[PATCH] D61458: [hip] Relax CUDA call restriction within `decltype` context.

[Justin Lebar via cfe-commits]

> So, actually, I wonder if that's not the right answer. We generally allow
different overloads to have different return types. What if, for example,
the return type on the host is __float128 and on the device it's
`MyLongFloatTy`?

The problem is that conceptually compiling for host/device does not create
a new set of overloads.

When we compile for (say) host, we build a full AST for all functions,
including device functions, and that AST must pass sema checks.  This is
significant for example because when compiling for device we need to know
which kernel templates were instantiated on the host side, so we know which
kernels to emit.

Here's a contrived example.

```
 __host__ int8 bar();
__device__ int16 bar();
__host__ __device__ auto foo() -> decltype(bar()) {}

template <int N> __global__ kernel();

void launch_kernel() {
  kernel<sizeof(decltype(foo()))><<<...>>>();
}
```

This template instantiation had better be the same when compiling for host
and device.

That's contrived, but consider this much simpler case:

```
void host_fn() {
  static_assert(sizeof(decltype(foo())) == sizeof(int8));
}
```

If we let foo return int16 in device mode, this static_assert will fail
when compiling in *device* mode even though host_fn is never called on the
device.  https://gcc.godbolt.org/z/gYq901

Why are we doing sema checks on the host code when compiling for device?
See contrived example above, we need quite a bit of info about the host
code to infer those templates.

On Thu, May 2, 2019 at 7:05 PM Hal Finkel via Phabricator <
reviews at reviews.llvm.org> wrote:

> hfinkel added a comment.
>
> In D61458#1488970 <https://reviews.llvm.org/D61458#1488970>, @jlebar
> wrote:
>
> > Here's one for you:
> >
> >   __host__ float bar();
> >   __device__ int bar();
> >   __host__ __device__ auto foo() -> decltype(bar()) {}
> >
> >
> > What is the return type of `foo`?  :)
> >
> > I don't believe the right answer is, "float when compiling for host, int
> when compiling for device."
>
>
> So, actually, I wonder if that's not the right answer. We generally allow
> different overloads to have different return types. What if, for example,
> the return type on the host is __float128 and on the device it's
> `MyLongFloatTy`?
>
> > I'd be happy if we said this was an error, so long as it's well-defined
> what exactly we're disallowing.  But I bet @rsmith can come up with
> substantially more evil testcases than this.
>
>
>
>
> Repository:
>   rG LLVM Github Monorepo
>
> CHANGES SINCE LAST ACTION
>   https://reviews.llvm.org/D61458/new/
>
> https://reviews.llvm.org/D61458
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20190502/0c748d6d/attachment.html>