Introduce a section in the EOF format (EIP-3540) for storing the list of JUMPDESTs, validate the correctness of this list at the time of contract creation, and remove the need for JUMPDEST-analysis at execution time. In EOF contracts, the JUMPDEST instruction is not needed anymore and becomes invalid. Legacy contracts are entirely unaffected by this change.
Currently existing contracts require no validation of correctness, but every time they are executed, a list must be built containing all the valid jump-destinations. This is an overhead which can be avoided, albeit the effect of the overhead depends on the client implementation.
With the structure provided by EIP-3540 it is easy to store and transmit a table of valid jump-destinations instead of using designated JUMPDEST (0x5b) opcodes in the code.
The goal of this change is that we trade less complexity (and processing time) at execution time for more complexity at contract creation time. Through benchmarks we have identified that the mandatory execution preparation time is the same as before for extreme cases (i.e. deliberate edge cases), while it is ~10x faster for the average case.
Finally, this change puts an implicit bound on "initcode analysis" which is now limited to jumpdests section loading of max size of 0xffff. The legacy code remains vulnerable.
This feature is introduced on the very same block EIP-3540 is enabled, therefore every EOF1-compatible bytecode MUST have a JUMPDEST-table if it uses jumps.
Remark: We rely on the notation of initcode, code and creation as defined by EIP-3540, and extend validation rules of EIP-3670.
jumpdests (section_kind = 3) is introduced. It contains a sequence of n unsigned integers jumploci.| description | encoding |
|---|---|
| jumploc0 | unsigned LEB128 |
| jumploc1 | unsigned LEB128 |
| ... | |
| jumplocn | unsigned LEB128 |
python
def jumpdest(n: int, jumpdests_table: list[int]) -> int:
return sum(jumpdests_table[:n+1])This section extends contract creation validation rules (as defined in EIP-3540).
jumpdests section MUST be present if and only if the code section contains JUMP or JUMPI opcodes.jumpdests section is present it MUST directly precede the code section. In this case a valid EOF bytecode will have the form of format, magic, version, [jumpdests_section_header], code_section_header, [data_section_header], 0, [jumpdests_section_contents], code_section_contents, [data_section_contents].jumploc must be valid: the encoding must be complete and not read out of jumpdests section. As an additional constraint, the shortest possible encoding must be used.jumploc MUST NOT be 0.jumploc MUST point to a valid opcode. They MUST NOT point into PUSH-data or outside of the code section.JUMPDEST (0x5b) instruction becomes undefined (Note: According to rules of EIP-3670, deploying the code will fail if it contains JUMPDEST)JUMP or JUMPI instructions, the jump destination MUST be in the jumpdests table. Otherwise, the execution aborts with bad jump destination. In case of JUMPI, the check is done only when the jump is to be taken (no change to the previous behaviour).The length of the jumpdests section is bounded by the EOF maximum section size value 0xffff. Moreover, for deployed code this additionally limited by the max bytecode size 0x6000. Then any valid jumpdests section may not be more larger than 0x3000.
Delta-encoding is very efficient for this job. From a quick analysis of a small set of contracts JUMPDEST opcodes are relatively close to each other. In the delta-encoding the values almost never exceed 128. Combined with any form of variable-length quantity (VLQ) where values < 128 occupy one byte, encoding of single jumpdest takes ~1 byte. We also remove JUMPDEST opcodes from the code section therefore the total bytecode length remains the same if extreme examples are ignored.
By extreme examples we mean contracts having a distance between two subsequent JUMPDESTs larger than 128. Then the LEB128 encoding of such distance requires more than one byte and the total bytecode size will increase by the additional number of bytes used.
The LEB128 encoding is the most popular VLQ used in DWARF and WebAssembly.
LEB128 allows encoding a fixed value with arbitrary number of bytes by having zero payloads for most significant bits of the value. To ensure there exists only single encoding of a given value, we additionally require the shortest possible LEB128 encoding to be used. This constraint is also required by WebAssembly.
This is another option for encoding inspired by UTF-8. The benefit is that the number of following bytes is encoded in the first byte (the top two bits), so the expected length is known upfront.
A simple decoder is the following:
def decode(input: bytes) -> int:
size_prefix = input[0] >> 6
if size_prefix == 0:
return input[0] & 0x3f
elif size_prefix == 1:
return (input[0] & 0x3f) << 8 | input[1]
elif size_prefix == 2:
return (input[0] & 0x3f) << 16 | (input[1] << 8) | input[2]
# Do not support case 3
assert(False)
In case code does not use jumps, an empty JUMPDEST table is represented by omitting jumpdests section as opposed to a section that is always present, but allowed to be empty. This is consistent with the requirement of EIP-3540 for section size to be non-zero. Additionally, omitting the section saves 3 bytes of code storage.
The contents of jumpdests section are always needed to start EVM execution. For chunked and/or merkleized bytecode it is more efficient to have jumpdests just after the EOF header so they can share the same first chunk(s).
