Because the original issue was one of the Windows version of CSV_XS, I have composed a Benchmark similar to my orignal. This covers reading only, and uses the ASPN version of CSV_XS (as opposed to compiling it myself):

#!/usr/bin/perl
use strict;
use warnings;

use Benchmark qw(:all);
my @buffer;

cmpthese( 200, { 'xSV' => \&xSV, 'CSV' => \&CSV_XS } );

sub xSV {
    print STDERR '.';
    use Text::xSV;
    my $xsv = new Text::xSV(filename=>"sample.csv", sep=>';');
    while (my $row = $xsv->get_row) {
        @buffer = @$row; #just a "do something" instruction
    }
    undef @buffer;
    print STDERR '.';
}

sub CSV_XS {
    print STDERR '*';
    use Text::CSV_XS;
    use IO::File;
    my $io = new IO::File "< sample.csv";
    my $csv = new Text::CSV_XS({sep_char=>';'});
    while (my $row = $csv->getline($io)) {
        last unless @$row;
        @buffer = @$row; #just a "do something" instruction
    }
    undef @buffer;
    $io->close;
    print STDERR '*';
}
[download]

Results for three runs (I've trimmed the status indicators):

1.       Rate  xSV  CSV
   xSV 2.11/s   -- -75%
   CSV 8.51/s 304%   --
   

2.       Rate  xSV  CSV
   xSV 2.20/s   -- -74%
   CSV 8.52/s 287%   --
   

3.       Rate  xSV  CSV
   xSV 2.19/s   -- -74%
   CSV 8.40/s 284%   --
   

As you can see, Text::CSV_XS is noticably faster than Text::xSV, though the margins are not as large as on my Linux machines. I'm glad that someone pointed out that CSV_XS is now available through ASPN, as it allows me to write fast, portable CSV-handling code now. ;-)

I guess I don't know how to optimize a compiler for Windows... but I knew that already.