Empirical Validation of Shadow-Price-Guided Inlining in MIR

Bilar, Daniyel Yaacov

doi:10.5281/zenodo.18828679

Published March 1, 2026 | Version 1.0.0

Preprint Open

Empirical Validation of Shadow-Price-Guided Inlining in MIR

Bilar, Daniyel Yaacov (Researcher)¹

1. Chokmah LLC

Full experimental validation of dual decomposition protocol for JIT compiler inlining on the MIR lightweight JIT compiler. Implements a three-phase pipeline: Phase 1 synthetic profiling (execution frequencies derived from benchmark structure), Phase 2 shadow-price-guided mutation pass (MIR_CALL → MIR_INLINE promotion), Phase 3 wall-clock benchmarking across five conditions (no inlining, blind inline-all, random 50%, shadow-price, inverted-price). Includes two benchmark suites (3-function and 8-function workloads), raw timing data (20 runs × 5 conditions per benchmark), statistical analysis (Cohen's d effect sizes), and publication-quality figures (PNG + SVG). Key results: P3 (shadow vs. inverted-price) confirmed with d = 2.65–4.42 across both benchmarks; shadow-price matches blind inline-all performance with half the mutations on the 8-function benchmark (4 vs. 8 promotions, d = −0.02); P1 (shadow vs. random) falsified on 3-function benchmark due to fixed RNG seed. Companion to theory paper (DOI: 10.5281/zenodo.18715390) and Paper 1 empirical consistency results (DOI: 10.5281/zenodo.18735516).

Notes (English)

This research was conducted with AI assistance from multiple large language models operating under the author's direction. Anthropic's Claude (Opus) served as the primary research partner for source code analysis, experimental design review, statistical methodology, and manuscript preparation. Anthropic's Claude (Sonnet) provided independent manuscript review, identifying methodological concerns including the fixed-RNG-seed limitation in the random condition, the oracle nature of synthetic profiling, and the distinction between prospective intervention and post-hoc reinterpretation in the Kelly/TCP analogy; these critiques led to substantive revisions in §5.2, §5.4, and §5.6. Google's Gemini ("Nova the Optimal AI") contributed proof-of-concept code generation, MIR IR construction, and engineering brainstorming. Additional AI systems contributed to early-stage ideation and literature review. All experimental decisions, hypothesis formulation, statistical interpretations, and scientific claims are the sole responsibility of the author. AI contributions are disclosed transparently in accordance with scientific integrity norms.

Files