[HIP] Do not unbundle object files for -fno-gpu-rdc

[yaxunl]

Summary

When -fno-gpu-rdc is set, device code is compiled, linked, and assembled into fat binary
and embedded as string in object files. The object files are normal object files which
can be linked by host linker. In the linking stage, the object files should not be unbundled
when -fno-gpu-rdc is set since they are normal object files, not bundles. The object files
only need to be unbundled when -fgpu-rdc is set.

Currently clang always unbundles object files, disregarding -fgpu-rdc option.

This patch fixes that.

[tra]

The change looks OK as far as regular CUDA is concerned.
That said, I'm not quite familiar with the use of bundling/unbundling actions and you should probably get someone who uses/depends on them to take a look. I think OpenMP uses them. Perhaps @ABataev would be the right person.

[ABataev]

LG with a nit

lib/Driver/Driver.cpp

2298

Set the default initializer for the field

[yaxunl]

lib/Driver/Driver.cpp

2298

will do. thanks.

lib/Driver/Driver.cpp

protected:
     /// The CUDA actions for the current input.
     ActionList CudaDeviceActions;
 
     /// The CUDA fat binary if it was generated for the current input.
     Action *CudaFatBinary = nullptr;
 
     /// Flag that is set to true if this builder acted on the current input.
     bool IsActive = false;
+
+    /// Flag for -fgpu-rdc.
+    bool Relocatable;
   public:
     CudaActionBuilderBase(Compilation &C, DerivedArgList &Args,
                           const Driver::InputList &Inputs,
                           Action::OffloadKind OFKind)
-        : DeviceActionBuilder(C, Args, Inputs, OFKind) {}
+        : DeviceActionBuilder(C, Args, Inputs, OFKind),
+          Relocatable(false) {}
 
     ActionBuilderReturnCode addDeviceDepences(Action *HostAction) override {
       // While generating code for CUDA, we only depend on the host input action
       // to trigger the creation of all the CUDA device actions.
 
       // If we are dealing with an input action, replicate it for each GPU
       // architecture. If we are in host-only mode we return 'success' so that
       // the host uses the CUDA offload kind.
ActionBuilderReturnCode addDeviceDepences(Action *HostAction) override {
               C.MakeAction<InputAction>(IA->getInputArg(), Ty));
         }
 
         return ABRT_Success;
       }
 
       // If this is an unbundling action use it as is for each CUDA toolchain.
       if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(HostAction)) {
+
+        // If -fgpu-rdc is disabled, should not unbundle since there is no
+        // device code to link.
+        if (!Relocatable)
+          return ABRT_Inactive;
+
         CudaDeviceActions.clear();
         auto *IA = cast<InputAction>(UA->getInputs().back());
         std::string FileName = IA->getInputArg().getAsString(Args);
         // Check if the type of the file is the same as the action. Do not
         // unbundle it if it is not. Do not unbundle .so files, for example,
         // which are not object files.
         if (IA->getType() == types::TY_Object &&
             (!llvm::sys::path::has_extension(FileName) ||
bool initialize() override {
           !C.hasOffloadToolChain<Action::OFK_Cuda>())
         return false;
 
       // We don't need to support HIP.
       if (AssociatedOffloadKind == Action::OFK_HIP &&
           !C.hasOffloadToolChain<Action::OFK_HIP>())
         return false;
 
+      Relocatable = Args.hasFlag(options::OPT_fgpu_rdc,
+          options::OPT_fno_gpu_rdc, /*Default=*/false);
+
       const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>();
       assert(HostTC && "No toolchain for host compilation.");
       if (HostTC->getTriple().isNVPTX() ||
           HostTC->getTriple().getArch() == llvm::Triple::amdgcn) {
         // We do not support targeting NVPTX/AMDGCN for host compilation. Throw
         // an error and abort pipeline construction early so we don't trip
         // asserts that assume device-side compilation.
         C.getDriver().Diag(diag::err_drv_cuda_host_arch)
getDeviceDependences(OffloadAction::DeviceDependences &DA,
       return ABRT_Success;
     }
   };
   /// \brief HIP action builder. It injects device code in the host backend
   /// action.
   class HIPActionBuilder final : public CudaActionBuilderBase {
     /// The linker inputs obtained for each device arch.
     SmallVector<ActionList, 8> DeviceLinkerInputs;
-    bool Relocatable;
 
   public:
     HIPActionBuilder(Compilation &C, DerivedArgList &Args,
                      const Driver::InputList &Inputs)
-        : CudaActionBuilderBase(C, Args, Inputs, Action::OFK_HIP),
-          Relocatable(false) {}
+        : CudaActionBuilderBase(C, Args, Inputs, Action::OFK_HIP) {}
 
