arXiv subjects

Enea Zaffanella

Publications and source records attributed to Enea Zaffanella.

17 recordsLinked to original sources

arXiv2025

Helping LLMs Improve Code Generation Using Feedback from Testing and Static Analysis

Large Language Models (LLMs) are one of the most promising developments in the field of artificial intelligence, and the software engineering community has readily noticed their potential role in the software development life-cycle. Developers routinely ask LLMs to generate code snippets, increasing productivity but also potentially introducing ownership, privacy, correctness, and security issues. Previous work highlighted how code generated by mainstream commercial LLMs is often not safe, containing vulnerabilities, bugs, and code smells. In this paper, we present a framework that leverages testing and static analysis to assess the quality, and guide the self-improvement, of code generated by general-purpose, open-source LLMs. First, we ask LLMs to generate C code to solve a number of programming tasks. Then we employ ground-truth tests to assess the (in)correctness of the generated code, and a static analysis tool to detect potential safety vulnerabilities. Next, we assess the models ability to evaluate the generated code, by asking them to detect errors and vulnerabilities. Finally, we test the models ability to fix the generated code, providing the reports produced during the static analysis and incorrectness evaluation phases as feedback. Our results show that models often produce incorrect code, and that the generated code can include safety issues. Moreover, they perform very poorly at detecting either issue. On the positive side, we observe a substantial ability to fix flawed code when provided with information about failed tests or potential vulnerabilities, indicating a promising avenue for improving the safety of LLM-based code generation tools.

cs.SE↗

arXiv2022

Decoupling the ascending and descending phases in Abstract Interpretation

Abstract Interpretation approximates the semantics of a program by mimicking its concrete fixpoint computation on an abstract domain $\mathbb{A}$. The abstract (post-) fixpoint computation is classically divided into two phases: the ascending phase, using widenings as extrapolation operators to enforce termination, is followed by a descending phase, using narrowings as interpolation operators, so as to mitigate the effect of the precision losses introduced by widenings. In this paper we propose a simple variation of this classical approach where, to more effectively recover precision, we decouple the two phases: in particular, before starting the descending phase, we replace the domain $\mathbb{A}$ with a more precise abstract domain $\mathbb{D}$. The correctness of the approach is justified by casting it as an instance of the A$^2$I framework. After demonstrating the new technique on a simple example, we summarize the results of a preliminary experimental evaluation, showing that it is able to obtain significant precision improvements for several choices of the domains $\mathbb{A}$ and $\mathbb{D}$.

cs.PL↗

arXiv2017

A Conversion Procedure for NNC Polyhedra

We present an alternative Double Description representation for the domain of NNC (not necessarily topologically closed) polyhedra, together with the corresponding Chernikova-like conversion procedure. The representation differs from the ones adopted in the currently available implementations of the Double Description method in that it uses no slack variable at all: this new approach provides a solution to a few technical issues caused by the encoding of an NNC polyhedron as a closed polyhedron in a higher dimension space. A preliminary experimental evaluation shows that the new conversion algorithm is able to achieve significant efficiency improvements with respect to state-of-the-art implementations.

cs.CG↗

arXiv2012

The Automatic Synthesis of Linear Ranking Functions: The Complete Unabridged Version

The classical technique for proving termination of a generic sequential computer program involves the synthesis of a ranking function for each loop of the program. Linear ranking functions are particularly interesting because many terminating loops admit one and algorithms exist to automatically synthesize it. In this paper we present two such algorithms: one based on work dated 1991 by Sohn and Van Gelder; the other, due to Podelski and Rybalchenko, dated 2004. Remarkably, while the two algorithms will synthesize a linear ranking function under exactly the same set of conditions, the former is mostly unknown to the community of termination analysis and its general applicability has never been put forward before the present paper. In this paper we thoroughly justify both algorithms, we prove their correctness, we compare their worst-case complexity and experimentally evaluate their efficiency, and we present an open-source implementation of them that will make it very easy to include termination-analysis capabilities in automatic program verifiers.

cs.PL↗

arXiv2009

Exact Join Detection for Convex Polyhedra and Other Numerical Abstractions

Deciding whether the union of two convex polyhedra is itself a convex polyhedron is a basic problem in polyhedral computations; having important applications in the field of constrained control and in the synthesis, analysis, verification and optimization of hardware and software systems. In such application fields though, general convex polyhedra are just one among many, so-called, numerical abstractions, which range from restricted families of (not necessarily closed) convex polyhedra to non-convex geometrical objects. We thus tackle the problem from an abstract point of view: for a wide range of numerical abstractions that can be modeled as bounded join-semilattices --that is, partial orders where any finite set of elements has a least upper bound--, we show necessary and sufficient conditions for the equivalence between the lattice-theoretic join and the set-theoretic union. For the case of closed convex polyhedra --which, as far as we know, is the only one already studied in the literature-- we improve upon the state-of-the-art by providing a new algorithm with a better worst-case complexity. The results and algorithms presented for the other numerical abstractions are new to this paper. All the algorithms have been implemented, experimentally validated, and made available in the Parma Polyhedra Library.

