Loading...
Home / Research / diving-into-decompilation /
  Share   | 1.53 k

Diving Into Smart Contract Decompilation

January 19, 2023    |    By Jonathan Becker

Heimdall Header

The Heimdall-rs decompilation module is a powerful tool for understanding the inner workings of Ethereum smart contracts. It allows users to convert raw bytecode into human-readable Solidity code and its corresponding Application Binary Interface (ABI). In this article, we will delve deep into the inner workings of the decompilation module, examining how it performs this conversion at a low level and exploring its various features and capabilities.

Please keep in mind that this article talks about how heimdall-rs performs decompilation. This may not be the same as how other decompilers work. If you have any suggestions or corrections, please feel free to open an issue or pull request on GitHub! Thank you!

0x01. Introduction

Decompilation is the process of converting machine code or bytecode into a higher-level, human-readable representation. However, it is not a straightforward task for several reasons:

  • Machine code or bytecode is designed to be executed by a computer, not read by humans. As a result, it can be ambiguous and difficult to interpret.
  • Bytecode does not contain information about variable and function names, making it difficult to understand the purpose of different parts of the code.
  • Bytecode does not contain all the information that the original source code had, such as comments, variable and function names, types, etc.
  • Bytecode is not a linear representation of the source code, for multiple reasons such as compiler optimizations, making it even more challenging to decompile effectively.

In order to make this complex, convoluted process easier to understand, I've broken the overall decompilation process into four main steps, which we will explore in detail in the following sections:

  • Disassembly: The process of converting bytecode into its assembly representation.
  • Symbolic Execution: The process of generating a branch-like control flow graph (CFG) from the disassembled code.
  • Branch Analysis: The process of analyzing and translating the CFG into a higher-level representation.
  • Post-Processing: The process of cleaning up the output and making it more readable.

0x02. Disassembly

The first step in the decompilation process is to convert the bytecode into a more human-readable assembly representation. This is done to allow the decompiler to find and analyze the different parts of the smart-contract's overall structure.

Assembly is a low-level representation of the bytecode, where each instruction is represented by a mnemonic and its corresponding arguments. Each instruction has roughly 3 parts:

  <program_counter> <opcode> <arguments>

The program counter is the index of the instruction in the bytecode. The opcode is the instruction itself, and the arguments are the values that the instruction operates on. For advanced information on Ethereum opcodes, check out EVM Codes.

For example, the following bytecode from 0x1bf797219482a29013d804ad96d1c6f84fba4c45:

  731bf797219482a29013d804ad96d1c6f84fba4c45301460806040...9d5ef505ae7230ffc3d88c49ceeb7441e0029

Would be converted into the following assembly:

  20 PUSH20 1bf797219482a29013d804ad96d1c6f84fba4c45
  21 ADDRESS 
  22 EQ 
  24 PUSH1 80
  26 PUSH1 40
  27 MSTORE 
  29 PUSH1 04
  30 CALLDATASIZE 
  31 LT 
  34 PUSH2 0058
  35 JUMPI 
  37 PUSH1 00
  38 CALLDATALOAD 
  ...

The Solidity Dispatcher

Heimdall-rs uses this disassembled code to search for JUMPI statements in the dispatch lookup table, which is used to determine the function that is being called. In the Ethereum Virtual Machine (EVM), functions are called by passing a function selector as the first 4 bytes of the calldata. The function selector is the first 4 bytes of the keccak256 hash of the function signature, which is the function name and its corresponding arguments. For example, the function signature transfer(address,uint256) would be converted into the function selector 0xa9059cbb.

The dispatch lookup table is a mapping of function selectors to the corresponding function address (as a program counter). It is used to determine which function is being called by the first 4 bytes of the calldata. Typically, AND(PUSH4(0xFFFFFFFF), CALLDATALOAD(0)) is used to load the first 4 bytes of the calldata, and then the function selector is compared to the function selectors in the dispatch lookup table to determine which function is being called. If a match is found, the program jumps to the corresponding location in the bytecode.

