How do I sanitize invalid UTF-8 in Perl?
My Perl program takes some text from a disk file as input, wraps it in some XML, then outputs it to STDOUT. The input is nominally UTF-8, but sometimes has junk inserted. I need to sanitize the output such that no invalid UTF-8 octets are emitted, otherwise the downstream consumer (Sphinx) will blow up.
At the very least I would like to know if the data is invalid so I can avoid passing it on; ideally I could remove just the offending bytes. However, enabling all the fatalisms I can find doesn't quite get me there with perl 5.12 (FWIW, use v5.12; use warnings qw( FATAL utf8 ); is in effect).
I'm specifically having trouble with the sequence "\xFE\xBF\xBE". If I create a file containing only these three bytes (perl -e 'print "\xEF\xBF\xBE"' > bad.txt), trying to read the file with mode :encoding(UTF-8) errors out with utf8 "\xFFFE" does not map to Unicode, but only under 5.14.0. 5.12.3 and earlier are perfectly fine reading and later writing that sequence. I'm unsure where it's getting the \xFFFE (illegal reverse-BOM) from, but at least having a complaint is consistent with Sphinx.
Unfortunately, decode_utf8("\xEF\xBF\xBE", 1) causes no errors under 5.12 or 5.14. I'd prefer a detection method that didn't require an encoded I/O layer, as that will just leave me with an error message and no way to sanitize the raw octets.
I'm sure there are more sequences that I need to address, but just handling this one would be a start. So my questions are: can I reliably detect this kind of problem data with a perl before 5.14? What substitution routine can generally sanitize almost-UTF-8 into strict UTF-8?
Answer
You should read the UTF-8 vs. utf8 vs. UTF8 section of the Encode docs.
To summarize, Perl has two different UTF-8 encodings. Its native encoding is called utf8, and basically allows any codepoint, regardless of what the Unicode standard says about that codepoint.
The other encoding is called utf-8 (a.k.a. utf-8-strict). This allows only codepoints that are listed as legal for interchange by the Unicode standard.
"\xEF\xBF\xBE", when interpreted as UTF-8, decodes to the codepoint U+FFFE. But that's not legal for interchange according to Unicode, so programs that are strict about such things complain.
Instead of using decode_utf8 (which uses the lax utf8 encoding), use decode with the utf-8 encoding. And read the Handling Malformed Data section to see the different ways you can handle or complain about problems.
Update: It does appear that some versions of Perl don't complain about U+FFFE, even when using the utf-8-strict encoding. This appears to be a bug. You may just have to build a list of codepoints that Sphinx complains about and filter them out manually (e.g. with tr).