Abstract
Array-based programming languages have shown significant promise for improving performance of column-based in-memory database systems, allowing elegant representation of query execution plans that are also amenable to standard compiler optimization techniques. Use of loop fusion, however, is not straightforward, due to the complexity of built-in functions for implementing complex database operators. In this work, we apply a compiler approach to optimize SQL query execution plans that are expressed in an array-based intermediate representation. We analyze this code to determine shape properties of the data being processed, and use a subsequent optimization phase to fuse multiple database operators into single, compound operations, reducing the need for separate computation and storage of intermediate values. Experimental results on a range of TPC-H queries show that our fusion technique is effective in generating efficient code, improving query time over a baseline system.

Abstract
Recent advances in execution environments for JavaScript and WebAssembly that run on a broad range of devices, from workstations and mobile phones to IoT devices, provide new opportunities for portable and web-based numerical computing. Indeed, numerous numerical libraries and applications are emerging on the web, including Tensorflow.js, JSMapReduce, and the NLG Protein Viewer. This paper evaluates the current performance of numerical computing on the web, including both JavaScript and WebAssembly, over a wide range of devices from workstations to IoT devices. We developed a new benchmarking approach, which allowed us to perform centralized benchmarking, including benchmarking on mobile and IoT devices. Using this approach we performed four performance studies using the Ostrich benchmark suite, a collection of numerical programs representing the numerical dwarf categories identified by Colella. We studied the performance evolution of JavaScript, the relative performance of WebAssembly, the performance of server-side Node.js, and a comprehensive performance showdown for a wide range of devices.

Abstract
Relational database management systems (RDBMS) are operationally similar to a dynamic language processor. They take SQL queries as input, dynamically generate an optimized execution plan, and then execute it. In recent decades, the emergence of in-memory databases with columnar storage, which use array-like storage structures, has shifted the focus on optimizations from the traditional I/O bottleneck to CPU and memory. However, database research so far has primarily focused on CPU cache optimizations. The similarity in the computational characteristics of such database workloads and array programming language optimizations are largely unexplored. We believe that these database implementations can benefit from merging database optimizations with dynamic array-based programming language approaches. Therefore, in this paper, we propose a novel approach to optimize database query execution using a new array-based intermediate representation, HorseIR, that resides between database queries and compiled code. Furthermore, we provide a translator to generate HorseIR from database execution plans and a compiler that optimizes HorseIR and generates efficient code. We compare HorseIR with the MonetDB RDBMS, by testing standard SQL queries, and show how our approach and compiler optimizations improve the runtime of complex queries.

Abstract
Compiling MATLAB --- a dynamic, array-based language --- to JavaScript is an attractive proposal: the output code can be deployed on a platform used by billions and can leverage the countless hours that have gone into making JavaScript JIT engines fast. But before that can happen, the original MATLAB code must be properly translated, making sure to bridge the semantic gaps of the two languages.

An important area where MATLAB and JavaScript differ is in their handling of arrays: for example, in MATLAB, arrays are one-indexed and writing at an index beyond the end of an array extends it; in JavaScript, typed arrays are zero-indexed and writing out of bounds is a no-op. A MATLAB-to-JavaScript compiler must address these mismatches. Another salient and pervasive difference between the two languages is the assignment of arrays to variables: in MATLAB, this operation has value semantics, while in JavaScript is has reference semantics.

In this paper, we present MatJuice --- a source-to-source, ahead-of-time compiler back-end for MATLAB --- and how it deals efficiently with this last issue. We present an intra-procedural data-flow analysis to track where each array variable may point to and which variables are possibly aliased. We also present the associated copy insertion transformation that uses the points-to information to insert explicit copies when necessary. The resulting JavaScript program respects the MATLAB value semantics and we show that it performs fewer run-time copies than some alternative approaches.

Abstract
Thread-level Speculation (TLS) is a technique for automatic parallelization. The complexity of even prototype implementations, however, limits the ability to explore and compare the wide variety of possible design choices, and also makes understanding performance characteristics difficult. In this work we build a general analytical model of the method-level variant of TLS which we can use for determining program speedup under a wide range of TLS designs. Our approach is exhaustive, and using either simple brute force or more efficient dynamic programming implementations we are able to show how performance is strongly limited by program structure, as well as core choices in speculation design, irrespective of and complementary to the impact of data-dependencies. These results provide new, high-level insight into where and how thread-level speculation can and should be applied in order to produce practical speedup.

Abstract
Dynamic array-based languages such as MATLAB provide a wide range of built-in operations which can be efficiently applied to all elements of an array. Historically, MATLAB and Octave programmers have been advised to manually transform loops to equivalent "vectorized" computations in order to maximize performance. In this paper we present the techniques and tools to perform automatic vectorization, including handling for loops with calls to user-defined functions. We evaluate the technique on 9 benchmarks using two interpreters and two JIT-based platforms and show that automatic vectorization is extremely effective for the interpreters on most benchmarks, and moderately effective on some benchmarks in the JIT context.

Abstract
Software-based, automatic parallelization through Thread-Level Speculation (TLS) has significant practical potential, but also high overhead costs. Traditional "lazy" buffering mechanisms enable strong isolation of speculative threads, but imply large memory overheads, while more recent "eager" mechanisms improve scalability, but are more sensitive to data dependencies and have higher rollback costs. We here describe an integrated system that incorporates the best of both designs, automatically selecting the best buffering mechanism. Our approach builds on well-optimized designs for both techniques, and we describe specific optimizations that improve both lazy and eager buffer management as well. We implement our design within MUTLS, a software-TLS system based on the LLVM compiler framework. Results show that we can get 75% geometric mean performance of OpenMP versions on 9 memory intensive benchmarks. Application of these optimizations is thus a useful part of the optimization stack needed for effective and practical software TLS.

Abstract
We present a toolkit called Velociraptor that can be used by compiler writers to quickly build compilers and other tools for array-based languages. Velociraptor operates on its own unique intermediate representation (IR) designed to support a variety of array-based languages. The toolkit also provides some novel analysis and transformations such as region detection and specialization, as well as a dynamic backend with CPU and GPU code generation. We discuss the components of the toolkit and also present case-studies illustrating the use of the toolkit.

