[Driver][SYCL] Add initial SYCL offload compilation support

Introduces the SYCL based toolchain and initial toolchain construction
when using the '-fsycl' option.  This option will enable SYCL based
offloading, creating a SPIR-V based IR file packaged into the compiled
host object.

This includes early support for creating the host/device object using
the new offloading model.  The device object is created using the
spir64-unknown-unknown target triple.

New/Updated Options:
 -fsycl  Enables SYCL offloading for host and device
 -fsycl-device-only
         Enables device only compilation for SYCL
 -fsycl-host-only
         Enables host only compilation for SYCL

RFC Reference: https://discourse.llvm.org/t/rfc-sycl-driver-enhancements/74092

Author: mdtoguchi

clang/include/clang/Driver/Action.h

@@ -94,6 +94,7 @@ class Action {
     OFK_Cuda = 0x02,
     OFK_OpenMP = 0x04,
     OFK_HIP = 0x08,
+    OFK_SYCL = 0x10,
   };
 
   static const char *getClassName(ActionClass AC);

clang/include/clang/Driver/Driver.h

@@ -582,6 +582,10 @@ class Driver {
   /// @name Helper Methods
   /// @{
 
+  /// MakeSYCLDeviceTriple - Returns the SYCL device triple for the
+  /// specified ArchType.
+  llvm::Triple MakeSYCLDeviceTriple(StringRef TargetArch = "spir64") const;
+
   /// PrintActions - Print the list of actions.
   void PrintActions(const Compilation &C) const;
 

clang/include/clang/Driver/Options.td

@@ -182,7 +182,8 @@ def opencl_Group : OptionGroup<"<opencl group>">, Group<f_Group>,
                    DocName<"OpenCL options">;
 
 def sycl_Group : OptionGroup<"<SYCL group>">, Group<f_Group>,
-                 DocName<"SYCL options">;
+                 DocName<"SYCL options">,
+                 Visibility<[ClangOption, CLOption]>;
 
 def cuda_Group : OptionGroup<"<CUDA group>">, Group<f_Group>,
                    DocName<"CUDA options">,
@@ -6691,12 +6692,16 @@ defm : FlangIgnoredDiagOpt<"frontend-loop-interchange">;
 defm : FlangIgnoredDiagOpt<"target-lifetime">;
 
 // C++ SYCL options
+let Group = sycl_Group in {
 def fsycl : Flag<["-"], "fsycl">,
-  Visibility<[ClangOption, CLOption]>,
-  Group<sycl_Group>, HelpText<"Enables SYCL kernels compilation for device">;
+  HelpText<"Enables SYCL kernels compilation for device">;
 def fno_sycl : Flag<["-"], "fno-sycl">,
-  Visibility<[ClangOption, CLOption]>,
-  Group<sycl_Group>, HelpText<"Disables SYCL kernels compilation for device">;
+  HelpText<"Disables SYCL kernels compilation for device">;
+def fsycl_device_only : Flag<["-"], "fsycl-device-only">,
+  Alias<offload_device_only>, HelpText<"Compile SYCL kernels for device only">;
+def fsycl_host_only : Flag<["-"], "fsycl-host-only">,
+  Alias<offload_host_only>, HelpText<"Compile SYCL kernels for host only">;
+} // let Group = sycl_Group
 
 // OS-specific options
 let Flags = [TargetSpecific] in {

clang/include/clang/Driver/ToolChain.h

@@ -762,6 +762,10 @@ class ToolChain {
   virtual void AddHIPIncludeArgs(const llvm::opt::ArgList &DriverArgs,
                                  llvm::opt::ArgStringList &CC1Args) const;
 
+  /// Add arguments to use SYCL specific includes.
+  virtual void AddSYCLIncludeArgs(const llvm::opt::ArgList &DriverArgs,
+                                  llvm::opt::ArgStringList &CC1Args) const;
+
   /// Add arguments to use MCU GCC toolchain includes.
   virtual void AddIAMCUIncludeArgs(const llvm::opt::ArgList &DriverArgs,
                                    llvm::opt::ArgStringList &CC1Args) const;

clang/lib/Driver/Action.cpp

@@ -111,6 +111,8 @@ std::string Action::getOffloadingKindPrefix() const {
     return "device-openmp";
   case OFK_HIP:
     return "device-hip";
+  case OFK_SYCL:
+    return "device-sycl";
 
