This adds SIL-level support and LLVM codegen for normal results of a coroutine.
The main user of this will be autodiff as VJP of a coroutine must be a coroutine itself (in order to produce the yielded result) and return a pullback closure as a normal result.
For now only direct results are supported, but this seems to be enough for autodiff purposes.
When an actual instance of a distributed actor is on the local node, it is
has the capabilities of `Actor`. This isn't expressible directly in the type
system, because not all `DistributedActor`s are `Actor`s, nor is the
opposite true.
Instead, provide an API `DistributedActor.asLocalActor` that can only
be executed when the distributed actor is known to be local (because
this API is not itself `distributed`), and produces an existential
`any Actor` referencing that actor. The resulting existential value
carries with it a special witness table that adapts any type
conforming to the DistributedActor protocol into a type that conforms
to the Actor protocol. It is "as if" one had written something like this:
extension DistributedActor: Actor { }
which, of course, is not permitted in the language. Nonetheless, we
lovingly craft such a witness table:
* The "type" being extended is represented as an extension context,
rather than as a type context. This hasn't been done before, all Swift
runtimes support it uniformly.
* A special witness is provided in the Distributed library to implement
the `Actor.unownedExecutor` operation. This witness back-deploys to the
Swift version were distributed actors were introduced (5.7). On Swift
5.9 runtimes (and newer), it will use
`DistributedActor.unownedExecutor` to support custom executors.
* The conformance of `Self: DistributedActor` is represented as a
conditional requirement, which gets satisfied by the witness table
that makes the type a `DistributedActor`. This makes the special
witness work.
* The witness table is *not* visible via any of the normal runtime
lookup tables, because doing so would allow any
`DistributedActor`-conforming type to conform to `Actor`, which would
break the safety model.
* The witness table is emitted on demand in any client that needs it.
In back-deployment configurations, there may be several witness tables
for the same concrete distributed actor conforming to `Actor`.
However, this duplication can only be observed under fairly extreme
circumstances (where one is opening the returned existential and
instantiating generic types with the distributed actor type as an
`Actor`, then performing dynamic type equivalence checks), and will
not be present with a new Swift runtime.
All of these tricks together mean that we need no runtime changes, and
`asLocalActor` back-deploys as far as distributed actors, allowing it's
use in `#isolation` and the async for...in loop.
```
let c = SomeClass()
```
is turned into
```
private let outlinedVariable = SomeClass() // statically initialized and allocated in the data section
let c = outlinedVariable
```
rdar://111021230
rdar://115502043
Also, make the ObjectOutliner work for OSSA. Though, it currently doesn't run in the OSSA pipeline.
Concurrency runtime expects discarding task operation entrypoint
function not to have result type, but the current SILGen
implementation generates reabstraction thunk to convert `() -> Void`
to `() -> T` for the operation function.
Since the `T` is always `Void` for DiscardingTG, the mismatch of result
type expectation does not cause any problem on most platforms, but the
signature mismatch causes a problem on WebAssembly.
This patch introduces new builtin operations for creating discarding
task, which always takes `() -> Void` as the operation function type.
* `alloc_vector`: allocates an uninitialized vector of elements on the stack or in a statically initialized global
* `vector`: creates an initialized vector in a statically initialized global
Values produced by address-only `enum` instructions have non-none
ownership. And because `enum` is representation-changing, they have
`owned` ownership.
This corresponds at the opaque values SIL stage to the fact that at the
address-lowered SIL stage the storage location has non-trivial
initialization which must be `destroy_addr`d, regardless of whether an
empty case was stored to it.
This instruction was given forwarding ownership in the original OSSA
implementation. That will obviously lead to memory leaks. Remove
ownership from this instruction and verify that it is never used for
non-trivial types.
This instructions marks the point where all let-fields of a class are initialized.
This is important to ensure the correctness of ``ref_element_addr [immutable]`` for let-fields,
because in the initializer of a class, its let-fields are not immutable, yet.
Codegen is the same, but `begin_dealloc_ref` consumes the operand and produces a new SSA value.
This cleanly splits the liferange to the region before and within the destructor of a class.
I was originally hoping to reuse mark_must_check for multiple types of checkers.
In practice, this is not what happened... so giving it a name specifically to do
with non copyable types makes more sense and makes the code clearer.
Just a pure rename.
When the differentiating a function containing loops, we allocate a linear map context object on the heap. This context object may store non-trivial objects, such as closures, that need to be released explicitly. Fix the autodiff linear map context allocation builtins to correctly release such objects and not just free the memory they occupy.
The new instruction is needed for opaque values mode to allow values to
be extracted from tuples containing packs which will appear for example
as function arguments.
