[NativeAOT] Save full ARM64 SIMD arg registers in UniversalTransition

[VSadov]

On arm64 entire SIMD registers (q0-q7) could be used for argument passing.

I am not sure if JIT uses HVA argument passing for NativeAOT, maybe it is disabled, but it looks like our save/restore of arguments in UniversalTransition frames could be incomplete.

[VSadov]

Re: PreStub in CoreClr.

It looks like PreStub on ARM64 stores full q registers

ThePreStub uses PROLOG_WITH_TRANSITION_BLOCK macro

NESTED_ENTRY ThePreStub

        PROLOG_WITH_TRANSITION_BLOCK
        ... 

PROLOG_WITH_TRANSITION_BLOCK uses SAVE_FLOAT_ARGUMENT_REGISTERS

        IF "$SaveFPArgs" != ""
__PWTB_SaveFPArgs SETL $SaveFPArgs
        ELSE
__PWTB_SaveFPArgs SETL {true}
        ENDIF

. . . 

        ; Spill argument registers.
        SAVE_ARGUMENT_REGISTERS        sp, __PWTB_ArgumentRegisters

        IF __PWTB_SaveFPArgs
        SAVE_FLOAT_ARGUMENT_REGISTERS  sp, __PWTB_FloatArgumentRegisters
        ENDIF

and SAVE_FLOAT_ARGUMENT_REGISTERS saves q registers

runtime/src/coreclr/vm/arm64/asmmacros.h

Line 200 in d365ff2

; Reserve 128 bytes of memory before calling SAVE_FLOAT_ARGUMENT_REGISTERS

[VSadov]

Interestingly nativeaot has SAVE_FLOAT_ARGUMENT_REGISTERS defined on unix, but it could be dead code, I could not track it to its use.

runtime/src/coreclr/nativeaot/Runtime/unix/unixasmmacrosarm64.inc

Line 182 in 536f34d

// Reserve 64 bytes of memory before calling SAVE_FLOAT_ARGUMENT_REGISTERS

[VSadov]

There is also use of d0, d1 in RhCallDescrWorker

runtime/src/coreclr/nativeaot/Runtime/arm64/CallDescrWorker.asm

Line 114 in b2dc37b

stp d0, d1, [x19, #00]

It looks like this corresponds to CoreClr CallDescrWorkerInternal, which has an additional conditional handling of HVA returns

runtime/src/coreclr/vm/arm64/CallDescrWorkerARM64.asm

Line 114 in 4e61fb1

;;VectorHFAReturn return case

, and the float argument part deals with q registers (only d portions in nativeaot counterpart)

runtime/src/coreclr/vm/arm64/CallDescrWorkerARM64.asm

Line 53 in 4e61fb1

ldp q0, q1, [x9]

I do not know if nativeaot and CoreClr should match here

[jkotas]

I do not know if nativeaot and CoreClr should match here

RhCallDescrWorker is probably a dead code.

[jkotas]

I think this is candidate for release/7.0 backport. The arm64 calling convention is passing HFA in these registers.

[VSadov]

I will check once more on the nativeaot+unix version of SAVE_FLOAT_ARGUMENT_REGISTERS - to be sure it is unused.

[VSadov]

I will check once more on the nativeaot+unix version of SAVE_FLOAT_ARGUMENT_REGISTERS - to be sure it is unused.

it is unused. I will just remove it to be less confusing, since it seems to be subtly incorrect.

[VSadov]

Tests are not starting. :-/

[VSadov]

Although the new commit just deletes code, I’d like to see it builds on everything.

[jkotas]

Tests are not starting. :-/

Yes, it is a planned eng system downtime. Check your email.

[VSadov]

Yes, it is a planned eng system downtime. Check your email.

I am one of those who miss the email from the airline about 1 hour earlier departure for the last flight to the mainland. (they did not wait :)

[VSadov]

Thanks!!

[VSadov]

/backport to release/7.0

[github-actions]

Started backporting to release/7.0: https://github.com/dotnet/runtime/actions/runs/2974601463