Link devicelib during device code compilation

clang/include/clang/Basic/DiagnosticDriverKinds.td

@@ -123,6 +123,9 @@ def err_drv_cuda_host_arch : Error<
 def err_drv_no_sycl_libspirv : Error<
   "cannot find '%0'; provide path to libspirv library via '-fsycl-libspirv-path', or "
   "pass '-fno-sycl-libspirv' to build without linking with libspirv">;
+def err_drv_no_sycl_libdevice : Error<
+  "cannot find '%0'"
+  "pass '-fno-sycl-libdevice' to build without linking with sycl libdevice">;
 def warn_flag_no_sycl_libspirv
     : Warning<"'-fno-sycl-libspirv' should not be used with target '%0'; "
               "libspirv is required for correct behavior">,

clang/include/clang/Driver/Options.td

@@ -6881,6 +6881,8 @@ def fsycl_libspirv_path_EQ : Joined<["-"], "fsycl-libspirv-path=">,
   HelpText<"Path to libspirv library">;
 def fno_sycl_libspirv : Flag<["-"], "fno-sycl-libspirv">,
   HelpText<"Disable check for libspirv">;
+def fno_sycl_libdevice : Flag<["-"], "fno-sycl-libdevice">,
+  HelpText<"Disable check for sycl libdevice">;
 def fsycl_use_spirv_backend_for_spirv_gen : Flag<["-"], "fsycl-use-spirv-backend-for-spirv-gen">,
   Flags<[HelpHidden]>,
   HelpText<"Use the SPIR-V backend for SPIR-V code generation. "

clang/lib/Driver/Driver.cpp

@@ -5533,7 +5533,7 @@ class OffloadingActionBuilder final {
         // AOT compilation.
         bool SYCLDeviceLibLinked = false;
         Action *NativeCPULib = nullptr;
-        if (IsSPIR || IsNVPTX || IsSYCLNativeCPU) {
+        if (IsSPIR || IsSYCLNativeCPU) {
           bool UseJitLink =
               IsSPIR &&
               Args.hasFlag(options::OPT_fsycl_device_lib_jit_link,
@@ -5848,10 +5848,9 @@ class OffloadingActionBuilder final {
             ++NumOfDeviceLibLinked;
             Arg *InputArg = MakeInputArg(Args, C.getDriver().getOpts(),
                                          Args.MakeArgString(LibName));
-            if (TC->getTriple().isNVPTX() ||
-                (TC->getTriple().isSPIR() &&
-                 TC->getTriple().getSubArch() ==
-                     llvm::Triple::SPIRSubArch_fpga)) {
+            if (TC->getTriple().isSPIR() &&
+                TC->getTriple().getSubArch() ==
+                    llvm::Triple::SPIRSubArch_fpga) {
               auto *SYCLDeviceLibsInputAction =
                   C.MakeAction<InputAction>(*InputArg, types::TY_Object);
               auto *SYCLDeviceLibsUnbundleAction =
@@ -5892,10 +5891,9 @@ class OffloadingActionBuilder final {
         }
       }
 
-      // For NVPTX backend we need to also link libclc and CUDA libdevice
-      // at the same stage that we link all of the unbundled SYCL libdevice
-      // objects together.
-      if ((TC->getTriple().isNVPTX() || isNativeCPU) && NumOfDeviceLibLinked) {
+      // For Natice CPU backend we need to also link libclc at the same stage
+      // that we link all of the unbundled SYCL libdevice objects together.
+      if (isNativeCPU && NumOfDeviceLibLinked) {
         std::string LibSpirvFile;
         if (Args.hasArg(options::OPT_fsycl_libspirv_path_EQ)) {
           auto ProvidedPath =
@@ -5950,23 +5948,6 @@ class OffloadingActionBuilder final {
           // return here to not generate cuda actions
           return NumOfDeviceLibLinked != 0;
         }
-
-        const toolchains::CudaToolChain *CudaTC =
-            static_cast<const toolchains::CudaToolChain *>(TC);
-        for (const auto &LinkInputEnum : enumerate(DeviceLinkerInputs)) {
-          const char *BoundArch =
-              SYCLTargetInfoList[LinkInputEnum.index()].BoundArch;
-          std::string LibDeviceFile =
-              CudaTC->CudaInstallation.getLibDeviceFile(BoundArch);
-          if (!LibDeviceFile.empty()) {
-            Arg *CudaDeviceLibInputArg =
-                MakeInputArg(Args, C.getDriver().getOpts(),
-                             Args.MakeArgString(LibDeviceFile));
-            auto *SYCLDeviceLibInputAction = C.MakeAction<InputAction>(
-                *CudaDeviceLibInputArg, types::TY_LLVM_BC);
-            DeviceLinkObjects.push_back(SYCLDeviceLibInputAction);
-          }
-        }
       }
       return NumOfDeviceLibLinked != 0;
     }