     // TODO: Add other programming models here.
   }
@@ -128,6 +130,8 @@ std::string Action::getOffloadingKindPrefix() const {
     Res += "-hip";
   if (ActiveOffloadKindMask & OFK_OpenMP)
     Res += "-openmp";
+  if (ActiveOffloadKindMask & OFK_SYCL)
+    Res += "-sycl";
 
   // TODO: Add other programming models here.
 
@@ -164,6 +168,8 @@ StringRef Action::GetOffloadKindName(OffloadKind Kind) {
     return "openmp";
   case OFK_HIP:
     return "hip";
+  case OFK_SYCL:
+    return "sycl";
 
     // TODO: Add other programming models here.
   }
@@ -320,7 +326,7 @@ void OffloadAction::DeviceDependences::add(Action &A, const ToolChain &TC,
   DeviceBoundArchs.push_back(BoundArch);
 
   // Add each active offloading kind from a mask.
-  for (OffloadKind OKind : {OFK_OpenMP, OFK_Cuda, OFK_HIP})
+  for (OffloadKind OKind : {OFK_OpenMP, OFK_Cuda, OFK_HIP, OFK_SYCL})
     if (OKind & OffloadKindMask)
       DeviceOffloadKinds.push_back(OKind);
 }

clang/lib/Driver/CMakeLists.txt

@@ -85,6 +85,7 @@ add_clang_library(clangDriver
   ToolChains/PPCFreeBSD.cpp
   ToolChains/InterfaceStubs.cpp
   ToolChains/ZOS.cpp
+  ToolChains/SYCL.cpp
   Types.cpp
   XRayArgs.cpp
 

clang/lib/Driver/Compilation.cpp

@@ -217,10 +217,11 @@ static bool ActionFailed(const Action *A,
   if (FailingCommands.empty())
     return false;
 
-  // CUDA/HIP can have the same input source code compiled multiple times so do
-  // not compiled again if there are already failures. It is OK to abort the
-  // CUDA pipeline on errors.
-  if (A->isOffloading(Action::OFK_Cuda) || A->isOffloading(Action::OFK_HIP))
+  // CUDA/HIP/SYCL can have the same input source code compiled multiple times
+  // so do not compile again if there are already failures. It is OK to abort
+  // the CUDA pipeline on errors.
+  if (A->isOffloading(Action::OFK_Cuda) || A->isOffloading(Action::OFK_HIP) ||
+      A->isOffloading(Action::OFK_SYCL))
     return true;
 
   for (const auto &CI : FailingCommands)

clang/lib/Driver/Driver.cpp

@@ -43,6 +43,7 @@
 #include "ToolChains/PS4CPU.h"
 #include "ToolChains/RISCVToolchain.h"
 #include "ToolChains/SPIRV.h"
+#include "ToolChains/SYCL.h"
 #include "ToolChains/Solaris.h"
 #include "ToolChains/TCE.h"
 #include "ToolChains/VEToolchain.h"
@@ -767,6 +768,26 @@ Driver::OpenMPRuntimeKind Driver::getOpenMPRuntime(const ArgList &Args) const {
   return RT;
 }
 
+static const char *getDefaultSYCLArch(Compilation &C) {
+  if (C.getDefaultToolChain().getTriple().getArch() == llvm::Triple::x86)
+    return "spir";
+  return "spir64";
+}
+
+static bool addSYCLDefaultTriple(Compilation &C,
+                                 SmallVectorImpl<llvm::Triple> &SYCLTriples) {
+  for (const auto &SYCLTriple : SYCLTriples) {
+    if (SYCLTriple.getSubArch() == llvm::Triple::NoSubArch &&
+        SYCLTriple.isSPIROrSPIRV())
+      return false;
+  }
+  // Add the default triple as it was not found.
+  llvm::Triple DefaultTriple =
+      C.getDriver().MakeSYCLDeviceTriple(getDefaultSYCLArch(C));
+  SYCLTriples.insert(SYCLTriples.begin(), DefaultTriple);
+  return true;
+}
+
 void Driver::CreateOffloadingDeviceToolChains(Compilation &C,
                                               InputList &Inputs) {
 
