Add upstream driver steps.

The default driver is the current driver i.e. the one using the opaque linker.

To use the upstream driver i.e. the one using the clang-linker-wrapper set the following environment variable to 1:

export CLANG_USE_LINKER_WRAPPER=1

Change-Id: I13f542fd1afc16eedfcf4c4c7d1b9d0d72daa059

Author: doru1004

clang/lib/Driver/ToolChains/AMDGPU.cpp

@@ -665,7 +665,11 @@ void amdgpu::getAMDGPUTargetFeatures(const Driver &D,
                                      StringRef TcTargetID) {
   // Add target ID features to -target-feature options. No diagnostics should
   // be emitted here since invalid target ID is diagnosed at other places.
-  StringRef TargetID = Args.getLastArgValue(options::OPT_mcpu_EQ);
+  StringRef TargetID;
+  if (Args.hasArg(options::OPT_mcpu_EQ))
+    TargetID = Args.getLastArgValue(options::OPT_mcpu_EQ);
+  else if (Args.hasArg(options::OPT_march_EQ))
+    TargetID = Args.getLastArgValue(options::OPT_march_EQ);
 
   // Use this toolchain's TargetID if mcpu is not defined
   if (TargetID.empty() && !TcTargetID.empty())

clang/lib/Driver/ToolChains/AMDGPUOpenMP.cpp

@@ -402,8 +402,8 @@ void AMDGPUOpenMPToolChain::addClangTargetOptions(
   StringRef GPUArch = DriverArgs.getLastArgValue(options::OPT_march_EQ);
   assert(!GPUArch.empty() && "Must have an explicit GPU arch.");
 
-  CC1Args.push_back("-target-cpu");
-  CC1Args.push_back(DriverArgs.MakeArgStringRef(GPUArch));
+  assert(DeviceOffloadingKind == Action::OFK_OpenMP &&
+         "Only OpenMP offloading kinds are supported.");
 
   // Extract all the -m options
   std::vector<llvm::StringRef> Features;
@@ -435,7 +435,10 @@ void AMDGPUOpenMPToolChain::addClangTargetOptions(
                          options::OPT_fno_gpu_allow_device_init, false))
     CC1Args.push_back("-fgpu-allow-device-init");
 
-  CC1Args.push_back("-fcuda-allow-variadic-functions");
+  // TODO: check if flag is needed for the opaque linker case
+  const char *UseLinkerWrapper = std::getenv("CLANG_USE_LINKER_WRAPPER");
+  if (!UseLinkerWrapper || atoi(UseLinkerWrapper) == 0)
+    CC1Args.push_back("-fcuda-allow-variadic-functions");
 
   // Default to "hidden" visibility, as object level linking will not be
   // supported for the foreseeable future.

clang/lib/Driver/ToolChains/Clang.cpp

@@ -9577,9 +9577,15 @@ void LinkerWrapper::ConstructJob(Compilation &C, const JobAction &JA,
     }
   }
 
-  if (!OpenMPTCs.empty() &&
-      Args.hasFlag(options::OPT_opaque_offload_linker,
-                   options::OPT_no_opaque_offload_linker, isAMDGPU)) {
+  // Overwrite the default driver choice using an env var:
+  // CLANG_USE_LINKER_WRAPPER which can be 0 or 1.
+  bool UseOpaqueOffloadLinker = isAMDGPU;
+  if (const char *UseLinkerWrapper = std::getenv("CLANG_USE_LINKER_WRAPPER"))
+    UseOpaqueOffloadLinker = !(atoi(UseLinkerWrapper) == 1);
+
+  if (!OpenMPTCs.empty() && Args.hasFlag(options::OPT_opaque_offload_linker,
+                                         options::OPT_no_opaque_offload_linker,
+                                         UseOpaqueOffloadLinker)) {
     ConstructOpaqueJob(C, JA, Output, Inputs, Args, TC->getTriple(),
                        LinkingOutput);
     return;

clang/test/Driver/amdgpu-openmp-toolchain.c

@@ -94,7 +94,8 @@
 
 // RUN: %clang -### -target x86_64-pc-linux-gnu -fopenmp --offload-arch=gfx90a:sramecc-:xnack+ \
 // RUN:   -nogpulib %s 2>&1 | FileCheck %s --check-prefix=CHECK-TARGET-ID
-// CHECK-TARGET-ID: "-cc1" "-triple" "amdgcn-amd-amdhsa" {{.*}} "-target-cpu" "gfx90a:sramecc-:xnack+" "-target-feature" "-sramecc" "-target-feature" "+xnack"
+// CHECK-TARGET-ID: "-cc1" "-triple" "amdgcn-amd-amdhsa" {{.*}} "-target-cpu" "gfx90a" "-target-feature" "-sramecc" "-target-feature" "+xnack"
+
 // CHECK-TARGET-ID: clang-offload-packager{{.*}}arch=gfx90a:sramecc-:xnack+,kind=openmp,feature=-sramecc,feature=+xnack
 
