JIT: Remove PUTARG_SPLIT

[jakobbotsch]

Instead replace arguments that are passed split across registers and stack by two arguments in lowering. With recent ABI and class layout work this is now straightforward.

fgTryMorphStructArg can be simplified after this, which I will do in a future change.

[dotnet-policy-service]

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

[jakobbotsch]

/azp run runtime

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[jakobbotsch]

/azp run runtime-coreclr superpmi-replay, runtime-coreclr superpmi-diffs

[azure-pipelines]

Azure Pipelines successfully started running 2 pipeline(s).

[jakobbotsch]

cc @dotnet/jit-contrib PTAL @kunalspathak

Diffs. Some MinOpts size regressions, mostly in tests, and also a few FullOpts ones. The FullOpts ones are generally the result of how we would previously create

PUTARG_SPLIT(BLK(some_address))

and the codegen for that would be able to contain address modes to load the stack part. We now create

PUTARG_STK(BLK(ADD(some_address, <size of register parts>)))

for the part that is pushed on the stack, and PUTARG_STK cannot contain the ADD(some_address, <size of register parts>). So we end up with additional codegen for the ADD.
Since it is still an overall size improvement for FullOpts I think it's ok to not bother trying to optimize this.

Other than that just some TP improvements from no longer having to check for PUTARG_SPLIT in various multireg paths.

[Copilot]

Pull Request Overview

This PR simplifies the JIT lowering by entirely removing support for the PUTARG_SPLIT operation. In its key changes, the PR renames LowerPutArgStkOrSplit to LowerPutArgStk, removes all references and code paths related to GT_PUTARG_SPLIT, and updates related headers and codegen routines accordingly.

• Rename and simplify lowering functions by consolidating split and stack cases.
• Remove obsolete GT_PUTARG_SPLIT branches and related code paths from multiple files.
• Update associated header declarations and documentation to reflect the change.

Reviewed Changes

Copilot reviewed 32 out of 32 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
src/coreclr/jit/lowerriscv64.cpp	Renamed LowerPutArgStkOrSplit to LowerPutArgStk; updated comment accordingly.
src/coreclr/jit/lowerloongarch64.cpp	Consistent renaming of lowering function and comment update.
src/coreclr/jit/lowerarmarch.cpp	Updated comments and function names to remove PUTARG_SPLIT references.
src/coreclr/jit/lower.h	Removed LowerPutArgStkOrSplit declaration and updated lowering API.
src/coreclr/jit/lower.cpp	Removed GT_PUTARG_SPLIT-specific code and updated call sites.
src/coreclr/jit/gentree.h	Removed obsolete API methods and affected call sites regarding PUTARG_SPLIT.
Other codegen and related files	Cleaned up switch cases and function definitions to remove PUTARG_SPLIT handling.

Comments suppressed due to low confidence (3)

src/coreclr/jit/codegenriscv64.cpp:4287

• The removal of the GT_PUTARG_SPLIT case from the switch statement is appropriate. Please verify that no architecture-specific invariants rely on this case.

case GT_PUTARG_SPLIT:

src/coreclr/jit/lowerriscv64.cpp:477

• The comment now clearly reflects the removal of the PUTARG_SPLIT case; please ensure that any associated documentation is updated similarly.

// LowerPutArgStk: Lower a GT_PUTARG_STK.

src/coreclr/jit/lower.h:396

• The header now only declares LowerPutArgStk; double-check that any outdated references or documentation regarding PUTARG_SPLIT have been removed.

void     LowerPutArgStk(GenTreePutArgStk* putArgNode);

[jakobbotsch]

Ping @kunalspathak

[kunalspathak]

LGTM. Did you already do a round of Fuzzlyn/Antigen?

[jakobbotsch]

LGTM. Did you already do a round of Fuzzlyn/Antigen?

No, but ABIStress tests should already have targeted fuzzing coverage of this.

[jakobbotsch]

/ba-g Timeout

[MichalStrehovsky]

@jakobbotsch I think this broke native AOT outerloops. I'm now seeing:

2025-06-04T11:38:26.7563674Z     ILC: /__w/1/s/src/coreclr/jit/compiler.cpp:10593
2025-06-04T11:38:26.7577952Z     ILC: Assertion failed '(result >= 0) && ((unsigned)result < ArrLen(str))' in 'Internal.CompilerGenerated.<Module>:DynamicInvoke(uint,byref,byref,uint):byref' during 'Lowering nodeinfo' (IL size 28; hash 0xda6f5354; FullOpts)
2025-06-04T11:38:26.7580904Z     
2025-06-04T11:38:27.2170713Z /__w/1/s/artifacts/bin/coreclr/linux.arm.Checked/build/Microsoft.NETCore.Native.targets(332,5): error MSB3073: The command ""/__w/1/s/artifacts/bin/coreclr/linux.arm.Checked/x64/ilc/ilc" @"/__w/1/s/artifacts/tests/coreclr/obj/linux.arm.Checked/Managed/CoreMangLib/system/span/RefStructWithSpan/native/RefStructWithSpan.ilc.rsp"" exited with code 134. [/__w/1/s/src/tests/CoreMangLib/system/span/RefStructWithSpan.csproj] [/__w/1/s/src/tests/build.proj]

The test in question uses recursive generics and from what I can gather, this PR added a call to printfAlloc that uses type names. Looks like printfAlloc has a 512 character limit on strings.

[jakobbotsch]

@jakobbotsch I think this broke native AOT outerloops. I'm now seeing:

2025-06-04T11:38:26.7563674Z     ILC: /__w/1/s/src/coreclr/jit/compiler.cpp:10593
2025-06-04T11:38:26.7577952Z     ILC: Assertion failed '(result >= 0) && ((unsigned)result < ArrLen(str))' in 'Internal.CompilerGenerated.<Module>:DynamicInvoke(uint,byref,byref,uint):byref' during 'Lowering nodeinfo' (IL size 28; hash 0xda6f5354; FullOpts)
2025-06-04T11:38:26.7580904Z     
2025-06-04T11:38:27.2170713Z /__w/1/s/artifacts/bin/coreclr/linux.arm.Checked/build/Microsoft.NETCore.Native.targets(332,5): error MSB3073: The command ""/__w/1/s/artifacts/bin/coreclr/linux.arm.Checked/x64/ilc/ilc" @"/__w/1/s/artifacts/tests/coreclr/obj/linux.arm.Checked/Managed/CoreMangLib/system/span/RefStructWithSpan/native/RefStructWithSpan.ilc.rsp"" exited with code 134. [/__w/1/s/src/tests/CoreMangLib/system/span/RefStructWithSpan.csproj] [/__w/1/s/src/tests/build.proj]

The test in question uses recursive generics and from what I can gather, this PR added a call to printfAlloc that uses type names. Looks like printfAlloc has a 512 character limit on strings.

Sorry about that. Will submit a fix shortly...