For example, the following assembly shows this dispatch table in action for the WETH contract 0xc02aaa39b223fe8d0a0e5c4f27ead9083c756cc2:

  14 PUSH1 00
  15 CALLDATALOAD 
  45 PUSH29 0100000000000000000000000000000000000000000000000000000000
  46 SWAP1 
  47 DIV 
  52 PUSH4 ffffffff
  53 AND 
  54 DUP1 
  59 PUSH4 06fdde03
  60 EQ 
  63 PUSH2 00b9
  64 JUMPI 
  65 DUP1 
  70 PUSH4 095ea7b3
  71 EQ 
  74 PUSH2 0147
  75 JUMPI 

From this, we can see that the function selector for 0x06fdde03 (name()) tells the EVM to jump to program counter 0xb9, and the function selector for 0x095ea7b3 (approve(address,uint256)) tells the EVM to jump to program counter 0x147. Heimdall-rs' decompile module searches through this dispatch lookup table to find all function selectors and their corresponding locations in bytecode, allowing for the symbolic execution and branch analysis steps to be performed.

0x03. Symbolic Execution

The second step in the decompilation process is to generate a control flow graph (CFG) from the disassembled code. A CFG is a directed graph that represents the control flow of a program. Each node in the graph represents a block of instructions that are executed sequentially. Each edge in the graph represents a jump or branch, which is a conditional jump to another basic block. The CFG is used to represent the control flow of the program and is used to determine the different paths that the program can take.

This CFG is generated by heimdall-rs by executing the bytecode in a custom EVM implementation specifically designed for decompilation and bytecode analysis. Whenever the EVM encounters a JUMPI instruction, a new branch is created in the CFG. A new VM is created for each branch, and the program is executed until terminated. This process is repeated until all branches have been explored, and the CFG is complete.

This process can be seen in the following code snippet from heimdall-rs:
  pub fn recursive_map(
      evm: &VM,
      handled_jumpdests: &mut Vec<String>,
      path: &mut String,
  ) -> VMTrace {
      let mut vm = evm.clone();

      // create a new VMTrace object
      let mut vm_trace = VMTrace {
          instruction: vm.instruction,
          operations: Vec::new(),
          children: Vec::new(),
          loop_detected: false,
          depth: 0,
      };

      // cap the number of branches to prevent infinite loops. Needs to be fixed in the future.
      if handled_jumpdests.len() >= 1000 { return vm_trace }

      // step through the bytecode until we find a JUMPI instruction
      while vm.bytecode.len() >= (vm.instruction * 2 + 2) as usize {
          let state = vm.step();
          vm_trace.operations.push(state.clone());

          // if we encounter a JUMPI, create children taking both paths and break
          if state.last_instruction.opcode == "57" {
              vm_trace.depth += 1;

              path.push_str(&format!("{}->{};", state.last_instruction.instruction, state.last_instruction.inputs[0]));

              // we need to create a trace for the path that wasn't taken.
              if state.last_instruction.inputs[1] == U256::from(0) {

                  // break out of loops
                  match LOOP_DETECTION_REGEX.is_match(&path) {
                      Ok(result) => {
                          if result {
                              vm_trace.loop_detected = true;
                              break;
                          }
                      }
                      Err(_) => {
                          return vm_trace
                      }
                  }

                  handled_jumpdests.push(format!("{}@{}", vm_trace.depth, state.last_instruction.instruction));

                  // push a new vm trace to the children
                  let mut trace_vm = vm.clone();
                  trace_vm.instruction = state.last_instruction.inputs[0].as_u128() + 1;
                  vm_trace.children.push(recursive_map(
                      &trace_vm,
                      handled_jumpdests,
                      &mut path.clone()
                  ));

                  // push the current path onto the stack
                  vm_trace.children.push(recursive_map(
                      &vm.clone(),
                      handled_jumpdests,
                      &mut path.clone()
                  ));
                  break;
              } else {

                  // break out of loops
                  match LOOP_DETECTION_REGEX.is_match(&path) {
                      Ok(result) => {
                          if result {
                              vm_trace.loop_detected = true;
                              break;
                          }
                      }
                      Err(_) => {
                          return vm_trace
                      }
                  }

                  handled_jumpdests.push(format!("{}@{}", vm_trace.depth, state.last_instruction.instruction));

                  // push a new vm trace to the children
                  let mut trace_vm = vm.clone();
                  trace_vm.instruction = state.last_instruction.instruction + 1;
                  vm_trace.children.push(recursive_map(
                      &trace_vm,
                      handled_jumpdests,
                      &mut path.clone()
                  ));

                  // push the current path onto the stack
                  vm_trace.children.push(recursive_map(
                      &vm.clone(),
                      handled_jumpdests,
                      &mut path.clone()
                  ));
                  break;
              }
          }

          if vm.exitcode != 255 || vm.returndata.len() > 0 {
              break;
          }
      }

      vm_trace
  }

Aside from the loop detection, this code is fairly straightforward. The VM is stepped through until a JUMPI instruction is encountered. The VM is then cloned, and the instruction pointer is set to the jump destination. The cloned VM is then passed to the recursive_map function, which will recursively step through the bytecode until it reaches a JUMPI instruction. This process is repeated until all branches have been explored, and the CFG is complete.