Abstract
In this paper we present extensions to an aspect oriented compiler developed for MATLAB. These extensions are intended to support important functionality for scientists, and include pattern matching on annotations, and types of variables, as well as new manners of exposing context. We provide use-cases of these features in the form of several general-use aspects which focus on solving issues that arise from use of dynamically-typed languages. We also detail performance enhancements to the ASPECTMATLAB compiler which result in an order of magnitude in performance gains.

Abstract
Developing just-in-time (JIT) compilers that that allow scientific programmers to efficiently target both CPUs and GPUs is of increasing interest. However building such compilers requires considerable effort. We present a reusable and embeddable compiler toolkit called Velociraptor that can be used to easily build compilers for numerical programs targeting multicores and GPUs.

Velociraptor provides a new high-level IR called VRIR which has been specifically designed for numeric computations, with rich support for arrays, plus support for high-level parallel and GPU constructs. A compiler developer uses Velociraptor by generating VRIR for key parts of an input program. Velociraptor provides an optimizing compiler toolkit for generating CPU and GPU code and also provides a smart runtime system to manage the GPU.

To demonstrate Velociraptor in action, we present two proof-of-concept case studies: a GPU extension for a JIT implementation of MATLAB language, and a JIT compiler for Python targeting CPUs and GPUs.

Abstract
MATLAB is a dynamic numerical scripting language widely used by scientists, engineers and students. While MATLAB's high-level syntax and dynamic types make it ideal for prototyping, programmers often prefer using high-performance static languages such as FORTRAN for their final distributable code. Rather than rewriting the code by hand, our solution is to provide a tool that automatically translates the original MATLAB program to an equivalent FORTRAN program. There are several important challenges for automatically translating MATLAB to FORTRAN, such as correctly estimating the static type characteristics of all the variables in a MATLAB program, mapping MATLAB built-in functions, and effectively mapping MATLAB constructs to equivalent FORTRAN constructs.

In this paper, we introduce Mc2FOR, a tool which automatically translates MATLAB to FORTRAN. This tool consists of two major parts. The first part is an interprocedural analysis component to estimate the static type characteristics, such as the shape of arrays and the range of scalars, which are used to generate variable declarations and to remove unnecessary array bounds checking in the translated FORTRAN program. The second part is an extensible FORTRAN code generation framework automatically transforming MATLAB constructs to FORTRAN. This work has been implemented within the McLab framework, and we demonstrate the performance of the translated FORTRAN code on a collection of MATLAB benchmarks.

Abstract
MATLAB is a popular dynamic array-based language used by engineers, scientists and students worldwide. The built-in function feval is an important MATLAB feature for certain classes of numerical programs and solvers which benefit from having functions as parameters. Programmers may pass a function name or function handle to the solver and then the solver uses feval to indirectly call the function. In this paper, we show that there are significant performance overheads for function calls via feval, in both MATLAB interpreters and JITs. The paper then proposes, implements and compares two on-the-fly mechanisms for specialization of feval calls. The first approach uses on-stack replacement technology, as supported by McVM/McOSR. The second approach specializes calls of functions with feval using a combination of runtime input argument types and values. Experimental results on seven numerical solvers show that the techniques provide good performance improvements.

Abstract
Software approaches to Thread-Level Speculation (TLS) have been recently explored, bypassing the need for specialized hardware designs. These approaches, however, tend to focus on source or VM-level implementations aimed at specific language and runtime environments. In addition, previous software approaches tend to make use of a simple thread forking model, reducing their ability to extract substantial parallelism from tree-form recursion programs such as depth-first search and divide-and-conquer. This paper proposes a Mixed forking model Universal software-TLS (MUTLS) system to overcome these limitations. MUTLS is purely based on the LLVM intermediate representation (IR), a language and architecture independent IR that supports more than 10 source languages and target architectures by many projects. MUTLS maximizes parallel coverage by applying a mixed forking model that allows all threads to speculate, forming a tree of threads. We evaluate MUTLS using several C/C++ and Fortran benchmarks on a 64-core machine. On 3 computation intensive applications we achieve speedups of 30 to 50 and 20 to 50 for the C and Fortran versions, respectively. We also observe speedups of 2 to 7 for memory intensive applications. Our experiments indicate that a mixed model is preferable for parallelization of tree-form recursion applications over the simple forking models used by previous software-TLS approaches. Our work also demonstrates that actual speedup is achievable on existing, commodity multi-core processors while maintaining the flexibility of a highly generic implementation context.

Abstract
Fork-heuristics play a key role in software Thread-Level Speculation (TLS). Current fork-heuristics either lack real parallel execution environment information to accurately evaluate fork points and/or focus on hardware-TLS implementation which cannot be directly applied to software TLS. This paper proposes adaptive fork-heuristics as well as a feedback-based selection technique to overcome the problems. Adaptive fork-heuristics insert and speculate on all potential fork/join points and purely rely on the runtime system to disable inappropriate ones. Feedback-based selection produces parallelized programs with ideal speedups using log files generated by adaptive heuristics. Experiments of three scientific computing benchmarks on a 64-core machine show that feedback-based selection and adaptive heuristics achieve more than 88% and 50% speedups of the manual-parallel version, respectively. For the Barnes-Hut benchmark, feedback-based selection is 49% faster than the manual-parallel version.

Abstract
MATLAB is a very popular dynamic "scripting" language for numerical computations used by scientists, engineers and students world-wide. MATLAB programs are often developed incrementally using a mixture of MATLAB scripts and functions, and frequently build upon existing code which may use outdated features. This results in programs that could benefit from refactoring, especially if the code will be reused and/or distributed. Despite the need for refactoring, there appear to be no MATLAB refactoring tools available. Furthermore, correct refactoring of MATLAB is quite challenging because of its non-standard rules for binding identifiers. Even simple refactorings are non-trivial.

This paper presents the important challenges of refactoring MATLAB along with automated techniques to handle a collection of refactorings for MATLAB functions and scripts including: converting scripts to functions, extracting functions, and converting dynamic function calls to static ones. The refactorings have been implemented using the McLAB compiler framework, and an evaluation is given on a large set of MATLAB benchmarks which demonstrates the effectiveness of our approach.

