[SYCL][NATIVECPU] Add device library and initial subgroup support

clang/lib/Driver/Driver.cpp

@@ -5524,18 +5524,17 @@ class OffloadingActionBuilder final {
         // device libraries are only needed when current toolchain is using
         // AOT compilation.
         bool SYCLDeviceLibLinked = false;
-        if (IsSPIR || IsNVPTX) {
+        Action *NativeCPULib = nullptr;
+        if (IsSPIR || IsNVPTX || IsSYCLNativeCPU) {
           bool UseJitLink =
               IsSPIR &&
               Args.hasFlag(options::OPT_fsycl_device_lib_jit_link,
                            options::OPT_fno_sycl_device_lib_jit_link, false);
           bool UseAOTLink = IsSPIR && (IsSpirvAOT || !UseJitLink);
           SYCLDeviceLibLinked = addSYCLDeviceLibs(
               TC, SYCLDeviceLibs, UseAOTLink,
-              C.getDefaultToolChain().getTriple().isWindowsMSVCEnvironment());
-        }
-        if (IsSYCLNativeCPU) {
-          SYCLDeviceLibLinked |= addSYCLNativeCPULibs(TC, SYCLDeviceLibs);
+              C.getDefaultToolChain().getTriple().isWindowsMSVCEnvironment(),
+              IsSYCLNativeCPU, NativeCPULib);
         }
         JobAction *LinkSYCLLibs =
             C.MakeAction<LinkJobAction>(SYCLDeviceLibs, types::TY_LLVM_BC);
@@ -5618,6 +5617,15 @@ class OffloadingActionBuilder final {
           };
           Action *PostLinkAction = createPostLinkAction();
           if (IsSYCLNativeCPU) {
+            if (NativeCPULib) {
+              // The native cpu device lib is linked without --only-needed
+              // as it contains builtins not referenced in source code but
+              // needed by the native cpu backend.
+              clang::driver::ActionList AllLibs = {FullDeviceLinkAction,
+                                                   NativeCPULib};
+              FullDeviceLinkAction =
+                  C.MakeAction<LinkJobAction>(AllLibs, types::TY_LLVM_BC);
+            }
             // for SYCL Native CPU, we just take the linked device
             // modules, lower them to an object file , and link it to the host
             // object file.
@@ -5802,60 +5810,9 @@ class OffloadingActionBuilder final {
       }
     }
 
