[PATCH] D30283: [ubsan] Reduce alignment checking of C++ object pointers

[Vedant Kumar via Phabricator via cfe-commits]

vsk added a comment.

Hi Eli, thanks for your feedback :).

In https://reviews.llvm.org/D30283#702085, @efriedma wrote:

> I'm not sure we actually want to skip these checks for DeclRefExps.  I mean, you can rely on the backend to correctly align a local variable (assuming the stack is correctly aligned), but it's very easy to misalign a global.


I did some experimenting, and think that we can safely skip the alignment check for DeclRefExps. Here is what I found:

1. It seems hard to declare a misaligned global. I could not come up with an example where this happens. FWIW, I tried playing tricks with the linker but it refused to do the wrong thing (`-Wl,-sectalign,__DATA,__common,1` ; `-Wl,-order_file,insert_pad_byte_to_misalign_global.order_file`).
2. It's easy to declare a global pointer or reference to misaligned data. UBSan already misses this bug: this patch doesn't affect the situation (I filed llvm.org/PR32364 to track the issue).

When you said "it's very easy to misalign a global", did you mean that it's easy for the compiler to mess up? If so, my thinking is that we should write better verifiers, either in our linkers or in LLVM proper. UBSan shouldn't be used as a compiler verifier: if it is, end-users will find its diagnostics to be less actionable. We shouldn't break user bots because of compiler bugs.

If, OTOH, you meant that it's easy for users to declare misaligned globals, I'd appreciate an example.


https://reviews.llvm.org/D30283