Right now, we reject dashes in package names. I've hit this a few times and it annoys
me each time. (It would seem to makes sense to permit hyphens in package names, given
that [almost?] every other package manager on Earth does...)
No more! Hyphens welcome!
This change introduces decorator support for CocoJS and the corresponding
IL/AST changes to store them on definition nodes. Nothing consumes these
at the moment, however, I am looking at leveraging this to indicate that
certain program fragments are "code" and should be serialized specially
(in support of Functions-as-lambdas).
This change includes a first basic whack at implementing S3 bucket
objects. It leverages the assets infrastructure put in place in the
last commit, supporting uploads from text, files, or arbitrary URIs.
Most of the interesting object properties remain unsupported for now,
but with this we can upload and delete basic S3 objects, sufficient
for a lot of the lambda functions management we need to implement.
This change introduces the basic concept of assets. It is far from
fully featured, however, it is enough to start adding support for various
storage kinds that require access to I/O-backed data (files, etc).
The challenge is that Coconut is deterministic by design, and so you
cannot simply read a file in an ad-hoc manner and present the bytes to
a resource provider. Instead, we will model "assets" as first class
entities whose data source is described to the system in a more declarative
manner, so that the system and resource providers can manage them.
There are three ways to create an asset at the moment:
1. A constant, in-memory string.
2. A path to a file on the local filesystem.
3. A URI, whose scheme is extensible.
Eventually, we want to support byte blobs, but due to our use of a
"JSON-like" type system, this isn't easily expressible just yet.
The URI scheme is extensible in that file://, http://, and https://
are supported "out of the box", but individual providers are free to
recognize their own schemes and support them. For instance, copying
one S3 object to another will be supported simply by passing a URI
with the s3:// protocol in the usual way.
Many utility functions are yet to be written, but this is a start.
This change implements simple index-based CocoPy subscripts (and
not the more fully featured slicing ones).
Alongside this, we relax a binder-time check that all dynamic
access types must be strings. The eval code already handles
numeric (array) accesses, so we will permit these to flow through.
This permits non-nil `this` objects for dynamically loaded properties
of classes and modules, provided they are the prototype or module
object for the target function, respectively.
This change emits all CocoJS interfaces as records. This allows us to
safely construct instances of them from Python using anonymous properties,
essentially emulating object literals.
For example, given a CocoJs interface defined as such:
interface Foo {
x;
y?;
z;
}
we can easily construct fresh instances using normal JS literals:
let f = { x: 42, z: "bar" };
But, Python doesn't have the equivalent literal syntax, wedging us.
It's now possible to initialize an instance from CocoPy as follows:
let f = Foo(x=42, z="bar")
This leverages the notion of records and primary properties, as
described in our CocoPack/CocoIL design documents.
* Pushing a class scope should permit subclasses (assertion).
* Returning nothing is legal for voids *and* dynamic (the latter was missing).
* The dynamic load readonly lval check is redundant; we check at the site of
assignment and doing it here as well led to two errors per usage.
Due to Python's ... interesting ... scoping rules, we want to avoid
forcing block scopes in certain places. Instead, we will let arbitrary
statements take their place. Of course, this could be a block, but it
very well could be a multi-statement (essentially a block that doesn't
imply a new lexical scope), or anything, really.
* Move checking the validity of a derived class with no constructor
from evlauation to binding (verification). Furthermore, make it
a verification error instead of an assert, and make the checking
complete (i.e., don't just check for the existence of a base class,
also check for the existence of a ctor, recursively).
* Refactor the constructor accessor out of the evaluator and make it
a method, Ctor, on the Function abstraction.
* Add a recursive Module population case to property initialization.
* Only treat the top-most frame in a module's initializer as belonging
to the module scope. This avoids interpreting block scopes within
that initializer as belonging to the module scope. Ultimately, it's
up to the language compiler to decide where to place scopes, but this
gives it more precise control over module scoping.
* Assert against unsupported named/star/keyword args in CocoPy.
If we encounter a dynamic invocation of a prototype object, we will
interpret it as an object allocation. This corresponds to code like
import ec2 from aws
instance = ec2.Instance(...)
where the second line dynamically loads the prototype object for the
Instance class from the module object for the aws/ec2 module, and
invokes it.
This change adds support for naming imports. At the moment, this simply
makes the names dynamically accessible for languages that do dynamic loads
against module objects, versus strongly typed tokens. The basic scheme
is to keep two objects per module: one that contains the globals and its
prototype parent that contains just the exports. This ensures we can
share the same slots while attaining the desired information hiding
(e.g., when handing out an object for dynamic access, we give out this
parent objects, while when loading globals from within a module, we use
the childmost one that contains all private and exported variables).
The old model for imports was to use top-level declarations on the
enclosing module itself. This was a laudible attempt to simplify
matters, but just doesn't work.
For one, the order of initialization doesn't precisely correspond
to the imports as they appear in the source code. This could incur
some weird module initialization problems that lead to differing
behavior between a language and its Coconut variant.
But more pressing as we work on CocoPy support, it doesn't give
us an opportunity to dynamically bind names in a correct way. For
example, "import aws" now needs to actually translate into a variable
declaration and assignment of sorts. Furthermore, that variable name
should be visible in the environment block in which it occurs.
This change switches imports to act like statements. For the most
part this doesn't change much compared to the old model. The common
pattern of declaring imports at the top of a file will translate to
the imports happening at the top of the module's initializer. This
has the effect of initializing the transitive closure just as it
happened previously. But it enables alternative models, like imports
inside of functions, and -- per the above -- dynamic name binding.
Previously, it was an error to look up a local that didn't exist.
Now it is common, thanks to dynamic lookups. A few code-paths didn't
previously handle this adequately; now they do.
This rearranges the way dynamic loads work a bit. Previously, they¬
required an object, and did a dynamic lookup in the object's property¬
map. For real dynamic loads -- of the kind Python uses, obviously,¬
but also ECMAScript -- we need to search the "environment".
This change searches the environment by looking first in the lexical¬
scope in the current function. If a variable exists, we will use it.¬
If that misses, we then look in the module scope. If a variable exists¬
there, we will use it. Otherwise, if the variable is used in a non-lval
position, an dynamic error will be raised ("name not declared"). If
an lval, however, we will lazily allocate a slot for it.
Note that Python doesn't use block scoping in the same way that most
languages do. This behavior is simply achieved by Python not emitting
any lexically scoped blocks other than at the function level.
This doesn't perfectly achieve the scoping behavior, because we don't
yet bind every name in a way that they can be dynamically discovered.
The two obvious cases are class names and import names. Those will be
covered in a subsequent commit.
Also note that we are getting lucky here that class static/instance
variables aren't accessible in Python or ECMAScript "ambiently" like
they are in some languages (e.g., C#, Java); as a result, we don't need
to introduce a class scope in the dynamic lookup. Some day, when we
want to support such languages, we'll need to think about how to let
languages control the environment probe order; for instance, perhaps
the LoadDynamicExpression node can have an "environment" property.