-    bool addSYCLNativeCPULibs(const ToolChain *TC,
-                              ActionList &DeviceLinkObjects) {
-      std::string LibSpirvFile;
-      if (Args.hasArg(options::OPT_fsycl_libspirv_path_EQ)) {
-        auto ProvidedPath =
-            Args.getLastArgValue(options::OPT_fsycl_libspirv_path_EQ).str();
-        if (llvm::sys::fs::exists(ProvidedPath))
-          LibSpirvFile = ProvidedPath;
-      } else {
-        SmallVector<StringRef, 8> LibraryPaths;
-
-        // Expected path w/out install.
-        SmallString<256> WithoutInstallPath(C.getDriver().ResourceDir);
-        llvm::sys::path::append(WithoutInstallPath, Twine("../../clc"));
-        LibraryPaths.emplace_back(WithoutInstallPath.c_str());
-
-        // Expected path w/ install.
-        SmallString<256> WithInstallPath(C.getDriver().ResourceDir);
-        llvm::sys::path::append(WithInstallPath, Twine("../../../share/clc"));
-        LibraryPaths.emplace_back(WithInstallPath.c_str());
-
-        // Select libclc variant based on target triple.
-        // On Windows long is 32 bits, so we have to select the right remangled
-        // libclc version.
-        std::string LibSpirvTargetName =
-            (TC->getAuxTriple()->isOSWindows())
-                ? "remangled-l32-signed_char.libspirv-"
-                : "remangled-l64-signed_char.libspirv-";
-        LibSpirvTargetName.append(TC->getTripleString() + ".bc");
-
-        for (StringRef LibraryPath : LibraryPaths) {
-          SmallString<128> LibSpirvTargetFile(LibraryPath);
-          llvm::sys::path::append(LibSpirvTargetFile, LibSpirvTargetName);
-          if (llvm::sys::fs::exists(LibSpirvTargetFile) ||
-              Args.hasArg(options::OPT__HASH_HASH_HASH)) {
-            LibSpirvFile = std::string(LibSpirvTargetFile.str());
-            break;
-          }
-        }
-      }
-
-      if (!LibSpirvFile.empty()) {
-        Arg *LibClcInputArg = MakeInputArg(Args, C.getDriver().getOpts(),
-                                           Args.MakeArgString(LibSpirvFile));
-        auto *SYCLLibClcInputAction =
-            C.MakeAction<InputAction>(*LibClcInputArg, types::TY_LLVM_BC);
-        DeviceLinkObjects.push_back(SYCLLibClcInputAction);
-        return true;
-      }
-      return false;
-    }
-
     bool addSYCLDeviceLibs(const ToolChain *TC, ActionList &DeviceLinkObjects,
-                           bool isSpirvAOT, bool isMSVCEnv) {
+                           bool isSpirvAOT, bool isMSVCEnv, bool isNativeCPU,
+                           Action *&NativeCPULib) {
       int NumOfDeviceLibLinked = 0;
       SmallVector<SmallString<128>, 4> LibLocCandidates;
       SYCLInstallation.getSYCLDeviceLibPath(LibLocCandidates);
@@ -5872,6 +5829,14 @@ class OffloadingActionBuilder final {
           SmallString<128> LibName(LLCandidate);
           llvm::sys::path::append(LibName, DeviceLib);
           if (llvm::sys::fs::exists(LibName)) {
+
+            // NativeCPU currently only needs libsycl-nativecpu_utils and
+            // libclc, so temporarily skip other device libs in invocation.
+            // Todo: remove once NativeCPU tests the other libraries.
+            if (isNativeCPU &&
+                !LibName.str().contains("libsycl-nativecpu_utils"))
+              continue;
+
             ++NumOfDeviceLibLinked;
             Arg *InputArg = MakeInputArg(Args, C.getDriver().getOpts(),
                                          Args.MakeArgString(LibName));
@@ -5905,14 +5870,24 @@ class OffloadingActionBuilder final {
             }
             if (!LibLocSelected)
               LibLocSelected = !LibLocSelected;
+
+            // The device link stage may remove symbols not referenced in the
+            // source code. Since libsycl-nativecpu_utils contains such symbols
+            // which are later needed by the NativeCPU backend passes we link
+            // that library separately afterwards without --only-needed.
+            if (isNativeCPU) {
+              assert(!NativeCPULib);
+              NativeCPULib = DeviceLinkObjects.back();
+              DeviceLinkObjects.pop_back();
+            }
           }
         }
       }
 
       // 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() && NumOfDeviceLibLinked) {
+      if ((TC->getTriple().isNVPTX() || isNativeCPU) && NumOfDeviceLibLinked) {
         std::string LibSpirvFile;
         if (Args.hasArg(options::OPT_fsycl_libspirv_path_EQ)) {
           auto ProvidedPath =
@@ -5932,13 +5907,18 @@ class OffloadingActionBuilder final {
           llvm::sys::path::append(WithInstallPath, Twine("../../../share/clc"));
           LibraryPaths.emplace_back(WithInstallPath.c_str());
 
+          // TODO: check if the isNVPTX() path can also use
+          // TC->getTripleString() so that the conditional could be removed
+          const std::string TrStr =
+              isNativeCPU ? TC->getTripleString() : "nvptx64-nvidia-cuda";
+
           // Select remangled libclc variant
-          std::string LibSpirvTargetName =
-              (TC->getAuxTriple()->isOSWindows())
-                  ? "remangled-l32-signed_char.libspirv-nvptx64-nvidia-cuda."
-                    "bc"
-                  : "remangled-l64-signed_char.libspirv-nvptx64-nvidia-cuda."
-                    "bc";
+          StringRef LibSpirvTargetNamePref =
+              TC->getAuxTriple()->isOSWindows()
+                  ? "remangled-l32-signed_char.libspirv-"
+                  : "remangled-l64-signed_char.libspirv-";
+          llvm::Twine LibSpirvTargetNameTemp = LibSpirvTargetNamePref + TrStr;
+          llvm::Twine LibSpirvTargetName = LibSpirvTargetNameTemp + ".bc";
 
           for (StringRef LibraryPath : LibraryPaths) {
             SmallString<128> LibSpirvTargetFile(LibraryPath);
@@ -5950,7 +5930,6 @@ class OffloadingActionBuilder final {
             }
           }
         }
-
         if (!LibSpirvFile.empty()) {
           Arg *LibClcInputArg = MakeInputArg(Args, C.getDriver().getOpts(),
                                              Args.MakeArgString(LibSpirvFile));
@@ -5959,6 +5938,11 @@ class OffloadingActionBuilder final {
           DeviceLinkObjects.push_back(SYCLLibClcInputAction);
         }
 
