<?xml version="1.0" standalone="yes"?>
<Paper uid="P04-3032">
  <Title>Dyna: A Declarative Language for Implementing Dynamic Programs[?]</Title>
  <Section position="6" start_page="0" end_page="0" type="evalu">
    <SectionTitle>
5 Speed and Code Size
</SectionTitle>
    <Paragraph position="0"> One of our future priorities is speed. Comparing informally to the best hand-written C++ code we found online for inside-outside and Dijkstra's algorithms, Dyna (like Java) currently runs up to 5 times slower. We mainly understand the reasons (memory layout and overreliance on hashing) and are working actively to close the gap.12 Programmer time is also worth considering. Our inside-outside and Dijkstra's algorithms are each about 5 lines of Dyna code (plus a short C driver program), but were compared in the previous paragraph against efficient C++ implementations of 5500 and 900 lines.13 Our colleague Markus Dreyer, as his first Dyna program, decided to replicate the Collins parser (3400 lines of C). His implementation used under 40 lines of Dyna code, plus a 300-line C++ driver program that mostly dealt with I/O. One of us (Smith) has written substantially more complex Dyna programs (e.g., 56 types + 46 inference rules), enabling research that he would not have been willing to undertake in another language.</Paragraph>
  </Section>
class="xml-element"></Paper>