This changes the CocoPy default load type from static to dynamic,
since we don't have enough information at compile-time to emit
fully qualified tokens. Previously, Coconut only supported dynamic
loads with object targets, however we will need to support the full
scope search (class, module, global, etc).
This change makes node optional in the lookupBasicType function, which is
necessary in cases where diagnostics information isn't available (such as
with configuration application). This eliminates an assert when you fat-
finger a configuration key. The associated message was missing apostrophes.
This change eliminates the need to constantly type in the environment
name when performing major commands like configuration, planning, and
deployment. It's probably due to my age, however, I keep fat-fingering
simple commands in front of investors and I am embarrassed!
In the new model, there is a notion of a "current environment", and
I have modeled it kinda sorta just like Git's notion of "current branch."
By default, the current environment is set when you `init` something.
Otherwise, there is the `coco env select <env>` command to change it.
(Running this command w/out a new <env> will show you the current one.)
The major commands `config`, `plan`, `deploy`, and `destroy` will prefer
to use the current environment, unless it is overridden by using the
--env flag. All of the `coco env <cmd> <env>` commands still require the
explicit passing of an environment which seems reasonable since they are,
after all, about manipulating environments.
As part of this, I've overhauled the aging workspace settings cruft,
which had fallen into disrepair since the initial prototype.
This changes the object mapper infrastructure to offer more fine-grained
reporting of errors, and control over verification, during the mapping from
an untyped payload to a typed one. As a result, we can eliminate a bit of
the explicit unmarshaling goo in the AWS providers (but not all of it; I'm
sure there is more we can, and should, be doing here...)
This change adds the ability to specify analyzers in two ways:
1) By listing them in the project file, for example:
analyzers:
- acmecorp/security
- acmecorp/gitflow
2) By explicitly listing them on the CLI, as a "one off":
$ coco deploy <env> \
--analyzer=acmecorp/security \
--analyzer=acmecorp/gitflow
This closes out pulumi/coconut#119.
This change introduces the basic requirements for analyzers, as per
pulumi/coconut#119. In particular, an analyzer can implement either,
or both, of the RPC methods, Analyze and AnalyzeResource. The former
is meant to check an overall deployment (e.g., to ensure it has been
signed off on) and the latter is to check individual resources (e.g.,
to ensure properties of them are correct, such as checking style,
security, etc. rules). These run simultaneous to overall checking.
Analyzers are loaded as plugins just like providers are. The difference
is mainly in their naming ("analyzer-" prefix, rather than "resource-"),
and the RPC methods that they support.
This isn't 100% functional since we need a way to specify at the CLI
that a particular analyzer should be run, in addition to a way of
recording which analyzers certain projects should use in their manifests.
This changes a few naming things:
* Rename "husk" to "environment" (`coco env` for short).
* Rename NutPack/NutIL to CocoPack/CocoIL.
* Rename the primary Nut.yaml/json project file to Coconut.yaml/json.
* Rename the compiled Nutpack.yaml/json file to Cocopack.yaml/json.
* Rename the package asset directory from nutpack/ to .coconut/.
This change adds a new Check RPC method on the provider interface,
permitting resource providers to perform arbitrary verification on
the values of properties. This is useful for validating things
that might be difficult to express in the type system, and it runs
before *any* modifications are run (so failures can be caight early
before it's too late). My favorite motivating example is verifying
that an AWS EC2 instance's AMI is available within the target region.
This resolvespulumi/coconut#107, although we aren't using this
in any resource providers just yet. I'll add a work item now for that...
This change is mostly just a rename of Moniker to URN. It does also
prefix resource URNs to have a standard URN namespace; in other words,
"urn🥥<name>", where <name> is the same as the prior Moniker.
This is a minor step that helps to prepare us for pulumi/coconut#109.
This change adds a new resource.NewUniqueHex API, that simply generates
a unique hex string with the given prefix, with a specific count of
random bytes, and optionally capped to a maximum length.
This is used in the AWS SecurityGroup resource provider to avoid name
collisions, which is especially important during replacements (otherwise
we cannot possibly create a new instance before deleting the old one).
This resolvespulumi/coconut#108.
This change tracks which updates triggered a replacement. This enables
better output and diagnostics. For example, we now colorize those
properties differently in the output. This makes it easier to diagnose
why an unexpected resource might be getting deleted and recreated.
This change, part of pulumi/coconut#105, rearranges support for
resource replacement. The old model didn't properly account for
the cascading updates and possible replacement of dependencies.
Namely, we need to model a replacement as a creation followed by
a deletion, inserted into the overall DAG correctly so that any
resources that must be updated are updated after the creation but
prior to the deletion. This is done by inserting *three* nodes
into the graph per replacement: a physical creation step, a
physical deletion step, and a logical replacement step. The logical
step simply makes it nicer in the output (the plan output shows
a single "replacement" rather than the fine-grained outputs, unless
they are requested with --show-replace-steps). It also makes it
easier to fold all of the edges into a single linchpin node.
As part of this, the update step no longer gets to choose whether
to recreate the resource. Instead, the engine takes care of
orchestrating the replacement through actual create and delete calls.
This change introduces a new RPC function to the provider interface;
in pseudo-code:
UpdateImpact(id ID, t Type, olds PropertyMap, news PropertyMap)
(bool, PropertyMap, error)
Essentially, during the planning phase, we will consult each provider
about the nature of a proposed update. This update includes a set of
old properties and the new ones and, if the resource provider will need
to replace the property as a result of the update, it will return true;
in general, the PropertyMap will eventually contain a list of all
properties that will be modified as a result of the operation (see below).
The planning phase reacts to this by propagating the change to dependent
resources, so that they know that the ID will change (and so that they
can recalculate their own state accordingly, possibly leading to a ripple
effect). This ensures the overall DAG / schedule is ordered correctly.
This change is most of pulumi/coconut#105. The only missing piece
is to generalize replacing the "ID" property with replacing arbitrary
properties; there are hooks in here for this, but until pulumi/coconut#90
is addressed, it doesn't make sense to make much progress on this.
For cerain update shapes, we will need to recover an ID of an already-deleted,
or soon-to-be-deleted resource; in those cases, we have a moniker but want to
serialize an ID. This change implements support for remembering/recovering them.
This change fixes a couple issues that prevented restarting a
deployment after partial failure; this was due to the fact that
unchanged resources didn't propagate IDs from old to new. This
is remedied by making unchanged a map from new to old, and making
ID propagation the first thing plan application does.
This change does a few things:
* First and foremost, it tracks configuration variables that are
initialized, and optionally prints them out as part of the
prelude/header (based on --show-config), both in a dry-run (plan)
and in an actual deployment (apply).
* It tidies up some of the colorization and messages, and includes
nice banners like "Deploying changes:", etc.
* Fix an assertion.
* Issue a new error
"One or more errors occurred while applying X's configuration"
just to make it easier to distinguish configuration-specific
failures from ordinary ones.
* Change config keys to tokens.Token, not tokens.ModuleMember,
since it is legal for keys to represent class members (statics).
This change adds support for configuration maps.
This is a new feature that permits initialization code to come from markup,
after compilation, but before evaluation. There is nothing special with this
code as it could have been authored by a user. But it offers a convenient
way to specialize configuration settings per target husk, without needing
to write code to specialize each of those husks (which is needlessly complex).