clang/lib/Driver/ToolChains/Cuda.cpp

@@ -875,6 +875,9 @@ static const char *getLibSpirvTargetName(const ToolChain &HostTC) {
   return "remangled-l64-signed_char.libspirv-nvptx64-nvidia-cuda.bc";
 }
 
+// Select DeviceLib file.
+static const char *getDeviceLibTargetName() { return "devicelib--cuda.bc"; }
+
 void NVPTXToolChain::addClangTargetOptions(
     const llvm::opt::ArgList &DriverArgs, llvm::opt::ArgStringList &CC1Args,
     Action::OffloadKind DeviceOffloadingKind) const {
@@ -1033,6 +1036,41 @@ void CudaToolChain::addClangTargetOptions(
     CC1Args.push_back(DriverArgs.MakeArgString(LibSpirvFile));
   }
 
+  auto NoDeviceLib = DriverArgs.hasArg(options::OPT_fno_sycl_libdevice) ||
+                     getDriver().offloadDeviceOnly();
+
+  if (DeviceOffloadingKind == Action::OFK_SYCL && !NoDeviceLib) {
+    std::string DeviceLibFile;
+
+    SmallVector<StringRef, 8> LibraryPaths;
+
+    // Expected path w/out install.
+    SmallString<256> WithoutInstallPath(getDriver().ResourceDir);
+    llvm::sys::path::append(WithoutInstallPath, Twine("../../"));
+    LibraryPaths.emplace_back(WithoutInstallPath.c_str());
+
+    // Select remangled libclc variant
+    std::string DeviceLibTargetName = getDeviceLibTargetName();
+
+    for (StringRef LibraryPath : LibraryPaths) {
+      SmallString<128> DeviceLibTargetFile(LibraryPath);
+      llvm::sys::path::append(DeviceLibTargetFile, DeviceLibTargetName);
+      if (llvm::sys::fs::exists(DeviceLibTargetFile) ||
+          DriverArgs.hasArg(options::OPT__HASH_HASH_HASH)) {
+        DeviceLibFile = std::string(DeviceLibTargetFile.str());
+        break;
+      }
+    }
+
+    if (DeviceLibFile.empty()) {
+      getDriver().Diag(diag::err_drv_no_sycl_libdevice)
+          << getDeviceLibTargetName();
+      return;
+    }
+
+    CC1Args.push_back("-mlink-builtin-bitcode");
+    CC1Args.push_back(DriverArgs.MakeArgString(DeviceLibFile));
+  }
   if (DriverArgs.hasArg(options::OPT_nogpulib))
     return;
 

clang/lib/Driver/ToolChains/SYCL.cpp

@@ -212,6 +212,10 @@ SYCL::getDeviceLibraries(const Compilation &C, const llvm::Triple &TargetTriple,
   SmallVector<std::string, 8> LibraryList;
   const llvm::opt::ArgList &Args = C.getArgs();
 
