Merge remote-tracking branch 'intel/sycl' into steffen/enable_threading_windows

Author: steffenlarsen

.github/workflows/sycl-linux-precommit.yml

@@ -117,11 +117,11 @@ jobs:
       env: ${{ matrix.env || '{}' }}
 
       # Do not install drivers on AMD and CUDA runners.
-      install_igc_driver: |
+      install_igc_driver: >-
         ${{ !contains(matrix.target_devices, 'ext_oneapi_cuda') &&
         !contains(matrix.target_devices, 'ext_oneapi_hip') &&
         contains(needs.detect_changes.outputs.filters, 'drivers') }}
-      install_dev_igc_driver: |
+      install_dev_igc_driver: >-
         ${{ !contains(matrix.target_devices, 'ext_oneapi_cuda') &&
         !contains(matrix.target_devices, 'ext_oneapi_hip') &&
         matrix.use_igc_dev && contains(needs.detect_changes.outputs.filters, 'devigccfg') ||

.github/workflows/sycl-linux-run-tests.yml

@@ -151,7 +151,7 @@ permissions:
 
 jobs:
   run:
-    if: inputs.skip_run == 'false'
+    if: github.event_name == 'workflow_dispatch' || inputs.skip_run == 'false'
     name: ${{ inputs.name }}
     runs-on: ${{ fromJSON(inputs.runner) }}
     container:

.github/workflows/sycl-windows-run-tests.yml

@@ -86,6 +86,17 @@ jobs:
         # Run E2E tests.
         export LIT_OPTS="-v --no-progress-bar --show-unsupported --show-pass --show-xfail --max-time 3600 --time-tests ${{ inputs.extra_lit_opts }}"
         cmake --build build-e2e --target check-sycl-e2e
+    - name: Detect hung tests
+      shell: powershell
+      run: |
+        $exitCode = 0
+        $hungTests = Get-Process | Where-Object { ($_.Path -match "llvm\\install") -or ($_.Path -match "llvm\\build-e2e") }
+        $hungTests | Foreach-Object {
+         $exitCode = 1
+         echo "Test $($_.Path) hung!"
+         Stop-Process -Force $_
+        }
+        exit $exitCode
     - name: Cleanup
       shell: cmd
       if: always()

clang/include/clang/Basic/LangOptions.def

@@ -307,7 +307,6 @@ LANGOPT(SYCLAllowFuncPtr  , 1, 0, "Allow function pointers in SYCL device code")
 LANGOPT(SYCLStdLayoutKernelParams, 1, 0, "Enable standard layout requirement for SYCL kernel parameters")
 LANGOPT(SYCLUnnamedLambda , 1, 0, "Allow unnamed lambda SYCL kernels")
 LANGOPT(SYCLForceInlineKernelLambda , 1, 0, "Force inline SYCL kernel lambdas in entry point")
-LANGOPT(SYCLAllowAllFeaturesInConstexpr, 1, 0, "Allow all C++ features in SYCL device code in manifestly constant-evaluated expressions")
 LANGOPT(SYCLESIMDForceStatelessMem, 1, 0, "Make accessors use USM memory in ESIMD kernels")
 LANGOPT(SYCLESIMDBuildHostCode, 1, 1, "Build the host implementation of ESIMD functions")
 ENUM_LANGOPT(SYCLVersion  , SYCLMajorVersion, 2, SYCL_None, "Version of the SYCL standard used")

clang/include/clang/Driver/Options.td

@@ -8628,11 +8628,6 @@ def fsycl_is_native_cpu : Flag<["-"], "fsycl-is-native-cpu">,
   HelpText<"Perform device compilation for Native CPU.">,
   Visibility<[CC1Option]>,
   MarshallingInfoFlag<LangOpts<"SYCLIsNativeCPU">>;
-defm sycl_allow_all_features_in_constexpr
-    : BoolFOption<
-          "sycl-allow-all-features-in-constexpr", LangOpts<"SYCLAllowAllFeaturesInConstexpr">,
-          DefaultFalse,
-          PosFlag<SetTrue, [], [CC1Option], "Allow all C++ features in SYCL device code in manifestly constant-evaluated expressions">, NegFlag<SetFalse>>;
 
 } // let Visibility = [CC1Option]
 

clang/include/clang/Driver/ToolChain.h

@@ -781,6 +781,10 @@ 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/Driver.cpp

