The encoding for the delimiter is the encoding in effect on the schema component carrying the property. Making them take on contextual encodings makes things much too complicated.
So yeah, I think in your case, if we're scanning for that "§" but we're using a decoder for ASCII, that's incorrect. These mixed encoding cases are all corner cases anyway, so they don't have to be natural or easy. The rules simply have to be easy to interpret. So your root element defines a terminator. That terminator's encoding has *nothing* to do with the encoding specified for a contained element within root. It is not that contained element's terminator, it is the root's terminator. The semantics of delimiter scanning in DFDL is in fact something that requires lowering the delimiters to byte patterns. This is required based on mixed scenarios like this, but also based on features like Byte-Value entities e.g., %#rHH; which specifies a hex byte that can appear, even in the middle of characters, when that byte makes no sense. <element name="foo" type="xs:int" dfdl:terminator="11%#r88;99" dfdl:encoding="utf-16BE"/> So the terminator of the above is bytes 0031 0031 88 0039 0039. See how that 88 is just thrown in there. Makes no sense in ANY encoding. We're even screwing up the character alignment here. That means scanning for delimiters in DFDL requires us to lower the scanning to bytes. Of course Daffodil doesn't implement %#rHH; byte-value (aka raw-bytes) entities except for one special case which is specifying the fill byte property. And our scanning is currently character oriented. So, what does it mean to scan for say a UTF-16 character '1' as terminator of an element that is in say, ASCII ? It means you are searching through the bytes ignoring ASCII, decoding it as UTF-16, looking for '1' (which is bytes 00 31). Then having found a 0031, the preceding bytes are then decoded as ASCII. Pretty sure Daffodil scanning isn't doing that. ________________________________ From: Sloane, Brandon <bslo...@tresys.com> Sent: Thursday, April 30, 2020 12:01 PM To: dev@daffodil.apache.org <dev@daffodil.apache.org> Subject: Re: Incorrect delimiter scanning when mixed encodings? Without looking at the spec, I would expect that delimiters be defined by the encoding the the element that defines the delimeter; so Daffodil is buggy in the case you describe. However, there are a couple of complications we have to consider: 1) What if instead of a terminator, we had a separator; and the separator is a valid character in both encoding; but has a different bytecode <xs:element name="root" > <xs:complexType> <xs:sequence dfdl:separator=","> <xs:element name="name" type="xs:string" maxOccurs="2" encoding="FOO"/> <xs:element name="address" type="xs:string" maxOccurs="2" encoding="BAR"/> </xs:sequence> </xs:complexType> </xs:element> In this case, I would expect the separator to be interperated based on the encoding of the individual elements, which is obviously not consistent with my expectation from your example. There is also the instance of the separator occurring between the two element types. So even in this case my naive expectation is not consistent. The correct answer here is probably to say that this example schema is wrong, and there should be 2 sequences, each defining their own separator. 2) What if the encodings have a different alignment? For instance, if the outer encoding that defines the delimiter is 8-bit and byte alligned, with a 7-bit inner encoding, should we look forward to the next byte boundary after every 7 bit character? 3) How does this interact with escape sequences? The solution here might be to think through some restrictions on where encoding changes are allowed to occur. I am not sure it is possible to give reasonable semantics for everything over a region that spans multiple encodings. ________________________________ From: Steve Lawrence <slawre...@apache.org> Sent: Thursday, April 30, 2020 11:15 AM To: dev@daffodil.apache.org <dev@daffodil.apache.org> Subject: Incorrect delimiter scanning when mixed encodings? Say we have a schema like this: <xs:schema xmlns:xs="http://www.w3.org/2001/XMLSchema" xmlns:dfdl="http://www.ogf.org/dfdl/dfdl-1.0/"> <xs:include schemaLocation="org/apache/daffodil/xsd/DFDLGeneralFormat.dfdl.xsd" /> <xs:annotation> <xs:appinfo source="http://www.ogf.org/dfdl/"> <dfdl:format ref="GeneralFormat" lengthKind="delimited" encoding="ISO-8859-1" /> </xs:appinfo> </xs:annotation> <xs:element name="root" dfdl:terminator="§"> <xs:complexType> <xs:sequence> <xs:element name="name" type="xs:string" /> </xs:sequence> </xs:complexType> </xs:element> </xs:schema> So we have a format that is all ISO-8859-1, and a delimited string called "name", and the root is terminated by "§" in the ISO-8859-1 encoding. If we have data that looks like this: text§ It will parse to this: <root> <name>text</name> </root> Now say we want just the "name" element to have a different encoding, so we change it to this: <xs:element name="name" type="xs:string" dfdl:encoding="US-ASCII" /> Now the terminator defined on the root element is in a different encoding than the delimited element. Note that the terminator § isn't even valid in this encoding. Currently, Daffodil does not successfully parse this. It scans the data, decoding a single character at a time looking for a delimiter. Eventually it gets to the § ata, the decoder says it's not valid in our ASCII encoding and converts it to the unicode replacement character. This of course doesn't match the delimiter we're looking for and continue on. The delimeter scanner then hits the end of data, and errors when it never finds the root termiantor. Is this the correct behavior, or is our delimiter scanning fundamentally broken? I wonder if the correct behavior is when the terminator comes into scope we should immediately encode it into its bytes. Delimiter scanning only looks for these bytes and doesn't actually decode any data. And only when bytes that match a delimiter are found do we decode all the bytes up until that point? Is this reasonable, or is this type of thing just not allowed? Note that this is somewhat hypothetical. I don't know of any formats that mix encodings like this, but this popped into my head looking at DAFFODIL-2323 which complains about encoding property and a sequence, which can have a siilar issue if the sequence encoding differs from element encodings.