Speed up Matrix4x4.CreateScale

[SingleAccretion]

In #76491 the Jit became a little more eager to optimize code using vectors as a whole. Unfortunately, this regressed Matrix4x4.CreateScale on some platforms because code for it is written in a way that favors promotion over whole-register access.

This change fixes the regression by using style that is more in line with what the Jit expects of vectors. It also optimizes related overloads by avoiding the relatively expensive get_Identity property call (that is not even inlined due to loading a value type field).

Linux x64 benchmark results:

Method	Job	Toolchain	Mean	Error	StdDev	Median	Min	Max	Ratio	RatioSD	Allocated	Alloc Ratio
CreateScaleFromVectorBenchmark	Job-AFMBXB	Root/corerun	12.575 ns	0.1031 ns	0.0861 ns	12.555 ns	12.452 ns	12.770 ns	0.74	0.01	-	NA
CreateScaleFromVectorBenchmark	Job-KPUHXS	/crb/corerun	16.988 ns	0.1283 ns	0.1138 ns	16.959 ns	16.789 ns	17.189 ns	1.00	0.00	-	NA
CreateScaleFromVectorWithCenterBenchmark	Job-AFMBXB	Root/corerun	9.548 ns	0.0628 ns	0.0588 ns	9.529 ns	9.476 ns	9.663 ns	0.44	0.02	-	NA
CreateScaleFromVectorWithCenterBenchmark	Job-KPUHXS	/crb/corerun	21.963 ns	0.9062 ns	1.0435 ns	21.291 ns	21.155 ns	24.310 ns	1.00	0.00	-	NA
CreateScaleFromScalarBenchmark	Job-AFMBXB	Root/corerun	9.321 ns	0.0701 ns	0.0655 ns	9.329 ns	9.207 ns	9.442 ns	0.54	0.00	-	NA
CreateScaleFromScalarBenchmark	Job-KPUHXS	/crb/corerun	17.138 ns	0.0886 ns	0.0828 ns	17.154 ns	16.954 ns	17.255 ns	1.00	0.00	-	NA
CreateScaleFromScalarWithCenterBenchmark	Job-AFMBXB	Root/corerun	12.807 ns	0.1044 ns	0.0977 ns	12.789 ns	12.689 ns	12.999 ns	0.74	0.01	-	NA
CreateScaleFromScalarWithCenterBenchmark	Job-KPUHXS	/crb/corerun	17.203 ns	0.0963 ns	0.0901 ns	17.216 ns	16.980 ns	17.342 ns	1.00	0.00	-	NA
CreateScaleFromScalarXYZBenchmark	Job-AFMBXB	Root/corerun	9.454 ns	0.0688 ns	0.0609 ns	9.433 ns	9.389 ns	9.577 ns	0.46	0.00	-	NA
CreateScaleFromScalarXYZBenchmark	Job-KPUHXS	/crb/corerun	20.481 ns	0.1567 ns	0.1466 ns	20.431 ns	20.296 ns	20.769 ns	1.00	0.00	-	NA
CreateScaleFromScalarXYZWithCenterBenchmark	Job-AFMBXB	Root/corerun	9.434 ns	0.0380 ns	0.0297 ns	9.438 ns	9.384 ns	9.477 ns	0.55	0.00	-	NA
CreateScaleFromScalarXYZWithCenterBenchmark	Job-KPUHXS	/crb/corerun	17.160 ns	0.1791 ns	0.1588 ns	17.171 ns	16.901 ns	17.433 ns	1.00	0.00	-	NA

Where Root/corerun is this change and /crb/corerun is main.

Fixes #78977.

[ghost]

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

Issue Details

---

In #76491 the Jit became a little more eager to optimize code using vectors as a whole. Unfortunately, this regressed Matrix4x4.CreateScale on some platforms because code for it is written in a way that favors promotion over whole-register access.

This change fixes the regression by using style that is more in line with what the Jit expects of vectors. It also optimizes related overloads by avoiding the relatively expensive get_Identity property call (that is not even inlined due to loading a value type field).

Linux x64 benchmark results:

