OTBN Random Instruction Generator

This directory contains a random instruction generator for OTBN called otbn-rig. This is intended to be run in multiple phases. If you just want to generate some random binaries, it might be easier to call the wrapper at ../uvm/gen-binaries.py.

At the moment, there are two sub-commands: gen and asm. In future, we might add more (to do things like test case mutation or shrinking).

The gen command

The gen command is in charge of actually generating a random program. This program is written to stdout in an JSON format. This format may change over time, but it is understood by the asm command.

Example usage:

hw/ip/otbn/dv/rig/otbn-rig gen --seed 123 --size 1000 >foo.json

To control random generation, there is a --seed parameter. The output should be stable for a fixed seed. If not specified, the seed is zero.

The size parameter

The --size parameter is used to control how big the program grows. A snippet with a single unconditionally executed instruction has a size of one. A snippet containing a sequence of N instructions has a size of N. For a more interesting example, consider an if/else branch that would look something like this in C:

    if (A) {
        B;
    } else {
        C;
    }

If calculating A takes a cycles and B and C have a size of b and c respectively, then the entire snippet has size a + max(b, c). The general idea is that the size taken by a snippet is an upper bound on the number of instructions that it causes the processor to execute. This might not be a strict upper bound: for example, it won’t be strict if a != b in the if/else snippet above.

Snippets

The random program is built from blocks called “snippets”. The simplest possible snippet is a sequence of straight-line instructions. There are also more complicated snippets like conditional branches and loops. The JSON output format represents this tree of snippets.

Initialised data

The generated program is designed never to trigger architecturally unspecified behaviour. It might trigger errors, but its execution will never depend on things like the initial contents of the register file or uninitialised memories.

To provide some data so that RIG can generate load instructions even near the start of the run, the generated program also includes a few words of (randomly) initialised data, scattered around dmem.

The asm command

Once a random program has been generated, the asm command can be used to convert it to an assembly listing. This, in turn, can be assembled and linked using the toolchain in hw/ip/otbn/util.

Unlike the gen command, this step does no random generation: it’s a deterministic translation from the JSON input to assembly and linker script output.

Example usage:

hw/ip/otbn/dv/rig/otbn-rig asm --output foo foo.json

When given the --output parameter, this command generates two output files. With --output foo, it generates foo.s and foo.ld. These are an assembly listing and a linker script, respectively.

To assemble and link, use commands like:

hw/ip/otbn/util/otbn-as -o foo.o foo.s
hw/ip/otbn/util/otbn-ld -o foo.elf -T foo.ld foo.o

This is automated in the gen-binaries.py wrapper described above.

Occasionally, it’s helpful to just read the assembly listing for a JSON file. To do this, run the command with no --output parameter to see the assembly listing on stdout. The linker script will not be generated.