EOF - Static relative jumps

Abstract

Three new EVM jump instructions are introduced (RJUMP, RJUMPI and RJUMPV) which encode destinations as signed immediate values. These can be useful in the majority of (but not all) use cases and offer a cost reduction.

Motivation

A recurring discussion topic is that EVM only has a mechanism for dynamic jumps. They provide a very flexible architecture with only 2 (!) instructions. This flexibility comes at a cost however: it makes analysis of code more complicated and it also (partially) resulted in the need to have the JUMPDEST marker.

In a great many cases control flow is actually static and there is no need for any dynamic behaviour, though not every use case can be solved by static jumps.

There are various ways to reduce the need for dynamic jumps, some examples:

1. With native support for functions / subroutines
2. A "return to caller" instruction
3. A "switch-case" table with dynamic indexing

This change does not attempt to solve these, but instead introduces a minimal feature set to allow compilers to decide which is the most adequate option for a given use case. It is expected that compilers will use RJUMP/RJUMPI almost exclusively, with the exception of returning to the caller continuing to use JUMP.

This functionality does not preclude the EVM from introducing other forms of control flow later on. RJUMP/RJUMPI can efficiently co-exists with a higher-level declaration of functions, where static relative jumps should be used for intra-function control flow.

The main benefit of these instruction is reduced gas cost (both at deploy and execution time) and better analysis properties.

Specification

We introduce three new instructions on the same block number EIP-3540 is activated on:

1. RJUMP (0xe0) - relative jump
2. RJUMPI (0xe1) - conditional relative jump
3. RJUMPV (0xe2) - relative jump via jump table

If the code is legacy bytecode, all of these instructions result in an exceptional halt. (Note: This means no change to behaviour.)

If the code is valid EOF1:

1. RJUMP relative_offset sets the PC to PC_post_instruction + relative_offset.
2. RJUMPI relative_offset pops a value (condition) from the stack, and sets the PC to PC_post_instruction + ((condition == 0) ? 0 : relative_offset).
3. RJUMPV max_index relative_offset+ pops a value (case) from the stack, and sets the PC to PC_post_instruction + ((case > max_index) ? 0 : relative_offset[case]).

The immediate argument relative_offset is encoded as a 16-bit signed (two's-complement) big-endian value. Under PC_post_instruction we mean the PC position after the entire immediate value.

The immediate encoding of RJUMPV is more special: the unsigned 8-bit max_index value determines the maximum index in the jump table. The number of relative_offset values following is max_index+1. This allows table sizes up to 256. The encoding of RJUMPV must have at least one relative_offset and thus it will take at minimum 4 bytes. Furthermore, the case > max_index condition falling through means that in many use cases, one would place the default path following the RJUMPV instruction. An interesting feature is that RJUMPV 0 relative_offset is an inverted-RJUMPI, which can be used in many cases instead of ISZERO RJUMPI relative_offset.

We also extend the validation algorithm of EIP-3670 to verify that each RJUMP/RJUMPI/RJUMPV has a relative_offset pointing to an instruction. This means it cannot point to an immediate data of PUSHn/RJUMP/RJUMPI/RJUMPV. It cannot point outside of code bounds. It is allowed to point to a JUMPDEST, but is not required to.

Because the destinations are validated upfront, the cost of these instructions are less than their dynamic counterparts: RJUMP should cost 2, and RJUMPI and RJUMPV should cost 4.

Rationale

Relative addressing

We chose relative addressing in order to support code which is relocatable. This also means a code snippet can be injected. A technique seen used prior to this EIP to achieve the same goal was to inject code like PUSHn PC ADD JUMPI.

We do not see any significant downside to relative addressing and it allows us to also deprecate the PC instruction.

Immediate size

The signed 16-bit immediate means that the largest jump distance possible is 32767. In the case the bytecode at PC=0 starts with an RJUMP, it will be possible to jump as far as PC=32770.

Given MAX_CODE_SIZE = 24576 (in EIP-170) and MAX_INITCODE_SIZE = 49152 (in EIP-3860), we think the 16-bit immediate is large enough.

