[SYCL][Driver] -foffload-static-lib improvements when no source is given (#1072)

Provides improvements to behaviors of -foffload-static-lib when no source
or object file is also provided on the command line.

 - Adds new -foffload-whole-static-lib option that is similar to
   -foffload-static-lib, except during the partial linkage stage.  That is
   where the whole-archive is used to preserve all items in the archive
   before being unbundled.
 - Adds the arguments to -foffload-static-lib to be considered as true inputs
   to the compiler.  This triggers proper link steps on the offload side as
   well as places the library directly on the link.
 - Moves the static-lib tests into a static-lib specific test

Signed-off-by: Michael D Toguchi <[email protected]>

Author: mdtoguchi

clang/include/clang/Driver/Action.h

@@ -613,7 +613,8 @@ class OffloadUnbundlingJobAction final : public JobAction {
 
 public:
   // Offloading unbundling doesn't change the type of output.
-  OffloadUnbundlingJobAction(ActionList &Inputs);
+  OffloadUnbundlingJobAction(Action *Input);
+  OffloadUnbundlingJobAction(ActionList &Inputs, types::ID Type);
 
   /// Register information about a dependent action.
   void registerDependentActionInfo(const ToolChain *TC, StringRef BoundArch,

clang/include/clang/Driver/Options.td

@@ -197,6 +197,8 @@ def u_Group : OptionGroup<"<u group>">, Group<Link_Group>, DocFlatten;
 def reserved_lib_Group : OptionGroup<"<reserved libs group>">,
                          Flags<[Unsupported]>;
 
+def offload_lib_Group : OptionGroup<"<offload libs group>">;
+
 // Temporary groups for clang options which we know we don't support,
 // but don't want to verbosely warn the user about.
 def clang_ignored_f_Group : OptionGroup<"<clang ignored f group>">,
@@ -1674,7 +1676,8 @@ def fobjc_nonfragile_abi : Flag<["-"], "fobjc-nonfragile-abi">, Group<f_Group>;
 def fno_objc_nonfragile_abi : Flag<["-"], "fno-objc-nonfragile-abi">, Group<f_Group>;
 
 def fobjc_sender_dependent_dispatch : Flag<["-"], "fobjc-sender-dependent-dispatch">, Group<f_Group>;
-def foffload_static_lib_EQ : CommaJoined<["-"], "foffload-static-lib=">, Flags<[DriverOption, CoreOption]>;
+def foffload_static_lib_EQ : CommaJoined<["-"], "foffload-static-lib=">, Flags<[DriverOption, CoreOption]>, Group<offload_lib_Group>;
+def foffload_whole_static_lib_EQ : CommaJoined<["-"], "foffload-whole-static-lib=">, Flags<[DriverOption, CoreOption]>, Group<offload_lib_Group>;
 def fomit_frame_pointer : Flag<["-"], "fomit-frame-pointer">, Group<f_Group>;
 def fopenmp : Flag<["-"], "fopenmp">, Group<f_Group>, Flags<[CC1Option, NoArgumentUnused]>,
   HelpText<"Parse OpenMP pragmas and generate parallel code.">;

clang/include/clang/Driver/Types.def

@@ -106,6 +106,7 @@ TYPE("tempfilelist",             Tempfilelist, INVALID,         "txt",    phases
 TYPE("tempentriesfilelist",      TempEntriesfilelist, INVALID,  "txt",    phases::Compile, phases::Backend, phases::Assemble, phases::Link)
 TYPE("tempAOCOfilelist",         TempAOCOfilelist, INVALID,     "txt",    phases::Compile, phases::Backend, phases::Assemble, phases::Link)
 TYPE("archive",                  Archive,      INVALID,         "a",      phases::Compile, phases::Backend, phases::Assemble, phases::Link)
+TYPE("wholearchive",             WholeArchive, INVALID,         "a",      phases::Compile, phases::Backend, phases::Assemble, phases::Link)
 TYPE("fpga_aocx",                FPGA_AOCX,    INVALID,         "aocx",   phases::Compile, phases::Backend, phases::Assemble, phases::Link)
 TYPE("fpga_aocr",                FPGA_AOCR,    INVALID,         "aocr",   phases::Compile, phases::Backend, phases::Assemble, phases::Link)
 TYPE("fpga_aoco",                FPGA_AOCO,    INVALID,         "aoco",   phases::Compile, phases::Backend, phases::Assemble, phases::Link)