Loop detection is still something that needs to be improved upon. The current method of detecting loops is to check if the path taken so far matches a regular expression that matches the pattern of a loop. This method is not perfect and can be fooled by certain bytecode patterns.

0x04. Branch Analysis

Once the CFG is generated, the next step is to analyze the branches of the CFG. This is where the real decompilation begins, and translation from opcode to solidity begins.

The WrappedOpcode Struct

As stated earlier, decompilation and symbolic execution are powered by my custom EVM implementation. This implementation introduces WrappedOpcodes which are essentially opcodes with additional information. This information includes the instruction pointer, the inputs, and the outputs of the opcode. This information is used to generate the CFG and is also used to generate the decompiled code.

This struct can be seen in the following code snippet from heimdall-rs:

  // enum allows for Wrapped Opcodes to contain both raw U256 and Opcodes as inputs
  #[derive(Clone, Debug, PartialEq)]
  pub enum WrappedInput {
      Raw(U256),
      Opcode(WrappedOpcode),
  }

  // represents an opcode with its direct inputs as WrappedInputs
  #[derive(Clone, Debug, PartialEq)]
  pub struct WrappedOpcode {
      pub opcode: Opcode,
      pub inputs: Vec<WrappedInput>,
  }

The VM implementation keeps track of these WrappedOpcodes as it steps through the bytecode. Allowing for each input and output for any instruction to be traced back to its origin. This is essential for decompilation, as it allows for the decompiler to know exactly what each instruction is doing and where its inputs are coming from, eventually allowing for the decompiler to generate the correct solidity code.

Raw WrappedOpcode
  [
      WrappedOpcode {
          opcode: Opcode {
              name: "SUB",
              mingas: 3,
              inputs: 2,
              outputs: 1,
          },
          inputs: [
              Opcode(
                  WrappedOpcode {
                      opcode: Opcode {
                          name: "ADD",
                          mingas: 3,
                          inputs: 2,
                          outputs: 1,
                      },
                      inputs: [
                          Opcode(
                              WrappedOpcode {
                                  opcode: Opcode {
                                      name: "PUSH1",
                                      mingas: 3,
                                      inputs: 0,
                                      outputs: 1,
                                  },
                                  inputs: [
                                      Raw(
                                          32,
                                      ),
                                  ],
                              },
                          ),
                          Opcode(
                              WrappedOpcode {
                                  opcode: Opcode {
                                      name: "MLOAD",
                                      mingas: 3,
                                      inputs: 1,
                                      outputs: 1,
                                  },
                                  inputs: [
                                      Opcode(
                                          WrappedOpcode {
                                              opcode: Opcode {
                                                  name: "PUSH1",
                                                  mingas: 3,
                                                  inputs: 0,
                                                  outputs: 1,
                                              },
                                              inputs: [
                                                  Raw(
                                                      64,
                                                  ),
                                              ],
                                          },
                                      ),
                                  ],
                              },
                          ),
                      ],
                  },
              ),
              Opcode(
                  WrappedOpcode {
                      opcode: Opcode {
                          name: "SLOAD",
                          mingas: 3,
                          inputs: 1,
                          outputs: 1,
                      },
                      inputs: [
                          Opcode(
                              WrappedOpcode {
                                  opcode: Opcode {
                                      name: "PUSH1",
                                      mingas: 3,
                                      inputs: 0,
                                      outputs: 1,
                                  },
                                  inputs: [
                                      Raw(
                                          1,
                                      ),
                                  ],
                              },
                          ),
                      ],
                  },
              ),
          ],
      },
  ]

Solidifying a WrappedOpcode

The WrappedOpcode above essentially breaks down to

  SUB(
      ADD(
          PUSH1(32),
          MLOAD(
              PUSH1(64)
          )
      ),
      SLOAD(
          PUSH1(1)
      )
  )

Which can be simplified further to

  SUB(ADD(32, MLOAD(64)), SLOAD(1))

And could be represented in solidity* as

  (32 + memory[64]) - storage[1]

This process of solidifying (converting to solidity) a WrappedOpcode is done recursively and is done for every WrappedOpcode in the CFG. This allows for the decompiler to generate the solidity code for any given contract.

