) in precompiles.as_slice() {
	precompile_acc.insert(*add, db.basic(*add));
	}
	subroutine.load_precompiles(precompile_acc);
	}

was replaced with this

revm/crates/revm/src/evm_impl.rs

Lines 204 to 208 in aa39d64

	let journaled_state = if GSPEC::enabled(SpecId::SPURIOUS_DRAGON) {
	JournaledState::new(precompiles.len())
	} else {
	JournaledState::new_legacy(precompiles.len())
	};

Cached precompile HashMap here gave speed up: https://github.com/bluealloy/revm/commits/main

This is good comparison of simple SM call that does nothing that is x4 faster than 1.8: https://github.com/cl2089/rust-evm-bench

An additional change is related to Bytecode analysis. You can read about it here: #121

[ImJeremyHe]

Thanks for your reply! I will take a further look for it.

[rakita]

If you want to see where exactly the processing went I would recommend doing cargo flamegraph and comparing commits from this branch: https://github.com/cl2089/rust-evm-bench it is very visually evident