 // RUN: not %clang -### -target x86_64-pc-linux-gnu -fopenmp --offload-arch=gfx90a,gfx90a:xnack+ \

clang/test/Driver/openmp-runtimelib.c

@@ -1,8 +1,8 @@
 // REQUIRES: amdgpu-registered-target
 
-// Asan-Debug: /lib-debug/asan/libomptarget
-// Asan-Devel: /lib/asan/libomptarget
-// Asan-Perf: /lib-perf/asan/libomptarget
+// Asan-Debug: /lib-debug/asan
+// Asan-Devel: /lib/asan
+// Asan-Perf: /lib-perf/asan
 
 // RUN:  %clang -### -fopenmp -nogpuinc -nogpulib  --offload-arch=gfx90a -fopenmp-runtimelib=lib-debug %s -O3 2>&1 \
 // RUN:   | FileCheck -check-prefixes=Debug %s
@@ -31,8 +31,8 @@
 // RUN: %clang -### -fopenmp -nogpuinc -nogpulib  --offload-arch=gfx90a:xnack+ -fopenmp-target-fast -fsanitize=address -shared-libasan %s -O3 2>&1 \
 // RUN:   | FileCheck -check-prefix=Asan-Devel %s
 
-// Debug: /lib-debug/libomptarget
-// Perf: /lib-perf/libomptarget
-// Devel: /lib/../runtimes/runtimes-bins/offload/libomptarget
-// Default: /lib/../runtimes/runtimes-bins/offload/libomptarget
-// Error: clang: error: unsupported argument
+// Debug: /lib-debug
+// Perf: /lib-perf
+// Devel: /../lib
+// Default: /../lib
+// Error: clang: error: unsupported argument 'oopsy' to option '-fopenmp-runtimelib='

clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp

@@ -1211,27 +1211,10 @@ DerivedArgList getLinkerArgs(ArrayRef<OffloadFile> Input,
 
   // Set the subarchitecture and target triple for this compilation.
   const OptTable &Tbl = getOptTable();
-  StringRef Triple = Input.front().getBinary()->getTriple();
-  std::string AMDGPUFeatures;
-
-  if (llvm::Triple(Triple).isAMDGPU()) {
-    // Extract Features from the binary and append them in arch
-    auto Features = getTargetFeatures(Input);
-    for (auto feature : Features) {
-      AMDGPUFeatures.append(feature.substr(1, feature.size()) +
-                            feature.substr(0, 1) + ":");
-    }
-  }
+  DAL.AddJoinedArg(nullptr, Tbl.getOption(OPT_arch_EQ),
+                   Args.MakeArgString(Input.front().getBinary()->getArch()));
   DAL.AddJoinedArg(nullptr, Tbl.getOption(OPT_triple_EQ),
-                   Args.MakeArgString(Triple));
-  if (llvm::Triple(Triple).isAMDGPU() && !AMDGPUFeatures.empty())
-    DAL.AddJoinedArg(nullptr, Tbl.getOption(OPT_arch_EQ),
-                     Args.MakeArgString(
-                         Input.front().getBinary()->getArch() + ":" +
-                         AMDGPUFeatures.substr(0, AMDGPUFeatures.size() - 1)));
-  else
-    DAL.AddJoinedArg(nullptr, Tbl.getOption(OPT_arch_EQ),
-                     Args.MakeArgString(Input.front().getBinary()->getArch()));
+                   Args.MakeArgString(Input.front().getBinary()->getTriple()));
 
   // If every input file is bitcode we have whole program visibility as we
   // do only support static linking with bitcode.

llvm/lib/Transforms/IPO/OpenMPOpt.cpp

@@ -5598,13 +5598,12 @@ void OpenMPOpt::registerAAsForFunction(Attributor &A, const Function &F) {
                              UsedAssumedInformation, AA::Interprocedural);
       continue;
     }
-#if 0 // fixme snap2 mi-teams nest_call_par2
     if (auto *CI = dyn_cast<CallBase>(&I)) {
-      if (CI->isIndirectCall())
+      if (CI->isIndirectCall() && !F.getName().contains("__kmpc_parallel_51") &&
+          !F.getName().contains("__kmpc_parallel_spmd"))
         A.getOrCreateAAFor<AAIndirectCallInfo>(
             IRPosition::callsite_function(*CI));
     }
-#endif
     if (auto *SI = dyn_cast<StoreInst>(&I)) {
       A.getOrCreateAAFor<AAIsDead>(IRPosition::value(*SI));
       continue;

