tarunnagarajan/compiler-opt

How tarunnagarajan/compiler-opt works

This is a research framework for hierarchical reinforcement learning of LLVM optimization policies. The main workflow selects an IR benchmark, extracts state, chooses a strategic macro-action and tactical pass refinement, applies compiler passes, then measures the result for the next decision and reward. Training and evaluation scripts support the research loop, including world-model prediction and baseline comparisons. Relationships among components are limited to those stated in the README or visible in indexed imports; several model and measurement internals were not sampled.

Main entry points: compiler_env.py RL compiler environment · hrl_agent.py Hierarchical policy · world_model.py World model · scripts Policy and model training

Components 25

Compiler environment

RL compiler environment src/env/compiler_env.py
LLVM and action config src/config.py
Hardware profiler src/env/hardware_profiler.py

IR representation

IR parser and CFG src/features/ir_parser.py
Scalar features src/features/feature_vector.py
IR graph extraction src/features/ir_graph_extractor.py
Loop-aware extraction src/features/ir_graph_extractor_v5.py
Graph encoder src/features/graph_encoder.py

Learning models

World model src/models/world_model.py
Scale-aware world model src/models/world_model_v7.py
Model graph encoder src/models/graph_encoder.py
Graph encoder V6 src/models/graph_encoder_v6.py
Hierarchical policy src/models/hrl_agent.py
Tactical worker src/models/hrl_worker.py
Pass history encoder src/models/history_encoder.py
Meta calibrator src/models/meta_calibrator.py

Passes and metrics

Macro actions src/actions/macro_actions.py
Micro actions src/actions/micro_actions.py
Pass executor src/passes/pass_executor.py
Metrics and runtime src/passes/metrics.py

Research workflows

Policy and model training scripts
Evaluation and audits scripts

Other components

Researcher
Benchmark inputs
LLVM toolchain

Connections 26

Researcher → RL compiler environment: starts episode README.md
Benchmark inputs → RL compiler environment: supplies IR README.md
RL compiler environment → LLVM and action config: uses configuration src/env/compiler_env.py
RL compiler environment → IR parser and CFG: parses IR src/env/compiler_env.py
RL compiler environment → Pass executor: applies passes src/env/compiler_env.py
RL compiler environment → Metrics and runtime: measures outcomes README.md
RL compiler environment → LLVM toolchain: invokes compiler README.md
RL compiler environment → Hardware profiler: profiles hardware
IR parser and CFG → Scalar features: provides IR stats src/features/feature_vector.py
RL compiler environment → Graph encoder V6: uses graph encoder src/env/compiler_env.py
World model → Scale-aware world model: uses V7 model src/models/world_model.py
Scale-aware world model → Graph encoder V6: uses encoder src/models/world_model_v7.py
Hierarchical policy → Tactical worker: selects refinements README.md
Hierarchical policy → Pass history encoder: encodes pass history README.md
Hierarchical policy → World model: predicts action effects
Hierarchical policy → Macro actions: chooses objectives README.md
Tactical worker → Micro actions: selects pass refinements src/models/hrl_worker.py
RL compiler environment → Macro actions: accepts macro actions src/env/compiler_env.py
RL compiler environment → Micro actions: accepts refined passes src/env/compiler_env.py
Metrics and runtime → LLVM toolchain: compiles benchmarks src/passes/metrics.py
Meta calibrator → World model: calibrates predictions
Policy and model training → RL compiler environment: trains against README.md
Policy and model training → Hierarchical policy: trains policy README.md
Policy and model training → World model: trains model README.md
Evaluation and audits → Hierarchical policy: evaluates behavior README.md
Evaluation and audits → LLVM toolchain: compares baselines README.md

More LLVM diagrams: tpde2/tpde · arm/arm-toolchain · inside-compiler/inside-llvm-code-gen · grqz/llvm-project · assemblew/hikari-llvm

Drawn by GitDiagram on September 30, 2026 from the repository's file tree, README and source files. Read as Markdown ·