[clang] 921fdd3 - [Clang][SYCL] Reject an NVIDIA offload target for SYCL (#229937)
via cfe-commits
cfe-commits at lists.llvm.org
Tue Nov 3 10:58:21 PST 2026
Author: Srividya Sundaram
Date: 2026-11-03T10:58:15-08:00
New Revision: 921fdd3ec5c0e4510e8f8c7473630dae7dbc54b7
URL: https://github.com/llvm/llvm-project/commit/921fdd3ec5c0e4510e8f8c7473630dae7dbc54b7
DIFF: https://github.com/llvm/llvm-project/commit/921fdd3ec5c0e4510e8f8c7473630dae7dbc54b7.diff
LOG: [Clang][SYCL] Reject an NVIDIA offload target for SYCL (#229937)
Reject an NVIDIA offload target for SYCL.
Previously, `clang++ -fsycl --offload-arch=sm_80`, or
`--offload-targets=nvptx64-nvidia-cuda`, hit an assertion in the CUDA
toolchain, which supports only CUDA and OpenMP. Clang does not support
SYCL offloading to NVIDIA GPUs without `-foffload-via-llvm`, so the
driver now reports "invalid or unsupported offload target" instead. With
`-foffload-via-llvm`, SYCL offloading to an NVIDIA GPU is unchanged.
Added:
Modified:
clang/lib/Driver/Driver.cpp
clang/test/Driver/sycl-offload-arch-igca.cpp
clang/test/Driver/sycl-offload-jit.cpp
Removed:
################################################################################
diff --git a/clang/lib/Driver/Driver.cpp b/clang/lib/Driver/Driver.cpp
index 37fb69baf6a3e..ce28ca4934a9a 100644
--- a/clang/lib/Driver/Driver.cpp
+++ b/clang/lib/Driver/Driver.cpp
@@ -1167,6 +1167,8 @@ void Driver::CreateOffloadingDeviceToolChains(Compilation &C,
// Build an offloading toolchain for every requested target and kind.
llvm::StringMap<StringRef> FoundNormalizedTriples;
+ bool UsesLLVMOffloading = C.getInputArgs().hasFlag(
+ options::OPT_foffload_via_llvm, options::OPT_fno_offload_via_llvm, false);
for (const llvm::Triple &Target : Triples) {
// OpenMP offloading requires a compatible libomp.
if (Kinds.contains(Action::OFK_OpenMP)) {
@@ -1184,6 +1186,14 @@ void Driver::CreateOffloadingDeviceToolChains(Compilation &C,
if (Arg *IncompatArg = C.getInputArgs().getLastArg(ID))
Diag(clang::diag::err_drv_argument_not_allowed_with)
<< IncompatArg->getSpelling() << "-fsycl";
+
+ // The CUDA toolchain supports only CUDA and OpenMP, so diagnose SYCL
+ // offloading to an NVIDIA GPU here, unless it uses -foffload-via-llvm.
+ if (Target.isNVPTX() && !UsesLLVMOffloading) {
+ Diag(diag::err_drv_invalid_or_unsupported_offload_target)
+ << Target.str();
+ continue;
+ }
}
// Create a device toolchain for every specified kind and triple.
diff --git a/clang/test/Driver/sycl-offload-arch-igca.cpp b/clang/test/Driver/sycl-offload-arch-igca.cpp
index 7a3a2b4f7d7ac..f8646c0faef4d 100644
--- a/clang/test/Driver/sycl-offload-arch-igca.cpp
+++ b/clang/test/Driver/sycl-offload-arch-igca.cpp
@@ -20,12 +20,13 @@
// CHK-FEATURE: error: unsupported IGCA gpu architecture: igca_20ca:foo
// CHK-FEATURE-NOT: llvm-offload-binary
-/// A valid IGCA target is still rejected on NVPTX and AMDGPU device triples,
-/// with the existing CUDA/HIP diagnostic:
+/// A valid IGCA target is still rejected on NVPTX and AMDGPU device triples.
+/// SYCL rejects an NVIDIA target itself, and AMDGPU uses the existing HIP
+/// diagnostic:
// RUN: not %clang -### --target=x86_64-unknown-linux-gnu -fsycl \
// RUN: --offload-targets=nvptx64-nvidia-cuda --offload-arch=igca_20ca \
// RUN: -c %s 2>&1 | FileCheck -check-prefix=CHK-NVPTX %s
-// CHK-NVPTX: error: unsupported CUDA gpu architecture: igca_20ca
+// CHK-NVPTX: error: invalid or unsupported offload target: 'nvptx64-nvidia-cuda'
// CHK-NVPTX-NOT: llvm-offload-binary
// RUN: not %clang -### --target=x86_64-unknown-linux-gnu -fsycl \
diff --git a/clang/test/Driver/sycl-offload-jit.cpp b/clang/test/Driver/sycl-offload-jit.cpp
index ab38aad61ca6d..40b6a9ede9ed8 100644
--- a/clang/test/Driver/sycl-offload-jit.cpp
+++ b/clang/test/Driver/sycl-offload-jit.cpp
@@ -95,6 +95,21 @@
// CHK-PHASES-NORDC-ARCHS-NEXT: 13: llvm-offload-binary, {7, 12}, image, (device-sycl)
// CHK-PHASES-NORDC-ARCHS-NEXT: 14: clang-linker-wrapper, {13}, sycl-fatbin, (device-sycl)
+/// Clang does not support SYCL offloading to NVIDIA GPUs without
+/// -foffload-via-llvm.
+// RUN: not %clang -### --target=x86_64-unknown-linux-gnu -fsycl \
+// RUN: --offload-arch=sm_80 -c %s 2>&1 \
+// RUN: | FileCheck -check-prefix=CHK-NVPTX %s
+// RUN: not %clang -### --target=x86_64-unknown-linux-gnu -fsycl \
+// RUN: --offload-targets=nvptx64-nvidia-cuda -c %s 2>&1 \
+// RUN: | FileCheck -check-prefix=CHK-NVPTX %s
+// CHK-NVPTX: error: invalid or unsupported offload target: 'nvptx64-nvidia-cuda'
+// RUN: %clang -### --target=x86_64-unknown-linux-gnu -fsycl -foffload-via-llvm \
+// RUN: --offload-arch=sm_80 -c %s 2>&1 \
+// RUN: | FileCheck -check-prefix=CHK-NVPTX-VIA-LLVM %s
+// CHK-NVPTX-VIA-LLVM-NOT: error:
+// CHK-NVPTX-VIA-LLVM: "-cc1" "-triple" "nvptx64-nvidia-cuda-llvm"
+
/// Multiple device triples are not supported today in non-RDC mode.
// RUN: not %clang -### --target=x86_64-unknown-linux-gnu -fsycl -fno-gpu-rdc \
// RUN: --offload-targets=spirv64-unknown-unknown,spirv32-unknown-unknown -c %s 2>&1 \
More information about the cfe-commits
mailing list