EIP-8337: Validated EVM Code

Code prefixed with MAGIC bytes is proven at CREATE time to have fully static control flow and to never underflow the data stack.

Abstract

Code can be prefixed with MAGIC bytes. MAGIC code is validated at CREATE time to ensure that it cannot execute invalid instructions, jump to invalid destinations, or underflow the data stack, that every return has a call to return from, and that within each subroutine the depth of the data stack at every instruction is the same on every execution. Stack overflow remains checked at run time, as the data stack is for all code today.

The complete control flow of MAGIC code can be traversed in time and space linear in the size of the code, enabling validation, ahead-of-time (AOT) and just-in-time (JIT) compilation, automated proofs of correctness, formal analysis and more.

Validation is optional: the instructions of EIP-7979 behave identically in validated and ordinary code.

Note: Significant assistance from AI is acknowledged, primarily for the reference implementation and its tests.

Motivation

Dynamic jumps obscure the flow of control. A destination is a runtime value, so analysis must assume that any jump can reach any JUMPDEST. The possible control transfers therefore grow as jumps times destinations: quadratic in the size of the code. For tools that run when code deploys or executes, like validators and compilers to machine code, that is a denial-of-service vulnerability. For analyses done offline — formal analysis, even gas estimates — the state space can grow exponentially or worse. EIP-8173 lays out these foundations, with worked examples and the literature.

Validation ends this: code that forgoes dynamic control flow is proven safe once, at CREATE time, and every downstream tool can rely on the proof.

Specification

The key words MUST and MUST NOT in this Specification are to be interpreted as described in RFC 2119 and RFC 8174.

MAGIC (0xEF....)

After this EIP has been activated, code beginning with the MAGIC bytes MUST be a valid program. Execution begins immediately after the MAGIC bytes; positions in the code — PC values and destinations — count from the first MAGIC byte.

Note: MAGIC values are still to be determined, but the MAGIC bytes will begin with 0xEF.

Validity

valid code has fully static control flow: every transfer of control lands where the code says it does. Validation begins with the JUMPDEST analysis clients already run — the sequential scan, from the first byte after the MAGIC header, that tells instructions from PUSH immediate data — and proves more:

• the destination of every jump and call,
• that every return has a call to return from,
• that within each subroutine the depth of the data stack at every instruction is the same on every execution, and
• that no instruction can find fewer items on the data stack than it removes and underflow it.

Execution is defined in the Yellow Paper as a sequence of changes to the state of the Ethereum Virtual Machine (EVM). Exceptional halting conditions are properties of the machine state, checked before each instruction executes: if executing would violate one, execution halts instead — as the Yellow Paper puts it (§9.4.2), “no instruction can, through its execution, cause an exceptional halt.” Proving that no reachable state violates a condition therefore proves that no execution can reach an exceptional halting state. That is what validation does — once, at CREATE time. The Yellow Paper defines six conditions:

• State modification during a static call
• Insufficient gas
• More than 1024 stack items
• Insufficient stack items
• Invalid jump destination
• Invalid instruction

Validation proves the last three before the code ever runs: invalid instructions by Constraint 1, invalid jump destinations by Constraints 2 and 3, and insufficient stack items by Constraint 5. The first three remain runtime matters: whether the call is static is not knowable from the code, neither is the gas, and with recursion the depth of the stacks depends on data — see “Why not prove overflow?” in the Rationale.

Validation considers only the code’s control flow and stack offsets, never its data and computations. It follows both arms of every JUMPI, so it may traverse paths that data values would never let execute, and it rejects code with invalid paths even if they could never be taken.

Definitions

Graph theory, compilers, languages, virtual machines, and security all meet here, each with its own names for these ideas; the terms below are plain English, defined once and used exactly. The constraints below, and the discussion that follows, are stated in terms of these definitions. The first three are the machine state of EIP-7979.

• data stack: the EVM’s stack of 256-bit words, at most 1024 deep. Its depth is the number of items on it.

• return stack: the stack of return addresses, pushed only by CALLSUB, popped only by RETURNSUB, and not otherwise accessible to EVM code.

• PC: the program counter — the position in the code, the index of the next instruction to execute.

• instruction: a byte of code that is not PUSH immediate data, as the sequential scan from the first byte after the MAGIC header — JUMPDEST analysis — determines.