A version with an 8-bit immediate would only allow moving PC backward by 125 or forward by 127 bytes. While that seems to be a good enough distance for many for-loops, it is likely not good enough for cross-function jumps, and since the 16-bit immediate is the same size as what a dynamic jump would take in such cases (3 bytes: JUMP PUSH1 n), we think having less instructions is better.

Should there be a need to have immediate encodings of other size (such as 8-bits, 24-bits or 32-bits), it would be possible to introduce new opcodes, similarly to how multiple PUSH instructions exist.

PUSHn JUMP sequences

If we chose absolute addressing, then RJUMP could be viewed similar to the sequence PUSHn JUMP (and RJUMPI similar to PUSHn JUMPI). In that case one could argue that instead of introducing a new instruction, such sequences should get a discount, because EVMs could optimise them.

We think this is a bad direction to go:

1. It further complicates the already complex rules of gas calculation.
2. And it either requires a consensus defined internal representation for EVM code, or forces EVM implementations to do optimisations on their own.

Both of these are risky. Furthermore we think that EVM implementations should be free to chose what optimisations they apply, and the savings do not need to be passed down at all cost.

Additionally it requires a potentially significant change to the current implementations which depend on a streaming one-by-one execution without a lookahead.

Relation to dynamic jumps

The goal was not to completely replace the current control flow system of the EVM, but to augment it. There are many cases where dynamic jumps are useful, such as returning to the caller.

It is possible to introduce a new mechanism for having a pre-defined table of valid jump destinations, and dynamically supplying the index within this table to accomplish some form of dynamic jumps. This is very useful for efficiently encoding a form of "switch-cases" statements. It could also be used for "return to caller" cases, however it is likely inefficient or awkward.

Lack of JUMPDEST

JUMPDEST serves two purposes:

1. To efficiently partition code -- this can be useful for pre-calculating total gas usage for a given block (i.e. instructions between JUMPDESTs), and for JIT/AOT translation.
2. To explicitly show valid locations (otherwise any non-data location would be valid).

This functionality is not needed for static jumps, as the analysers can easily tell destinations from the static jump immediates during jumpdest-analysis.

There are two benefits here:

1. Not wasting a byte for a JUMPDEST also means a saving of 200 gas during deployment, for each jump destination.
2. Saving an extra 1 gas per jump during execution, given JUMPDEST itself cost 1 gas and is "executed" during jumping.

RJUMPV fallback case

If no match is found (i.e. the default case) in the RJUMPV instruction execution will continue without branching. This allows for gaps in the arguments to be filled with 0s, and a choice of implementation by the programmer. Alternate options would include exceptional aborts in case of no match.

Backwards Compatibility

This change poses no risk to backwards compatibility, as it is introduced at the same time EIP-3540 is. The new instructions are not introduced for legacy bytecode (code which is not EOF formatted).

Test Cases

Validation

Valid cases

• RJUMP/RJUMPI/RJUMPV with JUMPDEST as target
  • relative_offset is positive/negative/0
• RJUMP/RJUMPI/RJUMPV with instruction other than JUMPDEST as target
  • relative_offset is positive/negative/0
• RJUMPV with various valid table sizes from 1 to 256
• RJUMP as a final instruction in code section

Invalid cases

• RJUMP/RJUMPI/RJUMPV with truncated immediate
• RJUMPI/RJUMPV as a final instruction in code section
• RJUMP/RJUMPI/RJUMPV target outside of code section bounds
• RJUMP/RJUMPI/RJUMPV target push data
• RJUMP/RJUMPI/RJUMPV target another RJUMP/RJUMPI/RJUMPV immediate argument

Execution

• RJUMP/RJUMPI/RJUMPV in legacy code aborts execution
• RJUMP
  • relative_offset is positive/negative/0
• RJUMPI
  • relative_offset is positive/negative/0
    • condition equals 0
    • condition does not equal 0
• RJUMPV 0 relative_offset
  • case equals 0
  • case does not equal 0
• RJUMPV with table containing positive, negative, 0 offsets
  • case equals 0
  • case does not equal 0
  • case outside of table bounds (case > max_index, fallback case)
  • case > 255

Security Considerations

TBA

Copyright

Copyright and related rights waived via CC0.