Method	Job	Toolchain	Mean	Error	StdDev	Median	Min	Max	Ratio	RatioSD	Allocated	Alloc Ratio
CreateScaleFromVectorBenchmark	Job-AFMBXB	Root/corerun	12.575 ns	0.1031 ns	0.0861 ns	12.555 ns	12.452 ns	12.770 ns	0.74	0.01	-	NA
CreateScaleFromVectorBenchmark	Job-KPUHXS	/crb/corerun	16.988 ns	0.1283 ns	0.1138 ns	16.959 ns	16.789 ns	17.189 ns	1.00	0.00	-	NA
CreateScaleFromVectorWithCenterBenchmark	Job-AFMBXB	Root/corerun	9.548 ns	0.0628 ns	0.0588 ns	9.529 ns	9.476 ns	9.663 ns	0.44	0.02	-	NA
CreateScaleFromVectorWithCenterBenchmark	Job-KPUHXS	/crb/corerun	21.963 ns	0.9062 ns	1.0435 ns	21.291 ns	21.155 ns	24.310 ns	1.00	0.00	-	NA
CreateScaleFromScalarBenchmark	Job-AFMBXB	Root/corerun	9.321 ns	0.0701 ns	0.0655 ns	9.329 ns	9.207 ns	9.442 ns	0.54	0.00	-	NA
CreateScaleFromScalarBenchmark	Job-KPUHXS	/crb/corerun	17.138 ns	0.0886 ns	0.0828 ns	17.154 ns	16.954 ns	17.255 ns	1.00	0.00	-	NA
CreateScaleFromScalarWithCenterBenchmark	Job-AFMBXB	Root/corerun	12.807 ns	0.1044 ns	0.0977 ns	12.789 ns	12.689 ns	12.999 ns	0.74	0.01	-	NA
CreateScaleFromScalarWithCenterBenchmark	Job-KPUHXS	/crb/corerun	17.203 ns	0.0963 ns	0.0901 ns	17.216 ns	16.980 ns	17.342 ns	1.00	0.00	-	NA
CreateScaleFromScalarXYZBenchmark	Job-AFMBXB	Root/corerun	9.454 ns	0.0688 ns	0.0609 ns	9.433 ns	9.389 ns	9.577 ns	0.46	0.00	-	NA
CreateScaleFromScalarXYZBenchmark	Job-KPUHXS	/crb/corerun	20.481 ns	0.1567 ns	0.1466 ns	20.431 ns	20.296 ns	20.769 ns	1.00	0.00	-	NA
CreateScaleFromScalarXYZWithCenterBenchmark	Job-AFMBXB	Root/corerun	9.434 ns	0.0380 ns	0.0297 ns	9.438 ns	9.384 ns	9.477 ns	0.55	0.00	-	NA
CreateScaleFromScalarXYZWithCenterBenchmark	Job-KPUHXS	/crb/corerun	17.160 ns	0.1791 ns	0.1588 ns	17.171 ns	16.901 ns	17.433 ns	1.00	0.00	-	NA

Where Root/corerun is this change and /crb/corerun is main.

Fixes #78977.

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

[ghost]

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

Issue Details

---

In #76491 the Jit became a little more eager to optimize code using vectors as a whole. Unfortunately, this regressed Matrix4x4.CreateScale on some platforms because code for it is written in a way that favors promotion over whole-register access.

This change fixes the regression by using style that is more in line with what the Jit expects of vectors. It also optimizes related overloads by avoiding the relatively expensive get_Identity property call (that is not even inlined due to loading a value type field).

Linux x64 benchmark results:

Method	Job	Toolchain	Mean	Error	StdDev	Median	Min	Max	Ratio	RatioSD	Allocated	Alloc Ratio
CreateScaleFromVectorBenchmark	Job-AFMBXB	Root/corerun	12.575 ns	0.1031 ns	0.0861 ns	12.555 ns	12.452 ns	12.770 ns	0.74	0.01	-	NA
CreateScaleFromVectorBenchmark	Job-KPUHXS	/crb/corerun	16.988 ns	0.1283 ns	0.1138 ns	16.959 ns	16.789 ns	17.189 ns	1.00	0.00	-	NA
CreateScaleFromVectorWithCenterBenchmark	Job-AFMBXB	Root/corerun	9.548 ns	0.0628 ns	0.0588 ns	9.529 ns	9.476 ns	9.663 ns	0.44	0.02	-	NA
CreateScaleFromVectorWithCenterBenchmark	Job-KPUHXS	/crb/corerun	21.963 ns	0.9062 ns	1.0435 ns	21.291 ns	21.155 ns	24.310 ns	1.00	0.00	-	NA
CreateScaleFromScalarBenchmark	Job-AFMBXB	Root/corerun	9.321 ns	0.0701 ns	0.0655 ns	9.329 ns	9.207 ns	9.442 ns	0.54	0.00	-	NA
CreateScaleFromScalarBenchmark	Job-KPUHXS	/crb/corerun	17.138 ns	0.0886 ns	0.0828 ns	17.154 ns	16.954 ns	17.255 ns	1.00	0.00	-	NA
CreateScaleFromScalarWithCenterBenchmark	Job-AFMBXB	Root/corerun	12.807 ns	0.1044 ns	0.0977 ns	12.789 ns	12.689 ns	12.999 ns	0.74	0.01	-	NA
CreateScaleFromScalarWithCenterBenchmark	Job-KPUHXS	/crb/corerun	17.203 ns	0.0963 ns	0.0901 ns	17.216 ns	16.980 ns	17.342 ns	1.00	0.00	-	NA
CreateScaleFromScalarXYZBenchmark	Job-AFMBXB	Root/corerun	9.454 ns	0.0688 ns	0.0609 ns	9.433 ns	9.389 ns	9.577 ns	0.46	0.00	-	NA
CreateScaleFromScalarXYZBenchmark	Job-KPUHXS	/crb/corerun	20.481 ns	0.1567 ns	0.1466 ns	20.431 ns	20.296 ns	20.769 ns	1.00	0.00	-	NA
CreateScaleFromScalarXYZWithCenterBenchmark	Job-AFMBXB	Root/corerun	9.434 ns	0.0380 ns	0.0297 ns	9.438 ns	9.384 ns	9.477 ns	0.55	0.00	-	NA
CreateScaleFromScalarXYZWithCenterBenchmark	Job-KPUHXS	/crb/corerun	17.160 ns	0.1791 ns	0.1588 ns	17.171 ns	16.901 ns	17.433 ns	1.00	0.00	-	NA

Where Root/corerun is this change and /crb/corerun is main.

Fixes #78977.

Author:	SingleAccretion
Assignees:	-
Labels:	area-System.Numerics, community-contribution
Milestone:	-

[EgorBo]

Thanks for the workaround!