• execution: as the Yellow Paper defines it, a sequence of changes to the EVM state: one instruction at a time, beginning at the first instruction after the MAGIC bytes, ending in a halt.

• control flow: the possible transfers of control: which instruction may execute after which — the next in sequence, or a jump’s or call’s destination. Control flow is static when every destination is fixed by the code alone.

• control-flow graph: the control flow drawn as a map: each instruction a node, each possible transfer an edge.

• path: a possible sequence of executed instructions, starting at the first instruction; validation follows both arms of every JUMPI, without regard to data values.

• reachable: on some path. Bytes on no path are data, and are ignored.

• entry: a CALLDEST. Control may arrive at an entry by a call, by a jump, or by falling through.

• subroutine: the code an entry owns: everything reachable from it without passing through another entry. It is a subgraph of the control-flow graph, not a contiguous stretch of bytecode. Constraint 4 requires every reachable instruction to belong to exactly one subroutine.

• frame: the span of a path from a CALLSUB to the RETURNSUB that pops the return address it pushed. Every frame begun by a CALLSUB starts at an entry. Frames nest; execution begins in no frame — the top level is unframed. A frame is one execution of a call, not the code it runs: a subroutine called twice gives two frames, and, since a jump to a CALLDEST pushes no return address, one frame may pass through several subroutines.

• unframed entry: the top level, and any entry reached by a jump or fall-through — rather than by a CALLSUB — from an unframed entry. A CALLSUB frames its callee; a jump does not. In valid code no unframed entry contains a RETURNSUB: there it could find no return address.

• stack offset: the depth of the data stack, minus its depth at the most recent entry in the current frame — or at the start of execution, if there is none.

• net stack effect: of an entry, the stack offset at the RETURNSUB that closes a frame begun at that entry. Items pushed minus items popped: positive, negative, or zero.

• demand: of an entry, the deepest the data stack can reach below its depth at the entry, along the subroutine’s paths: the items that must already be on the stack when it is entered.

Constraints on valid code

Code beginning with MAGIC MUST be valid. Constraint 1 requires every instruction to be defined. Constraints 2 and 3 prove the destinations of jumps and calls. Constraint 4 gives every instruction one stack offset, and makes a single linear traversal sufficient. Constraint 5 keeps both stacks safe from underflow.

1. Every reachable instruction MUST be a valid opcode:

   • it MUST have been defined in the Yellow Paper or a deployed EIP,
   • it MUST NOT have been deprecated by a subsequent deployed EIP, and
   • the INVALID opcode is valid.

2. Every reachable JUMP and JUMPI MUST be immediately preceded by a PUSH, whose immediate value is the destination. The destination MUST be a JUMPDEST or a CALLDEST instruction — immediate data is not an instruction. A jump to a CALLDEST enters the subroutine.

3. Every reachable CALLSUB MUST be immediately preceded by a PUSH, whose immediate value is the destination. The destination MUST be a CALLDEST instruction. A call to a CALLDEST enters the subroutine.

4. Stack offsets MUST be path-independent:

   • every path reaching an instruction MUST arrive with the same stack offset — its depth above the subroutine’s entry;
   • every frame begun at a given entry MUST close with the same net stack effect — items pushed minus items popped by the call; and
   • every reachable instruction MUST belong to exactly one subroutine.

5. On every path: each instruction MUST find at least as many items on the data stack as it removes, and each RETURNSUB MUST find a return address on the return stack.

A subroutine is validated once, in terms of stack offsets, no matter how many call sites — CALLSUBs that target it — invoke it at different absolute depths, and no matter how many jumps enter it without a call. Code that several subroutines share is entered as a subroutine of its own; a CALLDEST costs what a JUMPDEST costs. Each entry’s net stack effect is a constant, so the stack offset at every return point is static. Loops must be stack-neutral per iteration — a backward branch arrives at the stack offset it left — so each instruction need be visited only once.

Validation

MAGIC code MUST be validated against the constraints above at CREATE time, in time and space linear in the size of the code. Validation MUST be run on the output of the initialization code, before that output can ever be executed by the interpreter, and failure of validation is an exceptional halting state of the CREATE.

