[PATCH] D119651: [clang] Fix evaluation context type for consteval function calls in instantiations

[Utkarsh Saxena via Phabricator via cfe-commits]

usaxena95 added a comment.
Herald added a project: All.

Hi, I have been trying to get consteval in a better shape.

In D119651#3327212 <https://reviews.llvm.org/D119651#3327212>, @Izaron wrote:

> After an investigation in GDB I can say that the assert seems to be wrong. Since Clang instantiates classes and functions "on the fly" where appropriate, we indeed can get a run-time evaluation context after a compile-time evaluation context. I was sure that evaluation contexts were made to represent a clean hierarchy of context, because they're interrelated, but the case with instantiations confuses me.

Can you rebase and verify if there are still assertion failures. I think the assertion is fair. Dependent immediate function invocations must be done during instantiation. I suspect https://reviews.llvm.org/D132031 might have fixed the issue.

> This code ...
>
>   template<int N>
>   struct good_struct {
>       // we are in run-time eval context!
>       static consteval int evalconst() {
>           // we are in compile-time eval context!
>           return N * N;
>       }
>   
>       void foo();
>       void bar();
>   };
>   
>   //int good_struct_100 = good_struct<100>::evalconst();
>   //int good_struct_200 = good_struct<200>::evalconst();
>   
>   consteval int consteval_foo() {
>       // we are in compile-time eval context!
>       return good_struct<100>::evalconst();
>   }
>   
>   template</* we are in compile-time eval context! */ int N = good_struct<200>::evalconst()>
>   constexpr int templated_foo() {
>       return N;
>   }
>
> ... hits the assert two times, unless we uncomment the lines with `good_struct_100` and `good_struct_200`. That's because Clang instantiates the classes when it "sees" them, straight from consteval/template contexts. I couldn't come up with a correct code that breaks though.
>
> I am now less sure if the patch (without the assert) is acceptable, what if the concept of "evaluation contexts" needs a revision?..

Can you also give a reproducer for which you see the assertion failure if you still see this.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D119651/new/

https://reviews.llvm.org/D119651