Then again, if you eliminate the optimisation advantage;

#!/usr/bin/perl

use strict;
use warnings;

use Benchmark 'cmpthese';
use List::Util 'shuffle';

our @data = shuffle 1 .. 10000;

cmpthese -1, {
    '<=>'  => q[@s1=sort sub{ $b<=>$a }, @data],
    ' - '  => q[@s2=sort sub{ $b - $a }, @data],
};

__END__
P:\test>junk
      Rate <=>  -
<=> 36.9/s  -- -2%
 -  37.4/s  2%  --

P:\test>junk
      Rate  -  <=>
 -  36.3/s  -- -3%
<=> 37.4/s  3%  --

P:\test>junk
      Rate  -  <=>
 -  35.8/s  -- -0%
<=> 35.8/s  0%  --

P:\test>junk
      Rate <=>  -
<=> 36.8/s  -- -0%
 -  36.9/s  0%  --
[download]

... it makes no difference whether the block returns (-1,0,1) or (-n,0,n), which is more relavent to the original discussion.

However, in C or asm, it might avoid an unnecessary booleanisation of a subtraction?