offload/DeviceRTL/CMakeLists.txt

@@ -123,7 +123,7 @@ set(src_files
 # propagation. That said, we will run the vectorizer again after the runtime
 # has been linked into the user program.
 set(clang_opt_flags -O3 -mllvm -openmp-opt-disable -DSHARED_SCRATCHPAD_SIZE=512 -mllvm -vectorize-slp=false )
-set(link_opt_flags  -O3        -openmp-opt-disable -vectorize-slp=false )
+set(link_opt_flags  -O3        -openmp-opt-disable -attributor-enable=module -vectorize-slp=false )
 set(link_export_flag -passes=internalize -internalize-public-api-file=${source_directory}/exports)
 
 # If the user built with the GPU C library enabled we will use that instead.

offload/DeviceRTL/include/State.h

@@ -116,7 +116,7 @@ struct ThreadStateTy {
   }
 };
 
-extern ThreadStateTy *ThreadStates[mapping::MaxThreadsPerTeam];
+extern ThreadStateTy **ThreadStates;
 #pragma omp allocate(ThreadStates) allocator(omp_pteam_mem_alloc)
 
 /// Initialize the state machinery. Must be called by all threads.

offload/DeviceRTL/include/extra_allocators.h

@@ -88,7 +88,6 @@ void __kmpc_flush_acqrel(IdentTy *Loc);
 void *__kmpc_task_allow_completion_event(IdentTy *loc_ref, uint32_t gtid,
                                          TaskDescriptorTy *task);
 ///}
-
 } // extern "C"
 
 /// Extra API exposed by ROCm

offload/DeviceRTL/src/State.cpp

@@ -267,10 +267,7 @@ void state::TeamStateTy::assertEqual(TeamStateTy &Other) const {
 }
 
 state::TeamStateTy SHARED(ompx::state::TeamState);
-
-__attribute__((loader_uninitialized))
-state::ThreadStateTy *ompx::state::ThreadStates[mapping::MaxThreadsPerTeam];
-#pragma omp allocate(ompx::state::ThreadStates) allocator(omp_pteam_mem_alloc)
+state::ThreadStateTy **SHARED(ompx::state::ThreadStates);
 
 namespace {
 
@@ -294,11 +291,10 @@ void state::init(bool IsSPMD, KernelEnvironmentTy &KernelEnvironment,
   SharedMemorySmartStack.init(IsSPMD);
   if (mapping::isInitialThreadInLevel0(IsSPMD)) {
     TeamState.init(IsSPMD);
+    ThreadStates = nullptr;
     KernelEnvironmentPtr = &KernelEnvironment;
     KernelLaunchEnvironmentPtr = &KernelLaunchEnvironment;
   }
-
-  ThreadStates[mapping::getThreadIdInBlock()] = nullptr;
 }
 
 KernelEnvironmentTy &state::getKernelEnvironment() {
@@ -312,11 +308,12 @@ KernelLaunchEnvironmentTy &state::getKernelLaunchEnvironment() {
 void state::enterDataEnvironment(IdentTy *Ident) {
   ASSERT(config::mayUseThreadStates(),
          "Thread state modified while explicitly disabled!");
+  if (!config::mayUseThreadStates())
+    return;
 
   unsigned TId = mapping::getThreadIdInBlock();
   ThreadStateTy *NewThreadState = static_cast<ThreadStateTy *>(
       memory::allocGlobal(sizeof(ThreadStateTy), "ThreadStates alloc"));
-#ifdef FIXME // breaks snap_red nested_par3 nest_call_par2
   uintptr_t *ThreadStatesBitsPtr = reinterpret_cast<uintptr_t *>(&ThreadStates);
   if (!atomic::load(ThreadStatesBitsPtr, atomic::seq_cst)) {
     uint32_t Bytes =
@@ -332,7 +329,6 @@ void state::enterDataEnvironment(IdentTy *Ident) {
     ASSERT(atomic::load(ThreadStatesBitsPtr, atomic::seq_cst),
            "Expected valid thread states bit!");
   }
- #endif
   NewThreadState->init(ThreadStates[TId]);
   TeamState.HasThreadState = true;
   ThreadStates[TId] = NewThreadState;
@@ -347,6 +343,9 @@ void state::exitDataEnvironment() {
 }
 
 void state::resetStateForThread(uint32_t TId) {
+  if (!config::mayUseThreadStates())
+    return;
+
   if (OMP_LIKELY(!TeamState.HasThreadState || !ThreadStates[TId]))
     return;
 

