EIP-5656: Memory copying instruction

Abstract #

Provide an efficient EVM instruction for copying memory areas.

Motivation #

Memory copying is a basic operation, yet implementing it on the EVM comes with overhead.

This was recognised and alleviated early on with the introduction of the "identity" precompile, which accomplishes memory copying by the use of CALL's input and output memory offsets. Its cost is 15 + 3 * (length / 32) gas, plus the call overhead. The identity precompile was rendered ineffective by the raise of the cost of CALL to 700, but subsequently the reduction by EIP-2929 made it slightly more economical.

Copying exact words can be accomplished with <offset> MLOAD <offset> MSTORE or <offset> DUP1 MLOAD DUP2 MSTORE, at a cost of at least 12 gas per word. This is fairly efficient if the offsets are known upfront and the copying can be unrolled. In case copying is implemented at runtime with arbitrary starting offsets, besides the control flow overhead, the offset will need to be incremented using 32 ADD, adding at least 6 gas per word.

Copying non-exact words is more tricky, as for the last partial word, both the source and destination needs to be loaded, masked, or'd, and stored again. This overhead is significant. One edge case is if the last "partial word" is a single byte, it can be efficiently stored using MSTORE8.

As example use case, copying 256 bytes costs:

• at least 757 gas pre-EIP-2929 using the identity precompile
• at least 157 gas post-EIP-2929 using the identity precompile
• at least 96 gas using unrolled MLOAD/MSTORE instructions
• 27 gas using this EIP

According to an analysis of blocks 10537502 to 10538702, roughly 10.5% of memory copies would have had improved performance with the availability of an MCOPY instruction.

Memory copying is used by languages like Solidity and Vyper, where we expect this improvement to provide efficient means of building data structures, including efficient sliced access and copies of memory objects. Having a dedicated MCOPY instruction would also add forward protection against future gas cost changes to CALL instructions in general.

Having a special MCOPY instruction makes the job of static analyzers and optimizers easier, since the effects of a CALL in general have to be fenced, whereas an MCOPY instruction would be known to only have memory effects. Even if special cases are added for precompiles, a future hard fork could change CALL effects, and so any analysis of code using the identity precompile would only be valid for a certain range of blocks.

Finally, we expect memory copying to be immensely useful for various computationally heavy operations, such as EVM384, where it is identified as a significant overhead.

Specification #

The instruction MCOPY is introduced at 0x5c.

Input stack #

This ordering matches the other copying instructions, i.e. CALLDATACOPY, RETURNDATACOPY.

Output stack #

This instructions returns no stack items.

Semantics #

It copies length bytes from the offset pointed at src to the offset pointed at dst in memory. Copying takes place as if an intermediate buffer were used, allowing the destination and source to overlap.

If length > 0 and (src + length or dst + length) is beyond the current memory length, the memory is extended with respective gas cost applied.

The gas cost of this instruction mirrors that of other Wcopy instructions and is Gverylow + Gcopy * ceil(length / 32).

Rationale #

Production implementation of exact-word memory copying and partial-word memory copying can be found in the Solidity, Vyper and Fe compilers.

With EIP-2929 the call overhead using the identity precompile was reduced from 700 to 100 gas. This is still prohibitive for making the precompile a reasonable alternative again.

Backwards Compatibility #

This EIP introduces a new instruction which did not exists previously. Already deployed contracts using this instruction could change their behaviour after this EIP.

Test Cases #

TBA

Security Considerations #

TBA

Copyright #

Copyright and related rights waived via CC0.