[clang] Ensure we don't process OpenCL kernels as CUDA kernels

[sarnex]

We've decided the OpenCL spelling of the DeviceKernel attr needs separate handling, and add a case to that special handing for NVPTX target.

I wasn't able to come up with a test with any actual difference in behavior.

It's possible this is NFC but I'm not sure so I'm not adding the tag.

[llvmbot]

@llvm/pr-subscribers-clang

Author: Nick Sarnie (sarnex)

Changes

We've decided the OpenCL spelling of the DeviceKernel attr needs separate handling, and add a case to that special handing for NVPTX target.

I wasn't able to come up with a test with any actual difference in behavior.

It's possible this is NFC but I'm not sure so I'm not adding the tag.

---

Full diff: https://github.com/llvm/llvm-project/pull/163859.diff

1 Files Affected:

• (modified) clang/lib/Sema/SemaDeclAttr.cpp (+2-1)

diff --git a/clang/lib/Sema/SemaDeclAttr.cpp b/clang/lib/Sema/SemaDeclAttr.cpp
index e6f8748db7644..6978299734ece 100644
--- a/clang/lib/Sema/SemaDeclAttr.cpp
+++ b/clang/lib/Sema/SemaDeclAttr.cpp
@@ -5206,7 +5206,8 @@ static void handleCallConvAttr(Sema &S, Decl *D, const ParsedAttr &AL) {
 static void handleDeviceKernelAttr(Sema &S, Decl *D, const ParsedAttr &AL) {
   const auto *FD = dyn_cast_or_null<FunctionDecl>(D);
   bool IsFunctionTemplate = FD && FD->getDescribedFunctionTemplate();
-  if (S.getASTContext().getTargetInfo().getTriple().isNVPTX()) {
+  if (S.getASTContext().getTargetInfo().getTriple().isNVPTX() &&
+      !DeviceKernelAttr::isOpenCLSpelling(AL)) {
     handleGlobalAttr(S, D, AL);
   } else {
     // OpenCL C++ will throw a more specific error.

[tahonermann]

I think the conditions for each branch can be expressed more cleanly per my inline comment.

[tahonermann]

I think this can be made easier to read and maintain by reversing the order of the if/else branches.

if (DeviceKernelAttr::isOpenCLSpelling(AL)) {
  // Do OpenCL things...
} else if (S.getASTContext().getTargetInfo().getTriple().isNVPTX()) {
  // Do presumed CUDA/HIP kernel things, but only on NVPTX for some reason...
}

[sarnex]

Thanks for the review. We go into the else branch for OpenCL spelling or non-NVPTX, so I don't think we can only check for the OpenCL spelling if we swap that condition to be the if. I made an attempt in my latest commit, please take a look when you have a chance.