If you run out of ideas of how to make re2c better...
Reading multiple characters at a time is problematic due to the absence of alignment guarantees: even if unaligned reads are supported by the CPU, they are still likely to be slow.
But it should be possible to read a few characters, combine them into one using shifts and bitwise operations, and then switch on the combined character. It can be done with existing API; no new primitives are needed.
Of course it will only help in cases when the underlying DFA has long "chains" of transitions without branches that can be collapsed into one transition. But the optimization can be local (applied to some parts of the DFA).
The idea originated in a linux.org.ru thread (in Russian): https://www.linux.org.ru/news/development/16759268?cid=16762746
The original idea was to provide an API to render the generated AST into code and let the user implement it in the form of a plugin/library. This is problematic because 1) there is no stable AST, and 2) there are some language- agnostic optimizations and decisions that re2c takes way before the final rendering phase (so it is not as simple as providing a bunch or functions to generate code for different AST nodes).
There may be another way to provide any-language support: let the user define a set of primitives/configurations that re2c will use for language-agnostic decisions and for code generation (some of them exist already, but not enough to fully redefine the target language). This is akin to the idea of Midnight Commander's skins or syntax files. re2c can provide default syntax files for some languages (it already provides "standard" Unicode categories).
It would be good to support these operations on regular expressions, but it requires some research. Does TDFA present any special difficulties?
It would be good to have something like Bison locations (automatic tracking of line/column/filename inforamtion). This is primarily useful for tokenizers and large lexers.