You could reparse the string, if you really wanted you, by generating a reentrant parser. You would probably want a reentrant scanner, as well, although I suppose you could kludge something together with a default scanner, using flex's buffer stack. Indeed, it's worth learning how to build reentrant parsers and scanners on the general principle of avoiding unnecessary globals, whether or not you need them for this particular purpose.
But you don't really need to reparse anything; you can do the entire parse in one pass. You just need enough smarts in your scanner so that it knows about nested interpolations.
The basic idea is to let the scanner split the string literal with interpolations into a series of tokens, which can easily be assembled into an appropriate AST by the parser. Since the scanner may return more than one token out of a single string literal, we'll need to introduce a start condition to keep track of whether the scan is currently inside a string literal or not. And since interpolations can, presumably, be nested we'll use flex's optional start condition stack, enabled with %option stack, to keep track of the nested contexts.
So here's a rough sketch.
As mentioned, the scanner has extra start conditions: SC_PROGRAM, the default, which is in effect while the scanner is scanning regular program text, and SC_STRING, in effect while the scanner is scanning a string. SC_PROGRAM is only needed because flex does not provide an official interface to check whether the start condition stack is empty; aside from nesting, it is identical to the INITIAL top-level start condition. The start condition stack is used to keep track of interpolation markers ([ and ] in this example), and it is needed because an interpolated expression might use brackets (as array subscripts, for example) or might even include a nested interpolated string. Since SC_PROGRAM is, with one exception, identical to INITIAL, we'll make it an inclusive rule.
%option stack
%s SC_PROGRAM
%x SC_STRING
%%
Since we're using a separate start condition to analyse string literals, we can also normalise escape sequences as we parse. Not all applications will want to do this, but it's pretty common. But since that's not really the point of this answer, I've left out most of the details. More interesting is the way that embedded interpolation expressions are handled, particularly deeply nested ones.
The end result will be to turn string literals into a series of tokens, possibly representing a nested structure. In order to avoid actually parsing in the scanner, we don't make any attempt to create AST nodes or otherwise rewrite the string literal; instead, we just pass the quote characters themselves through to the parser, delimiting the sequence of string literal pieces:
["] { yy_push_state(SC_STRING); return '"'; }
<SC_STRING>["] { yy_pop_state(); return '"'; }
A very similar set of rules is used for interpolation markers:
<*>"[" { yy_push_state(SC_PROGRAM); return '['; }
<INITIAL>"]" { return ']'; }
<*>"]" { yy_pop_state(); return ']'; }
The second rule above avoids popping the start condition stack if it is empty (as it will be in the INITIAL state). It's not necessary to issue an error message in the scanner; we can just pass the unmatched close bracket through to the parser, which will then do whatever error recovery seems necessary.
To finish off the SC_STRING state, we need to return tokens for pieces of the string, possibly including escape sequences:
<SC_STRING>{
[^[\\"]+ { yylval.str = strdup(yytext); return T_STRING; }
\\n { yylval.chr = '\n'; return T_CHAR; }
\\t { yylval.chr = '\t'; return T_CHAR; }
/* ... Etc. */
\\x[[:xdigit]]{2} { yylval.chr = strtoul(yytext, NULL, 16);
return T_CHAR; }
\\. { yylval.chr = yytext[1]; return T_CHAR; }
}
Returning escaped characters like that to the parser is probably not the best strategy; normally I would use an internal scanner buffer to accumulate the entire string. But it was simple for illustrative purposes. (Some error handling is omitted here; there are various corner cases, including newline handling and the annoying case where the last character in the program is a backslash inside an unterminated string literal.)
In the parser, we just need to insert a concatenation node for interpolated strings. The only complication is that we don't want to insert such a node for the common case of a string literal without any interpolations, so we use two syntax productions, one for a string with exactly one contained piece, and one for a string with two or more pieces:
string : '"' piece '"' { $$ = $2; }
| '"' piece piece_list '"' { $$ = make_concat_node(
prepend_to_list($2, $3));
}
piece : T_STRING { $$ = make_literal_node($1); }
| '[' expr ']' { $$ = $2; }
piece_list
: piece { $$ = new_list($1); }
| piece_list piece { $$ = append_to_list($1, $2); }
