[OpenMP] Consider LIBRARY_PATH when selecting library paths for NVPTX targets in OpenMP mode.

[gtbercea]

Summary

Consider LIBRARY_PATH when selecting library paths for NVPTX targets in OpenMP device offloading mode.

[gtbercea]

Update patch to reflect latest source code changes.

[tstellar]

Why is this necessary?

[ABataev]

test/Driver/openmp-offload.c

622

Will it work on Windows?

[hfinkel]

In D29651#720179, @tstellar wrote:

Why is this necessary?

To put it another way, what is special about LIBRARY_PATH in this regard? Do we already pass paths specified with -L? I'm also slightly concerned here with the custom parsing code. Can the logic from tools::addDirectoryList not be reused here?

lib/Driver/ToolChains/Cuda.cpp

405 ↗

(On Diff #93188)

A comment is needed here explaining why this is necessary/desirable.

412 ↗

(On Diff #93188)

Is this the only use of this variable? How is this not dead code?

test/Driver/openmp-offload.c

626

Can this be a little more specific so that you actually make sure you're getting the offloading linking step here?

[gtbercea]

Not needed.
These changes are related to looking up the .bc library for inlining purposes. I believe @arpith-jacob has already handled this in trunk. Therefore this is obsolete code.

lib/Driver/ToolChains.cpp

void CudaToolChain::addClangTargetOptions(
   CC1Args.push_back("-mlink-cuda-bitcode");
   CC1Args.push_back(DriverArgs.MakeArgString(LibDeviceFile));
 
   // Libdevice in CUDA-7.0 requires PTX version that's more recent
   // than LLVM defaults to. Use PTX4.2 which is the PTX version that
   // came with CUDA-7.0.
   CC1Args.push_back("-target-feature");
   CC1Args.push_back("+ptx42");
+
+  if (DeviceOffloadingKind == Action::OFK_OpenMP) {
+    SmallVector<std::string, 8> LibraryPaths;
+    if (char *Env = ::getenv("LIBRARY_PATH")) {
+      StringRef CompilerPath = Env;
+      while (!CompilerPath.empty()) {
+        std::pair<StringRef, StringRef> Split =
+            CompilerPath.split(llvm::sys::EnvPathSeparator);
+        LibraryPaths.push_back(Split.first);
+        CompilerPath = Split.second;
+      }
+    }
+  }
 }
 
 void CudaToolChain::AddCudaIncludeArgs(const ArgList &DriverArgs,
                                        ArgStringList &CC1Args) const {
   // Check our CUDA version if we're going to include the CUDA headers.
   if (!DriverArgs.hasArg(options::OPT_nocudainc) &&
       !DriverArgs.hasArg(options::OPT_no_cuda_version_check)) {
     StringRef Arch = DriverArgs.getLastArgValue(options::OPT_march_EQ);

test/Driver/openmp-offload.c

 
 /// Check -fopenmp-is-device is also passed when generating the *.i and *.s intermediate files.
 // RUN:   %clang -### -fopenmp=libomp -fopenmp-targets=powerpc64le-ibm-linux-gnu -save-temps -no-canonical-prefixes %s 2>&1 \
 // RUN:   | FileCheck -check-prefix=CHK-FOPENMP-IS-DEVICE %s
 
 // CHK-FOPENMP-IS-DEVICE: clang{{.*}}.i" {{.*}}" "-fopenmp-is-device"
 // CHK-FOPENMP-IS-DEVICE-NEXT: clang{{.*}}.bc" {{.*}}.i" "-fopenmp-is-device" "-fopenmp-host-ir-file-path"
 // CHK-FOPENMP-IS-DEVICE-NEXT: clang{{.*}}.s" {{.*}}.bc" "-fopenmp-is-device"
+
+/// ###########################################################################
+
+/// Check that the lib folder pointed to by the LIBRARY_PATH is correctly passsed to the loader script.
+// RUN:   LIBRARY_PATH=/a/b/c/lib %clang -### -fopenmp=libomp -fopenmp-targets=powerpc64le-ibm-linux-gnu -save-temps -no-canonical-prefixes %s 2>&1 \
+// RUN:   | FileCheck -check-prefix=CHK-LIB-PATH %s
+
+// CHK-LIB-PATH: ld{{.*}}.o" "-L/a/b/c/lib"
+// CHK-LIB-PATH: ld{{.*}}.o" "-L/a/b/c/lib"