[AutoDiff] Simplify varied propagation in activity analysis.

[dan-zheng]

• Remove setVaried.
  • setVaried has only one user and can be inlined. Addresses feedback from @marcrasi.
• Remove setVariedAcrossArrayInitialization.
  • Activity info does not change when removing special variedness propagation
    support for array.uninitialized_intrinsic applications.
  • Verified with array initialization activity info test case.

[dan-zheng]

@swift-ci Please test tensorflow

[rxwei]

Activity info does not change when removing special variedness propagation
support for array.uninitialized_intrinsic applications.
Verified with array initialization activity info test case.

Could you please add a test to make sure this still holds true for indirectly initialized array literals? Indirectness should always be considered in these sorts of test cases.

func testArrayUninitializedIntrinsic<T>(_ x: T, _ y: T) -> [T] {
   return [x, y]
}

[dan-zheng]

Could you please add a test to make sure this still holds true for indirectly initialized array literals? Indirectness should always be considered in these sorts of test cases.

Done in 05122fb. Verified that activity info for array.uninitialized_intrinsic applications didn't change before/after this patch.

There already exist runtime tests for array.uninitialized_intrinsic applications in test/AutoDiff/array.swift for these cases and more.

[dan-zheng]

@swift-ci Please test tensorflow

[rxwei]

Great. To throw another example here: Consider testing an indirect type (e.g. struct S<T> { var x: T }) that conforms to ExpressibleByArrayLiteral. This is different from Array because Array is always directly passed.

[dan-zheng]

Great. To throw another example here: Consider testing an indirect type (e.g. struct S<T> { var x: T }) that conforms to ExpressibleByArrayLiteral. This is different from Array because Array is always directly passed.

Added an "indirect ExpressibleByArrayLiteral type" test in bb57241.

I didn't add a corresponding activity analysis test because activity info around the array.uninitialized_intrinsic application is the same.

[NONE]   // function_ref _allocateUninitializedArray<A>(_:)
  %6 = function_ref @$ss27_allocateUninitializedArrayySayxG_BptBwlF : $@convention(thin) <τ_0_0> (Builtin.Word) -> (@owned Array<τ_0_0>, Builtin.RawPointer) // user: %7
[ACTIVE]   %7 = apply %6<T>(%5) : $@convention(thin) <τ_0_0> (Builtin.Word) -> (@owned Array<τ_0_0>, Builtin.RawPointer) // user: %8
[ACTIVE] (**%8**, %9) = destructure_tuple %7 : $(Array<T>, Builtin.RawPointer) // user: %17
[VARIED] (%8, **%9**) = destructure_tuple %7 : $(Array<T>, Builtin.RawPointer) // user: %10
[VARIED]   %10 = pointer_to_address %9 : $Builtin.RawPointer to [strict] $*T // users: %13, %11
[VARIED]   %12 = integer_literal $Builtin.Word, 1          // user: %13
[VARIED]   %13 = index_addr %10 : $*T, %12 : $Builtin.Word // user: %14

[dan-zheng]

@swift-ci Please test tensorflow