Thank you for posting code, but without input, replicating your problem is more of a challenge than it should be. See How do I post a question effectively?.

The problem is that your iterator gets reset every time you change your string, so you end up accumulating an infinite number of your modifications in your first match. A much cleaner way of accomplishing this would be to use the e modified -- see Search and replace in perlretut. You can then get it all done in one clean shot:

my $counter = 0;
$str =~ s/<tag1[^>]*>\s*<tag2>\s*<tag3>[^<]*\K(?=<\/endtag3>)/sprintf 
+" [ $c-%d]", ++$counter/egsi;
[download]

I've used the \K control character (Character Classes and other Special Escapes) and a look-ahead (Looking ahead and looking behind) to save from having to put back unnecessarily removed material. It also makes the result much clearer, I think.

As a side note, you should not use $a and $b as normal variables, since they have special meaning in a sorting context; see $a. It's usually considered poor form to use single character variable names, except in some very specific, common scenarios, e.g. $i, $j... for counters and $x, $y... for coordinates.