GitLab

This commit does not pass CI or even fully compile, and while I usually
try to avoid those, A) I need a checkpoint!, and B) I think this one is
interestingly illustrative, and I'll probably want to refer to this
diff and the one that will follow it in the future as part of
architecture design records (or even possibly experience reports about
golang syntax).

In this commit: we have three packages:

- schema: full of interfaces (and only interfaces)
- schema/compiler: creates values matching schema interfaces
- schema/dmt: contains codegen'd types that parse schema documents.

The dmt package feeds data to the compiler package, and the compiler
package emits values matching the schema interface.
This all works very nicely and avoids import cycles.

(Avoiding import cycles has been nontrivial, here, unfortunately.
The schema/schema2 package (which is still present in this commit,
but will be removed shortly -- I've scraped most of it over into
this new 'compiler' package already, just not a bunch of the validation
rules stuff, yet) was a dream of making this all work by just having
thin wrapper types around the dmt types.  This didn't fly...
because codegen'd nodes comply with `schema.TypedNode`, and complying
with `schema.TypedNode` means they have a function which references
`schema.Type`... and that means we really must depend on that
interface and the package it's in.  Ooof.)

The big downer with this state, and why things are currently
non-compiling at this checkpoint I've made here, is that we have to
replicate a *lot* of methods into single-use interfaces in the schema
package for this to work.  This belies the meaning of "interface".
The reason we'd do this -- the reason to split 'compiler' into its own
package -- is most because I wanted to keep all the constructor
mechanisms for schema values out of the direct path of the user's eye,
because most users shouldn't be using the compiler directly at all.

But... I'm shifting to thinking this attempt to segregate the compiler
details isn't worth it.  A whole separate package costs too much.
Most concretely, it would make it impossible to make the `schema.Type`
interface "closed" (e.g. by having an unexported method), and I think
at that point we would be straying quite far from desired semantics.

As discussed on the issue thread, ipld.Kind and schema.TypeKind are more
intuitive, closer to the spec wording, and just generally better in the
long run.

The changes are almost entirely automated via the commands below. Very
minor changes were needed in some of the generators, and then gofmt.

	sed -ri 's/\<Kind\(\)/TypeKind()/g' **/*.go
	git checkout fluent # since it uses reflect.Value.Kind

	sed -ri 's/\<Kind_/TypeKind_/g' **/*.go
	sed -i 's/\<Kind\>/TypeKind/g' **/*.go
	sed -i 's/ReprKind/Kind/g' **/*.go

Plus manually undoing a few renames, as per Eric's review.

Fixes #94.

So far, the generation itself runs, but the result will not yet compile.
I just want a checkpoint here.  (Most of this was written a few days ago already.)

This is full of mundane plonking away at adding stars and ampersands,
of which approximately none are interesting.

(I'm particularly frustrated by this because these are all placeholder
types, and we're getting *very* close to replacing them, as we get
closer and closer to self-hosting... at which point all of this bonking
about will be made totally irrelevant.  And yet to close the last mile,
these "small" fixes are surprisingly irritating.  Ah well.)

The bits that *are* interesting:

- the Spawn functions for type info now take strings rather than types
(so that they don't provoke a cycle problem for the user when
constructing the information to describe cyclic type info);

- all of the Type info structure hold a pointer to the TypeSystem, and
use that to look up reified Type info for related types, so that their
methods don't force the caller to do that themselves.

(The TypeSystem pointer was already there, amusingly; just never before
initialized, because it hadn't turned out to be load-bearing yet.)

It also would've been possible to just change all the methods on the
Type types to return TypeName rather than full Type info.  That would
avoid the need to use a TypeSystem pointer.  I didn't because:

Overall, this was done in such a way as to minimize the diff that
impacts within the templates.  This was a goal because updating
templates is a fair bit more work than other code due to the weak
compiler support.  And we'll end up reviewing and changing these
methods when we get to our goal of self-hosting generation of the
schema types from the schema-schema, so, it's not worth pushing around
diffs in that same area when they'd be sure to be churned under soon.

I can't quite claim tests passed on the *first* shot... but,
the first shot after mostly syntactical (rather than semantic) fixes?
Yeah, actually.  That was pretty fun.

Snuck in a bit of DRY'ing up.  The repetition in BeginMap methods
got to me, and was low hanging fruit, so I extract that from unions
and also backported it to structs.

