I did something similar for word boundary zero-width assertions (i.e. \b and \B); that is, rewrite the PCRE by expanding permutations inline. The purpose was to compile as many PCREs as possible to DFAs using Ragel.
Fortunately, Ragel supported a limited form of assertions called semantic conditions, which it implements by expanding the code space of each state and allows you to associate a C boolean expression as a predicate for entering that state. For various reasons the semantic condition needed to be attached to a terminal, not to subexpressions (i.e. groupings and alternations).
Also, expressions like (cat?|dog?)\b needed to be permuted similar to the backreference problem, resulting in something that might look like (ca\b|cat\b|do\b|dog\b). This strategy didn't work for kleene stars (*), but usually a semantic condition could be attached to trailing terminals instead of leading terminals.
The actual code to perform these AST transformations was surprisingly simple. Excluding parsing of PCREs into an annotated AST and generation of the code, less than 200 lines of Lua code, maybe.
The PCREs weren't simple, but most were only moderately complex. For a corpus of about 10,000 PCREs about 1/2 used word boundary assertions. The vast majority of those (possibly even all of them... I can't remember) could be successfully transformed. With nested subexpressions (groups, alternations) the number of permutations rapidly grows, but I don't think our production corpus ever caused a problem in that regard, and only the most insane regular expressions nest subexpressions more than a few deep.
Before we tackled back references and other gnarly PCRE features I think we managed to coax better than 90% of our PCRE corpus to transform into semantically equivalent Ragel-based DFAs. In other words, back references were rare. With 90%+ compiled completely to native machine code and the remainder simplified into prefilters we saw greater than 10x performance improvement (> 1000%) over libpcre in terms of production throughput, not just benchmarks. (Ragel is awesome!) RE2 was usually faster than libpcre, but RE2's performance wasn't even remotely in the same league as the native machine code DFA Ragel produced, so I quickly ditched the idea of using RE2 anywhere in the pipeline.
That last 10% or so took much longer to tackle, and eventually most of my original code was dumped. But getting to about 90% was surprisingly easy. We did eventually get (and continue to maintain) 100% transformation to native machine code, but a small percentage still require backtracking. All-in-all I _think_ we're at 20x to 30x over baseline libpcre; that is, tacking the last 10% put us 2x or 3x atop the 10x. It wasn't a simple case of diminishing returns as the most difficult expressions to transform also tended to be the most costly at runtime, but if we had stopped early on at 90% it still would have been a resounding success.
Regarding RE2, it's important to note once an expression was transformed into a proper regular expression, they could be easily joined into a single union. Setting aside the bytecode vs machine code differences, RE2 just isn't capable of compiling huge unions of expressions (on the order of 100s or 1000s), whereas for Ragel it was a relative breeze. Another team was using re2c (not RE2) for a project, and it turned out that what it literally took re2c DAYS to compile only took SECONDS for Ragel to compile. GCC and clang also became bottlenecks compiling the Ragel generated code. And interestingly they both, coincidentally, exhibited (and still exhibit, AFAIK) quadratic complexity when parsing certain C constructs, such as initializer lists. So that necessitated some workarounds.