[HIP] Fix unbundling

[yaxunl]

Summary

HIP uses clang-offload-bundler to bundle intermediate files for host
and different gpu archs together. When a file is unbundled,
clang-offload-bundler should be called only once, and the objects
for host and different gpu archs should be passed to the next
jobs. This is because Driver maintains CachedResults which maps
triple-arch string to output files for each job.

This patch fixes a bug in Driver::BuildJobsForActionNoCache which
uses incorrect key for CachedResults for HIP which causes
clang-offload-bundler being called mutiple times and incorrect
output files being used.

It only affects HIP.

[tra]

Few minor nits/suggestions. LGTM otherwise.

lib/Driver/Driver.cpp

3895

Should it be something more descriptive? E.g. "all" or "combined".

test/Driver/hip-binding.hip

6–8

Nit: you could probably just use /dev/null instead of a real temp file.

12–13

There's currently a bundler invocation in-between these two linker commands.
I think you need a negative check for bundler here, too.

[yaxunl]

lib/Driver/Driver.cpp

3895

will use "all"

test/Driver/hip-binding.hip

6–8

I tried /dev/null but I got

error: no input files

I think we need a real input file here, even though an empty one.

12–13

will do

lib/Driver/Driver.cpp

bool addHostDependenceToDeviceActions(Action *&HostAction,
     // the input is not a bundle.
     if (CanUseBundler && isa<InputAction>(HostAction) &&
         InputArg->getOption().getKind() == llvm::opt::Option::InputClass &&
         !types::isSrcFile(HostAction->getType())) {
       auto UnbundlingHostAction =
           C.MakeAction<OffloadUnbundlingJobAction>(HostAction);
       UnbundlingHostAction->registerDependentActionInfo(
           C.getSingleOffloadToolChain<Action::OFK_Host>(),
-          /*BoundArch=*/StringRef(), Action::OFK_Host);
+          /*BoundArch=*/"all", Action::OFK_Host);
       HostAction = UnbundlingHostAction;
     }
 
     assert(HostAction && "Invalid host action!");
 
     // Register the offload kinds that are used.
     auto &OffloadKind = InputArgToOffloadKindMap[InputArg];
     for (auto *SB : SpecializedBuilders) {
for (auto &UI : UA->getDependentActionsInfo()) {
                                  UI.DependentOffloadKind == Action::OFK_HIP,
                              OffloadingPrefix),
           BaseInput);
       // Save the unbundling result.
       UnbundlingResults.push_back(CurI);
 
       // Get the unique string identifier for this dependence and cache the
       // result.
-      CachedResults[{A, GetTriplePlusArchString(
-                            UI.DependentToolChain, BoundArch,
-                            UI.DependentOffloadKind)}] = CurI;
+      StringRef Arch;
+      if (TargetDeviceOffloadKind == Action::OFK_HIP) {
+        if (UI.DependentOffloadKind == Action::OFK_Host)
+          Arch = "all";
+        else
+          Arch = UI.DependentBoundArch;
+      } else
+        Arch = BoundArch;
+
+      CachedResults[{A, GetTriplePlusArchString(UI.DependentToolChain, Arch,
+                                                UI.DependentOffloadKind)}] =
+          CurI;
     }
 
     // Now that we have all the results generated, select the one that should be
     // returned for the current depending action.
     std::pair<const Action *, std::string> ActionTC = {
         A, GetTriplePlusArchString(TC, BoundArch, TargetDeviceOffloadKind)};
     assert(CachedResults.find(ActionTC) != CachedResults.end() &&
            "Result does not exist??");
     Result = CachedResults[ActionTC];

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: 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--linux-gnu" - "GNU::Linker", inputs: ["[[OBJ1]]", "[[IMG2]]", "[[IMG3]]"], output: "a.out"