Anyone willing to improve the benchmark?

They couldn't make it much worse :)

Incorporating one-off set-ups -- or even a test for one-off setups -- into the benchmark subs is like incorporating the build-time of a car in its race time.

If the substr code is meant to reflect my second option, you've completely misunderstood the purpose of the substr refs and assigning through a fixed scalar buffer.

Its also traditional to post the results of a typical run.

I may have a go at producing a more realistic benchmark later tonight. Key ingredients are that you must not exclude the IO, buffer and memory handling when benchmarking IO processing of large files. Yours excludes all of these.

Hint: You cannot do IO filter bechmarks using the Benchmark module. The only realistic test is to time processing actual files that are big enough that they do not fit into the filecache. And you must ensure that the cache is flushed between runs.