Just-in-time shape inference for array-based languages

Abstract

In dynamic array-based languages, the most computationally intensive parts of the program often involve either explicit loops or vector operations. These loops and vector operations can be better optimized if the compiler has accurate information about array shapes and loop-bounds. However, accurate shape information about loops and arrays may not be known until runtime. We present a method of performing shape inference in a just-in-time compiler by using information obtained at runtime. We have implemented our method in a compiler toolkit for array-based languages that is integrated into two different compilers: a prototype compiler for Python, and McVM, a virtual machine for MATLAB. We present results obtained from these two compiler integrations on a variety of benchmarks to evaluate the accuracy of shape inference of our method and the runtime overhead of our implementation.

Velociraptor: a compiler toolkit for array-based languages targeting CPUs and GPUs

“…libVRIR implements a region detection analysis, which we described in a previous publication [7]. The idea is to identify regions which the compiler may be able to optimize or specialize using information about values or shapes of certain variables at runtime.…”

“…Velociraptor is an optimizing dynamic compilation framework that generates LLVM [10] for CPUs and OpenCL for GPUs respectively. We have described Velociraptor in previous publications [7,8]. In this paper, our goal is to describe the software architecture and discuss how it may be used by the compiler community to build compiler infrastructure for array-based languages.…”

Velociraptor

“…When the outlined functions are called, a shape inference pass determines the shape information and is used for code generation. Details of our JIT specialization system can be found elsewhere [9].…”