Make errors/error reporting more lightweight

[saethlin]

Together, I measure these commits to be a 13% improvement in Message parsing speed. If there are changes in here that make anyone uncomfortable I'm happy to back them out. If they are indeed improvements, they'll become more obvious as I continue to knock time off the benchmark.

One other question I have is how much the quality of the error messages regresses once we no longer have the actual data in decode errors.

It looks like dropping the error payloads from DecodeError does not result in a measurable improvement, so I've put them back in.

[codecov]

Codecov Report

Merging #1409 (b0cc115) into main (ebaba8b) will decrease coverage by 0.06%.
The diff coverage is 59.18%.

@@            Coverage Diff             @@
##             main    #1409      +/-   ##
==========================================
- Coverage   85.16%   85.10%   -0.06%     
==========================================
  Files         153      153              
  Lines       15035    15042       +7     
==========================================
- Hits        12804    12801       -3     
- Misses       2231     2241      +10     

[bluejekyll]

Overall I really like how simple these changes are. Based on your research, we probably want to make all Errors Box their Kinds?

Let's discuss the one defensive thing that was removed.

[saethlin]

Overall I really like how simple these changes are. Based on your research, we probably want to make all Errors Box their Kinds?

I've been toying with ways to shrink the sizes of the error types, and yeah I think this is probably the best approach. There isn't really one type or pattern that's responsible for the error size situation; as far as I can tell it's mostly a bit of nesting here and a bit of nesting there. Nest just 2 levels and suddenly you've blown up the size at the top level. One could run around Boxing all the types that are responsible for the ErrorKind size, but doing that all over significantly alters the public interface, makes things harder to match on (no box patterns), and doesn't even shrink the size as much as possible by boxing the whole ErrorKind (without backtraces, usually 24 bytes by boxing members, 8 bytes with boxed ErrorKind).

[bluejekyll]

Yup, I agree.

@djc, do you have any concnerns? I know you've been reviewing most of these changes, but this looks good to me.

[djc]

Most of this looks good to me. I'd probably have put the changes around read_character() in a separate commit, since they seem logically unrelated to the error types and it'd be good to record some rationale for these changes specifically in a commit message, but maybe not worth it now. Let's merge this!