@@ -6786,13 +6786,10 @@ class OffloadingActionBuilder final {
     // Do not use unbundler if the Host does not depend on device action.
     // Now that we have unbundled the object, when doing -fsycl-link we
     // want to continue the host link with the input object.
-    // For unbundling of an FPGA AOCX binary, we want to link with the original
-    // FPGA device archive.
     if ((OffloadKind == Action::OFK_None && CanUseBundler) ||
         (Args.hasArg(options::OPT_fsycl_link_EQ) && !HasFPGATarget) ||
         (HasFPGATarget && ((Args.hasArg(options::OPT_fsycl_link_EQ) &&
-                            HostAction->getType() == types::TY_Object) ||
-                           HostAction->getType() == types::TY_FPGA_AOCX)))
+                            HostAction->getType() == types::TY_Object))))
       if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(HostAction))
         HostAction = UA->getInputs().back();
 
@@ -7443,16 +7440,36 @@ void Driver::BuildActions(Compilation &C, DerivedArgList &Args,
 
   // For an FPGA archive, we add the unbundling step above to take care of
   // the device side, but also unbundle here to extract the host side
-  bool EarlyLink = false;
-  if (const Arg *A = Args.getLastArg(options::OPT_fsycl_link_EQ))
-    EarlyLink = A->getValue() == StringRef("early");
   for (auto &LI : LinkerInputs) {
     Action *UnbundlerInput = nullptr;
     auto wrapObject = [&] {
-      if (EarlyLink && Args.hasArg(options::OPT_fintelfpga)) {
-        // Only wrap the object with -fsycl-link=early
-        auto *BC = C.MakeAction<OffloadWrapperJobAction>(LI, types::TY_LLVM_BC);
-        auto *ASM = C.MakeAction<BackendJobAction>(BC, types::TY_PP_Asm);
+      if (Args.hasArg(options::OPT_fsycl_link_EQ) &&
+          Args.hasArg(options::OPT_fintelfpga)) {
+        // Wrap the object when creating an FPGA AOCX or AOCR binary.
+        // When the input file is an AOCR (early) archive, the unbundled host
+        // binary consists of a list of objects. We cannot directly wrap that
+        // binary to be consumed later - this has to go through each listed
+        // object.
+        bool FPGAEarly = true;
+        if (auto *A = C.getInputArgs().getLastArg(options::OPT_fsycl_link_EQ))
+          FPGAEarly = A->getValue() == StringRef("early");
+
+        Action *WrapperAction;
+        if ((LI->getType() == types::TY_FPGA_AOCR ||
+             LI->getType() == types::TY_FPGA_AOCR_EMU) &&
+            !FPGAEarly) {
+          auto *RenameAction = C.MakeAction<FileTableTformJobAction>(
+              LI, types::TY_Tempfilelist, types::TY_Tempfilelist);
+          RenameAction->addRenameColumnTform(FileTableTformJobAction::COL_ZERO,
+                                             FileTableTformJobAction::COL_CODE);
+          ActionList WrapperItems({RenameAction});
+          WrapperAction = C.MakeAction<OffloadWrapperJobAction>(
+              WrapperItems, types::TY_LLVM_BC);
+        } else
+          WrapperAction =
+              C.MakeAction<OffloadWrapperJobAction>(LI, types::TY_LLVM_BC);
+        auto *ASM =
+            C.MakeAction<BackendJobAction>(WrapperAction, types::TY_PP_Asm);
         auto *OBJ = C.MakeAction<AssembleJobAction>(ASM, types::TY_Object);
         OffloadAction::HostDependence HDep(
             *OBJ, *C.getSingleOffloadToolChain<Action::OFK_Host>(),

clang/lib/Driver/ToolChain.cpp

@@ -1519,6 +1519,9 @@ 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,

clang/lib/Driver/ToolChains/Clang.cpp

@@ -1160,7 +1160,7 @@ void Clang::AddPreprocessingOptions(Compilation &C, const JobAction &JA,
     getToolChain().AddHIPIncludeArgs(Args, CmdArgs);
 
   if (JA.isOffloading(Action::OFK_SYCL)) {
-    toolchains::SYCLToolChain::AddSYCLIncludeArgs(D, Args, CmdArgs);
+    getToolChain().AddSYCLIncludeArgs(Args, CmdArgs);
     if (Inputs[0].getType() == types::TY_CUDA) {
       // Include __clang_cuda_runtime_wrapper.h in .cu SYCL compilation.
       getToolChain().AddCudaIncludeArgs(Args, CmdArgs);
@@ -10216,7 +10216,7 @@ void OffloadWrapper::ConstructJob(Compilation &C, const JobAction &JA,
 
     if (I.getType() == types::TY_Tempfiletable ||
         I.getType() == types::TY_Tempfilelist || IsEmbeddedIR)
-      // wrapper actual input files are passed via the batch job file table:
+      // Input files are passed via the batch job file table.
       WrapperArgs.push_back(C.getArgs().MakeArgString("-batch"));
     WrapperArgs.push_back(C.getArgs().MakeArgString(I.getFilename()));
 
