The errors in the line buffer and tokenizer now have diagnostics. The
line number is trivial to keep track of due to the line buffer, but
the column index requires quite a bit of juggling, as we pass
successively trimmed down buffers to the internals of the parser.
There will probably be some column index counting problems in the
future. Also, handling the diagnostics is a bit awkward, since it's a
mandatory out-parameter of the parse functions now. The user must
provide a valid diagnostics object that survives for the life of the
parser.
Since the tokenizer is decoupled from the parser, there's no good way
to do this. Also without attempting to parse the last line, it's
impossible to say if it is junk data or simply a missing trailing new
line.
When the buffer was separated from the tokenizer, we lost some
validation, including really aggressive carriage return detection.
This brings this back in full force and adds some additional
validation on top of it.
With my pathological 50MiB 10_000 line nested list test, this is
definitely slower than the one shot parser, but it has peak memory
usage of 5MiB compared to the 120MiB of the one-shot parsing. Not bad.
Obviously this result is largely dependent on the fact that this
particular benchmark is 99% whitespace, which does not get copied into
the resulting document. A (significantly) smaller improvement will be
observed in files that are mostly data with little indentation or
empty lines.
But a win is a win.
finally the flow parser has been "integrated" with the main parser in
that they now share a stack. The bigger thing is that the parsing has
been decoupled from the tokenization, which will allow parsing
documents without loading them fully into memory first.
I've been calling this the streaming parser, but it's worth noting that
I am referring to streaming input, not streaming output. It would
certainly be possible to do streaming output, but I am not interested
in that at the moment (it would be the lowest-memory-overhead
approach, but it's a lot of work for little gain, and it is less
flexible for converting input to objects).
I don't like big monolithic source files, so let's restructure a bit.
parser.zig is still bigger than I would like it to be, but there isn't
a good way to break up the two state machine parsers, which take up
most of the space. This is the last junk commit before I am seriously
going to implement the "streaming" parser. Which is the last change
before implementing deserialization to object. I am definitely not
just spinning my wheels here.
This is a simplification, but the main motivation is that the flow
parser stack can be integrated with the main parser stack because they
are not disparate types any more.
There's still a fair amount lurking in here, but I believe this logic
is sound. Rather than duplicating the map/list logic under the
opposing key, we set the logic up to use the second loop around
(this is was how dedents worked, and now it also works for indents).
I'm not convinced this is as easy to follow, and it did lead me to add
some additional unreachables to the code, which should maybe be turned
into error returns instead. It does reduce the odds of a code change
missing a copied instance, which I think is a good thing.
I didn't do an exhaustive search, but it seems that the managed
hashmaps only allocates space for the structure of the map itself, not
its keys or values. This mostly makes sense, but it also means that
this was only working due to the fact that I am currently not freeing
the input buffer until after iterating through the parse result.
Looking through this, I'm also reasonably surprised by how many times
this is assigned in the normal parsing vs the flow parsing. There is a
lot more repetition in the code of the normal parser, I think because
it does not have a granular state machine. It may be worth revisiting
the structure to see if a more detailed state machine, like the one
used for parsing the flow-style objects, would reduce the amount of
code repetition here. I suspect it certainly could be better than it
currently is, since it seems unlikely that there really are four
different scenarios where we need to be parsing a dictionary key.
Taking a quick glance at it, it looks like I could be taking better
advantage of the flipflop loop on indent as well as dedent. This might
be a bit less efficient due to essentially being less loop unrolling,
but it would also potentially make more maintainable code by having
less manual repetition.
As I was thinking about this, I realized that data serialization is
much more of a bear than deserialization. Or, more accurately, trying
to make stable round trip serialization a goal puts heavier demands on
deserialization, including preserving input order.
I think there may be a mountain hiding under this molehill, though,
because the goals of having a format that is designed to be
handwritten and also machine written are at odds with each other.
Right now, the parser does not preserve comments at all. But even if
we did (they could easily become a special type of string), comment
indentation is ignored. Comments are not directly a child of any other
part of the document, they're awkward text that exists interspersed
throughout it.
With the current design, there are some essentially unsolvable
problems, like comments interspersed throughout multiline strings. The
string is processed into a single object in the output, so there can't
be weird magic data interleaved with it because it loses the concept
of being interleaved entirely (this is a bigger issue for space
strings, which don't even preserve a unique way to reserialize them.
Line strings at least contain a character (the newline) that can
appear nowhere else but at a break in the string). Obviously this isn't
technically impossible, but it would require a change to the way that
values are modeled.
And even if we did take the approach of associating a comment with,
say, the value that follows it (which I think is a reasonable thing to
do, ignoring the interleaved comment situation described above), if
software reads in data, changes it, and writes it back out, how do we
account for deleted items? Does the comment get deleted with the item?
Does it become a dangling comment that just gets shoved somewhere in
the document? How are comments that come after everything else in the
document handled?
From a pure data perspective, it's fairly obvious why JSON omits
comments: they're trivial to parse, but there's not a strategy for
emitting them that will always be correct, especially in a format that
doesn't give a hoot about linebreaks. It may be interesting to look at
fancy TOML (barf) parsers to see how they handle comments, though I
assume the general technique is to store their row position in the
original document and track when a line is added or removed.
Ultimately, I think the use case of a format to be written by humans
and read by computers is still useful. That's my intended use case for
this and why I started it, but its application as a configuration file
format is probably hamstrung muchly by software not being able to
write it back. On the other hand, there's a lot of successful software
I use where the config files are not written directly by the software
at all, so maybe it's entirely fine to declare this as being out of
scope and not worrying about it further. At the very least it's almost
certainly less of an issue than erroring on carriage returns. Also the
fact that certain keys are simply unrepresentable.
