[libc] UB in llvm-project/libc/src/string/memory_utils/utils.h:assume_aligned

[petrhosek]

UBSan running on RISC-V (Fuchsia) is reporting an undefined behavior at

llvm-project/libc/src/string/memory_utils/utils.h

Line 87 in e6cd409

return reinterpret_cast<T *>(__builtin_assume_aligned(ptr, alignment));

assumption of 2 byte alignment for pointer of type 'const __llvm_libc::cpp::byte *' failed, address is 1 aligned

[llvmbot]

@llvm/issue-subscribers-libc

[gchatelet]

I somewhat missed this issue. @petrhosek do you have more context about the failure or a way I can reproduce?

[petrhosek]

We believe the issue is this function:

llvm-project/libc/src/string/memory_utils/generic/aligned_access.h

Lines 35 to 46 in a0029a8

	[[maybe_unused]] LIBC_INLINE uint64_t load64_aligned(CPtr ptr, size_t offset,
	size_t alignment) {
	if (alignment == 0)
	return load64_aligned<uint64_t>(ptr, offset);
	else if (alignment == 4)
	return load64_aligned<uint32_t, uint32_t>(ptr, offset);
	else if (alignment == 2)
	return load64_aligned<uint16_t, uint16_t, uint16_t, uint16_t>(ptr, offset);
	else
	return load64_aligned<uint8_t, uint16_t, uint16_t, uint16_t, uint8_t>(
	ptr, offset);
	}

This doesn't handle the case when alignment is 6. The else case handles all odd numbered alignments, and 6 just happens to fall into it. The following fixes the issue:

 else if (alignment == 6)
    return load64_aligned<uint16_t, uint32_t, uint16_t>(ptr, offset);

This doesn't happen on other architectures as they don't use these generic functions for memcpy.

[gchatelet]

Thx a lot for pinpointing this down, I've created #71391 to fix it.