@@ -979,6 +1000,44 @@ void Driver::CreateOffloadingDeviceToolChains(Compilation &C,
     return;
   }
 
+  //
+  // SYCL
+  //
+  // We need to generate a SYCL toolchain if the user specified -fsycl.
+  // If -fsycl is supplied we will assume SPIR-V.
+  bool IsSYCL =
+      C.getInputArgs().hasFlag(options::OPT_fsycl, options::OPT_fno_sycl,
+                               false);
+
+  auto argSYCLIncompatible = [&](OptSpecifier OptId) {
+    if (!IsSYCL)
+      return;
+    if (Arg *IncompatArg = C.getInputArgs().getLastArg(OptId))
+      Diag(clang::diag::err_drv_argument_not_allowed_with)
+          << IncompatArg->getSpelling() << "-fsycl";
+  };
+  // -static-libstdc++ is not compatible with -fsycl.
+  argSYCLIncompatible(options::OPT_static_libstdcxx);
+  // -ffreestanding cannot be used with -fsycl
+  argSYCLIncompatible(options::OPT_ffreestanding);
+
+  llvm::SmallVector<llvm::Triple, 4> UniqueSYCLTriplesVec;
+
+  // If -fsycl is supplied we will assume SPIR-V.
+  if (IsSYCL) {
+    addSYCLDefaultTriple(C, UniqueSYCLTriplesVec);
+
+    // We'll need to use the SYCL and host triples as the key into
+    // getOffloadingDeviceToolChain, because the device toolchains we're
+    // going to create will depend on both.
+    const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>();
+    for (const auto &TT : UniqueSYCLTriplesVec) {
+      auto SYCLTC = &getOffloadingDeviceToolChain(C.getInputArgs(), TT, *HostTC,
+                                                  Action::OFK_SYCL);
+      C.addOffloadDeviceToolChain(SYCLTC, Action::OFK_SYCL);
+    }
+  }
+
   //
   // TODO: Add support for other offloading programming models here.
   //
@@ -2001,6 +2060,19 @@ void Driver::PrintHelp(bool ShowHidden) const {
                       VisibilityMask);
 }
 
+llvm::Triple Driver::MakeSYCLDeviceTriple(StringRef TargetArch) const {
+  SmallVector<StringRef, 5> SYCLAlias = { "spir", "spir64", "spirv32", "spirv64"};
+  if (std::find(SYCLAlias.begin(), SYCLAlias.end(), TargetArch) !=
+      SYCLAlias.end()) {
+    llvm::Triple TT;
+    TT.setArchName(TargetArch);
+    TT.setVendor(llvm::Triple::UnknownVendor);
+    TT.setOS(llvm::Triple::UnknownOS);
+    return TT;
+  }
+  return llvm::Triple(TargetArch);
+}
+
 void Driver::PrintVersion(const Compilation &C, raw_ostream &OS) const {
   if (IsFlangMode()) {
     OS << getClangToolFullVersion("flang-new") << '\n';
@@ -4155,6 +4227,7 @@ void Driver::BuildActions(Compilation &C, DerivedArgList &Args,
 