Abstract
On-stack replacement (OSR) is a technique that allows a virtual machine to interrupt running code during the execution of a function/method, to re-optimize the function on-the-fly using an optimizing JIT compiler, and then to resume the interrupted function at the point and state at which it was interrupted. OSR is particularly useful for programs with potentially long-running loops, as it allows dynamic optimization of those loops as soon as they become hot.

This paper presents a modular approach to implementing OSR for the LLVM compiler infrastructure. This is an important step forward because LLVM is gaining popular support, and adding the OSR capability allows compiler developers to develop new dynamic techniques. In particular, it will enable more sophisticated LLVM-based JIT compiler approaches. Indeed, other compiler/VM developers can use our approach because it is a clean modular addition to the standard LLVM distribution. Further, our approach is defined completely at the LLVM-IR level and thus does not require any modifications to the target code generation.

The OSR implementation can be used by different compilers to support a variety of dynamic optimizations. As a demonstration of our OSR approach, we have used it to support dynamic inlining in McVM. McVM is a virtual machine for MATLAB which uses a LLVM-based JIT compiler. MATLAB is a popular dynamic language for scientific and engineering applications that typically manipulate large matrices and often contain long-running loops, and is thus an ideal target for dynamic JIT compilation and OSRs. Using our McVM example, we demonstrate reasonable overheads for our benchmark set, and performance improvements when using it to perform dynamic inlining.

Abstract
MATLAB is a dynamic scientific language used by scientists, engineers and students worldwide. Although MATLAB is very suitable for rapid prototyping and development, MATLAB users often want to convert their final MATLAB programs to a static language such as FORTRAN. This paper presents an extensible object-oriented toolkit for supporting the generation of static programs from dynamic MATLAB programs. Our open source toolkit, called the MATLAB Tamer, identifies a large tame subset of MATLAB, supports the generation of a specialized Tame IR for that subset, provides a principled approach to handling the large number of builtin MATLAB functions, and supports an extensible interprocedural value analysis for estimating MATLAB types and call graphs.

Abstract

MATLAB is a popular dynamic programming language used for scientific and numerical programming. As a language, it has evolved from a small scripting language intended as an interactive interface to numerical libraries, to a very popular language supporting many language features and libraries. The overloaded syntax and dynamic nature of the language, plus the somewhat organic addition of language features over the years, makes static analysis of modern MATLAB quite challenging.

A fundamental problem in MATLAB is determining the kind of an identifier. Does an identifier refer to a variable, a named function or a prefix? Although this is a trivial problem for most programming languages, it was not clear how to do this properly in MATLAB. Furthermore, there was no simple explanation of kind analysis suitable for MATLAB programmers, nor a publicly-available implementation suitable for compiler researchers.

This paper explains the required background of MATLAB, clarifies the kind assignment program, and proposes some general guidelines for developing good kind analyses. Based on these foundations we present our design and implementation of a variety of kind analyses, including an approach that matches the intended behaviour of modern MATLAB 7 and two potentially better alternatives.

We have implemented all the variations of the kind analysis in McLAB, our extensible compiler framework, and we present an empirical evaluation of the various analyses on a large set of benchmark programs.

Abstract

Soot is a successful framework for experimenting with compiler and software engineering techniques for Java programs. Researchers from around the world have implemented a wide range of research tools which build on Soot, and Soot has been widely used by students for both courses and thesis research. In this paper, we describe relevant features of Soot, summarize its development process, and discuss useful features for future program analysis frameworks.

Abstract

Memory models are used in concurrent systems to specify visibility properties of shared data. A practical memory model, however, must permit code optimization as well as provide a useful semantics for programmers. Here we extend recent observations that the current Java memory model imposes significant restrictions on the ability to optimize code. Beyond the known and potentially correctable proof concerns illustrated by others we show that major constraints on code generation and optimization can in fact be derived from fundamental properties and guarantees provided by the memory model. To address this and accommodate a better balance between programmability and optimization we present ideas for a simple concurrency semantics for Java that avoids basic problems at a cost of backward compatibility.

Abstract

Compiler toolkits make it possible to rapidly develop compilers and translators for new programming languages. Although there exist elegant toolkits for modular and extensible parsers, compiler developers must often resort to ad-hoc solutions when extending or composing lexers. This paper presents MetaLexer, a new modular lexical specification language and associated tool.

MetaLexer allows programmers to define lexers in a modular fashion. MetaLexer modules can be used to break the lexical specification of a language into a collection smaller modular lexical specifications. Control is passed between the modules using the concept of meta-tokens and meta-lexing. MetaLexer modules are also extensible.

MetaLexer has three key features: it abstracts lexical state transitions out of semantic actions, it makes modules extensible by introducing multiple inheritance, and it provides platform agnostic support for a variety of programming languages and compiler front-end toolchains.

We have constructed a MetaLexer tool which converts MetaLexer specifications to the popular JFlex lexical specification language and we have used our tool to create lexers for three real programming languages and their extensions: AspectJ (and two AspectJ extensions), MATLAB, and MetaLexer itself. The new specifications are easier to read, are extensible, and require much less action code than the originals.

Abstract

The MATLAB programming language is heavily used in many scientific and engineering domains. Part of the appeal of the language is that one can quickly prototype numerical algorithms without requiring any static type declarations. However, this lack of type information is detrimental to both the programmer in terms of software reliability and understanding, and to the compiler in terms of generating efficient code.

This paper introduces the idea of adding typing aspects to MATLAB programs. A typing aspect can be used to: (1) capture the run-time types of variables, and (2) to check run-time types against either a declared type or against a previously captured run-time type. Typings aspects can be deployed at three different levels,they can be used: (1) solely as documentation, (2) to log type errors or (3) to catch type errors at run-time.

Abstract

MATLAB has gained widespread acceptance among scientists.Several dynamic aspects of the language contribute to its appeal, but also provide many challenges. One such problem is caused by the copy semantics of MATLAB. Existing MATLAB systems rely on reference-counting schemes to create copies only when a shared array representation is updated. This reduces array copies, but requires runtime checks.