For example, let's say we want to have two husks, one in AWS's us-west-1
region, and the other in us-east-2. From the same source package, we can
just create two husks, let's say "prod-west" and "prod-east":
prod-west.json:
{
"husk": "prod-west",
"config": {
"aws:config:region": "us-west-1"
}
}
prod-east.json:
{
"husk": "prod-east",
"config": {
"aws:config:region": "us-east-2"
}
}
Now when we evaluate these packages, they will automatically poke the
right configuration variables in the AWS package *before* actually
evaluating the CocoJS package contents. As a result, the static variable
"region" in the "aws:config" package will have the desired value.
This is obviously fairly general purpose, but will allow us to experiment
with different schemes and patterns. Also, I need to whip up support
for secrets, but that is a task for another day (perhaps tomorrow).
* Delete husks if err == nil, not err != nil.
* Swizzle the formatting padding on array elements so that the
diff modifier + or - binds more tightly to the [N] part.
* Print the un-doubly-indented padding for array element headers.
* Add some additional logging to step application (it helped).
* Remember unchanged resources even when glogging is off.
This change adds a --show-sames flag to `coco husk deploy`. This is
useful as I'm working on updates, to show what resources haven't changed
during a deployment.
This change checkpoints deployments properly. That is, even in the
face of partial failure, we should keep the huskfile up to date. This
accomplishes that by tracking the state during plan application.
There are still ways in which this can go wrong, however. Please see
pulumi/coconut#101 for additional thoughts on what we might do here
in the future to make checkpointing more robust in the face of failure.
This command is handy for development, so I whipped up a quick implementation.
All it does is print all known husks with their associated deployment time
and resource count (if any, or "n/a" for initialized husks with no deployments).
This change recognizes TextUnmarshaler during object mapping, and
will defer to it when we have a string but are assigning to a
non-string target that implements the interface.
As part of pulumi/coconut#94 -- adding targeting capabilities -- I've
decided to (yet again) reorganize the deployment commands a bit. This
makes targets ("husks") more of a first class thing.
Namely, you must first initialize a husk before using it:
$ coco husk init staging
Coconut husk 'staging' initialized; ready for deployments
Eventually, this is when you will be given a choice to configure it.
Afterwards, you can perform deployments. The first one is like a create,
but subsequent ones just figure out the right thing to do and do it:
$ ... make some changes ...
$ coco husk deploy staging
... standard deployment progress spew ...
Finally, should you want to teardown an entire environment:
$ coco husk destroy staging
... standard deletion progress spew for all resources ...
Coconut husk 'staging' has been destroyed!
This change partially implements pulumi/coconut#94, by adding the
ability to name targets during creation and reuse those names during
deletion and update. This simplifies the management of deployment
records, checkpoints, and snapshots.
I've opted to call these things "husks" (perhaps going overboard with
joy after our recent renaming). The basic idea is that for any
executable Nut that will be deployed, you have a nutpack/ directory
whose layout looks roughly as follows:
nutpack/
bin/
Nutpack.json
... any other compiled artifacts ...
husks/
... one snapshot per husk ...
For example, if we had a stage and prod husk, we would have:
nutpack/
bin/...
husks/
prod.json
stage.json
In the prod.json and stage.json files, we'd have the most recent
deployment record for that environment. These would presumably get
checked in and versioned along with the overall Nut, so that we
can use Git history for rollbacks, etc.
The create, update, and delete commands look in the right place for
these files automatically, so you don't need to manually supply them.
* Eliminate some superfluous "\n"s.
* Remove the redundant properties stored on AWS resources.
* Compute array diff lengths properly (+1).
* Display object property changes from null to non-null as
adds; and from non-null to null as deletes.
* Fix a boolean expression from ||s to &&s. (Bone-headed).
* Print a pretty message if the plan has nothing to do:
"info: nothing to do -- resources are up to date"
* Add an extra validation step after reading in a snapshot,
so that we detect more errors sooner. For example, I've
fed in the wrong file several times, and it just chugs
along as though it were actually a snapshot.
* Skip printing nulls in most plan outputs. These just
clutter up the output.
This change detects duplicate object names (monikers) and issues a nice
error message with source context include. For example:
index.ts(260,22): error MU2006: Duplicate objects with the same name:
prod::ec2instance:index::aws:ec2/securityGroup:SecurityGroup::group
The prior code asserted and failed abruptly, whereas this actually points
us to the offending line of code:
let group1 = new aws.ec2.SecurityGroup("group", { ... });
let group2 = new aws.ec2.SecurityGroup("group", { ... });
^^^^^^^^^^^^^^^^^^^^^^^^^
This change overhauls the way we do object monikers. The old mechanism,
generating monikers using graph paths, was far too brittle and prone to
collisions. The new approach mixes some amount of "automatic scoping"
plus some "explicit naming." Although there is some explicitness, this
is arguably a good thing, as the monikers will be relatable back to the
source more readily by developers inspecting the graph and resource state.
Each moniker has four parts:
<Namespace>::<AllocModule>::<Type>::<Name>
wherein each element is the following:
<Namespace> The namespace being deployed into
<AllocModule> The module in which the object was allocated
<Type> The type of the resource
<Name> The assigned name of the resource
The <Namespace> is essentially the deployment target -- so "prod",
"stage", etc -- although it is more general purpose to allow for future
namespacing within a target (e.g., "prod/customer1", etc); for now
this is rudimentary, however, see marapongo/mu#94.
The <AllocModule> is the token for the code that contained the 'new'
that led to this object being created. In the future, we may wish to
extend this to also track the module under evaluation. (This is a nice
aspect of monikers; they can become arbitrarily complex, so long as
they are precise, and not prone to false positives/negatives.)
The <Name> warrants more discussion. The resource provider is consulted
via a new gRPC method, Name, that fetches the name. How the provider
does this is entirely up to it. For some resource types, the resource
may have properties that developers must set (e.g., `new Bucket("foo")`);
for other providers, perhaps the resource intrinsically has a property
that explicitly and uniquely qualifies the object (e.g., AWS SecurityGroups,
via `new SecurityGroup({groupName: "my-sg"}`); and finally, it's conceivable
that a provider might auto-generate the name (e.g., such as an AWS Lambda
whose name could simply be a hash of the source code contents).
This should overall produce better results with respect to moniker
collisions, ability to match resources, and the usability of the system.
This change is a first whack at implementing updates.
Creation and deletion plans are pretty straightforward; we just take
a single graph, topologically sort it, and perform the operations in
the right order. For creation, this is in dependency order (things
that are depended upon must be created before dependents); for deletion,
this is in reverse-dependency order (things that depend on others must
be deleted before dependencies). These are just special cases of the more
general idea of performing DAG operations in dependency order.
Updates must work in terms of this more general notion. For example:
* It is an error to delete a resource while another refers to it; thus,
resources are deleted after deleting dependents, or after updating
dependent properties that reference the resource to new values.
* It is an error to depend on a create a resource before it is created;
thus, resources must be created before dependents are created, and/or
before updates to existing resource properties that would cause them
to refer to the new resource.