The new instruction wraps a value in a `@sil_weak` box and produces an
owned value. It is only legal in opaque values mode and is transformed
by `AddressLowering` to `store_weak`.
The new instruction unwraps an `@sil_weak` box and produces an owned
value. It is only legal in opaque values mode and is transformed by
`AddressLowering` to `load_weak`.
This is phase-1 of switching from llvm::Optional to std::optional in the
next rebranch. llvm::Optional was removed from upstream LLVM, so we need
to migrate off rather soon. On Darwin, std::optional, and llvm::Optional
have the same layout, so we don't need to be as concerned about ABI
beyond the name mangling. `llvm::Optional` is only returned from one
function in
```
getStandardTypeSubst(StringRef TypeName,
bool allowConcurrencyManglings);
```
It's the return value, so it should not impact the mangling of the
function, and the layout is the same as `std::optional`, so it should be
mostly okay. This function doesn't appear to have users, and the ABI was
already broken 2 years ago for concurrency and no one seemed to notice
so this should be "okay".
I'm doing the migration incrementally so that folks working on main can
cherry-pick back to the release/5.9 branch. Once 5.9 is done and locked
away, then we can go through and finish the replacement. Since `None`
and `Optional` show up in contexts where they are not `llvm::None` and
`llvm::Optional`, I'm preparing the work now by going through and
removing the namespace unwrapping and making the `llvm` namespace
explicit. This should make it fairly mechanical to go through and
replace llvm::Optional with std::optional, and llvm::None with
std::nullopt. It's also a change that can be brought onto the
release/5.9 with minimal impact. This should be an NFC change.
Just the $*T -> $*@moveOnly T variant for addresses. Unlike the object version
this acts like a cast rather than something that provides semantics from the
frontend to the optimizer.
The reason why I am using a different instruction for addresses and objects here
is that the object checker doesnt have to deal with things like initialization.
The new alloc_pack_metadata and dealloc_pack_metadata are inserted as
part of IRGen lowering. The former indicates that the next instruction
might result in on-stack pack metadata being emitted. The latter
indicates that this is the position at which metadata emitted on behalf
of its operand should be cleaned up.
* [Executors][Distributed] custom executors for distributed actor
* harden ordering guarantees of synthesised fields
* the issue was that a non-default actor must implement the is remote check differently
* NonDefaultDistributedActor to complete support and remote flag handling
* invoke nonDefaultDistributedActorInitialize when necessary in SILGen
* refactor inline assertion into method
* cleanup
* [Executors][Distributed] Update module version for NonDefaultDistributedActor
* Minor docs cleanup
* we solved those fixme's
* add mangling test for non-def-dist-actor
Although nonescaping closures are representationally trivial pointers to their
on-stack context, it is useful to model them as borrowing their captures, which
allows for checking correct use of move-only values across the closure, and
lets us model the lifetime dependence between a closure and its captures without
an ad-hoc web of `mark_dependence` instructions.
During ownership elimination, We eliminate copy/destroy_value instructions and
end the partial_apply's lifetime with an explicit dealloc_stack as before,
for compatibility with existing IRGen and non-OSSA aware passes.
This allows dynamically indexing into tuples. IRGen not yet
implemented.
I think I'm going to need a type_refine_addr instruction in
order to handle substitutions into the operand type that
eliminate the outer layer of tuple-ness. Gonna handle that
in a follow-up commit.
Having added these, I'm not entirely sure we couldn't just use
alloc_stack and dealloc_stack. Well, if we find ourselves adding
a lot of redundancy with those instructions (e.g. around DI), we
can always go back and rip these out.
The variants are produced by SILGen when opaque values are enabled.
They are necessary because otherwise SILGen would produce
address_to_pointer of values.
They will be lowered by AddressLowering.
This is a dedicated instruction for incrementing a
profiler counter, which lowers to the
`llvm.instrprof.increment` intrinsic. This
replaces the builtin instruction that was
previously used, and ensures that its arguments
are statically known. This ensures that SIL
optimization passes do not invalidate the
instruction, fixing some code coverage cases in
`-O`.
rdar://39146527
By using the keyword instead of the function, we actually get a much simpler
implementation since we avoid all of the machinery of SILGenApply. Given that we
are going down that path, I am removing the old builtin implementation since it
is dead code.
The reason why I am removing this now is that in a subsequent commit, I want to
move all of the ownership checking passes to run /before/ mandatory inlining. I
originally placed the passes after mandatory inlining since the function version
of the move keyword was transparent and needing to be inlined before we could
process it. Since we use the keyword now, that is no longer an issue.
