[PATCH] D56411: [CUDA][HIP][Sema] Fix template kernel with function as template parameter

[Yaxun Liu via Phabricator via cfe-commits]

yaxunl added a comment.

In D56411#1410153 <https://reviews.llvm.org/D56411#1410153>, @rjmccall wrote:

> In D56411#1406212 <https://reviews.llvm.org/D56411#1406212>, @yaxunl wrote:
>
> > I would like to fix the validation issue only and leave the overload resolution issue for future.
>
>
> As I understand it, the "validation issue" is just that you'd like a diagnostic to be emitted when resolving the template argument in order to force SFINAE to pick a different template.  I think that's actually just the overload-resolution issue.


Currently there are two host-ness related issues about function type template arguments:

1. when there are two or more candidates for the template argument, clang goes through host-ness based overloading resolution, which does not work properly

2. when there is only one candidate for the template argument, clang does not go through overloading resolution, therefore the first issue does not show up. However, clang still checks host-ness of template argument. As discussed before, clang should not check host-ness in non-evaluation or constant-evaluation context. Instead, clang should check host-ness in template instantiation.

I refer the first issue as host-ness overloading resolution issue and the second issue as host-ness validation issue. They are related but separate.

The first issue only happens when host-ness based overloading resolution is used. For applications which can be compiled with nvcc, this cannot happen, therefore it is less common and less urgent.

The second issue can happen to applications which can be compiled with nvcc, therefore is more imminent.

Fixing the second issue is relatively straightforward. It does not need to introduce new AST types for host-ness. Also it is orthogonal to fixing the first issue.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D56411/new/

https://reviews.llvm.org/D56411