slothy 0.2.1

SLOTHY: Assembly superoptimization via constraint solving

SLOTHY: Assembly optimization via constraint solving

About SLOTHY

SLOTHY - Super (Lazy) Optimization of Tricky Handwritten assemblY - is an assembly-level superoptimizer for:

1. Instruction scheduling
2. Register allocation
3. Software pipelining (= periodic loop interleaving)

SLOTHY is generic in the target architecture and microarchitecture. This repository provides instantiations for:

• Armv8.1-M+Helium: Cortex-M55, Cortex-M85
• AArch64: Cortex-A55, and experimentally Cortex-A72, Cortex-X/Neoverse-V, Apple M1 (Firestorm, Icestorm)
• RISC-V (RV{32,64}IMB): XuanTie C908

SLOTHY is discussed in Fast and Clean: Auditable high-performance assembly via constraint solving.

Goal

SLOTHY enables a development workflow where developers write 'clean' assembly by hand, emphasizing the logic of the computation, while SLOTHY automates microarchitecture-specific micro-optimizations. This accelerates development, keeps manually written code artifacts maintainable, and allows to split efforts for formal verification into the separate verification of the clean code and the micro-optimizations.

How it works

SLOTHY is essentially a constraint solver frontend: It converts the input source into a data flow graph and builds a constraint model capturing valid instruction schedulings, register renamings, and periodic loop interleavings. The model is passed to an external constraint solver and, upon success, a satisfying assignment converted back into the final code. Currently, SLOTHY uses Google OR-Tools as its constraint solver backend.

Performance

As a rough rule of thumb, SLOTHY typically optimizes workloads of <50 instructions in seconds to minutes, workloads up to 150 instructions in minutes to hours, while for larger kernels some heuristics are necessary.

Applications

SLOTHY has been used to provide the fastest known implementations of various cryptographic and DSP primitives: For example, the SLOTHY paper discusses the NTTs underlying ML-KEM and ML-DSA for Cortex-{A55, A72, M55, M85}, the FFT for Cortex-{M55,M85}, and the X25519 scalar multiplication for Cortex-A55. You find the clean and optimized source code for those examples in paper/.

Getting started

Have a look at the SLOTHY tutorial for a hands-on and example-based introduction to SLOTHY.

A more extensive tutorial including assignments was taught at CHES 2025 and can be found here with slides available here.

Real world uses

• AWS libcrypto (AWS-LC): SLOTHY-optimized X25519 code based on our un-interleaved form of the original code by Emil Lenngren has been formally verified and included in s2n-bignum (the bignum component of AWS-LC) and merged into AWS-LC. This was the topic of a Real World Crypto 2024 talk.

• s2n-bignum routinely employs SLOTHY for finding further highly optimized ECC implementations (e.g., P256, P384, P521 and verifies them through automated equivalence-checking in HOL-Light.

• Arm EndpointAI: SLOTHY-optimized code has been deployed to the CMSIS DSP Library for the radix-4 CFFT routines as part of the Arm EndpointAI project in this commit.

• mlkem-native: AArch64 assembly routines of ML-KEM are automatically optimized using SLOTHY.

• pqm7: Benchmarking framework for the Arm Cortex-M7 which has been created as a case study for automated microarchitectural migrations of software libraries using SLOTHY. See our paper for more details.

Installation

Option 1: Install from PyPI

The easiest way to install SLOTHY is via pip:

pip install slothy

Quick Start

Here's a minimal example of using SLOTHY to optimize assembly code:

import slothy
import slothy.targets.aarch64.aarch64_neon as AArch64_Neon
import slothy.targets.aarch64.cortex_a55 as Target_CortexA55

# Create SLOTHY instance for ARM Cortex-A55
s = slothy.Slothy(AArch64_Neon, Target_CortexA55)

# Load assembly code from file
s.load_source_from_file('example.s')

# Optimize the code
s.optimize(start='start_label', end='end_label')

# Write optimized assembly to file
s.write_source_to_file('example_optimized.s')

print("Optimization complete! Check example_optimized.s")

Example assembly file (example.s):

start_label:
    ldr x0, [x1]
    ldr x2, [x3]
    add x4, x0, x2
    str x4, [x5]
end_label:

Option 2: Development Installation

For development or to run the examples in this repository, first clone the SLOTHY development repository:

git clone https://github.com/slothy-optimizer/slothy.git

Requirements

SLOTHY has been successfully used on

• Ubuntu-21.10 and up (64-bit),
• macOS Monterey 12.6 and up.

SLOTHY supports Python 3.10 up to 3.14. For development Python >= 3.11 is required. See requirements.txt for package requirements, and install via pip install -r requirements.txt.

Note: requirements.txt pins versions for reproducibility. If you already have newer versions of some dependencies installed and don't want them downgraded, consider using a virtual environment:

python3 -m venv venv
./venv/bin/python3 -m pip install -r requirements.txt

Then, enter the virtual environment via source venv/bin/activate prior to running SLOTHY. Finally, adjust your PATH environment variable to include the directories containining the slothy-cli script and the LLVM llvm-mca tool.

Verify Development Installation

To check that your development setup is complete, try the following from the base directory:

% python3 test.py --tests aarch64_simple0_a55

You should see something like the following:

* Example: aarch64_simple0_a55...
INFO:aarch64_simple0_a55:SLOTHY version: 0.1.0
INFO:aarch64_simple0_a55:Instructions in body: 20
INFO:aarch64_simple0_a55.slothy:Perform internal binary search for minimal number of stalls...
INFO:aarch64_simple0_a55.slothy:Attempt optimization with max 32 stalls...
INFO:aarch64_simple0_a55.slothy:Objective: minimize number of stalls
INFO:aarch64_simple0_a55.slothy:Invoking external constraint solver (OR-Tools CP-SAT v9.7.2996) ...
INFO:aarch64_simple0_a55.slothy:[0.0721s]: Found 1 solutions so far... objective 19.0, bound 8.0 (minimize number of stalls)
INFO:aarch64_simple0_a55.slothy:[0.0765s]: Found 2 solutions so far... objective 18.0, bound 12.0 (minimize number of stalls)
INFO:aarch64_simple0_a55.slothy:OPTIMAL, wall time: 0.155224 s
INFO:aarch64_simple0_a55.slothy:Booleans in result: 509
INFO:aarch64_simple0_a55.slothy.selfcheck:OK!
INFO:aarch64_simple0_a55.slothy:Minimum number of stalls: 18

Option 3: Docker

A dockerfile for an Ubuntu-22.04 based Docker image with all dependencies of SLOTHY and the PQMX+PQAX test environments setup can be found in paper/artifact/slothy.dockerfile. See paper/artifact/README.md for instructions.

Examples

The SLOTHY Tutorial and the examples directory contain numerous exemplary assembly snippets. To try them, use python3 example.py --examples={YOUR_EXAMPLE}. See python3 example.py --help for the list of all available examples.

The use of SLOTHY from the command line is illustrated in paper/scripts/ supporting the real-world optimizations for the NTT, FFT and X25519 discussed in Fast and Clean: Auditable high-performance assembly via constraint solving.

Contributing

For information on how to contribute to SLOTHY, please see CONTRIBUTING.md.