As a side note, I guess what they say about commit message length being
inversely proportional to the change length is true. Hope you enjoyed
the blog over this 5 character change.
The goal here is to support a streaming parser. However, I did decide
the leave the flow item parser state machine as fully buffered
(i.e. not streaming). This is not JSON and in general documents should
be many, shorter lines, so this buffering strategy should work
reasonably well. I have not actually tried the streaming
implementation of this, yet.
This makes handling Value very slightly more work, but it provides
useful metadata that can be used to perform better conversion and
serialization.
The motivation behind the "scalar" type is that in general, only
scalars can be coerced to other types. For example, a scalar `null`
and a string `> null` have the same in-memory representation. If they
are treated identically, this precludes unambiguously converting an
optional string whose contents are "null". With the two disambiguated,
we can choose to convert `null` to the null object and `> null` to a
string of contents "null". This ambiguity does not necessary exist for
the standard boolean values `true` and `false`, but it does allow the
conversion to be more strict, and it will theoretically result in
documents that read more naturally.
The motivation behind exposing flow_list and flow_map is that it will
allow preserving document formatting round trip (well, this isn't
strictly true: single line explicit strings neither remember whether
they were line strings or space strings, and they don't remember if
they were indented. However, that is much less information to lose).
The following formulations will parse to the same indistinguishable
value:
key: > value
key:
> value
key: | value
key:
| value
I think that's okay. It's a lot easier to chose a canonical form for
this case than it is for a map/list without any hints regarding its
origin.
I was pretty sloppy with the code organization while writing out the
state machines because my focus was on thinking through the parsing
process and logic there. However, The code was not in good shape to
continue implementing code features (not document features). This is
the first of probably several commits that will work on cleaning up
some things.
Value has been promoted to the top level namespace, and Document has an
initializer function. Referencing Value.List and Value.Map are much
cleaner now. Type aliases are good.
For the flow parser, `popStack` does not have to access anything except
the current stack. This can be passed in as a parameter. This means
that `parse` is ready to be refactored to take a buffer and an
allocator.
The main next steps for code improvement are:
1. reentrant/streaming parser. I am planning to leave it as
line-buffered, though I could go further. Line-buffered has two main
benefits: the tokenizer doesn't need to be refactored significantly,
and the flow parser doesn't need to be made reentrant. I may
reevaluate this as I am implementing it, however, as those changes
may be simpler than I think.
2. Actually implement the error diagnostics info. I have some skeleton
structure in place for this, so it should just be doing the work of
getting it hooked up.
3. Parse into object. Metaprogramming, let's go. It will be interesting
to try to do this non-recursively, as well (curious to see if it
results in code bloat).
4. Object to Document. This is probably going to be annoying, since
there are a variety of edge cases that will have to be handled. And
lots of objects that cannot be represented as documents.
5. Serialize Document. One thing the parser does not preserve is
whether a Value was flow-style or not, so it will be impossible to
do round-trip formatting preservation. That's currently a non-goal,
and I haven't decided yet if flow-style output should be based on
some heuristic (number/length of values in container) or just never
emitted. Lack of round-trip preservation does make using this as a
general purpose config format a lot more dubious, so I will have to
think about this some more.
6. Document to JSON. Why not? I will hand roll this and it will suck.
And then everything will be perfect and never need to be touched again.
This was the final feature I wanted to add to the format. Also some
other things have been cleaned up a little bit (for example, the
inline parser does not need the dangling key to be attached to each
stack level just like the normal parser doesn't). There was also an
off-by-one error that bugged out detecting the pathological case of a
flow list consisting of only an empty string (`[ ]`, not to be
mistaken for the empty list `[]`).
Mixed multiline strings are a bit confusing but internally consistent.
> what character does this string end with?
|
ends with a newline character because that's the style of the
second-to-last line. However, seeing | last makes my brain think it
should end with a space. The reason it ends with a newline is because
our concatenation strategy consists of appending to the string early
(as soon as a line is added) rather than lazily. This is a tradeoff,
though. while lazy appending would make this result more intuitive
(the string would end with a space) and it would allow us to remove
the self-proclaimed cheesy hack, it would make the opposite boundary
condition confusing:
>
| what character does this string start with?
With lazy appending, this string would start with a space
(despite > making it look like it should have a leading newline).
While both of these are likely to be uncommon edge cases, it doesn't
seem we can have it both ways. Of the two options, I think the current
logic is a little bit more clear.
This was kind of a pain in the butt to implement because it basically
required a second full state machine parser (though this one is a bit
simpler since there are less possible value types). It seems likely to
me that I will probably shove this directly into the main parser
struct at some point in the near future.
There was no actual check that lines weren't being indented too far.
Inline strings weren't having their trailing newline get chopped.
Printing is still janky, but it's better than it was.
This hand-rolled wonder of switch statements is capable of parsing a 5
byte document in less than a gigasecond.
This was an interesting exercise in writing a non-recursive parser for
a nested structure format. There's a lot of very slightly different
repetition, which I'm not wild about, but it can handle deeply nested
documents. I tested with a 50 mb indented list tree document (10_000
lines of nesting) and a ReleaseFast build was able to parse it in
approximately 50 ms with a peak memory footprint of about 100 MB (of
which, half was the contents of the document itself, as the file is
read into a single allocated buffer that does not get freed until
program exit). I don't consider myself to be someone who writes high
performance software, but I think those results are quite acceptable,
and I doubt any recursive implementation would even be able to parse
that document at all (the python NestedText implementation smashes
directly into a RecursionError, unsurprisingly).
Anyway, let's call this a success. I will actually probably export this
to a separate project soon. The main problem is coming up with a name.
I also strongly suspect there are some lurking bugs still, and I think
I do want to add nested inline map/list support (and also parsing
directly into objects).