+  // For NVPTX we only use one single bitcode library and ignore
+  // manually specified SYCL device libraries.
+  bool IgnoreSingleLibs = TargetTriple.isNVPTX();
+
   struct DeviceLibOptInfo {
     StringRef DeviceLibName;
     StringRef DeviceLibOption;
@@ -233,10 +237,13 @@ SYCL::getDeviceLibraries(const Compilation &C, const llvm::Triple &TargetTriple,
       if (A->getOption().matches(options::OPT_fno_sycl_device_lib_EQ))
         NoDeviceLibs = true;
 
+      bool PrintUnusedLibWarning = false;
       for (StringRef Val : A->getValues()) {
         if (Val == "all") {
           for (const auto &K : DeviceLibLinkInfo.keys())
-            DeviceLibLinkInfo[K] = true && (!NoDeviceLibs || K == "internal");
+            DeviceLibLinkInfo[K] = (!IgnoreSingleLibs && !NoDeviceLibs) ||
+                                   (K == "internal" && NoDeviceLibs);
+          PrintUnusedLibWarning = false;
           break;
         }
         auto LinkInfoIter = DeviceLibLinkInfo.find(Val);
@@ -247,10 +254,18 @@ SYCL::getDeviceLibraries(const Compilation &C, const llvm::Triple &TargetTriple,
           C.getDriver().Diag(diag::err_drv_unsupported_option_argument)
               << A->getSpelling() << Val;
         }
-        DeviceLibLinkInfo[Val] = true && !NoDeviceLibs;
+        DeviceLibLinkInfo[Val] = !NoDeviceLibs && !IgnoreSingleLibs;
+        PrintUnusedLibWarning = IgnoreSingleLibs && !NoDeviceLibs;
       }
+      if (PrintUnusedLibWarning)
+        C.getDriver().Diag(diag::warn_ignored_clang_option)
+            << A->getSpelling() << A->getAsString(Args);
     }
   }
+
+  if (IgnoreSingleLibs && !NoDeviceLibs)
+    return LibraryList;
+
   using SYCLDeviceLibsList = SmallVector<DeviceLibOptInfo, 5>;
 
   const SYCLDeviceLibsList SYCLDeviceWrapperLibs = {
@@ -304,10 +319,9 @@ SYCL::getDeviceLibraries(const Compilation &C, const llvm::Triple &TargetTriple,
       C.getDefaultToolChain().getTriple().isWindowsMSVCEnvironment();
   bool IsNewOffload = C.getDriver().getUseNewOffloadingDriver();
   StringRef LibSuffix = ".bc";
-  if (TargetTriple.isNVPTX() ||
-      (TargetTriple.isSPIR() &&
-       TargetTriple.getSubArch() == llvm::Triple::SPIRSubArch_fpga))
-    // For NVidia or FPGA, we are unbundling objects.
+  if (TargetTriple.isSPIR() &&
+      TargetTriple.getSubArch() == llvm::Triple::SPIRSubArch_fpga)
+    // For FPGA, we are unbundling objects.
     LibSuffix = IsWindowsMSVCEnv ? ".obj" : ".o";
   if (IsNewOffload)
     // For new offload model, we use packaged .bc files.
@@ -323,7 +337,7 @@ SYCL::getDeviceLibraries(const Compilation &C, const llvm::Triple &TargetTriple,
   };
 
   addLibraries(SYCLDeviceWrapperLibs);
-  if (IsSpirvAOT || TargetTriple.isNVPTX())
+  if (IsSpirvAOT)
     addLibraries(SYCLDeviceFallbackLibs);
 
   bool NativeBfloatLibs;
@@ -551,7 +565,7 @@ const char *SYCL::Linker::constructLLVMLinkCommand(
                           this->getToolChain().getTriple().getSubArch() ==
                               llvm::Triple::SPIRSubArch_fpga;
       StringRef LibPostfix = ".bc";
-      if (IsNVPTX || IsFPGA) {
+      if (IsFPGA) {
         LibPostfix = ".o";
         if (HostTC->getTriple().isWindowsMSVCEnvironment() &&
             C.getDriver().IsCLMode())

clang/test/CodeGenSYCL/sycl-libdevice-cmath.cpp

@@ -5,7 +5,7 @@
 // intrinsics. This allows the driver to link in the libdevice definitions for
 // cosf etc. later in the driver flow.
 
