Surprising whitespace sensitivity in type parameters

[matklad]

Consider this code:

struct Ok <T: Clone + = ()> { t: T }

struct Err<T: Clone +=  ()> { t: T }

The Ok struct compiles just right, but the Err one fails with

error: expected one of `(`, `+`, `,`, `::`, `<`, `=`, or `>`, found `+=`
 --> main.rs:2:21
  |
2 | struct Err<T: Clone +=  ()> { t: T }
  |                     ^^ expected one of 7 possible tokens here

error: aborting due to previous error

I think this behavior is wrong: for tokens like >> or || use use "context sensitive" lexing.

[petrochenkov]

Yeah, "token splitting" (&& -> & & or += -> + =) is done very inconsistently.
It's not done automatically, but manually if the code author expects the split version to result in valid code sometimes. It looks like nobody expected + = to be a fragment of valid code in this position.

[Crazycolorz5]

I assume no one would mind if I took a look at this.