Of course, all of this is tangled up in a graph of dependencies. As a
result, we must create a DAG of the dependencies between creates, updates,
and deletes, and then topologically sort this DAG, in order to determine
the proper order of update operations.
To do this, we slightly generalize the existing graph infrastructure,
while also specializing two kinds of graphs; the existing one becomes a
heapstate.ObjectGraph, while this new one is resource.planGraph (internal).
This change introduces a simple retry package which will be used in
resource providers in order to perform waits, backoffs, etc. It's
pretty basic right now but leverages the fancy new Golang context
support for deadlines and timeouts. Soon we will layer AWS-specific
acceptors, etc. atop this more general purpose thing.
This change introduces a new informational message category to the
overall diagnostics infrastructure, and then wires up the resource
provider plugins stdout/stderr streams to it. In particular, a
write to stdout implies an informational message, whereas a write to
stderr implies an error. This is just a very simple and convenient
way for plugins to provide progress reporting; eventually we may
need something more complex, due to parallel evaluation of resource
graphs, however I hope we don't have to deviate too much from this.
This change adds custom serialization and deserialization to preserve
the ordering of resources in snapshot files. Unfortunately, because
Go maps are unordered, the results are scrambled (actually, it turns out,
Go will sort by the key). An alternative would be to resort the results
after reading in the file, or storing as an array, but this would be a
change to the MuGL file format and is less clear than simply keying each
resource by its moniker as we are doing today.
This change repivots the plan/apply commands slightly. This is largely
in preparation for performing deletes and updates of existing environments.
The old way was slightly confusing and made things appear more "magical"
than they actually are. Namely, different things are needed for different
kinds of deployment operations, and trying to present them each underneath
a single pair of CLI commands just leads to weird modality and options.
The new way is to offer three commands: create, update, and delete. Each
does what it says on the tin: create provisions a new environment, update
makes resource updates to an existing one, and delete tears down an existing
one entirely. The arguments are what make this interesting: create demands
a MuPackage to evaluate (producing the new desired state snapshot), update
takes *both* an existing snapshot file plus a MuPackage to evaluate (producing
the new desired state snapshot to diff against the existing one), and delete
merely takes an existing snapshot file and no MuPackage, since all it must
do is tear down an existing known environment.
Replacing the plan functionality is the --dry-run (-n) flag that may be
passed to any of the above commands. This will print out the plan without
actually performing any opterations.
All commands produce serializable resource files in the MuGL file format,
and attempt to do smart things with respect to backups, etc., to support the
intended "Git-oriented" workflow of the pure CLI dev experience.
Per Eric's suggestion, this moves the colors we use into a central
location so that it'll be easier someday down the road to reconfigure
and/or disable them, etc. This does not include a --no-colors option
although we should really include this soon before it gets too hairy.
This change adds support for serializing snapshots in MuGL, per the
design document in docs/design/mugl.md. At the moment, it is only
exposed from the `mu plan` command, which allows you to specify an
output location using `mu plan --output=file.json` (or `-o=file.json`
for short). This serializes the snapshot with monikers, resources,
and so on. Deserialization is not yet supported; that comes next.
* Specify MinCount/MaxCount when creating an EC2 instance. These
are required properties on the RunInstances API.
* Only attempt to unmarshal egressArray/ingressArray when non-nil.
* Remember the context object on the instanceProvider.
* Move the moniker and object maps into the shared context object.
* Marshal object monikers as the resource IDs to which they refer,
since monikers are useless on "the other side" of the RPC boundary.
This ensures that, for example, the AWS provider gets IDs it can use.
* Add some paranoia assertions.
This change does two things:
First, and foremost, it adds interface members to prototypes. This
ensures that interface members are available statically on all objects
of types that implement those interfaces.
Second, we permit dynamic loads through LoadLocationExpression when
the `this` object is `dynamic`. This is a convenient shortcut compared
to creating different LoadDynamicExpressions when the `this` happens to
be of a dynamic type.
This commit includes a basic AWS resource provider. Mostly it is just
scaffolding, however, it also includes prototype implementations for EC2
instance and security group resource creation operations.
This change implements `mu apply`, by driving compilation, evaluation,
planning, and then walking the plan and evaluating it. This is the bulk
of marapongo/mu#21, except that there's a ton of testing/hardening to
perform, in addition to things like progress reporting.
This change adds basic support for discovering, loading, binding to,
and invoking RPC methods on, resource provider plugins.
In a nutshell, we add a new context object that will share cached
state such as loaded plugins and connections to them. It will be
a policy decision in server scenarios how much state to share and
between whom. This context also controls per-resource context
allocation, which in the future will allow us to perform structured
cancellation and teardown amongst entire groups of requests.
Plugins are loaded based on their name, and can be found in one of
two ways: either simply by having them on your path (with a name of
"mu-ressrv-<pkg>", where "<pkg>" is the resource package name with
any "/"s replaced with "_"s); or by placing them in the standard
library installation location, which need not be on the path for this
to work (since we know precisely where to look).
If we find a protocol, we will load it as a child process.
The protocol for plugins is that they will choose a port on their
own -- to eliminate races that'd be involved should Mu attempt to
pre-pick one for them -- and then write that out as the first line
to STDOUT (terminated by a "\n"). This is the only STDERR/STDOUT
that Mu cares about; from there, the plugin is free to write all it
pleases (e.g., for logging, debugging purposes, etc).
Afterwards, we then bind our gRPC connection to that port, and create
a typed resource provider client. The CRUD operations that get driven
by plan application are then simple wrappers atop the underlying gRPC
calls. For now, we interpret all errors as catastrophic; in the near
future, we will probably want to introduce a "structured error"
mechanism in the gRPC interface for "transactional errors"; that is,
errors for which the server was able to recover to a safe checkpoint,
which can be interpreted as ResourceOK rather than ResourceUnknown.
This moves us one step closer from planning to application (marapongo/mu#21).
Namely, we now drive the right resource provider operations in response to
the plan's steps. Those providers, however, are still empty shells.
This change adds a flag to `plan` so that we can create deletion plans:
$ mu plan --delete
This will have an equivalent in the `apply` command, achieving the ability
to delete entire sets of resources altogether (see marapongo/mu#58).
This change introduces object monikers. These are unique, serializable
names that refer to resources created during the execution of a MuIL
program. They are pretty darned ugly at the moment, but at least they
serve their desired purpose. I suspect we will eventually want to use
more information (like edge "labels" (variable names and what not)),
but this should suffice for the time being. The names right now are
particularly sensitive to simple refactorings.
This is enough for marapongo/mu#69 during the current sprint, although
I will keep the work item (in a later sprint) to think more about how
to make these more stable. I'd prefer to do that with a bit of
experience under our belts first.
This change introduces a new package, pkg/resource, that will form
the foundation for actually performing deployment plans and applications.
It contains the following key abstractions:
* resource.Provider is a wrapper around the CRUD operations exposed by
underlying resource plugins. It will eventually defer to resource.Plugin,
which itself defers -- over an RPC interface -- to the actual plugin, one
per package exposing resources. The provider will also understand how to
load, cache, and overall manage the lifetime of each plugin.