   bool UseNewOffloadingDriver =
       C.isOffloadingHostKind(Action::OFK_OpenMP) ||
+      C.isOffloadingHostKind(Action::OFK_SYCL) ||
       Args.hasFlag(options::OPT_foffload_via_llvm,
                    options::OPT_fno_offload_via_llvm, false) ||
       Args.hasFlag(options::OPT_offload_new_driver,
@@ -4565,6 +4638,8 @@ Driver::getOffloadArchs(Compilation &C, const llvm::opt::DerivedArgList &Args,
       Archs.insert(OffloadArchToString(OffloadArch::HIPDefault));
     else if (Kind == Action::OFK_OpenMP)
       Archs.insert(StringRef());
+    else if (Kind == Action::OFK_SYCL)
+      Archs.insert(StringRef());
   } else {
     Args.ClaimAllArgs(options::OPT_offload_arch_EQ);
     Args.ClaimAllArgs(options::OPT_no_offload_arch_EQ);
@@ -4589,7 +4664,7 @@ Action *Driver::BuildOffloadingActions(Compilation &C,
   OffloadAction::DeviceDependences DDeps;
 
   const Action::OffloadKind OffloadKinds[] = {
-      Action::OFK_OpenMP, Action::OFK_Cuda, Action::OFK_HIP};
+      Action::OFK_OpenMP, Action::OFK_Cuda, Action::OFK_HIP, Action::OFK_SYCL};
 
   for (Action::OffloadKind Kind : OffloadKinds) {
     SmallVector<const ToolChain *, 2> ToolChains;
@@ -4634,6 +4709,11 @@ Action *Driver::BuildOffloadingActions(Compilation &C,
         break;
       }
 
+      // Assemble actions are not used for the SYCL device side.  Both compile
+      // and backend actions are used to generate IR and textual IR if needed.
+      if (Kind == Action::OFK_SYCL && Phase == phases::Assemble)
+        continue;
+
       auto TCAndArch = TCAndArchs.begin();
       for (Action *&A : DeviceActions) {
         if (A->getType() == types::TY_Nothing)
@@ -4867,12 +4947,13 @@ Action *Driver::ConstructPhaseAction(
       return C.MakeAction<BackendJobAction>(Input, Output);
     }
     if (Args.hasArg(options::OPT_emit_llvm) ||
-        (((Input->getOffloadingToolChain() &&
-           Input->getOffloadingToolChain()->getTriple().isAMDGPU()) ||
-          TargetDeviceOffloadKind == Action::OFK_HIP) &&
-         (Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc,
-                       false) ||
-          TargetDeviceOffloadKind == Action::OFK_OpenMP))) {
+        (TargetDeviceOffloadKind == Action::OFK_SYCL ||
+         (((Input->getOffloadingToolChain() &&
+            Input->getOffloadingToolChain()->getTriple().isAMDGPU()) ||
+           TargetDeviceOffloadKind == Action::OFK_HIP) &&
+          (Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc,
+                        false) ||
+           TargetDeviceOffloadKind == Action::OFK_OpenMP)))) {
       types::ID Output =
           Args.hasArg(options::OPT_S) &&
                   (TargetDeviceOffloadKind == Action::OFK_None ||
@@ -6553,6 +6634,20 @@ const ToolChain &Driver::getOffloadingDeviceToolChain(
                                                            HostTC, Args);
       break;
     }
+    case Action::OFK_SYCL:
+      // TODO: Add additional Arch values for Ahead of Time support for SYCL.
+      switch (Target.getArch()) {
+      case llvm::Triple::spir:
+      case llvm::Triple::spir64:
+      case llvm::Triple::spirv32:
+      case llvm::Triple::spirv64:
+        TC = std::make_unique<toolchains::SYCLToolChain>(*this, Target, HostTC,
+                                                         Args);
+        break;
+      default:
+        break;
+      }
+      break;
     default:
       break;
     }

clang/lib/Driver/ToolChain.cpp

@@ -1441,6 +1441,9 @@ void ToolChain::AddCudaIncludeArgs(const ArgList &DriverArgs,
 void ToolChain::AddHIPIncludeArgs(const ArgList &DriverArgs,
                                   ArgStringList &CC1Args) const {}
 
+void ToolChain::AddSYCLIncludeArgs(const ArgList &DriverArgs,
+                                   ArgStringList &CC1Args) const {}
+
 llvm::SmallVector<ToolChain::BitCodeLibraryInfo, 12>
 ToolChain::getDeviceLibs(const ArgList &DriverArgs) const {
   return {};