We present a staged static analysis approach to determine when copies are not required. The first stage uses two simple, intraprocedural analyses, while the second stage combines a forward necessary copy analysis with a backward copy placement analysis. Our approach eliminates unneeded array copies without requiring reference counting or frequent runtime checks.

We have implemented our approach in the McVM JIT. Our results demonstrate that, for our benchmark set, there are significant overheads for both existing reference-counted and naive copy-insertion approaches, and that our staged approach is effective in avoiding unnecessary copies.

Abstract

Parallelization and optimization of the MATLAB programming language presents several challenges due to the dynamic nature of MATLAB. Since MATLAB does not have static type declarations, neither the shape and size of arrays, nor the loop bounds are known at compile-time. This means that many standard array dependence tests and associated transformations cannot be applied straight-forwardly. On the other hand, many MATLAB programs operate on arrays using loops and thus are ideal candidates for loop transformations and possibly loop vectorization/parallelization.

This paper presents a new framework,McFLAT,which uses profile-based training runs to determine likely loop-bounds ranges for which specialized versions of the loops may be generated. The main idea is to collect information about observed loop bounds and hot loops using training data which is then used to heuristically decide upon which loops and which ranges are worth specializing using a variety of loop transformations.

Our McFLAT framework has been implemented as part of the McLAB extensible compiler toolkit. Currently, McFLAT, is used to automatically transform ordinary MATLAB code into specialized MATLAB code with transformations applied to it. This specialized code can be executed on any MATLAB system, and we report results for four execution engines, Mathwork’s proprietary MATLAB system, the GNU Octave open-source interpreter, McLAB’s McVM interpreter and the McVM JIT. For several benchmarks, we observed significant speedups for the specialized versions, and noted that loop transformations had different impacts depending on the loop range and execution engine.

Abstract

Scientists are increasingly using dynamic programming languages like Matlab for prototyping and implementation. Effectively compiling Matlab raises many challenges due to the dynamic and complex nature of Matlab types. This paper presents a new JIT-based approach which specializes and optimizes functions on-the-fly based on the current types of function arguments.

A key component of our approach is a new type inference algorithm which uses the run-time argument types to infer further type and shape information, which in turn provides new optimization opportunities. These techniques are implemented in McVM, our open implementation of a Matlab virtual machine. As this is the first paper reporting on McVM, a brief introduction to McVM is also given.

We have experimented with our implementation and compared it to several other Matlab implementations, including the Mathworks proprietary system, McVM without specialization, the Octave open-source interpreter and the McFor static compiler. The results are quite encouraging and indicate that specialization is an effective optimization—McVM with specialization outperforms Octave by a large margin and also sometimes outperforms the Mathworks implementation.

Abstract

This paper introduces a new aspect-oriented programming language, AspectMatlab. Matlab is a dynamic scientific programming language that is commonly used by scientists because of its convenient and high-level syntax for arrays, the fact that type declarations are not required, and the availability of a rich set of application libraries.

AspectMatlab introduces key aspect-oriented features in a way that is both accessible to scientists and where the aspect-oriented features concentrate on array accesses and loops, the core computation elements in scientific programs.

Introducing aspects into a dynamic language such asMatlab also provides some new challenges. In particular, it is difficult to statically determine precisely where patterns match, resulting in many dynamic checks in the woven code. Our compiler includes flow analyses which are used to eliminate many of those dynamic checks.

This paper reports on the language design of AspectMatlab, the amc compiler implementation and related optimizations, and also provides an overview of use cases that are specific to scientific programming.

Abstract

Many aspects for runtime monitoring are history-based: they contain pieces of advice that execute conditionally, based on the observed execution history. History-based aspects are notorious for causing high runtime overhead. Compilers can apply powerful optimizations to history-based aspects using domain knowledge. Unfortunately, current aspect languages like AspectJ impede optimizations, as they provide no means to express this domain knowledge.

In this paper we present dependent advice, a novel AspectJ language extension. A dependent advice contains dependency annotations that preserve crucial domain knowledge: a dependent advice needs to execute only when its dependencies are fulfilled. Optimizations can exploit this knowledge: we present a whole-program analysis that removes advice-dispatch code from program locations at which an advice's dependencies cannot be fulfilled.

Programmers often opt to have history-based aspects generated automatically, from formal specifications from model-driven development or runtime monitoring. As we show using code-generation tools for two runtime-monitoring approaches, tracematches and JavaMOP, such tools can use knowledge contained in the specification to automatically generate dependency annotations as well.

Our extensive evaluation using the DaCapo benchmark suite shows that the use of dependent advice can significantly lower, sometimes even completely eliminate, the runtime overhead caused by history-based aspects, independently of the specification formalism.

Abstract

Runtime monitoring allows programmers to validate, for instance, the proper use of application interfaces. Given a property specification, a runtime monitor tracks appropriate runtime events to detect violations and possibly execute recovery code. Although powerful, runtime monitoring inspects only one program run at a time and so may require many program runs to find errors. Therefore, in this paper, we present ahead-of-time techniques that can (1) prove the absence of property violations on all program runs, or (2) flag locations where violations are likely to occur.

Our work focuses on tracematches, an expressive runtime monitoring notation for reasoning about groups of correlated objects. We describe a novel flow-sensitive static analysis for analyzing monitor states. Our abstraction captures both positive information (a set of objects could be in a particular monitor state) and negative information (the set is known not to be in a state). The analysis resolves heap references by combining the results of three points-to and alias analyses. We also propose a machine learning phase to filter out likely false positives.

We applied a set of 13 tracematches to the DaCapo benchmark suite and SciMark2. Our static analysis rules out all potential points of failure in 50% of the cases, and 75% of false positives on average. Our machine learning algorithm correctly classifies the remaining potential points of failure in all but three of 461 cases. The approach revealed defects and suspicious code in three benchmark programs.