Errors got some work in this commit, because it turns out I've
straightjacketed myself by not allowing "fmt.Errorf" due to imports.
There's a lot to do there, and I only tackled what was directly
critical to get this commit about unions across the finish line,
but there's a few remarks in comments about where to go next.

Some more comments about future work on the type info holder types
also appears; I'm starting to skid on those placeholders and their
issues more and more.  I really hope we can get to replace those
sooner than later.

And... also, yes, the idea of not having a "focus" state field in
assemblers really bit it, hard, as speculated in the previous
commit message.  I ended up using 'ca' in more switches than I
expected, simply because it's easier to use that than have the
conditonal templating branches that would be necessary to use the other
tagging mechanisms that do also have sufficient info.

One big fixme in the core interfaces for nodebuilders (wince):
the ValuePrototype method can error sometimes, and that hasn't been
accounted for.  Need to make a decision about what to do there.
It's not really an exercised path in practice, but it shouldn't
contain caltrops, regardless of how frequently used it is.
(This probably would've come up earlier, except there's a bunch of
stubs about ValuePrototype in other parts of codegen already; all of
them need backfill, but haven't yet made it to top of the todo heap.)

But despite all the fixmes accumulated, this does bring unions
to be a usable thing, and definitively proves out that the design
still works, even for what turns out to be one of the most complicated
parts of the schema system!  It's very, *very* exciting to add the
checkmarks to this part of the feature table -- it's one of the places
I most feared "unknown unknowns", now it's put to bed.  Woot!

Also make some of the placeholder construction functions.
(Not all of them.  Just the ones I intend to use first.)

There was some very unfinished placeholder stuff going on there.
(And ironically, it wasn't written in a very clear union-like way;
it gets a lot less icky when it's rewritten so that it's uniony.)

The appearance of the word "any" there has started to perturb me;
"any" is a concept that we also need to describe in schemas, and that
type has nothing to do with it.  It's more of a base mixin.

Complex keys are a behemoth of a feature, and it's gonna take
substantial effort to sort all of that out.

This diff is pretty fun.  It's our first alternative representation,
and it all works out nicely (no internal syntactical complications
encountered or any other unexpected unpleasantness).

It's also our first representation that involves potentially recursive
destructuring work over a scalar!  So, you can see the new 'construct'
method conventions appearing to handle that.  Works out neatly.
This kind of recursive destructuring happens all within the span of
processing one "token" so to speak, so naturally it can work without
any stateful machinery.

Some utility functions are added to the mixins package.
(It's sort of betraying in the name of the package, but, well,
it's an extremely expedient choice.  See comments about import
management.  tl;dr: consistency makes things easier.)

Tests for a recursive case of this will be coming soon, but requires
addressing some other design flaws with the AdjunctConfig maps.
(StructField is... not a good key.  It's not always hashable... such
as when the StructField.Type field is inhabited by TypeStruct.)
Easy enough to fix, just not going to cram it into this commit.

The way null and the 'fcb' are handled here differs from previous
generators: here, we *always* have an 'fcb', and just make a special
one for the root builder.  This is... well, it works.  It was an
attempt to simplify things and generate fewer method overrides, and
it did.  However, I'm not going to dwell on this too long, because...

The 'fcb' system is not working out well overall for other reasons:
it's causing costly allocations.  (Taking a function pointer from
a method requires at least two words: one for the function pointer
and one for the value to apply it on.  That means an allocation.)
So a serious rewrite of anything involing the 'fcb' strategy is needed.

And that is going to be a little tricky.  The 'fcb' idea was itself
a trying-slightly-too-hard-to-be-clever attempt to avoid an alternative
solution that involves generating additional types per distinct child
value slot (and I'm still unwilling to pursue that one -- it suggests
far too many types).  The next best idea I have now seems to involve a
lot of pointers into other assembler's state machines; this will surely
be a bit touchy.  But no worse than any of the rest of all this, I
suppose: it's *all* "a bit touchy".  Buckle up for fun diffs ahead.)

So, considering these 'fcb' notes, this feels a bit mixed...
"Two steps forward, one step back", so to speak.  Still: progress!
And it's very exciting that we got through several new categories of
behavior without hitting any other new roadbumps.

Using golang's plugin feature, we can... well, *do* this.

To date, testing codegen has involved running the "test" in the gen
package to get it to emit code; and then switching to the emitted
package and _manually_ running the tests there.

