Exploration of custom awaitables; typical results:
for 1 inner-op (best to show memory delta):
| Method | Mean | Error | StdDev | Ratio | Gen 0 | Gen 1 | Gen 2 | Allocated |
|---|---|---|---|---|---|---|---|---|
| Task | 1.419 us | 0.0345 us | 0.0019 us | 1.00 | 0.0381 | - | - | 248 B |
| ValueTask | 1.427 us | 0.0825 us | 0.0045 us | 1.01 | 0.0420 | - | - | 264 B |
| TaskLike | 1.317 us | 0.0380 us | 0.0021 us | 0.93 | 0.0095 | - | - | 64 B |
or for 100 inner-ops (best to show performance delta):
| Method | Mean | Error | StdDev | Ratio | Gen 0 | Gen 1 | Gen 2 | Allocated |
|---|---|---|---|---|---|---|---|---|
| Task | 1,431.03 ns | 126.987 ns | 6.9606 ns | 1.00 | 0.0195 | - | - | 3 B |
| ValueTask | 1,424.12 ns | 9.393 ns | 0.5149 ns | 1.00 | 0.0195 | - | - | 3 B |
| TaskLike | 66.64 ns | 55.237 ns | 3.0277 ns | 0.05 | 0.0006 | 0.0002 | - | 1 B |
The 3 tests do the exact same thing; the only thing that changes is the return type, i.e.
public async Task<int> ViaTask() {...}
public async ValueTask<int> ViaValueTask() {...}
public async TaskLike<int> ViaTaskLike() {...}All of them have the same threading/execution-context/sync-context semantics; there's no cheating.