[Driver][SYCL] Refactor -fsycl-device-only behaviors (#2713)

Update -fsycl-device-only to go through the offload path, effectively
removing the host dependencies when it is used.  This allows for a
cleaner representation of SYCL device specific compilation needs.

Author: mdtoguchi

clang/lib/Driver/Driver.cpp

@@ -792,9 +792,10 @@ void Driver::CreateOffloadingDeviceToolChains(Compilation &C,
   // the -fsycl-targets, -fsycl-add-targets or -fsycl-link-targets option.
   // If -fsycl is supplied without any of these we will assume SPIR-V.
   // Use of -fsycl-device-only overrides -fsycl.
-  bool HasValidSYCLRuntime = (C.getInputArgs().hasFlag(options::OPT_fsycl,
-      options::OPT_fno_sycl, false) &&
-      !C.getInputArgs().hasArg(options::OPT_fsycl_device_only));
+  bool HasValidSYCLRuntime =
+      (C.getInputArgs().hasFlag(options::OPT_fsycl, options::OPT_fno_sycl,
+                                false) ||
+       C.getInputArgs().hasArg(options::OPT_fsycl_device_only));
 
   // A mechanism for retrieving SYCL-specific options, erroring out
   // if SYCL offloading wasn't enabled prior to that
@@ -913,11 +914,18 @@ void Driver::CreateOffloadingDeviceToolChains(Compilation &C,
   } else {
     // If -fsycl is supplied without -fsycl-*targets we will assume SPIR-V
     // unless -fintelfpga is supplied, which uses SPIR-V with fpga AOT.
-    if (HasValidSYCLRuntime) {
+    // For -fsycl-device-only, we also setup the implied triple as needed.
+    StringRef SYCLTargetArch;
+    if (C.getInputArgs().hasArg(options::OPT_fsycl_device_only))
+      if (C.getDefaultToolChain().getTriple().getArch() == llvm::Triple::x86)
+        SYCLTargetArch = "spir";
+      else
+        SYCLTargetArch = "spir64";
+    else if (HasValidSYCLRuntime)
       // Triple for -fintelfpga is spir64_fpga-unknown-unknown-sycldevice.
-      const char *SYCLTargetArch = SYCLfpga ? "spir64_fpga" : "spir64";
+      SYCLTargetArch = SYCLfpga ? "spir64_fpga" : "spir64";
+    if (!SYCLTargetArch.empty())
       UniqueSYCLTriplesVec.push_back(MakeSYCLDeviceTriple(SYCLTargetArch));
-    }
   }
   // We'll need to use the SYCL and host triples as the key into
   // getOffloadingDeviceToolChain, because the device toolchains we're
@@ -3700,10 +3708,24 @@ class OffloadingActionBuilder final {
       // The host depends on the generated integrated header from the device
       // compilation.
       if (CurPhase == phases::Compile) {
+        bool SYCLDeviceOnly = Args.hasArg(options::OPT_fsycl_device_only);
         for (Action *&A : SYCLDeviceActions) {
           DeviceCompilerInput =
               C.MakeAction<CompileJobAction>(A, types::TY_SYCL_Header);
-          A = C.MakeAction<CompileJobAction>(A, types::TY_LLVM_BC);
+          types::ID OutputType = types::TY_LLVM_BC;
+          if (SYCLDeviceOnly) {
+            if (Args.hasArg(options::OPT_S))
+              OutputType = types::TY_LLVM_IR;
+            if (Args.hasFlag(options::OPT_fno_sycl_use_bitcode,
+                             options::OPT_fsycl_use_bitcode, false)) {
+              auto *CompileAction =
+                  C.MakeAction<CompileJobAction>(A, types::TY_LLVM_BC);
+              A = C.MakeAction<SPIRVTranslatorJobAction>(CompileAction,
+                                                         types::TY_SPIRV);
+              continue;
+            }
+          }
+          A = C.MakeAction<CompileJobAction>(A, OutputType);
         }
         const auto *TC = ToolChains.front();
         const char *BoundArch = nullptr;
@@ -3713,7 +3735,7 @@ class OffloadingActionBuilder final {
         // Clear the input file, it is already a dependence to a host
         // action.
         DeviceCompilerInput = nullptr;
-        return ABRT_Success;
+        return SYCLDeviceOnly ? ABRT_Ignore_Host : ABRT_Success;
       }
 
