The topic of this thesis was challenging, requiring the student to study the state of the art in both compilers and machine learning. The student met all the goals of the thesis assignment and successfully delivered a working prototype. However, the results indicate that further work is needed before the prototype can be integrated into the official release of GraalVM Native Image. That being said, I am overall satisfied with the outcome of the work.

The resulting tool has been open-sourced on GitHub and is expected to serve as an inspiration for future work in this area.

The student demonstrated a responsible approach, consistently attended meetings, and was always well prepared.

Stupeň hodnocení: zadání splněno

The presentation level of the report is not very good. The text is sometimes quite hard to read, many parts are confusing and not well linked together. Often, there are figures and tables which are not referenced from the text or are referenced only several pages later. The division into chapters is weird, for instance Chapters 5 and 6 describe the two different neural network layouts that the student experimented with, however, Chapter 5 is titled "Implementation" while Chapter 6 is titled "Inlining Based on Graph Neural Networks". Also, the description of the underlying concepts (Native Image, function inlining, machine learning principles and algorithms) is not unified and is scattered among the individual chapters. Overall, the structure of the text could use a lot of improvement which would greatly help understanding the thesis and would improve the overall impression of the thesis.

The student has explored a decent amount of related work, especially on compiler optimizations and on using machine learning within their context. However, I would expect and welcome a more detailed description (preferably an entire section) of the relevant approaches, possibly including a comparison with the author's proposed approach.

Unfortunately, the proposed results do not seem to be usable in practice. It turned out that the simple naive feed-forward network performs pretty much the same as the complex graph convolutional network which should, at least in theory, be much better. From my perspective, this is caused by a combination of the selection of a very simple fitness function in the genetic algorithm and the introduction of artificial guardrails which did not allow the algorithm to explore a sufficient variety of network configurations. On top of that, it is unclear whether changing some of these properties of the algorithm would lead to inferrence of networks which would perform better than the reference inlining heuristic. Despite that, since this field is very much unexplored, even this result can have certain value for further work.