[PATCH] D62825: [C++2a] Add __builtin_bit_cast, used to implement std::bit_cast

[John McCall via Phabricator via cfe-commits]

rjmccall added a comment.

In D62825#1542639 <https://reviews.llvm.org/D62825#1542639>, @rsmith wrote:

> In D62825#1542597 <https://reviews.llvm.org/D62825#1542597>, @rjmccall wrote:
>
> > In D62825#1542374 <https://reviews.llvm.org/D62825#1542374>, @rsmith wrote:
> >
> > > In D62825#1542309 <https://reviews.llvm.org/D62825#1542309>, @rjmccall wrote:
> > >
> > > > In D62825#1542301 <https://reviews.llvm.org/D62825#1542301>, @rsmith wrote:
> > > >
> > > > > In D62825#1542247 <https://reviews.llvm.org/D62825#1542247>, @rjmccall wrote:
> > > > >
> > > > > > In what sense is the bit-pattern of a null pointer indeterminate?
> > > > >
> > > > >
> > > > > The problem is not null pointers, it's `nullptr_t`, which is required to have the same size and alignment as `void*` but which comprises only padding bits. (Loads of `nullptr_t` are not even permitted to touch memory...).
> > > >
> > > >
> > > > I mean, I know this is C++ and the committee loves tying itself in knots to make the language unnecessarily unusable, but surely the semantics of bitcasting an r-value of type `nullptr_t` are intended to be equivalent to bitcasting an r-value of type `void*` that happens to be a null pointer.
> > >
> > >
> > > I don't follow -- why would they be? `bit_cast` reads the object representation, which for `nullptr_t` is likely to be uninitialized, because the type contains only padding bits. (Note that there is formally no such thing as "bitcasting an rvalue". `bit_cast` takes an lvalue, and reinterprets its storage.)
> >
> >
> > I agree that the problem is that the object representation of `nullptr_t` is wrong, but it seems absurd to me that we're going to bake in an absurd special case (from the user's perspective) to `bit_cast` because we consider that representation unfixable.
>
>
> Note that `bit_cast` supports casting not just from fundamental types but also from aggregates (which might contain a `nullptr_t`). At runtime, we're going to `memcpy` from the source object, which will leave the bits in the destination corresponding to `nullptr_t` subobjects in the source uninitialized (if they were in fact uninitialized in the source object, which they might be). It would be unreasonable to initialize the bits in the `bit_cast` result during compile-time evaluation but not during runtime evaluation -- compile-time evaluation is supposed to diagnose cases that would be undefined at runtime, not define them.
>
> In my view, the mistake was specifying `nullptr_t` to have the same size and alignment as `void*`; it should instead be an empty type. Only confusion results from making it "look like" a pointer type rather than just being an empty tag type.


Perhaps, but that's clearly unfixable without breaking ABI, whereas the object-representation issue is fixable by, at most, requiring a few stores that might be difficult to eliminate in some fanciful situations.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D62825/new/

https://reviews.llvm.org/D62825