Update fgMorphHWIntrinsic to more closely follow fgMorphSmpOp

[tannergooding]

This updates fgMorphHWIntrinsic to have many of the same arithmetic and logical transforms as fgMorphSmpOp and more closely matches the conditions under which those transforms are done by the latter. This results in a some decent diffs due to other transformations that are unlocked.

As part of this, I also ensured that fgMorphSmpOp was lighting up certain optimizations for floating-point where it is safe to do. For example, the following transforms are IEEE 754 compliant (many of these are only done when one input is a constant as the transform isn't beneficial otherwise):

• a - b becoming a + (-b)
• (-a) - (-b) becoming b - a
• -(a * b) becoming a * (-b)
• (-a) * b becoming a * (-b)
• -(a / b) becoming a / (-b)
• (-a) / b becoming a / (-b)

Most importantly it covers handling ~(a cmp b) and transforming it to a cmp* b; that is, it reverses the comparison. This allows significantly smaller codegen around various branches and allows other optimizations to better light up.

The few number of codegen regressions are in T0 code where lowering ends up placing a value in a local for reuse and this causes a spill/reload of the value on the stack.

[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 enhances fgMorphHWIntrinsic by introducing an optional “effective operation” mode in GetOperForHWIntrinsicId—bringing its behavior in line with fgMorphSmpOp—and splits the morphing logic into required and optional paths in the compiler.

• Changed the GetOperForHWIntrinsicId signature in gentree.h and implemented the getEffectiveOp logic in gentree.cpp.
• Added fgMorphHWIntrinsicRequired and fgMorphHWIntrinsicOptional declarations in compiler.h.
• Removed the old single-parameter overload for GetOperForHWIntrinsicId.

Reviewed Changes

Copilot reviewed 3 out of 4 changed files in this pull request and generated no comments.

File	Description
src/coreclr/jit/gentree.h	Updated GetOperForHWIntrinsicId signature and removed the old overload
src/coreclr/jit/gentree.cpp	Added getEffectiveOp handling for GT_SUB and GT_XOR patterns
src/coreclr/jit/compiler.h	Declared new fgMorphHWIntrinsicRequired and fgMorphHWIntrinsicOptional methods

Comments suppressed due to low confidence (2)

[kunalspathak]

comment not related to this PR in particular. But I wish we had infrastructure to pattern match and replace this in morph....currently we are just adding this manually and making it more bug prone.

[kunalspathak]

Overall looks good, but since there is quite a bit of code churn, would recommend running some jitstress pipelines. I have already kicked off antigen/fuzzlyn

[kunalspathak]

looks like there are several failures in Antigen and Fuzzlyn

[tannergooding]

looks like there are several failures in Antigen and Fuzzlyn

Yep, there was an edge case missed where we didn't create a needed CvtMaskToVector node and a fallback path (which shouldn't have really existed) caused the wrong type to be created.

I've removed the fallback, replacing it with an assert instead, and updated the relevant path to correctly reuse the existing conversion intrinsic.

[tannergooding]

/azp run Antigen, Fuzzlyn

[azure-pipelines]

Azure Pipelines successfully started running 2 pipeline(s).

[tannergooding]

/azp run runtime-coreclr jitstress-isas-x86, runtime-coreclr jitstress, runtime-coreclr jitstressregs

[azure-pipelines]

Azure Pipelines successfully started running 3 pipeline(s).

[tannergooding]

/azp run Antigen, Fuzzlyn

[azure-pipelines]

Azure Pipelines successfully started running 2 pipeline(s).

[tannergooding]

Logged kunalspathak/Antigen#12 for some of the Antigen failures. The other failures are the Arm64 Sve issue being looked at and are unrelated.