In heimdall-rs, this process is done in the solidify function of the WrappedOpcode struct. This function can be seen in the following code snippet from heimdall-rs:

  pub fn solidify(&self) -> String {
      let mut solidified_wrapped_opcode = String::new();

      match self.opcode.name.as_str() {
          "ADD" => {
              solidified_wrapped_opcode.push_str(
                  format!(
                      "{} + {}",
                      self.inputs[0]._solidify(),
                      self.inputs[1]._solidify()
                  ).as_str()
              );
          },
          "MUL" => {
              solidified_wrapped_opcode.push_str(
                  format!(
                      "{} * {}",
                      self.inputs[0]._solidify(),
                      self.inputs[1]._solidify()
                  ).as_str()
              );
          },
          "SUB" => {
              solidified_wrapped_opcode.push_str(
                  format!(
                      "{} - {}",
                      self.inputs[0]._solidify(),
                      self.inputs[1]._solidify()
                  ).as_str()
              );
          },
          ...
      }
  }

Analyzing a CFG branch

The CFG branch analysis is done by the analyze function of the VMTrace struct. This function can be viewed in the following code snippet from analyze.rs. This function iterates over the operations in each branch of the VMTrace generated by symbolic execution and performs the bulk of the decompilation through the following steps:

Determining Function Visibility

The decompiler first determines whether or not the current opcode will modify the visibility of the function. In solidity, functions that are pure cannot read or write from the state via the following opcodes:

  BALANCE, ORIGIN, CALLER, GASPRICE, EXTCODESIZE, EXTCODECOPY, BLOCKHASH, COINBASE, TIMESTAMP, NUMBER, DIFFICULTY, GASLIMIT, CHAINID, SELFBALANCE, BASEFEE, SLOAD, SSTORE, CREATE, SELFDESTRUCT, CALL, CALLCODE, DELEGATECALL, STATICCALL, CREATE2

If the current opcode is within this set of state-modifying opcodes, the Function struct has its property pure set to false.

Similarly, functions that are view cannot write to the state via the following opcodes:

  SSTORE, CREATE, SELFDESTRUCT, CALL, CALLCODE, DELEGATECALL, STATICCALL, CREATE2

Again, if the current opcode is within this set of state-modifying opcodes, the Function struct has its property view set to false. This will also set the pure property to false.

Translating Opcodes

The main decompilation process is the translation of opcodes to their corresponding solidity code. The translation only happens for opcodes that directly modify memory or storage, as stack operations will be handled by WrappedOpcode solidifying.

LOGN

The LOG0, LOG1, LOG2, LOG3, and LOG4 opcodes are translated to the corresponding emit statement in solidity.

  1. First, the event is saved to the Function struct for ABI generation.

  2. The data field of the log is decoded into variables via WrappedOpcode::solidify.

     let data_mem_ops: Vec<StorageFrame> = function.get_memory_range(instruction.inputs[0], instruction.inputs[1]);
     let data_mem_ops_solidified: String = data_mem_ops.iter().map(|x| x.operations.solidify()).collect::<Vec<String>>().join(", ");

    This yields us a string of the form memory[0], (1 + memory[1]), ....

  3. Each topic is then solidified via WrappedOpcode::solidify.

    let mut solidified_topics: Vec<String> = Vec::new();
    for (i, _) in topics.iter().enumerate() {
        solidified_topics.push(instruction.input_operations[i+3].solidify());
    }
  4. The proper emit statement is generated with the following format:

     emit Event_<selector>(<solidified_topics>, <data_mem_ops_solidified>);

    The typings and event naming are resolved at a later stage.

JUMPI

The JUMPI opcode is translated to the corresponding if statement in solidity. It's also used to handle pseudo require() statements, which will be fixed in a later version.

  1. First, we check if the JUMPI was not taken in this branch. If it was not taken and the branch REVERTs, we can assume that the JUMPI was a require() statement.
  2. Otherwise, solidify the condition of the JUMPI and generate the if statement.
  3. We also check the conditional to determine if the function is payable. If the conditional is !msg.value, then the function may be payable.

