@rintaro taught me that it was the behavior that was intentionally introduced in #20152.

#20693 improves the diagnostics.

Yes, this is correct behavior. The syntax with parentheses will work on previous versions of Swift as well.

@eeckstein, @xedin, do we already have a bug making sure the performance is the same?

@belkadan I'm not sure, but at least I didn't add any myself, will look into that.

@eeckstein Looked a the benchmarks some and I'm not really sure how to make it so performance of two different calls is compared. What do you think the best approach here would be?

Do we even expect a change in the generated code? This sound to me like a lit test would be better than a benchmark

I have checked-in multiple test-cases already including some into test/Compatibility, the difference I think is that call is going to be wrapped into an implicit closure where before it would just be a function conversion on SIL level. John and Slava agree that such forwarding isn't something which is supposed to be correct.

Actually, what I meant with lit test is a test which compiles with -O (and maybe also -Osize) and checks if the generated SIL (or llvm IR) code is what we expect.
For small, isolated code patterns, where we expect the optimizer to remove almost everything, this is a better approach than a benchmark.
For example: https://github.com/apple/swift/blob/master/test/SILOptimizer/character_literals.swift or https://github.com/apple/swift/blob/master/test/SILOptimizer/unused_containers.swift

Ah interesting, I'll take a look, thank you Erik!