clang/include/clang/Driver/Types.h

@@ -84,6 +84,9 @@ namespace types {
   /// isFPGA - Is this FPGA input.
   bool isFPGA(ID Id);
 
+  /// isArchive - Is this an archive input.
+  bool isArchive(ID Id);
+
   /// isObjC - Is this an "ObjC" input (Obj-C and Obj-C++ sources and headers).
   bool isObjC(ID Id);
 

clang/lib/Driver/Action.cpp

@@ -421,8 +421,12 @@ OffloadBundlingJobAction::OffloadBundlingJobAction(ActionList &Inputs)
 
 void OffloadUnbundlingJobAction::anchor() {}
 
-OffloadUnbundlingJobAction::OffloadUnbundlingJobAction(ActionList &Inputs)
-    : JobAction(OffloadUnbundlingJobClass, Inputs, Inputs.back()->getType()) {}
+OffloadUnbundlingJobAction::OffloadUnbundlingJobAction(Action *Input)
+    : JobAction(OffloadUnbundlingJobClass, Input, Input->getType()) {}
+
+OffloadUnbundlingJobAction::OffloadUnbundlingJobAction(ActionList &Inputs,
+                                                       types:: ID Type)
+    : JobAction(OffloadUnbundlingJobClass, Inputs, Type) {}
 
 void OffloadWrapperJobAction::anchor() {}
 

clang/lib/Driver/Driver.cpp

@@ -2469,6 +2469,14 @@ void Driver::BuildInputs(const ToolChain &TC, DerivedArgList &Args,
         Diag(diag::note_use_dashdash);
       }
     }
+    else if (A->getOption().matches(options::OPT_offload_lib_Group)) {
+      // Add the foffload-static-lib library to the command line to allow
+      // processing when no source or object is supplied as well as proper
+      // host link.
+      Arg *InputArg = MakeInputArg(Args, Opts, A->getValue());
+      Inputs.push_back(std::make_pair(types::TY_Object, InputArg));
+      A->claim();
+    }
   }
   if (CCCIsCPP() && Inputs.empty()) {
     // If called as standalone preprocessor, stdin is processed
@@ -3336,7 +3344,7 @@ class OffloadingActionBuilder final {
 
         std::string InputName = IA->getInputArg().getAsString(Args);
         // Objects should already be consumed with -foffload-static-lib
-        if (Args.hasArg(options::OPT_foffload_static_lib_EQ) &&
+        if (Args.hasArg(options::OPT_offload_lib_Group) &&
             IA->getType() == types::TY_Object && isObjectFile(InputName))
           return ABRT_Inactive;
 
@@ -3870,7 +3878,7 @@ class OffloadingActionBuilder final {
       if (C.getDefaultToolChain().getTriple().isWindowsMSVCEnvironment() ||
           !(HostAction->getType() == types::TY_Object &&
             isObjectFile(InputName) &&
-            Args.hasArg(options::OPT_foffload_static_lib_EQ))) {
+            Args.hasArg(options::OPT_offload_lib_Group))) {
         ActionList HostActionList;
         Action *A(HostAction);
         // Only check for FPGA device information when using fpga SubArch.
@@ -3887,15 +3895,11 @@ class OffloadingActionBuilder final {
           else if (hasFPGABinary(C, InputName, types::TY_FPGA_AOCR))
             A = C.MakeAction<InputAction>(*InputArg, types::TY_FPGA_AOCR);
         }
-        HostActionList.push_back(A);
-        if (!HostActionList.empty()) {
-          auto UnbundlingHostAction =
-            C.MakeAction<OffloadUnbundlingJobAction>(HostActionList);
-          UnbundlingHostAction->registerDependentActionInfo(
+        auto UnbundlingHostAction = C.MakeAction<OffloadUnbundlingJobAction>(A);
+        UnbundlingHostAction->registerDependentActionInfo(
             C.getSingleOffloadToolChain<Action::OFK_Host>(),
             /*BoundArch=*/StringRef(), Action::OFK_Host);
-          HostAction = UnbundlingHostAction;
-        }
+        HostAction = UnbundlingHostAction;
       }
     }
 
