Keep object locals only once in AST
For example this reduces the size of stdlib ast file
roughly 3x. Note that this change doesn't regenerate the stdlib,
so that the diff here is sane.
It is likely to slightly improve performance of code using
a lot of locals (~10% on bench.05.gen.jsonnet).
The desugaring is more strightforward now, and we're back
to desugaring each node exactly once.
My understanding of origin of this bug was that once we
were creating thunks for binary operator arguments. So the environment
was no longer needed once they were created, so they could be tailcalls.
Now we're calling i.evaluate directly (for performance) and now
the environment cannot be destroyed during the evaluation of the first
argument.
* Add test support for multi-file output.
* Add -update support for multi-file output tests.
* Add support for string output in multi-file output mode.
* Rename 'stringOutput' to 'stringOutputMode' to better express what it does
* Refactor main_test to make it less nested.
This also causes the -update flag to output a list of files which
have been updated. This does not include the paths which are deleted
for multi-file tests.
Currently, `jsonnet.VM#NativeFunction` takes a single argument of type
`jsonnet.nativeFunction`. This is ok for internal use, but because this
type is private to the `jsonnet` package, it is not possible for a
third party to call this function (since it can't instantiate the type).
This commit makes this type public to remedy this problem.
* Location, error formatting and stack trace improvements
* Static context for AST nodes
* Thunks no longer need `name`
* Prototype for showing snippets in error messages (old format still
available)
* Use ast.Function to represent methods and local function sugar.
* Change tests so that the error output is pretty
This is necessary for example for native functions
(which take json as arguments).
Standard "encoding/json" representation is used, but I have
mixed feeling about it. Not sure if treating json values as interface{}
is the right trade-off in our case.
* Improve imports
* Nonexistent file is now a runtime error (used to be ignored)
* Permission denied is a runtime error (used to be an internal error)
* Basic tests added for imports
* Renamed tests from .input to .jsonnet
* It's standard
* It's compatible with C++ test suite
* Editors enable syntax highlighting based on extension
* Importing .input looks weird
* Interpreter & runtime - minimal usable version
Accomplishments in this commit:
* Majority of language features implemented:
* Unary operators
* Binary operators
* Conditionals
* Errors
* Indexing arrays
* Indexing objects
* Object inheritance
* Imports
* Functions
* Function calls
* There is a quite nice way for creating builtins
* Static analyzer is there with most of the functionality
* Standard library is included and parts unaffected by missing features
work
* Some bugs in existing parts fixed
* Most positive tests from C++ version pass, the rest is failing mostly
due to missing builtins and comprehensions.
* Some initial structure was created that should allow more incremental
and focused changes in the future PRs.
* Some comments/explanations added
* Panics translated to a little bit more gentle internal errors (with a
link to issues on github).
What still sucks:
* Stack traces & error messages (there's some stuff in place)
* Almost everything is in the same package
* Stuff is exported or unexporeted randomly (see above)
* Missing a few lexing/parsing features
* Missing builtins
* Missing support for extvars and top-level-args
* Checking function arguments is missing
* No clean Go API that commandline and compatibility layer to C can use
* No compatibility layer to C
* Assertions don't work (desugaring level, assertEquals works).
* Manifestation stack traces (and generally it could use some work).
* The way environments are constructed is sometimes suboptimal/clumsy.
