Add `linear_function` and `linear_function_extract` instructions.
`linear_function` creates a `@differentiable(linear)` function-typed value from
an original function operand and a transpose function operand (optional).
`linear_function_extract` extracts either the original or transpose function
value from a `@differentiable(linear)` function.
Resolves TF-1142 and TF-1143.
Add `differentiable_function` and `differentiable_function_extract`
instructions.
`differentiable_function` creates a `@differentiable` function-typed
value from an original function operand and derivative function operands
(optional).
`differentiable_function_extract` extracts either the original or
derivative function value from a `@differentiable` function.
The differentiation transform canonicalizes `differentiable_function`
instructions, filling in derivative function operands if missing.
Resolves TF-1139 and TF-1140.
`@differentiable` attribute on protocol requirements and non-final class members
will produce derivative function entries in witness tables and vtables.
This patch adds an optional derivative function configuration
(`AutoDiffDerivativeFunctionIdentifier`) to `SILDeclRef` to represent these
derivative function entries.
Derivative function configurations consist of:
- A derivative function kind (JVP or VJP).
- Differentiability parameter indices.
Resolves TF-1209.
Enables TF-1212: upstream derivative function entries in witness tables/vtables.
In order to allow this, I've had to rework the syntax of substituted function types; what was previously spelled `<T> in () -> T for <X>` is now spelled `@substituted <T> () -> T for <X>`. I think this is a nice improvement for readability, but it did require me to churn a lot of test cases.
Distinguishing the substitutions has two chief advantages over the existing representation. First, the semantics seem quite a bit clearer at use points; the `implicit` bit was very subtle and not always obvious how to use. More importantly, it allows the expression of generic function types that must satisfy a particular generic abstraction pattern, which was otherwise impossible to express.
As an example of the latter, consider the following protocol conformance:
```
protocol P { func foo() }
struct A<T> : P { func foo() {} }
```
The lowered signature of `P.foo` is `<Self: P> (@in_guaranteed Self) -> ()`. Without this change, the lowered signature of `A.foo`'s witness would be `<T> (@in_guaranteed A<T>) -> ()`, which does not preserve information about the conformance substitution in any useful way. With this change, the lowered signature of this witness could be `<T> @substituted <Self: P> (@in_guaranteed Self) -> () for <A<T>>`, which nicely preserves the exact substitutions which relate the witness to the requirement.
When we adopt this, it will both obviate the need for the special witness-table conformance field in SILFunctionType and make it far simpler for the SILOptimizer to devirtualize witness methods. This patch does not actually take that step, however; it merely makes it possible to do so.
As another piece of unfinished business, while `SILFunctionType::substGenericArgs()` conceptually ought to simply set the given substitutions as the invocation substitutions, that would disturb a number of places that expect that method to produce an unsubstituted type. This patch only set invocation arguments when the generic type is a substituted type, which we currently never produce in type-lowering.
My plan is to start by producing substituted function types for accessors. Accessors are an important case because the coroutine continuation function is essentially an implicit component of the function type which the current substitution rules simply erase the intended abstraction of. They're also used in narrower ways that should exercise less of the optimizer.
The `differentiability_witness_function` instruction looks up a
differentiability witness function (JVP, VJP, or transpose) for a referenced
function via SIL differentiability witnesses.
Add round-trip parsing/serialization and IRGen tests.
Notes:
- Differentiability witnesses for linear functions require more support.
`differentiability_witness_function [transpose]` instructions do not yet
have IRGen.
- Nothing currently generates `differentiability_witness_function` instructions.
The differentiation transform does, but it hasn't been upstreamed yet.
Resolves TF-1141.
SIL differentiability witnesses are a new top-level SIL construct mapping
"original" SIL functions to derivative SIL functions.
SIL differentiability witnesses have the following components:
- "Original" `SILFunction`.
- SIL linkage.
- Differentiability parameter indices (`IndexSubset`).
- Differentiability result indices (`IndexSubset`).
