[MLton] caching elaboration

[Daniel C. Wang]

This is a general gripe I have with compilers of all sorts. It always 
annoyed me that errors are reported in the order they are detected during 
parsing and elaboration of a program.  This means errors get reported based 
on the wims of the implementation.

The order of error reporting should be decoupled from the order of error 
detection.

I think, it be nice to go to a buffered error message architecture. i.e. all 
the error messages for an invocation are collected into a data structure. 
There should be enough meta-information so you can post-process the error 
buffer into what every sensible reporting order makes sense. Users as well 
as testing scripts can rely on a particular order of error messages even if 
the order in which errors are detected radically changes within the 
implementation from version to version.

Henry Cejtin wrote:
> With regards to getting parse errors from later files before type errors from
> earlier ones, I would agree with Stephen, but more strongly, that it is better
> to see all errors from the first file first.  It just matches more how I
> think of compilation working.
>