RFR: 8263512: [macos_aarch64] issues with calling va_args functions from invoke_native

[Jorn Vernee]

On Mon, 3 May 2021 15:31:53 GMT, Jorn Vernee <jvernee at openjdk.org> wrote:

>> macOS on Apple silicon uses slightly different ABI conventions to the
>> standard AArch64 ABI.  The differences are outlined in [1].  In
>> particular in the standard (AAPCS) ABI, variadic arguments may be passed
>> in either registers or on the stack following the normal calling
>> convention.  To handle this, va_list is a struct containing separate
>> pointers for arguments located in integer registers, floating point
>> registers, and on the stack.  Apple's ABI simplifies this by passing all
>> variadic arguments on the stack and the va_list type becomes a simple
>> char* pointer.
>> 
>> This patch adds a new MacOsAArch64 CABI type and MacOsAArch64Linker to
>> represent the new ABI variant on macOS.  StackVaList is based on
>> WinVaList lightly modified to handle the different TypeClasses on
>> AArch64.  The original AArch64Linker is renamed to AapcsLinker and is
>> currently used for all non-Mac platforms.  I think we also need to add a
>> WinAArch64 CABI but I haven't yet been able to test on a Windows system
>> so will do that later.
>> 
>> The macOS ABI also uses a different method of spilling arguments to the
>> stack (the standard ABI pads each argument to a multiple of 8 byte stack
>> slots, but the Mac ABI packs arguments according to their natural
>> alignment).  None of the existing tests exercise this so I'll open a new
>> JBS issue and work on that separately.
>> 
>> Tested jdk_foreign on macOS AArch64, Linux AArch64, and Linux X86_64.
>> 
>> [1] https://developer.apple.com/documentation/xcode/writing_arm64_code_for_apple_platforms
>
> src/jdk.incubator.foreign/share/classes/jdk/internal/foreign/abi/aarch64/macos/StackVaList.java line 131:
> 
>> 129:                     MemorySegment struct = allocator.allocate(layout);
>> 130:                     struct.copyFrom(segment.asSlice(0L, layout.byteSize()));
>> 131:                     segment = segment.asSlice(VA_SLOT_SIZE_BYTES);
> 
> Since arguments are packed according to alignment, I guess the offset could be larger or smaller than 8 bytes as well?

This is using `alignUp(arg.layout.byteSize(), VA_SLOT_SIZE_BYTES)` in the writing code, so I think it should be the same here?

-------------

PR: https://git.openjdk.java.net/jdk/pull/3617