       // Backend/Assemble actions are obsolete for the SYCL device side
@@ -4351,8 +4373,8 @@ class OffloadingActionBuilder final {
       Arg *SYCLTargets =
               C.getInputArgs().getLastArg(options::OPT_fsycl_targets_EQ);
       Arg *SYCLAddTargets = Args.getLastArg(options::OPT_fsycl_add_targets_EQ);
-      bool HasValidSYCLRuntime = C.getInputArgs().hasFlag(options::OPT_fsycl,
-                                              options::OPT_fno_sycl, false);
+      bool HasValidSYCLRuntime = C.getInputArgs().hasFlag(
+          options::OPT_fsycl, options::OPT_fno_sycl, false);
       bool SYCLfpgaTriple = false;
       if (SYCLTargets || SYCLAddTargets) {
         if (SYCLTargets) {
@@ -5411,19 +5433,6 @@ Action *Driver::ConstructPhaseAction(
           Args.hasArg(options::OPT_S) ? types::TY_LLVM_IR : types::TY_LLVM_BC;
       return C.MakeAction<BackendJobAction>(Input, Output);
     }
-    if (Args.hasArg(options::OPT_fsycl_device_only)) {
-      types::ID OutputType =
-          Args.hasArg(options::OPT_S) ? types::TY_LLVM_IR : types::TY_LLVM_BC;
-      if (Args.hasFlag(options::OPT_fsycl_use_bitcode,
-                       options::OPT_fno_sycl_use_bitcode, true))
-        return C.MakeAction<BackendJobAction>(Input, OutputType);
-      // Use of -fsycl-device-only creates a bitcode file, we need to translate
-      // that to a SPIR-V file with -fno-sycl-use-bitcode
-      auto *BackendAction =
-          C.MakeAction<BackendJobAction>(Input, types::TY_LLVM_BC);
-      return C.MakeAction<SPIRVTranslatorJobAction>(BackendAction,
-                                                    types::TY_SPIRV);
-    }
     return C.MakeAction<BackendJobAction>(Input, types::TY_PP_Asm);
   }
   case phases::Assemble:

clang/lib/Driver/ToolChains/Clang.cpp

@@ -1238,8 +1238,7 @@ void Clang::AddPreprocessingOptions(Compilation &C, const JobAction &JA,
   if (JA.isOffloading(Action::OFK_HIP))
     getToolChain().AddHIPIncludeArgs(Args, CmdArgs);
 
-  if (JA.isOffloading(Action::OFK_SYCL) ||
-      Args.hasArg(options::OPT_fsycl_device_only))
+  if (JA.isOffloading(Action::OFK_SYCL))
     toolchains::SYCLToolChain::AddSYCLIncludeArgs(D, Args, CmdArgs);
 
   // If we are offloading to a target via OpenMP we need to include the
@@ -4035,24 +4034,14 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
                          const ArgList &Args, const char *LinkingOutput) const {
   const auto &TC = getToolChain();
   const llvm::Triple &RawTriple = TC.getTriple();
-  llvm::Triple Triple = TC.getEffectiveTriple();
+  const llvm::Triple &Triple = TC.getEffectiveTriple();
+  const std::string &TripleStr = Triple.getTriple();
 
   bool KernelOrKext =
       Args.hasArg(options::OPT_mkernel, options::OPT_fapple_kext);
   const Driver &D = TC.getDriver();
   ArgStringList CmdArgs;
 
-  // -fsycl-device-only implies a SPIRV arch triple.  Do not set if current
-  // effective triple is SYCLDevice
-  if (Args.hasArg(options::OPT_fsycl_device_only) &&
-      Triple.getEnvironment() != llvm::Triple::SYCLDevice) {
-    const char *SYCLTargetArch = "spir64";
-    if (C.getDefaultToolChain().getTriple().getArch() == llvm::Triple::x86)
-      SYCLTargetArch = "spir";
-    Triple = C.getDriver().MakeSYCLDeviceTriple(SYCLTargetArch);
-  }
-  const std::string &TripleStr = Triple.getTriple();
-
   // Check number of inputs for sanity. We need at least one input.
   assert(Inputs.size() >= 1 && "Must have at least one input.");
   // CUDA/HIP compilation may have multiple inputs (source file + results of
@@ -4067,8 +4056,7 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
   bool IsHIP = JA.isOffloading(Action::OFK_HIP);
   bool IsOpenMPDevice = JA.isDeviceOffloading(Action::OFK_OpenMP);
   bool IsSYCLOffloadDevice = JA.isDeviceOffloading(Action::OFK_SYCL);
-  bool IsSYCL = JA.isOffloading(Action::OFK_SYCL) ||
-                Args.hasArg(options::OPT_fsycl_device_only);
+  bool IsSYCL = JA.isOffloading(Action::OFK_SYCL);
   bool IsHeaderModulePrecompile = isa<HeaderModulePrecompileJobAction>(JA);
   assert((IsCuda || IsHIP || (IsOpenMPDevice && Inputs.size() == 2) || IsSYCL ||
           IsHeaderModulePrecompile || Inputs.size() == 1) &&
@@ -4124,9 +4112,7 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
   // step, and being part of the SYCL tool chain causes the incorrect target.
   // FIXME - Is it possible to retain host environment when on a target
   // device toolchain.
-  bool UseSYCLTriple =
-      IsSYCLDevice && (!IsSYCL || IsSYCLOffloadDevice ||
-                       Args.hasArg(options::OPT_fsycl_device_only));
+  bool UseSYCLTriple = IsSYCLDevice && (!IsSYCL || IsSYCLOffloadDevice);
 
