To mangle Mark Twain: "There are four kinds of lies: Lies, Damn Lies, Statistics, and Benchmarks".
This is a good case for learning to use the Benchmark
module. I found that creating an anonymous sub ref was slightly faster than a full sub, and that a bare block had less efficiency gains (although still outstanding for this simple function) the larger the sample size. If you are sensitive to milliseconds of difference, or have extremely complex algorithms you should measure them in situ to determine whether a bare block is better than a subroutine.
To me, the main routine should be not unlike an outline. Some works, like the standard five paragraph essay, don't really need this. For reference books or novels, on the other hand, it can be an essential aid to the writer.
You can always paste the text of a subroutine into other parts of the script for production code where milliseconds count. This will assist the compiler in streamlining the code (although repeated functions will make the whole executable larger and perhaps increase compile time if you are compiling for each script load). But when you are designing the code, it would make sense to abstract most of your larger blocks, just like you would abstract chapters, sections, and paragraph sets when writing.