Fine-grain data-level parallelism can be exploited using SIMD or short-
vector architectures, which have become ubiquitous in DSP's, desktops and
servers. Programmers have benefited from such instructions for some time
using both programming language extensions as well as automatic
vectorization or simdization by compilers. Yet some limitations still hinder
wider usage and efficiency of SIMD extensions, mostly involved with
gathering data to feed the SIMD processing engine, including its
programmability, compilability and portability. In this short talk we'll discuss
several approaches that aim to overcome these limitations.