----------------------------------------------------------
Metrics-based Phase Ordering in Optimizing Compilers
Dr. Christian W. Probst

The success of optimizing compilers mainly is based on the number
of optimizations and their effects that the compilers can apply to
an input program. Much previous work has been done on the development
of new and clever analyses and optimizations. However, in almost all
available compilers the order in which analyses and optimizations 
are applied to an input program is fixed. Often the user can pick 
one of several optimization levels, each of which represents a certain
sequence of compiler phases. However, these orderings are determined
once by the compiler writer and are hard coded in the compiler. 

In this talk I will introduce the notion of a metric on an analysis/
optimization pair. Metrics can be used to rank the available phases 
based on the input program. The ranking basically determines which 
phases have a chance for a high impact, that is are supposed to be 
benifical with respect to the result of the compilation. In order to
allow an efficient compilation, there are two properties metrics 
should or must have. First, they must be easily computable, preferably 
in linear time.  Second, they should allow an online re-computation 
based on transformations applied to the intermediate representation, 
in order to avert the continuous re-computation of the metrics result 
after each change to the IR.

-------------------------------------------------------------------