* resource.Resource is the actual resource object. This is created from
the overall evaluation object graph, but is simplified. It contains only
serializable properties, for example. Inter-resource references are
translated into serializable monikers as part of creating the resource.
* resource.Moniker is a serializable string that uniquely identifies
a resource in the Mu system. This is in contrast to resource IDs, which
are generated by resource providers and generally opaque to the Mu
system. See marapongo/mu#69 for more information about monikers and some
of their challenges (namely, designing a stable algorithm).
* resource.Snapshot is a "snapshot" taken from a graph of resources. This
is a transitive closure of state representing one possible configuration
of a given environment. This is what plans are created from. Eventually,
two snapshots will be diffable, in order to perform incremental updates.
One way of thinking about this is that a snapshot of the old world's state
is advanced, one step at a time, until it reaches a desired snapshot of
the new world's state.
* resource.Plan is a plan for carrying out desired CRUD operations on a target
environment. Each plan consists of zero-to-many Steps, each of which has
a CRUD operation type, a resource target, and a next step. This is an
enumerator because it is possible the plan will evolve -- and introduce new
steps -- as it is carried out (hence, the Next() method). At the moment, this
is linearized; eventually, we want to make this more "graph-like" so that we
can exploit available parallelism within the dependencies.
There are tons of TODOs remaining. However, the `mu plan` command is functioning
with these new changes -- including colorization FTW -- so I'm landing it now.
This is part of marapongo/mu#38 and marapongo/mu#41.
This change fixes two aspects of binding names to objects. First, we need
to adjust pointers from prototype functions, since they will have the wrong
"this" (it points to the prototype instance and not the actual object).
Second, we need to bind `super` accesses using the correct prototype object.
* Improve some logging and assertions.
* Initialize the prototype map (it was always empty before).
* Actually memoize the prototype objects in this map.
* During class initialization, ensure the base class is also initialized.
* Trigger class initialization anytime a property or function is loaded
from that class. Similarly, trigger module initialization on loads too.
* Treat For/ForIn/ForOfStatements as legal parents for local variables.
This ensures for loops at the top-level in a module have their variables
treated as locals, rather than module properties. Add a test also.
* Track source locations for more nodes.
* Mark auto-generated constructors as public so that they may be called.
* Fix the naming of class init methods; it should be ".init", not ".ctor".
* Clone the loop test cases into both function and module top-level variants.
One guiding principle for what makes it into the MuIL AST is that
the gap between source language and AST should not be too great; the
projection of switch statements from MuJS into MuIL clearly violated
that principle, particularly considering that the logic wasn't even
right due to the incorrect emulation of conditional breaks.
Instead of digging deeper into the hole, I've encoded switch logic
in the AST, and implemented support in the evaluator.
* Add a TODO as a reminder to implement number toString formatting.
* Change the Loreley delimiters to something obscure ("<{%" and "%}>")
to avoid conflicting with actual characters we might use in messages.
Also, make the assertions more descriptive should Loreley fail.
* Rip out a debug.PrintStack() in verbose logging.
* Check the underlying pointer's object type for +=, not the pointer type.
This change implements intrinsic function support in the runtime.
The basic idea is to swap out references to MuIL functions with
runtime-implemented functionality, for operations that cannot be
expressed in the MuIL-subset. For example, this change includes
two such cases: 1) a mu.runtime.isFunction(obj) function to check if
the target object is a function; and 2) a
mu.runtime.dynamicInvoke(obj,obj,obj[]) function to dynamically
invoke a target object.
These will be used most immediately to implement ECMAScript toString
functionality in the MuJS runtime library, however it will also come
in handy shortly to implement things like printf (marapongo/mu#86),
string and array functions (marapongo/mu#80), and, eventually, the
ECMAScript-specific operators and behavior (marapongo/mu#61).
This change generalizes the support for default modules (added as
part of marapongo/mu#57), to use an "alias" table. This is just a
table of alternative module names to the real defining module name.
The default module is now just represented as a mapping from the
special default module token, ".default", to the actual defining
module. This generalization was necessary to support Node.js-style
alternative names for modules, like "lib" mapping to "lib/index".
Previously, I had thought we would ask MetaMu compilers to map their
own exception/error types to ours. That causes some complexity, however,
in that the exception types in each language (if they even exist) are
"different". So we would need to wrap things, etc., which seems cumbersome.
Partly I had done this after having been burned on the CLR by permitting
throws of any types; e.g., C++/CLI could throw integers, which would rip
through C# code unknowingly, because all C# throws had to derive from the
Exception base class. This actually caused some security issues!
But, we're in a different place, and a different time. The first three
languages I envision supporting -- ECMAScript, Python, and Ruby -- each
permit you to throw anything. And the only other languages I am seriously
contemplating right now -- Go -- doesn't even have exceptions.
So this change backs off and is less opinionated about what is thrown.
Instead, we track stack traces, etc., in the unwind information, and happily
propagate whatever object is thrown.
This change more accurately implements ECMAScript prototype chains.
This includes using the prototype chain to lookup properties when
necessary, and copying them down upon writes.
This still isn't 100% faithful -- for example, classes and
constructor functions should be represented as real objects with
the prototype link -- so that examples like those found here will
work: https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Objects/Object/constructor.
I've updated marapongo/mu#70 with additional details about this.
I'm sure we'll be forced to fix this as we encounter more "dynamic"
JavaScript. (In fact, it would be interesting to start running the
pre-ES6 output of TypeScript through the compiler as test cases.)
See http://www.ecma-international.org/ecma-262/6.0/#sec-objects
for additional details on the prototype chaining semantics.
The complexity of the last round of property initialization code was
starting to bug me. I've switched away from lazily initializing them
and now eagerly initialize at the proper times: module initialization,
class initialization (for statics), and object allocation (for instances).
This includes the logic around readonly and freezing. The code is a
lot simpler now and actually enforces some invariants that we'd like to
enforce, like not silently adding new properties when a supposedly static
member load is happening (all such accesses should be dynamic ones).
This change accurately enforces readonly properties. Namely, they
should not be written to anywhere outside of the respective initializer,
but writes must be allowed within them. This means the module initializer
for module properties, the class initializer for statics, and the object
constructor for instance properties.
Now that some name lookups are context-sensitive (namely, whether
or not we should use the export or member table for inter vs. intra
module token references), we need to faithfully track the context.
This change fixes up some issues in the big export refactoring;
namely, we need to chase down references one link at a time during
binding, before they settle. This is because an export might depend
on another export, which might depend on ...
This change redoes the way module exports are represented. The old
mechanism -- although laudible for its attempt at consistency -- was
wrong. For example, consider this case:
let v = 42;
export { v };
The old code would silently add *two* members, both with the name "v",
one of which would be dropped since the entries in the map collided.
It would be easy enough just to detect collisions, and update the
above to mark "v" as public, when the export was encountered. That
doesn't work either, as the following two examples demonstrate:
let v = 42;
export { v as w };
let x = w; // error!
This demonstrates:
* Exporting "v" with a different name, "w" to consumers of the
module. In particular, it should not be possible for module
consumers to access the member through the name "v".