Validity is defined against an instruction set, and instruction sets change. Each fork that changes the instructions available MUST define validation over the full instruction set of that fork. At CREATE, code beginning with MAGIC MUST have a header naming the validation rules of the creating fork, and MUST pass them; either failure — a wrong header, or invalid code — is a validation failure. Contracts created under earlier forks remain deployed, their headers naming the rules that admitted them; reaching an opcode deprecated by a later fork is an exceptional halting state, as reaching an undefined opcode is today.

Validation is a one-time cost, charged to the creator: CREATE MUST charge VALIDATION_BYTE_COST gas for each byte of code. VALIDATION_BYTE_COST is provisionally 64, sized to the measured worst case of native validation; the final value is deferred to the gas schedule of the adopting fork. The basis is under “What does validation cost?” in the Rationale.

The layout of the MAGIC header is explicitly deferred to a follow-on specification: other EIPs also need to carry information in the header, and the layouts will need to be reconciled. Only the encoding is deferred: what the header identifies — the validation rules that admitted the code — is fixed above.

Clients MUST implement validation, natively or by any equivalent means. The requirement is modest — the validate() function of the executable reference below is circa 170 lines, with shared test vectors.

Note: The Java Virtual Machine, WebAssembly, and .NET’s Common Language Runtime enforce similar constraints for similar reasons.

Semantics versus Syntax

The above is a purely semantic specification, placing no constraints on the syntax of bytecode beyond being an array of opcodes and immediate data. Subroutines are defined as subgraphs of the control-flow graph, not contiguous sequences of bytecode. The EVM is a simple state machine: each instruction advances it one step, and the control-flow graph maps where it can step. We only promise that valid code will not, as it were, jam up the gears of the machine.

Rather than enforce semantic constraints via syntax — as is done by higher-level languages — this proposal enforces them via validation: MAGIC code is proven valid at CREATE time.

With no syntactic constraints and minimal semantic constraints, we maximize opportunities for optimization — call elimination, shared epilogues (one exit sequence shared by many paths), and other cross-subroutine techniques — and for patterns like mutual recursion, multiple entry, and state-machine dispatch. Since we want to support compilation of EVM code to native code on the node — as code deploys or runs — it is crucial that the EVM code be as well optimized as possible by high-level-language compilers — upfront and offline.

Rationale

Why validation?

By marking MAGIC contracts as valid we are promising that their control flow is static, and the many tools that traverse the control flow can know this without inspecting the code for themselves. Today each side does the best it can with the machine as it is: compilers synthesize calls and returns from dynamic jumps, and tools work to recover what the compilers synthesized — effort on both sides that no one chose. With this proposal it will at least and at last become possible to write static EVM code, and the workarounds can retire with the problem.

As a demonstration, extract_cfg.py, provided with the reference implementation, recovers the complete control-flow graph of validated code in a single linear pass. It trusts validity and performs no checks of its own.

Validation also retires runtime work: for valid code, JUMPDEST analysis, the per-jump destination check, and the per-instruction underflow check are all unnecessary — each was proven at CREATE time.

More important, validation makes translation legal: with every destination proven and every instruction at one static stack offset, valid code can be translated — once, at deploy, in linear time — to forms a client executes far more cheaply. Measured with the demonstration compiler and translators among the assets of EIP-7979, on benchmark kernels bracketing arithmetic-heavy and call-heavy code: a register intermediate code, interpreted, runs in 1.5x to 2.1x fewer machine instructions than interpreted bytecode, and 3.2x to 3.3x composed with 64-bit arithmetic instructions; RISC-V machine code runs in 4.7x to 7x fewer, and 19x to 51x composed. The same binaries executed in a 64-bit RISC-V zero-knowledge virtual machine (zkVM) cost the prover exactly the measured instruction counts, plus a constant startup. Gas metering and the runtime checks this proposal cannot retire — stack overflow, return depth — are included in every measurement.

Why is validation optional?

Backwards compatibility. We cannot require existing code to become valid, and in a large, open ecosystem we cannot force an adoption schedule.

Why not prove overflow?

