I did another test, including a wrapped memchr as
I stated in another reply. Including array handling,
I found a factor of 3 (to index/regex) even on my non-
mem alignment (compared to Core2) critical system:

1769416 in
            Rate  by_regex  by_index by_memchr
by_regex  51.2/s        --       -3%      -68%
by_index  52.9/s        3%        --      -67%
by_memchr  162/s      217%      207%        --
[download]

The code used here was:

use strict;
use warnings;
use Benchmark qw( cmpthese );

my $fn = '/boot/vmlinux-2.6.18.8-96-default.gz';
open my $fh, '<', $fn or die $!;
read $fh, my $buffer, 2_000_000 or die $!;
print length $buffer, " in\n";
close $fh;

my $subs = {
    by_index => sub {
       my ($p0, @offs)=(-1, ());
       push @offs, $p0 while +($p0=index $buffer, "\xaa", $p0+1) != -1
+;
       push @offs, $p0 while +($p0=index $buffer, "\xbb", $p0+1) != -1
+;
       push @offs, $p0 while +($p0=index $buffer, "\xcc", $p0+1) != -1
+;
       return 0 + @offs
    },
    by_regex => sub {
       my @offs=();
       push @offs, pos($buffer) while $buffer =~ /\xaa/g;
       push @offs, pos($buffer) while $buffer =~ /\xbb/g;
       push @offs, pos($buffer) while $buffer =~ /\xcc/g;
       return 0 + @offs
    },
    by_memchr => sub {
       my @offs=();
       my_memchr( \@offs, $buffer, "\xaa" );
       my_memchr( \@offs, $buffer, "\xbb" );
       my_memchr( \@offs, $buffer, "\xcc" );
       return 0 + @offs
    }
};

cmpthese -3, $subs;

use Inline C => qq[
 /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
  void my_memchr(SV* rvav, SV* sv, SV *ch)
 {
  STRLEN srclen;
  char byte = *SvPV(ch, PL_na) ;
  char *svc = SvPV(sv, srclen);
  char *p   = svc, *end  = svc + srclen;
  AV   *av  = (AV*)SvRV(rvav);  // if(SvTYPE(SvRV(rvav)) == SVt_PVAV)
  
  while((p=memchr(p, (int)byte, end-p)) !=0 && p<end) {
     av_push(av, newSViv(p-svc));
     ++p;
  }
 }
 /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
];
[download]

Regards
mwa