Merge remote-tracking branch 'upstream/sycl' into fallback_sort

Author: jinge90

.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:

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

@@ -7446,8 +7446,30 @@ void Driver::BuildActions(Compilation &C, DerivedArgList &Args,
       if (Args.hasArg(options::OPT_fsycl_link_EQ) &&
           Args.hasArg(options::OPT_fintelfpga)) {
         // Wrap the object when creating an FPGA AOCX or AOCR binary.
-        auto *BC = C.MakeAction<OffloadWrapperJobAction>(LI, types::TY_LLVM_BC);
-        auto *ASM = C.MakeAction<BackendJobAction>(BC, types::TY_PP_Asm);
+        // 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);
 

clang/lib/Driver/ToolChains/Cuda.h

@@ -260,6 +260,8 @@ class LLVM_LIBRARY_VISIBILITY CudaToolChain : public NVPTXToolChain {
   Tool *SelectTool(const JobAction &JA) const override;
   const ToolChain &HostTC;
 
+  SYCLInstallationDetector SYCLInstallation;
+
 protected:
   Tool *buildAssembler() const override; // ptxas
   Tool *buildLinker() const override;    // fatbinary (ok, not really a linker)

clang/lib/Driver/ToolChains/Gnu.cpp

@@ -3190,7 +3190,8 @@ bool Generic_GCC::GCCInstallationDetector::ScanGentooGccConfig(
 Generic_GCC::Generic_GCC(const Driver &D, const llvm::Triple &Triple,
                          const ArgList &Args)
     : ToolChain(D, Triple, Args), GCCInstallation(D),
-      CudaInstallation(D, Triple, Args), RocmInstallation(D, Triple, Args) {
+      CudaInstallation(D, Triple, Args), RocmInstallation(D, Triple, Args),
+      SYCLInstallation(D) {
   getProgramPaths().push_back(getDriver().Dir);
 }
 

clang/lib/Driver/ToolChains/Gnu.h

@@ -294,6 +294,7 @@ class LLVM_LIBRARY_VISIBILITY Generic_GCC : public ToolChain {
   GCCInstallationDetector GCCInstallation;
   LazyDetector<CudaInstallationDetector> CudaInstallation;
   LazyDetector<RocmInstallationDetector> RocmInstallation;
+  SYCLInstallationDetector SYCLInstallation;
 
 public:
   Generic_GCC(const Driver &D, const llvm::Triple &Triple,

clang/lib/Driver/ToolChains/HIPAMD.cpp

@@ -253,7 +253,8 @@ void AMDGCN::Linker::ConstructJob(Compilation &C, const JobAction &JA,
 HIPAMDToolChain::HIPAMDToolChain(const Driver &D, const llvm::Triple &Triple,
                                  const ToolChain &HostTC, const ArgList &Args,
                                  const Action::OffloadKind OK)
-    : ROCMToolChain(D, Triple, Args), HostTC(HostTC), OK(OK) {
+    : ROCMToolChain(D, Triple, Args), HostTC(HostTC), SYCLInstallation(D),
+      OK(OK) {
   // Lookup binaries into the driver directory, this is used to
   // discover the clang-offload-bundler executable.
   getProgramPaths().push_back(getDriver().Dir);
@@ -318,8 +319,7 @@ void HIPAMDToolChain::addClangTargetOptions(
     CC1Args.push_back("-fembed-bitcode=marker");
 
   if (DeviceOffloadingKind == Action::OFK_SYCL) {
-    toolchains::SYCLToolChain::AddSYCLIncludeArgs(getDriver(), DriverArgs,
-                                                  CC1Args);
+    SYCLInstallation.AddSYCLIncludeArgs(DriverArgs, CC1Args);
   }
 
   auto NoLibSpirv = DriverArgs.hasArg(options::OPT_fno_sycl_libspirv) ||