forcemerge 15107 44861
stop

Suggested patch attached. A small test suite for replace-regexp-in-string has already been pushed to master -- very rudimentary, but better than nothing -- and the patch amends it with some new relevant cases that didn't work before.

It is basically your patch but slightly optimised; it turned out that the function call and allocation overhead of the original patch made it a tad too expensive (a pity, because it was very neat). Now performance is about the same as before when the pattern contains no submatches, and slightly above (< 10% slower) with one submatch. It seems worth the correctness.