Optimise long multiply + add/sub/neg on arm64.

[c272]

This PR optimises an extending multiply of two integers out to a long followed by an addition, subtraction or negation operation down to a single [s/u]maddl, [s/u]subl or smnegl instruction. It also adds a generic ternary operator which can be used in the future to support other three argument nodes.

The spmidiff results for this patch can be found at the below gist (performed on win-arm64 since linux-arm64 is currently broken as per #91257):
https://gist.github.com/c272/944721d2d083660e0ba7ec9fddb24de4

Minor regressions seen here are due to additional mov instructions changing the register allocation in cases where the add value is containable, preventing some previously performed ldp optimisations. Eliminating all cases of containable adds also removes many valid optimisations, so I have currently left it as is.

Contributes to #68028

[ghost]

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

Issue Details

---

This PR optimises an extending multiply of two integers out to a long followed by an addition, subtraction or negation operation down to a single [s/u]maddl, [s/u]subl or smnegl instruction. It also adds a generic ternary operator which can be used in the future to support other three argument nodes.

The spmidiff results for this patch can be found at the below gist (performed on win-arm64 since linux-arm64 is currently broken as per #91257):
https://gist.github.com/c272/944721d2d083660e0ba7ec9fddb24de4

Minor regressions seen here are due to additional mov instructions changing the register allocation in cases where the add value is containable, preventing some previously performed ldp optimisations. Eliminating all cases of containable adds also removes many valid optimisations, so I have currently left it as is.

Author:	c272
Assignees:	-
Labels:	area-CodeGen-coreclr, community-contribution
Milestone:	-

[kunalspathak]

In the diffs, I see smull+add is combined into umaddl instead of smaddl. Also, can you please fix the build errors?

[kunalspathak]

It seems adding GenTreeTernOp can be avoided. In the past, we directly recognized the pattern in codegen e.g. madd, msub. Did you try that approach and faced any problem?

[kunalspathak]

@dotnet/jit-contrib @TIHan

[a74nh]

It seems adding GenTreeTernOp can be avoided. In the past, we directly recognized the pattern in codegen e.g. madd, msub. Did you try that approach and faced any problem?

I guess that would be possible, but it feels like it should be done during lowering (I have no overly strong opinion though).

My thought for GenTreeTernOp is that a follow on patch could move SELECT, CMPXCHG over to use it, which should simplify things a little.

[tannergooding]

Rather than GenTreeTernOp, we already have GenTreeHWIntrinsic which supports multi-op and which is really oriented around these types of specialized instructions for both SIMD and Scalar code.

We can just introduce an ArmBase_LongMultiplyAdd or similar for internal use only and have lowering generate that node. The existing metadata table will make this very simple to support and likely require less overall code changes.

[c272]

@tannergooding Thank you for the comment, I have reworked the patch to utilise GT_HWINTRINSIC instead of creating a new ternary node. The updated patch should also resolve the testing failures.

Updated spmidiff results are here (re-run on win-arm64):
https://gist.github.com/c272/5f7275fdff6b829cdc6d002b29f5bb9e

[kunalspathak]

Need to resolve merge conflict and also fix the test failure:

Assert failure(PID 275 [0x00000113], Thread: 275 [0x0113]): Assertion failed 'compiler->compIsaSupportedDebugOnly(HWIntrinsicInfo::lookupIsa(intrin.id))' in 'System.Collections.Generic.Dictionary`2[System.__Canon,System.__Canon]:TryInsert(System.__Canon,System.__Canon,ubyte):ubyte:this' during 'Generate code' (IL size 668; hash 0x26ab038f; FullOpts)

    File: /__w/1/s/src/coreclr/jit/hwintrinsiccodegenarm64.cpp Line: 210
    Image: /root/helix/work/correlation/corerun

[tannergooding]

This looks correct to me. It should get sign off from someone from @dotnet/jit-contrib before being merged

[kunalspathak]

LGTM