Optimizing assembly code

[namasikanam]

Hi, I'm wondering if there is any attempt or experience for optimizing assembly with AlphaEvolve?

[codelion]

Yes! OpenEvolve supports multi-language evolution. Check out the
https://github.com/algorithmicsuperintelligence/openevolve/tree/main/examples/rust_adaptive_sort and
https://github.com/algorithmicsuperintelligence/openevolve/tree/main/examples/r_robust_regression for reference implementations.

Possible setup for Assembly:

Since assembly uses ; for comments (not #), you'll need to use full rewrite mode instead of diff-based evolution:

1. Config (config.yaml):

file_suffix: ".asm"  # or .s for GNU assembler
diff_based_evolution: false  # REQUIRED for assembly - full rewrites instead of diffs

llm:
  system_message: |
    You are an expert in x86-64 assembly optimization.
    Rewrite the entire program to optimize for performance.
    Focus on: instruction-level parallelism, cache optimization,
    register allocation, SIMD instructions, and modern CPU
    microarchitecture (out-of-order execution, branch prediction).

2. Initial Program (initial_program.asm):

; Baseline assembly routine to optimize
section .text
global optimize_me
optimize_me:
    ; your initial implementation
    ret

3. Evaluator (evaluator.py):

# Assemble with nasm/gas
subprocess.run(["nasm", "-f", "elf64", program_path, "-o", "output.o"])
# Link with test harness
subprocess.run(["gcc", "output.o", "test_harness.c", "-o", "benchmark"])
# Run and measure: cycles, instructions, cache misses, throughput

[namasikanam]

Thanks, I see it's technically possible. I guess my question is that: is there any successful cases for optimizing assembly with AlphaEvolve-style approach. The reason is that, assembly language exposes more details and intuitively it's more difficult to optimize for people. So, I'm thinking maybe it's also more difficult here and maybe there're some interesting obstacles which only emerges at assembly level (not at source language level). And, I think it will be really cool if anyone has already had a try in this direction.