+        if (isNativeCPU) {
+          // 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)) {
@@ -9233,7 +9217,10 @@ InputInfoList Driver::BuildJobsForActionNoCache(
       Action::OffloadKind DependentOffloadKind;
       if (UI.DependentOffloadKind == Action::OFK_SYCL &&
           TargetDeviceOffloadKind == Action::OFK_None &&
-          !(isSYCLNativeCPU(Args) &&  isSYCLNativeCPU(C.getDefaultToolChain().getTriple(), TC->getTriple())))
+          !(isSYCLNativeCPU(Args) &&
+            isSYCLNativeCPU(C.getDefaultToolChain().getTriple(),
+                            TC->getTriple()) &&
+            UA->getDependentActionsInfo().size() > 1))
         DependentOffloadKind = Action::OFK_Host;
       else
         DependentOffloadKind = UI.DependentOffloadKind;

clang/lib/Driver/ToolChains/Clang.cpp

@@ -5605,10 +5605,6 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
       // Let the FE know we are doing a SYCL offload compilation, but we are
       // doing the host pass.
       CmdArgs.push_back("-fsycl-is-host");
-      if (IsSYCLNativeCPU) {
-        CmdArgs.push_back("-D");
-        CmdArgs.push_back("__SYCL_NATIVE_CPU__");
-      }
 
       if (!D.IsCLMode()) {
         // SYCL library is guaranteed to work correctly only with dynamic

clang/lib/Driver/ToolChains/SYCL.cpp

@@ -290,6 +290,15 @@ SYCL::getDeviceLibraries(const Compilation &C, const llvm::Triple &TargetTriple,
   const SYCLDeviceLibsList SYCLDeviceSanitizerLibs = {
       {"libsycl-sanitizer", "internal"}};
 #endif
+
+  const SYCLDeviceLibsList SYCLNativeCpuDeviceLibs = {
+      {"libsycl-nativecpu_utils", "internal"}};
+
+  const bool isNativeCPU =
+      (driver::isSYCLNativeCPU(Args) &&
+       driver::isSYCLNativeCPU(C.getDefaultToolChain().getTriple(),
+                               TargetTriple));
+
   bool IsWindowsMSVCEnv =
       C.getDefaultToolChain().getTriple().isWindowsMSVCEnvironment();
   bool IsNewOffload = C.getDriver().getUseNewOffloadingDriver();
@@ -368,6 +377,10 @@ SYCL::getDeviceLibraries(const Compilation &C, const llvm::Triple &TargetTriple,
       addLibraries(SYCLDeviceSanitizerLibs);
   }
 #endif
+
+  if (isNativeCPU)
+    addLibraries(SYCLNativeCpuDeviceLibs);
+
   return LibraryList;
 }
 