Because in general it cannot be done. With recursion, the depth of the stacks depends on data, so no static proof exists. Without recursion overflow can be proven — the accompanying overflow_validator.py proves it — but then every safety claim carries the qualifier “in the absence of recursion”, and this proposal prefers unqualified claims. So overflow keeps the runtime check all code has today: a bounds comparison per push, in our estimate a small percent of an interpreter’s time, and one comparison per call for the return stack. And nothing is lost to tools: each subroutine’s growth above its entry is computable offline, in linear time, so a client can replace the per-push bounds check with one check per call, cached per code hash as JUMPDEST analysis is cached today, with no consensus needed.

How can you prove underflow?

Because, unlike overflow, it is a property of the code alone. Underflow is an instruction finding fewer items on the stack than it removes. For a subroutine, the items at risk are the ones below its own starting depth — items already on the stack when it is entered: its demand counts them, and because stack offsets are static, that count is a fixed number.

So the validator computes each subroutine’s demand, then checks every way in: each call site, and each jump or fall-through into the entry. If the subroutine arriving there has pushed enough items of its own, the demand is met. If it has pushed too few, the missing count is added to its own demand, to be met in turn where it is entered — if a subroutine needs three items and its caller has pushed only one, the caller now needs two. The checking repeats until nothing changes, and it must stop, recursion or not, because demands are counts that never go down and can never exceed 1024. Execution begins with an empty stack, so nothing lies below top-level code’s start: it validates only if its demand is zero.

Is the validation algorithm sound and complete?

A validator is sound if it accepts only code that meets the constraints, and complete if it accepts all code that does. Throughout this proposal the constraints are read over every path of the control-flow graph, whether or not data would ever take it. The reference algorithm is both sound and complete, constraint by constraint.

• Constraints 1, 2 and 3 are properties of single instructions: the opcode is defined; a PUSH immediately precedes each jump and call; and its value is a JUMPDEST or CALLDEST that is not immediate data. The traversal visits exactly the reachable instructions and checks each one on the spot, so it accepts precisely the code that satisfies them.

• Constraint 4 is what the traversal records. On first arrival at an instruction it records the stack offset and the entry, and every later arrival must match. On the first RETURNSUB that closes a frame begun at an entry it records the entry’s net stack effect, and every later one must match. A mismatch violates one of the three clauses, and every violation is a mismatch on some arrival, so the check is exact. The third clause is what lets the rest be computed per subroutine: nets and demands are kept per entry, and an instruction reached from two entries would have no one entry to charge it to.

• Constraint 5 has two halves.

  • The *data stack.* With path-independent offsets, an instruction’s depth relative to its entry is a single number. So “finds at least as many items as it removes, on every path” reduces to the demand computation: the deepest an entry’s own code reaches below its base, less the items each parent supplies at the point of entrance, propagated up to the top level, which has nothing to supply. This is a restatement, not an approximation. The limit turns away only demands that no stack could satisfy, which underflow on every path.

  • The *return stack.* Call an entry unframed if it is the top level, or is reached by a jump or fall-through from an unframed entry. A CALLSUB frames its callee, and a return point keeps its caller’s framing. The algorithm rejects exactly when an unframed entry contains a RETURNSUB. If it rejects, then a chain of jumps from the top level reaches that entry with no CALLSUB on it, so some path reaches the RETURNSUB with every CALLSUB already returned: a violation. Conversely, if some path reaches a RETURNSUB with no unreturned CALLSUB, delete its returned call–return pairs. What remains is a chain of jumps and fall-throughs from the start into the entry containing the RETURNSUB, through entries reached the same way, each unframed by definition. So that entry is unframed and contains a RETURNSUB, and the algorithm rejects. Code with no RETURNSUB may be shared between framed and unframed entries, as the constraint allows.

What does validation cost?

Little. Measured natively (native-cost/) on one 2.6 GHz Intel core, ordinary code validates at 12–14 nanoseconds per byte, the sequential scan included. The worst case is code built to pump demands around a recursive cycle — invalid code, paying to be refused — at a measured 587 nanoseconds per byte, bounded as “Why it is linear” argues. VALIDATION_BYTE_COST at its provisional 64 covers the worst case with margin: about thirty percent on top of today’s roughly 216 gas per byte of deployment, charged by CREATE and paid by the creator. There is nothing for an attacker to buy — the worst case is the priced case. A contract that fails validation consumes its gas, as reverting initialization code does.

Why are the arguments to JUMP and JUMPI restricted in MAGIC code?

