Here's my notes. As usual all mistakes are my own:
are boxed doubles identical?
There is a bug where doubles with the same value may return false for identical() because they have been boxed to different objects.
Lars said it's been discussed. For doubles, identical() should return true based on value. We don't want the language spec to have to mention boxing so that the VM is free to optimize how it wants.
Gilad asked if NaN is identical to NaN?
Lars says yes. It's identical but not equal. Gilad will fix the spec.
change uninitialized field error to warning?
Uninitialized final fields are currently an error in the language. Kasper suggests making it a warning. It seems in line with other stuff in language. It's easy to associate some value with an uninitialized final.
Gilad says we can do this and asked why Kasper ran into it. Kasper saw some bugs where implementations behaved differently in some related corner cases.
Lars doesn't have a problem with it.
const instance variables
Gilad's view is that they should work like statics except for scoping. Apparently, though, it's complicating the VM implementation of instance metadata. Three solutions:
1. No const instance fields.
2. Metadata is statically scoped.
3. Try to do it correctly.
Lars likes 1. I say 1 simplifies things for users. Right now, people get confused with static final const etc. Gilad is OK with 1.
I asked if the syntax would be "static const" or just "const"? Users get confused when having to do "static" with constants.
Lars says they are confused because they don't understand the system. Requiring "static" will help them understand what's going on.
what liberties can editor take with type system
[I didn't have a lot of context here, so I'm fuzzy on the details.]
Dan Rubel asked how flexible the analyzer can be with the Dart type system and how it can extend it.
Gilad is OK with things like type inference for auto-complete. What other things should we allow?
Lars says step one is to do exactly what the spec says. Going beyond that and helping user with refactoring and stuff is great. Using it for warnings gets strange. If you go from the Editor to command line, you would get different warnings.
Dan's concern is more about false positives. We should look at code and see if they use constructs like guarded type promotion.
Lars says if users are using these constructs a lot, we should change the language to support them. If you want to have a type guard match thing we should have a different construct. But for now, the analyzer should go with the current spec.
Kasper says we have to be careful if we report fewer warnings because users will get used to that and then get confused if other tools follow the spec more closely and have more warnings.
Lars had lots of discussions back and forth with some internal Dart users. One issue is about catching exceptions and capturing stack traces and how its painful for some. Lars is OK if severe errors like noSuchMethod automatically get a stack trace.
Kasper says that would mean two ways to access stack traces. I note that with async, it's three. Lars says Florian has some idea of a constant flag to enable/disable stack trace capturing.
Gilad asks if the spec would have to lay out which errors get stack traces and which don't? Lars says they'll come up with a proposal.
As always, view the changelog for the full list of changes, and to get started with the Editor see our tutorial.