clang/test/Driver/sycl-native-cpu-fsycl.cpp

@@ -1,3 +1,4 @@
+//UNSUPPORTED: (system-windows && !native_cpu)
 //RUN: %clang -fsycl -fsycl-targets=native_cpu -fsycl-libspirv-path=%S/Inputs/SYCL/libspirv.bc -ccc-print-phases %s 2>&1 | FileCheck %s --check-prefix=CHECK_ACTIONS
 //RUN: %clang -fsycl -fsycl-targets=native_cpu -fsycl-libspirv-path=%S/Inputs/SYCL/libspirv.bc -ccc-print-bindings %s 2>&1 | FileCheck %s --check-prefix=CHECK_BINDINGS
 //RUN: %clang -fsycl -fsycl-targets=native_cpu -fsycl-libspirv-path=%S/Inputs/SYCL/libspirv.bc -### %s 2>&1 | FileCheck %s --check-prefix=CHECK_INVO
@@ -19,24 +20,29 @@
 //CHECK_ACTIONS:         +- 7: compiler, {6}, ir, (host-sycl)
 //CHECK_ACTIONS:      +- 8: backend, {7}, assembler, (host-sycl)
 //CHECK_ACTIONS:   +- 9: assembler, {8}, object, (host-sycl)
-
-//CHECK_ACTIONS:|           +- 10: linker, {5}, ir, (device-sycl)
-//CHECK_ACTIONS:|           |- 11: input, "{{.*}}libspirv{{.*}}", ir, (device-sycl)
-//CHECK_ACTIONS:|        +- 12: linker, {10, 11}, ir, (device-sycl)
-//CHECK_ACTIONS:|     +- 13: backend, {12}, assembler, (device-sycl)
-//CHECK_ACTIONS:|  +- 14: assembler, {13}, object, (device-sycl)
-//CHECK_ACTIONS:|- 15: offload, "device-sycl ({{.*}})" {14}, object
-//CHECK_ACTIONS:|     +- 16: sycl-post-link, {12}, tempfiletable, (device-sycl)
-//CHECK_ACTIONS:|  +- 17: clang-offload-wrapper, {16}, object, (device-sycl)
-//CHECK_ACTIONS:|- 18: offload, "device-sycl ({{.*}})" {17}, object
-//CHECK_ACTIONS:19: linker, {9, 15, 18}, image, (host-sycl)
+//CHECK_ACTIONS:        +- 10: linker, {5}, ir, (device-sycl)
+//CHECK_ACTIONS:        |- [[SPIRVLIB:.*]]: input, "{{.*}}libspirv{{.*}}", ir, (device-sycl)
+//different libraries may be linked on different platforms, so just check the common stages
+//CHECK_ACTIONS:      +- [[LINKALL:.*]]: linker, {10, [[SPIRVLIB]]}, ir, (device-sycl)
+//CHECK_ACTIONS:      |- [[NCPUINP:.*]]: input, "{{.*}}nativecpu{{.*}}", ir, (device-sycl)
+//CHECK_ACTIONS:    +- [[NCPULINK:.*]]: linker, {[[LINKALL]], [[NCPUINP]]}, ir, (device-sycl)
+//this is where we compile the device code to a shared lib, and we link the host shared lib and the device shared lib
+//CHECK_ACTIONS:|  +- [[VAL81:.*]]: backend, {[[NCPULINK]]}, assembler, (device-sycl)
+//CHECK_ACTIONS:| +- [[VAL82:.*]]: assembler, {[[VAL81]]}, object, (device-sycl)
+//CHECK_ACTIONS:|- [[VAL822:.*]]: offload, "device-sycl ({{.*}})" {[[VAL82]]}, object
+//call sycl-post-link and clang-offload-wrapper
+//CHECK_ACTIONS:|  +- [[VAL83:.*]]: sycl-post-link, {[[LINKALL]]}, tempfiletable, (device-sycl)
+//CHECK_ACTIONS:| +- [[VAL84:.*]]: clang-offload-wrapper, {[[VAL83]]}, object, (device-sycl)
+//CHECK_ACTIONS:|- [[VAL85:.*]]: offload, "device-sycl ({{.*}})" {[[VAL84]]}, object
+//CHECK_ACTIONS:[[VAL86:.*]]: linker, {9, [[VAL822]], [[VAL85]]}, image, (host-sycl)
 
 //CHECK_BINDINGS:# "{{.*}}" - "clang", inputs: ["{{.*}}sycl-native-cpu-fsycl.cpp"], output: "[[KERNELIR:.*]].bc"
 //CHECK_BINDINGS:# "{{.*}}" - "Append Footer to source", inputs: ["{{.*}}sycl-native-cpu-fsycl.cpp"], output: "[[SRCWFOOTER:.*]].cpp"
 //CHECK_BINDINGS:# "{{.*}}" - "clang", inputs: ["[[SRCWFOOTER]].cpp", "[[KERNELIR]].bc"], output: "[[HOSTOBJ:.*]].o"
 //CHECK_BINDINGS:# "{{.*}}" - "SYCL::Linker", inputs: ["[[KERNELIR]].bc"], output: "[[KERNELLINK:.*]].bc"
 //CHECK_BINDINGS:# "{{.*}}" - "SYCL::Linker", inputs: ["[[KERNELLINK]].bc", "{{.*}}.bc"], output: "[[KERNELLINKWLIB:.*]].bc"
-//CHECK_BINDINGS:# "{{.*}}" - "clang", inputs: ["[[KERNELLINKWLIB]].bc"], output: "[[KERNELOBJ:.*]].o"
+//CHECK_BINDINGS:# "{{.*}}" - "SYCL::Linker", inputs: ["[[KERNELLINKWLIB]].bc", "[[UNBUNDLEDNCPU:.*]].bc"], output: "[[KERNELLINKWLIB12:.*]].bc"
+//CHECK_BINDINGS:# "{{.*}}" - "clang", inputs: ["[[KERNELLINKWLIB12]].bc"], output: "[[KERNELOBJ:.*]].o"
 //CHECK_BINDINGS:# "{{.*}}" - "SYCL post link", inputs: ["[[KERNELLINKWLIB]].bc"], output: "[[TABLEFILE:.*]].table"
 //CHECK_BINDINGS:# "{{.*}}" - "offload wrapper", inputs: ["[[TABLEFILE]].table"], output: "[[WRAPPEROBJ:.*]].o"
 //CHECK_BINDINGS:# "{{.*}}" - "{{.*}}::Linker", inputs: ["[[HOSTOBJ]].o", "[[KERNELOBJ]].o", "[[WRAPPEROBJ]].o"], output: "a.{{.*}}"
@@ -49,8 +55,8 @@
 
