Path: utzoo!attcan!uunet!lll-winken!lll-tis!helios.ee.lbl.gov!nosc!humu!uhccux!mikem
From: mikem@uhccux.uhcc.hawaii.edu (Mike Morton)
Newsgroups: comp.sys.m68k
Subject: Re: addq.w #n,sp and a pop quiz
Keywords: C asm addq
Message-ID: <2833@uhccux.uhcc.hawaii.edu>
Date: 14 Dec 88 19:33:31 GMT
References: <5005@bsu-cs.UUCP> <5460@cbmvax.UUCP> <1100@ncar.ucar.edu> <196@xpiinc.UU.NET>
Distribution: na
Organization: University of Hawaii
Lines: 29



Tom Talpey is right that the CMPM.B does more memory access.  But it *is* the
shortest way, and fewest cycles.  I've never actually used this trick, so I
don't know if it's empirically faster.  [I use a Mac Plus, which has funny
timing anyway...]

I also bow to Tom's point that the Carry and eXtend bits must be the same
for "SUBX.B Dn,Dn" to replace "SCS".  Incidentally, I got this trick from
an article Bill Gates wrote ca 1975 on 8080 (?) hacking, reprinted in
"Programmers at Work".

Counting the cycles in Tom's Max(Dx,Dy) as a function of word or long operands
and whether the branch takes/doesn't.

                                 word               long
	mov 	Dx,result        4                  4
	cmp 	Dx,Dy            4                  6
	bles	.+4                10/8 (taken/not)
	mov 	Dy,result        4                  4
                                 18/20              20/22

The Superoptimizer's solution has no branches at all, so its time is constant
at 16 cycles for word operands (small win) or 32 for longword (big loss).

Anyone care to solve the 4-instruction signum function?

 -- Mike Morton // P.O. Box 11378, Honolulu, HI  96828, (808) 676-6966 HST
      Internet: msm@ceta.ics.hawaii.edu
    (anagrams): Mr. Machine Tool; Ethical Mormon; Chosen Immortal; etc.