Yes, the context-free idea isn't a logical necessity for expressing in BNF, because that alone can be conquered by expressing the different context types as alternative BNF definitions.
How so? Could you give a BNF snippet that would express how to generate a forward declaration of a subroutine that matches the signature of the actual subroutine definition?
Rather, the most compelling arguments that have been made in reply here come from those perl syntax elements which overpower the nearby context, such as (amongst other examples given by various people) the e-regexp modifier.
regexp modifier? In which way does that "overpower" nearby context? Couldn't that construct be parsed with the following BNF snippet:
non-a-slash: # string that doesn't contain a non escaped slash
SUBSTITION: non-eval-substition | eval-substitution
's' '/' not-a-slash '/' not-a-slash '/' non-e-modifiers
's' '/' not-a-slash '/' slashless-code '/' e-modifiers(
non-e-modifiers: 'x' | 's' | 'm' | 'g' | 'x' | 'i'
e-modifiers: 'e' | non-e-modifiers
Assuming you could create BNF for Perl code (and then you can create BNF for Perl code that doesn't contain a slash).
Grammars expressed in BNF can be parsed with recursive decent parsers (which do not have a fixed lookahead) - and if you can create a parser for all other Perl constructs, I can parse the difference between s/// and s///e.