OPTIMIZING THE JAVA VIRTUAL MACHINE INSTRUCTION SET BY ...

139
cache performance. As a result, several different multicode identification techniques
are considered.
Each technique is responsible for assigning a score to each candidate sequence.
The score is an integer value that specifies the candidate sequences’ relative value
as a multicode. It may be derived through the use of a formula or may be assigned
through less definite means such as intuition. Higher scores are assigned to sequences
that are expected to be good selections for replacement with a multicode while lower
scores are assigned to sequences that are expected to be poor selections. Using scores
in this manner provides a mechanism for estimating the value of ε x1 x 2 x 3 ...x n→ during
the multicode identification process. Several scoring systems are discussed in the
following sections.
7.3.1 Intuition
Intuition can be used to identify some short sequences that are expected to execute
with great frequency. An example of such a sequence is aload 0 getfield. It is
expected to execute with great frequency because it represents the process of reading
the value of a field within a receiver object. Consequently, a high score will normally
be assigned to this sequence while another sequence, intuitively expected to be less
useful, will be assigned a low score. Unfortunately this technique may not provide
good results because programs do not always behave the way one might intuitively
expect and because it is difficult to predict sequences of more than a few bytecodes.
This technique has been used by some of the related optimization techniques discussed
in Section 3.1, such as bytecode idioms.
7.3.2 Static Profiling
Techniques were developed that attempted to overcome the problems associated with
intuitive approaches. A static examination of the classes that make up the benchmarks
and the Java library was considered initially. The goal of this examination was
to identify which sequences of bytecodes occurred within those class files with greatest
frequency. Then the frequency information could be used in conjunction with other
information, such as the length of the sequence, in order to determine a score for each
candidate sequence.
A static analysis would have begun by extracting the code attribute for each
method in each class. Each code attribute would have been analyzed in order to