[clang] 7a19899 - [clang][CUDA] Search include/cccl of CUDA 13 installations (#235910)

via cfe-commits cfe-commits at lists.llvm.org
Thu Nov 5 15:49:43 PST 2026


Author: Caroline Rodewig
Date: 2026-11-06T00:49:38+01:00
New Revision: 7a1989913d44a9cad259806dd18d8a3c5a2d7342

URL: https://github.com/llvm/llvm-project/commit/7a1989913d44a9cad259806dd18d8a3c5a2d7342
DIFF: https://github.com/llvm/llvm-project/commit/7a1989913d44a9cad259806dd18d8a3c5a2d7342.diff

LOG: [clang][CUDA] Search include/cccl of CUDA 13 installations (#235910)

>From CUDA 13.0 on, the CUDA C++ Core Libraries (libcu++, CUB and Thrust)
are in `include/cccl` of the CUDA installation rather than in
`include/`. Some of the toolkit's own headers include them, so with CUDA
13 something as simple as `#include <cooperative_groups.h>` fails to
compile with Clang:

```
cooperative_groups/details/info.h: fatal error:
    'cuda/std/type_traits' file not found
```

nvcc searches `include/cccl` right after `include/`. This makes Clang do
the same, when the directory exists. `-nogpuinc` still leaves it out,
and the user's own include directories are still searched first.

The HIP toolchain already does the equivalent for
[`include/libhipcxx`](https://github.com/llvm/llvm-project/blob/fe84f6e3b0ae40ef28998cd9cee3220eec2ea261/clang/lib/Driver/ToolChains/AMDGPU.cpp#L617-L624).

Assisted-by: Devin

Added: 
    clang/test/Driver/Inputs/CUDA_130/usr/local/cuda/bin/.keep
    clang/test/Driver/Inputs/CUDA_130/usr/local/cuda/include/cccl/.keep
    clang/test/Driver/Inputs/CUDA_130/usr/local/cuda/include/cuda.h
    clang/test/Driver/Inputs/CUDA_130/usr/local/cuda/lib/.keep
    clang/test/Driver/Inputs/CUDA_130/usr/local/cuda/lib64/.keep
    clang/test/Driver/Inputs/CUDA_130/usr/local/cuda/nvvm/libdevice/libdevice.10.bc
    clang/test/Driver/cuda-cccl-include.cu

Modified: 
    clang/docs/ReleaseNotes.md
    clang/lib/Driver/ToolChains/Cuda.cpp

Removed: 
    


################################################################################
diff  --git a/clang/docs/ReleaseNotes.md b/clang/docs/ReleaseNotes.md
index 03c5b85c32c13e..ea9fbe00ae3fe2 100644
--- a/clang/docs/ReleaseNotes.md
+++ b/clang/docs/ReleaseNotes.md
@@ -1410,6 +1410,12 @@ documentation](CIR/index.md) for more information regarding ClangIR's status and
   functions with CUDA 13.4 and newer. Added `__nvvm_idp4a_*` and
   `__nvvm_idp2a_*` builtins. (#GH233434)
 
+- CUDA compilations now add the `include/cccl` directory of the selected CUDA
+  installation to the header search path when it exists. From CUDA 13.0 on,
+  libcu++, CUB and Thrust are in that directory, and toolkit headers such as
+  `cooperative_groups.h` include them. The `-nogpuinc` option disables this
+  path together with the other CUDA include paths.
+
 #### AIX Support
 
 #### NetBSD Support

diff  --git a/clang/lib/Driver/ToolChains/Cuda.cpp b/clang/lib/Driver/ToolChains/Cuda.cpp
index acbbbb3b72c048..57470c18db3930 100644
--- a/clang/lib/Driver/ToolChains/Cuda.cpp
+++ b/clang/lib/Driver/ToolChains/Cuda.cpp
@@ -328,6 +328,15 @@ void CudaInstallationDetector::AddCudaIncludeArgs(
 
   CC1Args.push_back("-include");
   CC1Args.push_back("__clang_cuda_runtime_wrapper.h");
+
+  // From CUDA 13.0 on, libcu++, CUB and Thrust are in include/cccl, and some
+  // of the toolkit's own headers, such as cooperative_groups.h, include them.
+  SmallString<128> CCCLPath(getIncludePath());
+  llvm::sys::path::append(CCCLPath, "cccl");
+  if (D.getVFS().exists(CCCLPath)) {
+    CC1Args.push_back("-internal-isystem");
+    CC1Args.push_back(DriverArgs.MakeArgString(CCCLPath));
+  }
 }
 