Abstract
Pointer analyses enable many subsequent program analyses and transformations by statically disambiguating references to the heap. However, different client analyses may have different sets of pointer analysis needs, and each must pick some pointer analysis along the cost/precision spectrum to meet those needs. Some analysis clients employ combinations of pointer analyses to obtain better precision with reduced analysis times. Our goal is to ease the task of developing client analyses by enabling composition and substitutability for pointer analyses. We therefore propose object representatives, which statically represent runtime objects. A representative encapsulates the notion of object identity, as observed through the representative's aliasing relations with other representatives. Object representatives enable pointer analysis clients to disambiguate references to the heap in a uniform yet flexible way. Representatives can be generated from many combinations of pointer analyses, and pointer analyses can be freely exchanged and combined without changing client code. We believe that the use of object representatives brings many software engineering benefits to compiler implementations because, at compile time, object representatives are Java objects. We discuss our motivating case for object representatives, namely, the development of an abstract interpreter for tracematches, a language feature for runtime monitoring. We explain one particular algorithm for computing object representatives which combines flow-sensitive intraprocedural must-alias and must-not-alias analyses with a flow-insensitive, context-sensitive whole-program points-to analysis. In our experience, client analysis implementations can almost directly substitute object representatives for runtime objects, simplifying the design and implementation of such analyses.

Abstract

The relationships between objects in an object-oriented program are an essential property of the program's design and implementation. Two previous approaches to implement relationships with aspects were association aspects, an AspectJ-based language extension, and the relationship aspects library. While those approaches greatly ease software development, we believe that they are not general enough. For instance, the library approach only works for binary relationships, while the language extension does not allow for the association of primitive values or values from non-weavable classes. Hence, in this work we propose a generalized alternative implementation via a direct reduction to tracematches, a language feature for executing an advice after having matched a sequence of events. This new implementation scheme yields multiple benefits. Firstly, our implementation is more general than existing ones, avoiding most previous limitations. It also yields a new language construct, relational tracematches. We provide an efficient implementation based on the AspectBench Compiler, along with test cases and microbenchmarks. Our empirical studies showed that our implementation, when compared to previous approaches, uses a similar memory footprint with no leaking, but the generality of our approach does lead to some runtime overhead. We believe that our implementation can provide a solid foundation for future research.

Abstract
Virtual Machine authors face a difficult choice between low performance, cheap interpreters, or specialized and costly compilers. A method able to bridge this wide gap is the existing \emph{code-copying} technique that reuses chunks of the VM's binary code to create a simple JIT. This technique is not reliable without a compiler guaranteeing that copied chunks are still functionally equivalent despite aggressive optimizations. We present a proof-of-concept, minimal-impact modification of a highly optimizing compiler, GCC. A VM programmer marks chunks of VM source code as {\em copyable}. The chunks of native code resulting from compilation of the marked source become addressable and self-contained. Chunks can be safely copied at VM runtime, concatenated and executed together. This allows code-copying VMs to safely achieve speedup up to 3 times, 1.67 on average, over the {\em direct} interpretation. This maintainable enhancement makes the code-copying technique reliable and thus practically usable.

Abstract
Modern JIT compilers often employ multi-level recompilation strategies as a means of ensuring the most used code is also the most highly optimized, balancing optimization costs and expected future performance. Accurate selection of code to compile and level of optimization to apply is thus important to performance. In this paper we investigate the effect of an improved recompilation strategy for a Java virtual machine. Our design makes use of a lightweight, low-level profiling mechanism to detect high-level, variable length phases in program execution. Phases are then used to guide adaptive recompilation choices, improving performance. We develop both an offline implementation based on trace data and a self-contained online version. Our offline study shows an average speedup of 8.7% and up to 21%, and our online system achieves an average speedup of 4.4%, up to 18%. We subject our results to extensive analysis and show that our design achieves good overall performance with high consistency despite the existence of many complex and interacting factors in such an environment.

Abstract
The allocation of lock objects to critical sections in concurrent programs affects both performance and correctness. Recent work explores automatic lock allocation, aiming primarily to minimize conflicts and maximize parallelism by allocating locks to individual critical section interferences. We investigate component-based lock allocation, which allocates locks to entire groups of interfering critical sections. Our allocator depends on a thread-based side effect analysis, and benefits from precise points-to and may happen in parallel information. Thread-local object information has a small impact, and dynamic locks do not improve significantly on static locks. We experiment with a range of small and large Java benchmarks on 2-way, 4-way, and 8-way machines, and find that a single static lock is sufficient for mtrt, that performance degrades by 10% for hsqldb, that jbb2000 becomes mostly serialized, and that for lusearch, xalan, and jbb2005, component-based lock allocation recovers the performance of the original program.

Abstract
The pure methods in a program are those that exhibit functional or side effect free behaviour, a useful property in many contexts.  However, existing purity investigations present primarily static results.  We perform a detailed examination of dynamic method purity in Java programs using a JVM-based analysis.  We evaluate multiple purity definitions that range from strong to weak, consider purity forms specific to dynamic execution, and accomodate constraints imposed by an example consumer application, memoization.  We show that while dynamic method purity is actually fairly consistent between programs, examining pure invocation counts and the percentage of the bytecode instruction stream contained within some pure method reveals great variation.  We also show that while weakening purity definitions exposes considerable dynamic purity, consumer requirements can limit the actual utility of this information. 

Abstract
Bytecode, Java's binary form, is relatively high-level and therefore susceptible to decompilation attacks. An obfuscator transforms code such that it becomes more complex and therefore harder to reverse engineer. We develop bytecode obfuscations that are complex to reverse engineer but also do not significantly degrade performance. We present three kinds of techniques that: (1) obscure intent at the operational level; (2) complicate control flow and object-oriented design (i.e. program structure); and (3) exploit the semantic gap between what is legal in source code and what is legal in bytecode. Obfuscations are applied to a benchmark suite to examine their affect on runtime performance, control flow graph complexity, and decompilation. These results show that most of the obfuscations have only minor negative performance impacts and many increase complexity. In almost all cases, tested decompilers fail to produce legal source code or crash completely. Those obfuscations that are decompilable greatly reduce the readability of the output source code.

