My suspicion was confirmed about PyPy (My PyPy here is Python 3.6.1 (784b254d6699, Apr 16 2019, 12:10:48) [PyPy 7.1.1-beta0 with MSC v.1910 32 bit] on win32). In what follows, "heapq2.py" had exactly the `class merge` Python implementation from PR 18427.

pypy -m pyperf timeit -s "from random import random; from heapq import merge; from collections import deque; iters = [sorted(random() for j in ra
nge(10_000)) for i in range(20)]" "deque(merge(*iters), maxlen=0)"

    Mean +- std dev: 191 ms +- 3 ms

pypy -m pyperf timeit -s "from random import random; from heapq2 import merge; from collections import deque; iters = [sorted(random() for j in r
ange(10_000)) for i in range(20)]" "deque(merge(*iters), maxlen=0)"

    Mean +- std dev: 69.3 ms +- 2.2 ms

Another test: PyPy merging 2 iterables of size 100_000:

    heapq: Mean +- std dev: 91.4 ms +- 2.4 ms
    heapq2: Mean +- std dev: 52.4 ms +- 2.1 ms

PyPy merging 10_000 iterables of size 100:

    heapq: Mean +- std dev: 2.74 sec +- 0.07 sec
    heapq2: Mean +- std dev: 781 ms +- 19 ms