[cfe-dev] Alignment problems in Clang's internal data structures?

[Ahmed Bougacha]

It turns out the alignment sanitizer, though "unsupported", works fine
on ARM when building with clang (I was using gcc as the stage1
compiler, which actively complains).

I stopped because it's not even possible to build llvm-tblgen, since
there is some undefined behavior in libstdc++, uncovered by the
sanitizer..  I'll try to give it another look.

In the meantime, filed
  http://llvm.org/bugs/show_bug.cgi?id=22728

-Ahmed


On Mon, Feb 16, 2015 at 4:38 PM, Ahmed Bougacha
<ahmed.bougacha at gmail.com> wrote:
> Hi all,
>
> I've been looking into what might be alignment issues throughout clang.
>
> Consider "DeclRefExpr::Create(.., NameInfo, ..)".  It allocates itself
> and a few other structs in a single allocation, using the well-known
> pattern:
>
>   std::size_t Size = sizeof(DeclRefExpr);
>   if (...)
>     Size += sizeof(...);
>   ...
>   void *Mem = Context.Allocate(Size, llvm::alignOf<DeclRefExpr>());
>
> The alignment should also take into account the trailing classes,
> really.  In every Allocate call that has a non-trivial size, the
> alignment should be correspondingly non-trivial.  Some calls just use
> ASTContext's default, 8 (enough for most purposes, at least; better
> than the explicit alignOf, which might be 4).
>
>
> [skip forward if you don't care about the investigation]
>
> Currently, I think we get away with this because 1) X86 is lenient, 2)
> most people don't self-host on another architecture, and 3) at least
> on ARM, alignment checking has to be explicitly enabled, statically or
> dynamically.
>
> On ARM, if you try running clang with either:
> - SCTRL.A==1 (alignment checking on non-strict-aligned instructions)
> (I should say I couldn't get this to work); or
> - different LLVM CodeGen picking stricter instructions that actually
> enforce alignment (what made me investigate),
>
> it should crash, at some point.  On the LIT-tests, the first (of many)
> to fail was
>   test/Analysis/dead-stores.cpp
> Indeed, on trunk clang, the "this" pointer in
>   ASTTemplateArgumentListInfo::initializeFrom(Info, b, b, b)
> is sometimes dynamically 4-aligned, whereas it is supposed to be
> 8-aligned (according to alignof and the IR we generate.)
>
>
>
> This pattern is pervasive throughout clang, and even though a few
> classes try to get it right (with a trailing AlignedCharArray), most
> don't.  What do clang developers think?  Did I miss something,
> perhaps?
>
> UBSan's alignment sanitizer would be *very* useful here, but my
> understanding is, ARM isn't currently supported, and running it on X86
> is futile (at least x86_64, maybe i386 could reproduce though; that's
> on my todo list.)
>
> Thanks!
>
> -Ahmed