optimized s_mp_sqr

[czurnieden]

s_mp_sqr is significantly slower than s_mp_mul here (without comba). Changing the addition r + r to r<<1 made it already faster and changing the computation in the carry propagating loop to a manual multi-word doubled the speed-up already gained. The relation between s_mp_mul and s_mp_sqr follows the theoretical runtimes much closer now.
A test for s_mp_sqr against s_mp_mul has been added to etc/test.c.

Those advantages depend on compiler and/or architectur, of course, but the changes should not make it worse.

[czurnieden]

I'm trying to clean up my PRs (quite a mess), so:
Is it useful or can it go?

[sjaeckel]

Thanks for reminding me of those PR's!

I rebased after merging #502 and fixed the merge conflict ... Does that merge make sense? Both tests test_s_mp_sqr() and test_s_mp_sqr_comba() do something slightly different, or should there only be one?

[czurnieden]

I rebased after merging #502 and fixed the merge conflict ... Does that merge make sense?

As far as I can see: yes.

Both tests test_s_mp_sqr() and test_s_mp_sqr_comba() do something slightly different, or should there only be one?

The outer loops are different, that is intentional but test_s_mp_sqr_comba() only works if Comba is switched off at compile time. Add a round to the CI-tests?

And, no, I don't have the faintest idea anymore why I added that small inner loop in test_s_mp_sqr_comba(). The known edge case is the main point, one test at a random point for the unknown edge cases should suffice.

[sjaeckel]

Add a round to the CI-tests?

And, no, I don't have the faintest idea anymore why I added that small inner loop in test_s_mp_sqr_comba(). The known edge case is the main point, one test at a random point for the unknown edge cases should suffice.

you wanna update these tests or should we merge as is?

[czurnieden]

you wanna update these tests or should we merge as is?

Merge as it is, because I know me and if I start fiddling around I will find this…and that…and those…and at the end a week is gone ;-)