I wrote:
> I'm inclined to accept 0001, reject 0002, and move on. This doesn't
> seem like an area that's worth a huge amount of discussion.
Done that way. I made a couple more cosmetic changes and added
test cases for the double-backslash code path (which hadn't been
covered in byteaout either, I see now).
BTW, in my hands the microbenchmark that Stepan suggested shows the
committed version to be about 40% faster than before for long input.
So apparently the StringInfo-ification suggested in 0002 gave back
just about all the performance gain from 0001.
regards, tom lane