CC @AaronBallman

@llvm/issue-subscribers-clang-codegen

Author: None (m-gupta)

struct vb2_hash {
	/* enum vb2_hash_algorithm. Fixed width for serialization.
	   Single byte to avoid endianness issues. */
	uint8_t algo;
	/* Padding to align and to match existing CBFS attribute. */
	uint8_t reserved[3];
	/* The actual digest. Can add new types here as required. */
	union {
		uint8_t raw[0]; // triggers ubsan oob checks
#if VB2_SUPPORT_SHA1
		uint8_t sha1[VB2_SHA1_DIGEST_SIZE];
#endif
#if VB2_SUPPORT_SHA256
		uint8_t sha256[VB2_SHA256_DIGEST_SIZE];
#endif
#if VB2_SUPPORT_SHA512
		uint8_t sha512[VB2_SHA512_DIGEST_SIZE];
#endif
	} bytes;  /* This has a name so that it's easy to sizeof(). */
};

I think any attempted read or write of a zero-sized object should be caught by UBSan because that access is trying to do something out of bounds. CC @zygoloid for additional opinions

We have explicit code to test whether we're indexing a flexible array member and disable the UBSan check, and we unconditionally treat an array of 0 elements in a union as being a flexible array member. I also can't reproduce the problem with a simple testcase. So maybe this is fixed already?

Your simple test case doesn't reproduce an issue in any version of Clang, so perhaps the issue requires a particular test case?

It looks to me like that testcase reproduces the problem with Clang 10 and before, and not with Clang 11 onwards. So I think this was fixed in Clang 11.

Thanks! Closing the issue as resolved, but if the issue does still reproduce for you with Clang 20 (or better yet, on main), please reopen.