GCC: Impliment Microsoft variadic ABI

[ZacWalk]

Windows used the aarch64 abi for everything except variadic functions. The assumption is that this will be an issue when calling variadic Methods in the Microsoft CRT like printf. However, in most cases looks like the mingw implementation of printf is used.

I have marked this as an enhancement because I am not sure if is essential for our initial upstream. The Micorosft aarch64 ABI follows AArch64 EABI for everything except variadic functions. We need to investigate what problems there are in real-world scenarios caused by this.

Relevant documentation:

• https://learn.microsoft.com/en-us/cpp/build/arm64-windows-abi-conventions?view=msvc-170#addendum-variadic-functions

[puetzk]

Another difference (that you're probably aware of, but I don't see it mentioned) is that Microsoft reserves the optional platform register from the aapcs64 (x18) to always point to the TEB: https://learn.microsoft.com/en-us/cpp/build/arm64-windows-abi-conventions?view=msvc-170#integer-registers

Which means GCC needs to keep track of which functions are -fabi=ms or__attribute__((ms_abi)) (like it currently does for x86 things like stack alignment). And ms_abi functions need to not use it as scratch, and should caller-save when calling functions that aren't ms_abi (since e.g. in winelib one can actually mix ms_abi code calling ELF functions that won't necessarily preserve it).

[ZacWalk]

Thanks for the feedback @puetzk. There is a simple fix for x18/TEB in this commit. Simply configure x18 not to be scheduled. I need to rework it a little to make it configurable depending on the target. Adding an issue to track that.

The ABI is an interesting problem. My understanding is that the Micorosft aarch64 ABI follows AArch64 EABI for everything except variadic functions. I started working on implementing a ms-variadic ABI but have not had time to complete it yet. WIP is in this branch. Felt like I might be going down a rabbit hole with it.

[ZacWalk]

Added some test code for this here:
https://github.com/ZacWalk/mingw-woarm64-build/blob/main/tests/test-varargs.c
Looks like there is some corruption with va_list but __builtin_va_arg_pack did look ok.

[Blackhex]

Currently there are two separate issues related to variadic:

• Calling e.g. a variadic function with float arguments stores the float argument into a wrong register. Let's keep this issue focused on that issue.
• va_list structure is corrupted. Let's move this to another issue GCC: va_list structure is corrupted #70