offload/test/api/assert.c

@@ -1,4 +1,3 @@
-// XFAIL: amdgcn-amd-amdhsa
 // RUN: %libomptarget-compile-run-and-check-generic
 // RUN: %libomptarget-compileopt-run-and-check-generic
 

offload/test/api/omp_dynamic_shared_memory_mixed.c

@@ -1,4 +1,3 @@
-// XFAIL: amdgcn-amd-amdhsa
 // clang-format off
 // RUN: %libomptarget-compileopt-generic -fopenmp-target-jit \
 // RUN:     -DTGT1_DIRECTIVE="target"

offload/test/hsa/xnack-check.cpp

@@ -1,4 +1,3 @@
-// XFAIL: amdgcn-amd-amdhsa
 // clang-format off
 //
 // RUN: %libomptarget-compileopt-generic -fopenmp-target-jit

offload/test/jit/empty_kernel_lvl1.c

@@ -34,6 +34,9 @@ if 'OMP_TARGET_OFFLOAD' in os.environ:
 if 'HSA_ENABLE_SDMA' in os.environ:
     config.environment['HSA_ENABLE_SDMA'] = os.environ['HSA_ENABLE_SDMA']
 
+if 'CLANG_USE_LINKER_WRAPPER' in os.environ:
+    config.environment['CLANG_USE_LINKER_WRAPPER'] = os.environ['CLANG_USE_LINKER_WRAPPER']
+
 # Architectures like gfx942 may or may not be APUs so an additional environment
 # variable is required as some tests can be APU specific.
 if 'IS_APU' in os.environ:
@@ -146,11 +149,6 @@ elif config.libomptarget_current_target.startswith('amdgcn'):
         config.available_features.add('apu')
     if is_mi200:
         config.available_features.add('mi200')
-    hsa_inc_dir = config.llvm_lib_directory + "/../../../rocr-runtime/src/inc"
-    hsa_lib_dir = config.llvm_lib_directory + "/../../llvm-project/runtimes/rocr-runtime-prefix/src/rocr-runtime-build"
-    if os.path.isdir(hsa_inc_dir) and os.path.isdir(hsa_lib_dir):
-        config.available_features.add('hsa')
-        config.test_flags = config.test_flags + " -I " + hsa_inc_dir + " -L " + hsa_lib_dir
 
 
 # Setup environment to find dynamic library at runtime
@@ -205,7 +203,7 @@ def add_libraries(source):
     #       error:
     # error: Linking globals named '_ZN4ompx5state9TeamStateE': symbol multiply defined!
     # return source + " " + config.llvm_library_intdir + "/libomptarget.devicertl.a"
-    return source
+    return source + " " + config.llvm_library_intdir + "/../runtimes/runtimes-bins/offload/libomptarget.devicertl.a"
 
 def get_arch_from_target(libomptarget_target):
     if libomptarget_target.startswith('amdgcn'):
@@ -265,6 +263,8 @@ for libomptarget_target in config.libomptarget_all_targets:
             "%libomptarget-compilexx-" + libomptarget_target))
         config.substitutions.append(("%libomptarget-compilexxx-generic-force-usm",
             "%libomptarget-compilexxx-force-usm-" + libomptarget_target))
+        config.substitutions.append(("%libomptarget-compilexxx-generic-cuda",
+            "%clangxxx-cuda-" + libomptarget_target))
         config.substitutions.append(("%libomptarget-compile-generic",
             "%libomptarget-compile-" + libomptarget_target))
         config.substitutions.append(("%libomptarget-compile-fortran-generic",
@@ -377,6 +377,8 @@ for libomptarget_target in config.libomptarget_all_targets:
         config.substitutions.append(("%flang-" + libomptarget_target, \
                                      "%flang %openmp_flags %flags %flags_flang --offload-arch=" +\
                                      get_arch_from_target(libomptarget_target)))
+        config.substitutions.append(("%clangxxx-cuda-" + libomptarget_target, \
+                                     "%clangxx %flags %flags_clang -foffload-via-llvm --offload-arch=native"))
         config.substitutions.append(("%fcheck-" + libomptarget_target, \
             config.libomptarget_filecheck + " %s"))
     else:

offload/test/jit/type_punning.c

@@ -1,4 +1,3 @@
-// XFAIL: amdgcn-amd-amdhsa
 // RUN: %libomptarget-compile-generic -DLIBRARY -c -o %t.o
 // RUN: ar rcs %t.a %t.o
 // RUN: %libomptarget-compile-generic %t.a && %libomptarget-run-generic 2>&1 | %fcheck-generic