This speeds up pickling large bytes objects.

$ ./python -m timeit -s 'import pickle; a = [bytes([i%256])*1000000 for i in range(256)]' 'with open("/dev/null", "wb") as f: pickle._dump(a, f)'
Unpatched:  10 loops, best of 5: 20.7 msec per loop
Patched:    200 loops, best of 5: 1.12 msec per loop

But slows down pickling short bytes objects longer than 256 bytes (up to 40%).

$ ./python -m timeit -s 'import pickle; a = [bytes([i%256])*1000 for i in range(25600)]' 'with open("/dev/null", "wb") as f: pickle._dump(a, f)'
Unpatched:  5 loops, best of 5: 77.8 msec per loop
Patched:    2 loops, best of 5: 98.5 msec per loop

$ ./python -m timeit -s 'import pickle; a = [bytes([i%256])*256 for i in range(100000)]' 'with open("/dev/null", "wb") as f: pickle._dump(a, f)'
Unpatched:  1 loop, best of 5: 278 msec per loop
Patched:    1 loop, best of 5: 382 msec per loop

Compare with:

$ ./python -m timeit -s 'import pickle; a = [bytes([i%256])*255 for i in range(100000)]' 'with open("/dev/null", "wb") as f: pickle._dump(a, f)'
Unpatched:  1 loop, best of 5: 277 msec per loop
Patched:    1 loop, best of 5: 273 msec per loop

I think the code should be optimized for decreasing an overhead of _write_many().