Constraint 2 requires that JUMP and JUMPI in MAGIC code be immediately preceded by a PUSH instruction, making their destinations compile-time constants.

The destination may be a JUMPDEST or a CALLDEST. A jump to a CALLDEST enters the subroutine without a call — call elimination, discussed in EIP-7979’s Rationale. Within a subroutine, jumps go to JUMPDESTs; between subroutines, control enters at an entry, by call or by jump. Computed jumps stay out: destinations remain compile-time constants, so the path explosions described below cannot occur. A jump into another subroutine remains invalid except at an entry. Making a point enterable costs one CALLDEST byte, and keeps every cross-subroutine transfer visible to one-pass analysis.

How can dynamic jumps explode control flow analysis?

Recovering a program’s control flow is a fundamental first step for many analyses. When all jumps are static, each jump has a fixed set of successors, and the analysis is linear in the size of the code. With dynamic jumps every destination must be considered at every jump: the possible transfers grow quadratically, and analyses that must follow the paths through them grow exponentially. EIP-8173 develops these costs with worked examples. For Ethereum they are a denial-of-service vulnerability for tools that run at CREATE time or at runtime, and even offline they render many analyses impractical, intractable, or impossible.

Backwards Compatibility

Validation is opt-in and changes no semantics. No deployed contract begins with the MAGIC bytes — EIP-3541 reserved them — so no existing code is affected. Opcode behavior is not affected by the prefix: the same code runs identically with or without it. Validation of MAGIC code is done before the interpreter runs, so the interpreter never sees MAGIC code that is not valid. Clients need not maintain two interpreters.

Test Cases

Note: the bytecode strings in these tests use the opcode values 0xBB=CALLDEST, 0xBC=CALLSUB, 0xBF=RETURNSUB.

Validation

The following bytecodes exercise the validator itself: the last column is the expected result of validate(). They are run by test_validator.py, which accompanies the reference implementation, with STACK_LIMIT reduced to 16 so that the overflow examples stay small; the algorithm is independent of the limit.

The tables group the cases by the constraint they exercise, and three verdicts carry most of the lessons. The PUSH-data impostors — JUMPDEST or CALLDEST bytes buried in immediate data — are refused whatever their values: the sequential scan already ruled them out. Recursion validates even with no base case, because validation follows control flow, not data. And the overflow rows all validate, because overflow is the runtime check — see “Why not prove overflow?”.

The runtime test cases of EIP-7979

Constraint 1: opcodes

Constraints 2 and 3: destinations

Constraint 4: offsets and net effects

Constraint 5: underflow and the return stack

Reuse and multiple entry points

Recursion

Jumps to a CALLDEST (call elimination)

Framing is a property of entries

Overflow is not validated

Reference Implementation

The reference implementation below is Python: executable and tested. It validates EVM bytecode against the five constraints defined above, in time and space linear in the size of the code, in circa 210 lines — the validate() function itself is circa 170.

It is embedded here so that this EIP reads as a single document, and provided as a separate file, validator.py. The reference and its test vectors carry no MAGIC header; behind a header, every position shifts by the header’s length. Beside it are its opcode table, opcodes.py; the test suite test_validator.py, which runs all the validation test cases above, its vectors shared with test_overflow.py; the one-pass control-flow-graph extractor extract_cfg.py; and, in native-cost/, the C port that measures validation’s native cost. The validator that additionally proves overflow for non-recursive code is overflow_validator.py — kept for comparison; see “Why not prove overflow?”.

Algorithm

The idea

Every client already scans deployed code, before executing it, to tell instructions from PUSH immediate data: JUMPDEST analysis. Validation runs that scan first, then traverses the code the way execution would: starting at the first instruction, following every jump and call, except that where execution takes one arm of a JUMPI, the traversal takes both. The traversal is confined to the instructions the scan found — PUSH immediate data can never be executed or jumped to, whatever its bytes. Bytes the traversal never reaches are data, just as unreachable bytes are today.

The traversal visits every reachable instruction exactly once. What makes once enough is measuring the data stack relative to the subroutine being traversed: at a CALLDEST the depth count resets to zero, and every check inside is phrased as an offset from that start. A subroutine therefore looks the same from all the CALLSUBs that invoke it, however deep their stacks, and checking it once covers them all. Constraint 4 completes the argument: every path to an instruction must arrive at the same offset, so a second arrival adds nothing new. For the same reason a loop is traversed once, since its backward branch must arrive at the offset it left.

