Modern compilers are responsible for translating the idealistic operational semantics of the source program into a form that makes efficient use of a highly complex heterogeneous machine. Since optimization problems are associated with huge and unstructured search spaces, this combinational task is poorly achieved in general, resulting in weak scalability and disappointing sustained performance. We address this challenge by working on the program representation itself, using a semi-automatic optimization approach to demonstrate that current compilers offen suffer from unnecessary constraints and intricacies that can be avoided in a semantically richer transformation framework. Technically, the purpose of this paper is threefold: (1) to show that syntactic code representations close to the operational semantics lead to rigid phase ordering and cumbersome expression of architecture-aware loop transformations, (2) to illustrate how complex transformation sequences may be needed to achieve significant performance benefits, (3) to facilitate the automatic search for program transformation sequences, improving on classical polyhedral representations to better support operation research strategies in a simpler, structured search space. The proposed framework relies on a unified polyhedral representation of loops and statements, using normalization rules to allow flexible and expressive transformation sequencing. This representation allows to extend the scalability of polyhedral dependence analysis, and to delay the (automatic) legality checks until the end of a transformation sequence. Our work leverages on algorithmic advances in polyhedral code generation and has been implemented in a modern research compiler.

Static compiler optimizations can hardly cope with the complex run-time behavior and hardware components interplay of modern processor architectures. Multiple architectural phenomena occur and interact simultaneously, which requires the optimizer to combine multiple program transformations. Whether these transformations are selected through static analysis and models, runtime feedback, or both, the underlying infrastructure must have the ability to perform long and complex compositions of program transformations in a flexible manner. Existing compilers are ill-equipped to perform that task because of rigid phase ordering, fragile selection rules using pattern matching, and cumbersome expression of loop transformations on syntax trees. Moreover, iterative optimization emerges as a pragmatic and general means to select an optimization strategy via machine learning and operations research. Searching for the composition of dozens of complex, dependent, parameterized transformations is a challenge for iterative approaches.

Scientific programs often include multiple loops over the same data; interleaving parts of different loops may greatly improve performance. We exploit this in a compiler for Titanium, a dialect of Java. Our compiler combines reordering optimizations such as loop fusion and tiling with storage optimizations such as array contraction (eliminating or reducing the size of temporary arrays).

As processor complexity increases compilers tend to deliver suboptimal performance. Library generators such as ATLAS, FFTW and SPIRAL overcome this issue by empirically searching in the space of possible program versions for the one that performs the best. Empirical search can also be applied by programmers, but because they lack a tool to automate the process, programmers need to manually re-write the application in terms of several parameters whose best value will be determined by the empirical search in the target machine.

This paper describes a platform independent optimisation approach based on feedback-directed program restructuring. We have developed two strategies that search the optimisation space by means of profiling to find the best possible program variant. These strategies have no a priori knowledge of the target machine and can be run on any platform. In this paper our approach is evaluated on three full SPEC benchmarks, rather than the kernels evaluated in earlier studies where the optimisation space is relatively small. This approach was evaluated on six different platforms, where it is shown that we obtain on average a 20.5% reduction in execution time compared to the native compiler with full optimisation. By using training data instead of reference data for the search procedure, we can reduce compilation time and still give on average a 16.5% reduction in time when running on reference data. We show that our approach is able to give similar significant reductions in execution time over a state of the art high level restructurer based on static analysis and a platform specific profile feedback directed compiler that employs the same transformations as our iterative system.

Loop fusion and tiling are both recognized as effective transformations for improving memory performance of scientific applications. However, because of their sensitivity to the underlying cache architecture and their interaction with each other it is difficult to determine a good heuristic for applying these transformations profitably across architectures. In this paper, we present a model-guided empirical tuning strategy for profitable application of loop fusion and tiling. Our strategy consists of a detailed cost model that characterizes the interaction between the two transformations at different levels of the memory hierarchy. The novelty of our approach is in exposing key architectural parameters within the model for automatic tuning through empirical search. Preliminary experiments with a set of applications on four different platforms show that our strategy achieves significant performance improvement over fully optimized code generated by state-of-the-art commercial compilers. The time spent in searching for the best parameters is considerably less than with other search strategies.

We present the design and implementation of compiler optimizations that choreograph the use of data in scientific programs. Scientific programs often include multiple loops over the same data, where completing each loop before starting the next discards opportunities for fine-grained data reuse. Interleaving parts of different loops may greatly improve performance. Our approach combines reordering optimizations such as loop fusion and tiling with storage optimizations such as eliminating or reducing the size of temporary arrays. The programmers

Program-specific or function-specific optimization phase sequences are universally accepted to achieve better overall performance than any fixed optimization phase ordering. A number of heuristic phase order space search algorithms have been devised to find customized phase orderings achieving high performance for each function. However, to make this approach of iterative compilation more widely accepted and deployed in mainstream compilers, it is essential to modify existing algorithms, or develop new ones that find near-optimal solutions quickly. As a step in this direction, in this paper we attempt to identify and understand the important properties of some commonly employed heuristic search methods by using information collected during an exhaustive exploration of the phase order search space. We compare the performance obtained by each algorithm with all others, as well as with the optimal phase ordering performance. Finally, we show how we can use the features of the phase order space to improve existing algorithms as well as devise new, and better performing search algorithms. 1.

Software designers face many challenges when developing applications for embedded systems. One major challenge is meeting the conflicting constraints of speed, code size and power consumption. Embedded application developers often resort to hand-coded assembly language to meet these constraints since traditional optimizing compiler technology is usually of little help in addressing this challenge. The results are software systems that are not portable, less robust and more costly to develop and maintain. Another limitation is that compilers traditionally apply the optimizations to a program in a fixed order. However, it has long been known that a single ordering of optimization phases will not produce the best code for every application. In fact, the smallest unit of compilation in most compilers is typically a function and the programmer has no control over the code improvement process other than setting flags to enable or disable certain optimization phases. This paper describes a new code improvement paradigm implemented in a system called VISTA that can help achieve the cost/performance trade-offs that embedded applications demand. The VISTA system opens the code improvement process and gives the application programmer, when necessary, the ability to finely control it. VISTA also provides support for finding effective sequences of optimization phases. This support includes the ability to interactively get

Abstract. In this paper, we describe transformation recipes, which provide a high-level interface to the code transformation and code generation capability of a compiler. These recipes can be generated by compiler decision algorithms or savvy software developers. This interface is part of an auto-tuning framework that explores a set of different implementations of the same computation and automatically selects the best-performing implementation. Along with the original computation, a transformation recipe specifies a range of implementations of the computation resulting from composing a set of high-level code transformations. In our system, an underlying polyhedral framework coupled with transformation algorithms takes this set of transformations, composes them and automatically generates correct code. We first describe an abstract interface for transformation recipes, which we propose to facilitate interoperability with other transformation frameworks. We then focus on the specific transformation recipe interface used in our compiler and present performance results on its application to kernel and library tuning and tuning of key computations in high-end applications. We also show how this framework can be used to generate and auto-tune parallel OpenMP or CUDA code from a high-level specification. 1