1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
|
Lock is_always_lock free test fails on i386 because std::atomic is aligned
to 8 bytes while long long is aligned to 4 bytes. clang can't generate inline
code for unaligned 8 byte atomics even tough instruction set and gcc support
it.
That makes it expected thaqt ATOMIC_LLONG_LOCK_FREE and
std::atomic<long long>::is_always_lock_free don't match on i386. Correct test
for std::atomic<long long> is to check if target cpu support cmpxchg8 instruction.
To set instruction support one can check __GCC_HAVE_SYNC_COMPARE_AND_SWAP_8 define.
Bug: https://llvm.org/bugs/show_bug.cgi?id=19355
Index: llvm-toolchain-snapshot_11~++20200326111000+4673699a470/libcxx/test/std/atomics/atomics.lockfree/isalwayslockfree.pass.cpp
===================================================================
--- llvm-toolchain-snapshot_11~++20200326111000+4673699a470.orig/libcxx/test/std/atomics/atomics.lockfree/isalwayslockfree.pass.cpp
+++ llvm-toolchain-snapshot_11~++20200326111000+4673699a470/libcxx/test/std/atomics/atomics.lockfree/isalwayslockfree.pass.cpp
@@ -21,6 +21,14 @@
# error Feature test macro missing.
#endif
+#if defined(__i386__) && defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_8)
+/* Fix for clang setting __GCC_ATOMIC_LLONG_LOCK_FREE incorecctly for x86
+ * https://llvm.org/bugs/show_bug.cgi?id=19355
+ */
+#undef ATOMIC_LLONG_LOCK_FREE
+#define ATOMIC_LLONG_LOCK_FREE 2
+#endif
+
template <typename T> void checkAlwaysLockFree() {
if (std::atomic<T>::is_always_lock_free)
assert(std::atomic<T>().is_lock_free());
|
