Math.Round[F](x, MidpointRounding) doesn't use VROUNDS[SD] outside of "to even"

[colejohnson66]

Description

It appears that Math[F].Round doesn't always use VROUNDS[SD] with an embedded round control. It does use VROUNDS[SD] for "to even" as that follows MXCSR

The "round control" bits of VROUNDS[SD] are as follows:

• 00: nearest/to even
• 01: down/to negative infinity
• 10: up/to positive infinity
• 11: truncate/to zero

Bit three of the immediate is set to use MXCSR.RC and cleared to use the embedded round control bits. It looks like the JIT only uses VROUND[SD] with "use MXCSR" set.

Data

Expected behavior occurs for "to even" rounding, but with the "use MXCSR" bit set - hence the suspicion. In this case, no harm, no foul.
https://sharplab.io/#v2:EYLgxg9gTgpgtADwGwBYA0AXEBDAzgWwB8ABAJgEYBYAKBuIGYACM58gdhoG8bHfmmAJhACuwADYxGAFRi4MACiGiJjAG7YxASkYBeAHw8+RgLLYMACwB0AJREA7AfPVi0jYwEsBABwju7GW2EHPwBzSykIAFFVGDtNAG4aAF8gA

using System;

public class Program
{
    public double Test(double val) =>
        Math.Round(val, MidpointRounding.ToEven);
}

Program.Test(Double)
    L0000: vzeroupper
    L0003: vroundsd xmm0, xmm0, xmm1, 4
    L0009: ret

However, using a mode that is not set in MXCSR calls into the framework for help instead of using an embedding rounding mode. For the below, VROUNDSD should be used with an immediate of b010 (ignore MXCSR, and use "round up/to infinity").

https://sharplab.io/#v2:EYLgxg9gTgpgtADwGwBYA0AXEBDAzgWwB8ABAJgEYBYAKBuIGYACM58gdhoG8bHfmmAJhACuwADYxGAFRi4MACiGiJjAG7YxASkYBeAHw8+RgLLYMACwB0AJREA7AfPVi0jYwEsBABwju7GW2EHPwBzSykIAAUIXHcMd1UYAEk7ADM/OIBPTQBuGgBfIA===

using System;

public class Program
{
    public double Test(double val) =>
        Math.Round(val, MidpointRounding.ToPositiveInfinity);
}

Program.Test(Double)
    L0000: vzeroupper
    L0003: vmovaps xmm0, xmm1
    L0007: xor edx, edx
    L0009: mov r8d, 4
    L000f: mov rax, 0x7ff7e0257618
    L0019: jmp qword ptr [rax]

Expected behavior for this would be:

Program.Test(Double)
    L0000: vzeroupper
    L0003: vroundsd xmm0, xmm0, xmm1, 2
    L0009: ret

[ghost]

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

Issue Details

---

Description

It appears that Math[F].Round doesn't always use VROUNDS[SD] with an embedded round control. It does use VROUNDS[SD] for "to even" as that follows MXCSR

The "round control" bits of VROUNDS[SD] are as follows:

• 00: nearest/to even
• 01: down/to negative infinity
• 10: up/to positive infinity
• 11: truncate/to zero

Bit three of the immediate is set to use MXCSR.RC and cleared to use the embedded round control bits. It looks like the JIT only uses VROUND[SD] with "use MXCSR" set.

Data

Expected behavior occurs for "to even" rounding, but with the "use MXCSR" bit set instead of "use embedded with 00", hence the suspicion. In this case, no harm, no foul.
https://sharplab.io/#v2:EYLgxg9gTgpgtADwGwBYA0AXEBDAzgWwB8ABAJgEYBYAKBuIGYACM58gdhoG8bHfmmAJhACuwADYxGAFRi4MACiGiJjAG7YxASkYBeAHw8+RgLLYMACwB0AJREA7AfPVi0jYwEsBABwju7GW2EHPwBzSykIAFFVGDtNAG4aAF8gA

using System;
public class Program
{
    public double Test(double val) =>
        Math.Round(val, MidpointRounding.ToEven);
}

Program.Test(Double)
    L0000: vzeroupper
    L0003: vroundsd xmm0, xmm0, xmm1, 4
    L0009: ret

However, using a mode that is not set in MXCSR calls into the framework for help instead of using an embedding rounding mode. For the below, VROUNDSD should be used with an immediate of b010 (ignore MXCSR, and use "round up/to infinity").

https://sharplab.io/#v2:EYLgxg9gTgpgtADwGwBYA0AXEBDAzgWwB8ABAJgEYBYAKBuIGYACM58gdhoG8bHfmmAJhACuwADYxGAFRi4MACiGiJjAG7YxASkYBeAHw8+RgLLYMACwB0AJREA7AfPVi0jYwEsBABwju7GW2EHPwBzSykIAAUIXHcMd1UYAEk7ADM/OIBPTQBuGgBfIA===

using System;
public class Program
{
    public double Test(double val) =>
        Math.Round(val, MidpointRounding.ToPositiveInfinity);
}

Program.Test(Double)
    L0000: vzeroupper
    L0003: vmovaps xmm0, xmm1
    L0007: xor edx, edx
    L0009: mov r8d, 4
    L000f: mov rax, 0x7ff7e0257618
    L0019: jmp qword ptr [rax]

Author:	colejohnson66
Assignees:	-
Labels:	area-System.Numerics, tenet-performance, untriaged
Milestone:	-

[tannergooding]

It's typically preferred to just use the simpler dedicated API instead:

• nearest/to even - Round
• down/to negative infinity - Floor
• up/to positive infinity - Ceiling
• truncate/to zero - Truncate

We can of course optimize the explicit enum modes as well, but its not as high of a priority.

[colejohnson66]

I do see that the dedicated methods use the embedded round control. For example, Ceiling uses b1010 (inexact exception suppression, to positive infinity).

Would it be preferable to just have the enum variant forward to the other methods like it already does for "to even"? I personally prefer Round with an explicit direction, as "to positive infinity" is more clear than "ceiling" when talking about negative values.

[tannergooding]

If you'd like to put up a PR, then feel free.

[MichalPetryka]

Would it be preferable to just have the enum variant forward to the other methods like it already does for "to even"? I personally prefer Round with an explicit direction, as "to positive infinity" is more clear than "ceiling" when talking about negative values.

The issue with that was that iirc those methods forward to the enum ones on platforms that aren't accelerated.
EDIT: I guess it'd be enough then to just make it inlineable then.