Abstract
Just-in-Time (JIT) compilation in web browsers is critical to the performance of modern web 
applications. These compilers typically rely on runtime profiling to identify frequently executed 
(hot) functions before applying optimizations, yet this process introduces overhead especially on 
first execution. Although widely used techniques—such as per-application caching and simple 
threshold-based recompilation—partially mitigate these issues, they often fall short for
loop-intensive workloads with unpredictable runtime characteristics. To address these challenges, 
this thesis proposes a lightweight speculative optimization framework—LoopOracle. The framework 
embeds lightweight machine learning models and survival analysis directly into inline caches (ICs). 
The thesis considers two optimization targets: JIT recompilation and inlining decisions. Similar to 
how traditional techniques sample code execution to guide compilation, LoopOracle observes loop 
patterns and probabilistically infers whether a function is likely to run often enough to warrant 
an alternative strategy compared to the baseline compiler. By integrating these predictive models 
within the existing IC mechanism, the framework integrates training, monitoring, inference and 
model switching seamlessly into the compiler pipeline. This thesis presents a concrete 
implementation in Mozilla’s SpiderMonkey compiler. To evaluate the framework, empirical evaluations 
on two industry-standard benchmarks—Jetstream 2.2 (emphasizing computationally expensive web 
applications) and Speedometer 3.0 (focusing on DOM interaction)—are presented in contrast to 
industry established approaches such as JitHints.
Committee:

● Prof. Santosh Pande – School of Computing (advisor)
● Prof. Ada Gravilovska Habl – School of Computing
● Matthew Gaudet – Staff Engineer, SpiderMonkey Team, Mozilla C