host-utils: Avoid using __builtin_subcll on buggy versions of Apple Clang

We use __builtin_subcll() to do a 64-bit subtract with borrow-in and
borrow-out when the host compiler supports it.  Unfortunately some
versions of Apple Clang have a bug in their implementation of this
intrinsic which means it returns the wrong value.  The effect is that
a QEMU built with the affected compiler will hang when emulating x86
or m68k float80 division.

The upstream LLVM issue is:
https://github.com/llvm/llvm-project/issues/55253

The commit that introduced the bug apparently never made it into an
upstream LLVM release without the subsequent fix
fffb6e6afd
but unfortunately it did make it into Apple Clang 14.0, as shipped
in Xcode 14.3 (14.2 is reported to be OK). The Apple bug number is
FB12210478.

Add ifdefs to avoid use of __builtin_subcll() on Apple Clang version
14 or greater.  There is not currently a version of Apple Clang which
has the bug fix -- when one appears we should be able to add an upper
bound to the ifdef condition so we can start using the builtin again.
We make the lower bound a conservative "any Apple clang with major
version 14 or greater" because the consequences of incorrectly
disabling the builtin when it would work are pretty small and the
consequences of not disabling it when we should are pretty bad.

Many thanks to those users who both reported this bug and also
did a lot of work in identifying the root cause; in particular
to Daniel Bertalan and osy.

Cc: qemu-stable@nongnu.org
Resolves: https://gitlab.com/qemu-project/qemu/-/issues/1631
Resolves: https://gitlab.com/qemu-project/qemu/-/issues/1659
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Tested-by: Daniel Bertalan <dani@danielbertalan.dev>
Tested-by: Tested-By: Solra Bizna <solra@bizna.name>
Message-id: 20230622130823.1631719-1-peter.maydell@linaro.org
This commit is contained in:
Peter Maydell 2023-06-22 14:08:23 +01:00
parent 57223a4c24
commit b0438861ef
2 changed files with 14 additions and 1 deletions

View File

@ -184,4 +184,17 @@
#define QEMU_DISABLE_CFI
#endif
/*
* Apple clang version 14 has a bug in its __builtin_subcll(); define
* BUILTIN_SUBCLL_BROKEN for the offending versions so we can avoid it.
* When a version of Apple clang which has this bug fixed is released
* we can add an upper bound to this check.
* See https://gitlab.com/qemu-project/qemu/-/issues/1631
* and https://gitlab.com/qemu-project/qemu/-/issues/1659 for details.
* The bug never made it into any upstream LLVM releases, only Apple ones.
*/
#if defined(__apple_build_version__) && __clang_major__ >= 14
#define BUILTIN_SUBCLL_BROKEN
#endif
#endif /* COMPILER_H */

View File

@ -649,7 +649,7 @@ static inline uint64_t uadd64_carry(uint64_t x, uint64_t y, bool *pcarry)
*/
static inline uint64_t usub64_borrow(uint64_t x, uint64_t y, bool *pborrow)
{
#if __has_builtin(__builtin_subcll)
#if __has_builtin(__builtin_subcll) && !defined(BUILTIN_SUBCLL_BROKEN)
unsigned long long b = *pborrow;
x = __builtin_subcll(x, y, b, &b);
*pborrow = b & 1;