This talk explains how the Broadway system offers a promising mechanism
for creating parallel domain-specific libraries.  The key idea is to
encapsulate domain-specific information that a compiler can exploit.
The current Broadway compiler accepts as input the source code of an
application, the source code of a library, and a set of annotations that
describe the library.  With this information, the Broadway compiler can
perform domain-specific data-flow analyses and transformations.

   This talk summarizes empirical results that show that Broadway
successfully optimizes programs written using the PLAPACK parallel linear
algebra package, matching manually-optimized code that requires deep
knowledge of the PLAPACK implementation.  This talk also shows how the
same compiler and annotation language, configured with a different set of
annotations, can be used to detect a number of security vulnerabilities
in a set of 18 Open Source C programs, consisting primarily of systems
utilities and servers.  Finally, this talk conjectures that the Broadway
architecture is ideal for conveying domain-specific information that
could be used to create parallel libraries from sequential libraries.
This new goal is significantly more ambitious and would require changes    
to both the Broadway annotation language and the Broadway compiler.