@@ -10283,6 +10283,11 @@ void OffloadWrapper::ConstructJob(Compilation &C, const JobAction &JA,
     CmdArgs.push_back(
         TCArgs.MakeArgString(Twine("-target=") + Triple.getTriple()));
 
+    if (Inputs[0].getType() == types::TY_Tempfiletable ||
+        Inputs[0].getType() == types::TY_Tempfilelist)
+      // Input files are passed via the batch job file table.
+      CmdArgs.push_back(C.getArgs().MakeArgString("-batch"));
+
     // Add input.
     assert(Inputs[0].isFilename() && "Invalid input.");
     CmdArgs.push_back(TCArgs.MakeArgString(Inputs[0].getFilename()));
@@ -10320,7 +10325,7 @@ void OffloadWrapper::ConstructJob(Compilation &C, const JobAction &JA,
 
     if (Inputs[I].getType() == types::TY_Tempfiletable ||
         Inputs[I].getType() == types::TY_Tempfilelist)
-      // wrapper actual input files are passed via the batch job file table:
+      // Input files are passed via the batch job file table.
       CmdArgs.push_back(C.getArgs().MakeArgString("-batch"));
 
     // Add input.

clang/lib/Driver/ToolChains/Cuda.cpp

@@ -924,7 +924,7 @@ CudaToolChain::CudaToolChain(const Driver &D, const llvm::Triple &Triple,
                              const ToolChain &HostTC, const ArgList &Args,
                              const Action::OffloadKind OK)
     : NVPTXToolChain(D, Triple, HostTC.getTriple(), Args), HostTC(HostTC),
-      OK(OK) {}
+      SYCLInstallation(D), OK(OK) {}
 
 void CudaToolChain::addClangTargetOptions(
     const llvm::opt::ArgList &DriverArgs, llvm::opt::ArgStringList &CC1Args,
@@ -941,11 +941,19 @@ void CudaToolChain::addClangTargetOptions(
   // If we are compiling SYCL kernels for Nvidia GPUs, we do not support Cuda
   // device code compatability, hence we do not set Cuda mode in that instance.
   if (DeviceOffloadingKind == Action::OFK_SYCL) {
-    toolchains::SYCLToolChain::AddSYCLIncludeArgs(getDriver(), DriverArgs,
-                                                  CC1Args);
+    SYCLInstallation.AddSYCLIncludeArgs(DriverArgs, CC1Args);
 
     if (DriverArgs.hasArg(options::OPT_fsycl_fp32_prec_sqrt))
       CC1Args.push_back("-fcuda-prec-sqrt");
+
+    bool FastRelaxedMath = DriverArgs.hasFlag(
+        options::OPT_ffast_math, options::OPT_fno_fast_math, false);
+    bool UnsafeMathOpt =
+        DriverArgs.hasFlag(options::OPT_funsafe_math_optimizations,
+                           options::OPT_fno_unsafe_math_optimizations, false);
+    if (FastRelaxedMath || UnsafeMathOpt)
+      CC1Args.append({"-mllvm", "--nvptx-prec-divf32=0", "-mllvm",
+                      "--nvptx-prec-sqrtf32=0"});
   } else {
     CC1Args.append(
         {"-fcuda-is-device", "-mllvm", "-enable-memcpyopt-without-libcalls"});
@@ -1196,8 +1204,7 @@ CudaToolChain::GetCXXStdlibType(const ArgList &Args) const {
 void CudaToolChain::AddClangSystemIncludeArgs(const ArgList &DriverArgs,
                                               ArgStringList &CC1Args) const {
   if (DriverArgs.hasArg(options::OPT_fsycl)) {
-    toolchains::SYCLToolChain::AddSYCLIncludeArgs(getDriver(), DriverArgs,
-                                                  CC1Args);
+    SYCLInstallation.AddSYCLIncludeArgs(DriverArgs, CC1Args);
   }
   HostTC.AddClangSystemIncludeArgs(DriverArgs, CC1Args);