About the EVM


The Ethereum Virtual Machine (or EVM) is a stack-based computer. It means that all instructions take their parameters from the stack, and write their results on the stack. Each instruction thus has stack inputs, the parameters that it needs (if any), stack outputs, and the return values (if any). All instructions are encoded on 1 byte, with the exception of the PUSH instructions, which allows to put an arbitrary value on the stack and encode the value directly after the instruction. The list of instructions available, with their opcodes, is shown in the reference.

An instruction is assigned a value between 0 and 255 (or FF in hexadecimal), called the opcode, and a mnemonic, which is a text representation that helps us human read the instruction. A smart contract is a set of instructions. When the EVM executes a smart contract, it reads and executes each instruction one by one. If an instruction cannot be executed (for example because there are not enough values on the stack), the execution reverts.

Execution environment

When the EVM executes a smart contract, a context is created for it. The context is made of several memory regions, each with its own purpose.

The code

The code is the region where the instructions are stored. It is persistent and part of an account properties. During smart contract execution, these are the bytes that the EVM will read, interpret and execute. This is a region that cannot be modified, but can be read with the instructions CODESIZE and CODECOPY. Other contracts code can also be read with EXTCODESIZE and EXTCODECOPY. The program counter (PC) encodes which instruction should be read next by the EVM in this region. An externally owned account (or EOA) has an empty code region.

The stack

The stack is a list of 32-byte elements. The stack is used to put the parameters that are needed by the instructions, and their result values. When a new value is put on the stack, it is put on top, and only the top values are used by the instructions. The stack currently has a maximum limit of 1024 values. All instructions interact with the stack, but it can be directly manipulated with instructions like PUSH1, POP, DUP1, or SWAP1.

The memory

The memory is a region that only exists during the smart contract execution, and is accessed with a byte offset. While all the 32-byte address space is available and initialized to 0, the size is counted with the highest address that was accessed. It is generally read and written with MLOAD and MSTORE instructions, but is also used by other instructions like CREATE or EXTCODECOPY.

The storage

The storage is the persistent memory of the smart contract. It is a map of the 32-byte slot to 32-byte value, and each value written is kept until it is set to 0 or the contract self-destruction. Reading from an unset key also returns 0. It is read and written with the instructions SLOAD and SSTORE.

The calldata

The calldata region is the data that is sent with a transaction. In the case of contract creation, it would be the constructor code. This region is immutable and can be read with the instructions CALLDATALOAD, CALLDATASIZE, and CALLDATACOPY.

The return data

The return data region is the way a smart contract can return a value after a call. It can be set by external contract calls through the RETURN and REVERT instructions and can be read by the calling contract with RETURNDATASIZE and RETURNDATACOPY.

Gas costs

As an incentive to provide resources to run transactions, a fee is paid to send and execute a transaction. The fee is determined by several factors, including the amount of data sent or the amount of work that a transaction requires. That fee is calculated through two mechanisms. The first is a fixed cost that is defined by Ethereum, depending on what is executed. The unit for that is called gas, and remains the same for all transactions, though it can be changed by a hardfork. The second component is the price of one gas unit, which varies over time according to what people are willing to pay to run their transactions. Its unit is ETH per gas.

Each opcode has its own gas cost. There are two parts to their gas cost. The first one is the static cost, which has to be paid for running that opcode. The second is a dynamic cost, which depends on several factors during the execution of a transaction. The factors that are used to determine the dynamic gas cost can change from fork to fork. Use our reference to learn about the specific computations per opcode and fork. And finally, remember that the gas cost still has other components, for example the transaction fee, or the size of the calldata. To get a complete estimation of the gas cost for your programm, with your compiler options and specific state and inputs, use a tool like Remix or Truffle.

Intrinsic Gas

Each transaction has an intrinsic cost of 21000 gas. Creating a contract costs 32000 gas, on top of the transaction cost. And finally, the calldata costs 4 gas per byte equal to 0, and 16 gas for the others (64 before the hardfork Istanbul). This cost is paid from the transaction before any opcode or transfer is executed.

Memory expansion

During an execution, the whole memory is accessible, but not for free. When an offset is accessed for the first time (either read or write), it may trigger a memory expansion, which will cost gas. A memory expansion may be triggered when the offset used is bigger than any used before. When that happens, the cost of accessing that higher offset is computed and removed from the total gas available in the current context.

The total cost for a given memory size is computed as follows:

memory_size_word = (memory_byte_size + 31) / 32
memory_cost = (memory_size_word ** 2) / 512 + (3 * memory_size_word)

When a memory expansion is triggered however, only the additional chunk of memory has to be paid. The cost of memory expansion for a specific opcode is thus:

memory_expansion_cost = new_memory_cost - last_memory_cost

The memory_byte_size can be obtained with MSIZE. We can see that the cost grows quadratically with the size, making higher offsets more costly and discouraging to use too much memory. Any opcode accessing memory may trigger an expansion (including, for example, MLOAD, RETURN or CALLDATACOPY). Each opcode that can is mentioned in the reference. Note also that an opcode with a byte size parameter of 0 will not trigger a memory expansion, regardless of its offset parameters.

Access sets

Access sets have been introduced in the hardfork Berlin. They are kept per transaction (and not per call context). Two of them exist: the touched contracts addresses, and the touched contract storage slots. If an address or slot is present in the set, it is called 'warm', otherwise it is 'cold'. The dynamic cost of some opcodes depends on whether the address or slot is warm or cold.

  • Addresses: a set of contract addresses that have been touched in the current transaction. It is initialized with the sender and receiver (or the new contract address in case of a creation) of the transaction, as well as all the precompiled contracts. When an opcode accesses an address that is not present in the set, it adds it in it. The relevant opcodes are EXTCODESIZE, EXTCODECOPY, EXTCODEHASH, BALANCE, CALL, CALLCODE, DELEGATECALL, STATICCALL, CREATE, CREATE2 and SELFDESTRUCT.
  • Slots: a set of contract address and their storage slot keys that have been accessed. It is initialized to empty. When an opcode accesses a slot that is not present in the set, it adds it to it. The relevant opcodes are SLOAD and SSTORE
When a context reverts, the sets are also reverted to the state they had before that context.

Gas refunds

Some opcodes can trigger gas refunds, which reduces the gas cost of a transaction. However the gas refund is applied at the end of a transaction, meaning that a transaction always need enough gas to run as if there was no refunds. The amount of gas that can be refunded is also limited, to half of the total transaction cost before the hardfork London, otherwise to a fifth. Starting from the hardfork London also, only SSTORE may trigger refunds. Before that, SELFDESTRUCT could also trigger refunds.

Other EVM Related Resources

Ethereum EVM Illustrated (2018), The History and Forks of Ethereum, EVM: From Solidity to bytecode, memory and storage, EVM Deep Dives by noxx, The EVM Chapter in the Mastering Ethereum book

Acknowledgment to wolflo for the cost descriptions.