EIP-2330 - EXTSLOAD opcode

Created	2019-10-29
Status	Stagnant
Category	Core
Type	Standards Track
Authors	Dominic Letz (@dominicletz)
Requires	EIP-2929

Abstract

This proposal adds a new opcode EXTSLOAD at 0x5c which pops two items from the stack: <account address> <storage key> and pushes one item: <storage value>. The gas cost is sum of account access cost and storage read based on EIP-2929 Access Lists.

Motivation

While any off-chain application can read all contract storage data of all contracts, this is not possible for deployed smart contracts themselves. These are bound to use contract calls for any interaction including reading data from other contracts. This EIP adds an EVM opcode to directly read external contract storage.

The gas cost when reading from registry style contract such as EIP-20s, ENS and other data contracts is very high, because they incur cross contract call cost, cost for ABI encoding, decoding and dispatching and finally loading the data. In many cases the underlying storage that is being queried is though just a simple mapping. On top of that, the view function may SLOAD many other slots which caller may not be interested in, which further adds to the gas costs. In these cases a new EXTSLOAD call directly accessing the mapping in storage could not only reduce the gas cost of the interaction more than 10x, but also it would make the gas cost predictable for the reading contract.

Specification

A new EVM instruction EXTSLOAD (0x5c) that works like SLOAD (0x54) but an additional parameter representing the contract that is to be read from.

EXTSLOAD (0x5c)

The EXTSLOAD instruction pops 2 values from the stack, first contract a contract address and then second slot a storage address within contract. As result EXTSLOAD pushes on the stack the value from the contract storage of contract at the storage slot address or 0 in case the account contract does not exist.

Gas cost pre-verkle

Gas to be charged before Verkle Tree change is specified as ACCOUNT_ACCESS_COST + STORAGE_READ_COST where:

ACCOUNT_ACCESS_COST is 0 if the account address is already in accessed_addresses set, otherwise COLD_ACCOUNT_ACCESS_COST.
STORAGE_READ_COST is WARM_STORAGE_READ_COST if storage key is already in accessed_storage_keys set, otherwise COLD_STORAGE_READ_COST.

Gas cost post-verkle

It is important to consider that post Verkle tree change, ACCOUNT_ACCESS_COST will not be needed since a single account's storage would be spread across the entire global trie. Hence gas to be charged post Verkle Tree change is just STORAGE_READ_COST, which is as specified in Gas cost pre-verkle.

Rationale

Without this EIP, a contract can still opt-in to make their entire state public, by having a method that simply SLOADs and returns the values (example). The complexity of the gas cost can be seen as 1x CALL cost + Nx SLOAD cost. Hence, the gas cost specified for using EXTSLOAD opcode on an account for N times, the charge of 1x COLD_ACCOUNT_ACCESS_COST and Nx STORAGE_READ_COST is hereby justified.
Without this EIP, a contract can still use internal state of other contracts. An external party can supply a value and proof to a contract, which the contract can verify using BLOCKHASH. This is only possible for the previous blocks and not the latest state (since current blockhash cannot be determined before execution).
This opcode can be seen as breaking object-oriented (OO) model because it allows to read storage of other contracts. In usual systems using OO is net positive, because there is no limit on machine code and it hardly adds any cost to add more methods or use single method to get a ton of data while the caller needs to just a small portion of data. However on EVM, there are visible costs, i.e. about $0.2 per SLOAD (20 gwei and ETHUSD 2000). Also, OO has caused misleading assumptions for developers where variables marked as "private" in smart contracts are encrypted in some way/impossible to read which has resulted bad designs. Hence, this EIP can be beneficial in terms of making smart contract systems more efficient as well as preventing misconceptions as well.

Backwards Compatibility

This change is fully backwards compatible since it adds a new instruction.

Security Considerations

Since the opcode is similar to SLOAD, it should be easy to implement in various clients.
This opcode allows the callee A to re-enter a caller contract B and read state of B and B cannot stop A from doing that. Since this does not change any state, it should not be a security issue. Contracts generally use re-entrancy guards, but that is only added to write methods. So even currently without EXTSLOAD, A can re-enter B and read their state exposed by any view methods and it has not been an issue.