Optimize Math.Round for arm64 and AwayFromZero for x64

[EgorBo]

I noticed that people quite often explicitly set AwayFromZero rounding for Math.Round (btw in C++ round(double) uses AwayFromZero by default while we use "accountant-friendly" ToEven) to remind you the difference:

	1.5	2.5
ToEven	2.0	2.0
AwayFromZero	2.0	3.0

so I decided to optimize it.

double Test(double x) => Math.Round(x, MidpointRounding.AwayFromZero);

Current codegen:

; Method Program:Test(double):double:this
            stp     fp, lr, [sp,#-16]!
            mov     fp, sp

            mov     w0, #0
            mov     w1, #1
            ldp     fp, lr, [sp],#16
            b       System.Math:Round(double,int,int):double ;; call, much slower
; Total bytes of code: 24

New codegen:

; Method Program:Test(double):double:this
            stp     fp, lr, [sp,#-16]!
            mov     fp, sp

            frinta  d0, d0

            ldp     fp, lr, [sp],#16
            ret     lr
; Total bytes of code: 20

cc @tannergooding

PS: I can guard my change with RuntimeHelpers.IsConstArgument but I think it worth inlining as is.

[ghost]

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

Issue Details

---

I noticed that people quite often explicitly set AwayFromZero rounding for Math.Round (btw in C++ round(double) uses AwayFromZero by default while we use "accountant-friendly" ToEven) so I decided to optimize it.

double Test(double x) => Math.Round(x, MidpointRounding.AwayFromZero);

Current codegen:

       vzeroupper 
       vmovaps  xmm0, xmm1
       xor      edx, edx
       mov      r8d, 1
       jmp      System.Math:Round(double,int,int):double ;; not inlined, is not hw accelerated for AwayFromZero
; Total bytes of code: 20

New codegen:

       vzeroupper 
       vandpd   xmm0, xmm1, xmmword ptr [reloc @RWD00]
       vorpd    xmm0, xmm0, xmmword ptr [reloc @RWD16]
       vaddpd   xmm0, xmm0, xmm1
       vroundsd xmm0, xmm0, xmm0, 3
       ret      
RWD00  	dq	8000000000000000h, 8000000000000000h
RWD16  	dq	3FDFFFFFFFFFFFFFh, 3FDFFFFFFFFFFFFFh
; Total bytes of code: 30

On ARM64 it's a single instruction for both modes.
cc @tannergooding

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

[stephentoub]

I noticed that people quite often explicitly set AwayFromZero rounding for Math.Round

Can you share your data source?

[stephentoub]

Src looks good to me. Please ensure there are tests that will exercise the new code paths. Thanks.

[EgorBo]

Failed job is unrelated.

@tannergooding are the tests I added sufficient? Also, it seems like Math.Round(var = 0.49999999999999994) returns 1 on ARM32 😐

[EgorBo]

@tannergooding ping 🙂