Is asm code selection optimal?

Torbjörn Granlund tg at gmplib.org
Tue May 16 21:27:28 UTC 2017


"Marco Bodrato" <bodrato at mail.dm.unipi.it> writes:

  I'm not sure I'm able to understand what exactly the columns report...
  
  I expected the "default" column always equivalent to one of the other ones...
  
  But it sometimes is worst than all alternatives...
  
  mpn_divrem_1.3
      default      x86_64        coreisbr      core2
                   divrem_1.asm  divrem_1.asm  divrem_1.asm
  1   70.0224      34.3025       37.2641       37.6082
  2   42.3045      22.0282       23.7997       24.3283
  3   30.4712      19.5858       19.5641       19.8066
  4   24.7977      16.3102       17.6666       17.7169
  
  Do you think it may make sense to also add a column with the C version?
  
Yes.  :-)

And in the case you quote I'd assume the default is actually the C code,
i.e., that the config did not pick up any asm.

I'm bored by the manual asm config, and hopefully this new
measure-and-report stuff could be improved to suggest reconfig.

-- 
Torbjörn
Please encrypt, key id 0xC8601622


More information about the gmp-devel mailing list