Good points.

It's just that when balancing elegance & performance & dependencies, in some cases I end up favouring the "one big regex" approach - assuming I can work around its technical limitations.

For example, in one current use-case I want to search a large input stream for occurrences of any of several "things", and extract relevant information from each match. This is not exactly a "parser" use-case (since I don't care about most of the input stream, and don't want to build a unified AST from it) but it shares some similarities.

In my prototype design (which seems to be working out fine so far), I do it by defining an array entry for each "thing" like so:

my @things = (

    { parse  => qr/ regex0 /sx,
      action => sub { ... }        },
      
    { parse  => qr/ regex1 /sx,
      action => sub { ... }        },
    
    ...
);
[download]

And then when the program starts, it dynamically compiles all of those individual regexes into a big branching one of the form:

qr{ (?| regex0 (*MARK:0)
      | regex1 (*MARK:1)
      | regex2 (*MARK:2)
      | ...              ) }sx
[download]

...and starts matching this big regex against the input stream using a sliding window approach inspired by this old post by BrowserUk.

And whenever a match is found, it calls the corresponding action function roughly like so:

$things[$main::REGMARK]{action}->();
[download]

This approach is attractive to me because:

• Defining each thing's parsing and processing code separately but closely together, provides the best kind of encapsulation IMO.
• It is efficient. Especially if each branch starts with a literal part (rather than a capture etc.) – apparently the regex compiler applies an optimization in that case which allows it to blaze through parts of the input stream with no matches.
• It is cleanly extensible - I can easily add more entries to @things.
• It has no non-core dependencies.

Each branch/thing can have its own capture groups etc, and process them in its action function.
However, if their parsing regexes need more complex quantified parts, then the advanced techniques I described above become necessary in order to be able to keep this design.

If those techniques were built-in instead of requiring scary embedded code blocks, I think I would use this kind of approach more often.