The instruction after a CALLSUB is reached when the frame begun at the callee returns, at a depth that depends on what the callee did: the call-site offset plus the callee’s net stack effect. So a return point cannot be visited until its callee’s net is known. Return points wait on a pending list until that net is first fixed — by a RETURNSUB reached from the callee’s entry, in whatever subroutine it lies. If a callee never returns, its return points are never visited. That is correct: they are unreachable.

The details

Each step of the traversal visits one instruction, knowing four things: its position; the stack offset on arrival; which subroutine it is in — that is, which CALLDEST it was reached from, or top-level code; and the value of the immediately preceding PUSH, if any, since that is where JUMP, JUMPI, and CALLSUB destinations come from.

The first visit to an instruction records the offset and the subroutine; every later arrival must match both, or the code is invalid. Offsets must match by Constraint 4. Subroutines must match because the net stack effects are kept per subroutine: if paths from two different CALLDESTs could rejoin, there would be no one subroutine to charge the instructions to.

Whether a RETURNSUB will find a return address is a property of the subroutine it is in, not of the path that reached it. The traversal records the jump and fall-through links between entries — a CALLSUB makes no such link, since it frames its callee — and which entries contain a RETURNSUB. Afterwards, framing is propagated over those links: the top level is unframed, and so is every entry that an unframed entry jumps or falls into. An unframed entry containing a RETURNSUB is invalid, since a path through the unframed entry reaches the RETURNSUB with no unreturned CALLSUB on it. Code shared between framed and unframed entries is otherwise unconstrained: a revert block reached both from the top level and from inside subroutines is valid, since nothing about it touches the return stack.

Destinations are checked in whichever order the traversal reaches them. If the destination is already visited, its opcode is checked on the spot: a JUMPDEST or CALLDEST for jumps, a CALLDEST for calls. If it is not yet visited, the requirement is recorded and checked when the traversal first arrives there. A destination inside PUSH data can never satisfy either check: the scan has already ruled such bytes out, whatever their values.

A CALLDEST is always visited as the start of its own subroutine, at offset zero. Reaching one some other way — by falling through, or by a jump — links the two subroutines: whatever the entered subroutine’s net stack effect turns out to be, the entering subroutine’s is the offset at the entrance plus that, since from there on their fates are the same.

The reference implementation below follows this section, and each check in it is annotated with the constraint it enforces.

Python

"""Reference validator for EIP-8337 MAGIC code.

Abridged: the complete file, validator.py, accompanies this
EIP; it imports its opcode table from opcodes.py.
"""
from collections import defaultdict, deque

from opcodes import (JUMP, JUMPI, JUMPDEST, CALLSUB, CALLDEST, RETURNSUB,
                     PUSH0, PUSH32, opcode_info, push_value)

STACK_LIMIT = 1024
OUTER = None          # stands for the entry of top-level code
LABEL = "label"       # destination must be a JUMPDEST or a CALLDEST
ENTRY = "calldest"    # destination must be a CALLDEST


def validate(code, stack_limit=STACK_LIMIT):
    """True iff the code satisfies the five constraints of EIP-8337
    validation: valid opcodes, proven destinations, one static stack
    offset per instruction, framed returns, and no underflow.
    Framing is a property of entries: no entry reachable by jumps from
    the top level may contain a RETURNSUB."""
    if len(code) == 0:
        return False

    # JUMPDEST analysis: the sequential scan from position 0 that every
    # client runs today.  Its instructions are the only bytes that may
    # be executed or jumped to; PUSH immediate data never qualifies,
    # whatever its values (Constraints 2 and 3).
    instructions = set()
    i = 0
    while i < len(code):
        instructions.add(i)
        i += 1 + (code[i] - PUSH0 if PUSH0 < code[i] <= PUSH32 else 0)

