[PATCH] arm64: Use __builtin_ms_va_list and __attribute__((ms_abi)) on arm64

[Martin Storsjö]

On Mon, 14 Aug 2017, André Hentschel wrote:

> Am 14.08.2017 um 12:35 schrieb Martin Storsjo:
>> Windows uses a different ABI for va_list on arm64 just like on x86_64.
>>
>> On x86_64, the calling convention for windows functions is completely
>> different from the one on other platforms. On arm64, they're mostly the
>> same, with the only exception being variadic functions (where all float
>> arguments are passed in integer registers, since the va_list is a single
>> pointer).
>>
>> Any functions using __builtin_ms_va_start need to be decorated with
>> __attribute__((ms_abi)).
>>
>> Add a check in configure for the support of __builtin_ms_va_list
>> (both features appared in clang at the same time, in the 5.0 release).
>>
>> This fixes running binaries that use e.g. printf style functions
>> (e.g. midl.exe from the windows 10 sdk), fixing
>> https://bugs.winehq.org/show_bug.cgi?id=38886.
>>
>> Signed-off-by: Martin Storsjo <martin at martin.st>
>> ---
>> clang 5.0, which this requires, isn't yet released, but is in RC2 and
>> will probably be released pretty soon.
>> ---
>>  configure.ac             | 7 +++++++
>>  include/msvcrt/crtdefs.h | 6 +++++-
>>  include/windef.h         | 6 +++++-
>>  include/wine/test.h      | 4 ++--
>>  4 files changed, 19 insertions(+), 4 deletions(-)
>>
>> diff --git a/configure.ac b/configure.ac
>> index 7d52b24..c33221f 100644
>> --- a/configure.ac
>> +++ b/configure.ac
>> @@ -179,6 +179,13 @@ case $host in
>>      CFLAGS="$CFLAGS -marm"
>>      AC_SUBST(TARGETFLAGS,"-marm")
>>      ;;
>> +  aarch64*)
>> +    AC_MSG_CHECKING([whether $CC supports __builtin_ms_va_list])
>> +    AC_COMPILE_IFELSE([AC_LANG_PROGRAM([[#include <stdarg.h>]], [[void func(__builtin_ms_va_list *args);]])],
>> +                      [AC_MSG_RESULT([yes])],
>> +                      [AC_MSG_RESULT([no])
>> +                       AC_MSG_ERROR([You need clang >= 5.0 to build Wine for arm64.])])
>
> Hi,
>
> The upstreamed variadic function handling is terrific, thanks for that!
> I also like this patch as logical next step, I'm just a bit worried because it locks out current builds without upgrading to an RC...
> At the moment I'm doing winelib builds with gcc for Linux/ARM64 and clang builds for Android/ARM64.
> It's quite easy on Linux to pick up a new clang version, not sure how long it takes until clang5 is picked up in the Android NDK...
> On the other hand, there are not much people building Wine for ARM64, and those can pretty easily revert your patch until they are at clang5

Yeah, I was a little worried about this, but I thought that making this 
part conditional in the headers would make it quite messy, for something 
that is mostly transitional anyway.

Now the other question is who would be willing to implement the same in 
gcc. Since it requires changes to the implementation of va_start (which, 
at least in llvm, is pretty closely tied with the implementation of stack 
frame handling and callee register saving), it's a much larger task than I 
originally thought when I embarked on it. In June, I tried grepping the 
gcc tree for the relevant keywords, but I decided to skip it for now.

// Martin