-// RUN: %clang_cc1 %s -fsycl-is-device -triple nvptx64-nvidia-cuda -emit-llvm -o - | FileCheck %s 
+// RUN: %clang_cc1 %s -fsycl-is-device -triple nvptx64-nvidia-cuda -emit-llvm -o - | FileCheck %s
 // RUN: %clang_cc1 %s -fsycl-is-device -triple nvptx64-nvidia-cuda -ffast-math -emit-llvm -o - | FileCheck %s
 
 #include "Inputs/sycl.hpp"

clang/test/Driver/sycl-offload-nvptx.cpp

@@ -53,28 +53,22 @@
 // CHK-PHASES-NO-CC: 7: backend, {6}, assembler, (host-sycl)
 // CHK-PHASES-NO-CC: 8: assembler, {7}, object, (host-sycl)
 // CHK-PHASES-NO-CC: 9: linker, {4}, ir, (device-sycl, sm_50)
-// CHK-PHASES-NO-CC: 10: input, "{{.*}}libsycl-itt-user-wrappers.o{{.*}}", object
-// CHK-PHASES-NO-CC: 11: clang-offload-unbundler, {10}, object
-// CHK-PHASES-NO-CC: 12: offload, " (nvptx64-nvidia-cuda)" {11}, object
-// CHK-PHASES-NO-CC: 13: input, "{{.*}}libsycl-itt-compiler-wrappers.o{{.*}}", object
-// CHK-PHASES-NO-CC: 14: clang-offload-unbundler, {13}, object
-// CHK-PHASES-NO-CC: 15: offload, " (nvptx64-nvidia-cuda)" {14}, object
-// CHK-PHASES-NO-CC: 16: input, "{{.*}}libsycl-itt-stubs.o{{.*}}", object
-// CHK-PHASES-NO-CC: 17: clang-offload-unbundler, {16}, object
-// CHK-PHASES-NO-CC: 18: offload, " (nvptx64-nvidia-cuda)" {17}, object
-// CHK-PHASES-NO-CC: 19: input, "{{.*}}nvidiacl{{.*}}", ir, (device-sycl, sm_50)
-// CHK-PHASES-NO-CC: 20: input, "{{.*}}libdevice{{.*}}", ir, (device-sycl, sm_50)
-// CHK-PHASES-NO-CC: 21: linker, {9, 12, 15, 18, 19, 20}, ir, (device-sycl, sm_50)
-// CHK-PHASES-NO-CC: 22: sycl-post-link, {21}, ir, (device-sycl, sm_50)
-// CHK-PHASES-NO-CC: 23: file-table-tform, {22}, ir, (device-sycl, sm_50)
-// CHK-PHASES-NO-CC: 24: backend, {23}, assembler, (device-sycl, sm_50)
-// CHK-PHASES-NO-CC: 25: assembler, {24}, object, (device-sycl, sm_50)
-// CHK-PHASES-NO-CC: 26: linker, {24, 25}, cuda-fatbin, (device-sycl, sm_50)
-// CHK-PHASES-NO-CC: 27: foreach, {23, 26}, cuda-fatbin, (device-sycl, sm_50)
-// CHK-PHASES-NO-CC: 28: file-table-tform, {22, 27}, tempfiletable, (device-sycl, sm_50)
-// CHK-PHASES-NO-CC: 29: clang-offload-wrapper, {28}, object, (device-sycl, sm_50)
-// CHK-PHASES-NO-CC: 30: offload, "device-sycl (nvptx64-nvidia-cuda:sm_50)" {29}, object
-// CHK-PHASES-NO-CC: 31: linker, {8, 30}, image, (host-sycl)
+// CHK-PHASES-NO-CC: 10: input, "{{.*}}libsycl-itt-user-wrappers.bc", ir, (device-sycl, sm_50)
+// CHK-PHASES-NO-CC: 11: input, "{{.*}}libsycl-itt-compiler-wrappers.bc", ir, (device-sycl, sm_50)