REVERT

The REVERT opcode is translated to the corresponding revert statement in solidity. This straightforward opcode has a few special cases that are handled by the decompiler:

  1. If the REVERT data starts with 08c379a0 (the Error(string) signature), then the revert statement is generated with the corresponding error message. We can decode the error message and generate the revert statement with the following format:

     if (!<condition>) revert("Error message");
  2. If the REVERT data starts with 4e487b71 (the Panic(uint256) signature), the statement is ignored. Since symbolic execution is guaranteed to find all branches, the panics will be included and can be ignored.

  3. If the above two cases are not met, this is a custom error. The decompiler saves these to the function logic as the following:

     if (!<condition>) revert CustomError_<selector>();

    Again, these are saved for ABI generation and will be resolved at a later stage.

RETURN

The RETURN opcode is translated to the corresponding return statement in solidity. We use this opcode to determine the return type of the function with the following heuristics:

  1. If the return data is checked with an ISZERO we can assume that the return type is bool.
  2. If there are bitwise operations on the return data, we can perform our variable size checks to determine potential typings.
  3. For return data over 32 bytes, we can assume that the return type is bytes or string. The return type is also memory.
  4. If all else fails, we assume that the return type is uint256.

SELFDESTRUCT

This is very straightforward, and is translated to the corresponding selfdestruct statement in solidity:

  selfdestruct(<address>);

SSTORE and MSTORE

The SSTORE opcode is translated to the corresponding storage statement in solidity. These are translated as:

  storage[<key>] = <value>;

Since MSTORE essentially does the same thing, we can translate it to the corresponding memory statement in solidity:

  memory[<offset>] = <value>;

CALL, CALLCODE, DELEGATECALL, and STATICCALL

These opcodes are translated to the corresponding call statement in solidity. These are translated as:

  (bool success, bytes memory ret0) = address(<address>).<opcode>{gas: <gas>, value: <value>}(<solidified_memory>);

CREATE and CREATE2

For simplicity, these opcodes are translated as assembly:

  • CREATE:
  assembly { addr := create(<value>, <offset>, <size>) }
  • CREATE2:
  assembly { addr := create(<value>, <offset>, <size>, <salt>) }

CALLDATALOAD

This opcode is used to determine arguments for the function. The following formula is used to determine the argument slot:

  let calldata_slot = (instruction.inputs[0].as_usize() - 4) / 32;

So, for example, if the CALLDATALOAD is at offset 4, then the argument slot is 0. If the CALLDATALOAD is at offset 36, then the argument slot is 1.

We then add this argument to the Function struct, along with some default potential types for the argument (e.g. uint256, bytes32, int256, string, bytes, uint, int). These potential types will be narrowed down later.

ISZERO

If the ISZERO is used on a CALLDATALOAD operation, we can assume that the argument could be within the set {bool, bytes1, uint8, int8}. We add these potential types to the argument.

AND and OR

These operations are checked if they modify the size of an argument via CALLDATALOAD, and will update the potential types of the argument accordingly.

Determining Variable Types

In most programming languages, variables and arguments have a type associated with them. When a program or smart contract is compiled, these types are often removed and replaced with bitwise masking operations. For example, an address in solidity is a 20-byte value. When compiled, the bytecode will often use bitwise masking operations to ensure that the value is exactly 20 bytes. This is done to save space in the bytecode and to make the bytecode more efficient.

We can use this heuristic to infer the types of variables and arguments in a smart contract, for example:

  AND(PUSH20(0xFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF), CALLDATALOAD(4))

is understood by the decompiler to be an argument in the first slot of calldata (since the first 4 bytes of calldata are used for the function selector). We know that this argument is likely a variable with type existing in {address, uint160, bytes20}.

We can use this heuristic to infer the size of any variable or argument in the smart contract. In order to settle on a type within the set of possibilities, we need to watch for how the smart-contract interacts with the value.

For example, the decompiler assumes that the following is an address since it's used as an address further in the program.

  STATICCALL(GAS(), AND(PUSH20(0xFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF), CALLDATALOAD(4)), ...)

Additionally, the decompiler assumes that the following is a uint256 since it's used in arithmetic operations further in the program.

