[AutoDiff] Fix @differentiable(linear) type-checking.

[dan-zheng]

For all non-@noDerivative parameter and result types, @differentiable(linear) function types should require and imply T: Differentiable, T == T.TangentVector requirements instead of T: Differentiable & AdditiveArithmetic.

Update tests.

---

Type-checking example:

// linear.swift
struct D: Differentiable {}
// `D` does not conform to `AdditiveArithmetic`.
// `D` is not equal to `D.TangentVector`.
func test(_: @differentiable(linear) (D) -> D) {}

Before:

$ swift linear.swift
# Missing error.

After:

$ swift linear.swift
linear.swift:5:39: error: parameter type 'D' does not conform to 'Differentiable' and satisfy 'D == D.TangentVector', but the enclosing function type is '@differentiable(linear)'; did you want to add '@noDerivative' to this parameter?
func test(_: @differentiable(linear) (D) -> D) {}
                                      ^
                                      @noDerivative
linear.swift:5:45: error: result type 'D' does not conform to 'Differentiable' and satisfy 'D == D.TangentVector', but the enclosing function type is '@differentiable(linear)'
func test(_: @differentiable(linear) (D) -> D) {}
                                            ^

---

Inference example:

func inferred<T, U>(_: @differentiable(linear) (T) -> U) {}

Before:

$ swiftc -print-ast linear.swift
internal func inferred<T, U>(_: @differentiable(linear) (T) -> U) where T : AdditiveArithmetic, T : Differentiable, U : AdditiveArithmetic, U : Differentiable

After:

$ swiftc -print-ast linear.swift
internal func inferred<T, U>(_: @differentiable(linear) (T) -> U) where T : Differentiable, T == T.TangentVector, U : Differentiable, U == U.TangentVector

[dan-zheng]

@swift-ci Please test tensorflow

[rxwei]

linear.swift:5:39: error: parameter type 'D' does not conform to 'Differentiable' and satisfy 'D == D.TangentVector', but the enclosing function type is '@differentiable(linear)'; did you want to add '@noDerivative' to this parameter?
func test(_: @differentiable(linear) (D) -> D) {}
                                      ^
                                      @noDerivative
linear.swift:5:45: error: result type 'D' does not conform to 'Differentiable' and satisfy 'D == D.TangentVector', but the enclosing function type is '@differentiable(linear)'
func test(_: @differentiable(linear) (D) -> D) {}
                                            ^

Two issues:

• Why are there duplicate errors?
• The fixit is not appropriate the there's only one argument, because the function will still be illformed when you add @noDerivative.

[dan-zheng]

• Why are there duplicate errors?

The errors are not duplicate: one's for the invalid parameter and one's for the invalid result.

• The fixit is not appropriate the there's only one argument, because the function will still be illformed when you add @noDerivative.

This was the existing behavior. We could fix the fix-it so it appears only if at least one differentiability parameter remains after adding @noDerivative.

Edit: fixed in 642f2e4. Emit @noDerivative fix-it only when there is at least one valid
differentiability/linearity parameter.

[rxwei]

Not adding T.TangentVector == T was by design, but adding this constraint is okay with me. Did you encounter any concrete use cases where the absence of T.TangentVector == T requires extra user attention?

[dan-zheng]

Not adding T.TangentVector == T was by design, but adding this constraint is okay with me. Did you encounter any concrete use cases where the absence of T.TangentVector == T requires extra user attention?

Ah, interesting. I assumed that T == T.TangentVector is necessary since @transpose attribute type-checking requires it for linearity parameters - this patch makes @differentiable(linear) type-checking consistent. But I guess it may not be necessary.

---

Do we want to consistently require T: Differentiable, T == T.TangentVector for @transpose and @differentiable(linear) function types, or to relax it to T: Differentiable & AdditiveArithmetic?

Relaxing the requirement to make @differentiable(linear) function types work with more types seems preferable to me, if possible. For example, relaxing the requirement makes default transposes for protocol requirements like AdditiveArithmetic.+ work for more conforming types.

[dan-zheng]

Consensus after chatting: let's consistently require T == T.TangentVector for now.

[dan-zheng]

@swift-ci Please test tensorflow