@@ -3945,11 +3949,11 @@ class OffloadingActionBuilder final {
     if (!IsValid || InputActionList.empty())
       return true;
 
-    auto *DeviceUnbundlingAction =
-              C.MakeAction<OffloadUnbundlingJobAction>(InputActionList);
+    auto *DeviceUnbundlingAction = C.MakeAction<OffloadUnbundlingJobAction>(
+        InputActionList, types::TY_Object);
     DeviceUnbundlingAction->registerDependentActionInfo(
-          C.getSingleOffloadToolChain<Action::OFK_Host>(),
-          /*BoundArch=*/StringRef(), Action::OFK_Host);
+        C.getSingleOffloadToolChain<Action::OFK_Host>(),
+        /*BoundArch=*/StringRef(), Action::OFK_Host);
     HostAction = DeviceUnbundlingAction;
 
     // Register the offload kinds that are used.
@@ -4357,59 +4361,69 @@ void Driver::BuildActions(Compilation &C, DerivedArgList &Args,
   // When a static fat archive is provided, create a new unbundling step
   // for all of the objects.
   if (!C.getDefaultToolChain().getTriple().isWindowsMSVCEnvironment() &&
-      Args.hasArg(options::OPT_foffload_static_lib_EQ) &&
-      !LinkerInputs.empty()) {
+      Args.hasArg(options::OPT_offload_lib_Group)) {
     ActionList UnbundlerInputs;
-    ActionList TempLinkerInputs;
     for (const auto &LI : LinkerInputs) {
       // Unbundler only handles objects.
       if (auto *IA = dyn_cast<InputAction>(LI)) {
         std::string FileName = IA->getInputArg().getAsString(Args);
         if ((IA->getType() == types::TY_Object && !isObjectFile(FileName)) ||
             IA->getInputArg().getOption().hasFlag(options::LinkerInput))
-          // Pass the Input along to linker.
-          TempLinkerInputs.push_back(LI);
-        else
-          // Add to unbundler.
-          UnbundlerInputs.push_back(LI);
-      } else
+          // Pass the Input along to linker only.
+          continue;
         UnbundlerInputs.push_back(LI);
+      }
     }
-    LinkerInputs.clear();
+    const Arg *LastArg;
+    auto addUnbundlerInput = [&](types::ID T, const Arg *A) {
+      const llvm::opt::OptTable &Opts = getOpts();
+      Arg *InputArg = MakeInputArg(Args, Opts, A->getValue());
+      LastArg = InputArg;
+      Action *Current = C.MakeAction<InputAction>(*InputArg, T);
+      UnbundlerInputs.push_back(Current);
+    };
+    for (const auto *A : Args.filtered(options::OPT_foffload_static_lib_EQ))
+      addUnbundlerInput(types::TY_Archive, A);
+    for (const auto *A :
+        Args.filtered(options::OPT_foffload_whole_static_lib_EQ))
+      addUnbundlerInput(types::TY_WholeArchive, A);
     if (!UnbundlerInputs.empty()) {
-      Action *Current;
-      const Arg *LastArg = Args.getLastArg(options::OPT_foffload_static_lib_EQ);
+      Action *Current = C.MakeAction<InputAction>(*LastArg, types::TY_Archive);
       OffloadBuilder.addHostDependenceToUnbundlingAction(Current,
-                                                    UnbundlerInputs, LastArg);
+          UnbundlerInputs, LastArg);
       Current = OffloadBuilder.addDeviceDependencesToHostAction(Current,
-                                       LastArg, phases::Link, PL.back(), PL);
-      LinkerInputs.push_back(Current);
+          LastArg, phases::Link, PL.back(), PL);
     }
-    for (const auto &TLI : TempLinkerInputs)
-      LinkerInputs.push_back(TLI);
   }
   const llvm::opt::OptTable &Opts = getOpts();
+  auto unbundleStaticLib = [&](types::ID T, const Arg *A) {
+    Arg *InputArg = MakeInputArg(Args, Opts, A->getValue());
+    Action *Current = C.MakeAction<InputAction>(*InputArg, T);
