commit 7f266408e95c50af158df1108fe6bffc3030eec7
author adonovan <adonovan@google.com> Mon Apr 20 10:15:47 2020 -0700
committer Copybara-Service <copybara-worker@google.com> Mon Apr 20 10:17:08 2020 -0700
tree dd9d35cd3861175d77f939438d05fa8e5ed8747b
parent 8a312b666b098c43a99a8ee628cc64809ab3dec4

bazel syntax: hide BaseFunction and matchSignature

commit c568d844b6c73160273e60096fa9507d6755acde replaced most uses of BaseFunction with StarlarkCallable,
which abstracts all function-like values. This change removes BaseFunction
from the API, and a follow-up will merge it into its sole remaining subclass, StarlarkFunction.

Most of the classes that implemented BaseFunction did so only for its
trivial toString, repr, and isImmutable methods, which have been copied
as needed, and in some cases tweaked.

SkylarkProvider formerly used BaseFunction and matchSignature
to match arguments to schema fields, but provider instantiation can be
done more efficiently by permuting the argument array directly, without
allocating any hash tables or indeed any additional memory at all.
Thanks to Murali Ganapathy for the algorithm.
This also unifies the schemaful and schemaless cases.

This sets the stage for removing FunctionSignature from the API.
Its only remaining uses are in the doc tools.

---

Also, this CL causes ParamDescriptor to no longer use SkylarkType internally
to validate arguments and to describe types for error messages.
Parameter types are now represented as a union of Class symbols.
Validation uses instanceof. Error messages are produced using
getDataTypeNameFromClass.

The reason this change is bundled in this CL is that SkylarkType.of
(a conceptual mess) uses fragile heuristics to decide whether a value is
a "function" based on whether it implements BaseFunction; this breaks
when changed to use StarlarkCallable.

Type errors in the function call machinery now report only what they test,
that is, the top-level constructor: "want list", not "want list of strings".
The old messages were a bit of a lie because the call would happily accept
lists of other types unless the function implementation performed an explicit
element type check. Such checks continue to produce detailed errors:
"got list of X, want list of Y".

The 'generic1' annotation no longer affects the interpreter;
it is purely for documentation.

The only remaining use of SkylarkType is for Depsets.
A follow-up change will make the necessary renamings and simplifications.

PiperOrigin-RevId: 307424087