bpo-42911: Addition chains

[jneb]

I put the code that I can only test with 3.8 here, to see actually works in 3.10.
Short description: speeding up pow(int,int) especially for exponents with between 20 and 500 bits (until the fiveary method kicks in) , but is also saves time outside this range.
Speedup is between 2 % (for very high exponents) to 15% (for exponents of a few hundred bits).
And, I speeded up _Py_bit_length for generic processors just because I could :-)

https://bugs.python.org/issue42911

[jneb]

Seems to work under Python 3.10. That's a relief!

[jneb]

So it coredumps with negative exponents. Must be refcount problem. Working on it.

[jneb]

Finally i figured out how to test this on my own machine.
One should learn to operate the fire extinguisher before lighting a fire...
So, I am now convinced I'll get it to work (eventually :-) )

[jneb]

I found the refcount bug (Yay), and a few bugs that were slowing down things unnecessarily. (That's the disadvantage of using heuristics: it still works even if you do it wrong).
Ready to do some speed measurements and move those changes from my 38_fast_pow to the addition chain branch

[jneb]

Interesting. Now the test failed because apparently ** 1 must be an int. My routine optimizes this case and returns the original value.
I guess this had to be "fixed" even though I disagree :-)
Of course, adding zero to something also makes it an int. Interesting edge case.

[jneb]

Finally!
I'll put some speed measurement code and results (measured on 3.8 due to lack of resources) in the file additionchain.txt. If someone wants to verify the same is true on 3.10, please do.

[jneb]

See file Misc/additionchain.txt for some timings I just made

[mdickinson]

Closing following discussion in the issue.