* An inability to access the exported name "w" from within the
module itself. This is solely for external consumption.
Because of this, we will use an export table approach. The exports
live alongside the members, and we are smart about when to consult
the export table, versus the member table, during name binding.
This commit implements array l-values (for dynamic loads only, since we do
not yet ever produce static ones). Note that there are some ECMAScript
compliance noteworthies about this change, which are captured in comments.
Namely, we are emulating ECMAScript's ability to index into an array
anywhere (except for negative numbers, which is a problem). This is in
contrast to the usual approach of throwing an error on out-of-bounds access,
which will crop up when we move on to other languages like Python. And yet we
are usuing a real array as the backing store, which can cause problems with
some real ECMAScript programs that use sparse arrays and expect the "bag of
properties" approach to keep memory usage reasonable.
The work item marapongo/mu#70 tracks this among other things.
This change plucks array element types out of more AST nodes, tightening
up the MuIL that MuJS produces. It also adds a few cases where we should
be emitting "object", and permits more dynamic conversions (which are
technically unsafe, but necessary for the sort of duck typing we anticipate).
In select few cases (right now, just nulls and boolean true/false),
we use predefined object constants to avoid allocating wastefully.
In the future, it's possible we will do this for other types, like
interned strings. So, the graph generator should tolerate this.
This change starts tracking all objects in our MuGL graph. The reason is
that resources can be buried deep within a nest of objects, and unless we
do "on the fly" reachability analysis, we can't know a priori whether any
given object will be of interest or not. So, we track 'em all. For large
programs, this would obviously create space leak problems, so we'll
eventually, I assume, want to prune the graph at some point.
With this change, the EC2instance example produces a (gigantic) graph!
This change implements stack traces. This is primarily so that we
can print out a full stack trace in the face of an unhandled exception,
and is done simply by recording the full trace during evaluation
alongside the existing local variable scopes.
In dynamic scenarios, property keys could be arbitrary strings, including
invalid identifiers. This change reflects that in the runtime representation
of objects and also in the object literal and overall indexing logic.
This change "kicks the can" of dealing with `dynamic` down the road.
Namely, the binder will be permissive about letting dynamic objects
flow through to the evaluation phase. At that point, the evaluator
will be responsible for performing duck typing, etc.
This change introduces support for dynamic object literals. If the
type of a literal is dynamic, then the property initializers for that
literal are naked strings; if the type of a literal is anything else,
however, we expect that the property names are full blown member tokens.
This change prepares to mimick the TypeScript behavior of `any` which,
it turns out, is closer to Python and Ruby duck typing than what we have
here. It also introduces `object` to mean "the base of all types".
MuJS has been updated to map `any` to `dynamic`.
At this point, the conversion logic now distinguishes between NoConversion,
ImplicitConversion, and AutoCastConversion cases, the latter being a case
in which `dynamic` is the source and so a duck-type-style cast must occur.
Right now we do not honor the AutoCastConversion case anywhere, so we still
have not achieved the duck-typing semantics of the source language.
This is part of fixing marapongo/mu#79.
After a bit more thinking, we will create new SDK packages for each
of the languages we wish to support writing resource providers in.
This is where the RPC goo will live, so I have created a new sdk/
directory, moved the Protobuf/gRPC definitions underneath sdk/proto/,
and put the generated code into sdk/go/ and sdk/js/.
This change moves the RPC definitions to the pkg/murpc package,
underneath the proto folder. This is to ensure that the resulting
Protobufs get a good package name "murpc" that is unique within the
overall toolchain. It also includes a `generate.sh` script that
can be used to manually regenerate client/server code. Finally,
we are actually checking in the generated files underneath pkg/murpc.
This change adds a --dot option to the eval command, which will simply
output the MuGL graph using the DOT language. This allows you to use
tools like Graphviz to inspect the resulting graph, including using the
`dot` command to generate images (like PNGs and whatnot).
For example, the simple MuGL program:
class C extends mu.Resource {...}
class B extends mu.Resource {...}
class A extends mu.Resource {
private b: B;
private c: C;
constructor() {
this.b = new B();
this.c = new C();
}
}
let a = new A();
Results in the following DOT file, from `mu eval --dot`:
strict digraph {
Resource0 [label="A"];
Resource0 -> {Resource1 Resource2}
Resource1 [label="B"];
Resource2 [label="C"];
}
Eventually the auto-generated ResourceN identifiers will go away in
favor of using true object monikers (marapongo/mu#76).
Glog doesn't actually print out the stack traces for all goroutines,
when --logtostderr is enabled, the same way it normally does. This
makes debugging more complex in some cases. So, we'll manually do it.
This change renames the old Error type to Exception -- more consistent
with our AST, etc. nodes anyway -- and introduces a new Error type ("<error>")
to use when something during typechecking or binding fails.
The old way led to errors like:
error: MU504: tags.ts:32:18: Symbol 'Tag:push' not found
error: MU522: tags.ts:32:8: Cannot invoke a non-function; 'any' is not a function
This is because of cascading errors during type-checking; the symbol not found
error means we cannot produce the right type for the function invoke that
consumes it. But the 'any' part is weird. Instead, this new change produces:
error: MU504: tags.ts:32:18: Symbol 'Tag:push' not found
error: MU522: tags.ts:32:8: Cannot invoke a non-function; '<error>' is not a function
It's slightly better. And furthermore, it gives us a leg to stand on someday
should we decide to get smarter about detecting cascades and avoiding issuing
the secondary error messages (we can just check for the Error type).
This change fixes a few more phasing issues in the compiler. Namely,
it now splits all passes into three distinct phases:
1. Declarations: simply populating names.
2. Definitions: chasing down any references to other names from those
declared entities. For instance, base classes, other modules, etc.
3. Bodies: fully type-checking everything else, which will depend
upon both declarations and definitions being fully present.
This changes a few things with dependency probing:
1) Probe for Mupack files, not Mufiles.
2) Substitute defaults in the PackageURL before probing.
3) Trace the full search paths when an import fails to resolve.
This will help diagnose dependency resolution issues.
This adds support for sugared "*" semvers, which is a shortcut for
">=0.0.0" (in other words, any version matches). This is a minor part
of marapongo/mu#18. I'm just doing it now as a convenience for dev
scenarios, as we start to do real packages and dependencies between them.
This change permits object literals with the type of `any`; although
they will typically get coerced to record/interface types prior to use,
there is no reason to ban `any`. In fact, it is the idiomatic way of
encoding "objects as bags of properties" that is common to dynamic
languages like JavaScript, Python, etc.
This change adds support for loading from export locations. All we
need to do is keep chasing the referent pointer until we bottom out
on an actual method, property, etc.
This change fixes a whole host of issues with our current token binding
logic. There are two primary aspects of this change:
First, the prior token syntax was ambiguous, due to our choice of
delimiter characters. For instance, "/" could be used both as a module
member delimiter, in addition to being a valid character for sub-modules.
The result is that we could not look at a token and know for certain
which kind it is. There was also some annoyance with "." being the
delimiter for class members in addition to being the leading character
for special names like ".this", ".super", and ".ctor". Now, we just use
":" as the delimiter character for everything. The result is unambiguous.