+    OffloadBuilder.addHostDependenceToDeviceActions(Current, InputArg, Args);
+    OffloadBuilder.addDeviceDependencesToHostAction(
+        Current, InputArg, phases::Link, PL.back(), PL);
+  };
   for (const auto *A : Args.filtered(options::OPT_foffload_static_lib_EQ)) {
-    auto unbundleStaticLib = [&](types::ID T) {
-      Arg *InputArg = MakeInputArg(Args, Opts, A->getValue());
-      Action *Current = C.MakeAction<InputAction>(*InputArg, T);
-      OffloadBuilder.addHostDependenceToDeviceActions(Current, InputArg, Args);
-      OffloadBuilder.addDeviceDependencesToHostAction(
-          Current, InputArg, phases::Link, PL.back(), PL);
-    };
     // In MSVC environment offload-static-libs are handled slightly different
     // because of missing support for partial linking in the linker. We add an
     // unbundling action for each static archive which produces list files with
     // extracted objects. Device lists are then added to the appropriate device
     // link actions and host list is ignored since we are adding
     // offload-static-libs as normal libraries to the host link command.
     if (C.getDefaultToolChain().getTriple().isWindowsMSVCEnvironment())
-      unbundleStaticLib(types::TY_Archive);
+      unbundleStaticLib(types::TY_Archive, A);
     // Pass along the -foffload-static-lib values to check if we need to
     // add them for unbundling for FPGA AOT static lib usage.  Uses FPGA
     // aoco type to differentiate if aoco unbundling is needed.
     if (Args.hasArg(options::OPT_fintelfpga))
-      unbundleStaticLib(types::TY_FPGA_AOCO);
+      unbundleStaticLib(types::TY_FPGA_AOCO, A);
+  }
+  for (const auto *A :
+      Args.filtered(options::OPT_foffload_whole_static_lib_EQ)) {
+    if (C.getDefaultToolChain().getTriple().isWindowsMSVCEnvironment())
+      unbundleStaticLib(types::TY_WholeArchive, A);
+    if (Args.hasArg(options::OPT_fintelfpga))
+      unbundleStaticLib(types::TY_FPGA_AOCO, A);
   }
 
   // For an FPGA archive, we add the unbundling step above to take care of
@@ -5282,13 +5296,11 @@ InputInfo Driver::BuildJobsForActionNoCache(
       bool IsFPGAObjLink = (JA->getType() == types::TY_Object &&
           C.getInputArgs().hasArg(options::OPT_fintelfpga) &&
           C.getInputArgs().hasArg(options::OPT_fsycl_link_EQ));
-      if (C.getInputArgs().hasArg(options::OPT_foffload_static_lib_EQ) &&
+      if (C.getInputArgs().hasArg(options::OPT_offload_lib_Group) &&
           ((JA->getType() == types::TY_Archive && IsMSVCEnv) ||
-           (UI.DependentOffloadKind != Action::OFK_Host &&
-            (JA->getType() == types::TY_Object && !IsMSVCEnv)))) {
+           (JA->getType() == types::TY_Object && !IsMSVCEnv))) {
         // Host part of the unbundled static archive is not used.
-        if (UI.DependentOffloadKind == Action::OFK_Host &&
-            JA->getType() == types::TY_Archive && IsMSVCEnv)
+        if (UI.DependentOffloadKind == Action::OFK_Host)
           continue;
         // Host part of the unbundled object when -fintelfpga -fsycl-link is
         // enabled is not used

clang/lib/Driver/ToolChains/Clang.cpp

@@ -7080,14 +7080,17 @@ void OffloadBundler::ConstructJobMultipleOutputs(
   InputInfo Input = Inputs.front();
   const char *TypeArg = types::getTypeTempSuffix(Input.getType());
   const char *InputFileName = Input.getFilename();
+  bool IsMSVCEnv =
+      C.getDefaultToolChain().getTriple().isWindowsMSVCEnvironment();
+  types::ID InputType(Input.getType());
 
