0.3.2 Aug 23, 2022
0.3.1 Aug 13, 2022
0.2.1 Jul 22, 2022
0.1.6 Jul 5, 2022
0.1.4 May 24, 2022

#57 in #verifier

Apache-2.0

1.5MB
32K SLoC


id: bytecode-test-generation title: Bytecode Test Generation Tool custom_edit_url: https://github.com/move-language/move/edit/main/language/testing-infra/test-generation/README.md

Bytecode Test Generation Tool

Overview

This tool can generate known-valid or known-invalid Move bytecode programs. Known-valid means that the program is valid by construction; it was constructed according to a formal specification of the Move bytecode language. Known-invalid bytecode programs are those that diverge from the specification in a controlled way.

The generated modules are checked by the Move bytecode verifier and then run on the VM runtime. If the module contains known-valid bytecode, we should expect that the program will pass the bytecode verifier. If it does not pass the verifier then this indicates that either there is a bug in the bytecode verifier or the formal specification of the failing instruction is incorrect. Likewise, it should also run successfully on the VM runtime.

If the module contains known-invalid bytecode, we should expect that the verifier will reject the module. If the verifier does not reject the module this indicates again that either there is a bug in the bytecode verifier or the formal specification is incorrect. Likewise for the VM runtime.

Usage

Building and Configuration

To build the tool

  • cargo build

There are configuration options (and explanations for their behavior) in src/config.rs. Most of the time these should be left as the provided defaults.

Running

The tool expects up to two arguments:

  • --iterations: The number of programs that should be generated and tested
  • --output: (Optional) If provided, this is the path to which modules that result in errors will be serialied and saved. If not provided, failing cases will be logged to the console.

Additionally, there is an optional flag RUST_LOG that controls the verbosity of debug information. It goes from error, the least information, to debug the most information. The most common setting for this flag is info which will print stats such as the number of iterations, the number of verified and executed programs, etc.

To run the tool

  • RUST_LOG=info cargo run -- --iterations N --output PATH

Architecture

This tool works by modeling the state of the VM abstractly, and by modeling the bytecode instructions in terms of that abstract state. The abstract state is defined in abstact_state.rs. It consists of type-level modeling of the VM stack, locals, and borrow graph.

Instructions are defined in terms of their preconditions and effects. These definitions are found in summaries.rs and use macros defined in transitions.rs. The preconditions of an instruction are predicates that are true or false for a given abstract state. For example the Bytecode::Add instruction requires the stack to contain two integers. This is modeled by saying that the preconditions of Bytecode::Add are

  • state_stack_has!(0, Some(SignatureToken::U64))
  • state_stack_has!(1, Some(SignatureToken::U64)) where indexes 0 and 1 refer to the top two elements of the stack.

The effects of an instruction describe how the instruction modifies the abstract state. For Bytecode::Add the effects are that it performs two pops on the stack and pushes a SignatureToken::U64 to the stack.

In this way, we are able to fully capture what each instruction needs and does. This information is used to generate valid bytecode programs.

Generation of bytecode programs proceeds as follows:

  1. In lib.rs the generator loop begins by initializing a ModuleBuilder
  2. The ModuleBuilder, defined in ../utils/src/module_generator.rs, generates a module definition
  3. The ModuleBuilder calls the generator defined in bytecode_generator.rs to fill in function bodies within the module
  4. The generator builds a control flow graph (CFG) in control_flow_graph.rs. Each block of the CFG is assigned a valid starting and ending abstract state.
  5. The generator fills in blocks of the CFG according to the following algorithm:
    1. Given starting abstract state AS1, let candidates be the list of all instructions whose preconditions are all satisfied in AS1
      • If invalid generation is desired, then let x preconditions be false
    2. Select candidate instr from candidates according to the stack height heuristic
      • The stack height heuristic selects instructions that add to the stack when the height is small, and instructions that subtract from the stack when the height is large
    3. Apply the effects of instr to AS1, producing AS2
    4. If the stack is empty, terminate, otherwise repeat from step a with AS2

This results in the generation of one module. The module is then given to the bytecode verifier and the bytecode verifier's behavior (i.e. no verification errors, some verification errors, panic, crash) is recorded. Likewise with the VM runtime (except with runtime errors rather than verification errors). Depending on the configuration that this tool is built with, if the module caused a verification error, panic, or crash the module will then be printed out or serialized to disk.

This will continue for the number of iterations specified when invoking the tool.

Other files:

  • error.rs defines an error struct which is used to pass error messages.
  • tests/ contains a set of files that test the preconditions and effects of each bytecode instruction

Extending the tool

The most common change or extension to this tool will probably be changing instruction preconditions and effects. To do that follow these steps:

  1. See if there already a macro defined in transitions.rs that captures your desired precondition/effect
  2. If the macro is already defined, just add it to the summary of the instruction being changed in summaries.rs
  3. If a suitable macro does not exist, define it in transitions.rs. Look at other macros in that file for examples.
  4. Macros in transitions.rs have access to the public fields and functions of the AbstractState. If your macro needs access to something more, add a new helper method in abstract_state.rs and then invoke it in the macro.

Dependencies

~28–42MB
~645K SLoC