Fuel VM Specification

• Introduction
• Parameters
• Semantics
• Flags
• Instruction Set
• VM Initialization
• Contexts
• Predicate Estimation
• Predicate Verification
• Script Execution
• Call Frames
• Ownership

Introduction

This document provides the specification for the Fuel Virtual Machine (FuelVM). The specification covers the types, instruction set, and execution semantics.

Parameters

Semantics

FuelVM instructions are exactly 32 bits (4 bytes) wide and comprise of a combination of:

• Opcode: 8 bits
• Register/special register (see below) identifier: 6 bits
• Immediate value: 12, 18, or 24 bits, depending on operation

Of the 64 registers (6-bit register address space), the first 16 are reserved:

Integers are represented in big-endian format, and all operations are unsigned. Boolean false is 0 and Boolean true is 1.

Registers are 64 bits (8 bytes) wide. Words are the same width as registers.

Persistent state (i.e. storage) is a key-value store with 32-byte keys and 32-byte values. Each contract has its own persistent state that is independent of other contracts. This is committed to in a Sparse Binary Merkle Tree.

Flags

All other flags are reserved, any must be set to zero.

Instruction Set

A complete instruction set of the Fuel VM is documented in the following page.

VM Initialization

Every time the VM runs, a single monolithic memory of size VM_MAX_RAM bytes is allocated, indexed by individual byte. A stack and heap memory model is used, allowing for dynamic memory allocation in higher-level languages. The stack begins at 0 and grows upward. The heap begins at VM_MAX_RAM and grows downward.

To initialize the VM, the following is pushed on the stack sequentially:

1. Transaction hash (byte[32], word-aligned), computed as defined here.
2. Base asset ID (byte[32], word-aligned)
3. MAX_INPUTS pairs of (asset_id: byte[32], balance: uint64), of:
   1. For predicate estimation and predicate verification, zeroes.
   2. For script execution, the free balance for each asset ID seen in the transaction's inputs, ordered in ascending order. If there are fewer than MAX_INPUTS asset IDs, the pair has a value of zero.
4. Transaction length, in bytes (uint64, word-aligned).
5. The transaction, serialized.

Then the following registers are initialized (without explicit initialization, all registers are initialized to zero):

1. $ssp = 32 + 32 + MAX_INPUTS*(32+8) + size(tx)): the writable stack area starts immediately after the serialized transaction in memory (see above).
2. $sp = $ssp: writable stack area is empty to start.
3. $hp = VM_MAX_RAM: the heap area begins at the top and is empty to start.

Contexts

There are 4 contexts in the FuelVM: predicate estimation, predicate verification, scripts, and calls. A context is an isolated execution environment with defined memory ownership and can be external or internal:

• External: predicate and script. $fp will be zero.
• Internal: call. $fp will be non-zero.

Returning from a context behaves differently depending on whether the context is external or internal.

Predicate Estimation

For any input of type InputType.Coin or InputType.Message, a non-zero predicateLength field means the UTXO being spent is a P2SH rather than a P2PKH output.

For each such input in the transaction, the VM is initialized, then:

1. $pc and $is are set to the start of the input's predicate field.
2. $ggas and $cgas are set to MAX_GAS_PER_PREDICATE.

Predicate estimation will fail if gas is exhausted during execution.

During predicate mode, hitting any of the following instructions causes predicate estimation to halt, returning Boolean false:

1. Any contract instruction.

In addition, during predicate mode if $pc is set to a value greater than the end of predicate bytecode (this would allow bytecode outside the actual predicate), predicate estimation halts returning Boolean false.

A predicate that halts without returning Boolean true would not pass verification, making the entire transaction invalid. Note that predicate validity is monotonic with respect to time (i.e. if a predicate evaluates to true then it will always evaluate to true in the future).

After successful execution, predicateGasUsed is set to MAX_GAS_PER_PREDICATE - $ggas.

Predicate Verification

For any input of type InputType.Coin or InputType.Message, a non-zero predicateLength field means the UTXO being spent is a P2SH rather than a P2PKH output.

For each such input in the transaction, the VM is initialized, then:

1. $pc and $is are set to the start of the input's predicate field.
2. $ggas and $cgas are set to predicateGasUsed.

Predicate verification will fail if gas is exhausted during execution.

During predicate mode, hitting any contract instruction causes predicate verification to halt, returning Boolean false.

In addition, during predicate mode if $pc is set to a value greater than the end of predicate bytecode (this would allow bytecode outside the actual predicate), predicate verification halts returning Boolean false.

A predicate that halts without returning Boolean true does not pass verification, making the entire transaction invalid. Note that predicate validity is monotonic with respect to time (i.e. if a predicate evaluates to true then it will always evaluate to true in the future).

After execution, if $ggas is non-zero, predicate verification fails.

Script Execution

If script bytecode is present, transaction validation requires execution.

The VM is initialized, then:

1. $pc and $is are set to the start of the transaction's script bytecode.
2. $ggas and $cgas are set to tx.scriptGasLimit.

Following initialization, execution begins.

For each instruction, its gas cost gc is first computed. If gc > $cgas, deduct $cgas from $ggas and $cgas (i.e. spend all of $cgas and no more), then revert immediately without actually executing the instruction. Otherwise, deduct gc from $ggas and $cgas.

After the script has been executed, tx.receiptsRoot is updated to contain the Merkle root of the receipts, as described in the TransactionScript spec.

Call Frames

Cross-contract calls push a call frame onto the stack, similar to a stack frame used in regular languages for function calls (which may be used by a high-level language that targets the FuelVM). The distinction is as follows:

1. Stack frames: store metadata across trusted internal (i.e. intra-contract) function calls. Not supported natively by the FuelVM, but may be used as an abstraction at a higher layer.
2. Call frames: store metadata across untrusted external (i.e. inter-contract) calls. Supported natively by the FuelVM.

Call frames are needed to ensure that the called contract cannot mutate the running state of the current executing contract. They segment access rights for memory: the currently-executing contracts may only write to their own call frame and their own heap.

A call frame consists of the following, word-aligned:

Access rights

Only memory that has been allocated is accessible. Attempting to read or write memory that has not been allocated will result in VM panic. Similarly reads or writes that cross from the stack to the heap will panic. Note stack remains readable even after stack frame has been shrunk. However, if heap is alter expanded to cover that area, the crossing read prohibition still remains. In other word, memory between highest-ever $sp value and current $hp is inaccessible.

Ownership

Whenever memory is written to (i.e. with SB or SW), or write access is granted (i.e. with CALL), ownership must be checked.

If the context is external, the owned memory range is:

1. [$ssp, $sp): the writable stack area.
2. [$hp, VM_MAX_RAM): the heap area allocated by this script or predicate.

If the context is internal, the owned memory range for a call frame is:

1. [$ssp, $sp): the writable stack area of the call frame.
2. [$hp, $fp->$hp): the heap area allocated by this call frame.

Executablity

Memory is only executable in range [$is, $ssp). Attempting to execute instructions outside these boundaries will panic. This area never overlaps with writable registers, essentially providing W^X protection.