clang doesn't emit the same warnings for global variables

[nickdesaulniers]

https://llvm.org/pr38789

[nickdesaulniers]

I plan on revisiting this later this quarter.

[nickdesaulniers]

Posting what @zygoloid sent me about this:

I think it should not require major invasive changes to Clang to fix this. While we can't built a CFG for the initializer of a global variable, we don't really need to, since there's never any control flow there (you can't have a statement-expression at global scope). So we could just walk the initializer after attaching it, skip obviously-dead branches in ?: expressions, and emit deferred runtime-behavior diagnostics for the rest. (And ideally we'd skip all of that if we don't have any pending runtime-behavior diagnostics when we reach the end of a global's initializer.)

[nickdesaulniers]

https://reviews.llvm.org/D63889

[nathanchance]

I believe this has come up again: https://lore.kernel.org/r/8c252429-5000-0649-c49f-8225d911241b@gmail.com/

As I noted in my reply though, it seems Huck's patch does not fix this particular instance but it certainly seems related.

[nickdesaulniers]

@arndb sent a workaround in https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=b27aca2e555f87d384cdf60059dd1a1b61c215ef, but we can still repro with that reverted for ARCH=arm allyesconfig if we revert that, so this is still an issue.

[nathanchance]

Another instance:

https://lore.kernel.org/r/202112181827.o3X7GmHz-lkp@intel.com/

https://godbolt.org/z/srK9cfvPT

$ cat test.c
#define DMA_BIT_MASK(n) (((n) == 64) ? ~0ULL : ((1ULL<<(n))-1))

static unsigned long long foo = DMA_BIT_MASK(64);

static unsigned long long bar(void)
{
        return DMA_BIT_MASK(64);
}

$ clang -fsyntax-only -Wshift-count-overflow test.c
test.c:3:33: warning: shift count >= width of type [-Wshift-count-overflow]
static unsigned long long foo = DMA_BIT_MASK(64);
                                ^~~~~~~~~~~~~~~~
test.c:1:54: note: expanded from macro 'DMA_BIT_MASK'
#define DMA_BIT_MASK(n) (((n) == 64) ? ~0ULL : ((1ULL<<(n))-1))
                                                     ^ ~~~
1 warning generated.

[nickdesaulniers]

Another instance:

kernel patch: https://lore.kernel.org/llvm/1644176216-12531-1-git-send-email-mikelley@microsoft.com/

[nathanchance]

The Hyper-V patch has been merged into mainline: https://git.kernel.org/linus/6bf625a4140f24b490766043b307f8252519578b

I think that fixing this should be relatively high priority so we do not get bit by this again but we are good for now.

[nickdesaulniers]

further reduced test case:

unsigned long long foo = 1 ? 0 : 1ULL << 64;

unsigned long long bar (void) {
    return 1 ? 0 : 1ULL << 64;
}

Maybe we can view the CFG to see what looks different between the two.
ClangBuiltLinux/llvm-dev-conf-2020#3 (comment)

https://reviews.llvm.org/D63889 didn't seem fundamentally against the approach, maybe could use a dust off.