Abstract
Infinite recursion is a known problem of aspect-oriented programming with AspectJ: if no special precautions are taken, aspects which advise other aspects can easily and unintentionally advise themselves. We present a compiler for an extension of the AspectJ programming language that avoids self reference by associating aspects with levels, and by automatically restricting the scope of pointcuts used by an aspect to join points of lower levels. We report on a case study using our language extension and quantify the changes necessary for migrating existing applications to it. Our results suggest that we can make programming with AspectJ simpler and safer, without restricting its expressive power unduly.

Abstract
Java decompilers convert Java class files to Java source. Java class files may be created by a number of different tools including standard Java compilers, compilers for other languages such as AspectJ, or other tools such as optimizers or obfuscators. There are two kinds of Java decompilers, javac-specific decompilers that assume that the class file was created by a standard javac compiler and tool-independent decompilers that can decompile arbitrary class files, independent of the tool that created the class files. Typically javac-specific decompilers produce more readable code, but they fail to decompile many class files produced by other tools.

This paper tackles the problem of how to make a toolindependent decompiler, Dava, produce Java source code that is programmer-friendly. In past work it has been shown that Dava can decompile arbitrary class files, but often the output, although correct, is very different from what a programmer would write and is hard to understand. Furthermore, tools like obfuscators intentionally confuse the class files and this also leads to confusing decompiled source files.

Given that Dava already produces correct Java abstract syntax trees (ASTs) for arbitrary class files, we provide a new back-end for Dava. The back-end rewrites the ASTs to semantically equivalent ASTs that correspond to code that is easier for programmers to understand. Our new backend includes a new AST traversal framework, a set of simple pattern-based transformations, a structure-based data flow analysis framework and a collection of more advanced AST transformations that use flow analysis information. We include several illustrative examples including the use of advanced transformations to clean up obfuscated code.

Abstract
We present the results of an empirical study evaluating the precision of subset­based points­to analysis with several variations of context sensitivity on Java benchmarks of significant size. We compare the use of call site strings as the context abstraction, object sensitivity, and the BDD­based context­sensitive algo­ rithm proposed by Zhu and Calman, and by Whaley and Lam. Our study includes analyses that context­sensitively specialize only pointer variables, as well as ones that also specialize the heap abstraction. We measure both characteristics of the points­to sets themselves, as well as effects on the precision of client analyses. To guide development of efficient analysis implementations, we measure the number of contexts, the number of distinct contexts, and the number of distinct points­to sets that arise with each context sensitivity variation. To evaluate precision, we measure the size of the call graph in terms of methods and edges, the number of devirtualizable call sites, and the number of casts statically provable to be safe. The results of our study indicate that object­sensitive analysis implementations are likely to scale better and more predictably than the other approaches; that object­ sensitive analyses are more precise than comparable variations of the other ap­ proaches; that specializing the heap abstraction improves precision more than ex­ tending the length of context strings; and that the profusion of cycles in Java call graphs severely reduces precision of analyses that forsake context sensitivity in cyclic regions.

Abstract
Analysis of dynamic data structure usage is useful for both program understanding and for improving the accuracy of other program analyses. Static analysis techniques, however, suffer from reduced accuracy in complex situations, and do not necessarily give a clear picture of runtime heap activity. We have designed and implemented a dynamic heap analysis system that allows one to examine and analyze how Java programs build and modify data structures. Using a complete execution trace from a profiled run of the program, we build a internal representation that mirrors the evolving runtime data structures. The resulting series of representations can then be analyzed and visualized, and we show how to use our approach to help understand how programs use data structures, the precise effect of garbage collection, and to establish limits on static data structure analysis. A deep understanding of dynamic data structures is particularly important for modern, object-oriented languages that make extensive use of heapbased data structures.

Abstract
Many new Java runtime optimizations report relatively small, single-digit performance improvements. On modern virtual and actual hardware, however, the performance impact of an optimization can be influenced by a variety of factors in the underlying systems. Using a case study of a new garbage collection optimization in two different Java virtual machines, we show the relative effects of issues that must be taken into consideration when claiming an improvement. We examine the specific and overall performance changes due to our optimization and show how unintended side-effects can contribute to, and distort the final assessment. Our experience shows that VM and hardware concerns can generate variances of up to 9.5% in whole program execution time. Consideration of these confounding effects is critical to a good, objective understanding of Java performance and optimization.

Abstract
Thread level speculation (TLS) has shown great promise as a strategy for fine to medium grain automatic parallelisation, and in a hardware context techniques to ensure correct TLS behaviour are now well established.  Software and virtual machine TLS designs, however, require adherence to high level language semantics, and this can impose many additional constraints on TLS behaviour, as well as open up new opportunities to exploit language-specific information.  We present a detailed design for a Java-specific, software TLS system that operates at the bytecode level, and fully addresses the problems and requirements imposed by the Java language and VM environment.  Using SableSpMT, our research TLS framework, we provide experimental data on the corresponding costs and benefits; we find that exceptions, GC, and dynamic class loading have only a small impact, but that concurrency, native methods, and memory model concerns do play an important role, as does an appropriate, language-specific runtime TLS support system.  Full consideration of language and execution semantics is critical to correct and efficient execution of high level TLS designs, and our work here provides a baseline for future Java or Java virtual machine implementations.

Abstract
Speculative multithreading (SpMT) is a promising optimisation technique for achieving faster execution of sequential programs on multiprocessor hardware.  Analysis of and data acquisition from such systems is however difficult and complex, and is typically limited to a specific hardware design and simulation environment.  We have implemented a flexible, software-based speculative multithreading architecture within the context of a full-featured Java virtual machine.  We consider the entire Java language and provide a complete set of support features for speculative execution, including return value prediction.  Using our system we are able to generate extensive dynamic analysis information, analyse the effects of runtime feedback, and determine the impact of incorporating static, offline information.  Our approach allows for accurate analysis of Java SpMT on existing, commodity multiprocessor hardware, and provides a vehicle for further experimentation with speculative approaches and optimisations.

