I had a generalized syslog processing tool that I hacked up for
a special case, removing a lot of code unnecessary for the purpose,
lifting tests out to a higher level so as to avoid expensive
operations such as 'split', and so on.

The special case tool then ran several times *slower* than the
original :(

I thought it was due to a difference in how I was handling I/O, but
testing showed that not to be the case, so I used -dSmallDprof .

The expensive line turned out to be a pattern match of the
form X.*Y.*Z where X was a plain string, and Y and Z had simple (?:a|b)
semantics.

I replaced this with two pattern matches in succession, matching
against first Y and then Z (and assumed X would always be present for my
purpose).

Even though the new code had two matches instead of one, the result
was twice as fast -- probably because perl didn't have to be bothered
with expensive back-up semantics on the .*'s.

So... the tip for today is that if you are doing a lot of matches,
processing large files, then multiple matches in succession might be
MUCH faster for you then using .*'s to join required subsequences.