    visited = {}                  # pc -> (offset, entry) at first visit
    required = {}                 # pc -> LABEL or ENTRY, set by jumps and calls
    net_effect = {}               # entry -> its *net stack effect*, once known
    inputs = defaultdict(int)     # entry -> its demand: the items it needs from its caller
    # Two parent lists with different lifetimes.  parents is permanent
    # and complete — every call and enter — for propagating demands in
    # the demand checking.  enter_parents holds only arrivals whose entry's
    # net is still unknown; each record is consumed exactly once, when
    # that net is first set.
    parents = defaultdict(list)   # child -> [(parent, offset)]
    enter_parents = defaultdict(list)  # entry -> [(parent, offset)]
    pending = defaultdict(list)   # entry -> return points waiting on its net
    jump_children = defaultdict(list)  # entry -> entries it jumps or falls into
    returning = set()             # entries that contain a RETURNSUB
    work_items = [(0, 0, OUTER, None)]

    def resolve(entry, value):
        """Record an entry's net: release the return points waiting on
        it, and settle the entries that jump or fall into it, whose
        nets follow from this one.  False on a conflict."""
        settle = [(entry, value)]
        while settle:
            e, v = settle.pop()
            if e in net_effect:
                if net_effect[e] != v:
                    return False  # Constraint 4: one net per entry
                continue
            net_effect[e] = v
            for ret_pc, offset, caller in pending.pop(e, ()):
                work_items.append((ret_pc, offset + v, caller, None))
            for parent, d in enter_parents[e]:
                settle.append((parent, d + v))
        return True

    while work_items:
        pc, offset, entry, push = work_items.pop()
        if pc >= len(code):
            continue                     # implicit STOP: a valid end
        if pc not in instructions:
            return False                 # Constraints 2, 3: immediate data
        op = code[pc]
        size, pops, pushes, term = opcode_info(op)
        if size == 0:
            return False                 # Constraint 1: not a valid opcode

        # A CALLDEST is visited at offset 0, as its own entry; arriving
        # any other way first records the link between the subroutines.
        if op == CALLDEST and (entry != pc or offset != 0):
            parents[pc].append((entry, offset))
            jump_children[entry].append(pc)
            if pc in net_effect:      # settled: the arriving net follows
                if not resolve(entry, offset + net_effect[pc]):
                    return False
            else:                     # each record is consumed exactly once
                enter_parents[pc].append((entry, offset))
            offset, entry = 0, pc

        if pc in visited:
            # Constraint 4: paths must agree.
            if visited[pc] != (offset, entry):
                return False
            continue
        visited[pc] = (offset, entry)

        # Constraints 2 and 3: a required destination type, if any.
        if required.get(pc) == LABEL and op not in (JUMPDEST, CALLDEST):
            return False
        if required.get(pc) == ENTRY and op != CALLDEST:
            return False

        # Constraint 5: items used from below the subroutine's start.
        need = pops - offset
        if need > inputs[entry]:
            if need > stack_limit:
                return False
            inputs[entry] = need
        offset += pushes - pops
        nxt = pc + size

        if op == CALLSUB:
            if push is None:
                return False             # Constraint 3: PUSH before CALLSUB
            dest = push
            if dest >= len(code):
                return False
            if dest in visited and code[dest] != CALLDEST:
                return False
            required[dest] = ENTRY       # a jump's LABEL upgrades to ENTRY
            parents[dest].append((entry, offset))
            work_items.append((dest, 0, dest, None))
            if dest in net_effect:
                # Return point: call-site offset plus the callee's net.
                work_items.append((nxt, offset + net_effect[dest], entry, None))
            else:
                pending[dest].append((nxt, offset, entry))
        elif op == RETURNSUB:
            returning.add(entry)
            if not resolve(entry, offset):
                return False
        elif op in (JUMP, JUMPI):
            if push is None:
                return False             # Constraint 2: PUSH before JUMP/JUMPI
            dest = push
            if dest >= len(code):
                return False
            if dest in visited and code[dest] not in (JUMPDEST, CALLDEST):
                return False
            required.setdefault(dest, LABEL)
            work_items.append((dest, offset, entry, None))
            if op == JUMPI:              # and the fall-through arm
                work_items.append((nxt, offset, entry, None))
        elif not term:                   # everything else falls through
            value = push_value(code, pc) if PUSH0 <= op <= PUSH32 else None
            work_items.append((nxt, offset, entry, value))

