Subtract-variant of `VectorXXX.MultiplyAddEstimate` does not light up for negated constant local

[TechPizzaDev]

Description

public static Vector128<double> BadCode(Vector128<double> a) 
{
    Vector128<double> half = Vector128.Create(0.5);
    return Vector128.MultiplyAddEstimate(a, Vector128.Create(2.0), -half) * half;
}

public static Vector128<double> GoodCode(Vector128<double> a, Vector128<double> half) 
{
    // Fma.MultiplyAdd lits up too!
    return Vector128.MultiplyAddEstimate(a, Vector128.Create(2.0), -half) * half;
}

Regression?

No

Data

// coreclr trunk-20251102+e8812e7419db9137f20b990786a53ed71e27e11e

C:BadCode(System.Runtime.Intrinsics.Vector128`1[double]):System.Runtime.Intrinsics.Vector128`1[double] (FullOpts):
       vmovddup xmm0, qword ptr [reloc @RWD00]
       vmovddup xmm1, qword ptr [reloc @RWD08]
       vmovaps  xmm2, xmmword ptr [rsp+0x08]
       vfmadd213pd xmm2, xmm1, xmmword ptr [reloc @RWD16]
       vmulpd   xmm0, xmm2, xmm0
       vmovups  xmmword ptr [rdi], xmm0
       mov      rax, rdi
       ret      
RWD00  	dq	3FE0000000000000h
RWD08  	dq	4000000000000000h
RWD16  	dq	BFE0000000000000h, BFE0000000000000h

C:GoodCode(System.Runtime.Intrinsics.Vector128`1[double],System.Runtime.Intrinsics.Vector128`1[double]):System.Runtime.Intrinsics.Vector128`1[double] (FullOpts):
       vmovups  xmm0, xmmword ptr [rsp+0x18]
       vmovaps  xmm1, xmmword ptr [rsp+0x08]
       vfmsub132pd xmm1, xmm0, xmmword ptr [reloc @RWD00]
       vmulpd   xmm0, xmm1, xmm0
       vmovups  xmmword ptr [rdi], xmm0
       mov      rax, rdi
       ret      
RWD00  	dq	4000000000000000h, 4000000000000000h

Analysis

Turning locals into constant data seems to take priority over negated FMA.

[dotnet-policy-service]

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

[tannergooding]

The negated constant local case isn't "lighting up" because the constant itself is already negative and x + (-y) is the same as x - y.

There's potentially some more complex analysis that could be done to recognize +/- constant pairs or normalize certain -constants into NEG(PositiveCns), but it's not clear that will actually be worth the expense required.

Someone is more than welcome to investigate, but this will likely remain "low priority".

[TechPizzaDev]

This seems to occur quite often for non-constant locals as well. I just hit a case where ten calls to my custom lerp turned into nine pairs of vxorps vfmadd213ss and only one vfnmadd213ss.

This may suggest that there is an underlying problem that affects "constants" too.

Small repro that negates left param and lights up partially.
Negating right param lights up in this example, but may not after inlining.

static float lerp(float t, float a, float b) => float.MultiplyAddEstimate(t, b, float.MultiplyAddEstimate(-t, a, a));
static float lerp2(float d, float e, float f, float g, float h, float i) => lerp(e, lerp(d, f, g), lerp(d, h, i));
static float lerp3(float d, float e, float f, float g, float h, float i, float j, float k, float l, float m, float n) => lerp(f, lerp2(d, e, g, h, i, j), lerp2(d, e, k, l, m, n));

[tannergooding]

That looks like likely a case of CSE, where the copy is instead hoisted and decided to be more profitable to optimize.

Will take a look in a bit.