[Bug] error out in compilation when encounter an unknown annotation/attribute

[lightmark]

🚀 Feature Request

Compiler ignores unknown annotations but actually it should just error out. Otherwise, typo must be introduced while the author doesn't know. For example -> #8866 for examples.

Motivation

#[testonly] in #8866 for examples.

Pitch

Additional context

[brmataptos]

list of allowed attributes is implicitly in the constants defined here.

Oops, extended_checks.rs also has a few known attributes ("legacy_entry_fun", "view", "group", "scope", "resource_group", "resource_group_member", "group"), with an odd variety of constant names associated with them.

We’re also missing #[deprecated{

---

More from Slack discussion:

Kevin Hoang
Another thing to note is we need a good error message in case users already use some custom attributes

Wolfgang Grieskamp
The cli needs a flag to skip those checks

Kevin Hoang
We might also consider giving a harsher warning for variants of test_only such as testonly as those are the most risky ones (accidentally releasing test code)

[brmataptos]

Specific Requirements:

• mechanism to register and check known attribute names from various code
• register attributes
  • third_party/move/move-compiler/src/shared/mod.rs#L518
  • aptos-move/framework/src/extended_checks.rs#L24
  • #[deprecated]
• good error message in case users use custom attributes
  • CLI needs a flag to skip those checks
  • maybe a flag to allow registering more attributes
• also consider harsher warning for variants of test_only such as testonly
• tests

[brmataptos]

Wolfgang is apparently greatly proposed to a flag registering more attributes, so we'll just have a flag to skip checks.

[wrwg]

Yes, I do think we should leave user registerable attributes to the compiler v2 and language extensions. Specifically, I propose to do attributes similar as in C#/Java, so they can be declared in the language, together with typing of their parameters, as in:

#[define_attribute] public fun my_attribute(param1: u64, param2: String)

Besides this being more general, the flag approach doesn't seem to cover what people expect from compiler command line options. A user would have to every single time specify this options, if he forgets, then compilation fails. Moreover, the allowed parameters of the attributes cannot be easily defined this way.

I assume at this point not many people are out there which actually use there own attributes. Those few can be grandfathered by the the single flag --skip-attribute-checks.