-  // For objects, we have initial support for fat archives (archives which
+  // For Linux, we have initial support for fat archives (archives which
   // contain bundled objects). We will perform partial linking against the
-  // object and specific offload target archives which will be sent to the
-  // unbundler to produce a list of target objects.
-  if (!C.getDefaultToolChain().getTriple().isWindowsMSVCEnvironment() &&
-      Input.getType() == types::TY_Object &&
-      TCArgs.hasArg(options::OPT_foffload_static_lib_EQ)) {
+  // specific offload target archives which will be sent to the unbundler to
+  // produce a list of target objects.
+  // FIXME: This should be a separate job in the toolchain.
+  if (!IsMSVCEnv && TCArgs.hasArg(options::OPT_offload_lib_Group) &&
+      (types::isArchive(InputType) || InputType == types::TY_Object)) {
     TypeArg = "oo";
     ArgStringList LinkArgs;
     LinkArgs.push_back("-r");
@@ -7097,28 +7100,47 @@ void OffloadBundler::ConstructJobMultipleOutputs(
           llvm::sys::path::stem(Input.getFilename()).str() + "-prelink", "o");
     InputFileName = C.addTempFile(C.getArgs().MakeArgString(TmpName));
     LinkArgs.push_back(InputFileName);
-    // Input files consist of fat libraries and the object(s) to be unbundled.
-    for (const auto &I : Inputs)
-      LinkArgs.push_back(I.getFilename());
+    const ToolChain *HTC = C.getSingleOffloadToolChain<Action::OFK_Host>();
+    // Add crt objects
+    LinkArgs.push_back(TCArgs.MakeArgString(HTC->GetFilePath("crt1.o")));
+    LinkArgs.push_back(TCArgs.MakeArgString(HTC->GetFilePath("crti.o")));
     // Add -L<dir> search directories.
     TCArgs.AddAllArgs(LinkArgs, options::OPT_L);
-    for (const auto &A :
-            TCArgs.getAllArgValues(options::OPT_foffload_static_lib_EQ))
-      LinkArgs.push_back(TCArgs.MakeArgString(A));
+
+    // TODO - We can potentially go through the args and add the known linker
+    // pass through args of --whole-archive and --no-whole-archive.  This
+    // would allow to support user commands like: -Wl,--whole-archive
+    // -foffload-static-lib=<lib> -Wl,--no-whole-archive
+    // Input files consist of fat libraries and the object(s) to be unbundled.
+    bool IsWholeArchive = false;
+    for (const auto &I : Inputs) {
+      if (I.getType() == types::TY_WholeArchive && !IsWholeArchive) {
+        LinkArgs.push_back("--whole-archive");
+        IsWholeArchive = true;
+      } else if (I.getType() == types::TY_Archive && IsWholeArchive) {
+        LinkArgs.push_back("--no-whole-archive");
+        IsWholeArchive = false;
+      }
+      LinkArgs.push_back(I.getFilename());
+    }
+    // Disable whole archive if it was enabled for the previous inputs.
+    if (IsWholeArchive)
+      LinkArgs.push_back("--no-whole-archive");
+    // Add crt objects
+    LinkArgs.push_back(TCArgs.MakeArgString(HTC->GetFilePath("crtn.o")));
     const char *Exec = TCArgs.MakeArgString(getToolChain().GetLinkerPath());
     C.addCommand(std::make_unique<Command>(JA, *this, Exec, LinkArgs, Inputs));
-  } else if (Input.getType() == types::TY_FPGA_AOCX ||
-             Input.getType() == types::TY_FPGA_AOCR) {
+  } else if (InputType == types::TY_FPGA_AOCX ||
+             InputType == types::TY_FPGA_AOCR) {
     // Override type with archive object
     if (getToolChain().getTriple().getSubArch() ==
         llvm::Triple::SPIRSubArch_fpga)
       TypeArg = "ao";
     else
       TypeArg = "aoo";
   }
-  if (Input.getType() == types::TY_FPGA_AOCO ||
-      (C.getDefaultToolChain().getTriple().isWindowsMSVCEnvironment() &&
-       Input.getType() == types::TY_Archive))
+  if (InputType == types::TY_FPGA_AOCO ||
+      (IsMSVCEnv && types::isArchive(InputType)))
     TypeArg = "aoo";
 