Second, the simplistic token table lookup really doesn't work. This is
for three reasons: 1) decorated types like arrays, maps, pointers, and
functions shouldn't need token lookup in the classical sense; 2) largely
because of decorated naming, the mapping of token pieces to symbolic
information isn't straightforward and requires parsing; 3) default modules
need to be expanded and the old method only worked for simple cases and,
in particular, would not work when combined with decorated names.
Previously we asserted that typemap entries are never nil. It turns
out we represent "void" as the absence of a type in MuIL, and so we need
to permit these for constructors, etc.
This change further rearranges the phasing of binding to account for
the fact that class definitions may freely reference exports, which won't
be known until after binding class names. Therefore, we bind class names
first and then the definitions (function signatures and variables).
This brings the AWS MuPackage's number of verification errors down to 84
from 136.
The old method of specifying a default module for a package was using
a bit on the module definition itself. The new method is to specify the
module name as part of the package definition itself.
The new approach makes more sense for a couple reasons. 1) it doesn't
make sense to have more than one "default" for a given package, and yet
the old model didn't prevent this; the new one prevents it by construction.
2) The defaultness of a module is really an aspect of the package, not the
module, who is generally unaware of the package containing it.
The other reason is that I'm auditing the code for nondeterministic map
enumerations, and this came up, which simply pushed me over the edge.
The prior module binding logic was naive. It turns out that, because of inter-module
references within the same package, the order of binding certain modules and its members
within a package matters. This leads to a more Java/C#/JavaScript/etc. multi-pass model,
versus a classical single-pass C/C++ model (which requires forward declarations).
To do this, we bind things in this order within a package:
* First, we add an entry for each module. It is largely empty, but at least names resolve.
* Second, we bind all imports. Since all modules are known, all inter-module tokens resolve.
* Next, we can bind class symbols. Note that we must do this prior to module properties,
methods, and exports, since they might be referenced and need to exist first. As before,
we do not yet bind function bodies, since those can reference anything.
* Now we bind all module "members", meaning any module scoped properties and methods. Again,
we do not bind function bodies just yet. This is just a symbolic binding step.
* Next, we can finally bind exports. Because exports may reference any members, all of the
above must have been done first. But all possibly exported symbols are now resolved.
* Finally, we can bind function bodies.
This adds a `mu verify` command that simply runs the verification
pass against a MuPackage and its MuIL. This is handy for compiler
authors to verify that the right stuff is getting emitted.
This is pretty worthless, but will help me debug some issues locally.
Eventually we want MuGL to be fully serializable, including the option
to emit DOT files.
This change actually invokes the OnVariableAssign interpreter hook
at the right places. (Renamed from OnAssignProperty, as it will now
handle all variable assignments, and not just properties.) This
requires tracking a bit more information about l-values so that we
can accurately convey the target object and symbol associated with
the assignment (resulting in the new "location" struct type).
This change lowers the information collected about resource allocations
and dependencies into the MuGL graph representation.
As part of this, we've moved pkg/compiler/eval out into its own top-level
package, pkg/eval, and split up its innards into a smaller sub-package,
pkg/eval/rt, that contains the Object and Pointer abstractions. This
permits the graph generation logic to use it without introducing cycles.
This change refactors the interpreter hooks into a first class interface
with many relevant event handlers (including enter/leave functions for
packages, modules, and functions -- something necessary to generate object
monikers). It also includes a rudimentary start for tracking actual object
allocations and their dependencies, a step towards creating a MuGL graph.
This change adds a new Allocator type that will handle all object allocations
that occur during evaluation. This allows us to hook into interesting lifecycle
events that are required in order to create a MuGL resource graph.
This change mirrors the same change we made for local variable scope
addressing. GetPropertyPointer is all about getting the property's
address; instead of GetPropertyPointer, which could be confused as
meaning the property's value is a pointer, we will call this method
GetPropertyAddr instead.
This change dumps the evaluation state after evaluation completes, at
log-level 5. This includes which modules and classes were initialized,
in addition to the values for all global variables.
In addition to this, we rename a few things:
* Rename Object's Data field to Value.
* Rename the Object.T() methods to Object.TValue(). This more clearly
indicates what they are doing (i.e., fetching the value from the object)
and also avoids object.String() conflicting with fmt.Stringer's String().
* Rename Reference to Pointer, so it's consistent with everything else.
* Rename the GetValueReference/InitValueReference/etc. family of methods
to GetValueAddr/InitValueAddr/etc., since this reflects what they are
actually doing: manipulating a variable slot's address.
* Assert that module/class initializer functions returned nil.
* Evaluate the LHS as an l-value for binary assignment operators.
* Also consider two nulls to be equal, always.
This change implements unary operator evaluation, including prefix/postfix
assignment operators. Doing this required implementing l-values properly
in the interpreter; namely, load location and dereference, when used in an
l-value position, result in a pointer to the location, while otherwise they
(implicitly, in load location's case) deference the location and yield a value.
This change uses an enum rather than set of bools for unwind's inner
state, makes its fields private, and adds accessor methods. This makes
it easier to protect invariants and wonky states like being a break and
a throw simultaneously. Now that Unwind is part of the public API (vs
being private as it used to), this is really a requirement (and was
obviously a good idea regardless). Thanks to Eric for CR feedback.
This change adds some fail statements into places that aren't yet
implemented during binding and evaluation so I can qucikly catch
stuff that isn't working as expected yet.
Three minor blocking issues fixed, discovered as I am running tests
locally to stress the interpretr in a few different ways:
* Properly initiaize the local scope.
* Log pkg.Name(), not pkg.Name, so we get a string and not an address.
* Don't dereference uw after doing a call, when it could be nil.
This change populates the special `this` and `super` variables in
instance method activation frames. This occurs both during binding --
so that we can properly typecheck and verify references to them -- and
during evaluation -- so we can actually operate on their values.
Now that we have introduced a full blown token map -- new as of just
a few changes ago -- we can start using it for all of our symbol binding.
This also addresses some order-dependent issues, like intra-module
references looking up symbols that have been registered in the token map
but not necessarily stored in the relevant parent symbols just yet.
Plus, frankly, it's much simpler and uses a hashmap lookup instead of
a fairly complex recursive tree walk.
I've kept the tree walk case, however, to improve diagnostics upon
failure. This allows us to tell developers, for example, that the reason
a binding failed was due to a missing package.
This change revives some compiler tests that are still lingering around
from the old architecture, before our latest round of ship burning.
It also fixes up some bugs uncovered during this:
* Don't claim that a symbol's kind is incorrect in the binder error
message when it wasn't found. Instead, say that it was missing.
* Do not attempt to compile if an error was issued during workspace
resolution and/or loading of the Mufile. This leads to trying to
load an empty path and badness quickly ensues (crash).
* Issue an error if the Mufile wasn't found (this got lost apparently).
* Rename the ErrorMissingPackageName message to ErrorInvalidPackageName,
since missing names are now caught by our new fancy decoder that
understands required versus optional fields. We still need to guard
against illegal characters in the name, including the empty string "".
