Order of reported errors

One important philosophy of the error reporting is that the very first message should be the root cause of the problem whenever possible: users should not be forced to search a hundred messages to find the root cause. In this way, the procedure:

solve the first error
reconvert
solving the new first error
reconvert
etc.

should always deterministically lead to a resolution of all problems.

Error messages are normally sorted by file, line and column, regardless of which conversion stage they happened (e.g. a tokeniser error first gets reported before a parser error).

There is however one important exception to that: broken cross references are always reported last.

For example, consider the following syntactically wrong document:

= a

\x[b]

``
== b

Here we have an unterminated code block at line 5.

However, this unterminated code block leads the header b not to be seen, and therefore the reference \x[b] on line 3 to fail.

Therefore, if we sorted naively by line, the broken reference would shoe up first:

error: tmp.bigb:3:3: cross reference to unknown id: "b"
error: tmp.bigb:5:1: unterminated literal argument

But in a big document, this case could lead to hundreds of undefined references to show up before the actual root unterminated literal problem.:

error: tmp.bigb:3:3: cross reference to unknown id: "b"
error: tmp.bigb:4:3: cross reference to unknown id: "b"
error: tmp.bigb:5:3: cross reference to unknown id: "b"
...
error: tmp.bigb:1000:1: unterminated literal argument

Therefore, we force undefined references to show up last to prevent this common problem:

error: tmp.bigb:1000:1: unterminated literal argument
error: tmp.bigb:3:3: cross reference to unknown id: "b"
error: tmp.bigb:4:3: cross reference to unknown id: "b"
error: tmp.bigb:5:3: cross reference to unknown id: "b"
...

Order of reported errors

 Ancestors