Optimize DateTime.Equals

[EgorBo]

A tiny improvement for DateTime.Equals:

bool Test1(DateTime d1, DateTime d2) => d1 == d2;

// cmp against a constant
bool Test2(DateTime d1, DateTime d2) => d1 == new DateTime(637686893321649705);

Codegen diff: https://www.diffchecker.com/v4BfSbey
Could be done in JIT but it needs forward-sub feature for that first.

Self-contained benchmark: https://gist.github.com/EgorBo/e5642d9c2d362bb53880d3c9c87e5b95

|         Method |         d1 |         d2 |     Mean |    Error |   StdDev |
|--------------- |----------- |----------- |---------:|---------:|---------:|
|     Equals_new | MyDt[1000] | MyDt[1000] | 599.9 ns |  0.46 ns |  0.38 ns |
|     Equals_old | MyDt[1000] | MyDt[1000] | 651.8 ns |  0.70 ns |  0.62 ns |

|     Equals_new | MyDt[1000] | MyDt[1000] | 806.2 ns | 15.59 ns | 16.68 ns |
|     Equals_old | MyDt[1000] | MyDt[1000] | 859.2 ns |  2.34 ns |  2.19 ns |

| Equals_new_cns | MyDt[1000] | MyDt[1000] | 436.3 ns |  0.90 ns |  0.75 ns |
| Equals_old_cns | MyDt[1000] | MyDt[1000] | 855.7 ns | 13.78 ns | 12.89 ns |

| Equals_new_cns | MyDt[1000] | MyDt[1000] | 436.0 ns |  2.50 ns |  2.22 ns |
| Equals_old_cns | MyDt[1000] | MyDt[1000] | 875.0 ns | 16.63 ns | 17.80 ns |

[ghost]

I couldn't figure out the best area label to add to this PR. If you have write-permissions please help me learn by adding exactly one area label.

[ghost]

Tagging subscribers to this area: @dotnet/area-system-runtime
See info in area-owners.md if you want to be subscribed.

Issue Details

---

A tiny improvement for DateTime.Equals:

bool Test1(DateTime d1, DateTime d2) => d1 == d2;

// cmp against a constant
bool Test2(DateTime d1, DateTime d2) => d1 == new DateTime(637686893321649705);

Codegen diff: https://www.diffchecker.com/v4BfSbey
Could be done in JIT but it needs forward-sub feature for that first.

Self-contained benchmark: https://gist.github.com/EgorBo/e5642d9c2d362bb53880d3c9c87e5b95

|         Method |         d1 |         d2 |     Mean |    Error |   StdDev |
|--------------- |----------- |----------- |---------:|---------:|---------:|
|     Equals_new | MyDt[1000] | MyDt[1000] | 599.9 ns |  0.46 ns |  0.38 ns |
|     Equals_old | MyDt[1000] | MyDt[1000] | 651.8 ns |  0.70 ns |  0.62 ns |

|     Equals_new | MyDt[1000] | MyDt[1000] | 806.2 ns | 15.59 ns | 16.68 ns |
|     Equals_old | MyDt[1000] | MyDt[1000] | 859.2 ns |  2.34 ns |  2.19 ns |

| Equals_new_cns | MyDt[1000] | MyDt[1000] | 436.3 ns |  0.90 ns |  0.75 ns |
| Equals_old_cns | MyDt[1000] | MyDt[1000] | 855.7 ns | 13.78 ns | 12.89 ns |

| Equals_new_cns | MyDt[1000] | MyDt[1000] | 436.0 ns |  2.50 ns |  2.22 ns |
| Equals_old_cns | MyDt[1000] | MyDt[1000] | 875.0 ns | 16.63 ns | 17.80 ns |

Author:	EgorBo
Assignees:	-
Labels:	area-System.Runtime
Milestone:	-

[am11]

Looks like we have similar pattern in TimeOnly class:

runtime/src/libraries/System.Private.CoreLib/src/System/TimeOnly.cs

Line 317 in c9e2675

public bool Equals(TimeOnly value) => _ticks == value._ticks;

[EgorBo]

TimeOnly's one compares two fields so it's ok. DateTime compared two computed properties

[jkotas]

What does it do to perf on 32-bit platforms? (I think it is ok to take this even it makes 32-bit slower, but it would be useful to have it as a datapoint.)

[jkotas]

Just a guess - masking the top two bits instead of shifting may be significantly faster on 32-bit, and about the same on 64-bit.

[EgorBo]

What does it do to perf on 32-bit platforms? (I think it is ok to take this even it makes 32-bit slower, but it would be useful to have it as a datapoint.)

It seems it's still faster on 32bit, here is the codegen diff: https://www.diffchecker.com/vexK8vcb

|         Method |       Job |              Toolchain | IterationCount | LaunchCount | WarmupCount |         d1 |         d2 |    eq |       Mean |    Error |   StdDev |
|--------------- |---------- |----------------------- |--------------- |------------ |------------ |----------- |----------- |------ |-----------:|---------:|---------:|

// equals always return false
|     Equals_new | InProcess | InProcessEmitToolchain |        Default |     Default |     Default | MyDt[1000] | MyDt[1000] | False | 1,242.1 ns |  2.13 ns |  1.67 ns |
|     Equals_old | InProcess | InProcessEmitToolchain |        Default |     Default |     Default | MyDt[1000] | MyDt[1000] | False | 1,426.8 ns | 11.18 ns | 10.46 ns |

| Equals_new_cns | InProcess | InProcessEmitToolchain |        Default |     Default |     Default | MyDt[1000] | MyDt[1000] | False |   750.8 ns |  7.30 ns |  6.83 ns |
| Equals_old_cns | InProcess | InProcessEmitToolchain |        Default |     Default |     Default | MyDt[1000] | MyDt[1000] | False |   861.4 ns |  1.02 ns |  0.85 ns |


// equals always return true
|     Equals_new | InProcess | InProcessEmitToolchain |        Default |     Default |     Default | MyDt[1000] | MyDt[1000] |  True | 1,409.0 ns |  3.83 ns |  3.40 ns |
|     Equals_old | InProcess | InProcessEmitToolchain |        Default |     Default |     Default | MyDt[1000] | MyDt[1000] |  True | 1,605.4 ns | 13.36 ns | 11.16 ns |

| Equals_new_cns | InProcess | InProcessEmitToolchain |        Default |     Default |     Default | MyDt[1000] | MyDt[1000] |  True |   768.4 ns |  1.91 ns |  1.79 ns |
| Equals_old_cns | InProcess | InProcessEmitToolchain |        Default |     Default |     Default | MyDt[1000] | MyDt[1000] |  True |   861.5 ns |  1.81 ns |  1.60 ns |

Just a guess - masking the top two bits instead of shifting may be significantly faster on 32-bit, and about the same on 64-bit.

Looks like it hurts x64 because << 2 2 is contained and for the mask we have to request an additional reg. Diff for Equals_new_cns on x64: https://www.diffchecker.com/1pJUMY6t (<<2 is on the left)

[kunalspathak]

Self-contained benchmark: https://gist.github.com/EgorBo/e5642d9c2d362bb53880d3c9c87e5b95

Can you add this to dotnet/performance ?

[EgorBo]

Self-contained benchmark: https://gist.github.com/EgorBo/e5642d9c2d362bb53880d3c9c87e5b95

Can you add this to dotnet/performance ?

Sure, it seems we don't cover these APIs there