Thanks, you are right, I missunderstood my own results. Measured now on the reduced case, best of five, list rebuilt each iteration: with the thread 1.8 / 5.1 / 21.0 ms for 4 / 16 / 48 entries, with -fno-thread-jumps 2.1 / 5.8 / 23.9 ms. So it buys 12 to 14 % on this loop for 58 -> 82 x86_64 insns, and the path is in the inner loop at estimated frequency 10 to 20.
Sorry for the noise.
