Win98 is much like Unix¹ here. If you delete a file while a process has it open, then the directory entry for the file is removed but the contents of the file (and the FAT² or similar entries for the file) hang around until the last process to have the file open closes it.

Win98 is different from Unix in couple of ways that are related to this. First, when you open a file in Win98, you specify whether you want other processes to be able to read, write, and/or delete the file while you have it open. The "default"³ is to not allow any of this "sharing". But when using the C RTL⁴ (as Perl does) the default is to allow all sharing.

Secondly, Win98 doesn't support a file system that allows multiple directory entries per file⁵ (like most Unix file systems do). This is only tangentially related -- deleting a "hard link" for a file also doesn't delete the file contents just like deleting an open file doesn't.

¹ Yes, I know it is officially "UNIX", but I just prefer "Unix" and always will. Tough noogies. q-:
² File Allocation Table
³ I shy away from calling it a real default because you have to specify what type of sharing you want to allow, it is just much easier to specify no sharing (by specifying "0") than all sharing (by specifying FILE_SHARE_READ|FILE_SHARE_WRITE|FILE_SHARE_DELETE)
⁴ Run-Time Library
⁵ NTFS supports this, if in a rather awkward way

While playing with this I suggest taking CGI out of the picture. That will greatly simplify what is going on and let you understand better what is happening. (Random note, rather than writing a temp file then calling system, try doing an, open (PIPE, "| perl) then writing out your Perl script directly to the pipe.)

Incidentally the idea of polymorphic code has come up before. For instance camel code, and Genetic Programming or breeding Perls. There is more discussion at Code that writes code. And if you wanted to get truly bizarre, look at Lingua::Romana::Perligata.

Thanks tilly, i was confused about that. I still don't really expect that behavior. I narrowed down the source of my confusion with a perl to english translation:
1. Parent creates child and (system) executes it.
2. Child modifies parent (file).
3. Child calls parent URL with a CGI parameter.
   (at this point i would expect the parent to reload and ignore it's last line but)
4. Parent executes its last line of code which deletes the child.
5. Changed parent reloads with parameter supplied by the child.
This seems strange to me. I experimented with $| in both parent and child with no change (because sometimes this radically effects chained scripts, like when there's multiple children each with print statements, which is different but...

I noticed that the parent waits for the child to finish before executing any number of remaining lines in itself except print statements (which are just ignored!), before the browser responds to the Location print of the child.

Thanks for summarizing all the nodes on perlmonks relating to this. I've seen most of them but except for the camel code spoiler didn't learn much (because my humps aren't that big yet ;-). Other references include the Camel pages 44 and 140, Cookbook recipe 7.6, and:

• perlman:perldata
• perlman:lib:SelfLoader
• Perl Module Reference, Volume 1: Chapter 2
• Perl-Users Digest, Issue: 5919 Volume: 8

  and maybe even

• recursive self modifying eval japh
Thank you tye for clearing up the filesystem issue. I'll have to come back to this node to tackle Dominus' contribution someday after i've read a few perl books with carnivores on the cover.

        package Devel::Dumpcode;

        sub DB::DB { }    # Do nothing special
        
        sub main::source_of_function {
          my $package = caller;
          $function = $package . '::' . shift();
          my ($file, $start, $end) =
            $DB::sub{$function} =~ /(.*):(\d+)-(\d+)/;
          @{"::_<$file"}[$start..$end];
        }

        1;
[download]

        #!/usr/bin/perl

        for (;;) {
          print "Gen = ", Gen_val(), "\n";
          my @code = source_of_function('Gen_val');
          for (@code) {
            last if s{(\d+)(?=;\s+\# EDIT ME)}
                     {(($1 * 1103515245 + 12345) / 65536) % 32768}e;
          }
          eval join '', @code;
        }

        sub Gen_val {
          1;     # EDIT ME
        }
[download]

        perl -d:Dumpcode tryme.pl
[download]

The nice thing about this is that although it is a genuinely useful and effective technique, it's also extremely bizarre.