Optimize _double_var #164

1. 
   peterdettman commented at 4:16 pm on December 14, 2014: contributor

   • Save 1 _fe_negate
   • Prefer operations on local variables
   • Reorder code somewhat to group inline function calls together

   I am definitely looking for others to test this one. I measure up to 1% improvement for bench_verify, but of the three things I listed above, the “reorder code” measures as the most significant, so it’s possible I have only optimized for my machine/compiler somehow.

2. 
   sipa commented at 7:36 pm on December 14, 2014: contributor

   I can’t measure any improvement (with or without endomorphism optimization).
3. 
   peterdettman commented at 2:29 pm on December 16, 2014: contributor

   @sipa I was afraid you’d say that! Still, I don’t think I’ve gone crazy; I’ve been round and around with the benchmarking and there’s still a 1% difference for bench_verify (in my setup).

   Hopefully we could hear from a few others (maybe platform/compiler is a factor?). For the record, I’m on an i7, OSX 10.10.1, using gcc 4.9.2 from macports (but with the clang assembler via -Wa,-q), –with-field=64bit –enable-endomorphism=yes, -O3 -march=native.

   Actually, with endo=no, the difference widens to about 1.5%, as “expected” I guess. If I use -O2, the difference is much smaller, if actually there; similarly if I remove -march=native.

4. peterdettman force-pushed on Dec 16, 2014
5. peterdettman force-pushed on Dec 23, 2014
6. 
   sipa commented at 1:02 am on February 13, 2015: contributor

   This results in a measurable 0.2% slowdown for me…
7. 
   peterdettman commented at 5:42 am on February 13, 2015: contributor

   @sipa I’ll re-test when I get a chance.
8. Optimize _double_var

   - Save 1 _fe_negate
   - Prefer operations on local variables
   - Reorder code somewhat to group inline function calls together

   9f6aecbf23
9. peterdettman force-pushed on Feb 13, 2015
10. 
    peterdettman commented at 1:08 pm on February 13, 2015: contributor

    @sipa Well, I still see a difference (bench_verify), but it’s smaller now, ~0.2us/55us. Since it’s inconsistent at best… closed!
11. peterdettman closed this on Feb 13, 2015

    ---

Contributors
peterdettman sipa