[PATCH] include: Don't use the MSVC intrinsic _InterlockedCompareExchange128 on clang in MSVC mode

[Martin Storsjo]

Clang doesn't consider the currently declared function as the proper
intrinsic function (a few parameters differ vaguely), and even if
declared properly, clang requires building for a target with the cx16
flag enabled to be able to use that intrinsic.

Instead just use the inline assembly or GCC builtin version instead.

This fixes building in PE mode with Clang in MSVC mode for both x86_64
and aarch64.

Signed-off-by: Martin Storsjo <martin(a)martin.st>
---
 include/winnt.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/winnt.h b/include/winnt.h
index 04353d9405d..9d797a85099 100644
--- a/include/winnt.h
+++ b/include/winnt.h
@@ -6894,7 +6894,7 @@ static inline BOOLEAN BitScanReverse(DWORD *index, DWORD mask)
 
 #ifdef _WIN64
 
-#if defined(_MSC_VER)
+#if defined(_MSC_VER) && !defined(__clang__)
 
 #define InterlockedCompareExchange128 _InterlockedCompareExchange128
 static inline unsigned char _InterlockedCompareExchange128(__int64 *dest, __int64 xchg_high, __int64 xchg_low, __int64 *compare);
@@ -6913,7 +6913,7 @@ static inline unsigned char InterlockedCompareExchange128(__int64 *dest, __int64
     return ret;
 }
 
-#elif defined(__GNUC__)
+#elif defined(__GNUC__) || defined(__clang__)
 
 static inline unsigned char InterlockedCompareExchange128(__int64 *dest, __int64 xchg_high, __int64 xchg_low, __int64 *compare)
 {
-- 
2.17.1

[Jacek Caban]

Hi Martin,

On 18.12.2020 23:35, Martin Storsjo wrote:
> This fixes building in PE mode with Clang in MSVC mode for both x86_64
> and aarch64.

I'd expect x86_64 to be fixed by [1], we could do the same for aarch64.


If changing header is preferable for some reason, we should reflect that 
in configure and maybe reorder #ifdefs to simplify it a bit.


Thanks,

Jacek


[1] 
https://source.winehq.org/git/wine.git/commitdiff/dc199ae8b8b38ac088eb0bcd05...

[Martin Storsjö]

Hi Jacek,

On Sat, 19 Dec 2020, Jacek Caban wrote:
> Hi Martin,
>
> On 18.12.2020 23:35, Martin Storsjo wrote:
> > This fixes building in PE mode with Clang in MSVC mode for both x86_64
> > and aarch64.
>
> I'd expect x86_64 to be fixed by [1], we could do the same for aarch64.

Ah, I see - that explains it. Yeah I was a bit surprised to see that it 
didn't work as expected, as I presumed that you had tested building in 
MSVC mode.

The actual issue here seems to be that the signature of the intrinsic 
_InterlockedCompareExchange128 is slightly wrong. Clang 11 and older (and 
MSVC itself) doesn't mind, but I was testing with a fresh trunk version of 
clang, where it needs to match better. See https://godbolt.org/z/TWbMs8 
for an example of the difference.

FWIW, I bisected the change in behaviour down to 
https://reviews.llvm.org/D77491.

In that case, the fix is simply to correct the signature of the 
declaration - sending a new patch for that instead.

// Martin