N_N_equal - an assembly lesson

[Sam S. <sd...@gn...>]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Bruno,
I found the way comelem.d:N_N_equal is written (see cmp1.c, attached) to
be counterintuitive, ugly, and unreadable, so I rewrote it in the more
algebraic way, using a single return (see cmp1.c, attached).
Then I thought that you probably has a reason why you wrote it the way
you did, and the only reason I could think of was optimization, so I
decided to compare the assembly for the 2 version.
Since I can hardly read assembly (I can guess what "jmp" does, but not
what "movzbl" does), all I did was compare the assembly length:
  165   378  2502 cmp1-O0.s
  163   386  2527 cmp1-O2.s
  181   417  2769 cmp2-O0.s
  163   382  2487 cmp2-O2.s
i.e.: the algebraic version optimizes to slightly smaller size than the
ugly one, but the unoptimized algebraic version is about 10% larger than
the ugly one.
Can you provide some insight into this issue?
Specifically, why did you use the ugly version - is it really better?
Thanks!
Sam.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.6 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iD8DBQFHWEtMPp1Qsf2qnMcRAmOaAJ0U/vzBnaQ7oU3z2ZVkWqAhGRcWLgCdHV1A
L7PlpFGcy5o4F477G2FVn0A=
=DN/4
-----END PGP SIGNATURE-----