[llvm-branch-commits] [clang] [Clang] Improve type traits recognition in `__has_builtin` (#111516) (PR #111660)

[Erich Keane via llvm-branch-commits]

erichkeane wrote:

Note that the NVCC thing here is that NVCC has an interesting programming model for its offload.

My understanding is that: 
-Uses the (in this case) Clang Preprocessor to process the code.
-Does some messing around with that preprocessed output, generating the device code, plus a modified 'host' preprocessed file.
-Compiles that preprocessed/modified file with the host (Clang in this case).

Those errors LOOK like they came out of Clang, and the bug report has the EDG Frontend of NVCC trying to emulate Clang 20 (`--clang_version=200000`).

That said, if Godbolt is correct, this might be a regression in our handling of that builtin: https://godbolt.org/z/r8od54M1P
Or is this patch FIXING that regression?

This makes me wonder if NVCC is somehow using 2 different versions of Clang here, 1 to do preprocessing that enables this, and an older version of 19.1.x to do the actual compilation?

@t-lutz is a member at Nvidia on the NVCC team that might be able to help.

https://github.com/llvm/llvm-project/pull/111660