   // Adjust IsWindowsXYZ for CUDA/HIP/SYCL compilations.  Even when compiling in
   // device mode (i.e., getToolchain().getTriple() is NVPTX/AMDGCN, not
@@ -4228,9 +4214,6 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
 
     // Pass the triple of host when doing SYCL
     llvm::Triple AuxT = C.getDefaultToolChain().getTriple();
-    if (Args.hasArg(options::OPT_fsycl_device_only) &&
-        RawTriple.getEnvironment() == llvm::Triple::SYCLDevice)
-      AuxT = llvm::Triple(llvm::sys::getProcessTriple());
     std::string NormalizedTriple = AuxT.normalize();
     CmdArgs.push_back("-aux-triple");
     CmdArgs.push_back(Args.MakeArgString(NormalizedTriple));
@@ -6371,8 +6354,6 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
         // SYCL library is guaranteed to work correctly only with dynamic
         // MSVC runtime.
         llvm::Triple AuxT = C.getDefaultToolChain().getTriple();
-        if (Args.hasFlag(options::OPT_fsycl_device_only, OptSpecifier(), false))
-          AuxT = llvm::Triple(llvm::sys::getProcessTriple());
         if (AuxT.isWindowsMSVCEnvironment()) {
           CmdArgs.push_back("-D_MT");
           CmdArgs.push_back("-D_DLL");
@@ -6897,7 +6878,6 @@ void Clang::AddClangCLArgs(const ArgList &Args, types::ID InputType,
   unsigned RTOptionID = options::OPT__SLASH_MT;
   bool isNVPTX = getToolChain().getTriple().isNVPTX();
   bool isSYCLDevice =
-      Args.hasArg(options::OPT_fsycl_device_only) ||
       getToolChain().getTriple().getEnvironment() == llvm::Triple::SYCLDevice;
   bool isSYCL = Args.hasArg(options::OPT_fsycl) || isSYCLDevice;
   // For SYCL Windows, /MD is the default.
@@ -7912,8 +7892,7 @@ void SPIRVTranslator::ConstructJob(Compilation &C, const JobAction &JA,
 
   TranslatorArgs.push_back("-o");
   TranslatorArgs.push_back(Output.getFilename());
-  if (getToolChain().getTriple().isSYCLDeviceEnvironment() ||
-      TCArgs.hasArg(options::OPT_fsycl_device_only)) {
+  if (getToolChain().getTriple().isSYCLDeviceEnvironment()) {
     TranslatorArgs.push_back("-spirv-max-version=1.1");
     TranslatorArgs.push_back("-spirv-debug-info-version=legacy");
     // Prevent crash in the translator if input IR contains DIExpression

clang/lib/Driver/ToolChains/Cuda.cpp

@@ -642,7 +642,8 @@ void CudaToolChain::addClangTargetOptions(
                                                   CC1Args);
   }
 
-  auto NoLibSpirv = DriverArgs.hasArg(options::OPT_fno_sycl_libspirv);
+  auto NoLibSpirv = DriverArgs.hasArg(options::OPT_fno_sycl_libspirv,
+                                      options::OPT_fsycl_device_only);
   if (DeviceOffloadingKind == Action::OFK_SYCL && !NoLibSpirv) {
     std::string LibSpirvFile;
 

clang/test/Driver/sycl.c

@@ -73,10 +73,10 @@
 
 /// Verify -fsycl-device-only phases
 // RUN: %clang -### -ccc-print-phases -fsycl-device-only %s 2>&1 | FileCheck %s --check-prefix=DEFAULT-PHASES
-// DEFAULT-PHASES: 0: input, "{{.*}}", c++
-// DEFAULT-PHASES: 1: preprocessor, {0}, c++-cpp-output
-// DEFAULT-PHASES: 2: compiler, {1}, ir
-// DEFAULT-PHASES: 3: backend, {2}, ir
+// DEFAULT-PHASES: 0: input, "{{.*}}", c++, (device-sycl)
+// DEFAULT-PHASES: 1: preprocessor, {0}, c++-cpp-output, (device-sycl)
+// DEFAULT-PHASES: 2: compiler, {1}, ir, (device-sycl)
+// DEFAULT-PHASES: 3: offload, "device-sycl (spir64-unknown-unknown-sycldevice)" {2}, ir
 // DEFAULT-PHASES-NOT: linker
 
 // -fsycl-help tests

libclc/cmake/modules/AddLibclc.cmake

@@ -219,7 +219,7 @@ function(add_libclc_sycl_binding OUT_LIST)
                          COMMAND ${CMAKE_COMMAND} -E make_directory
                          ${CMAKE_CURRENT_BINARY_DIR}/sycldevice-binding-${ARG_TRIPLE}
                          COMMAND ${LLVM_CLANG}
-                         -target ${ARG_TRIPLE}-sycldevice
+                         -fsycl-targets=${ARG_TRIPLE}-sycldevice
                          -fsycl
                          -fsycl-device-only
                          -Dcl_khr_fp64