+// CHK-PHASES-NO-CC: 12: input, "{{.*}}libsycl-itt-stubs.bc", ir, (device-sycl, sm_50)
+// CHK-PHASES-NO-CC: 13: input, "{{.*}}nvidiacl{{.*}}", ir, (device-sycl, sm_50)
+// CHK-PHASES-NO-CC: 14: input, "{{.*}}libdevice{{.*}}", ir, (device-sycl, sm_50)
+// CHK-PHASES-NO-CC: 15: linker, {9, 10, 11, 12, 13, 14}, ir, (device-sycl, sm_50)
+// CHK-PHASES-NO-CC: 16: sycl-post-link, {15}, ir, (device-sycl, sm_50)
+// CHK-PHASES-NO-CC: 17: file-table-tform, {16}, ir, (device-sycl, sm_50)
+// CHK-PHASES-NO-CC: 18: backend, {17}, assembler, (device-sycl, sm_50)
+// CHK-PHASES-NO-CC: 19: assembler, {18}, object, (device-sycl, sm_50)
+// CHK-PHASES-NO-CC: 20: linker, {18, 19}, cuda-fatbin, (device-sycl, sm_50)
+// CHK-PHASES-NO-CC: 21: foreach, {17, 20}, cuda-fatbin, (device-sycl, sm_50)
+// CHK-PHASES-NO-CC: 22: file-table-tform, {16, 21}, tempfiletable, (device-sycl, sm_50)
+// CHK-PHASES-NO-CC: 23: clang-offload-wrapper, {22}, object, (device-sycl, sm_50)
+// CHK-PHASES-NO-CC: 24: offload, "device-sycl (nvptx64-nvidia-cuda:sm_50)" {23}, object
+// CHK-PHASES-NO-CC: 25: linker, {8, 24}, image, (host-sycl)
 //
 /// Check phases specifying a compute capability.
 // RUN: %clangxx -ccc-print-phases --sysroot=%S/Inputs/SYCL -std=c++11 \
@@ -97,28 +91,22 @@
 // CHK-PHASES: 7: backend, {6}, assembler, (host-sycl)
 // CHK-PHASES: 8: assembler, {7}, object, (host-sycl)
 // CHK-PHASES: 9: linker, {4}, ir, (device-sycl, sm_35)
-// CHK-PHASES: 10: input, "{{.*}}libsycl-itt-user-wrappers.o", object
-// CHK-PHASES: 11: clang-offload-unbundler, {10}, object
-// CHK-PHASES: 12: offload, " (nvptx64-nvidia-cuda)" {11}, object
-// CHK-PHASES: 13: input, "{{.*}}libsycl-itt-compiler-wrappers.o", object
-// CHK-PHASES: 14: clang-offload-unbundler, {13}, object
-// CHK-PHASES: 15: offload, " (nvptx64-nvidia-cuda)" {14}, object
-// CHK-PHASES: 16: input, "{{.*}}libsycl-itt-stubs.o", object
-// CHK-PHASES: 17: clang-offload-unbundler, {16}, object
-// CHK-PHASES: 18: offload, " (nvptx64-nvidia-cuda)" {17}, object
-// CHK-PHASES: 19: input, "{{.*}}nvidiacl{{.*}}", ir, (device-sycl, sm_35)
-// CHK-PHASES: 20: input, "{{.*}}libdevice{{.*}}", ir, (device-sycl, sm_35)
-// CHK-PHASES: 21: linker, {9, 12, 15, 18, 19, 20}, ir, (device-sycl, sm_35)
- // CHK-PHASES: 22: sycl-post-link, {21}, ir, (device-sycl, sm_35)
-// CHK-PHASES: 23: file-table-tform, {22}, ir, (device-sycl, sm_35)
-// CHK-PHASES: 24: backend, {23}, assembler, (device-sycl, sm_35)
-// CHK-PHASES: 25: assembler, {24}, object, (device-sycl, sm_35)