In code chunking the contract code is split into (fixed size) chunks. Due to the requirement of jumpdest-analysis, it must be known where the first instruction starts in a given chunk, in case the split happened within a PUSH-data. This is commonly accomplished with reserving the first byte of the chunk as the "first instruction offset" (FIO) field.
With this EIP, code chunking does not need to have such a field. However, the jumpdest table must be provided instead (for all the chunks up until the last jump location used during execution).
We compared the performance of jumpdests section loading to JUMPDEST analysis in evmone/Baseline interpreter. In both cases a bitset of valid jumpdest positions is built.
We used the worst case for jumpdests section as the benchmark baseline. This is the case where every position in the code section is valid jumpdest. I.e. the bytecode has as many jumpdests as possible making the jumpdests section as large as possible. The encoded representation is 0x00, 0x01, 0x01, 0x01, ....
This also happen to be the worst case for the JUMPDEST analysis.
Further, we picked 5 popular contracts from the Ethereum mainnet.
| case | size | num JUMPDESTs | JUMPDEST analysis (cycles/byte) | jumpdests load (cycles/byte) | performance change |
|---|---|---|---|---|---|
| worst | 65535 | 65535 | 9.11 | 9.36 | 2.75% |
| RoninBridge | 1760 | 71 | 3.57 | -89.41% | |
| UniswapV2ERC20 | 2319 | 61 | 2.10 | -88.28% | |
| DepositContract | 6358 | 123 | 1.86 | -90.24% | |
| TetherToken | 11075 | 236 | 1.91 | -89.58% | |
| UniswapV2Router02 | 21943 | 468 | 2.26 | -91.17% |
For the worst case the performance difference between JUMPDEST analysis and jumpdests section loading is very small. The performance very slow compared to memory copy (0.15 cycles/byte).
However, the maximum length for the worst cases is different. For JUMPDEST analysis this is 24576 (0x6000) for deployed contracts and only limited by EVM memory cost in case of initcode (can be over 1MB). For jumpdests sections, the limit is 12288 for deployed contracts (the deployed bytecode length limit must be split equally between jumpdests and code sections). For initcode case, the limit is 65535 because this is the maximum section size allowed by EOF.
For "popular" contracts the gained efficiency is ~10x because the jumpdests section is relatively small compared to the code section and therefore there is much less bytes to loop over than in JUMPDEST analysis.
We extend the validate_code() function of EIP-3670:
# The same table as in EIP-3670
valid_opcodes = ...
# Remove JUMPDEST from the list of valid opcodes
valid_opcodes.remove(0x5b)
# This helper decodes a single unsigned LEB128 encoded value
# This will abort on truncated (short) input
def leb128u_decode(input: bytes) -> (int, int):
ret = 0
shift = 0
consumed_bytes = 0
while True:
# Check for truncated input
assert(consumed_bytes < len(input))
# Only allow up to 4-byte long leb128 encodings
assert(consumed_bytes <= 3)
input_byte = input[consumed_bytes]
consumed_bytes += 1
ret |= (input_byte & 0x7f) << shift
if (input_byte & 0x80) == 0:
# Do not allow additional leading zero bits.
assert(input_byte != 0 || consumed_bytes == 0)
break
shift += 7
return (ret, consumed_bytes)
# This helper parses the jumpdest table into a list of relative offsets
# This will abort on truncated (short) input
def parse_table(input: bytes) -> list[int]:
jumpdests = []
pos = 0
while pos < len(input):
value, consumed_bytes = leb128u_decode(input[pos:])
jumpdests.append(value)
pos += consumed_bytes
return jumpdests
# This helper translates the delta offsets into absolute ones
# This will abort on invalid 0-value entries
def process_jumpdests(delta: list[int]) -> list[int]:
jumpdests = []
partial_sum = 0
first = True
for d in delta:
if first:
first = False
else:
assert(d != 0)
partial_sum += d
jumpdests.append(partial_sum)
return jumpdests
# Fails with assertion on invalid code
# Expects list of absolute jumpdest offsets
def validate_code(code: bytes, jumpdests: list[int]):
pos = 0
while pos < len(code):
# Ensure the opcode is valid
opcode = code[pos]
pos += 1
assert(opcode in valid_opcodes)
# Remove touched offset
try:
jumpdests.remove(pos)
except ValueError:
pass
# Skip pushdata
if opcode >= 0x60 and opcode <= 0x7f:
pos += opcode - 0x60 + 1
# Ensure last PUSH doesn't go over code end
assert(pos == len(code))
# The table is invalid if there are untouched locations
assert(len(jumpdests) == 0)
JUMP but no jumpdest tableJUMPI but no jumpdest tableJUMP/JUMPIJUMPDESTThis change poses no risk to backwards compatibility, as it is introduced at the same time EIP-3540 is. The requirement of a JUMPDEST table does not cover legacy bytecode.
The authors are not aware of any security or DoS risks posed by this change.
Copyright and related rights waived via CC0.