   // Get the type.
@@ -7133,12 +7155,12 @@ void OffloadBundler::ConstructJobMultipleOutputs(
     // FPGA device triples are 'transformed' for the bundler when creating
     // aocx or aocr type bundles.  Also, we only do a specific target
     // unbundling, skipping the host side or device side.
-    if (types::isFPGA(Input.getType())) {
+    if (types::isFPGA(InputType)) {
       if (getToolChain().getTriple().getSubArch() ==
               llvm::Triple::SPIRSubArch_fpga &&
           Dep.DependentOffloadKind == Action::OFK_SYCL) {
         llvm::Triple TT;
-        TT.setArchName(types::getTypeName(Input.getType()));
+        TT.setArchName(types::getTypeName(InputType));
         TT.setVendorName("intel");
         TT.setOS(getToolChain().getTriple().getOS());
         TT.setEnvironment(llvm::Triple::SYCLDevice);
@@ -7152,10 +7174,10 @@ void OffloadBundler::ConstructJobMultipleOutputs(
         Triples += Dep.DependentToolChain->getTriple().normalize();
       }
       continue;
-    } else if (Input.getType() == types::TY_Archive ||
-               (Input.getType() == types::TY_Object &&
-                TCArgs.hasArg(options::OPT_fintelfpga) &&
-                TCArgs.hasArg(options::OPT_fsycl_link_EQ))) {
+    } else if (types::isArchive(InputType) || (InputType == types::TY_Object &&
+               ((!IsMSVCEnv && TCArgs.hasArg(options::OPT_offload_lib_Group)) ||
+                (TCArgs.hasArg(options::OPT_fintelfpga) &&
+                 TCArgs.hasArg(options::OPT_fsycl_link_EQ))))) {
       // Do not extract host part if we are unbundling archive on Windows
       // because it is not needed. Static offload libraries are added to the
       // host link command just as normal libraries.  Do not extract the host

clang/lib/Driver/ToolChains/MSVC.cpp

@@ -366,6 +366,9 @@ void visualstudio::Linker::ConstructJob(Compilation &C, const JobAction &JA,
   for (const auto *A : Args.filtered(options::OPT_foffload_static_lib_EQ))
     CmdArgs.push_back(
         Args.MakeArgString(Twine("-defaultlib:") + A->getValue()));
+  for (const auto *A : Args.filtered(options::OPT_foffload_whole_static_lib_EQ))
+    CmdArgs.push_back(
+        Args.MakeArgString(Twine("-wholearchive:") + A->getValue()));
 
   if (!llvm::sys::Process::GetEnv("LIB")) {
     // If the VC environment hasn't been configured (perhaps because the user

clang/lib/Driver/Types.cpp

@@ -235,6 +235,16 @@ bool types::isFPGA(ID Id) {
   }
 }
 
+bool types::isArchive(ID Id) {
+  switch (Id) {
+  default:
+    return false;
+  case TY_Archive:
+  case TY_WholeArchive:
+    return true;
+  }
+}
+
 bool types::isSrcFile(ID Id) {
   return Id != TY_Object && getPreprocessedType(Id) != TY_INVALID;
 }