[clang] 2323019 - [clang-repl] Pass the target features to the device TargetMachine (#235754)
via cfe-commits
cfe-commits at lists.llvm.org
Sat Oct 31 00:37:15 PDT 2026
Author: Aaron Jomy
Date: 2026-10-31T07:37:10Z
New Revision: 2323019b73220fd87aaf28175e830dca67fa4cdd
URL: https://github.com/llvm/llvm-project/commit/2323019b73220fd87aaf28175e830dca67fa4cdd
DIFF: https://github.com/llvm/llvm-project/commit/2323019b73220fd87aaf28175e830dca67fa4cdd.diff
LOG: [clang-repl] Pass the target features to the device TargetMachine (#235754)
IncrementalCUDADeviceParser creates the NVPTX TargetMachine with an
empty feature string. The +ptxNN feature that the driver derives from
the CUDA version is dropped, causing the backend to emit the minimum PTX
version for the target SM instead. A kernel using a newer instruction
then fails to load with cudaErrorInvalidPtx, for example shfl.sync below
sm_70.
Join TargetOptions::Features into the feature string, as done in
EmitAssemblyHelper::CreateTargetMachine.
Add a unittest that passes --cuda-feature=+ptx80 to the device
compilation and checks the .version of the embedded PTX.
Added:
clang/test/Interpreter/CUDA/ptx-version.cu
Modified:
clang/lib/Interpreter/DeviceOffload.cpp
Removed:
################################################################################
diff --git a/clang/lib/Interpreter/DeviceOffload.cpp b/clang/lib/Interpreter/DeviceOffload.cpp
index 94fec23b1c0b2f..d7ae386466f40f 100644
--- a/clang/lib/Interpreter/DeviceOffload.cpp
+++ b/clang/lib/Interpreter/DeviceOffload.cpp
@@ -19,6 +19,7 @@
#include "clang/Frontend/CompilerInstance.h"
#include "clang/Interpreter/PartialTranslationUnit.h"
+#include "llvm/ADT/StringExtras.h"
#include "llvm/IR/LegacyPassManager.h"
#include "llvm/IR/Module.h"
#include "llvm/MC/TargetRegistry.h"
@@ -122,8 +123,8 @@ llvm::Expected<llvm::StringRef> IncrementalCUDADeviceParser::GeneratePTX() {
std::error_code());
llvm::TargetOptions TO = llvm::TargetOptions();
std::unique_ptr<llvm::TargetMachine> TM(Target->createTargetMachine(
- PTU.TheModule->getTargetTriple(), TargetOpts.CPU, "", TO,
- llvm::Reloc::Model::PIC_));
+ PTU.TheModule->getTargetTriple(), TargetOpts.CPU,
+ llvm::join(TargetOpts.Features, ","), TO, llvm::Reloc::Model::PIC_));
PTXCode.clear();
llvm::raw_svector_ostream dest(PTXCode);
diff --git a/clang/test/Interpreter/CUDA/ptx-version.cu b/clang/test/Interpreter/CUDA/ptx-version.cu
new file mode 100644
index 00000000000000..536fd37de458ea
--- /dev/null
+++ b/clang/test/Interpreter/CUDA/ptx-version.cu
@@ -0,0 +1,14 @@
+// The PTX version of the embedded fatbin must follow the pinned feature.
+// RUN: rm -rf %t
+// RUN: mkdir -p %t/lib
+// RUN: %clang --target=%host-jit-triple -nostdlib -fPIC -shared \
+// RUN: %S/Inputs/cudart.c -o %t/lib/libcudart.so
+// RUN: cat %s | clang-repl --cuda --cuda-path=%t -Xcc -nocudainc \
+// RUN: -Xcc -nocudalib -Xcc -include -Xcc %S/Inputs/cuda.h \
+// RUN: -Xcc -Xarch_device -Xcc --cuda-feature=+ptx80 \
+// RUN: -Xcc -Xclang -Xcc -emit-llvm | FileCheck %s
+
+__global__ void kernel() {}
+// CHECK: private constant {{.*}}.version 8.0{{.*}}.target sm_
+
+%quit
More information about the cfe-commits
mailing list