Score: 3  Confidence: 1
       Interest: 4 | Accessibility: 2 | Quality: 3 | Contribution: 3

This paper augments the Proebsting-Watterson algorithm for fusing pipelines of
filters of network protocols in two ways: by using a lazy symbolic execution,
and by developing three optimization techniques.  The paper is extremely poorly
presented.  The application area appears to be layered network protocols, but
this is not made clear until almost the end of the paper.  The paper assumes
that the reader is familiar with both the application domain and the P-W fusion
algorithm.  The paper is written in vague generalities that do not help with
understanding the paper.  More seriously, it appears that the techniques
proposed in the paper are not very effective anyway.  The performance results
indicate little or no performance improvement in most of the benchmarks, and a
performance loss in some.  Given this largely negative result, the paper cannot
be considered seriously.

////////////////////////////////////////////////////////////////////////////

Score: 4  Confidence: 1
       Interest: 5 | Accessibility: 4 | Quality: 5 | Contribution: 4

This paper describes how the Proebsting-Watterson filter fusion algorithm can
be changed in order to generate code which is easier to optimise. Three
optimisations are given: filter loop fusion, binary loop fusion and
subroutines. These transformations have been implemented, and the results are
given for a number of test inputs.  My comments are as follows: There is not
enough description of the filter fusion algorithm itself to set the scene for
the later work. The code generation techniques are also not described very well
- perhaps some diagrams would help here. It is also not made clear what is
previous work, and what are the new contributions of this paper.  None of the
transformations are described very well. For example, consider the example of
filter loop fusion in Figure 1. C0 and F0 are included within this figure, but
they are not included in the regular expression describing it. The second box
in Figure 1b should contain F1 rather than F2. F3 is included in Figure 1b, but
there is no indication as to what it represents. Also, a basic block B1 is
referred to within Figure 1, but it is not labelled. The third step in filter
fusion, and the other two transformations are not described in sufficient
detail - more examples should be given. There are also no proofs of correctness
for each of the transformations.  In the results section, the legends in the
graphs can actually be confused with the data itself, and it is not made clear
what the information in the last graph actually represents. The conclusions
drawn from these graphs are not particularly convincing.  Other minor points:
Page 2, column 1 - third last line: remove one occurrence of the 'write' Page
2, column 2 - line 6: reading data stream -> reading a data stream Page 4,
column 1 - line 13: in fused function -> in a fused function Page 4, column 2 -
line 1: filters need to be -> filters needs to be

////////////////////////////////////////////////////////////////////////////

[The third reviewer signed the review, so I've deleted it in preference to
 getting permission to republish.]

////////////////////////////////////////////////////////////////////////////