Selectively-Amortized Resource Bounding

Abstract

Managing resource usage has become a critical concern for software developers as software development continues to grow in complexity. This thesis focuses on the practical aspects of automatically proving resource bounds by investigating how to bound an integer-valued resource variable by a given program expression. The automatic resource bound analysis has become increasingly important in detecting performance bugs, preventing algorithmic-complexity attacks, and identifying side-channel vulnerabilities. This thesis puts forth a novel paradigm of selectively-amortized resource bounding for practical yet sound analysis of resource usage.The most simple and common approach, which we refer to as the worst-case reasoning, involves reducing the problem to the reachability bound problem. This method infers the maximum number of times a location can be visited, multiplies it by the worst-case cost at this location among all visits, and sums up such products for all locations. However, this approach fails to work for non-terminating programs like reactive programs, as the number of times a location can be reached may be unbounded.To address this limitation, we adapt worst-case reasoning to work with non-terminating programs by introducing counters-auxiliary variables that track the number of times a location is reached-and specifying inductive relations that describe the worst-case reasoning using these counters. To handle unboundedness, we infer additional invariants about counters that describe the control flow. This approach is type-directed, and the inductive relations can be viewed as refinement types that can be validated by a type checking system. However, worst-case reasoning, including the aforementioned type-directed approach, can become highly imprecise when the costs at a location vary widely among various iterations. This situation is particularly common in the design and analysis of leading data structures and algorithms iii that take advantage of amortization to achieve better overall performance, such as dynamic arrays and search structures. To address the imprecision that arises due to wide cost variations, one remedy is to use the most precise reasoning, which we dub fully-amortized reasoning. This approach accurately approximates the total resource usage before and after a transition by inferring flow-sensitive invariants. However, the practicality of this approach is limited since the required invariants are often complex polynomials that can be challenging to reliably infer. This thesis introduces and focuses around a novel paradigm called selectively-amortized reasoning, which combines the simplicity of worst-case reasoning with the precision of fullyamortized reasoning. Selectively-amortized reasoning defines a spectrum of analysis by adjusting the analysis precision and simplicity, with worst-case reasoning being the simplest and fullyamortized reasoning being the most precise. To analyze a trace's total resource usage, one decomposes the trace into subtraces, applies fully-amortized reasoning wit...

Complexity-guided container replacement synthesis

“…Checking the complexity superiority requires precise reasoning of program complexity. However, deriving a tight bound of program complexity is stunningly difficult [Wilhelm et al 2008] and far from practical for a real-world program [Gulwani et al 2009a;Xie et al 2016], especially for programs involving sophisticated manipulations of data structures [Fiedor et al 2018;Gulwani et al 2009c;Lu et al 2021;Srikanth et al 2017]. Although the container complexity superiority does not imply the complexity superiority, it provides the effective guidance to find the opportunity of synthesizing the replacements to improve program efficiency, as evidenced by our evaluation in ğ 7.…”

Differential Cost Analysis with Simultaneous Potentials and Anti-potentials

“…Given that Infer also focuses on imperative programming languages, it may be possible to couple it with our approach in order to obtain a tool which scales to large codebases, but which allows a sound and precise differential cost analysis for parts of programs that are deemed performance critical. One possible strategy is to use our approach for precise amortized reasoning locally within a global worst-case reasoning framework [37].…”