Taking a Closer Look: An Outlier-Driven Approach to Compilation-Time Optimization
This program is tentative and subject to change.
Improving compilation time in optimizing compilers is challenging due to their large number of interconnected components. This includes compiler optimizations, compiler tiers, heuristics, and profiling information. Despite this complexity, research in compilation-time optimization is often guided by analyzing metrics of entire program runs, such as the total compilation time and overall memory footprint. This coarse-grained perspective hides relevant information, such as source program functions for which the compiler allocates a lot of memory or compiler optimizations with a high impact on the total compilation time. This leaves high-level metrics as the only reference point for driving optimization design. Consequently, compilation-time regressions in one program function that are obscured by improvements in other functions stay undetected, while the impacts of compiler changes on untouched parts of the compiler are mainly unknown. Furthermore, developers overlook long-standing compiler defects because their high-level metrics do not change over time.
To address these limitations, we propose ICON, a new data-driven approach to compilation-time optimization that breaks up high-level metrics into individual source program functions, compiler optimizations, or even into individual instructions in the compiler source code. Our methodology enables an iterative in-depth compilation-time analysis, focusing on outliers to identify optimization opportunities. We show that outliers, both in terms of time spent in a particular compiler optimization, and in terms of individual compilations that take substantially longer, can reveal potential problems in the compiler implementation. We applied our approach to GraalVM and extracted data for multiple of its language runtimes. We analyzed the resulting data, present the first detailed look into the distribution of compilation time in the GraalVM compiler, a state-of-the-art multi-language compiler, and identified defects that led to regressions in overall compilation time or the compilation time of specific languages. We furthermore designed two optimizations based on the identified outliers that improve compilation time between 2.25% and 9.45%. We believe that our approach can guide compiler developers in finding usually overlooked optimization potential and defects, and focus future research efforts in making compilers more efficient.
This program is tentative and subject to change.
Mon 16 SepDisplayed time zone: Amsterdam, Berlin, Bern, Rome, Stockholm, Vienna change
10:30 - 12:00 | |||
10:30 15mTalk | Static Basic Block Versioning Technical Papers Manuel Serrano Inria; University of Côte d'Azur, Olivier Melançon Université de Montréal, Marc Feeley Université de Montréal | ||
10:45 15mTalk | Cross Module Quickening - The Curious Case of C Extensions Technical Papers Felix Berlakovich μCSRL, CODE Research Institute, University of the Bundeswehr Munich, Stefan Brunthaler Bundeswehr University Munich | ||
11:00 15mTalk | Compiling with Arrays Technical Papers David Richter Technical University of Darmstadt, Timon Böhler Technical University of Darmstadt, Pascal Weisenburger University of St. Gallen, Mira Mezini TU Darmstadt Pre-print | ||
11:15 15mTalk | The Performance Effects of Virtual-Machine Instruction Pointer Updates Technical Papers | ||
11:30 15mTalk | Taking a Closer Look: An Outlier-Driven Approach to Compilation-Time Optimization Technical Papers Florian Huemer JKU Linz, David Leopoldseder Oracle Labs, Aleksandar Prokopec Oracle Labs, Raphael Mosaner JKU Linz, Hanspeter Mössenböck JKU Linz | ||
11:45 15mTalk | Optimizing Layout of Recursive Datatypes with Marmoset Technical Papers Vidush Singhal Purdue University, Chaitanya S. Koparkar Indiana University, Joseph Zullo Purdue University, Artem Pelenitsyn Purdue University, Michael Vollmer University of Kent, Mike Rainey Carnegie Mellon University, Ryan R. Newton Purdue University, Milind Kulkarni Purdue University Pre-print |