Some notes:
1. I put in both a swiftpm like test case and a library evolution test case. I
also updated the moveonly_deinit serialization swift test to show that we
actually serialize the deinit.
2. I changed when we emit the deinit table to only be when we have a type with
an actual value type destructor. Notably this doesn't include classes today so
as a side-effect, we no longer attempt to devirtualize moveonly class deinits.
This doesn't affect anything we are trying to actually do since we do not
support noncopyable classes today. With that in mind, I changed one test that
was showing that deinit devirtualization worked to use a struct with deinit
instead of a class.
rdar://109679168
The reason why we are doing this is that:
1. For non-copyable types, switches are always at +1 for now.
2. non-copyable enums with deinits cannot be switched upon since that would
invalidate the deinit.
So deinits on non-copyable enums are just not useful at this point since you
cannot open the enum.
Once we make it so that you can bind a non-copyable enum at +0, we will
remove this check.
I added an experimental feature MoveOnlyEnumDeinits so tests that validate the
codegen/etc will still work.
rdar://101651138
the main things still left behind the experimental flag(s) are
- move-only classes (guarded by MoveOnlyClasses feature)
- noimplicitcopy
- the _borrow operator
The Swift Simplification pass can do more than the old MandatoryCombine pass: simplification of more instruction types and dead code elimination.
The result is a better -Onone performance while still keeping debug info consistent.
Currently following code patterns are simplified:
* `struct` -> `struct_extract`
* `enum` -> `unchecked_enum_data`
* `partial_apply` -> `apply`
* `br` to a 1:1 related block
* `cond_br` with a constant condition
* `isConcrete` and `is_same_metadata` builtins
More simplifications can be added in the future.
rdar://96708429
rdar://104562580