- Derivative `GenericSignature` representing differentiability generic
requirements (optional).
- JVP derivative `SILFunction` (optional).
- VJP derivative `SILFunction` (optional).
- "Is serialized?" bit.
This patch adds the `SILDifferentiabilityWitness` data structure, with
documentation, parsing, and printing.
Resolves TF-911.
Todos:
- TF-1136: upstream `SILDifferentiabilityWitness` serialization.
- TF-1137: upstream `SILDifferentiabilityWitness` verification.
- TF-1138: upstream `SILDifferentiabilityWitness` SILGen from
`@differentiable` and `@derivative` attributes.
- TF-20: robust mangling for `SILDifferentiabilityWitness` names.
`destroy_addr` is never semantically a "no-op". Just because an
instruction can be safely eliminated does not mean the instruction has
no semantics! That would be like saying we could move an unknown
memory read below an otherwise dead store!
Even if a destroy_addr of a trivial type is a no-op, we must not end up with using such a value after a destroy_addr.
The fix is to also handle aggregate fields of trivial types in MemoryLifetime.
rdar://problem/55125020
This commit adds docs for copy_unmanaged_value to SIL.rst as requested by
@eeckstein in #26839. I noticed while doing the PR that copy_unowned_value was
also not documented, so I added docs for it as well as a bonus. = ).
In case of a tuple as value type, the initializer and setter takes the tuple elements as separate arguments. This was just not handled in the assign_by_wrapper instruction lowering.
rdar://problem/53866473
`partial_apply` does not own `@inout_aliasable` arguments, so the original description "the closure does however take ownership of the partially applied arguments" is not accurate. This patch makes things clear.
It does not take ownership of its non-trivial arguments, is a trivial
function type and therefore must not be destroyed. The compiler must
make sure to extend the lifetime of non-trivial arguments beyond the
last use of the closure.
%objc = copy_value %0 : $AnObject
%closure = partial_apply [stack] [callee_guaranteed] %16(%obj) : $@convention(thin) (@guaranteed AnObject) -> ()
%closure2 = mark_dependence %closure : $@noescape @callee_guaranteed () -> () on %obj : $AnObject
%user = function_ref @useClosure : $@convention(thin) (@noescape @callee_guaranteed () -> ()) -> ()
apply %user(%closure2) : $@convention(thin) (@noescape @callee_guaranteed () -> ()) -> ()
dealloc_stack %closure : $() ->()
destroy_value %obj : $AnObject // noescape closure does not take ownership
SR-904
rdar://35590578
This undoes some of Joe's work in 8665342 to add a guarantee: if an
@objc convenience initializer only calls other @objc initializers that
eventually call a designated initializer, it won't result in an extra
allocation. While Objective-C /allows/ returning a different object
from an initializer than the allocation you were given, doing so
doesn't play well with some very hairy implementation details of
compiled nib files (or NSCoding archives with cyclic references in
general).
This guarantee only applies to
(1) calling `self.init`
(2) where the delegated-to initializer is @objc
because convenience initializers must do dynamic dispatch when they
delegate, and Swift only stores allocating entry points for
initializers in a class's vtable. To dynamically find an initializing
entry point, ObjC dispatch must be used instead.
(It's worth noting that this patch does NOT check that the calling
initializer is a convenience initializer when deciding whether to use
ObjC dispatch for `self.init`. If we ever add peer delegation to
designated initializers, which is totally a valid feature, that should
use static dispatch and therefore should not go through objc_msgSend.)
This change doesn't /always/ result in fewer allocations; if the
delegated-to initializer ends up returning a different object after
all, the original allocation was wasted. Objective-C has the same
problem (one of the reasons why factory methods exist for things like
NSNumber and NSArray).
We do still get most of the benefits of Joe's original change. In
particular, vtables only ever contain allocating initializer entry
points, never the initializing ones, and never /both/ (which was a
thing that could happen with 'required' before).
rdar://problem/46823518
