[libc++][PowerPC] align.pass.cpp test failure for 13.0.0-rc1 on PowerPC

[amy-kwan]

Bugzilla Link	51548
Resolution	FIXED
Resolved on	Oct 11, 2021 20:29
Version	trunk
OS	Linux
Blocks	#51489
CC	@ldionne,@tstellar
Fixed by commit(s)	f640c31 198edf0

Extended Description

The following libc++ LIT test fails on 13.0.0-rc1 for Ubuntu on Power:

libc++ :: libcxx/atomics/atomics.align/align.pass.cpp

---

An assertion of the following is produced for this test:

command stderr:

t.tmp.exe: /home/conanap/llvm/community/llvm-project/libcxx/test/libcxx/atomics/atomics.align/align.pass.cpp:41: atomic_test::atomic_test() [T = LLIArr2]: Assertion `alignof(AtomicImpl) >= sizeof(AtomicImpl) && "expected natural alignment for lock-free type"' failed.

error: command failed with exit status: 250

This failure is still being investigated.

[amy-kwan]

This failure was revealed through this commit:

commit 6427c53
Author: Louis Dionne ldionne.2@gmail.com
Date: Wed Mar 24 16:45:55 2021 -0400

[libc++] Use add_lit_testsuite to register the libc++ test suite

The Runtimes build uses variables set by add_lit_testsuite to discover
tests suites to run.

Differential Revision: https://reviews.llvm.org/D97913

If I understand correctly, this commit runs the libc++ test cases on our side now, whereas they weren't being run for us on Power previously.

[amy-kwan]

This failure is Linux specific for Power.

The align.pass.cpp test case tests the alignment of multiple different atomic types. If the atomic type is "lock-free", then the type must be properly aligned.

The issue is that alignof() should be >= sizeof() if the type is considered lock-free, but is not for two of the types found in the test. A reduced test case showing the two types that are not adequately aligned:

---

#include
#include
#include <stdio.h>

template
struct atomic_test : public std::__atomic_base {
atomic_test() {
if (this->is_lock_free()) {
using AtomicImpl = decltype(this->_a);
printf("alignof() = %lu, sizeof() = %lu\n",
alignof(AtomicImpl), sizeof(AtomicImpl));
//assert(alignof(AtomicImpl) >= sizeof(AtomicImpl) &&
// "expected natural alignment for lock-free type");
}
}
};

int main(int, char**) {

#define CHECK_ALIGNMENT(T)
do {
typedef T type;
atomic_test t;
} while (0)

CHECK_ALIGNMENT(struct LLIArr2 { long long int i[2]; });
CHECK_ALIGNMENT(struct Padding { char c; /* padding */ long long int i; });

return 0;
}

---

When ran, this produces:

alignof() = 8, sizeof() = 16
alignof() = 8, sizeof() = 16

As they do not match, the test would normally fail due to the assert() for expected natural alignment for lock-free type (but is commented out in the reduced test above).

[amy-kwan]

For now, I have a patch that XFAILs this test case on Power: https://reviews.llvm.org/D108668

[ldionne]

Shipped to main and cherry-picked to release/13.x as 198edf0.

[tstellar]

Merged: 198edf0

[tstellar]

mentioned in issue #51489