  ADD(PUSH1(0xFF), AND(PUSH20(0xFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF), CALLDATALOAD(4)))

The same can be said for bytes types. For example, the decompiler assumes that the following is a bytes32 since it's used in bitwise operations further in the program.

  BYTE(0, AND(PUSH20(0xFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF), CALLDATALOAD(4)))

Using these heuristics together, we can accurately infer the types of most variables and arguments in any given smart contract. If no heuristics can be applied to a variable, the decompiler looks for any other variables within the expression that can have their type inherited. For example:

  function() external payable {
      uint256 a = 0;
      b = a + 1;
  }

Since a is a typed variable and is used in the instantiation of b, it's typically safe for b to inherit the type of a:

  function() external payable {
      uint256 a = 0;
      uint256 b = a + 1;
  }

Where no type can be inferred, the decompiler defaults to either bytes32 or uint256, as these are the most common 32-byte types in solidity.

Handling Precompiled Contracts

The EVM has a number of precompiled contracts that can perform certain operations, such as recovering the signer of a message. Heimdall-rs has support for these precompiled contracts and can convert the external calls to their corresponding solidity functions.

Handling these is very simple. Whenever the decompiler comes across an external call, it's checked if the to address is a precompiled contract. If it is, the decompiler can decode the calldata and convert the call to a solidity function call.

For example, this simple ECRecovery contract compiles almost perfectly to the following solidity code:

Click to expand
  // SPDX-License-Identifier: MIT
  pragma solidity >=0.8.0;

  /// @title            Decompiled Contract
  /// @author           Jonathan Becker <jonathan@jbecker.dev>
  /// @custom:version   heimdall-rs v0.2.2
  ///
  /// @notice           This contract was decompiled using the heimdall-rs decompiler.
  ///                     It was generated directly by tracing the EVM opcodes from this contract.
  ///                     As a result, it may not compile or even be valid solidity code.
  ///                     Despite this, it should be obvious what each function does. Overall
  ///                     logic should have been preserved throughout decompiling.
  ///
  /// @custom:github    You can find the open-source decompiler here:
  ///                       https://github.com/Jon-Becker/heimdall-rs

  contract DecompiledContract {

      /// @custom:selector    0x19045a25
      /// @custom:name        Unresolved_19045a25
      /// @param              arg0 ["bytes", "bytes32", "int", "int256", "string", "uint", "uint256"]
      /// @param              arg1 ["bytes", "uint256", "int256", "string", "bytes32", "uint", "int"]
      function Unresolved_19045a25(bytes memory arg0, bytes memory arg1) public payable returns (address) {
          bytes memory var_a = var_a + (0x20 + ((0x1f + (arg1) / 0x20) * 0x20));
          if (var_a.length == 0x41) {
              if (!(var_a[0x60]) < 0x1b) {
                  if (var_a[0x60] == 0x1b) {
                      if (var_a[0x60] == 0x1b) {
                          var_a = 0x20 + var_a;
                          uint256 var_d = arg0;
                          bytes1 var_e = var_a[0x60];
                          uint256 var_f = var_a[0x20];
                          uint256 var_g = var_a[0x40];
                          address var_h = ecrecover(var_d, var_e, var_f, var_g);
                          if (!var_h) { revert(); } else {
                              address var_d = var_h;
                              return(var_d);
                          }
                          return(0);
                      }
                      if (var_a[0x60] == 0x1c) {
                          var_a = 0x20 + var_a;
                          uint256 var_d = arg0;
                          bytes1 var_e = var_a[0x60];
                          uint256 var_f = var_a[0x20];
                          uint256 var_g = var_a[0x40];
                          address var_h = ecrecover(var_d, var_e, var_f, var_g);
                          if (!var_h) { revert(); } else {
                              address var_d = var_h;
                              return(var_d);
                          }
                          return(0);
                      }
                  }
                  if (var_a[0x60] + 0x1b == 0x1b) {
                      if (var_a[0x60] + 0x1b == 0x1c) {
                          var_a = 0x20 + var_a;
                          uint256 var_d = arg0;
                          bytes1 var_e = var_a[0x60] + 0x1b;
                          uint256 var_f = var_a[0x20];
                          uint256 var_g = var_a[0x40];
                          address var_h = ecrecover(var_d, var_e, var_f, var_g);
                          if (!var_h) { revert(); } else {
                              address var_d = var_h;
                              return(var_d);
                          }
                          return(0);
                      }
                      if (var_a[0x60] + 0x1b == 0x1b) {
                          return(0);
                          var_a = 0x20 + var_a;
                          uint256 var_d = arg0;
                          bytes1 var_e = var_a[0x60] + 0x1b;
                          uint256 var_f = var_a[0x20];
                          uint256 var_g = var_a[0x40];
                          address var_h = ecrecover(var_d, var_e, var_f, var_g);
                          if (!var_h) { revert(); } else {
                              address var_d = var_h;
                              return(var_d);
                          }
                      }
                  }
              }
              return(0);
          }
      }
  }

A full list of precompiled contracts can be found here.

