Allow the vector and scalar code paths for min, max, and fma APIs to share logic

[tannergooding]

This fixes #116803
This fixes #115381

[dotnet-policy-service]

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

[Copilot]

Pull Request Overview

This PR consolidates the vector and scalar code paths for math intrinsics (min, max, and FMA) by sharing common logic across different platforms and updating the intrinsic lowering functions accordingly. Key changes include:

• Removal of duplicate intrinsic case blocks in ValueNumStore and selective compilation using TARGET_RISCV64.
• Renaming and refactoring of the FMA lowering method from LowerFusedMultiplyAdd to LowerFusedMultiplyOp.
• Updates to intrinsic lists and helper APIs to support unified handling for min/max variants.

Reviewed Changes

Copilot reviewed 11 out of 13 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
src/coreclr/jit/valuenum.cpp	Removed redundant intrinsic cases for max/min and encapsulated platform‑specific handling.
src/coreclr/jit/lowerxarch.cpp	Renamed and refactored the FMA lowering routine to share logic among different intrinsics.
src/coreclr/jit/namedintrinsiclist.h	Added new intrinsic entries (e.g. NI_System_Math_MaxNative/MinNative) for unified API support.
src/coreclr/jit/hwintrinsiclistxarch.h	Updated intrinsic declarations to include new min/max variants.
src/coreclr/jit/compiler.h	Consolidated helper API signatures for SIMD min/max nodes into unified functions.
Other files	Adjusted intrinsic processing and lowering logic across xarch, arm64, and codegen modules.

Comments suppressed due to low confidence (3)

src/coreclr/jit/lowerxarch.cpp:1386

• Renaming the intrinsic lowering method to LowerFusedMultiplyOp improves clarity; please ensure that all call sites and related documentation are updated accordingly.

void Lowering::LowerFusedMultiplyOp(GenTreeHWIntrinsic* node)

src/coreclr/jit/compiler.h:3338

• With the introduction of unified APIs for min/max intrinsics, please confirm that all consumers of the previous gtNewSimdMaxNode/gtNewSimdMinNativeNode functions have been updated to use the new unified APIs.

                                 bool        isNumber);

[tannergooding]

CC. @dotnet/jit-contrib, @jakobbotsch, @kunalspathak for review.

diffs are very positive. We see -77.3k bytes for windows arm64 in full opts due to the significantly more compact codegen as compared to the naive inlined version. There is a +11.6k in minopts regression, which is due to us having expanded intrinsics instead of calls.

The diffs for x64 are much smaller since we were already accelerating most scalar scenarios already. There are some opportunities to improve the codegen more, but it would involve more work than what this did, which was mostly centralizing the existing logic.

Throughput ranges from a slight improvement (-0.07% on the high end for Windows Arm64 full opts) to a slight regression (+0.04% on the high end for Windows x64 full opts)

[kunalspathak]

do you know why linux/x64 is also seeing regression?

[tannergooding]

do you know why linux/x64 is also seeing regression?

Looks like it generated code for AVX2 rather than AVX512 (likely just a difference in Azure machine allocation):

; Windows 64 - System.MathBenchmarks.Single:MaxTest() (Instrumented Tier0)
        vaddsd   xmm0, xmm0, xmm2
+       vmovaps  xmm5, xmm3
+       vmovaps  xmm16, xmm0
+       vrangesd xmm17, xmm5, xmm16, 5
+       vfixupimmsd xmm5, xmm16, xmm4, 0
+       vfixupimmsd xmm17, xmm5, xmm4, 0
        vaddsd   xmm6, xmm17, xmm6

; Linux x64 - System.MathBenchmarks.Single:MinTest() (FullOpts)
+       vsubss   xmm1, xmm1, xmm2
+       vmovaps  xmm4, xmm3
+       vmovaps  xmm5, xmm1
+       vcmpps   xmm6, xmm4, xmm5, 0
+       vxorps   xmm7, xmm7, xmm7
+       vpcmpgtd xmm7, xmm7, xmm4
+       vandps   xmm6, xmm7, xmm6
+       vcmpps   xmm7, xmm4, xmm4, 4
+       vorps    xmm6, xmm7, xmm6
+       vcmpps   xmm7, xmm4, xmm5, 1
+       vorps    xmm6, xmm7, xmm6
+       vblendvps xmm4, xmm5, xmm4, xmm6
+       vaddss   xmm0, xmm4, xmm0

The code is still much better than the previous pattern, it's just larger than the naive code with 4+ branches. This also means more places are inlined, where-as previously they were calls.

[kunalspathak]

/azp run Antigen, Fuzzlyn

[azure-pipelines]

Azure Pipelines successfully started running 2 pipeline(s).

[kunalspathak]

LGTM as long as Antigen/Fuzzlyn is green.

[tannergooding]

/azp run Antigen, Fuzzlyn

[azure-pipelines]

Azure Pipelines successfully started running 2 pipeline(s).