OPTIMIZING THE JAVA VIRTUAL MACHINE INSTRUCTION SET BY ...

147
number of bytecodes that will be impacted by performing the substitution. However,
this algorithm can use any scoring system that assigns a non-negative integer number
to each candidate sequence, with better sequences receiving a higher number than
poorer sequences. The overall effectiveness of the multicode identification algorithm
is based on the accuracy of the scoring function. Section 7.3.5 describes one of many
possible alternative scoring functions that could be employed.
As was observed in the previous section, once a multicode has been identified the
set of candidate sequences must be recomputed before the next multicode can be identified.
However, this recomputation can be done incrementally and quite efficiently.
In particular, a set of update operations are performed that remove occurrences of
target candidate sequences from the candidate sequence data-structure. In addition,
all occurrences of sequences that either partially or completely overlap with this sequence
must also be removed from future consideration. Most importantly, the counts
of remaining candidate sequences must be adjusted to reflect the removal of the target
candidate sequence. The algorithm used to accomplish these tasks is described in
Figure 7.5.
The algorithm identifies all of the sequences which contain at least one occurrence
of candidate sequence BestSeq, presumably identified previously using the algorithm
presented in Figure 7.4. Each sequence containing BestSeq is processed, starting
with the longest sequence and proceeding to shorter and shorter sequences. When
a sequence contains BestSeq, the count for that sequence, e, and the counts for all
subsequences of e, denoted by e ′ , are reduced by the count for e. However, while this
successfully removes all of the bytecodes used to represent occurrences of BestSeq,
it also reduces the counts for some subsequences of e that are not impacted by the
selection of BestSeq.
In order to correct the counts for those sequences which should not have changed,
prefix and suffix sequences are determined which represent those bytecodes that occur
before and after the first occurrence of BestSeq in e respectively. The count that was
associated with e is added back to each of the subsequences of the prefix and suffix,
resulting in a net change of zero in their counts. As a result, this algorithm successfully
updates the data structures generate during the determination of BestSeq,
removing all occurrences of that sequence and its subsequences without impacting the
counts associated with any other bytecode sequences. When e contains two or more
occurrences of BestSeq, the additional occurrences will reside in the suffix sequence.
The count for this sequence will be increment just like any other suffix sequence.
This is not a problem because the occurrence of BestSeq contained within the suffix