Abstract
Complex computer game narratives can suffer from logical consistency and playability problems if not carefully constructed, and current, state of the art design tools do little to help analysis or ensure good narrative properties.  A formally-grounded system that allows for relatively easy design and analysis is therefore desireable.  We present a language and an environment for expressing game narratives based on a structured form of Petri Net, the Narrative Flow Graph.  Our "(P)NFG" system provides a simple, high level view of narrative programming that maps onto a low level representation suitable for expressing and analysing game properties.  The (P)NFG framework is demonstrated experimentally by modelling narratives based on non-trivial interactive fiction games, and integrates with the NuSMV model checker.  Our system provides a necessary component for systematic analysis of computer game narratives, and lays the foundation for all-around improvements to game quality.

Abstract
Method inlining is one of most important optimizations to achieve a high performance JIT compiler in Java virtual machines. A type analysis allows the compiler directly inline monomorphic calls. At runtime, the compiler and type analysis have to handle dynamic class loading properly because the analysis result is only correct at compile time. Loading of new classes could invalidate previous analysis results and optimizations. Class hierarchy analysis (CHA) has been used successfully in JIT compilers for speculative inlining with various invalidation techniques as backup.

In this paper, we present the results of a limit study of method inlining using dynamic type analysis on a set of standard Java benchmarks. We developed a general type analysis framework for measure the effectiveness of several well-known type analysis, including CHA, RTA, XTA and VTA. Surprisingly, the simple dynamic CHA is nearly as good as an ideal type analysis for inlining virtual method calls. It leaves no room for other type analysis to improve. On the other hand, only reachability-based interprocedural type analysis (VTA) is able to capture the majority of monomorphic interface calls. We measured the runtime overhead of interprocedural type analysis in the JIT environment. To overcome the memory overhead of dynamic whole-program analysis, we outlined the design of a demand-driven inter-procedural type analysis for inlining hot interface calls.

Abstract
Inter-procedural analyses such as side-effect analysis can provide information useful for performing aggressive optimizations. We present a study of whether side-effect information improves performance in just-in-time (JIT) compilers, and if so, what level of analysis precision is needed.

We used Spark, the inter-procedural analysis component of the Soot Java analysis and optimization framework, to compute side-effect information and encode it in class files. We modified Jikes RVM, a research JIT, to make use of side-effect analysis in local common sub-expression elimination, heap SSA, redundant load elimination and loop-invariant code motion. On the SpecJVM98 benchmarks, we measured the static number of memory operations removed, the dynamic counts of memory reads eliminated, and the execution time.

Our results show that the use of side-effect analysis increases the number of static opportunities for load elimination by up to 98%, and reduces dynamic field read instructions by up to 27%. Side-effect information enabled speedups in the range of 1.08x to 1.20x for some benchmarks. Finally, among the different levels of precision of side-effect information, a simple side-effect analysis is usually sufficient to obtain most of these speedups.

Abstract
Research in the design of aspect-oriented programming languages requires a workbench that facilitates easy experimentation with new language features and implementation techniques. In particular, new features for AspectJ have been proposed that require extensions in many dimensions: syntax, type checking and code generation, as well as data flow and control flow analyses.

The AspectBench Compiler (abc) is an implementation of such a workbench. The base version of abc implements the full AspectJ language. Its frontend is built, using the Polyglot framework, as a modular extension of the Java language. The use of Polyglot gives flexibility of syntax and type checking. The backend is built using the Soot framework, to give modular code generation and analyses.

In this paper, we outline the design of abc, focusing mostly on how the design supports extensibility. We then provide a general overview of how to use abc to implement an extension. Finally, we illustrate the extension mechanisms of abc through a number of small, but non-trivial, examples. abc is freely available under the GNU LGPL.

Abstract
We describe the effect of a particular form of "noise" in benchmarking.  We investigate the source of anomalous measurement data in a series of optimization strategies that attempt to improve runtime performance in the garbage collector of a Java virtual machine.  The results of our experiments can be explained in terms of the difference in code layout, and hence instruction and data cache behaviour.  We show that unintended changes in code layout due to code modifications as trivial as symbol renaming can contribute up to 2.7% of measured machine cycle cost, 20% in data cache misses, and 37% in instruction cache misses.

Abstract
We present the design and implementation of return value prediction in SableVM, a Java Virtual Machine.  We give detailed results for the full SPEC JVM98 benchmark suite, and compare our results with previous, more limited data.  At the performance limit of existing last value, stride, 2-delta stride, parameter stride, and context (FCM) sub-predictors in a hybrid, we achieve an average accuracy of 72%.  We describe and characterize a new table-based memoization predictor that complements these predictors nicely, yielding an increased average hybrid accuracy of 81%.  VM level information about data widths provides a 35% reduction in space, and dynamic allocation and expansion of per-callsite hashtables allows for highly accurate prediction with an average per-benchmark requirement of 119 MB for the context predictor and 43 MB for the memoization predictor.  As far as we know, the is the first implementation of non-trace-based return value prediction within a JVM.

Abstract
In this paper we present an implementation of May Happen in Parallel analysis for Java that attempts to address some of the practical implementation concerns of the original work.  We describe a design that incorporates techniques for aiding a feasible implementation and expanding the range of acceptable inputs.  We provide experimental results showing the utility and impact of our approach and optimizations using a variety of concurrent benchmarks.

Abstract
In this paper we present Jedd, a language extension to Java that supports a convenient way of programming with Binary Decision Diagrams (BDDs). The Jedd language abstracts BDDs as database-style relations and operations on relations, and provides static type rules to ensure that relational operations are used correctly.

The paper provides a description of the Jedd language and reports on the design and implementation of the Jedd translator and associated runtime system. Of particular interest is the approach to assigning attributes from the high-level relations to physical domains in the underlying BDDs, which is done by expressing the constraints as a SAT problem and using a modern SAT solver to compute the solution. Further, a runtime system is defined that handles memory management issues and supports a browsable profiling tool for tuning the key BDD operations.

The motivation for designing Jedd was to support the development of whole program analyses based on BDDs, and we have used Jedd to express five key interrelated whole program analyses in our Soot compiler framework. We provide some examples of this application and discuss our experiences using Jedd.

Abstract
This paper presents a new, inexpensive, mechanism for constructing a complete call graph for Java programs at runtime, and provides an example of using the mechanism for implementing a dynamic reachability-based interprocedural analysis (IPA), namely dynamic XTA.