* During decoding, reject !src.IsValid elements. This represents the
zero value and should be treated equivalently to a missing field.
* Do not permit empty strings "" as Names or QNames. The old logic
accidentally permitted them because regexp.FindString("") == "", no
matter the regex!
* Move the TestDiagSink abstraction to a new pkg/util/testutil package,
allowing us to share this common code across multiple package tests.
* Fix up a few messages that needed tidying or to use Infof vs. Info.
The binder tests -- deleted in this -- are about to come back, however,
I am splitting up the changes, since this represents a passing fixed point.
This change looks up the main module from a package, and that module's
entrypoint, when performing evaluation. In any case, the arguments are
validated and bound to the resulting function's parameters.
This change eliminates all traces of the old, legacy errors and warnings.
It also renumbers everything so that we don't have any "gaps". Finally,
it introduces a factory method for new errors and warnings, private to
this package, that ensures we don't accidentally register duplicate IDs.
This change deletes a bunch of old legacy code. I've tagged the
prior changelist as "0.1" in case we need to go back and recover
some of the code (as I expect some of the constraint type logic
and AWS-specific code to come in handy down the road).
🔥🔥🔥
This adds support for LoadLocationExpression, which will be key to
getting virtually everything else working, including class and module
member accesses, local variable gets/sets, function calls, and more.
In particular, an extra level of indirection is added to the globals,
locals, and properties maps. This new thing is called a *Reference,
and it is simply a mutable slot that contains an *Object. Instead
of storing and retrieving objects directly, this extra level of
indirection enables us to create *Object instances that themselves
simply refer to other objects (for subsequent loads and stores).
For now, we do not permit overwriting of method properties. This
is a future work item (see marapongo/mu#56).
This change adds support for function calls. We do not yet wire it
into the various function call expressions, however, that should be
relatively trivial after this change. (It is dependent on figuring
out the runtime representation of load location objects and subsequent
uses of them). This change also updates the scoping logic to respect
"activation frame" style of lexical scoping, where an inner function
as a result of a call should not have access to the callee context.
This change also includes logic to detect unhandled exceptions and to
print an error as a result. Eventually, we will want stack traces here.
This change eliminates the scope-based symbol table. Because we now
require that all module, type, function, and variable elements are
encoded as fully qualified tokens, there is no need for the scope-based
lookups. Instead, the languages themselves decide how the names bind
to locations and just encode that information directly.
The scope is still required for local variables, however, since those
don't have a well-defined "fixed" notion of name. This is also how
we will ensure the evaluator stores values correctly -- including
discarding them -- in a lexically scoped manner.
This change checks in an enormously rudimentary interpreter. There is
a lot left to do, as evidenced by the countless TODOs scattered throughout
pkg/compiler/eval/eval.go. Nevertheless, the scaffolding and some important
pieces are included in this change.
In particular, to evaluate a package, we must locate its entrypoint and then,
using the results of binding, systematically walk the full AST. As we do
so, we will assert that aspects of the AST match the expected shape,
including symbols and their types, and produce value objects for expressions.
An *unwind structure is used for returns, throws, breaks, and continues (both
labeled and unlabeled). Each statement or expression visitation may optionally
return one and its presence indicates that control flow transfer is occurring.
The visitation logic then needs to propagate this; usually just by bailing out
of the current context immediately, but sometimes -- as is the case with
TryCatchBlock statements, for instance -- the unwind is manipulated in more
"interesting" ways.
An *Object structure is used for expressions yielding values. This is a
"runtime object" in our system and is comprised of three things: 1) a Type
(essentially its v-table), 2) an optional data pointer (for tagged primitives),
and 3) an optional bag of properties (for complex object property values).
I am still on the fence about whether to unify the data representations.
The hokiest aspect of this change is the scoping and value management. I am
trying to avoid needing to implement any sort of "garbage collection", which
means our bag-of-values approach will not work. Instead, we will need to
arrange for scopes to be erected and discarded in the correct order during
evaluation. I will probably tackle that next, along with fleshing out the
many missing statement and expression cases (...and tests, of course).
This change introduces a binder.Context structure, with a
*core.Context embedded, that carries additional semantic information
forward to future passes in the compiler. In particular, this is how
the evaluation/interpretation phase will gain access to types, scopes,
and symbols.
This change memoizes decorated types to avoid creating boatloads
of redundant symbols at runtime. We want it to be the case that you
can create instances of decorated types as needed throughout the
compiler without worry for excess garbage (e.g., arrays, pointers, etc).
During evaluation, we need to perform very delicate walking of the AST,
and are likely to use the ability to substitute visitors "in situ." As
a result, I'd like to make anonymous visitors easier to produce. This
change permits both single function forms of Visit/After, in addition to
anonymous structures that pair up a Visit and an After, to produce a Visitor.
This change completes my testing of decorator parsing for now. It tests the token
`*[]map[string]map[()*(bool,string,test/package:test/module/Crazy)number][][]test/package:test/module/Crazy`.
This turned up some bugs, most notably in the way we returned the "full" token for
the parsed types. We need to extract the subset of the token consumed by the parsing
routine, rather than the entire thing. To do this, we introduce a tokenBuffer type
that allows for convenient parsing of tokens (eating, advancing, extraction, etc).
This isn't comprehensive yet, however it caught two bugs:
1. parseNextType should operate on "rest" in most cases, not "tok".
2. We must eat the "]" map separator before moving on to the element type.
Part of the token grammar permits so-called "decorated" types. These
are tokens that are pointer, array, map, or function types. For example:
* `*any`: a pointer to anything.
* `[]string`: an array of primitive strings.
* `map[string]number`: a map from strings to numbers.
* `(string,string)bool`: a function with two string parameters and a
boolean return type.
* `[]aws:s3/Bucket`: an array of objects whose class is `Bucket` from
the package `aws` and its module `s3`.
This change introduces this notion into the parsing and handling of
type tokens. In particular, it uses recursive parsing to handle complex
nested structures, and the binder.bindTypeToken routine has been updated
to call out to these as needed, in order to produce the correct symbol.
This changes a few things around binding logic, as part of eliminating
all of the legacy logic and weaving it into the new codebase:
* Give Scopes access to the Context object. Related, add a TryRegister
method to Scope that is like RequireRegister, except that instead of
fail-fast upon encountering a duplicate entry, it will issue an error.
* Move all typecheck visitation functions out of the big honkin' switch
and into their own member functions. As this stuff gets more complex,
having everything in one routine was starting to irk my sensibilities.
* Validate that packages have names.
* Store both the package symbol, plus the canonicalized URL used to
resolve it, in the package map. This will help us verify that versions
match for multiple package references resolving to the same symbol.
* Add nice inquiry methods to the Class symbol (Sealed, Abstract, Record,
Interface) that simplify accessing the modifiers on the underlying node.
The options structure will be shared between multiple passes of
compilation, including evaluation and graph generation. Therefore,
it must not be in the pkg/compile package, else we would create
package cycles. Now that the options structure is barebones --
and, in particular, no more "backend" settings pollute it -- this
refactoring actually works.