cs.CG↗

arXiv2008

Applications of Polyhedral Computations to the Analysis and Verification of Hardware and Software Systems

Convex polyhedra are the basis for several abstractions used in static analysis and computer-aided verification of complex and sometimes mission critical systems. For such applications, the identification of an appropriate complexity-precision trade-off is a particularly acute problem, so that the availability of a wide spectrum of alternative solutions is mandatory. We survey the range of applications of polyhedral computations in this area; give an overview of the different classes of polyhedra that may be adopted; outline the main polyhedral operations required by automatic analyzers and verifiers; and look at some possible combinations of polyhedra with other numerical abstractions that have the potential to improve the precision of the analysis. Areas where further theoretical investigations can result in important contributions are highlighted.

cs.CG↗

arXiv2007

A Prolog-based Environment for Reasoning about Programming Languages (Extended abstract)

ECLAIR is a Prolog-based prototype system aiming to provide a functionally complete environment for the study, development and evaluation of programming language analysis and implementation tools. In this paper, we sketch the overall structure of the system, outlining the main methodologies and technologies underlying its components. We also discuss the appropriateness of Prolog as the implementation language for the system: besides highlighting its strengths, we also point out a few potential weaknesses, hinting at possible solutions.

cs.PL↗

arXiv2007

On the Design of Generic Static Analyzers for Modern Imperative Languages

The design and implementation of precise static analyzers for significant fragments of modern imperative languages like C, C++, Java and Python is a challenging problem. In this paper, we consider a core imperative language that has several features found in mainstream languages such as those including recursive functions, run-time system and user-defined exceptions, and a realistic data and memory model. For this language we provide a concrete semantics --characterizing both finite and infinite computations-- and a generic abstract semantics that we prove sound with respect to the concrete one. We say the abstract semantics is generic since it is designed to be completely parametric on the analysis domains: in particular, it provides support for \emph{relational} domains (i.e., abstract domains that can capture the relationships between different data objects). We also sketch how the proposed methodology can be extended to accommodate a larger language that includes pointers, compound data objects and non-structured control flow mechanisms. The approach, which is based on structured, big-step $\mathrm{G}^\infty\mathrm{SOS}$ operational semantics and on abstract interpretation, is modular in that the overall static analyzer is naturally partitioned into components with clearly identified responsibilities and interfaces, something that greatly simplifies both the proof of correctness and the implementation.

cs.PL↗

arXiv2007

An Improved Tight Closure Algorithm for Integer Octagonal Constraints

Integer octagonal constraints (a.k.a. ``Unit Two Variables Per Inequality'' or ``UTVPI integer constraints'') constitute an interesting class of constraints for the representation and solution of integer problems in the fields of constraint programming and formal analysis and verification of software and hardware systems, since they couple algorithms having polynomial complexity with a relatively good expressive power. The main algorithms required for the manipulation of such constraints are the satisfiability check and the computation of the inferential closure of a set of constraints. The latter is called `tight' closure to mark the difference with the (incomplete) closure algorithm that does not exploit the integrality of the variables. In this paper we present and fully justify an O(n^3) algorithm to compute the tight closure of a set of UTVPI integer constraints.

cs.DS↗

arXiv2006

The Parma Polyhedra Library: Toward a Complete Set of Numerical Abstractions for the Analysis and Verification of Hardware and Software Systems

Since its inception as a student project in 2001, initially just for the handling (as the name implies) of convex polyhedra, the Parma Polyhedra Library has been continuously improved and extended by joining scrupulous research on the theoretical foundations of (possibly non-convex) numerical abstractions to a total adherence to the best available practices in software development. Even though it is still not fully mature and functionally complete, the Parma Polyhedra Library already offers a combination of functionality, reliability, usability and performance that is not matched by similar, freely available libraries. In this paper, we present the main features of the current version of the library, emphasizing those that distinguish it from other similar libraries and those that are important for applications in the field of analysis and verification of hardware and software systems.

cs.MS↗

arXiv2005

PURRS: Towards Computer Algebra Support for Fully Automatic Worst-Case Complexity Analysis

Fully automatic worst-case complexity analysis has a number of applications in computer-assisted program manipulation. A classical and powerful approach to complexity analysis consists in formally deriving, from the program syntax, a set of constraints expressing bounds on the resources required by the program, which are then solved, possibly applying safe approximations. In several interesting cases, these constraints take the form of recurrence relations. While techniques for solving recurrences are known and implemented in several computer algebra systems, these do not completely fulfill the needs of fully automatic complexity analysis: they only deal with a somewhat restricted class of recurrence relations, or sometimes require user intervention, or they are restricted to the computation of exact solutions that are often so complex to be unmanageable, and thus useless in practice. In this paper we briefly describe PURRS, a system and software library aimed at providing all the computer algebra services needed by applications performing or exploiting the results of worst-case complexity analyses. The capabilities of the system are illustrated by means of examples derived from the analysis of programs written in a domain-specific functional programming language for real-time embedded systems.