-// CHK-PHASES: 26: linker, {24, 25}, cuda-fatbin, (device-sycl, sm_35)
-// CHK-PHASES: 27: foreach, {23, 26}, cuda-fatbin, (device-sycl, sm_35)
-// CHK-PHASES: 28: file-table-tform, {22, 27}, tempfiletable, (device-sycl, sm_35)
-// CHK-PHASES: 29: clang-offload-wrapper, {28}, object, (device-sycl, sm_35)
-// CHK-PHASES: 30: offload, "device-sycl (nvptx64-nvidia-cuda:sm_35)" {29}, object
-// CHK-PHASES: 31: linker, {8, 30}, image, (host-sycl)
+// CHK-PHASES: 10: input, "{{.*}}libsycl-itt-user-wrappers.bc", ir, (device-sycl, sm_35)
+// CHK-PHASES: 11: input, "{{.*}}libsycl-itt-compiler-wrappers.bc", ir, (device-sycl, sm_35)
+// CHK-PHASES: 12: input, "{{.*}}libsycl-itt-stubs.bc", ir, (device-sycl, sm_35)
+// CHK-PHASES: 13: input, "{{.*}}nvidiacl{{.*}}", ir, (device-sycl, sm_35)
+// CHK-PHASES: 14: input, "{{.*}}libdevice{{.*}}", ir, (device-sycl, sm_35)
+// CHK-PHASES: 15: linker, {9, 10, 11, 12, 13, 14}, ir, (device-sycl, sm_35)
+// CHK-PHASES: 16: sycl-post-link, {15}, ir, (device-sycl, sm_35)
+// CHK-PHASES: 17: file-table-tform, {16}, ir, (device-sycl, sm_35)
+// CHK-PHASES: 18: backend, {17}, assembler, (device-sycl, sm_35)
+// CHK-PHASES: 19: assembler, {18}, object, (device-sycl, sm_35)
+// CHK-PHASES: 20: linker, {18, 19}, cuda-fatbin, (device-sycl, sm_35)
+// CHK-PHASES: 21: foreach, {17, 20}, cuda-fatbin, (device-sycl, sm_35)
+// CHK-PHASES: 22: file-table-tform, {16, 21}, tempfiletable, (device-sycl, sm_35)
+// CHK-PHASES: 23: clang-offload-wrapper, {22}, object, (device-sycl, sm_35)
+// CHK-PHASES: 24: offload, "device-sycl (nvptx64-nvidia-cuda:sm_35)" {23}, object
+// CHK-PHASES: 25: linker, {8, 24}, image, (host-sycl)
 
 /// Check calling preprocessor only
 // RUN: %clangxx -E -fsycl -fsycl-targets=nvptx64-nvidia-cuda -ccc-print-phases %s 2>&1 \

libclc/CMakeLists.txt

@@ -233,6 +233,7 @@ if( ENABLE_RUNTIME_SUBNORMAL )
   foreach( file subnormal_use_default subnormal_disable )
     link_bc(
        TARGET ${file}
+       RSP_DIR ${LIBCLC_ARCH_OBJFILE_DIR}
        INPUTS ${CMAKE_CURRENT_SOURCE_DIR}/generic/lib/${file}.ll
     )
     install( FILES $<TARGET_PROPERTY:${file},TARGET_FILE> ARCHIVE
@@ -405,7 +406,6 @@ foreach( t ${LIBCLC_TARGETS_TO_BUILD} )
     # Enable SPIR-V builtin function declarations, so they don't
     # have to be explicity declared in the soruce.
     list( APPEND flags -Xclang -fdeclare-spirv-builtins)
-
     set( LIBCLC_ARCH_OBJFILE_DIR "${LIBCLC_OBJFILE_DIR}/${arch_suffix}" )
     file( MAKE_DIRECTORY ${LIBCLC_ARCH_OBJFILE_DIR} )