     bool canUseBundlerUnbundler() const override { return true; }
 
     ActionBuilderReturnCode
     getDeviceDependences(OffloadAction::DeviceDependences &DA,
                          phases::ID CurPhase, phases::ID FinalPhase,
                          PhasesTy &Phases) override {
       // amdgcn does not support linking of object files, therefore we skip
void appendLinkDependences(OffloadAction::DeviceDependences &DA) override {
       for (auto &LI : DeviceLinkerInputs) {
         auto *DeviceLinkAction =
             C.MakeAction<LinkJobAction>(LI, types::TY_Image);
         DA.add(*DeviceLinkAction, *ToolChains[0],
                CudaArchToString(GpuArchList[I]), AssociatedOffloadKind);
         ++I;
       }
     }
-
-    bool initialize() override {
-      Relocatable = Args.hasFlag(options::OPT_fgpu_rdc,
-          options::OPT_fno_gpu_rdc, /*Default=*/false);
-
-      return CudaActionBuilderBase::initialize();
-    }
   };
 
   /// OpenMP action builder. The host bitcode is passed to the device frontend
   /// and all the device linked images are passed to the host link phase.
   class OpenMPActionBuilder final : public DeviceActionBuilder {
     /// The OpenMP actions for the current input.
     ActionList OpenMPDeviceActions;
 

test/Driver/hip-binding.hip

 // REQUIRES: clang-driver
 // REQUIRES: x86-registered-target
 // REQUIRES: amdgpu-registered-target
 
 // RUN: touch %t.o
 // RUN: %clang --hip-link -ccc-print-bindings -target x86_64-linux-gnu \
-// RUN:   --cuda-gpu-arch=gfx803 --cuda-gpu-arch=gfx900 %t.o\
+// RUN:   --cuda-gpu-arch=gfx803 --cuda-gpu-arch=gfx900 -fgpu-rdc %t.o\
 // RUN: 2>&1 | FileCheck %s
 
 // CHECK: # "amdgcn-amd-amdhsa" - "offload bundler", inputs: ["[[IN:.*o]]"], outputs: ["[[OBJ1:.*o]]", "[[OBJ2:.*o]]", "[[OBJ3:.*o]]"] 
 // CHECK: # "amdgcn-amd-amdhsa" - "AMDGCN::Linker", inputs: ["[[OBJ2]]"], output: "[[IMG2:.*out]]"
 // CHECK-NOT: offload bundler
 // CHECK: # "amdgcn-amd-amdhsa" - "AMDGCN::Linker", inputs: ["[[OBJ3]]"], output: "[[IMG3:.*out]]"
 // CHECK-NOT: offload bundler
 // CHECK: # "x86_64-unknown-linux-gnu" - "GNU::Linker", inputs: ["[[OBJ1]]", "[[IMG2]]", "[[IMG3]]"], output: "a.out"
+
+// RUN: %clang --hip-link -ccc-print-bindings -target x86_64-linux-gnu \
+// RUN:   --cuda-gpu-arch=gfx803 --cuda-gpu-arch=gfx900 %t.o\
+// RUN: 2>&1 | FileCheck -check-prefix=NORDC %s
+
+// NORDC-NOT: offload bundler
+// NORDC: # "x86_64-unknown-linux-gnu" - "GNU::Linker", inputs: ["{{.*o}}"], output: "a.out"

test/Driver/hip-link-shared-library.hip

 // RUN: touch %t.o
 // RUN: %clang --hip-link -ccc-print-bindings -target x86_64-linux-gnu \
 // RUN:   --cuda-gpu-arch=gfx803 --cuda-gpu-arch=gfx900 %t.o %S/Inputs/in.so \
-// RUN: 2>&1 | FileCheck %s
+// RUN:   -fgpu-rdc 2>&1 | FileCheck %s
 
 // CHECK: # "amdgcn-amd-amdhsa" - "offload bundler", inputs: ["[[IN:.*o]]"], outputs: ["[[OBJ1:.*o]]", "[[OBJ2:.*o]]", "[[OBJ3:.*o]]"]
 // CHECK: # "amdgcn-amd-amdhsa" - "AMDGCN::Linker", inputs: ["[[OBJ2]]"], output: "[[IMG2:.*out]]"
 // CHECK-NOT: offload bundler
 // CHECK: # "amdgcn-amd-amdhsa" - "AMDGCN::Linker", inputs: ["[[OBJ3]]"], output: "[[IMG3:.*out]]"
 // CHECK-NOT: offload bundler
 // CHECK: # "x86_64-unknown-linux-gnu" - "GNU::Linker", inputs: ["[[OBJ1]]", "{{.*}}/Inputs/in.so", "[[IMG2]]", "[[IMG3]]"], output: "a.out"