Use faster byteswapping and avoid alignment-increasing casts #216

1. 
   sipa commented at 0:25 am on February 15, 2015: contributor
2. 
   sipa commented at 4:18 am on February 15, 2015: contributor

   9% faster SHA256 and RFC6979.
3. 
   gmaxwell commented at 8:08 am on February 15, 2015: contributor

   ACK. (remind me to test on BE if I forget to remove this request.)
4. 
   sipa commented at 7:31 pm on February 16, 2015: contributor

   clang seems to warn about changed alignment when casting. I don’t see why that is a problem, so perhaps we should try to avoid that warning?
5. 
   sipa commented at 3:28 am on February 18, 2015: contributor

   @theuni opinion?
6. 
   theuni commented at 6:49 am on February 19, 2015: contributor

   @sipa I took a quick stab at avoiding changing alignment requirements: https://github.com/theuni/secp256k1/commit/78df7673f9d31627ae6425face109a0df6ac08ac just to see how much trouble it’d be. Assuming I didn’t muck it up too badly, overhead seems minimal. It’s hard to imagine this causing any real issues, though.
7. 
   sipa commented at 9:48 am on February 21, 2015: contributor

   Hmm, a 1% slowdown. This shouldn’t be necessary at all…
8. sipa force-pushed on Feb 22, 2015
9. sipa renamed this:
   Use byteswapping where available, instead of manual shifting
   Use faster byteswapping and avoid alignment-increasing casts
   on Feb 22, 2015
10. 
    sipa commented at 2:31 pm on February 22, 2015: contributor

    I’ve rewritten this based on @theuni’s suggestion, taking it a step further. Oddly enough, it’s even faster than the previous version.

    Also, I’ve just switched to a different byteswapping macro, which recent GCC at -O2 and above automatically convert to a bswap instruction.

11. 
    sipa commented at 11:14 am on February 23, 2015: contributor

    @theuni clang-3.5 also optimizes the construct to a bswap; I haven’t tried any other implementation.
12. 
    sipa commented at 11:21 am on February 23, 2015: contributor

    10.5% faster hashing now.
13. Use faster byteswapping and avoid alignment-increasing casts. ae55e850c7
14. sipa force-pushed on Feb 23, 2015
15. 
    theuni commented at 11:29 pm on February 23, 2015: contributor

    Nice! Tested ACK. Roughly an 8% speedup here from a very quick test.
16. sipa merged this on Feb 24, 2015
17. sipa closed this on Feb 24, 2015

    ---

18. sipa referenced this in commit 137a8ec45b on Feb 24, 2015

Contributors
sipa gmaxwell theuni