Recursively Analyze Child Branches

The final step of branch analysis is to analyze the child branches of each VMTrace. Once all branches are analyzed, the logic of the contract should be fully extracted and ready for post-processing.

0x05. Post-Processing

The final step of decompiling is post-processing. This step is responsible for cleaning up the decompiled code and making it more readable. It's also responsible for assigning readable names to variables, as well as resolving function, event, and error selectors.

Resolving Selectors

The first step of post-processing is to resolve selectors for functions, errors, and events. This is done by using Samczsun's Ethereum Signature Database API to resolve the selectors and find their matching signatures. Once we have a list of potential signatures for each selector, we check if they match the arguments of the function, event, or error. If they do, we replace the selector with the signature.

For example, the following selector:

  function Unresolved_19045a25(bytes memory arg0, bytes memory arg1) public payable returns (address) {

would be replaced with:

  function recover(bytes32 arg0, bytes memory arg1) public payable returns (address) {

after resolving the selector.

Building the ABI

Now that we have a list of resolved and unresolved selectors, we can build the ABI for the contract. This is very straightforward, as we just need to build a JSON file with the following structs:

  #[derive(Serialize, Deserialize)]
  struct FunctionABI {
      #[serde(rename = "type")]
      type_: String,
      name: String,
      inputs: Vec<ABIToken>,
      outputs: Vec<ABIToken>,
      #[serde(rename = "stateMutability")]
      state_mutability: String,
      constant: bool,
  }

  #[derive(Serialize, Deserialize)]
  struct ErrorABI {
      #[serde(rename = "type")]
      type_: String,
      name: String,
      inputs: Vec<ABIToken>
  }


  #[derive(Serialize, Deserialize)]
  struct EventABI {
      #[serde(rename = "type")]
      type_: String,
      name: String,
      inputs: Vec<ABIToken>
  }

After building these structs from the list of resolved and unresolved selectors, we can serialize them to JSON and write them to a file, giving us a beautiful and extremely accurate ABI for the contract.

Cleaning Up the Code

We can now assemble the solidity code from the contract logic. Before being finalized, each line is passed through a series of post-processing steps to clean up the code and make it more readable.

  1. Convert all bitwise masks to casts.

    For example:

     (0xffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff) & (arg0);

    would be converted to:

     uint256(arg0);
  2. Simplify Casts

    For example:

     ecrecover(uint256(uint256(arg0)), uint256(uint256(arg0)), uint256(uint256(uint256(arg0))));

    would be simplified as:

     ecrecover(uint256(arg0), uint256(arg0), uint256(arg0));
  3. Convert iszero(...) to !(...)

  4. Simplify parentheses

    For example:

     if (((((((((((((((cast(((((((((((arg0 * (((((arg1))))))))))))) + 1)) / 10)))))))))))))))) {

    would be simplified as:

     if (cast((arg0 * (arg1)) + 1 / 10)) {
  5. Convert all memory accesses to variables. For example:

     memory[0x20] = 0;
     memory[0x40] = memory[0x20] + 0x20;

    would be converted to:

     var_a = 0;
     var_b = var_a + 0x20;
  6. Remove expressions where existing variables can be used

  7. Move all outermost type castings to variable declarations. For example:

     var_a = uint256(arg0);

    would be converted to:

     uint256 var_a = arg0;
  8. Inherit or infer types from existing variables within expressions.

  9. Replace all resolved selectors with their signatures.

  10. Remove unused variable assignments.

Conclusion

That's it! We've now successfully decompiled a contract. The final step is to write the decompiled code to a file, and we're done! Hopefully, this article has shed some light on how decompilers work, and how they can be used to analyze smart contracts. If you have any questions, feel free to reach out to me on Twitter.