Now, running `go test` in the gen package is sufficient to do
*everything*: both the generation, and the result compiling,
and we can even write tests against the interfaces and run those,
all in one step.

There's also lots of stuff that becomes possible now that we can easily
generate multiple separate packages with various codegen outputs:

- Overall: everything is granular.  We can test selections of things,
  rather than needing to have everything fall into place at once.
- Generally more organized results.
- We can more easily inspect the size of generated code.
- We can more easily inspect the size of the compiled result of gen!
  (Okay, not really.  I'm seeing a '.so' file that's 4MB is coming out
  from 200sloc of "String".  I don't think that's exactly informative.
  Some constant factor is thoroughly obscuring the data of interest.
  Nice idea in theory though, and maybe we'll get there later.)
- We can diff the generated type for variations in adjunct config!
  (Probably not something that will end up tested, but neat to be able
  to do during dev.)

Doing this with go plugins seemed like the neatest way to do this.
It's certainly not the only way, though.  And in particular, I will
confess that this will probably make developing from a windows host
pretty painful: go plugins aren't supported on windows.  Mind,
this doesn't mean you can't *use* codegen or its results on windows.
It just means the tests won't work.  So, someone doing development
_on the codegen itself_ would have to wait for the CI server to run
the tests on their behalf.  Hopefully this is survivable.

(There's also a fun wee wiggle in that loading a plugin has the
requirement that it be built with the same "runtime".  The definition
of "runtime" here happens to include whether or not things have been
built in "race" mode.  So, the '-race' flag disappears from our CI
config file in this diff; otherwise, CI will get the (rather confusing)
error "plugin was built with a different version of package runtime".
This isn't really worrying to ditch, though.  I'm... not even sure why
the '-race' was in our CI script in the first place.  Must've arrived
via cargo cult; we don't _have_ any concurrency in this library.)

An alternative way to go about all this would be to have the tests for
gen invoke `go test` (rather than `go build` in plugin mode) on each of
the generated packages.  It strikes me as similar but worse.
We still have to invoke the go tools from inside the test;
we'd have *more* work to do to either copy tests into the gen'd package
or else generate calls back to the parent package for test functions
(which still have to be written against interfaces, so that they can
compile even when the gen isn't done, as it indeed isn't when you
freshly check out the repo -- exact same as with the plugin approach);
and in exchange for the extra work, we get markedly worse output
('go test' doesn't nest nicely, afaict), and we can't separate the
compiling of the generated code from the evaluation of tests on it,
and we'd have no possibility of passing information via closures should
we wish to develop table-driven tests where this would be useful.
tl;dr: Highest cost, uglier, and worse.

No matter which way we go about this, there *is* a speed trade-off.
Invoking the compiler per package adds at least a half second of time
for each package, in practice.  Worth it, though.

And on the off chance that this plugin approach does burn later,
and we do want to switch to child process 'go test' invocations...
the good news is: we shouldn't actually have to rewrite very much.
The everything-starts-from-NodeStyle-and-tests-against-Node work is
exactly the same for making the plugin approach work, and will
trivially pivot to working fine in for child 'go test' approaches,
should we find it necessary to do so in the future.  So!  With our
butts covered: a plugin'ing we shall go!

Some of the code here still needs cleanup; this is a proof of concept
checkpointing commit.  (The real thing probably won't have such
function names as "TestFancier".)  But, we do get to see here:
plugins work; more than one in the process works; and they work even
when the same type names are in the generated packages.  All good.

Some touches to ImplicitValue came along for the ride, but aren't
exercised yet.

Tests are getting unwieldy.  I think something must be done about
this before proceding much further or it's going to start resulting
in increasingly significant velocity loss.

fill in generated type for list

Fill in generated types for int, link, and bytes nodes

Signed-off-by: Eric Myhre <hash@exultant.us>

I'm solving the decorum problem by making "temporary" builder methods for the
schema.Type values as necessary.  This will allow the gengo package to use
them freely for now.  When we're done with all this, I want to restrict building
of schema.Type values to a single method in the schema package which takes the
"ast" types -- the ones we're hoping to codegen (!) -- and do lots of checks
and cross-linking; the temporary builder methods will do *none*, and *of course*
they can't take the "ast" types yet since they don't exist yet.

Boom, cycle broken.

Also: fix generateStringKind from taking a TypeName redundantly.