Reachability-based IPAs, such as points-to analysis and escape analysis, require a context-insensitive call graph of the analyzed program. Computing a call graph at runtime presents several challenges. First, the overhead must be low. Second, when implementing the mechanism for languages such as Java, both polymorphism and lazy class loading must be dealt with correctly and efficiently. We propose a new, low-cost, mechanism for constructing runtime call graphs in a JIT environment. The mechanism uses a profiling code stub to capture the first execution of a call edge, and adds at most one more instruction to repeated call edge invocations. Polymorphism and lazy class loading are handled transparently. The call graph is constructed incrementally, and it supports optimistic analysis and speculative optimizations with invalidations.

We also developed a dynamic, reachability-based type analysis, dynamic XTA, as an application of runtime call graphs. It also serves as an example of handling lazy class loading in dynamic IPAs.

The dynamic call graph construction algorithm and dynamic version of XTA have been implemented in Jikes RVM. We present empirical measurements of the overhead of call graph profiling and compare the characteristics of call graphs built using our profiling code stubs with conservative ones constructed by using dynamic class hierarchy analysis (CHA).

Abstract
This paper presents the integration of Soot, a byte-code analysis and transformation framework, with an integrated development environment (IDE), Eclipse. Such an integrated toolkit is useful for both the compiler developer, to aid in understanding and debugging new analyses, and also for the end-user of the IDE, to aid in program understanding by exposing semantic information gathered by the advanced compiler analyses. The paper discusses these advantages and provides concrete examples of its usefulness.

There are several major challenges to overcome in developing the integrated toolkit, and the paper discusses three major challenges and the solutions to those challenges. An overview of Soot and the integrated toolkit is given, followed by a more detailed discussion of the fundamental components. The paper concludes with several illustrative examples of using the integrated toolkit along with a discussion of future plans and research.

Abstract
Our integration of the Soot bytecode manipulation framework into the Eclipse IDE forms a powerful tool for graphically visualizing both the progress and output of program analyses. We demonstrate several examples of the visualizations that we have developed, and explain how they are useful for both compiler research and teaching.

Abstract
In order to perform meaningful experiments in optimizing compilation and run-time system design, researchers usually rely on a suite of benchmark programs of interest to the optimization technique under consideration. Programs are described as numeric, memory-intensive, concurrent, or object-oriented, based on a qualitative appraisal, in some cases with little justification. We believe it is beneficial to quantify the behaviour of programs with a concise and precisely defined set of metrics, in order to make these intuitive notions of program behaviour more concrete and subject to experimental validation. We therefore define and measure a set of unambiguous, dynamic, robust and architecture-independent metrics that can be used to categorize programs according to their dynamic behaviour in five areas: size, data structure, memory use, concurrency, and polymorphism. A framework computing some of these metrics for Java programs is presented along with specific results demonstrating how to use metric data to understand a program's behaviour, and both guide and evaluate compiler optimizations.

Abstract
Existing visualization tools typically do not allow easy extension by new visualization techniques, and are often coupled with inflexible data input mechanisms. This paper presents EVolve, a flexible and extensible framework for visualizing program characteristics and behaviour. The framework is flexible in the sense that it can visualize many kinds of data, and it is extensible in the sense that it is quite straightforward to add new kinds of visualizations.

The overall architecture of the framwork consists of the core EVolve platform that communicates with data sources via a well defined data protocal and which communicates with visualization methods via a visualization protocol.

Given a data source, an end-user can use EVolve as a stand-alone tool by interactively creating, configuring and modifying visualizations. A variety of visualizations are provided in the current EVolve library, with features that facilitate the comparison of multiple views on the same execution data. We demonstrate EVolve in the context of visualizing execution behaviour of Java programs.

Abstract
This paper reports on a new approach to solving a subset-based points-to analysis for Java using Binary Decision Diagrams (BDDs). In the model checking community, BDDs have been shown very effective for representing large sets and solving very large verification problems. Our work shows that BDDs can also be very effective for developing a points-to analysis that is simple to implement and that scales well, in both space and time, to large programs.

The paper first introduces BDDs and operations on BDDs using some simple points-to examples. Then, a complete subset-based points-to algorithm is presented, expressed completely using BDDs and BDD operations. This algorithm is then refined by finding appropriate variable orderings and by making the algorithm propagate sets incrementally, in order to arrive at a very efficient algorithm. Experimental results are given to justify the choice of variable ordering, to demonstrate the improvement due to incrementalization, and to compare the performance of the BDD-based solver to an efficient hand-coded graph-based solver. Finally, based on the results of the BDD-based solver, a variety of BDD-based queries are presented, including the points-to query.

Abstract
Dynamic program optimization is increasingly important for achieving good runtime performance. A key issue is how to select which code to optimize. One approach is to dynamically detect traces, long sequences of instructions spanning multiple methods, which are likely to execute to completion. Traces are easy to optimize and have been shown to be a good unit for optimization.

This paper reports on a new approach for dynamically detecting, creating and storing traces in a Java virtual machine. We first describe four important criteria for a successful trace strategy: good instruction stream coverage, low dispatch rate, cache stability, and optimizability of traces. We then present our approach based on branch correlation graphs. A branch correlation graph stores information about the correlation between pairs of branches, as weel as additional state information.

We present the complete design for an efficient implementation of the system, including a detailed discussion of the trace cache and profiling mechanisms. We have implemented an experimental framework to measure the traces generated by our approach in a direct-threaded Java VM(SableVM) and we presnet experimental results to show that the trace we generate meet the design criteria.

Abstract
Inline-threaded interpretation is a recent technique that improves performance by eliminating dispatch overhead within basic blocks for interpreters written in C. The dynamic class loading, lazy class initialization, and multi-threading features of Java reduce the effectiveness of a straight-forward implementation of this technique within Java interpreters. In this paper, we introduce preparation sequences, a new technique that solves the particular challenge of effectively inline-threading Java. We have implemented our technique in the SableVM Java virtual machine, and our experimental results show that using our technique, inline-threaded interpretation of Java, on a set of benchmarks, achieves a speedup ranging from 1.20 to 2.41 over switch-based interpretation, and a speedup ranging from 1.15 to 2.14 over direct-threaded interpretation.