cs.MS↗

arXiv2004

Widening Operators for Weakly-Relational Numeric Abstractions (Extended Abstract)

We discuss the divergence problems recently identified in some extrapolation operators for weakly-relational numeric domains. We identify the cause of the divergences and point out that resorting to more concrete, syntactic domains can be avoided by researching suitable algorithms for the elimination of redundant constraints in the chosen representation.

cs.PL↗

arXiv2004

Finite-Tree Analysis for Constraint Logic-Based Languages: The Complete Unabridged Version

Logic languages based on the theory of rational, possibly infinite, trees have much appeal in that rational trees allow for faster unification (due to the safe omission of the occurs-check) and increased expressivity (cyclic terms can provide very efficient representations of grammars and other useful objects). Unfortunately, the use of infinite rational trees has problems. For instance, many of the built-in and library predicates are ill-defined for such trees and need to be supplemented by run-time checks whose cost may be significant. Moreover, some widely-used program analysis and manipulation techniques are correct only for those parts of programs working over finite trees. It is thus important to obtain, automatically, a knowledge of the program variables (the finite variables) that, at the program points of interest, will always be bound to finite terms. For these reasons, we propose here a new data-flow analysis, based on abstract interpretation, that captures such information.

cs.PL↗

arXiv2004

A correct, precise and efficient integration of set-sharing, freeness and linearity for the analysis of finite and rational tree languages

It is well-known that freeness and linearity information positively interact with aliasing information, allowing both the precision and the efficiency of the sharing analysis of logic programs to be improved. In this paper we present a novel combination of set-sharing with freeness and linearity information, which is characterized by an improved abstract unification operator. We provide a new abstraction function and prove the correctness of the analysis for both the finite tree and the rational tree cases. Moreover, we show that the same notion of redundant information as identified in (Bagnara et al. 2002; Zaffanella et al. 2002) also applies to this abstract domain combination: this allows for the implementation of an abstract unification operator running in polynomial time and achieving the same precision on all the considered observable properties.

cs.PL↗

arXiv2004

Enhanced sharing analysis techniques: a comprehensive evaluation

Sharing, an abstract domain developed by D. Jacobs and A. Langen for the analysis of logic programs, derives useful aliasing information. It is well-known that a commonly used core of techniques, such as the integration of Sharing with freeness and linearity information, can significantly improve the precision of the analysis. However, a number of other proposals for refined domain combinations have been circulating for years. One feature that is common to these proposals is that they do not seem to have undergone a thorough experimental evaluation even with respect to the expected precision gains. In this paper we experimentally evaluate: helping Sharing with the definitely ground variables found using Pos, the domain of positive Boolean formulas; the incorporation of explicit structural information; a full implementation of the reduced product of Sharing and Pos; the issue of reordering the bindings in the computation of the abstract mgu; an original proposal for the addition of a new mode recording the set of variables that are deemed to be ground or free; a refined way of using linearity to improve the analysis; the recovery of hidden information in the combination of Sharing with freeness information. Finally, we discuss the issue of whether tracking compoundness allows the computation of more sharing information.

cs.PL↗

arXiv2001

Soundness, Idempotence and Commutativity of Set-Sharing

It is important that practical data-flow analyzers are backed by reliably proven theoretical results. Abstract interpretation provides a sound mathematical framework and necessary generic properties for an abstract domain to be well-defined and sound with respect to the concrete semantics. In logic programming, the abstract domain Sharing is a standard choice for sharing analysis for both practical work and further theoretical study. In spite of this, we found that there were no satisfactory proofs for the key properties of commutativity and idempotence that are essential for Sharing to be well-defined and that published statements of the soundness of Sharing assume the occurs-check. This paper provides a generalization of the abstraction function for Sharing that can be applied to any language, with or without the occurs-check. Results for soundness, idempotence and commutativity for abstract unification using this abstraction function are proven.

cs.PL↗

arXiv2001

Decomposing Non-Redundant Sharing by Complementation

Complementation, the inverse of the reduced product operation, is a technique for systematically finding minimal decompositions of abstract domains. File' and Ranzato advanced the state of the art by introducing a simple method for computing a complement. As an application, they considered the extraction by complementation of the pair-sharing domain PS from the Jacobs and Langen's set-sharing domain SH. However, since the result of this operation was still SH, they concluded that PS was too abstract for this. Here, we show that the source of this result lies not with PS but with SH and, more precisely, with the redundant information contained in SH with respect to ground-dependencies and pair-sharing. In fact, a proper decomposition is obtained if the non-redundant version of SH, PSD, is substituted for SH. To establish the results for PSD, we define a general schema for subdomains of SH that includes PSD and Def as special cases. This sheds new light on the structure of PSD and exposes a natural though unexpected connection between Def and PSD. Moreover, we substantiate the claim that complementation alone is not sufficient to obtain truly minimal decompositions of domains. The right solution to this problem is to first remove redundancies by computing the quotient of the domain with respect to the observable behavior, and only then decompose it by complementation.

cs.PL↗