Net::IDN::Punycode versions before 2.590 for Perl allow CPU exhaustion via quadratic insertion cost when decoding a long label in decode_punycode.
The XS backend inserts each decoded code point into a UTF-8 buffer and finds the insertion point by scanning that buffer from the start, one character at a time. The scan runs once per code point over the output built so far, so the cost is quadratic in the label length. The pure-Perl backend downgrades its input to bytes so that substr can index it directly, but takes its working copy before the downgrade, so when the input carries the UTF-8 flag every substr on the copy scans from the start, with the same quadratic cost.
Nothing bounds the label length in the to-Unicode direction. The 63-byte DNS limit is checked only when converting to ASCII, so domain_to_unicode and uts46_to_unicode pass an attacker-supplied label of any length to the decoder.
Upgrade to Net-IDN-Encode 2.590-TRIAL or later.