| name | mlir-evolve |
| description | MLIR/LLVM/IREE workflow for the mlirAgent repo. Use when working in this repo to build or configure toolchains, analyze/debug passes (provenance/rlm), run evolutionary optimization, manage the Neo4j/LanceDB knowledge graph, or mine compiler recipes. |
mlirAgent
Overview
Use this skill to navigate and extend the mlirAgent toolchain and research workflows.
Quick Start
- Reuse existing wrappers in
src/mlirAgent/tools/*.py before creating new ones.
- Use
src/mlirAgent/config.py for paths and external endpoints (Neo4j, LanceDB).
- Install the package with
pip install -e . for proper imports (no sys.path hacks).
Core Capabilities
1) Build + Toolchain Management
- Use
src/mlirAgent/tools/build.py for Ninja/CMake builds.
- Prefer Config-driven paths and flags.
2) Pass Analysis + Provenance
- Prefer the structural tracer in
src/mlirAgent/tools/provenance.py.
- Use
src/mlirAgent/tools/trace_provenance.py only when bindings are unavailable.
- For RLM-based analyses, see
src/mlirAgent/rlm/analysis.py.
3) Evolutionary Optimization
- Evolve tasks live under
src/mlirAgent/evolve/tasks/.
- Available tasks:
llvm_inlining/, llvm_bench.py, regalloc_priority/.
- Configuration via
configs/ directory.
4) Recipe Mining
- Mining pipeline in
src/mlirAgent/mining/.
- Synthesize recipes with
synthesize_recipes.py (requires a model_fn callable).
5) Knowledge Graph (SCIP + Neo4j)
- SCIP ingestion scripts in
src/mlirAgent/scip/.
- Graph status, inspection, and impact analysis tools included.
Conventions
- Write outputs to
data/ or data/artifacts/ to keep experiments reproducible.
- All file paths should come from
Config or environment variables, not hardcoded.
