Jasper Galton-Fenzi

← All projects

Arm MCP Server: open-source fix

October 2026 · merged into arm/mcp main

• Python
• Docker
• LLVM (llvm-mca)
• MCP Inspector
• pytest
• Git

What I did

Arm publishes an MCP server: a set of tools an AI assistant can call, including mca, which predicts how fast a piece of assembly runs on an Arm CPU. It was my first time using Arm software, so I ran it locally in Docker and called each tool directly with MCP Inspector, with no AI in the loop, to see what it really does.

The bug

Filling in the tool's triple option (which Arm CPU family to target) always failed:

llvm-mca: Unknown command line argument '--triple'.
llvm-mca: Did you mean '--mtriple'?
llvm-mca: Too many positional arguments specified!

The response showed the exact command the server had run, which led me to one line in llvm_mca_tool.py: it passed --triple, but llvm-mca's flag is --mtriple. Because the flag wasn't recognised, the value after it was read as a second input file. The option had never worked since the first release, and no test covered it.

The fix and the test

• Changed one word, --triple to --mtriple, and opened a pull request.
• Arm's reviewers asked for a regression test, so I added one to their integration suite. It calls mca with triple="aarch64" and cpu="neoverse-v2", and checks both that the call succeeds and that --mtriple is in the command.
• Built their Docker image locally with my change and ran their full test suite: all passed, including 50/50 knowledge-base smoke questions.
• Ran the new test against the released image without my fix: it failed as expected. A test that can't fail isn't worth much.
• Arm moved the change through their internal review process and merged it into main.

Using the tool: -O0 vs -O2

I also used llvm-mca to compare a simple C loop compiled with optimisation off (-O0) and on (-O2), on a Neoverse V2 model:

Comparison	-O0	-O2
Whole function	474 cycles	403 cycles (looks ~15% faster)
Loop body, per number added	3.08 cycles	0.26 cycles (~12× faster)

The whole-function number hides the real gain, because llvm-mca runs every line once per pass and the -O2 code has extra setup around a vector loop. One surprise: IPC went down while the code got faster, because one vector instruction does four additions. IPC measures how busy the CPU is, not how much useful work it does.

What I learned

• Test the tool, not the description of the tool.
• Fork, branch, pull request, merge conflicts, and how a large project reviews outside contributions.
• A good regression test fails without the fix and passes with it.