When C implementation of the unpickle reads a data, it also prefetches some data using peek() (if available) and then concatenates read and peeked chunks in the one input buffer. This causes an additional copying when a large data is read. The proposed patch gets rid of concatenating by moving a peeking before reading.

Do you have benchmark results to show the code with the patch is faster?

No, I have no any interesting results. ;) If an effect exists, it should reveal itself only in rare cases and be very small. This patch may be considered rather as a tiny refactoring (it decreases a number of code lines).

See issue18073 for test case.

I'd prefer to see efforts go towards finishing the PEP 3154 implementation...

This patch is orthogonal to the PEP 3154 framing. The affecting of issue18073 is only an unexpected side effect.

Patch is synchronized with tip (it was desynchronized since 23459df0753e).

Microbenchmark:

$ ./python -c "import pickle; d = b'x' * 10**6; f = open('test.pickle3', 'wb'); pickle.dump(d, f, 3); f.close()"
$ ./python -m timeit -s "from pickle import load"  "with open('test.pickle3', 'rb') as f: load(f)"

Unpatched:  100 loops, best of 3: 7.27 msec per loop
Patched:    100 loops, best of 3: 4.87 msec per loop

Looks good to me! Feel free to commit.

New changeset d565310e7ae3 by Serhiy Storchaka in branch 'default':
Issue #17897: Optimized unpickle prefetching.
http://hg.python.org/cpython/rev/d565310e7ae3

Thank you Alexandre.