Shrink Task.Delay when used without Cancellation

[benaadams]

Task.Delay is commonly used without cancellation:

Task Delay(int) - API Port 22.6%
Task Delay(TimeSpan) - API Port 9.8%

Task Delay(int, CancellationToken) - API Port 5.6%
Task Delay(TimeSpan, CancellationToken) - API Port 9.3%

This shaves 24 bytes of allocation off that common path.

Using @SergeyTeplyakov's ObjectLayoutInspector and local copies of the internal objects to make it easier to inspect.

ObjectLayoutInspector.TypeLayout.PrintLayout<DelayPromise>(recursively: true);
ObjectLayoutInspector.TypeLayout.PrintLayout<DelayPromiseWithoutCancellation>(recursively: true);

Type layout for 'DelayPromise'
Size: 88 bytes. Paddings: 7 bytes (%7 of empty space)
|==============================================================|
| Object Header (8 bytes)                                      |
|--------------------------------------------------------------|
| Method Table Ptr (8 bytes)                                   |
|==============================================================|
|   0-7: Delegate m_action (8 bytes)                           |
|--------------------------------------------------------------|
|  8-15: Object m_stateObject (8 bytes)                        |
|--------------------------------------------------------------|
| 16-23: TaskScheduler m_taskScheduler (8 bytes)               |
|--------------------------------------------------------------|
| 24-31: Object m_continuationObject (8 bytes)                 |
|--------------------------------------------------------------|
| 32-39: ContingentProperties m_contingentProperties (8 bytes) |
|--------------------------------------------------------------|
| 40-43: Int32 m_taskId (4 bytes)                              |
|--------------------------------------------------------------|
| 44-47: Int32 m_stateFlags (4 bytes)                          |
|--------------------------------------------------------------|
|    48: VoidTaskResult m_result (1 byte)                      |
|--------------------------------------------------------------|
| 49-55: padding (7 bytes)                                     |
|--------------------------------------------------------------|
| 56-63: TimerQueueTimer Timer (8 bytes)                       |
|--------------------------------------------------------------|
| 64-71: CancellationToken Token (8 bytes)                     |
|--------------------------------------------------------------|
| 72-87: CancellationTokenRegistration Registration (16 bytes) |
| |=====================================|                      |
| |   0-7: CallbackNode _node (8 bytes) |                      |
| |-------------------------------------|                      |
| |  8-15: Int64 _id (8 bytes)          |                      |
| |=====================================|                      |
|==============================================================|

To

Type layout for 'DelayPromiseWithoutCancellation'
Size: 64 bytes. Paddings: 7 bytes (%10 of empty space)
|==============================================================|
| Object Header (8 bytes)                                      |
|--------------------------------------------------------------|
| Method Table Ptr (8 bytes)                                   |
|==============================================================|
|   0-7: Delegate m_action (8 bytes)                           |
|--------------------------------------------------------------|
|  8-15: Object m_stateObject (8 bytes)                        |
|--------------------------------------------------------------|
| 16-23: TaskScheduler m_taskScheduler (8 bytes)               |
|--------------------------------------------------------------|
| 24-31: Object m_continuationObject (8 bytes)                 |
|--------------------------------------------------------------|
| 32-39: ContingentProperties m_contingentProperties (8 bytes) |
|--------------------------------------------------------------|
| 40-43: Int32 m_taskId (4 bytes)                              |
|--------------------------------------------------------------|
| 44-47: Int32 m_stateFlags (4 bytes)                          |
|--------------------------------------------------------------|
|    48: VoidTaskResult m_result (1 byte)                      |
|--------------------------------------------------------------|
| 49-55: padding (7 bytes)                                     |
|--------------------------------------------------------------|
| 56-63: TimerQueueTimer Timer (8 bytes)                       |
|==============================================================|

/cc @stephentoub

[benaadams]

@dotnet test OSX10.12 x64 Checked Innerloop Build and Test

[Wraith2]

That 7 bytes of padding bothers me. Is there a way to mask the voidresult into the state flags?

[stephentoub]

Is there a way to mask the voidresult into the state flags?

No. That's just the generic TResult. The right answer there is to move the promise functionality down from the generic task into the non-generic one. It's on my to-do list to explore that.

[benaadams]

Is there a way to mask the voidresult into the state flags?

No. That's just the generic TResult. The right answer there is to move the promise functionality down from the generic task into the non-generic one. It's on my to-do list to explore that.

Could move m_taskId into m_contingentProperties, then it would drop to 3 bytes ;)

Though it would go back to 4 bytes after making it non-generic (e.g. moving m_taskId would be a wash at that point)

[Wraith2]

I just hadn't realised it was T result, it isn't really wasted padding.

[stephentoub]

The right answer there is to move the promise functionality down from the generic task into the non-generic one. It's on my to-do list to explore that.

#22238

[benaadams]

Fixed commits; change should be the same

[benaadams]

Though I think I have them reversed O_o

[benaadams]

@dotnet-bot test OSX10.12 x64 Checked Innerloop Build and Test

[jkotas]

@benaadams Could you please resolve the conflict?

[stephentoub]

Resolved.