/TestRIG

Testing processors with Random Instruction Generation

Primary LanguagePythonOtherNOASSERTION

TestRIG

Framework for testing RISC-V processors with Random Instruction Generation.

Glorious vision

TestRIG is a framework for RISC-V processor verification using the RVFI-DII (pronounced "rividy") interface. TestRIG supports two types of components:

  1. Vengines (verification engines)
  2. Implementations (including models, simulators, and SoCs)

Vengines generate one or more DII streams of instruction traces, and consume one or more RVFI streams of execution traces. Implementations consume a DII instruction trace, and generate an RVFI execution trace. A Vengine can, for example, produce two equivalent DII streams for two implementations and check that the two generated RVFI streams are equivalent.

TestRIG eliminates the "test gap" between specification and implementation. If the specification is an executable model (see Sail, L3, and many other efforts), then TestRIG allows automated verification of any specified property without passing through human interpretation and hand-writing tests. While TestRIG should require work to construct trace generators for classes of ISA behaviour, new instructions in a class can automatically be included in new traces and will be run in many more variations than hand-written tests would allow.

TestRIG verifies the pipeline as well as specification compliance. Test suites are designed to test as many aspects of an instruction set specification as possible, but cannot be expected to provide any reasonable verification of complex pipeline behaviour. TestRIG can verify every register value read in the pipeline under random sequence generation, while a test suite will only report a prescribed test result.

TestRIG greatly increases debugging efficiency. In-memory test suites require a significant amount of boiler plate in order to construct a valid test state which cannot be automatically reduced without disturbing instruction layout in complex ways. As TestRIG relies on direct instruction injection, bypassing fetch through PC, a sequence of instructions can easily be shortened by simply eliminating instructions from the trace to see if we still find divergence. As a result we can expect automatically reduced counterexamples on the order of a handful of instructions.

We hope that TestRIG will prove fertile ground to inspire innovation in instruction trace generation, counterexample reduction, and model construction such that trace-based verification can approach formal verification for many practical purposes with a much friendlier user experience than either test-suite verification or formal verification currently enable.

Vengines

A vengine typically includes instruction trace (itrace) generators targeting various classes of behaviour. The vengine then feeds these itraces into both a model and an implementation through two TCP sockets in the RVFI-DII format. The model and implementation return an RVFI-DII execution trace (etrace) that details the state observation and state change of each instruction. The vengine compares these two traces and identifies any divergence between the model and the implementation. Any failure is reported nicely to the user with a means of conveniently replaying the failing itrace for debugging. A capable vengine will also attempt to reduce the failing trace to a minimal example that diverges.

Implementations (including Models)

Any implementation that wants to be verified using TestRIG will need to support RVFI-DII itrace format as an instruction source and etrace reporting format. The implementation should have a mode where instructions are consumed exclusively from the RVFI-DII instruction port, bypassing any actual instruction memory, ignoring the architectural program counter. An implementation should then deliver a trace report at the end of execution detailing implementation behaviour in response to that instruction in the RVFI-DII format. The RVFI-DII communication uses a single socket with the itrace consumed and the etrace be delivered over the same socket. Please look at the RVFI-DII specification for more details

Currently, the provided modules are:

Getting started

In order to get the different submodules provided by TestRIG, run the following command:

$ git submodule update --init --recursive

The root makefile can currently build the QuickCheck Verification Engine, Spike, and the Sail implementation. Both Spike and Sail are built without CHERI support by default and Sail is built as a 32-bit version.

Dependencies

The dependencies for the QuickCheck Verification Engine are cabal and ghc, both of which you can install using GHCup. For cabal you must install version 3.4 or later to use the list-bin command.

The dependencies for Spike are:

  • sudo apt-get install device-tree-compiler

The dependencies for RVBS are the Bluespec compiler bsc. You can follow the build instructions on the B-Lang-org github.

The dependencies for the Sail model can be installed using opam by following the instructions from the Sail wiki. You may have to run opam update and opam upgrade occasionally if sail gets updated.

The dependencies for Ibex are verilator:

  • sudo apt-get install verilator

Toooba depends on the Bluespec compiler (see the dependencies for RVBS) and verilator. Toooba needs the version of verilator to be greater than: Verilator 3.922 2018-03-17 rev verilator_3_920-32-gdf3d1a4

Default Configuration

You can verify a default configuration by executing:

$ make
$ utils/scripts/runTestRIG.py

Custom Configurations

Look at the Makefile to see different targets to compare against each other. Also use the following command to figure out the different options for running TestRIG:

$ utils/scripts/runTestRIG.py --help

64-bit: Sail vs Toooba

Executing the following commands will compare Sail and Toooba with the 64-bit version of the RISC-V instruction set, assuming that you've initialized the submodules and have installed all the dependencies described above.

$ make vengines
$ make sail-rv64
$ make toooba-rv64
$ utils/scripts/runTestRIG.py -a sail -b toooba -r rv64i

CHERI 32-bit: Sail vs Ibex

Executing the following commands will compare Sail and Ibex with the 32-bit version of the RISC-V instruction set (including compressed) and CHERI extensions enabled, assuming that you've initialized the submodules and have installed all the dependencies described above.

$ make vengines
$ make sail-rv32-cheri
$ make ibex-rv32ic-cheri
$ utils/scripts/runTestRIG.py -a sail -b ibex -r rv32icxcheri