From eb6d4fd97ea93729bc36f4a5df33cf63f8dcc4e3 Mon Sep 17 00:00:00 2001 From: Ivan Maidanski Date: Sat, 20 Aug 2016 18:39:25 +0300 Subject: [PATCH] Workaround a bug in double-wide CAS intrinsic of Apple clang-600/x64 The bug causes test_stack failure. * src/atomic_ops/sysdeps/gcc/x86.h [AO_GCC_ATOMIC_TEST_AND_SET && __APPLE_CC__ && __x86_64__]: Fall-back to asm-based implementation of double-wide CAS (AO_compare_double_and_swap_double_full); add comment explaining the reason. * src/atomic_ops/sysdeps/gcc/x86.h [__GCC_HAVE_SYNC_COMPARE_AND_SWAP_16 && AO_GCC_ATOMIC_TEST_AND_SET && __APPLE_CC__] (AO_SKIPATOMIC_double_compare_and_swap_ANY): Define. --- src/atomic_ops/sysdeps/gcc/x86.h | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) diff --git a/src/atomic_ops/sysdeps/gcc/x86.h b/src/atomic_ops/sysdeps/gcc/x86.h index 344b78a..041b75a 100644 --- a/src/atomic_ops/sysdeps/gcc/x86.h +++ b/src/atomic_ops/sysdeps/gcc/x86.h @@ -293,11 +293,15 @@ AO_fetch_compare_and_swap_full(volatile AO_t *addr, AO_t old_val, #if (defined(AO_PREFER_BUILTIN_ATOMICS) || !defined(__clang__) \ || defined(__x86_64__) || defined(__APPLE_CC__) || defined(__CYGWIN__)) \ - && defined(AO_GCC_ATOMIC_TEST_AND_SET) + && defined(AO_GCC_ATOMIC_TEST_AND_SET) \ + && !(defined(__APPLE_CC__) && defined(__x86_64__) && !defined(__ILP32__) \ + && defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_16)) /* As of clang-3.8 i686 (NDK r11c), it requires -latomic for all the */ /* double-wide operations. For now, we fall back to the */ /* non-intrinsic implementation if clang/x86. */ + /* As of Apple clang-600 (based on LLVM 3.5svn), it has some bug in */ + /* double-wide CAS implementation for x64 target. */ /* TODO: Refine for newer clang releases. */ # if defined(__ILP32__) || !defined(__x86_64__) /* 32-bit AO_t */ \ @@ -436,6 +440,8 @@ AO_fetch_compare_and_swap_full(volatile AO_t *addr, AO_t old_val, /* doesn't work. And the emulation is unsafe by our usual rules. */ /* However both are clearly useful in certain cases. */ +# define AO_SKIPATOMIC_double_compare_and_swap_ANY + AO_INLINE int AO_compare_double_and_swap_double_full(volatile AO_double_t *addr, AO_t old_val1, AO_t old_val2, -- 2.40.0