Improve handling of reused constants in the register allocator

[tannergooding]

Today https://github.com/dotnet/runtime/blob/main/docs/design/coreclr/jit/lsra-heuristic-tuning.md documents that it does FREE and then CONST_AVAILABLE. This ends up not being able to reuse an already enregistered constant if it wasn't "last use".

A simple example is the following where C81 and C82 both represent the CNS_DBL of 0.0:

68.#46  C81   Def    ORDER(A) mm0  |     |V13 a|     |V2  a|V3  a|V0  a|V1  a|     |     |V10 a|V4  a|V7  a|C81 a|     |     |V101a|     |
70.#47  C82   Def    ORDER(A) mm1  |     |V13 a|     |V2  a|V3  a|V0  a|V1  a|     |     |V10 a|V4  a|V7  a|C81 a|C82 a|     |V101a|     |

While these values aren't last use, the last use does happen before the constant is overwritten and in some cases we can end up with 4-5 registers all being initialized (as is the case for Matrix4x4.Decompose) to the same constant value (often zero so emitting xorps many times in a row).

It would be beneficial if the register allocator could improve this scenario and keep the constant in one register for such scenarios. This in particular appears to impact Zero and AllBitsSet since they are "cheap to compute" and so typically don't undergo CSE.

category:cq
theme:register-allocator

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Today https://github.com/dotnet/runtime/blob/main/docs/design/coreclr/jit/lsra-heuristic-tuning.md documents that it does FREE and then CONST_AVAILABLE. This ends up not being able to reuse an already enregistered constant if it wasn't "last use".

A simple example is the following where C81 and C82 both represent the CNS_DBL of 0.0:

68.#46  C81   Def    ORDER(A) mm0  |     |V13 a|     |V2  a|V3  a|V0  a|V1  a|     |     |V10 a|V4  a|V7  a|C81 a|     |     |V101a|     |
70.#47  C82   Def    ORDER(A) mm1  |     |V13 a|     |V2  a|V3  a|V0  a|V1  a|     |     |V10 a|V4  a|V7  a|C81 a|C82 a|     |V101a|     |

While these values aren't last use, the last use does happen before the constant is overwritten and in some cases we can end up with 4-5 registers all being initialized to the same constant value (often zero so emitting xorps many times in a row).

It would be beneficial if the register allocator could improve this scenario and keep the constant in one register for such scenarios. This in particular appears to impact Zero and AllBitsSet since they are "cheap to compute" and so typically don't undergo CSE.

Author:	tannergooding
Assignees:	-
Labels:	area-CodeGen-coreclr
Milestone:	-

[kunalspathak]

There is an assumption in LSRA that materializing constants are cheap and hence we don't want to tie it to a register. May be if we have an information of the occurances of particular constants (zeros for sure), we can have it assigned a register. I tried experimenting by switching the order of CONST_AVAILABLE and FREE and that (not surprisingly) shows lot of regressions:

[15:20:52]
[15:20:52] Top method regressions (bytes):
[15:20:52]           54 (20.00% of base) : 36178.dasm - System.Numerics.Tests.Perf_Matrix4x4:CreateConstrainedBillboardBenchmark():System.Numerics.Matrix4x4:this
[15:20:52]           43 ( 2.55% of base) : 34909.dasm - System.Numerics.Matrix4x4:Decompose(System.Numerics.Matrix4x4,byref,byref,byref):bool
[15:20:52]           28 (19.18% of base) : 10260.dasm - System.Numerics.Tests.Perf_Matrix4x4:CreateLookAtBenchmark():System.Numerics.Matrix4x4:this
[15:20:52]           28 (19.18% of base) : 17977.dasm - System.Numerics.Tests.Perf_Matrix4x4:CreateWorldBenchmark():System.Numerics.Matrix4x4:this
[15:20:52]           26 (13.27% of base) : 35808.dasm - System.Numerics.Tests.Perf_Matrix4x4:CreateBillboardBenchmark():System.Numerics.Matrix4x4:this
[15:20:52]           22 ( 6.83% of base) : 17425.dasm - System.Numerics.Tests.Perf_Plane:CreateFromVerticesBenchmark():System.Numerics.Plane:this
[15:20:52]           18 ( 1.91% of base) : 3355.dasm - Benchmarks.SIMD.RayTracer.RayTracer:CreateDefaultScene():Benchmarks.SIMD.RayTracer.Scene
[15:20:52]           10 ( 4.44% of base) : 33917.dasm - Benchstone.MDBenchI.MDXposMatrix:Test():bool:this
[15:20:52]            5 ( 0.38% of base) : 15958.dasm - Benchstone.MDBenchF.MDRomber:Test():bool:this
[15:20:52]            5 ( 1.61% of base) : 16620.dasm - Benchstone.MDBenchF.MDSqMtx:Test():bool:this
[15:20:52]
[15:20:52] Top method regressions (percentages):
[15:20:52]           54 (20.00% of base) : 36178.dasm - System.Numerics.Tests.Perf_Matrix4x4:CreateConstrainedBillboardBenchmark():System.Numerics.Matrix4x4:this
[15:20:52]           28 (19.18% of base) : 10260.dasm - System.Numerics.Tests.Perf_Matrix4x4:CreateLookAtBenchmark():System.Numerics.Matrix4x4:this
[15:20:52]           28 (19.18% of base) : 17977.dasm - System.Numerics.Tests.Perf_Matrix4x4:CreateWorldBenchmark():System.Numerics.Matrix4x4:this
[15:20:52]           26 (13.27% of base) : 35808.dasm - System.Numerics.Tests.Perf_Matrix4x4:CreateBillboardBenchmark():System.Numerics.Matrix4x4:this
[15:20:52]           22 ( 6.83% of base) : 17425.dasm - System.Numerics.Tests.Perf_Plane:CreateFromVerticesBenchmark():System.Numerics.Plane:this
[15:20:52]           10 ( 4.44% of base) : 33917.dasm - Benchstone.MDBenchI.MDXposMatrix:Test():bool:this
[15:20:52]           43 ( 2.55% of base) : 34909.dasm - System.Numerics.Matrix4x4:Decompose(System.Numerics.Matrix4x4,byref,byref,byref):bool
[15:20:52]           18 ( 1.91% of base) : 3355.dasm - Benchmarks.SIMD.RayTracer.RayTracer:CreateDefaultScene():Benchmarks.SIMD.RayTracer.Scene
[15:20:52]            5 ( 1.61% of base) : 16620.dasm - Benchstone.MDBenchF.MDSqMtx:Test():bool:this
[15:20:52]            5 ( 0.38% of base) : 15958.dasm - Benchstone.MDBenchF.MDRomber:Test():bool:this
[15:20:52]
[15:20:52] 10 total methods with Code Size differences (0 improved, 10 regressed), 0 unchanged.

windows/x64 coreclr:

[kunalspathak]

This will most likely not happen in .NET 7