Introduce an upper bound on the inputs of the MODEXP precompile. This can reduce the number of potential bugs, because the testing surface is not infinite anymore, and makes it easier to be replaced using EVMMAX.
The MODEXP precompile has been a source of numerous consensus bugs. Many of them were due to specifically crafted cases using impractical input lengths.
Its pricing function is also quite complex given its nature of unbounded inputs. While we don't suggest to rework the pricing function, it may be possible in a future upgrade once the limits are in place.
Furthermore this limitation makes it more feasible to have the precompile replaced with EVM code through features like EVMMAX.
Recap from EIP-198:
At address
0x00……05, add a precompile that expects input in the following format:
<length_of_BASE> <length_of_EXPONENT> <length_of_MODULUS> <BASE> <EXPONENT> <MODULUS>
We introduce an upper bound to the inputs of the precompile, each of the length inputs (length_of_BASE, length_of_EXPONENT and length_of_MODULUS) MUST be less than or equal to 8192 bits (1024 bytes).
If any of these inputs are larger than the limit, the precompile execution stops, returns an error, and consumes all gas.
This upper bound allows the existing use cases of MODEXP:
Replacing the precompile with EVM code using an instruction set like EVMMAX would be made simpler with this limit: Common cases (256, 381, 1024, 2048) could be implemented in special fast paths, while a slow fallback could be provided for the rest. Or even special, frequently used, moduli could have their own paths.
Furthermore one could consider limiting the lengths to certain inputs only.
Since MODEXP was introduced in the Byzantium hard fork, an analysis has been conducted between block 5472266 (April 20, 2018) and block 21550926 (January 4th, 2025). All lengths of inputs are expressed in bytes.
input_of_BASE | count |
|---|---|
| 32 | 2439595 |
| 128 | 4167 |
| 256 | 2969 |
| 160 | 436 |
| 512 | 36 |
| 0 | 13 |
| 64 | 7 |
| 78 | 2 |
| 513 | 2 |
| 129 | 1 |
| 385 | 1 |
input_of_EXPONENT | count |
|---|---|
| 32 | 2442255 |
| 3 | 4771 |
| 1 | 159 |
| 128 | 29 |
| 0 | 13 |
| 5 | 2 |
input_of_MODULUS | count |
|---|---|
| 32 | 2439594 |
| 128 | 4167 |
| 256 | 2968 |
| 160 | 436 |
| 512 | 38 |
| 0 | 13 |
| 64 | 8 |
| 78 | 2 |
| 129 | 1 |
| 384 | 1 |
| 257 | 1 |
This shows that no past successful use case exceeded an input length of 513 bytes, and the majority uses 32/128/256 byte inputs.
Besides these, there were a few invocations with invalid inputs:
0x9e5faafc or 0x854747280x9e281a98000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000021e19e0c9bab2400000This is a backwards incompatible change. However, based on analysis until block 21550926 (see above), no past transaction would have behaved differently after this change.
Since only the accepted input range is reduced, no new security surface area is expected.
Copyright and related rights waived via CC0.