🚀 go-pugleaf

In article <1995Oct1102.48.02.1823@silverton.berkeley.edu>,
D. J. Bernstein <djb@silverton.berkeley.edu> wrote:
>Joe Buck <jbuck@synopsys.com> wrote:
>> 	/* now res is at most 32767+4 */
>> 	return (res < 32767) ? res : res - 32767;

>Branching is bad. Look at the routines that use res, and rewrite them so
>that they won't mind inputs up to 32771.

Branches aren't necesssary, nor are they necessarily bad.  On some
machines, if well-scheduled, they have effectively zero cost, or very
close to it.  If you've got conditional move (e.g., Sparc V9), I think
you could do that with something like:

  sethi %hi(32767),%g1
  or    %g1,%lo(32767),%g1
  cmp   res,%g1
  movl  %icc,0,
  sub   res,%g1,res

It would be helpful to spread this out more (in scheduling terms) since each
instruction depends on its predecessor.  Inlining would be a win.

There are also branchless idioms for doing this in Sparc V8, but
they're probably not a win over the cost of the branches.

speaking for myself,

David Chase