Benchmarking is worthwhile in this instance. The regex backtracking turns an N*log(n) problem (assuming the sort dominates) into an N^2 problem. Here's the result of applying the two algorithms to the Net-Howto (which is 100 times smaller than the data set I initially used).

greg@spark:~/test$ cat sleepingsquirrel 
#!/usr/bin/perl

print "$_\n" for sort keys %{{map {$_,()} grep /^[a-z]+$/, (split /\s/
+, join(" ",<>))}};
greg@spark:~/test$ time sleepingsquirrel Net-HOWTO >words.txt

real    0m0.178s
user    0m0.158s
sys     0m0.016s

greg@spark:~/test$ cat jasper 
#!/usr/bin/perl

$/=undef;
print "$_\n" for sort <> =~ /\b([a-z]+)\b(?!.*\b\1\b)/sg
greg@spark:~/test$ time jasper Net-HOWTO >words2.txt

real    1m8.477s
user    1m8.471s
sys     0m0.003s
[download]

...only about 350x slower. YMMV