patch: avoid ubsan errors in implicit copy constructors

[Richard Smith]

Use a ctor-initializer for OldSanOpts.

Do you need to handle the assignment operator too? A test for move
constructors / move assignments would be great, too (not that I expect any
problems there).

Do you care that the UBSan checks will still be enabled
when CGF.getLangOpts().getGC() != LangOptions::NonGC? We seem to also fail
to memcpy non-GC'd fields in that case too =)


On Mon, Sep 9, 2013 at 3:03 PM, Nick Lewycky <nlewycky at google.com> wrote:

> The attached patch disables the bool and enum sanitizers when emitting the
> implicitly-defined copy constructor.
>
> To start with an example:
>   struct X { X(); X(const X&); };
>   struct Y { X x; bool b; };
> if you don't initialize Y::b then try to copy an object of Y type, ubsan
> will complain. Technically, with the standard as written, ubsan is correct.
> However, this is a useful thing to do -- you may have a discriminator which
> decides which elements are not interesting and therefore never initialize
> or read them. Secondly, it's a departure from the rules in C, making
> well-defined C code have undefined behaviour in C++ (structs are never trap
> values, see C99 6.2.6.1p6). Thirdly, it's checked incompletely right now --
> if you make subtle changes (f.e. add an "int i;" member to Y) ubsan will
> stop complaining. The semantic I'm implementing is as if the implicit copy
> constructor is copying the value representation (the bytes) not the object
> representation (the meaning of those bytes).
>
> Nick
>
> _______________________________________________
> cfe-commits mailing list
> cfe-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20130909/380d3db5/attachment.html>