    # Constraint 5, framing: the top level is unframed, and so is every
    # entry that an unframed entry jumps or falls into; no unframed entry
    # may contain a RETURNSUB.  A worklist walks each jump link at most
    # once — a CALLSUB makes no link, since it frames its callee.
    unframed = {OUTER}
    work = [OUTER]
    while work:
        for e in jump_children[work.pop()]:
            if e not in unframed:
                unframed.add(e)
                work.append(e)
    if unframed & returning:
        return False

    # The demand checking: a subroutine's demand for caller items, less the
    # depth already on the stack at the entrance, becomes its parent's
    # demand.  Demands only rise and the limit caps them, so this ends.
    queue = deque(e for e in inputs if inputs[e])
    queued = set(queue)
    while queue:
        e = queue.popleft()
        queued.discard(e)
        for parent, d in parents[e]:
            need = inputs[e] - d
            if need > inputs[parent]:
                if need > stack_limit:
                    return False
                inputs[parent] = need
                if parent not in queued:
                    queue.append(parent)
                    queued.add(parent)

    # Top-level code has no caller to take items from.
    return inputs[OUTER] == 0

Why it is linear

Space first. Every table has one row per instruction or one per entry, and the parent lists hold one record, a few numbers, for each call and each entrance the traversal finds. Space is O(n) for code of n bytes.

Time is counted in work items. A work item says: visit this instruction, at this stack offset, in this subroutine. Each costs a bounded amount of work, a few table lookups, so the question is how many items there are. The first visit to an instruction creates at most two. A JUMPI creates one per arm. A CALLSUB creates one for the callee and one for the return point, which is held back until the callee’s net stack effect is known and released once, when the callee’s first RETURNSUB fixes it. Fixing a net also settles, once each, the entries that jumped or fell into that subroutine. A later arrival at a visited instruction compares offsets and creates nothing. So the items number at most about twice the instructions, and the traversal is O(n). The framing propagation that follows is a worklist over the jump links between entries. An entry is unframed at most once, and its links are walked only then, so each link is visited at most once: O(n) again.

The demand checking — “How can you prove underflow?” in the Rationale — repeats. When a subroutine comes up short at some entrance, the shortfall raises the demand of the code entering there, and that code must be checked again. How much can the repetition cost? A subroutine is checked again only because its own demand just rose, and a demand is a whole number of stack items that never falls and cannot pass 1024. So no subroutine is checked more than 1024 times. Each check walks that subroutine’s entrances once, and the entrances number O(n): one record per call, jump, or fall-through the traversal found. Time is O(1024 × n), which is still linear in the size of the code, because 1024 is the protocol’s constant and not the adversary’s choice. Only code built to pump demands around a recursive cycle approaches the bound: subroutines in a ring, each drawing items from the next, so that each round of checking raises every demand by one. That is invalid code, paying to be refused. Ordinary code settles in a round or two.

Measured on a C port of the validator (native-cost/), on one 2.6 GHz Intel core, ordinary patterns — straight-line code, branch-dense code, subroutines, deep call chains — cost 12–14 nanoseconds per byte, and the demand pump 587. The Python reference runs at 3–4 microseconds per byte on the same ordinary patterns. “What does validation cost?” in the Rationale prices these numbers.

Security Considerations

Validated contracts cannot execute invalid instructions, jump to invalid destinations, or underflow the data stack, and their return addresses are isolated from the data stack, so code cannot corrupt its own control flow. Stack overflow is checked at run time, as the data stack is for all code today.

Validation is consensus-critical, and this proposal asks every client to implement it. The defenses are an executable specification, shared test vectors, and differential testing; the risks that remain are a divergence between implementations, and a fault in the specification itself admitting code that a later fork must deal with. A remediation mechanism for the latter is deferred, like the header layout. (Fork-driven changes to validity — see Validation — need no such mechanism: contracts validated under earlier forks simply remain deployed.)

Validation is linear in the size of the code, its worst case measured and priced — see “What does validation cost?”. An attacker cannot impose more validation work on the network than they pay for.

Copyright

Copyright and related rights waived via CC0.

Citation

Please cite this document as:

Greg Colvin (@gcolvin) <greg@colvin.org>, Martin Holst Swende (@holiman), Brooklyn Zelenka (@expede), John Max Skaller, "EIP-8337: Validated EVM Code [DRAFT]," Ethereum Improvement Proposals, no. 8337, July 2026. Available: https://eips.ethereum.org/EIPS/eip-8337.