Emit cmp (extended register) on ARM64 to simplify cast-then-compare expressions

[snickolls-arm]

Working towards #68028

Matching on GT_EQ/NE/... nodes containing GT_CAST.

Example:

<==
uxtb w1, w0
cmp w2, w1
==>
cmp w2, w1, UXTB

[snickolls-arm]

@a74nh @kunalspathak

[dotnet-policy-service]

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

[kunalspathak]

ClearContained

so what happens if the cast containment is not cleared? I am trying to understand how it works currently for this scenario.

[snickolls-arm]

Some assertions hit related to register allocation if we don't clear it here. The problem is with chains such as:

ADD(
  op1,
  CAST(
    LCL_VAR
  )
)

Usually the LCL_VAR node will be contained by the CAST node, this allows a later optimization using LDRB/H to perform the cast with load. The register will be allocated to the CAST node which is then used by the ADD. If we contain the CAST and the LCL_VAR spills, then both the CAST and LCL_VAR nodes won't have a register allocated when we try to generate the ADD. If I understand correctly, it's not possible to tell whether a LCL_VAR will be spilled until after register allocation, but you can force the casted node into a register and contain the CAST node instead.

In most cases the casted node is a register value and the containment from CAST -> LCL_VAR had no effect, so this will result in removing the cast instruction. However, if the casted node is spilled then you would see LDR used (no cast), and the ADD node will use the cast extension, which is the behavior I'm preserving here. I've moved these containment settings from ContainCheckBinary to here because they were causing side effects with the comparison nodes.

This problem also applies to the comparison nodes, which is why I clear the RegOptional flag on some nodes.

[kunalspathak]

looks like there is a test failure . Can you please double check if it is caused with this PR?

[snickolls-arm]

looks like there is a test failure . Can you please double check if it is caused with this PR?

It looks like this failed on "Send to Helix", is this a network error? I can't see any failing build or test output.

[jakobbotsch]

It looks like this failed on "Send to Helix", is this a network error? I can't see any failing build or test output.

You can see the test output by going to Tests -> System.Runtime.Intrinsics.Tests Work item -> Artifacts -> Console Logs. That takes you to this link: https://helixr1107v0xdcypoyl9e7f.blob.core.windows.net/dotnet-runtime-refs-pull-112411-merge-30e75b278b4d4c6eae/System.Runtime.Intrinsics.Tests/1/console.cb1ccbdb.log?helixlogtype=result

In this case the failure looks like known failure #111922.

[jakobbotsch]

It looks odd to me that a function called IsContainableUnaryOrBinaryOp mutates random trees. Can this be done by the caller instead?

[snickolls-arm]

Thanks, fixed this

[kunalspathak]

LGTM