Math/MathF.Truncate isn't an intrinsic and results in inefficient codegen

[MichalPetryka]

Description

Math/MathF.Truncate results in bad codegen that calls into native modf which is pretty slow. Instead, it should be an intrinsic for vroundsd/vroundss like Ceiling and Floor are (and clang also does do it for truncate).

Configuration

Sharplab Core CLR 5.0.721.25508 on amd64

Regression?

No idea.

Data

Sharplab for Math
Sharplab for MathF
Godbolt clang for double
Godbolt clang for float

[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

Math/MathF.Truncate results in bad codegen that calls into native modf which is pretty slow. Instead, it should be an intrinsic for vroundsd/vroundss like Ceiling and Floor are (and clang also does do it for truncate).

Configuration

Sharplab Core CLR 5.0.721.25508 on amd64

Regression?

No idea.

Data

Sharplab for Math
Sharplab for MathF
Godbolt clang for double
Godbolt clang for float

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

[tannergooding]

This should be a relatively simple fix but is going to be for 7.0.0 at the earliest.

[danmoseley]

@MichalPetryka any interest in offering a PR?

[MichalPetryka]

I've never worked with the JIT so I'd prefer somebody else to do it since I don't really know how sadly.

[danmoseley]

Fair enough. Still, if you have an interest I expect @tannergooding could give pointers. 😄

[AraHaan]

Sure thing, here you go: char*, int*, uint*, void*, short*, ushort*, long*, ulong*, float*.

[tannergooding]

This would be similar to the handling that exists for Ceiling, Floor, and Round.

Basically if you search for NI_System_Math_Ceiling, you'll find a number of entries and for the most part its just adding an additional entry to cover NI_System_Math_Truncate. You'd need to also ensure the managed side API is marked with [Intrinsic].

Basically the only places that aren't adding in a new nearly identical path for NI_System_Math_Truncate is in codegenarmarch.cpp and codegenxarch.cpp where you need to pick the right instruction and the right constant for the rounding mode, respectively.

It should be relatively straightforward overall and I'm happy to help walk you through it over DMs, the C# community discord (https://aka.ms/csharp-discord, specifically the #lowlevel channel), or more ad-hoc on the issue here.