 // checks that the device and host triple is correct in the generated actions when it is set explicitly
 //CHECK_ACTIONS-AARCH64:            +- 6: offload, "host-sycl (aarch64-unknown-linux-gnu)" {2}, "device-sycl (aarch64-unknown-linux-gnu)" {5}, c++-cpp-output
-//CHECK_ACTIONS-AARCH64:|- 15: offload, "device-sycl (aarch64-unknown-linux-gnu)" {14}, object
-//CHECK_ACTIONS-AARCH64:|- 18: offload, "device-sycl (aarch64-unknown-linux-gnu)" {17}, object
+//CHECK_ACTIONS-AARCH64:|- 17: offload, "device-sycl (aarch64-unknown-linux-gnu)" {16}, object
+//CHECK_ACTIONS-AARCH64:|- 20: offload, "device-sycl (aarch64-unknown-linux-gnu)" {19}, object
 
 // checks that bindings are correct when linking together multiple TUs on native cpu
 //CHECK_BINDINGS_MULTI_TU:# "{{.*}}" - "offload bundler", inputs: ["{{.*}}.o"], outputs: ["[[FILE1HOST:.*]].o", "[[FILE1DEV:.*]].o"] 
@@ -59,8 +65,7 @@
 //CHECK_BINDINGS_MULTI_TU:# "{{.*}}" - "Convert SPIR-V to LLVM-IR if needed", inputs: ["[[FILE2DEV]].o"], output: "[[FILE2SPV:.*]].bc"
 //CHECK_BINDINGS_MULTI_TU:# "{{.*}}" - "SYCL::Linker", inputs: ["[[FILE1SPV]].bc", "[[FILE2SPV]].bc"], output: "[[LINK1:.*]].bc"
 //CHECK_BINDINGS_MULTI_TU:# "{{.*}}" - "SYCL::Linker", inputs: ["[[LINK1]].bc", "{{.*}}.bc"], output: "[[LINK2:.*]].bc"
-//CHECK_BINDINGS_MULTI_TU:# "{{.*}}" - "clang", inputs: ["[[LINK2]].bc"], output: "[[KERNELO:.*]].o"
+//CHECK_BINDINGS_MULTI_TU:# "{{.*}}" - "clang", inputs: ["{{.*}}.bc"], output: "[[KERNELO:.*]].o"
 //CHECK_BINDINGS_MULTI_TU:# "{{.*}}" - "SYCL post link", inputs: ["[[LINK2]].bc"], output: "[[POSTL:.*]].table"
 //CHECK_BINDINGS_MULTI_TU:# "{{.*}}" - "offload wrapper", inputs: ["[[POSTL]].table"], output: "[[WRAP:.*]].o"
 //CHECK_BINDINGS_MULTI_TU:# "{{.*}}" - "{{.*}}::Linker", inputs: ["[[FILE1HOST]].o", "[[FILE2HOST]].o", "[[KERNELO]].o", "[[WRAP]].o"], output: "{{.*}}"
-

libdevice/cmake/modules/SYCLLibdevice.cmake

@@ -167,6 +167,16 @@ if (NOT MSVC)
     sycl-compiler)
 endif()
 
+if("native_cpu" IN_LIST SYCL_ENABLE_PLUGINS)
+  if (NOT DEFINED NATIVE_CPU_DIR)
+    message( FATAL_ERROR "Undefined UR variable NATIVE_CPU_DIR. The name may have changed." )
+  endif()
+  # Include NativeCPU UR adapter path to enable finding header file with state struct.
+  # libsycl-nativecpu_utils is only needed as BC file by NativeCPU.
+  # Todo: add versions for other targets (for cross-compilation)
+  add_devicelib_bc(libsycl-nativecpu_utils SRC nativecpu_utils.cpp DEP ${itt_obj_deps} EXTRA_ARGS -I ${NATIVE_CPU_DIR} -fsycl-targets=native_cpu)
+endif()
+
 add_devicelib(libsycl-itt-stubs SRC itt_stubs.cpp DEP ${itt_obj_deps})
 add_devicelib(libsycl-itt-compiler-wrappers SRC itt_compiler_wrappers.cpp DEP ${itt_obj_deps})
 add_devicelib(libsycl-itt-user-wrappers SRC itt_user_wrappers.cpp DEP ${itt_obj_deps})