 void CudaInstallationDetector::CheckCudaVersionSupportsArch(

diff  --git a/clang/test/Driver/Inputs/CUDA_130/usr/local/cuda/bin/.keep b/clang/test/Driver/Inputs/CUDA_130/usr/local/cuda/bin/.keep
new file mode 100644
index 00000000000000..e69de29bb2d1d6

diff  --git a/clang/test/Driver/Inputs/CUDA_130/usr/local/cuda/include/cccl/.keep b/clang/test/Driver/Inputs/CUDA_130/usr/local/cuda/include/cccl/.keep
new file mode 100644
index 00000000000000..e69de29bb2d1d6

diff  --git a/clang/test/Driver/Inputs/CUDA_130/usr/local/cuda/include/cuda.h b/clang/test/Driver/Inputs/CUDA_130/usr/local/cuda/include/cuda.h
new file mode 100644
index 00000000000000..9ddfd9b011fdac
--- /dev/null
+++ b/clang/test/Driver/Inputs/CUDA_130/usr/local/cuda/include/cuda.h
@@ -0,0 +1,7 @@
+//
+// Placeholder file for testing CUDA version detection
+//
+
+#define CUDA_VERSION 13000
+
+//

diff  --git a/clang/test/Driver/Inputs/CUDA_130/usr/local/cuda/lib/.keep b/clang/test/Driver/Inputs/CUDA_130/usr/local/cuda/lib/.keep
new file mode 100644
index 00000000000000..e69de29bb2d1d6

diff  --git a/clang/test/Driver/Inputs/CUDA_130/usr/local/cuda/lib64/.keep b/clang/test/Driver/Inputs/CUDA_130/usr/local/cuda/lib64/.keep
new file mode 100644
index 00000000000000..e69de29bb2d1d6

diff  --git a/clang/test/Driver/Inputs/CUDA_130/usr/local/cuda/nvvm/libdevice/libdevice.10.bc b/clang/test/Driver/Inputs/CUDA_130/usr/local/cuda/nvvm/libdevice/libdevice.10.bc
new file mode 100644
index 00000000000000..e69de29bb2d1d6

diff  --git a/clang/test/Driver/cuda-cccl-include.cu b/clang/test/Driver/cuda-cccl-include.cu
new file mode 100644
index 00000000000000..72406d5ff5a057
--- /dev/null
+++ b/clang/test/Driver/cuda-cccl-include.cu
@@ -0,0 +1,27 @@
+// From CUDA 13.0 on, the CUDA C++ Core Libraries are in include/cccl of the
+// CUDA installation. Check that it is a system include directory for both the
+// device and the host compilation.
+
+// RUN: %clang -### --target=x86_64-unknown-linux-gnu --cuda-gpu-arch=sm_75 \
+// RUN:   --cuda-path=%S/Inputs/CUDA_130/usr/local/cuda %s 2>&1 \
+// RUN:   | FileCheck %s --check-prefix=CCCL
+
+// CCCL: "-cc1" "-triple" "nvptx64-nvidia-cuda"
+// CCCL-SAME: "-internal-isystem" "[[CUDA:[^"]*Inputs/CUDA_130/usr/local/cuda]]{{[/\\]+}}include{{[/\\]+}}cccl" {{.*}}"-internal-isystem" "[[CUDA]]{{[/\\]+}}include"
+// CCCL: "-cc1" "-triple" "x86_64-unknown-linux-gnu"
+// CCCL-SAME: "-internal-isystem" "[[CUDA]]{{[/\\]+}}include{{[/\\]+}}cccl" {{.*}}"-internal-isystem" "[[CUDA]]{{[/\\]+}}include"
+
+// Older CUDA installations have no include/cccl.
+// RUN: %clang -### --target=x86_64-unknown-linux-gnu --cuda-gpu-arch=sm_75 \
+// RUN:   --cuda-path=%S/Inputs/CUDA_111/usr/local/cuda %s 2>&1 \
+// RUN:   | FileCheck %s --check-prefix=NOCCCL
+
+// -nogpuinc leaves out all of the CUDA installation's include directories.
+// RUN: %clang -### --target=x86_64-unknown-linux-gnu --cuda-gpu-arch=sm_75 \
+// RUN:   -nogpuinc --cuda-path=%S/Inputs/CUDA_130/usr/local/cuda %s 2>&1 \
+// RUN:   | FileCheck %s --check-prefix=NOCCCL
+
+// NOCCCL: "-cc1" "-triple" "nvptx64-nvidia-cuda"
+// NOCCCL-NOT: include{{[/\\]+}}cccl"
+// NOCCCL: "-cc1" "-triple" "x86_64-unknown-linux-gnu"
+// NOCCCL-NOT: include{{[/\\]+}}cccl"


        


More information about the cfe-commits mailing list