Over the past several decades, the compiler research community has developed 
a number of sophisticated and powerful algorithms for a variety of code 
improvements. Although there are still promising directions for particular 
aspects of code optimization, research on optimizations is nearing the point 
of diminishing returns and other approaches are needed to achieve further 
performance improvements. Our research aims to address this challenge by 
investigating and developing an innovative framework and system for 
continuously and adaptively applying code improvements. Our system, the 
Continuous Compiler (CoCo), determines "optimization plans" at compile-time 
that describe the best way in which to apply both static and dynamic code 
transformations. The plans consider program and machine context, interaction 
among optimizations, and performance profit. Through such planning, CoCo can 
tailor and adapt its decisions to more synergistically apply a whole suite of 
code transformations.

This talk will describe the Continuous Compilation approach and present 
initial results. These results include novel analytic models that can 
accurately predict the performance benefit (profit) of applying an 
optimization without actually doing it or running the resulting program code.
 Using the analytic models, we have developed several planners that can guide 
a static optimizer. Initial results are very encouraging and show that 
optimizations can be effectively directed by planning. The talk will conclude 
with a brief discussion of CoCo's run-time system, including its dynamic code 
translator, instrumentation optimizer, and source-level debugger for 
dynamically translated code.